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. Any non-store instruction
1442  // must have exactly one use, and we check later that use is by a single,
1443  // common PHI instruction in the successor.
1444  for (auto *I : Insts) {
1445  // These instructions may change or break semantics if moved.
1446  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1447  I->getType()->isTokenTy())
1448  return false;
1449 
1450  // Conservatively return false if I is an inline-asm instruction. Sinking
1451  // and merging inline-asm instructions can potentially create arguments
1452  // that cannot satisfy the inline-asm constraints.
1453  if (const auto *C = dyn_cast<CallBase>(I))
1454  if (C->isInlineAsm())
1455  return false;
1456 
1457  // Everything must have only one use too, apart from stores which
1458  // have no uses.
1459  if (!isa<StoreInst>(I) && !I->hasOneUse())
1460  return false;
1461  }
1462 
1463  const Instruction *I0 = Insts.front();
1464  for (auto *I : Insts)
1465  if (!I->isSameOperationAs(I0))
1466  return false;
1467 
1468  // All instructions in Insts are known to be the same opcode. If they aren't
1469  // stores, check the only user of each is a PHI or in the same block as the
1470  // instruction, because if a user is in the same block as an instruction
1471  // we're contemplating sinking, it must already be determined to be sinkable.
1472  if (!isa<StoreInst>(I0)) {
1473  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1474  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1475  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1476  auto *U = cast<Instruction>(*I->user_begin());
1477  return (PNUse &&
1478  PNUse->getParent() == Succ &&
1479  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1480  U->getParent() == I->getParent();
1481  }))
1482  return false;
1483  }
1484 
1485  // Because SROA can't handle speculating stores of selects, try not
1486  // to sink loads or stores of allocas when we'd have to create a PHI for
1487  // the address operand. Also, because it is likely that loads or stores
1488  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1489  // This can cause code churn which can have unintended consequences down
1490  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1491  // FIXME: This is a workaround for a deficiency in SROA - see
1492  // https://llvm.org/bugs/show_bug.cgi?id=30188
1493  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1494  return isa<AllocaInst>(I->getOperand(1));
1495  }))
1496  return false;
1497  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1498  return isa<AllocaInst>(I->getOperand(0));
1499  }))
1500  return false;
1501 
1502  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1503  if (I0->getOperand(OI)->getType()->isTokenTy())
1504  // Don't touch any operand of token type.
1505  return false;
1506 
1507  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1508  assert(I->getNumOperands() == I0->getNumOperands());
1509  return I->getOperand(OI) == I0->getOperand(OI);
1510  };
1511  if (!all_of(Insts, SameAsI0)) {
1512  if (!canReplaceOperandWithVariable(I0, OI))
1513  // We can't create a PHI from this GEP.
1514  return false;
1515  // Don't create indirect calls! The called value is the final operand.
1516  if (isa<CallBase>(I0) && OI == OE - 1) {
1517  // FIXME: if the call was *already* indirect, we should do this.
1518  return false;
1519  }
1520  for (auto *I : Insts)
1521  PHIOperands[I].push_back(I->getOperand(OI));
1522  }
1523  }
1524  return true;
1525 }
1526 
1527 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1528 // instruction of every block in Blocks to their common successor, commoning
1529 // into one instruction.
1531  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1532 
1533  // canSinkLastInstruction returning true guarantees that every block has at
1534  // least one non-terminator instruction.
1536  for (auto *BB : Blocks) {
1537  Instruction *I = BB->getTerminator();
1538  do {
1539  I = I->getPrevNode();
1540  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1541  if (!isa<DbgInfoIntrinsic>(I))
1542  Insts.push_back(I);
1543  }
1544 
1545  // The only checking we need to do now is that all users of all instructions
1546  // are the same PHI node. canSinkLastInstruction should have checked this but
1547  // it is slightly over-aggressive - it gets confused by commutative instructions
1548  // so double-check it here.
1549  Instruction *I0 = Insts.front();
1550  if (!isa<StoreInst>(I0)) {
1551  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1552  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1553  auto *U = cast<Instruction>(*I->user_begin());
1554  return U == PNUse;
1555  }))
1556  return false;
1557  }
1558 
1559  // We don't need to do any more checking here; canSinkLastInstruction should
1560  // have done it all for us.
1561  SmallVector<Value*, 4> NewOperands;
1562  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1563  // This check is different to that in canSinkLastInstruction. There, we
1564  // cared about the global view once simplifycfg (and instcombine) have
1565  // completed - it takes into account PHIs that become trivially
1566  // simplifiable. However here we need a more local view; if an operand
1567  // differs we create a PHI and rely on instcombine to clean up the very
1568  // small mess we may make.
1569  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1570  return I->getOperand(O) != I0->getOperand(O);
1571  });
1572  if (!NeedPHI) {
1573  NewOperands.push_back(I0->getOperand(O));
1574  continue;
1575  }
1576 
1577  // Create a new PHI in the successor block and populate it.
1578  auto *Op = I0->getOperand(O);
1579  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1580  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1581  Op->getName() + ".sink", &BBEnd->front());
1582  for (auto *I : Insts)
1583  PN->addIncoming(I->getOperand(O), I->getParent());
1584  NewOperands.push_back(PN);
1585  }
1586 
1587  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1588  // and move it to the start of the successor block.
1589  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1590  I0->getOperandUse(O).set(NewOperands[O]);
1591  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1592 
1593  // Update metadata and IR flags, and merge debug locations.
1594  for (auto *I : Insts)
1595  if (I != I0) {
1596  // The debug location for the "common" instruction is the merged locations
1597  // of all the commoned instructions. We start with the original location
1598  // of the "common" instruction and iteratively merge each location in the
1599  // loop below.
1600  // This is an N-way merge, which will be inefficient if I0 is a CallInst.
1601  // However, as N-way merge for CallInst is rare, so we use simplified API
1602  // instead of using complex API for N-way merge.
1603  I0->applyMergedLocation(I0->getDebugLoc(), I->getDebugLoc());
1604  combineMetadataForCSE(I0, I, true);
1605  I0->andIRFlags(I);
1606  }
1607 
1608  if (!isa<StoreInst>(I0)) {
1609  // canSinkLastInstruction checked that all instructions were used by
1610  // one and only one PHI node. Find that now, RAUW it to our common
1611  // instruction and nuke it.
1612  assert(I0->hasOneUse());
1613  auto *PN = cast<PHINode>(*I0->user_begin());
1614  PN->replaceAllUsesWith(I0);
1615  PN->eraseFromParent();
1616  }
1617 
1618  // Finally nuke all instructions apart from the common instruction.
1619  for (auto *I : Insts)
1620  if (I != I0)
1621  I->eraseFromParent();
1622 
1623  return true;
1624 }
1625 
1626 namespace {
1627 
1628  // LockstepReverseIterator - Iterates through instructions
1629  // in a set of blocks in reverse order from the first non-terminator.
1630  // For example (assume all blocks have size n):
1631  // LockstepReverseIterator I([B1, B2, B3]);
1632  // *I-- = [B1[n], B2[n], B3[n]];
1633  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1634  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1635  // ...
1636  class LockstepReverseIterator {
1637  ArrayRef<BasicBlock*> Blocks;
1639  bool Fail;
1640 
1641  public:
1642  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) : Blocks(Blocks) {
1643  reset();
1644  }
1645 
1646  void reset() {
1647  Fail = false;
1648  Insts.clear();
1649  for (auto *BB : Blocks) {
1650  Instruction *Inst = BB->getTerminator();
1651  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1652  Inst = Inst->getPrevNode();
1653  if (!Inst) {
1654  // Block wasn't big enough.
1655  Fail = true;
1656  return;
1657  }
1658  Insts.push_back(Inst);
1659  }
1660  }
1661 
1662  bool isValid() const {
1663  return !Fail;
1664  }
1665 
1666  void operator--() {
1667  if (Fail)
1668  return;
1669  for (auto *&Inst : Insts) {
1670  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1671  Inst = Inst->getPrevNode();
1672  // Already at beginning of block.
1673  if (!Inst) {
1674  Fail = true;
1675  return;
1676  }
1677  }
1678  }
1679 
1681  return Insts;
1682  }
1683  };
1684 
1685 } // end anonymous namespace
1686 
1687 /// Check whether BB's predecessors end with unconditional branches. If it is
1688 /// true, sink any common code from the predecessors to BB.
1689 /// We also allow one predecessor to end with conditional branch (but no more
1690 /// than one).
1692  // We support two situations:
1693  // (1) all incoming arcs are unconditional
1694  // (2) one incoming arc is conditional
1695  //
1696  // (2) is very common in switch defaults and
1697  // else-if patterns;
1698  //
1699  // if (a) f(1);
1700  // else if (b) f(2);
1701  //
1702  // produces:
1703  //
1704  // [if]
1705  // / \
1706  // [f(1)] [if]
1707  // | | \
1708  // | | |
1709  // | [f(2)]|
1710  // \ | /
1711  // [ end ]
1712  //
1713  // [end] has two unconditional predecessor arcs and one conditional. The
1714  // conditional refers to the implicit empty 'else' arc. This conditional
1715  // arc can also be caused by an empty default block in a switch.
1716  //
1717  // In this case, we attempt to sink code from all *unconditional* arcs.
1718  // If we can sink instructions from these arcs (determined during the scan
1719  // phase below) we insert a common successor for all unconditional arcs and
1720  // connect that to [end], to enable sinking:
1721  //
1722  // [if]
1723  // / \
1724  // [x(1)] [if]
1725  // | | \
1726  // | | \
1727  // | [x(2)] |
1728  // \ / |
1729  // [sink.split] |
1730  // \ /
1731  // [ end ]
1732  //
1733  SmallVector<BasicBlock*,4> UnconditionalPreds;
1734  Instruction *Cond = nullptr;
1735  for (auto *B : predecessors(BB)) {
1736  auto *T = B->getTerminator();
1737  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1738  UnconditionalPreds.push_back(B);
1739  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1740  Cond = T;
1741  else
1742  return false;
1743  }
1744  if (UnconditionalPreds.size() < 2)
1745  return false;
1746 
1747  bool Changed = false;
1748  // We take a two-step approach to tail sinking. First we scan from the end of
1749  // each block upwards in lockstep. If the n'th instruction from the end of each
1750  // block can be sunk, those instructions are added to ValuesToSink and we
1751  // carry on. If we can sink an instruction but need to PHI-merge some operands
1752  // (because they're not identical in each instruction) we add these to
1753  // PHIOperands.
1754  unsigned ScanIdx = 0;
1755  SmallPtrSet<Value*,4> InstructionsToSink;
1757  LockstepReverseIterator LRI(UnconditionalPreds);
1758  while (LRI.isValid() &&
1759  canSinkInstructions(*LRI, PHIOperands)) {
1760  LLVM_DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0]
1761  << "\n");
1762  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1763  ++ScanIdx;
1764  --LRI;
1765  }
1766 
1767  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1768  unsigned NumPHIdValues = 0;
1769  for (auto *I : *LRI)
1770  for (auto *V : PHIOperands[I])
1771  if (InstructionsToSink.count(V) == 0)
1772  ++NumPHIdValues;
1773  LLVM_DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1774  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1775  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1776  NumPHIInsts++;
1777 
1778  return NumPHIInsts <= 1;
1779  };
1780 
1781  if (ScanIdx > 0 && Cond) {
1782  // Check if we would actually sink anything first! This mutates the CFG and
1783  // adds an extra block. The goal in doing this is to allow instructions that
1784  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1785  // (such as trunc, add) can be sunk and predicated already. So we check that
1786  // we're going to sink at least one non-speculatable instruction.
1787  LRI.reset();
1788  unsigned Idx = 0;
1789  bool Profitable = false;
1790  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1791  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1792  Profitable = true;
1793  break;
1794  }
1795  --LRI;
1796  ++Idx;
1797  }
1798  if (!Profitable)
1799  return false;
1800 
1801  LLVM_DEBUG(dbgs() << "SINK: Splitting edge\n");
1802  // We have a conditional edge and we're going to sink some instructions.
1803  // Insert a new block postdominating all blocks we're going to sink from.
1804  if (!SplitBlockPredecessors(BB, UnconditionalPreds, ".sink.split"))
1805  // Edges couldn't be split.
1806  return false;
1807  Changed = true;
1808  }
1809 
1810  // Now that we've analyzed all potential sinking candidates, perform the
1811  // actual sink. We iteratively sink the last non-terminator of the source
1812  // blocks into their common successor unless doing so would require too
1813  // many PHI instructions to be generated (currently only one PHI is allowed
1814  // per sunk instruction).
1815  //
1816  // We can use InstructionsToSink to discount values needing PHI-merging that will
1817  // actually be sunk in a later iteration. This allows us to be more
1818  // aggressive in what we sink. This does allow a false positive where we
1819  // sink presuming a later value will also be sunk, but stop half way through
1820  // and never actually sink it which means we produce more PHIs than intended.
1821  // This is unlikely in practice though.
1822  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1823  LLVM_DEBUG(dbgs() << "SINK: Sink: "
1824  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1825  << "\n");
1826 
1827  // Because we've sunk every instruction in turn, the current instruction to
1828  // sink is always at index 0.
1829  LRI.reset();
1830  if (!ProfitableToSinkInstruction(LRI)) {
1831  // Too many PHIs would be created.
1832  LLVM_DEBUG(
1833  dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1834  break;
1835  }
1836 
1837  if (!sinkLastInstruction(UnconditionalPreds))
1838  return Changed;
1839  NumSinkCommons++;
1840  Changed = true;
1841  }
1842  return Changed;
1843 }
1844 
1845 /// Determine if we can hoist sink a sole store instruction out of a
1846 /// conditional block.
1847 ///
1848 /// We are looking for code like the following:
1849 /// BrBB:
1850 /// store i32 %add, i32* %arrayidx2
1851 /// ... // No other stores or function calls (we could be calling a memory
1852 /// ... // function).
1853 /// %cmp = icmp ult %x, %y
1854 /// br i1 %cmp, label %EndBB, label %ThenBB
1855 /// ThenBB:
1856 /// store i32 %add5, i32* %arrayidx2
1857 /// br label EndBB
1858 /// EndBB:
1859 /// ...
1860 /// We are going to transform this into:
1861 /// BrBB:
1862 /// store i32 %add, i32* %arrayidx2
1863 /// ... //
1864 /// %cmp = icmp ult %x, %y
1865 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1866 /// store i32 %add.add5, i32* %arrayidx2
1867 /// ...
1868 ///
1869 /// \return The pointer to the value of the previous store if the store can be
1870 /// hoisted into the predecessor block. 0 otherwise.
1872  BasicBlock *StoreBB, BasicBlock *EndBB) {
1873  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1874  if (!StoreToHoist)
1875  return nullptr;
1876 
1877  // Volatile or atomic.
1878  if (!StoreToHoist->isSimple())
1879  return nullptr;
1880 
1881  Value *StorePtr = StoreToHoist->getPointerOperand();
1882 
1883  // Look for a store to the same pointer in BrBB.
1884  unsigned MaxNumInstToLookAt = 9;
1885  for (Instruction &CurI : reverse(BrBB->instructionsWithoutDebug())) {
1886  if (!MaxNumInstToLookAt)
1887  break;
1888  --MaxNumInstToLookAt;
1889 
1890  // Could be calling an instruction that affects memory like free().
1891  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1892  return nullptr;
1893 
1894  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1895  // Found the previous store make sure it stores to the same location.
1896  if (SI->getPointerOperand() == StorePtr)
1897  // Found the previous store, return its value operand.
1898  return SI->getValueOperand();
1899  return nullptr; // Unknown store.
1900  }
1901  }
1902 
1903  return nullptr;
1904 }
1905 
1906 /// Speculate a conditional basic block flattening the CFG.
1907 ///
1908 /// Note that this is a very risky transform currently. Speculating
1909 /// instructions like this is most often not desirable. Instead, there is an MI
1910 /// pass which can do it with full awareness of the resource constraints.
1911 /// However, some cases are "obvious" and we should do directly. An example of
1912 /// this is speculating a single, reasonably cheap instruction.
1913 ///
1914 /// There is only one distinct advantage to flattening the CFG at the IR level:
1915 /// it makes very common but simplistic optimizations such as are common in
1916 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1917 /// modeling their effects with easier to reason about SSA value graphs.
1918 ///
1919 ///
1920 /// An illustration of this transform is turning this IR:
1921 /// \code
1922 /// BB:
1923 /// %cmp = icmp ult %x, %y
1924 /// br i1 %cmp, label %EndBB, label %ThenBB
1925 /// ThenBB:
1926 /// %sub = sub %x, %y
1927 /// br label BB2
1928 /// EndBB:
1929 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1930 /// ...
1931 /// \endcode
1932 ///
1933 /// Into this IR:
1934 /// \code
1935 /// BB:
1936 /// %cmp = icmp ult %x, %y
1937 /// %sub = sub %x, %y
1938 /// %cond = select i1 %cmp, 0, %sub
1939 /// ...
1940 /// \endcode
1941 ///
1942 /// \returns true if the conditional block is removed.
1944  const TargetTransformInfo &TTI) {
1945  // Be conservative for now. FP select instruction can often be expensive.
1946  Value *BrCond = BI->getCondition();
1947  if (isa<FCmpInst>(BrCond))
1948  return false;
1949 
1950  BasicBlock *BB = BI->getParent();
1951  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1952 
1953  // If ThenBB is actually on the false edge of the conditional branch, remember
1954  // to swap the select operands later.
1955  bool Invert = false;
1956  if (ThenBB != BI->getSuccessor(0)) {
1957  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1958  Invert = true;
1959  }
1960  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1961 
1962  // Keep a count of how many times instructions are used within ThenBB when
1963  // they are candidates for sinking into ThenBB. Specifically:
1964  // - They are defined in BB, and
1965  // - They have no side effects, and
1966  // - All of their uses are in ThenBB.
1967  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1968 
1969  SmallVector<Instruction *, 4> SpeculatedDbgIntrinsics;
1970 
1971  unsigned SpeculationCost = 0;
1972  Value *SpeculatedStoreValue = nullptr;
1973  StoreInst *SpeculatedStore = nullptr;
1974  for (BasicBlock::iterator BBI = ThenBB->begin(),
1975  BBE = std::prev(ThenBB->end());
1976  BBI != BBE; ++BBI) {
1977  Instruction *I = &*BBI;
1978  // Skip debug info.
1979  if (isa<DbgInfoIntrinsic>(I)) {
1980  SpeculatedDbgIntrinsics.push_back(I);
1981  continue;
1982  }
1983 
1984  // Only speculatively execute a single instruction (not counting the
1985  // terminator) for now.
1986  ++SpeculationCost;
1987  if (SpeculationCost > 1)
1988  return false;
1989 
1990  // Don't hoist the instruction if it's unsafe or expensive.
1991  if (!isSafeToSpeculativelyExecute(I) &&
1992  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1993  I, BB, ThenBB, EndBB))))
1994  return false;
1995  if (!SpeculatedStoreValue &&
1996  ComputeSpeculationCost(I, TTI) >
1998  return false;
1999 
2000  // Store the store speculation candidate.
2001  if (SpeculatedStoreValue)
2002  SpeculatedStore = cast<StoreInst>(I);
2003 
2004  // Do not hoist the instruction if any of its operands are defined but not
2005  // used in BB. The transformation will prevent the operand from
2006  // being sunk into the use block.
2007  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
2008  Instruction *OpI = dyn_cast<Instruction>(*i);
2009  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
2010  continue; // Not a candidate for sinking.
2011 
2012  ++SinkCandidateUseCounts[OpI];
2013  }
2014  }
2015 
2016  // Consider any sink candidates which are only used in ThenBB as costs for
2017  // speculation. Note, while we iterate over a DenseMap here, we are summing
2018  // and so iteration order isn't significant.
2020  I = SinkCandidateUseCounts.begin(),
2021  E = SinkCandidateUseCounts.end();
2022  I != E; ++I)
2023  if (I->first->hasNUses(I->second)) {
2024  ++SpeculationCost;
2025  if (SpeculationCost > 1)
2026  return false;
2027  }
2028 
2029  // Check that the PHI nodes can be converted to selects.
2030  bool HaveRewritablePHIs = false;
2031  for (PHINode &PN : EndBB->phis()) {
2032  Value *OrigV = PN.getIncomingValueForBlock(BB);
2033  Value *ThenV = PN.getIncomingValueForBlock(ThenBB);
2034 
2035  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
2036  // Skip PHIs which are trivial.
2037  if (ThenV == OrigV)
2038  continue;
2039 
2040  // Don't convert to selects if we could remove undefined behavior instead.
2041  if (passingValueIsAlwaysUndefined(OrigV, &PN) ||
2042  passingValueIsAlwaysUndefined(ThenV, &PN))
2043  return false;
2044 
2045  HaveRewritablePHIs = true;
2046  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
2047  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
2048  if (!OrigCE && !ThenCE)
2049  continue; // Known safe and cheap.
2050 
2051  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
2052  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2053  return false;
2054  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2055  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2056  unsigned MaxCost =
2058  if (OrigCost + ThenCost > MaxCost)
2059  return false;
2060 
2061  // Account for the cost of an unfolded ConstantExpr which could end up
2062  // getting expanded into Instructions.
2063  // FIXME: This doesn't account for how many operations are combined in the
2064  // constant expression.
2065  ++SpeculationCost;
2066  if (SpeculationCost > 1)
2067  return false;
2068  }
2069 
2070  // If there are no PHIs to process, bail early. This helps ensure idempotence
2071  // as well.
2072  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2073  return false;
2074 
2075  // If we get here, we can hoist the instruction and if-convert.
2076  LLVM_DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2077 
2078  // Insert a select of the value of the speculated store.
2079  if (SpeculatedStoreValue) {
2080  IRBuilder<NoFolder> Builder(BI);
2081  Value *TrueV = SpeculatedStore->getValueOperand();
2082  Value *FalseV = SpeculatedStoreValue;
2083  if (Invert)
2084  std::swap(TrueV, FalseV);
2085  Value *S = Builder.CreateSelect(
2086  BrCond, TrueV, FalseV, "spec.store.select", BI);
2087  SpeculatedStore->setOperand(0, S);
2088  SpeculatedStore->applyMergedLocation(BI->getDebugLoc(),
2089  SpeculatedStore->getDebugLoc());
2090  }
2091 
2092  // Metadata can be dependent on the condition we are hoisting above.
2093  // Conservatively strip all metadata on the instruction.
2094  for (auto &I : *ThenBB)
2096 
2097  // Hoist the instructions.
2098  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2099  ThenBB->begin(), std::prev(ThenBB->end()));
2100 
2101  // Insert selects and rewrite the PHI operands.
2102  IRBuilder<NoFolder> Builder(BI);
2103  for (PHINode &PN : EndBB->phis()) {
2104  unsigned OrigI = PN.getBasicBlockIndex(BB);
2105  unsigned ThenI = PN.getBasicBlockIndex(ThenBB);
2106  Value *OrigV = PN.getIncomingValue(OrigI);
2107  Value *ThenV = PN.getIncomingValue(ThenI);
2108 
2109  // Skip PHIs which are trivial.
2110  if (OrigV == ThenV)
2111  continue;
2112 
2113  // Create a select whose true value is the speculatively executed value and
2114  // false value is the preexisting value. Swap them if the branch
2115  // destinations were inverted.
2116  Value *TrueV = ThenV, *FalseV = OrigV;
2117  if (Invert)
2118  std::swap(TrueV, FalseV);
2119  Value *V = Builder.CreateSelect(
2120  BrCond, TrueV, FalseV, "spec.select", BI);
2121  PN.setIncomingValue(OrigI, V);
2122  PN.setIncomingValue(ThenI, V);
2123  }
2124 
2125  // Remove speculated dbg intrinsics.
2126  // FIXME: Is it possible to do this in a more elegant way? Moving/merging the
2127  // dbg value for the different flows and inserting it after the select.
2128  for (Instruction *I : SpeculatedDbgIntrinsics)
2129  I->eraseFromParent();
2130 
2131  ++NumSpeculations;
2132  return true;
2133 }
2134 
2135 /// Return true if we can thread a branch across this block.
2137  unsigned Size = 0;
2138 
2139  for (Instruction &I : BB->instructionsWithoutDebug()) {
2140  if (Size > 10)
2141  return false; // Don't clone large BB's.
2142  ++Size;
2143 
2144  // We can only support instructions that do not define values that are
2145  // live outside of the current basic block.
2146  for (User *U : I.users()) {
2147  Instruction *UI = cast<Instruction>(U);
2148  if (UI->getParent() != BB || isa<PHINode>(UI))
2149  return false;
2150  }
2151 
2152  // Looks ok, continue checking.
2153  }
2154 
2155  return true;
2156 }
2157 
2158 /// If we have a conditional branch on a PHI node value that is defined in the
2159 /// same block as the branch and if any PHI entries are constants, thread edges
2160 /// corresponding to that entry to be branches to their ultimate destination.
2161 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2162  AssumptionCache *AC) {
2163  BasicBlock *BB = BI->getParent();
2164  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2165  // NOTE: we currently cannot transform this case if the PHI node is used
2166  // outside of the block.
2167  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2168  return false;
2169 
2170  // Degenerate case of a single entry PHI.
2171  if (PN->getNumIncomingValues() == 1) {
2173  return true;
2174  }
2175 
2176  // Now we know that this block has multiple preds and two succs.
2178  return false;
2179 
2180  // Can't fold blocks that contain noduplicate or convergent calls.
2181  if (any_of(*BB, [](const Instruction &I) {
2182  const CallInst *CI = dyn_cast<CallInst>(&I);
2183  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2184  }))
2185  return false;
2186 
2187  // Okay, this is a simple enough basic block. See if any phi values are
2188  // constants.
2189  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2191  if (!CB || !CB->getType()->isIntegerTy(1))
2192  continue;
2193 
2194  // Okay, we now know that all edges from PredBB should be revectored to
2195  // branch to RealDest.
2196  BasicBlock *PredBB = PN->getIncomingBlock(i);
2197  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2198 
2199  if (RealDest == BB)
2200  continue; // Skip self loops.
2201  // Skip if the predecessor's terminator is an indirect branch.
2202  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2203  continue;
2204 
2205  // The dest block might have PHI nodes, other predecessors and other
2206  // difficult cases. Instead of being smart about this, just insert a new
2207  // block that jumps to the destination block, effectively splitting
2208  // the edge we are about to create.
2209  BasicBlock *EdgeBB =
2210  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2211  RealDest->getParent(), RealDest);
2212  BranchInst::Create(RealDest, EdgeBB);
2213 
2214  // Update PHI nodes.
2215  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2216 
2217  // BB may have instructions that are being threaded over. Clone these
2218  // instructions into EdgeBB. We know that there will be no uses of the
2219  // cloned instructions outside of EdgeBB.
2220  BasicBlock::iterator InsertPt = EdgeBB->begin();
2221  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2222  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2223  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2224  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2225  continue;
2226  }
2227  // Clone the instruction.
2228  Instruction *N = BBI->clone();
2229  if (BBI->hasName())
2230  N->setName(BBI->getName() + ".c");
2231 
2232  // Update operands due to translation.
2233  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2234  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2235  if (PI != TranslateMap.end())
2236  *i = PI->second;
2237  }
2238 
2239  // Check for trivial simplification.
2240  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2241  if (!BBI->use_empty())
2242  TranslateMap[&*BBI] = V;
2243  if (!N->mayHaveSideEffects()) {
2244  N->deleteValue(); // Instruction folded away, don't need actual inst
2245  N = nullptr;
2246  }
2247  } else {
2248  if (!BBI->use_empty())
2249  TranslateMap[&*BBI] = N;
2250  }
2251  // Insert the new instruction into its new home.
2252  if (N)
2253  EdgeBB->getInstList().insert(InsertPt, N);
2254 
2255  // Register the new instruction with the assumption cache if necessary.
2256  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2257  if (II->getIntrinsicID() == Intrinsic::assume)
2258  AC->registerAssumption(II);
2259  }
2260 
2261  // Loop over all of the edges from PredBB to BB, changing them to branch
2262  // to EdgeBB instead.
2263  Instruction *PredBBTI = PredBB->getTerminator();
2264  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2265  if (PredBBTI->getSuccessor(i) == BB) {
2266  BB->removePredecessor(PredBB);
2267  PredBBTI->setSuccessor(i, EdgeBB);
2268  }
2269 
2270  // Recurse, simplifying any other constants.
2271  return FoldCondBranchOnPHI(BI, DL, AC) || true;
2272  }
2273 
2274  return false;
2275 }
2276 
2277 /// Given a BB that starts with the specified two-entry PHI node,
2278 /// see if we can eliminate it.
2280  const DataLayout &DL) {
2281  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2282  // statement", which has a very simple dominance structure. Basically, we
2283  // are trying to find the condition that is being branched on, which
2284  // subsequently causes this merge to happen. We really want control
2285  // dependence information for this check, but simplifycfg can't keep it up
2286  // to date, and this catches most of the cases we care about anyway.
2287  BasicBlock *BB = PN->getParent();
2288  const Function *Fn = BB->getParent();
2289  if (Fn && Fn->hasFnAttribute(Attribute::OptForFuzzing))
2290  return false;
2291 
2292  BasicBlock *IfTrue, *IfFalse;
2293  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2294  if (!IfCond ||
2295  // Don't bother if the branch will be constant folded trivially.
2296  isa<ConstantInt>(IfCond))
2297  return false;
2298 
2299  // Okay, we found that we can merge this two-entry phi node into a select.
2300  // Doing so would require us to fold *all* two entry phi nodes in this block.
2301  // At some point this becomes non-profitable (particularly if the target
2302  // doesn't support cmov's). Only do this transformation if there are two or
2303  // fewer PHI nodes in this block.
2304  unsigned NumPhis = 0;
2305  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2306  if (NumPhis > 2)
2307  return false;
2308 
2309  // Loop over the PHI's seeing if we can promote them all to select
2310  // instructions. While we are at it, keep track of the instructions
2311  // that need to be moved to the dominating block.
2312  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2313  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2314  MaxCostVal1 = PHINodeFoldingThreshold;
2315  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2316  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2317 
2318  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2319  PHINode *PN = cast<PHINode>(II++);
2320  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2321  PN->replaceAllUsesWith(V);
2322  PN->eraseFromParent();
2323  continue;
2324  }
2325 
2326  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, AggressiveInsts,
2327  MaxCostVal0, TTI) ||
2328  !DominatesMergePoint(PN->getIncomingValue(1), BB, AggressiveInsts,
2329  MaxCostVal1, TTI))
2330  return false;
2331  }
2332 
2333  // If we folded the first phi, PN dangles at this point. Refresh it. If
2334  // we ran out of PHIs then we simplified them all.
2335  PN = dyn_cast<PHINode>(BB->begin());
2336  if (!PN)
2337  return true;
2338 
2339  // Don't fold i1 branches on PHIs which contain binary operators. These can
2340  // often be turned into switches and other things.
2341  if (PN->getType()->isIntegerTy(1) &&
2342  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2343  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2344  isa<BinaryOperator>(IfCond)))
2345  return false;
2346 
2347  // If all PHI nodes are promotable, check to make sure that all instructions
2348  // in the predecessor blocks can be promoted as well. If not, we won't be able
2349  // to get rid of the control flow, so it's not worth promoting to select
2350  // instructions.
2351  BasicBlock *DomBlock = nullptr;
2352  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2353  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2354  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2355  IfBlock1 = nullptr;
2356  } else {
2357  DomBlock = *pred_begin(IfBlock1);
2358  for (BasicBlock::iterator I = IfBlock1->begin(); !I->isTerminator(); ++I)
2359  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2360  // This is not an aggressive instruction that we can promote.
2361  // Because of this, we won't be able to get rid of the control flow, so
2362  // the xform is not worth it.
2363  return false;
2364  }
2365  }
2366 
2367  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2368  IfBlock2 = nullptr;
2369  } else {
2370  DomBlock = *pred_begin(IfBlock2);
2371  for (BasicBlock::iterator I = IfBlock2->begin(); !I->isTerminator(); ++I)
2372  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2373  // This is not an aggressive instruction that we can promote.
2374  // Because of this, we won't be able to get rid of the control flow, so
2375  // the xform is not worth it.
2376  return false;
2377  }
2378  }
2379 
2380  LLVM_DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond
2381  << " T: " << IfTrue->getName()
2382  << " F: " << IfFalse->getName() << "\n");
2383 
2384  // If we can still promote the PHI nodes after this gauntlet of tests,
2385  // do all of the PHI's now.
2386  Instruction *InsertPt = DomBlock->getTerminator();
2387  IRBuilder<NoFolder> Builder(InsertPt);
2388 
2389  // Move all 'aggressive' instructions, which are defined in the
2390  // conditional parts of the if's up to the dominating block.
2391  if (IfBlock1)
2392  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock1);
2393  if (IfBlock2)
2394  hoistAllInstructionsInto(DomBlock, InsertPt, IfBlock2);
2395 
2396  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2397  // Change the PHI node into a select instruction.
2398  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2399  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2400 
2401  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2402  PN->replaceAllUsesWith(Sel);
2403  Sel->takeName(PN);
2404  PN->eraseFromParent();
2405  }
2406 
2407  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2408  // has been flattened. Change DomBlock to jump directly to our new block to
2409  // avoid other simplifycfg's kicking in on the diamond.
2410  Instruction *OldTI = DomBlock->getTerminator();
2411  Builder.SetInsertPoint(OldTI);
2412  Builder.CreateBr(BB);
2413  OldTI->eraseFromParent();
2414  return true;
2415 }
2416 
2417 /// If we found a conditional branch that goes to two returning blocks,
2418 /// try to merge them together into one return,
2419 /// introducing a select if the return values disagree.
2421  IRBuilder<> &Builder) {
2422  assert(BI->isConditional() && "Must be a conditional branch");
2423  BasicBlock *TrueSucc = BI->getSuccessor(0);
2424  BasicBlock *FalseSucc = BI->getSuccessor(1);
2425  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2426  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2427 
2428  // Check to ensure both blocks are empty (just a return) or optionally empty
2429  // with PHI nodes. If there are other instructions, merging would cause extra
2430  // computation on one path or the other.
2431  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2432  return false;
2433  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2434  return false;
2435 
2436  Builder.SetInsertPoint(BI);
2437  // Okay, we found a branch that is going to two return nodes. If
2438  // there is no return value for this function, just change the
2439  // branch into a return.
2440  if (FalseRet->getNumOperands() == 0) {
2441  TrueSucc->removePredecessor(BI->getParent());
2442  FalseSucc->removePredecessor(BI->getParent());
2443  Builder.CreateRetVoid();
2445  return true;
2446  }
2447 
2448  // Otherwise, figure out what the true and false return values are
2449  // so we can insert a new select instruction.
2450  Value *TrueValue = TrueRet->getReturnValue();
2451  Value *FalseValue = FalseRet->getReturnValue();
2452 
2453  // Unwrap any PHI nodes in the return blocks.
2454  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2455  if (TVPN->getParent() == TrueSucc)
2456  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2457  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2458  if (FVPN->getParent() == FalseSucc)
2459  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2460 
2461  // In order for this transformation to be safe, we must be able to
2462  // unconditionally execute both operands to the return. This is
2463  // normally the case, but we could have a potentially-trapping
2464  // constant expression that prevents this transformation from being
2465  // safe.
2466  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2467  if (TCV->canTrap())
2468  return false;
2469  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2470  if (FCV->canTrap())
2471  return false;
2472 
2473  // Okay, we collected all the mapped values and checked them for sanity, and
2474  // defined to really do this transformation. First, update the CFG.
2475  TrueSucc->removePredecessor(BI->getParent());
2476  FalseSucc->removePredecessor(BI->getParent());
2477 
2478  // Insert select instructions where needed.
2479  Value *BrCond = BI->getCondition();
2480  if (TrueValue) {
2481  // Insert a select if the results differ.
2482  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2483  } else if (isa<UndefValue>(TrueValue)) {
2484  TrueValue = FalseValue;
2485  } else {
2486  TrueValue =
2487  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2488  }
2489  }
2490 
2491  Value *RI =
2492  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2493 
2494  (void)RI;
2495 
2496  LLVM_DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2497  << "\n " << *BI << "NewRet = " << *RI << "TRUEBLOCK: "
2498  << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2499 
2501 
2502  return true;
2503 }
2504 
2505 /// Return true if the given instruction is available
2506 /// in its predecessor block. If yes, the instruction will be removed.
2508  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2509  return false;
2510  for (Instruction &I : *PB) {
2511  Instruction *PBI = &I;
2512  // Check whether Inst and PBI generate the same value.
2513  if (Inst->isIdenticalTo(PBI)) {
2514  Inst->replaceAllUsesWith(PBI);
2515  Inst->eraseFromParent();
2516  return true;
2517  }
2518  }
2519  return false;
2520 }
2521 
2522 /// Return true if either PBI or BI has branch weight available, and store
2523 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2524 /// not have branch weight, use 1:1 as its weight.
2526  uint64_t &PredTrueWeight,
2527  uint64_t &PredFalseWeight,
2528  uint64_t &SuccTrueWeight,
2529  uint64_t &SuccFalseWeight) {
2530  bool PredHasWeights =
2531  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2532  bool SuccHasWeights =
2533  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2534  if (PredHasWeights || SuccHasWeights) {
2535  if (!PredHasWeights)
2536  PredTrueWeight = PredFalseWeight = 1;
2537  if (!SuccHasWeights)
2538  SuccTrueWeight = SuccFalseWeight = 1;
2539  return true;
2540  } else {
2541  return false;
2542  }
2543 }
2544 
2545 /// If this basic block is simple enough, and if a predecessor branches to us
2546 /// and one of our successors, fold the block into the predecessor and use
2547 /// logical operations to pick the right destination.
2548 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2549  BasicBlock *BB = BI->getParent();
2550 
2551  const unsigned PredCount = pred_size(BB);
2552 
2553  Instruction *Cond = nullptr;
2554  if (BI->isConditional())
2555  Cond = dyn_cast<Instruction>(BI->getCondition());
2556  else {
2557  // For unconditional branch, check for a simple CFG pattern, where
2558  // BB has a single predecessor and BB's successor is also its predecessor's
2559  // successor. If such pattern exists, check for CSE between BB and its
2560  // predecessor.
2561  if (BasicBlock *PB = BB->getSinglePredecessor())
2562  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2563  if (PBI->isConditional() &&
2564  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2565  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2566  for (auto I = BB->instructionsWithoutDebug().begin(),
2567  E = BB->instructionsWithoutDebug().end();
2568  I != E;) {
2569  Instruction *Curr = &*I++;
2570  if (isa<CmpInst>(Curr)) {
2571  Cond = Curr;
2572  break;
2573  }
2574  // Quit if we can't remove this instruction.
2575  if (!tryCSEWithPredecessor(Curr, PB))
2576  return false;
2577  }
2578  }
2579 
2580  if (!Cond)
2581  return false;
2582  }
2583 
2584  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2585  Cond->getParent() != BB || !Cond->hasOneUse())
2586  return false;
2587 
2588  // Make sure the instruction after the condition is the cond branch.
2589  BasicBlock::iterator CondIt = ++Cond->getIterator();
2590 
2591  // Ignore dbg intrinsics.
2592  while (isa<DbgInfoIntrinsic>(CondIt))
2593  ++CondIt;
2594 
2595  if (&*CondIt != BI)
2596  return false;
2597 
2598  // Only allow this transformation if computing the condition doesn't involve
2599  // too many instructions and these involved instructions can be executed
2600  // unconditionally. We denote all involved instructions except the condition
2601  // as "bonus instructions", and only allow this transformation when the
2602  // number of the bonus instructions we'll need to create when cloning into
2603  // each predecessor does not exceed a certain threshold.
2604  unsigned NumBonusInsts = 0;
2605  for (auto I = BB->begin(); Cond != &*I; ++I) {
2606  // Ignore dbg intrinsics.
2607  if (isa<DbgInfoIntrinsic>(I))
2608  continue;
2609  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2610  return false;
2611  // I has only one use and can be executed unconditionally.
2613  if (User == nullptr || User->getParent() != BB)
2614  return false;
2615  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2616  // to use any other instruction, User must be an instruction between next(I)
2617  // and Cond.
2618 
2619  // Account for the cost of duplicating this instruction into each
2620  // predecessor.
2621  NumBonusInsts += PredCount;
2622  // Early exits once we reach the limit.
2623  if (NumBonusInsts > BonusInstThreshold)
2624  return false;
2625  }
2626 
2627  // Cond is known to be a compare or binary operator. Check to make sure that
2628  // neither operand is a potentially-trapping constant expression.
2629  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2630  if (CE->canTrap())
2631  return false;
2632  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2633  if (CE->canTrap())
2634  return false;
2635 
2636  // Finally, don't infinitely unroll conditional loops.
2637  BasicBlock *TrueDest = BI->getSuccessor(0);
2638  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2639  if (TrueDest == BB || FalseDest == BB)
2640  return false;
2641 
2642  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2643  BasicBlock *PredBlock = *PI;
2644  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2645 
2646  // Check that we have two conditional branches. If there is a PHI node in
2647  // the common successor, verify that the same value flows in from both
2648  // blocks.
2650  if (!PBI || PBI->isUnconditional() ||
2651  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2652  (!BI->isConditional() &&
2653  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2654  continue;
2655 
2656  // Determine if the two branches share a common destination.
2657  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2658  bool InvertPredCond = false;
2659 
2660  if (BI->isConditional()) {
2661  if (PBI->getSuccessor(0) == TrueDest) {
2662  Opc = Instruction::Or;
2663  } else if (PBI->getSuccessor(1) == FalseDest) {
2664  Opc = Instruction::And;
2665  } else if (PBI->getSuccessor(0) == FalseDest) {
2666  Opc = Instruction::And;
2667  InvertPredCond = true;
2668  } else if (PBI->getSuccessor(1) == TrueDest) {
2669  Opc = Instruction::Or;
2670  InvertPredCond = true;
2671  } else {
2672  continue;
2673  }
2674  } else {
2675  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2676  continue;
2677  }
2678 
2679  LLVM_DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2680  IRBuilder<> Builder(PBI);
2681 
2682  // If we need to invert the condition in the pred block to match, do so now.
2683  if (InvertPredCond) {
2684  Value *NewCond = PBI->getCondition();
2685 
2686  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2687  CmpInst *CI = cast<CmpInst>(NewCond);
2688  CI->setPredicate(CI->getInversePredicate());
2689  } else {
2690  NewCond =
2691  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2692  }
2693 
2694  PBI->setCondition(NewCond);
2695  PBI->swapSuccessors();
2696  }
2697 
2698  // If we have bonus instructions, clone them into the predecessor block.
2699  // Note that there may be multiple predecessor blocks, so we cannot move
2700  // bonus instructions to a predecessor block.
2701  ValueToValueMapTy VMap; // maps original values to cloned values
2702  // We already make sure Cond is the last instruction before BI. Therefore,
2703  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2704  // instructions.
2705  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2706  if (isa<DbgInfoIntrinsic>(BonusInst))
2707  continue;
2708  Instruction *NewBonusInst = BonusInst->clone();
2709  RemapInstruction(NewBonusInst, VMap,
2711  VMap[&*BonusInst] = NewBonusInst;
2712 
2713  // If we moved a load, we cannot any longer claim any knowledge about
2714  // its potential value. The previous information might have been valid
2715  // only given the branch precondition.
2716  // For an analogous reason, we must also drop all the metadata whose
2717  // semantics we don't understand.
2718  NewBonusInst->dropUnknownNonDebugMetadata();
2719 
2720  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2721  NewBonusInst->takeName(&*BonusInst);
2722  BonusInst->setName(BonusInst->getName() + ".old");
2723  }
2724 
2725  // Clone Cond into the predecessor basic block, and or/and the
2726  // two conditions together.
2727  Instruction *CondInPred = Cond->clone();
2728  RemapInstruction(CondInPred, VMap,
2730  PredBlock->getInstList().insert(PBI->getIterator(), CondInPred);
2731  CondInPred->takeName(Cond);
2732  Cond->setName(CondInPred->getName() + ".old");
2733 
2734  if (BI->isConditional()) {
2735  Instruction *NewCond = cast<Instruction>(
2736  Builder.CreateBinOp(Opc, PBI->getCondition(), CondInPred, "or.cond"));
2737  PBI->setCondition(NewCond);
2738 
2739  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2740  bool HasWeights =
2741  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2742  SuccTrueWeight, SuccFalseWeight);
2743  SmallVector<uint64_t, 8> NewWeights;
2744 
2745  if (PBI->getSuccessor(0) == BB) {
2746  if (HasWeights) {
2747  // PBI: br i1 %x, BB, FalseDest
2748  // BI: br i1 %y, TrueDest, FalseDest
2749  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2750  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2751  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2752  // TrueWeight for PBI * FalseWeight for BI.
2753  // We assume that total weights of a BranchInst can fit into 32 bits.
2754  // Therefore, we will not have overflow using 64-bit arithmetic.
2755  NewWeights.push_back(PredFalseWeight *
2756  (SuccFalseWeight + SuccTrueWeight) +
2757  PredTrueWeight * SuccFalseWeight);
2758  }
2759  AddPredecessorToBlock(TrueDest, PredBlock, BB);
2760  PBI->setSuccessor(0, TrueDest);
2761  }
2762  if (PBI->getSuccessor(1) == BB) {
2763  if (HasWeights) {
2764  // PBI: br i1 %x, TrueDest, BB
2765  // BI: br i1 %y, TrueDest, FalseDest
2766  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2767  // FalseWeight for PBI * TrueWeight for BI.
2768  NewWeights.push_back(PredTrueWeight *
2769  (SuccFalseWeight + SuccTrueWeight) +
2770  PredFalseWeight * SuccTrueWeight);
2771  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2772  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2773  }
2774  AddPredecessorToBlock(FalseDest, PredBlock, BB);
2775  PBI->setSuccessor(1, FalseDest);
2776  }
2777  if (NewWeights.size() == 2) {
2778  // Halve the weights if any of them cannot fit in an uint32_t
2779  FitWeights(NewWeights);
2780 
2781  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2782  NewWeights.end());
2783  setBranchWeights(PBI, MDWeights[0], MDWeights[1]);
2784  } else
2785  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2786  } else {
2787  // Update PHI nodes in the common successors.
2788  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2789  ConstantInt *PBI_C = cast<ConstantInt>(
2790  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2791  assert(PBI_C->getType()->isIntegerTy(1));
2792  Instruction *MergedCond = nullptr;
2793  if (PBI->getSuccessor(0) == TrueDest) {
2794  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2795  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2796  // is false: !PBI_Cond and BI_Value
2797  Instruction *NotCond = cast<Instruction>(
2798  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2799  MergedCond = cast<Instruction>(
2800  Builder.CreateBinOp(Instruction::And, NotCond, CondInPred,
2801  "and.cond"));
2802  if (PBI_C->isOne())
2803  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2804  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2805  } else {
2806  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2807  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2808  // is false: PBI_Cond and BI_Value
2809  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2810  Instruction::And, PBI->getCondition(), CondInPred, "and.cond"));
2811  if (PBI_C->isOne()) {
2812  Instruction *NotCond = cast<Instruction>(
2813  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2814  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2815  Instruction::Or, NotCond, MergedCond, "or.cond"));
2816  }
2817  }
2818  // Update PHI Node.
2819  PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2820  MergedCond);
2821  }
2822  // Change PBI from Conditional to Unconditional.
2823  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2825  PBI = New_PBI;
2826  }
2827 
2828  // If BI was a loop latch, it may have had associated loop metadata.
2829  // We need to copy it to the new latch, that is, PBI.
2830  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2831  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2832 
2833  // TODO: If BB is reachable from all paths through PredBlock, then we
2834  // could replace PBI's branch probabilities with BI's.
2835 
2836  // Copy any debug value intrinsics into the end of PredBlock.
2837  for (Instruction &I : *BB)
2838  if (isa<DbgInfoIntrinsic>(I))
2839  I.clone()->insertBefore(PBI);
2840 
2841  return true;
2842  }
2843  return false;
2844 }
2845 
2846 // If there is only one store in BB1 and BB2, return it, otherwise return
2847 // nullptr.
2849  StoreInst *S = nullptr;
2850  for (auto *BB : {BB1, BB2}) {
2851  if (!BB)
2852  continue;
2853  for (auto &I : *BB)
2854  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2855  if (S)
2856  // Multiple stores seen.
2857  return nullptr;
2858  else
2859  S = SI;
2860  }
2861  }
2862  return S;
2863 }
2864 
2866  Value *AlternativeV = nullptr) {
2867  // PHI is going to be a PHI node that allows the value V that is defined in
2868  // BB to be referenced in BB's only successor.
2869  //
2870  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2871  // doesn't matter to us what the other operand is (it'll never get used). We
2872  // could just create a new PHI with an undef incoming value, but that could
2873  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2874  // other PHI. So here we directly look for some PHI in BB's successor with V
2875  // as an incoming operand. If we find one, we use it, else we create a new
2876  // one.
2877  //
2878  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2879  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2880  // where OtherBB is the single other predecessor of BB's only successor.
2881  PHINode *PHI = nullptr;
2882  BasicBlock *Succ = BB->getSingleSuccessor();
2883 
2884  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2885  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2886  PHI = cast<PHINode>(I);
2887  if (!AlternativeV)
2888  break;
2889 
2890  assert(Succ->hasNPredecessors(2));
2891  auto PredI = pred_begin(Succ);
2892  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2893  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2894  break;
2895  PHI = nullptr;
2896  }
2897  if (PHI)
2898  return PHI;
2899 
2900  // If V is not an instruction defined in BB, just return it.
2901  if (!AlternativeV &&
2902  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2903  return V;
2904 
2905  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2906  PHI->addIncoming(V, BB);
2907  for (BasicBlock *PredBB : predecessors(Succ))
2908  if (PredBB != BB)
2909  PHI->addIncoming(
2910  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2911  return PHI;
2912 }
2913 
2915  BasicBlock *QTB, BasicBlock *QFB,
2916  BasicBlock *PostBB, Value *Address,
2917  bool InvertPCond, bool InvertQCond,
2918  const DataLayout &DL) {
2919  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2920  return Operator::getOpcode(&I) == Instruction::BitCast &&
2921  I.getType()->isPointerTy();
2922  };
2923 
2924  // If we're not in aggressive mode, we only optimize if we have some
2925  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2926  auto IsWorthwhile = [&](BasicBlock *BB) {
2927  if (!BB)
2928  return true;
2929  // Heuristic: if the block can be if-converted/phi-folded and the
2930  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2931  // thread this store.
2932  unsigned N = 0;
2933  for (auto &I : BB->instructionsWithoutDebug()) {
2934  // Cheap instructions viable for folding.
2935  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2936  isa<StoreInst>(I))
2937  ++N;
2938  // Free instructions.
2939  else if (I.isTerminator() || IsaBitcastOfPointerType(I))
2940  continue;
2941  else
2942  return false;
2943  }
2944  // The store we want to merge is counted in N, so add 1 to make sure
2945  // we're counting the instructions that would be left.
2946  return N <= (PHINodeFoldingThreshold + 1);
2947  };
2948 
2950  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2951  !IsWorthwhile(QFB)))
2952  return false;
2953 
2954  // For every pointer, there must be exactly two stores, one coming from
2955  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2956  // store (to any address) in PTB,PFB or QTB,QFB.
2957  // FIXME: We could relax this restriction with a bit more work and performance
2958  // testing.
2959  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2960  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2961  if (!PStore || !QStore)
2962  return false;
2963 
2964  // Now check the stores are compatible.
2965  if (!QStore->isUnordered() || !PStore->isUnordered())
2966  return false;
2967 
2968  // Check that sinking the store won't cause program behavior changes. Sinking
2969  // the store out of the Q blocks won't change any behavior as we're sinking
2970  // from a block to its unconditional successor. But we're moving a store from
2971  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2972  // So we need to check that there are no aliasing loads or stores in
2973  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2974  // operations between PStore and the end of its parent block.
2975  //
2976  // The ideal way to do this is to query AliasAnalysis, but we don't
2977  // preserve AA currently so that is dangerous. Be super safe and just
2978  // check there are no other memory operations at all.
2979  for (auto &I : *QFB->getSinglePredecessor())
2980  if (I.mayReadOrWriteMemory())
2981  return false;
2982  for (auto &I : *QFB)
2983  if (&I != QStore && I.mayReadOrWriteMemory())
2984  return false;
2985  if (QTB)
2986  for (auto &I : *QTB)
2987  if (&I != QStore && I.mayReadOrWriteMemory())
2988  return false;
2989  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2990  I != E; ++I)
2991  if (&*I != PStore && I->mayReadOrWriteMemory())
2992  return false;
2993 
2994  // If PostBB has more than two predecessors, we need to split it so we can
2995  // sink the store.
2996  if (std::next(pred_begin(PostBB), 2) != pred_end(PostBB)) {
2997  // We know that QFB's only successor is PostBB. And QFB has a single
2998  // predecessor. If QTB exists, then its only successor is also PostBB.
2999  // If QTB does not exist, then QFB's only predecessor has a conditional
3000  // branch to QFB and PostBB.
3001  BasicBlock *TruePred = QTB ? QTB : QFB->getSinglePredecessor();
3002  BasicBlock *NewBB = SplitBlockPredecessors(PostBB, { QFB, TruePred},
3003  "condstore.split");
3004  if (!NewBB)
3005  return false;
3006  PostBB = NewBB;
3007  }
3008 
3009  // OK, we're going to sink the stores to PostBB. The store has to be
3010  // conditional though, so first create the predicate.
3011  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
3012  ->getCondition();
3013  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
3014  ->getCondition();
3015 
3017  PStore->getParent());
3019  QStore->getParent(), PPHI);
3020 
3021  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
3022 
3023  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
3024  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
3025 
3026  if (InvertPCond)
3027  PPred = QB.CreateNot(PPred);
3028  if (InvertQCond)
3029  QPred = QB.CreateNot(QPred);
3030  Value *CombinedPred = QB.CreateOr(PPred, QPred);
3031 
3032  auto *T =
3033  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
3034  QB.SetInsertPoint(T);
3035  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
3036  AAMDNodes AAMD;
3037  PStore->getAAMetadata(AAMD, /*Merge=*/false);
3038  PStore->getAAMetadata(AAMD, /*Merge=*/true);
3039  SI->setAAMetadata(AAMD);
3040  unsigned PAlignment = PStore->getAlignment();
3041  unsigned QAlignment = QStore->getAlignment();
3042  unsigned TypeAlignment =
3044  unsigned MinAlignment;
3045  unsigned MaxAlignment;
3046  std::tie(MinAlignment, MaxAlignment) = std::minmax(PAlignment, QAlignment);
3047  // Choose the minimum alignment. If we could prove both stores execute, we
3048  // could use biggest one. In this case, though, we only know that one of the
3049  // stores executes. And we don't know it's safe to take the alignment from a
3050  // store that doesn't execute.
3051  if (MinAlignment != 0) {
3052  // Choose the minimum of all non-zero alignments.
3053  SI->setAlignment(MinAlignment);
3054  } else if (MaxAlignment != 0) {
3055  // Choose the minimal alignment between the non-zero alignment and the ABI
3056  // default alignment for the type of the stored value.
3057  SI->setAlignment(std::min(MaxAlignment, TypeAlignment));
3058  } else {
3059  // If both alignments are zero, use ABI default alignment for the type of
3060  // the stored value.
3061  SI->setAlignment(TypeAlignment);
3062  }
3063 
3064  QStore->eraseFromParent();
3065  PStore->eraseFromParent();
3066 
3067  return true;
3068 }
3069 
3071  const DataLayout &DL) {
3072  // The intention here is to find diamonds or triangles (see below) where each
3073  // conditional block contains a store to the same address. Both of these
3074  // stores are conditional, so they can't be unconditionally sunk. But it may
3075  // be profitable to speculatively sink the stores into one merged store at the
3076  // end, and predicate the merged store on the union of the two conditions of
3077  // PBI and QBI.
3078  //
3079  // This can reduce the number of stores executed if both of the conditions are
3080  // true, and can allow the blocks to become small enough to be if-converted.
3081  // This optimization will also chain, so that ladders of test-and-set
3082  // sequences can be if-converted away.
3083  //
3084  // We only deal with simple diamonds or triangles:
3085  //
3086  // PBI or PBI or a combination of the two
3087  // / \ | \
3088  // PTB PFB | PFB
3089  // \ / | /
3090  // QBI QBI
3091  // / \ | \
3092  // QTB QFB | QFB
3093  // \ / | /
3094  // PostBB PostBB
3095  //
3096  // We model triangles as a type of diamond with a nullptr "true" block.
3097  // Triangles are canonicalized so that the fallthrough edge is represented by
3098  // a true condition, as in the diagram above.
3099  BasicBlock *PTB = PBI->getSuccessor(0);
3100  BasicBlock *PFB = PBI->getSuccessor(1);
3101  BasicBlock *QTB = QBI->getSuccessor(0);
3102  BasicBlock *QFB = QBI->getSuccessor(1);
3103  BasicBlock *PostBB = QFB->getSingleSuccessor();
3104 
3105  // Make sure we have a good guess for PostBB. If QTB's only successor is
3106  // QFB, then QFB is a better PostBB.
3107  if (QTB->getSingleSuccessor() == QFB)
3108  PostBB = QFB;
3109 
3110  // If we couldn't find a good PostBB, stop.
3111  if (!PostBB)
3112  return false;
3113 
3114  bool InvertPCond = false, InvertQCond = false;
3115  // Canonicalize fallthroughs to the true branches.
3116  if (PFB == QBI->getParent()) {
3117  std::swap(PFB, PTB);
3118  InvertPCond = true;
3119  }
3120  if (QFB == PostBB) {
3121  std::swap(QFB, QTB);
3122  InvertQCond = true;
3123  }
3124 
3125  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3126  // and QFB may not. Model fallthroughs as a nullptr block.
3127  if (PTB == QBI->getParent())
3128  PTB = nullptr;
3129  if (QTB == PostBB)
3130  QTB = nullptr;
3131 
3132  // Legality bailouts. We must have at least the non-fallthrough blocks and
3133  // the post-dominating block, and the non-fallthroughs must only have one
3134  // predecessor.
3135  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3136  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3137  };
3138  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3139  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3140  return false;
3141  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3142  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3143  return false;
3144  if (!QBI->getParent()->hasNUses(2))
3145  return false;
3146 
3147  // OK, this is a sequence of two diamonds or triangles.
3148  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3149  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3150  for (auto *BB : {PTB, PFB}) {
3151  if (!BB)
3152  continue;
3153  for (auto &I : *BB)
3154  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3155  PStoreAddresses.insert(SI->getPointerOperand());
3156  }
3157  for (auto *BB : {QTB, QFB}) {
3158  if (!BB)
3159  continue;
3160  for (auto &I : *BB)
3161  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3162  QStoreAddresses.insert(SI->getPointerOperand());
3163  }
3164 
3165  set_intersect(PStoreAddresses, QStoreAddresses);
3166  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3167  // clear what it contains.
3168  auto &CommonAddresses = PStoreAddresses;
3169 
3170  bool Changed = false;
3171  for (auto *Address : CommonAddresses)
3172  Changed |= mergeConditionalStoreToAddress(
3173  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond, DL);
3174  return Changed;
3175 }
3176 
3177 /// If we have a conditional branch as a predecessor of another block,
3178 /// this function tries to simplify it. We know
3179 /// that PBI and BI are both conditional branches, and BI is in one of the
3180 /// successor blocks of PBI - PBI branches to BI.
3182  const DataLayout &DL) {
3183  assert(PBI->isConditional() && BI->isConditional());
3184  BasicBlock *BB = BI->getParent();
3185 
3186  // If this block ends with a branch instruction, and if there is a
3187  // predecessor that ends on a branch of the same condition, make
3188  // this conditional branch redundant.
3189  if (PBI->getCondition() == BI->getCondition() &&
3190  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3191  // Okay, the outcome of this conditional branch is statically
3192  // knowable. If this block had a single pred, handle specially.
3193  if (BB->getSinglePredecessor()) {
3194  // Turn this into a branch on constant.
3195  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3196  BI->setCondition(
3197  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3198  return true; // Nuke the branch on constant.
3199  }
3200 
3201  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3202  // in the constant and simplify the block result. Subsequent passes of
3203  // simplifycfg will thread the block.
3205  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3206  PHINode *NewPN = PHINode::Create(
3207  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3208  BI->getCondition()->getName() + ".pr", &BB->front());
3209  // Okay, we're going to insert the PHI node. Since PBI is not the only
3210  // predecessor, compute the PHI'd conditional value for all of the preds.
3211  // Any predecessor where the condition is not computable we keep symbolic.
3212  for (pred_iterator PI = PB; PI != PE; ++PI) {
3213  BasicBlock *P = *PI;
3214  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3215  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3216  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3217  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3218  NewPN->addIncoming(
3219  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3220  P);
3221  } else {
3222  NewPN->addIncoming(BI->getCondition(), P);
3223  }
3224  }
3225 
3226  BI->setCondition(NewPN);
3227  return true;
3228  }
3229  }
3230 
3231  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3232  if (CE->canTrap())
3233  return false;
3234 
3235  // If both branches are conditional and both contain stores to the same
3236  // address, remove the stores from the conditionals and create a conditional
3237  // merged store at the end.
3238  if (MergeCondStores && mergeConditionalStores(PBI, BI, DL))
3239  return true;
3240 
3241  // If this is a conditional branch in an empty block, and if any
3242  // predecessors are a conditional branch to one of our destinations,
3243  // fold the conditions into logical ops and one cond br.
3244 
3245  // Ignore dbg intrinsics.
3246  if (&*BB->instructionsWithoutDebug().begin() != BI)
3247  return false;
3248 
3249  int PBIOp, BIOp;
3250  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3251  PBIOp = 0;
3252  BIOp = 0;
3253  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3254  PBIOp = 0;
3255  BIOp = 1;
3256  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3257  PBIOp = 1;
3258  BIOp = 0;
3259  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3260  PBIOp = 1;
3261  BIOp = 1;
3262  } else {
3263  return false;
3264  }
3265 
3266  // Check to make sure that the other destination of this branch
3267  // isn't BB itself. If so, this is an infinite loop that will
3268  // keep getting unwound.
3269  if (PBI->getSuccessor(PBIOp) == BB)
3270  return false;
3271 
3272  // Do not perform this transformation if it would require
3273  // insertion of a large number of select instructions. For targets
3274  // without predication/cmovs, this is a big pessimization.
3275 
3276  // Also do not perform this transformation if any phi node in the common
3277  // destination block can trap when reached by BB or PBB (PR17073). In that
3278  // case, it would be unsafe to hoist the operation into a select instruction.
3279 
3280  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3281  unsigned NumPhis = 0;
3282  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3283  ++II, ++NumPhis) {
3284  if (NumPhis > 2) // Disable this xform.
3285  return false;
3286 
3287  PHINode *PN = cast<PHINode>(II);
3288  Value *BIV = PN->getIncomingValueForBlock(BB);
3289  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3290  if (CE->canTrap())
3291  return false;
3292 
3293  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3294  Value *PBIV = PN->getIncomingValue(PBBIdx);
3295  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3296  if (CE->canTrap())
3297  return false;
3298  }
3299 
3300  // Finally, if everything is ok, fold the branches to logical ops.
3301  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3302 
3303  LLVM_DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3304  << "AND: " << *BI->getParent());
3305 
3306  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3307  // branch in it, where one edge (OtherDest) goes back to itself but the other
3308  // exits. We don't *know* that the program avoids the infinite loop
3309  // (even though that seems likely). If we do this xform naively, we'll end up
3310  // recursively unpeeling the loop. Since we know that (after the xform is
3311  // done) that the block *is* infinite if reached, we just make it an obviously
3312  // infinite loop with no cond branch.
3313  if (OtherDest == BB) {
3314  // Insert it at the end of the function, because it's either code,
3315  // or it won't matter if it's hot. :)
3316  BasicBlock *InfLoopBlock =
3317  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3318  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3319  OtherDest = InfLoopBlock;
3320  }
3321 
3322  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3323 
3324  // BI may have other predecessors. Because of this, we leave
3325  // it alone, but modify PBI.
3326 
3327  // Make sure we get to CommonDest on True&True directions.
3328  Value *PBICond = PBI->getCondition();
3329  IRBuilder<NoFolder> Builder(PBI);
3330  if (PBIOp)
3331  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3332 
3333  Value *BICond = BI->getCondition();
3334  if (BIOp)
3335  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3336 
3337  // Merge the conditions.
3338  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3339 
3340  // Modify PBI to branch on the new condition to the new dests.
3341  PBI->setCondition(Cond);
3342  PBI->setSuccessor(0, CommonDest);
3343  PBI->setSuccessor(1, OtherDest);
3344 
3345  // Update branch weight for PBI.
3346  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3347  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3348  bool HasWeights =
3349  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3350  SuccTrueWeight, SuccFalseWeight);
3351  if (HasWeights) {
3352  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3353  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3354  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3355  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3356  // The weight to CommonDest should be PredCommon * SuccTotal +
3357  // PredOther * SuccCommon.
3358  // The weight to OtherDest should be PredOther * SuccOther.
3359  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3360  PredOther * SuccCommon,
3361  PredOther * SuccOther};
3362  // Halve the weights if any of them cannot fit in an uint32_t
3363  FitWeights(NewWeights);
3364 
3365  setBranchWeights(PBI, NewWeights[0], NewWeights[1]);
3366  }
3367 
3368  // OtherDest may have phi nodes. If so, add an entry from PBI's
3369  // block that are identical to the entries for BI's block.
3370  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3371 
3372  // We know that the CommonDest already had an edge from PBI to
3373  // it. If it has PHIs though, the PHIs may have different
3374  // entries for BB and PBI's BB. If so, insert a select to make
3375  // them agree.
3376  for (PHINode &PN : CommonDest->phis()) {
3377  Value *BIV = PN.getIncomingValueForBlock(BB);
3378  unsigned PBBIdx = PN.getBasicBlockIndex(PBI->getParent());
3379  Value *PBIV = PN.getIncomingValue(PBBIdx);
3380  if (BIV != PBIV) {
3381  // Insert a select in PBI to pick the right value.
3382  SelectInst *NV = cast<SelectInst>(
3383  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3384  PN.setIncomingValue(PBBIdx, NV);
3385  // Although the select has the same condition as PBI, the original branch
3386  // weights for PBI do not apply to the new select because the select's
3387  // 'logical' edges are incoming edges of the phi that is eliminated, not
3388  // the outgoing edges of PBI.
3389  if (HasWeights) {
3390  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3391  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3392  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3393  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3394  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3395  // The weight to PredOtherDest should be PredOther * SuccCommon.
3396  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3397  PredOther * SuccCommon};
3398 
3399  FitWeights(NewWeights);
3400 
3401  setBranchWeights(NV, NewWeights[0], NewWeights[1]);
3402  }
3403  }
3404  }
3405 
3406  LLVM_DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3407  LLVM_DEBUG(dbgs() << *PBI->getParent()->getParent());
3408 
3409  // This basic block is probably dead. We know it has at least
3410  // one fewer predecessor.
3411  return true;
3412 }
3413 
3414 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3415 // true or to FalseBB if Cond is false.
3416 // Takes care of updating the successors and removing the old terminator.
3417 // Also makes sure not to introduce new successors by assuming that edges to
3418 // non-successor TrueBBs and FalseBBs aren't reachable.
3419 static bool SimplifyTerminatorOnSelect(Instruction *OldTerm, Value *Cond,
3420  BasicBlock *TrueBB, BasicBlock *FalseBB,
3421  uint32_t TrueWeight,
3422  uint32_t FalseWeight) {
3423  // Remove any superfluous successor edges from the CFG.
3424  // First, figure out which successors to preserve.
3425  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3426  // successor.
3427  BasicBlock *KeepEdge1 = TrueBB;
3428  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3429 
3430  // Then remove the rest.
3431  for (BasicBlock *Succ : successors(OldTerm)) {
3432  // Make sure only to keep exactly one copy of each edge.
3433  if (Succ == KeepEdge1)
3434  KeepEdge1 = nullptr;
3435  else if (Succ == KeepEdge2)
3436  KeepEdge2 = nullptr;
3437  else
3438  Succ->removePredecessor(OldTerm->getParent(),
3439  /*KeepOneInputPHIs=*/true);
3440  }
3441 
3442  IRBuilder<> Builder(OldTerm);
3443  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3444 
3445  // Insert an appropriate new terminator.
3446  if (!KeepEdge1 && !KeepEdge2) {
3447  if (TrueBB == FalseBB)
3448  // We were only looking for one successor, and it was present.
3449  // Create an unconditional branch to it.
3450  Builder.CreateBr(TrueBB);
3451  else {
3452  // We found both of the successors we were looking for.
3453  // Create a conditional branch sharing the condition of the select.
3454  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3455  if (TrueWeight != FalseWeight)
3456  setBranchWeights(NewBI, TrueWeight, FalseWeight);
3457  }
3458  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3459  // Neither of the selected blocks were successors, so this
3460  // terminator must be unreachable.
3461  new UnreachableInst(OldTerm->getContext(), OldTerm);
3462  } else {
3463  // One of the selected values was a successor, but the other wasn't.
3464  // Insert an unconditional branch to the one that was found;
3465  // the edge to the one that wasn't must be unreachable.
3466  if (!KeepEdge1)
3467  // Only TrueBB was found.
3468  Builder.CreateBr(TrueBB);
3469  else
3470  // Only FalseBB was found.
3471  Builder.CreateBr(FalseBB);
3472  }
3473 
3474  EraseTerminatorAndDCECond(OldTerm);
3475  return true;
3476 }
3477 
3478 // Replaces
3479 // (switch (select cond, X, Y)) on constant X, Y
3480 // with a branch - conditional if X and Y lead to distinct BBs,
3481 // unconditional otherwise.
3483  // Check for constant integer values in the select.
3484  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3485  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3486  if (!TrueVal || !FalseVal)
3487  return false;
3488 
3489  // Find the relevant condition and destinations.
3490  Value *Condition = Select->getCondition();
3491  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3492  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3493 
3494  // Get weight for TrueBB and FalseBB.
3495  uint32_t TrueWeight = 0, FalseWeight = 0;
3496  SmallVector<uint64_t, 8> Weights;
3497  bool HasWeights = HasBranchWeights(SI);
3498  if (HasWeights) {
3499  GetBranchWeights(SI, Weights);
3500  if (Weights.size() == 1 + SI->getNumCases()) {
3501  TrueWeight =
3502  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3503  FalseWeight =
3504  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3505  }
3506  }
3507 
3508  // Perform the actual simplification.
3509  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3510  FalseWeight);
3511 }
3512 
3513 // Replaces
3514 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3515 // blockaddress(@fn, BlockB)))
3516 // with
3517 // (br cond, BlockA, BlockB).
3519  // Check that both operands of the select are block addresses.
3522  if (!TBA || !FBA)
3523  return false;
3524 
3525  // Extract the actual blocks.
3526  BasicBlock *TrueBB = TBA->getBasicBlock();
3527  BasicBlock *FalseBB = FBA->getBasicBlock();
3528 
3529  // Perform the actual simplification.
3530  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3531  0);
3532 }
3533 
3534 /// This is called when we find an icmp instruction
3535 /// (a seteq/setne with a constant) as the only instruction in a
3536 /// block that ends with an uncond branch. We are looking for a very specific
3537 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3538 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3539 /// default value goes to an uncond block with a seteq in it, we get something
3540 /// like:
3541 ///
3542 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3543 /// DEFAULT:
3544 /// %tmp = icmp eq i8 %A, 92
3545 /// br label %end
3546 /// end:
3547 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3548 ///
3549 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3550 /// the PHI, merging the third icmp into the switch.
3551 bool SimplifyCFGOpt::tryToSimplifyUncondBranchWithICmpInIt(
3552  ICmpInst *ICI, IRBuilder<> &Builder) {
3553  BasicBlock *BB = ICI->getParent();
3554 
3555  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3556  // complex.
3557  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3558  return false;
3559 
3560  Value *V = ICI->getOperand(0);
3561  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3562 
3563  // The pattern we're looking for is where our only predecessor is a switch on
3564  // 'V' and this block is the default case for the switch. In this case we can
3565  // fold the compared value into the switch to simplify things.
3566  BasicBlock *Pred = BB->getSinglePredecessor();
3567  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3568  return false;
3569 
3570  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3571  if (SI->getCondition() != V)
3572  return false;
3573 
3574  // If BB is reachable on a non-default case, then we simply know the value of
3575  // V in this block. Substitute it and constant fold the icmp instruction
3576  // away.
3577  if (SI->getDefaultDest() != BB) {
3578  ConstantInt *VVal = SI->findCaseDest(BB);
3579  assert(VVal && "Should have a unique destination value");
3580  ICI->setOperand(0, VVal);
3581 
3582  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3583  ICI->replaceAllUsesWith(V);
3584  ICI->eraseFromParent();
3585  }
3586  // BB is now empty, so it is likely to simplify away.
3587  return requestResimplify();
3588  }
3589 
3590  // Ok, the block is reachable from the default dest. If the constant we're
3591  // comparing exists in one of the other edges, then we can constant fold ICI
3592  // and zap it.
3593  if (SI->findCaseValue(Cst) != SI->case_default()) {
3594  Value *V;
3595  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3596  V = ConstantInt::getFalse(BB->getContext());
3597  else
3598  V = ConstantInt::getTrue(BB->getContext());
3599 
3600  ICI->replaceAllUsesWith(V);
3601  ICI->eraseFromParent();
3602  // BB is now empty, so it is likely to simplify away.
3603  return requestResimplify();
3604  }
3605 
3606  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3607  // the block.
3608  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3609  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3610  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3611  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3612  return false;
3613 
3614  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3615  // true in the PHI.
3616  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3617  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3618 
3619  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3620  std::swap(DefaultCst, NewCst);
3621 
3622  // Replace ICI (which is used by the PHI for the default value) with true or
3623  // false depending on if it is EQ or NE.
3624  ICI->replaceAllUsesWith(DefaultCst);
3625  ICI->eraseFromParent();
3626 
3627  // Okay, the switch goes to this block on a default value. Add an edge from
3628  // the switch to the merge point on the compared value.
3629  BasicBlock *NewBB =
3630  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3631  SmallVector<uint64_t, 8> Weights;
3632  bool HasWeights = HasBranchWeights(SI);
3633  if (HasWeights) {
3634  GetBranchWeights(SI, Weights);
3635  if (Weights.size() == 1 + SI->getNumCases()) {
3636  // Split weight for default case to case for "Cst".
3637  Weights[0] = (Weights[0] + 1) >> 1;
3638  Weights.push_back(Weights[0]);
3639 
3640  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3641  setBranchWeights(SI, MDWeights);
3642  }
3643  }
3644  SI->addCase(Cst, NewBB);
3645 
3646  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3647  Builder.SetInsertPoint(NewBB);
3648  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3649  Builder.CreateBr(SuccBlock);
3650  PHIUse->addIncoming(NewCst, NewBB);
3651  return true;
3652 }
3653 
3654 /// The specified branch is a conditional branch.
3655 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3656 /// fold it into a switch instruction if so.
3658  const DataLayout &DL) {
3659  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3660  if (!Cond)
3661  return false;
3662 
3663  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3664  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3665  // 'setne's and'ed together, collect them.
3666 
3667  // Try to gather values from a chain of and/or to be turned into a switch
3668  ConstantComparesGatherer ConstantCompare(Cond, DL);
3669  // Unpack the result
3670  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3671  Value *CompVal = ConstantCompare.CompValue;
3672  unsigned UsedICmps = ConstantCompare.UsedICmps;
3673  Value *ExtraCase = ConstantCompare.Extra;
3674 
3675  // If we didn't have a multiply compared value, fail.
3676  if (!CompVal)
3677  return false;
3678 
3679  // Avoid turning single icmps into a switch.
3680  if (UsedICmps <= 1)
3681  return false;
3682 
3683  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3684 
3685  // There might be duplicate constants in the list, which the switch
3686  // instruction can't handle, remove them now.
3687  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3688  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3689 
3690  // If Extra was used, we require at least two switch values to do the
3691  // transformation. A switch with one value is just a conditional branch.
3692  if (ExtraCase && Values.size() < 2)
3693  return false;
3694 
3695  // TODO: Preserve branch weight metadata, similarly to how
3696  // FoldValueComparisonIntoPredecessors preserves it.
3697 
3698  // Figure out which block is which destination.
3699  BasicBlock *DefaultBB = BI->getSuccessor(1);
3700  BasicBlock *EdgeBB = BI->getSuccessor(0);
3701  if (!TrueWhenEqual)
3702  std::swap(DefaultBB, EdgeBB);
3703 
3704  BasicBlock *BB = BI->getParent();
3705 
3706  LLVM_DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3707  << " cases into SWITCH. BB is:\n"
3708  << *BB);
3709 
3710  // If there are any extra values that couldn't be folded into the switch
3711  // then we evaluate them with an explicit branch first. Split the block
3712  // right before the condbr to handle it.
3713  if (ExtraCase) {
3714  BasicBlock *NewBB =
3715  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3716  // Remove the uncond branch added to the old block.
3717  Instruction *OldTI = BB->getTerminator();
3718  Builder.SetInsertPoint(OldTI);
3719 
3720  if (TrueWhenEqual)
3721  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3722  else
3723  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3724 
3725  OldTI->eraseFromParent();
3726 
3727  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3728  // for the edge we just added.
3729  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3730 
3731  LLVM_DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3732  << "\nEXTRABB = " << *BB);
3733  BB = NewBB;
3734  }
3735 
3736  Builder.SetInsertPoint(BI);
3737  // Convert pointer to int before we switch.
3738  if (CompVal->getType()->isPointerTy()) {
3739  CompVal = Builder.CreatePtrToInt(
3740  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3741  }
3742 
3743  // Create the new switch instruction now.
3744  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3745 
3746  // Add all of the 'cases' to the switch instruction.
3747  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3748  New->addCase(Values[i], EdgeBB);
3749 
3750  // We added edges from PI to the EdgeBB. As such, if there were any
3751  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3752  // the number of edges added.
3753  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3754  PHINode *PN = cast<PHINode>(BBI);
3755  Value *InVal = PN->getIncomingValueForBlock(BB);
3756  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3757  PN->addIncoming(InVal, BB);
3758  }
3759 
3760  // Erase the old branch instruction.
3762 
3763  LLVM_DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3764  return true;
3765 }
3766 
3767 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3768  if (isa<PHINode>(RI->getValue()))
3769  return SimplifyCommonResume(RI);
3770  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3771  RI->getValue() == RI->getParent()->getFirstNonPHI())
3772  // The resume must unwind the exception that caused control to branch here.
3773  return SimplifySingleResume(RI);
3774 
3775  return false;
3776 }
3777 
3778 // Simplify resume that is shared by several landing pads (phi of landing pad).
3779 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3780  BasicBlock *BB = RI->getParent();
3781 
3782  // Check that there are no other instructions except for debug intrinsics
3783  // between the phi of landing pads (RI->getValue()) and resume instruction.
3784  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3785  E = RI->getIterator();
3786  while (++I != E)
3787  if (!isa<DbgInfoIntrinsic>(I))
3788  return false;
3789 
3790  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3791  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3792 
3793  // Check incoming blocks to see if any of them are trivial.
3794  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3795  Idx++) {
3796  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3797  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3798 
3799  // If the block has other successors, we can not delete it because
3800  // it has other dependents.
3801  if (IncomingBB->getUniqueSuccessor() != BB)
3802  continue;
3803 
3804  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3805  // Not the landing pad that caused the control to branch here.
3806  if (IncomingValue != LandingPad)
3807  continue;
3808 
3809  bool isTrivial = true;
3810 
3811  I = IncomingBB->getFirstNonPHI()->getIterator();
3812  E = IncomingBB->getTerminator()->getIterator();
3813  while (++I != E)
3814  if (!isa<DbgInfoIntrinsic>(I)) {
3815  isTrivial = false;
3816  break;
3817  }
3818 
3819  if (isTrivial)
3820  TrivialUnwindBlocks.insert(IncomingBB);
3821  }
3822 
3823  // If no trivial unwind blocks, don't do any simplifications.
3824  if (TrivialUnwindBlocks.empty())
3825  return false;
3826 
3827  // Turn all invokes that unwind here into calls.
3828  for (auto *TrivialBB : TrivialUnwindBlocks) {
3829  // Blocks that will be simplified should be removed from the phi node.
3830  // Note there could be multiple edges to the resume block, and we need
3831  // to remove them all.
3832  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3833  BB->removePredecessor(TrivialBB, true);
3834 
3835  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3836  PI != PE;) {
3837  BasicBlock *Pred = *PI++;
3838  removeUnwindEdge(Pred);
3839  }
3840 
3841  // In each SimplifyCFG run, only the current processed block can be erased.
3842  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3843  // of erasing TrivialBB, we only remove the branch to the common resume
3844  // block so that we can later erase the resume block since it has no
3845  // predecessors.
3846  TrivialBB->getTerminator()->eraseFromParent();
3847  new UnreachableInst(RI->getContext(), TrivialBB);
3848  }
3849 
3850  // Delete the resume block if all its predecessors have been removed.
3851  if (pred_empty(BB))
3852  BB->eraseFromParent();
3853 
3854  return !TrivialUnwindBlocks.empty();
3855 }
3856 
3857 // Simplify resume that is only used by a single (non-phi) landing pad.
3858 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3859  BasicBlock *BB = RI->getParent();
3861  assert(RI->getValue() == LPInst &&
3862  "Resume must unwind the exception that caused control to here");
3863 
3864  // Check that there are no other instructions except for debug intrinsics.
3865  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3866  while (++I != E)
3867  if (!isa<DbgInfoIntrinsic>(I))
3868  return false;
3869 
3870  // Turn all invokes that unwind here into calls and delete the basic block.
3871  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3872  BasicBlock *Pred = *PI++;
3873  removeUnwindEdge(Pred);
3874  }
3875 
3876  // The landingpad is now unreachable. Zap it.
3877  if (LoopHeaders)
3878  LoopHeaders->erase(BB);
3879  BB->eraseFromParent();
3880  return true;
3881 }
3882 
3884  // If this is a trivial cleanup pad that executes no instructions, it can be
3885  // eliminated. If the cleanup pad continues to the caller, any predecessor
3886  // that is an EH pad will be updated to continue to the caller and any
3887  // predecessor that terminates with an invoke instruction will have its invoke
3888  // instruction converted to a call instruction. If the cleanup pad being
3889  // simplified does not continue to the caller, each predecessor will be
3890  // updated to continue to the unwind destination of the cleanup pad being
3891  // simplified.
3892  BasicBlock *BB = RI->getParent();
3893  CleanupPadInst *CPInst = RI->getCleanupPad();
3894  if (CPInst->getParent() != BB)
3895  // This isn't an empty cleanup.
3896  return false;
3897 
3898  // We cannot kill the pad if it has multiple uses. This typically arises
3899  // from unreachable basic blocks.
3900  if (!CPInst->hasOneUse())
3901  return false;
3902 
3903  // Check that there are no other instructions except for benign intrinsics.
3904  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3905  while (++I != E) {
3906  auto *II = dyn_cast<IntrinsicInst>(I);
3907  if (!II)
3908  return false;
3909 
3910  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3911  switch (IntrinsicID) {
3912  case Intrinsic::dbg_declare:
3913  case Intrinsic::dbg_value:
3914  case Intrinsic::dbg_label:
3915  case Intrinsic::lifetime_end:
3916  break;
3917  default:
3918  return false;
3919  }
3920  }
3921 
3922  // If the cleanup return we are simplifying unwinds to the caller, this will
3923  // set UnwindDest to nullptr.
3924  BasicBlock *UnwindDest = RI->getUnwindDest();
3925  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3926 
3927  // We're about to remove BB from the control flow. Before we do, sink any
3928  // PHINodes into the unwind destination. Doing this before changing the
3929  // control flow avoids some potentially slow checks, since we can currently
3930  // be certain that UnwindDest and BB have no common predecessors (since they
3931  // are both EH pads).
3932  if (UnwindDest) {
3933  // First, go through the PHI nodes in UnwindDest and update any nodes that
3934  // reference the block we are removing
3935  for (BasicBlock::iterator I = UnwindDest->begin(),
3936  IE = DestEHPad->getIterator();
3937  I != IE; ++I) {
3938  PHINode *DestPN = cast<PHINode>(I);
3939 
3940  int Idx = DestPN->getBasicBlockIndex(BB);
3941  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3942  assert(Idx != -1);
3943  // This PHI node has an incoming value that corresponds to a control
3944  // path through the cleanup pad we are removing. If the incoming
3945  // value is in the cleanup pad, it must be a PHINode (because we
3946  // verified above that the block is otherwise empty). Otherwise, the
3947  // value is either a constant or a value that dominates the cleanup
3948  // pad being removed.
3949  //
3950  // Because BB and UnwindDest are both EH pads, all of their
3951  // predecessors must unwind to these blocks, and since no instruction
3952  // can have multiple unwind destinations, there will be no overlap in
3953  // incoming blocks between SrcPN and DestPN.
3954  Value *SrcVal = DestPN->getIncomingValue(Idx);
3955  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3956 
3957  // Remove the entry for the block we are deleting.
3958  DestPN->removeIncomingValue(Idx, false);
3959 
3960  if (SrcPN && SrcPN->getParent() == BB) {
3961  // If the incoming value was a PHI node in the cleanup pad we are
3962  // removing, we need to merge that PHI node's incoming values into
3963  // DestPN.
3964  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3965  SrcIdx != SrcE; ++SrcIdx) {
3966  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3967  SrcPN->getIncomingBlock(SrcIdx));
3968  }
3969  } else {
3970  // Otherwise, the incoming value came from above BB and
3971  // so we can just reuse it. We must associate all of BB's
3972  // predecessors with this value.
3973  for (auto *pred : predecessors(BB)) {
3974  DestPN->addIncoming(SrcVal, pred);
3975  }
3976  }
3977  }
3978 
3979  // Sink any remaining PHI nodes directly into UnwindDest.
3980  Instruction *InsertPt = DestEHPad;
3981  for (BasicBlock::iterator I = BB->begin(),
3982  IE = BB->getFirstNonPHI()->getIterator();
3983  I != IE;) {
3984  // The iterator must be incremented here because the instructions are
3985  // being moved to another block.
3986  PHINode *PN = cast<PHINode>(I++);
3987  if (PN->use_empty())
3988  // If the PHI node has no uses, just leave it. It will be erased
3989  // when we erase BB below.
3990  continue;
3991 
3992  // Otherwise, sink this PHI node into UnwindDest.
3993  // Any predecessors to UnwindDest which are not already represented
3994  // must be back edges which inherit the value from the path through
3995  // BB. In this case, the PHI value must reference itself.
3996  for (auto *pred : predecessors(UnwindDest))
3997  if (pred != BB)
3998  PN->addIncoming(PN, pred);
3999  PN->moveBefore(InsertPt);
4000  }
4001  }
4002 
4003  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
4004  // The iterator must be updated here because we are removing this pred.
4005  BasicBlock *PredBB = *PI++;
4006  if (UnwindDest == nullptr) {
4007  removeUnwindEdge(PredBB);
4008  } else {
4009  Instruction *TI = PredBB->getTerminator();
4010  TI->replaceUsesOfWith(BB, UnwindDest);
4011  }
4012  }
4013 
4014  // The cleanup pad is now unreachable. Zap it.
4015  BB->eraseFromParent();
4016  return true;
4017 }
4018 
4019 // Try to merge two cleanuppads together.
4021  // Skip any cleanuprets which unwind to caller, there is nothing to merge
4022  // with.
4023  BasicBlock *UnwindDest = RI->getUnwindDest();
4024  if (!UnwindDest)
4025  return false;
4026 
4027  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
4028  // be safe to merge without code duplication.
4029  if (UnwindDest->getSinglePredecessor() != RI->getParent())
4030  return false;
4031 
4032  // Verify that our cleanuppad's unwind destination is another cleanuppad.
4033  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
4034  if (!SuccessorCleanupPad)
4035  return false;
4036 
4037  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
4038  // Replace any uses of the successor cleanupad with the predecessor pad
4039  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
4040  // funclet bundle operands.
4041  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
4042  // Remove the old cleanuppad.
4043  SuccessorCleanupPad->eraseFromParent();
4044  // Now, we simply replace the cleanupret with a branch to the unwind
4045  // destination.
4046  BranchInst::Create(UnwindDest, RI->getParent());
4047  RI->eraseFromParent();
4048 
4049  return true;
4050 }
4051 
4052 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
4053  // It is possible to transiantly have an undef cleanuppad operand because we
4054  // have deleted some, but not all, dead blocks.
4055  // Eventually, this block will be deleted.
4056  if (isa<UndefValue>(RI->getOperand(0)))
4057  return false;
4058 
4059  if (mergeCleanupPad(RI))
4060  return true;
4061 
4062  if (removeEmptyCleanup(RI))
4063  return true;
4064 
4065  return false;
4066 }
4067 
4068 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
4069  BasicBlock *BB = RI->getParent();
4070  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
4071  return false;
4072 
4073  // Find predecessors that end with branches.
4074  SmallVector<BasicBlock *, 8> UncondBranchPreds;
4075  SmallVector<BranchInst *, 8> CondBranchPreds;
4076  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
4077  BasicBlock *P = *PI;
4078  Instruction *PTI = P->getTerminator();
4079  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
4080  if (BI->isUnconditional())
4081  UncondBranchPreds.push_back(P);
4082  else
4083  CondBranchPreds.push_back(BI);
4084  }
4085  }
4086 
4087  // If we found some, do the transformation!
4088  if (!UncondBranchPreds.empty() && DupRet) {
4089  while (!UncondBranchPreds.empty()) {
4090  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4091  LLVM_DEBUG(dbgs() << "FOLDING: " << *BB
4092  << "INTO UNCOND BRANCH PRED: " << *Pred);
4093  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4094  }
4095 
4096  // If we eliminated all predecessors of the block, delete the block now.
4097  if (pred_empty(BB)) {
4098  // We know there are no successors, so just nuke the block.
4099  if (LoopHeaders)
4100  LoopHeaders->erase(BB);
4101  BB->eraseFromParent();
4102  }
4103 
4104  return true;
4105  }
4106 
4107  // Check out all of the conditional branches going to this return
4108  // instruction. If any of them just select between returns, change the
4109  // branch itself into a select/return pair.
4110  while (!CondBranchPreds.empty()) {
4111  BranchInst *BI = CondBranchPreds.pop_back_val();
4112 
4113  // Check to see if the non-BB successor is also a return block.
4114  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4115  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4116  SimplifyCondBranchToTwoReturns(BI, Builder))
4117  return true;
4118  }
4119  return false;
4120 }
4121 
4122 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4123  BasicBlock *BB = UI->getParent();
4124 
4125  bool Changed = false;
4126 
4127  // If there are any instructions immediately before the unreachable that can
4128  // be removed, do so.
4129  while (UI->getIterator() != BB->begin()) {
4130  BasicBlock::iterator BBI = UI->getIterator();
4131  --BBI;
4132  // Do not delete instructions that can have side effects which might cause
4133  // the unreachable to not be reachable; specifically, calls and volatile
4134  // operations may have this effect.
4135  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4136  break;
4137 
4138  if (BBI->mayHaveSideEffects()) {
4139  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4140  if (SI->isVolatile())
4141  break;
4142  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4143  if (LI->isVolatile())
4144  break;
4145  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4146  if (RMWI->isVolatile())
4147  break;
4148  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4149  if (CXI->isVolatile())
4150  break;
4151  } else if (isa<CatchPadInst>(BBI)) {
4152  // A catchpad may invoke exception object constructors and such, which
4153  // in some languages can be arbitrary code, so be conservative by
4154  // default.
4155  // For CoreCLR, it just involves a type test, so can be removed.
4158  break;
4159  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4160  !isa<LandingPadInst>(BBI)) {
4161  break;
4162  }
4163  // Note that deleting LandingPad's here is in fact okay, although it
4164  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4165  // all the predecessors of this block will be the unwind edges of Invokes,
4166  // and we can therefore guarantee this block will be erased.
4167  }
4168 
4169  // Delete this instruction (any uses are guaranteed to be dead)
4170  if (!BBI->use_empty())
4171  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4172  BBI->eraseFromParent();
4173  Changed = true;
4174  }
4175 
4176  // If the unreachable instruction is the first in the block, take a gander
4177  // at all of the predecessors of this instruction, and simplify them.
4178  if (&BB->front() != UI)
4179  return Changed;
4180 
4182  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4183  Instruction *TI = Preds[i]->getTerminator();
4184  IRBuilder<> Builder(TI);
4185  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4186  if (BI->isUnconditional()) {
4187  if (BI->getSuccessor(0) == BB) {
4188  new UnreachableInst(TI->getContext(), TI);
4189  TI->eraseFromParent();
4190  Changed = true;
4191  }
4192  } else {
4193  if (BI->getSuccessor(0) == BB) {
4194  Builder.CreateBr(BI->getSuccessor(1));
4196  } else if (BI->getSuccessor(1) == BB) {
4197  Builder.CreateBr(BI->getSuccessor(0));
4199  Changed = true;
4200  }
4201  }
4202  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4203  for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4204  if (i->getCaseSuccessor() != BB) {
4205  ++i;
4206  continue;
4207  }
4208  BB->removePredecessor(SI->getParent());
4209  i = SI->removeCase(i);
4210  e = SI->case_end();
4211  Changed = true;
4212  }
4213  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4214  if (II->getUnwindDest() == BB) {
4215  removeUnwindEdge(TI->getParent());
4216  Changed = true;
4217  }
4218  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4219  if (CSI->getUnwindDest() == BB) {
4220  removeUnwindEdge(TI->getParent());
4221  Changed = true;
4222  continue;
4223  }
4224 
4225  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4226  E = CSI->handler_end();
4227  I != E; ++I) {
4228  if (*I == BB) {
4229  CSI->removeHandler(I);
4230  --I;
4231  --E;
4232  Changed = true;
4233  }
4234  }
4235  if (CSI->getNumHandlers() == 0) {
4236  BasicBlock *CatchSwitchBB = CSI->getParent();
4237  if (CSI->hasUnwindDest()) {
4238  // Redirect preds to the unwind dest
4239  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4240  } else {
4241  // Rewrite all preds to unwind to caller (or from invoke to call).
4242  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4243  for (BasicBlock *EHPred : EHPreds)
4244  removeUnwindEdge(EHPred);
4245  }
4246  // The catchswitch is no longer reachable.
4247  new UnreachableInst(CSI->getContext(), CSI);
4248  CSI->eraseFromParent();
4249  Changed = true;
4250  }
4251  } else if (isa<CleanupReturnInst>(TI)) {
4252  new UnreachableInst(TI->getContext(), TI);
4253  TI->eraseFromParent();
4254  Changed = true;
4255  }
4256  }
4257 
4258  // If this block is now dead, remove it.
4259  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4260  // We know there are no successors, so just nuke the block.
4261  if (LoopHeaders)
4262  LoopHeaders->erase(BB);
4263  BB->eraseFromParent();
4264  return true;
4265  }
4266 
4267  return Changed;
4268 }
4269 
4271  assert(Cases.size() >= 1);
4272 
4273  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4274  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4275  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4276  return false;
4277  }
4278  return true;
4279 }
4280 
4281 /// Turn a switch with two reachable destinations into an integer range
4282 /// comparison and branch.
4284  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4285 
4286  bool HasDefault =
4287  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4288 
4289  // Partition the cases into two sets with different destinations.
4290  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4291  BasicBlock *DestB = nullptr;
4294 
4295  for (auto Case : SI->cases()) {
4296  BasicBlock *Dest = Case.getCaseSuccessor();
4297  if (!DestA)
4298  DestA = Dest;
4299  if (Dest == DestA) {
4300  CasesA.push_back(Case.getCaseValue());
4301  continue;
4302  }
4303  if (!DestB)
4304  DestB = Dest;
4305  if (Dest == DestB) {
4306  CasesB.push_back(Case.getCaseValue());
4307  continue;
4308  }
4309  return false; // More than two destinations.
4310  }
4311 
4312  assert(DestA && DestB &&
4313  "Single-destination switch should have been folded.");
4314  assert(DestA != DestB);
4315  assert(DestB != SI->getDefaultDest());
4316  assert(!CasesB.empty() && "There must be non-default cases.");
4317  assert(!CasesA.empty() || HasDefault);
4318 
4319  // Figure out if one of the sets of cases form a contiguous range.
4320  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4321  BasicBlock *ContiguousDest = nullptr;
4322  BasicBlock *OtherDest = nullptr;
4323  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4324  ContiguousCases = &CasesA;
4325  ContiguousDest = DestA;
4326  OtherDest = DestB;
4327  } else if (CasesAreContiguous(CasesB)) {
4328  ContiguousCases = &CasesB;
4329  ContiguousDest = DestB;
4330  OtherDest = DestA;
4331  } else
4332  return false;
4333 
4334  // Start building the compare and branch.
4335 
4336  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4337  Constant *NumCases =
4338  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4339 
4340  Value *Sub = SI->getCondition();
4341  if (!Offset->isNullValue())
4342  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4343 
4344  Value *Cmp;
4345  // If NumCases overflowed, then all possible values jump to the successor.
4346  if (NumCases->isNullValue() && !ContiguousCases->empty())
4347  Cmp = ConstantInt::getTrue(SI->getContext());
4348  else
4349  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4350  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4351 
4352  // Update weight for the newly-created conditional branch.
4353  if (HasBranchWeights(SI)) {
4354  SmallVector<uint64_t, 8> Weights;
4355  GetBranchWeights(SI, Weights);
4356  if (Weights.size() == 1 + SI->getNumCases()) {
4357  uint64_t TrueWeight = 0;
4358  uint64_t FalseWeight = 0;
4359  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4360  if (SI->getSuccessor(I) == ContiguousDest)
4361  TrueWeight += Weights[I];
4362  else
4363  FalseWeight += Weights[I];
4364  }
4365  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4366  TrueWeight /= 2;
4367  FalseWeight /= 2;
4368  }
4369  setBranchWeights(NewBI, TrueWeight, FalseWeight);
4370  }
4371  }
4372 
4373  // Prune obsolete incoming values off the successors' PHI nodes.
4374  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4375  unsigned PreviousEdges = ContiguousCases->size();
4376  if (ContiguousDest == SI->getDefaultDest())
4377  ++PreviousEdges;
4378  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4379  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4380  }
4381  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4382  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4383  if (OtherDest == SI->getDefaultDest())
4384  ++PreviousEdges;
4385  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4386  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4387  }
4388 
4389  // Drop the switch.
4390  SI->eraseFromParent();
4391 
4392  return true;
4393 }
4394 
4395 /// Compute masked bits for the condition of a switch
4396 /// and use it to remove dead cases.
4398  const DataLayout &DL) {
4399  Value *Cond = SI->getCondition();
4400  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4401  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4402 
4403  // We can also eliminate cases by determining that their values are outside of
4404  // the limited range of the condition based on how many significant (non-sign)
4405  // bits are in the condition value.
4406  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4407  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4408 
4409  // Gather dead cases.
4411  for (auto &Case : SI->cases()) {
4412  const APInt &CaseVal = Case.getCaseValue()->getValue();
4413  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4414  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4415  DeadCases.push_back(Case.getCaseValue());
4416  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal
4417  << " is dead.\n");
4418  }
4419  }
4420 
4421  // If we can prove that the cases must cover all possible values, the
4422  // default destination becomes dead and we can remove it. If we know some
4423  // of the bits in the value, we can use that to more precisely compute the
4424  // number of possible unique case values.
4425  bool HasDefault =
4426  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4427  const unsigned NumUnknownBits =
4428  Bits - (Known.Zero | Known.One).countPopulation();
4429  assert(NumUnknownBits <= Bits);
4430  if (HasDefault && DeadCases.empty() &&
4431  NumUnknownBits < 64 /* avoid overflow */ &&
4432  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4433  LLVM_DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4434  BasicBlock *NewDefault =
4436  SI->setDefaultDest(&*NewDefault);
4437  SplitBlock(&*NewDefault, &NewDefault->front());
4438  auto *OldTI = NewDefault->getTerminator();
4439  new UnreachableInst(SI->getContext(), OldTI);
4441  return true;
4442  }
4443 
4444  SmallVector<uint64_t, 8> Weights;
4445  bool HasWeight = HasBranchWeights(SI);
4446  if (HasWeight) {
4447  GetBranchWeights(SI, Weights);
4448  HasWeight = (Weights.size() == 1 + SI->getNumCases());
4449  }
4450 
4451  // Remove dead cases from the switch.
4452  for (ConstantInt *DeadCase : DeadCases) {
4453  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4454  assert(CaseI != SI->case_default() &&
4455  "Case was not found. Probably mistake in DeadCases forming.");
4456  if (HasWeight) {
4457  std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4458  Weights.pop_back();
4459  }
4460 
4461  // Prune unused values from PHI nodes.
4462  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4463  SI->removeCase(CaseI);
4464  }
4465  if (HasWeight && Weights.size() >= 2) {
4466  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4467  setBranchWeights(SI, MDWeights);
4468  }
4469 
4470  return !DeadCases.empty();
4471 }
4472 
4473 /// If BB would be eligible for simplification by
4474 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4475 /// by an unconditional branch), look at the phi node for BB in the successor
4476 /// block and see if the incoming value is equal to CaseValue. If so, return
4477 /// the phi node, and set PhiIndex to BB's index in the phi node.
4479  BasicBlock *BB, int *PhiIndex) {
4480  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4481  return nullptr; // BB must be empty to be a candidate for simplification.
4482  if (!BB->getSinglePredecessor())
4483  return nullptr; // BB must be dominated by the switch.
4484 
4486  if (!Branch || !Branch->isUnconditional())
4487  return nullptr; // Terminator must be unconditional branch.
4488 
4489  BasicBlock *Succ = Branch->getSuccessor(0);
4490 
4491  for (PHINode &PHI : Succ->phis()) {
4492  int Idx = PHI.getBasicBlockIndex(BB);
4493  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4494 
4495  Value *InValue = PHI.getIncomingValue(Idx);
4496  if (InValue != CaseValue)
4497  continue;
4498 
4499  *PhiIndex = Idx;
4500  return &PHI;
4501  }
4502 
4503  return nullptr;
4504 }
4505 
4506 /// Try to forward the condition of a switch instruction to a phi node
4507 /// dominated by the switch, if that would mean that some of the destination
4508 /// blocks of the switch can be folded away. Return true if a change is made.
4510  using ForwardingNodesMap = DenseMap<PHINode *, SmallVector<int, 4>>;
4511 
4512  ForwardingNodesMap ForwardingNodes;
4513  BasicBlock *SwitchBlock = SI->getParent();
4514  bool Changed = false;
4515  for (auto &Case : SI->cases()) {
4516  ConstantInt *CaseValue = Case.getCaseValue();
4517  BasicBlock *CaseDest = Case.getCaseSuccessor();
4518 
4519  // Replace phi operands in successor blocks that are using the constant case
4520  // value rather than the switch condition variable:
4521  // switchbb:
4522  // switch i32 %x, label %default [
4523  // i32 17, label %succ
4524  // ...
4525  // succ:
4526  // %r = phi i32 ... [ 17, %switchbb ] ...
4527  // -->
4528  // %r = phi i32 ... [ %x, %switchbb ] ...
4529 
4530  for (PHINode &Phi : CaseDest->phis()) {
4531  // This only works if there is exactly 1 incoming edge from the switch to
4532  // a phi. If there is >1, that means multiple cases of the switch map to 1
4533  // value in the phi, and that phi value is not the switch condition. Thus,
4534  // this transform would not make sense (the phi would be invalid because
4535  // a phi can't have different incoming values from the same block).
4536  int SwitchBBIdx = Phi.getBasicBlockIndex(SwitchBlock);
4537  if (Phi.getIncomingValue(SwitchBBIdx) == CaseValue &&
4538  count(Phi.blocks(), SwitchBlock) == 1) {
4539  Phi.setIncomingValue(SwitchBBIdx, SI->getCondition());
4540  Changed = true;
4541  }
4542  }
4543 
4544  // Collect phi nodes that are indirectly using this switch's case constants.
4545  int PhiIdx;
4546  if (auto *Phi = FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIdx))
4547  ForwardingNodes[Phi].push_back(PhiIdx);
4548  }
4549 
4550  for (auto &ForwardingNode : ForwardingNodes) {
4551  PHINode *Phi = ForwardingNode.first;
4552  SmallVectorImpl<int> &Indexes = ForwardingNode.second;
4553  if (Indexes.size() < 2)
4554  continue;
4555 
4556  for (int Index : Indexes)
4557  Phi->setIncomingValue(Index, SI->getCondition());
4558  Changed = true;
4559  }
4560 
4561  return Changed;
4562 }
4563 
4564 /// Return true if the backend will be able to handle
4565 /// initializing an array of constants like C.
4567  if (C->isThreadDependent())
4568  return false;
4569  if (C->isDLLImportDependent())
4570  return false;
4571 
4572  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4573  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4574  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4575  return false;
4576 
4577  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4578  if (!CE->isGEPWithNoNotionalOverIndexing())
4579  return false;
4580  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4581  return false;
4582  }
4583 
4585  return false;
4586 
4587  return true;
4588 }
4589 
4590 /// If V is a Constant, return it. Otherwise, try to look up
4591 /// its constant value in ConstantPool, returning 0 if it's not there.
4592 static Constant *
4595  if (Constant *C = dyn_cast<Constant>(V))
4596  return C;
4597  return ConstantPool.lookup(V);
4598 }
4599 
4600 /// Try to fold instruction I into a constant. This works for
4601 /// simple instructions such as binary operations where both operands are
4602 /// constant or can be replaced by constants from the ConstantPool. Returns the
4603 /// resulting constant on success, 0 otherwise.
4604 static Constant *
4607  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4608  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4609  if (!A)
4610  return nullptr;
4611  if (A->isAllOnesValue())
4612  return LookupConstant(Select->getTrueValue(), ConstantPool);
4613  if (A->isNullValue())
4614  return LookupConstant(Select->getFalseValue(), ConstantPool);
4615  return nullptr;
4616  }
4617 
4619  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4621  COps.push_back(A);
4622  else
4623  return nullptr;
4624  }
4625 
4626  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4627  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4628  COps[1], DL);
4629  }
4630 
4631  return ConstantFoldInstOperands(I, COps, DL);
4632 }
4633 
4634 /// Try to determine the resulting constant values in phi nodes
4635 /// at the common destination basic block, *CommonDest, for one of the case
4636 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4637 /// case), of a switch instruction SI.
4638 static bool
4640  BasicBlock **CommonDest,
4641  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4642  const DataLayout &DL, const TargetTransformInfo &TTI) {
4643  // The block from which we enter the common destination.
4644  BasicBlock *Pred = SI->getParent();
4645 
4646  // If CaseDest is empty except for some side-effect free instructions through
4647  // which we can constant-propagate the CaseVal, continue to its successor.
4649  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4650  for (Instruction &I :CaseDest->instructionsWithoutDebug()) {
4651  if (I.isTerminator()) {
4652  // If the terminator is a simple branch, continue to the next block.
4653  if (I.getNumSuccessors() != 1 || I.isExceptionalTerminator())
4654  return false;
4655  Pred = CaseDest;
4656  CaseDest = I.getSuccessor(0);
4657  } else if (Constant *C = ConstantFold(&I, DL, ConstantPool)) {
4658  // Instruction is side-effect free and constant.
4659 
4660  // If the instruction has uses outside this block or a phi node slot for
4661  // the block, it is not safe to bypass the instruction since it would then
4662  // no longer dominate all its uses.
4663  for (auto &Use : I.uses()) {
4664  User *User = Use.getUser();
4665  if (Instruction *I = dyn_cast<Instruction>(User))
4666  if (I->getParent() == CaseDest)
4667  continue;
4668  if (PHINode *Phi = dyn_cast<PHINode>(User))
4669  if (Phi->getIncomingBlock(Use) == CaseDest)
4670  continue;
4671  return false;
4672  }
4673 
4674  ConstantPool.insert(std::make_pair(&I, C));
4675  } else {
4676  break;
4677  }
4678  }
4679 
4680  // If we did not have a CommonDest before, use the current one.
4681  if (!*CommonDest)
4682  *CommonDest = CaseDest;
4683  // If the destination isn't the common one, abort.
4684  if (CaseDest != *CommonDest)
4685  return false;
4686 
4687  // Get the values for this case from phi nodes in the destination block.
4688  for (PHINode &PHI : (*CommonDest)->phis()) {
4689  int Idx = PHI.getBasicBlockIndex(Pred);
4690  if (Idx == -1)
4691  continue;
4692 
4693  Constant *ConstVal =
4694  LookupConstant(PHI.getIncomingValue(Idx), ConstantPool);
4695  if (!ConstVal)
4696  return false;
4697 
4698  // Be conservative about which kinds of constants we support.
4699  if (!ValidLookupTableConstant(ConstVal, TTI))
4700  return false;
4701 
4702  Res.push_back(std::make_pair(&PHI, ConstVal));
4703  }
4704 
4705  return Res.size() > 0;
4706 }
4707 
4708 // Helper function used to add CaseVal to the list of cases that generate
4709 // Result. Returns the updated number of cases that generate this result.
4710 static uintptr_t MapCaseToResult(ConstantInt *CaseVal,
4711  SwitchCaseResultVectorTy &UniqueResults,
4712  Constant *Result) {
4713  for (auto &I : UniqueResults) {
4714  if (I.first == Result) {
4715  I.second.push_back(CaseVal);
4716  return I.second.size();
4717  }
4718  }
4719  UniqueResults.push_back(
4720  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4721  return 1;
4722 }
4723 
4724 // Helper function that initializes a map containing
4725 // results for the PHI node of the common destination block for a switch
4726 // instruction. Returns false if multiple PHI nodes have been found or if
4727 // there is not a common destination block for the switch.
4728 static bool
4730  SwitchCaseResultVectorTy &UniqueResults,
4731  Constant *&DefaultResult, const DataLayout &DL,
4732  const TargetTransformInfo &TTI,
4733  uintptr_t MaxUniqueResults, uintptr_t MaxCasesPerResult) {
4734  for (auto &I : SI->cases()) {
4735  ConstantInt *CaseVal = I.getCaseValue();
4736 
4737  // Resulting value at phi nodes for this case value.
4738  SwitchCaseResultsTy Results;
4739  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4740  DL, TTI))
4741  return false;
4742 
4743  // Only one value per case is permitted.
4744  if (Results.size() > 1)
4745  return false;
4746 
4747  // Add the case->result mapping to UniqueResults.
4748  const uintptr_t NumCasesForResult =
4749  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4750 
4751  // Early out if there are too many cases for this result.
4752  if (NumCasesForResult > MaxCasesPerResult)
4753  return false;
4754 
4755  // Early out if there are too many unique results.
4756  if (UniqueResults.size() > MaxUniqueResults)
4757  return false;
4758 
4759  // Check the PHI consistency.
4760  if (!PHI)
4761  PHI = Results[0].first;
4762  else if (PHI != Results[0].first)
4763  return false;
4764  }
4765  // Find the default result value.
4767  BasicBlock *DefaultDest = SI->getDefaultDest();
4768  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4769  DL, TTI);
4770  // If the default value is not found abort unless the default destination
4771  // is unreachable.
4772  DefaultResult =
4773  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4774  if ((!DefaultResult &&
4775  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4776  return false;
4777 
4778  return true;
4779 }
4780 
4781 // Helper function that checks if it is possible to transform a switch with only
4782 // two cases (or two cases + default) that produces a result into a select.
4783 // Example:
4784 // switch (a) {
4785 // case 10: %0 = icmp eq i32 %a, 10
4786 // return 10; %1 = select i1 %0, i32 10, i32 4
4787 // case 20: ----> %2 = icmp eq i32 %a, 20
4788 // return 2; %3 = select i1 %2, i32 2, i32 %1
4789 // default:
4790 // return 4;
4791 // }
4792 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4793  Constant *DefaultResult, Value *Condition,
4794  IRBuilder<> &Builder) {
4795  assert(ResultVector.size() == 2 &&
4796  "We should have exactly two unique results at this point");
4797  // If we are selecting between only two cases transform into a simple
4798  // select or a two-way select if default is possible.
4799  if (ResultVector[0].second.size() == 1 &&
4800  ResultVector[1].second.size() == 1) {
4801  ConstantInt *const FirstCase = ResultVector[0].second[0];
4802  ConstantInt *const SecondCase = ResultVector[1].second[0];
4803 
4804  bool DefaultCanTrigger = DefaultResult;
4805  Value *SelectValue = ResultVector[1].first;
4806  if (DefaultCanTrigger) {
4807  Value *const ValueCompare =
4808  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4809  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4810  DefaultResult, "switch.select");
4811  }
4812  Value *const ValueCompare =
4813  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4814  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4815  SelectValue, "switch.select");
4816  }
4817 
4818  return nullptr;
4819 }
4820 
4821 // Helper function to cleanup a switch instruction that has been converted into
4822 // a select, fixing up PHI nodes and basic blocks.
4824  Value *SelectValue,
4825  IRBuilder<> &Builder) {
4826  BasicBlock *SelectBB = SI->getParent();
4827  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4828  PHI->removeIncomingValue(SelectBB);
4829  PHI->addIncoming(SelectValue, SelectBB);
4830 
4831  Builder.CreateBr(PHI->getParent());
4832 
4833  // Remove the switch.
4834  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4835  BasicBlock *Succ = SI->getSuccessor(i);
4836 
4837  if (Succ == PHI->getParent())
4838  continue;
4839  Succ->removePredecessor(SelectBB);
4840  }
4841  SI->eraseFromParent();
4842 }
4843 
4844 /// If the switch is only used to initialize one or more
4845 /// phi nodes in a common successor block with only two different
4846 /// constant values, replace the switch with select.
4847 static bool switchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4848  const DataLayout &DL,
4849  const TargetTransformInfo &TTI) {
4850  Value *const Cond = SI->getCondition();
4851  PHINode *PHI = nullptr;
4852  BasicBlock *CommonDest = nullptr;
4853  Constant *DefaultResult;
4854  SwitchCaseResultVectorTy UniqueResults;
4855  // Collect all the cases that will deliver the same value from the switch.
4856  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4857  DL, TTI, 2, 1))
4858  return false;
4859  // Selects choose between maximum two values.
4860  if (UniqueResults.size() != 2)
4861  return false;
4862  assert(PHI != nullptr && "PHI for value select not found");
4863 
4864  Builder.SetInsertPoint(SI);
4865  Value *SelectValue =
4866  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4867  if (SelectValue) {
4868  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4869  return true;
4870  }
4871  // The switch couldn't be converted into a select.
4872  return false;
4873 }
4874 
4875 namespace {
4876 
4877 /// This class represents a lookup table that can be used to replace a switch.
4878 class SwitchLookupTable {
4879 public:
4880  /// Create a lookup table to use as a switch replacement with the contents
4881  /// of Values, using DefaultValue to fill any holes in the table.
4882  SwitchLookupTable(
4883  Module &M, uint64_t TableSize, ConstantInt *Offset,
4884  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4885  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4886 
4887  /// Build instructions with Builder to retrieve the value at
4888  /// the position given by Index in the lookup table.
4889  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4890 
4891  /// Return true if a table with TableSize elements of
4892  /// type ElementType would fit in a target-legal register.
4893  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4894  Type *ElementType);
4895 
4896 private:
4897  // Depending on the contents of the table, it can be represented in
4898  // different ways.
4899  enum {
4900  // For tables where each element contains the same value, we just have to
4901  // store that single value and return it for each lookup.
4902  SingleValueKind,
4903 
4904  // For tables where there is a linear relationship between table index
4905  // and values. We calculate the result with a simple multiplication
4906  // and addition instead of a table lookup.
4907  LinearMapKind,
4908 
4909  // For small tables with integer elements, we can pack them into a bitmap
4910  // that fits into a target-legal register. Values are retrieved by
4911  // shift and mask operations.
4912  BitMapKind,
4913 
4914  // The table is stored as an array of values. Values are retrieved by load
4915  // instructions from the table.
4916  ArrayKind
4917  } Kind;
4918 
4919  // For SingleValueKind, this is the single value.
4920  Constant *SingleValue = nullptr;
4921 
4922  // For BitMapKind, this is the bitmap.
4923  ConstantInt *BitMap = nullptr;
4924  IntegerType *BitMapElementTy = nullptr;
4925 
4926  // For LinearMapKind, these are the constants used to derive the value.
4927  ConstantInt *LinearOffset = nullptr;
4928  ConstantInt *LinearMultiplier = nullptr;
4929 
4930  // For ArrayKind, this is the array.
4931  GlobalVariable *Array = nullptr;
4932 };
4933 
4934 } // end anonymous namespace
4935 
4936 SwitchLookupTable::SwitchLookupTable(
4937  Module &M, uint64_t TableSize, ConstantInt *Offset,
4938  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4939  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName) {
4940  assert(Values.size() && "Can't build lookup table without values!");
4941  assert(TableSize >= Values.size() && "Can't fit values in table!");
4942 
4943  // If all values in the table are equal, this is that value.
4944  SingleValue = Values.begin()->second;
4945 
4946  Type *ValueType = Values.begin()->second->getType();
4947 
4948  // Build up the table contents.
4949  SmallVector<Constant *, 64> TableContents(TableSize);
4950  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4951  ConstantInt *CaseVal = Values[I].first;
4952  Constant *CaseRes = Values[I].second;
4953  assert(CaseRes->getType() == ValueType);
4954 
4955  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4956  TableContents[Idx] = CaseRes;
4957 
4958  if (CaseRes != SingleValue)
4959  SingleValue = nullptr;
4960  }
4961 
4962  // Fill in any holes in the table with the default result.
4963  if (Values.size() < TableSize) {
4964  assert(DefaultValue &&
4965  "Need a default value to fill the lookup table holes.");
4966  assert(DefaultValue->getType() == ValueType);
4967  for (uint64_t I = 0; I < TableSize; ++I) {
4968  if (!TableContents[I])
4969  TableContents[I] = DefaultValue;
4970  }
4971 
4972  if (DefaultValue != SingleValue)
4973  SingleValue = nullptr;
4974  }
4975 
4976  // If each element in the table contains the same value, we only need to store
4977  // that single value.
4978  if (SingleValue) {
4979  Kind = SingleValueKind;
4980  return;
4981  }
4982 
4983  // Check if we can derive the value with a linear transformation from the
4984  // table index.
4985  if (isa<IntegerType>(ValueType)) {
4986  bool LinearMappingPossible = true;
4987  APInt PrevVal;
4988  APInt DistToPrev;
4989  assert(TableSize >= 2 && "Should be a SingleValue table.");
4990  // Check if there is the same distance between two consecutive values.
4991  for (uint64_t I = 0; I < TableSize; ++I) {
4992  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4993  if (!ConstVal) {
4994  // This is an undef. We could deal with it, but undefs in lookup tables
4995  // are very seldom. It's probably not worth the additional complexity.
4996  LinearMappingPossible = false;
4997  break;
4998  }
4999  const APInt &Val = ConstVal->getValue();
5000  if (I != 0) {
5001  APInt Dist = Val - PrevVal;
5002  if (I == 1) {
5003  DistToPrev = Dist;
5004  } else if (Dist != DistToPrev) {
5005  LinearMappingPossible = false;
5006  break;
5007  }
5008  }
5009  PrevVal = Val;
5010  }
5011  if (LinearMappingPossible) {
5012  LinearOffset = cast<ConstantInt>(TableContents[0]);
5013  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
5014  Kind = LinearMapKind;
5015  ++NumLinearMaps;
5016  return;
5017  }
5018  }
5019 
5020  // If the type is integer and the table fits in a register, build a bitmap.
5021  if (WouldFitInRegister(DL, TableSize, ValueType)) {
5022  IntegerType *IT = cast<IntegerType>(ValueType);
5023  APInt TableInt(TableSize * IT->getBitWidth(), 0);
5024  for (uint64_t I = TableSize; I > 0; --I) {
5025  TableInt <<= IT->getBitWidth();
5026  // Insert values into the bitmap. Undef values are set to zero.
5027  if (!isa<UndefValue>(TableContents[I - 1])) {
5028  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
5029  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
5030  }
5031  }
5032  BitMap = ConstantInt::get(M.getContext(), TableInt);
5033  BitMapElementTy = IT;
5034  Kind = BitMapKind;
5035  ++NumBitMaps;
5036  return;
5037  }
5038 
5039  // Store the table in an array.
5040  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
5041  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
5042 
5043  Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
5044  GlobalVariable::PrivateLinkage, Initializer,
5045  "switch.table." + FuncName);
5046  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
5047  // Set the alignment to that of an array items. We will be only loading one
5048  // value out of it.
5049  Array->setAlignment(DL.getPrefTypeAlignment(ValueType));
5050  Kind = ArrayKind;
5051 }
5052 
5053 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
5054  switch (Kind) {
5055  case SingleValueKind:
5056  return SingleValue;
5057  case LinearMapKind: {
5058  // Derive the result value from the input value.
5059  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
5060  false, "switch.idx.cast");
5061  if (!LinearMultiplier->isOne())
5062  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
5063  if (!LinearOffset->isZero())
5064  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
5065  return Result;
5066  }
5067  case BitMapKind: {
5068  // Type of the bitmap (e.g. i59).
5069  IntegerType *MapTy = BitMap->getType();
5070 
5071  // Cast Index to the same type as the bitmap.
5072  // Note: The Index is <= the number of elements in the table, so
5073  // truncating it to the width of the bitmask is safe.
5074  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
5075 
5076  // Multiply the shift amount by the element width.
5077  ShiftAmt = Builder.CreateMul(
5078  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
5079  "switch.shiftamt");
5080 
5081  // Shift down.
5082  Value *DownShifted =
5083  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
5084  // Mask off.
5085  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
5086  }
5087  case ArrayKind: {
5088  // Make sure the table index will not overflow when treated as signed.
5089  IntegerType *IT = cast<IntegerType>(Index->getType());
5090  uint64_t TableSize =
5091  Array->getInitializer()->getType()->getArrayNumElements();
5092  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
5093  Index = Builder.CreateZExt(
5094  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
5095  "switch.tableidx.zext");
5096 
5097  Value *GEPIndices[] = {Builder.getInt32(0), Index};
5098  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
5099  GEPIndices, "switch.gep");
5100  return Builder.CreateLoad(
5101  cast<ArrayType>(Array->getValueType())->getElementType(), GEP,
5102  "switch.load");
5103  }
5104  }
5105  llvm_unreachable("Unknown lookup table kind!");
5106 }
5107 
5108 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5109  uint64_t TableSize,
5110  Type *ElementType) {
5111  auto *IT = dyn_cast<IntegerType>(ElementType);
5112  if (!IT)
5113  return false;
5114  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5115  // are <= 15, we could try to narrow the type.
5116 
5117  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5118  if (TableSize >= UINT_MAX / IT->getBitWidth())
5119  return false;
5120  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5121 }
5122 
5123 /// Determine whether a lookup table should be built for this switch, based on
5124 /// the number of cases, size of the table, and the types of the results.
5125 static bool
5126 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5127  const TargetTransformInfo &TTI, const DataLayout &DL,
5128  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5129  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5130  return false; // TableSize overflowed, or mul below might overflow.
5131 
5132  bool AllTablesFitInRegister = true;
5133  bool HasIllegalType = false;
5134  for (const auto &I : ResultTypes) {
5135  Type *Ty = I.second;
5136 
5137  // Saturate this flag to true.
5138  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5139 
5140  // Saturate this flag to false.
5141  AllTablesFitInRegister =
5142  AllTablesFitInRegister &&
5143  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5144 
5145  // If both flags saturate, we're done. NOTE: This *only* works with
5146  // saturating flags, and all flags have to saturate first due to the
5147  // non-deterministic behavior of iterating over a dense map.
5148  if (HasIllegalType && !AllTablesFitInRegister)
5149  break;
5150  }
5151 
5152  // If each table would fit in a register, we should build it anyway.
5153  if (AllTablesFitInRegister)
5154  return true;
5155 
5156  // Don't build a table that doesn't fit in-register if it has illegal types.
5157  if (HasIllegalType)
5158  return false;
5159 
5160  // The table density should be at least 40%. This is the same criterion as for
5161  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5162  // FIXME: Find the best cut-off.
5163  return SI->getNumCases() * 10 >= TableSize * 4;
5164 }
5165 
5166 /// Try to reuse the switch table index compare. Following pattern:
5167 /// \code
5168 /// if (idx < tablesize)
5169 /// r = table[idx]; // table does not contain default_value
5170 /// else
5171 /// r = default_value;
5172 /// if (r != default_value)
5173 /// ...
5174 /// \endcode
5175 /// Is optimized to:
5176 /// \code
5177 /// cond = idx < tablesize;
5178 /// if (cond)
5179 /// r = table[idx];
5180 /// else
5181 /// r = default_value;
5182 /// if (cond)
5183 /// ...
5184 /// \endcode
5185 /// Jump threading will then eliminate the second if(cond).
5186 static void reuseTableCompare(
5187  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5188  Constant *DefaultValue,
5189  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5190  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5191  if (!CmpInst)
5192  return;
5193 
5194  // We require that the compare is in the same block as the phi so that jump
5195  // threading can do its work afterwards.
5196  if (CmpInst->getParent() != PhiBlock)
5197  return;
5198 
5199  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5200  if (!CmpOp1)
5201  return;
5202 
5203  Value *RangeCmp = RangeCheckBranch->getCondition();
5204  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5205  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5206 
5207  // Check if the compare with the default value is constant true or false.
5208  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5209  DefaultValue, CmpOp1, true);
5210  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5211  return;
5212 
5213  // Check if the compare with the case values is distinct from the default
5214  // compare result.
5215  for (auto ValuePair : Values) {
5216  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5217  ValuePair.second, CmpOp1, true);
5218  if (!CaseConst || CaseConst == DefaultConst || isa<UndefValue>(CaseConst))
5219  return;
5220  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5221  "Expect true or false as compare result.");
5222  }
5223 
5224  // Check if the branch instruction dominates the phi node. It's a simple
5225  // dominance check, but sufficient for our needs.
5226  // Although this check is invariant in the calling loops, it's better to do it
5227  // at this late stage. Practically we do it at most once for a switch.
5228  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5229  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5230  BasicBlock *Pred = *PI;
5231  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5232  return;
5233  }
5234 
5235  if (DefaultConst == FalseConst) {
5236  // The compare yields the same result. We can replace it.
5237  CmpInst->replaceAllUsesWith(RangeCmp);
5238  ++NumTableCmpReuses;
5239  } else {
5240  // The compare yields the same result, just inverted. We can replace it.
5241  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5242  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5243  RangeCheckBranch);
5244  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5245  ++NumTableCmpReuses;
5246  }
5247 }
5248 
5249 /// If the switch is only used to initialize one or more phi nodes in a common
5250 /// successor block with different constant values, replace the switch with
5251 /// lookup tables.
5253  const DataLayout &DL,
5254  const TargetTransformInfo &TTI) {
5255  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5256 
5257  Function *Fn = SI->getParent()->getParent();
5258  // Only build lookup table when we have a target that supports it or the
5259  // attribute is not set.
5260  if (!TTI.shouldBuildLookupTables() ||
5261  (Fn->getFnAttribute("no-jump-tables").getValueAsString() == "true"))
5262  return false;
5263 
5264  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5265  // split off a dense part and build a lookup table for that.
5266 
5267  // FIXME: This creates arrays of GEPs to constant strings, which means each
5268  // GEP needs a runtime relocation in PIC code. We should just build one big
5269  // string and lookup indices into that.
5270 
5271  // Ignore switches with less than three cases. Lookup tables will not make
5272  // them faster, so we don't analyze them.
5273  if (SI->getNumCases() < 3)
5274  return false;
5275 
5276  // Figure out the corresponding result for each case value and phi node in the
5277  // common destination, as well as the min and max case values.
5278  assert(!empty(SI->cases()));
5279  SwitchInst::CaseIt CI = SI->case_begin();
5280  ConstantInt *MinCaseVal = CI->getCaseValue();
5281  ConstantInt *MaxCaseVal = CI->getCaseValue();
5282 
5283  BasicBlock *CommonDest = nullptr;
5284 
5285  using ResultListTy = SmallVector<std::pair<ConstantInt *, Constant *>, 4>;
5287 
5288  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5291 
5292  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5293  ConstantInt *CaseVal = CI->getCaseValue();
5294  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5295  MinCaseVal = CaseVal;
5296  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5297  MaxCaseVal = CaseVal;
5298 
5299  // Resulting value at phi nodes for this case value.
5300  using ResultsTy = SmallVector<std::pair<PHINode *, Constant *>, 4>;
5301  ResultsTy Results;
5302  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5303  Results, DL, TTI))
5304  return false;
5305 
5306  // Append the result from this case to the list for each phi.
5307  for (const auto &I : Results) {
5308  PHINode *PHI = I.first;
5309  Constant *Value = I.second;
5310  if (!ResultLists.count(PHI))
5311  PHIs.push_back(PHI);
5312  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5313  }
5314  }
5315 
5316  // Keep track of the result types.
5317  for (PHINode *PHI : PHIs) {
5318  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5319  }
5320 
5321  uint64_t NumResults = ResultLists[PHIs[0]].size();
5322  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5323  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5324  bool TableHasHoles = (NumResults < TableSize);
5325 
5326  // If the table has holes, we need a constant result for the default case
5327  // or a bitmask that fits in a register.
5328  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5329  bool HasDefaultResults =
5330  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5331  DefaultResultsList, DL, TTI);
5332 
5333  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5334  if (NeedMask) {
5335  // As an extra penalty for the validity test we require more cases.
5336  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5337  return false;
5338  if (!DL.fitsInLegalInteger(TableSize))
5339  return false;
5340  }
5341 
5342  for (const auto &I : DefaultResultsList) {
5343  PHINode *PHI = I.first;
5344  Constant *Result = I.second;
5345  DefaultResults[PHI] = Result;
5346  }
5347 
5348  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5349  return false;
5350 
5351  // Create the BB that does the lookups.
5352  Module &Mod = *CommonDest->getParent()->getParent();
5353  BasicBlock *LookupBB = BasicBlock::Create(
5354  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5355 
5356  // Compute the table index value.
5357  Builder.SetInsertPoint(SI);
5358  Value *TableIndex;
5359  if (MinCaseVal->isNullValue())
5360  TableIndex = SI->getCondition();
5361  else
5362  TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
5363  "switch.tableidx");
5364 
5365  // Compute the maximum table size representable by the integer type we are
5366  // switching upon.
5367  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5368  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5369  assert(MaxTableSize >= TableSize &&
5370  "It is impossible for a switch to have more entries than the max "
5371  "representable value of its input integer type's size.");
5372 
5373  // If the default destination is unreachable, or if the lookup table covers
5374  // all values of the conditional variable, branch directly to the lookup table
5375  // BB. Otherwise, check that the condition is within the case range.
5376  const bool DefaultIsReachable =
5377  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5378  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5379  BranchInst *RangeCheckBranch = nullptr;
5380 
5381  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5382  Builder.CreateBr(LookupBB);
5383  // Note: We call removeProdecessor later since we need to be able to get the
5384  // PHI value for the default case in case we're using a bit mask.
5385  } else {
5386  Value *Cmp = Builder.CreateICmpULT(
5387  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5388  RangeCheckBranch =
5389  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5390  }
5391 
5392  // Populate the BB that does the lookups.
5393  Builder.SetInsertPoint(LookupBB);
5394 
5395  if (NeedMask) {
5396  // Before doing the lookup, we do the hole check. The LookupBB is therefore
5397  // re-purposed to do the hole check, and we create a new LookupBB.
5398  BasicBlock *MaskBB = LookupBB;
5399  MaskBB->setName("switch.hole_check");
5400  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5401  CommonDest->getParent(), CommonDest);
5402 
5403  // Make the mask's bitwidth at least 8-bit and a power-of-2 to avoid
5404  // unnecessary illegal types.
5405  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5406  APInt MaskInt(TableSizePowOf2, 0);
5407  APInt One(TableSizePowOf2, 1);
5408  // Build bitmask; fill in a 1 bit for every case.
5409  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5410  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5411  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5412  .getLimitedValue();
5413  MaskInt |= One << Idx;
5414  }
5415  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5416 
5417  // Get the TableIndex'th bit of the bitmask.
5418  // If this bit is 0 (meaning hole) jump to the default destination,
5419  // else continue with table lookup.
5420  IntegerType *MapTy = TableMask->getType();
5421  Value *MaskIndex =
5422  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5423  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5424  Value *LoBit = Builder.CreateTrunc(
5425  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5426  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5427 
5428  Builder.SetInsertPoint(LookupBB);
5429  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5430  }
5431 
5432  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5433  // We cached PHINodes in PHIs. To avoid accessing deleted PHINodes later,
5434  // do not delete PHINodes here.
5436  /*KeepOneInputPHIs=*/true);
5437  }
5438 
5439  bool ReturnedEarly = false;
5440  for (PHINode *PHI : PHIs) {
5441  const ResultListTy &ResultList = ResultLists[PHI];
5442 
5443  // If using a bitmask, use any value to fill the lookup table holes.
5444  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5445  StringRef FuncName = Fn->getName();
5446  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5447  FuncName);
5448 
5449  Value *Result = Table.BuildLookup(TableIndex, Builder);
5450 
5451  // If the result is used to return immediately from the function, we want to
5452  // do that right here.
5453  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5454  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5455  Builder.CreateRet(Result);
5456  ReturnedEarly = true;
5457  break;
5458  }
5459 
5460  // Do a small peephole optimization: re-use the switch table compare if
5461  // possible.
5462  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5463  BasicBlock *PhiBlock = PHI->getParent();
5464  // Search for compare instructions which use the phi.
5465  for (auto *User : PHI->users()) {
5466  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5467  }
5468  }
5469 
5470  PHI->addIncoming(Result, LookupBB);
5471  }
5472 
5473  if (!ReturnedEarly)
5474  Builder.CreateBr(CommonDest);
5475 
5476  // Remove the switch.
5477  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5478  BasicBlock *Succ = SI->getSuccessor(i);
5479 
5480  if (Succ == SI->getDefaultDest())
5481  continue;
5482  Succ->removePredecessor(SI->getParent());
5483  }
5484  SI->eraseFromParent();
5485 
5486  ++NumLookupTables;
5487  if (NeedMask)
5488  ++NumLookupTablesHoles;
5489  return true;
5490 }
5491 
5492 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5493  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5494  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5495  uint64_t Range = Diff + 1;
5496  uint64_t NumCases = Values.size();
5497  // 40% is the default density for building a jump table in optsize/minsize mode.
5498  uint64_t MinDensity = 40;
5499 
5500  return NumCases * 100 >= Range * MinDensity;
5501 }
5502 
5503 /// Try to transform a switch that has "holes" in it to a contiguous sequence
5504 /// of cases.
5505 ///
5506 /// A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5507 /// range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5508 ///
5509 /// This converts a sparse switch into a dense switch which allows better
5510 /// lowering and could also allow transforming into a lookup table.
5512  const DataLayout &DL,
5513  const TargetTransformInfo &TTI) {
5514  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5515  if (CondTy->getIntegerBitWidth() > 64 ||
5516  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5517  return false;
5518  // Only bother with this optimization if there are more than 3 switch cases;
5519  // SDAG will only bother creating jump tables for 4 or more cases.
5520  if (SI->getNumCases() < 4)
5521  return false;
5522 
5523  // This transform is agnostic to the signedness of the input or case values. We
5524  // can treat the case values as signed or unsigned. We can optimize more common
5525  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5526  // as signed.
5527  SmallVector<int64_t,4> Values;
5528  for (auto &C : SI->cases())
5529  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5530  llvm::sort(Values);
5531 
5532  // If the switch is already dense, there's nothing useful to do here.
5533  if (isSwitchDense(Values))
5534  return false;
5535 
5536  // First, transform the values such that they start at zero and ascend.
5537  int64_t Base = Values[0];
5538  for (auto &V : Values)
5539  V -= (uint64_t)(Base);
5540 
5541  // Now we have signed numbers that have been shifted so that, given enough
5542  // precision, there are no negative values. Since the rest of the transform
5543  // is bitwise only, we switch now to an unsigned representation.
5544  uint64_t GCD = 0;
5545  for (auto &V : Values)
5546  GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5547 
5548  // This transform can be done speculatively because it is so cheap - it results
5549  // in a single rotate operation being inserted. This can only happen if the
5550  // factor extracted is a power of 2.
5551  // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5552  // inverse of GCD and then perform this transform.
5553  // FIXME: It's possible that optimizing a switch on powers of two might also
5554  // be beneficial - flag values are often powers of two and we could use a CLZ
5555  // as the key function.
5556  if (GCD <= 1 || !isPowerOf2_64(GCD))
5557  // No common divisor found or too expensive to compute key function.
5558  return false;
5559 
5560  unsigned Shift = Log2_64(GCD);
5561  for (auto &V : Values)
5562  V = (int64_t)((uint64_t)V >> Shift);
5563 
5564  if (!isSwitchDense(Values))
5565  // Transform didn't create a dense switch.
5566  return false;
5567 
5568  // The obvious transform is to shift the switch condition right and emit a
5569  // check that the condition actually cleanly divided by GCD, i.e.
5570  // C & (1 << Shift - 1) == 0
5571  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5572  //
5573  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5574  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5575  // are nonzero then the switch condition will be very large and will hit the
5576  // default case.
5577 
5578  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5579  Builder.SetInsertPoint(SI);
5580  auto *ShiftC = ConstantInt::get(Ty, Shift);
5581  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5582  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5583  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5584  auto *Rot = Builder.CreateOr(LShr, Shl);
5585  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5586 
5587  for (auto Case : SI->cases()) {
5588  auto *Orig = Case.getCaseValue();
5589  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5590  Case.setValue(
5591  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5592  }
5593  return true;
5594 }
5595 
5596 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5597  BasicBlock *BB = SI->