LLVM  6.0.0svn
SimplifyCFG.cpp
Go to the documentation of this file.
1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // Peephole optimize the CFG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SetOperations.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringRef.h"
32 #include "llvm/IR/Attributes.h"
33 #include "llvm/IR/BasicBlock.h"
34 #include "llvm/IR/CFG.h"
35 #include "llvm/IR/CallSite.h"
36 #include "llvm/IR/Constant.h"
37 #include "llvm/IR/ConstantRange.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalValue.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/IRBuilder.h"
45 #include "llvm/IR/InstrTypes.h"
46 #include "llvm/IR/Instruction.h"
47 #include "llvm/IR/Instructions.h"
48 #include "llvm/IR/IntrinsicInst.h"
49 #include "llvm/IR/Intrinsics.h"
50 #include "llvm/IR/LLVMContext.h"
51 #include "llvm/IR/MDBuilder.h"
52 #include "llvm/IR/Metadata.h"
53 #include "llvm/IR/Module.h"
54 #include "llvm/IR/NoFolder.h"
55 #include "llvm/IR/Operator.h"
56 #include "llvm/IR/PatternMatch.h"
57 #include "llvm/IR/Type.h"
58 #include "llvm/IR/Use.h"
59 #include "llvm/IR/User.h"
60 #include "llvm/IR/Value.h"
61 #include "llvm/Support/Casting.h"
63 #include "llvm/Support/Debug.h"
65 #include "llvm/Support/KnownBits.h"
71 #include <algorithm>
72 #include <cassert>
73 #include <climits>
74 #include <cstddef>
75 #include <cstdint>
76 #include <iterator>
77 #include <map>
78 #include <set>
79 #include <tuple>
80 #include <utility>
81 #include <vector>
82 
83 using namespace llvm;
84 using namespace PatternMatch;
85 
86 #define DEBUG_TYPE "simplifycfg"
87 
88 // Chosen as 2 so as to be cheap, but still to have enough power to fold
89 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
90 // To catch this, we need to fold a compare and a select, hence '2' being the
91 // minimum reasonable default.
93  "phi-node-folding-threshold", cl::Hidden, cl::init(2),
94  cl::desc(
95  "Control the amount of phi node folding to perform (default = 2)"));
96 
97 static cl::opt<bool> DupRet(
98  "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
99  cl::desc("Duplicate return instructions into unconditional branches"));
100 
101 static cl::opt<bool>
102  SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
103  cl::desc("Sink common instructions down to the end block"));
104 
106  "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
107  cl::desc("Hoist conditional stores if an unconditional store precedes"));
108 
110  "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
111  cl::desc("Hoist conditional stores even if an unconditional store does not "
112  "precede - hoist multiple conditional stores into a single "
113  "predicated store"));
114 
116  "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
117  cl::desc("When merging conditional stores, do so even if the resultant "
118  "basic blocks are unlikely to be if-converted as a result"));
119 
121  "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
122  cl::desc("Allow exactly one expensive instruction to be speculatively "
123  "executed"));
124 
126  "max-speculation-depth", cl::Hidden, cl::init(10),
127  cl::desc("Limit maximum recursion depth when calculating costs of "
128  "speculatively executed instructions"));
129 
130 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
131 STATISTIC(NumLinearMaps,
132  "Number of switch instructions turned into linear mapping");
133 STATISTIC(NumLookupTables,
134  "Number of switch instructions turned into lookup tables");
135 STATISTIC(
136  NumLookupTablesHoles,
137  "Number of switch instructions turned into lookup tables (holes checked)");
138 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
139 STATISTIC(NumSinkCommons,
140  "Number of common instructions sunk down to the end block");
141 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
142 
143 namespace {
144 
145 // The first field contains the value that the switch produces when a certain
146 // case group is selected, and the second field is a vector containing the
147 // cases composing the case group.
148 using SwitchCaseResultVectorTy =
150 
151 // The first field contains the phi node that generates a result of the switch
152 // and the second field contains the value generated for a certain case in the
153 // switch for that PHI.
154 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
155 
156 /// ValueEqualityComparisonCase - Represents a case of a switch.
157 struct ValueEqualityComparisonCase {
159  BasicBlock *Dest;
160 
161  ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
162  : Value(Value), Dest(Dest) {}
163 
164  bool operator<(ValueEqualityComparisonCase RHS) const {
165  // Comparing pointers is ok as we only rely on the order for uniquing.
166  return Value < RHS.Value;
167  }
168 
169  bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
170 };
171 
172 class SimplifyCFGOpt {
173  const TargetTransformInfo &TTI;
174  const DataLayout &DL;
175  SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
176  const SimplifyCFGOptions &Options;
177 
178  Value *isValueEqualityComparison(TerminatorInst *TI);
179  BasicBlock *GetValueEqualityComparisonCases(
180  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
181  bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
182  BasicBlock *Pred,
183  IRBuilder<> &Builder);
184  bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
185  IRBuilder<> &Builder);
186 
187  bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
188  bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
189  bool SimplifySingleResume(ResumeInst *RI);
190  bool SimplifyCommonResume(ResumeInst *RI);
191  bool SimplifyCleanupReturn(CleanupReturnInst *RI);
192  bool SimplifyUnreachable(UnreachableInst *UI);
193  bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
194  bool SimplifyIndirectBr(IndirectBrInst *IBI);
195  bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
196  bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
197 
198 public:
199  SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
200  SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
201  const SimplifyCFGOptions &Opts)
202  : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
203 
204  bool run(BasicBlock *BB);
205 };
206 
207 } // end anonymous namespace
208 
209 /// Return true if it is safe to merge these two
210 /// terminator instructions together.
211 static bool
213  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
214  if (SI1 == SI2)
215  return false; // Can't merge with self!
216 
217  // It is not safe to merge these two switch instructions if they have a common
218  // successor, and if that successor has a PHI node, and if *that* PHI node has
219  // conflicting incoming values from the two switch blocks.
220  BasicBlock *SI1BB = SI1->getParent();
221  BasicBlock *SI2BB = SI2->getParent();
222 
223  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
224  bool Fail = false;
225  for (BasicBlock *Succ : successors(SI2BB))
226  if (SI1Succs.count(Succ))
227  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
228  PHINode *PN = cast<PHINode>(BBI);
229  if (PN->getIncomingValueForBlock(SI1BB) !=
230  PN->getIncomingValueForBlock(SI2BB)) {
231  if (FailBlocks)
232  FailBlocks->insert(Succ);
233  Fail = true;
234  }
235  }
236 
237  return !Fail;
238 }
239 
240 /// Return true if it is safe and profitable to merge these two terminator
241 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
242 /// store all PHI nodes in common successors.
243 static bool
245  Instruction *Cond,
246  SmallVectorImpl<PHINode *> &PhiNodes) {
247  if (SI1 == SI2)
248  return false; // Can't merge with self!
249  assert(SI1->isUnconditional() && SI2->isConditional());
250 
251  // We fold the unconditional branch if we can easily update all PHI nodes in
252  // common successors:
253  // 1> We have a constant incoming value for the conditional branch;
254  // 2> We have "Cond" as the incoming value for the unconditional branch;
255  // 3> SI2->getCondition() and Cond have same operands.
256  CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
257  if (!Ci2)
258  return false;
259  if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
260  Cond->getOperand(1) == Ci2->getOperand(1)) &&
261  !(Cond->getOperand(0) == Ci2->getOperand(1) &&
262  Cond->getOperand(1) == Ci2->getOperand(0)))
263  return false;
264 
265  BasicBlock *SI1BB = SI1->getParent();
266  BasicBlock *SI2BB = SI2->getParent();
267  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
268  for (BasicBlock *Succ : successors(SI2BB))
269  if (SI1Succs.count(Succ))
270  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
271  PHINode *PN = cast<PHINode>(BBI);
272  if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
273  !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
274  return false;
275  PhiNodes.push_back(PN);
276  }
277  return true;
278 }
279 
280 /// Update PHI nodes in Succ to indicate that there will now be entries in it
281 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
282 /// will be the same as those coming in from ExistPred, an existing predecessor
283 /// of Succ.
284 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
285  BasicBlock *ExistPred) {
286  if (!isa<PHINode>(Succ->begin()))
287  return; // Quick exit if nothing to do
288 
289  PHINode *PN;
290  for (BasicBlock::iterator I = Succ->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
291  PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
292 }
293 
294 /// Compute an abstract "cost" of speculating the given instruction,
295 /// which is assumed to be safe to speculate. TCC_Free means cheap,
296 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
297 /// expensive.
298 static unsigned ComputeSpeculationCost(const User *I,
299  const TargetTransformInfo &TTI) {
301  "Instruction is not safe to speculatively execute!");
302  return TTI.getUserCost(I);
303 }
304 
305 /// If we have a merge point of an "if condition" as accepted above,
306 /// return true if the specified value dominates the block. We
307 /// don't handle the true generality of domination here, just a special case
308 /// which works well enough for us.
309 ///
310 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
311 /// see if V (which must be an instruction) and its recursive operands
312 /// that do not dominate BB have a combined cost lower than CostRemaining and
313 /// are non-trapping. If both are true, the instruction is inserted into the
314 /// set and true is returned.
315 ///
316 /// The cost for most non-trapping instructions is defined as 1 except for
317 /// Select whose cost is 2.
318 ///
319 /// After this function returns, CostRemaining is decreased by the cost of
320 /// V plus its non-dominating operands. If that cost is greater than
321 /// CostRemaining, false is returned and CostRemaining is undefined.
323  SmallPtrSetImpl<Instruction *> *AggressiveInsts,
324  unsigned &CostRemaining,
325  const TargetTransformInfo &TTI,
326  unsigned Depth = 0) {
327  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
328  // so limit the recursion depth.
329  // TODO: While this recursion limit does prevent pathological behavior, it
330  // would be better to track visited instructions to avoid cycles.
331  if (Depth == MaxSpeculationDepth)
332  return false;
333 
335  if (!I) {
336  // Non-instructions all dominate instructions, but not all constantexprs
337  // can be executed unconditionally.
338  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
339  if (C->canTrap())
340  return false;
341  return true;
342  }
343  BasicBlock *PBB = I->getParent();
344 
345  // We don't want to allow weird loops that might have the "if condition" in
346  // the bottom of this block.
347  if (PBB == BB)
348  return false;
349 
350  // If this instruction is defined in a block that contains an unconditional
351  // branch to BB, then it must be in the 'conditional' part of the "if
352  // statement". If not, it definitely dominates the region.
354  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
355  return true;
356 
357  // If we aren't allowing aggressive promotion anymore, then don't consider
358  // instructions in the 'if region'.
359  if (!AggressiveInsts)
360  return false;
361 
362  // If we have seen this instruction before, don't count it again.
363  if (AggressiveInsts->count(I))
364  return true;
365 
366  // Okay, it looks like the instruction IS in the "condition". Check to
367  // see if it's a cheap instruction to unconditionally compute, and if it
368  // only uses stuff defined outside of the condition. If so, hoist it out.
370  return false;
371 
372  unsigned Cost = ComputeSpeculationCost(I, TTI);
373 
374  // Allow exactly one instruction to be speculated regardless of its cost
375  // (as long as it is safe to do so).
376  // This is intended to flatten the CFG even if the instruction is a division
377  // or other expensive operation. The speculation of an expensive instruction
378  // is expected to be undone in CodeGenPrepare if the speculation has not
379  // enabled further IR optimizations.
380  if (Cost > CostRemaining &&
381  (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
382  return false;
383 
384  // Avoid unsigned wrap.
385  CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
386 
387  // Okay, we can only really hoist these out if their operands do
388  // not take us over the cost threshold.
389  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
390  if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
391  Depth + 1))
392  return false;
393  // Okay, it's safe to do this! Remember this instruction.
394  AggressiveInsts->insert(I);
395  return true;
396 }
397 
398 /// Extract ConstantInt from value, looking through IntToPtr
399 /// and PointerNullValue. Return NULL if value is not a constant int.
400 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
401  // Normal constant int.
402  ConstantInt *CI = dyn_cast<ConstantInt>(V);
403  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
404  return CI;
405 
406  // This is some kind of pointer constant. Turn it into a pointer-sized
407  // ConstantInt if possible.
408  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
409 
410  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
411  if (isa<ConstantPointerNull>(V))
412  return ConstantInt::get(PtrTy, 0);
413 
414  // IntToPtr const int.
415  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
416  if (CE->getOpcode() == Instruction::IntToPtr)
417  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
418  // The constant is very likely to have the right type already.
419  if (CI->getType() == PtrTy)
420  return CI;
421  else
422  return cast<ConstantInt>(
423  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
424  }
425  return nullptr;
426 }
427 
428 namespace {
429 
430 /// Given a chain of or (||) or and (&&) comparison of a value against a
431 /// constant, this will try to recover the information required for a switch
432 /// structure.
433 /// It will depth-first traverse the chain of comparison, seeking for patterns
434 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
435 /// representing the different cases for the switch.
436 /// Note that if the chain is composed of '||' it will build the set of elements
437 /// that matches the comparisons (i.e. any of this value validate the chain)
438 /// while for a chain of '&&' it will build the set elements that make the test
439 /// fail.
440 struct ConstantComparesGatherer {
441  const DataLayout &DL;
442 
443  /// Value found for the switch comparison
444  Value *CompValue = nullptr;
445 
446  /// Extra clause to be checked before the switch
447  Value *Extra = nullptr;
448 
449  /// Set of integers to match in switch
451 
452  /// Number of comparisons matched in the and/or chain
453  unsigned UsedICmps = 0;
454 
455  /// Construct and compute the result for the comparison instruction Cond
456  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL) : DL(DL) {
457  gather(Cond);
458  }
459 
460  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
461  ConstantComparesGatherer &
462  operator=(const ConstantComparesGatherer &) = delete;
463 
464 private:
465  /// Try to set the current value used for the comparison, it succeeds only if
466  /// it wasn't set before or if the new value is the same as the old one
467  bool setValueOnce(Value *NewVal) {
468  if (CompValue && CompValue != NewVal)
469  return false;
470  CompValue = NewVal;
471  return (CompValue != nullptr);
472  }
473 
474  /// Try to match Instruction "I" as a comparison against a constant and
475  /// populates the array Vals with the set of values that match (or do not
476  /// match depending on isEQ).
477  /// Return false on failure. On success, the Value the comparison matched
478  /// against is placed in CompValue.
479  /// If CompValue is already set, the function is expected to fail if a match
480  /// is found but the value compared to is different.
481  bool matchInstruction(Instruction *I, bool isEQ) {
482  // If this is an icmp against a constant, handle this as one of the cases.
483  ICmpInst *ICI;
484  ConstantInt *C;
485  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
486  (C = GetConstantInt(I->getOperand(1), DL)))) {
487  return false;
488  }
489 
490  Value *RHSVal;
491  const APInt *RHSC;
492 
493  // Pattern match a special case
494  // (x & ~2^z) == y --> x == y || x == y|2^z
495  // This undoes a transformation done by instcombine to fuse 2 compares.
496  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
497  // It's a little bit hard to see why the following transformations are
498  // correct. Here is a CVC3 program to verify them for 64-bit values:
499 
500  /*
501  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
502  x : BITVECTOR(64);
503  y : BITVECTOR(64);
504  z : BITVECTOR(64);
505  mask : BITVECTOR(64) = BVSHL(ONE, z);
506  QUERY( (y & ~mask = y) =>
507  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
508  );
509  QUERY( (y | mask = y) =>
510  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
511  );
512  */
513 
514  // Please note that each pattern must be a dual implication (<--> or
515  // iff). One directional implication can create spurious matches. If the
516  // implication is only one-way, an unsatisfiable condition on the left
517  // side can imply a satisfiable condition on the right side. Dual
518  // implication ensures that satisfiable conditions are transformed to
519  // other satisfiable conditions and unsatisfiable conditions are
520  // transformed to other unsatisfiable conditions.
521 
522  // Here is a concrete example of a unsatisfiable condition on the left
523  // implying a satisfiable condition on the right:
524  //
525  // mask = (1 << z)
526  // (x & ~mask) == y --> (x == y || x == (y | mask))
527  //
528  // Substituting y = 3, z = 0 yields:
529  // (x & -2) == 3 --> (x == 3 || x == 2)
530 
531  // Pattern match a special case:
532  /*
533  QUERY( (y & ~mask = y) =>
534  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
535  );
536  */
537  if (match(ICI->getOperand(0),
538  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
539  APInt Mask = ~*RHSC;
540  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
541  // If we already have a value for the switch, it has to match!
542  if (!setValueOnce(RHSVal))
543  return false;
544 
545  Vals.push_back(C);
546  Vals.push_back(
548  C->getValue() | Mask));
549  UsedICmps++;
550  return true;
551  }
552  }
553 
554  // Pattern match a special case:
555  /*
556  QUERY( (y | mask = y) =>
557  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
558  );
559  */
560  if (match(ICI->getOperand(0),
561  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
562  APInt Mask = *RHSC;
563  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
564  // If we already have a value for the switch, it has to match!
565  if (!setValueOnce(RHSVal))
566  return false;
567 
568  Vals.push_back(C);
570  C->getValue() & ~Mask));
571  UsedICmps++;
572  return true;
573  }
574  }
575 
576  // If we already have a value for the switch, it has to match!
577  if (!setValueOnce(ICI->getOperand(0)))
578  return false;
579 
580  UsedICmps++;
581  Vals.push_back(C);
582  return ICI->getOperand(0);
583  }
584 
585  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
587  ICI->getPredicate(), C->getValue());
588 
589  // Shift the range if the compare is fed by an add. This is the range
590  // compare idiom as emitted by instcombine.
591  Value *CandidateVal = I->getOperand(0);
592  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
593  Span = Span.subtract(*RHSC);
594  CandidateVal = RHSVal;
595  }
596 
597  // If this is an and/!= check, then we are looking to build the set of
598  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
599  // x != 0 && x != 1.
600  if (!isEQ)
601  Span = Span.inverse();
602 
603  // If there are a ton of values, we don't want to make a ginormous switch.
604  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
605  return false;
606  }
607 
608  // If we already have a value for the switch, it has to match!
609  if (!setValueOnce(CandidateVal))
610  return false;
611 
612  // Add all values from the range to the set
613  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
614  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
615 
616  UsedICmps++;
617  return true;
618  }
619 
620  /// Given a potentially 'or'd or 'and'd together collection of icmp
621  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
622  /// the value being compared, and stick the list constants into the Vals
623  /// vector.
624  /// One "Extra" case is allowed to differ from the other.
625  void gather(Value *V) {
627  bool isEQ = (I->getOpcode() == Instruction::Or);
628 
629  // Keep a stack (SmallVector for efficiency) for depth-first traversal
631  SmallPtrSet<Value *, 8> Visited;
632 
633  // Initialize
634  Visited.insert(V);
635  DFT.push_back(V);
636 
637  while (!DFT.empty()) {
638  V = DFT.pop_back_val();
639 
640  if (Instruction *I = dyn_cast<Instruction>(V)) {
641  // If it is a || (or && depending on isEQ), process the operands.
642  if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
643  if (Visited.insert(I->getOperand(1)).second)
644  DFT.push_back(I->getOperand(1));
645  if (Visited.insert(I->getOperand(0)).second)
646  DFT.push_back(I->getOperand(0));
647  continue;
648  }
649 
650  // Try to match the current instruction
651  if (matchInstruction(I, isEQ))
652  // Match succeed, continue the loop
653  continue;
654  }
655 
656  // One element of the sequence of || (or &&) could not be match as a
657  // comparison against the same value as the others.
658  // We allow only one "Extra" case to be checked before the switch
659  if (!Extra) {
660  Extra = V;
661  continue;
662  }
663  // Failed to parse a proper sequence, abort now
664  CompValue = nullptr;
665  break;
666  }
667  }
668 };
669 
670 } // end anonymous namespace
671 
673  Instruction *Cond = nullptr;
674  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
675  Cond = dyn_cast<Instruction>(SI->getCondition());
676  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
677  if (BI->isConditional())
678  Cond = dyn_cast<Instruction>(BI->getCondition());
679  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
680  Cond = dyn_cast<Instruction>(IBI->getAddress());
681  }
682 
683  TI->eraseFromParent();
684  if (Cond)
686 }
687 
688 /// Return true if the specified terminator checks
689 /// to see if a value is equal to constant integer value.
690 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
691  Value *CV = nullptr;
692  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
693  // Do not permit merging of large switch instructions into their
694  // predecessors unless there is only one predecessor.
695  if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
696  pred_end(SI->getParent())) <=
697  128)
698  CV = SI->getCondition();
699  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
700  if (BI->isConditional() && BI->getCondition()->hasOneUse())
701  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
702  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
703  CV = ICI->getOperand(0);
704  }
705 
706  // Unwrap any lossless ptrtoint cast.
707  if (CV) {
708  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
709  Value *Ptr = PTII->getPointerOperand();
710  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
711  CV = Ptr;
712  }
713  }
714  return CV;
715 }
716 
717 /// Given a value comparison instruction,
718 /// decode all of the 'cases' that it represents and return the 'default' block.
719 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
720  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
721  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
722  Cases.reserve(SI->getNumCases());
723  for (auto Case : SI->cases())
724  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
725  Case.getCaseSuccessor()));
726  return SI->getDefaultDest();
727  }
728 
729  BranchInst *BI = cast<BranchInst>(TI);
730  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
731  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
732  Cases.push_back(ValueEqualityComparisonCase(
733  GetConstantInt(ICI->getOperand(1), DL), Succ));
734  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
735 }
736 
737 /// Given a vector of bb/value pairs, remove any entries
738 /// in the list that match the specified block.
739 static void
741  std::vector<ValueEqualityComparisonCase> &Cases) {
742  Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
743 }
744 
745 /// Return true if there are any keys in C1 that exist in C2 as well.
746 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
747  std::vector<ValueEqualityComparisonCase> &C2) {
748  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
749 
750  // Make V1 be smaller than V2.
751  if (V1->size() > V2->size())
752  std::swap(V1, V2);
753 
754  if (V1->empty())
755  return false;
756  if (V1->size() == 1) {
757  // Just scan V2.
758  ConstantInt *TheVal = (*V1)[0].Value;
759  for (unsigned i = 0, e = V2->size(); i != e; ++i)
760  if (TheVal == (*V2)[i].Value)
761  return true;
762  }
763 
764  // Otherwise, just sort both lists and compare element by element.
765  array_pod_sort(V1->begin(), V1->end());
766  array_pod_sort(V2->begin(), V2->end());
767  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
768  while (i1 != e1 && i2 != e2) {
769  if ((*V1)[i1].Value == (*V2)[i2].Value)
770  return true;
771  if ((*V1)[i1].Value < (*V2)[i2].Value)
772  ++i1;
773  else
774  ++i2;
775  }
776  return false;
777 }
778 
779 // Set branch weights on SwitchInst. This sets the metadata if there is at
780 // least one non-zero weight.
782  // Check that there is at least one non-zero weight. Otherwise, pass
783  // nullptr to setMetadata which will erase the existing metadata.
784  MDNode *N = nullptr;
785  if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
786  N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
788 }
789 
790 // Similar to the above, but for branch and select instructions that take
791 // exactly 2 weights.
792 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
793  uint32_t FalseWeight) {
794  assert(isa<BranchInst>(I) || isa<SelectInst>(I));
795  // Check that there is at least one non-zero weight. Otherwise, pass
796  // nullptr to setMetadata which will erase the existing metadata.
797  MDNode *N = nullptr;
798  if (TrueWeight || FalseWeight)
799  N = MDBuilder(I->getParent()->getContext())
800  .createBranchWeights(TrueWeight, FalseWeight);
802 }
803 
804 /// If TI is known to be a terminator instruction and its block is known to
805 /// only have a single predecessor block, check to see if that predecessor is
806 /// also a value comparison with the same value, and if that comparison
807 /// determines the outcome of this comparison. If so, simplify TI. This does a
808 /// very limited form of jump threading.
809 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
810  TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
811  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
812  if (!PredVal)
813  return false; // Not a value comparison in predecessor.
814 
815  Value *ThisVal = isValueEqualityComparison(TI);
816  assert(ThisVal && "This isn't a value comparison!!");
817  if (ThisVal != PredVal)
818  return false; // Different predicates.
819 
820  // TODO: Preserve branch weight metadata, similarly to how
821  // FoldValueComparisonIntoPredecessors preserves it.
822 
823  // Find out information about when control will move from Pred to TI's block.
824  std::vector<ValueEqualityComparisonCase> PredCases;
825  BasicBlock *PredDef =
826  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
827  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
828 
829  // Find information about how control leaves this block.
830  std::vector<ValueEqualityComparisonCase> ThisCases;
831  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
832  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
833 
834  // If TI's block is the default block from Pred's comparison, potentially
835  // simplify TI based on this knowledge.
836  if (PredDef == TI->getParent()) {
837  // If we are here, we know that the value is none of those cases listed in
838  // PredCases. If there are any cases in ThisCases that are in PredCases, we
839  // can simplify TI.
840  if (!ValuesOverlap(PredCases, ThisCases))
841  return false;
842 
843  if (isa<BranchInst>(TI)) {
844  // Okay, one of the successors of this condbr is dead. Convert it to a
845  // uncond br.
846  assert(ThisCases.size() == 1 && "Branch can only have one case!");
847  // Insert the new branch.
848  Instruction *NI = Builder.CreateBr(ThisDef);
849  (void)NI;
850 
851  // Remove PHI node entries for the dead edge.
852  ThisCases[0].Dest->removePredecessor(TI->getParent());
853 
854  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
855  << "Through successor TI: " << *TI << "Leaving: " << *NI
856  << "\n");
857 
859  return true;
860  }
861 
862  SwitchInst *SI = cast<SwitchInst>(TI);
863  // Okay, TI has cases that are statically dead, prune them away.
864  SmallPtrSet<Constant *, 16> DeadCases;
865  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
866  DeadCases.insert(PredCases[i].Value);
867 
868  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
869  << "Through successor TI: " << *TI);
870 
871  // Collect branch weights into a vector.
872  SmallVector<uint32_t, 8> Weights;
874  bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
875  if (HasWeight)
876  for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
877  ++MD_i) {
878  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
879  Weights.push_back(CI->getValue().getZExtValue());
880  }
881  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
882  --i;
883  if (DeadCases.count(i->getCaseValue())) {
884  if (HasWeight) {
885  std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
886  Weights.pop_back();
887  }
888  i->getCaseSuccessor()->removePredecessor(TI->getParent());
889  SI->removeCase(i);
890  }
891  }
892  if (HasWeight && Weights.size() >= 2)
893  setBranchWeights(SI, Weights);
894 
895  DEBUG(dbgs() << "Leaving: " << *TI << "\n");
896  return true;
897  }
898 
899  // Otherwise, TI's block must correspond to some matched value. Find out
900  // which value (or set of values) this is.
901  ConstantInt *TIV = nullptr;
902  BasicBlock *TIBB = TI->getParent();
903  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
904  if (PredCases[i].Dest == TIBB) {
905  if (TIV)
906  return false; // Cannot handle multiple values coming to this block.
907  TIV = PredCases[i].Value;
908  }
909  assert(TIV && "No edge from pred to succ?");
910 
911  // Okay, we found the one constant that our value can be if we get into TI's
912  // BB. Find out which successor will unconditionally be branched to.
913  BasicBlock *TheRealDest = nullptr;
914  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
915  if (ThisCases[i].Value == TIV) {
916  TheRealDest = ThisCases[i].Dest;
917  break;
918  }
919 
920  // If not handled by any explicit cases, it is handled by the default case.
921  if (!TheRealDest)
922  TheRealDest = ThisDef;
923 
924  // Remove PHI node entries for dead edges.
925  BasicBlock *CheckEdge = TheRealDest;
926  for (BasicBlock *Succ : successors(TIBB))
927  if (Succ != CheckEdge)
928  Succ->removePredecessor(TIBB);
929  else
930  CheckEdge = nullptr;
931 
932  // Insert the new branch.
933  Instruction *NI = Builder.CreateBr(TheRealDest);
934  (void)NI;
935 
936  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
937  << "Through successor TI: " << *TI << "Leaving: " << *NI
938  << "\n");
939 
941  return true;
942 }
943 
944 namespace {
945 
946 /// This class implements a stable ordering of constant
947 /// integers that does not depend on their address. This is important for
948 /// applications that sort ConstantInt's to ensure uniqueness.
949 struct ConstantIntOrdering {
950  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
951  return LHS->getValue().ult(RHS->getValue());
952  }
953 };
954 
955 } // end anonymous namespace
956 
957 static int ConstantIntSortPredicate(ConstantInt *const *P1,
958  ConstantInt *const *P2) {
959  const ConstantInt *LHS = *P1;
960  const ConstantInt *RHS = *P2;
961  if (LHS == RHS)
962  return 0;
963  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
964 }
965 
966 static inline bool HasBranchWeights(const Instruction *I) {
968  if (ProfMD && ProfMD->getOperand(0))
969  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
970  return MDS->getString().equals("branch_weights");
971 
972  return false;
973 }
974 
975 /// Get Weights of a given TerminatorInst, the default weight is at the front
976 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
977 /// metadata.
979  SmallVectorImpl<uint64_t> &Weights) {
981  assert(MD);
982  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
983  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
984  Weights.push_back(CI->getValue().getZExtValue());
985  }
986 
987  // If TI is a conditional eq, the default case is the false case,
988  // and the corresponding branch-weight data is at index 2. We swap the
989  // default weight to be the first entry.
990  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
991  assert(Weights.size() == 2);
992  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
993  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
994  std::swap(Weights.front(), Weights.back());
995  }
996 }
997 
998 /// Keep halving the weights until all can fit in uint32_t.
999 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
1000  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
1001  if (Max > UINT_MAX) {
1002  unsigned Offset = 32 - countLeadingZeros(Max);
1003  for (uint64_t &I : Weights)
1004  I >>= Offset;
1005  }
1006 }
1007 
1008 /// The specified terminator is a value equality comparison instruction
1009 /// (either a switch or a branch on "X == c").
1010 /// See if any of the predecessors of the terminator block are value comparisons
1011 /// on the same value. If so, and if safe to do so, fold them together.
1012 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
1013  IRBuilder<> &Builder) {
1014  BasicBlock *BB = TI->getParent();
1015  Value *CV = isValueEqualityComparison(TI); // CondVal
1016  assert(CV && "Not a comparison?");
1017  bool Changed = false;
1018 
1020  while (!Preds.empty()) {
1021  BasicBlock *Pred = Preds.pop_back_val();
1022 
1023  // See if the predecessor is a comparison with the same value.
1024  TerminatorInst *PTI = Pred->getTerminator();
1025  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1026 
1027  if (PCV == CV && TI != PTI) {
1028  SmallSetVector<BasicBlock*, 4> FailBlocks;
1029  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1030  for (auto *Succ : FailBlocks) {
1031  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1032  return false;
1033  }
1034  }
1035 
1036  // Figure out which 'cases' to copy from SI to PSI.
1037  std::vector<ValueEqualityComparisonCase> BBCases;
1038  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1039 
1040  std::vector<ValueEqualityComparisonCase> PredCases;
1041  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1042 
1043  // Based on whether the default edge from PTI goes to BB or not, fill in
1044  // PredCases and PredDefault with the new switch cases we would like to
1045  // build.
1046  SmallVector<BasicBlock *, 8> NewSuccessors;
1047 
1048  // Update the branch weight metadata along the way
1049  SmallVector<uint64_t, 8> Weights;
1050  bool PredHasWeights = HasBranchWeights(PTI);
1051  bool SuccHasWeights = HasBranchWeights(TI);
1052 
1053  if (PredHasWeights) {
1054  GetBranchWeights(PTI, Weights);
1055  // branch-weight metadata is inconsistent here.
1056  if (Weights.size() != 1 + PredCases.size())
1057  PredHasWeights = SuccHasWeights = false;
1058  } else if (SuccHasWeights)
1059  // If there are no predecessor weights but there are successor weights,
1060  // populate Weights with 1, which will later be scaled to the sum of
1061  // successor's weights
1062  Weights.assign(1 + PredCases.size(), 1);
1063 
1064  SmallVector<uint64_t, 8> SuccWeights;
1065  if (SuccHasWeights) {
1066  GetBranchWeights(TI, SuccWeights);
1067  // branch-weight metadata is inconsistent here.
1068  if (SuccWeights.size() != 1 + BBCases.size())
1069  PredHasWeights = SuccHasWeights = false;
1070  } else if (PredHasWeights)
1071  SuccWeights.assign(1 + BBCases.size(), 1);
1072 
1073  if (PredDefault == BB) {
1074  // If this is the default destination from PTI, only the edges in TI
1075  // that don't occur in PTI, or that branch to BB will be activated.
1076  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1077  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1078  if (PredCases[i].Dest != BB)
1079  PTIHandled.insert(PredCases[i].Value);
1080  else {
1081  // The default destination is BB, we don't need explicit targets.
1082  std::swap(PredCases[i], PredCases.back());
1083 
1084  if (PredHasWeights || SuccHasWeights) {
1085  // Increase weight for the default case.
1086  Weights[0] += Weights[i + 1];
1087  std::swap(Weights[i + 1], Weights.back());
1088  Weights.pop_back();
1089  }
1090 
1091  PredCases.pop_back();
1092  --i;
1093  --e;
1094  }
1095 
1096  // Reconstruct the new switch statement we will be building.
1097  if (PredDefault != BBDefault) {
1098  PredDefault->removePredecessor(Pred);
1099  PredDefault = BBDefault;
1100  NewSuccessors.push_back(BBDefault);
1101  }
1102 
1103  unsigned CasesFromPred = Weights.size();
1104  uint64_t ValidTotalSuccWeight = 0;
1105  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1106  if (!PTIHandled.count(BBCases[i].Value) &&
1107  BBCases[i].Dest != BBDefault) {
1108  PredCases.push_back(BBCases[i]);
1109  NewSuccessors.push_back(BBCases[i].Dest);
1110  if (SuccHasWeights || PredHasWeights) {
1111  // The default weight is at index 0, so weight for the ith case
1112  // should be at index i+1. Scale the cases from successor by
1113  // PredDefaultWeight (Weights[0]).
1114  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1115  ValidTotalSuccWeight += SuccWeights[i + 1];
1116  }
1117  }
1118 
1119  if (SuccHasWeights || PredHasWeights) {
1120  ValidTotalSuccWeight += SuccWeights[0];
1121  // Scale the cases from predecessor by ValidTotalSuccWeight.
1122  for (unsigned i = 1; i < CasesFromPred; ++i)
1123  Weights[i] *= ValidTotalSuccWeight;
1124  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1125  Weights[0] *= SuccWeights[0];
1126  }
1127  } else {
1128  // If this is not the default destination from PSI, only the edges
1129  // in SI that occur in PSI with a destination of BB will be
1130  // activated.
1131  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1132  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1133  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1134  if (PredCases[i].Dest == BB) {
1135  PTIHandled.insert(PredCases[i].Value);
1136 
1137  if (PredHasWeights || SuccHasWeights) {
1138  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1139  std::swap(Weights[i + 1], Weights.back());
1140  Weights.pop_back();
1141  }
1142 
1143  std::swap(PredCases[i], PredCases.back());
1144  PredCases.pop_back();
1145  --i;
1146  --e;
1147  }
1148 
1149  // Okay, now we know which constants were sent to BB from the
1150  // predecessor. Figure out where they will all go now.
1151  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1152  if (PTIHandled.count(BBCases[i].Value)) {
1153  // If this is one we are capable of getting...
1154  if (PredHasWeights || SuccHasWeights)
1155  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1156  PredCases.push_back(BBCases[i]);
1157  NewSuccessors.push_back(BBCases[i].Dest);
1158  PTIHandled.erase(
1159  BBCases[i].Value); // This constant is taken care of
1160  }
1161 
1162  // If there are any constants vectored to BB that TI doesn't handle,
1163  // they must go to the default destination of TI.
1164  for (ConstantInt *I : PTIHandled) {
1165  if (PredHasWeights || SuccHasWeights)
1166  Weights.push_back(WeightsForHandled[I]);
1167  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1168  NewSuccessors.push_back(BBDefault);
1169  }
1170  }
1171 
1172  // Okay, at this point, we know which new successor Pred will get. Make
1173  // sure we update the number of entries in the PHI nodes for these
1174  // successors.
1175  for (BasicBlock *NewSuccessor : NewSuccessors)
1176  AddPredecessorToBlock(NewSuccessor, Pred, BB);
1177 
1178  Builder.SetInsertPoint(PTI);
1179  // Convert pointer to int before we switch.
1180  if (CV->getType()->isPointerTy()) {
1181  CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1182  "magicptr");
1183  }
1184 
1185  // Now that the successors are updated, create the new Switch instruction.
1186  SwitchInst *NewSI =
1187  Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1188  NewSI->setDebugLoc(PTI->getDebugLoc());
1189  for (ValueEqualityComparisonCase &V : PredCases)
1190  NewSI->addCase(V.Value, V.Dest);
1191 
1192  if (PredHasWeights || SuccHasWeights) {
1193  // Halve the weights if any of them cannot fit in an uint32_t
1194  FitWeights(Weights);
1195 
1196  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1197 
1198  setBranchWeights(NewSI, MDWeights);
1199  }
1200 
1202 
1203  // Okay, last check. If BB is still a successor of PSI, then we must
1204  // have an infinite loop case. If so, add an infinitely looping block
1205  // to handle the case to preserve the behavior of the code.
1206  BasicBlock *InfLoopBlock = nullptr;
1207  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1208  if (NewSI->getSuccessor(i) == BB) {
1209  if (!InfLoopBlock) {
1210  // Insert it at the end of the function, because it's either code,
1211  // or it won't matter if it's hot. :)
1212  InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1213  BB->getParent());
1214  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1215  }
1216  NewSI->setSuccessor(i, InfLoopBlock);
1217  }
1218 
1219  Changed = true;
1220  }
1221  }
1222  return Changed;
1223 }
1224 
1225 // If we would need to insert a select that uses the value of this invoke
1226 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1227 // can't hoist the invoke, as there is nowhere to put the select in this case.
1229  Instruction *I1, Instruction *I2) {
1230  for (BasicBlock *Succ : successors(BB1)) {
1231  PHINode *PN;
1232  for (BasicBlock::iterator BBI = Succ->begin();
1233  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1234  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1235  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1236  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1237  return false;
1238  }
1239  }
1240  }
1241  return true;
1242 }
1243 
1245 
1246 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1247 /// in the two blocks up into the branch block. The caller of this function
1248 /// guarantees that BI's block dominates BB1 and BB2.
1250  const TargetTransformInfo &TTI) {
1251  // This does very trivial matching, with limited scanning, to find identical
1252  // instructions in the two blocks. In particular, we don't want to get into
1253  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1254  // such, we currently just scan for obviously identical instructions in an
1255  // identical order.
1256  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1257  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1258 
1259  BasicBlock::iterator BB1_Itr = BB1->begin();
1260  BasicBlock::iterator BB2_Itr = BB2->begin();
1261 
1262  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1263  // Skip debug info if it is not identical.
1266  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1267  while (isa<DbgInfoIntrinsic>(I1))
1268  I1 = &*BB1_Itr++;
1269  while (isa<DbgInfoIntrinsic>(I2))
1270  I2 = &*BB2_Itr++;
1271  }
1272  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1273  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1274  return false;
1275 
1276  BasicBlock *BIParent = BI->getParent();
1277 
1278  bool Changed = false;
1279  do {
1280  // If we are hoisting the terminator instruction, don't move one (making a
1281  // broken BB), instead clone it, and remove BI.
1282  if (isa<TerminatorInst>(I1))
1283  goto HoistTerminator;
1284 
1285  if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1286  return Changed;
1287 
1288  // For a normal instruction, we just move one to right before the branch,
1289  // then replace all uses of the other with the first. Finally, we remove
1290  // the now redundant second instruction.
1291  BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1292  if (!I2->use_empty())
1293  I2->replaceAllUsesWith(I1);
1294  I1->andIRFlags(I2);
1295  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1305  combineMetadata(I1, I2, KnownIDs);
1306 
1307  // I1 and I2 are being combined into a single instruction. Its debug
1308  // location is the merged locations of the original instructions.
1309  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1310 
1311  I2->eraseFromParent();
1312  Changed = true;
1313 
1314  I1 = &*BB1_Itr++;
1315  I2 = &*BB2_Itr++;
1316  // Skip debug info if it is not identical.
1319  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1320  while (isa<DbgInfoIntrinsic>(I1))
1321  I1 = &*BB1_Itr++;
1322  while (isa<DbgInfoIntrinsic>(I2))
1323  I2 = &*BB2_Itr++;
1324  }
1325  } while (I1->isIdenticalToWhenDefined(I2));
1326 
1327  return true;
1328 
1329 HoistTerminator:
1330  // It may not be possible to hoist an invoke.
1331  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1332  return Changed;
1333 
1334  for (BasicBlock *Succ : successors(BB1)) {
1335  PHINode *PN;
1336  for (BasicBlock::iterator BBI = Succ->begin();
1337  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1338  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1339  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1340  if (BB1V == BB2V)
1341  continue;
1342 
1343  // Check for passingValueIsAlwaysUndefined here because we would rather
1344  // eliminate undefined control flow then converting it to a select.
1345  if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1347  return Changed;
1348 
1349  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1350  return Changed;
1351  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1352  return Changed;
1353  }
1354  }
1355 
1356  // Okay, it is safe to hoist the terminator.
1357  Instruction *NT = I1->clone();
1358  BIParent->getInstList().insert(BI->getIterator(), NT);
1359  if (!NT->getType()->isVoidTy()) {
1360  I1->replaceAllUsesWith(NT);
1361  I2->replaceAllUsesWith(NT);
1362  NT->takeName(I1);
1363  }
1364 
1365  IRBuilder<NoFolder> Builder(NT);
1366  // Hoisting one of the terminators from our successor is a great thing.
1367  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1368  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1369  // nodes, so we insert select instruction to compute the final result.
1370  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1371  for (BasicBlock *Succ : successors(BB1)) {
1372  PHINode *PN;
1373  for (BasicBlock::iterator BBI = Succ->begin();
1374  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1375  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1376  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1377  if (BB1V == BB2V)
1378  continue;
1379 
1380  // These values do not agree. Insert a select instruction before NT
1381  // that determines the right value.
1382  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1383  if (!SI)
1384  SI = cast<SelectInst>(
1385  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1386  BB1V->getName() + "." + BB2V->getName(), BI));
1387 
1388  // Make the PHI node use the select for all incoming values for BB1/BB2
1389  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1390  if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1391  PN->setIncomingValue(i, SI);
1392  }
1393  }
1394 
1395  // Update any PHI nodes in our new successors.
1396  for (BasicBlock *Succ : successors(BB1))
1397  AddPredecessorToBlock(Succ, BIParent, BB1);
1398 
1400  return true;
1401 }
1402 
1403 // All instructions in Insts belong to different blocks that all unconditionally
1404 // branch to a common successor. Analyze each instruction and return true if it
1405 // would be possible to sink them into their successor, creating one common
1406 // instruction instead. For every value that would be required to be provided by
1407 // PHI node (because an operand varies in each input block), add to PHIOperands.
1410  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1411  // Prune out obviously bad instructions to move. Any non-store instruction
1412  // must have exactly one use, and we check later that use is by a single,
1413  // common PHI instruction in the successor.
1414  for (auto *I : Insts) {
1415  // These instructions may change or break semantics if moved.
1416  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1417  I->getType()->isTokenTy())
1418  return false;
1419 
1420  // Conservatively return false if I is an inline-asm instruction. Sinking
1421  // and merging inline-asm instructions can potentially create arguments
1422  // that cannot satisfy the inline-asm constraints.
1423  if (const auto *C = dyn_cast<CallInst>(I))
1424  if (C->isInlineAsm())
1425  return false;
1426 
1427  // Everything must have only one use too, apart from stores which
1428  // have no uses.
1429  if (!isa<StoreInst>(I) && !I->hasOneUse())
1430  return false;
1431  }
1432 
1433  const Instruction *I0 = Insts.front();
1434  for (auto *I : Insts)
1435  if (!I->isSameOperationAs(I0))
1436  return false;
1437 
1438  // All instructions in Insts are known to be the same opcode. If they aren't
1439  // stores, check the only user of each is a PHI or in the same block as the
1440  // instruction, because if a user is in the same block as an instruction
1441  // we're contemplating sinking, it must already be determined to be sinkable.
1442  if (!isa<StoreInst>(I0)) {
1443  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1444  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1445  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1446  auto *U = cast<Instruction>(*I->user_begin());
1447  return (PNUse &&
1448  PNUse->getParent() == Succ &&
1449  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1450  U->getParent() == I->getParent();
1451  }))
1452  return false;
1453  }
1454 
1455  // Because SROA can't handle speculating stores of selects, try not
1456  // to sink loads or stores of allocas when we'd have to create a PHI for
1457  // the address operand. Also, because it is likely that loads or stores
1458  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1459  // This can cause code churn which can have unintended consequences down
1460  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1461  // FIXME: This is a workaround for a deficiency in SROA - see
1462  // https://llvm.org/bugs/show_bug.cgi?id=30188
1463  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1464  return isa<AllocaInst>(I->getOperand(1));
1465  }))
1466  return false;
1467  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1468  return isa<AllocaInst>(I->getOperand(0));
1469  }))
1470  return false;
1471 
1472  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1473  if (I0->getOperand(OI)->getType()->isTokenTy())
1474  // Don't touch any operand of token type.
1475  return false;
1476 
1477  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1478  assert(I->getNumOperands() == I0->getNumOperands());
1479  return I->getOperand(OI) == I0->getOperand(OI);
1480  };
1481  if (!all_of(Insts, SameAsI0)) {
1482  if (!canReplaceOperandWithVariable(I0, OI))
1483  // We can't create a PHI from this GEP.
1484  return false;
1485  // Don't create indirect calls! The called value is the final operand.
1486  if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1487  // FIXME: if the call was *already* indirect, we should do this.
1488  return false;
1489  }
1490  for (auto *I : Insts)
1491  PHIOperands[I].push_back(I->getOperand(OI));
1492  }
1493  }
1494  return true;
1495 }
1496 
1497 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1498 // instruction of every block in Blocks to their common successor, commoning
1499 // into one instruction.
1501  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1502 
1503  // canSinkLastInstruction returning true guarantees that every block has at
1504  // least one non-terminator instruction.
1506  for (auto *BB : Blocks) {
1507  Instruction *I = BB->getTerminator();
1508  do {
1509  I = I->getPrevNode();
1510  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1511  if (!isa<DbgInfoIntrinsic>(I))
1512  Insts.push_back(I);
1513  }
1514 
1515  // The only checking we need to do now is that all users of all instructions
1516  // are the same PHI node. canSinkLastInstruction should have checked this but
1517  // it is slightly over-aggressive - it gets confused by commutative instructions
1518  // so double-check it here.
1519  Instruction *I0 = Insts.front();
1520  if (!isa<StoreInst>(I0)) {
1521  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1522  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1523  auto *U = cast<Instruction>(*I->user_begin());
1524  return U == PNUse;
1525  }))
1526  return false;
1527  }
1528 
1529  // We don't need to do any more checking here; canSinkLastInstruction should
1530  // have done it all for us.
1531  SmallVector<Value*, 4> NewOperands;
1532  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1533  // This check is different to that in canSinkLastInstruction. There, we
1534  // cared about the global view once simplifycfg (and instcombine) have
1535  // completed - it takes into account PHIs that become trivially
1536  // simplifiable. However here we need a more local view; if an operand
1537  // differs we create a PHI and rely on instcombine to clean up the very
1538  // small mess we may make.
1539  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1540  return I->getOperand(O) != I0->getOperand(O);
1541  });
1542  if (!NeedPHI) {
1543  NewOperands.push_back(I0->getOperand(O));
1544  continue;
1545  }
1546 
1547  // Create a new PHI in the successor block and populate it.
1548  auto *Op = I0->getOperand(O);
1549  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1550  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1551  Op->getName() + ".sink", &BBEnd->front());
1552  for (auto *I : Insts)
1553  PN->addIncoming(I->getOperand(O), I->getParent());
1554  NewOperands.push_back(PN);
1555  }
1556 
1557  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1558  // and move it to the start of the successor block.
1559  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1560  I0->getOperandUse(O).set(NewOperands[O]);
1561  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1562 
1563  // Update metadata and IR flags, and merge debug locations.
1564  for (auto *I : Insts)
1565  if (I != I0) {
1566  // The debug location for the "common" instruction is the merged locations
1567  // of all the commoned instructions. We start with the original location
1568  // of the "common" instruction and iteratively merge each location in the
1569  // loop below.
1570  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1571  // However, as N-way merge for CallInst is rare, so we use simplified API
1572  // instead of using complex API for N-way merge.
1573  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1574  combineMetadataForCSE(I0, I);
1575  I0->andIRFlags(I);
1576  }
1577 
1578  if (!isa<StoreInst>(I0)) {
1579  // canSinkLastInstruction checked that all instructions were used by
1580  // one and only one PHI node. Find that now, RAUW it to our common
1581  // instruction and nuke it.
1582  assert(I0->hasOneUse());
1583  auto *PN = cast<PHINode>(*I0->user_begin());
1584  PN->replaceAllUsesWith(I0);
1585  PN->eraseFromParent();
1586  }
1587 
1588  // Finally nuke all instructions apart from the common instruction.
1589  for (auto *I : Insts)
1590  if (I != I0)
1591  I->eraseFromParent();
1592 
1593  return true;
1594 }
1595 
1596 namespace {
1597 
1598  // LockstepReverseIterator - Iterates through instructions
1599  // in a set of blocks in reverse order from the first non-terminator.
1600  // For example (assume all blocks have size n):
1601  // LockstepReverseIterator I([B1, B2, B3]);
1602  // *I-- = [B1[n], B2[n], B3[n]];
1603  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1604  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1605  // ...
1606  class LockstepReverseIterator {
1607  ArrayRef<BasicBlock*> Blocks;
1609  bool Fail;
1610 
1611  public:
1612  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1613  reset();
1614  }
1615 
1616  void reset() {
1617  Fail = false;
1618  Insts.clear();
1619  for (auto *BB : Blocks) {
1620  Instruction *Inst = BB->getTerminator();
1621  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1622  Inst = Inst->getPrevNode();
1623  if (!Inst) {
1624  // Block wasn't big enough.
1625  Fail = true;
1626  return;
1627  }
1628  Insts.push_back(Inst);
1629  }
1630  }
1631 
1632  bool isValid() const {
1633  return !Fail;
1634  }
1635 
1636  void operator--() {
1637  if (Fail)
1638  return;
1639  for (auto *&Inst : Insts) {
1640  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1641  Inst = Inst->getPrevNode();
1642  // Already at beginning of block.
1643  if (!Inst) {
1644  Fail = true;
1645  return;
1646  }
1647  }
1648  }
1649 
1651  return Insts;
1652  }
1653  };
1654 
1655 } // end anonymous namespace
1656 
1657 /// Given an unconditional branch that goes to BBEnd,
1658 /// check whether BBEnd has only two predecessors and the other predecessor
1659 /// ends with an unconditional branch. If it is true, sink any common code
1660 /// in the two predecessors to BBEnd.
1662  assert(BI1->isUnconditional());
1663  BasicBlock *BBEnd = BI1->getSuccessor(0);
1664 
1665  // We support two situations:
1666  // (1) all incoming arcs are unconditional
1667  // (2) one incoming arc is conditional
1668  //
1669  // (2) is very common in switch defaults and
1670  // else-if patterns;
1671  //
1672  // if (a) f(1);
1673  // else if (b) f(2);
1674  //
1675  // produces:
1676  //
1677  // [if]
1678  // / \
1679  // [f(1)] [if]
1680  // | | \
1681  // | | |
1682  // | [f(2)]|
1683  // \ | /
1684  // [ end ]
1685  //
1686  // [end] has two unconditional predecessor arcs and one conditional. The
1687  // conditional refers to the implicit empty 'else' arc. This conditional
1688  // arc can also be caused by an empty default block in a switch.
1689  //
1690  // In this case, we attempt to sink code from all *unconditional* arcs.
1691  // If we can sink instructions from these arcs (determined during the scan
1692  // phase below) we insert a common successor for all unconditional arcs and
1693  // connect that to [end], to enable sinking:
1694  //
1695  // [if]
1696  // / \
1697  // [x(1)] [if]
1698  // | | \
1699  // | | \
1700  // | [x(2)] |
1701  // \ / |
1702  // [sink.split] |
1703  // \ /
1704  // [ end ]
1705  //
1706  SmallVector<BasicBlock*,4> UnconditionalPreds;
1707  Instruction *Cond = nullptr;
1708  for (auto *B : predecessors(BBEnd)) {
1709  auto *T = B->getTerminator();
1710  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1711  UnconditionalPreds.push_back(B);
1712  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1713  Cond = T;
1714  else
1715  return false;
1716  }
1717  if (UnconditionalPreds.size() < 2)
1718  return false;
1719 
1720  bool Changed = false;
1721  // We take a two-step approach to tail sinking. First we scan from the end of
1722  // each block upwards in lockstep. If the n'th instruction from the end of each
1723  // block can be sunk, those instructions are added to ValuesToSink and we
1724  // carry on. If we can sink an instruction but need to PHI-merge some operands
1725  // (because they're not identical in each instruction) we add these to
1726  // PHIOperands.
1727  unsigned ScanIdx = 0;
1728  SmallPtrSet<Value*,4> InstructionsToSink;
1730  LockstepReverseIterator LRI(UnconditionalPreds);
1731  while (LRI.isValid() &&
1732  canSinkInstructions(*LRI, PHIOperands)) {
1733  DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1734  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1735  ++ScanIdx;
1736  --LRI;
1737  }
1738 
1739  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1740  unsigned NumPHIdValues = 0;
1741  for (auto *I : *LRI)
1742  for (auto *V : PHIOperands[I])
1743  if (InstructionsToSink.count(V) == 0)
1744  ++NumPHIdValues;
1745  DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1746  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1747  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1748  NumPHIInsts++;
1749 
1750  return NumPHIInsts <= 1;
1751  };
1752 
1753  if (ScanIdx > 0 && Cond) {
1754  // Check if we would actually sink anything first! This mutates the CFG and
1755  // adds an extra block. The goal in doing this is to allow instructions that
1756  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1757  // (such as trunc, add) can be sunk and predicated already. So we check that
1758  // we're going to sink at least one non-speculatable instruction.
1759  LRI.reset();
1760  unsigned Idx = 0;
1761  bool Profitable = false;
1762  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1763  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1764  Profitable = true;
1765  break;
1766  }
1767  --LRI;
1768  ++Idx;
1769  }
1770  if (!Profitable)
1771  return false;
1772 
1773  DEBUG(dbgs() << "SINK: Splitting edge\n");
1774  // We have a conditional edge and we're going to sink some instructions.
1775  // Insert a new block postdominating all blocks we're going to sink from.
1776  if (!SplitBlockPredecessors(BI1->getSuccessor(0), UnconditionalPreds,
1777  ".sink.split"))
1778  // Edges couldn't be split.
1779  return false;
1780  Changed = true;
1781  }
1782 
1783  // Now that we've analyzed all potential sinking candidates, perform the
1784  // actual sink. We iteratively sink the last non-terminator of the source
1785  // blocks into their common successor unless doing so would require too
1786  // many PHI instructions to be generated (currently only one PHI is allowed
1787  // per sunk instruction).
1788  //
1789  // We can use InstructionsToSink to discount values needing PHI-merging that will
1790  // actually be sunk in a later iteration. This allows us to be more
1791  // aggressive in what we sink. This does allow a false positive where we
1792  // sink presuming a later value will also be sunk, but stop half way through
1793  // and never actually sink it which means we produce more PHIs than intended.
1794  // This is unlikely in practice though.
1795  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1796  DEBUG(dbgs() << "SINK: Sink: "
1797  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1798  << "\n");
1799 
1800  // Because we've sunk every instruction in turn, the current instruction to
1801  // sink is always at index 0.
1802  LRI.reset();
1803  if (!ProfitableToSinkInstruction(LRI)) {
1804  // Too many PHIs would be created.
1805  DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1806  break;
1807  }
1808 
1809  if (!sinkLastInstruction(UnconditionalPreds))
1810  return Changed;
1811  NumSinkCommons++;
1812  Changed = true;
1813  }
1814  return Changed;
1815 }
1816 
1817 /// \brief Determine if we can hoist sink a sole store instruction out of a
1818 /// conditional block.
1819 ///
1820 /// We are looking for code like the following:
1821 /// BrBB:
1822 /// store i32 %add, i32* %arrayidx2
1823 /// ... // No other stores or function calls (we could be calling a memory
1824 /// ... // function).
1825 /// %cmp = icmp ult %x, %y
1826 /// br i1 %cmp, label %EndBB, label %ThenBB
1827 /// ThenBB:
1828 /// store i32 %add5, i32* %arrayidx2
1829 /// br label EndBB
1830 /// EndBB:
1831 /// ...
1832 /// We are going to transform this into:
1833 /// BrBB:
1834 /// store i32 %add, i32* %arrayidx2
1835 /// ... //
1836 /// %cmp = icmp ult %x, %y
1837 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1838 /// store i32 %add.add5, i32* %arrayidx2
1839 /// ...
1840 ///
1841 /// \return The pointer to the value of the previous store if the store can be
1842 /// hoisted into the predecessor block. 0 otherwise.
1844  BasicBlock *StoreBB, BasicBlock *EndBB) {
1845  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1846  if (!StoreToHoist)
1847  return nullptr;
1848 
1849  // Volatile or atomic.
1850  if (!StoreToHoist->isSimple())
1851  return nullptr;
1852 
1853  Value *StorePtr = StoreToHoist->getPointerOperand();
1854 
1855  // Look for a store to the same pointer in BrBB.
1856  unsigned MaxNumInstToLookAt = 9;
1857  for (Instruction &CurI : reverse(*BrBB)) {
1858  if (!MaxNumInstToLookAt)
1859  break;
1860  // Skip debug info.
1861  if (isa<DbgInfoIntrinsic>(CurI))
1862  continue;
1863  --MaxNumInstToLookAt;
1864 
1865  // Could be calling an instruction that affects memory like free().
1866  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1867  return nullptr;
1868 
1869  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1870  // Found the previous store make sure it stores to the same location.
1871  if (SI->getPointerOperand() == StorePtr)
1872  // Found the previous store, return its value operand.
1873  return SI->getValueOperand();
1874  return nullptr; // Unknown store.
1875  }
1876  }
1877 
1878  return nullptr;
1879 }
1880 
1881 /// \brief Speculate a conditional basic block flattening the CFG.
1882 ///
1883 /// Note that this is a very risky transform currently. Speculating
1884 /// instructions like this is most often not desirable. Instead, there is an MI
1885 /// pass which can do it with full awareness of the resource constraints.
1886 /// However, some cases are "obvious" and we should do directly. An example of
1887 /// this is speculating a single, reasonably cheap instruction.
1888 ///
1889 /// There is only one distinct advantage to flattening the CFG at the IR level:
1890 /// it makes very common but simplistic optimizations such as are common in
1891 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1892 /// modeling their effects with easier to reason about SSA value graphs.
1893 ///
1894 ///
1895 /// An illustration of this transform is turning this IR:
1896 /// \code
1897 /// BB:
1898 /// %cmp = icmp ult %x, %y
1899 /// br i1 %cmp, label %EndBB, label %ThenBB
1900 /// ThenBB:
1901 /// %sub = sub %x, %y
1902 /// br label BB2
1903 /// EndBB:
1904 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1905 /// ...
1906 /// \endcode
1907 ///
1908 /// Into this IR:
1909 /// \code
1910 /// BB:
1911 /// %cmp = icmp ult %x, %y
1912 /// %sub = sub %x, %y
1913 /// %cond = select i1 %cmp, 0, %sub
1914 /// ...
1915 /// \endcode
1916 ///
1917 /// \returns true if the conditional block is removed.
1919  const TargetTransformInfo &TTI) {
1920  // Be conservative for now. FP select instruction can often be expensive.
1921  Value *BrCond = BI->getCondition();
1922  if (isa<FCmpInst>(BrCond))
1923  return false;
1924 
1925  BasicBlock *BB = BI->getParent();
1926  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1927 
1928  // If ThenBB is actually on the false edge of the conditional branch, remember
1929  // to swap the select operands later.
1930  bool Invert = false;
1931  if (ThenBB != BI->getSuccessor(0)) {
1932  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1933  Invert = true;
1934  }
1935  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1936 
1937  // Keep a count of how many times instructions are used within CondBB when
1938  // they are candidates for sinking into CondBB. Specifically:
1939  // - They are defined in BB, and
1940  // - They have no side effects, and
1941  // - All of their uses are in CondBB.
1942  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1943 
1944  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1945 
1946  unsigned SpeculationCost = 0;
1947  Value *SpeculatedStoreValue = nullptr;
1948  StoreInst *SpeculatedStore = nullptr;
1949  for (BasicBlock::iterator BBI = ThenBB->begin(),
1950  BBE = std::prev(ThenBB->end());
1951  BBI != BBE; ++BBI) {
1952  Instruction *I = &*BBI;
1953  // Skip debug info.
1954  if (isa<DbgInfoIntrinsic>(I)) {
1955  SpeculatedDbgIntrinsics.push_back(I);
1956  continue;
1957  }
1958 
1959  // Only speculatively execute a single instruction (not counting the
1960  // terminator) for now.
1961  ++SpeculationCost;
1962  if (SpeculationCost > 1)
1963  return false;
1964 
1965  // Don't hoist the instruction if it's unsafe or expensive.
1966  if (!isSafeToSpeculativelyExecute(I) &&
1967  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1968  I, BB, ThenBB, EndBB))))
1969  return false;
1970  if (!SpeculatedStoreValue &&
1971  ComputeSpeculationCost(I, TTI) >
1973  return false;
1974 
1975  // Store the store speculation candidate.
1976  if (SpeculatedStoreValue)
1977  SpeculatedStore = cast<StoreInst>(I);
1978 
1979  // Do not hoist the instruction if any of its operands are defined but not
1980  // used in BB. The transformation will prevent the operand from
1981  // being sunk into the use block.
1982  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1983  Instruction *OpI = dyn_cast<Instruction>(*i);
1984  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1985  continue; // Not a candidate for sinking.
1986 
1987  ++SinkCandidateUseCounts[OpI];
1988  }
1989  }
1990 
1991  // Consider any sink candidates which are only used in CondBB as costs for
1992  // speculation. Note, while we iterate over a DenseMap here, we are summing
1993  // and so iteration order isn't significant.
1995  I = SinkCandidateUseCounts.begin(),
1996  E = SinkCandidateUseCounts.end();
1997  I != E; ++I)
1998  if (I->first->getNumUses() == I->second) {
1999  ++SpeculationCost;
2000  if (SpeculationCost > 1)
2001  return false;
2002  }
2003 
2004  // Check that the PHI nodes can be converted to selects.
2005  bool HaveRewritablePHIs = false;
2006  for (BasicBlock::iterator I = EndBB->begin();
2007  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2008  Value *OrigV = PN->getIncomingValueForBlock(BB);
2009  Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
2010 
2011  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2012  // Skip PHIs which are trivial.
2013  if (ThenV == OrigV)
2014  continue;
2015 
2016  // Don't convert to selects if we could remove undefined behavior instead.
2017  if (passingValueIsAlwaysUndefined(OrigV, PN) ||
2018  passingValueIsAlwaysUndefined(ThenV, PN))
2019  return false;
2020 
2021  HaveRewritablePHIs = true;
2022  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2023  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2024  if (!OrigCE && !ThenCE)
2025  continue; // Known safe and cheap.
2026 
2027  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2028  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2029  return false;
2030  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2031  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2032  unsigned MaxCost =
2034  if (OrigCost + ThenCost > MaxCost)
2035  return false;
2036 
2037  // Account for the cost of an unfolded ConstantExpr which could end up
2038  // getting expanded into Instructions.
2039  // FIXME: This doesn't account for how many operations are combined in the
2040  // constant expression.
2041  ++SpeculationCost;
2042  if (SpeculationCost > 1)
2043  return false;
2044  }
2045 
2046  // If there are no PHIs to process, bail early. This helps ensure idempotence
2047  // as well.
2048  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2049  return false;
2050 
2051  // If we get here, we can hoist the instruction and if-convert.
2052  DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2053 
2054  // Insert a select of the value of the speculated store.
2055  if (SpeculatedStoreValue) {
2056  IRBuilder<NoFolder> Builder(BI);
2057  Value *TrueV = SpeculatedStore->getValueOperand();
2058  Value *FalseV = SpeculatedStoreValue;
2059  if (Invert)
2060  std::swap(TrueV, FalseV);
2061  Value *S = Builder.CreateSelect(
2062  BrCond, TrueV, FalseV, "spec.store.select", BI);
2063  SpeculatedStore->setOperand(0, S);
2064  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2065  SpeculatedStore->getDebugLoc());
2066  }
2067 
2068  // Metadata can be dependent on the condition we are hoisting above.
2069  // Conservatively strip all metadata on the instruction.
2070  for (auto &I : *ThenBB)
2072 
2073  // Hoist the instructions.
2074  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2075  ThenBB->begin(), std::prev(ThenBB->end()));
2076 
2077  // Insert selects and rewrite the PHI operands.
2078  IRBuilder<NoFolder> Builder(BI);
2079  for (BasicBlock::iterator I = EndBB->begin();
2080  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2081  unsigned OrigI = PN->getBasicBlockIndex(BB);
2082  unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
2083  Value *OrigV = PN->getIncomingValue(OrigI);
2084  Value *ThenV = PN->getIncomingValue(ThenI);
2085 
2086  // Skip PHIs which are trivial.
2087  if (OrigV == ThenV)
2088  continue;
2089 
2090  // Create a select whose true value is the speculatively executed value and
2091  // false value is the preexisting value. Swap them if the branch
2092  // destinations were inverted.
2093  Value *TrueV = ThenV, *FalseV = OrigV;
2094  if (Invert)
2095  std::swap(TrueV, FalseV);
2096  Value *V = Builder.CreateSelect(
2097  BrCond, TrueV, FalseV, "spec.select", BI);
2098  PN->setIncomingValue(OrigI, V);
2099  PN->setIncomingValue(ThenI, V);
2100  }
2101 
2102  // Remove speculated dbg intrinsics.
2103  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2104  // dbg value for the different flows and inserting it after the select.
2105  for (Instruction *I : SpeculatedDbgIntrinsics)
2106  I->eraseFromParent();
2107 
2108  ++NumSpeculations;
2109  return true;
2110 }
2111 
2112 /// Return true if we can thread a branch across this block.
2114  BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2115  unsigned Size = 0;
2116 
2117  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2118  if (isa<DbgInfoIntrinsic>(BBI))
2119  continue;
2120  if (Size > 10)
2121  return false; // Don't clone large BB's.
2122  ++Size;
2123 
2124  // We can only support instructions that do not define values that are
2125  // live outside of the current basic block.
2126  for (User *U : BBI->users()) {
2127  Instruction *UI = cast<Instruction>(U);
2128  if (UI->getParent() != BB || isa<PHINode>(UI))
2129  return false;
2130  }
2131 
2132  // Looks ok, continue checking.
2133  }
2134 
2135  return true;
2136 }
2137 
2138 /// If we have a conditional branch on a PHI node value that is defined in the
2139 /// same block as the branch and if any PHI entries are constants, thread edges
2140 /// corresponding to that entry to be branches to their ultimate destination.
2141 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2142  AssumptionCache *AC) {
2143  BasicBlock *BB = BI->getParent();
2144  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2145  // NOTE: we currently cannot transform this case if the PHI node is used
2146  // outside of the block.
2147  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2148  return false;
2149 
2150  // Degenerate case of a single entry PHI.
2151  if (PN->getNumIncomingValues() == 1) {
2153  return true;
2154  }
2155 
2156  // Now we know that this block has multiple preds and two succs.
2158  return false;
2159 
2160  // Can't fold blocks that contain noduplicate or convergent calls.
2161  if (any_of(*BB, [](const Instruction &I) {
2162  const CallInst *CI = dyn_cast<CallInst>(&I);
2163  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2164  }))
2165  return false;
2166 
2167  // Okay, this is a simple enough basic block. See if any phi values are
2168  // constants.
2169  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2171  if (!CB || !CB->getType()->isIntegerTy(1))
2172  continue;
2173 
2174  // Okay, we now know that all edges from PredBB should be revectored to
2175  // branch to RealDest.
2176  BasicBlock *PredBB = PN->getIncomingBlock(i);
2177  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2178 
2179  if (RealDest == BB)
2180  continue; // Skip self loops.
2181  // Skip if the predecessor's terminator is an indirect branch.
2182  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2183  continue;
2184 
2185  // The dest block might have PHI nodes, other predecessors and other
2186  // difficult cases. Instead of being smart about this, just insert a new
2187  // block that jumps to the destination block, effectively splitting
2188  // the edge we are about to create.
2189  BasicBlock *EdgeBB =
2190  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2191  RealDest->getParent(), RealDest);
2192  BranchInst::Create(RealDest, EdgeBB);
2193 
2194  // Update PHI nodes.
2195  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2196 
2197  // BB may have instructions that are being threaded over. Clone these
2198  // instructions into EdgeBB. We know that there will be no uses of the
2199  // cloned instructions outside of EdgeBB.
2200  BasicBlock::iterator InsertPt = EdgeBB->begin();
2201  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2202  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2203  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2204  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2205  continue;
2206  }
2207  // Clone the instruction.
2208  Instruction *N = BBI->clone();
2209  if (BBI->hasName())
2210  N->setName(BBI->getName() + ".c");
2211 
2212  // Update operands due to translation.
2213  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2214  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2215  if (PI != TranslateMap.end())
2216  *i = PI->second;
2217  }
2218 
2219  // Check for trivial simplification.
2220  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2221  if (!BBI->use_empty())
2222  TranslateMap[&*BBI] = V;
2223  if (!N->mayHaveSideEffects()) {
2224  N->deleteValue(); // Instruction folded away, don't need actual inst
2225  N = nullptr;
2226  }
2227  } else {
2228  if (!BBI->use_empty())
2229  TranslateMap[&*BBI] = N;
2230  }
2231  // Insert the new instruction into its new home.
2232  if (N)
2233  EdgeBB->getInstList().insert(InsertPt, N);
2234 
2235  // Register the new instruction with the assumption cache if necessary.
2236  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2237  if (II->getIntrinsicID() == Intrinsic::assume)
2238  AC->registerAssumption(II);
2239  }
2240 
2241  // Loop over all of the edges from PredBB to BB, changing them to branch
2242  // to EdgeBB instead.
2243  TerminatorInst *PredBBTI = PredBB->getTerminator();
2244  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2245  if (PredBBTI->getSuccessor(i) == BB) {
2246  BB->removePredecessor(PredBB);
2247  PredBBTI->setSuccessor(i, EdgeBB);
2248  }
2249 
2250  // Recurse, simplifying any other constants.
2251  return FoldCondBranchOnPHI(BI, DL, AC) | true;
2252  }
2253 
2254  return false;
2255 }
2256 
2257 /// Given a BB that starts with the specified two-entry PHI node,
2258 /// see if we can eliminate it.
2260  const DataLayout &DL) {
2261  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2262  // statement", which has a very simple dominance structure. Basically, we
2263  // are trying to find the condition that is being branched on, which
2264  // subsequently causes this merge to happen. We really want control
2265  // dependence information for this check, but simplifycfg can't keep it up
2266  // to date, and this catches most of the cases we care about anyway.
2267  BasicBlock *BB = PN->getParent();
2268  BasicBlock *IfTrue, *IfFalse;
2269  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2270  if (!IfCond ||
2271  // Don't bother if the branch will be constant folded trivially.
2272  isa<ConstantInt>(IfCond))
2273  return false;
2274 
2275  // Okay, we found that we can merge this two-entry phi node into a select.
2276  // Doing so would require us to fold *all* two entry phi nodes in this block.
2277  // At some point this becomes non-profitable (particularly if the target
2278  // doesn't support cmov's). Only do this transformation if there are two or
2279  // fewer PHI nodes in this block.
2280  unsigned NumPhis = 0;
2281  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2282  if (NumPhis > 2)
2283  return false;
2284 
2285  // Loop over the PHI's seeing if we can promote them all to select
2286  // instructions. While we are at it, keep track of the instructions
2287  // that need to be moved to the dominating block.
2288  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2289  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2290  MaxCostVal1 = PHINodeFoldingThreshold;
2291  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2292  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2293 
2294  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2295  PHINode *PN = cast<PHINode>(II++);
2296  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2297  PN->replaceAllUsesWith(V);
2298  PN->eraseFromParent();
2299  continue;
2300  }
2301 
2302  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2303  MaxCostVal0, TTI) ||
2304  !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2305  MaxCostVal1, TTI))
2306  return false;
2307  }
2308 
2309  // If we folded the first phi, PN dangles at this point. Refresh it. If
2310  // we ran out of PHIs then we simplified them all.
2311  PN = dyn_cast<PHINode>(BB->begin());
2312  if (!PN)
2313  return true;
2314 
2315  // Don't fold i1 branches on PHIs which contain binary operators. These can
2316  // often be turned into switches and other things.
2317  if (PN->getType()->isIntegerTy(1) &&
2318  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2319  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2320  isa<BinaryOperator>(IfCond)))
2321  return false;
2322 
2323  // If all PHI nodes are promotable, check to make sure that all instructions
2324  // in the predecessor blocks can be promoted as well. If not, we won't be able
2325  // to get rid of the control flow, so it's not worth promoting to select
2326  // instructions.
2327  BasicBlock *DomBlock = nullptr;
2328  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2329  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2330  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2331  IfBlock1 = nullptr;
2332  } else {
2333  DomBlock = *pred_begin(IfBlock1);
2334  for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2335  ++I)
2336  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2337  // This is not an aggressive instruction that we can promote.
2338  // Because of this, we won't be able to get rid of the control flow, so
2339  // the xform is not worth it.
2340  return false;
2341  }
2342  }
2343 
2344  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2345  IfBlock2 = nullptr;
2346  } else {
2347  DomBlock = *pred_begin(IfBlock2);
2348  for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2349  ++I)
2350  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2351  // This is not an aggressive instruction that we can promote.
2352  // Because of this, we won't be able to get rid of the control flow, so
2353  // the xform is not worth it.
2354  return false;
2355  }
2356  }
2357 
2358  DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2359  << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2360 
2361  // If we can still promote the PHI nodes after this gauntlet of tests,
2362  // do all of the PHI's now.
2363  Instruction *InsertPt = DomBlock->getTerminator();
2364  IRBuilder<NoFolder> Builder(InsertPt);
2365 
2366  // Move all 'aggressive' instructions, which are defined in the
2367  // conditional parts of the if's up to the dominating block.
2368  if (IfBlock1) {
2369  for (auto &I : *IfBlock1)
2371  DomBlock->getInstList().splice(InsertPt->getIterator(),
2372  IfBlock1->getInstList(), IfBlock1->begin(),
2373  IfBlock1->getTerminator()->getIterator());
2374  }
2375  if (IfBlock2) {
2376  for (auto &I : *IfBlock2)
2378  DomBlock->getInstList().splice(InsertPt->getIterator(),
2379  IfBlock2->getInstList(), IfBlock2->begin(),
2380  IfBlock2->getTerminator()->getIterator());
2381  }
2382 
2383  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2384  // Change the PHI node into a select instruction.
2385  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2386  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2387 
2388  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2389  PN->replaceAllUsesWith(Sel);
2390  Sel->takeName(PN);
2391  PN->eraseFromParent();
2392  }
2393 
2394  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2395  // has been flattened. Change DomBlock to jump directly to our new block to
2396  // avoid other simplifycfg's kicking in on the diamond.
2397  TerminatorInst *OldTI = DomBlock->getTerminator();
2398  Builder.SetInsertPoint(OldTI);
2399  Builder.CreateBr(BB);
2400  OldTI->eraseFromParent();
2401  return true;
2402 }
2403 
2404 /// If we found a conditional branch that goes to two returning blocks,
2405 /// try to merge them together into one return,
2406 /// introducing a select if the return values disagree.
2408  IRBuilder<> &Builder) {
2409  assert(BI->isConditional() && "Must be a conditional branch");
2410  BasicBlock *TrueSucc = BI->getSuccessor(0);
2411  BasicBlock *FalseSucc = BI->getSuccessor(1);
2412  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2413  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2414 
2415  // Check to ensure both blocks are empty (just a return) or optionally empty
2416  // with PHI nodes. If there are other instructions, merging would cause extra
2417  // computation on one path or the other.
2418  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2419  return false;
2420  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2421  return false;
2422 
2423  Builder.SetInsertPoint(BI);
2424  // Okay, we found a branch that is going to two return nodes. If
2425  // there is no return value for this function, just change the
2426  // branch into a return.
2427  if (FalseRet->getNumOperands() == 0) {
2428  TrueSucc->removePredecessor(BI->getParent());
2429  FalseSucc->removePredecessor(BI->getParent());
2430  Builder.CreateRetVoid();
2432  return true;
2433  }
2434 
2435  // Otherwise, figure out what the true and false return values are
2436  // so we can insert a new select instruction.
2437  Value *TrueValue = TrueRet->getReturnValue();
2438  Value *FalseValue = FalseRet->getReturnValue();
2439 
2440  // Unwrap any PHI nodes in the return blocks.
2441  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2442  if (TVPN->getParent() == TrueSucc)
2443  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2444  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2445  if (FVPN->getParent() == FalseSucc)
2446  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2447 
2448  // In order for this transformation to be safe, we must be able to
2449  // unconditionally execute both operands to the return. This is
2450  // normally the case, but we could have a potentially-trapping
2451  // constant expression that prevents this transformation from being
2452  // safe.
2453  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2454  if (TCV->canTrap())
2455  return false;
2456  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2457  if (FCV->canTrap())
2458  return false;
2459 
2460  // Okay, we collected all the mapped values and checked them for sanity, and
2461  // defined to really do this transformation. First, update the CFG.
2462  TrueSucc->removePredecessor(BI->getParent());
2463  FalseSucc->removePredecessor(BI->getParent());
2464 
2465  // Insert select instructions where needed.
2466  Value *BrCond = BI->getCondition();
2467  if (TrueValue) {
2468  // Insert a select if the results differ.
2469  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2470  } else if (isa<UndefValue>(TrueValue)) {
2471  TrueValue = FalseValue;
2472  } else {
2473  TrueValue =
2474  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2475  }
2476  }
2477 
2478  Value *RI =
2479  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2480 
2481  (void)RI;
2482 
2483  DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2484  << "\n " << *BI << "NewRet = " << *RI
2485  << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2486 
2488 
2489  return true;
2490 }
2491 
2492 /// Return true if the given instruction is available
2493 /// in its predecessor block. If yes, the instruction will be removed.
2495  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2496  return false;
2497  for (Instruction &I : *PB) {
2498  Instruction *PBI = &I;
2499  // Check whether Inst and PBI generate the same value.
2500  if (Inst->isIdenticalTo(PBI)) {
2501  Inst->replaceAllUsesWith(PBI);
2502  Inst->eraseFromParent();
2503  return true;
2504  }
2505  }
2506  return false;
2507 }
2508 
2509 /// Return true if either PBI or BI has branch weight available, and store
2510 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2511 /// not have branch weight, use 1:1 as its weight.
2513  uint64_t &PredTrueWeight,
2514  uint64_t &PredFalseWeight,
2515  uint64_t &SuccTrueWeight,
2516  uint64_t &SuccFalseWeight) {
2517  bool PredHasWeights =
2518  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2519  bool SuccHasWeights =
2520  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2521  if (PredHasWeights || SuccHasWeights) {
2522  if (!PredHasWeights)
2523  PredTrueWeight = PredFalseWeight = 1;
2524  if (!SuccHasWeights)
2525  SuccTrueWeight = SuccFalseWeight = 1;
2526  return true;
2527  } else {
2528  return false;
2529  }
2530 }
2531 
2532 /// If this basic block is simple enough, and if a predecessor branches to us
2533 /// and one of our successors, fold the block into the predecessor and use
2534 /// logical operations to pick the right destination.
2535 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2536  BasicBlock *BB = BI->getParent();
2537 
2538  Instruction *Cond = nullptr;
2539  if (BI->isConditional())
2540  Cond = dyn_cast<Instruction>(BI->getCondition());
2541  else {
2542  // For unconditional branch, check for a simple CFG pattern, where
2543  // BB has a single predecessor and BB's successor is also its predecessor's
2544  // successor. If such pattern exists, check for CSE between BB and its
2545  // predecessor.
2546  if (BasicBlock *PB = BB->getSinglePredecessor())
2547  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2548  if (PBI->isConditional() &&
2549  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2550  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2551  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2552  Instruction *Curr = &*I++;
2553  if (isa<CmpInst>(Curr)) {
2554  Cond = Curr;
2555  break;
2556  }
2557  // Quit if we can't remove this instruction.
2558  if (!checkCSEInPredecessor(Curr, PB))
2559  return false;
2560  }
2561  }
2562 
2563  if (!Cond)
2564  return false;
2565  }
2566 
2567  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2568  Cond->getParent() != BB || !Cond->hasOneUse())
2569  return false;
2570 
2571  // Make sure the instruction after the condition is the cond branch.
2572  BasicBlock::iterator CondIt = ++Cond->getIterator();
2573 
2574  // Ignore dbg intrinsics.
2575  while (isa<DbgInfoIntrinsic>(CondIt))
2576  ++CondIt;
2577 
2578  if (&*CondIt != BI)
2579  return false;
2580 
2581  // Only allow this transformation if computing the condition doesn't involve
2582  // too many instructions and these involved instructions can be executed
2583  // unconditionally. We denote all involved instructions except the condition
2584  // as "bonus instructions", and only allow this transformation when the
2585  // number of the bonus instructions does not exceed a certain threshold.
2586  unsigned NumBonusInsts = 0;
2587  for (auto I = BB->begin(); Cond != &*I; ++I) {
2588  // Ignore dbg intrinsics.
2589  if (isa<DbgInfoIntrinsic>(I))
2590  continue;
2591  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2592  return false;
2593  // I has only one use and can be executed unconditionally.
2595  if (User == nullptr || User->getParent() != BB)
2596  return false;
2597  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2598  // to use any other instruction, User must be an instruction between next(I)
2599  // and Cond.
2600  ++NumBonusInsts;
2601  // Early exits once we reach the limit.
2602  if (NumBonusInsts > BonusInstThreshold)
2603  return false;
2604  }
2605 
2606  // Cond is known to be a compare or binary operator. Check to make sure that
2607  // neither operand is a potentially-trapping constant expression.
2608  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2609  if (CE->canTrap())
2610  return false;
2611  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2612  if (CE->canTrap())
2613  return false;
2614 
2615  // Finally, don't infinitely unroll conditional loops.
2616  BasicBlock *TrueDest = BI->getSuccessor(0);
2617  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2618  if (TrueDest == BB || FalseDest == BB)
2619  return false;
2620 
2621  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2622  BasicBlock *PredBlock = *PI;
2623  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2624 
2625  // Check that we have two conditional branches. If there is a PHI node in
2626  // the common successor, verify that the same value flows in from both
2627  // blocks.
2629  if (!PBI || PBI->isUnconditional() ||
2630  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2631  (!BI->isConditional() &&
2632  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2633  continue;
2634 
2635  // Determine if the two branches share a common destination.
2636  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2637  bool InvertPredCond = false;
2638 
2639  if (BI->isConditional()) {
2640  if (PBI->getSuccessor(0) == TrueDest) {
2641  Opc = Instruction::Or;
2642  } else if (PBI->getSuccessor(1) == FalseDest) {
2643  Opc = Instruction::And;
2644  } else if (PBI->getSuccessor(0) == FalseDest) {
2645  Opc = Instruction::And;
2646  InvertPredCond = true;
2647  } else if (PBI->getSuccessor(1) == TrueDest) {
2648  Opc = Instruction::Or;
2649  InvertPredCond = true;
2650  } else {
2651  continue;
2652  }
2653  } else {
2654  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2655  continue;
2656  }
2657 
2658  DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2659  IRBuilder<> Builder(PBI);
2660 
2661  // If we need to invert the condition in the pred block to match, do so now.
2662  if (InvertPredCond) {
2663  Value *NewCond = PBI->getCondition();
2664 
2665  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2666  CmpInst *CI = cast<CmpInst>(NewCond);
2667  CI->setPredicate(CI->getInversePredicate());
2668  } else {
2669  NewCond =
2670  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2671  }
2672 
2673  PBI->setCondition(NewCond);
2674  PBI->swapSuccessors();
2675  }
2676 
2677  // If we have bonus instructions, clone them into the predecessor block.
2678  // Note that there may be multiple predecessor blocks, so we cannot move
2679  // bonus instructions to a predecessor block.
2680  ValueToValueMapTy VMap; // maps original values to cloned values
2681  // We already make sure Cond is the last instruction before BI. Therefore,
2682  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2683  // instructions.
2684  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2685  if (isa<DbgInfoIntrinsic>(BonusInst))
2686  continue;
2687  Instruction *NewBonusInst = BonusInst->clone();
2688  RemapInstruction(NewBonusInst, VMap,
2690  VMap[&*BonusInst] = NewBonusInst;
2691 
2692  // If we moved a load, we cannot any longer claim any knowledge about
2693  // its potential value. The previous information might have been valid
2694  // only given the branch precondition.
2695  // For an analogous reason, we must also drop all the metadata whose
2696  // semantics we don't understand.
2697  NewBonusInst->dropUnknownNonDebugMetadata();
2698 
2699  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2700  NewBonusInst->takeName(&*BonusInst);
2701  BonusInst->setName(BonusInst->getName() + ".old");
2702  }
2703 
2704  // Clone Cond into the predecessor basic block, and or/and the
2705  // two conditions together.
2706  Instruction *New = Cond->clone();
2707  RemapInstruction(New, VMap,
2709  PredBlock->getInstList().insert(PBI->getIterator(), New);
2710  New->takeName(Cond);
2711  Cond->setName(New->getName() + ".old");
2712 
2713  if (BI->isConditional()) {
2714  Instruction *NewCond = cast<Instruction>(
2715  Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2716  PBI->setCondition(NewCond);
2717 
2718  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2719  bool HasWeights =
2720  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2721  SuccTrueWeight, SuccFalseWeight);
2722  SmallVector<uint64_t, 8> NewWeights;
2723 
2724  if (PBI->getSuccessor(0) == BB) {
2725  if (HasWeights) {
2726  // PBI: br i1 %x, BB, FalseDest
2727  // BI: br i1 %y, TrueDest, FalseDest
2728  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2729  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2730  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2731  // TrueWeight for PBI * FalseWeight for BI.
2732  // We assume that total weights of a BranchInst can fit into 32 bits.
2733  // Therefore, we will not have overflow using 64-bit arithmetic.
2734  NewWeights.push_back(PredFalseWeight *
2735  (SuccFalseWeight + SuccTrueWeight) +
2736  PredTrueWeight * SuccFalseWeight);
2737  }
2738  AddPredecessorToBlock(TrueDest, PredBlock, BB);
2739  PBI->setSuccessor(0, TrueDest);
2740  }
2741  if (PBI->getSuccessor(1) == BB) {
2742  if (HasWeights) {
2743  // PBI: br i1 %x, TrueDest, BB
2744  // BI: br i1 %y, TrueDest, FalseDest
2745  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2746  // FalseWeight for PBI * TrueWeight for BI.
2747  NewWeights.push_back(PredTrueWeight *
2748  (SuccFalseWeight + SuccTrueWeight) +
2749  PredFalseWeight * SuccTrueWeight);
2750  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2751  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2752  }
2753  AddPredecessorToBlock(FalseDest, PredBlock, BB);
2754  PBI->setSuccessor(1, FalseDest);
2755  }
2756  if (NewWeights.size() == 2) {
2757  // Halve the weights if any of them cannot fit in an uint32_t
2758  FitWeights(NewWeights);
2759 
2760  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2761  NewWeights.end());
2762  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2763  } else
2764  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2765  } else {
2766  // Update PHI nodes in the common successors.
2767  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2768  ConstantInt *PBI_C = cast<ConstantInt>(
2769  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2770  assert(PBI_C->getType()->isIntegerTy(1));
2771  Instruction *MergedCond = nullptr;
2772  if (PBI->getSuccessor(0) == TrueDest) {
2773  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2774  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2775  // is false: !PBI_Cond and BI_Value
2776  Instruction *NotCond = cast<Instruction>(
2777  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2778  MergedCond = cast<Instruction>(
2779  Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2780  if (PBI_C->isOne())
2781  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2782  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2783  } else {
2784  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2785  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2786  // is false: PBI_Cond and BI_Value
2787  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2788  Instruction::And, PBI->getCondition(), New, "and.cond"));
2789  if (PBI_C->isOne()) {
2790  Instruction *NotCond = cast<Instruction>(
2791  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2792  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2793  Instruction::Or, NotCond, MergedCond, "or.cond"));
2794  }
2795  }
2796  // Update PHI Node.
2797  PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2798  MergedCond);
2799  }
2800  // Change PBI from Conditional to Unconditional.
2801  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2803  PBI = New_PBI;
2804  }
2805 
2806  // If BI was a loop latch, it may have had associated loop metadata.
2807  // We need to copy it to the new latch, that is, PBI.
2808  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2809  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2810 
2811  // TODO: If BB is reachable from all paths through PredBlock, then we
2812  // could replace PBI's branch probabilities with BI's.
2813 
2814  // Copy any debug value intrinsics into the end of PredBlock.
2815  for (Instruction &I : *BB)
2816  if (isa<DbgInfoIntrinsic>(I))
2817  I.clone()->insertBefore(PBI);
2818 
2819  return true;
2820  }
2821  return false;
2822 }
2823 
2824 // If there is only one store in BB1 and BB2, return it, otherwise return
2825 // nullptr.
2827  StoreInst *S = nullptr;
2828  for (auto *BB : {BB1, BB2}) {
2829  if (!BB)
2830  continue;
2831  for (auto &I : *BB)
2832  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2833  if (S)
2834  // Multiple stores seen.
2835  return nullptr;
2836  else
2837  S = SI;
2838  }
2839  }
2840  return S;
2841 }
2842 
2844  Value *AlternativeV = nullptr) {
2845  // PHI is going to be a PHI node that allows the value V that is defined in
2846  // BB to be referenced in BB's only successor.
2847  //
2848  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2849  // doesn't matter to us what the other operand is (it'll never get used). We
2850  // could just create a new PHI with an undef incoming value, but that could
2851  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2852  // other PHI. So here we directly look for some PHI in BB's successor with V
2853  // as an incoming operand. If we find one, we use it, else we create a new
2854  // one.
2855  //
2856  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2857  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2858  // where OtherBB is the single other predecessor of BB's only successor.
2859  PHINode *PHI = nullptr;
2860  BasicBlock *Succ = BB->getSingleSuccessor();
2861 
2862  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2863  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2864  PHI = cast<PHINode>(I);
2865  if (!AlternativeV)
2866  break;
2867 
2868  assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2869  auto PredI = pred_begin(Succ);
2870  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2871  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2872  break;
2873  PHI = nullptr;
2874  }
2875  if (PHI)
2876  return PHI;
2877 
2878  // If V is not an instruction defined in BB, just return it.
2879  if (!AlternativeV &&
2880  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2881  return V;
2882 
2883  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2884  PHI->addIncoming(V, BB);
2885  for (BasicBlock *PredBB : predecessors(Succ))
2886  if (PredBB != BB)
2887  PHI->addIncoming(
2888  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2889  return PHI;
2890 }
2891 
2893  BasicBlock *QTB, BasicBlock *QFB,
2894  BasicBlock *PostBB, Value *Address,
2895  bool InvertPCond, bool InvertQCond,
2896  const DataLayout &DL) {
2897  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2898  return Operator::getOpcode(&I) == Instruction::BitCast &&
2899  I.getType()->isPointerTy();
2900  };
2901 
2902  // If we're not in aggressive mode, we only optimize if we have some
2903  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2904  auto IsWorthwhile = [&](BasicBlock *BB) {
2905  if (!BB)
2906  return true;
2907  // Heuristic: if the block can be if-converted/phi-folded and the
2908  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2909  // thread this store.
2910  unsigned N = 0;
2911  for (auto &I : *BB) {
2912  // Cheap instructions viable for folding.
2913  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2914  isa<StoreInst>(I))
2915  ++N;
2916  // Free instructions.
2917  else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2918  IsaBitcastOfPointerType(I))
2919  continue;
2920  else
2921  return false;
2922  }
2923  // The store we want to merge is counted in N, so add 1 to make sure
2924  // we're counting the instructions that would be left.
2925  return N <= (PHINodeFoldingThreshold + 1);
2926  };
2927 
2929  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2930  !IsWorthwhile(QFB)))
2931  return false;
2932 
2933  // For every pointer, there must be exactly two stores, one coming from
2934  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2935  // store (to any address) in PTB,PFB or QTB,QFB.
2936  // FIXME: We could relax this restriction with a bit more work and performance
2937  // testing.
2938  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2939  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2940  if (!PStore || !QStore)
2941  return false;
2942 
2943  // Now check the stores are compatible.
2944  if (!QStore->isUnordered() || !PStore->isUnordered())
2945  return false;
2946 
2947  // Check that sinking the store won't cause program behavior changes. Sinking
2948  // the store out of the Q blocks won't change any behavior as we're sinking
2949  // from a block to its unconditional successor. But we're moving a store from
2950  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2951  // So we need to check that there are no aliasing loads or stores in
2952  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2953  // operations between PStore and the end of its parent block.
2954  //
2955  // The ideal way to do this is to query AliasAnalysis, but we don't
2956  // preserve AA currently so that is dangerous. Be super safe and just
2957  // check there are no other memory operations at all.
2958  for (auto &I : *QFB->getSinglePredecessor())
2959  if (I.mayReadOrWriteMemory())
2960  return false;
2961  for (auto &I : *QFB)
2962  if (&I != QStore && I.mayReadOrWriteMemory())
2963  return false;
2964  if (QTB)
2965  for (auto &I : *QTB)
2966  if (&I != QStore && I.mayReadOrWriteMemory())
2967  return false;
2968  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2969  I != E; ++I)
2970  if (&*I != PStore && I->mayReadOrWriteMemory())
2971  return false;
2972 
2973  // OK, we're going to sink the stores to PostBB. The store has to be
2974  // conditional though, so first create the predicate.
2975  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2976  ->getCondition();
2977  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2978  ->getCondition();
2979 
2981  PStore->getParent());
2983  QStore->getParent(), PPHI);
2984 
2985  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2986 
2987  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2988  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2989 
2990  if (InvertPCond)
2991  PPred = QB.CreateNot(PPred);
2992  if (InvertQCond)
2993  QPred = QB.CreateNot(QPred);
2994  Value *CombinedPred = QB.CreateOr(PPred, QPred);
2995 
2996  auto *T =
2997  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2998  QB.SetInsertPoint(T);
2999  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3000  AAMDNodes AAMD;
3001  PStore->getAAMetadata(AAMD, /*Merge=*/false);
3002  PStore->getAAMetadata(AAMD, /*Merge=*/true);
3003  SI->setAAMetadata(AAMD);
3004  unsigned PAlignment = PStore->getAlignment();
3005  unsigned QAlignment = QStore->getAlignment();
3006  unsigned TypeAlignment =
3008  unsigned MinAlignment;
3009  unsigned MaxAlignment;
3010  std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3011  // Choose the minimum alignment. If we could prove both stores execute, we
3012  // could use biggest one. In this case, though, we only know that one of the
3013  // stores executes. And we don't know it's safe to take the alignment from a
3014  // store that doesn't execute.
3015  if (MinAlignment != 0) {
3016  // Choose the minimum of all non-zero alignments.
3017  SI->setAlignment(MinAlignment);
3018  } else if (MaxAlignment != 0) {
3019  // Choose the minimal alignment between the non-zero alignment and the ABI
3020  // default alignment for the type of the stored value.
3021  SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3022  } else {
3023  // If both alignments are zero, use ABI default alignment for the type of
3024  // the stored value.
3025  SI->setAlignment(TypeAlignment);
3026  }
3027 
3028  QStore->eraseFromParent();
3029  PStore->eraseFromParent();
3030 
3031  return true;
3032 }
3033 
3035  const DataLayout &DL) {
3036  // The intention here is to find diamonds or triangles (see below) where each
3037  // conditional block contains a store to the same address. Both of these
3038  // stores are conditional, so they can't be unconditionally sunk. But it may
3039  // be profitable to speculatively sink the stores into one merged store at the
3040  // end, and predicate the merged store on the union of the two conditions of
3041  // PBI and QBI.
3042  //
3043  // This can reduce the number of stores executed if both of the conditions are
3044  // true, and can allow the blocks to become small enough to be if-converted.
3045  // This optimization will also chain, so that ladders of test-and-set
3046  // sequences can be if-converted away.
3047  //
3048  // We only deal with simple diamonds or triangles:
3049  //
3050  // PBI or PBI or a combination of the two
3051  // / \ | \
3052  // PTB PFB | PFB
3053  // \ / | /
3054  // QBI QBI
3055  // / \ | \
3056  // QTB QFB | QFB
3057  // \ / | /
3058  // PostBB PostBB
3059  //
3060  // We model triangles as a type of diamond with a nullptr "true" block.
3061  // Triangles are canonicalized so that the fallthrough edge is represented by
3062  // a true condition, as in the diagram above.
3063  BasicBlock *PTB = PBI->getSuccessor(0);
3064  BasicBlock *PFB = PBI->getSuccessor(1);
3065  BasicBlock *QTB = QBI->getSuccessor(0);
3066  BasicBlock *QFB = QBI->getSuccessor(1);
3067  BasicBlock *PostBB = QFB->getSingleSuccessor();
3068 
3069  // Make sure we have a good guess for PostBB. If QTB's only successor is
3070  // QFB, then QFB is a better PostBB.
3071  if (QTB->getSingleSuccessor() == QFB)
3072  PostBB = QFB;
3073 
3074  // If we couldn't find a good PostBB, stop.
3075  if (!PostBB)
3076  return false;
3077 
3078  bool InvertPCond = false, InvertQCond = false;
3079  // Canonicalize fallthroughs to the true branches.
3080  if (PFB == QBI->getParent()) {
3081  std::swap(PFB, PTB);
3082  InvertPCond = true;
3083  }
3084  if (QFB == PostBB) {
3085  std::swap(QFB, QTB);
3086  InvertQCond = true;
3087  }
3088 
3089  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3090  // and QFB may not. Model fallthroughs as a nullptr block.
3091  if (PTB == QBI->getParent())
3092  PTB = nullptr;
3093  if (QTB == PostBB)
3094  QTB = nullptr;
3095 
3096  // Legality bailouts. We must have at least the non-fallthrough blocks and
3097  // the post-dominating block, and the non-fallthroughs must only have one
3098  // predecessor.
3099  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3100  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3101  };
3102  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3103  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3104  return false;
3105  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3106  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3107  return false;
3108  if (!PostBB->hasNUses(2) || !QBI->getParent()->hasNUses(2))
3109  return false;
3110 
3111  // OK, this is a sequence of two diamonds or triangles.
3112  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3113  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3114  for (auto *BB : {PTB, PFB}) {
3115  if (!BB)
3116  continue;
3117  for (auto &I : *BB)
3118  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3119  PStoreAddresses.insert(SI->getPointerOperand());
3120  }
3121  for (auto *BB : {QTB, QFB}) {
3122  if (!BB)
3123  continue;
3124  for (auto &I : *BB)
3125  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3126  QStoreAddresses.insert(SI->getPointerOperand());
3127  }
3128 
3129  set_intersect(PStoreAddresses, QStoreAddresses);
3130  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3131  // clear what it contains.
3132  auto &CommonAddresses = PStoreAddresses;
3133 
3134  bool Changed = false;
3135  for (auto *Address : CommonAddresses)
3136  Changed |= mergeConditionalStoreToAddress(
3137  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3138  return Changed;
3139 }
3140 
3141 /// If we have a conditional branch as a predecessor of another block,
3142 /// this function tries to simplify it. We know
3143 /// that PBI and BI are both conditional branches, and BI is in one of the
3144 /// successor blocks of PBI - PBI branches to BI.
3146  const DataLayout &DL) {
3147  assert(PBI->isConditional() && BI->isConditional());
3148  BasicBlock *BB = BI->getParent();
3149 
3150  // If this block ends with a branch instruction, and if there is a
3151  // predecessor that ends on a branch of the same condition, make
3152  // this conditional branch redundant.
3153  if (PBI->getCondition() == BI->getCondition() &&
3154  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3155  // Okay, the outcome of this conditional branch is statically
3156  // knowable. If this block had a single pred, handle specially.
3157  if (BB->getSinglePredecessor()) {
3158  // Turn this into a branch on constant.
3159  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3160  BI->setCondition(
3161  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3162  return true; // Nuke the branch on constant.
3163  }
3164 
3165  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3166  // in the constant and simplify the block result. Subsequent passes of
3167  // simplifycfg will thread the block.
3169  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3170  PHINode *NewPN = PHINode::Create(
3171  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3172  BI->getCondition()->getName() + ".pr", &BB->front());
3173  // Okay, we're going to insert the PHI node. Since PBI is not the only
3174  // predecessor, compute the PHI'd conditional value for all of the preds.
3175  // Any predecessor where the condition is not computable we keep symbolic.
3176  for (pred_iterator PI = PB; PI != PE; ++PI) {
3177  BasicBlock *P = *PI;
3178  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3179  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3180  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3181  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3182  NewPN->addIncoming(
3183  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3184  P);
3185  } else {
3186  NewPN->addIncoming(BI->getCondition(), P);
3187  }
3188  }
3189 
3190  BI->setCondition(NewPN);
3191  return true;
3192  }
3193  }
3194 
3195  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3196  if (CE->canTrap())
3197  return false;
3198 
3199  // If both branches are conditional and both contain stores to the same
3200  // address, remove the stores from the conditionals and create a conditional
3201  // merged store at the end.
3202  if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3203  return true;
3204 
3205  // If this is a conditional branch in an empty block, and if any
3206  // predecessors are a conditional branch to one of our destinations,
3207  // fold the conditions into logical ops and one cond br.
3208  BasicBlock::iterator BBI = BB->begin();
3209  // Ignore dbg intrinsics.
3210  while (isa<DbgInfoIntrinsic>(BBI))
3211  ++BBI;
3212  if (&*BBI != BI)
3213  return false;
3214 
3215  int PBIOp, BIOp;
3216  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3217  PBIOp = 0;
3218  BIOp = 0;
3219  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3220  PBIOp = 0;
3221  BIOp = 1;
3222  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3223  PBIOp = 1;
3224  BIOp = 0;
3225  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3226  PBIOp = 1;
3227  BIOp = 1;
3228  } else {
3229  return false;
3230  }
3231 
3232  // Check to make sure that the other destination of this branch
3233  // isn't BB itself. If so, this is an infinite loop that will
3234  // keep getting unwound.
3235  if (PBI->getSuccessor(PBIOp) == BB)
3236  return false;
3237 
3238  // Do not perform this transformation if it would require
3239  // insertion of a large number of select instructions. For targets
3240  // without predication/cmovs, this is a big pessimization.
3241 
3242  // Also do not perform this transformation if any phi node in the common
3243  // destination block can trap when reached by BB or PBB (PR17073). In that
3244  // case, it would be unsafe to hoist the operation into a select instruction.
3245 
3246  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3247  unsigned NumPhis = 0;
3248  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3249  ++II, ++NumPhis) {
3250  if (NumPhis > 2) // Disable this xform.
3251  return false;
3252 
3253  PHINode *PN = cast<PHINode>(II);
3254  Value *BIV = PN->getIncomingValueForBlock(BB);
3255  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3256  if (CE->canTrap())
3257  return false;
3258 
3259  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3260  Value *PBIV = PN->getIncomingValue(PBBIdx);
3261  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3262  if (CE->canTrap())
3263  return false;
3264  }
3265 
3266  // Finally, if everything is ok, fold the branches to logical ops.
3267  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3268 
3269  DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3270  << "AND: " << *BI->getParent());
3271 
3272  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3273  // branch in it, where one edge (OtherDest) goes back to itself but the other
3274  // exits. We don't *know* that the program avoids the infinite loop
3275  // (even though that seems likely). If we do this xform naively, we'll end up
3276  // recursively unpeeling the loop. Since we know that (after the xform is
3277  // done) that the block *is* infinite if reached, we just make it an obviously
3278  // infinite loop with no cond branch.
3279  if (OtherDest == BB) {
3280  // Insert it at the end of the function, because it's either code,
3281  // or it won't matter if it's hot. :)
3282  BasicBlock *InfLoopBlock =
3283  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3284  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3285  OtherDest = InfLoopBlock;
3286  }
3287 
3288  DEBUG(dbgs() << *PBI->getParent()->getParent());
3289 
3290  // BI may have other predecessors. Because of this, we leave
3291  // it alone, but modify PBI.
3292 
3293  // Make sure we get to CommonDest on True&True directions.
3294  Value *PBICond = PBI->getCondition();
3295  IRBuilder<NoFolder> Builder(PBI);
3296  if (PBIOp)
3297  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3298 
3299  Value *BICond = BI->getCondition();
3300  if (BIOp)
3301  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3302 
3303  // Merge the conditions.
3304  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3305 
3306  // Modify PBI to branch on the new condition to the new dests.
3307  PBI->setCondition(Cond);
3308  PBI->setSuccessor(0, CommonDest);
3309  PBI->setSuccessor(1, OtherDest);
3310 
3311  // Update branch weight for PBI.
3312  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3313  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3314  bool HasWeights =
3315  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3316  SuccTrueWeight, SuccFalseWeight);
3317  if (HasWeights) {
3318  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3319  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3320  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3321  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3322  // The weight to CommonDest should be PredCommon * SuccTotal +
3323  // PredOther * SuccCommon.
3324  // The weight to OtherDest should be PredOther * SuccOther.
3325  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3326  PredOther * SuccCommon,
3327  PredOther * SuccOther};
3328  // Halve the weights if any of them cannot fit in an uint32_t
3329  FitWeights(NewWeights);
3330 
3331  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3332  }
3333 
3334  // OtherDest may have phi nodes. If so, add an entry from PBI's
3335  // block that are identical to the entries for BI's block.
3336  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3337 
3338  // We know that the CommonDest already had an edge from PBI to
3339  // it. If it has PHIs though, the PHIs may have different
3340  // entries for BB and PBI's BB. If so, insert a select to make
3341  // them agree.
3342  PHINode *PN;
3343  for (BasicBlock::iterator II = CommonDest->begin();
3344  (PN = dyn_cast<PHINode>(II)); ++II) {
3345  Value *BIV = PN->getIncomingValueForBlock(BB);
3346  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3347  Value *PBIV = PN->getIncomingValue(PBBIdx);
3348  if (BIV != PBIV) {
3349  // Insert a select in PBI to pick the right value.
3350  SelectInst *NV = cast<SelectInst>(
3351  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3352  PN->setIncomingValue(PBBIdx, NV);
3353  // Although the select has the same condition as PBI, the original branch
3354  // weights for PBI do not apply to the new select because the select's
3355  // 'logical' edges are incoming edges of the phi that is eliminated, not
3356  // the outgoing edges of PBI.
3357  if (HasWeights) {
3358  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3359  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3360  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3361  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3362  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3363  // The weight to PredOtherDest should be PredOther * SuccCommon.
3364  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3365  PredOther * SuccCommon};
3366 
3367  FitWeights(NewWeights);
3368 
3369  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3370  }
3371  }
3372  }
3373 
3374  DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3375  DEBUG(dbgs() << *PBI->getParent()->getParent());
3376 
3377  // This basic block is probably dead. We know it has at least
3378  // one fewer predecessor.
3379  return true;
3380 }
3381 
3382 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3383 // true or to FalseBB if Cond is false.
3384 // Takes care of updating the successors and removing the old terminator.
3385 // Also makes sure not to introduce new successors by assuming that edges to
3386 // non-successor TrueBBs and FalseBBs aren't reachable.
3388  BasicBlock *TrueBB, BasicBlock *FalseBB,
3389  uint32_t TrueWeight,
3390  uint32_t FalseWeight) {
3391  // Remove any superfluous successor edges from the CFG.
3392  // First, figure out which successors to preserve.
3393  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3394  // successor.
3395  BasicBlock *KeepEdge1 = TrueBB;
3396  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3397 
3398  // Then remove the rest.
3399  for (BasicBlock *Succ : OldTerm->successors()) {
3400  // Make sure only to keep exactly one copy of each edge.
3401  if (Succ == KeepEdge1)
3402  KeepEdge1 = nullptr;
3403  else if (Succ == KeepEdge2)
3404  KeepEdge2 = nullptr;
3405  else
3406  Succ->removePredecessor(OldTerm->getParent(),
3407  /*DontDeleteUselessPHIs=*/true);
3408  }
3409 
3410  IRBuilder<> Builder(OldTerm);
3411  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3412 
3413  // Insert an appropriate new terminator.
3414  if (!KeepEdge1 && !KeepEdge2) {
3415  if (TrueBB == FalseBB)
3416  // We were only looking for one successor, and it was present.
3417  // Create an unconditional branch to it.
3418  Builder.CreateBr(TrueBB);
3419  else {
3420  // We found both of the successors we were looking for.
3421  // Create a conditional branch sharing the condition of the select.
3422  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3423  if (TrueWeight != FalseWeight)
3424  setBranchWeights(NewBI, TrueWeight, FalseWeight);
3425  }
3426  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3427  // Neither of the selected blocks were successors, so this
3428  // terminator must be unreachable.
3429  new UnreachableInst(OldTerm->getContext(), OldTerm);
3430  } else {
3431  // One of the selected values was a successor, but the other wasn't.
3432  // Insert an unconditional branch to the one that was found;
3433  // the edge to the one that wasn't must be unreachable.
3434  if (!KeepEdge1)
3435  // Only TrueBB was found.
3436  Builder.CreateBr(TrueBB);
3437  else
3438  // Only FalseBB was found.
3439  Builder.CreateBr(FalseBB);
3440  }
3441 
3443  return true;
3444 }
3445 
3446 // Replaces
3447 // (switch (select cond, X, Y)) on constant X, Y
3448 // with a branch - conditional if X and Y lead to distinct BBs,
3449 // unconditional otherwise.
3451  // Check for constant integer values in the select.
3452  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3453  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3454  if (!TrueVal || !FalseVal)
3455  return false;
3456 
3457  // Find the relevant condition and destinations.
3458  Value *Condition = Select->getCondition();
3459  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3460  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3461 
3462  // Get weight for TrueBB and FalseBB.
3463  uint32_t TrueWeight = 0, FalseWeight = 0;
3464  SmallVector<uint64_t, 8> Weights;
3465  bool HasWeights = HasBranchWeights(SI);
3466  if (HasWeights) {
3467  GetBranchWeights(SI, Weights);
3468  if (Weights.size() == 1 + SI->getNumCases()) {
3469  TrueWeight =
3470  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3471  FalseWeight =
3472  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3473  }
3474  }
3475 
3476  // Perform the actual simplification.
3477  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3478  FalseWeight);
3479 }
3480 
3481 // Replaces
3482 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3483 // blockaddress(@fn, BlockB)))
3484 // with
3485 // (br cond, BlockA, BlockB).
3487  // Check that both operands of the select are block addresses.
3490  if (!TBA || !FBA)
3491  return false;
3492 
3493  // Extract the actual blocks.
3494  BasicBlock *TrueBB = TBA->getBasicBlock();
3495  BasicBlock *FalseBB = FBA->getBasicBlock();
3496 
3497  // Perform the actual simplification.
3498  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3499  0);
3500 }
3501 
3502 /// This is called when we find an icmp instruction
3503 /// (a seteq/setne with a constant) as the only instruction in a
3504 /// block that ends with an uncond branch. We are looking for a very specific
3505 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3506 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3507 /// default value goes to an uncond block with a seteq in it, we get something
3508 /// like:
3509 ///
3510 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3511 /// DEFAULT:
3512 /// %tmp = icmp eq i8 %A, 92
3513 /// br label %end
3514 /// end:
3515 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3516 ///
3517 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3518 /// the PHI, merging the third icmp into the switch.
3520  ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3521  const TargetTransformInfo &TTI, const SimplifyCFGOptions &Options) {
3522  BasicBlock *BB = ICI->getParent();
3523 
3524  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3525  // complex.
3526  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3527  return false;
3528 
3529  Value *V = ICI->getOperand(0);
3530  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3531 
3532  // The pattern we're looking for is where our only predecessor is a switch on
3533  // 'V' and this block is the default case for the switch. In this case we can
3534  // fold the compared value into the switch to simplify things.
3535  BasicBlock *Pred = BB->getSinglePredecessor();
3536  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3537  return false;
3538 
3539  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3540  if (SI->getCondition() != V)
3541  return false;
3542 
3543  // If BB is reachable on a non-default case, then we simply know the value of
3544  // V in this block. Substitute it and constant fold the icmp instruction
3545  // away.
3546  if (SI->getDefaultDest() != BB) {
3547  ConstantInt *VVal = SI->findCaseDest(BB);
3548  assert(VVal && "Should have a unique destination value");
3549  ICI->setOperand(0, VVal);
3550 
3551  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3552  ICI->replaceAllUsesWith(V);
3553  ICI->eraseFromParent();
3554  }
3555  // BB is now empty, so it is likely to simplify away.
3556  return simplifyCFG(BB, TTI, Options) | true;
3557  }
3558 
3559  // Ok, the block is reachable from the default dest. If the constant we're
3560  // comparing exists in one of the other edges, then we can constant fold ICI
3561  // and zap it.
3562  if (SI->findCaseValue(Cst) != SI->case_default()) {
3563  Value *V;
3564  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3565  V = ConstantInt::getFalse(BB->getContext());
3566  else
3567  V = ConstantInt::getTrue(BB->getContext());
3568 
3569  ICI->replaceAllUsesWith(V);
3570  ICI->eraseFromParent();
3571  // BB is now empty, so it is likely to simplify away.
3572  return simplifyCFG(BB, TTI, Options) | true;
3573  }
3574 
3575  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3576  // the block.
3577  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3578  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3579  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3580  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3581  return false;
3582 
3583  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3584  // true in the PHI.
3585  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3586  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3587 
3588  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3589  std::swap(DefaultCst, NewCst);
3590 
3591  // Replace ICI (which is used by the PHI for the default value) with true or
3592  // false depending on if it is EQ or NE.
3593  ICI->replaceAllUsesWith(DefaultCst);
3594  ICI->eraseFromParent();
3595 
3596  // Okay, the switch goes to this block on a default value. Add an edge from
3597  // the switch to the merge point on the compared value.
3598  BasicBlock *NewBB =
3599  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3600  SmallVector<uint64_t, 8> Weights;
3601  bool HasWeights = HasBranchWeights(SI);
3602  if (HasWeights) {
3603  GetBranchWeights(SI, Weights);
3604  if (Weights.size() == 1 + SI->getNumCases()) {
3605  // Split weight for default case to case for "Cst".
3606  Weights[0] = (Weights[0] + 1) >> 1;
3607  Weights.push_back(Weights[0]);
3608 
3609  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3610  setBranchWeights(SI, MDWeights);
3611  }
3612  }
3613  SI->addCase(Cst, NewBB);
3614 
3615  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3616  Builder.SetInsertPoint(NewBB);
3617  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3618  Builder.CreateBr(SuccBlock);
3619  PHIUse->addIncoming(NewCst, NewBB);
3620  return true;
3621 }
3622 
3623 /// The specified branch is a conditional branch.
3624 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3625 /// fold it into a switch instruction if so.
3627  const DataLayout &DL) {
3628  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3629  if (!Cond)
3630  return false;
3631 
3632  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3633  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3634  // 'setne's and'ed together, collect them.
3635 
3636  // Try to gather values from a chain of and/or to be turned into a switch
3637  ConstantComparesGatherer ConstantCompare(Cond, DL);
3638  // Unpack the result
3639  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3640  Value *CompVal = ConstantCompare.CompValue;
3641  unsigned UsedICmps = ConstantCompare.UsedICmps;
3642  Value *ExtraCase = ConstantCompare.Extra;
3643 
3644  // If we didn't have a multiply compared value, fail.
3645  if (!CompVal)
3646  return false;
3647 
3648  // Avoid turning single icmps into a switch.
3649  if (UsedICmps <= 1)
3650  return false;
3651 
3652  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3653 
3654  // There might be duplicate constants in the list, which the switch
3655  // instruction can't handle, remove them now.
3656  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3657  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3658 
3659  // If Extra was used, we require at least two switch values to do the
3660  // transformation. A switch with one value is just a conditional branch.
3661  if (ExtraCase && Values.size() < 2)
3662  return false;
3663 
3664  // TODO: Preserve branch weight metadata, similarly to how
3665  // FoldValueComparisonIntoPredecessors preserves it.
3666 
3667  // Figure out which block is which destination.
3668  BasicBlock *DefaultBB = BI->getSuccessor(1);
3669  BasicBlock *EdgeBB = BI->getSuccessor(0);
3670  if (!TrueWhenEqual)
3671  std::swap(DefaultBB, EdgeBB);
3672 
3673  BasicBlock *BB = BI->getParent();
3674 
3675  DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3676  << " cases into SWITCH. BB is:\n"
3677  << *BB);
3678 
3679  // If there are any extra values that couldn't be folded into the switch
3680  // then we evaluate them with an explicit branch first. Split the block
3681  // right before the condbr to handle it.
3682  if (ExtraCase) {
3683  BasicBlock *NewBB =
3684  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3685  // Remove the uncond branch added to the old block.
3686  TerminatorInst *OldTI = BB->getTerminator();
3687  Builder.SetInsertPoint(OldTI);
3688 
3689  if (TrueWhenEqual)
3690  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3691  else
3692  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3693 
3694  OldTI->eraseFromParent();
3695 
3696  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3697  // for the edge we just added.
3698  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3699 
3700  DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3701  << "\nEXTRABB = " << *BB);
3702  BB = NewBB;
3703  }
3704 
3705  Builder.SetInsertPoint(BI);
3706  // Convert pointer to int before we switch.
3707  if (CompVal->getType()->isPointerTy()) {
3708  CompVal = Builder.CreatePtrToInt(
3709  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3710  }
3711 
3712  // Create the new switch instruction now.
3713  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3714 
3715  // Add all of the 'cases' to the switch instruction.
3716  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3717  New->addCase(Values[i], EdgeBB);
3718 
3719  // We added edges from PI to the EdgeBB. As such, if there were any
3720  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3721  // the number of edges added.
3722  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3723  PHINode *PN = cast<PHINode>(BBI);
3724  Value *InVal = PN->getIncomingValueForBlock(BB);
3725  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3726  PN->addIncoming(InVal, BB);
3727  }
3728 
3729  // Erase the old branch instruction.
3731 
3732  DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3733  return true;
3734 }
3735 
3736 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3737  if (isa<PHINode>(RI->getValue()))
3738  return SimplifyCommonResume(RI);
3739  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3740  RI->getValue() == RI->getParent()->getFirstNonPHI())
3741  // The resume must unwind the exception that caused control to branch here.
3742  return SimplifySingleResume(RI);
3743 
3744  return false;
3745 }
3746 
3747 // Simplify resume that is shared by several landing pads (phi of landing pad).
3748 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3749  BasicBlock *BB = RI->getParent();
3750 
3751  // Check that there are no other instructions except for debug intrinsics
3752  // between the phi of landing pads (RI->getValue()) and resume instruction.
3753  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3754  E = RI->getIterator();
3755  while (++I != E)
3756  if (!isa<DbgInfoIntrinsic>(I))
3757  return false;
3758 
3759  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3760  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3761 
3762  // Check incoming blocks to see if any of them are trivial.
3763  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3764  Idx++) {
3765  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3766  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3767 
3768  // If the block has other successors, we can not delete it because
3769  // it has other dependents.
3770  if (IncomingBB->getUniqueSuccessor() != BB)
3771  continue;
3772 
3773  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3774  // Not the landing pad that caused the control to branch here.
3775  if (IncomingValue != LandingPad)
3776  continue;
3777 
3778  bool isTrivial = true;
3779 
3780  I = IncomingBB->getFirstNonPHI()->getIterator();
3781  E = IncomingBB->getTerminator()->getIterator();
3782  while (++I != E)
3783  if (!isa<DbgInfoIntrinsic>(I)) {
3784  isTrivial = false;
3785  break;
3786  }
3787 
3788  if (isTrivial)
3789  TrivialUnwindBlocks.insert(IncomingBB);
3790  }
3791 
3792  // If no trivial unwind blocks, don't do any simplifications.
3793  if (TrivialUnwindBlocks.empty())
3794  return false;
3795 
3796  // Turn all invokes that unwind here into calls.
3797  for (auto *TrivialBB : TrivialUnwindBlocks) {
3798  // Blocks that will be simplified should be removed from the phi node.
3799  // Note there could be multiple edges to the resume block, and we need
3800  // to remove them all.
3801  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3802  BB->removePredecessor(TrivialBB, true);
3803 
3804  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3805  PI != PE;) {
3806  BasicBlock *Pred = *PI++;
3807  removeUnwindEdge(Pred);
3808  }
3809 
3810  // In each SimplifyCFG run, only the current processed block can be erased.
3811  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3812  // of erasing TrivialBB, we only remove the branch to the common resume
3813  // block so that we can later erase the resume block since it has no
3814  // predecessors.
3815  TrivialBB->getTerminator()->eraseFromParent();
3816  new UnreachableInst(RI->getContext(), TrivialBB);
3817  }
3818 
3819  // Delete the resume block if all its predecessors have been removed.
3820  if (pred_empty(BB))
3821  BB->eraseFromParent();
3822 
3823  return !TrivialUnwindBlocks.empty();
3824 }
3825 
3826 // Simplify resume that is only used by a single (non-phi) landing pad.
3827 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3828  BasicBlock *BB = RI->getParent();
3830  assert(RI->getValue() == LPInst &&
3831  "Resume must unwind the exception that caused control to here");
3832 
3833  // Check that there are no other instructions except for debug intrinsics.
3834  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3835  while (++I != E)
3836  if (!isa<DbgInfoIntrinsic>(I))
3837  return false;
3838 
3839  // Turn all invokes that unwind here into calls and delete the basic block.
3840  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3841  BasicBlock *Pred = *PI++;
3842  removeUnwindEdge(Pred);
3843  }
3844 
3845  // The landingpad is now unreachable. Zap it.
3846  BB->eraseFromParent();
3847  if (LoopHeaders)
3848  LoopHeaders->erase(BB);
3849  return true;
3850 }
3851 
3853  // If this is a trivial cleanup pad that executes no instructions, it can be
3854  // eliminated. If the cleanup pad continues to the caller, any predecessor
3855  // that is an EH pad will be updated to continue to the caller and any
3856  // predecessor that terminates with an invoke instruction will have its invoke
3857  // instruction converted to a call instruction. If the cleanup pad being
3858  // simplified does not continue to the caller, each predecessor will be
3859  // updated to continue to the unwind destination of the cleanup pad being
3860  // simplified.
3861  BasicBlock *BB = RI->getParent();
3862  CleanupPadInst *CPInst = RI->getCleanupPad();
3863  if (CPInst->getParent() != BB)
3864  // This isn't an empty cleanup.
3865  return false;
3866 
3867  // We cannot kill the pad if it has multiple uses. This typically arises
3868  // from unreachable basic blocks.
3869  if (!CPInst->hasOneUse())
3870  return false;
3871 
3872  // Check that there are no other instructions except for benign intrinsics.
3873  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3874  while (++I != E) {
3875  auto *II = dyn_cast<IntrinsicInst>(I);
3876  if (!II)
3877  return false;
3878 
3879  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3880  switch (IntrinsicID) {
3881  case Intrinsic::dbg_declare:
3882  case Intrinsic::dbg_value:
3883  case Intrinsic::lifetime_end:
3884  break;
3885  default:
3886  return false;
3887  }
3888  }
3889 
3890  // If the cleanup return we are simplifying unwinds to the caller, this will
3891  // set UnwindDest to nullptr.
3892  BasicBlock *UnwindDest = RI->getUnwindDest();
3893  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3894 
3895  // We're about to remove BB from the control flow. Before we do, sink any
3896  // PHINodes into the unwind destination. Doing this before changing the
3897  // control flow avoids some potentially slow checks, since we can currently
3898  // be certain that UnwindDest and BB have no common predecessors (since they
3899  // are both EH pads).
3900  if (UnwindDest) {
3901  // First, go through the PHI nodes in UnwindDest and update any nodes that
3902  // reference the block we are removing
3903  for (BasicBlock::iterator I = UnwindDest->begin(),
3904  IE = DestEHPad->getIterator();
3905  I != IE; ++I) {
3906  PHINode *DestPN = cast<PHINode>(I);
3907 
3908  int Idx = DestPN->getBasicBlockIndex(BB);
3909  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3910  assert(Idx != -1);
3911  // This PHI node has an incoming value that corresponds to a control
3912  // path through the cleanup pad we are removing. If the incoming
3913  // value is in the cleanup pad, it must be a PHINode (because we
3914  // verified above that the block is otherwise empty). Otherwise, the
3915  // value is either a constant or a value that dominates the cleanup
3916  // pad being removed.
3917  //
3918  // Because BB and UnwindDest are both EH pads, all of their
3919  // predecessors must unwind to these blocks, and since no instruction
3920  // can have multiple unwind destinations, there will be no overlap in
3921  // incoming blocks between SrcPN and DestPN.
3922  Value *SrcVal = DestPN->getIncomingValue(Idx);
3923  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3924 
3925  // Remove the entry for the block we are deleting.
3926  DestPN->removeIncomingValue(Idx, false);
3927 
3928  if (SrcPN && SrcPN->getParent() == BB) {
3929  // If the incoming value was a PHI node in the cleanup pad we are
3930  // removing, we need to merge that PHI node's incoming values into
3931  // DestPN.
3932  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3933  SrcIdx != SrcE; ++SrcIdx) {
3934  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3935  SrcPN->getIncomingBlock(SrcIdx));
3936  }
3937  } else {
3938  // Otherwise, the incoming value came from above BB and
3939  // so we can just reuse it. We must associate all of BB's
3940  // predecessors with this value.
3941  for (auto *pred : predecessors(BB)) {
3942  DestPN->addIncoming(SrcVal, pred);
3943  }
3944  }
3945  }
3946 
3947  // Sink any remaining PHI nodes directly into UnwindDest.
3948  Instruction *InsertPt = DestEHPad;
3949  for (BasicBlock::iterator I = BB->begin(),
3950  IE = BB->getFirstNonPHI()->getIterator();
3951  I != IE;) {
3952  // The iterator must be incremented here because the instructions are
3953  // being moved to another block.
3954  PHINode *PN = cast<PHINode>(I++);
3955  if (PN->use_empty())
3956  // If the PHI node has no uses, just leave it. It will be erased
3957  // when we erase BB below.
3958  continue;
3959 
3960  // Otherwise, sink this PHI node into UnwindDest.
3961  // Any predecessors to UnwindDest which are not already represented
3962  // must be back edges which inherit the value from the path through
3963  // BB. In this case, the PHI value must reference itself.
3964  for (auto *pred : predecessors(UnwindDest))
3965  if (pred != BB)
3966  PN->addIncoming(PN, pred);
3967  PN->moveBefore(InsertPt);
3968  }
3969  }
3970 
3971  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3972  // The iterator must be updated here because we are removing this pred.
3973  BasicBlock *PredBB = *PI++;
3974  if (UnwindDest == nullptr) {
3975  removeUnwindEdge(PredBB);
3976  } else {
3977  TerminatorInst *TI = PredBB->getTerminator();
3978  TI->replaceUsesOfWith(BB, UnwindDest);
3979  }
3980  }
3981 
3982  // The cleanup pad is now unreachable. Zap it.
3983  BB->eraseFromParent();
3984  return true;
3985 }
3986 
3987 // Try to merge two cleanuppads together.
3989  // Skip any cleanuprets which unwind to caller, there is nothing to merge
3990  // with.
3991  BasicBlock *UnwindDest = RI->getUnwindDest();
3992  if (!UnwindDest)
3993  return false;
3994 
3995  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3996  // be safe to merge without code duplication.
3997  if (UnwindDest->getSinglePredecessor() != RI->getParent())
3998  return false;
3999 
4000  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4001  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4002  if (!SuccessorCleanupPad)
4003  return false;
4004 
4005  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4006  // Replace any uses of the successor cleanupad with the predecessor pad
4007  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4008  // funclet bundle operands.
4009  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4010  // Remove the old cleanuppad.
4011  SuccessorCleanupPad->eraseFromParent();
4012  // Now, we simply replace the cleanupret with a branch to the unwind
4013  // destination.
4014  BranchInst::Create(UnwindDest, RI->getParent());
4015  RI->eraseFromParent();
4016 
4017  return true;
4018 }
4019 
4020 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4021  // It is possible to transiantly have an undef cleanuppad operand because we
4022  // have deleted some, but not all, dead blocks.
4023  // Eventually, this block will be deleted.
4024  if (isa<UndefValue>(RI->getOperand(0)))
4025  return false;
4026 
4027  if (mergeCleanupPad(RI))
4028  return true;
4029 
4030  if (removeEmptyCleanup(RI))
4031  return true;
4032 
4033  return false;
4034 }
4035 
4036 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4037  BasicBlock *BB = RI->getParent();
4038  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4039  return false;
4040 
4041  // Find predecessors that end with branches.
4042  SmallVector<BasicBlock *, 8> UncondBranchPreds;
4043  SmallVector<BranchInst *, 8> CondBranchPreds;
4044  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4045  BasicBlock *P = *PI;
4046  TerminatorInst *PTI = P->getTerminator();
4047  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4048  if (BI->isUnconditional())
4049  UncondBranchPreds.push_back(P);
4050  else
4051  CondBranchPreds.push_back(BI);
4052  }
4053  }
4054 
4055  // If we found some, do the transformation!
4056  if (!UncondBranchPreds.empty() && DupRet) {
4057  while (!UncondBranchPreds.empty()) {
4058  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4059  DEBUG(dbgs() << "FOLDING: " << *BB
4060  << "INTO UNCOND BRANCH PRED: " << *Pred);
4061  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4062  }
4063 
4064  // If we eliminated all predecessors of the block, delete the block now.
4065  if (pred_empty(BB)) {
4066  // We know there are no successors, so just nuke the block.
4067  BB->eraseFromParent();
4068  if (LoopHeaders)
4069  LoopHeaders->erase(BB);
4070  }
4071 
4072  return true;
4073  }
4074 
4075  // Check out all of the conditional branches going to this return
4076  // instruction. If any of them just select between returns, change the
4077  // branch itself into a select/return pair.
4078  while (!CondBranchPreds.empty()) {
4079  BranchInst *BI = CondBranchPreds.pop_back_val();
4080 
4081  // Check to see if the non-BB successor is also a return block.
4082  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4083  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4084  SimplifyCondBranchToTwoReturns(BI, Builder))
4085  return true;
4086  }
4087  return false;
4088 }
4089 
4090 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4091  BasicBlock *BB = UI->getParent();
4092 
4093  bool Changed = false;
4094 
4095  // If there are any instructions immediately before the unreachable that can
4096  // be removed, do so.
4097  while (UI->getIterator() != BB->begin()) {
4098  BasicBlock::iterator BBI = UI->getIterator();
4099  --BBI;
4100  // Do not delete instructions that can have side effects which might cause
4101  // the unreachable to not be reachable; specifically, calls and volatile
4102  // operations may have this effect.
4103  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4104  break;
4105 
4106  if (BBI->mayHaveSideEffects()) {
4107  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4108  if (SI->isVolatile())
4109  break;
4110  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4111  if (LI->isVolatile())
4112  break;
4113  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4114  if (RMWI->isVolatile())
4115  break;
4116  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4117  if (CXI->isVolatile())
4118  break;
4119  } else if (isa<CatchPadInst>(BBI)) {
4120  // A catchpad may invoke exception object constructors and such, which
4121  // in some languages can be arbitrary code, so be conservative by
4122  // default.
4123  // For CoreCLR, it just involves a type test, so can be removed.
4126  break;
4127  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4128  !isa<LandingPadInst>(BBI)) {
4129  break;
4130  }
4131  // Note that deleting LandingPad's here is in fact okay, although it
4132  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4133  // all the predecessors of this block will be the unwind edges of Invokes,
4134  // and we can therefore guarantee this block will be erased.
4135  }
4136 
4137  // Delete this instruction (any uses are guaranteed to be dead)
4138  if (!BBI->use_empty())
4139  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4140  BBI->eraseFromParent();
4141  Changed = true;
4142  }
4143 
4144  // If the unreachable instruction is the first in the block, take a gander
4145  // at all of the predecessors of this instruction, and simplify them.
4146  if (&BB->front() != UI)
4147  return Changed;
4148 
4150  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4151  TerminatorInst *TI = Preds[i]->getTerminator();
4152  IRBuilder<> Builder(TI);
4153  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4154  if (BI->isUnconditional()) {
4155  if (BI->getSuccessor(0) == BB) {
4156  new UnreachableInst(TI->getContext(), TI);
4157  TI->eraseFromParent();
4158  Changed = true;
4159  }
4160  } else {
4161  if (BI->getSuccessor(0) == BB) {
4162  Builder.CreateBr(BI->getSuccessor(1));
4164  } else if (BI->getSuccessor(1) == BB) {
4165  Builder.CreateBr(BI->getSuccessor(0));
4167  Changed = true;
4168  }
4169  }
4170  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4171  for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4172  if (i->getCaseSuccessor() != BB) {
4173  ++i;
4174  continue;
4175  }
4176  BB->removePredecessor(SI->getParent());
4177  i = SI->removeCase(i);
4178  e = SI->case_end();
4179  Changed = true;
4180  }
4181  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4182  if (II->getUnwindDest() == BB) {
4183  removeUnwindEdge(TI->getParent());
4184  Changed = true;
4185  }
4186  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4187  if (CSI->getUnwindDest() == BB) {
4188  removeUnwindEdge(TI->getParent());
4189  Changed = true;
4190  continue;
4191  }
4192 
4193  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4194  E = CSI->handler_end();
4195  I != E; ++I) {
4196  if (*I == BB) {
4197  CSI->removeHandler(I);
4198  --I;
4199  --E;
4200  Changed = true;
4201  }
4202  }
4203  if (CSI->getNumHandlers() == 0) {
4204  BasicBlock *CatchSwitchBB = CSI->getParent();
4205  if (CSI->hasUnwindDest()) {
4206  // Redirect preds to the unwind dest
4207  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4208  } else {
4209  // Rewrite all preds to unwind to caller (or from invoke to call).
4210  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4211  for (BasicBlock *EHPred : EHPreds)
4212  removeUnwindEdge(EHPred);
4213  }
4214  // The catchswitch is no longer reachable.
4215  new UnreachableInst(CSI->getContext(), CSI);
4216  CSI->eraseFromParent();
4217  Changed = true;
4218  }
4219  } else if (isa<CleanupReturnInst>(TI)) {
4220  new UnreachableInst(TI->getContext(), TI);
4221  TI->eraseFromParent();
4222  Changed = true;
4223  }
4224  }
4225 
4226  // If this block is now dead, remove it.
4227  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4228  // We know there are no successors, so just nuke the block.
4229  BB->eraseFromParent();
4230  if (LoopHeaders)
4231  LoopHeaders->erase(BB);
4232  return true;
4233  }
4234 
4235  return Changed;
4236 }
4237 
4239  assert(Cases.size() >= 1);
4240 
4241  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4242  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4243  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4244  return false;
4245  }
4246  return true;
4247 }
4248 
4249 /// Turn a switch with two reachable destinations into an integer range
4250 /// comparison and branch.
4252  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4253 
4254  bool HasDefault =
4255  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4256 
4257  // Partition the cases into two sets with different destinations.
4258  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4259  BasicBlock *DestB = nullptr;
4262 
4263  for (auto Case : SI->cases()) {
4264  BasicBlock *Dest = Case.getCaseSuccessor();
4265  if (!DestA)
4266  DestA = Dest;
4267  if (Dest == DestA) {
4268  CasesA.push_back(Case.getCaseValue());
4269  continue;
4270  }
4271  if (!DestB)
4272  DestB = Dest;
4273  if (Dest == DestB) {
4274  CasesB.push_back(Case.getCaseValue());
4275  continue;
4276  }
4277  return false; // More than two destinations.
4278  }
4279 
4280  assert(DestA && DestB &&
4281  "Single-destination switch should have been folded.");
4282  assert(DestA != DestB);
4283  assert(DestB != SI->getDefaultDest());
4284  assert(!CasesB.empty() && "There must be non-default cases.");
4285  assert(!CasesA.empty() || HasDefault);
4286 
4287  // Figure out if one of the sets of cases form a contiguous range.
4288  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4289  BasicBlock *ContiguousDest = nullptr;
4290  BasicBlock *OtherDest = nullptr;
4291  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4292  ContiguousCases = &CasesA;
4293  ContiguousDest = DestA;
4294  OtherDest = DestB;
4295  } else if (CasesAreContiguous(CasesB)) {
4296  ContiguousCases = &CasesB;
4297  ContiguousDest = DestB;
4298  OtherDest = DestA;
4299  } else
4300  return false;
4301 
4302  // Start building the compare and branch.
4303 
4304  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4305  Constant *NumCases =
4306  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4307 
4308  Value *Sub = SI->getCondition();
4309  if (!Offset->isNullValue())
4310  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4311 
4312  Value *Cmp;
4313  // If NumCases overflowed, then all possible values jump to the successor.
4314  if (NumCases->isNullValue() && !ContiguousCases->empty())
4315  Cmp = ConstantInt::getTrue(SI->getContext());
4316  else
4317  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4318  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4319 
4320  // Update weight for the newly-created conditional branch.
4321  if (HasBranchWeights(SI)) {
4322  SmallVector<uint64_t, 8> Weights;
4323  GetBranchWeights(SI, Weights);
4324  if (Weights.size() == 1 + SI->getNumCases()) {
4325  uint64_t TrueWeight = 0;
4326  uint64_t FalseWeight = 0;
4327  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4328  if (SI->getSuccessor(I) == ContiguousDest)
4329  TrueWeight += Weights[I];
4330  else
4331  FalseWeight += Weights[I];
4332  }
4333  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4334  TrueWeight /= 2;
4335  FalseWeight /= 2;
4336  }
4337  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4338  }
4339  }
4340 
4341  // Prune obsolete incoming values off the successors' PHI nodes.
4342  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4343  unsigned PreviousEdges = ContiguousCases->size();
4344  if (ContiguousDest == SI->getDefaultDest())
4345  ++PreviousEdges;
4346  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4347  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4348  }
4349  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4350  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4351  if (OtherDest == SI->getDefaultDest())
4352  ++PreviousEdges;
4353  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4354  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4355  }
4356 
4357  // Drop the switch.
4358  SI->eraseFromParent();
4359 
4360  return true;
4361 }
4362 
4363 /// Compute masked bits for the condition of a switch
4364 /// and use it to remove dead cases.
4366  const DataLayout &DL) {
4367  Value *Cond = SI->getCondition();
4368  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4369  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4370 
4371  // We can also eliminate cases by determining that their values are outside of
4372  // the limited range of the condition based on how many significant (non-sign)
4373  // bits are in the condition value.
4374  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4375  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4376 
4377  // Gather dead cases.
4379  for (auto &Case : SI->cases()) {
4380  const APInt &CaseVal = Case.getCaseValue()->getValue();
4381  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4382  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4383  DeadCases.push_back(Case.getCaseValue());
4384  DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4385  }
4386  }
4387 
4388  // If we can prove that the cases must cover all possible values, the
4389  // default destination becomes dead and we can remove it. If we know some
4390  // of the bits in the value, we can use that to more precisely compute the
4391  // number of possible unique case values.
4392  bool HasDefault =
4393  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4394  const unsigned NumUnknownBits =
4395  Bits - (Known.Zero | Known.One).countPopulation();
4396  assert(NumUnknownBits <= Bits);
4397  if (HasDefault && DeadCases.empty() &&
4398  NumUnknownBits < 64 /* avoid overflow */ &&
4399  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4400  DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4401  BasicBlock *NewDefault =
4403  SI->setDefaultDest(&*NewDefault);
4404  SplitBlock(&*NewDefault, &NewDefault->front());
4405  auto *OldTI = NewDefault->getTerminator();
4406  new UnreachableInst(SI->getContext(), OldTI);
4408  return true;
4409  }
4410 
4411  SmallVector<uint64_t, 8> Weights;
4412  bool HasWeight = HasBranchWeights(SI);
4413  if (HasWeight) {
4414  GetBranchWeights(SI, Weights);
4415  HasWeight = (Weights.size() == 1 + SI->getNumCases());
4416  }
4417 
4418  // Remove dead cases from the switch.
4419  for (ConstantInt *DeadCase : DeadCases) {
4420  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4421  assert(CaseI != SI->case_default() &&
4422  "Case was not found. Probably mistake in DeadCases forming.");
4423  if (HasWeight) {
4424  std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4425  Weights.pop_back();
4426  }
4427 
4428  // Prune unused values from PHI nodes.
4429  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4430  SI->removeCase(CaseI);
4431  }
4432  if (HasWeight && Weights.size() >= 2) {
4433  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4434  setBranchWeights(SI, MDWeights);
4435  }
4436 
4437  return !DeadCases.empty();
4438 }
4439 
4440 /// If BB would be eligible for simplification by
4441 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4442 /// by an unconditional branch), look at the phi node for BB in the successor
4443 /// block and see if the incoming value is equal to CaseValue. If so, return
4444 /// the phi node, and set PhiIndex to BB's index in the phi node.
4446  BasicBlock *BB, int *PhiIndex) {
4447  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4448  return nullptr; // BB must be empty to be a candidate for simplification.
4449  if (!BB->getSinglePredecessor())
4450  return nullptr; // BB must be dominated by the switch.
4451 
4453  if (!Branch || !Branch->isUnconditional())
4454  return nullptr; // Terminator must be unconditional branch.
4455 
4456  BasicBlock *Succ = Branch->getSuccessor(0);
4457 
4458  BasicBlock::iterator I = Succ->begin();
4459  while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4460  int Idx = PHI->getBasicBlockIndex(BB);
4461  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4462 
4463  Value *InValue = PHI->getIncomingValue(Idx);
4464  if (InValue != CaseValue)
4465  continue;
4466 
4467  *PhiIndex = Idx;
4468  return PHI;
4469  }
4470 
4471  return nullptr;
4472 }
4473 
4474 /// Try to forward the condition of a switch instruction to a phi node
4475 /// dominated by the switch, if that would mean that some of the destination
4476 /// blocks of the switch can be folded away. Return true if a change is made.
4478  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4479 
4480  ForwardingNodesMap ForwardingNodes;
4481  BasicBlock *SwitchBlock = SI->getParent();
4482  bool Changed = false;
4483  for (auto &Case : SI->cases()) {
4484  ConstantInt *CaseValue = Case.getCaseValue();
4485  BasicBlock *CaseDest = Case.getCaseSuccessor();
4486 
4487  // Replace phi operands in successor blocks that are using the constant case
4488  // value rather than the switch condition variable:
4489  // switchbb:
4490  // switch i32 %x, label %default [
4491  // i32 17, label %succ
4492  // ...
4493  // succ:
4494  // %r = phi i32 ... [ 17, %switchbb ] ...
4495  // -->
4496  // %r = phi i32 ... [ %x, %switchbb ] ...
4497 
4498  for (Instruction &InstInCaseDest : *CaseDest) {
4499  auto *Phi = dyn_cast<PHINode>(&InstInCaseDest);
4500  if (!Phi) break;
4501 
4502  // This only works if there is exactly 1 incoming edge from the switch to
4503  // a phi. If there is >1, that means multiple cases of the switch map to 1
4504  // value in the phi, and that phi value is not the switch condition. Thus,
4505  // this transform would not make sense (the phi would be invalid because
4506  // a phi can't have different incoming values from the same block).
4507  int SwitchBBIdx = Phi->getBasicBlockIndex(SwitchBlock);
4508  if (Phi->getIncomingValue(SwitchBBIdx) == CaseValue &&
4509  count(Phi->blocks(), SwitchBlock) == 1) {
4510  Phi->setIncomingValue(SwitchBBIdx, SI->getCondition());
4511  Changed = true;
4512  }
4513  }
4514 
4515  // Collect phi nodes that are indirectly using this switch's case constants.
4516  int PhiIdx;
4517  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4518  ForwardingNodes[Phi].push_back(PhiIdx);
4519  }
4520 
4521  for (auto &ForwardingNode : ForwardingNodes) {
4522  PHINode *Phi = ForwardingNode.first;
4523  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4524  if (Indexes.size() < 2)
4525  continue;
4526 
4527  for (int Index : Indexes)
4528  Phi->setIncomingValue(Index, SI->getCondition());
4529  Changed = true;
4530  }
4531 
4532  return Changed;
4533 }
4534 
4535 /// Return true if the backend will be able to handle
4536 /// initializing an array of constants like C.
4538  if (C->isThreadDependent())
4539  return false;
4540  if (C->isDLLImportDependent())
4541  return false;
4542 
4543  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4544  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4545  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4546  return false;
4547 
4548  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4549  if (!CE->isGEPWithNoNotionalOverIndexing())
4550  return false;
4551  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4552  return false;
4553  }
4554 
4556  return false;
4557 
4558  return true;
4559 }
4560 
4561 /// If V is a Constant, return it. Otherwise, try to look up
4562 /// its constant value in ConstantPool, returning 0 if it's not there.
4563 static Constant *
4566  if (Constant *C = dyn_cast<Constant>(V))
4567  return C;
4568  return ConstantPool.lookup(V);
4569 }
4570 
4571 /// Try to fold instruction I into a constant. This works for
4572 /// simple instructions such as binary operations where both operands are
4573 /// constant or can be replaced by constants from the ConstantPool. Returns the
4574 /// resulting constant on success, 0 otherwise.
4575 static Constant *
4578  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4579  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4580  if (!A)
4581  return nullptr;
4582  if (A->isAllOnesValue())
4583  return LookupConstant(Select->getTrueValue(), ConstantPool);
4584  if (A->isNullValue())
4585  return LookupConstant(Select->getFalseValue(), ConstantPool);
4586  return nullptr;
4587  }
4588 
4590  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4592  COps.push_back(A);
4593  else
4594  return nullptr;
4595  }
4596 
4597  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4598  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4599  COps[1], DL);
4600  }
4601 
4602  return ConstantFoldInstOperands(I, COps, DL);
4603 }
4604 
4605 /// Try to determine the resulting constant values in phi nodes
4606 /// at the common destination basic block, *CommonDest, for one of the case
4607 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4608 /// case), of a switch instruction SI.
4609 static bool
4611  BasicBlock **CommonDest,
4612  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4613  const DataLayout &DL, const TargetTransformInfo &TTI) {
4614  // The block from which we enter the common destination.
4615  BasicBlock *Pred = SI->getParent();
4616 
4617  // If CaseDest is empty except for some side-effect free instructions through
4618  // which we can constant-propagate the CaseVal, continue to its successor.
4620  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4621  for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4622  ++I) {
4623  if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4624  // If the terminator is a simple branch, continue to the next block.
4625  if (T->getNumSuccessors() != 1 || T->isExceptional())
4626  return false;
4627  Pred = CaseDest;
4628  CaseDest = T->getSuccessor(0);
4629  } else if (isa<DbgInfoIntrinsic>(I)) {
4630  // Skip debug intrinsic.
4631  continue;
4632  } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4633  // Instruction is side-effect free and constant.
4634 
4635  // If the instruction has uses outside this block or a phi node slot for
4636  // the block, it is not safe to bypass the instruction since it would then
4637  // no longer dominate all its uses.
4638  for (auto &Use : I->uses()) {
4639  User *User = Use.getUser();
4640  if (Instruction *I = dyn_cast<Instruction>(User))
4641  if (I->getParent() == CaseDest)
4642  continue;
4643  if (PHINode *Phi = dyn_cast<PHINode>(User))
4644  if (Phi->getIncomingBlock(Use) == CaseDest)
4645  continue;
4646  return false;
4647  }
4648 
4649  ConstantPool.insert(std::make_pair(&*I, C));
4650  } else {
4651  break;
4652  }
4653  }
4654 
4655  // If we did not have a CommonDest before, use the current one.
4656  if (!*CommonDest)
4657  *CommonDest = CaseDest;
4658  // If the destination isn't the common one, abort.
4659  if (CaseDest != *CommonDest)
4660  return false;
4661 
4662  // Get the values for this case from phi nodes in the destination block.
4663  BasicBlock::iterator I = (*CommonDest)->begin();
4664  while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4665  int Idx = PHI->getBasicBlockIndex(Pred);
4666  if (Idx == -1)
4667  continue;
4668 
4669  Constant *ConstVal =
4670  LookupConstant(PHI->getIncomingValue(Idx), ConstantPool);
4671  if (!ConstVal)
4672  return false;
4673 
4674  // Be conservative about which kinds of constants we support.
4675  if (!ValidLookupTableConstant(ConstVal, TTI))
4676  return false;
4677 
4678  Res.push_back(std::make_pair(PHI, ConstVal));
4679  }
4680 
4681  return Res.size() > 0;
4682 }
4683 
4684 // Helper function used to add CaseVal to the list of cases that generate
4685 // Result.
4686 static void MapCaseToResult(ConstantInt *CaseVal,
4687  SwitchCaseResultVectorTy &UniqueResults,
4688  Constant *Result) {
4689  for (auto &I : UniqueResults) {
4690  if (I.first == Result) {
4691  I.second.push_back(CaseVal);
4692  return;
4693  }
4694  }
4695  UniqueResults.push_back(
4696  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4697 }
4698 
4699 // Helper function that initializes a map containing
4700 // results for the PHI node of the common destination block for a switch
4701 // instruction. Returns false if multiple PHI nodes have been found or if
4702 // there is not a common destination block for the switch.
4704  BasicBlock *&CommonDest,
4705  SwitchCaseResultVectorTy &UniqueResults,
4706  Constant *&DefaultResult,
4707  const DataLayout &DL,
4708  const TargetTransformInfo &TTI) {
4709  for (auto &I : SI->cases()) {
4710  ConstantInt *CaseVal = I.getCaseValue();
4711 
4712  // Resulting value at phi nodes for this case value.
4713  SwitchCaseResultsTy Results;
4714  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4715  DL, TTI))
4716  return false;
4717 
4718  // Only one value per case is permitted
4719  if (Results.size() > 1)
4720  return false;
4721  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4722 
4723  // Check the PHI consistency.
4724  if (!PHI)
4725  PHI = Results[0].first;
4726  else if (PHI != Results[0].first)
4727  return false;
4728  }
4729  // Find the default result value.
4731  BasicBlock *DefaultDest = SI->getDefaultDest();
4732  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4733  DL, TTI);
4734  // If the default value is not found abort unless the default destination
4735  // is unreachable.
4736  DefaultResult =
4737  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4738  if ((!DefaultResult &&
4739  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4740  return false;
4741 
4742  return true;
4743 }
4744 
4745 // Helper function that checks if it is possible to transform a switch with only
4746 // two cases (or two cases + default) that produces a result into a select.
4747 // Example:
4748 // switch (a) {
4749 // case 10: %0 = icmp eq i32 %a, 10
4750 // return 10; %1 = select i1 %0, i32 10, i32 4
4751 // case 20: ----> %2 = icmp eq i32 %a, 20
4752 // return 2; %3 = select i1 %2, i32 2, i32 %1
4753 // default:
4754 // return 4;
4755 // }
4756 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4757  Constant *DefaultResult, Value *Condition,
4758  IRBuilder<> &Builder) {
4759  assert(ResultVector.size() == 2 &&
4760  "We should have exactly two unique results at this point");
4761  // If we are selecting between only two cases transform into a simple
4762  // select or a two-way select if default is possible.
4763  if (ResultVector[0].second.size() == 1 &&
4764  ResultVector[1].second.size() == 1) {
4765  ConstantInt *const FirstCase = ResultVector[0].second[0];
4766  ConstantInt *const SecondCase = ResultVector[1].second[0];
4767 
4768  bool DefaultCanTrigger = DefaultResult;
4769  Value *SelectValue = ResultVector[1].first;
4770  if (DefaultCanTrigger) {
4771  Value *const ValueCompare =
4772  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4773  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4774  DefaultResult, "switch.select");
4775  }
4776  Value *const ValueCompare =
4777  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4778  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4779  SelectValue, "switch.select");
4780  }
4781 
4782  return nullptr;
4783 }
4784 
4785 // Helper function to cleanup a switch instruction that has been converted into
4786 // a select, fixing up PHI nodes and basic blocks.
4788  Value *SelectValue,
4789  IRBuilder<> &Builder) {
4790  BasicBlock *SelectBB = SI->getParent();
4791  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4792  PHI->removeIncomingValue(SelectBB);
4793  PHI->addIncoming(SelectValue, SelectBB);
4794 
4795  Builder.CreateBr(PHI->getParent());
4796 
4797  // Remove the switch.
4798  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4799  BasicBlock *Succ = SI->getSuccessor(i);
4800 
4801  if (Succ == PHI->getParent())
4802  continue;
4803  Succ->removePredecessor(SelectBB);
4804  }
4805  SI->eraseFromParent();
4806 }
4807 
4808 /// If the switch is only used to initialize one or more
4809 /// phi nodes in a common successor block with only two different
4810 /// constant values, replace the switch with select.
4811 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4812  const DataLayout &DL,
4813  const TargetTransformInfo &TTI) {
4814  Value *const Cond = SI->getCondition();
4815  PHINode *PHI = nullptr;
4816  BasicBlock *CommonDest = nullptr;
4817  Constant *DefaultResult;
4818  SwitchCaseResultVectorTy UniqueResults;
4819  // Collect all the cases that will deliver the same value from the switch.
4820  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4821  DL, TTI))
4822  return false;
4823  // Selects choose between maximum two values.
4824  if (UniqueResults.size() != 2)
4825  return false;
4826  assert(PHI != nullptr && "PHI for value select not found");
4827 
4828  Builder.SetInsertPoint(SI);
4829  Value *SelectValue =
4830  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4831  if (SelectValue) {
4832  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4833  return true;
4834  }
4835  // The switch couldn't be converted into a select.
4836  return false;
4837 }
4838 
4839 namespace {
4840 
4841 /// This class represents a lookup table that can be used to replace a switch.
4842 class SwitchLookupTable {
4843 public:
4844  /// Create a lookup table to use as a switch replacement with the contents
4845  /// of Values, using DefaultValue to fill any holes in the table.
4846  SwitchLookupTable(
4847  Module &M, uint64_t TableSize, ConstantInt *Offset,
4848  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4849  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4850 
4851  /// Build instructions with Builder to retrieve the value at
4852  /// the position given by Index in the lookup table.
4853  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4854 
4855  /// Return true if a table with TableSize elements of
4856  /// type ElementType would fit in a target-legal register.
4857  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4858  Type *ElementType);
4859 
4860 private:
4861  // Depending on the contents of the table, it can be represented in
4862  // different ways.
4863  enum {
4864  // For tables where each element contains the same value, we just have to
4865  // store that single value and return it for each lookup.
4866  SingleValueKind,
4867 
4868  // For tables where there is a linear relationship between table index
4869  // and values. We calculate the result with a simple multiplication
4870  // and addition instead of a table lookup.
4871  LinearMapKind,
4872 
4873  // For small tables with integer elements, we can pack them into a bitmap
4874  // that fits into a target-legal register. Values are retrieved by
4875  // shift and mask operations.
4876  BitMapKind,
4877 
4878  // The table is stored as an array of values. Values are retrieved by load
4879  // instructions from the table.
4880  ArrayKind
4881  } Kind;
4882 
4883  // For SingleValueKind, this is the single value.
4884  Constant *SingleValue = nullptr;
4885 
4886  // For BitMapKind, this is the bitmap.
4887  ConstantInt *BitMap = nullptr;
4888  IntegerType *BitMapElementTy = nullptr;
4889 
4890  // For LinearMapKind, these are the constants used to derive the value.
4891  ConstantInt *LinearOffset = nullptr;
4892  ConstantInt *LinearMultiplier = nullptr;
4893 
4894  // For ArrayKind, this is the array.
4895  GlobalVariable *Array = nullptr;
4896 };
4897 
4898 } // end anonymous namespace
4899 
4900 SwitchLookupTable::SwitchLookupTable(
4901  Module &M, uint64_t TableSize, ConstantInt *Offset,
4902  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4903  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4904  assert(Values.size() && "Can't build lookup table without values!");
4905  assert(TableSize >= Values.size() && "Can't fit values in table!");
4906 
4907  // If all values in the table are equal, this is that value.
4908  SingleValue = Values.begin()->second;
4909 
4910  Type *ValueType = Values.begin()->second->getType();
4911 
4912  // Build up the table contents.
4913  SmallVector<Constant *, 64> TableContents(TableSize);
4914  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4915  ConstantInt *CaseVal = Values[I].first;
4916  Constant *CaseRes = Values[I].second;
4917  assert(CaseRes->getType() == ValueType);
4918 
4919  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4920  TableContents[Idx] = CaseRes;
4921 
4922  if (CaseRes != SingleValue)
4923  SingleValue = nullptr;
4924  }
4925 
4926  // Fill in any holes in the table with the default result.
4927  if (Values.size() < TableSize) {
4928  assert(DefaultValue &&
4929  "Need a default value to fill the lookup table holes.");
4930  assert(DefaultValue->getType() == ValueType);
4931  for (uint64_t I = 0; I < TableSize; ++I) {
4932  if (!TableContents[I])
4933  TableContents[I] = DefaultValue;
4934  }
4935 
4936  if (DefaultValue != SingleValue)
4937  SingleValue = nullptr;
4938  }
4939 
4940  // If each element in the table contains the same value, we only need to store
4941  // that single value.
4942  if (SingleValue) {
4943  Kind = SingleValueKind;
4944  return;
4945  }
4946 
4947  // Check if we can derive the value with a linear transformation from the
4948  // table index.
4949  if (isa<IntegerType>(ValueType)) {
4950  bool LinearMappingPossible = true;
4951  APInt PrevVal;
4952  APInt DistToPrev;
4953  assert(TableSize >= 2 && "Should be a SingleValue table.");
4954  // Check if there is the same distance between two consecutive values.
4955  for (uint64_t I = 0; I < TableSize; ++I) {
4956  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4957  if (!ConstVal) {
4958  // This is an undef. We could deal with it, but undefs in lookup tables
4959  // are very seldom. It's probably not worth the additional complexity.
4960  LinearMappingPossible = false;
4961  break;
4962  }
4963  const APInt &Val = ConstVal->getValue();
4964  if (I != 0) {
4965  APInt Dist = Val - PrevVal;
4966  if (I == 1) {
4967  DistToPrev = Dist;
4968  } else if (Dist != DistToPrev) {
4969  LinearMappingPossible = false;
4970  break;
4971  }
4972  }
4973  PrevVal = Val;
4974  }
4975  if (LinearMappingPossible) {
4976  LinearOffset = cast<ConstantInt>(TableContents[0]);
4977  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4978  Kind = LinearMapKind;
4979  ++NumLinearMaps;
4980  return;
4981  }
4982  }
4983 
4984  // If the type is integer and the table fits in a register, build a bitmap.
4985  if (WouldFitInRegister(DL, TableSize, ValueType)) {
4986  IntegerType *IT = cast<IntegerType>(ValueType);
4987  APInt TableInt(TableSize * IT->getBitWidth(), 0);
4988  for (uint64_t I = TableSize; I > 0; --I) {
4989  TableInt <<= IT->getBitWidth();
4990  // Insert values into the bitmap. Undef values are set to zero.
4991  if (!isa<UndefValue>(TableContents[I - 1])) {
4992  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4993  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4994  }
4995  }
4996  BitMap = ConstantInt::get(M.getContext(), TableInt);
4997  BitMapElementTy = IT;
4998  Kind = BitMapKind;
4999  ++NumBitMaps;
5000  return;
5001  }
5002 
5003  // Store the table in an array.
5004  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5005  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5006 
5007  Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
5008  GlobalVariable::PrivateLinkage, Initializer,
5009  "switch.table." + FuncName);
5010  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5011  Kind = ArrayKind;
5012 }
5013 
5014 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5015  switch (Kind) {
5016  case SingleValueKind:
5017  return SingleValue;
5018  case LinearMapKind: {
5019  // Derive the result value from the input value.
5020  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5021  false, "switch.idx.cast");
5022  if (!LinearMultiplier->isOne())
5023  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5024  if (!LinearOffset->isZero())
5025  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5026  return Result;
5027  }
5028  case BitMapKind: {
5029  // Type of the bitmap (e.g. i59).
5030  IntegerType *MapTy = BitMap->getType();
5031 
5032  // Cast Index to the same type as the bitmap.
5033  // Note: The Index is <= the number of elements in the table, so
5034  // truncating it to the width of the bitmask is safe.
5035  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5036 
5037  // Multiply the shift amount by the element width.
5038  ShiftAmt = Builder.CreateMul(
5039  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5040  "switch.shiftamt");
5041 
5042  // Shift down.
5043  Value *DownShifted =
5044  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5045  // Mask off.
5046  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5047  }
5048  case ArrayKind: {
5049  // Make sure the table index will not overflow when treated as signed.
5050  IntegerType *IT = cast<IntegerType>(Index->getType());
5051  uint64_t TableSize =
5052  Array->getInitializer()->getType()->getArrayNumElements();
5053  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5054  Index = Builder.CreateZExt(
5055  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5056  "switch.tableidx.zext");
5057 
5058  Value *GEPIndices[] = {Builder.getInt32(0), Index};
5059  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5060  GEPIndices, "switch.gep");
5061  return Builder.CreateLoad(GEP, "switch.load");
5062  }
5063  }
5064  llvm_unreachable("Unknown lookup table kind!");
5065 }
5066 
5067 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5068  uint64_t TableSize,
5069  Type *ElementType) {
5070  auto *IT = dyn_cast<IntegerType>(ElementType);
5071  if (!IT)
5072  return false;
5073  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5074  // are <= 15, we could try to narrow the type.
5075 
5076  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5077  if (TableSize >= UINT_MAX / IT->getBitWidth())
5078  return false;
5079  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5080 }
5081 
5082 /// Determine whether a lookup table should be built for this switch, based on
5083 /// the number of cases, size of the table, and the types of the results.
5084 static bool
5085 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5086  const TargetTransformInfo &TTI, const DataLayout &DL,
5087  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5088  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5089  return false; // TableSize overflowed, or mul below might overflow.
5090 
5091  bool AllTablesFitInRegister = true;
5092  bool HasIllegalType = false;
5093  for (const auto &I : ResultTypes) {
5094  Type *Ty = I.second;
5095 
5096  // Saturate this flag to true.
5097  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5098 
5099  // Saturate this flag to false.
5100  AllTablesFitInRegister =
5101  AllTablesFitInRegister &&
5102  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5103 
5104  // If both flags saturate, we're done. NOTE: This *only* works with
5105  // saturating flags, and all flags have to saturate first due to the
5106  // non-deterministic behavior of iterating over a dense map.
5107  if (HasIllegalType && !AllTablesFitInRegister)
5108  break;
5109  }
5110 
5111  // If each table would fit in a register, we should build it anyway.
5112  if (AllTablesFitInRegister)
5113  return true;
5114 
5115  // Don't build a table that doesn't fit in-register if it has illegal types.
5116  if (HasIllegalType)
5117  return false;
5118 
5119  // The table density should be at least 40%. This is the same criterion as for
5120  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5121  // FIXME: Find the best cut-off.
5122  return SI->getNumCases() * 10 >= TableSize * 4;
5123 }
5124 
5125 /// Try to reuse the switch table index compare. Following pattern:
5126 /// \code
5127 /// if (idx < tablesize)
5128 /// r = table[idx]; // table does not contain default_value
5129 /// else
5130 /// r = default_value;
5131 /// if (r != default_value)
5132 /// ...
5133 /// \endcode
5134 /// Is optimized to:
5135 /// \code
5136 /// cond = idx < tablesize;
5137 /// if (cond)
5138 /// r = table[idx];
5139 /// else
5140 /// r = default_value;
5141 /// if (cond)
5142 /// ...
5143 /// \endcode
5144 /// Jump threading will then eliminate the second if(cond).
5145 static void reuseTableCompare(
5146  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5147  Constant *DefaultValue,
5148  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5149  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5150  if (!CmpInst)
5151  return;
5152 
5153  // We require that the compare is in the same block as the phi so that jump
5154  // threading can do its work afterwards.
5155  if (CmpInst->getParent() != PhiBlock)
5156  return;
5157 
5158  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5159  if (!CmpOp1)
5160  return;
5161 
5162  Value *RangeCmp = RangeCheckBranch->getCondition();
5163  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5164  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5165 
5166  // Check if the compare with the default value is constant true or false.
5167  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5168  DefaultValue, CmpOp1, true);
5169  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5170  return;
5171 
5172  // Check if the compare with the case values is distinct from the default
5173  // compare result.
5174  for (auto ValuePair : Values) {
5175  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5176  ValuePair.second, CmpOp1, true);
5177  if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5178  return;
5179  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5180  "Expect true or false as compare result.");
5181  }
5182 
5183  // Check if the branch instruction dominates the phi node. It's a simple
5184  // dominance check, but sufficient for our needs.
5185  // Although this check is invariant in the calling loops, it's better to do it
5186  // at this late stage. Practically we do it at most once for a switch.
5187  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5188  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5189  BasicBlock *Pred = *PI;
5190  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5191  return;
5192  }
5193 
5194  if (DefaultConst == FalseConst) {
5195  // The compare yields the same result. We can replace it.
5196  CmpInst->replaceAllUsesWith(RangeCmp);
5197  ++NumTableCmpReuses;
5198  } else {
5199  // The compare yields the same result, just inverted. We can replace it.
5200  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5201  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5202  RangeCheckBranch);
5203  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5204  ++NumTableCmpReuses;
5205  }
5206 }
5207 
5208 /// If the switch is only used to initialize one or more phi nodes in a common
5209 /// successor block with different constant values, replace the switch with
5210 /// lookup tables.
5212  const DataLayout &DL,
5213  const TargetTransformInfo &TTI) {
5214  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5215 
5216  Function *Fn = SI->getParent()->getParent();
5217  // Only build lookup table when we have a target that supports it or the
5218  // attribute is not set.
5219  if (!TTI.shouldBuildLookupTables() ||
5220  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5221  return false;
5222 
5223  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5224  // split off a dense part and build a lookup table for that.
5225 
5226  // FIXME: This creates arrays of GEPs to constant strings, which means each
5227  // GEP needs a runtime relocation in PIC code. We should just build one big
5228  // string and lookup indices into that.
5229 
5230  // Ignore switches with less than three cases. Lookup tables will not make
5231  // them faster, so we don't analyze them.
5232  if (SI->getNumCases() < 3)
5233  return false;
5234 
5235  // Figure out the corresponding result for each case value and phi node in the
5236  // common destination, as well as the min and max case values.
5237  assert(SI->case_begin() != SI->case_end());
5238  SwitchInst::CaseIt CI = SI->case_begin();
5239  ConstantInt *MinCaseVal = CI->getCaseValue();
5240  ConstantInt *MaxCaseVal = CI->getCaseValue();
5241 
5242  BasicBlock *CommonDest = nullptr;
5243 
5244  using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5246 
5247  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5250 
5251  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5252  ConstantInt *CaseVal = CI->getCaseValue();
5253  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5254  MinCaseVal = CaseVal;
5255  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5256  MaxCaseVal = CaseVal;
5257 
5258  // Resulting value at phi nodes for this case value.
5259  using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5260  ResultsTy Results;
5261  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5262  Results, DL, TTI))
5263  return false;
5264 
5265  // Append the result from this case to the list for each phi.
5266  for (const auto &I : Results) {
5267  PHINode *PHI = I.first;
5268  Constant *Value = I.second;
5269  if (!ResultLists.count(PHI))
5270  PHIs.push_back(PHI);
5271  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5272  }
5273  }
5274 
5275  // Keep track of the result types.
5276  for (PHINode *PHI : PHIs) {
5277  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5278  }
5279 
5280  uint64_t NumResults = ResultLists[PHIs[0]].size();
5281  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5282  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5283  bool TableHasHoles = (NumResults < TableSize);
5284 
5285  // If the table has holes, we need a constant result for the default case
5286  // or a bitmask that fits in a register.
5287  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5288  bool HasDefaultResults =
5289  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5290  DefaultResultsList, DL, TTI);
5291 
5292  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5293  if (NeedMask) {
5294  // As an extra penalty for the validity test we require more cases.
5295  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5296  return false;
5297  if (!DL.fitsInLegalInteger(TableSize))
5298  return false;
5299  }
5300 
5301  for (const auto &I : DefaultResultsList) {
5302  PHINode *PHI = I.first;
5303  Constant *Result = I.second;
5304  DefaultResults[PHI] = Result;
5305  }
5306 
5307  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5308  return false;
5309 
5310  // Create the BB that does the lookups.
5311  Module &Mod = *CommonDest->getParent()->getParent();
5312  BasicBlock *LookupBB = BasicBlock::Create(
5313  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5314 
5315  // Compute the table index value.
5316  Builder.SetInsertPoint(SI);
5317  Value *TableIndex;
5318  if (MinCaseVal->isNullValue())
5319  TableIndex = SI->getCondition();
5320  else
5321  TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5322  "switch.tableidx");
5323 
5324  // Compute the maximum table size representable by the integer type we are
5325  // switching upon.
5326  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5327  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5328  assert(MaxTableSize >= TableSize &&
5329  "It is impossible for a switch to have more entries than the max "
5330  "representable value of its input integer type's size.");
5331 
5332  // If the default destination is unreachable, or if the lookup table covers
5333  // all values of the conditional variable, branch directly to the lookup table
5334  // BB. Otherwise, check that the condition is within the case range.
5335  const bool DefaultIsReachable =
5336  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5337  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5338  BranchInst *RangeCheckBranch = nullptr;
5339 
5340  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5341  Builder.CreateBr(LookupBB);
5342  // Note: We call removeProdecessor later since we need to be able to get the
5343  // PHI value for the default case in case we're using a bit mask.
5344  } else {
5345  Value *Cmp = Builder.CreateICmpULT(
5346  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5347  RangeCheckBranch =
5348  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5349  }
5350 
5351  // Populate the BB that does the lookups.
5352  Builder.SetInsertPoint(LookupBB);
5353 
5354  if (NeedMask) {
5355  // Before doing the lookup, we do the hole check. The LookupBB is therefore
5356  // re-purposed to do the hole check, and we create a new LookupBB.
5357  BasicBlock *MaskBB = LookupBB;
5358  MaskBB->setName("switch.hole_check");
5359  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5360  CommonDest->getParent(), CommonDest);
5361 
5362  // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5363  // unnecessary illegal types.
5364  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5365  APInt MaskInt(TableSizePowOf2, 0);
5366  APInt One(TableSizePowOf2, 1);
5367  // Build bitmask; fill in a 1 bit for every case.
5368  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5369  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5370  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5371  .getLimitedValue();
5372  MaskInt |= One << Idx;
5373  }
5374  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5375 
5376  // Get the TableIndex'th bit of the bitmask.
5377  // If this bit is 0 (meaning hole) jump to the default destination,
5378  // else continue with table lookup.
5379  IntegerType *MapTy = TableMask->getType();
5380  Value *MaskIndex =
5381  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5382  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5383  Value *LoBit = Builder.CreateTrunc(
5384  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5385  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5386 
5387  Builder.SetInsertPoint(LookupBB);
5388  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5389  }
5390 
5391  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5392  // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5393  // do not delete PHINodes here.
5395  /*DontDeleteUselessPHIs=*/true);
5396  }
5397 
5398  bool ReturnedEarly = false;
5399  for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5400  PHINode *PHI = PHIs[I];
5401  const ResultListTy &ResultList = ResultLists[PHI];
5402 
5403  // If using a bitmask, use any value to fill the lookup table holes.
5404  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5405  StringRef FuncName = Fn->getName();
5406  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5407  FuncName);
5408 
5409  Value *Result = Table.BuildLookup(TableIndex, Builder);
5410 
5411  // If the result is used to return immediately from the function, we want to
5412  // do that right here.
5413  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5414  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5415  Builder.CreateRet(Result);
5416  ReturnedEarly = true;
5417  break;
5418  }
5419 
5420  // Do a small peephole optimization: re-use the switch table compare if
5421  // possible.
5422  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5423  BasicBlock *PhiBlock = PHI->getParent();
5424  // Search for compare instructions which use the phi.
5425  for (auto *User : PHI->users()) {
5426  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5427  }
5428  }
5429 
5430  PHI->addIncoming(Result, LookupBB);
5431  }
5432 
5433  if (!ReturnedEarly)
5434  Builder.CreateBr(CommonDest);
5435 
5436  // Remove the switch.
5437  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5438  BasicBlock *Succ = SI->getSuccessor(i);
5439 
5440  if (Succ == SI->getDefaultDest())
5441  continue;
5442  Succ->removePredecessor(SI->getParent());
5443  }
5444  SI->eraseFromParent();
5445 
5446  ++NumLookupTables;
5447  if (NeedMask)
5448  ++NumLookupTablesHoles;
5449  return true;
5450 }
5451 
5452 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5453  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5454  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5455  uint64_t Range = Diff + 1;
5456  uint64_t NumCases = Values.size();
5457  // 40% is the default density for building a jump table in optsize/minsize mode.
5458  uint64_t MinDensity = 40;
5459 
5460  return NumCases * 100 >= Range * MinDensity;
5461 }
5462 
5463 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5464 /// of cases.
5465 ///
5466 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5467 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5468 ///
5469 /// This converts a sparse switch into a dense switch which allows better
5470 /// lowering and could also allow transforming into a lookup table.
5472  const DataLayout &DL,
5473  const TargetTransformInfo &TTI) {
5474  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5475  if (CondTy->getIntegerBitWidth() > 64 ||
5476  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5477  return false;
5478  // Only bother with this optimization if there are more than 3 switch cases;
5479  // SDAG will only bother creating jump tables for 4 or more cases.
5480  if (SI->getNumCases() < 4)
5481  return false;
5482 
5483  // This transform is agnostic to the signedness of the input or case values. We
5484  // can treat the case values as signed or unsigned. We can optimize more common
5485  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5486  // as signed.
5487  SmallVector<int64_t,4> Values;
5488  for (auto &C : SI->cases())
5489  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5490  std::sort(Values.begin(), Values.end());
5491 
5492  // If the switch is already dense, there's nothing useful to do here.
5493  if (isSwitchDense(Values))
5494  return false;
5495 
5496  // First, transform the values such that they start at zero and ascend.
5497  int64_t Base = Values[0];
5498  for (auto &V : Values)
5499  V -= (uint64_t)(Base);
5500 
5501  // Now we have signed numbers that have been shifted so that, given enough
5502  // precision, there are no negative values. Since the rest of the transform
5503  // is bitwise only, we switch now to an unsigned representation.
5504  uint64_t GCD = 0;
5505  for (auto &V : Values)
5506  GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5507 
5508  // This transform can be done speculatively because it is so cheap - it results
5509  // in a single rotate operation being inserted. This can only happen if the
5510  // factor extracted is a power of 2.
5511  // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5512  // inverse of GCD and then perform this transform.
5513  // FIXME: It's possible that optimizing a switch on powers of two might also
5514  // be beneficial - flag values are often powers of two and we could use a CLZ
5515  // as the key function.
5516  if (GCD <= 1 || !isPowerOf2_64(GCD))
5517  // No common divisor found or too expensive to compute key function.
5518  return false;
5519 
5520  unsigned Shift = Log2_64(GCD);
5521  for (auto &V : Values)
5522  V = (int64_t)((uint64_t)V >> Shift);
5523 
5524  if (!isSwitchDense(Values))
5525  // Transform didn't create a dense switch.
5526  return false;
5527 
5528  // The obvious transform is to shift the switch condition right and emit a
5529  // check that the condition actually cleanly divided by GCD, i.e.
5530  // C & (1 << Shift - 1) == 0
5531  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5532  //
5533  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5534  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5535  // are nonzero then the switch condition will be very large and will hit the
5536  // default case.
5537 
5538  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5539  Builder.SetInsertPoint(SI);
5540  auto *ShiftC = ConstantInt::get(Ty, Shift);
5541  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5542  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5543  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5544  auto *Rot = Builder.CreateOr(LShr, Shl);
5545  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5546 
5547  for (auto Case : SI->cases()) {
5548  auto *Orig = Case.getCaseValue();
5549  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5550  Case.setValue(
5551  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5552  }
5553  return true;
5554 }
5555 
5556 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5557  BasicBlock *BB = SI->getParent();
5558 
5559  if (isValueEqualityComparison(SI)) {
5560  // If we only have one predecessor, and if it is a branch on this value,
5561  // see if that predecessor totally determines the outcome of this switch.
5562  if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5563  if (SimplifyEqualityComp