LLVM  6.0.0svn
SimplifyCFG.cpp
Go to the documentation of this file.
1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // Peephole optimize the CFG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SetOperations.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/CFG.h"
33 #include "llvm/IR/CallSite.h"
34 #include "llvm/IR/Constant.h"
35 #include "llvm/IR/ConstantRange.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/DebugInfo.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/GlobalValue.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/InstrTypes.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Intrinsics.h"
48 #include "llvm/IR/LLVMContext.h"
49 #include "llvm/IR/MDBuilder.h"
50 #include "llvm/IR/Metadata.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/IR/NoFolder.h"
53 #include "llvm/IR/Operator.h"
54 #include "llvm/IR/PatternMatch.h"
55 #include "llvm/IR/Type.h"
56 #include "llvm/IR/User.h"
57 #include "llvm/IR/Value.h"
58 #include "llvm/Support/Casting.h"
60 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/KnownBits.h"
68 #include <algorithm>
69 #include <cassert>
70 #include <climits>
71 #include <cstddef>
72 #include <cstdint>
73 #include <iterator>
74 #include <map>
75 #include <set>
76 #include <utility>
77 #include <vector>
78 
79 using namespace llvm;
80 using namespace PatternMatch;
81 
82 #define DEBUG_TYPE "simplifycfg"
83 
84 // Chosen as 2 so as to be cheap, but still to have enough power to fold
85 // a select, so the "clamp" idiom (of a min followed by a max) will be caught.
86 // To catch this, we need to fold a compare and a select, hence '2' being the
87 // minimum reasonable default.
89  "phi-node-folding-threshold", cl::Hidden, cl::init(2),
90  cl::desc(
91  "Control the amount of phi node folding to perform (default = 2)"));
92 
93 static cl::opt<bool> DupRet(
94  "simplifycfg-dup-ret", cl::Hidden, cl::init(false),
95  cl::desc("Duplicate return instructions into unconditional branches"));
96 
97 static cl::opt<bool>
98  SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
99  cl::desc("Sink common instructions down to the end block"));
100 
102  "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
103  cl::desc("Hoist conditional stores if an unconditional store precedes"));
104 
106  "simplifycfg-merge-cond-stores", cl::Hidden, cl::init(true),
107  cl::desc("Hoist conditional stores even if an unconditional store does not "
108  "precede - hoist multiple conditional stores into a single "
109  "predicated store"));
110 
112  "simplifycfg-merge-cond-stores-aggressively", cl::Hidden, cl::init(false),
113  cl::desc("When merging conditional stores, do so even if the resultant "
114  "basic blocks are unlikely to be if-converted as a result"));
115 
117  "speculate-one-expensive-inst", cl::Hidden, cl::init(true),
118  cl::desc("Allow exactly one expensive instruction to be speculatively "
119  "executed"));
120 
122  "max-speculation-depth", cl::Hidden, cl::init(10),
123  cl::desc("Limit maximum recursion depth when calculating costs of "
124  "speculatively executed instructions"));
125 
126 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
127 STATISTIC(NumLinearMaps,
128  "Number of switch instructions turned into linear mapping");
129 STATISTIC(NumLookupTables,
130  "Number of switch instructions turned into lookup tables");
131 STATISTIC(
132  NumLookupTablesHoles,
133  "Number of switch instructions turned into lookup tables (holes checked)");
134 STATISTIC(NumTableCmpReuses, "Number of reused switch table lookup compares");
135 STATISTIC(NumSinkCommons,
136  "Number of common instructions sunk down to the end block");
137 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
138 
139 namespace {
140 
141 // The first field contains the value that the switch produces when a certain
142 // case group is selected, and the second field is a vector containing the
143 // cases composing the case group.
145  SwitchCaseResultVectorTy;
146 // The first field contains the phi node that generates a result of the switch
147 // and the second field contains the value generated for a certain case in the
148 // switch for that PHI.
149 typedef SmallVector<std::pair<PHINode *, Constant *>, 4> SwitchCaseResultsTy;
150 
151 /// ValueEqualityComparisonCase - Represents a case of a switch.
152 struct ValueEqualityComparisonCase {
154  BasicBlock *Dest;
155 
156  ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
157  : Value(Value), Dest(Dest) {}
158 
159  bool operator<(ValueEqualityComparisonCase RHS) const {
160  // Comparing pointers is ok as we only rely on the order for uniquing.
161  return Value < RHS.Value;
162  }
163 
164  bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
165 };
166 
167 class SimplifyCFGOpt {
168  const TargetTransformInfo &TTI;
169  const DataLayout &DL;
170  unsigned BonusInstThreshold;
171  AssumptionCache *AC;
172  SmallPtrSetImpl<BasicBlock *> *LoopHeaders;
173  // See comments in SimplifyCFGOpt::SimplifySwitch.
174  bool LateSimplifyCFG;
175  Value *isValueEqualityComparison(TerminatorInst *TI);
176  BasicBlock *GetValueEqualityComparisonCases(
177  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases);
178  bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
179  BasicBlock *Pred,
180  IRBuilder<> &Builder);
181  bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
182  IRBuilder<> &Builder);
183 
184  bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
185  bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
186  bool SimplifySingleResume(ResumeInst *RI);
187  bool SimplifyCommonResume(ResumeInst *RI);
188  bool SimplifyCleanupReturn(CleanupReturnInst *RI);
189  bool SimplifyUnreachable(UnreachableInst *UI);
190  bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
191  bool SimplifyIndirectBr(IndirectBrInst *IBI);
192  bool SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder);
193  bool SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder);
194 
195 public:
196  SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout &DL,
197  unsigned BonusInstThreshold, AssumptionCache *AC,
198  SmallPtrSetImpl<BasicBlock *> *LoopHeaders,
199  bool LateSimplifyCFG)
200  : TTI(TTI), DL(DL), BonusInstThreshold(BonusInstThreshold), AC(AC),
201  LoopHeaders(LoopHeaders), LateSimplifyCFG(LateSimplifyCFG) {}
202 
203  bool run(BasicBlock *BB);
204 };
205 
206 } // end anonymous namespace
207 
208 /// Return true if it is safe to merge these two
209 /// terminator instructions together.
210 static bool
212  SmallSetVector<BasicBlock *, 4> *FailBlocks = nullptr) {
213  if (SI1 == SI2)
214  return false; // Can't merge with self!
215 
216  // It is not safe to merge these two switch instructions if they have a common
217  // successor, and if that successor has a PHI node, and if *that* PHI node has
218  // conflicting incoming values from the two switch blocks.
219  BasicBlock *SI1BB = SI1->getParent();
220  BasicBlock *SI2BB = SI2->getParent();
221 
222  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
223  bool Fail = false;
224  for (BasicBlock *Succ : successors(SI2BB))
225  if (SI1Succs.count(Succ))
226  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
227  PHINode *PN = cast<PHINode>(BBI);
228  if (PN->getIncomingValueForBlock(SI1BB) !=
229  PN->getIncomingValueForBlock(SI2BB)) {
230  if (FailBlocks)
231  FailBlocks->insert(Succ);
232  Fail = true;
233  }
234  }
235 
236  return !Fail;
237 }
238 
239 /// Return true if it is safe and profitable to merge these two terminator
240 /// instructions together, where SI1 is an unconditional branch. PhiNodes will
241 /// store all PHI nodes in common successors.
242 static bool
244  Instruction *Cond,
245  SmallVectorImpl<PHINode *> &PhiNodes) {
246  if (SI1 == SI2)
247  return false; // Can't merge with self!
248  assert(SI1->isUnconditional() && SI2->isConditional());
249 
250  // We fold the unconditional branch if we can easily update all PHI nodes in
251  // common successors:
252  // 1> We have a constant incoming value for the conditional branch;
253  // 2> We have "Cond" as the incoming value for the unconditional branch;
254  // 3> SI2->getCondition() and Cond have same operands.
255  CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
256  if (!Ci2)
257  return false;
258  if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
259  Cond->getOperand(1) == Ci2->getOperand(1)) &&
260  !(Cond->getOperand(0) == Ci2->getOperand(1) &&
261  Cond->getOperand(1) == Ci2->getOperand(0)))
262  return false;
263 
264  BasicBlock *SI1BB = SI1->getParent();
265  BasicBlock *SI2BB = SI2->getParent();
266  SmallPtrSet<BasicBlock *, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
267  for (BasicBlock *Succ : successors(SI2BB))
268  if (SI1Succs.count(Succ))
269  for (BasicBlock::iterator BBI = Succ->begin(); isa<PHINode>(BBI); ++BBI) {
270  PHINode *PN = cast<PHINode>(BBI);
271  if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
272  !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
273  return false;
274  PhiNodes.push_back(PN);
275  }
276  return true;
277 }
278 
279 /// Update PHI nodes in Succ to indicate that there will now be entries in it
280 /// from the 'NewPred' block. The values that will be flowing into the PHI nodes
281 /// will be the same as those coming in from ExistPred, an existing predecessor
282 /// of Succ.
283 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
284  BasicBlock *ExistPred) {
285  if (!isa<PHINode>(Succ->begin()))
286  return; // Quick exit if nothing to do
287 
288  PHINode *PN;
289  for (BasicBlock::iterator I = Succ->begin(); (PN = dyn_cast<PHINode>(I)); ++I)
290  PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
291 }
292 
293 /// Compute an abstract "cost" of speculating the given instruction,
294 /// which is assumed to be safe to speculate. TCC_Free means cheap,
295 /// TCC_Basic means less cheap, and TCC_Expensive means prohibitively
296 /// expensive.
297 static unsigned ComputeSpeculationCost(const User *I,
298  const TargetTransformInfo &TTI) {
300  "Instruction is not safe to speculatively execute!");
301  return TTI.getUserCost(I);
302 }
303 
304 /// If we have a merge point of an "if condition" as accepted above,
305 /// return true if the specified value dominates the block. We
306 /// don't handle the true generality of domination here, just a special case
307 /// which works well enough for us.
308 ///
309 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
310 /// see if V (which must be an instruction) and its recursive operands
311 /// that do not dominate BB have a combined cost lower than CostRemaining and
312 /// are non-trapping. If both are true, the instruction is inserted into the
313 /// set and true is returned.
314 ///
315 /// The cost for most non-trapping instructions is defined as 1 except for
316 /// Select whose cost is 2.
317 ///
318 /// After this function returns, CostRemaining is decreased by the cost of
319 /// V plus its non-dominating operands. If that cost is greater than
320 /// CostRemaining, false is returned and CostRemaining is undefined.
322  SmallPtrSetImpl<Instruction *> *AggressiveInsts,
323  unsigned &CostRemaining,
324  const TargetTransformInfo &TTI,
325  unsigned Depth = 0) {
326  // It is possible to hit a zero-cost cycle (phi/gep instructions for example),
327  // so limit the recursion depth.
328  // TODO: While this recursion limit does prevent pathological behavior, it
329  // would be better to track visited instructions to avoid cycles.
330  if (Depth == MaxSpeculationDepth)
331  return false;
332 
334  if (!I) {
335  // Non-instructions all dominate instructions, but not all constantexprs
336  // can be executed unconditionally.
337  if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
338  if (C->canTrap())
339  return false;
340  return true;
341  }
342  BasicBlock *PBB = I->getParent();
343 
344  // We don't want to allow weird loops that might have the "if condition" in
345  // the bottom of this block.
346  if (PBB == BB)
347  return false;
348 
349  // If this instruction is defined in a block that contains an unconditional
350  // branch to BB, then it must be in the 'conditional' part of the "if
351  // statement". If not, it definitely dominates the region.
353  if (!BI || BI->isConditional() || BI->getSuccessor(0) != BB)
354  return true;
355 
356  // If we aren't allowing aggressive promotion anymore, then don't consider
357  // instructions in the 'if region'.
358  if (!AggressiveInsts)
359  return false;
360 
361  // If we have seen this instruction before, don't count it again.
362  if (AggressiveInsts->count(I))
363  return true;
364 
365  // Okay, it looks like the instruction IS in the "condition". Check to
366  // see if it's a cheap instruction to unconditionally compute, and if it
367  // only uses stuff defined outside of the condition. If so, hoist it out.
369  return false;
370 
371  unsigned Cost = ComputeSpeculationCost(I, TTI);
372 
373  // Allow exactly one instruction to be speculated regardless of its cost
374  // (as long as it is safe to do so).
375  // This is intended to flatten the CFG even if the instruction is a division
376  // or other expensive operation. The speculation of an expensive instruction
377  // is expected to be undone in CodeGenPrepare if the speculation has not
378  // enabled further IR optimizations.
379  if (Cost > CostRemaining &&
380  (!SpeculateOneExpensiveInst || !AggressiveInsts->empty() || Depth > 0))
381  return false;
382 
383  // Avoid unsigned wrap.
384  CostRemaining = (Cost > CostRemaining) ? 0 : CostRemaining - Cost;
385 
386  // Okay, we can only really hoist these out if their operands do
387  // not take us over the cost threshold.
388  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
389  if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining, TTI,
390  Depth + 1))
391  return false;
392  // Okay, it's safe to do this! Remember this instruction.
393  AggressiveInsts->insert(I);
394  return true;
395 }
396 
397 /// Extract ConstantInt from value, looking through IntToPtr
398 /// and PointerNullValue. Return NULL if value is not a constant int.
399 static ConstantInt *GetConstantInt(Value *V, const DataLayout &DL) {
400  // Normal constant int.
401  ConstantInt *CI = dyn_cast<ConstantInt>(V);
402  if (CI || !isa<Constant>(V) || !V->getType()->isPointerTy())
403  return CI;
404 
405  // This is some kind of pointer constant. Turn it into a pointer-sized
406  // ConstantInt if possible.
407  IntegerType *PtrTy = cast<IntegerType>(DL.getIntPtrType(V->getType()));
408 
409  // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
410  if (isa<ConstantPointerNull>(V))
411  return ConstantInt::get(PtrTy, 0);
412 
413  // IntToPtr const int.
414  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
415  if (CE->getOpcode() == Instruction::IntToPtr)
416  if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
417  // The constant is very likely to have the right type already.
418  if (CI->getType() == PtrTy)
419  return CI;
420  else
421  return cast<ConstantInt>(
422  ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
423  }
424  return nullptr;
425 }
426 
427 namespace {
428 
429 /// Given a chain of or (||) or and (&&) comparison of a value against a
430 /// constant, this will try to recover the information required for a switch
431 /// structure.
432 /// It will depth-first traverse the chain of comparison, seeking for patterns
433 /// like %a == 12 or %a < 4 and combine them to produce a set of integer
434 /// representing the different cases for the switch.
435 /// Note that if the chain is composed of '||' it will build the set of elements
436 /// that matches the comparisons (i.e. any of this value validate the chain)
437 /// while for a chain of '&&' it will build the set elements that make the test
438 /// fail.
439 struct ConstantComparesGatherer {
440  const DataLayout &DL;
441  Value *CompValue; /// Value found for the switch comparison
442  Value *Extra; /// Extra clause to be checked before the switch
443  SmallVector<ConstantInt *, 8> Vals; /// Set of integers to match in switch
444  unsigned UsedICmps; /// Number of comparisons matched in the and/or chain
445 
446  /// Construct and compute the result for the comparison instruction Cond
447  ConstantComparesGatherer(Instruction *Cond, const DataLayout &DL)
448  : DL(DL), CompValue(nullptr), Extra(nullptr), UsedICmps(0) {
449  gather(Cond);
450  }
451 
452  /// Prevent copy
453  ConstantComparesGatherer(const ConstantComparesGatherer &) = delete;
454  ConstantComparesGatherer &
455  operator=(const ConstantComparesGatherer &) = delete;
456 
457 private:
458  /// Try to set the current value used for the comparison, it succeeds only if
459  /// it wasn't set before or if the new value is the same as the old one
460  bool setValueOnce(Value *NewVal) {
461  if (CompValue && CompValue != NewVal)
462  return false;
463  CompValue = NewVal;
464  return (CompValue != nullptr);
465  }
466 
467  /// Try to match Instruction "I" as a comparison against a constant and
468  /// populates the array Vals with the set of values that match (or do not
469  /// match depending on isEQ).
470  /// Return false on failure. On success, the Value the comparison matched
471  /// against is placed in CompValue.
472  /// If CompValue is already set, the function is expected to fail if a match
473  /// is found but the value compared to is different.
474  bool matchInstruction(Instruction *I, bool isEQ) {
475  // If this is an icmp against a constant, handle this as one of the cases.
476  ICmpInst *ICI;
477  ConstantInt *C;
478  if (!((ICI = dyn_cast<ICmpInst>(I)) &&
479  (C = GetConstantInt(I->getOperand(1), DL)))) {
480  return false;
481  }
482 
483  Value *RHSVal;
484  const APInt *RHSC;
485 
486  // Pattern match a special case
487  // (x & ~2^z) == y --> x == y || x == y|2^z
488  // This undoes a transformation done by instcombine to fuse 2 compares.
489  if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
490 
491  // It's a little bit hard to see why the following transformations are
492  // correct. Here is a CVC3 program to verify them for 64-bit values:
493 
494  /*
495  ONE : BITVECTOR(64) = BVZEROEXTEND(0bin1, 63);
496  x : BITVECTOR(64);
497  y : BITVECTOR(64);
498  z : BITVECTOR(64);
499  mask : BITVECTOR(64) = BVSHL(ONE, z);
500  QUERY( (y & ~mask = y) =>
501  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
502  );
503  QUERY( (y | mask = y) =>
504  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
505  );
506  */
507 
508  // Please note that each pattern must be a dual implication (<--> or
509  // iff). One directional implication can create spurious matches. If the
510  // implication is only one-way, an unsatisfiable condition on the left
511  // side can imply a satisfiable condition on the right side. Dual
512  // implication ensures that satisfiable conditions are transformed to
513  // other satisfiable conditions and unsatisfiable conditions are
514  // transformed to other unsatisfiable conditions.
515 
516  // Here is a concrete example of a unsatisfiable condition on the left
517  // implying a satisfiable condition on the right:
518  //
519  // mask = (1 << z)
520  // (x & ~mask) == y --> (x == y || x == (y | mask))
521  //
522  // Substituting y = 3, z = 0 yields:
523  // (x & -2) == 3 --> (x == 3 || x == 2)
524 
525  // Pattern match a special case:
526  /*
527  QUERY( (y & ~mask = y) =>
528  ((x & ~mask = y) <=> (x = y OR x = (y | mask)))
529  );
530  */
531  if (match(ICI->getOperand(0),
532  m_And(m_Value(RHSVal), m_APInt(RHSC)))) {
533  APInt Mask = ~*RHSC;
534  if (Mask.isPowerOf2() && (C->getValue() & ~Mask) == C->getValue()) {
535  // If we already have a value for the switch, it has to match!
536  if (!setValueOnce(RHSVal))
537  return false;
538 
539  Vals.push_back(C);
540  Vals.push_back(
542  C->getValue() | Mask));
543  UsedICmps++;
544  return true;
545  }
546  }
547 
548  // Pattern match a special case:
549  /*
550  QUERY( (y | mask = y) =>
551  ((x | mask = y) <=> (x = y OR x = (y & ~mask)))
552  );
553  */
554  if (match(ICI->getOperand(0),
555  m_Or(m_Value(RHSVal), m_APInt(RHSC)))) {
556  APInt Mask = *RHSC;
557  if (Mask.isPowerOf2() && (C->getValue() | Mask) == C->getValue()) {
558  // If we already have a value for the switch, it has to match!
559  if (!setValueOnce(RHSVal))
560  return false;
561 
562  Vals.push_back(C);
564  C->getValue() & ~Mask));
565  UsedICmps++;
566  return true;
567  }
568  }
569 
570  // If we already have a value for the switch, it has to match!
571  if (!setValueOnce(ICI->getOperand(0)))
572  return false;
573 
574  UsedICmps++;
575  Vals.push_back(C);
576  return ICI->getOperand(0);
577  }
578 
579  // If we have "x ult 3", for example, then we can add 0,1,2 to the set.
581  ICI->getPredicate(), C->getValue());
582 
583  // Shift the range if the compare is fed by an add. This is the range
584  // compare idiom as emitted by instcombine.
585  Value *CandidateVal = I->getOperand(0);
586  if (match(I->getOperand(0), m_Add(m_Value(RHSVal), m_APInt(RHSC)))) {
587  Span = Span.subtract(*RHSC);
588  CandidateVal = RHSVal;
589  }
590 
591  // If this is an and/!= check, then we are looking to build the set of
592  // value that *don't* pass the and chain. I.e. to turn "x ugt 2" into
593  // x != 0 && x != 1.
594  if (!isEQ)
595  Span = Span.inverse();
596 
597  // If there are a ton of values, we don't want to make a ginormous switch.
598  if (Span.isSizeLargerThan(8) || Span.isEmptySet()) {
599  return false;
600  }
601 
602  // If we already have a value for the switch, it has to match!
603  if (!setValueOnce(CandidateVal))
604  return false;
605 
606  // Add all values from the range to the set
607  for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
608  Vals.push_back(ConstantInt::get(I->getContext(), Tmp));
609 
610  UsedICmps++;
611  return true;
612  }
613 
614  /// Given a potentially 'or'd or 'and'd together collection of icmp
615  /// eq/ne/lt/gt instructions that compare a value against a constant, extract
616  /// the value being compared, and stick the list constants into the Vals
617  /// vector.
618  /// One "Extra" case is allowed to differ from the other.
619  void gather(Value *V) {
621  bool isEQ = (I->getOpcode() == Instruction::Or);
622 
623  // Keep a stack (SmallVector for efficiency) for depth-first traversal
625  SmallPtrSet<Value *, 8> Visited;
626 
627  // Initialize
628  Visited.insert(V);
629  DFT.push_back(V);
630 
631  while (!DFT.empty()) {
632  V = DFT.pop_back_val();
633 
634  if (Instruction *I = dyn_cast<Instruction>(V)) {
635  // If it is a || (or && depending on isEQ), process the operands.
636  if (I->getOpcode() == (isEQ ? Instruction::Or : Instruction::And)) {
637  if (Visited.insert(I->getOperand(1)).second)
638  DFT.push_back(I->getOperand(1));
639  if (Visited.insert(I->getOperand(0)).second)
640  DFT.push_back(I->getOperand(0));
641  continue;
642  }
643 
644  // Try to match the current instruction
645  if (matchInstruction(I, isEQ))
646  // Match succeed, continue the loop
647  continue;
648  }
649 
650  // One element of the sequence of || (or &&) could not be match as a
651  // comparison against the same value as the others.
652  // We allow only one "Extra" case to be checked before the switch
653  if (!Extra) {
654  Extra = V;
655  continue;
656  }
657  // Failed to parse a proper sequence, abort now
658  CompValue = nullptr;
659  break;
660  }
661  }
662 };
663 
664 } // end anonymous namespace
665 
667  Instruction *Cond = nullptr;
668  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
669  Cond = dyn_cast<Instruction>(SI->getCondition());
670  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
671  if (BI->isConditional())
672  Cond = dyn_cast<Instruction>(BI->getCondition());
673  } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
674  Cond = dyn_cast<Instruction>(IBI->getAddress());
675  }
676 
677  TI->eraseFromParent();
678  if (Cond)
680 }
681 
682 /// Return true if the specified terminator checks
683 /// to see if a value is equal to constant integer value.
684 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
685  Value *CV = nullptr;
686  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
687  // Do not permit merging of large switch instructions into their
688  // predecessors unless there is only one predecessor.
689  if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
690  pred_end(SI->getParent())) <=
691  128)
692  CV = SI->getCondition();
693  } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
694  if (BI->isConditional() && BI->getCondition()->hasOneUse())
695  if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) {
696  if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), DL))
697  CV = ICI->getOperand(0);
698  }
699 
700  // Unwrap any lossless ptrtoint cast.
701  if (CV) {
702  if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
703  Value *Ptr = PTII->getPointerOperand();
704  if (PTII->getType() == DL.getIntPtrType(Ptr->getType()))
705  CV = Ptr;
706  }
707  }
708  return CV;
709 }
710 
711 /// Given a value comparison instruction,
712 /// decode all of the 'cases' that it represents and return the 'default' block.
713 BasicBlock *SimplifyCFGOpt::GetValueEqualityComparisonCases(
714  TerminatorInst *TI, std::vector<ValueEqualityComparisonCase> &Cases) {
715  if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
716  Cases.reserve(SI->getNumCases());
717  for (auto Case : SI->cases())
718  Cases.push_back(ValueEqualityComparisonCase(Case.getCaseValue(),
719  Case.getCaseSuccessor()));
720  return SI->getDefaultDest();
721  }
722 
723  BranchInst *BI = cast<BranchInst>(TI);
724  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
725  BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
726  Cases.push_back(ValueEqualityComparisonCase(
727  GetConstantInt(ICI->getOperand(1), DL), Succ));
728  return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
729 }
730 
731 /// Given a vector of bb/value pairs, remove any entries
732 /// in the list that match the specified block.
733 static void
735  std::vector<ValueEqualityComparisonCase> &Cases) {
736  Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
737 }
738 
739 /// Return true if there are any keys in C1 that exist in C2 as well.
740 static bool ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
741  std::vector<ValueEqualityComparisonCase> &C2) {
742  std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
743 
744  // Make V1 be smaller than V2.
745  if (V1->size() > V2->size())
746  std::swap(V1, V2);
747 
748  if (V1->empty())
749  return false;
750  if (V1->size() == 1) {
751  // Just scan V2.
752  ConstantInt *TheVal = (*V1)[0].Value;
753  for (unsigned i = 0, e = V2->size(); i != e; ++i)
754  if (TheVal == (*V2)[i].Value)
755  return true;
756  }
757 
758  // Otherwise, just sort both lists and compare element by element.
759  array_pod_sort(V1->begin(), V1->end());
760  array_pod_sort(V2->begin(), V2->end());
761  unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
762  while (i1 != e1 && i2 != e2) {
763  if ((*V1)[i1].Value == (*V2)[i2].Value)
764  return true;
765  if ((*V1)[i1].Value < (*V2)[i2].Value)
766  ++i1;
767  else
768  ++i2;
769  }
770  return false;
771 }
772 
773 /// If TI is known to be a terminator instruction and its block is known to
774 /// only have a single predecessor block, check to see if that predecessor is
775 /// also a value comparison with the same value, and if that comparison
776 /// determines the outcome of this comparison. If so, simplify TI. This does a
777 /// very limited form of jump threading.
778 bool SimplifyCFGOpt::SimplifyEqualityComparisonWithOnlyPredecessor(
779  TerminatorInst *TI, BasicBlock *Pred, IRBuilder<> &Builder) {
780  Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
781  if (!PredVal)
782  return false; // Not a value comparison in predecessor.
783 
784  Value *ThisVal = isValueEqualityComparison(TI);
785  assert(ThisVal && "This isn't a value comparison!!");
786  if (ThisVal != PredVal)
787  return false; // Different predicates.
788 
789  // TODO: Preserve branch weight metadata, similarly to how
790  // FoldValueComparisonIntoPredecessors preserves it.
791 
792  // Find out information about when control will move from Pred to TI's block.
793  std::vector<ValueEqualityComparisonCase> PredCases;
794  BasicBlock *PredDef =
795  GetValueEqualityComparisonCases(Pred->getTerminator(), PredCases);
796  EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
797 
798  // Find information about how control leaves this block.
799  std::vector<ValueEqualityComparisonCase> ThisCases;
800  BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
801  EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
802 
803  // If TI's block is the default block from Pred's comparison, potentially
804  // simplify TI based on this knowledge.
805  if (PredDef == TI->getParent()) {
806  // If we are here, we know that the value is none of those cases listed in
807  // PredCases. If there are any cases in ThisCases that are in PredCases, we
808  // can simplify TI.
809  if (!ValuesOverlap(PredCases, ThisCases))
810  return false;
811 
812  if (isa<BranchInst>(TI)) {
813  // Okay, one of the successors of this condbr is dead. Convert it to a
814  // uncond br.
815  assert(ThisCases.size() == 1 && "Branch can only have one case!");
816  // Insert the new branch.
817  Instruction *NI = Builder.CreateBr(ThisDef);
818  (void)NI;
819 
820  // Remove PHI node entries for the dead edge.
821  ThisCases[0].Dest->removePredecessor(TI->getParent());
822 
823  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
824  << "Through successor TI: " << *TI << "Leaving: " << *NI
825  << "\n");
826 
828  return true;
829  }
830 
831  SwitchInst *SI = cast<SwitchInst>(TI);
832  // Okay, TI has cases that are statically dead, prune them away.
833  SmallPtrSet<Constant *, 16> DeadCases;
834  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
835  DeadCases.insert(PredCases[i].Value);
836 
837  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
838  << "Through successor TI: " << *TI);
839 
840  // Collect branch weights into a vector.
841  SmallVector<uint32_t, 8> Weights;
843  bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
844  if (HasWeight)
845  for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
846  ++MD_i) {
847  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(MD_i));
848  Weights.push_back(CI->getValue().getZExtValue());
849  }
850  for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
851  --i;
852  if (DeadCases.count(i->getCaseValue())) {
853  if (HasWeight) {
854  std::swap(Weights[i->getCaseIndex() + 1], Weights.back());
855  Weights.pop_back();
856  }
857  i->getCaseSuccessor()->removePredecessor(TI->getParent());
858  SI->removeCase(i);
859  }
860  }
861  if (HasWeight && Weights.size() >= 2)
863  MDBuilder(SI->getParent()->getContext())
864  .createBranchWeights(Weights));
865 
866  DEBUG(dbgs() << "Leaving: " << *TI << "\n");
867  return true;
868  }
869 
870  // Otherwise, TI's block must correspond to some matched value. Find out
871  // which value (or set of values) this is.
872  ConstantInt *TIV = nullptr;
873  BasicBlock *TIBB = TI->getParent();
874  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
875  if (PredCases[i].Dest == TIBB) {
876  if (TIV)
877  return false; // Cannot handle multiple values coming to this block.
878  TIV = PredCases[i].Value;
879  }
880  assert(TIV && "No edge from pred to succ?");
881 
882  // Okay, we found the one constant that our value can be if we get into TI's
883  // BB. Find out which successor will unconditionally be branched to.
884  BasicBlock *TheRealDest = nullptr;
885  for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
886  if (ThisCases[i].Value == TIV) {
887  TheRealDest = ThisCases[i].Dest;
888  break;
889  }
890 
891  // If not handled by any explicit cases, it is handled by the default case.
892  if (!TheRealDest)
893  TheRealDest = ThisDef;
894 
895  // Remove PHI node entries for dead edges.
896  BasicBlock *CheckEdge = TheRealDest;
897  for (BasicBlock *Succ : successors(TIBB))
898  if (Succ != CheckEdge)
899  Succ->removePredecessor(TIBB);
900  else
901  CheckEdge = nullptr;
902 
903  // Insert the new branch.
904  Instruction *NI = Builder.CreateBr(TheRealDest);
905  (void)NI;
906 
907  DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
908  << "Through successor TI: " << *TI << "Leaving: " << *NI
909  << "\n");
910 
912  return true;
913 }
914 
915 namespace {
916 
917 /// This class implements a stable ordering of constant
918 /// integers that does not depend on their address. This is important for
919 /// applications that sort ConstantInt's to ensure uniqueness.
920 struct ConstantIntOrdering {
921  bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
922  return LHS->getValue().ult(RHS->getValue());
923  }
924 };
925 
926 } // end anonymous namespace
927 
928 static int ConstantIntSortPredicate(ConstantInt *const *P1,
929  ConstantInt *const *P2) {
930  const ConstantInt *LHS = *P1;
931  const ConstantInt *RHS = *P2;
932  if (LHS == RHS)
933  return 0;
934  return LHS->getValue().ult(RHS->getValue()) ? 1 : -1;
935 }
936 
937 static inline bool HasBranchWeights(const Instruction *I) {
939  if (ProfMD && ProfMD->getOperand(0))
940  if (MDString *MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
941  return MDS->getString().equals("branch_weights");
942 
943  return false;
944 }
945 
946 /// Get Weights of a given TerminatorInst, the default weight is at the front
947 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
948 /// metadata.
950  SmallVectorImpl<uint64_t> &Weights) {
952  assert(MD);
953  for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
954  ConstantInt *CI = mdconst::extract<ConstantInt>(MD->getOperand(i));
955  Weights.push_back(CI->getValue().getZExtValue());
956  }
957 
958  // If TI is a conditional eq, the default case is the false case,
959  // and the corresponding branch-weight data is at index 2. We swap the
960  // default weight to be the first entry.
961  if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
962  assert(Weights.size() == 2);
963  ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
964  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
965  std::swap(Weights.front(), Weights.back());
966  }
967 }
968 
969 /// Keep halving the weights until all can fit in uint32_t.
970 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
971  uint64_t Max = *std::max_element(Weights.begin(), Weights.end());
972  if (Max > UINT_MAX) {
973  unsigned Offset = 32 - countLeadingZeros(Max);
974  for (uint64_t &I : Weights)
975  I >>= Offset;
976  }
977 }
978 
979 /// The specified terminator is a value equality comparison instruction
980 /// (either a switch or a branch on "X == c").
981 /// See if any of the predecessors of the terminator block are value comparisons
982 /// on the same value. If so, and if safe to do so, fold them together.
983 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
984  IRBuilder<> &Builder) {
985  BasicBlock *BB = TI->getParent();
986  Value *CV = isValueEqualityComparison(TI); // CondVal
987  assert(CV && "Not a comparison?");
988  bool Changed = false;
989 
991  while (!Preds.empty()) {
992  BasicBlock *Pred = Preds.pop_back_val();
993 
994  // See if the predecessor is a comparison with the same value.
995  TerminatorInst *PTI = Pred->getTerminator();
996  Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
997 
998  if (PCV == CV && TI != PTI) {
1000  if (!SafeToMergeTerminators(TI, PTI, &FailBlocks)) {
1001  for (auto *Succ : FailBlocks) {
1002  if (!SplitBlockPredecessors(Succ, TI->getParent(), ".fold.split"))
1003  return false;
1004  }
1005  }
1006 
1007  // Figure out which 'cases' to copy from SI to PSI.
1008  std::vector<ValueEqualityComparisonCase> BBCases;
1009  BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
1010 
1011  std::vector<ValueEqualityComparisonCase> PredCases;
1012  BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
1013 
1014  // Based on whether the default edge from PTI goes to BB or not, fill in
1015  // PredCases and PredDefault with the new switch cases we would like to
1016  // build.
1017  SmallVector<BasicBlock *, 8> NewSuccessors;
1018 
1019  // Update the branch weight metadata along the way
1020  SmallVector<uint64_t, 8> Weights;
1021  bool PredHasWeights = HasBranchWeights(PTI);
1022  bool SuccHasWeights = HasBranchWeights(TI);
1023 
1024  if (PredHasWeights) {
1025  GetBranchWeights(PTI, Weights);
1026  // branch-weight metadata is inconsistent here.
1027  if (Weights.size() != 1 + PredCases.size())
1028  PredHasWeights = SuccHasWeights = false;
1029  } else if (SuccHasWeights)
1030  // If there are no predecessor weights but there are successor weights,
1031  // populate Weights with 1, which will later be scaled to the sum of
1032  // successor's weights
1033  Weights.assign(1 + PredCases.size(), 1);
1034 
1035  SmallVector<uint64_t, 8> SuccWeights;
1036  if (SuccHasWeights) {
1037  GetBranchWeights(TI, SuccWeights);
1038  // branch-weight metadata is inconsistent here.
1039  if (SuccWeights.size() != 1 + BBCases.size())
1040  PredHasWeights = SuccHasWeights = false;
1041  } else if (PredHasWeights)
1042  SuccWeights.assign(1 + BBCases.size(), 1);
1043 
1044  if (PredDefault == BB) {
1045  // If this is the default destination from PTI, only the edges in TI
1046  // that don't occur in PTI, or that branch to BB will be activated.
1047  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1048  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1049  if (PredCases[i].Dest != BB)
1050  PTIHandled.insert(PredCases[i].Value);
1051  else {
1052  // The default destination is BB, we don't need explicit targets.
1053  std::swap(PredCases[i], PredCases.back());
1054 
1055  if (PredHasWeights || SuccHasWeights) {
1056  // Increase weight for the default case.
1057  Weights[0] += Weights[i + 1];
1058  std::swap(Weights[i + 1], Weights.back());
1059  Weights.pop_back();
1060  }
1061 
1062  PredCases.pop_back();
1063  --i;
1064  --e;
1065  }
1066 
1067  // Reconstruct the new switch statement we will be building.
1068  if (PredDefault != BBDefault) {
1069  PredDefault->removePredecessor(Pred);
1070  PredDefault = BBDefault;
1071  NewSuccessors.push_back(BBDefault);
1072  }
1073 
1074  unsigned CasesFromPred = Weights.size();
1075  uint64_t ValidTotalSuccWeight = 0;
1076  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1077  if (!PTIHandled.count(BBCases[i].Value) &&
1078  BBCases[i].Dest != BBDefault) {
1079  PredCases.push_back(BBCases[i]);
1080  NewSuccessors.push_back(BBCases[i].Dest);
1081  if (SuccHasWeights || PredHasWeights) {
1082  // The default weight is at index 0, so weight for the ith case
1083  // should be at index i+1. Scale the cases from successor by
1084  // PredDefaultWeight (Weights[0]).
1085  Weights.push_back(Weights[0] * SuccWeights[i + 1]);
1086  ValidTotalSuccWeight += SuccWeights[i + 1];
1087  }
1088  }
1089 
1090  if (SuccHasWeights || PredHasWeights) {
1091  ValidTotalSuccWeight += SuccWeights[0];
1092  // Scale the cases from predecessor by ValidTotalSuccWeight.
1093  for (unsigned i = 1; i < CasesFromPred; ++i)
1094  Weights[i] *= ValidTotalSuccWeight;
1095  // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
1096  Weights[0] *= SuccWeights[0];
1097  }
1098  } else {
1099  // If this is not the default destination from PSI, only the edges
1100  // in SI that occur in PSI with a destination of BB will be
1101  // activated.
1102  std::set<ConstantInt *, ConstantIntOrdering> PTIHandled;
1103  std::map<ConstantInt *, uint64_t> WeightsForHandled;
1104  for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1105  if (PredCases[i].Dest == BB) {
1106  PTIHandled.insert(PredCases[i].Value);
1107 
1108  if (PredHasWeights || SuccHasWeights) {
1109  WeightsForHandled[PredCases[i].Value] = Weights[i + 1];
1110  std::swap(Weights[i + 1], Weights.back());
1111  Weights.pop_back();
1112  }
1113 
1114  std::swap(PredCases[i], PredCases.back());
1115  PredCases.pop_back();
1116  --i;
1117  --e;
1118  }
1119 
1120  // Okay, now we know which constants were sent to BB from the
1121  // predecessor. Figure out where they will all go now.
1122  for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1123  if (PTIHandled.count(BBCases[i].Value)) {
1124  // If this is one we are capable of getting...
1125  if (PredHasWeights || SuccHasWeights)
1126  Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1127  PredCases.push_back(BBCases[i]);
1128  NewSuccessors.push_back(BBCases[i].Dest);
1129  PTIHandled.erase(
1130  BBCases[i].Value); // This constant is taken care of
1131  }
1132 
1133  // If there are any constants vectored to BB that TI doesn't handle,
1134  // they must go to the default destination of TI.
1135  for (ConstantInt *I : PTIHandled) {
1136  if (PredHasWeights || SuccHasWeights)
1137  Weights.push_back(WeightsForHandled[I]);
1138  PredCases.push_back(ValueEqualityComparisonCase(I, BBDefault));
1139  NewSuccessors.push_back(BBDefault);
1140  }
1141  }
1142 
1143  // Okay, at this point, we know which new successor Pred will get. Make
1144  // sure we update the number of entries in the PHI nodes for these
1145  // successors.
1146  for (BasicBlock *NewSuccessor : NewSuccessors)
1147  AddPredecessorToBlock(NewSuccessor, Pred, BB);
1148 
1149  Builder.SetInsertPoint(PTI);
1150  // Convert pointer to int before we switch.
1151  if (CV->getType()->isPointerTy()) {
1152  CV = Builder.CreatePtrToInt(CV, DL.getIntPtrType(CV->getType()),
1153  "magicptr");
1154  }
1155 
1156  // Now that the successors are updated, create the new Switch instruction.
1157  SwitchInst *NewSI =
1158  Builder.CreateSwitch(CV, PredDefault, PredCases.size());
1159  NewSI->setDebugLoc(PTI->getDebugLoc());
1160  for (ValueEqualityComparisonCase &V : PredCases)
1161  NewSI->addCase(V.Value, V.Dest);
1162 
1163  if (PredHasWeights || SuccHasWeights) {
1164  // Halve the weights if any of them cannot fit in an uint32_t
1165  FitWeights(Weights);
1166 
1167  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1168 
1169  NewSI->setMetadata(
1171  MDBuilder(BB->getContext()).createBranchWeights(MDWeights));
1172  }
1173 
1175 
1176  // Okay, last check. If BB is still a successor of PSI, then we must
1177  // have an infinite loop case. If so, add an infinitely looping block
1178  // to handle the case to preserve the behavior of the code.
1179  BasicBlock *InfLoopBlock = nullptr;
1180  for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1181  if (NewSI->getSuccessor(i) == BB) {
1182  if (!InfLoopBlock) {
1183  // Insert it at the end of the function, because it's either code,
1184  // or it won't matter if it's hot. :)
1185  InfLoopBlock = BasicBlock::Create(BB->getContext(), "infloop",
1186  BB->getParent());
1187  BranchInst::Create(InfLoopBlock, InfLoopBlock);
1188  }
1189  NewSI->setSuccessor(i, InfLoopBlock);
1190  }
1191 
1192  Changed = true;
1193  }
1194  }
1195  return Changed;
1196 }
1197 
1198 // If we would need to insert a select that uses the value of this invoke
1199 // (comments in HoistThenElseCodeToIf explain why we would need to do this), we
1200 // can't hoist the invoke, as there is nowhere to put the select in this case.
1202  Instruction *I1, Instruction *I2) {
1203  for (BasicBlock *Succ : successors(BB1)) {
1204  PHINode *PN;
1205  for (BasicBlock::iterator BBI = Succ->begin();
1206  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1207  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1208  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1209  if (BB1V != BB2V && (BB1V == I1 || BB2V == I2)) {
1210  return false;
1211  }
1212  }
1213  }
1214  return true;
1215 }
1216 
1218 
1219 /// Given a conditional branch that goes to BB1 and BB2, hoist any common code
1220 /// in the two blocks up into the branch block. The caller of this function
1221 /// guarantees that BI's block dominates BB1 and BB2.
1223  const TargetTransformInfo &TTI) {
1224  // This does very trivial matching, with limited scanning, to find identical
1225  // instructions in the two blocks. In particular, we don't want to get into
1226  // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1227  // such, we currently just scan for obviously identical instructions in an
1228  // identical order.
1229  BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1230  BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1231 
1232  BasicBlock::iterator BB1_Itr = BB1->begin();
1233  BasicBlock::iterator BB2_Itr = BB2->begin();
1234 
1235  Instruction *I1 = &*BB1_Itr++, *I2 = &*BB2_Itr++;
1236  // Skip debug info if it is not identical.
1239  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1240  while (isa<DbgInfoIntrinsic>(I1))
1241  I1 = &*BB1_Itr++;
1242  while (isa<DbgInfoIntrinsic>(I2))
1243  I2 = &*BB2_Itr++;
1244  }
1245  if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1246  (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1247  return false;
1248 
1249  BasicBlock *BIParent = BI->getParent();
1250 
1251  bool Changed = false;
1252  do {
1253  // If we are hoisting the terminator instruction, don't move one (making a
1254  // broken BB), instead clone it, and remove BI.
1255  if (isa<TerminatorInst>(I1))
1256  goto HoistTerminator;
1257 
1258  if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
1259  return Changed;
1260 
1261  // For a normal instruction, we just move one to right before the branch,
1262  // then replace all uses of the other with the first. Finally, we remove
1263  // the now redundant second instruction.
1264  BIParent->getInstList().splice(BI->getIterator(), BB1->getInstList(), I1);
1265  if (!I2->use_empty())
1266  I2->replaceAllUsesWith(I1);
1267  I1->andIRFlags(I2);
1268  unsigned KnownIDs[] = {LLVMContext::MD_tbaa,
1278  combineMetadata(I1, I2, KnownIDs);
1279 
1280  // I1 and I2 are being combined into a single instruction. Its debug
1281  // location is the merged locations of the original instructions.
1282  if (!isa<CallInst>(I1))
1283  I1->setDebugLoc(
1284  DILocation::getMergedLocation(I1->getDebugLoc(), I2->getDebugLoc()));
1285 
1286  I2->eraseFromParent();
1287  Changed = true;
1288 
1289  I1 = &*BB1_Itr++;
1290  I2 = &*BB2_Itr++;
1291  // Skip debug info if it is not identical.
1294  if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1295  while (isa<DbgInfoIntrinsic>(I1))
1296  I1 = &*BB1_Itr++;
1297  while (isa<DbgInfoIntrinsic>(I2))
1298  I2 = &*BB2_Itr++;
1299  }
1300  } while (I1->isIdenticalToWhenDefined(I2));
1301 
1302  return true;
1303 
1304 HoistTerminator:
1305  // It may not be possible to hoist an invoke.
1306  if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1307  return Changed;
1308 
1309  for (BasicBlock *Succ : successors(BB1)) {
1310  PHINode *PN;
1311  for (BasicBlock::iterator BBI = Succ->begin();
1312  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1313  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1314  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1315  if (BB1V == BB2V)
1316  continue;
1317 
1318  // Check for passingValueIsAlwaysUndefined here because we would rather
1319  // eliminate undefined control flow then converting it to a select.
1320  if (passingValueIsAlwaysUndefined(BB1V, PN) ||
1322  return Changed;
1323 
1324  if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1325  return Changed;
1326  if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1327  return Changed;
1328  }
1329  }
1330 
1331  // Okay, it is safe to hoist the terminator.
1332  Instruction *NT = I1->clone();
1333  BIParent->getInstList().insert(BI->getIterator(), NT);
1334  if (!NT->getType()->isVoidTy()) {
1335  I1->replaceAllUsesWith(NT);
1336  I2->replaceAllUsesWith(NT);
1337  NT->takeName(I1);
1338  }
1339 
1340  IRBuilder<NoFolder> Builder(NT);
1341  // Hoisting one of the terminators from our successor is a great thing.
1342  // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1343  // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1344  // nodes, so we insert select instruction to compute the final result.
1345  std::map<std::pair<Value *, Value *>, SelectInst *> InsertedSelects;
1346  for (BasicBlock *Succ : successors(BB1)) {
1347  PHINode *PN;
1348  for (BasicBlock::iterator BBI = Succ->begin();
1349  (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1350  Value *BB1V = PN->getIncomingValueForBlock(BB1);
1351  Value *BB2V = PN->getIncomingValueForBlock(BB2);
1352  if (BB1V == BB2V)
1353  continue;
1354 
1355  // These values do not agree. Insert a select instruction before NT
1356  // that determines the right value.
1357  SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1358  if (!SI)
1359  SI = cast<SelectInst>(
1360  Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1361  BB1V->getName() + "." + BB2V->getName(), BI));
1362 
1363  // Make the PHI node use the select for all incoming values for BB1/BB2
1364  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1365  if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1366  PN->setIncomingValue(i, SI);
1367  }
1368  }
1369 
1370  // Update any PHI nodes in our new successors.
1371  for (BasicBlock *Succ : successors(BB1))
1372  AddPredecessorToBlock(Succ, BIParent, BB1);
1373 
1375  return true;
1376 }
1377 
1378 // All instructions in Insts belong to different blocks that all unconditionally
1379 // branch to a common successor. Analyze each instruction and return true if it
1380 // would be possible to sink them into their successor, creating one common
1381 // instruction instead. For every value that would be required to be provided by
1382 // PHI node (because an operand varies in each input block), add to PHIOperands.
1385  DenseMap<Instruction *, SmallVector<Value *, 4>> &PHIOperands) {
1386  // Prune out obviously bad instructions to move. Any non-store instruction
1387  // must have exactly one use, and we check later that use is by a single,
1388  // common PHI instruction in the successor.
1389  for (auto *I : Insts) {
1390  // These instructions may change or break semantics if moved.
1391  if (isa<PHINode>(I) || I->isEHPad() || isa<AllocaInst>(I) ||
1392  I->getType()->isTokenTy())
1393  return false;
1394 
1395  // Conservatively return false if I is an inline-asm instruction. Sinking
1396  // and merging inline-asm instructions can potentially create arguments
1397  // that cannot satisfy the inline-asm constraints.
1398  if (const auto *C = dyn_cast<CallInst>(I))
1399  if (C->isInlineAsm())
1400  return false;
1401 
1402  // Everything must have only one use too, apart from stores which
1403  // have no uses.
1404  if (!isa<StoreInst>(I) && !I->hasOneUse())
1405  return false;
1406  }
1407 
1408  const Instruction *I0 = Insts.front();
1409  for (auto *I : Insts)
1410  if (!I->isSameOperationAs(I0))
1411  return false;
1412 
1413  // All instructions in Insts are known to be the same opcode. If they aren't
1414  // stores, check the only user of each is a PHI or in the same block as the
1415  // instruction, because if a user is in the same block as an instruction
1416  // we're contemplating sinking, it must already be determined to be sinkable.
1417  if (!isa<StoreInst>(I0)) {
1418  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1419  auto *Succ = I0->getParent()->getTerminator()->getSuccessor(0);
1420  if (!all_of(Insts, [&PNUse,&Succ](const Instruction *I) -> bool {
1421  auto *U = cast<Instruction>(*I->user_begin());
1422  return (PNUse &&
1423  PNUse->getParent() == Succ &&
1424  PNUse->getIncomingValueForBlock(I->getParent()) == I) ||
1425  U->getParent() == I->getParent();
1426  }))
1427  return false;
1428  }
1429 
1430  // Because SROA can't handle speculating stores of selects, try not
1431  // to sink loads or stores of allocas when we'd have to create a PHI for
1432  // the address operand. Also, because it is likely that loads or stores
1433  // of allocas will disappear when Mem2Reg/SROA is run, don't sink them.
1434  // This can cause code churn which can have unintended consequences down
1435  // the line - see https://llvm.org/bugs/show_bug.cgi?id=30244.
1436  // FIXME: This is a workaround for a deficiency in SROA - see
1437  // https://llvm.org/bugs/show_bug.cgi?id=30188
1438  if (isa<StoreInst>(I0) && any_of(Insts, [](const Instruction *I) {
1439  return isa<AllocaInst>(I->getOperand(1));
1440  }))
1441  return false;
1442  if (isa<LoadInst>(I0) && any_of(Insts, [](const Instruction *I) {
1443  return isa<AllocaInst>(I->getOperand(0));
1444  }))
1445  return false;
1446 
1447  for (unsigned OI = 0, OE = I0->getNumOperands(); OI != OE; ++OI) {
1448  if (I0->getOperand(OI)->getType()->isTokenTy())
1449  // Don't touch any operand of token type.
1450  return false;
1451 
1452  auto SameAsI0 = [&I0, OI](const Instruction *I) {
1453  assert(I->getNumOperands() == I0->getNumOperands());
1454  return I->getOperand(OI) == I0->getOperand(OI);
1455  };
1456  if (!all_of(Insts, SameAsI0)) {
1457  if (!canReplaceOperandWithVariable(I0, OI))
1458  // We can't create a PHI from this GEP.
1459  return false;
1460  // Don't create indirect calls! The called value is the final operand.
1461  if ((isa<CallInst>(I0) || isa<InvokeInst>(I0)) && OI == OE - 1) {
1462  // FIXME: if the call was *already* indirect, we should do this.
1463  return false;
1464  }
1465  for (auto *I : Insts)
1466  PHIOperands[I].push_back(I->getOperand(OI));
1467  }
1468  }
1469  return true;
1470 }
1471 
1472 // Assuming canSinkLastInstruction(Blocks) has returned true, sink the last
1473 // instruction of every block in Blocks to their common successor, commoning
1474 // into one instruction.
1476  auto *BBEnd = Blocks[0]->getTerminator()->getSuccessor(0);
1477 
1478  // canSinkLastInstruction returning true guarantees that every block has at
1479  // least one non-terminator instruction.
1481  for (auto *BB : Blocks) {
1482  Instruction *I = BB->getTerminator();
1483  do {
1484  I = I->getPrevNode();
1485  } while (isa<DbgInfoIntrinsic>(I) && I != &BB->front());
1486  if (!isa<DbgInfoIntrinsic>(I))
1487  Insts.push_back(I);
1488  }
1489 
1490  // The only checking we need to do now is that all users of all instructions
1491  // are the same PHI node. canSinkLastInstruction should have checked this but
1492  // it is slightly over-aggressive - it gets confused by commutative instructions
1493  // so double-check it here.
1494  Instruction *I0 = Insts.front();
1495  if (!isa<StoreInst>(I0)) {
1496  auto *PNUse = dyn_cast<PHINode>(*I0->user_begin());
1497  if (!all_of(Insts, [&PNUse](const Instruction *I) -> bool {
1498  auto *U = cast<Instruction>(*I->user_begin());
1499  return U == PNUse;
1500  }))
1501  return false;
1502  }
1503 
1504  // We don't need to do any more checking here; canSinkLastInstruction should
1505  // have done it all for us.
1506  SmallVector<Value*, 4> NewOperands;
1507  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O) {
1508  // This check is different to that in canSinkLastInstruction. There, we
1509  // cared about the global view once simplifycfg (and instcombine) have
1510  // completed - it takes into account PHIs that become trivially
1511  // simplifiable. However here we need a more local view; if an operand
1512  // differs we create a PHI and rely on instcombine to clean up the very
1513  // small mess we may make.
1514  bool NeedPHI = any_of(Insts, [&I0, O](const Instruction *I) {
1515  return I->getOperand(O) != I0->getOperand(O);
1516  });
1517  if (!NeedPHI) {
1518  NewOperands.push_back(I0->getOperand(O));
1519  continue;
1520  }
1521 
1522  // Create a new PHI in the successor block and populate it.
1523  auto *Op = I0->getOperand(O);
1524  assert(!Op->getType()->isTokenTy() && "Can't PHI tokens!");
1525  auto *PN = PHINode::Create(Op->getType(), Insts.size(),
1526  Op->getName() + ".sink", &BBEnd->front());
1527  for (auto *I : Insts)
1528  PN->addIncoming(I->getOperand(O), I->getParent());
1529  NewOperands.push_back(PN);
1530  }
1531 
1532  // Arbitrarily use I0 as the new "common" instruction; remap its operands
1533  // and move it to the start of the successor block.
1534  for (unsigned O = 0, E = I0->getNumOperands(); O != E; ++O)
1535  I0->getOperandUse(O).set(NewOperands[O]);
1536  I0->moveBefore(&*BBEnd->getFirstInsertionPt());
1537 
1538  // The debug location for the "common" instruction is the merged locations of
1539  // all the commoned instructions. We start with the original location of the
1540  // "common" instruction and iteratively merge each location in the loop below.
1541  const DILocation *Loc = I0->getDebugLoc();
1542 
1543  // Update metadata and IR flags, and merge debug locations.
1544  for (auto *I : Insts)
1545  if (I != I0) {
1546  Loc = DILocation::getMergedLocation(Loc, I->getDebugLoc());
1547  combineMetadataForCSE(I0, I);
1548  I0->andIRFlags(I);
1549  }
1550  if (!isa<CallInst>(I0))
1551  I0->setDebugLoc(Loc);
1552 
1553  if (!isa<StoreInst>(I0)) {
1554  // canSinkLastInstruction checked that all instructions were used by
1555  // one and only one PHI node. Find that now, RAUW it to our common
1556  // instruction and nuke it.
1557  assert(I0->hasOneUse());
1558  auto *PN = cast<PHINode>(*I0->user_begin());
1559  PN->replaceAllUsesWith(I0);
1560  PN->eraseFromParent();
1561  }
1562 
1563  // Finally nuke all instructions apart from the common instruction.
1564  for (auto *I : Insts)
1565  if (I != I0)
1566  I->eraseFromParent();
1567 
1568  return true;
1569 }
1570 
1571 namespace {
1572 
1573  // LockstepReverseIterator - Iterates through instructions
1574  // in a set of blocks in reverse order from the first non-terminator.
1575  // For example (assume all blocks have size n):
1576  // LockstepReverseIterator I([B1, B2, B3]);
1577  // *I-- = [B1[n], B2[n], B3[n]];
1578  // *I-- = [B1[n-1], B2[n-1], B3[n-1]];
1579  // *I-- = [B1[n-2], B2[n-2], B3[n-2]];
1580  // ...
1581  class LockstepReverseIterator {
1582  ArrayRef<BasicBlock*> Blocks;
1584  bool Fail;
1585  public:
1586  LockstepReverseIterator(ArrayRef<BasicBlock*> Blocks) :
1587  Blocks(Blocks) {
1588  reset();
1589  }
1590 
1591  void reset() {
1592  Fail = false;
1593  Insts.clear();
1594  for (auto *BB : Blocks) {
1595  Instruction *Inst = BB->getTerminator();
1596  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1597  Inst = Inst->getPrevNode();
1598  if (!Inst) {
1599  // Block wasn't big enough.
1600  Fail = true;
1601  return;
1602  }
1603  Insts.push_back(Inst);
1604  }
1605  }
1606 
1607  bool isValid() const {
1608  return !Fail;
1609  }
1610 
1611  void operator -- () {
1612  if (Fail)
1613  return;
1614  for (auto *&Inst : Insts) {
1615  for (Inst = Inst->getPrevNode(); Inst && isa<DbgInfoIntrinsic>(Inst);)
1616  Inst = Inst->getPrevNode();
1617  // Already at beginning of block.
1618  if (!Inst) {
1619  Fail = true;
1620  return;
1621  }
1622  }
1623  }
1624 
1626  return Insts;
1627  }
1628  };
1629 
1630 } // end anonymous namespace
1631 
1632 /// Given an unconditional branch that goes to BBEnd,
1633 /// check whether BBEnd has only two predecessors and the other predecessor
1634 /// ends with an unconditional branch. If it is true, sink any common code
1635 /// in the two predecessors to BBEnd.
1637  assert(BI1->isUnconditional());
1638  BasicBlock *BBEnd = BI1->getSuccessor(0);
1639 
1640  // We support two situations:
1641  // (1) all incoming arcs are unconditional
1642  // (2) one incoming arc is conditional
1643  //
1644  // (2) is very common in switch defaults and
1645  // else-if patterns;
1646  //
1647  // if (a) f(1);
1648  // else if (b) f(2);
1649  //
1650  // produces:
1651  //
1652  // [if]
1653  // / \
1654  // [f(1)] [if]
1655  // | | \
1656  // | | |
1657  // | [f(2)]|
1658  // \ | /
1659  // [ end ]
1660  //
1661  // [end] has two unconditional predecessor arcs and one conditional. The
1662  // conditional refers to the implicit empty 'else' arc. This conditional
1663  // arc can also be caused by an empty default block in a switch.
1664  //
1665  // In this case, we attempt to sink code from all *unconditional* arcs.
1666  // If we can sink instructions from these arcs (determined during the scan
1667  // phase below) we insert a common successor for all unconditional arcs and
1668  // connect that to [end], to enable sinking:
1669  //
1670  // [if]
1671  // / \
1672  // [x(1)] [if]
1673  // | | \
1674  // | | \
1675  // | [x(2)] |
1676  // \ / |
1677  // [sink.split] |
1678  // \ /
1679  // [ end ]
1680  //
1681  SmallVector<BasicBlock*,4> UnconditionalPreds;
1682  Instruction *Cond = nullptr;
1683  for (auto *B : predecessors(BBEnd)) {
1684  auto *T = B->getTerminator();
1685  if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
1686  UnconditionalPreds.push_back(B);
1687  else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
1688  Cond = T;
1689  else
1690  return false;
1691  }
1692  if (UnconditionalPreds.size() < 2)
1693  return false;
1694 
1695  bool Changed = false;
1696  // We take a two-step approach to tail sinking. First we scan from the end of
1697  // each block upwards in lockstep. If the n'th instruction from the end of each
1698  // block can be sunk, those instructions are added to ValuesToSink and we
1699  // carry on. If we can sink an instruction but need to PHI-merge some operands
1700  // (because they're not identical in each instruction) we add these to
1701  // PHIOperands.
1702  unsigned ScanIdx = 0;
1703  SmallPtrSet<Value*,4> InstructionsToSink;
1705  LockstepReverseIterator LRI(UnconditionalPreds);
1706  while (LRI.isValid() &&
1707  canSinkInstructions(*LRI, PHIOperands)) {
1708  DEBUG(dbgs() << "SINK: instruction can be sunk: " << *(*LRI)[0] << "\n");
1709  InstructionsToSink.insert((*LRI).begin(), (*LRI).end());
1710  ++ScanIdx;
1711  --LRI;
1712  }
1713 
1714  auto ProfitableToSinkInstruction = [&](LockstepReverseIterator &LRI) {
1715  unsigned NumPHIdValues = 0;
1716  for (auto *I : *LRI)
1717  for (auto *V : PHIOperands[I])
1718  if (InstructionsToSink.count(V) == 0)
1719  ++NumPHIdValues;
1720  DEBUG(dbgs() << "SINK: #phid values: " << NumPHIdValues << "\n");
1721  unsigned NumPHIInsts = NumPHIdValues / UnconditionalPreds.size();
1722  if ((NumPHIdValues % UnconditionalPreds.size()) != 0)
1723  NumPHIInsts++;
1724 
1725  return NumPHIInsts <= 1;
1726  };
1727 
1728  if (ScanIdx > 0 && Cond) {
1729  // Check if we would actually sink anything first! This mutates the CFG and
1730  // adds an extra block. The goal in doing this is to allow instructions that
1731  // couldn't be sunk before to be sunk - obviously, speculatable instructions
1732  // (such as trunc, add) can be sunk and predicated already. So we check that
1733  // we're going to sink at least one non-speculatable instruction.
1734  LRI.reset();
1735  unsigned Idx = 0;
1736  bool Profitable = false;
1737  while (ProfitableToSinkInstruction(LRI) && Idx < ScanIdx) {
1738  if (!isSafeToSpeculativelyExecute((*LRI)[0])) {
1739  Profitable = true;
1740  break;
1741  }
1742  --LRI;
1743  ++Idx;
1744  }
1745  if (!Profitable)
1746  return false;
1747 
1748  DEBUG(dbgs() << "SINK: Splitting edge\n");
1749  // We have a conditional edge and we're going to sink some instructions.
1750  // Insert a new block postdominating all blocks we're going to sink from.
1751  if (!SplitBlockPredecessors(BI1->getSuccessor(0), UnconditionalPreds,
1752  ".sink.split"))
1753  // Edges couldn't be split.
1754  return false;
1755  Changed = true;
1756  }
1757 
1758  // Now that we've analyzed all potential sinking candidates, perform the
1759  // actual sink. We iteratively sink the last non-terminator of the source
1760  // blocks into their common successor unless doing so would require too
1761  // many PHI instructions to be generated (currently only one PHI is allowed
1762  // per sunk instruction).
1763  //
1764  // We can use InstructionsToSink to discount values needing PHI-merging that will
1765  // actually be sunk in a later iteration. This allows us to be more
1766  // aggressive in what we sink. This does allow a false positive where we
1767  // sink presuming a later value will also be sunk, but stop half way through
1768  // and never actually sink it which means we produce more PHIs than intended.
1769  // This is unlikely in practice though.
1770  for (unsigned SinkIdx = 0; SinkIdx != ScanIdx; ++SinkIdx) {
1771  DEBUG(dbgs() << "SINK: Sink: "
1772  << *UnconditionalPreds[0]->getTerminator()->getPrevNode()
1773  << "\n");
1774 
1775  // Because we've sunk every instruction in turn, the current instruction to
1776  // sink is always at index 0.
1777  LRI.reset();
1778  if (!ProfitableToSinkInstruction(LRI)) {
1779  // Too many PHIs would be created.
1780  DEBUG(dbgs() << "SINK: stopping here, too many PHIs would be created!\n");
1781  break;
1782  }
1783 
1784  if (!sinkLastInstruction(UnconditionalPreds))
1785  return Changed;
1786  NumSinkCommons++;
1787  Changed = true;
1788  }
1789  return Changed;
1790 }
1791 
1792 /// \brief Determine if we can hoist sink a sole store instruction out of a
1793 /// conditional block.
1794 ///
1795 /// We are looking for code like the following:
1796 /// BrBB:
1797 /// store i32 %add, i32* %arrayidx2
1798 /// ... // No other stores or function calls (we could be calling a memory
1799 /// ... // function).
1800 /// %cmp = icmp ult %x, %y
1801 /// br i1 %cmp, label %EndBB, label %ThenBB
1802 /// ThenBB:
1803 /// store i32 %add5, i32* %arrayidx2
1804 /// br label EndBB
1805 /// EndBB:
1806 /// ...
1807 /// We are going to transform this into:
1808 /// BrBB:
1809 /// store i32 %add, i32* %arrayidx2
1810 /// ... //
1811 /// %cmp = icmp ult %x, %y
1812 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1813 /// store i32 %add.add5, i32* %arrayidx2
1814 /// ...
1815 ///
1816 /// \return The pointer to the value of the previous store if the store can be
1817 /// hoisted into the predecessor block. 0 otherwise.
1819  BasicBlock *StoreBB, BasicBlock *EndBB) {
1820  StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1821  if (!StoreToHoist)
1822  return nullptr;
1823 
1824  // Volatile or atomic.
1825  if (!StoreToHoist->isSimple())
1826  return nullptr;
1827 
1828  Value *StorePtr = StoreToHoist->getPointerOperand();
1829 
1830  // Look for a store to the same pointer in BrBB.
1831  unsigned MaxNumInstToLookAt = 9;
1832  for (Instruction &CurI : reverse(*BrBB)) {
1833  if (!MaxNumInstToLookAt)
1834  break;
1835  // Skip debug info.
1836  if (isa<DbgInfoIntrinsic>(CurI))
1837  continue;
1838  --MaxNumInstToLookAt;
1839 
1840  // Could be calling an instruction that affects memory like free().
1841  if (CurI.mayHaveSideEffects() && !isa<StoreInst>(CurI))
1842  return nullptr;
1843 
1844  if (auto *SI = dyn_cast<StoreInst>(&CurI)) {
1845  // Found the previous store make sure it stores to the same location.
1846  if (SI->getPointerOperand() == StorePtr)
1847  // Found the previous store, return its value operand.
1848  return SI->getValueOperand();
1849  return nullptr; // Unknown store.
1850  }
1851  }
1852 
1853  return nullptr;
1854 }
1855 
1856 /// \brief Speculate a conditional basic block flattening the CFG.
1857 ///
1858 /// Note that this is a very risky transform currently. Speculating
1859 /// instructions like this is most often not desirable. Instead, there is an MI
1860 /// pass which can do it with full awareness of the resource constraints.
1861 /// However, some cases are "obvious" and we should do directly. An example of
1862 /// this is speculating a single, reasonably cheap instruction.
1863 ///
1864 /// There is only one distinct advantage to flattening the CFG at the IR level:
1865 /// it makes very common but simplistic optimizations such as are common in
1866 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1867 /// modeling their effects with easier to reason about SSA value graphs.
1868 ///
1869 ///
1870 /// An illustration of this transform is turning this IR:
1871 /// \code
1872 /// BB:
1873 /// %cmp = icmp ult %x, %y
1874 /// br i1 %cmp, label %EndBB, label %ThenBB
1875 /// ThenBB:
1876 /// %sub = sub %x, %y
1877 /// br label BB2
1878 /// EndBB:
1879 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1880 /// ...
1881 /// \endcode
1882 ///
1883 /// Into this IR:
1884 /// \code
1885 /// BB:
1886 /// %cmp = icmp ult %x, %y
1887 /// %sub = sub %x, %y
1888 /// %cond = select i1 %cmp, 0, %sub
1889 /// ...
1890 /// \endcode
1891 ///
1892 /// \returns true if the conditional block is removed.
1894  const TargetTransformInfo &TTI) {
1895  // Be conservative for now. FP select instruction can often be expensive.
1896  Value *BrCond = BI->getCondition();
1897  if (isa<FCmpInst>(BrCond))
1898  return false;
1899 
1900  BasicBlock *BB = BI->getParent();
1901  BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1902 
1903  // If ThenBB is actually on the false edge of the conditional branch, remember
1904  // to swap the select operands later.
1905  bool Invert = false;
1906  if (ThenBB != BI->getSuccessor(0)) {
1907  assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1908  Invert = true;
1909  }
1910  assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1911 
1912  // Keep a count of how many times instructions are used within CondBB when
1913  // they are candidates for sinking into CondBB. Specifically:
1914  // - They are defined in BB, and
1915  // - They have no side effects, and
1916  // - All of their uses are in CondBB.
1917  SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1918 
1919  unsigned SpeculationCost = 0;
1920  Value *SpeculatedStoreValue = nullptr;
1921  StoreInst *SpeculatedStore = nullptr;
1922  for (BasicBlock::iterator BBI = ThenBB->begin(),
1923  BBE = std::prev(ThenBB->end());
1924  BBI != BBE; ++BBI) {
1925  Instruction *I = &*BBI;
1926  // Skip debug info.
1927  if (isa<DbgInfoIntrinsic>(I))
1928  continue;
1929 
1930  // Only speculatively execute a single instruction (not counting the
1931  // terminator) for now.
1932  ++SpeculationCost;
1933  if (SpeculationCost > 1)
1934  return false;
1935 
1936  // Don't hoist the instruction if it's unsafe or expensive.
1937  if (!isSafeToSpeculativelyExecute(I) &&
1938  !(HoistCondStores && (SpeculatedStoreValue = isSafeToSpeculateStore(
1939  I, BB, ThenBB, EndBB))))
1940  return false;
1941  if (!SpeculatedStoreValue &&
1942  ComputeSpeculationCost(I, TTI) >
1944  return false;
1945 
1946  // Store the store speculation candidate.
1947  if (SpeculatedStoreValue)
1948  SpeculatedStore = cast<StoreInst>(I);
1949 
1950  // Do not hoist the instruction if any of its operands are defined but not
1951  // used in BB. The transformation will prevent the operand from
1952  // being sunk into the use block.
1953  for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) {
1954  Instruction *OpI = dyn_cast<Instruction>(*i);
1955  if (!OpI || OpI->getParent() != BB || OpI->mayHaveSideEffects())
1956  continue; // Not a candidate for sinking.
1957 
1958  ++SinkCandidateUseCounts[OpI];
1959  }
1960  }
1961 
1962  // Consider any sink candidates which are only used in CondBB as costs for
1963  // speculation. Note, while we iterate over a DenseMap here, we are summing
1964  // and so iteration order isn't significant.
1966  I = SinkCandidateUseCounts.begin(),
1967  E = SinkCandidateUseCounts.end();
1968  I != E; ++I)
1969  if (I->first->getNumUses() == I->second) {
1970  ++SpeculationCost;
1971  if (SpeculationCost > 1)
1972  return false;
1973  }
1974 
1975  // Check that the PHI nodes can be converted to selects.
1976  bool HaveRewritablePHIs = false;
1977  for (BasicBlock::iterator I = EndBB->begin();
1978  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1979  Value *OrigV = PN->getIncomingValueForBlock(BB);
1980  Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1981 
1982  // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1983  // Skip PHIs which are trivial.
1984  if (ThenV == OrigV)
1985  continue;
1986 
1987  // Don't convert to selects if we could remove undefined behavior instead.
1988  if (passingValueIsAlwaysUndefined(OrigV, PN) ||
1989  passingValueIsAlwaysUndefined(ThenV, PN))
1990  return false;
1991 
1992  HaveRewritablePHIs = true;
1993  ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1994  ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1995  if (!OrigCE && !ThenCE)
1996  continue; // Known safe and cheap.
1997 
1998  if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1999  (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
2000  return false;
2001  unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE, TTI) : 0;
2002  unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE, TTI) : 0;
2003  unsigned MaxCost =
2005  if (OrigCost + ThenCost > MaxCost)
2006  return false;
2007 
2008  // Account for the cost of an unfolded ConstantExpr which could end up
2009  // getting expanded into Instructions.
2010  // FIXME: This doesn't account for how many operations are combined in the
2011  // constant expression.
2012  ++SpeculationCost;
2013  if (SpeculationCost > 1)
2014  return false;
2015  }
2016 
2017  // If there are no PHIs to process, bail early. This helps ensure idempotence
2018  // as well.
2019  if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
2020  return false;
2021 
2022  // If we get here, we can hoist the instruction and if-convert.
2023  DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
2024 
2025  // Insert a select of the value of the speculated store.
2026  if (SpeculatedStoreValue) {
2027  IRBuilder<NoFolder> Builder(BI);
2028  Value *TrueV = SpeculatedStore->getValueOperand();
2029  Value *FalseV = SpeculatedStoreValue;
2030  if (Invert)
2031  std::swap(TrueV, FalseV);
2032  Value *S = Builder.CreateSelect(
2033  BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2034  SpeculatedStore->setOperand(0, S);
2035  SpeculatedStore->setDebugLoc(
2037  BI->getDebugLoc(), SpeculatedStore->getDebugLoc()));
2038  }
2039 
2040  // Metadata can be dependent on the condition we are hoisting above.
2041  // Conservatively strip all metadata on the instruction.
2042  for (auto &I : *ThenBB)
2044 
2045  // Hoist the instructions.
2046  BB->getInstList().splice(BI->getIterator(), ThenBB->getInstList(),
2047  ThenBB->begin(), std::prev(ThenBB->end()));
2048 
2049  // Insert selects and rewrite the PHI operands.
2050  IRBuilder<NoFolder> Builder(BI);
2051  for (BasicBlock::iterator I = EndBB->begin();
2052  PHINode *PN = dyn_cast<PHINode>(I); ++I) {
2053  unsigned OrigI = PN->getBasicBlockIndex(BB);
2054  unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
2055  Value *OrigV = PN->getIncomingValue(OrigI);
2056  Value *ThenV = PN->getIncomingValue(ThenI);
2057 
2058  // Skip PHIs which are trivial.
2059  if (OrigV == ThenV)
2060  continue;
2061 
2062  // Create a select whose true value is the speculatively executed value and
2063  // false value is the preexisting value. Swap them if the branch
2064  // destinations were inverted.
2065  Value *TrueV = ThenV, *FalseV = OrigV;
2066  if (Invert)
2067  std::swap(TrueV, FalseV);
2068  Value *V = Builder.CreateSelect(
2069  BrCond, TrueV, FalseV, TrueV->getName() + "." + FalseV->getName(), BI);
2070  PN->setIncomingValue(OrigI, V);
2071  PN->setIncomingValue(ThenI, V);
2072  }
2073 
2074  ++NumSpeculations;
2075  return true;
2076 }
2077 
2078 /// Return true if we can thread a branch across this block.
2080  BranchInst *BI = cast<BranchInst>(BB->getTerminator());
2081  unsigned Size = 0;
2082 
2083  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2084  if (isa<DbgInfoIntrinsic>(BBI))
2085  continue;
2086  if (Size > 10)
2087  return false; // Don't clone large BB's.
2088  ++Size;
2089 
2090  // We can only support instructions that do not define values that are
2091  // live outside of the current basic block.
2092  for (User *U : BBI->users()) {
2093  Instruction *UI = cast<Instruction>(U);
2094  if (UI->getParent() != BB || isa<PHINode>(UI))
2095  return false;
2096  }
2097 
2098  // Looks ok, continue checking.
2099  }
2100 
2101  return true;
2102 }
2103 
2104 /// If we have a conditional branch on a PHI node value that is defined in the
2105 /// same block as the branch and if any PHI entries are constants, thread edges
2106 /// corresponding to that entry to be branches to their ultimate destination.
2107 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout &DL,
2108  AssumptionCache *AC) {
2109  BasicBlock *BB = BI->getParent();
2110  PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
2111  // NOTE: we currently cannot transform this case if the PHI node is used
2112  // outside of the block.
2113  if (!PN || PN->getParent() != BB || !PN->hasOneUse())
2114  return false;
2115 
2116  // Degenerate case of a single entry PHI.
2117  if (PN->getNumIncomingValues() == 1) {
2119  return true;
2120  }
2121 
2122  // Now we know that this block has multiple preds and two succs.
2124  return false;
2125 
2126  // Can't fold blocks that contain noduplicate or convergent calls.
2127  if (any_of(*BB, [](const Instruction &I) {
2128  const CallInst *CI = dyn_cast<CallInst>(&I);
2129  return CI && (CI->cannotDuplicate() || CI->isConvergent());
2130  }))
2131  return false;
2132 
2133  // Okay, this is a simple enough basic block. See if any phi values are
2134  // constants.
2135  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
2137  if (!CB || !CB->getType()->isIntegerTy(1))
2138  continue;
2139 
2140  // Okay, we now know that all edges from PredBB should be revectored to
2141  // branch to RealDest.
2142  BasicBlock *PredBB = PN->getIncomingBlock(i);
2143  BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
2144 
2145  if (RealDest == BB)
2146  continue; // Skip self loops.
2147  // Skip if the predecessor's terminator is an indirect branch.
2148  if (isa<IndirectBrInst>(PredBB->getTerminator()))
2149  continue;
2150 
2151  // The dest block might have PHI nodes, other predecessors and other
2152  // difficult cases. Instead of being smart about this, just insert a new
2153  // block that jumps to the destination block, effectively splitting
2154  // the edge we are about to create.
2155  BasicBlock *EdgeBB =
2156  BasicBlock::Create(BB->getContext(), RealDest->getName() + ".critedge",
2157  RealDest->getParent(), RealDest);
2158  BranchInst::Create(RealDest, EdgeBB);
2159 
2160  // Update PHI nodes.
2161  AddPredecessorToBlock(RealDest, EdgeBB, BB);
2162 
2163  // BB may have instructions that are being threaded over. Clone these
2164  // instructions into EdgeBB. We know that there will be no uses of the
2165  // cloned instructions outside of EdgeBB.
2166  BasicBlock::iterator InsertPt = EdgeBB->begin();
2167  DenseMap<Value *, Value *> TranslateMap; // Track translated values.
2168  for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
2169  if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
2170  TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
2171  continue;
2172  }
2173  // Clone the instruction.
2174  Instruction *N = BBI->clone();
2175  if (BBI->hasName())
2176  N->setName(BBI->getName() + ".c");
2177 
2178  // Update operands due to translation.
2179  for (User::op_iterator i = N->op_begin(), e = N->op_end(); i != e; ++i) {
2180  DenseMap<Value *, Value *>::iterator PI = TranslateMap.find(*i);
2181  if (PI != TranslateMap.end())
2182  *i = PI->second;
2183  }
2184 
2185  // Check for trivial simplification.
2186  if (Value *V = SimplifyInstruction(N, {DL, nullptr, nullptr, AC})) {
2187  if (!BBI->use_empty())
2188  TranslateMap[&*BBI] = V;
2189  if (!N->mayHaveSideEffects()) {
2190  N->deleteValue(); // Instruction folded away, don't need actual inst
2191  N = nullptr;
2192  }
2193  } else {
2194  if (!BBI->use_empty())
2195  TranslateMap[&*BBI] = N;
2196  }
2197  // Insert the new instruction into its new home.
2198  if (N)
2199  EdgeBB->getInstList().insert(InsertPt, N);
2200 
2201  // Register the new instruction with the assumption cache if necessary.
2202  if (auto *II = dyn_cast_or_null<IntrinsicInst>(N))
2203  if (II->getIntrinsicID() == Intrinsic::assume)
2204  AC->registerAssumption(II);
2205  }
2206 
2207  // Loop over all of the edges from PredBB to BB, changing them to branch
2208  // to EdgeBB instead.
2209  TerminatorInst *PredBBTI = PredBB->getTerminator();
2210  for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
2211  if (PredBBTI->getSuccessor(i) == BB) {
2212  BB->removePredecessor(PredBB);
2213  PredBBTI->setSuccessor(i, EdgeBB);
2214  }
2215 
2216  // Recurse, simplifying any other constants.
2217  return FoldCondBranchOnPHI(BI, DL, AC) | true;
2218  }
2219 
2220  return false;
2221 }
2222 
2223 /// Given a BB that starts with the specified two-entry PHI node,
2224 /// see if we can eliminate it.
2226  const DataLayout &DL) {
2227  // Ok, this is a two entry PHI node. Check to see if this is a simple "if
2228  // statement", which has a very simple dominance structure. Basically, we
2229  // are trying to find the condition that is being branched on, which
2230  // subsequently causes this merge to happen. We really want control
2231  // dependence information for this check, but simplifycfg can't keep it up
2232  // to date, and this catches most of the cases we care about anyway.
2233  BasicBlock *BB = PN->getParent();
2234  BasicBlock *IfTrue, *IfFalse;
2235  Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
2236  if (!IfCond ||
2237  // Don't bother if the branch will be constant folded trivially.
2238  isa<ConstantInt>(IfCond))
2239  return false;
2240 
2241  // Okay, we found that we can merge this two-entry phi node into a select.
2242  // Doing so would require us to fold *all* two entry phi nodes in this block.
2243  // At some point this becomes non-profitable (particularly if the target
2244  // doesn't support cmov's). Only do this transformation if there are two or
2245  // fewer PHI nodes in this block.
2246  unsigned NumPhis = 0;
2247  for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
2248  if (NumPhis > 2)
2249  return false;
2250 
2251  // Loop over the PHI's seeing if we can promote them all to select
2252  // instructions. While we are at it, keep track of the instructions
2253  // that need to be moved to the dominating block.
2254  SmallPtrSet<Instruction *, 4> AggressiveInsts;
2255  unsigned MaxCostVal0 = PHINodeFoldingThreshold,
2256  MaxCostVal1 = PHINodeFoldingThreshold;
2257  MaxCostVal0 *= TargetTransformInfo::TCC_Basic;
2258  MaxCostVal1 *= TargetTransformInfo::TCC_Basic;
2259 
2260  for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
2261  PHINode *PN = cast<PHINode>(II++);
2262  if (Value *V = SimplifyInstruction(PN, {DL, PN})) {
2263  PN->replaceAllUsesWith(V);
2264  PN->eraseFromParent();
2265  continue;
2266  }
2267 
2268  if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
2269  MaxCostVal0, TTI) ||
2270  !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
2271  MaxCostVal1, TTI))
2272  return false;
2273  }
2274 
2275  // If we folded the first phi, PN dangles at this point. Refresh it. If
2276  // we ran out of PHIs then we simplified them all.
2277  PN = dyn_cast<PHINode>(BB->begin());
2278  if (!PN)
2279  return true;
2280 
2281  // Don't fold i1 branches on PHIs which contain binary operators. These can
2282  // often be turned into switches and other things.
2283  if (PN->getType()->isIntegerTy(1) &&
2284  (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
2285  isa<BinaryOperator>(PN->getIncomingValue(1)) ||
2286  isa<BinaryOperator>(IfCond)))
2287  return false;
2288 
2289  // If all PHI nodes are promotable, check to make sure that all instructions
2290  // in the predecessor blocks can be promoted as well. If not, we won't be able
2291  // to get rid of the control flow, so it's not worth promoting to select
2292  // instructions.
2293  BasicBlock *DomBlock = nullptr;
2294  BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
2295  BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
2296  if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
2297  IfBlock1 = nullptr;
2298  } else {
2299  DomBlock = *pred_begin(IfBlock1);
2300  for (BasicBlock::iterator I = IfBlock1->begin(); !isa<TerminatorInst>(I);
2301  ++I)
2302  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2303  // This is not an aggressive instruction that we can promote.
2304  // Because of this, we won't be able to get rid of the control flow, so
2305  // the xform is not worth it.
2306  return false;
2307  }
2308  }
2309 
2310  if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
2311  IfBlock2 = nullptr;
2312  } else {
2313  DomBlock = *pred_begin(IfBlock2);
2314  for (BasicBlock::iterator I = IfBlock2->begin(); !isa<TerminatorInst>(I);
2315  ++I)
2316  if (!AggressiveInsts.count(&*I) && !isa<DbgInfoIntrinsic>(I)) {
2317  // This is not an aggressive instruction that we can promote.
2318  // Because of this, we won't be able to get rid of the control flow, so
2319  // the xform is not worth it.
2320  return false;
2321  }
2322  }
2323 
2324  DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
2325  << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
2326 
2327  // If we can still promote the PHI nodes after this gauntlet of tests,
2328  // do all of the PHI's now.
2329  Instruction *InsertPt = DomBlock->getTerminator();
2330  IRBuilder<NoFolder> Builder(InsertPt);
2331 
2332  // Move all 'aggressive' instructions, which are defined in the
2333  // conditional parts of the if's up to the dominating block.
2334  if (IfBlock1) {
2335  for (auto &I : *IfBlock1)
2337  DomBlock->getInstList().splice(InsertPt->getIterator(),
2338  IfBlock1->getInstList(), IfBlock1->begin(),
2339  IfBlock1->getTerminator()->getIterator());
2340  }
2341  if (IfBlock2) {
2342  for (auto &I : *IfBlock2)
2344  DomBlock->getInstList().splice(InsertPt->getIterator(),
2345  IfBlock2->getInstList(), IfBlock2->begin(),
2346  IfBlock2->getTerminator()->getIterator());
2347  }
2348 
2349  while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
2350  // Change the PHI node into a select instruction.
2351  Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
2352  Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
2353 
2354  Value *Sel = Builder.CreateSelect(IfCond, TrueVal, FalseVal, "", InsertPt);
2355  PN->replaceAllUsesWith(Sel);
2356  Sel->takeName(PN);
2357  PN->eraseFromParent();
2358  }
2359 
2360  // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
2361  // has been flattened. Change DomBlock to jump directly to our new block to
2362  // avoid other simplifycfg's kicking in on the diamond.
2363  TerminatorInst *OldTI = DomBlock->getTerminator();
2364  Builder.SetInsertPoint(OldTI);
2365  Builder.CreateBr(BB);
2366  OldTI->eraseFromParent();
2367  return true;
2368 }
2369 
2370 /// If we found a conditional branch that goes to two returning blocks,
2371 /// try to merge them together into one return,
2372 /// introducing a select if the return values disagree.
2374  IRBuilder<> &Builder) {
2375  assert(BI->isConditional() && "Must be a conditional branch");
2376  BasicBlock *TrueSucc = BI->getSuccessor(0);
2377  BasicBlock *FalseSucc = BI->getSuccessor(1);
2378  ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
2379  ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
2380 
2381  // Check to ensure both blocks are empty (just a return) or optionally empty
2382  // with PHI nodes. If there are other instructions, merging would cause extra
2383  // computation on one path or the other.
2384  if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
2385  return false;
2386  if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
2387  return false;
2388 
2389  Builder.SetInsertPoint(BI);
2390  // Okay, we found a branch that is going to two return nodes. If
2391  // there is no return value for this function, just change the
2392  // branch into a return.
2393  if (FalseRet->getNumOperands() == 0) {
2394  TrueSucc->removePredecessor(BI->getParent());
2395  FalseSucc->removePredecessor(BI->getParent());
2396  Builder.CreateRetVoid();
2398  return true;
2399  }
2400 
2401  // Otherwise, figure out what the true and false return values are
2402  // so we can insert a new select instruction.
2403  Value *TrueValue = TrueRet->getReturnValue();
2404  Value *FalseValue = FalseRet->getReturnValue();
2405 
2406  // Unwrap any PHI nodes in the return blocks.
2407  if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
2408  if (TVPN->getParent() == TrueSucc)
2409  TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
2410  if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
2411  if (FVPN->getParent() == FalseSucc)
2412  FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
2413 
2414  // In order for this transformation to be safe, we must be able to
2415  // unconditionally execute both operands to the return. This is
2416  // normally the case, but we could have a potentially-trapping
2417  // constant expression that prevents this transformation from being
2418  // safe.
2419  if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
2420  if (TCV->canTrap())
2421  return false;
2422  if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
2423  if (FCV->canTrap())
2424  return false;
2425 
2426  // Okay, we collected all the mapped values and checked them for sanity, and
2427  // defined to really do this transformation. First, update the CFG.
2428  TrueSucc->removePredecessor(BI->getParent());
2429  FalseSucc->removePredecessor(BI->getParent());
2430 
2431  // Insert select instructions where needed.
2432  Value *BrCond = BI->getCondition();
2433  if (TrueValue) {
2434  // Insert a select if the results differ.
2435  if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2436  } else if (isa<UndefValue>(TrueValue)) {
2437  TrueValue = FalseValue;
2438  } else {
2439  TrueValue =
2440  Builder.CreateSelect(BrCond, TrueValue, FalseValue, "retval", BI);
2441  }
2442  }
2443 
2444  Value *RI =
2445  !TrueValue ? Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2446 
2447  (void)RI;
2448 
2449  DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2450  << "\n " << *BI << "NewRet = " << *RI
2451  << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: " << *FalseSucc);
2452 
2454 
2455  return true;
2456 }
2457 
2458 /// Return true if the given instruction is available
2459 /// in its predecessor block. If yes, the instruction will be removed.
2461  if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2462  return false;
2463  for (Instruction &I : *PB) {
2464  Instruction *PBI = &I;
2465  // Check whether Inst and PBI generate the same value.
2466  if (Inst->isIdenticalTo(PBI)) {
2467  Inst->replaceAllUsesWith(PBI);
2468  Inst->eraseFromParent();
2469  return true;
2470  }
2471  }
2472  return false;
2473 }
2474 
2475 /// Return true if either PBI or BI has branch weight available, and store
2476 /// the weights in {Pred|Succ}{True|False}Weight. If one of PBI and BI does
2477 /// not have branch weight, use 1:1 as its weight.
2479  uint64_t &PredTrueWeight,
2480  uint64_t &PredFalseWeight,
2481  uint64_t &SuccTrueWeight,
2482  uint64_t &SuccFalseWeight) {
2483  bool PredHasWeights =
2484  PBI->extractProfMetadata(PredTrueWeight, PredFalseWeight);
2485  bool SuccHasWeights =
2486  BI->extractProfMetadata(SuccTrueWeight, SuccFalseWeight);
2487  if (PredHasWeights || SuccHasWeights) {
2488  if (!PredHasWeights)
2489  PredTrueWeight = PredFalseWeight = 1;
2490  if (!SuccHasWeights)
2491  SuccTrueWeight = SuccFalseWeight = 1;
2492  return true;
2493  } else {
2494  return false;
2495  }
2496 }
2497 
2498 /// If this basic block is simple enough, and if a predecessor branches to us
2499 /// and one of our successors, fold the block into the predecessor and use
2500 /// logical operations to pick the right destination.
2501 bool llvm::FoldBranchToCommonDest(BranchInst *BI, unsigned BonusInstThreshold) {
2502  BasicBlock *BB = BI->getParent();
2503 
2504  Instruction *Cond = nullptr;
2505  if (BI->isConditional())
2506  Cond = dyn_cast<Instruction>(BI->getCondition());
2507  else {
2508  // For unconditional branch, check for a simple CFG pattern, where
2509  // BB has a single predecessor and BB's successor is also its predecessor's
2510  // successor. If such pattern exisits, check for CSE between BB and its
2511  // predecessor.
2512  if (BasicBlock *PB = BB->getSinglePredecessor())
2513  if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2514  if (PBI->isConditional() &&
2515  (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2516  BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2517  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2518  Instruction *Curr = &*I++;
2519  if (isa<CmpInst>(Curr)) {
2520  Cond = Curr;
2521  break;
2522  }
2523  // Quit if we can't remove this instruction.
2524  if (!checkCSEInPredecessor(Curr, PB))
2525  return false;
2526  }
2527  }
2528 
2529  if (!Cond)
2530  return false;
2531  }
2532 
2533  if (!Cond || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2534  Cond->getParent() != BB || !Cond->hasOneUse())
2535  return false;
2536 
2537  // Make sure the instruction after the condition is the cond branch.
2538  BasicBlock::iterator CondIt = ++Cond->getIterator();
2539 
2540  // Ignore dbg intrinsics.
2541  while (isa<DbgInfoIntrinsic>(CondIt))
2542  ++CondIt;
2543 
2544  if (&*CondIt != BI)
2545  return false;
2546 
2547  // Only allow this transformation if computing the condition doesn't involve
2548  // too many instructions and these involved instructions can be executed
2549  // unconditionally. We denote all involved instructions except the condition
2550  // as "bonus instructions", and only allow this transformation when the
2551  // number of the bonus instructions does not exceed a certain threshold.
2552  unsigned NumBonusInsts = 0;
2553  for (auto I = BB->begin(); Cond != &*I; ++I) {
2554  // Ignore dbg intrinsics.
2555  if (isa<DbgInfoIntrinsic>(I))
2556  continue;
2557  if (!I->hasOneUse() || !isSafeToSpeculativelyExecute(&*I))
2558  return false;
2559  // I has only one use and can be executed unconditionally.
2561  if (User == nullptr || User->getParent() != BB)
2562  return false;
2563  // I is used in the same BB. Since BI uses Cond and doesn't have more slots
2564  // to use any other instruction, User must be an instruction between next(I)
2565  // and Cond.
2566  ++NumBonusInsts;
2567  // Early exits once we reach the limit.
2568  if (NumBonusInsts > BonusInstThreshold)
2569  return false;
2570  }
2571 
2572  // Cond is known to be a compare or binary operator. Check to make sure that
2573  // neither operand is a potentially-trapping constant expression.
2574  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2575  if (CE->canTrap())
2576  return false;
2577  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2578  if (CE->canTrap())
2579  return false;
2580 
2581  // Finally, don't infinitely unroll conditional loops.
2582  BasicBlock *TrueDest = BI->getSuccessor(0);
2583  BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : nullptr;
2584  if (TrueDest == BB || FalseDest == BB)
2585  return false;
2586 
2587  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2588  BasicBlock *PredBlock = *PI;
2589  BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2590 
2591  // Check that we have two conditional branches. If there is a PHI node in
2592  // the common successor, verify that the same value flows in from both
2593  // blocks.
2595  if (!PBI || PBI->isUnconditional() ||
2596  (BI->isConditional() && !SafeToMergeTerminators(BI, PBI)) ||
2597  (!BI->isConditional() &&
2598  !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2599  continue;
2600 
2601  // Determine if the two branches share a common destination.
2602  Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2603  bool InvertPredCond = false;
2604 
2605  if (BI->isConditional()) {
2606  if (PBI->getSuccessor(0) == TrueDest) {
2607  Opc = Instruction::Or;
2608  } else if (PBI->getSuccessor(1) == FalseDest) {
2609  Opc = Instruction::And;
2610  } else if (PBI->getSuccessor(0) == FalseDest) {
2611  Opc = Instruction::And;
2612  InvertPredCond = true;
2613  } else if (PBI->getSuccessor(1) == TrueDest) {
2614  Opc = Instruction::Or;
2615  InvertPredCond = true;
2616  } else {
2617  continue;
2618  }
2619  } else {
2620  if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2621  continue;
2622  }
2623 
2624  DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2625  IRBuilder<> Builder(PBI);
2626 
2627  // If we need to invert the condition in the pred block to match, do so now.
2628  if (InvertPredCond) {
2629  Value *NewCond = PBI->getCondition();
2630 
2631  if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2632  CmpInst *CI = cast<CmpInst>(NewCond);
2633  CI->setPredicate(CI->getInversePredicate());
2634  } else {
2635  NewCond =
2636  Builder.CreateNot(NewCond, PBI->getCondition()->getName() + ".not");
2637  }
2638 
2639  PBI->setCondition(NewCond);
2640  PBI->swapSuccessors();
2641  }
2642 
2643  // If we have bonus instructions, clone them into the predecessor block.
2644  // Note that there may be multiple predecessor blocks, so we cannot move
2645  // bonus instructions to a predecessor block.
2646  ValueToValueMapTy VMap; // maps original values to cloned values
2647  // We already make sure Cond is the last instruction before BI. Therefore,
2648  // all instructions before Cond other than DbgInfoIntrinsic are bonus
2649  // instructions.
2650  for (auto BonusInst = BB->begin(); Cond != &*BonusInst; ++BonusInst) {
2651  if (isa<DbgInfoIntrinsic>(BonusInst))
2652  continue;
2653  Instruction *NewBonusInst = BonusInst->clone();
2654  RemapInstruction(NewBonusInst, VMap,
2656  VMap[&*BonusInst] = NewBonusInst;
2657 
2658  // If we moved a load, we cannot any longer claim any knowledge about
2659  // its potential value. The previous information might have been valid
2660  // only given the branch precondition.
2661  // For an analogous reason, we must also drop all the metadata whose
2662  // semantics we don't understand.
2663  NewBonusInst->dropUnknownNonDebugMetadata();
2664 
2665  PredBlock->getInstList().insert(PBI->getIterator(), NewBonusInst);
2666  NewBonusInst->takeName(&*BonusInst);
2667  BonusInst->setName(BonusInst->getName() + ".old");
2668  }
2669 
2670  // Clone Cond into the predecessor basic block, and or/and the
2671  // two conditions together.
2672  Instruction *New = Cond->clone();
2673  RemapInstruction(New, VMap,
2675  PredBlock->getInstList().insert(PBI->getIterator(), New);
2676  New->takeName(Cond);
2677  Cond->setName(New->getName() + ".old");
2678 
2679  if (BI->isConditional()) {
2680  Instruction *NewCond = cast<Instruction>(
2681  Builder.CreateBinOp(Opc, PBI->getCondition(), New, "or.cond"));
2682  PBI->setCondition(NewCond);
2683 
2684  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2685  bool HasWeights =
2686  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
2687  SuccTrueWeight, SuccFalseWeight);
2688  SmallVector<uint64_t, 8> NewWeights;
2689 
2690  if (PBI->getSuccessor(0) == BB) {
2691  if (HasWeights) {
2692  // PBI: br i1 %x, BB, FalseDest
2693  // BI: br i1 %y, TrueDest, FalseDest
2694  // TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2695  NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2696  // FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2697  // TrueWeight for PBI * FalseWeight for BI.
2698  // We assume that total weights of a BranchInst can fit into 32 bits.
2699  // Therefore, we will not have overflow using 64-bit arithmetic.
2700  NewWeights.push_back(PredFalseWeight *
2701  (SuccFalseWeight + SuccTrueWeight) +
2702  PredTrueWeight * SuccFalseWeight);
2703  }
2704  AddPredecessorToBlock(TrueDest, PredBlock, BB);
2705  PBI->setSuccessor(0, TrueDest);
2706  }
2707  if (PBI->getSuccessor(1) == BB) {
2708  if (HasWeights) {
2709  // PBI: br i1 %x, TrueDest, BB
2710  // BI: br i1 %y, TrueDest, FalseDest
2711  // TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2712  // FalseWeight for PBI * TrueWeight for BI.
2713  NewWeights.push_back(PredTrueWeight *
2714  (SuccFalseWeight + SuccTrueWeight) +
2715  PredFalseWeight * SuccTrueWeight);
2716  // FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2717  NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2718  }
2719  AddPredecessorToBlock(FalseDest, PredBlock, BB);
2720  PBI->setSuccessor(1, FalseDest);
2721  }
2722  if (NewWeights.size() == 2) {
2723  // Halve the weights if any of them cannot fit in an uint32_t
2724  FitWeights(NewWeights);
2725 
2726  SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),
2727  NewWeights.end());
2728  PBI->setMetadata(
2730  MDBuilder(BI->getContext()).createBranchWeights(MDWeights));
2731  } else
2732  PBI->setMetadata(LLVMContext::MD_prof, nullptr);
2733  } else {
2734  // Update PHI nodes in the common successors.
2735  for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2736  ConstantInt *PBI_C = cast<ConstantInt>(
2737  PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2738  assert(PBI_C->getType()->isIntegerTy(1));
2739  Instruction *MergedCond = nullptr;
2740  if (PBI->getSuccessor(0) == TrueDest) {
2741  // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2742  // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2743  // is false: !PBI_Cond and BI_Value
2744  Instruction *NotCond = cast<Instruction>(
2745  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2746  MergedCond = cast<Instruction>(
2747  Builder.CreateBinOp(Instruction::And, NotCond, New, "and.cond"));
2748  if (PBI_C->isOne())
2749  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2750  Instruction::Or, PBI->getCondition(), MergedCond, "or.cond"));
2751  } else {
2752  // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2753  // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2754  // is false: PBI_Cond and BI_Value
2755  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2756  Instruction::And, PBI->getCondition(), New, "and.cond"));
2757  if (PBI_C->isOne()) {
2758  Instruction *NotCond = cast<Instruction>(
2759  Builder.CreateNot(PBI->getCondition(), "not.cond"));
2760  MergedCond = cast<Instruction>(Builder.CreateBinOp(
2761  Instruction::Or, NotCond, MergedCond, "or.cond"));
2762  }
2763  }
2764  // Update PHI Node.
2765  PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2766  MergedCond);
2767  }
2768  // Change PBI from Conditional to Unconditional.
2769  BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2771  PBI = New_PBI;
2772  }
2773 
2774  // If BI was a loop latch, it may have had associated loop metadata.
2775  // We need to copy it to the new latch, that is, PBI.
2776  if (MDNode *LoopMD = BI->getMetadata(LLVMContext::MD_loop))
2777  PBI->setMetadata(LLVMContext::MD_loop, LoopMD);
2778 
2779  // TODO: If BB is reachable from all paths through PredBlock, then we
2780  // could replace PBI's branch probabilities with BI's.
2781 
2782  // Copy any debug value intrinsics into the end of PredBlock.
2783  for (Instruction &I : *BB)
2784  if (isa<DbgInfoIntrinsic>(I))
2785  I.clone()->insertBefore(PBI);
2786 
2787  return true;
2788  }
2789  return false;
2790 }
2791 
2792 // If there is only one store in BB1 and BB2, return it, otherwise return
2793 // nullptr.
2795  StoreInst *S = nullptr;
2796  for (auto *BB : {BB1, BB2}) {
2797  if (!BB)
2798  continue;
2799  for (auto &I : *BB)
2800  if (auto *SI = dyn_cast<StoreInst>(&I)) {
2801  if (S)
2802  // Multiple stores seen.
2803  return nullptr;
2804  else
2805  S = SI;
2806  }
2807  }
2808  return S;
2809 }
2810 
2812  Value *AlternativeV = nullptr) {
2813  // PHI is going to be a PHI node that allows the value V that is defined in
2814  // BB to be referenced in BB's only successor.
2815  //
2816  // If AlternativeV is nullptr, the only value we care about in PHI is V. It
2817  // doesn't matter to us what the other operand is (it'll never get used). We
2818  // could just create a new PHI with an undef incoming value, but that could
2819  // increase register pressure if EarlyCSE/InstCombine can't fold it with some
2820  // other PHI. So here we directly look for some PHI in BB's successor with V
2821  // as an incoming operand. If we find one, we use it, else we create a new
2822  // one.
2823  //
2824  // If AlternativeV is not nullptr, we care about both incoming values in PHI.
2825  // PHI must be exactly: phi <ty> [ %BB, %V ], [ %OtherBB, %AlternativeV]
2826  // where OtherBB is the single other predecessor of BB's only successor.
2827  PHINode *PHI = nullptr;
2828  BasicBlock *Succ = BB->getSingleSuccessor();
2829 
2830  for (auto I = Succ->begin(); isa<PHINode>(I); ++I)
2831  if (cast<PHINode>(I)->getIncomingValueForBlock(BB) == V) {
2832  PHI = cast<PHINode>(I);
2833  if (!AlternativeV)
2834  break;
2835 
2836  assert(std::distance(pred_begin(Succ), pred_end(Succ)) == 2);
2837  auto PredI = pred_begin(Succ);
2838  BasicBlock *OtherPredBB = *PredI == BB ? *++PredI : *PredI;
2839  if (PHI->getIncomingValueForBlock(OtherPredBB) == AlternativeV)
2840  break;
2841  PHI = nullptr;
2842  }
2843  if (PHI)
2844  return PHI;
2845 
2846  // If V is not an instruction defined in BB, just return it.
2847  if (!AlternativeV &&
2848  (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB))
2849  return V;
2850 
2851  PHI = PHINode::Create(V->getType(), 2, "simplifycfg.merge", &Succ->front());
2852  PHI->addIncoming(V, BB);
2853  for (BasicBlock *PredBB : predecessors(Succ))
2854  if (PredBB != BB)
2855  PHI->addIncoming(
2856  AlternativeV ? AlternativeV : UndefValue::get(V->getType()), PredBB);
2857  return PHI;
2858 }
2859 
2861  BasicBlock *QTB, BasicBlock *QFB,
2862  BasicBlock *PostBB, Value *Address,
2863  bool InvertPCond, bool InvertQCond) {
2864  auto IsaBitcastOfPointerType = [](const Instruction &I) {
2865  return Operator::getOpcode(&I) == Instruction::BitCast &&
2866  I.getType()->isPointerTy();
2867  };
2868 
2869  // If we're not in aggressive mode, we only optimize if we have some
2870  // confidence that by optimizing we'll allow P and/or Q to be if-converted.
2871  auto IsWorthwhile = [&](BasicBlock *BB) {
2872  if (!BB)
2873  return true;
2874  // Heuristic: if the block can be if-converted/phi-folded and the
2875  // instructions inside are all cheap (arithmetic/GEPs), it's worthwhile to
2876  // thread this store.
2877  unsigned N = 0;
2878  for (auto &I : *BB) {
2879  // Cheap instructions viable for folding.
2880  if (isa<BinaryOperator>(I) || isa<GetElementPtrInst>(I) ||
2881  isa<StoreInst>(I))
2882  ++N;
2883  // Free instructions.
2884  else if (isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
2885  IsaBitcastOfPointerType(I))
2886  continue;
2887  else
2888  return false;
2889  }
2890  return N <= PHINodeFoldingThreshold;
2891  };
2892 
2894  (!IsWorthwhile(PTB) || !IsWorthwhile(PFB) || !IsWorthwhile(QTB) ||
2895  !IsWorthwhile(QFB)))
2896  return false;
2897 
2898  // For every pointer, there must be exactly two stores, one coming from
2899  // PTB or PFB, and the other from QTB or QFB. We don't support more than one
2900  // store (to any address) in PTB,PFB or QTB,QFB.
2901  // FIXME: We could relax this restriction with a bit more work and performance
2902  // testing.
2903  StoreInst *PStore = findUniqueStoreInBlocks(PTB, PFB);
2904  StoreInst *QStore = findUniqueStoreInBlocks(QTB, QFB);
2905  if (!PStore || !QStore)
2906  return false;
2907 
2908  // Now check the stores are compatible.
2909  if (!QStore->isUnordered() || !PStore->isUnordered())
2910  return false;
2911 
2912  // Check that sinking the store won't cause program behavior changes. Sinking
2913  // the store out of the Q blocks won't change any behavior as we're sinking
2914  // from a block to its unconditional successor. But we're moving a store from
2915  // the P blocks down through the middle block (QBI) and past both QFB and QTB.
2916  // So we need to check that there are no aliasing loads or stores in
2917  // QBI, QTB and QFB. We also need to check there are no conflicting memory
2918  // operations between PStore and the end of its parent block.
2919  //
2920  // The ideal way to do this is to query AliasAnalysis, but we don't
2921  // preserve AA currently so that is dangerous. Be super safe and just
2922  // check there are no other memory operations at all.
2923  for (auto &I : *QFB->getSinglePredecessor())
2924  if (I.mayReadOrWriteMemory())
2925  return false;
2926  for (auto &I : *QFB)
2927  if (&I != QStore && I.mayReadOrWriteMemory())
2928  return false;
2929  if (QTB)
2930  for (auto &I : *QTB)
2931  if (&I != QStore && I.mayReadOrWriteMemory())
2932  return false;
2933  for (auto I = BasicBlock::iterator(PStore), E = PStore->getParent()->end();
2934  I != E; ++I)
2935  if (&*I != PStore && I->mayReadOrWriteMemory())
2936  return false;
2937 
2938  // OK, we're going to sink the stores to PostBB. The store has to be
2939  // conditional though, so first create the predicate.
2940  Value *PCond = cast<BranchInst>(PFB->getSinglePredecessor()->getTerminator())
2941  ->getCondition();
2942  Value *QCond = cast<BranchInst>(QFB->getSinglePredecessor()->getTerminator())
2943  ->getCondition();
2944 
2946  PStore->getParent());
2948  QStore->getParent(), PPHI);
2949 
2950  IRBuilder<> QB(&*PostBB->getFirstInsertionPt());
2951 
2952  Value *PPred = PStore->getParent() == PTB ? PCond : QB.CreateNot(PCond);
2953  Value *QPred = QStore->getParent() == QTB ? QCond : QB.CreateNot(QCond);
2954 
2955  if (InvertPCond)
2956  PPred = QB.CreateNot(PPred);
2957  if (InvertQCond)
2958  QPred = QB.CreateNot(QPred);
2959  Value *CombinedPred = QB.CreateOr(PPred, QPred);
2960 
2961  auto *T =
2962  SplitBlockAndInsertIfThen(CombinedPred, &*QB.GetInsertPoint(), false);
2963  QB.SetInsertPoint(T);
2964  StoreInst *SI = cast<StoreInst>(QB.CreateStore(QPHI, Address));
2965  AAMDNodes AAMD;
2966  PStore->getAAMetadata(AAMD, /*Merge=*/false);
2967  PStore->getAAMetadata(AAMD, /*Merge=*/true);
2968  SI->setAAMetadata(AAMD);
2969 
2970  QStore->eraseFromParent();
2971  PStore->eraseFromParent();
2972 
2973  return true;
2974 }
2975 
2977  // The intention here is to find diamonds or triangles (see below) where each
2978  // conditional block contains a store to the same address. Both of these
2979  // stores are conditional, so they can't be unconditionally sunk. But it may
2980  // be profitable to speculatively sink the stores into one merged store at the
2981  // end, and predicate the merged store on the union of the two conditions of
2982  // PBI and QBI.
2983  //
2984  // This can reduce the number of stores executed if both of the conditions are
2985  // true, and can allow the blocks to become small enough to be if-converted.
2986  // This optimization will also chain, so that ladders of test-and-set
2987  // sequences can be if-converted away.
2988  //
2989  // We only deal with simple diamonds or triangles:
2990  //
2991  // PBI or PBI or a combination of the two
2992  // / \ | \
2993  // PTB PFB | PFB
2994  // \ / | /
2995  // QBI QBI
2996  // / \ | \
2997  // QTB QFB | QFB
2998  // \ / | /
2999  // PostBB PostBB
3000  //
3001  // We model triangles as a type of diamond with a nullptr "true" block.
3002  // Triangles are canonicalized so that the fallthrough edge is represented by
3003  // a true condition, as in the diagram above.
3004  //
3005  BasicBlock *PTB = PBI->getSuccessor(0);
3006  BasicBlock *PFB = PBI->getSuccessor(1);
3007  BasicBlock *QTB = QBI->getSuccessor(0);
3008  BasicBlock *QFB = QBI->getSuccessor(1);
3009  BasicBlock *PostBB = QFB->getSingleSuccessor();
3010 
3011  // Make sure we have a good guess for PostBB. If QTB's only successor is
3012  // QFB, then QFB is a better PostBB.
3013  if (QTB->getSingleSuccessor() == QFB)
3014  PostBB = QFB;
3015 
3016  // If we couldn't find a good PostBB, stop.
3017  if (!PostBB)
3018  return false;
3019 
3020  bool InvertPCond = false, InvertQCond = false;
3021  // Canonicalize fallthroughs to the true branches.
3022  if (PFB == QBI->getParent()) {
3023  std::swap(PFB, PTB);
3024  InvertPCond = true;
3025  }
3026  if (QFB == PostBB) {
3027  std::swap(QFB, QTB);
3028  InvertQCond = true;
3029  }
3030 
3031  // From this point on we can assume PTB or QTB may be fallthroughs but PFB
3032  // and QFB may not. Model fallthroughs as a nullptr block.
3033  if (PTB == QBI->getParent())
3034  PTB = nullptr;
3035  if (QTB == PostBB)
3036  QTB = nullptr;
3037 
3038  // Legality bailouts. We must have at least the non-fallthrough blocks and
3039  // the post-dominating block, and the non-fallthroughs must only have one
3040  // predecessor.
3041  auto HasOnePredAndOneSucc = [](BasicBlock *BB, BasicBlock *P, BasicBlock *S) {
3042  return BB->getSinglePredecessor() == P && BB->getSingleSuccessor() == S;
3043  };
3044  if (!HasOnePredAndOneSucc(PFB, PBI->getParent(), QBI->getParent()) ||
3045  !HasOnePredAndOneSucc(QFB, QBI->getParent(), PostBB))
3046  return false;
3047  if ((PTB && !HasOnePredAndOneSucc(PTB, PBI->getParent(), QBI->getParent())) ||
3048  (QTB && !HasOnePredAndOneSucc(QTB, QBI->getParent(), PostBB)))
3049  return false;
3050  if (!PostBB->hasNUses(2) || !QBI->getParent()->hasNUses(2))
3051  return false;
3052 
3053  // OK, this is a sequence of two diamonds or triangles.
3054  // Check if there are stores in PTB or PFB that are repeated in QTB or QFB.
3055  SmallPtrSet<Value *, 4> PStoreAddresses, QStoreAddresses;
3056  for (auto *BB : {PTB, PFB}) {
3057  if (!BB)
3058  continue;
3059  for (auto &I : *BB)
3060  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3061  PStoreAddresses.insert(SI->getPointerOperand());
3062  }
3063  for (auto *BB : {QTB, QFB}) {
3064  if (!BB)
3065  continue;
3066  for (auto &I : *BB)
3067  if (StoreInst *SI = dyn_cast<StoreInst>(&I))
3068  QStoreAddresses.insert(SI->getPointerOperand());
3069  }
3070 
3071  set_intersect(PStoreAddresses, QStoreAddresses);
3072  // set_intersect mutates PStoreAddresses in place. Rename it here to make it
3073  // clear what it contains.
3074  auto &CommonAddresses = PStoreAddresses;
3075 
3076  bool Changed = false;
3077  for (auto *Address : CommonAddresses)
3078  Changed |= mergeConditionalStoreToAddress(
3079  PTB, PFB, QTB, QFB, PostBB, Address, InvertPCond, InvertQCond);
3080  return Changed;
3081 }
3082 
3083 /// If we have a conditional branch as a predecessor of another block,
3084 /// this function tries to simplify it. We know
3085 /// that PBI and BI are both conditional branches, and BI is in one of the
3086 /// successor blocks of PBI - PBI branches to BI.
3088  const DataLayout &DL) {
3089  assert(PBI->isConditional() && BI->isConditional());
3090  BasicBlock *BB = BI->getParent();
3091 
3092  // If this block ends with a branch instruction, and if there is a
3093  // predecessor that ends on a branch of the same condition, make
3094  // this conditional branch redundant.
3095  if (PBI->getCondition() == BI->getCondition() &&
3096  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3097  // Okay, the outcome of this conditional branch is statically
3098  // knowable. If this block had a single pred, handle specially.
3099  if (BB->getSinglePredecessor()) {
3100  // Turn this into a branch on constant.
3101  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3102  BI->setCondition(
3103  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue));
3104  return true; // Nuke the branch on constant.
3105  }
3106 
3107  // Otherwise, if there are multiple predecessors, insert a PHI that merges
3108  // in the constant and simplify the block result. Subsequent passes of
3109  // simplifycfg will thread the block.
3111  pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
3112  PHINode *NewPN = PHINode::Create(
3113  Type::getInt1Ty(BB->getContext()), std::distance(PB, PE),
3114  BI->getCondition()->getName() + ".pr", &BB->front());
3115  // Okay, we're going to insert the PHI node. Since PBI is not the only
3116  // predecessor, compute the PHI'd conditional value for all of the preds.
3117  // Any predecessor where the condition is not computable we keep symbolic.
3118  for (pred_iterator PI = PB; PI != PE; ++PI) {
3119  BasicBlock *P = *PI;
3120  if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) && PBI != BI &&
3121  PBI->isConditional() && PBI->getCondition() == BI->getCondition() &&
3122  PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
3123  bool CondIsTrue = PBI->getSuccessor(0) == BB;
3124  NewPN->addIncoming(
3125  ConstantInt::get(Type::getInt1Ty(BB->getContext()), CondIsTrue),
3126  P);
3127  } else {
3128  NewPN->addIncoming(BI->getCondition(), P);
3129  }
3130  }
3131 
3132  BI->setCondition(NewPN);
3133  return true;
3134  }
3135  }
3136 
3137  if (auto *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
3138  if (CE->canTrap())
3139  return false;
3140 
3141  // If both branches are conditional and both contain stores to the same
3142  // address, remove the stores from the conditionals and create a conditional
3143  // merged store at the end.
3144  if (MergeCondStores && mergeConditionalStores(PBI, BI))
3145  return true;
3146 
3147  // If this is a conditional branch in an empty block, and if any
3148  // predecessors are a conditional branch to one of our destinations,
3149  // fold the conditions into logical ops and one cond br.
3150  BasicBlock::iterator BBI = BB->begin();
3151  // Ignore dbg intrinsics.
3152  while (isa<DbgInfoIntrinsic>(BBI))
3153  ++BBI;
3154  if (&*BBI != BI)
3155  return false;
3156 
3157  int PBIOp, BIOp;
3158  if (PBI->getSuccessor(0) == BI->getSuccessor(0)) {
3159  PBIOp = 0;
3160  BIOp = 0;
3161  } else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) {
3162  PBIOp = 0;
3163  BIOp = 1;
3164  } else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) {
3165  PBIOp = 1;
3166  BIOp = 0;
3167  } else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) {
3168  PBIOp = 1;
3169  BIOp = 1;
3170  } else {
3171  return false;
3172  }
3173 
3174  // Check to make sure that the other destination of this branch
3175  // isn't BB itself. If so, this is an infinite loop that will
3176  // keep getting unwound.
3177  if (PBI->getSuccessor(PBIOp) == BB)
3178  return false;
3179 
3180  // Do not perform this transformation if it would require
3181  // insertion of a large number of select instructions. For targets
3182  // without predication/cmovs, this is a big pessimization.
3183 
3184  // Also do not perform this transformation if any phi node in the common
3185  // destination block can trap when reached by BB or PBB (PR17073). In that
3186  // case, it would be unsafe to hoist the operation into a select instruction.
3187 
3188  BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
3189  unsigned NumPhis = 0;
3190  for (BasicBlock::iterator II = CommonDest->begin(); isa<PHINode>(II);
3191  ++II, ++NumPhis) {
3192  if (NumPhis > 2) // Disable this xform.
3193  return false;
3194 
3195  PHINode *PN = cast<PHINode>(II);
3196  Value *BIV = PN->getIncomingValueForBlock(BB);
3197  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BIV))
3198  if (CE->canTrap())
3199  return false;
3200 
3201  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3202  Value *PBIV = PN->getIncomingValue(PBBIdx);
3203  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(PBIV))
3204  if (CE->canTrap())
3205  return false;
3206  }
3207 
3208  // Finally, if everything is ok, fold the branches to logical ops.
3209  BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
3210 
3211  DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
3212  << "AND: " << *BI->getParent());
3213 
3214  // If OtherDest *is* BB, then BB is a basic block with a single conditional
3215  // branch in it, where one edge (OtherDest) goes back to itself but the other
3216  // exits. We don't *know* that the program avoids the infinite loop
3217  // (even though that seems likely). If we do this xform naively, we'll end up
3218  // recursively unpeeling the loop. Since we know that (after the xform is
3219  // done) that the block *is* infinite if reached, we just make it an obviously
3220  // infinite loop with no cond branch.
3221  if (OtherDest == BB) {
3222  // Insert it at the end of the function, because it's either code,
3223  // or it won't matter if it's hot. :)
3224  BasicBlock *InfLoopBlock =
3225  BasicBlock::Create(BB->getContext(), "infloop", BB->getParent());
3226  BranchInst::Create(InfLoopBlock, InfLoopBlock);
3227  OtherDest = InfLoopBlock;
3228  }
3229 
3230  DEBUG(dbgs() << *PBI->getParent()->getParent());
3231 
3232  // BI may have other predecessors. Because of this, we leave
3233  // it alone, but modify PBI.
3234 
3235  // Make sure we get to CommonDest on True&True directions.
3236  Value *PBICond = PBI->getCondition();
3237  IRBuilder<NoFolder> Builder(PBI);
3238  if (PBIOp)
3239  PBICond = Builder.CreateNot(PBICond, PBICond->getName() + ".not");
3240 
3241  Value *BICond = BI->getCondition();
3242  if (BIOp)
3243  BICond = Builder.CreateNot(BICond, BICond->getName() + ".not");
3244 
3245  // Merge the conditions.
3246  Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
3247 
3248  // Modify PBI to branch on the new condition to the new dests.
3249  PBI->setCondition(Cond);
3250  PBI->setSuccessor(0, CommonDest);
3251  PBI->setSuccessor(1, OtherDest);
3252 
3253  // Update branch weight for PBI.
3254  uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
3255  uint64_t PredCommon, PredOther, SuccCommon, SuccOther;
3256  bool HasWeights =
3257  extractPredSuccWeights(PBI, BI, PredTrueWeight, PredFalseWeight,
3258  SuccTrueWeight, SuccFalseWeight);
3259  if (HasWeights) {
3260  PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3261  PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3262  SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3263  SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3264  // The weight to CommonDest should be PredCommon * SuccTotal +
3265  // PredOther * SuccCommon.
3266  // The weight to OtherDest should be PredOther * SuccOther.
3267  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther) +
3268  PredOther * SuccCommon,
3269  PredOther * SuccOther};
3270  // Halve the weights if any of them cannot fit in an uint32_t
3271  FitWeights(NewWeights);
3272 
3274  MDBuilder(BI->getContext())
3275  .createBranchWeights(NewWeights[0], NewWeights[1]));
3276  }
3277 
3278  // OtherDest may have phi nodes. If so, add an entry from PBI's
3279  // block that are identical to the entries for BI's block.
3280  AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
3281 
3282  // We know that the CommonDest already had an edge from PBI to
3283  // it. If it has PHIs though, the PHIs may have different
3284  // entries for BB and PBI's BB. If so, insert a select to make
3285  // them agree.
3286  PHINode *PN;
3287  for (BasicBlock::iterator II = CommonDest->begin();
3288  (PN = dyn_cast<PHINode>(II)); ++II) {
3289  Value *BIV = PN->getIncomingValueForBlock(BB);
3290  unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
3291  Value *PBIV = PN->getIncomingValue(PBBIdx);
3292  if (BIV != PBIV) {
3293  // Insert a select in PBI to pick the right value.
3294  SelectInst *NV = cast<SelectInst>(
3295  Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName() + ".mux"));
3296  PN->setIncomingValue(PBBIdx, NV);
3297  // Although the select has the same condition as PBI, the original branch
3298  // weights for PBI do not apply to the new select because the select's
3299  // 'logical' edges are incoming edges of the phi that is eliminated, not
3300  // the outgoing edges of PBI.
3301  if (HasWeights) {
3302  uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
3303  uint64_t PredOther = PBIOp ? PredTrueWeight : PredFalseWeight;
3304  uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
3305  uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
3306  // The weight to PredCommonDest should be PredCommon * SuccTotal.
3307  // The weight to PredOtherDest should be PredOther * SuccCommon.
3308  uint64_t NewWeights[2] = {PredCommon * (SuccCommon + SuccOther),
3309  PredOther * SuccCommon};
3310 
3311  FitWeights(NewWeights);
3312 
3314  MDBuilder(BI->getContext())
3315  .createBranchWeights(NewWeights[0], NewWeights[1]));
3316  }
3317  }
3318  }
3319 
3320  DEBUG(dbgs() << "INTO: " << *PBI->getParent());
3321  DEBUG(dbgs() << *PBI->getParent()->getParent());
3322 
3323  // This basic block is probably dead. We know it has at least
3324  // one fewer predecessor.
3325  return true;
3326 }
3327 
3328 // Simplifies a terminator by replacing it with a branch to TrueBB if Cond is
3329 // true or to FalseBB if Cond is false.
3330 // Takes care of updating the successors and removing the old terminator.
3331 // Also makes sure not to introduce new successors by assuming that edges to
3332 // non-successor TrueBBs and FalseBBs aren't reachable.
3334  BasicBlock *TrueBB, BasicBlock *FalseBB,
3335  uint32_t TrueWeight,
3336  uint32_t FalseWeight) {
3337  // Remove any superfluous successor edges from the CFG.
3338  // First, figure out which successors to preserve.
3339  // If TrueBB and FalseBB are equal, only try to preserve one copy of that
3340  // successor.
3341  BasicBlock *KeepEdge1 = TrueBB;
3342  BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : nullptr;
3343 
3344  // Then remove the rest.
3345  for (BasicBlock *Succ : OldTerm->successors()) {
3346  // Make sure only to keep exactly one copy of each edge.
3347  if (Succ == KeepEdge1)
3348  KeepEdge1 = nullptr;
3349  else if (Succ == KeepEdge2)
3350  KeepEdge2 = nullptr;
3351  else
3352  Succ->removePredecessor(OldTerm->getParent(),
3353  /*DontDeleteUselessPHIs=*/true);
3354  }
3355 
3356  IRBuilder<> Builder(OldTerm);
3357  Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
3358 
3359  // Insert an appropriate new terminator.
3360  if (!KeepEdge1 && !KeepEdge2) {
3361  if (TrueBB == FalseBB)
3362  // We were only looking for one successor, and it was present.
3363  // Create an unconditional branch to it.
3364  Builder.CreateBr(TrueBB);
3365  else {
3366  // We found both of the successors we were looking for.
3367  // Create a conditional branch sharing the condition of the select.
3368  BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
3369  if (TrueWeight != FalseWeight)
3371  MDBuilder(OldTerm->getContext())
3372  .createBranchWeights(TrueWeight, FalseWeight));
3373  }
3374  } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
3375  // Neither of the selected blocks were successors, so this
3376  // terminator must be unreachable.
3377  new UnreachableInst(OldTerm->getContext(), OldTerm);
3378  } else {
3379  // One of the selected values was a successor, but the other wasn't.
3380  // Insert an unconditional branch to the one that was found;
3381  // the edge to the one that wasn't must be unreachable.
3382  if (!KeepEdge1)
3383  // Only TrueBB was found.
3384  Builder.CreateBr(TrueBB);
3385  else
3386  // Only FalseBB was found.
3387  Builder.CreateBr(FalseBB);
3388  }
3389 
3391  return true;
3392 }
3393 
3394 // Replaces
3395 // (switch (select cond, X, Y)) on constant X, Y
3396 // with a branch - conditional if X and Y lead to distinct BBs,
3397 // unconditional otherwise.
3399  // Check for constant integer values in the select.
3400  ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
3401  ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
3402  if (!TrueVal || !FalseVal)
3403  return false;
3404 
3405  // Find the relevant condition and destinations.
3406  Value *Condition = Select->getCondition();
3407  BasicBlock *TrueBB = SI->findCaseValue(TrueVal)->getCaseSuccessor();
3408  BasicBlock *FalseBB = SI->findCaseValue(FalseVal)->getCaseSuccessor();
3409 
3410  // Get weight for TrueBB and FalseBB.
3411  uint32_t TrueWeight = 0, FalseWeight = 0;
3412  SmallVector<uint64_t, 8> Weights;
3413  bool HasWeights = HasBranchWeights(SI);
3414  if (HasWeights) {
3415  GetBranchWeights(SI, Weights);
3416  if (Weights.size() == 1 + SI->getNumCases()) {
3417  TrueWeight =
3418  (uint32_t)Weights[SI->findCaseValue(TrueVal)->getSuccessorIndex()];
3419  FalseWeight =
3420  (uint32_t)Weights[SI->findCaseValue(FalseVal)->getSuccessorIndex()];
3421  }
3422  }
3423 
3424  // Perform the actual simplification.
3425  return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB, TrueWeight,
3426  FalseWeight);
3427 }
3428 
3429 // Replaces
3430 // (indirectbr (select cond, blockaddress(@fn, BlockA),
3431 // blockaddress(@fn, BlockB)))
3432 // with
3433 // (br cond, BlockA, BlockB).
3435  // Check that both operands of the select are block addresses.
3438  if (!TBA || !FBA)
3439  return false;
3440 
3441  // Extract the actual blocks.
3442  BasicBlock *TrueBB = TBA->getBasicBlock();
3443  BasicBlock *FalseBB = FBA->getBasicBlock();
3444 
3445  // Perform the actual simplification.
3446  return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB, 0,
3447  0);
3448 }
3449 
3450 /// This is called when we find an icmp instruction
3451 /// (a seteq/setne with a constant) as the only instruction in a
3452 /// block that ends with an uncond branch. We are looking for a very specific
3453 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
3454 /// this case, we merge the first two "or's of icmp" into a switch, but then the
3455 /// default value goes to an uncond block with a seteq in it, we get something
3456 /// like:
3457 ///
3458 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
3459 /// DEFAULT:
3460 /// %tmp = icmp eq i8 %A, 92
3461 /// br label %end
3462 /// end:
3463 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
3464 ///
3465 /// We prefer to split the edge to 'end' so that there is a true/false entry to
3466 /// the PHI, merging the third icmp into the switch.
3468  ICmpInst *ICI, IRBuilder<> &Builder, const DataLayout &DL,
3469  const TargetTransformInfo &TTI, unsigned BonusInstThreshold,
3470  AssumptionCache *AC) {
3471  BasicBlock *BB = ICI->getParent();
3472 
3473  // If the block has any PHIs in it or the icmp has multiple uses, it is too
3474  // complex.
3475  if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse())
3476  return false;
3477 
3478  Value *V = ICI->getOperand(0);
3479  ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
3480 
3481  // The pattern we're looking for is where our only predecessor is a switch on
3482  // 'V' and this block is the default case for the switch. In this case we can
3483  // fold the compared value into the switch to simplify things.
3484  BasicBlock *Pred = BB->getSinglePredecessor();
3485  if (!Pred || !isa<SwitchInst>(Pred->getTerminator()))
3486  return false;
3487 
3488  SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
3489  if (SI->getCondition() != V)
3490  return false;
3491 
3492  // If BB is reachable on a non-default case, then we simply know the value of
3493  // V in this block. Substitute it and constant fold the icmp instruction
3494  // away.
3495  if (SI->getDefaultDest() != BB) {
3496  ConstantInt *VVal = SI->findCaseDest(BB);
3497  assert(VVal && "Should have a unique destination value");
3498  ICI->setOperand(0, VVal);
3499 
3500  if (Value *V = SimplifyInstruction(ICI, {DL, ICI})) {
3501  ICI->replaceAllUsesWith(V);
3502  ICI->eraseFromParent();
3503  }
3504  // BB is now empty, so it is likely to simplify away.
3505  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3506  }
3507 
3508  // Ok, the block is reachable from the default dest. If the constant we're
3509  // comparing exists in one of the other edges, then we can constant fold ICI
3510  // and zap it.
3511  if (SI->findCaseValue(Cst) != SI->case_default()) {
3512  Value *V;
3513  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3514  V = ConstantInt::getFalse(BB->getContext());
3515  else
3516  V = ConstantInt::getTrue(BB->getContext());
3517 
3518  ICI->replaceAllUsesWith(V);
3519  ICI->eraseFromParent();
3520  // BB is now empty, so it is likely to simplify away.
3521  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
3522  }
3523 
3524  // The use of the icmp has to be in the 'end' block, by the only PHI node in
3525  // the block.
3526  BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
3527  PHINode *PHIUse = dyn_cast<PHINode>(ICI->user_back());
3528  if (PHIUse == nullptr || PHIUse != &SuccBlock->front() ||
3529  isa<PHINode>(++BasicBlock::iterator(PHIUse)))
3530  return false;
3531 
3532  // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
3533  // true in the PHI.
3534  Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
3535  Constant *NewCst = ConstantInt::getFalse(BB->getContext());
3536 
3537  if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
3538  std::swap(DefaultCst, NewCst);
3539 
3540  // Replace ICI (which is used by the PHI for the default value) with true or
3541  // false depending on if it is EQ or NE.
3542  ICI->replaceAllUsesWith(DefaultCst);
3543  ICI->eraseFromParent();
3544 
3545  // Okay, the switch goes to this block on a default value. Add an edge from
3546  // the switch to the merge point on the compared value.
3547  BasicBlock *NewBB =
3548  BasicBlock::Create(BB->getContext(), "switch.edge", BB->getParent(), BB);
3549  SmallVector<uint64_t, 8> Weights;
3550  bool HasWeights = HasBranchWeights(SI);
3551  if (HasWeights) {
3552  GetBranchWeights(SI, Weights);
3553  if (Weights.size() == 1 + SI->getNumCases()) {
3554  // Split weight for default case to case for "Cst".
3555  Weights[0] = (Weights[0] + 1) >> 1;
3556  Weights.push_back(Weights[0]);
3557 
3558  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3559  SI->setMetadata(
3561  MDBuilder(SI->getContext()).createBranchWeights(MDWeights));
3562  }
3563  }
3564  SI->addCase(Cst, NewBB);
3565 
3566  // NewBB branches to the phi block, add the uncond branch and the phi entry.
3567  Builder.SetInsertPoint(NewBB);
3568  Builder.SetCurrentDebugLocation(SI->getDebugLoc());
3569  Builder.CreateBr(SuccBlock);
3570  PHIUse->addIncoming(NewCst, NewBB);
3571  return true;
3572 }
3573 
3574 /// The specified branch is a conditional branch.
3575 /// Check to see if it is branching on an or/and chain of icmp instructions, and
3576 /// fold it into a switch instruction if so.
3578  const DataLayout &DL) {
3579  Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
3580  if (!Cond)
3581  return false;
3582 
3583  // Change br (X == 0 | X == 1), T, F into a switch instruction.
3584  // If this is a bunch of seteq's or'd together, or if it's a bunch of
3585  // 'setne's and'ed together, collect them.
3586 
3587  // Try to gather values from a chain of and/or to be turned into a switch
3588  ConstantComparesGatherer ConstantCompare(Cond, DL);
3589  // Unpack the result
3590  SmallVectorImpl<ConstantInt *> &Values = ConstantCompare.Vals;
3591  Value *CompVal = ConstantCompare.CompValue;
3592  unsigned UsedICmps = ConstantCompare.UsedICmps;
3593  Value *ExtraCase = ConstantCompare.Extra;
3594 
3595  // If we didn't have a multiply compared value, fail.
3596  if (!CompVal)
3597  return false;
3598 
3599  // Avoid turning single icmps into a switch.
3600  if (UsedICmps <= 1)
3601  return false;
3602 
3603  bool TrueWhenEqual = (Cond->getOpcode() == Instruction::Or);
3604 
3605  // There might be duplicate constants in the list, which the switch
3606  // instruction can't handle, remove them now.
3607  array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
3608  Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
3609 
3610  // If Extra was used, we require at least two switch values to do the
3611  // transformation. A switch with one value is just a conditional branch.
3612  if (ExtraCase && Values.size() < 2)
3613  return false;
3614 
3615  // TODO: Preserve branch weight metadata, similarly to how
3616  // FoldValueComparisonIntoPredecessors preserves it.
3617 
3618  // Figure out which block is which destination.
3619  BasicBlock *DefaultBB = BI->getSuccessor(1);
3620  BasicBlock *EdgeBB = BI->getSuccessor(0);
3621  if (!TrueWhenEqual)
3622  std::swap(DefaultBB, EdgeBB);
3623 
3624  BasicBlock *BB = BI->getParent();
3625 
3626  DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
3627  << " cases into SWITCH. BB is:\n"
3628  << *BB);
3629 
3630  // If there are any extra values that couldn't be folded into the switch
3631  // then we evaluate them with an explicit branch first. Split the block
3632  // right before the condbr to handle it.
3633  if (ExtraCase) {
3634  BasicBlock *NewBB =
3635  BB->splitBasicBlock(BI->getIterator(), "switch.early.test");
3636  // Remove the uncond branch added to the old block.
3637  TerminatorInst *OldTI = BB->getTerminator();
3638  Builder.SetInsertPoint(OldTI);
3639 
3640  if (TrueWhenEqual)
3641  Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
3642  else
3643  Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
3644 
3645  OldTI->eraseFromParent();
3646 
3647  // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
3648  // for the edge we just added.
3649  AddPredecessorToBlock(EdgeBB, BB, NewBB);
3650 
3651  DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
3652  << "\nEXTRABB = " << *BB);
3653  BB = NewBB;
3654  }
3655 
3656  Builder.SetInsertPoint(BI);
3657  // Convert pointer to int before we switch.
3658  if (CompVal->getType()->isPointerTy()) {
3659  CompVal = Builder.CreatePtrToInt(
3660  CompVal, DL.getIntPtrType(CompVal->getType()), "magicptr");
3661  }
3662 
3663  // Create the new switch instruction now.
3664  SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
3665 
3666  // Add all of the 'cases' to the switch instruction.
3667  for (unsigned i = 0, e = Values.size(); i != e; ++i)
3668  New->addCase(Values[i], EdgeBB);
3669 
3670  // We added edges from PI to the EdgeBB. As such, if there were any
3671  // PHI nodes in EdgeBB, they need entries to be added corresponding to
3672  // the number of edges added.
3673  for (BasicBlock::iterator BBI = EdgeBB->begin(); isa<PHINode>(BBI); ++BBI) {
3674  PHINode *PN = cast<PHINode>(BBI);
3675  Value *InVal = PN->getIncomingValueForBlock(BB);
3676  for (unsigned i = 0, e = Values.size() - 1; i != e; ++i)
3677  PN->addIncoming(InVal, BB);
3678  }
3679 
3680  // Erase the old branch instruction.
3682 
3683  DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
3684  return true;
3685 }
3686 
3687 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
3688  if (isa<PHINode>(RI->getValue()))
3689  return SimplifyCommonResume(RI);
3690  else if (isa<LandingPadInst>(RI->getParent()->getFirstNonPHI()) &&
3691  RI->getValue() == RI->getParent()->getFirstNonPHI())
3692  // The resume must unwind the exception that caused control to branch here.
3693  return SimplifySingleResume(RI);
3694 
3695  return false;
3696 }
3697 
3698 // Simplify resume that is shared by several landing pads (phi of landing pad).
3699 bool SimplifyCFGOpt::SimplifyCommonResume(ResumeInst *RI) {
3700  BasicBlock *BB = RI->getParent();
3701 
3702  // Check that there are no other instructions except for debug intrinsics
3703  // between the phi of landing pads (RI->getValue()) and resume instruction.
3704  BasicBlock::iterator I = cast<Instruction>(RI->getValue())->getIterator(),
3705  E = RI->getIterator();
3706  while (++I != E)
3707  if (!isa<DbgInfoIntrinsic>(I))
3708  return false;
3709 
3710  SmallSetVector<BasicBlock *, 4> TrivialUnwindBlocks;
3711  auto *PhiLPInst = cast<PHINode>(RI->getValue());
3712 
3713  // Check incoming blocks to see if any of them are trivial.
3714  for (unsigned Idx = 0, End = PhiLPInst->getNumIncomingValues(); Idx != End;
3715  Idx++) {
3716  auto *IncomingBB = PhiLPInst->getIncomingBlock(Idx);
3717  auto *IncomingValue = PhiLPInst->getIncomingValue(Idx);
3718 
3719  // If the block has other successors, we can not delete it because
3720  // it has other dependents.
3721  if (IncomingBB->getUniqueSuccessor() != BB)
3722  continue;
3723 
3724  auto *LandingPad = dyn_cast<LandingPadInst>(IncomingBB->getFirstNonPHI());
3725  // Not the landing pad that caused the control to branch here.
3726  if (IncomingValue != LandingPad)
3727  continue;
3728 
3729  bool isTrivial = true;
3730 
3731  I = IncomingBB->getFirstNonPHI()->getIterator();
3732  E = IncomingBB->getTerminator()->getIterator();
3733  while (++I != E)
3734  if (!isa<DbgInfoIntrinsic>(I)) {
3735  isTrivial = false;
3736  break;
3737  }
3738 
3739  if (isTrivial)
3740  TrivialUnwindBlocks.insert(IncomingBB);
3741  }
3742 
3743  // If no trivial unwind blocks, don't do any simplifications.
3744  if (TrivialUnwindBlocks.empty())
3745  return false;
3746 
3747  // Turn all invokes that unwind here into calls.
3748  for (auto *TrivialBB : TrivialUnwindBlocks) {
3749  // Blocks that will be simplified should be removed from the phi node.
3750  // Note there could be multiple edges to the resume block, and we need
3751  // to remove them all.
3752  while (PhiLPInst->getBasicBlockIndex(TrivialBB) != -1)
3753  BB->removePredecessor(TrivialBB, true);
3754 
3755  for (pred_iterator PI = pred_begin(TrivialBB), PE = pred_end(TrivialBB);
3756  PI != PE;) {
3757  BasicBlock *Pred = *PI++;
3758  removeUnwindEdge(Pred);
3759  }
3760 
3761  // In each SimplifyCFG run, only the current processed block can be erased.
3762  // Otherwise, it will break the iteration of SimplifyCFG pass. So instead
3763  // of erasing TrivialBB, we only remove the branch to the common resume
3764  // block so that we can later erase the resume block since it has no
3765  // predecessors.
3766  TrivialBB->getTerminator()->eraseFromParent();
3767  new UnreachableInst(RI->getContext(), TrivialBB);
3768  }
3769 
3770  // Delete the resume block if all its predecessors have been removed.
3771  if (pred_empty(BB))
3772  BB->eraseFromParent();
3773 
3774  return !TrivialUnwindBlocks.empty();
3775 }
3776 
3777 // Simplify resume that is only used by a single (non-phi) landing pad.
3778 bool SimplifyCFGOpt::SimplifySingleResume(ResumeInst *RI) {
3779  BasicBlock *BB = RI->getParent();
3781  assert(RI->getValue() == LPInst &&
3782  "Resume must unwind the exception that caused control to here");
3783 
3784  // Check that there are no other instructions except for debug intrinsics.
3785  BasicBlock::iterator I = LPInst->getIterator(), E = RI->getIterator();
3786  while (++I != E)
3787  if (!isa<DbgInfoIntrinsic>(I))
3788  return false;
3789 
3790  // Turn all invokes that unwind here into calls and delete the basic block.
3791  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3792  BasicBlock *Pred = *PI++;
3793  removeUnwindEdge(Pred);
3794  }
3795 
3796  // The landingpad is now unreachable. Zap it.
3797  BB->eraseFromParent();
3798  if (LoopHeaders)
3799  LoopHeaders->erase(BB);
3800  return true;
3801 }
3802 
3804  // If this is a trivial cleanup pad that executes no instructions, it can be
3805  // eliminated. If the cleanup pad continues to the caller, any predecessor
3806  // that is an EH pad will be updated to continue to the caller and any
3807  // predecessor that terminates with an invoke instruction will have its invoke
3808  // instruction converted to a call instruction. If the cleanup pad being
3809  // simplified does not continue to the caller, each predecessor will be
3810  // updated to continue to the unwind destination of the cleanup pad being
3811  // simplified.
3812  BasicBlock *BB = RI->getParent();
3813  CleanupPadInst *CPInst = RI->getCleanupPad();
3814  if (CPInst->getParent() != BB)
3815  // This isn't an empty cleanup.
3816  return false;
3817 
3818  // We cannot kill the pad if it has multiple uses. This typically arises
3819  // from unreachable basic blocks.
3820  if (!CPInst->hasOneUse())
3821  return false;
3822 
3823  // Check that there are no other instructions except for benign intrinsics.
3824  BasicBlock::iterator I = CPInst->getIterator(), E = RI->getIterator();
3825  while (++I != E) {
3826  auto *II = dyn_cast<IntrinsicInst>(I);
3827  if (!II)
3828  return false;
3829 
3830  Intrinsic::ID IntrinsicID = II->getIntrinsicID();
3831  switch (IntrinsicID) {
3832  case Intrinsic::dbg_declare:
3833  case Intrinsic::dbg_value:
3834  case Intrinsic::lifetime_end:
3835  break;
3836  default:
3837  return false;
3838  }
3839  }
3840 
3841  // If the cleanup return we are simplifying unwinds to the caller, this will
3842  // set UnwindDest to nullptr.
3843  BasicBlock *UnwindDest = RI->getUnwindDest();
3844  Instruction *DestEHPad = UnwindDest ? UnwindDest->getFirstNonPHI() : nullptr;
3845 
3846  // We're about to remove BB from the control flow. Before we do, sink any
3847  // PHINodes into the unwind destination. Doing this before changing the
3848  // control flow avoids some potentially slow checks, since we can currently
3849  // be certain that UnwindDest and BB have no common predecessors (since they
3850  // are both EH pads).
3851  if (UnwindDest) {
3852  // First, go through the PHI nodes in UnwindDest and update any nodes that
3853  // reference the block we are removing
3854  for (BasicBlock::iterator I = UnwindDest->begin(),
3855  IE = DestEHPad->getIterator();
3856  I != IE; ++I) {
3857  PHINode *DestPN = cast<PHINode>(I);
3858 
3859  int Idx = DestPN->getBasicBlockIndex(BB);
3860  // Since BB unwinds to UnwindDest, it has to be in the PHI node.
3861  assert(Idx != -1);
3862  // This PHI node has an incoming value that corresponds to a control
3863  // path through the cleanup pad we are removing. If the incoming
3864  // value is in the cleanup pad, it must be a PHINode (because we
3865  // verified above that the block is otherwise empty). Otherwise, the
3866  // value is either a constant or a value that dominates the cleanup
3867  // pad being removed.
3868  //
3869  // Because BB and UnwindDest are both EH pads, all of their
3870  // predecessors must unwind to these blocks, and since no instruction
3871  // can have multiple unwind destinations, there will be no overlap in
3872  // incoming blocks between SrcPN and DestPN.
3873  Value *SrcVal = DestPN->getIncomingValue(Idx);
3874  PHINode *SrcPN = dyn_cast<PHINode>(SrcVal);
3875 
3876  // Remove the entry for the block we are deleting.
3877  DestPN->removeIncomingValue(Idx, false);
3878 
3879  if (SrcPN && SrcPN->getParent() == BB) {
3880  // If the incoming value was a PHI node in the cleanup pad we are
3881  // removing, we need to merge that PHI node's incoming values into
3882  // DestPN.
3883  for (unsigned SrcIdx = 0, SrcE = SrcPN->getNumIncomingValues();
3884  SrcIdx != SrcE; ++SrcIdx) {
3885  DestPN->addIncoming(SrcPN->getIncomingValue(SrcIdx),
3886  SrcPN->getIncomingBlock(SrcIdx));
3887  }
3888  } else {
3889  // Otherwise, the incoming value came from above BB and
3890  // so we can just reuse it. We must associate all of BB's
3891  // predecessors with this value.
3892  for (auto *pred : predecessors(BB)) {
3893  DestPN->addIncoming(SrcVal, pred);
3894  }
3895  }
3896  }
3897 
3898  // Sink any remaining PHI nodes directly into UnwindDest.
3899  Instruction *InsertPt = DestEHPad;
3900  for (BasicBlock::iterator I = BB->begin(),
3901  IE = BB->getFirstNonPHI()->getIterator();
3902  I != IE;) {
3903  // The iterator must be incremented here because the instructions are
3904  // being moved to another block.
3905  PHINode *PN = cast<PHINode>(I++);
3906  if (PN->use_empty())
3907  // If the PHI node has no uses, just leave it. It will be erased
3908  // when we erase BB below.
3909  continue;
3910 
3911  // Otherwise, sink this PHI node into UnwindDest.
3912  // Any predecessors to UnwindDest which are not already represented
3913  // must be back edges which inherit the value from the path through
3914  // BB. In this case, the PHI value must reference itself.
3915  for (auto *pred : predecessors(UnwindDest))
3916  if (pred != BB)
3917  PN->addIncoming(PN, pred);
3918  PN->moveBefore(InsertPt);
3919  }
3920  }
3921 
3922  for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
3923  // The iterator must be updated here because we are removing this pred.
3924  BasicBlock *PredBB = *PI++;
3925  if (UnwindDest == nullptr) {
3926  removeUnwindEdge(PredBB);
3927  } else {
3928  TerminatorInst *TI = PredBB->getTerminator();
3929  TI->replaceUsesOfWith(BB, UnwindDest);
3930  }
3931  }
3932 
3933  // The cleanup pad is now unreachable. Zap it.
3934  BB->eraseFromParent();
3935  return true;
3936 }
3937 
3938 // Try to merge two cleanuppads together.
3940  // Skip any cleanuprets which unwind to caller, there is nothing to merge
3941  // with.
3942  BasicBlock *UnwindDest = RI->getUnwindDest();
3943  if (!UnwindDest)
3944  return false;
3945 
3946  // This cleanupret isn't the only predecessor of this cleanuppad, it wouldn't
3947  // be safe to merge without code duplication.
3948  if (UnwindDest->getSinglePredecessor() != RI->getParent())
3949  return false;
3950 
3951  // Verify that our cleanuppad's unwind destination is another cleanuppad.
3952  auto *SuccessorCleanupPad = dyn_cast<CleanupPadInst>(&UnwindDest->front());
3953  if (!SuccessorCleanupPad)
3954  return false;
3955 
3956  CleanupPadInst *PredecessorCleanupPad = RI->getCleanupPad();
3957  // Replace any uses of the successor cleanupad with the predecessor pad
3958  // The only cleanuppad uses should be this cleanupret, it's cleanupret and
3959  // funclet bundle operands.
3960  SuccessorCleanupPad->replaceAllUsesWith(PredecessorCleanupPad);
3961  // Remove the old cleanuppad.
3962  SuccessorCleanupPad->eraseFromParent();
3963  // Now, we simply replace the cleanupret with a branch to the unwind
3964  // destination.
3965  BranchInst::Create(UnwindDest, RI->getParent());
3966  RI->eraseFromParent();
3967 
3968  return true;
3969 }
3970 
3971 bool SimplifyCFGOpt::SimplifyCleanupReturn(CleanupReturnInst *RI) {
3972  // It is possible to transiantly have an undef cleanuppad operand because we
3973  // have deleted some, but not all, dead blocks.
3974  // Eventually, this block will be deleted.
3975  if (isa<UndefValue>(RI->getOperand(0)))
3976  return false;
3977 
3978  if (mergeCleanupPad(RI))
3979  return true;
3980 
3981  if (removeEmptyCleanup(RI))
3982  return true;
3983 
3984  return false;
3985 }
3986 
3987 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
3988  BasicBlock *BB = RI->getParent();
3989  if (!BB->getFirstNonPHIOrDbg()->isTerminator())
3990  return false;
3991 
3992  // Find predecessors that end with branches.
3993  SmallVector<BasicBlock *, 8> UncondBranchPreds;
3994  SmallVector<BranchInst *, 8> CondBranchPreds;
3995  for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
3996  BasicBlock *P = *PI;
3997  TerminatorInst *PTI = P->getTerminator();
3998  if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
3999  if (BI->isUnconditional())
4000  UncondBranchPreds.push_back(P);
4001  else
4002  CondBranchPreds.push_back(BI);
4003  }
4004  }
4005 
4006  // If we found some, do the transformation!
4007  if (!UncondBranchPreds.empty() && DupRet) {
4008  while (!UncondBranchPreds.empty()) {
4009  BasicBlock *Pred = UncondBranchPreds.pop_back_val();
4010  DEBUG(dbgs() << "FOLDING: " << *BB
4011  << "INTO UNCOND BRANCH PRED: " << *Pred);
4012  (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
4013  }
4014 
4015  // If we eliminated all predecessors of the block, delete the block now.
4016  if (pred_empty(BB)) {
4017  // We know there are no successors, so just nuke the block.
4018  BB->eraseFromParent();
4019  if (LoopHeaders)
4020  LoopHeaders->erase(BB);
4021  }
4022 
4023  return true;
4024  }
4025 
4026  // Check out all of the conditional branches going to this return
4027  // instruction. If any of them just select between returns, change the
4028  // branch itself into a select/return pair.
4029  while (!CondBranchPreds.empty()) {
4030  BranchInst *BI = CondBranchPreds.pop_back_val();
4031 
4032  // Check to see if the non-BB successor is also a return block.
4033  if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
4034  isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
4035  SimplifyCondBranchToTwoReturns(BI, Builder))
4036  return true;
4037  }
4038  return false;
4039 }
4040 
4041 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
4042  BasicBlock *BB = UI->getParent();
4043 
4044  bool Changed = false;
4045 
4046  // If there are any instructions immediately before the unreachable that can
4047  // be removed, do so.
4048  while (UI->getIterator() != BB->begin()) {
4049  BasicBlock::iterator BBI = UI->getIterator();
4050  --BBI;
4051  // Do not delete instructions that can have side effects which might cause
4052  // the unreachable to not be reachable; specifically, calls and volatile
4053  // operations may have this effect.
4054  if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI))
4055  break;
4056 
4057  if (BBI->mayHaveSideEffects()) {
4058  if (auto *SI = dyn_cast<StoreInst>(BBI)) {
4059  if (SI->isVolatile())
4060  break;
4061  } else if (auto *LI = dyn_cast<LoadInst>(BBI)) {
4062  if (LI->isVolatile())
4063  break;
4064  } else if (auto *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
4065  if (RMWI->isVolatile())
4066  break;
4067  } else if (auto *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
4068  if (CXI->isVolatile())
4069  break;
4070  } else if (isa<CatchPadInst>(BBI)) {
4071  // A catchpad may invoke exception object constructors and such, which
4072  // in some languages can be arbitrary code, so be conservative by
4073  // default.
4074  // For CoreCLR, it just involves a type test, so can be removed.
4077  break;
4078  } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
4079  !isa<LandingPadInst>(BBI)) {
4080  break;
4081  }
4082  // Note that deleting LandingPad's here is in fact okay, although it
4083  // involves a bit of subtle reasoning. If this inst is a LandingPad,
4084  // all the predecessors of this block will be the unwind edges of Invokes,
4085  // and we can therefore guarantee this block will be erased.
4086  }
4087 
4088  // Delete this instruction (any uses are guaranteed to be dead)
4089  if (!BBI->use_empty())
4090  BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
4091  BBI->eraseFromParent();
4092  Changed = true;
4093  }
4094 
4095  // If the unreachable instruction is the first in the block, take a gander
4096  // at all of the predecessors of this instruction, and simplify them.
4097  if (&BB->front() != UI)
4098  return Changed;
4099 
4101  for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
4102  TerminatorInst *TI = Preds[i]->getTerminator();
4103  IRBuilder<> Builder(TI);
4104  if (auto *BI = dyn_cast<BranchInst>(TI)) {
4105  if (BI->isUnconditional()) {
4106  if (BI->getSuccessor(0) == BB) {
4107  new UnreachableInst(TI->getContext(), TI);
4108  TI->eraseFromParent();
4109  Changed = true;
4110  }
4111  } else {
4112  if (BI->getSuccessor(0) == BB) {
4113  Builder.CreateBr(BI->getSuccessor(1));
4115  } else if (BI->getSuccessor(1) == BB) {
4116  Builder.CreateBr(BI->getSuccessor(0));
4118  Changed = true;
4119  }
4120  }
4121  } else if (auto *SI = dyn_cast<SwitchInst>(TI)) {
4122  for (auto i = SI->case_begin(), e = SI->case_end(); i != e;) {
4123  if (i->getCaseSuccessor() != BB) {
4124  ++i;
4125  continue;
4126  }
4127  BB->removePredecessor(SI->getParent());
4128  i = SI->removeCase(i);
4129  e = SI->case_end();
4130  Changed = true;
4131  }
4132  } else if (auto *II = dyn_cast<InvokeInst>(TI)) {
4133  if (II->getUnwindDest() == BB) {
4134  removeUnwindEdge(TI->getParent());
4135  Changed = true;
4136  }
4137  } else if (auto *CSI = dyn_cast<CatchSwitchInst>(TI)) {
4138  if (CSI->getUnwindDest() == BB) {
4139  removeUnwindEdge(TI->getParent());
4140  Changed = true;
4141  continue;
4142  }
4143 
4144  for (CatchSwitchInst::handler_iterator I = CSI->handler_begin(),
4145  E = CSI->handler_end();
4146  I != E; ++I) {
4147  if (*I == BB) {
4148  CSI->removeHandler(I);
4149  --I;
4150  --E;
4151  Changed = true;
4152  }
4153  }
4154  if (CSI->getNumHandlers() == 0) {
4155  BasicBlock *CatchSwitchBB = CSI->getParent();
4156  if (CSI->hasUnwindDest()) {
4157  // Redirect preds to the unwind dest
4158  CatchSwitchBB->replaceAllUsesWith(CSI->getUnwindDest());
4159  } else {
4160  // Rewrite all preds to unwind to caller (or from invoke to call).
4161  SmallVector<BasicBlock *, 8> EHPreds(predecessors(CatchSwitchBB));
4162  for (BasicBlock *EHPred : EHPreds)
4163  removeUnwindEdge(EHPred);
4164  }
4165  // The catchswitch is no longer reachable.
4166  new UnreachableInst(CSI->getContext(), CSI);
4167  CSI->eraseFromParent();
4168  Changed = true;
4169  }
4170  } else if (isa<CleanupReturnInst>(TI)) {
4171  new UnreachableInst(TI->getContext(), TI);
4172  TI->eraseFromParent();
4173  Changed = true;
4174  }
4175  }
4176 
4177  // If this block is now dead, remove it.
4178  if (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()) {
4179  // We know there are no successors, so just nuke the block.
4180  BB->eraseFromParent();
4181  if (LoopHeaders)
4182  LoopHeaders->erase(BB);
4183  return true;
4184  }
4185 
4186  return Changed;
4187 }
4188 
4190  assert(Cases.size() >= 1);
4191 
4192  array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
4193  for (size_t I = 1, E = Cases.size(); I != E; ++I) {
4194  if (Cases[I - 1]->getValue() != Cases[I]->getValue() + 1)
4195  return false;
4196  }
4197  return true;
4198 }
4199 
4200 /// Turn a switch with two reachable destinations into an integer range
4201 /// comparison and branch.
4202 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
4203  assert(SI->getNumCases() > 1 && "Degenerate switch?");
4204 
4205  bool HasDefault =
4206  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4207 
4208  // Partition the cases into two sets with different destinations.
4209  BasicBlock *DestA = HasDefault ? SI->getDefaultDest() : nullptr;
4210  BasicBlock *DestB = nullptr;
4213 
4214  for (auto Case : SI->cases()) {
4215  BasicBlock *Dest = Case.getCaseSuccessor();
4216  if (!DestA)
4217  DestA = Dest;
4218  if (Dest == DestA) {
4219  CasesA.push_back(Case.getCaseValue());
4220  continue;
4221  }
4222  if (!DestB)
4223  DestB = Dest;
4224  if (Dest == DestB) {
4225  CasesB.push_back(Case.getCaseValue());
4226  continue;
4227  }
4228  return false; // More than two destinations.
4229  }
4230 
4231  assert(DestA && DestB &&
4232  "Single-destination switch should have been folded.");
4233  assert(DestA != DestB);
4234  assert(DestB != SI->getDefaultDest());
4235  assert(!CasesB.empty() && "There must be non-default cases.");
4236  assert(!CasesA.empty() || HasDefault);
4237 
4238  // Figure out if one of the sets of cases form a contiguous range.
4239  SmallVectorImpl<ConstantInt *> *ContiguousCases = nullptr;
4240  BasicBlock *ContiguousDest = nullptr;
4241  BasicBlock *OtherDest = nullptr;
4242  if (!CasesA.empty() && CasesAreContiguous(CasesA)) {
4243  ContiguousCases = &CasesA;
4244  ContiguousDest = DestA;
4245  OtherDest = DestB;
4246  } else if (CasesAreContiguous(CasesB)) {
4247  ContiguousCases = &CasesB;
4248  ContiguousDest = DestB;
4249  OtherDest = DestA;
4250  } else
4251  return false;
4252 
4253  // Start building the compare and branch.
4254 
4255  Constant *Offset = ConstantExpr::getNeg(ContiguousCases->back());
4256  Constant *NumCases =
4257  ConstantInt::get(Offset->getType(), ContiguousCases->size());
4258 
4259  Value *Sub = SI->getCondition();
4260  if (!Offset->isNullValue())
4261  Sub = Builder.CreateAdd(Sub, Offset, Sub->getName() + ".off");
4262 
4263  Value *Cmp;
4264  // If NumCases overflowed, then all possible values jump to the successor.
4265  if (NumCases->isNullValue() && !ContiguousCases->empty())
4266  Cmp = ConstantInt::getTrue(SI->getContext());
4267  else
4268  Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
4269  BranchInst *NewBI = Builder.CreateCondBr(Cmp, ContiguousDest, OtherDest);
4270 
4271  // Update weight for the newly-created conditional branch.
4272  if (HasBranchWeights(SI)) {
4273  SmallVector<uint64_t, 8> Weights;
4274  GetBranchWeights(SI, Weights);
4275  if (Weights.size() == 1 + SI->getNumCases()) {
4276  uint64_t TrueWeight = 0;
4277  uint64_t FalseWeight = 0;
4278  for (size_t I = 0, E = Weights.size(); I != E; ++I) {
4279  if (SI->getSuccessor(I) == ContiguousDest)
4280  TrueWeight += Weights[I];
4281  else
4282  FalseWeight += Weights[I];
4283  }
4284  while (TrueWeight > UINT32_MAX || FalseWeight > UINT32_MAX) {
4285  TrueWeight /= 2;
4286  FalseWeight /= 2;
4287  }
4289  MDBuilder(SI->getContext())
4290  .createBranchWeights((uint32_t)TrueWeight,
4291  (uint32_t)FalseWeight));
4292  }
4293  }
4294 
4295  // Prune obsolete incoming values off the successors' PHI nodes.
4296  for (auto BBI = ContiguousDest->begin(); isa<PHINode>(BBI); ++BBI) {
4297  unsigned PreviousEdges = ContiguousCases->size();
4298  if (ContiguousDest == SI->getDefaultDest())
4299  ++PreviousEdges;
4300  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4301  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4302  }
4303  for (auto BBI = OtherDest->begin(); isa<PHINode>(BBI); ++BBI) {
4304  unsigned PreviousEdges = SI->getNumCases() - ContiguousCases->size();
4305  if (OtherDest == SI->getDefaultDest())
4306  ++PreviousEdges;
4307  for (unsigned I = 0, E = PreviousEdges - 1; I != E; ++I)
4308  cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
4309  }
4310 
4311  // Drop the switch.
4312  SI->eraseFromParent();
4313 
4314  return true;
4315 }
4316 
4317 /// Compute masked bits for the condition of a switch
4318 /// and use it to remove dead cases.
4320  const DataLayout &DL) {
4321  Value *Cond = SI->getCondition();
4322  unsigned Bits = Cond->getType()->getIntegerBitWidth();
4323  KnownBits Known = computeKnownBits(Cond, DL, 0, AC, SI);
4324 
4325  // We can also eliminate cases by determining that their values are outside of
4326  // the limited range of the condition based on how many significant (non-sign)
4327  // bits are in the condition value.
4328  unsigned ExtraSignBits = ComputeNumSignBits(Cond, DL, 0, AC, SI) - 1;
4329  unsigned MaxSignificantBitsInCond = Bits - ExtraSignBits;
4330 
4331  // Gather dead cases.
4333  for (auto &Case : SI->cases()) {
4334  const APInt &CaseVal = Case.getCaseValue()->getValue();
4335  if (Known.Zero.intersects(CaseVal) || !Known.One.isSubsetOf(CaseVal) ||
4336  (CaseVal.getMinSignedBits() > MaxSignificantBitsInCond)) {
4337  DeadCases.push_back(Case.getCaseValue());
4338  DEBUG(dbgs() << "SimplifyCFG: switch case " << CaseVal << " is dead.\n");
4339  }
4340  }
4341 
4342  // If we can prove that the cases must cover all possible values, the
4343  // default destination becomes dead and we can remove it. If we know some
4344  // of the bits in the value, we can use that to more precisely compute the
4345  // number of possible unique case values.
4346  bool HasDefault =
4347  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
4348  const unsigned NumUnknownBits =
4349  Bits - (Known.Zero | Known.One).countPopulation();
4350  assert(NumUnknownBits <= Bits);
4351  if (HasDefault && DeadCases.empty() &&
4352  NumUnknownBits < 64 /* avoid overflow */ &&
4353  SI->getNumCases() == (1ULL << NumUnknownBits)) {
4354  DEBUG(dbgs() << "SimplifyCFG: switch default is dead.\n");
4355  BasicBlock *NewDefault =
4357  SI->setDefaultDest(&*NewDefault);
4358  SplitBlock(&*NewDefault, &NewDefault->front());
4359  auto *OldTI = NewDefault->getTerminator();
4360  new UnreachableInst(SI->getContext(), OldTI);
4362  return true;
4363  }
4364 
4365  SmallVector<uint64_t, 8> Weights;
4366  bool HasWeight = HasBranchWeights(SI);
4367  if (HasWeight) {
4368  GetBranchWeights(SI, Weights);
4369  HasWeight = (Weights.size() == 1 + SI->getNumCases());
4370  }
4371 
4372  // Remove dead cases from the switch.
4373  for (ConstantInt *DeadCase : DeadCases) {
4374  SwitchInst::CaseIt CaseI = SI->findCaseValue(DeadCase);
4375  assert(CaseI != SI->case_default() &&
4376  "Case was not found. Probably mistake in DeadCases forming.");
4377  if (HasWeight) {
4378  std::swap(Weights[CaseI->getCaseIndex() + 1], Weights.back());
4379  Weights.pop_back();
4380  }
4381 
4382  // Prune unused values from PHI nodes.
4383  CaseI->getCaseSuccessor()->removePredecessor(SI->getParent());
4384  SI->removeCase(CaseI);
4385  }
4386  if (HasWeight && Weights.size() >= 2) {
4387  SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
4389  MDBuilder(SI->getParent()->getContext())
4390  .createBranchWeights(MDWeights));
4391  }
4392 
4393  return !DeadCases.empty();
4394 }
4395 
4396 /// If BB would be eligible for simplification by
4397 /// TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
4398 /// by an unconditional branch), look at the phi node for BB in the successor
4399 /// block and see if the incoming value is equal to CaseValue. If so, return
4400 /// the phi node, and set PhiIndex to BB's index in the phi node.
4402  BasicBlock *BB, int *PhiIndex) {
4403  if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
4404  return nullptr; // BB must be empty to be a candidate for simplification.
4405  if (!BB->getSinglePredecessor())
4406  return nullptr; // BB must be dominated by the switch.
4407 
4409  if (!Branch || !Branch->isUnconditional())
4410  return nullptr; // Terminator must be unconditional branch.
4411 
4412  BasicBlock *Succ = Branch->getSuccessor(0);
4413 
4414  BasicBlock::iterator I = Succ->begin();
4415  while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4416  int Idx = PHI->getBasicBlockIndex(BB);
4417  assert(Idx >= 0 && "PHI has no entry for predecessor?");
4418 
4419  Value *InValue = PHI->getIncomingValue(Idx);
4420  if (InValue != CaseValue)
4421  continue;
4422 
4423  *PhiIndex = Idx;
4424  return PHI;
4425  }
4426 
4427  return nullptr;
4428 }
4429 
4430 /// Try to forward the condition of a switch instruction to a phi node
4431 /// dominated by the switch, if that would mean that some of the destination
4432 /// blocks of the switch can be folded away.
4433 /// Returns true if a change is made.
4435  typedef DenseMap<PHINode *, SmallVector<int, 4>> ForwardingNodesMap;
4436  ForwardingNodesMap ForwardingNodes;
4437 
4438  for (auto Case : SI->cases()) {
4439  ConstantInt *CaseValue = Case.getCaseValue();
4440  BasicBlock *CaseDest = Case.getCaseSuccessor();
4441 
4442  int PhiIndex;
4443  PHINode *PHI =
4444  FindPHIForConditionForwarding(CaseValue, CaseDest, &PhiIndex);
4445  if (!PHI)
4446  continue;
4447 
4448  ForwardingNodes[PHI].push_back(PhiIndex);
4449  }
4450 
4451  bool Changed = false;
4452 
4453  for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
4454  E = ForwardingNodes.end();
4455  I != E; ++I) {
4456  PHINode *Phi = I->first;
4457  SmallVectorImpl<int> &Indexes = I->second;
4458 
4459  if (Indexes.size() < 2)
4460  continue;
4461 
4462  for (size_t I = 0, E = Indexes.size(); I != E; ++I)
4463  Phi->setIncomingValue(Indexes[I], SI->getCondition());
4464  Changed = true;
4465  }
4466 
4467  return Changed;
4468 }
4469 
4470 /// Return true if the backend will be able to handle
4471 /// initializing an array of constants like C.
4473  if (C->isThreadDependent())
4474  return false;
4475  if (C->isDLLImportDependent())
4476  return false;
4477 
4478  if (!isa<ConstantFP>(C) && !isa<ConstantInt>(C) &&
4479  !isa<ConstantPointerNull>(C) && !isa<GlobalValue>(C) &&
4480  !isa<UndefValue>(C) && !isa<ConstantExpr>(C))
4481  return false;
4482 
4483  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
4484  if (!CE->isGEPWithNoNotionalOverIndexing())
4485  return false;
4486  if (!ValidLookupTableConstant(CE->getOperand(0), TTI))
4487  return false;
4488  }
4489 
4491  return false;
4492 
4493  return true;
4494 }
4495 
4496 /// If V is a Constant, return it. Otherwise, try to look up
4497 /// its constant value in ConstantPool, returning 0 if it's not there.
4498 static Constant *
4501  if (Constant *C = dyn_cast<Constant>(V))
4502  return C;
4503  return ConstantPool.lookup(V);
4504 }
4505 
4506 /// Try to fold instruction I into a constant. This works for
4507 /// simple instructions such as binary operations where both operands are
4508 /// constant or can be replaced by constants from the ConstantPool. Returns the
4509 /// resulting constant on success, 0 otherwise.
4510 static Constant *
4513  if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
4514  Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
4515  if (!A)
4516  return nullptr;
4517  if (A->isAllOnesValue())
4518  return LookupConstant(Select->getTrueValue(), ConstantPool);
4519  if (A->isNullValue())
4520  return LookupConstant(Select->getFalseValue(), ConstantPool);
4521  return nullptr;
4522  }
4523 
4525  for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
4527  COps.push_back(A);
4528  else
4529  return nullptr;
4530  }
4531 
4532  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
4533  return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
4534  COps[1], DL);
4535  }
4536 
4537  return ConstantFoldInstOperands(I, COps, DL);
4538 }
4539 
4540 /// Try to determine the resulting constant values in phi nodes
4541 /// at the common destination basic block, *CommonDest, for one of the case
4542 /// destionations CaseDest corresponding to value CaseVal (0 for the default
4543 /// case), of a switch instruction SI.
4544 static bool
4546  BasicBlock **CommonDest,
4547  SmallVectorImpl<std::pair<PHINode *, Constant *>> &Res,
4548  const DataLayout &DL, const TargetTransformInfo &TTI) {
4549  // The block from which we enter the common destination.
4550  BasicBlock *Pred = SI->getParent();
4551 
4552  // If CaseDest is empty except for some side-effect free instructions through
4553  // which we can constant-propagate the CaseVal, continue to its successor.
4555  ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
4556  for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
4557  ++I) {
4558  if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
4559  // If the terminator is a simple branch, continue to the next block.
4560  if (T->getNumSuccessors() != 1 || T->isExceptional())
4561  return false;
4562  Pred = CaseDest;
4563  CaseDest = T->getSuccessor(0);
4564  } else if (isa<DbgInfoIntrinsic>(I)) {
4565  // Skip debug intrinsic.
4566  continue;
4567  } else if (Constant *C = ConstantFold(&*I, DL, ConstantPool)) {
4568  // Instruction is side-effect free and constant.
4569 
4570  // If the instruction has uses outside this block or a phi node slot for
4571  // the block, it is not safe to bypass the instruction since it would then
4572  // no longer dominate all its uses.
4573  for (auto &Use : I->uses()) {
4574  User *User = Use.getUser();
4575  if (Instruction *I = dyn_cast<Instruction>(User))
4576  if (I->getParent() == CaseDest)
4577  continue;
4578  if (PHINode *Phi = dyn_cast<PHINode>(User))
4579  if (Phi->getIncomingBlock(Use) == CaseDest)
4580  continue;
4581  return false;
4582  }
4583 
4584  ConstantPool.insert(std::make_pair(&*I, C));
4585  } else {
4586  break;
4587  }
4588  }
4589 
4590  // If we did not have a CommonDest before, use the current one.
4591  if (!*CommonDest)
4592  *CommonDest = CaseDest;
4593  // If the destination isn't the common one, abort.
4594  if (CaseDest != *CommonDest)
4595  return false;
4596 
4597  // Get the values for this case from phi nodes in the destination block.
4598  BasicBlock::iterator I = (*CommonDest)->begin();
4599  while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
4600  int Idx = PHI->getBasicBlockIndex(Pred);
4601  if (Idx == -1)
4602  continue;
4603 
4604  Constant *ConstVal =
4605  LookupConstant(PHI->getIncomingValue(Idx), ConstantPool);
4606  if (!ConstVal)
4607  return false;
4608 
4609  // Be conservative about which kinds of constants we support.
4610  if (!ValidLookupTableConstant(ConstVal, TTI))
4611  return false;
4612 
4613  Res.push_back(std::make_pair(PHI, ConstVal));
4614  }
4615 
4616  return Res.size() > 0;
4617 }
4618 
4619 // Helper function used to add CaseVal to the list of cases that generate
4620 // Result.
4621 static void MapCaseToResult(ConstantInt *CaseVal,
4622  SwitchCaseResultVectorTy &UniqueResults,
4623  Constant *Result) {
4624  for (auto &I : UniqueResults) {
4625  if (I.first == Result) {
4626  I.second.push_back(CaseVal);
4627  return;
4628  }
4629  }
4630  UniqueResults.push_back(
4631  std::make_pair(Result, SmallVector<ConstantInt *, 4>(1, CaseVal)));
4632 }
4633 
4634 // Helper function that initializes a map containing
4635 // results for the PHI node of the common destination block for a switch
4636 // instruction. Returns false if multiple PHI nodes have been found or if
4637 // there is not a common destination block for the switch.
4639  BasicBlock *&CommonDest,
4640  SwitchCaseResultVectorTy &UniqueResults,
4641  Constant *&DefaultResult,
4642  const DataLayout &DL,
4643  const TargetTransformInfo &TTI) {
4644  for (auto &I : SI->cases()) {
4645  ConstantInt *CaseVal = I.getCaseValue();
4646 
4647  // Resulting value at phi nodes for this case value.
4648  SwitchCaseResultsTy Results;
4649  if (!GetCaseResults(SI, CaseVal, I.getCaseSuccessor(), &CommonDest, Results,
4650  DL, TTI))
4651  return false;
4652 
4653  // Only one value per case is permitted
4654  if (Results.size() > 1)
4655  return false;
4656  MapCaseToResult(CaseVal, UniqueResults, Results.begin()->second);
4657 
4658  // Check the PHI consistency.
4659  if (!PHI)
4660  PHI = Results[0].first;
4661  else if (PHI != Results[0].first)
4662  return false;
4663  }
4664  // Find the default result value.
4666  BasicBlock *DefaultDest = SI->getDefaultDest();
4667  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest, DefaultResults,
4668  DL, TTI);
4669  // If the default value is not found abort unless the default destination
4670  // is unreachable.
4671  DefaultResult =
4672  DefaultResults.size() == 1 ? DefaultResults.begin()->second : nullptr;
4673  if ((!DefaultResult &&
4674  !isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg())))
4675  return false;
4676 
4677  return true;
4678 }
4679 
4680 // Helper function that checks if it is possible to transform a switch with only
4681 // two cases (or two cases + default) that produces a result into a select.
4682 // Example:
4683 // switch (a) {
4684 // case 10: %0 = icmp eq i32 %a, 10
4685 // return 10; %1 = select i1 %0, i32 10, i32 4
4686 // case 20: ----> %2 = icmp eq i32 %a, 20
4687 // return 2; %3 = select i1 %2, i32 2, i32 %1
4688 // default:
4689 // return 4;
4690 // }
4691 static Value *ConvertTwoCaseSwitch(const SwitchCaseResultVectorTy &ResultVector,
4692  Constant *DefaultResult, Value *Condition,
4693  IRBuilder<> &Builder) {
4694  assert(ResultVector.size() == 2 &&
4695  "We should have exactly two unique results at this point");
4696  // If we are selecting between only two cases transform into a simple
4697  // select or a two-way select if default is possible.
4698  if (ResultVector[0].second.size() == 1 &&
4699  ResultVector[1].second.size() == 1) {
4700  ConstantInt *const FirstCase = ResultVector[0].second[0];
4701  ConstantInt *const SecondCase = ResultVector[1].second[0];
4702 
4703  bool DefaultCanTrigger = DefaultResult;
4704  Value *SelectValue = ResultVector[1].first;
4705  if (DefaultCanTrigger) {
4706  Value *const ValueCompare =
4707  Builder.CreateICmpEQ(Condition, SecondCase, "switch.selectcmp");
4708  SelectValue = Builder.CreateSelect(ValueCompare, ResultVector[1].first,
4709  DefaultResult, "switch.select");
4710  }
4711  Value *const ValueCompare =
4712  Builder.CreateICmpEQ(Condition, FirstCase, "switch.selectcmp");
4713  return Builder.CreateSelect(ValueCompare, ResultVector[0].first,
4714  SelectValue, "switch.select");
4715  }
4716 
4717  return nullptr;
4718 }
4719 
4720 // Helper function to cleanup a switch instruction that has been converted into
4721 // a select, fixing up PHI nodes and basic blocks.
4723  Value *SelectValue,
4724  IRBuilder<> &Builder) {
4725  BasicBlock *SelectBB = SI->getParent();
4726  while (PHI->getBasicBlockIndex(SelectBB) >= 0)
4727  PHI->removeIncomingValue(SelectBB);
4728  PHI->addIncoming(SelectValue, SelectBB);
4729 
4730  Builder.CreateBr(PHI->getParent());
4731 
4732  // Remove the switch.
4733  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
4734  BasicBlock *Succ = SI->getSuccessor(i);
4735 
4736  if (Succ == PHI->getParent())
4737  continue;
4738  Succ->removePredecessor(SelectBB);
4739  }
4740  SI->eraseFromParent();
4741 }
4742 
4743 /// If the switch is only used to initialize one or more
4744 /// phi nodes in a common successor block with only two different
4745 /// constant values, replace the switch with select.
4746 static bool SwitchToSelect(SwitchInst *SI, IRBuilder<> &Builder,
4747  AssumptionCache *AC, const DataLayout &DL,
4748  const TargetTransformInfo &TTI) {
4749  Value *const Cond = SI->getCondition();
4750  PHINode *PHI = nullptr;
4751  BasicBlock *CommonDest = nullptr;
4752  Constant *DefaultResult;
4753  SwitchCaseResultVectorTy UniqueResults;
4754  // Collect all the cases that will deliver the same value from the switch.
4755  if (!InitializeUniqueCases(SI, PHI, CommonDest, UniqueResults, DefaultResult,
4756  DL, TTI))
4757  return false;
4758  // Selects choose between maximum two values.
4759  if (UniqueResults.size() != 2)
4760  return false;
4761  assert(PHI != nullptr && "PHI for value select not found");
4762 
4763  Builder.SetInsertPoint(SI);
4764  Value *SelectValue =
4765  ConvertTwoCaseSwitch(UniqueResults, DefaultResult, Cond, Builder);
4766  if (SelectValue) {
4767  RemoveSwitchAfterSelectConversion(SI, PHI, SelectValue, Builder);
4768  return true;
4769  }
4770  // The switch couldn't be converted into a select.
4771  return false;
4772 }
4773 
4774 namespace {
4775 
4776 /// This class represents a lookup table that can be used to replace a switch.
4777 class SwitchLookupTable {
4778 public:
4779  /// Create a lookup table to use as a switch replacement with the contents
4780  /// of Values, using DefaultValue to fill any holes in the table.
4781  SwitchLookupTable(
4782  Module &M, uint64_t TableSize, ConstantInt *Offset,
4783  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4784  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName);
4785 
4786  /// Build instructions with Builder to retrieve the value at
4787  /// the position given by Index in the lookup table.
4788  Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
4789 
4790  /// Return true if a table with TableSize elements of
4791  /// type ElementType would fit in a target-legal register.
4792  static bool WouldFitInRegister(const DataLayout &DL, uint64_t TableSize,
4793  Type *ElementType);
4794 
4795 private:
4796  // Depending on the contents of the table, it can be represented in
4797  // different ways.
4798  enum {
4799  // For tables where each element contains the same value, we just have to
4800  // store that single value and return it for each lookup.
4801  SingleValueKind,
4802 
4803  // For tables where there is a linear relationship between table index
4804  // and values. We calculate the result with a simple multiplication
4805  // and addition instead of a table lookup.
4806  LinearMapKind,
4807 
4808  // For small tables with integer elements, we can pack them into a bitmap
4809  // that fits into a target-legal register. Values are retrieved by
4810  // shift and mask operations.
4811  BitMapKind,
4812 
4813  // The table is stored as an array of values. Values are retrieved by load
4814  // instructions from the table.
4815  ArrayKind
4816  } Kind;
4817 
4818  // For SingleValueKind, this is the single value.
4819  Constant *SingleValue;
4820 
4821  // For BitMapKind, this is the bitmap.
4822  ConstantInt *BitMap;
4823  IntegerType *BitMapElementTy;
4824 
4825  // For LinearMapKind, these are the constants used to derive the value.
4826  ConstantInt *LinearOffset;
4827  ConstantInt *LinearMultiplier;
4828 
4829  // For ArrayKind, this is the array.
4830  GlobalVariable *Array;
4831 };
4832 
4833 } // end anonymous namespace
4834 
4835 SwitchLookupTable::SwitchLookupTable(
4836  Module &M, uint64_t TableSize, ConstantInt *Offset,
4837  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values,
4838  Constant *DefaultValue, const DataLayout &DL, const StringRef &FuncName)
4839  : SingleValue(nullptr), BitMap(nullptr), BitMapElementTy(nullptr),
4840  LinearOffset(nullptr), LinearMultiplier(nullptr), Array(nullptr) {
4841  assert(Values.size() && "Can't build lookup table without values!");
4842  assert(TableSize >= Values.size() && "Can't fit values in table!");
4843 
4844  // If all values in the table are equal, this is that value.
4845  SingleValue = Values.begin()->second;
4846 
4847  Type *ValueType = Values.begin()->second->getType();
4848 
4849  // Build up the table contents.
4850  SmallVector<Constant *, 64> TableContents(TableSize);
4851  for (size_t I = 0, E = Values.size(); I != E; ++I) {
4852  ConstantInt *CaseVal = Values[I].first;
4853  Constant *CaseRes = Values[I].second;
4854  assert(CaseRes->getType() == ValueType);
4855 
4856  uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
4857  TableContents[Idx] = CaseRes;
4858 
4859  if (CaseRes != SingleValue)
4860  SingleValue = nullptr;
4861  }
4862 
4863  // Fill in any holes in the table with the default result.
4864  if (Values.size() < TableSize) {
4865  assert(DefaultValue &&
4866  "Need a default value to fill the lookup table holes.");
4867  assert(DefaultValue->getType() == ValueType);
4868  for (uint64_t I = 0; I < TableSize; ++I) {
4869  if (!TableContents[I])
4870  TableContents[I] = DefaultValue;
4871  }
4872 
4873  if (DefaultValue != SingleValue)
4874  SingleValue = nullptr;
4875  }
4876 
4877  // If each element in the table contains the same value, we only need to store
4878  // that single value.
4879  if (SingleValue) {
4880  Kind = SingleValueKind;
4881  return;
4882  }
4883 
4884  // Check if we can derive the value with a linear transformation from the
4885  // table index.
4886  if (isa<IntegerType>(ValueType)) {
4887  bool LinearMappingPossible = true;
4888  APInt PrevVal;
4889  APInt DistToPrev;
4890  assert(TableSize >= 2 && "Should be a SingleValue table.");
4891  // Check if there is the same distance between two consecutive values.
4892  for (uint64_t I = 0; I < TableSize; ++I) {
4893  ConstantInt *ConstVal = dyn_cast<ConstantInt>(TableContents[I]);
4894  if (!ConstVal) {
4895  // This is an undef. We could deal with it, but undefs in lookup tables
4896  // are very seldom. It's probably not worth the additional complexity.
4897  LinearMappingPossible = false;
4898  break;
4899  }
4900  const APInt &Val = ConstVal->getValue();
4901  if (I != 0) {
4902  APInt Dist = Val - PrevVal;
4903  if (I == 1) {
4904  DistToPrev = Dist;
4905  } else if (Dist != DistToPrev) {
4906  LinearMappingPossible = false;
4907  break;
4908  }
4909  }
4910  PrevVal = Val;
4911  }
4912  if (LinearMappingPossible) {
4913  LinearOffset = cast<ConstantInt>(TableContents[0]);
4914  LinearMultiplier = ConstantInt::get(M.getContext(), DistToPrev);
4915  Kind = LinearMapKind;
4916  ++NumLinearMaps;
4917  return;
4918  }
4919  }
4920 
4921  // If the type is integer and the table fits in a register, build a bitmap.
4922  if (WouldFitInRegister(DL, TableSize, ValueType)) {
4923  IntegerType *IT = cast<IntegerType>(ValueType);
4924  APInt TableInt(TableSize * IT->getBitWidth(), 0);
4925  for (uint64_t I = TableSize; I > 0; --I) {
4926  TableInt <<= IT->getBitWidth();
4927  // Insert values into the bitmap. Undef values are set to zero.
4928  if (!isa<UndefValue>(TableContents[I - 1])) {
4929  ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
4930  TableInt |= Val->getValue().zext(TableInt.getBitWidth());
4931  }
4932  }
4933  BitMap = ConstantInt::get(M.getContext(), TableInt);
4934  BitMapElementTy = IT;
4935  Kind = BitMapKind;
4936  ++NumBitMaps;
4937  return;
4938  }
4939 
4940  // Store the table in an array.
4941  ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
4942  Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
4943 
4944  Array = new GlobalVariable(M, ArrayTy, /*constant=*/true,
4945  GlobalVariable::PrivateLinkage, Initializer,
4946  "switch.table." + FuncName);
4947  Array->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
4948  Kind = ArrayKind;
4949 }
4950 
4951 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
4952  switch (Kind) {
4953  case SingleValueKind:
4954  return SingleValue;
4955  case LinearMapKind: {
4956  // Derive the result value from the input value.
4957  Value *Result = Builder.CreateIntCast(Index, LinearMultiplier->getType(),
4958  false, "switch.idx.cast");
4959  if (!LinearMultiplier->isOne())
4960  Result = Builder.CreateMul(Result, LinearMultiplier, "switch.idx.mult");
4961  if (!LinearOffset->isZero())
4962  Result = Builder.CreateAdd(Result, LinearOffset, "switch.offset");
4963  return Result;
4964  }
4965  case BitMapKind: {
4966  // Type of the bitmap (e.g. i59).
4967  IntegerType *MapTy = BitMap->getType();
4968 
4969  // Cast Index to the same type as the bitmap.
4970  // Note: The Index is <= the number of elements in the table, so
4971  // truncating it to the width of the bitmask is safe.
4972  Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
4973 
4974  // Multiply the shift amount by the element width.
4975  ShiftAmt = Builder.CreateMul(
4976  ShiftAmt, ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
4977  "switch.shiftamt");
4978 
4979  // Shift down.
4980  Value *DownShifted =
4981  Builder.CreateLShr(BitMap, ShiftAmt, "switch.downshift");
4982  // Mask off.
4983  return Builder.CreateTrunc(DownShifted, BitMapElementTy, "switch.masked");
4984  }
4985  case ArrayKind: {
4986  // Make sure the table index will not overflow when treated as signed.
4987  IntegerType *IT = cast<IntegerType>(Index->getType());
4988  uint64_t TableSize =
4989  Array->getInitializer()->getType()->getArrayNumElements();
4990  if (TableSize > (1ULL << (IT->getBitWidth() - 1)))
4991  Index = Builder.CreateZExt(
4992  Index, IntegerType::get(IT->getContext(), IT->getBitWidth() + 1),
4993  "switch.tableidx.zext");
4994 
4995  Value *GEPIndices[] = {Builder.getInt32(0), Index};
4996  Value *GEP = Builder.CreateInBoundsGEP(Array->getValueType(), Array,
4997  GEPIndices, "switch.gep");
4998  return Builder.CreateLoad(GEP, "switch.load");
4999  }
5000  }
5001  llvm_unreachable("Unknown lookup table kind!");
5002 }
5003 
5004 bool SwitchLookupTable::WouldFitInRegister(const DataLayout &DL,
5005  uint64_t TableSize,
5006  Type *ElementType) {
5007  auto *IT = dyn_cast<IntegerType>(ElementType);
5008  if (!IT)
5009  return false;
5010  // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
5011  // are <= 15, we could try to narrow the type.
5012 
5013  // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
5014  if (TableSize >= UINT_MAX / IT->getBitWidth())
5015  return false;
5016  return DL.fitsInLegalInteger(TableSize * IT->getBitWidth());
5017 }
5018 
5019 /// Determine whether a lookup table should be built for this switch, based on
5020 /// the number of cases, size of the table, and the types of the results.
5021 static bool
5022 ShouldBuildLookupTable(SwitchInst *SI, uint64_t TableSize,
5023  const TargetTransformInfo &TTI, const DataLayout &DL,
5024  const SmallDenseMap<PHINode *, Type *> &ResultTypes) {
5025  if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
5026  return false; // TableSize overflowed, or mul below might overflow.
5027 
5028  bool AllTablesFitInRegister = true;
5029  bool HasIllegalType = false;
5030  for (const auto &I : ResultTypes) {
5031  Type *Ty = I.second;
5032 
5033  // Saturate this flag to true.
5034  HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
5035 
5036  // Saturate this flag to false.
5037  AllTablesFitInRegister =
5038  AllTablesFitInRegister &&
5039  SwitchLookupTable::WouldFitInRegister(DL, TableSize, Ty);
5040 
5041  // If both flags saturate, we're done. NOTE: This *only* works with
5042  // saturating flags, and all flags have to saturate first due to the
5043  // non-deterministic behavior of iterating over a dense map.
5044  if (HasIllegalType && !AllTablesFitInRegister)
5045  break;
5046  }
5047 
5048  // If each table would fit in a register, we should build it anyway.
5049  if (AllTablesFitInRegister)
5050  return true;
5051 
5052  // Don't build a table that doesn't fit in-register if it has illegal types.
5053  if (HasIllegalType)
5054  return false;
5055 
5056  // The table density should be at least 40%. This is the same criterion as for
5057  // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
5058  // FIXME: Find the best cut-off.
5059  return SI->getNumCases() * 10 >= TableSize * 4;
5060 }
5061 
5062 /// Try to reuse the switch table index compare. Following pattern:
5063 /// \code
5064 /// if (idx < tablesize)
5065 /// r = table[idx]; // table does not contain default_value
5066 /// else
5067 /// r = default_value;
5068 /// if (r != default_value)
5069 /// ...
5070 /// \endcode
5071 /// Is optimized to:
5072 /// \code
5073 /// cond = idx < tablesize;
5074 /// if (cond)
5075 /// r = table[idx];
5076 /// else
5077 /// r = default_value;
5078 /// if (cond)
5079 /// ...
5080 /// \endcode
5081 /// Jump threading will then eliminate the second if(cond).
5082 static void reuseTableCompare(
5083  User *PhiUser, BasicBlock *PhiBlock, BranchInst *RangeCheckBranch,
5084  Constant *DefaultValue,
5085  const SmallVectorImpl<std::pair<ConstantInt *, Constant *>> &Values) {
5086 
5087  ICmpInst *CmpInst = dyn_cast<ICmpInst>(PhiUser);
5088  if (!CmpInst)
5089  return;
5090 
5091  // We require that the compare is in the same block as the phi so that jump
5092  // threading can do its work afterwards.
5093  if (CmpInst->getParent() != PhiBlock)
5094  return;
5095 
5096  Constant *CmpOp1 = dyn_cast<Constant>(CmpInst->getOperand(1));
5097  if (!CmpOp1)
5098  return;
5099 
5100  Value *RangeCmp = RangeCheckBranch->getCondition();
5101  Constant *TrueConst = ConstantInt::getTrue(RangeCmp->getType());
5102  Constant *FalseConst = ConstantInt::getFalse(RangeCmp->getType());
5103 
5104  // Check if the compare with the default value is constant true or false.
5105  Constant *DefaultConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5106  DefaultValue, CmpOp1, true);
5107  if (DefaultConst != TrueConst && DefaultConst != FalseConst)
5108  return;
5109 
5110  // Check if the compare with the case values is distinct from the default
5111  // compare result.
5112  for (auto ValuePair : Values) {
5113  Constant *CaseConst = ConstantExpr::getICmp(CmpInst->getPredicate(),
5114  ValuePair.second, CmpOp1, true);
5115  if (!CaseConst || CaseConst == DefaultConst)
5116  return;
5117  assert((CaseConst == TrueConst || CaseConst == FalseConst) &&
5118  "Expect true or false as compare result.");
5119  }
5120 
5121  // Check if the branch instruction dominates the phi node. It's a simple
5122  // dominance check, but sufficient for our needs.
5123  // Although this check is invariant in the calling loops, it's better to do it
5124  // at this late stage. Practically we do it at most once for a switch.
5125  BasicBlock *BranchBlock = RangeCheckBranch->getParent();
5126  for (auto PI = pred_begin(PhiBlock), E = pred_end(PhiBlock); PI != E; ++PI) {
5127  BasicBlock *Pred = *PI;
5128  if (Pred != BranchBlock && Pred->getUniquePredecessor() != BranchBlock)
5129  return;
5130  }
5131 
5132  if (DefaultConst == FalseConst) {
5133  // The compare yields the same result. We can replace it.
5134  CmpInst->replaceAllUsesWith(RangeCmp);
5135  ++NumTableCmpReuses;
5136  } else {
5137  // The compare yields the same result, just inverted. We can replace it.
5138  Value *InvertedTableCmp = BinaryOperator::CreateXor(
5139  RangeCmp, ConstantInt::get(RangeCmp->getType(), 1), "inverted.cmp",
5140  RangeCheckBranch);
5141  CmpInst->replaceAllUsesWith(InvertedTableCmp);
5142  ++NumTableCmpReuses;
5143  }
5144 }
5145 
5146 /// If the switch is only used to initialize one or more phi nodes in a common
5147 /// successor block with different constant values, replace the switch with
5148 /// lookup tables.
5149 static bool SwitchToLookupTable(SwitchInst *SI, IRBuilder<> &Builder,
5150  const DataLayout &DL,
5151  const TargetTransformInfo &TTI) {
5152  assert(SI->getNumCases() > 1 && "Degenerate switch?");
5153 
5154  // Only build lookup table when we have a target that supports it.
5155  if (!TTI.shouldBuildLookupTables())
5156  return false;
5157 
5158  // FIXME: If the switch is too sparse for a lookup table, perhaps we could
5159  // split off a dense part and build a lookup table for that.
5160 
5161  // FIXME: This creates arrays of GEPs to constant strings, which means each
5162  // GEP needs a runtime relocation in PIC code. We should just build one big
5163  // string and lookup indices into that.
5164 
5165  // Ignore switches with less than three cases. Lookup tables will not make
5166  // them
5167  // faster, so we don't analyze them.
5168  if (SI->getNumCases() < 3)
5169  return false;
5170 
5171  // Figure out the corresponding result for each case value and phi node in the
5172  // common destination, as well as the min and max case values.
5173  assert(SI->case_begin() != SI->case_end());
5174  SwitchInst::CaseIt CI = SI->case_begin();
5175  ConstantInt *MinCaseVal = CI->getCaseValue();
5176  ConstantInt *MaxCaseVal = CI->getCaseValue();
5177 
5178  BasicBlock *CommonDest = nullptr;
5179  typedef SmallVector<std::pair<ConstantInt *, Constant *>, 4> ResultListTy;
5181  SmallDenseMap<PHINode *, Constant *> DefaultResults;
5184 
5185  for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
5186  ConstantInt *CaseVal = CI->getCaseValue();
5187  if (CaseVal->getValue().slt(MinCaseVal->getValue()))
5188  MinCaseVal = CaseVal;
5189  if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
5190  MaxCaseVal = CaseVal;
5191 
5192  // Resulting value at phi nodes for this case value.
5193  typedef SmallVector<std::pair<PHINode *, Constant *>, 4> ResultsTy;
5194  ResultsTy Results;
5195  if (!GetCaseResults(SI, CaseVal, CI->getCaseSuccessor(), &CommonDest,
5196  Results, DL, TTI))
5197  return false;
5198 
5199  // Append the result from this case to the list for each phi.
5200  for (const auto &I : Results) {
5201  PHINode *PHI = I.first;
5202  Constant *Value = I.second;
5203  if (!ResultLists.count(PHI))
5204  PHIs.push_back(PHI);
5205  ResultLists[PHI].push_back(std::make_pair(CaseVal, Value));
5206  }
5207  }
5208 
5209  // Keep track of the result types.
5210  for (PHINode *PHI : PHIs) {
5211  ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
5212  }
5213 
5214  uint64_t NumResults = ResultLists[PHIs[0]].size();
5215  APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
5216  uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
5217  bool TableHasHoles = (NumResults < TableSize);
5218 
5219  // If the table has holes, we need a constant result for the default case
5220  // or a bitmask that fits in a register.
5221  SmallVector<std::pair<PHINode *, Constant *>, 4> DefaultResultsList;
5222  bool HasDefaultResults =
5223  GetCaseResults(SI, nullptr, SI->getDefaultDest(), &CommonDest,
5224  DefaultResultsList, DL, TTI);
5225 
5226  bool NeedMask = (TableHasHoles && !HasDefaultResults);
5227  if (NeedMask) {
5228  // As an extra penalty for the validity test we require more cases.
5229  if (SI->getNumCases() < 4) // FIXME: Find best threshold value (benchmark).
5230  return false;
5231  if (!DL.fitsInLegalInteger(TableSize))
5232  return false;
5233  }
5234 
5235  for (const auto &I : DefaultResultsList) {
5236  PHINode *PHI = I.first;
5237  Constant *Result = I.second;
5238  DefaultResults[PHI] = Result;
5239  }
5240 
5241  if (!ShouldBuildLookupTable(SI, TableSize, TTI, DL, ResultTypes))
5242  return false;
5243 
5244  // Create the BB that does the lookups.
5245  Module &Mod = *CommonDest->getParent()->getParent();
5246  BasicBlock *LookupBB = BasicBlock::Create(
5247  Mod.getContext(), "switch.lookup", CommonDest->getParent(), CommonDest);
5248 
5249  // Compute the table index value.
5250  Builder.SetInsertPoint(SI);
5251  Value *TableIndex =
5252  Builder.CreateSub(SI->getCondition(), MinCaseVal, "switch.tableidx");
5253 
5254  // Compute the maximum table size representable by the integer type we are
5255  // switching upon.
5256  unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
5257  uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
5258  assert(MaxTableSize >= TableSize &&
5259  "It is impossible for a switch to have more entries than the max "
5260  "representable value of its input integer type's size.");
5261 
5262  // If the default destination is unreachable, or if the lookup table covers
5263  // all values of the conditional variable, branch directly to the lookup table
5264  // BB. Otherwise, check that the condition is within the case range.
5265  const bool DefaultIsReachable =
5266  !isa<UnreachableInst>(SI->getDefaultDest()->getFirstNonPHIOrDbg());
5267  const bool GeneratingCoveredLookupTable = (MaxTableSize == TableSize);
5268  BranchInst *RangeCheckBranch = nullptr;
5269 
5270  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5271  Builder.CreateBr(LookupBB);
5272  // Note: We call removeProdecessor later since we need to be able to get the
5273  // PHI value for the default case in case we're using a bit mask.
5274  } else {
5275  Value *Cmp = Builder.CreateICmpULT(
5276  TableIndex, ConstantInt::get(MinCaseVal->getType(), TableSize));
5277  RangeCheckBranch =
5278  Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
5279  }
5280 
5281  // Populate the BB that does the lookups.
5282  Builder.SetInsertPoint(LookupBB);
5283 
5284  if (NeedMask) {
5285  // Before doing the lookup we do the hole check.
5286  // The LookupBB is therefore re-purposed to do the hole check
5287  // and we create a new LookupBB.
5288  BasicBlock *MaskBB = LookupBB;
5289  MaskBB->setName("switch.hole_check");
5290  LookupBB = BasicBlock::Create(Mod.getContext(), "switch.lookup",
5291  CommonDest->getParent(), CommonDest);
5292 
5293  // Make the mask's bitwidth at least 8bit and a power-of-2 to avoid
5294  // unnecessary illegal types.
5295  uint64_t TableSizePowOf2 = NextPowerOf2(std::max(7ULL, TableSize - 1ULL));
5296  APInt MaskInt(TableSizePowOf2, 0);
5297  APInt One(TableSizePowOf2, 1);
5298  // Build bitmask; fill in a 1 bit for every case.
5299  const ResultListTy &ResultList = ResultLists[PHIs[0]];
5300  for (size_t I = 0, E = ResultList.size(); I != E; ++I) {
5301  uint64_t Idx = (ResultList[I].first->getValue() - MinCaseVal->getValue())
5302  .getLimitedValue();
5303  MaskInt |= One << Idx;
5304  }
5305  ConstantInt *TableMask = ConstantInt::get(Mod.getContext(), MaskInt);
5306 
5307  // Get the TableIndex'th bit of the bitmask.
5308  // If this bit is 0 (meaning hole) jump to the default destination,
5309  // else continue with table lookup.
5310  IntegerType *MapTy = TableMask->getType();
5311  Value *MaskIndex =
5312  Builder.CreateZExtOrTrunc(TableIndex, MapTy, "switch.maskindex");
5313  Value *Shifted = Builder.CreateLShr(TableMask, MaskIndex, "switch.shifted");
5314  Value *LoBit = Builder.CreateTrunc(
5315  Shifted, Type::getInt1Ty(Mod.getContext()), "switch.lobit");
5316  Builder.CreateCondBr(LoBit, LookupBB, SI->getDefaultDest());
5317 
5318  Builder.SetInsertPoint(LookupBB);
5319  AddPredecessorToBlock(SI->getDefaultDest(), MaskBB, SI->getParent());
5320  }
5321 
5322  if (!DefaultIsReachable || GeneratingCoveredLookupTable) {
5323  // We cached PHINodes in PHIs, to avoid accessing deleted PHINodes later,
5324  // do not delete PHINodes here.
5326  /*DontDeleteUselessPHIs=*/true);
5327  }
5328 
5329  bool ReturnedEarly = false;
5330  for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
5331  PHINode *PHI = PHIs[I];
5332  const ResultListTy &ResultList = ResultLists[PHI];
5333 
5334  // If using a bitmask, use any value to fill the lookup table holes.
5335  Constant *DV = NeedMask ? ResultLists[PHI][0].second : DefaultResults[PHI];
5336  StringRef FuncName = SI->getParent()->getParent()->getName();
5337  SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultList, DV, DL,
5338  FuncName);
5339 
5340  Value *Result = Table.BuildLookup(TableIndex, Builder);
5341 
5342  // If the result is used to return immediately from the function, we want to
5343  // do that right here.
5344  if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->user_begin()) &&
5345  PHI->user_back() == CommonDest->getFirstNonPHIOrDbg()) {
5346  Builder.CreateRet(Result);
5347  ReturnedEarly = true;
5348  break;
5349  }
5350 
5351  // Do a small peephole optimization: re-use the switch table compare if
5352  // possible.
5353  if (!TableHasHoles && HasDefaultResults && RangeCheckBranch) {
5354  BasicBlock *PhiBlock = PHI->getParent();
5355  // Search for compare instructions which use the phi.
5356  for (auto *User : PHI->users()) {
5357  reuseTableCompare(User, PhiBlock, RangeCheckBranch, DV, ResultList);
5358  }
5359  }
5360 
5361  PHI->addIncoming(Result, LookupBB);
5362  }
5363 
5364  if (!ReturnedEarly)
5365  Builder.CreateBr(CommonDest);
5366 
5367  // Remove the switch.
5368  for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
5369  BasicBlock *Succ = SI->getSuccessor(i);
5370 
5371  if (Succ == SI->getDefaultDest())
5372  continue;
5373  Succ->removePredecessor(SI->getParent());
5374  }
5375  SI->eraseFromParent();
5376 
5377  ++NumLookupTables;
5378  if (NeedMask)
5379  ++NumLookupTablesHoles;
5380  return true;
5381 }
5382 
5383 static bool isSwitchDense(ArrayRef<int64_t> Values) {
5384  // See also SelectionDAGBuilder::isDense(), which this function was based on.
5385  uint64_t Diff = (uint64_t)Values.back() - (uint64_t)Values.front();
5386  uint64_t Range = Diff + 1;
5387  uint64_t NumCases = Values.size();
5388  // 40% is the default density for building a jump table in optsize/minsize mode.
5389  uint64_t MinDensity = 40;
5390 
5391  return NumCases * 100 >= Range * MinDensity;
5392 }
5393 
5394 // Try and transform a switch that has "holes" in it to a contiguous sequence
5395 // of cases.
5396 //
5397 // A switch such as: switch(i) {case 5: case 9: case 13: case 17:} can be
5398 // range-reduced to: switch ((i-5) / 4) {case 0: case 1: case 2: case 3:}.
5399 //
5400 // This converts a sparse switch into a dense switch which allows better
5401 // lowering and could also allow transforming into a lookup table.
5402 static bool ReduceSwitchRange(SwitchInst *SI, IRBuilder<> &Builder,
5403  const DataLayout &DL,
5404  const TargetTransformInfo &TTI) {
5405  auto *CondTy = cast<IntegerType>(SI->getCondition()->getType());
5406  if (CondTy->getIntegerBitWidth() > 64 ||
5407  !DL.fitsInLegalInteger(CondTy->getIntegerBitWidth()))
5408  return false;
5409  // Only bother with this optimization if there are more than 3 switch cases;
5410  // SDAG will only bother creating jump tables for 4 or more cases.
5411  if (SI->getNumCases() < 4)
5412  return false;
5413 
5414  // This transform is agnostic to the signedness of the input or case values. We
5415  // can treat the case values as signed or unsigned. We can optimize more common
5416  // cases such as a sequence crossing zero {-4,0,4,8} if we interpret case values
5417  // as signed.
5418  SmallVector<int64_t,4> Values;
5419  for (auto &C : SI->cases())
5420  Values.push_back(C.getCaseValue()->getValue().getSExtValue());
5421  std::sort(Values.begin(), Values.end());
5422 
5423  // If the switch is already dense, there's nothing useful to do here.
5424  if (isSwitchDense(Values))
5425  return false;
5426 
5427  // First, transform the values such that they start at zero and ascend.
5428  int64_t Base = Values[0];
5429  for (auto &V : Values)
5430  V -= Base;
5431 
5432  // Now we have signed numbers that have been shifted so that, given enough
5433  // precision, there are no negative values. Since the rest of the transform
5434  // is bitwise only, we switch now to an unsigned representation.
5435  uint64_t GCD = 0;
5436  for (auto &V : Values)
5437  GCD = GreatestCommonDivisor64(GCD, (uint64_t)V);
5438 
5439  // This transform can be done speculatively because it is so cheap - it results
5440  // in a single rotate operation being inserted. This can only happen if the
5441  // factor extracted is a power of 2.
5442  // FIXME: If the GCD is an odd number we can multiply by the multiplicative
5443  // inverse of GCD and then perform this transform.
5444  // FIXME: It's possible that optimizing a switch on powers of two might also
5445  // be beneficial - flag values are often powers of two and we could use a CLZ
5446  // as the key function.
5447  if (GCD <= 1 || !isPowerOf2_64(GCD))
5448  // No common divisor found or too expensive to compute key function.
5449  return false;
5450 
5451  unsigned Shift = Log2_64(GCD);
5452  for (auto &V : Values)
5453  V = (int64_t)((uint64_t)V >> Shift);
5454 
5455  if (!isSwitchDense(Values))
5456  // Transform didn't create a dense switch.
5457  return false;
5458 
5459  // The obvious transform is to shift the switch condition right and emit a
5460  // check that the condition actually cleanly divided by GCD, i.e.
5461  // C & (1 << Shift - 1) == 0
5462  // inserting a new CFG edge to handle the case where it didn't divide cleanly.
5463  //
5464  // A cheaper way of doing this is a simple ROTR(C, Shift). This performs the
5465  // shift and puts the shifted-off bits in the uppermost bits. If any of these
5466  // are nonzero then the switch condition will be very large and will hit the
5467  // default case.
5468 
5469  auto *Ty = cast<IntegerType>(SI->getCondition()->getType());
5470  Builder.SetInsertPoint(SI);
5471  auto *ShiftC = ConstantInt::get(Ty, Shift);
5472  auto *Sub = Builder.CreateSub(SI->getCondition(), ConstantInt::get(Ty, Base));
5473  auto *LShr = Builder.CreateLShr(Sub, ShiftC);
5474  auto *Shl = Builder.CreateShl(Sub, Ty->getBitWidth() - Shift);
5475  auto *Rot = Builder.CreateOr(LShr, Shl);
5476  SI->replaceUsesOfWith(SI->getCondition(), Rot);
5477 
5478  for (auto Case : SI->cases()) {
5479  auto *Orig = Case.getCaseValue();
5480  auto Sub = Orig->getValue() - APInt(Ty->getBitWidth(), Base);
5481  Case.setValue(
5482  cast<ConstantInt>(ConstantInt::get(Ty, Sub.lshr(ShiftC->getValue()))));
5483  }
5484  return true;
5485 }
5486 
5487 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
5488  BasicBlock *BB = SI->getParent();
5489 
5490  if (isValueEqualityComparison(SI)) {
5491  // If we only have one predecessor, and if it is a branch on this value,
5492  // see if that predecessor totally determines the outcome of this switch.
5493  if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
5494  if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
5495  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5496 
5497  Value *Cond = SI->getCondition();
5498  if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
5499  if (SimplifySwitchOnSelect(SI, Select))
5500  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5501 
5502  // If the block only contains the switch, see if we can fold the block
5503  // away into any preds.
5504  BasicBlock::iterator BBI = BB->begin();
5505  // Ignore dbg intrinsics.
5506  while (isa<DbgInfoIntrinsic>(BBI))
5507  ++BBI;
5508  if (SI == &*BBI)
5509  if (FoldValueComparisonIntoPredecessors(SI, Builder))
5510  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5511  }
5512 
5513  // Try to transform the switch into an icmp and a branch.
5514  if (TurnSwitchRangeIntoICmp(SI, Builder))
5515  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5516 
5517  // Remove unreachable cases.
5518  if (EliminateDeadSwitchCases(SI, AC, DL))
5519  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5520 
5521  if (SwitchToSelect(SI, Builder, AC, DL, TTI))
5522  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5523 
5525  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5526 
5527  // The conversion from switch to lookup tables results in difficult
5528  // to analyze code and makes pruning branches much harder.
5529  // This is a problem of the switch expression itself can still be
5530  // restricted as a result of inlining or CVP. There only apply this
5531  // transformation during late steps of the optimisation chain.
5532  if (LateSimplifyCFG && SwitchToLookupTable(SI, Builder, DL, TTI))
5533  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5534 
5535  if (ReduceSwitchRange(SI, Builder, DL, TTI))
5536  return SimplifyCFG(BB, TTI, BonusInstThreshold, AC) | true;
5537 
5538  return false;
5539 }
5540 
5541 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
5542  BasicBlock *BB = IBI->getParent();
5543  bool Changed = false;
5544 
5545  // Eliminate redundant destinations.
5547  for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
5548  BasicBlock *Dest = IBI->getDestination(i);
5549  if (!Dest->hasAddressTaken() || !Succs.insert(Dest).second) {
5550  Dest->removePredecessor(BB);
5551  IBI->removeDestination(i);
5552  --i;
5553  --e;
5554  Changed = true;
5555  }
5556  }
5557 
5558  if (IBI->getNumDestinations() == 0) {
5559  // If the indirectbr has no successors, change it to unreachable.
5560  new UnreachableInst(IBI->getContext(), IBI);
5562  return true;
5563  }
5564 
5565  if (IBI->getNumDestinations() == 1) {
5566  // If the indirectbr has one successor, change it to a direct branch.
5567  BranchInst::Create(IBI->getDestination(0), IBI);
5569  return true;
5570