LLVM  9.0.0svn
SimplifyCFG.cpp
Go to the documentation of this file.
1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Peephole optimize the CFG.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/ADT/APInt.h"
14 #include "llvm/ADT/ArrayRef.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/ADT/StringRef.h"
31 #include "llvm/IR/Attributes.h"
32 #include "llvm/IR/BasicBlock.h"
33 #include "llvm/IR/CFG.h"
34 #include "llvm/IR/CallSite.h"
35 #include "llvm/IR/Constant.h"
36 #include "llvm/IR/ConstantRange.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/GlobalValue.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/IRBuilder.h"
44 #include "llvm/IR/InstrTypes.h"
45 #include "llvm/IR/Instruction.h"
46 #include "llvm/IR/Instructions.h"
47 #include "llvm/IR/IntrinsicInst.h"
48 #include "llvm/IR/Intrinsics.h"
49 #include "llvm/IR/LLVMContext.h"
50 #include "llvm/IR/MDBuilder.h"
51 #include "llvm/IR/Metadata.h"
52 #include "llvm/IR/Module.h"
53 #include "llvm/IR/NoFolder.h"
54 #include "llvm/IR/Operator.h"
55 #include "llvm/IR/PatternMatch.h"
56 #include "llvm/IR/Type.h"
57 #include "llvm/IR/Use.h"
58 #include "llvm/IR/User.h"
59 #include "llvm/IR/Value.h"
60 #include "llvm/Support/Casting.h"
62 #include "llvm/Support/Debug.h"
64 #include "llvm/Support/KnownBits.h"
69 #include <algorithm>
70 #include <cassert>
71 #include <climits>
72 #include <cstddef>
73 #include <cstdint>
74 #include <iterator>
75 #include <map>
76 #include <set>
77 #include <tuple>
78 #include <utility>
79 #include <vector>
80 
81 using namespace llvm;
82 using namespace PatternMatch;
83 
84 #define DEBUG_TYPE "simplifycfg"
85 
86 // Chosen as 2 so as to be cheap, but still to have enough power to fold
87 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
88 // To catch this, we need to fold a compare and a select, hence '2' being the
89 // minimum reasonable default.
91  "phi-node-folding-threshold", cl::Hidden, cl::init(2),
92  cl::desc(
93  "Control the amount of phi node folding to perform (default = 2)"));
94 
95 static cl::opt<bool> DupRet(
96  "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
97  cl::desc("Duplicate return instructions into unconditional branches"));
98 
99 static cl::opt<bool>
100  SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
101  cl::desc("Sink common instructions down to the end block"));
102 
104  "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
105  cl::desc("Hoist conditional stores if an unconditional store precedes"));
106 
108  "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
109  cl::desc("Hoist conditional stores even if an unconditional store does not "
110  "precede - hoist multiple conditional stores into a single "
111  "predicated store"));
112 
114  "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
115  cl::desc("When merging conditional stores, do so even if the resultant "
116  "basic blocks are unlikely to be if-converted as a result"));
117 
119  "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
120  cl::desc("Allow exactly one expensive instruction to be speculatively "
121  "executed"));
122 
124  "max-speculation-depth", cl::Hidden, cl::init(10),
125  cl::desc("Limit maximum recursion depth when calculating costs of "
126  "speculatively executed instructions"));
127 
128 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
129 STATISTIC(NumLinearMaps,
130  "Number of switch instructions turned into linear mapping");
131 STATISTIC(NumLookupTables,
132  "Number of switch instructions turned into lookup tables");
133 STATISTIC(
134  NumLookupTablesHoles,
135  "Number of switch instructions turned into lookup tables (holes checked)");
136 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
137 STATISTIC(NumSinkCommons,
138  "Number of common instructions sunk down to the end block");
139 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
140 
141 namespace {
142 
143 // The first field contains the value that the switch produces when a certain
144 // case group is selected, and the second field is a vector containing the
145 // cases composing the case group.
146 using SwitchCaseResultVectorTy =
148 
149 // The first field contains the phi node that generates a result of the switch
150 // and the second field contains the value generated for a certain case in the
151 // switch for that PHI.
152 using SwitchCaseResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
153 
154 /// ValueEqualityComparisonCase - Represents a case of a switch.
155 struct ValueEqualityComparisonCase {
157  BasicBlock *Dest;
158 
159  ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
160  : Value(Value), Dest(Dest) {}
161 
162  bool operator<(ValueEqualityComparisonCase RHS) const {
163  // Comparing pointers is ok as we only rely on the order for uniquing.
164  return Value < RHS.Value;
165  }
166 
167  bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
168 };
169 
170 class SimplifyCFGOpt {
171  const TargetTransformInfo &TTI;
172  const DataLayout &DL;
173  SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
174  const SimplifyCFGOptions &Options;
175  bool Resimplify;
176 
177  Value *isValueEqualityComparison(Instruction *TI);
178  BasicBlock *GetValueEqualityComparisonCases(
179  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases);
180  bool SimplifyEqualityComparisonWithOnlyPredecessor(Instruction *TI,
181  BasicBlock *Pred,
182  IRBuilder<> &Builder);
183  bool FoldValueComparisonIntoPredecessors(Instruction *TI,
184  IRBuilder<> &Builder);
185 
186  bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
187  bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
188  bool SimplifySingleResume(ResumeInst *RI);
189  bool SimplifyCommonResume(ResumeInst *RI);
190  bool SimplifyCleanupReturn(CleanupReturnInst *RI);
191  bool SimplifyUnreachable(UnreachableInst *UI);
192  bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
193  bool SimplifyIndirectBr(IndirectBrInst *IBI);
194  bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
195  bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
196 
197  bool tryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
198  IRBuilder<> &Builder);
199 
200 public:
201  SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
202  SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
203  const SimplifyCFGOptions &Opts)
204  : TTI(TTI), DL(DL), LoopHeaders(LoopHeaders), Options(Opts) {}
205 
206  bool run(BasicBlock *BB);
207  bool simplifyOnce(BasicBlock *BB);
208 
209  // Helper to set Resimplify and return change indication.
210  bool requestResimplify() {
211  Resimplify = true;
212  return true;
213  }
214 };
215 
216 } // end anonymous namespace
217 
218 /// Return true if it is safe to merge these two
219 /// terminator instructions together.
220 static bool
222  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
223  if (SI1 == SI2)
224  return false; // Can't merge with self!
225 
226  // It is not safe to merge these two switch instructions if they have a common
227  // successor, and if that successor has a PHI node, and if *that* PHI node has
228  // conflicting incoming values from the two switch blocks.
229  BasicBlock *SI1BB = SI1->getParent();
230  BasicBlock *SI2BB = SI2->getParent();
231 
232  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
233  bool Fail = false;
234  for (BasicBlock *Succ : successors(SI2BB))
235  if (SI1Succs.count(Succ))
236  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
237  PHINode *PN = cast<PHINode>(BBI);
238  if (PN->getIncomingValueForBlock(SI1BB) !=
239  PN->getIncomingValueForBlock(SI2BB)) {
240  if (FailBlocks)
241  FailBlocks->insert(Succ);
242  Fail = true;
243  }
244  }
245 
246  return !Fail;
247 }
248 
249 /// Return true if it is safe and profitable to merge these two terminator
250 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
251 /// store all PHI nodes in common successors.
252 static bool
254  Instruction *Cond,
255  SmallVectorImpl<PHINode *> &PhiNodes) {
256  if (SI1 == SI2)
257  return false; // Can't merge with self!
258  assert(SI1->isUnconditional() && SI2->isConditional());
259 
260  // We fold the unconditional branch if we can easily update all PHI nodes in
261  // common successors:
262  // 1> We have a constant incoming value for the conditional branch;
263  // 2> We have "Cond" as the incoming value for the unconditional branch;
264  // 3> SI2->getCondition() and Cond have same operands.
265  CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
266  if (!Ci2)
267  return false;
268  if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
269  Cond->getOperand(1) == Ci2->getOperand(1)) &&
270  !(Cond->getOperand(0) == Ci2->getOperand(1) &&
271  Cond->getOperand(1) == Ci2->getOperand(0)))
272  return false;
273 
274  BasicBlock *SI1BB = SI1->getParent();
275  BasicBlock *SI2BB = SI2->getParent();
276  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
277  for (BasicBlock *Succ : successors(SI2BB))
278  if (SI1Succs.count(Succ))
279  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
280  PHINode *PN = cast<PHINode>(BBI);
281  if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
282  !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
283  return false;
284  PhiNodes.push_back(PN);
285  }
286  return true;
287 }
288 
289 /// Update PHI nodes in Succ to indicate that there will now be entries in it
290 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
291 /// will be the same as those coming in from ExistPred, an existing predecessor
292 /// of Succ.
293 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
294  BasicBlock *ExistPred) {
295  for (PHINode &PN : Succ->phis())
296  PN.addIncoming(PN.getIncomingValueForBlock(ExistPred), NewPred);
297 }
298 
299 /// Compute an abstract "cost" of speculating the given instruction,
300 /// which is assumed to be safe to speculate. TCC_Free means cheap,
301 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
302 /// expensive.
303 static unsigned ComputeSpeculationCost(const User *I,
304  const TargetTransformInfo &TTI) {
306  "Instruction is not safe to speculatively execute!");
307  return TTI.getUserCost(I);
308 }
309 
310 /// If we have a merge point of an "if condition" as accepted above,
311 /// return true if the specified value dominates the block. We
312 /// don't handle the true generality of domination here, just a special case
313 /// which works well enough for us.
314 ///
315 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
316 /// see if V (which must be an instruction) and its recursive operands
317 /// that do not dominate BB have a combined cost lower than CostRemaining and
318 /// are non-trapping. If both are true, the instruction is inserted into the
319 /// set and true is returned.
320 ///
321 /// The cost for most non-trapping instructions is defined as 1 except for
322 /// Select whose cost is 2.
323 ///
324 /// After this function returns, CostRemaining is decreased by the cost of
325 /// V plus its non-dominating operands. If that cost is greater than
326 /// CostRemaining, false is returned and CostRemaining is undefined.
328  SmallPtrSetImpl<Instruction *> &AggressiveInsts,
329  unsigned &CostRemaining,
330  const TargetTransformInfo &TTI,
331  unsigned Depth = 0) {
332  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
333  // so limit the recursion depth.
334  // TODO: While this recursion limit does prevent pathological behavior, it
335  // would be better to track visited instructions to avoid cycles.
336  if (Depth == MaxSpeculationDepth)
337  return false;
338 
340  if (!I) {
341  // Non-instructions all dominate instructions, but not all constantexprs
342  // can be executed unconditionally.
343  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
344  if (C->canTrap())
345  return false;
346  return true;
347  }
348  BasicBlock *PBB = I->getParent();
349 
350  // We don't want to allow weird loops that might have the "if condition" in
351  // the bottom of this block.
352  if (PBB == BB)
353  return false;
354 
355  // If this instruction is defined in a block that contains an unconditional
356  // branch to BB, then it must be in the 'conditional' part of the "if
357  // statement". If not, it definitely dominates the region.
359  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
360  return true;
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(Instruction *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->getParent()->hasNPredecessorsOrMore(128 / SI->getNumSuccessors()))
696  CV = SI->getCondition();
697  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
698  if (BI->isConditional() && BI->getCondition()->hasOneUse())
699  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
700  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
701  CV = ICI->getOperand(0);
702  }
703 
704  // Unwrap any lossless ptrtoint cast.
705  if (CV) {
706  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
707  Value *Ptr = PTII->getPointerOperand();
708  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
709  CV = Ptr;
710  }
711  }
712  return CV;
713 }
714 
715 /// Given a value comparison instruction,
716 /// decode all of the 'cases' that it represents and return the 'default' block.
717 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
718  Instruction *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
719  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
720  Cases.reserve(SI->getNumCases());
721  for (auto Case : SI->cases())
722  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
723  Case.getCaseSuccessor()));
724  return SI->getDefaultDest();
725  }
726 
727  BranchInst *BI = cast<BranchInst>(TI);
728  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
729  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
730  Cases.push_back(ValueEqualityComparisonCase(
731  GetConstantInt(ICI->getOperand(1), DL), Succ));
732  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
733 }
734 
735 /// Given a vector of bb/value pairs, remove any entries
736 /// in the list that match the specified block.
737 static void
739  std::vector<ValueEqualityComparisonCase> &Cases) {
740  Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
741 }
742 
743 /// Return true if there are any keys in C1 that exist in C2 as well.
744 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
745  std::vector<ValueEqualityComparisonCase> &C2) {
746  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
747 
748  // Make V1 be smaller than V2.
749  if (V1->size() > V2->size())
750  std::swap(V1, V2);
751 
752  if (V1->empty())
753  return false;
754  if (V1->size() == 1) {
755  // Just scan V2.
756  ConstantInt *TheVal = (*V1)[0].Value;
757  for (unsigned i = 0, e = V2->size(); i != e; ++i)
758  if (TheVal == (*V2)[i].Value)
759  return true;
760  }
761 
762  // Otherwise, just sort both lists and compare element by element.
763  array_pod_sort(V1->begin(), V1->end());
764  array_pod_sort(V2->begin(), V2->end());
765  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
766  while (i1 != e1 && i2 != e2) {
767  if ((*V1)[i1].Value == (*V2)[i2].Value)
768  return true;
769  if ((*V1)[i1].Value < (*V2)[i2].Value)
770  ++i1;
771  else
772  ++i2;
773  }
774  return false;
775 }
776 
777 // Set branch weights on SwitchInst. This sets the metadata if there is at
778 // least one non-zero weight.
780  // Check that there is at least one non-zero weight. Otherwise, pass
781  // nullptr to setMetadata which will erase the existing metadata.
782  MDNode *N = nullptr;
783  if (llvm::any_of(Weights, [](uint32_t W) { return W != 0; }))
784  N = MDBuilder(SI->getParent()->getContext()).createBranchWeights(Weights);
786 }
787 
788 // Similar to the above, but for branch and select instructions that take
789 // exactly 2 weights.
790 static void setBranchWeights(Instruction *I, uint32_t TrueWeight,
791  uint32_t FalseWeight) {
792  assert(isa<BranchInst>(I) || isa<SelectInst>(I));
793  // Check that there is at least one non-zero weight. Otherwise, pass
794  // nullptr to setMetadata which will erase the existing metadata.
795  MDNode *N = nullptr;
796  if (TrueWeight || FalseWeight)
797  N = MDBuilder(I->getParent()->getContext())
798  .createBranchWeights(TrueWeight, FalseWeight);
800 }
801 
802 /// If TI is known to be a terminator instruction and its block is known to
803 /// only have a single predecessor block, check to see if that predecessor is
804 /// also a value comparison with the same value, and if that comparison
805 /// determines the outcome of this comparison. If so, simplify TI. This does a
806 /// very limited form of jump threading.
807 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
808  Instruction *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
809  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
810  if (!PredVal)
811  return false; // Not a value comparison in predecessor.
812 
813  Value *ThisVal = isValueEqualityComparison(TI);
814  assert(ThisVal && "This isn't a value comparison!!");
815  if (ThisVal != PredVal)
816  return false; // Different predicates.
817 
818  // TODO: Preserve branch weight metadata, similarly to how
819  // FoldValueComparisonIntoPredecessors preserves it.
820 
821  // Find out information about when control will move from Pred to TI's block.
822  std::vector<ValueEqualityComparisonCase> PredCases;
823  BasicBlock *PredDef =
824  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
825  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
826 
827  // Find information about how control leaves this block.
828  std::vector<ValueEqualityComparisonCase> ThisCases;
829  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
830  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
831 
832  // If TI's block is the default block from Pred's comparison, potentially
833  // simplify TI based on this knowledge.
834  if (PredDef == TI->getParent()) {
835  // If we are here, we know that the value is none of those cases listed in
836  // PredCases. If there are any cases in ThisCases that are in PredCases, we
837  // can simplify TI.
838  if (!ValuesOverlap(PredCases, ThisCases))
839  return false;
840 
841  if (isa<BranchInst>(TI)) {
842  // Okay, one of the successors of this condbr is dead. Convert it to a
843  // uncond br.
844  assert(ThisCases.size() == 1 && "Branch can only have one case!");
845  // Insert the new branch.
846  Instruction *NI = Builder.CreateBr(ThisDef);
847  (void)NI;
848 
849  // Remove PHI node entries for the dead edge.
850  ThisCases[0].Dest->removePredecessor(TI->getParent());
851 
852  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
853  << "Through successor TI: " << *TI << "Leaving: " << *NI
854  << "\n");
855 
857  return true;
858  }
859 
860  SwitchInst *SI = cast<SwitchInst>(TI);
861  // Okay, TI has cases that are statically dead, prune them away.
862  SmallPtrSet<Constant *, 16> DeadCases;
863  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
864  DeadCases.insert(PredCases[i].Value);
865 
866  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
867  << "Through successor TI: " << *TI);
868 
869  // Collect branch weights into a vector.
870  SmallVector<uint32_t, 8> Weights;
872  bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
873  if (HasWeight)
874  for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
875  ++MD_i) {
876  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
877  Weights.push_back(CI->getValue().getZExtValue());
878  }
879  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
880  --i;
881  if (DeadCases.count(i->getCaseValue())) {
882  if (HasWeight) {
883  std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
884  Weights.pop_back();
885  }
886  i->getCaseSuccessor()->removePredecessor(TI->getParent());
887  SI->removeCase(i);
888  }
889  }
890  if (HasWeight && Weights.size() >= 2)
891  setBranchWeights(SI, Weights);
892 
893  LLVM_DEBUG(dbgs() << "Leaving: " << *TI << "\n");
894  return true;
895  }
896 
897  // Otherwise, TI's block must correspond to some matched value. Find out
898  // which value (or set of values) this is.
899  ConstantInt *TIV = nullptr;
900  BasicBlock *TIBB = TI->getParent();
901  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
902  if (PredCases[i].Dest == TIBB) {
903  if (TIV)
904  return false; // Cannot handle multiple values coming to this block.
905  TIV = PredCases[i].Value;
906  }
907  assert(TIV && "No edge from pred to succ?");
908 
909  // Okay, we found the one constant that our value can be if we get into TI's
910  // BB. Find out which successor will unconditionally be branched to.
911  BasicBlock *TheRealDest = nullptr;
912  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
913  if (ThisCases[i].Value == TIV) {
914  TheRealDest = ThisCases[i].Dest;
915  break;
916  }
917 
918  // If not handled by any explicit cases, it is handled by the default case.
919  if (!TheRealDest)
920  TheRealDest = ThisDef;
921 
922  // Remove PHI node entries for dead edges.
923  BasicBlock *CheckEdge = TheRealDest;
924  for (BasicBlock *Succ : successors(TIBB))
925  if (Succ != CheckEdge)
926  Succ->removePredecessor(TIBB);
927  else
928  CheckEdge = nullptr;
929 
930  // Insert the new branch.
931  Instruction *NI = Builder.CreateBr(TheRealDest);
932  (void)NI;
933 
934  LLVM_DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
935  << "Through successor TI: " << *TI << "Leaving: " << *NI
936  << "\n");
937 
939  return true;
940 }
941 
942 namespace {
943 
944 /// This class implements a stable ordering of constant
945 /// integers that does not depend on their address. This is important for
946 /// applications that sort ConstantInt's to ensure uniqueness.
947 struct ConstantIntOrdering {
948  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
949  return LHS->getValue().ult(RHS->getValue());
950  }
951 };
952 
953 } // end anonymous namespace
954 
955 static int ConstantIntSortPredicate(ConstantInt *const *P1,
956  ConstantInt *const *P2) {
957  const ConstantInt *LHS = *P1;
958  const ConstantInt *RHS = *P2;
959  if (LHS == RHS)
960  return 0;
961  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
962 }
963 
964 static inline bool HasBranchWeights(const Instruction *I) {
966  if (ProfMD && ProfMD->getOperand(0))
967  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
968  return MDS->getString().equals("branch_weights");
969 
970  return false;
971 }
972 
973 /// Get Weights of a given terminator, the default weight is at the front
974 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
975 /// metadata.
977  SmallVectorImpl<uint64_t> &Weights) {
979  assert(MD);
980  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
981  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
982  Weights.push_back(CI->getValue().getZExtValue());
983  }
984 
985  // If TI is a conditional eq, the default case is the false case,
986  // and the corresponding branch-weight data is at index 2. We swap the
987  // default weight to be the first entry.
988  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
989  assert(Weights.size() == 2);
990  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
991  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
992  std::swap(Weights.front(), Weights.back());
993  }
994 }
995 
996 /// Keep halving the weights until all can fit in uint32_t.
997 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
998  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
999  if (Max > UINT_MAX) {
1000  unsigned Offset = 32 - countLeadingZeros(Max);
1001  for (uint64_t &I : Weights)
1002  I >>= Offset;
1003  }
1004 }
1005 
1006 /// The specified terminator is a value equality comparison instruction
1007 /// (either a switch or a branch on "X == c").
1008 /// See if any of the predecessors of the terminator block are value comparisons
1009 /// on the same value. If so, and if safe to do so, fold them together.
1010 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(Instruction *TI,
1011  IRBuilder<> &Builder) {
1012  BasicBlock *BB = TI->getParent();
1013  Value *CV = isValueEqualityComparison(TI); // CondVal
1014  assert(CV && "Not a comparison?");
1015  bool Changed = false;
1016 
1018  while (!Preds.empty()) {
1019  BasicBlock *Pred = Preds.pop_back_val();
1020 
1021  // See if the predecessor is a comparison with the same value.
1022  Instruction *PTI = Pred->getTerminator();
1023  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
1024 
1025  if (PCV == CV && TI != PTI) {
1026  SmallSetVector<BasicBlock*, 4> FailBlocks;
1027  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1028  for (auto *Succ : FailBlocks) {
1029  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1030  return false;
1031  }
1032  }
1033 
1034  // Figure out which 'cases' to copy from SI to PSI.
1035  std::vector<ValueEqualityComparisonCase> BBCases;
1036  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1037 
1038  std::vector<ValueEqualityComparisonCase> PredCases;
1039  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1040 
1041  // Based on whether the default edge from PTI goes to BB or not, fill in
1042  // PredCases and PredDefault with the new switch cases we would like to
1043  // build.
1044  SmallVector<BasicBlock *, 8> NewSuccessors;
1045 
1046  // Update the branch weight metadata along the way
1047  SmallVector<uint64_t, 8> Weights;
1048  bool PredHasWeights = HasBranchWeights(PTI);
1049  bool SuccHasWeights = HasBranchWeights(TI);
1050 
1051  if (PredHasWeights) {
1052  GetBranchWeights(PTI, Weights);
1053  // branch-weight metadata is inconsistent here.
1054  if (Weights.size() != 1 + PredCases.size())
1055  PredHasWeights = SuccHasWeights = false;
1056  } else if (SuccHasWeights)
1057  // If there are no predecessor weights but there are successor weights,
1058  // populate Weights with 1, which will later be scaled to the sum of
1059  // successor's weights
1060  Weights.assign(1 + PredCases.size(), 1);
1061 
1062  SmallVector<uint64_t, 8> SuccWeights;
1063  if (SuccHasWeights) {
1064  GetBranchWeights(TI, SuccWeights);
1065  // branch-weight metadata is inconsistent here.
1066  if (SuccWeights.size() != 1 + BBCases.size())
1067  PredHasWeights = SuccHasWeights = false;
1068  } else if (PredHasWeights)
1069  SuccWeights.assign(1 + BBCases.size(), 1);
1070 
1071  if (PredDefault == BB) {
1072  // If this is the default destination from PTI, only the edges in TI
1073  // that don't occur in PTI, or that branch to BB will be activated.
1074  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1075  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1076  if (PredCases[i].Dest != BB)
1077  PTIHandled.insert(PredCases[i].Value);
1078  else {
1079  // The default destination is BB, we don't need explicit targets.
1080  std::swap(PredCases[i], PredCases.back());
1081 
1082  if (PredHasWeights || SuccHasWeights) {
1083  // Increase weight for the default case.
1084  Weights[0] += Weights[i + 1];
1085  std::swap(Weights[i + 1], Weights.back());
1086  Weights.pop_back();
1087  }
1088 
1089  PredCases.pop_back();
1090  --i;
1091  --e;
1092  }
1093 
1094  // Reconstruct the new switch statement we will be building.
1095  if (PredDefault != BBDefault) {
1096  PredDefault->removePredecessor(Pred);
1097  PredDefault = BBDefault;
1098  NewSuccessors.push_back(BBDefault);
1099  }
1100 
1101  unsigned CasesFromPred = Weights.size();
1102  uint64_t ValidTotalSuccWeight = 0;
1103  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1104  if (!PTIHandled.count(BBCases[i].Value) &&
1105  BBCases[i].Dest != BBDefault) {
1106  PredCases.push_back(BBCases[i]);
1107  NewSuccessors.push_back(BBCases[i].Dest);
1108  if (SuccHasWeights || PredHasWeights) {
1109  // The default weight is at index 0, so weight for the ith case
1110  // should be at index i+1. Scale the cases from successor by
1111  // PredDefaultWeight (Weights[0]).
1112  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1113  ValidTotalSuccWeight += SuccWeights[i + 1];
1114  }
1115  }
1116 
1117  if (SuccHasWeights || PredHasWeights) {
1118  ValidTotalSuccWeight += SuccWeights[0];
1119  // Scale the cases from predecessor by ValidTotalSuccWeight.
1120  for (unsigned i = 1; i < CasesFromPred; ++i)
1121  Weights[i] *= ValidTotalSuccWeight;
1122  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1123  Weights[0] *= SuccWeights[0];
1124  }
1125  } else {
1126  // If this is not the default destination from PSI, only the edges
1127  // in SI that occur in PSI with a destination of BB will be
1128  // activated.
1129  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1130  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1131  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1132  if (PredCases[i].Dest == BB) {
1133  PTIHandled.insert(PredCases[i].Value);
1134 
1135  if (PredHasWeights || SuccHasWeights) {
1136  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1137  std::swap(Weights[i + 1], Weights.back());
1138  Weights.pop_back();
1139  }
1140 
1141  std::swap(PredCases[i], PredCases.back());
1142  PredCases.pop_back();
1143  --i;
1144  --e;
1145  }
1146 
1147  // Okay, now we know which constants were sent to BB from the
1148  // predecessor. Figure out where they will all go now.
1149  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1150  if (PTIHandled.count(BBCases[i].Value)) {
1151  // If this is one we are capable of getting...
1152  if (PredHasWeights || SuccHasWeights)
1153  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1154  PredCases.push_back(BBCases[i]);
1155  NewSuccessors.push_back(BBCases[i].Dest);
1156  PTIHandled.erase(
1157  BBCases[i].Value); // This constant is taken care of
1158  }
1159 
1160  // If there are any constants vectored to BB that TI doesn't handle,
1161  // they must go to the default destination of TI.
1162  for (ConstantInt *I : PTIHandled) {
1163  if (PredHasWeights || SuccHasWeights)
1164  Weights.push_back(WeightsForHandled[I]);
1165  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1166  NewSuccessors.push_back(BBDefault);
1167  }
1168  }
1169 
1170  // Okay, at this point, we know which new successor Pred will get. Make
1171  // sure we update the number of entries in the PHI nodes for these
1172  // successors.
1173  for (BasicBlock *NewSuccessor : NewSuccessors)
1174  AddPredecessorToBlock(NewSuccessor, Pred, BB);
1175 
1176  Builder.SetInsertPoint(PTI);
1177  // Convert pointer to int before we switch.
1178  if (CV->getType()->isPointerTy()) {
1179  CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1180  "magicptr");
1181  }
1182 
1183  // Now that the successors are updated, create the new Switch instruction.
1184  SwitchInst *NewSI =
1185  Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1186  NewSI->setDebugLoc(PTI->getDebugLoc());
1187  for (ValueEqualityComparisonCase &V : PredCases)
1188  NewSI->addCase(V.Value, V.Dest);
1189 
1190  if (PredHasWeights || SuccHasWeights) {
1191  // Halve the weights if any of them cannot fit in an uint32_t
1192  FitWeights(Weights);
1193 
1194  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1195 
1196  setBranchWeights(NewSI, MDWeights);
1197  }
1198 
1200 
1201  // Okay, last check. If BB is still a successor of PSI, then we must
1202  // have an infinite loop case. If so, add an infinitely looping block
1203  // to handle the case to preserve the behavior of the code.
1204  BasicBlock *InfLoopBlock = nullptr;
1205  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1206  if (NewSI->getSuccessor(i) == BB) {
1207  if (!InfLoopBlock) {
1208  // Insert it at the end of the function, because it's either code,
1209  // or it won't matter if it's hot. :)
1210  InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1211  BB->getParent());
1212  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1213  }
1214  NewSI->setSuccessor(i, InfLoopBlock);
1215  }
1216 
1217  Changed = true;
1218  }
1219  }
1220  return Changed;
1221 }
1222 
1223 // If we would need to insert a select that uses the value of this invoke
1224 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1225 // can't hoist the invoke, as there is nowhere to put the select in this case.
1227  Instruction *I1, Instruction *I2) {
1228  for (BasicBlock *Succ : successors(BB1)) {
1229  for (const PHINode &PN : Succ->phis()) {
1230  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1231  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1232  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1233  return false;
1234  }
1235  }
1236  }
1237  return true;
1238 }
1239 
1241 
1242 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1243 /// in the two blocks up into the branch block. The caller of this function
1244 /// guarantees that BI's block dominates BB1 and BB2.
1246  const TargetTransformInfo &TTI) {
1247  // This does very trivial matching, with limited scanning, to find identical
1248  // instructions in the two blocks. In particular, we don't want to get into
1249  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1250  // such, we currently just scan for obviously identical instructions in an
1251  // identical order.
1252  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1253  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1254 
1255  BasicBlock::iterator BB1_Itr = BB1->begin();
1256  BasicBlock::iterator BB2_Itr = BB2->begin();
1257 
1258  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1259  // Skip debug info if it is not identical.
1262  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1263  while (isa<DbgInfoIntrinsic>(I1))
1264  I1 = &*BB1_Itr++;
1265  while (isa<DbgInfoIntrinsic>(I2))
1266  I2 = &*BB2_Itr++;
1267  }
1268  // FIXME: Can we define a safety predicate for CallBr?
1269  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1270  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)) ||
1271  isa<CallBrInst>(I1))
1272  return false;
1273 
1274  BasicBlock *BIParent = BI->getParent();
1275 
1276  bool Changed = false;
1277  do {
1278  // If we are hoisting the terminator instruction, don't move one (making a
1279  // broken BB), instead clone it, and remove BI.
1280  if (I1->isTerminator())
1281  goto HoistTerminator;
1282 
1283  // If we're going to hoist a call, make sure that the two instructions we're
1284  // commoning/hoisting are both marked with musttail, or neither of them is
1285  // marked as such. Otherwise, we might end up in a situation where we hoist
1286  // from a block where the terminator is a `ret` to a block where the terminator
1287  // is a `br`, and `musttail` calls expect to be followed by a return.
1288  auto *C1 = dyn_cast<CallInst>(I1);
1289  auto *C2 = dyn_cast<CallInst>(I2);
1290  if (C1 && C2)
1291  if (C1->isMustTailCall() != C2->isMustTailCall())
1292  return Changed;
1293 
1294  if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1295  return Changed;
1296 
1297  if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
1298  assert (isa<DbgInfoIntrinsic>(I1) && isa<DbgInfoIntrinsic>(I2));
1299  // The debug location is an integral part of a debug info intrinsic
1300  // and can't be separated from it or replaced. Instead of attempting
1301  // to merge locations, simply hoist both copies of the intrinsic.
1302  BIParent->getInstList().splice(BI->getIterator(),
1303  BB1->getInstList(), I1);
1304  BIParent->getInstList().splice(BI->getIterator(),
1305  BB2->getInstList(), I2);
1306  Changed = true;
1307  } else {
1308  // For a normal instruction, we just move one to right before the branch,
1309  // then replace all uses of the other with the first. Finally, we remove
1310  // the now redundant second instruction.
1311  BIParent->getInstList().splice(BI->getIterator(),
1312  BB1->getInstList(), I1);
1313  if (!I2->use_empty())
1314  I2->replaceAllUsesWith(I1);
1315  I1->andIRFlags(I2);
1316  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1327  combineMetadata(I1, I2, KnownIDs, true);
1328 
1329  // I1 and I2 are being combined into a single instruction. Its debug
1330  // location is the merged locations of the original instructions.
1331  I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1332 
1333  I2->eraseFromParent();
1334  Changed = true;
1335  }
1336 
1337  I1 = &*BB1_Itr++;
1338  I2 = &*BB2_Itr++;
1339  // Skip debug info if it is not identical.
1342  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1343  while (isa<DbgInfoIntrinsic>(I1))
1344  I1 = &*BB1_Itr++;
1345  while (isa<DbgInfoIntrinsic>(I2))
1346  I2 = &*BB2_Itr++;
1347  }
1348  } while (I1->isIdenticalToWhenDefined(I2));
1349 
1350  return true;
1351 
1352 HoistTerminator:
1353  // It may not be possible to hoist an invoke.
1354  // FIXME: Can we define a safety predicate for CallBr?
1355  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1356  return Changed;
1357 
1358  // TODO: callbr hoisting currently disabled pending further study.
1359  if (isa<CallBrInst>(I1))
1360  return Changed;
1361 
1362  for (BasicBlock *Succ : successors(BB1)) {
1363  for (PHINode &PN : Succ->phis()) {
1364  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1365  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1366  if (BB1V == BB2V)
1367  continue;
1368 
1369  // Check for passingValueIsAlwaysUndefined here because we would rather
1370  // eliminate undefined control flow then converting it to a select.
1371  if (passingValueIsAlwaysUndefined(BB1V, &PN) ||
1372  passingValueIsAlwaysUndefined(BB2V, &PN))
1373  return Changed;
1374 
1375  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1376  return Changed;
1377  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1378  return Changed;
1379  }
1380  }
1381 
1382  // Okay, it is safe to hoist the terminator.
1383  Instruction *NT = I1->clone();
1384  BIParent->getInstList().insert(BI->getIterator(), NT);
1385  if (!NT->getType()->isVoidTy()) {
1386  I1->replaceAllUsesWith(NT);
1387  I2->replaceAllUsesWith(NT);
1388  NT->takeName(I1);
1389  }
1390 
1391  // Ensure terminator gets a debug location, even an unknown one, in case
1392  // it involves inlinable calls.
1393  NT->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
1394 
1395  // PHIs created below will adopt NT's merged DebugLoc.
1396  IRBuilder<NoFolder> Builder(NT);
1397 
1398  // Hoisting one of the terminators from our successor is a great thing.
1399  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1400  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1401  // nodes, so we insert select instruction to compute the final result.
1402  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1403  for (BasicBlock *Succ : successors(BB1)) {
1404  for (PHINode &PN : Succ->phis()) {
1405  Value *BB1V = PN.getIncomingValueForBlock(BB1);
1406  Value *BB2V = PN.getIncomingValueForBlock(BB2);
1407  if (BB1V == BB2V)
1408  continue;
1409 
1410  // These values do not agree. Insert a select instruction before NT
1411  // that determines the right value.
1412  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1413  if (!SI)
1414  SI = cast<SelectInst>(
1415  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1416  BB1V->getName() + "." + BB2V->getName(), BI));
1417 
1418  // Make the PHI node use the select for all incoming values for BB1/BB2
1419  for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1420  if (PN.getIncomingBlock(i) == BB1 || PN.getIncomingBlock(i) == BB2)
1421  PN.setIncomingValue(i, SI);
1422  }
1423  }
1424 
1425  // Update any PHI nodes in our new successors.
1426  for (BasicBlock *Succ : successors(BB1))
1427  AddPredecessorToBlock(Succ, BIParent, BB1);
1428 
1430  return true;
1431 }
1432 
1433 // All instructions in Insts belong to different blocks that all unconditionally
1434 // branch to a common successor. Analyze each instruction and return true if it
1435 // would be possible to sink them into their successor, creating one common
1436 // instruction instead. For every value that would be required to be provided by
1437 // PHI node (because an operand varies in each input block), add to PHIOperands.
1440  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1441  // Prune out obviously bad instructions to move. Each instruction must have
1442  // exactly zero or one use, and we check later that use is by a single, common
1443  // PHI instruction in the successor.
1444  bool HasUse = !Insts.front()->user_empty();
1445  for (auto *I : Insts) {
1446  // These instructions may change or break semantics if moved.
1447  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1448  I->getType()->isTokenTy())
1449  return false;
1450 
1451  // Conservatively return false if I is an inline-asm instruction. Sinking
1452  // and merging inline-asm instructions can potentially create arguments
1453  // that cannot satisfy the inline-asm constraints.
1454  if (const auto *C = dyn_cast<CallBase>(I))
1455  if (C->isInlineAsm())
1456  return false;
1457 
1458  // Each instruction must have zero or one use.
1459  if (HasUse && !I->hasOneUse())
1460  return false;
1461  if (!HasUse && !I->user_empty())
1462  return false;
1463  }
1464 
1465  const Instruction *I0 = Insts.front();
1466  for (auto *I : Insts)
1467  if (!I->isSameOperationAs(I0))
1468  return false;
1469 
1470  // All instructions in Insts are known to be the same opcode. If they have a
1471  // use, check that the only user is a PHI or in the same block as the
1472  // instruction, because if a user is in the same block as an instruction we're
1473  // contemplating sinking, it must already be determined to be sinkable.
1474  if (HasUse) {
1475  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1476  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1477  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1478  auto *U = cast<Instruction>(*I->user_begin());
1479  return (PNUse &&
1480  PNUse->getParent() == Succ &&
1481  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1482  U->getParent() == I->getParent();
1483  }))
1484  return false;
1485  }
1486 
1487  // Because SROA can't handle speculating stores of selects, try not
1488  // to sink loads or stores of allocas when we'd have to create a PHI for
1489  // the address operand. Also, because it is likely that loads or stores
1490  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1491  // This can cause code churn which can have unintended consequences down
1492  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1493  // FIXME: This is a workaround for a deficiency in SROA - see
1494  // https://llvm.org/bugs/show_bug.cgi?id=30188
1495  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1496  return isa<AllocaInst>(I->getOperand(1));
1497  }))
1498  return false;
1499  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1500  return isa<AllocaInst>(I->getOperand(0));
1501  }))
1502  return false;
1503 
1504  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1505  if (I0->getOperand(OI)->getType()->isTokenTy())
1506  // Don't touch any operand of token type.
1507  return false;
1508 
1509  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1510  assert(I->getNumOperands() == I0->getNumOperands());
1511  return I->getOperand(OI) == I0->getOperand(OI);
1512  };
1513  if (!all_of(Insts, SameAsI0)) {
1514  if (!canReplaceOperandWithVariable(I0, OI))
1515  // We can't create a PHI from this GEP.
1516  return false;
1517  // Don't create indirect calls! The called value is the final operand.
1518  if (isa<CallBase>(I0) && OI == OE - 1) {
1519  // FIXME: if the call was *already* indirect, we should do this.
1520  return false;
1521  }
1522  for (auto *I : Insts)
1523  PHIOperands[I].push_back(I->getOperand(OI));
1524  }
1525  }
1526  return true;
1527 }
1528 
1529 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1530 // instruction of every block in Blocks to their common successor, commoning
1531 // into one instruction.
1533  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1534 
1535  // canSinkLastInstruction returning true guarantees that every block has at
1536  // least one non-terminator instruction.
1538  for (auto *BB : Blocks) {
1539  Instruction *I = BB->getTerminator();
1540  do {
1541  I = I->getPrevNode();
1542  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1543  if (!isa<DbgInfoIntrinsic>(I))
1544  Insts.push_back(I);
1545  }
1546 
1547  // The only checking we need to do now is that all users of all instructions
1548  // are the same PHI node. canSinkLastInstruction should have checked this but
1549  // it is slightly over-aggressive - it gets confused by commutative instructions
1550  // so double-check it here.
1551  Instruction *I0 = Insts.front();
1552  if (!I0->user_empty()) {
1553  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1554  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1555  auto *U = cast<Instruction>(*I->user_begin());
1556  return U == PNUse;
1557  }))
1558  return false;
1559  }
1560 
1561  // We don't need to do any more checking here; canSinkLastInstruction should
1562  // have done it all for us.
1563  SmallVector<Value*, 4> NewOperands;
1564  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1565  // This check is different to that in canSinkLastInstruction. There, we
1566  // cared about the global view once simplifycfg (and instcombine) have
1567  // completed - it takes into account PHIs that become trivially
1568  // simplifiable. However here we need a more local view; if an operand
1569  // differs we create a PHI and rely on instcombine to clean up the very
1570  // small mess we may make.
1571  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1572  return I->getOperand(O) != I0->getOperand(O);
1573  });
1574  if (!NeedPHI) {
1575  NewOperands.push_back(I0->getOperand(O));
1576  continue;
1577  }
1578 
1579  // Create a new PHI in the successor block and populate it.
1580  auto *Op = I0->getOperand(O);
1581  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1582  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1583  Op->getName() + ".sink", &BBEnd->front());
1584  for (auto *I : Insts)
1585  PN->addIncoming(I->getOperand(O), I->getParent());
1586  NewOperands.push_back(PN);
1587  }
1588 
1589  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1590  // and move it to the start of the successor block.
1591  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1592  I0->getOperandUse(O).set(NewOperands[O]);
1593  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1594 
1595  // Update metadata and IR flags, and merge debug locations.
1596  for (auto *I : Insts)
1597  if (I != I0) {
1598  // The debug location for the "common" instruction is the merged locations
1599  // of all the commoned instructions. We start with the original location
1600  // of the "common" instruction and iteratively merge each location in the
1601  // loop below.
1602  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1603  // However, as N-way merge for CallInst is rare, so we use simplified API
1604  // instead of using complex API for N-way merge.
1605  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1606  combineMetadataForCSE(I0, I, true);
1607  I0->andIRFlags(I);
1608  }
1609 
1610  if (!I0->user_empty()) {
1611  // canSinkLastInstruction checked that all instructions were used by
1612  // one and only one PHI node. Find that now, RAUW it to our common
1613  // instruction and nuke it.
1614  auto *PN = cast<PHINode>(*I0->user_begin());
1615  PN->replaceAllUsesWith(I0);
1616  PN->eraseFromParent();
1617  }
1618 
1619  // Finally nuke all instructions apart from the common instruction.
1620  for (auto *I : Insts)
1621  if (I != I0)
1622  I->eraseFromParent();
1623 
1624  return true;
1625 }
1626 
1627 namespace {
1628 
1629  // LockstepReverseIterator - Iterates through instructions
1630  // in a set of blocks in reverse order from the first non-terminator.
1631  // For example (assume all blocks have size n):
1632  // LockstepReverseIterator I([B1, B2, B3]);
1633  // *I-- = [B1[n], B2[n], B3[n]];
1634  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1635  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1636  // ...
1637  class LockstepReverseIterator {
1638  ArrayRef<BasicBlock*> Blocks;
1640  bool Fail;
1641 
1642  public:
1643  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1644  reset();
1645  }
1646 
1647  void reset() {
1648  Fail = false;
1649  Insts.clear();
1650  for (auto *BB : Blocks) {
1651  Instruction *Inst = BB->getTerminator();
1652  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1653  Inst = Inst->getPrevNode();
1654  if (!Inst) {
1655  // Block wasn't big enough.
1656  Fail = true;
1657  return;
1658  }
1659  Insts.push_back(Inst);
1660  }
1661  }
1662 
1663  bool isValid() const {
1664  return !Fail;
1665  }
1666 
1667  void operator--() {
1668  if (Fail)
1669  return;
1670  for (auto *&Inst : Insts) {
1671  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1672  Inst = Inst->getPrevNode();
1673  // Already at beginning of block.
1674  if (!Inst) {
1675  Fail = true;
1676  return;
1677  }
1678  }
1679  }
1680 
1682  return Insts;
1683  }
1684  };
1685 
1686 } // end anonymous namespace
1687 
1688 /// Check whether BB's predecessors end with unconditional branches. If it is
1689 /// true, sink any common code from the predecessors to BB.
1690 /// We also allow one predecessor to end with conditional branch (but no more
1691 /// than one).
1693  // We support two situations:
1694  // (1) all incoming arcs are unconditional
1695  // (2) one incoming arc is conditional
1696  //
1697  // (2) is very common in switch defaults and
1698  // else-if patterns;
1699  //
1700  // if (a) f(1);
1701  // else if (b) f(2);
1702  //
1703  // produces:
1704  //
1705  // [if]
1706  // / \
1707  // [f(1)] [if]
1708  // | | \
1709  // | | |
1710  // | [f(2)]|
1711  // \ | /
1712  // [ end ]
1713  //
1714  // [end] has two unconditional predecessor arcs and one conditional. The
1715  // conditional refers to the implicit empty 'else' arc. This conditional
1716  // arc can also be caused by an empty default block in a switch.
1717  //
1718  // In this case, we attempt to sink code from all *unconditional* arcs.
1719  // If we can sink instructions from these arcs (determined during the scan
1720  // phase below) we insert a common successor for all unconditional arcs and
1721  // connect that to [end], to enable sinking:
1722  //
1723  // [if]
1724  // / \
1725  // [x(1)] [if]
1726  // | | \
1727  // | | \
1728  // | [x(2)] |
1729  // \ / |
1730  // [sink.split] |
1731  // \ /
1732  // [ end ]
1733  //
1734  SmallVector<BasicBlock*,4> UnconditionalPreds;
1735  Instruction *Cond = nullptr;
1736  for (auto *B : predecessors(BB)) {
1737  auto *T = B->getTerminator();
1738  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1739  UnconditionalPreds.push_back(B);
1740  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1741  Cond = T;
1742  else
1743  return false;
1744  }
1745  if (UnconditionalPreds.size() < 2)
1746  return false;
1747 
1748  bool Changed = false;
1749  // We take a two-step approach to tail sinking. First we scan from the end of
1750  // each block upwards in lockstep. If the n'th instruction from the end of each
1751  // block can be sunk, those instructions are added to ValuesToSink and we
1752  // carry on. If we can sink an instruction but need to PHI-merge some operands
1753  // (because they're not identical in each instruction) we add these to
1754  // PHIOperands.
1755  unsigned ScanIdx = 0;
1756  SmallPtrSet<Value*,4> InstructionsToSink;
1758  LockstepReverseIterator LRI(UnconditionalPreds);
1759  while (LRI.isValid() &&
1760  canSinkInstructions(*LRI, PHIOperands)) {
1761  LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1762  << "\n");
1763  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1764  ++ScanIdx;
1765  --LRI;
1766  }
1767 
1768  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1769  unsigned NumPHIdValues = 0;
1770  for (auto *I : *LRI)
1771  for (auto *V : PHIOperands[I])
1772  if (InstructionsToSink.count(V) == 0)
1773  ++NumPHIdValues;
1774  LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1775  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1776  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1777  NumPHIInsts++;
1778 
1779  return NumPHIInsts <= 1;
1780  };
1781 
1782  if (ScanIdx > 0 && Cond) {
1783  // Check if we would actually sink anything first! This mutates the CFG and
1784  // adds an extra block. The goal in doing this is to allow instructions that
1785  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1786  // (such as trunc, add) can be sunk and predicated already. So we check that
1787  // we're going to sink at least one non-speculatable instruction.
1788  LRI.reset();
1789  unsigned Idx = 0;
1790  bool Profitable = false;
1791  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1792  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1793  Profitable = true;
1794  break;
1795  }
1796  --LRI;
1797  ++Idx;
1798  }
1799  if (!Profitable)
1800  return false;
1801 
1802  LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
1803  // We have a conditional edge and we're going to sink some instructions.
1804  // Insert a new block postdominating all blocks we're going to sink from.
1805  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1806  // Edges couldn't be split.
1807  return false;
1808  Changed = true;
1809  }
1810 
1811  // Now that we've analyzed all potential sinking candidates, perform the
1812  // actual sink. We iteratively sink the last non-terminator of the source
1813  // blocks into their common successor unless doing so would require too
1814  // many PHI instructions to be generated (currently only one PHI is allowed
1815  // per sunk instruction).
1816  //
1817  // We can use InstructionsToSink to discount values needing PHI-merging that will
1818  // actually be sunk in a later iteration. This allows us to be more
1819  // aggressive in what we sink. This does allow a false positive where we
1820  // sink presuming a later value will also be sunk, but stop half way through
1821  // and never actually sink it which means we produce more PHIs than intended.
1822  // This is unlikely in practice though.
1823  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1824  LLVM_DEBUG(dbgs() << "SINK: Sink: "
1825  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1826  << "\n");
1827 
1828  // Because we've sunk every instruction in turn, the current instruction to
1829  // sink is always at index 0.
1830  LRI.reset();
1831  if (!ProfitableToSinkInstruction(LRI)) {
1832  // Too many PHIs would be created.
1833  LLVM_DEBUG(
1834  dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1835  break;
1836  }
1837 
1838  if (!sinkLastInstruction(UnconditionalPreds))
1839  return Changed;
1840  NumSinkCommons++;
1841  Changed = true;
1842  }
1843  return Changed;
1844 }
1845 
1846 /// Determine if we can hoist sink a sole store instruction out of a
1847 /// conditional block.
1848 ///
1849 /// We are looking for code like the following:
1850 /// BrBB:
1851 /// store i32 %add, i32* %arrayidx2
1852 /// ... // No other stores or function calls (we could be calling a memory
1853 /// ... // function).
1854 /// %cmp = icmp ult %x, %y
1855 /// br i1 %cmp, label %EndBB, label %ThenBB
1856 /// ThenBB:
1857 /// store i32 %add5, i32* %arrayidx2
1858 /// br label EndBB
1859 /// EndBB:
1860 /// ...
1861 /// We are going to transform this into:
1862 /// BrBB:
1863 /// store i32 %add, i32* %arrayidx2
1864 /// ... //
1865 /// %cmp = icmp ult %x, %y
1866 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1867 /// store i32 %add.add5, i32* %arrayidx2
1868 /// ...
1869 ///
1870 /// \return The pointer to the value of the previous store if the store can be
1871 /// hoisted into the predecessor block. 0 otherwise.
1873  BasicBlock *StoreBB, BasicBlock *EndBB) {
1874  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1875  if (!StoreToHoist)
1876  return nullptr;
1877 
1878  // Volatile or atomic.
1879  if (!StoreToHoist->isSimple())
1880  return nullptr;
1881 
1882  Value *StorePtr = StoreToHoist->getPointerOperand();
1883 
1884  // Look for a store to the same pointer in BrBB.
1885  unsigned MaxNumInstToLookAt = 9;
1886  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1887  if (!MaxNumInstToLookAt)
1888  break;
1889  --MaxNumInstToLookAt;
1890 
1891  // Could be calling an instruction that affects memory like free().
1892  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1893  return nullptr;
1894 
1895  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1896  // Found the previous store make sure it stores to the same location.
1897  if (SI->getPointerOperand() == StorePtr)
1898  // Found the previous store, return its value operand.
1899  return SI->getValueOperand();
1900  return nullptr; // Unknown store.
1901  }
1902  }
1903 
1904  return nullptr;
1905 }
1906 
1907 /// Speculate a conditional basic block flattening the CFG.
1908 ///
1909 /// Note that this is a very risky transform currently. Speculating
1910 /// instructions like this is most often not desirable. Instead, there is an MI
1911 /// pass which can do it with full awareness of the resource constraints.
1912 /// However, some cases are "obvious" and we should do directly. An example of
1913 /// this is speculating a single, reasonably cheap instruction.
1914 ///
1915 /// There is only one distinct advantage to flattening the CFG at the IR level:
1916 /// it makes very common but simplistic optimizations such as are common in
1917 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1918 /// modeling their effects with easier to reason about SSA value graphs.
1919 ///
1920 ///
1921 /// An illustration of this transform is turning this IR:
1922 /// \code
1923 /// BB:
1924 /// %cmp = icmp ult %x, %y
1925 /// br i1 %cmp, label %EndBB, label %ThenBB
1926 /// ThenBB:
1927 /// %sub = sub %x, %y
1928 /// br label BB2
1929 /// EndBB:
1930 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1931 /// ...
1932 /// \endcode
1933 ///
1934 /// Into this IR:
1935 /// \code
1936 /// BB:
1937 /// %cmp = icmp ult %x, %y
1938 /// %sub = sub %x, %y
1939 /// %cond = select i1 %cmp, 0, %sub
1940 /// ...
1941 /// \endcode
1942 ///
1943 /// \returns true if the conditional block is removed.
1945  const TargetTransformInfo &TTI) {
1946  // Be conservative for now. FP select instruction can often be expensive.
1947  Value *BrCond = BI->getCondition();
1948  if (isa<FCmpInst>(BrCond))
1949  return false;
1950 
1951  BasicBlock *BB = BI->getParent();
1952  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1953 
1954  // If ThenBB is actually on the false edge of the conditional branch, remember
1955  // to swap the select operands later.
1956  bool Invert = false;
1957  if (ThenBB != BI->getSuccessor(0)) {
1958  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1959  Invert = true;
1960  }
1961  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1962 
1963  // Keep a count of how many times instructions are used within ThenBB when
1964  // they are candidates for sinking into ThenBB. Specifically:
1965  // - They are defined in BB, and
1966  // - They have no side effects, and
1967  // - All of their uses are in ThenBB.
1968  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1969 
1970  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1971 
1972  unsigned SpeculationCost = 0;
1973  Value *SpeculatedStoreValue = nullptr;
1974  StoreInst *SpeculatedStore = nullptr;
1975  for (BasicBlock::iterator BBI = ThenBB->begin(),
1976  BBE = std::prev(ThenBB->end());
1977  BBI != BBE; ++BBI) {
1978  Instruction *I = &*BBI;
1979  // Skip debug info.
1980  if (isa<DbgInfoIntrinsic>(I)) {
1981  SpeculatedDbgIntrinsics.push_back(I);
1982  continue;
1983  }
1984 
1985  // Only speculatively execute a single instruction (not counting the
1986  // terminator) for now.
1987  ++SpeculationCost;
1988  if (SpeculationCost > 1)
1989  return false;
1990 
1991  // Don't hoist the instruction if it's unsafe or expensive.
1992  if (!isSafeToSpeculativelyExecute(I) &&
1993  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1994  I, BB, ThenBB, EndBB))))
1995  return false;
1996  if (!SpeculatedStoreValue &&
1997  ComputeSpeculationCost(I, TTI) >
1999  return false;
2000 
2001  // Store the store speculation candidate.
2002  if (SpeculatedStoreValue)
2003  SpeculatedStore = cast<StoreInst>(I);
2004 
2005  // Do not hoist the instruction if any of its operands are defined but not
2006  // used in BB. The transformation will prevent the operand from
2007  // being sunk into the use block.
2008  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2009  Instruction *OpI = dyn_cast<Instruction>(*i);
2010  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2011  continue; // Not a candidate for sinking.
2012 
2013  ++SinkCandidateUseCounts[OpI];
2014  }
2015  }
2016 
2017  // Consider any sink candidates which are only used in ThenBB as costs for
2018  // speculation. Note, while we iterate over a DenseMap here, we are summing
2019  // and so iteration order isn't significant.
2021  I = SinkCandidateUseCounts.begin(),
2022  E = SinkCandidateUseCounts.end();
2023  I != E; ++I)
2024  if (I->first->hasNUses(I->second)) {
2025  ++SpeculationCost;
2026  if (SpeculationCost > 1)
2027  return false;
2028  }
2029 
2030  // Check that the PHI nodes can be converted to selects.
2031  bool HaveRewritablePHIs = false;
2032  for (PHINode &PN : EndBB->phis()) {
2033  Value *OrigV = PN.getIncomingValueForBlock(BB);
2034  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2035 
2036  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2037  // Skip PHIs which are trivial.
2038  if (ThenV == OrigV)
2039  continue;
2040 
2041  // Don't convert to selects if we could remove undefined behavior instead.
2042  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2043  passingValueIsAlwaysUndefined(ThenV, &PN))
2044  return false;
2045 
2046  HaveRewritablePHIs = true;
2047  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2048  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2049  if (!OrigCE && !ThenCE)
2050  continue; // Known safe and cheap.
2051 
2052  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2053  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2054  return false;
2055  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2056  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2057  unsigned MaxCost =
2059  if (OrigCost + ThenCost > MaxCost)
2060  return false;
2061 
2062  // Account for the cost of an unfolded ConstantExpr which could end up
2063  // getting expanded into Instructions.
2064  // FIXME: This doesn't account for how many operations are combined in the
2065  // constant expression.
2066  ++SpeculationCost;
2067  if (SpeculationCost > 1)
2068  return false;
2069  }
2070 
2071  // If there are no PHIs to process, bail early. This helps ensure idempotence
2072  // as well.
2073  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2074  return false;
2075 
2076  // If we get here, we can hoist the instruction and if-convert.
2077  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2078 
2079  // Insert a select of the value of the speculated store.
2080  if (SpeculatedStoreValue) {
2081  IRBuilder<NoFolder> Builder(BI);
2082  Value *TrueV = SpeculatedStore->getValueOperand();
2083  Value *FalseV = SpeculatedStoreValue;
2084  if (Invert)
2085  std::swap(TrueV, FalseV);
2086  Value *S = Builder.CreateSelect(
2087  BrCond, TrueV, FalseV, "spec.store.select", BI);
2088  SpeculatedStore->setOperand(0, S);
2089  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2090  SpeculatedStore->getDebugLoc());
2091  }
2092 
2093  // Metadata can be dependent on the condition we are hoisting above.
2094  // Conservatively strip all metadata on the instruction.
2095  for (auto &I : *ThenBB)
2097 
2098  // Hoist the instructions.
2099  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2100  ThenBB->begin(), std::prev(ThenBB->end()));
2101 
2102  // Insert selects and rewrite the PHI operands.
2103  IRBuilder<NoFolder> Builder(BI);
2104  for (PHINode &PN : EndBB->phis()) {
2105  unsigned OrigI = PN.getBasicBlockIndex(BB);
2106  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2107  Value *OrigV = PN.getIncomingValue(OrigI);
2108  Value *ThenV = PN.getIncomingValue(ThenI);
2109 
2110  // Skip PHIs which are trivial.
2111  if (OrigV == ThenV)
2112  continue;
2113 
2114  // Create a select whose true value is the speculatively executed value and
2115  // false value is the preexisting value. Swap them if the branch
2116  // destinations were inverted.
2117  Value *TrueV = ThenV, *FalseV = OrigV;
2118  if (Invert)
2119  std::swap(TrueV, FalseV);
2120  Value *V = Builder.CreateSelect(
2121  BrCond, TrueV, FalseV, "spec.select", BI);
2122  PN.setIncomingValue(OrigI, V);
2123  PN.setIncomingValue(ThenI, V);
2124  }
2125 
2126  // Remove speculated dbg intrinsics.
2127  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2128  // dbg value for the different flows and inserting it after the select.
2129  for (Instruction *I : SpeculatedDbgIntrinsics)
2130  I->eraseFromParent();
2131 
2132  ++NumSpeculations;
2133  return true;
2134 }
2135 
2136 /// Return true if we can thread a branch across this block.
2138  unsigned Size = 0;
2139 
2140  for (Instruction &I : BB->instructionsWithoutDebug()) {
2141  if (Size > 10)
2142  return false; // Don't clone large BB's.
2143  ++Size;
2144 
2145  // We can only support instructions that do not define values that are
2146  // live outside of the current basic block.
2147  for (User *U : I.users()) {
2148  Instruction *UI = cast<Instruction>(U);
2149  if (UI->getParent() != BB || isa<PHINode>(UI))
2150  return false;
2151  }
2152 
2153  // Looks ok, continue checking.
2154  }
2155 
2156  return true;
2157 }
2158 
2159 /// If we have a conditional branch on a PHI node value that is defined in the
2160 /// same block as the branch and if any PHI entries are constants, thread edges
2161 /// corresponding to that entry to be branches to their ultimate destination.
2162 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2163  AssumptionCache *AC) {
2164  BasicBlock *BB = BI->getParent();
2165  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2166  // NOTE: we currently cannot transform this case if the PHI node is used
2167  // outside of the block.
2168  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2169  return false;
2170 
2171  // Degenerate case of a single entry PHI.
2172  if (PN->getNumIncomingValues() == 1) {
2174  return true;
2175  }
2176 
2177  // Now we know that this block has multiple preds and two succs.
2179  return false;
2180 
2181  // Can't fold blocks that contain noduplicate or convergent calls.
2182  if (any_of(*BB, [](const Instruction &I) {
2183  const CallInst *CI = dyn_cast<CallInst>(&I);
2184  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2185  }))
2186  return false;
2187 
2188  // Okay, this is a simple enough basic block. See if any phi values are
2189  // constants.
2190  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2192  if (!CB || !CB->getType()->isIntegerTy(1))
2193  continue;
2194 
2195  // Okay, we now know that all edges from PredBB should be revectored to
2196  // branch to RealDest.
2197  BasicBlock *PredBB = PN->getIncomingBlock(i);
2198  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2199 
2200  if (RealDest == BB)
2201  continue; // Skip self loops.
2202  // Skip if the predecessor's terminator is an indirect branch.
2203  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2204  continue;
2205 
2206  // The dest block might have PHI nodes, other predecessors and other
2207  // difficult cases. Instead of being smart about this, just insert a new
2208  // block that jumps to the destination block, effectively splitting
2209  // the edge we are about to create.
2210  BasicBlock *EdgeBB =
2211  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2212  RealDest->getParent(), RealDest);
2213  BranchInst *CritEdgeBranch = BranchInst::Create(RealDest, EdgeBB);
2214  CritEdgeBranch->setDebugLoc(BI->getDebugLoc());
2215 
2216  // Update PHI nodes.
2217  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2218 
2219  // BB may have instructions that are being threaded over. Clone these
2220  // instructions into EdgeBB. We know that there will be no uses of the
2221  // cloned instructions outside of EdgeBB.
2222  BasicBlock::iterator InsertPt = EdgeBB->begin();
2223  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2224  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2225  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2226  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2227  continue;
2228  }
2229  // Clone the instruction.
2230  Instruction *N = BBI->clone();
2231  if (BBI->hasName())
2232  N->setName(BBI->getName() + ".c");
2233 
2234  // Update operands due to translation.
2235  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2236  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2237  if (PI != TranslateMap.end())
2238  *i = PI->second;
2239  }
2240 
2241  // Check for trivial simplification.
2242  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2243  if (!BBI->use_empty())
2244  TranslateMap[&*BBI] = V;
2245  if (!N->mayHaveSideEffects()) {
2246  N->deleteValue(); // Instruction folded away, don't need actual inst
2247  N = nullptr;
2248  }
2249  } else {
2250  if (!BBI->use_empty())
2251  TranslateMap[&*BBI] = N;
2252  }
2253  // Insert the new instruction into its new home.
2254  if (N)
2255  EdgeBB->getInstList().insert(InsertPt, N);
2256 
2257  // Register the new instruction with the assumption cache if necessary.
2258  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2259  if (II->getIntrinsicID() == Intrinsic::assume)
2260  AC->registerAssumption(II);
2261  }
2262 
2263  // Loop over all of the edges from PredBB to BB, changing them to branch
2264  // to EdgeBB instead.
2265  Instruction *PredBBTI = PredBB->getTerminator();
2266  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2267  if (PredBBTI->getSuccessor(i) == BB) {
2268  BB->removePredecessor(PredBB);
2269  PredBBTI->setSuccessor(i, EdgeBB);
2270  }
2271 
2272  // Recurse, simplifying any other constants.
2273  return FoldCondBranchOnPHI(BI, DL, AC) || true;
2274  }
2275 
2276  return false;
2277 }
2278 
2279 /// Given a BB that starts with the specified two-entry PHI node,
2280 /// see if we can eliminate it.
2282  const DataLayout &DL) {
2283  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2284  // statement", which has a very simple dominance structure. Basically, we
2285  // are trying to find the condition that is being branched on, which
2286  // subsequently causes this merge to happen. We really want control
2287  // dependence information for this check, but simplifycfg can't keep it up
2288  // to date, and this catches most of the cases we care about anyway.
2289  BasicBlock *BB = PN->getParent();
2290  const Function *Fn = BB->getParent();
2291  if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
2292  return false;
2293 
2294  BasicBlock *IfTrue, *IfFalse;
2295  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2296  if (!IfCond ||
2297  // Don't bother if the branch will be constant folded trivially.
2298  isa<ConstantInt>(IfCond))
2299  return false;
2300 
2301  // Okay, we found that we can merge this two-entry phi node into a select.
2302  // Doing so would require us to fold *all* two entry phi nodes in this block.
2303  // At some point this becomes non-profitable (particularly if the target
2304  // doesn't support cmov's). Only do this transformation if there are two or
2305  // fewer PHI nodes in this block.
2306  unsigned NumPhis = 0;
2307  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2308  if (NumPhis > 2)
2309  return false;
2310 
2311  // Loop over the PHI's seeing if we can promote them all to select
2312  // instructions. While we are at it, keep track of the instructions
2313  // that need to be moved to the dominating block.
2314  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2315  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2316  MaxCostVal1 = PHINodeFoldingThreshold;
2317  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2318  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2319 
2320  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2321  PHINode *PN = cast<PHINode>(II++);
2322  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2323  PN->replaceAllUsesWith(V);
2324  PN->eraseFromParent();
2325  continue;
2326  }
2327 
2328  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2329  MaxCostVal0, TTI) ||
2330  !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2331  MaxCostVal1, TTI))
2332  return false;
2333  }
2334 
2335  // If we folded the first phi, PN dangles at this point. Refresh it. If
2336  // we ran out of PHIs then we simplified them all.
2337  PN = dyn_cast<PHINode>(BB->begin());
2338  if (!PN)
2339  return true;
2340 
2341  // Don't fold i1 branches on PHIs which contain binary operators. These can
2342  // often be turned into switches and other things.
2343  if (PN->getType()->isIntegerTy(1) &&
2344  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2345  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2346  isa<BinaryOperator>(IfCond)))
2347  return false;
2348 
2349  // If all PHI nodes are promotable, check to make sure that all instructions
2350  // in the predecessor blocks can be promoted as well. If not, we won't be able
2351  // to get rid of the control flow, so it's not worth promoting to select
2352  // instructions.
2353  BasicBlock *DomBlock = nullptr;
2354  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2355  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2356  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2357  IfBlock1 = nullptr;
2358  } else {
2359  DomBlock = *pred_begin(IfBlock1);
2360  for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2361  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2362  // This is not an aggressive instruction that we can promote.
2363  // Because of this, we won't be able to get rid of the control flow, so
2364  // the xform is not worth it.
2365  return false;
2366  }
2367  }
2368 
2369  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2370  IfBlock2 = nullptr;
2371  } else {
2372  DomBlock = *pred_begin(IfBlock2);
2373  for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2374  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2375  // This is not an aggressive instruction that we can promote.
2376  // Because of this, we won't be able to get rid of the control flow, so
2377  // the xform is not worth it.
2378  return false;
2379  }
2380  }
2381 
2382  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2383  << " T: " << IfTrue->getName()
2384  << " F: " << IfFalse->getName() << "\n");
2385 
2386  // If we can still promote the PHI nodes after this gauntlet of tests,
2387  // do all of the PHI's now.
2388  Instruction *InsertPt = DomBlock->getTerminator();
2389  IRBuilder<NoFolder> Builder(InsertPt);
2390 
2391  // Move all 'aggressive' instructions, which are defined in the
2392  // conditional parts of the if's up to the dominating block.
2393  if (IfBlock1)
2394  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2395  if (IfBlock2)
2396  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2397 
2398  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2399  // Change the PHI node into a select instruction.
2400  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2401  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2402 
2403  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2404  PN->replaceAllUsesWith(Sel);
2405  Sel->takeName(PN);
2406  PN->eraseFromParent();
2407  }
2408 
2409  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2410  // has been flattened. Change DomBlock to jump directly to our new block to
2411  // avoid other simplifycfg's kicking in on the diamond.
2412  Instruction *OldTI = DomBlock->getTerminator();
2413  Builder.SetInsertPoint(OldTI);
2414  Builder.CreateBr(BB);
2415  OldTI->eraseFromParent();
2416  return true;
2417 }
2418 
2419 /// If we found a conditional branch that goes to two returning blocks,
2420 /// try to merge them together into one return,
2421 /// introducing a select if the return values disagree.
2423  IRBuilder<> &Builder) {
2424  assert(BI->isConditional() && "Must be a conditional branch");
2425  BasicBlock *TrueSucc = BI->getSuccessor(0);
2426  BasicBlock *FalseSucc = BI->getSuccessor(1);
2427  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2428  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2429 
2430  // Check to ensure both blocks are empty (just a return) or optionally empty
2431  // with PHI nodes. If there are other instructions, merging would cause extra
2432  // computation on one path or the other.
2433  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2434  return false;
2435  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2436  return false;
2437 
2438  Builder.SetInsertPoint(BI);
2439  // Okay, we found a branch that is going to two return nodes. If
2440  // there is no return value for this function, just change the
2441  // branch into a return.
2442  if (FalseRet->getNumOperands() == 0) {
2443  TrueSucc->removePredecessor(BI->getParent());
2444  FalseSucc->removePredecessor(BI->getParent());
2445  Builder.CreateRetVoid();
2447  return true;
2448  }
2449 
2450  // Otherwise, figure out what the true and false return values are
2451  // so we can insert a new select instruction.
2452  Value *TrueValue = TrueRet->getReturnValue();
2453  Value *FalseValue = FalseRet->getReturnValue();
2454 
2455  // Unwrap any PHI nodes in the return blocks.
2456  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2457  if (TVPN->getParent() == TrueSucc)
2458  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2459  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2460  if (FVPN->getParent() == FalseSucc)
2461  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2462 
2463  // In order for this transformation to be safe, we must be able to
2464  // unconditionally execute both operands to the return. This is
2465  // normally the case, but we could have a potentially-trapping
2466  // constant expression that prevents this transformation from being
2467  // safe.
2468  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2469  if (TCV->canTrap())
2470  return false;
2471  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2472  if (FCV->canTrap())
2473  return false;
2474 
2475  // Okay, we collected all the mapped values and checked them for sanity, and
2476  // defined to really do this transformation. First, update the CFG.
2477  TrueSucc->removePredecessor(BI->getParent());
2478  FalseSucc->removePredecessor(BI->getParent());
2479 
2480  // Insert select instructions where needed.
2481  Value *BrCond = BI->getCondition();
2482  if (TrueValue) {
2483  // Insert a select if the results differ.
2484  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2485  } else if (isa<UndefValue>(TrueValue)) {
2486  TrueValue = FalseValue;
2487  } else {
2488  TrueValue =
2489  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2490  }
2491  }
2492 
2493  Value *RI =
2494  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2495 
2496  (void)RI;
2497 
2498  LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2499  << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: "
2500  << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2501 
2503 
2504  return true;
2505 }
2506 
2507 /// Return true if the given instruction is available
2508 /// in its predecessor block. If yes, the instruction will be removed.
2510  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2511  return false;
2512  for (Instruction &I : *PB) {
2513  Instruction *PBI = &I;
2514  // Check whether Inst and PBI generate the same value.
2515  if (Inst->isIdenticalTo(PBI)) {
2516  Inst->replaceAllUsesWith(PBI);
2517  Inst->eraseFromParent();
2518  return true;
2519  }
2520  }
2521  return false;
2522 }
2523 
2524 /// Return true if either PBI or BI has branch weight available, and store
2525 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2526 /// not have branch weight, use 1:1 as its weight.
2528  uint64_t &PredTrueWeight,
2529  uint64_t &PredFalseWeight,
2530  uint64_t &SuccTrueWeight,
2531  uint64_t &SuccFalseWeight) {
2532  bool PredHasWeights =
2533  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2534  bool SuccHasWeights =
2535  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2536  if (PredHasWeights || SuccHasWeights) {
2537  if (!PredHasWeights)
2538  PredTrueWeight = PredFalseWeight = 1;
2539  if (!SuccHasWeights)
2540  SuccTrueWeight = SuccFalseWeight = 1;
2541  return true;
2542  } else {
2543  return false;
2544  }
2545 }
2546 
2547 /// If this basic block is simple enough, and if a predecessor branches to us
2548 /// and one of our successors, fold the block into the predecessor and use
2549 /// logical operations to pick the right destination.
2550 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2551  BasicBlock *BB = BI->getParent();
2552 
2553  const unsigned PredCount = pred_size(BB);
2554 
2555  Instruction *Cond = nullptr;
2556  if (BI->isConditional())
2557  Cond = dyn_cast<Instruction>(BI->getCondition());
2558  else {
2559  // For unconditional branch, check for a simple CFG pattern, where
2560  // BB has a single predecessor and BB's successor is also its predecessor's
2561  // successor. If such pattern exists, check for CSE between BB and its
2562  // predecessor.
2563  if (BasicBlock *PB = BB->getSinglePredecessor())
2564  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2565  if (PBI->isConditional() &&
2566  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2567  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2568  for (auto I = BB->instructionsWithoutDebug().begin(),
2569  E = BB->instructionsWithoutDebug().end();
2570  I != E;) {
2571  Instruction *Curr = &*I++;
2572  if (isa<CmpInst>(Curr)) {
2573  Cond = Curr;
2574  break;
2575  }
2576  // Quit if we can't remove this instruction.
2577  if (!tryCSEWithPredecessor(Curr, PB))
2578  return false;
2579  }
2580  }
2581 
2582  if (!Cond)
2583  return false;
2584  }
2585 
2586  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2587  Cond->getParent() != BB || !Cond->hasOneUse())
2588  return false;
2589 
2590  // Make sure the instruction after the condition is the cond branch.
2591  BasicBlock::iterator CondIt = ++Cond->getIterator();
2592 
2593  // Ignore dbg intrinsics.
2594  while (isa<DbgInfoIntrinsic>(CondIt))
2595  ++CondIt;
2596 
2597  if (&*CondIt != BI)
2598  return false;
2599 
2600  // Only allow this transformation if computing the condition doesn't involve
2601  // too many instructions and these involved instructions can be executed
2602  // unconditionally. We denote all involved instructions except the condition
2603  // as "bonus instructions", and only allow this transformation when the
2604  // number of the bonus instructions we'll need to create when cloning into
2605  // each predecessor does not exceed a certain threshold.
2606  unsigned NumBonusInsts = 0;
2607  for (auto I = BB->begin(); Cond != &*I; ++I) {
2608  // Ignore dbg intrinsics.
2609  if (isa<DbgInfoIntrinsic>(I))
2610  continue;
2611  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2612  return false;
2613  // I has only one use and can be executed unconditionally.
2615  if (User == nullptr || User->getParent() != BB)
2616  return false;
2617  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2618  // to use any other instruction, User must be an instruction between next(I)
2619  // and Cond.
2620 
2621  // Account for the cost of duplicating this instruction into each
2622  // predecessor.
2623  NumBonusInsts += PredCount;
2624  // Early exits once we reach the limit.
2625  if (NumBonusInsts > BonusInstThreshold)
2626  return false;
2627  }
2628 
2629  // Cond is known to be a compare or binary operator. Check to make sure that
2630  // neither operand is a potentially-trapping constant expression.
2631  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2632  if (CE->canTrap())
2633  return false;
2634  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2635  if (CE->canTrap())
2636  return false;
2637 
2638  // Finally, don't infinitely unroll conditional loops.
2639  BasicBlock *TrueDest = BI->getSuccessor(0);
2640  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2641  if (TrueDest == BB || FalseDest == BB)
2642  return false;
2643 
2644  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2645  BasicBlock *PredBlock = *PI;
2646  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2647 
2648  // Check that we have two conditional branches. If there is a PHI node in
2649  // the common successor, verify that the same value flows in from both
2650  // blocks.
2652  if (!PBI || PBI->isUnconditional() ||
2653  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2654  (!BI->isConditional() &&
2655  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2656  continue;
2657 
2658  // Determine if the two branches share a common destination.
2659  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2660  bool InvertPredCond = false;
2661 
2662  if (BI->isConditional()) {
2663  if (PBI->getSuccessor(0) == TrueDest) {
2664  Opc = Instruction::Or;
2665  } else if (PBI->getSuccessor(1) == FalseDest) {
2666  Opc = Instruction::And;
2667  } else if (PBI->getSuccessor(0) == FalseDest) {
2668  Opc = Instruction::And;
2669  InvertPredCond = true;
2670  } else if (PBI->getSuccessor(1) == TrueDest) {
2671  Opc = Instruction::Or;
2672  InvertPredCond = true;
2673  } else {
2674  continue;
2675  }
2676  } else {
2677  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2678  continue;
2679  }
2680 
2681  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2682  IRBuilder<> Builder(PBI);
2683 
2684  // If we need to invert the condition in the pred block to match, do so now.
2685  if (InvertPredCond) {
2686  Value *NewCond = PBI->getCondition();
2687 
2688  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2689  CmpInst *CI = cast<CmpInst>(NewCond);
2690  CI->setPredicate(CI->getInversePredicate());
2691  } else {
2692  NewCond =
2693  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2694  }
2695 
2696  PBI->setCondition(NewCond);
2697  PBI->swapSuccessors();
2698  }
2699 
2700  // If we have bonus instructions, clone them into the predecessor block.
2701  // Note that there may be multiple predecessor blocks, so we cannot move
2702  // bonus instructions to a predecessor block.
2703  ValueToValueMapTy VMap; // maps original values to cloned values
2704  // We already make sure Cond is the last instruction before BI. Therefore,
2705  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2706  // instructions.
2707  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2708  if (isa<DbgInfoIntrinsic>(BonusInst))
2709  continue;
2710  Instruction *NewBonusInst = BonusInst->clone();
2711  RemapInstruction(NewBonusInst, VMap,
2713  VMap[&*BonusInst] = NewBonusInst;
2714 
2715  // If we moved a load, we cannot any longer claim any knowledge about
2716  // its potential value. The previous information might have been valid
2717  // only given the branch precondition.
2718  // For an analogous reason, we must also drop all the metadata whose
2719  // semantics we don't understand.
2720  NewBonusInst->dropUnknownNonDebugMetadata();
2721 
2722  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2723  NewBonusInst->takeName(&*BonusInst);
2724  BonusInst->setName(BonusInst->getName() + ".old");
2725  }
2726 
2727  // Clone Cond into the predecessor basic block, and or/and the
2728  // two conditions together.
2729  Instruction *CondInPred = Cond->clone();
2730  RemapInstruction(CondInPred, VMap,
2732  PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2733  CondInPred->takeName(Cond);
2734  Cond->setName(CondInPred->getName() + ".old");
2735 
2736  if (BI->isConditional()) {
2737  Instruction *NewCond = cast<Instruction>(
2738  Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2739  PBI->setCondition(NewCond);
2740 
2741  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2742  bool HasWeights =
2743  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2744  SuccTrueWeight, SuccFalseWeight);
2745  SmallVector<uint64_t, 8> NewWeights;
2746 
2747  if (PBI->getSuccessor(0) == BB) {
2748  if (HasWeights) {
2749  // PBI: br i1 %x, BB, FalseDest
2750  // BI: br i1 %y, TrueDest, FalseDest
2751  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2752  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2753  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2754  // TrueWeight for PBI * FalseWeight for BI.
2755  // We assume that total weights of a BranchInst can fit into 32 bits.
2756  // Therefore, we will not have overflow using 64-bit arithmetic.
2757  NewWeights.push_back(PredFalseWeight *
2758  (SuccFalseWeight + SuccTrueWeight) +
2759  PredTrueWeight * SuccFalseWeight);
2760  }
2761  AddPredecessorToBlock(TrueDest, PredBlock, BB);
2762  PBI->setSuccessor(0, TrueDest);
2763  }
2764  if (PBI->getSuccessor(1) == BB) {
2765  if (HasWeights) {
2766  // PBI: br i1 %x, TrueDest, BB
2767  // BI: br i1 %y, TrueDest, FalseDest
2768  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2769  // FalseWeight for PBI * TrueWeight for BI.
2770  NewWeights.push_back(PredTrueWeight *
2771  (SuccFalseWeight + SuccTrueWeight) +
2772  PredFalseWeight * SuccTrueWeight);
2773  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2774  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2775  }
2776  AddPredecessorToBlock(FalseDest, PredBlock, BB);
2777  PBI->setSuccessor(1, FalseDest);
2778  }
2779  if (NewWeights.size() == 2) {
2780  // Halve the weights if any of them cannot fit in an uint32_t
2781  FitWeights(NewWeights);
2782 
2783  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2784  NewWeights.end());
2785  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2786  } else
2787  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2788  } else {
2789  // Update PHI nodes in the common successors.
2790  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2791  ConstantInt *PBI_C = cast<ConstantInt>(
2792  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2793  assert(PBI_C->getType()->isIntegerTy(1));
2794  Instruction *MergedCond = nullptr;
2795  if (PBI->getSuccessor(0) == TrueDest) {
2796  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2797  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2798  // is false: !PBI_Cond and BI_Value
2799  Instruction *NotCond = cast<Instruction>(
2800  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2801  MergedCond = cast<Instruction>(
2802  Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2803  "and.cond"));
2804  if (PBI_C->isOne())
2805  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2806  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2807  } else {
2808  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2809  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2810  // is false: PBI_Cond and BI_Value
2811  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2812  Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2813  if (PBI_C->isOne()) {
2814  Instruction *NotCond = cast<Instruction>(
2815  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2816  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2817  Instruction::Or, NotCond, MergedCond, "or.cond"));
2818  }
2819  }
2820  // Update PHI Node.
2821  PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2822  MergedCond);
2823  }
2824  // Change PBI from Conditional to Unconditional.
2825  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2827  PBI = New_PBI;
2828  }
2829 
2830  // If BI was a loop latch, it may have had associated loop metadata.
2831  // We need to copy it to the new latch, that is, PBI.
2832  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2833  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2834 
2835  // TODO: If BB is reachable from all paths through PredBlock, then we
2836  // could replace PBI's branch probabilities with BI's.
2837 
2838  // Copy any debug value intrinsics into the end of PredBlock.
2839  for (Instruction &I : *BB)
2840  if (isa<DbgInfoIntrinsic>(I))
2841  I.clone()->insertBefore(PBI);
2842 
2843  return true;
2844  }
2845  return false;
2846 }
2847 
2848 // If there is only one store in BB1 and BB2, return it, otherwise return
2849 // nullptr.
2851  StoreInst *S = nullptr;
2852  for (auto *BB : {BB1, BB2}) {
2853  if (!BB)
2854  continue;
2855  for (auto &I : *BB)
2856  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2857  if (S)
2858  // Multiple stores seen.
2859  return nullptr;
2860  else
2861  S = SI;
2862  }
2863  }
2864  return S;
2865 }
2866 
2868  Value *AlternativeV = nullptr) {
2869  // PHI is going to be a PHI node that allows the value V that is defined in
2870  // BB to be referenced in BB's only successor.
2871  //
2872  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2873  // doesn't matter to us what the other operand is (it'll never get used). We
2874  // could just create a new PHI with an undef incoming value, but that could
2875  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2876  // other PHI. So here we directly look for some PHI in BB's successor with V
2877  // as an incoming operand. If we find one, we use it, else we create a new
2878  // one.
2879  //
2880  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2881  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2882  // where OtherBB is the single other predecessor of BB's only successor.
2883  PHINode *PHI = nullptr;
2884  BasicBlock *Succ = BB->getSingleSuccessor();
2885 
2886  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2887  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2888  PHI = cast<PHINode>(I);
2889  if (!AlternativeV)
2890  break;
2891 
2892  assert(Succ->hasNPredecessors(2));
2893  auto PredI = pred_begin(Succ);
2894  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2895  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2896  break;
2897  PHI = nullptr;
2898  }
2899  if (PHI)
2900  return PHI;
2901 
2902  // If V is not an instruction defined in BB, just return it.
2903  if (!AlternativeV &&
2904  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2905  return V;
2906 
2907  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2908  PHI->addIncoming(V, BB);
2909  for (BasicBlock *PredBB : predecessors(Succ))
2910  if (PredBB != BB)
2911  PHI->addIncoming(
2912  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2913  return PHI;
2914 }
2915 
2917  BasicBlock *QTB, BasicBlock *QFB,
2918  BasicBlock *PostBB, Value *Address,
2919  bool InvertPCond, bool InvertQCond,
2920  const DataLayout &DL) {
2921  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2922  return Operator::getOpcode(&I) == Instruction::BitCast &&
2923  I.getType()->isPointerTy();
2924  };
2925 
2926  // If we're not in aggressive mode, we only optimize if we have some
2927  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2928  auto IsWorthwhile = [&](BasicBlock *BB) {
2929  if (!BB)
2930  return true;
2931  // Heuristic: if the block can be if-converted/phi-folded and the
2932  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2933  // thread this store.
2934  unsigned N = 0;
2935  for (auto &I : BB->instructionsWithoutDebug()) {
2936  // Cheap instructions viable for folding.
2937  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2938  isa<StoreInst>(I))
2939  ++N;
2940  // Free instructions.
2941  else if (I.isTerminator() || IsaBitcastOfPointerType(I))
2942  continue;
2943  else
2944  return false;
2945  }
2946  // The store we want to merge is counted in N, so add 1 to make sure
2947  // we're counting the instructions that would be left.
2948  return N <= (PHINodeFoldingThreshold + 1);
2949  };
2950 
2952  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2953  !IsWorthwhile(QFB)))
2954  return false;
2955 
2956  // For every pointer, there must be exactly two stores, one coming from
2957  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2958  // store (to any address) in PTB,PFB or QTB,QFB.
2959  // FIXME: We could relax this restriction with a bit more work and performance
2960  // testing.
2961  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2962  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2963  if (!PStore || !QStore)
2964  return false;
2965 
2966  // Now check the stores are compatible.
2967  if (!QStore->isUnordered() || !PStore->isUnordered())
2968  return false;
2969 
2970  // Check that sinking the store won't cause program behavior changes. Sinking
2971  // the store out of the Q blocks won't change any behavior as we're sinking
2972  // from a block to its unconditional successor. But we're moving a store from
2973  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2974  // So we need to check that there are no aliasing loads or stores in
2975  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2976  // operations between PStore and the end of its parent block.
2977  //
2978  // The ideal way to do this is to query AliasAnalysis, but we don't
2979  // preserve AA currently so that is dangerous. Be super safe and just
2980  // check there are no other memory operations at all.
2981  for (auto &I : *QFB->getSinglePredecessor())
2982  if (I.mayReadOrWriteMemory())
2983  return false;
2984  for (auto &I : *QFB)
2985  if (&I != QStore && I.mayReadOrWriteMemory())
2986  return false;
2987  if (QTB)
2988  for (auto &I : *QTB)
2989  if (&I != QStore && I.mayReadOrWriteMemory())
2990  return false;
2991  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2992  I != E; ++I)
2993  if (&*I != PStore && I->mayReadOrWriteMemory())
2994  return false;
2995 
2996  // If PostBB has more than two predecessors, we need to split it so we can
2997  // sink the store.
2998  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
2999  // We know that QFB's only successor is PostBB. And QFB has a single
3000  // predecessor. If QTB exists, then its only successor is also PostBB.
3001  // If QTB does not exist, then QFB's only predecessor has a conditional
3002  // branch to QFB and PostBB.
3003  BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3004  BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3005  "condstore.split");
3006  if (!NewBB)
3007  return false;
3008  PostBB = NewBB;
3009  }
3010 
3011  // OK, we're going to sink the stores to PostBB. The store has to be
3012  // conditional though, so first create the predicate.
3013  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3014  ->getCondition();
3015  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3016  ->getCondition();
3017 
3019  PStore->getParent());
3021  QStore->getParent(), PPHI);
3022 
3023  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3024 
3025  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3026  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3027 
3028  if (InvertPCond)
3029  PPred = QB.CreateNot(PPred);
3030  if (InvertQCond)
3031  QPred = QB.CreateNot(QPred);
3032  Value *CombinedPred = QB.CreateOr(PPred, QPred);
3033 
3034  auto *T =
3035  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3036  QB.SetInsertPoint(T);
3037  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3038  AAMDNodes AAMD;
3039  PStore->getAAMetadata(AAMD, /*Merge=*/false);
3040  PStore->getAAMetadata(AAMD, /*Merge=*/true);
3041  SI->setAAMetadata(AAMD);
3042  unsigned PAlignment = PStore->getAlignment();
3043  unsigned QAlignment = QStore->getAlignment();
3044  unsigned TypeAlignment =
3046  unsigned MinAlignment;
3047  unsigned MaxAlignment;
3048  std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3049  // Choose the minimum alignment. If we could prove both stores execute, we
3050  // could use biggest one. In this case, though, we only know that one of the
3051  // stores executes. And we don't know it's safe to take the alignment from a
3052  // store that doesn't execute.
3053  if (MinAlignment != 0) {
3054  // Choose the minimum of all non-zero alignments.
3055  SI->setAlignment(MinAlignment);
3056  } else if (MaxAlignment != 0) {
3057  // Choose the minimal alignment between the non-zero alignment and the ABI
3058  // default alignment for the type of the stored value.
3059  SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3060  } else {
3061  // If both alignments are zero, use ABI default alignment for the type of
3062  // the stored value.
3063  SI->setAlignment(TypeAlignment);
3064  }
3065 
3066  QStore->eraseFromParent();
3067  PStore->eraseFromParent();
3068 
3069  return true;
3070 }
3071 
3073  const DataLayout &DL) {
3074  // The intention here is to find diamonds or triangles (see below) where each
3075  // conditional block contains a store to the same address. Both of these
3076  // stores are conditional, so they can't be unconditionally sunk. But it may
3077  // be profitable to speculatively sink the stores into one merged store at the
3078  // end, and predicate the merged store on the union of the two conditions of
3079  // PBI and QBI.
3080  //
3081  // This can reduce the number of stores executed if both of the conditions are
3082  // true, and can allow the blocks to become small enough to be if-converted.
3083  // This optimization will also chain, so that ladders of test-and-set
3084  // sequences can be if-converted away.
3085  //
3086  // We only deal with simple diamonds or triangles:
3087  //
3088  // PBI or PBI or a combination of the two
3089  // / \ | \
3090  // PTB PFB | PFB
3091  // \ / | /
3092  // QBI QBI
3093  // / \ | \
3094  // QTB QFB | QFB
3095  // \ / | /
3096  // PostBB PostBB
3097  //
3098  // We model triangles as a type of diamond with a nullptr "true" block.
3099  // Triangles are canonicalized so that the fallthrough edge is represented by
3100  // a true condition, as in the diagram above.
3101  BasicBlock *PTB = PBI->getSuccessor(0);
3102  BasicBlock *PFB = PBI->getSuccessor(1);
3103  BasicBlock *QTB = QBI->getSuccessor(0);
3104  BasicBlock *QFB = QBI->getSuccessor(1);
3105  BasicBlock *PostBB = QFB->getSingleSuccessor();
3106 
3107  // Make sure we have a good guess for PostBB. If QTB's only successor is
3108  // QFB, then QFB is a better PostBB.
3109  if (QTB->getSingleSuccessor() == QFB)
3110  PostBB = QFB;
3111 
3112  // If we couldn't find a good PostBB, stop.
3113  if (!PostBB)
3114  return false;
3115 
3116  bool InvertPCond = false, InvertQCond = false;
3117  // Canonicalize fallthroughs to the true branches.
3118  if (PFB == QBI->getParent()) {
3119  std::swap(PFB, PTB);
3120  InvertPCond = true;
3121  }
3122  if (QFB == PostBB) {
3123  std::swap(QFB, QTB);
3124  InvertQCond = true;
3125  }
3126 
3127  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3128  // and QFB may not. Model fallthroughs as a nullptr block.
3129  if (PTB == QBI->getParent())
3130  PTB = nullptr;
3131  if (QTB == PostBB)
3132  QTB = nullptr;
3133 
3134  // Legality bailouts. We must have at least the non-fallthrough blocks and
3135  // the post-dominating block, and the non-fallthroughs must only have one
3136  // predecessor.
3137  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3138  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3139  };
3140  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3141  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3142  return false;
3143  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3144  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3145  return false;
3146  if (!QBI->getParent()->hasNUses(2))
3147  return false;
3148 
3149  // OK, this is a sequence of two diamonds or triangles.
3150  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3151  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3152  for (auto *BB : {PTB, PFB}) {
3153  if (!BB)
3154  continue;
3155  for (auto &I : *BB)
3156  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3157  PStoreAddresses.insert(SI->getPointerOperand());
3158  }
3159  for (auto *BB : {QTB, QFB}) {
3160  if (!BB)
3161  continue;
3162  for (auto &I : *BB)
3163  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3164  QStoreAddresses.insert(SI->getPointerOperand());
3165  }
3166 
3167  set_intersect(PStoreAddresses, QStoreAddresses);
3168  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3169  // clear what it contains.
3170  auto &CommonAddresses = PStoreAddresses;
3171 
3172  bool Changed = false;
3173  for (auto *Address : CommonAddresses)
3174  Changed |= mergeConditionalStoreToAddress(
3175  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3176  return Changed;
3177 }
3178 
3179 /// If we have a conditional branch as a predecessor of another block,
3180 /// this function tries to simplify it. We know
3181 /// that PBI and BI are both conditional branches, and BI is in one of the
3182 /// successor blocks of PBI - PBI branches to BI.
3184  const DataLayout &DL) {
3185  assert(PBI->isConditional() && BI->isConditional());
3186  BasicBlock *BB = BI->getParent();
3187 
3188  // If this block ends with a branch instruction, and if there is a
3189  // predecessor that ends on a branch of the same condition, make
3190  // this conditional branch redundant.
3191  if (PBI->getCondition() == BI->getCondition() &&
3192  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3193  // Okay, the outcome of this conditional branch is statically
3194  // knowable. If this block had a single pred, handle specially.
3195  if (BB->getSinglePredecessor()) {
3196  // Turn this into a branch on constant.
3197  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3198  BI->setCondition(
3199  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3200  return true; // Nuke the branch on constant.
3201  }
3202 
3203  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3204  // in the constant and simplify the block result. Subsequent passes of
3205  // simplifycfg will thread the block.
3207  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3208  PHINode *NewPN = PHINode::Create(
3209  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3210  BI->getCondition()->getName() + ".pr", &BB->front());
3211  // Okay, we're going to insert the PHI node. Since PBI is not the only
3212  // predecessor, compute the PHI'd conditional value for all of the preds.
3213  // Any predecessor where the condition is not computable we keep symbolic.
3214  for (pred_iterator PI = PB; PI != PE; ++PI) {
3215  BasicBlock *P = *PI;
3216  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3217  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3218  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3219  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3220  NewPN->addIncoming(
3221  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3222  P);
3223  } else {
3224  NewPN->addIncoming(BI->getCondition(), P);
3225  }
3226  }
3227 
3228  BI->setCondition(NewPN);
3229  return true;
3230  }
3231  }
3232 
3233  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3234  if (CE->canTrap())
3235  return false;
3236 
3237  // If both branches are conditional and both contain stores to the same
3238  // address, remove the stores from the conditionals and create a conditional
3239  // merged store at the end.
3240  if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3241  return true;
3242 
3243  // If this is a conditional branch in an empty block, and if any
3244  // predecessors are a conditional branch to one of our destinations,
3245  // fold the conditions into logical ops and one cond br.
3246 
3247  // Ignore dbg intrinsics.
3248  if (&*BB->instructionsWithoutDebug().begin() != BI)
3249  return false;
3250 
3251  int PBIOp, BIOp;
3252  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3253  PBIOp = 0;
3254  BIOp = 0;
3255  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3256  PBIOp = 0;
3257  BIOp = 1;
3258  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3259  PBIOp = 1;
3260  BIOp = 0;
3261  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3262  PBIOp = 1;
3263  BIOp = 1;
3264  } else {
3265  return false;
3266  }
3267 
3268  // Check to make sure that the other destination of this branch
3269  // isn't BB itself. If so, this is an infinite loop that will
3270  // keep getting unwound.
3271  if (PBI->getSuccessor(PBIOp) == BB)
3272  return false;
3273 
3274  // Do not perform this transformation if it would require
3275  // insertion of a large number of select instructions. For targets
3276  // without predication/cmovs, this is a big pessimization.
3277 
3278  // Also do not perform this transformation if any phi node in the common
3279  // destination block can trap when reached by BB or PBB (PR17073). In that
3280  // case, it would be unsafe to hoist the operation into a select instruction.
3281 
3282  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3283  unsigned NumPhis = 0;
3284  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3285  ++II, ++NumPhis) {
3286  if (NumPhis > 2) // Disable this xform.
3287  return false;
3288 
3289  PHINode *PN = cast<PHINode>(II);
3290  Value *BIV = PN->getIncomingValueForBlock(BB);
3291  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3292  if (CE->canTrap())
3293  return false;
3294 
3295  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3296  Value *PBIV = PN->getIncomingValue(PBBIdx);
3297  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3298  if (CE->canTrap())
3299  return false;
3300  }
3301 
3302  // Finally, if everything is ok, fold the branches to logical ops.
3303  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3304 
3305  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3306  << "AND: " << *BI->getParent());
3307 
3308  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3309  // branch in it, where one edge (OtherDest) goes back to itself but the other
3310  // exits. We don't *know* that the program avoids the infinite loop
3311  // (even though that seems likely). If we do this xform naively, we'll end up
3312  // recursively unpeeling the loop. Since we know that (after the xform is
3313  // done) that the block *is* infinite if reached, we just make it an obviously
3314  // infinite loop with no cond branch.
3315  if (OtherDest == BB) {
3316  // Insert it at the end of the function, because it's either code,
3317  // or it won't matter if it's hot. :)
3318  BasicBlock *InfLoopBlock =
3319  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3320  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3321  OtherDest = InfLoopBlock;
3322  }
3323 
3324  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3325 
3326  // BI may have other predecessors. Because of this, we leave
3327  // it alone, but modify PBI.
3328 
3329  // Make sure we get to CommonDest on True&True directions.
3330  Value *PBICond = PBI->getCondition();
3331  IRBuilder<NoFolder> Builder(PBI);
3332  if (PBIOp)
3333  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3334 
3335  Value *BICond = BI->getCondition();
3336  if (BIOp)
3337  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3338 
3339  // Merge the conditions.
3340  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3341 
3342  // Modify PBI to branch on the new condition to the new dests.
3343  PBI->setCondition(Cond);
3344  PBI->setSuccessor(0, CommonDest);
3345  PBI->setSuccessor(1, OtherDest);
3346 
3347  // Update branch weight for PBI.
3348  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3349  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3350  bool HasWeights =
3351  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3352  SuccTrueWeight, SuccFalseWeight);
3353  if (HasWeights) {
3354  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3355  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3356  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3357  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3358  // The weight to CommonDest should be PredCommon * SuccTotal +
3359  // PredOther * SuccCommon.
3360  // The weight to OtherDest should be PredOther * SuccOther.
3361  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3362  PredOther * SuccCommon,
3363  PredOther * SuccOther};
3364  // Halve the weights if any of them cannot fit in an uint32_t
3365  FitWeights(NewWeights);
3366 
3367  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3368  }
3369 
3370  // OtherDest may have phi nodes. If so, add an entry from PBI's
3371  // block that are identical to the entries for BI's block.
3372  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3373 
3374  // We know that the CommonDest already had an edge from PBI to
3375  // it. If it has PHIs though, the PHIs may have different
3376  // entries for BB and PBI's BB. If so, insert a select to make
3377  // them agree.
3378  for (PHINode &PN : CommonDest->phis()) {
3379  Value *BIV = PN.getIncomingValueForBlock(BB);
3380  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3381  Value *PBIV = PN.getIncomingValue(PBBIdx);
3382  if (BIV != PBIV) {
3383  // Insert a select in PBI to pick the right value.
3384  SelectInst *NV = cast<SelectInst>(
3385  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3386  PN.setIncomingValue(PBBIdx, NV);
3387  // Although the select has the same condition as PBI, the original branch
3388  // weights for PBI do not apply to the new select because the select's
3389  // 'logical' edges are incoming edges of the phi that is eliminated, not
3390  // the outgoing edges of PBI.
3391  if (HasWeights) {
3392  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3393  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3394  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3395  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3396  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3397  // The weight to PredOtherDest should be PredOther * SuccCommon.
3398  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3399  PredOther * SuccCommon};
3400 
3401  FitWeights(NewWeights);
3402 
3403  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3404  }
3405  }
3406  }
3407 
3408  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3409  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3410 
3411  // This basic block is probably dead. We know it has at least
3412  // one fewer predecessor.
3413  return true;
3414 }
3415 
3416 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3417 // true or to FalseBB if Cond is false.
3418 // Takes care of updating the successors and removing the old terminator.
3419 // Also makes sure not to introduce new successors by assuming that edges to
3420 // non-successor TrueBBs and FalseBBs aren't reachable.
3421 static bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
3422  BasicBlock *TrueBB, BasicBlock *FalseBB,
3423  uint32_t TrueWeight,
3424  uint32_t FalseWeight) {
3425  // Remove any superfluous successor edges from the CFG.
3426  // First, figure out which successors to preserve.
3427  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3428  // successor.
3429  BasicBlock *KeepEdge1 = TrueBB;
3430  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3431 
3432  // Then remove the rest.
3433  for (BasicBlock *Succ : successors(OldTerm)) {
3434  // Make sure only to keep exactly one copy of each edge.
3435  if (Succ == KeepEdge1)
3436  KeepEdge1 = nullptr;
3437  else if (Succ == KeepEdge2)
3438  KeepEdge2 = nullptr;
3439  else
3440  Succ->removePredecessor(OldTerm->getParent(),
3441  /*KeepOneInputPHIs=*/true);
3442  }
3443 
3444  IRBuilder<> Builder(OldTerm);
3445  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3446 
3447  // Insert an appropriate new terminator.
3448  if (!KeepEdge1 && !KeepEdge2) {
3449  if (TrueBB == FalseBB)
3450  // We were only looking for one successor, and it was present.
3451  // Create an unconditional branch to it.
3452  Builder.CreateBr(TrueBB);
3453  else {
3454  // We found both of the successors we were looking for.
3455  // Create a conditional branch sharing the condition of the select.
3456  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3457  if (TrueWeight != FalseWeight)
3458  setBranchWeights(NewBI, TrueWeight, FalseWeight);
3459  }
3460  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3461  // Neither of the selected blocks were successors, so this
3462  // terminator must be unreachable.
3463  new UnreachableInst(OldTerm->getContext(), OldTerm);
3464  } else {
3465  // One of the selected values was a successor, but the other wasn't.
3466  // Insert an unconditional branch to the one that was found;
3467  // the edge to the one that wasn't must be unreachable.
3468  if (!KeepEdge1)
3469  // Only TrueBB was found.
3470  Builder.CreateBr(TrueBB);
3471  else
3472  // Only FalseBB was found.
3473  Builder.CreateBr(FalseBB);
3474  }
3475 
3476  EraseTerminatorAndDCECond(OldTerm);
3477  return true;
3478 }
3479 
3480 // Replaces
3481 // (switch (select cond, X, Y)) on constant X, Y
3482 // with a branch - conditional if X and Y lead to distinct BBs,
3483 // unconditional otherwise.
3485  // Check for constant integer values in the select.
3486  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3487  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3488  if (!TrueVal || !FalseVal)
3489  return false;
3490 
3491  // Find the relevant condition and destinations.
3492  Value *Condition = Select->getCondition();
3493  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3494  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3495 
3496  // Get weight for TrueBB and FalseBB.
3497  uint32_t TrueWeight = 0, FalseWeight = 0;
3498  SmallVector<uint64_t, 8> Weights;
3499  bool HasWeights = HasBranchWeights(SI);
3500  if (HasWeights) {
3501  GetBranchWeights(SI, Weights);
3502  if (Weights.size() == 1 + SI->getNumCases()) {
3503  TrueWeight =
3504  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3505  FalseWeight =
3506  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3507  }
3508  }
3509 
3510  // Perform the actual simplification.
3511  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3512  FalseWeight);
3513 }
3514 
3515 // Replaces
3516 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3517 // blockaddress(@fn, BlockB)))
3518 // with
3519 // (br cond, BlockA, BlockB).
3521  // Check that both operands of the select are block addresses.
3524  if (!TBA || !FBA)
3525  return false;
3526 
3527  // Extract the actual blocks.
3528  BasicBlock *TrueBB = TBA->getBasicBlock();
3529  BasicBlock *FalseBB = FBA->getBasicBlock();
3530 
3531  // Perform the actual simplification.
3532  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3533  0);
3534 }
3535 
3536 /// This is called when we find an icmp instruction
3537 /// (a seteq/setne with a constant) as the only instruction in a
3538 /// block that ends with an uncond branch. We are looking for a very specific
3539 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3540 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3541 /// default value goes to an uncond block with a seteq in it, we get something
3542 /// like:
3543 ///
3544 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3545 /// DEFAULT:
3546 /// %tmp = icmp eq i8 %A, 92
3547 /// br label %end
3548 /// end:
3549 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3550 ///
3551 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3552 /// the PHI, merging the third icmp into the switch.
3553 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3554  ICmpInst *ICI, IRBuilder<> &Builder) {
3555  BasicBlock *BB = ICI->getParent();
3556 
3557  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3558  // complex.
3559  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3560  return false;
3561 
3562  Value *V = ICI->getOperand(0);
3563  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3564 
3565  // The pattern we're looking for is where our only predecessor is a switch on
3566  // 'V' and this block is the default case for the switch. In this case we can
3567  // fold the compared value into the switch to simplify things.
3568  BasicBlock *Pred = BB->getSinglePredecessor();
3569  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3570  return false;
3571 
3572  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3573  if (SI->getCondition() != V)
3574  return false;
3575 
3576  // If BB is reachable on a non-default case, then we simply know the value of
3577  // V in this block. Substitute it and constant fold the icmp instruction
3578  // away.
3579  if (SI->getDefaultDest() != BB) {
3580  ConstantInt *VVal = SI->findCaseDest(BB);
3581  assert(VVal && "Should have a unique destination value");
3582  ICI->setOperand(0, VVal);
3583 
3584  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3585  ICI->replaceAllUsesWith(V);
3586  ICI->eraseFromParent();
3587  }
3588  // BB is now empty, so it is likely to simplify away.
3589  return requestResimplify();
3590  }
3591 
3592  // Ok, the block is reachable from the default dest. If the constant we're
3593  // comparing exists in one of the other edges, then we can constant fold ICI
3594  // and zap it.
3595  if (SI->findCaseValue(Cst) != SI->case_default()) {
3596  Value *V;
3597  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3598  V = ConstantInt::getFalse(BB->getContext());
3599  else
3600  V = ConstantInt::getTrue(BB->getContext());
3601 
3602  ICI->replaceAllUsesWith(V);
3603  ICI->eraseFromParent();
3604  // BB is now empty, so it is likely to simplify away.
3605  return requestResimplify();
3606  }
3607 
3608  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3609  // the block.
3610  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3611  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3612  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3613  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3614  return false;
3615 
3616  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3617  // true in the PHI.
3618  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3619  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3620 
3621  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3622  std::swap(DefaultCst, NewCst);
3623 
3624  // Replace ICI (which is used by the PHI for the default value) with true or
3625  // false depending on if it is EQ or NE.
3626  ICI->replaceAllUsesWith(DefaultCst);
3627  ICI->eraseFromParent();
3628 
3629  // Okay, the switch goes to this block on a default value. Add an edge from
3630  // the switch to the merge point on the compared value.
3631  BasicBlock *NewBB =
3632  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3633  SmallVector<uint64_t, 8> Weights;
3634  bool HasWeights = HasBranchWeights(SI);
3635  if (HasWeights) {
3636  GetBranchWeights(SI, Weights);
3637  if (Weights.size() == 1 + SI->getNumCases()) {
3638  // Split weight for default case to case for "Cst".
3639  Weights[0] = (Weights[0] + 1) >> 1;
3640  Weights.push_back(Weights[0]);
3641 
3642  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3643  setBranchWeights(SI, MDWeights);
3644  }
3645  }
3646  SI->addCase(Cst, NewBB);
3647 
3648  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3649  Builder.SetInsertPoint(NewBB);
3650  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3651  Builder.CreateBr(SuccBlock);
3652  PHIUse->addIncoming(NewCst, NewBB);
3653  return true;
3654 }
3655 
3656 /// The specified branch is a conditional branch.
3657 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3658 /// fold it into a switch instruction if so.
3660  const DataLayout &DL) {
3661  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3662  if (!Cond)
3663  return false;
3664 
3665  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3666  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3667  // 'setne's and'ed together, collect them.
3668 
3669  // Try to gather values from a chain of and/or to be turned into a switch
3670  ConstantComparesGatherer ConstantCompare(Cond, DL);
3671  // Unpack the result
3672  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3673  Value *CompVal = ConstantCompare.CompValue;
3674  unsigned UsedICmps = ConstantCompare.UsedICmps;
3675  Value *ExtraCase = ConstantCompare.Extra;
3676 
3677  // If we didn't have a multiply compared value, fail.
3678  if (!CompVal)
3679  return false;
3680 
3681  // Avoid turning single icmps into a switch.
3682  if (UsedICmps <= 1)
3683  return false;
3684 
3685  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3686 
3687  // There might be duplicate constants in the list, which the switch
3688  // instruction can't handle, remove them now.
3689  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3690  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3691 
3692  // If Extra was used, we require at least two switch values to do the
3693  // transformation. A switch with one value is just a conditional branch.
3694  if (ExtraCase && Values.size() < 2)
3695  return false;
3696 
3697  // TODO: Preserve branch weight metadata, similarly to how
3698  // FoldValueComparisonIntoPredecessors preserves it.
3699 
3700  // Figure out which block is which destination.
3701  BasicBlock *DefaultBB = BI->getSuccessor(1);
3702  BasicBlock *EdgeBB = BI->getSuccessor(0);
3703  if (!TrueWhenEqual)
3704  std::swap(DefaultBB, EdgeBB);
3705 
3706  BasicBlock *BB = BI->getParent();
3707 
3708  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3709  << " cases into SWITCH. BB is:\n"
3710  << *BB);
3711 
3712  // If there are any extra values that couldn't be folded into the switch
3713  // then we evaluate them with an explicit branch first. Split the block
3714  // right before the condbr to handle it.
3715  if (ExtraCase) {
3716  BasicBlock *NewBB =
3717  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3718  // Remove the uncond branch added to the old block.
3719  Instruction *OldTI = BB->getTerminator();
3720  Builder.SetInsertPoint(OldTI);
3721 
3722  if (TrueWhenEqual)
3723  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3724  else
3725  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3726 
3727  OldTI->eraseFromParent();
3728 
3729  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3730  // for the edge we just added.
3731  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3732 
3733  LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3734  << "\nEXTRABB = " << *BB);
3735  BB = NewBB;
3736  }
3737 
3738  Builder.SetInsertPoint(BI);
3739  // Convert pointer to int before we switch.
3740  if (CompVal->getType()->isPointerTy()) {
3741  CompVal = Builder.CreatePtrToInt(
3742  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3743  }
3744 
3745  // Create the new switch instruction now.
3746  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3747 
3748  // Add all of the 'cases' to the switch instruction.
3749  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3750  New->addCase(Values[i], EdgeBB);
3751 
3752  // We added edges from PI to the EdgeBB. As such, if there were any
3753  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3754  // the number of edges added.
3755  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3756  PHINode *PN = cast<PHINode>(BBI);
3757  Value *InVal = PN->getIncomingValueForBlock(BB);
3758  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3759  PN->addIncoming(InVal, BB);
3760  }
3761 
3762  // Erase the old branch instruction.
3764 
3765  LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3766  return true;
3767 }
3768 
3769 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3770  if (isa<PHINode>(RI->getValue()))
3771  return SimplifyCommonResume(RI);
3772  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3773  RI->getValue() == RI->getParent()->getFirstNonPHI())
3774  // The resume must unwind the exception that caused control to branch here.
3775  return SimplifySingleResume(RI);
3776 
3777  return false;
3778 }
3779 
3780 // Simplify resume that is shared by several landing pads (phi of landing pad).
3781 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3782  BasicBlock *BB = RI->getParent();
3783 
3784  // Check that there are no other instructions except for debug intrinsics
3785  // between the phi of landing pads (RI->getValue()) and resume instruction.
3786  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3787  E = RI->getIterator();
3788  while (++I != E)
3789  if (!isa<DbgInfoIntrinsic>(I))
3790  return false;
3791 
3792  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3793  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3794 
3795  // Check incoming blocks to see if any of them are trivial.
3796  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3797  Idx++) {
3798  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3799  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3800 
3801  // If the block has other successors, we can not delete it because
3802  // it has other dependents.
3803  if (IncomingBB->getUniqueSuccessor() != BB)
3804  continue;
3805 
3806  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3807  // Not the landing pad that caused the control to branch here.
3808  if (IncomingValue != LandingPad)
3809  continue;
3810 
3811  bool isTrivial = true;
3812 
3813  I = IncomingBB->getFirstNonPHI()->getIterator();
3814  E = IncomingBB->getTerminator()->getIterator();
3815  while (++I != E)
3816  if (!isa<DbgInfoIntrinsic>(I)) {
3817  isTrivial = false;
3818  break;
3819  }
3820 
3821  if (isTrivial)
3822  TrivialUnwindBlocks.insert(IncomingBB);
3823  }
3824 
3825  // If no trivial unwind blocks, don't do any simplifications.
3826  if (TrivialUnwindBlocks.empty())
3827  return false;
3828 
3829  // Turn all invokes that unwind here into calls.
3830  for (auto *TrivialBB : TrivialUnwindBlocks) {
3831  // Blocks that will be simplified should be removed from the phi node.
3832  // Note there could be multiple edges to the resume block, and we need
3833  // to remove them all.
3834  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3835  BB->removePredecessor(TrivialBB, true);
3836 
3837  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3838  PI != PE;) {
3839  BasicBlock *Pred = *PI++;
3840  removeUnwindEdge(Pred);
3841  }
3842 
3843  // In each SimplifyCFG run, only the current processed block can be erased.
3844  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3845  // of erasing TrivialBB, we only remove the branch to the common resume
3846  // block so that we can later erase the resume block since it has no
3847  // predecessors.
3848  TrivialBB->getTerminator()->eraseFromParent();
3849  new UnreachableInst(RI->getContext(), TrivialBB);
3850  }
3851 
3852  // Delete the resume block if all its predecessors have been removed.
3853  if (pred_empty(BB))
3854  BB->eraseFromParent();
3855 
3856  return !TrivialUnwindBlocks.empty();
3857 }
3858 
3859 // Simplify resume that is only used by a single (non-phi) landing pad.
3860 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3861  BasicBlock *BB = RI->getParent();
3863  assert(RI->getValue() == LPInst &&
3864  "Resume must unwind the exception that caused control to here");
3865 
3866  // Check that there are no other instructions except for debug intrinsics.
3867  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3868  while (++I != E)
3869  if (!isa<DbgInfoIntrinsic>(I))
3870  return false;
3871 
3872  // Turn all invokes that unwind here into calls and delete the basic block.
3873  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3874  BasicBlock *Pred = *PI++;
3875  removeUnwindEdge(Pred);
3876  }
3877 
3878  // The landingpad is now unreachable. Zap it.
3879  if (LoopHeaders)
3880  LoopHeaders->erase(BB);
3881  BB->eraseFromParent();
3882  return true;
3883 }
3884 
3886  // If this is a trivial cleanup pad that executes no instructions, it can be
3887  // eliminated. If the cleanup pad continues to the caller, any predecessor
3888  // that is an EH pad will be updated to continue to the caller and any
3889  // predecessor that terminates with an invoke instruction will have its invoke
3890  // instruction converted to a call instruction. If the cleanup pad being
3891  // simplified does not continue to the caller, each predecessor will be
3892  // updated to continue to the unwind destination of the cleanup pad being
3893  // simplified.
3894  BasicBlock *BB = RI->getParent();
3895  CleanupPadInst *CPInst = RI->getCleanupPad();
3896  if (CPInst->getParent() != BB)
3897  // This isn't an empty cleanup.
3898  return false;
3899 
3900  // We cannot kill the pad if it has multiple uses. This typically arises
3901  // from unreachable basic blocks.
3902  if (!CPInst->hasOneUse())
3903  return false;
3904 
3905  // Check that there are no other instructions except for benign intrinsics.
3906  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3907  while (++I != E) {
3908  auto *II = dyn_cast<IntrinsicInst>(I);
3909  if (!II)
3910  return false;
3911 
3912  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3913  switch (IntrinsicID) {
3914  case Intrinsic::dbg_declare:
3915  case Intrinsic::dbg_value:
3916  case Intrinsic::dbg_label:
3917  case Intrinsic::lifetime_end:
3918  break;
3919  default:
3920  return false;
3921  }
3922  }
3923 
3924  // If the cleanup return we are simplifying unwinds to the caller, this will
3925  // set UnwindDest to nullptr.
3926  BasicBlock *UnwindDest = RI->getUnwindDest();
3927  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3928 
3929  // We're about to remove BB from the control flow. Before we do, sink any
3930  // PHINodes into the unwind destination. Doing this before changing the
3931  // control flow avoids some potentially slow checks, since we can currently
3932  // be certain that UnwindDest and BB have no common predecessors (since they
3933  // are both EH pads).
3934  if (UnwindDest) {
3935  // First, go through the PHI nodes in UnwindDest and update any nodes that
3936  // reference the block we are removing
3937  for (BasicBlock::iterator I = UnwindDest->begin(),
3938  IE = DestEHPad->getIterator();
3939  I != IE; ++I) {
3940  PHINode *DestPN = cast<PHINode>(I);
3941 
3942  int Idx = DestPN->getBasicBlockIndex(BB);
3943  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3944  assert(Idx != -1);
3945  // This PHI node has an incoming value that corresponds to a control
3946  // path through the cleanup pad we are removing. If the incoming
3947  // value is in the cleanup pad, it must be a PHINode (because we
3948  // verified above that the block is otherwise empty). Otherwise, the
3949  // value is either a constant or a value that dominates the cleanup
3950  // pad being removed.
3951  //
3952  // Because BB and UnwindDest are both EH pads, all of their
3953  // predecessors must unwind to these blocks, and since no instruction
3954  // can have multiple unwind destinations, there will be no overlap in
3955  // incoming blocks between SrcPN and DestPN.
3956  Value *SrcVal = DestPN->getIncomingValue(Idx);
3957  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3958 
3959  // Remove the entry for the block we are deleting.
3960  DestPN->removeIncomingValue(Idx, false);
3961 
3962  if (SrcPN && SrcPN->getParent() == BB) {
3963  // If the incoming value was a PHI node in the cleanup pad we are
3964  // removing, we need to merge that PHI node's incoming values into
3965  // DestPN.
3966  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3967  SrcIdx != SrcE; ++SrcIdx) {
3968  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3969  SrcPN->getIncomingBlock(SrcIdx));
3970  }
3971  } else {
3972  // Otherwise, the incoming value came from above BB and
3973  // so we can just reuse it. We must associate all of BB's
3974  // predecessors with this value.
3975  for (auto *pred : predecessors(BB)) {
3976  DestPN->addIncoming(SrcVal, pred);
3977  }
3978  }
3979  }
3980 
3981  // Sink any remaining PHI nodes directly into UnwindDest.
3982  Instruction *InsertPt = DestEHPad;
3983  for (BasicBlock::iterator I = BB->begin(),
3984  IE = BB->getFirstNonPHI()->getIterator();
3985  I != IE;) {
3986  // The iterator must be incremented here because the instructions are
3987  // being moved to another block.
3988  PHINode *PN = cast<PHINode>(I++);
3989  if (PN->use_empty())
3990  // If the PHI node has no uses, just leave it. It will be erased
3991  // when we erase BB below.
3992  continue;
3993 
3994  // Otherwise, sink this PHI node into UnwindDest.
3995  // Any predecessors to UnwindDest which are not already represented
3996  // must be back edges which inherit the value from the path through
3997  // BB. In this case, the PHI value must reference itself.
3998  for (auto *pred : predecessors(UnwindDest))
3999  if (pred != BB)
4000  PN->addIncoming(PN, pred);
4001  PN->moveBefore(InsertPt);
4002  }
4003  }
4004 
4005  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4006  // The iterator must be updated here because we are removing this pred.
4007  BasicBlock *PredBB = *PI++;
4008  if (UnwindDest == nullptr) {
4009  removeUnwindEdge(PredBB);
4010  } else {
4011  Instruction *TI = PredBB->getTerminator();
4012  TI->replaceUsesOfWith(BB, UnwindDest);
4013  }
4014  }
4015 
4016  // The cleanup pad is now unreachable. Zap it.
4017  BB->eraseFromParent();
4018  return true;
4019 }
4020 
4021 // Try to merge two cleanuppads together.
4023  // Skip any cleanuprets which unwind to caller, there is nothing to merge
4024  // with.
4025  BasicBlock *UnwindDest = RI->getUnwindDest();
4026  if (!UnwindDest)
4027  return false;
4028 
4029  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4030  // be safe to merge without code duplication.
4031  if (UnwindDest->getSinglePredecessor() != RI->getParent())
4032  return false;
4033 
4034  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4035  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4036  if (!SuccessorCleanupPad)
4037  return false;
4038 
4039  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4040  // Replace any uses of the successor cleanupad with the predecessor pad
4041  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4042  // funclet bundle operands.
4043  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4044  // Remove the old cleanuppad.
4045  SuccessorCleanupPad->eraseFromParent();
4046  // Now, we simply replace the cleanupret with a branch to the unwind
4047  // destination.
4048  BranchInst::Create(UnwindDest, RI->getParent());
4049  RI->eraseFromParent();
4050 
4051  return true;
4052 }
4053 
4054 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4055  // It is possible to transiantly have an undef cleanuppad operand because we
4056  // have deleted some, but not all, dead blocks.
4057  // Eventually, this block will be deleted.
4058  if (isa<UndefValue>(RI->getOperand(0)))
4059  return false;
4060 
4061  if (mergeCleanupPad(RI))
4062  return true;
4063 
4064  if (removeEmptyCleanup(RI))
4065  return true;
4066 
4067  return false;
4068 }
4069 
4070 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4071  BasicBlock *BB = RI->getParent();
4072  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4073  return false;
4074 
4075  // Find predecessors that end with branches.
4076  SmallVector<BasicBlock *, 8> UncondBranchPreds;
4077  SmallVector<BranchInst *, 8> CondBranchPreds;
4078  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4079  BasicBlock *P = *PI;
4080  Instruction *PTI = P->getTerminator();
4081  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4082  if (BI->isUnconditional())
4083  UncondBranchPreds.push_back(P);
4084  else
4085  CondBranchPreds.push_back(BI);
4086  }
4087  }
4088 
4089  // If we found some, do the transformation!
4090  if (!UncondBranchPreds.empty() && DupRet) {
4091  while (!UncondBranchPreds.empty()) {
4092  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4093  LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4094  << "INTO UNCOND BRANCH PRED: " << *Pred);
4095  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4096  }
4097 
4098  // If we eliminated all predecessors of the block, delete the block now.
4099  if (pred_empty(BB)) {
4100  // We know there are no successors, so just nuke the block.
4101  if (LoopHeaders)
4102  LoopHeaders->erase(BB);
4103  BB->eraseFromParent();
4104  }
4105 
4106  return true;
4107  }
4108 
4109  // Check out all of the conditional branches going to this return
4110  // instruction. If any of them just select between returns, change the
4111  // branch itself into a select/return pair.
4112  while (!CondBranchPreds.empty()) {
4113  BranchInst *BI = CondBranchPreds.pop_back_val();
4114 
4115  // Check to see if the non-BB successor is also a return block.
4116  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4117  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4118  SimplifyCondBranchToTwoReturns(BI, Builder))
4119  return true;
4120  }
4121  return false;
4122 }
4123 
4124 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4125  BasicBlock *BB = UI->getParent();
4126 
4127  bool Changed = false;
4128 
4129  // If there are any instructions immediately before the unreachable that can
4130  // be removed, do so.
4131  while (UI->getIterator() != BB->begin()) {
4132  BasicBlock::iterator BBI = UI->getIterator();
4133  --BBI;
4134  // Do not delete instructions that can have side effects which might cause
4135  // the unreachable to not be reachable; specifically, calls and volatile
4136  // operations may have this effect.
4137  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4138  break;
4139 
4140  if (BBI->mayHaveSideEffects()) {
4141  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4142  if (SI->isVolatile())
4143  break;
4144  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4145  if (LI->isVolatile())
4146  break;
4147  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4148  if (RMWI->isVolatile())
4149  break;
4150  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4151  if (CXI->isVolatile())
4152  break;
4153  } else if (isa<CatchPadInst>(BBI)) {
4154  // A catchpad may invoke exception object constructors and such, which
4155  // in some languages can be arbitrary code, so be conservative by
4156  // default.
4157  // For CoreCLR, it just involves a type test, so can be removed.
4160  break;
4161  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4162  !isa<LandingPadInst>(BBI)) {
4163  break;
4164  }
4165  // Note that deleting LandingPad's here is in fact okay, although it
4166  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4167  // all the predecessors of this block will be the unwind edges of Invokes,
4168  // and we can therefore guarantee this block will be erased.
4169  }
4170 
4171  // Delete this instruction (any uses are guaranteed to be dead)
4172  if (!BBI->use_empty())
4173  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4174  BBI->eraseFromParent();
4175  Changed = true;
4176  }
4177 
4178  // If the unreachable instruction is the first in the block, take a gander
4179  // at all of the predecessors of this instruction, and simplify them.
4180  if (&BB->front() != UI)
4181  return Changed;
4182 
4184  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4185  Instruction *TI = Preds[i]->getTerminator();
4186  IRBuilder<> Builder(TI);
4187  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4188  if (BI->isUnconditional()) {
4189  if (BI->getSuccessor(0) == BB) {
4190  new UnreachableInst(TI->getContext(), TI);
4191  TI->eraseFromParent();
4192  Changed = true;
4193  }
4194  } else {
4195  if (BI->getSuccessor(0) == BB) {
4196  Builder.CreateBr(BI->getSuccessor(1));
4198  } else if (BI->getSuccessor(1) == BB) {
4199  Builder.CreateBr(BI->getSuccessor(0));
4201  Changed = true;
4202  }
4203  }
4204  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4205  for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4206  if (i->getCaseSuccessor() != BB) {
4207  ++i;
4208  continue;
4209  }
4210  BB->removePredecessor(SI->getParent());
4211  i = SI->removeCase(i);
4212  e = SI->case_end();
4213  Changed = true;
4214  }
4215  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4216  if (II->getUnwindDest() == BB) {
4217  removeUnwindEdge(TI->getParent());
4218  Changed = true;
4219  }
4220  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4221  if (CSI->getUnwindDest() == BB) {
4222  removeUnwindEdge(TI->getParent());
4223  Changed = true;
4224  continue;
4225  }
4226 
4227  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4228  E = CSI->handler_end();
4229  I != E; ++I) {
4230  if (*I == BB) {
4231  CSI->removeHandler(I);
4232  --I;
4233  --E;
4234  Changed = true;
4235  }
4236  }
4237  if (CSI->getNumHandlers() == 0) {
4238  BasicBlock *CatchSwitchBB = CSI->getParent();
4239  if (CSI->hasUnwindDest()) {
4240  // Redirect preds to the unwind dest
4241  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4242  } else {
4243  // Rewrite all preds to unwind to caller (or from invoke to call).
4244  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4245  for (BasicBlock *EHPred : EHPreds)
4246  removeUnwindEdge(EHPred);
4247  }
4248  // The catchswitch is no longer reachable.
4249  new UnreachableInst(CSI->getContext(), CSI);
4250  CSI->eraseFromParent();
4251  Changed = true;
4252  }
4253  } else if (isa<CleanupReturnInst>(TI)) {
4254  new UnreachableInst(TI->getContext(), TI);
4255  TI->eraseFromParent();
4256  Changed = true;
4257  }
4258  }
4259 
4260  // If this block is now dead, remove it.
4261  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4262  // We know there are no successors, so just nuke the block.
4263  if (LoopHeaders)
4264  LoopHeaders->erase(BB);
4265  BB->eraseFromParent();
4266  return true;
4267  }
4268 
4269  return Changed;
4270 }
4271 
4273  assert(Cases.size() >= 1);
4274 
4275  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4276  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4277  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4278  return false;
4279  }
4280  return true;
4281 }
4282 
4283 /// Turn a switch with two reachable destinations into an integer range
4284 /// comparison and branch.
4286  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4287 
4288  bool HasDefault =
4289  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4290 
4291  // Partition the cases into two sets with different destinations.
4292  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4293  BasicBlock *DestB = nullptr;
4296 
4297  for (auto Case : SI->cases()) {
4298  BasicBlock *Dest = Case.getCaseSuccessor();
4299  if (!DestA)
4300  DestA = Dest;
4301  if (Dest == DestA) {
4302  CasesA.push_back(Case.getCaseValue());
4303  continue;
4304  }
4305  if (!DestB)
4306  DestB = Dest;
4307  if (Dest == DestB) {
4308  CasesB.push_back(Case.getCaseValue());
4309  continue;
4310  }
4311  return false; // More than two destinations.
4312  }
4313 
4314  assert(DestA && DestB &&
4315  "Single-destination switch should have been folded.");
4316  assert(DestA != DestB);
4317  assert(DestB != SI->getDefaultDest());
4318  assert(!CasesB.empty() && "There must be non-default cases.");
4319  assert(!CasesA.empty() || HasDefault);
4320 
4321  // Figure out if one of the sets of cases form a contiguous range.
4322  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4323  BasicBlock *ContiguousDest = nullptr;
4324  BasicBlock *OtherDest = nullptr;
4325  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4326  ContiguousCases = &CasesA;
4327  ContiguousDest = DestA;
4328  OtherDest = DestB;
4329  } else if (CasesAreContiguous(CasesB)) {
4330  ContiguousCases = &CasesB;
4331  ContiguousDest = DestB;
4332  OtherDest = DestA;
4333  } else
4334  return false;
4335 
4336  // Start building the compare and branch.
4337 
4338  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4339  Constant *NumCases =
4340  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4341 
4342  Value *Sub = SI->getCondition();
4343  if (!Offset->isNullValue())
4344  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4345 
4346  Value *Cmp;
4347  // If NumCases overflowed, then all possible values jump to the successor.
4348  if (NumCases->isNullValue() && !ContiguousCases->empty())
4349  Cmp = ConstantInt::getTrue(SI->getContext());
4350  else
4351  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4352  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4353 
4354  // Update weight for the newly-created conditional branch.
4355  if (HasBranchWeights(SI)) {
4356  SmallVector<uint64_t, 8> Weights;
4357  GetBranchWeights(SI, Weights);
4358  if (Weights.size() == 1 + SI->getNumCases()) {
4359  uint64_t TrueWeight = 0;
4360  uint64_t FalseWeight = 0;
4361  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4362  if (SI->getSuccessor(I) == ContiguousDest)
4363  TrueWeight += Weights[I];
4364  else
4365  FalseWeight += Weights[I];
4366  }
4367  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4368  TrueWeight /= 2;
4369  FalseWeight /= 2;
4370  }
4371  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4372  }
4373  }
4374 
4375  // Prune obsolete incoming values off the successors' PHI nodes.
4376  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4377  unsigned PreviousEdges = ContiguousCases->size();
4378  if (ContiguousDest == SI->getDefaultDest())
4379  ++PreviousEdges;
4380  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4381  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4382  }
4383  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4384  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4385  if (OtherDest == SI->getDefaultDest())
4386  ++PreviousEdges;
4387  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4388  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4389  }
4390 
4391  // Drop the switch.
4392  SI->eraseFromParent();
4393 
4394  return true;
4395 }
4396 
4397 /// Compute masked bits for the condition of a switch
4398 /// and use it to remove dead cases.
4400  const DataLayout &DL) {
4401  Value *Cond = SI->getCondition();
4402  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4403  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4404 
4405  // We can also eliminate cases by determining that their values are outside of
4406  // the limited range of the condition based on how many significant (non-sign)
4407  // bits are in the condition value.
4408  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4409  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4410 
4411  // Gather dead cases.
4413  for (auto &Case : SI->cases()) {
4414  const APInt &CaseVal = Case.getCaseValue()->getValue();
4415  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4416  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4417  DeadCases.push_back(Case.getCaseValue());
4418  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4419  << " is dead.\n");
4420  }
4421  }
4422 
4423  // If we can prove that the cases must cover all possible values, the
4424  // default destination becomes dead and we can remove it. If we know some
4425  // of the bits in the value, we can use that to more precisely compute the
4426  // number of possible unique case values.
4427  bool HasDefault =
4428  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4429  const unsigned NumUnknownBits =
4430  Bits - (Known.Zero | Known.One).countPopulation();
4431  assert(NumUnknownBits <= Bits);
4432  if (HasDefault && DeadCases.empty() &&
4433  NumUnknownBits < 64 /* avoid overflow */ &&
4434  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4435  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4436  BasicBlock *NewDefault =
4438  SI->setDefaultDest(&*NewDefault);
4439  SplitBlock(&*NewDefault, &NewDefault->front());
4440  auto *OldTI = NewDefault->getTerminator();
4441  new UnreachableInst(SI->getContext(), OldTI);
4443  return true;
4444  }
4445 
4446  SmallVector<uint64_t, 8> Weights;
4447  bool HasWeight = HasBranchWeights(SI);
4448  if (HasWeight) {
4449  GetBranchWeights(SI, Weights);
4450  HasWeight = (Weights.size() == 1 + SI->getNumCases());
4451  }
4452 
4453  // Remove dead cases from the switch.
4454  for (ConstantInt *DeadCase : DeadCases) {
4455  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4456  assert(CaseI != SI->case_default() &&
4457  "Case was not found. Probably mistake in DeadCases forming.");
4458  if (HasWeight) {
4459  std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4460  Weights.pop_back();
4461  }
4462 
4463  // Prune unused values from PHI nodes.
4464  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4465  SI->removeCase(CaseI);
4466  }
4467  if (HasWeight && Weights.size() >= 2) {
4468  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4469  setBranchWeights(SI, MDWeights);
4470  }
4471 
4472  return !DeadCases.empty();
4473 }
4474 
4475 /// If BB would be eligible for simplification by
4476 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4477 /// by an unconditional branch), look at the phi node for BB in the successor
4478 /// block and see if the incoming value is equal to CaseValue. If so, return
4479 /// the phi node, and set PhiIndex to BB's index in the phi node.
4481  BasicBlock *BB, int *PhiIndex) {
4482  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4483  return nullptr; // BB must be empty to be a candidate for simplification.
4484  if (!BB->getSinglePredecessor())
4485  return nullptr; // BB must be dominated by the switch.
4486 
4488  if (!Branch || !Branch->isUnconditional())
4489  return nullptr; // Terminator must be unconditional branch.
4490 
4491  BasicBlock *Succ = Branch->getSuccessor(0);
4492 
4493  for (PHINode &PHI : Succ->phis()) {
4494  int Idx = PHI.getBasicBlockIndex(BB);
4495  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4496 
4497  Value *InValue = PHI.getIncomingValue(Idx);
4498  if (InValue != CaseValue)
4499  continue;
4500 
4501  *PhiIndex = Idx;
4502  return &PHI;
4503  }
4504 
4505  return nullptr;
4506 }
4507 
4508 /// Try to forward the condition of a switch instruction to a phi node
4509 /// dominated by the switch, if that would mean that some of the destination
4510 /// blocks of the switch can be folded away. Return true if a change is made.
4512  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4513 
4514  ForwardingNodesMap ForwardingNodes;
4515  BasicBlock *SwitchBlock = SI->getParent();
4516  bool Changed = false;
4517  for (auto &Case : SI->cases()) {
4518  ConstantInt *CaseValue = Case.getCaseValue();
4519  BasicBlock *CaseDest = Case.getCaseSuccessor();
4520 
4521  // Replace phi operands in successor blocks that are using the constant case
4522  // value rather than the switch condition variable:
4523  // switchbb:
4524  // switch i32 %x, label %default [
4525  // i32 17, label %succ
4526  // ...
4527  // succ:
4528  // %r = phi i32 ... [ 17, %switchbb ] ...
4529  // -->
4530  // %r = phi i32 ... [ %x, %switchbb ] ...
4531 
4532  for (PHINode &Phi : CaseDest->phis()) {
4533  // This only works if there is exactly 1 incoming edge from the switch to
4534  // a phi. If there is >1, that means multiple cases of the switch map to 1
4535  // value in the phi, and that phi value is not the switch condition. Thus,
4536  // this transform would not make sense (the phi would be invalid because
4537  // a phi can't have different incoming values from the same block).
4538  int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4539  if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4540  count(Phi.blocks(), SwitchBlock) == 1) {
4541  Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4542  Changed = true;
4543  }
4544  }
4545 
4546  // Collect phi nodes that are indirectly using this switch's case constants.
4547  int PhiIdx;
4548  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4549  ForwardingNodes[Phi].push_back(PhiIdx);
4550  }
4551 
4552  for (auto &ForwardingNode : ForwardingNodes) {
4553  PHINode *Phi = ForwardingNode.first;
4554  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4555  if (Indexes.size() < 2)
4556  continue;
4557 
4558  for (int Index : Indexes)
4559  Phi->setIncomingValue(Index, SI->getCondition());
4560  Changed = true;
4561  }
4562 
4563  return Changed;
4564 }
4565 
4566 /// Return true if the backend will be able to handle
4567 /// initializing an array of constants like C.
4569  if (C->isThreadDependent())
4570  return false;
4571  if (C->isDLLImportDependent())
4572  return false;
4573 
4574  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4575  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4576  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4577  return false;
4578 
4579  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4580  if (!CE->isGEPWithNoNotionalOverIndexing())
4581  return false;
4582  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4583  return false;
4584  }
4585 
4587  return false;
4588 
4589  return true;
4590 }
4591 
4592 /// If V is a Constant, return it. Otherwise, try to look up
4593 /// its constant value in ConstantPool, returning 0 if it's not there.
4594 static Constant *
4597  if (Constant *C = dyn_cast<Constant>(V))
4598  return C;
4599  return ConstantPool.lookup(V);
4600 }
4601 
4602 /// Try to fold instruction I into a constant. This works for
4603 /// simple instructions such as binary operations where both operands are
4604 /// constant or can be replaced by constants from the ConstantPool. Returns the
4605 /// resulting constant on success, 0 otherwise.
4606 static Constant *
4609  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4610  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4611  if (!A)
4612  return nullptr;
4613  if (A->isAllOnesValue())
4614  return LookupConstant(Select->getTrueValue(), ConstantPool);
4615  if (A->isNullValue())
4616  return LookupConstant(Select->getFalseValue(), ConstantPool);
4617  return nullptr;
4618  }
4619 
4621  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4623  COps.push_back(A);
4624  else
4625  return nullptr;
4626  }
4627 
4628  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4629  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4630  COps[1], DL);
4631  }
4632 
4633  return ConstantFoldInstOperands(I, COps, DL);
4634 }
4635 
4636 /// Try to determine the resulting constant values in phi nodes
4637 /// at the common destination basic block, *CommonDest, for one of the case
4638 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4639 /// case), of a switch instruction SI.
4640 static bool
4642  BasicBlock **CommonDest,
4643  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4644  const DataLayout &DL, const TargetTransformInfo &TTI) {
4645  // The block from which we enter the common destination.
4646  BasicBlock *Pred = SI->getParent();
4647 
4648  // If CaseDest is empty except for some side-effect free instructions through
4649  // which we can constant-propagate the CaseVal, continue to its successor.
4651  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4652  for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4653  if (I.isTerminator()) {
4654  // If the terminator is a simple branch, continue to the next block.
4655  if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
4656  return false;
4657  Pred = CaseDest;
4658  CaseDest = I.getSuccessor(0);
4659  } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4660  // Instruction is side-effect free and constant.
4661 
4662  // If the instruction has uses outside this block or a phi node slot for
4663  // the block, it is not safe to bypass the instruction since it would then
4664  // no longer dominate all its uses.
4665  for (auto &Use : I.uses()) {
4666  User *User = Use.getUser();
4667  if (Instruction *I = dyn_cast<Instruction>(User))
4668  if (I->getParent() == CaseDest)
4669  continue;
4670  if (PHINode *Phi = dyn_cast<PHINode>(User))
4671  if (Phi->getIncomingBlock(Use) == CaseDest)
4672  continue;
4673  return false;
4674  }
4675 
4676  ConstantPool.insert(std::make_pair(&I, C));
4677  } else {
4678  break;
4679  }
4680  }
4681 
4682  // If we did not have a CommonDest before, use the current one.
4683  if (!*CommonDest)
4684  *CommonDest = CaseDest;
4685  // If the destination isn't the common one, abort.
4686  if (CaseDest != *CommonDest)
4687  return false;
4688 
4689  // Get the values for this case from phi nodes in the destination block.
4690  for (PHINode &PHI : (*CommonDest)->phis()) {
4691  int Idx = PHI.getBasicBlockIndex(Pred);
4692  if (Idx == -1)
4693  continue;
4694 
4695  Constant *ConstVal =
4696  LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4697  if (!ConstVal)
4698  return false;
4699 
4700  // Be conservative about which kinds of constants we support.
4701  if (!ValidLookupTableConstant(ConstVal, TTI))
4702  return false;
4703 
4704  Res.push_back(std::make_pair(&PHI, ConstVal));
4705  }
4706 
4707  return Res.size() > 0;
4708 }
4709 
4710 // Helper function used to add CaseVal to the list of cases that generate
4711 // Result. Returns the updated number of cases that generate this result.
4712 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4713  SwitchCaseResultVectorTy &UniqueResults,
4714  Constant *Result) {
4715  for (auto &I : UniqueResults) {
4716  if (I.first == Result) {
4717  I.second.push_back(CaseVal);
4718  return I.second.size();
4719  }
4720  }
4721  UniqueResults.push_back(
4722  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4723  return 1;
4724 }
4725 
4726 // Helper function that initializes a map containing
4727 // results for the PHI node of the common destination block for a switch
4728 // instruction. Returns false if multiple PHI nodes have been found or if
4729 // there is not a common destination block for the switch.
4730 static bool
4732  SwitchCaseResultVectorTy &UniqueResults,
4733  Constant *&DefaultResult, const DataLayout &DL,
4734  const TargetTransformInfo &TTI,
4735  uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4736  for (auto &I : SI->cases()) {
4737  ConstantInt *CaseVal = I.getCaseValue();
4738 
4739  // Resulting value at phi nodes for this case value.
4740  SwitchCaseResultsTy Results;
4741  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4742  DL, TTI))
4743  return false;
4744 
4745  // Only one value per case is permitted.
4746  if (Results.size() > 1)
4747  return false;
4748 
4749  // Add the case->result mapping to UniqueResults.
4750  const uintptr_t NumCasesForResult =
4751  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4752 
4753  // Early out if there are too many cases for this result.
4754  if (NumCasesForResult > MaxCasesPerResult)
4755  return false;
4756 
4757  // Early out if there are too many unique results.
4758  if (UniqueResults.size() > MaxUniqueResults)
4759  return false;
4760 
4761  // Check the PHI consistency.
4762  if (!PHI)
4763  PHI = Results[0].first;
4764  else if (PHI != Results[0].first)
4765  return false;
4766  }
4767  // Find the default result value.
4769  BasicBlock *DefaultDest = SI->getDefaultDest();
4770  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4771  DL, TTI);
4772  // If the default value is not found abort unless the default destination
4773  // is unreachable.
4774  DefaultResult =
4775  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4776  if ((!DefaultResult &&
4777  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4778  return false;
4779 
4780  return true;
4781 }
4782 
4783 // Helper function that checks if it is possible to transform a switch with only
4784 // two cases (or two cases + default) that produces a result into a select.
4785 // Example:
4786 // switch (a) {
4787 // case 10: %0 = icmp eq i32 %a, 10
4788 // return 10; %1 = select i1 %0, i32 10, i32 4
4789 // case 20: ----> %2 = icmp eq i32 %a, 20
4790 // return 2; %3 = select i1 %2, i32 2, i32 %1
4791 // default:
4792 // return 4;
4793 // }
4794 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4795  Constant *DefaultResult, Value *Condition,
4796  IRBuilder<> &Builder) {
4797  assert(ResultVector.size() == 2 &&
4798  "We should have exactly two unique results at this point");
4799  // If we are selecting between only two cases transform into a simple
4800  // select or a two-way select if default is possible.
4801  if (ResultVector[0].second.size() == 1 &&
4802  ResultVector[1].second.size() == 1) {
4803  ConstantInt *const FirstCase = ResultVector[0].second[0];
4804  ConstantInt *const SecondCase = ResultVector[1].second[0];
4805 
4806  bool DefaultCanTrigger = DefaultResult;
4807  Value *SelectValue = ResultVector[1].first;
4808  if (DefaultCanTrigger) {
4809  Value *const ValueCompare =
4810  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4811  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4812  DefaultResult, "switch.select");
4813  }
4814  Value *const ValueCompare =
4815  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4816  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4817  SelectValue, "switch.select");
4818  }
4819 
4820  return nullptr;
4821 }
4822 
4823 // Helper function to cleanup a switch instruction that has been converted into
4824 // a select, fixing up PHI nodes and basic blocks.
4826  Value *SelectValue,
4827  IRBuilder<> &Builder) {
4828  BasicBlock *SelectBB = SI->getParent();
4829  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4830  PHI->removeIncomingValue(SelectBB);
4831  PHI->addIncoming(SelectValue, SelectBB);
4832 
4833  Builder.CreateBr(PHI->getParent());
4834 
4835  // Remove the switch.
4836  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4837  BasicBlock *Succ = SI->getSuccessor(i);
4838 
4839  if (Succ == PHI->getParent())
4840  continue;
4841  Succ->removePredecessor(SelectBB);
4842  }
4843  SI->eraseFromParent();
4844 }
4845 
4846 /// If the switch is only used to initialize one or more
4847 /// phi nodes in a common successor block with only two different
4848 /// constant values, replace the switch with select.
4849 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4850  const DataLayout &DL,
4851  const TargetTransformInfo &TTI) {
4852  Value *const Cond = SI->getCondition();
4853  PHINode *PHI = nullptr;
4854  BasicBlock *CommonDest = nullptr;
4855  Constant *DefaultResult;
4856  SwitchCaseResultVectorTy UniqueResults;
4857  // Collect all the cases that will deliver the same value from the switch.
4858  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4859  DL, TTI, 2, 1))
4860  return false;
4861  // Selects choose between maximum two values.
4862  if (UniqueResults.size() != 2)
4863  return false;
4864  assert(PHI != nullptr && "PHI for value select not found");
4865 
4866  Builder.SetInsertPoint(SI);
4867  Value *SelectValue =
4868  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4869  if (SelectValue) {
4870  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4871  return true;
4872  }
4873  // The switch couldn't be converted into a select.
4874  return false;
4875 }
4876 
4877 namespace {
4878 
4879 /// This class represents a lookup table that can be used to replace a switch.
4880 class SwitchLookupTable {
4881 public:
4882  /// Create a lookup table to use as a switch replacement with the contents
4883  /// of Values, using DefaultValue to fill any holes in the table.
4884  SwitchLookupTable(
4885  Module &M, uint64_t TableSize, ConstantInt *Offset,
4886  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4887  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4888 
4889  /// Build instructions with Builder to retrieve the value at
4890  /// the position given by Index in the lookup table.
4891  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4892 
4893  /// Return true if a table with TableSize elements of
4894  /// type ElementType would fit in a target-legal register.
4895  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4896  Type *ElementType);
4897 
4898 private:
4899  // Depending on the contents of the table, it can be represented in
4900  // different ways.
4901  enum {
4902  // For tables where each element contains the same value, we just have to
4903  // store that single value and return it for each lookup.
4904  SingleValueKind,
4905 
4906  // For tables where there is a linear relationship between table index
4907  // and values. We calculate the result with a simple multiplication
4908  // and addition instead of a table lookup.
4909  LinearMapKind,
4910 
4911  // For small tables with integer elements, we can pack them into a bitmap
4912  // that fits into a target-legal register. Values are retrieved by
4913  // shift and mask operations.
4914  BitMapKind,
4915 
4916  // The table is stored as an array of values. Values are retrieved by load
4917  // instructions from the table.
4918  ArrayKind
4919  } Kind;
4920 
4921  // For SingleValueKind, this is the single value.
4922  Constant *SingleValue = nullptr;
4923 
4924  // For BitMapKind, this is the bitmap.
4925  ConstantInt *BitMap = nullptr;
4926  IntegerType *BitMapElementTy = nullptr;
4927 
4928  // For LinearMapKind, these are the constants used to derive the value.
4929  ConstantInt *LinearOffset = nullptr;
4930  ConstantInt *LinearMultiplier = nullptr;
4931 
4932  // For ArrayKind, this is the array.
4933  GlobalVariable *Array = nullptr;
4934 };
4935 
4936 } // end anonymous namespace
4937 
4938 SwitchLookupTable::SwitchLookupTable(
4939  Module &M, uint64_t TableSize, ConstantInt *Offset,
4940  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4941  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4942  assert(Values.size() && "Can't build lookup table without values!");
4943  assert(TableSize >= Values.size() && "Can't fit values in table!");
4944 
4945  // If all values in the table are equal, this is that value.
4946  SingleValue = Values.begin()->second;
4947 
4948  Type *ValueType = Values.begin()->second->getType();
4949 
4950  // Build up the table contents.
4951  SmallVector<Constant *, 64> TableContents(TableSize);
4952  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4953  ConstantInt *CaseVal = Values[I].first;
4954  Constant *CaseRes = Values[I].second;
4955  assert(CaseRes->getType() == ValueType);
4956 
4957  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4958  TableContents[Idx] = CaseRes;
4959 
4960  if (CaseRes != SingleValue)
4961  SingleValue = nullptr;
4962  }
4963 
4964  // Fill in any holes in the table with the default result.
4965  if (Values.size() < TableSize) {
4966  assert(DefaultValue &&
4967  "Need a default value to fill the lookup table holes.");
4968  assert(DefaultValue->getType() == ValueType);
4969  for (uint64_t I = 0; I < TableSize; ++I) {
4970  if (!TableContents[I])
4971  TableContents[I] = DefaultValue;
4972  }
4973 
4974  if (DefaultValue != SingleValue)
4975  SingleValue = nullptr;
4976  }
4977 
4978  // If each element in the table contains the same value, we only need to store
4979  // that single value.
4980  if (SingleValue) {
4981  Kind = SingleValueKind;
4982  return;
4983  }
4984 
4985  // Check if we can derive the value with a linear transformation from the
4986  // table index.
4987  if (isa<IntegerType>(ValueType)) {
4988  bool LinearMappingPossible = true;
4989  APInt PrevVal;
4990  APInt DistToPrev;
4991  assert(TableSize >= 2 && "Should be a SingleValue table.");
4992  // Check if there is the same distance between two consecutive values.
4993  for (uint64_t I = 0; I < TableSize; ++I) {
4994  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4995  if (!ConstVal) {
4996  // This is an undef. We could deal with it, but undefs in lookup tables
4997  // are very seldom. It's probably not worth the additional complexity.
4998  LinearMappingPossible = false;
4999  break;
5000  }
5001  const APInt &Val = ConstVal->getValue();
5002  if (I != 0) {
5003  APInt Dist = Val - PrevVal;
5004  if (I == 1) {
5005  DistToPrev = Dist;
5006  } else if (Dist != DistToPrev) {
5007  LinearMappingPossible = false;
5008  break;
5009  }
5010  }
5011  PrevVal = Val;
5012  }
5013  if (LinearMappingPossible) {
5014  LinearOffset = cast<ConstantInt>(TableContents[0]);
5015  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5016  Kind = LinearMapKind;
5017  ++NumLinearMaps;
5018  return;
5019  }
5020  }
5021 
5022  // If the type is integer and the table fits in a register, build a bitmap.
5023  if (WouldFitInRegister(DL, TableSize, ValueType)) {
5024  IntegerType *IT = cast<IntegerType>(ValueType);
5025  APInt TableInt(TableSize * IT->getBitWidth(), 0);
5026  for (uint64_t I = TableSize; I > 0; --I) {
5027  TableInt <<= IT->getBitWidth();
5028  // Insert values into the bitmap. Undef values are set to zero.
5029  if (!isa<UndefValue>(TableContents[I - 1])) {
5030  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5031  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5032  }
5033  }
5034  BitMap = ConstantInt::get(M.getContext(), TableInt);
5035  BitMapElementTy = IT;
5036  Kind = BitMapKind;
5037  ++NumBitMaps;
5038  return;
5039  }
5040 
5041  // Store the table in an array.
5042  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5043  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5044 
5045  Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
5046  GlobalVariable::PrivateLinkage, Initializer,
5047  "switch.table." + FuncName);
5048  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5049  // Set the alignment to that of an array items. We will be only loading one
5050  // value out of it.
5051  Array->setAlignment(DL.getPrefTypeAlignment(ValueType));
5052  Kind = ArrayKind;
5053 }
5054 
5055 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5056  switch (Kind) {
5057  case SingleValueKind:
5058  return SingleValue;
5059  case LinearMapKind: {
5060  // Derive the result value from the input value.
5061  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5062  false, "switch.idx.cast");
5063  if (!LinearMultiplier->isOne())
5064  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5065  if (!LinearOffset->isZero())
5066  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5067  return Result;
5068  }
5069  case BitMapKind: {
5070  // Type of the bitmap (e.g. i59).
5071  IntegerType *MapTy = BitMap->getType();
5072 
5073  // Cast Index to the same type as the bitmap.
5074  // Note: The Index is <= the number of elements in the table, so
5075  // truncating it to the width of the bitmask is safe.
5076  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5077 
5078  // Multiply the shift amount by the element width.
5079  ShiftAmt = Builder.CreateMul(
5080  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5081  "switch.shiftamt");
5082 
5083  // Shift down.
5084  Value *DownShifted =
5085  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5086  // Mask off.
5087  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5088  }
5089  case ArrayKind: {
5090  // Make sure the table index will not overflow when treated as signed.
5091  IntegerType *IT = cast<IntegerType>(Index->getType());
5092  uint64_t TableSize =
5093  Array->getInitializer()->getType()->getArrayNumElements();
5094  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5095  Index = Builder.CreateZExt(
5096  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5097  "switch.tableidx.zext");
5098 
5099  Value *GEPIndices[] = {Builder.getInt32(0), Index};
5100  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5101  GEPIndices, "switch.gep");
5102  return Builder.CreateLoad(
5103  cast<ArrayType>(Array->getValueType())->getElementType(), GEP,
5104  "switch.load");
5105  }
5106  }
5107  llvm_unreachable("Unknown lookup table kind!");
5108 }
5109 
5110 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5111  uint64_t TableSize,
5112  Type *ElementType) {
5113  auto *IT = dyn_cast<IntegerType>(ElementType);
5114  if (!IT)
5115  return false;
5116  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5117  // are <= 15, we could try to narrow the type.
5118 
5119  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5120  if (TableSize >= UINT_MAX / IT->getBitWidth())
5121  return false;
5122  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5123 }
5124 
5125 /// Determine whether a lookup table should be built for this switch, based on
5126 /// the number of cases, size of the table, and the types of the results.
5127 static bool
5128 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5129  const TargetTransformInfo &TTI, const DataLayout &DL,
5130  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5131  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5132  return false; // TableSize overflowed, or mul below might overflow.
5133 
5134  bool AllTablesFitInRegister = true;
5135  bool HasIllegalType = false;
5136  for (const auto &I : ResultTypes) {
5137  Type *Ty = I.second;
5138 
5139  // Saturate this flag to true.
5140  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5141 
5142  // Saturate this flag to false.
5143  AllTablesFitInRegister =
5144  AllTablesFitInRegister &&
5145  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5146 
5147  // If both flags saturate, we're done. NOTE: This *only* works with
5148  // saturating flags, and all flags have to saturate first due to the
5149  // non-deterministic behavior of iterating over a dense map.
5150  if (HasIllegalType && !AllTablesFitInRegister)
5151  break;
5152  }
5153 
5154  // If each table would fit in a register, we should build it anyway.
5155  if (AllTablesFitInRegister)
5156  return true;
5157 
5158  // Don't build a table that doesn't fit in-register if it has illegal types.
5159  if (HasIllegalType)
5160  return false;
5161 
5162  // The table density should be at least 40%. This is the same criterion as for
5163  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5164  // FIXME: Find the best cut-off.
5165  return SI->getNumCases() * 10 >= TableSize * 4;
5166 }
5167 
5168 /// Try to reuse the switch table index compare. Following pattern:
5169 /// \code
5170 /// if (idx < tablesize)
5171 /// r = table[idx]; // table does not contain default_value
5172 /// else
5173 /// r = default_value;
5174 /// if (r != default_value)
5175 /// ...
5176 /// \endcode
5177 /// Is optimized to:
5178 /// \code
5179 /// cond = idx < tablesize;
5180 /// if (cond)
5181 /// r = table[idx];
5182 /// else
5183 /// r = default_value;
5184 /// if (cond)
5185 /// ...
5186 /// \endcode
5187 /// Jump threading will then eliminate the second if(cond).
5188 static void reuseTableCompare(
5189  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5190  Constant *DefaultValue,
5191  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5192  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5193  if (!CmpInst)
5194  return;
5195 
5196  // We require that the compare is in the same block as the phi so that jump
5197  // threading can do its work afterwards.
5198  if (CmpInst->getParent() != PhiBlock)
5199  return;
5200 
5201  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5202  if (!CmpOp1)
5203  return;
5204 
5205  Value *RangeCmp = RangeCheckBranch->getCondition();
5206  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5207  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5208 
5209  // Check if the compare with the default value is constant true or false.
5210  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5211  DefaultValue, CmpOp1, true);
5212  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5213  return;
5214 
5215  // Check if the compare with the case values is distinct from the default
5216  // compare result.
5217  for (auto ValuePair : Values) {
5218  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5219  ValuePair.second, CmpOp1, true);
5220  if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5221  return;
5222  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5223  "Expect true or false as compare result.");
5224  }
5225 
5226  // Check if the branch instruction dominates the phi node. It's a simple
5227  // dominance check, but sufficient for our needs.
5228  // Although this check is invariant in the calling loops, it's better to do it
5229  // at this late stage. Practically we do it at most once for a switch.
5230  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5231  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5232  BasicBlock *Pred = *PI;
5233  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5234  return;
5235  }
5236 
5237  if (DefaultConst == FalseConst) {
5238  // The compare yields the same result. We can replace it.
5239  CmpInst->replaceAllUsesWith(RangeCmp);
5240  ++NumTableCmpReuses;
5241  } else {
5242  // The compare yields the same result, just inverted. We can replace it.
5243  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5244  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5245  RangeCheckBranch);
5246  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5247  ++NumTableCmpReuses;
5248  }
5249 }
5250 
5251 /// If the switch is only used to initialize one or more phi nodes in a common
5252 /// successor block with different constant values, replace the switch with
5253 /// lookup tables.
5255  const DataLayout &DL,
5256  const TargetTransformInfo &TTI) {
5257  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5258 
5259  Function *Fn = SI->getParent()->getParent();
5260  // Only build lookup table when we have a target that supports it or the
5261  // attribute is not set.
5262  if (!TTI.shouldBuildLookupTables() ||
5263  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5264  return false;
5265 
5266  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5267  // split off a dense part and build a lookup table for that.
5268 
5269  // FIXME: This creates arrays of GEPs to constant strings, which means each
5270  // GEP needs a runtime relocation in PIC code. We should just build one big
5271  // string and lookup indices into that.
5272 
5273  // Ignore switches with less than three cases. Lookup tables will not make
5274  // them faster, so we don't analyze them.
5275  if (SI->getNumCases() < 3)
5276  return false;
5277 
5278  // Figure out the corresponding result for each case value and phi node in the
5279  // common destination, as well as the min and max case values.
5280  assert(!empty(SI->cases()));
5281  SwitchInst::CaseIt CI = SI->case_begin();
5282  ConstantInt *MinCaseVal = CI->getCaseValue();
5283  ConstantInt *MaxCaseVal = CI->getCaseValue();
5284 
5285  BasicBlock *CommonDest = nullptr;
5286 
5287  using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5289 
5290  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5293 
5294  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5295  ConstantInt *CaseVal = CI->getCaseValue();
5296  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5297  MinCaseVal = CaseVal;
5298  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5299  MaxCaseVal = CaseVal;
5300 
5301  // Resulting value at phi nodes for this case value.
5302  using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5303  ResultsTy Results;
5304  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5305  Results, DL, TTI))
5306  return false;
5307 
5308  // Append the result from this case to the list for each phi.
5309  for (const auto &I : Results) {
5310  PHINode *PHI = I.first;
5311  Constant *Value = I.second;
5312  if (!ResultLists.count(PHI))
5313  PHIs.push_back(PHI);
5314  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5315  }
5316  }
5317 
5318  // Keep track of the result types.
5319  for (PHINode *PHI : PHIs) {
5320  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5321  }
5322 
5323  uint64_t NumResults = ResultLists[PHIs[0]].size();
5324  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5325  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5326  bool TableHasHoles = (NumResults < TableSize);
5327 
5328  // If the table has holes, we need a constant result for the default case
5329  // or a bitmask that fits in a register.
5330  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5331  bool HasDefaultResults =
5332  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5333  DefaultResultsList, DL, TTI);
5334 
5335  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5336  if (NeedMask) {
5337  // As an extra penalty for the validity test we require more cases.
5338  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5339  return false;
5340  if (!DL.fitsInLegalInteger(TableSize))
5341  return false;
5342  }
5343 
5344  for (const auto &I : DefaultResultsList) {
5345  PHINode *PHI = I.first;
5346  Constant *Result = I.second;
5347  DefaultResults[PHI] = Result;
5348  }
5349 
5350  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5351  return false;
5352 
5353  // Create the BB that does the lookups.
5354  Module &Mod = *CommonDest->getParent()->getParent();
5355  BasicBlock *LookupBB = BasicBlock::Create(
5356  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5357 
5358  // Compute the table index value.
5359  Builder.SetInsertPoint(SI);
5360  Value *TableIndex;
5361  if (MinCaseVal->isNullValue())
5362  TableIndex = SI->getCondition();
5363  else
5364  TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5365  "switch.tableidx");
5366 
5367  // Compute the maximum table size representable by the integer type we are
5368  // switching upon.
5369  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5370  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5371  assert(MaxTableSize >= TableSize &&
5372  "It is impossible for a switch to have more entries than the max "
5373  "representable value of its input integer type's size.");
5374 
5375  // If the default destination is unreachable, or if the lookup table covers
5376  // all values of the conditional variable, branch directly to the lookup table
5377  // BB. Otherwise, check that the condition is within the case range.
5378  const bool DefaultIsReachable =
5379  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5380  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5381  BranchInst *RangeCheckBranch = nullptr;
5382 
5383  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5384  Builder.CreateBr(LookupBB);
5385  // Note: We call removeProdecessor later since we need to be able to get the
5386  // PHI value for the default case in case we're using a bit mask.
5387  } else {
5388  Value *Cmp = Builder.CreateICmpULT(
5389  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5390  RangeCheckBranch =
5391  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5392  }
5393 
5394  // Populate the BB that does the lookups.
5395  Builder.SetInsertPoint(LookupBB);
5396 
5397  if (NeedMask) {
5398  // Before doing the lookup, we do the hole check. The LookupBB is therefore
5399  // re-purposed to do the hole check, and we create a new LookupBB.
5400  BasicBlock *MaskBB = LookupBB;
5401  MaskBB->setName("switch.hole_check");
5402  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5403  CommonDest->getParent(), CommonDest);
5404 
5405  // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5406  // unnecessary illegal types.
5407  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5408  APInt MaskInt(TableSizePowOf2, 0);
5409  APInt One(TableSizePowOf2, 1);
5410  // Build bitmask; fill in a 1 bit for every case.
5411  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5412  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5413  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5414  .getLimitedValue();
5415  MaskInt |= One << Idx;
5416  }
5417  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5418 
5419  // Get the TableIndex'th bit of the bitmask.
5420  // If this bit is 0 (meaning hole) jump to the default destination,
5421  // else continue with table lookup.
5422  IntegerType *MapTy = TableMask->getType();
5423  Value *MaskIndex =
5424  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5425  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5426  Value *LoBit = Builder.CreateTrunc(
5427  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5428  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5429 
5430  Builder.SetInsertPoint(LookupBB);
5431  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5432  }
5433 
5434  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5435  // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5436  // do not delete PHINodes here.
5438  /*KeepOneInputPHIs=*/true);
5439  }
5440 
5441  bool ReturnedEarly = false;
5442  for (PHINode *PHI : PHIs) {
5443  const ResultListTy &ResultList = ResultLists[PHI];
5444 
5445  // If using a bitmask, use any value to fill the lookup table holes.
5446  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5447  StringRef FuncName = Fn->getName();
5448  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5449  FuncName);
5450 
5451  Value *Result = Table.BuildLookup(TableIndex, Builder);
5452 
5453  // If the result is used to return immediately from the function, we want to
5454  // do that right here.
5455  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5456  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5457  Builder.CreateRet(Result);
5458  ReturnedEarly = true;
5459  break;
5460  }
5461 
5462  // Do a small peephole optimization: re-use the switch table compare if
5463  // possible.
5464  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5465  BasicBlock *PhiBlock = PHI->getParent();
5466  // Search for compare instructions which use the phi.
5467  for (auto *User : PHI->users()) {
5468  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5469  }
5470  }
5471 
5472  PHI->addIncoming(Result, LookupBB);
5473  }
5474 
5475  if (!ReturnedEarly)
5476  Builder.CreateBr(CommonDest);
5477 
5478  // Remove the switch.
5479  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5480  BasicBlock *Succ = SI->getSuccessor(i);
5481 
5482  if (Succ == SI->getDefaultDest())
5483  continue;
5484  Succ->removePredecessor(SI->getParent());
5485  }
5486  SI->eraseFromParent();
5487 
5488  ++NumLookupTables;
5489  if (NeedMask)
5490  ++NumLookupTablesHoles;
5491  return true;
5492 }
5493 
5494 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5495  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5496  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5497  uint64_t Range = Diff + 1;
5498  uint64_t NumCases = Values.size();
5499  // 40% is the default density for building a jump table in optsize/minsize mode.
5500  uint64_t MinDensity = 40;
5501 
5502  return NumCases * 100 >= Range * MinDensity;
5503 }
5504 
5505 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5506 /// of cases.
5507 ///
5508 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5509 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5510 ///
5511 /// This converts a sparse switch into a dense switch which allows better
5512 /// lowering and could also allow transforming into a lookup table.
5514  const DataLayout &DL,
5515  const TargetTransformInfo &TTI) {
5516  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5517  if (CondTy->getIntegerBitWidth() > 64 ||
5518  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5519  return false;
5520  // Only bother with this optimization if there are more than 3 switch cases;
5521  // SDAG will only bother creating jump tables for 4 or more cases.
5522  if (SI->getNumCases() < 4)
5523  return false;
5524 
5525  // This transform is agnostic to the signedness of the input or case values. We
5526  // can treat the case values as signed or unsigned. We can optimize more common
5527  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5528  // as signed.
5529  SmallVector<int64_t,4> Values;
5530  for (auto &C : SI->cases())
5531  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5532  llvm::sort(Values);
5533 
5534  // If the switch is already dense, there's nothing useful to do here.
5535  if (isSwitchDense(Values))
5536  return false;
5537 
5538  // First, transform the values such that they start at zero and ascend.
5539  int64_t Base = Values[0];
5540  for (auto &V : Values)
5541  V -= (uint64_t)(Base);
5542 
5543  // Now we have signed numbers that have been shifted so that, given enough
5544  // precision, there are no negative values. Since the rest of the transform
5545  // is bitwise only, we switch now to an unsigned representation.
5546  uint64_t GCD = 0;
5547  for (auto &V : Values)
5548  GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5549 
5550  // This transform can be done speculatively because it is so cheap - it results
5551  // in a single rotate operation being inserted. This can only happen if the
5552  // factor extracted is a power of 2.
5553  // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5554  // inverse of GCD and then perform this transform.
5555  // FIXME: It's possible that optimizing a switch on powers of two might also
5556  // be beneficial - flag values are often powers of two and we could use a CLZ
5557  // as the key function.
5558  if (GCD <= 1 || !isPowerOf2_64(GCD))
5559  // No common divisor found or too expensive to compute key function.
5560  return false;
5561 
5562  unsigned Shift = Log2_64(GCD);
5563  for (auto &V : Values)
5564  V = (int64_t)((uint64_t)V >> Shift);
5565 
5566  if (!isSwitchDense(Values))
5567  // Transform didn't create a dense switch.
5568  return false;
5569 
5570  // The obvious transform is to shift the switch condition right and emit a
5571  // check that the condition actually cleanly divided by GCD, i.e.
5572  // C & (1 << Shift - 1) == 0
5573  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5574  //
5575  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5576  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5577  // are nonzero then the switch condition will be very large and will hit the
5578  // default case.
5579 
5580  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5581  Builder.SetInsertPoint(SI);
5582  auto *ShiftC = ConstantInt::get(Ty, Shift);
5583  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5584  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5585  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5586  auto *Rot = Builder.CreateOr(LShr, Shl);
5587  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5588 
5589  for (auto Case : SI->cases()) {
5590  auto *Orig = Case.getCaseValue();
5591  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5592  Case.setValue(
5593  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5594  }
5595  return true;
5596 }