• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #define DEBUG_TYPE "simplifycfg"
15 #include "llvm/Transforms/Utils/Local.h"
16 #include "llvm/Constants.h"
17 #include "llvm/DerivedTypes.h"
18 #include "llvm/GlobalVariable.h"
19 #include "llvm/IRBuilder.h"
20 #include "llvm/Instructions.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/MDBuilder.h"
24 #include "llvm/Metadata.h"
25 #include "llvm/Module.h"
26 #include "llvm/Operator.h"
27 #include "llvm/Type.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/ADT/SetVector.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/Analysis/InstructionSimplify.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/Support/CFG.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/ConstantRange.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/NoFolder.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Target/TargetData.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include <algorithm>
45 #include <set>
46 #include <map>
47 using namespace llvm;
48 
49 static cl::opt<unsigned>
50 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
51    cl::desc("Control the amount of phi node folding to perform (default = 1)"));
52 
53 static cl::opt<bool>
54 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
55        cl::desc("Duplicate return instructions into unconditional branches"));
56 
57 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
58 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
59 
60 namespace {
61   /// ValueEqualityComparisonCase - Represents a case of a switch.
62   struct ValueEqualityComparisonCase {
63     ConstantInt *Value;
64     BasicBlock *Dest;
65 
ValueEqualityComparisonCase__anon08d918330111::ValueEqualityComparisonCase66     ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
67       : Value(Value), Dest(Dest) {}
68 
operator <__anon08d918330111::ValueEqualityComparisonCase69     bool operator<(ValueEqualityComparisonCase RHS) const {
70       // Comparing pointers is ok as we only rely on the order for uniquing.
71       return Value < RHS.Value;
72     }
73   };
74 
75 class SimplifyCFGOpt {
76   const TargetData *const TD;
77 
78   Value *isValueEqualityComparison(TerminatorInst *TI);
79   BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
80                                std::vector<ValueEqualityComparisonCase> &Cases);
81   bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
82                                                      BasicBlock *Pred,
83                                                      IRBuilder<> &Builder);
84   bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
85                                            IRBuilder<> &Builder);
86 
87   bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
88   bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
89   bool SimplifyUnreachable(UnreachableInst *UI);
90   bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
91   bool SimplifyIndirectBr(IndirectBrInst *IBI);
92   bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
93   bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
94 
95 public:
SimplifyCFGOpt(const TargetData * td)96   explicit SimplifyCFGOpt(const TargetData *td) : TD(td) {}
97   bool run(BasicBlock *BB);
98 };
99 }
100 
101 /// SafeToMergeTerminators - Return true if it is safe to merge these two
102 /// terminator instructions together.
103 ///
SafeToMergeTerminators(TerminatorInst * SI1,TerminatorInst * SI2)104 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
105   if (SI1 == SI2) return false;  // Can't merge with self!
106 
107   // It is not safe to merge these two switch instructions if they have a common
108   // successor, and if that successor has a PHI node, and if *that* PHI node has
109   // conflicting incoming values from the two switch blocks.
110   BasicBlock *SI1BB = SI1->getParent();
111   BasicBlock *SI2BB = SI2->getParent();
112   SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
113 
114   for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
115     if (SI1Succs.count(*I))
116       for (BasicBlock::iterator BBI = (*I)->begin();
117            isa<PHINode>(BBI); ++BBI) {
118         PHINode *PN = cast<PHINode>(BBI);
119         if (PN->getIncomingValueForBlock(SI1BB) !=
120             PN->getIncomingValueForBlock(SI2BB))
121           return false;
122       }
123 
124   return true;
125 }
126 
127 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
128 /// to merge these two terminator instructions together, where SI1 is an
129 /// unconditional branch. PhiNodes will store all PHI nodes in common
130 /// successors.
131 ///
isProfitableToFoldUnconditional(BranchInst * SI1,BranchInst * SI2,Instruction * Cond,SmallVectorImpl<PHINode * > & PhiNodes)132 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
133                                           BranchInst *SI2,
134                                           Instruction *Cond,
135                                           SmallVectorImpl<PHINode*> &PhiNodes) {
136   if (SI1 == SI2) return false;  // Can't merge with self!
137   assert(SI1->isUnconditional() && SI2->isConditional());
138 
139   // We fold the unconditional branch if we can easily update all PHI nodes in
140   // common successors:
141   // 1> We have a constant incoming value for the conditional branch;
142   // 2> We have "Cond" as the incoming value for the unconditional branch;
143   // 3> SI2->getCondition() and Cond have same operands.
144   CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
145   if (!Ci2) return false;
146   if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
147         Cond->getOperand(1) == Ci2->getOperand(1)) &&
148       !(Cond->getOperand(0) == Ci2->getOperand(1) &&
149         Cond->getOperand(1) == Ci2->getOperand(0)))
150     return false;
151 
152   BasicBlock *SI1BB = SI1->getParent();
153   BasicBlock *SI2BB = SI2->getParent();
154   SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
155   for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
156     if (SI1Succs.count(*I))
157       for (BasicBlock::iterator BBI = (*I)->begin();
158            isa<PHINode>(BBI); ++BBI) {
159         PHINode *PN = cast<PHINode>(BBI);
160         if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
161             !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
162           return false;
163         PhiNodes.push_back(PN);
164       }
165   return true;
166 }
167 
168 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
169 /// now be entries in it from the 'NewPred' block.  The values that will be
170 /// flowing into the PHI nodes will be the same as those coming in from
171 /// ExistPred, an existing predecessor of Succ.
AddPredecessorToBlock(BasicBlock * Succ,BasicBlock * NewPred,BasicBlock * ExistPred)172 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
173                                   BasicBlock *ExistPred) {
174   if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
175 
176   PHINode *PN;
177   for (BasicBlock::iterator I = Succ->begin();
178        (PN = dyn_cast<PHINode>(I)); ++I)
179     PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
180 }
181 
182 
183 /// GetIfCondition - Given a basic block (BB) with two predecessors (and at
184 /// least one PHI node in it), check to see if the merge at this block is due
185 /// to an "if condition".  If so, return the boolean condition that determines
186 /// which entry into BB will be taken.  Also, return by references the block
187 /// that will be entered from if the condition is true, and the block that will
188 /// be entered if the condition is false.
189 ///
190 /// This does no checking to see if the true/false blocks have large or unsavory
191 /// instructions in them.
GetIfCondition(BasicBlock * BB,BasicBlock * & IfTrue,BasicBlock * & IfFalse)192 static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue,
193                              BasicBlock *&IfFalse) {
194   PHINode *SomePHI = cast<PHINode>(BB->begin());
195   assert(SomePHI->getNumIncomingValues() == 2 &&
196          "Function can only handle blocks with 2 predecessors!");
197   BasicBlock *Pred1 = SomePHI->getIncomingBlock(0);
198   BasicBlock *Pred2 = SomePHI->getIncomingBlock(1);
199 
200   // We can only handle branches.  Other control flow will be lowered to
201   // branches if possible anyway.
202   BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator());
203   BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator());
204   if (Pred1Br == 0 || Pred2Br == 0)
205     return 0;
206 
207   // Eliminate code duplication by ensuring that Pred1Br is conditional if
208   // either are.
209   if (Pred2Br->isConditional()) {
210     // If both branches are conditional, we don't have an "if statement".  In
211     // reality, we could transform this case, but since the condition will be
212     // required anyway, we stand no chance of eliminating it, so the xform is
213     // probably not profitable.
214     if (Pred1Br->isConditional())
215       return 0;
216 
217     std::swap(Pred1, Pred2);
218     std::swap(Pred1Br, Pred2Br);
219   }
220 
221   if (Pred1Br->isConditional()) {
222     // The only thing we have to watch out for here is to make sure that Pred2
223     // doesn't have incoming edges from other blocks.  If it does, the condition
224     // doesn't dominate BB.
225     if (Pred2->getSinglePredecessor() == 0)
226       return 0;
227 
228     // If we found a conditional branch predecessor, make sure that it branches
229     // to BB and Pred2Br.  If it doesn't, this isn't an "if statement".
230     if (Pred1Br->getSuccessor(0) == BB &&
231         Pred1Br->getSuccessor(1) == Pred2) {
232       IfTrue = Pred1;
233       IfFalse = Pred2;
234     } else if (Pred1Br->getSuccessor(0) == Pred2 &&
235                Pred1Br->getSuccessor(1) == BB) {
236       IfTrue = Pred2;
237       IfFalse = Pred1;
238     } else {
239       // We know that one arm of the conditional goes to BB, so the other must
240       // go somewhere unrelated, and this must not be an "if statement".
241       return 0;
242     }
243 
244     return Pred1Br->getCondition();
245   }
246 
247   // Ok, if we got here, both predecessors end with an unconditional branch to
248   // BB.  Don't panic!  If both blocks only have a single (identical)
249   // predecessor, and THAT is a conditional branch, then we're all ok!
250   BasicBlock *CommonPred = Pred1->getSinglePredecessor();
251   if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor())
252     return 0;
253 
254   // Otherwise, if this is a conditional branch, then we can use it!
255   BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator());
256   if (BI == 0) return 0;
257 
258   assert(BI->isConditional() && "Two successors but not conditional?");
259   if (BI->getSuccessor(0) == Pred1) {
260     IfTrue = Pred1;
261     IfFalse = Pred2;
262   } else {
263     IfTrue = Pred2;
264     IfFalse = Pred1;
265   }
266   return BI->getCondition();
267 }
268 
269 /// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the
270 /// given instruction, which is assumed to be safe to speculate. 1 means
271 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
ComputeSpeculationCost(const User * I)272 static unsigned ComputeSpeculationCost(const User *I) {
273   assert(isSafeToSpeculativelyExecute(I) &&
274          "Instruction is not safe to speculatively execute!");
275   switch (Operator::getOpcode(I)) {
276   default:
277     // In doubt, be conservative.
278     return UINT_MAX;
279   case Instruction::GetElementPtr:
280     // GEPs are cheap if all indices are constant.
281     if (!cast<GEPOperator>(I)->hasAllConstantIndices())
282       return UINT_MAX;
283     return 1;
284   case Instruction::Load:
285   case Instruction::Add:
286   case Instruction::Sub:
287   case Instruction::And:
288   case Instruction::Or:
289   case Instruction::Xor:
290   case Instruction::Shl:
291   case Instruction::LShr:
292   case Instruction::AShr:
293   case Instruction::ICmp:
294   case Instruction::Trunc:
295   case Instruction::ZExt:
296   case Instruction::SExt:
297     return 1; // These are all cheap.
298 
299   case Instruction::Call:
300   case Instruction::Select:
301     return 2;
302   }
303 }
304 
305 /// DominatesMergePoint - If we have a merge point of an "if condition" as
306 /// accepted above, return true if the specified value dominates the block.  We
307 /// don't handle the true generality of domination here, just a special case
308 /// which works well enough for us.
309 ///
310 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
311 /// see if V (which must be an instruction) and its recursive operands
312 /// that do not dominate BB have a combined cost lower than CostRemaining and
313 /// are non-trapping.  If both are true, the instruction is inserted into the
314 /// set and true is returned.
315 ///
316 /// The cost for most non-trapping instructions is defined as 1 except for
317 /// Select whose cost is 2.
318 ///
319 /// After this function returns, CostRemaining is decreased by the cost of
320 /// V plus its non-dominating operands.  If that cost is greater than
321 /// CostRemaining, false is returned and CostRemaining is undefined.
DominatesMergePoint(Value * V,BasicBlock * BB,SmallPtrSet<Instruction *,4> * AggressiveInsts,unsigned & CostRemaining)322 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
323                                 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
324                                 unsigned &CostRemaining) {
325   Instruction *I = dyn_cast<Instruction>(V);
326   if (!I) {
327     // Non-instructions all dominate instructions, but not all constantexprs
328     // can be executed unconditionally.
329     if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
330       if (C->canTrap())
331         return false;
332     return true;
333   }
334   BasicBlock *PBB = I->getParent();
335 
336   // We don't want to allow weird loops that might have the "if condition" in
337   // the bottom of this block.
338   if (PBB == BB) return false;
339 
340   // If this instruction is defined in a block that contains an unconditional
341   // branch to BB, then it must be in the 'conditional' part of the "if
342   // statement".  If not, it definitely dominates the region.
343   BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
344   if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB)
345     return true;
346 
347   // If we aren't allowing aggressive promotion anymore, then don't consider
348   // instructions in the 'if region'.
349   if (AggressiveInsts == 0) return false;
350 
351   // If we have seen this instruction before, don't count it again.
352   if (AggressiveInsts->count(I)) return true;
353 
354   // Okay, it looks like the instruction IS in the "condition".  Check to
355   // see if it's a cheap instruction to unconditionally compute, and if it
356   // only uses stuff defined outside of the condition.  If so, hoist it out.
357   if (!isSafeToSpeculativelyExecute(I))
358     return false;
359 
360   unsigned Cost = ComputeSpeculationCost(I);
361 
362   if (Cost > CostRemaining)
363     return false;
364 
365   CostRemaining -= Cost;
366 
367   // Okay, we can only really hoist these out if their operands do
368   // not take us over the cost threshold.
369   for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
370     if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining))
371       return false;
372   // Okay, it's safe to do this!  Remember this instruction.
373   AggressiveInsts->insert(I);
374   return true;
375 }
376 
377 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
378 /// and PointerNullValue. Return NULL if value is not a constant int.
GetConstantInt(Value * V,const TargetData * TD)379 static ConstantInt *GetConstantInt(Value *V, const TargetData *TD) {
380   // Normal constant int.
381   ConstantInt *CI = dyn_cast<ConstantInt>(V);
382   if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy())
383     return CI;
384 
385   // This is some kind of pointer constant. Turn it into a pointer-sized
386   // ConstantInt if possible.
387   IntegerType *PtrTy = TD->getIntPtrType(V->getContext());
388 
389   // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
390   if (isa<ConstantPointerNull>(V))
391     return ConstantInt::get(PtrTy, 0);
392 
393   // IntToPtr const int.
394   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
395     if (CE->getOpcode() == Instruction::IntToPtr)
396       if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
397         // The constant is very likely to have the right type already.
398         if (CI->getType() == PtrTy)
399           return CI;
400         else
401           return cast<ConstantInt>
402             (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
403       }
404   return 0;
405 }
406 
407 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
408 /// collection of icmp eq/ne instructions that compare a value against a
409 /// constant, return the value being compared, and stick the constant into the
410 /// Values vector.
411 static Value *
GatherConstantCompares(Value * V,std::vector<ConstantInt * > & Vals,Value * & Extra,const TargetData * TD,bool isEQ,unsigned & UsedICmps)412 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
413                        const TargetData *TD, bool isEQ, unsigned &UsedICmps) {
414   Instruction *I = dyn_cast<Instruction>(V);
415   if (I == 0) return 0;
416 
417   // If this is an icmp against a constant, handle this as one of the cases.
418   if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
419     if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) {
420       if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
421         UsedICmps++;
422         Vals.push_back(C);
423         return I->getOperand(0);
424       }
425 
426       // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
427       // the set.
428       ConstantRange Span =
429         ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
430 
431       // If this is an and/!= check then we want to optimize "x ugt 2" into
432       // x != 0 && x != 1.
433       if (!isEQ)
434         Span = Span.inverse();
435 
436       // If there are a ton of values, we don't want to make a ginormous switch.
437       if (Span.getSetSize().ugt(8) || Span.isEmptySet())
438         return 0;
439 
440       for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
441         Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
442       UsedICmps++;
443       return I->getOperand(0);
444     }
445     return 0;
446   }
447 
448   // Otherwise, we can only handle an | or &, depending on isEQ.
449   if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
450     return 0;
451 
452   unsigned NumValsBeforeLHS = Vals.size();
453   unsigned UsedICmpsBeforeLHS = UsedICmps;
454   if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD,
455                                           isEQ, UsedICmps)) {
456     unsigned NumVals = Vals.size();
457     unsigned UsedICmpsBeforeRHS = UsedICmps;
458     if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
459                                             isEQ, UsedICmps)) {
460       if (LHS == RHS)
461         return LHS;
462       Vals.resize(NumVals);
463       UsedICmps = UsedICmpsBeforeRHS;
464     }
465 
466     // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
467     // set it and return success.
468     if (Extra == 0 || Extra == I->getOperand(1)) {
469       Extra = I->getOperand(1);
470       return LHS;
471     }
472 
473     Vals.resize(NumValsBeforeLHS);
474     UsedICmps = UsedICmpsBeforeLHS;
475     return 0;
476   }
477 
478   // If the LHS can't be folded in, but Extra is available and RHS can, try to
479   // use LHS as Extra.
480   if (Extra == 0 || Extra == I->getOperand(0)) {
481     Value *OldExtra = Extra;
482     Extra = I->getOperand(0);
483     if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
484                                             isEQ, UsedICmps))
485       return RHS;
486     assert(Vals.size() == NumValsBeforeLHS);
487     Extra = OldExtra;
488   }
489 
490   return 0;
491 }
492 
EraseTerminatorInstAndDCECond(TerminatorInst * TI)493 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
494   Instruction *Cond = 0;
495   if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
496     Cond = dyn_cast<Instruction>(SI->getCondition());
497   } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
498     if (BI->isConditional())
499       Cond = dyn_cast<Instruction>(BI->getCondition());
500   } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
501     Cond = dyn_cast<Instruction>(IBI->getAddress());
502   }
503 
504   TI->eraseFromParent();
505   if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
506 }
507 
508 /// isValueEqualityComparison - Return true if the specified terminator checks
509 /// to see if a value is equal to constant integer value.
isValueEqualityComparison(TerminatorInst * TI)510 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
511   Value *CV = 0;
512   if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
513     // Do not permit merging of large switch instructions into their
514     // predecessors unless there is only one predecessor.
515     if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
516                                              pred_end(SI->getParent())) <= 128)
517       CV = SI->getCondition();
518   } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
519     if (BI->isConditional() && BI->getCondition()->hasOneUse())
520       if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
521         if ((ICI->getPredicate() == ICmpInst::ICMP_EQ ||
522              ICI->getPredicate() == ICmpInst::ICMP_NE) &&
523             GetConstantInt(ICI->getOperand(1), TD))
524           CV = ICI->getOperand(0);
525 
526   // Unwrap any lossless ptrtoint cast.
527   if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext()))
528     if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV))
529       CV = PTII->getOperand(0);
530   return CV;
531 }
532 
533 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
534 /// decode all of the 'cases' that it represents and return the 'default' block.
535 BasicBlock *SimplifyCFGOpt::
GetValueEqualityComparisonCases(TerminatorInst * TI,std::vector<ValueEqualityComparisonCase> & Cases)536 GetValueEqualityComparisonCases(TerminatorInst *TI,
537                                 std::vector<ValueEqualityComparisonCase>
538                                                                        &Cases) {
539   if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
540     Cases.reserve(SI->getNumCases());
541     for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
542       Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
543                                                   i.getCaseSuccessor()));
544     return SI->getDefaultDest();
545   }
546 
547   BranchInst *BI = cast<BranchInst>(TI);
548   ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
549   BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
550   Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
551                                                              TD),
552                                               Succ));
553   return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
554 }
555 
556 
557 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
558 /// in the list that match the specified block.
EliminateBlockCases(BasicBlock * BB,std::vector<ValueEqualityComparisonCase> & Cases)559 static void EliminateBlockCases(BasicBlock *BB,
560                               std::vector<ValueEqualityComparisonCase> &Cases) {
561   for (unsigned i = 0, e = Cases.size(); i != e; ++i)
562     if (Cases[i].Dest == BB) {
563       Cases.erase(Cases.begin()+i);
564       --i; --e;
565     }
566 }
567 
568 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
569 /// well.
570 static bool
ValuesOverlap(std::vector<ValueEqualityComparisonCase> & C1,std::vector<ValueEqualityComparisonCase> & C2)571 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
572               std::vector<ValueEqualityComparisonCase > &C2) {
573   std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
574 
575   // Make V1 be smaller than V2.
576   if (V1->size() > V2->size())
577     std::swap(V1, V2);
578 
579   if (V1->size() == 0) return false;
580   if (V1->size() == 1) {
581     // Just scan V2.
582     ConstantInt *TheVal = (*V1)[0].Value;
583     for (unsigned i = 0, e = V2->size(); i != e; ++i)
584       if (TheVal == (*V2)[i].Value)
585         return true;
586   }
587 
588   // Otherwise, just sort both lists and compare element by element.
589   array_pod_sort(V1->begin(), V1->end());
590   array_pod_sort(V2->begin(), V2->end());
591   unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
592   while (i1 != e1 && i2 != e2) {
593     if ((*V1)[i1].Value == (*V2)[i2].Value)
594       return true;
595     if ((*V1)[i1].Value < (*V2)[i2].Value)
596       ++i1;
597     else
598       ++i2;
599   }
600   return false;
601 }
602 
603 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
604 /// terminator instruction and its block is known to only have a single
605 /// predecessor block, check to see if that predecessor is also a value
606 /// comparison with the same value, and if that comparison determines the
607 /// outcome of this comparison.  If so, simplify TI.  This does a very limited
608 /// form of jump threading.
609 bool SimplifyCFGOpt::
SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst * TI,BasicBlock * Pred,IRBuilder<> & Builder)610 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
611                                               BasicBlock *Pred,
612                                               IRBuilder<> &Builder) {
613   Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
614   if (!PredVal) return false;  // Not a value comparison in predecessor.
615 
616   Value *ThisVal = isValueEqualityComparison(TI);
617   assert(ThisVal && "This isn't a value comparison!!");
618   if (ThisVal != PredVal) return false;  // Different predicates.
619 
620   // TODO: Preserve branch weight metadata, similarly to how
621   // FoldValueComparisonIntoPredecessors preserves it.
622 
623   // Find out information about when control will move from Pred to TI's block.
624   std::vector<ValueEqualityComparisonCase> PredCases;
625   BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
626                                                         PredCases);
627   EliminateBlockCases(PredDef, PredCases);  // Remove default from cases.
628 
629   // Find information about how control leaves this block.
630   std::vector<ValueEqualityComparisonCase> ThisCases;
631   BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
632   EliminateBlockCases(ThisDef, ThisCases);  // Remove default from cases.
633 
634   // If TI's block is the default block from Pred's comparison, potentially
635   // simplify TI based on this knowledge.
636   if (PredDef == TI->getParent()) {
637     // If we are here, we know that the value is none of those cases listed in
638     // PredCases.  If there are any cases in ThisCases that are in PredCases, we
639     // can simplify TI.
640     if (!ValuesOverlap(PredCases, ThisCases))
641       return false;
642 
643     if (isa<BranchInst>(TI)) {
644       // Okay, one of the successors of this condbr is dead.  Convert it to a
645       // uncond br.
646       assert(ThisCases.size() == 1 && "Branch can only have one case!");
647       // Insert the new branch.
648       Instruction *NI = Builder.CreateBr(ThisDef);
649       (void) NI;
650 
651       // Remove PHI node entries for the dead edge.
652       ThisCases[0].Dest->removePredecessor(TI->getParent());
653 
654       DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
655            << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
656 
657       EraseTerminatorInstAndDCECond(TI);
658       return true;
659     }
660 
661     SwitchInst *SI = cast<SwitchInst>(TI);
662     // Okay, TI has cases that are statically dead, prune them away.
663     SmallPtrSet<Constant*, 16> DeadCases;
664     for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
665       DeadCases.insert(PredCases[i].Value);
666 
667     DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
668                  << "Through successor TI: " << *TI);
669 
670     for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
671       --i;
672       if (DeadCases.count(i.getCaseValue())) {
673         i.getCaseSuccessor()->removePredecessor(TI->getParent());
674         SI->removeCase(i);
675       }
676     }
677 
678     DEBUG(dbgs() << "Leaving: " << *TI << "\n");
679     return true;
680   }
681 
682   // Otherwise, TI's block must correspond to some matched value.  Find out
683   // which value (or set of values) this is.
684   ConstantInt *TIV = 0;
685   BasicBlock *TIBB = TI->getParent();
686   for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
687     if (PredCases[i].Dest == TIBB) {
688       if (TIV != 0)
689         return false;  // Cannot handle multiple values coming to this block.
690       TIV = PredCases[i].Value;
691     }
692   assert(TIV && "No edge from pred to succ?");
693 
694   // Okay, we found the one constant that our value can be if we get into TI's
695   // BB.  Find out which successor will unconditionally be branched to.
696   BasicBlock *TheRealDest = 0;
697   for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
698     if (ThisCases[i].Value == TIV) {
699       TheRealDest = ThisCases[i].Dest;
700       break;
701     }
702 
703   // If not handled by any explicit cases, it is handled by the default case.
704   if (TheRealDest == 0) TheRealDest = ThisDef;
705 
706   // Remove PHI node entries for dead edges.
707   BasicBlock *CheckEdge = TheRealDest;
708   for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
709     if (*SI != CheckEdge)
710       (*SI)->removePredecessor(TIBB);
711     else
712       CheckEdge = 0;
713 
714   // Insert the new branch.
715   Instruction *NI = Builder.CreateBr(TheRealDest);
716   (void) NI;
717 
718   DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
719             << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
720 
721   EraseTerminatorInstAndDCECond(TI);
722   return true;
723 }
724 
725 namespace {
726   /// ConstantIntOrdering - This class implements a stable ordering of constant
727   /// integers that does not depend on their address.  This is important for
728   /// applications that sort ConstantInt's to ensure uniqueness.
729   struct ConstantIntOrdering {
operator ()__anon08d918330211::ConstantIntOrdering730     bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
731       return LHS->getValue().ult(RHS->getValue());
732     }
733   };
734 }
735 
ConstantIntSortPredicate(const void * P1,const void * P2)736 static int ConstantIntSortPredicate(const void *P1, const void *P2) {
737   const ConstantInt *LHS = *(const ConstantInt*const*)P1;
738   const ConstantInt *RHS = *(const ConstantInt*const*)P2;
739   if (LHS->getValue().ult(RHS->getValue()))
740     return 1;
741   if (LHS->getValue() == RHS->getValue())
742     return 0;
743   return -1;
744 }
745 
HasBranchWeights(const Instruction * I)746 static inline bool HasBranchWeights(const Instruction* I) {
747   MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
748   if (ProfMD && ProfMD->getOperand(0))
749     if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
750       return MDS->getString().equals("branch_weights");
751 
752   return false;
753 }
754 
755 /// Tries to get a branch weight for the given instruction, returns NULL if it
756 /// can't. Pos starts at 0.
GetWeight(Instruction * I,int Pos)757 static ConstantInt* GetWeight(Instruction* I, int Pos) {
758   MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
759   if (ProfMD && ProfMD->getOperand(0)) {
760     if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) {
761       if (MDS->getString().equals("branch_weights")) {
762         assert(ProfMD->getNumOperands() >= 3);
763         return dyn_cast<ConstantInt>(ProfMD->getOperand(1 + Pos));
764       }
765     }
766   }
767 
768   return 0;
769 }
770 
771 /// Scale the given weights based on the successor TI's metadata. Scaling is
772 /// done by multiplying every weight by the sum of the successor's weights.
ScaleWeights(Instruction * STI,MutableArrayRef<uint64_t> Weights)773 static void ScaleWeights(Instruction* STI, MutableArrayRef<uint64_t> Weights) {
774   // Sum the successor's weights
775   assert(HasBranchWeights(STI));
776   unsigned Scale = 0;
777   MDNode* ProfMD = STI->getMetadata(LLVMContext::MD_prof);
778   for (unsigned i = 1; i < ProfMD->getNumOperands(); ++i) {
779     ConstantInt* CI = dyn_cast<ConstantInt>(ProfMD->getOperand(i));
780     assert(CI);
781     Scale += CI->getValue().getZExtValue();
782   }
783 
784   // Skip default, as it's replaced during the folding
785   for (unsigned i = 1; i < Weights.size(); ++i) {
786     Weights[i] *= Scale;
787   }
788 }
789 
790 /// Sees if any of the weights are too big for a uint32_t, and halves all the
791 /// weights if any are.
FitWeights(MutableArrayRef<uint64_t> Weights)792 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
793   bool Halve = false;
794   for (unsigned i = 0; i < Weights.size(); ++i)
795     if (Weights[i] > UINT_MAX) {
796       Halve = true;
797       break;
798     }
799 
800   if (! Halve)
801     return;
802 
803   for (unsigned i = 0; i < Weights.size(); ++i)
804     Weights[i] /= 2;
805 }
806 
807 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
808 /// equality comparison instruction (either a switch or a branch on "X == c").
809 /// See if any of the predecessors of the terminator block are value comparisons
810 /// on the same value.  If so, and if safe to do so, fold them together.
FoldValueComparisonIntoPredecessors(TerminatorInst * TI,IRBuilder<> & Builder)811 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
812                                                          IRBuilder<> &Builder) {
813   BasicBlock *BB = TI->getParent();
814   Value *CV = isValueEqualityComparison(TI);  // CondVal
815   assert(CV && "Not a comparison?");
816   bool Changed = false;
817 
818   SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
819   while (!Preds.empty()) {
820     BasicBlock *Pred = Preds.pop_back_val();
821 
822     // See if the predecessor is a comparison with the same value.
823     TerminatorInst *PTI = Pred->getTerminator();
824     Value *PCV = isValueEqualityComparison(PTI);  // PredCondVal
825 
826     if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
827       // Figure out which 'cases' to copy from SI to PSI.
828       std::vector<ValueEqualityComparisonCase> BBCases;
829       BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
830 
831       std::vector<ValueEqualityComparisonCase> PredCases;
832       BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
833 
834       // Based on whether the default edge from PTI goes to BB or not, fill in
835       // PredCases and PredDefault with the new switch cases we would like to
836       // build.
837       SmallVector<BasicBlock*, 8> NewSuccessors;
838 
839       // Update the branch weight metadata along the way
840       SmallVector<uint64_t, 8> Weights;
841       uint64_t PredDefaultWeight = 0;
842       bool PredHasWeights = HasBranchWeights(PTI);
843       bool SuccHasWeights = HasBranchWeights(TI);
844 
845       if (PredHasWeights) {
846         MDNode* MD = PTI->getMetadata(LLVMContext::MD_prof);
847         assert(MD);
848         for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
849           ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i));
850           assert(CI);
851           Weights.push_back(CI->getValue().getZExtValue());
852         }
853 
854         // If the predecessor is a conditional eq, then swap the default weight
855         // to be the first entry.
856         if (BranchInst* BI = dyn_cast<BranchInst>(PTI)) {
857           assert(Weights.size() == 2);
858           ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
859 
860           if (ICI->getPredicate() == ICmpInst::ICMP_EQ) {
861             std::swap(Weights.front(), Weights.back());
862           }
863         }
864 
865         PredDefaultWeight = Weights.front();
866       } else if (SuccHasWeights) {
867         // If there are no predecessor weights but there are successor weights,
868         // populate Weights with 1, which will later be scaled to the sum of
869         // successor's weights
870         Weights.assign(1 + PredCases.size(), 1);
871         PredDefaultWeight = 1;
872       }
873 
874       uint64_t SuccDefaultWeight = 0;
875       if (SuccHasWeights) {
876         int Index = 0;
877         if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
878           ICmpInst* ICI = dyn_cast<ICmpInst>(BI->getCondition());
879           assert(ICI);
880 
881           if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
882             Index = 1;
883         }
884 
885         SuccDefaultWeight = GetWeight(TI, Index)->getValue().getZExtValue();
886       }
887 
888       if (PredDefault == BB) {
889         // If this is the default destination from PTI, only the edges in TI
890         // that don't occur in PTI, or that branch to BB will be activated.
891         std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
892         for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
893           if (PredCases[i].Dest != BB)
894             PTIHandled.insert(PredCases[i].Value);
895           else {
896             // The default destination is BB, we don't need explicit targets.
897             std::swap(PredCases[i], PredCases.back());
898 
899             if (PredHasWeights) {
900               std::swap(Weights[i+1], Weights.back());
901               Weights.pop_back();
902             }
903 
904             PredCases.pop_back();
905             --i; --e;
906           }
907 
908         // Reconstruct the new switch statement we will be building.
909         if (PredDefault != BBDefault) {
910           PredDefault->removePredecessor(Pred);
911           PredDefault = BBDefault;
912           NewSuccessors.push_back(BBDefault);
913         }
914 
915         if (SuccHasWeights) {
916           ScaleWeights(TI, Weights);
917           Weights.front() *= SuccDefaultWeight;
918         } else if (PredHasWeights) {
919           Weights.front() /= (1 + BBCases.size());
920         }
921 
922         for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
923           if (!PTIHandled.count(BBCases[i].Value) &&
924               BBCases[i].Dest != BBDefault) {
925             PredCases.push_back(BBCases[i]);
926             NewSuccessors.push_back(BBCases[i].Dest);
927             if (SuccHasWeights) {
928               Weights.push_back(PredDefaultWeight *
929                                 GetWeight(TI, i)->getValue().getZExtValue());
930             } else if (PredHasWeights) {
931               // Split the old default's weight amongst the children
932               Weights.push_back(PredDefaultWeight / (1 + BBCases.size()));
933             }
934           }
935 
936       } else {
937         // FIXME: preserve branch weight metadata, similarly to the 'then'
938         // above. For now, drop it.
939         PredHasWeights = false;
940         SuccHasWeights = false;
941 
942         // If this is not the default destination from PSI, only the edges
943         // in SI that occur in PSI with a destination of BB will be
944         // activated.
945         std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
946         for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
947           if (PredCases[i].Dest == BB) {
948             PTIHandled.insert(PredCases[i].Value);
949             std::swap(PredCases[i], PredCases.back());
950             PredCases.pop_back();
951             --i; --e;
952           }
953 
954         // Okay, now we know which constants were sent to BB from the
955         // predecessor.  Figure out where they will all go now.
956         for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
957           if (PTIHandled.count(BBCases[i].Value)) {
958             // If this is one we are capable of getting...
959             PredCases.push_back(BBCases[i]);
960             NewSuccessors.push_back(BBCases[i].Dest);
961             PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
962           }
963 
964         // If there are any constants vectored to BB that TI doesn't handle,
965         // they must go to the default destination of TI.
966         for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
967                                     PTIHandled.begin(),
968                E = PTIHandled.end(); I != E; ++I) {
969           PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
970           NewSuccessors.push_back(BBDefault);
971         }
972       }
973 
974       // Okay, at this point, we know which new successor Pred will get.  Make
975       // sure we update the number of entries in the PHI nodes for these
976       // successors.
977       for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
978         AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
979 
980       Builder.SetInsertPoint(PTI);
981       // Convert pointer to int before we switch.
982       if (CV->getType()->isPointerTy()) {
983         assert(TD && "Cannot switch on pointer without TargetData");
984         CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()),
985                                     "magicptr");
986       }
987 
988       // Now that the successors are updated, create the new Switch instruction.
989       SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
990                                                PredCases.size());
991       NewSI->setDebugLoc(PTI->getDebugLoc());
992       for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
993         NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
994 
995       if (PredHasWeights || SuccHasWeights) {
996         // Halve the weights if any of them cannot fit in an uint32_t
997         FitWeights(Weights);
998 
999         SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1000 
1001         NewSI->setMetadata(LLVMContext::MD_prof,
1002                            MDBuilder(BB->getContext()).
1003                            createBranchWeights(MDWeights));
1004       }
1005 
1006       EraseTerminatorInstAndDCECond(PTI);
1007 
1008       // Okay, last check.  If BB is still a successor of PSI, then we must
1009       // have an infinite loop case.  If so, add an infinitely looping block
1010       // to handle the case to preserve the behavior of the code.
1011       BasicBlock *InfLoopBlock = 0;
1012       for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1013         if (NewSI->getSuccessor(i) == BB) {
1014           if (InfLoopBlock == 0) {
1015             // Insert it at the end of the function, because it's either code,
1016             // or it won't matter if it's hot. :)
1017             InfLoopBlock = BasicBlock::Create(BB->getContext(),
1018                                               "infloop", BB->getParent());
1019             BranchInst::Create(InfLoopBlock, InfLoopBlock);
1020           }
1021           NewSI->setSuccessor(i, InfLoopBlock);
1022         }
1023 
1024       Changed = true;
1025     }
1026   }
1027   return Changed;
1028 }
1029 
1030 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1031 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1032 // would need to do this), we can't hoist the invoke, as there is nowhere
1033 // to put the select in this case.
isSafeToHoistInvoke(BasicBlock * BB1,BasicBlock * BB2,Instruction * I1,Instruction * I2)1034 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1035                                 Instruction *I1, Instruction *I2) {
1036   for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1037     PHINode *PN;
1038     for (BasicBlock::iterator BBI = SI->begin();
1039          (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1040       Value *BB1V = PN->getIncomingValueForBlock(BB1);
1041       Value *BB2V = PN->getIncomingValueForBlock(BB2);
1042       if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1043         return false;
1044       }
1045     }
1046   }
1047   return true;
1048 }
1049 
1050 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1051 /// BB2, hoist any common code in the two blocks up into the branch block.  The
1052 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
HoistThenElseCodeToIf(BranchInst * BI)1053 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1054   // This does very trivial matching, with limited scanning, to find identical
1055   // instructions in the two blocks.  In particular, we don't want to get into
1056   // O(M*N) situations here where M and N are the sizes of BB1 and BB2.  As
1057   // such, we currently just scan for obviously identical instructions in an
1058   // identical order.
1059   BasicBlock *BB1 = BI->getSuccessor(0);  // The true destination.
1060   BasicBlock *BB2 = BI->getSuccessor(1);  // The false destination
1061 
1062   BasicBlock::iterator BB1_Itr = BB1->begin();
1063   BasicBlock::iterator BB2_Itr = BB2->begin();
1064 
1065   Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1066   // Skip debug info if it is not identical.
1067   DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1068   DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1069   if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1070     while (isa<DbgInfoIntrinsic>(I1))
1071       I1 = BB1_Itr++;
1072     while (isa<DbgInfoIntrinsic>(I2))
1073       I2 = BB2_Itr++;
1074   }
1075   if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1076       (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1077     return false;
1078 
1079   // If we get here, we can hoist at least one instruction.
1080   BasicBlock *BIParent = BI->getParent();
1081 
1082   do {
1083     // If we are hoisting the terminator instruction, don't move one (making a
1084     // broken BB), instead clone it, and remove BI.
1085     if (isa<TerminatorInst>(I1))
1086       goto HoistTerminator;
1087 
1088     // For a normal instruction, we just move one to right before the branch,
1089     // then replace all uses of the other with the first.  Finally, we remove
1090     // the now redundant second instruction.
1091     BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1092     if (!I2->use_empty())
1093       I2->replaceAllUsesWith(I1);
1094     I1->intersectOptionalDataWith(I2);
1095     I2->eraseFromParent();
1096 
1097     I1 = BB1_Itr++;
1098     I2 = BB2_Itr++;
1099     // Skip debug info if it is not identical.
1100     DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1101     DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1102     if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1103       while (isa<DbgInfoIntrinsic>(I1))
1104         I1 = BB1_Itr++;
1105       while (isa<DbgInfoIntrinsic>(I2))
1106         I2 = BB2_Itr++;
1107     }
1108   } while (I1->isIdenticalToWhenDefined(I2));
1109 
1110   return true;
1111 
1112 HoistTerminator:
1113   // It may not be possible to hoist an invoke.
1114   if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1115     return true;
1116 
1117   // Okay, it is safe to hoist the terminator.
1118   Instruction *NT = I1->clone();
1119   BIParent->getInstList().insert(BI, NT);
1120   if (!NT->getType()->isVoidTy()) {
1121     I1->replaceAllUsesWith(NT);
1122     I2->replaceAllUsesWith(NT);
1123     NT->takeName(I1);
1124   }
1125 
1126   IRBuilder<true, NoFolder> Builder(NT);
1127   // Hoisting one of the terminators from our successor is a great thing.
1128   // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1129   // them.  If they do, all PHI entries for BB1/BB2 must agree for all PHI
1130   // nodes, so we insert select instruction to compute the final result.
1131   std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1132   for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1133     PHINode *PN;
1134     for (BasicBlock::iterator BBI = SI->begin();
1135          (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1136       Value *BB1V = PN->getIncomingValueForBlock(BB1);
1137       Value *BB2V = PN->getIncomingValueForBlock(BB2);
1138       if (BB1V == BB2V) continue;
1139 
1140       // These values do not agree.  Insert a select instruction before NT
1141       // that determines the right value.
1142       SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1143       if (SI == 0)
1144         SI = cast<SelectInst>
1145           (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1146                                 BB1V->getName()+"."+BB2V->getName()));
1147 
1148       // Make the PHI node use the select for all incoming values for BB1/BB2
1149       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1150         if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1151           PN->setIncomingValue(i, SI);
1152     }
1153   }
1154 
1155   // Update any PHI nodes in our new successors.
1156   for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1157     AddPredecessorToBlock(*SI, BIParent, BB1);
1158 
1159   EraseTerminatorInstAndDCECond(BI);
1160   return true;
1161 }
1162 
1163 /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1
1164 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code
1165 /// (for now, restricted to a single instruction that's side effect free) from
1166 /// the BB1 into the branch block to speculatively execute it.
1167 ///
1168 /// Turn
1169 /// BB:
1170 ///     %t1 = icmp
1171 ///     br i1 %t1, label %BB1, label %BB2
1172 /// BB1:
1173 ///     %t3 = add %t2, c
1174 ///     br label BB2
1175 /// BB2:
1176 /// =>
1177 /// BB:
1178 ///     %t1 = icmp
1179 ///     %t4 = add %t2, c
1180 ///     %t3 = select i1 %t1, %t2, %t3
SpeculativelyExecuteBB(BranchInst * BI,BasicBlock * BB1)1181 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) {
1182   // Only speculatively execution a single instruction (not counting the
1183   // terminator) for now.
1184   Instruction *HInst = NULL;
1185   Instruction *Term = BB1->getTerminator();
1186   for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end();
1187        BBI != BBE; ++BBI) {
1188     Instruction *I = BBI;
1189     // Skip debug info.
1190     if (isa<DbgInfoIntrinsic>(I)) continue;
1191     if (I == Term) break;
1192 
1193     if (HInst)
1194       return false;
1195     HInst = I;
1196   }
1197 
1198   BasicBlock *BIParent = BI->getParent();
1199 
1200   // Check the instruction to be hoisted, if there is one.
1201   if (HInst) {
1202     // Don't hoist the instruction if it's unsafe or expensive.
1203     if (!isSafeToSpeculativelyExecute(HInst))
1204       return false;
1205     if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold)
1206       return false;
1207 
1208     // Do not hoist the instruction if any of its operands are defined but not
1209     // used in this BB. The transformation will prevent the operand from
1210     // being sunk into the use block.
1211     for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end();
1212          i != e; ++i) {
1213       Instruction *OpI = dyn_cast<Instruction>(*i);
1214       if (OpI && OpI->getParent() == BIParent &&
1215           !OpI->mayHaveSideEffects() &&
1216           !OpI->isUsedInBasicBlock(BIParent))
1217         return false;
1218     }
1219   }
1220 
1221   // Be conservative for now. FP select instruction can often be expensive.
1222   Value *BrCond = BI->getCondition();
1223   if (isa<FCmpInst>(BrCond))
1224     return false;
1225 
1226   // If BB1 is actually on the false edge of the conditional branch, remember
1227   // to swap the select operands later.
1228   bool Invert = false;
1229   if (BB1 != BI->getSuccessor(0)) {
1230     assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?");
1231     Invert = true;
1232   }
1233 
1234   // Collect interesting PHIs, and scan for hazards.
1235   SmallSetVector<std::pair<Value *, Value *>, 4> PHIs;
1236   BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0);
1237   for (BasicBlock::iterator I = BB2->begin();
1238        PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1239     Value *BB1V = PN->getIncomingValueForBlock(BB1);
1240     Value *BIParentV = PN->getIncomingValueForBlock(BIParent);
1241 
1242     // Skip PHIs which are trivial.
1243     if (BB1V == BIParentV)
1244       continue;
1245 
1246     // Check for saftey.
1247     if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) {
1248       // An unfolded ConstantExpr could end up getting expanded into
1249       // Instructions. Don't speculate this and another instruction at
1250       // the same time.
1251       if (HInst)
1252         return false;
1253       if (!isSafeToSpeculativelyExecute(CE))
1254         return false;
1255       if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold)
1256         return false;
1257     }
1258 
1259     // Ok, we may insert a select for this PHI.
1260     PHIs.insert(std::make_pair(BB1V, BIParentV));
1261   }
1262 
1263   // If there are no PHIs to process, bail early. This helps ensure idempotence
1264   // as well.
1265   if (PHIs.empty())
1266     return false;
1267 
1268   // If we get here, we can hoist the instruction and if-convert.
1269   DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";);
1270 
1271   // Hoist the instruction.
1272   if (HInst)
1273     BIParent->getInstList().splice(BI, BB1->getInstList(), HInst);
1274 
1275   // Insert selects and rewrite the PHI operands.
1276   IRBuilder<true, NoFolder> Builder(BI);
1277   for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
1278     Value *TrueV = PHIs[i].first;
1279     Value *FalseV = PHIs[i].second;
1280 
1281     // Create a select whose true value is the speculatively executed value and
1282     // false value is the previously determined FalseV.
1283     SelectInst *SI;
1284     if (Invert)
1285       SI = cast<SelectInst>
1286         (Builder.CreateSelect(BrCond, FalseV, TrueV,
1287                               FalseV->getName() + "." + TrueV->getName()));
1288     else
1289       SI = cast<SelectInst>
1290         (Builder.CreateSelect(BrCond, TrueV, FalseV,
1291                               TrueV->getName() + "." + FalseV->getName()));
1292 
1293     // Make the PHI node use the select for all incoming values for "then" and
1294     // "if" blocks.
1295     for (BasicBlock::iterator I = BB2->begin();
1296          PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1297       unsigned BB1I = PN->getBasicBlockIndex(BB1);
1298       unsigned BIParentI = PN->getBasicBlockIndex(BIParent);
1299       Value *BB1V = PN->getIncomingValue(BB1I);
1300       Value *BIParentV = PN->getIncomingValue(BIParentI);
1301       if (TrueV == BB1V && FalseV == BIParentV) {
1302         PN->setIncomingValue(BB1I, SI);
1303         PN->setIncomingValue(BIParentI, SI);
1304       }
1305     }
1306   }
1307 
1308   ++NumSpeculations;
1309   return true;
1310 }
1311 
1312 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1313 /// across this block.
BlockIsSimpleEnoughToThreadThrough(BasicBlock * BB)1314 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1315   BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1316   unsigned Size = 0;
1317 
1318   for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1319     if (isa<DbgInfoIntrinsic>(BBI))
1320       continue;
1321     if (Size > 10) return false;  // Don't clone large BB's.
1322     ++Size;
1323 
1324     // We can only support instructions that do not define values that are
1325     // live outside of the current basic block.
1326     for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1327          UI != E; ++UI) {
1328       Instruction *U = cast<Instruction>(*UI);
1329       if (U->getParent() != BB || isa<PHINode>(U)) return false;
1330     }
1331 
1332     // Looks ok, continue checking.
1333   }
1334 
1335   return true;
1336 }
1337 
1338 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1339 /// that is defined in the same block as the branch and if any PHI entries are
1340 /// constants, thread edges corresponding to that entry to be branches to their
1341 /// ultimate destination.
FoldCondBranchOnPHI(BranchInst * BI,const TargetData * TD)1342 static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) {
1343   BasicBlock *BB = BI->getParent();
1344   PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1345   // NOTE: we currently cannot transform this case if the PHI node is used
1346   // outside of the block.
1347   if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1348     return false;
1349 
1350   // Degenerate case of a single entry PHI.
1351   if (PN->getNumIncomingValues() == 1) {
1352     FoldSingleEntryPHINodes(PN->getParent());
1353     return true;
1354   }
1355 
1356   // Now we know that this block has multiple preds and two succs.
1357   if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1358 
1359   // Okay, this is a simple enough basic block.  See if any phi values are
1360   // constants.
1361   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1362     ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1363     if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1364 
1365     // Okay, we now know that all edges from PredBB should be revectored to
1366     // branch to RealDest.
1367     BasicBlock *PredBB = PN->getIncomingBlock(i);
1368     BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1369 
1370     if (RealDest == BB) continue;  // Skip self loops.
1371     // Skip if the predecessor's terminator is an indirect branch.
1372     if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1373 
1374     // The dest block might have PHI nodes, other predecessors and other
1375     // difficult cases.  Instead of being smart about this, just insert a new
1376     // block that jumps to the destination block, effectively splitting
1377     // the edge we are about to create.
1378     BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1379                                             RealDest->getName()+".critedge",
1380                                             RealDest->getParent(), RealDest);
1381     BranchInst::Create(RealDest, EdgeBB);
1382 
1383     // Update PHI nodes.
1384     AddPredecessorToBlock(RealDest, EdgeBB, BB);
1385 
1386     // BB may have instructions that are being threaded over.  Clone these
1387     // instructions into EdgeBB.  We know that there will be no uses of the
1388     // cloned instructions outside of EdgeBB.
1389     BasicBlock::iterator InsertPt = EdgeBB->begin();
1390     DenseMap<Value*, Value*> TranslateMap;  // Track translated values.
1391     for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1392       if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1393         TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1394         continue;
1395       }
1396       // Clone the instruction.
1397       Instruction *N = BBI->clone();
1398       if (BBI->hasName()) N->setName(BBI->getName()+".c");
1399 
1400       // Update operands due to translation.
1401       for (User::op_iterator i = N->op_begin(), e = N->op_end();
1402            i != e; ++i) {
1403         DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1404         if (PI != TranslateMap.end())
1405           *i = PI->second;
1406       }
1407 
1408       // Check for trivial simplification.
1409       if (Value *V = SimplifyInstruction(N, TD)) {
1410         TranslateMap[BBI] = V;
1411         delete N;   // Instruction folded away, don't need actual inst
1412       } else {
1413         // Insert the new instruction into its new home.
1414         EdgeBB->getInstList().insert(InsertPt, N);
1415         if (!BBI->use_empty())
1416           TranslateMap[BBI] = N;
1417       }
1418     }
1419 
1420     // Loop over all of the edges from PredBB to BB, changing them to branch
1421     // to EdgeBB instead.
1422     TerminatorInst *PredBBTI = PredBB->getTerminator();
1423     for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1424       if (PredBBTI->getSuccessor(i) == BB) {
1425         BB->removePredecessor(PredBB);
1426         PredBBTI->setSuccessor(i, EdgeBB);
1427       }
1428 
1429     // Recurse, simplifying any other constants.
1430     return FoldCondBranchOnPHI(BI, TD) | true;
1431   }
1432 
1433   return false;
1434 }
1435 
1436 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1437 /// PHI node, see if we can eliminate it.
FoldTwoEntryPHINode(PHINode * PN,const TargetData * TD)1438 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) {
1439   // Ok, this is a two entry PHI node.  Check to see if this is a simple "if
1440   // statement", which has a very simple dominance structure.  Basically, we
1441   // are trying to find the condition that is being branched on, which
1442   // subsequently causes this merge to happen.  We really want control
1443   // dependence information for this check, but simplifycfg can't keep it up
1444   // to date, and this catches most of the cases we care about anyway.
1445   BasicBlock *BB = PN->getParent();
1446   BasicBlock *IfTrue, *IfFalse;
1447   Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1448   if (!IfCond ||
1449       // Don't bother if the branch will be constant folded trivially.
1450       isa<ConstantInt>(IfCond))
1451     return false;
1452 
1453   // Okay, we found that we can merge this two-entry phi node into a select.
1454   // Doing so would require us to fold *all* two entry phi nodes in this block.
1455   // At some point this becomes non-profitable (particularly if the target
1456   // doesn't support cmov's).  Only do this transformation if there are two or
1457   // fewer PHI nodes in this block.
1458   unsigned NumPhis = 0;
1459   for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1460     if (NumPhis > 2)
1461       return false;
1462 
1463   // Loop over the PHI's seeing if we can promote them all to select
1464   // instructions.  While we are at it, keep track of the instructions
1465   // that need to be moved to the dominating block.
1466   SmallPtrSet<Instruction*, 4> AggressiveInsts;
1467   unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1468            MaxCostVal1 = PHINodeFoldingThreshold;
1469 
1470   for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1471     PHINode *PN = cast<PHINode>(II++);
1472     if (Value *V = SimplifyInstruction(PN, TD)) {
1473       PN->replaceAllUsesWith(V);
1474       PN->eraseFromParent();
1475       continue;
1476     }
1477 
1478     if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1479                              MaxCostVal0) ||
1480         !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1481                              MaxCostVal1))
1482       return false;
1483   }
1484 
1485   // If we folded the first phi, PN dangles at this point.  Refresh it.  If
1486   // we ran out of PHIs then we simplified them all.
1487   PN = dyn_cast<PHINode>(BB->begin());
1488   if (PN == 0) return true;
1489 
1490   // Don't fold i1 branches on PHIs which contain binary operators.  These can
1491   // often be turned into switches and other things.
1492   if (PN->getType()->isIntegerTy(1) &&
1493       (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1494        isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1495        isa<BinaryOperator>(IfCond)))
1496     return false;
1497 
1498   // If we all PHI nodes are promotable, check to make sure that all
1499   // instructions in the predecessor blocks can be promoted as well.  If
1500   // not, we won't be able to get rid of the control flow, so it's not
1501   // worth promoting to select instructions.
1502   BasicBlock *DomBlock = 0;
1503   BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1504   BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1505   if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1506     IfBlock1 = 0;
1507   } else {
1508     DomBlock = *pred_begin(IfBlock1);
1509     for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1510       if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1511         // This is not an aggressive instruction that we can promote.
1512         // Because of this, we won't be able to get rid of the control
1513         // flow, so the xform is not worth it.
1514         return false;
1515       }
1516   }
1517 
1518   if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1519     IfBlock2 = 0;
1520   } else {
1521     DomBlock = *pred_begin(IfBlock2);
1522     for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1523       if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1524         // This is not an aggressive instruction that we can promote.
1525         // Because of this, we won't be able to get rid of the control
1526         // flow, so the xform is not worth it.
1527         return false;
1528       }
1529   }
1530 
1531   DEBUG(dbgs() << "FOUND IF CONDITION!  " << *IfCond << "  T: "
1532                << IfTrue->getName() << "  F: " << IfFalse->getName() << "\n");
1533 
1534   // If we can still promote the PHI nodes after this gauntlet of tests,
1535   // do all of the PHI's now.
1536   Instruction *InsertPt = DomBlock->getTerminator();
1537   IRBuilder<true, NoFolder> Builder(InsertPt);
1538 
1539   // Move all 'aggressive' instructions, which are defined in the
1540   // conditional parts of the if's up to the dominating block.
1541   if (IfBlock1)
1542     DomBlock->getInstList().splice(InsertPt,
1543                                    IfBlock1->getInstList(), IfBlock1->begin(),
1544                                    IfBlock1->getTerminator());
1545   if (IfBlock2)
1546     DomBlock->getInstList().splice(InsertPt,
1547                                    IfBlock2->getInstList(), IfBlock2->begin(),
1548                                    IfBlock2->getTerminator());
1549 
1550   while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1551     // Change the PHI node into a select instruction.
1552     Value *TrueVal  = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1553     Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1554 
1555     SelectInst *NV =
1556       cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1557     PN->replaceAllUsesWith(NV);
1558     NV->takeName(PN);
1559     PN->eraseFromParent();
1560   }
1561 
1562   // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1563   // has been flattened.  Change DomBlock to jump directly to our new block to
1564   // avoid other simplifycfg's kicking in on the diamond.
1565   TerminatorInst *OldTI = DomBlock->getTerminator();
1566   Builder.SetInsertPoint(OldTI);
1567   Builder.CreateBr(BB);
1568   OldTI->eraseFromParent();
1569   return true;
1570 }
1571 
1572 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1573 /// to two returning blocks, try to merge them together into one return,
1574 /// introducing a select if the return values disagree.
SimplifyCondBranchToTwoReturns(BranchInst * BI,IRBuilder<> & Builder)1575 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1576                                            IRBuilder<> &Builder) {
1577   assert(BI->isConditional() && "Must be a conditional branch");
1578   BasicBlock *TrueSucc = BI->getSuccessor(0);
1579   BasicBlock *FalseSucc = BI->getSuccessor(1);
1580   ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1581   ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1582 
1583   // Check to ensure both blocks are empty (just a return) or optionally empty
1584   // with PHI nodes.  If there are other instructions, merging would cause extra
1585   // computation on one path or the other.
1586   if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1587     return false;
1588   if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1589     return false;
1590 
1591   Builder.SetInsertPoint(BI);
1592   // Okay, we found a branch that is going to two return nodes.  If
1593   // there is no return value for this function, just change the
1594   // branch into a return.
1595   if (FalseRet->getNumOperands() == 0) {
1596     TrueSucc->removePredecessor(BI->getParent());
1597     FalseSucc->removePredecessor(BI->getParent());
1598     Builder.CreateRetVoid();
1599     EraseTerminatorInstAndDCECond(BI);
1600     return true;
1601   }
1602 
1603   // Otherwise, figure out what the true and false return values are
1604   // so we can insert a new select instruction.
1605   Value *TrueValue = TrueRet->getReturnValue();
1606   Value *FalseValue = FalseRet->getReturnValue();
1607 
1608   // Unwrap any PHI nodes in the return blocks.
1609   if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1610     if (TVPN->getParent() == TrueSucc)
1611       TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1612   if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1613     if (FVPN->getParent() == FalseSucc)
1614       FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1615 
1616   // In order for this transformation to be safe, we must be able to
1617   // unconditionally execute both operands to the return.  This is
1618   // normally the case, but we could have a potentially-trapping
1619   // constant expression that prevents this transformation from being
1620   // safe.
1621   if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1622     if (TCV->canTrap())
1623       return false;
1624   if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1625     if (FCV->canTrap())
1626       return false;
1627 
1628   // Okay, we collected all the mapped values and checked them for sanity, and
1629   // defined to really do this transformation.  First, update the CFG.
1630   TrueSucc->removePredecessor(BI->getParent());
1631   FalseSucc->removePredecessor(BI->getParent());
1632 
1633   // Insert select instructions where needed.
1634   Value *BrCond = BI->getCondition();
1635   if (TrueValue) {
1636     // Insert a select if the results differ.
1637     if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1638     } else if (isa<UndefValue>(TrueValue)) {
1639       TrueValue = FalseValue;
1640     } else {
1641       TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1642                                        FalseValue, "retval");
1643     }
1644   }
1645 
1646   Value *RI = !TrueValue ?
1647     Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1648 
1649   (void) RI;
1650 
1651   DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1652                << "\n  " << *BI << "NewRet = " << *RI
1653                << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1654 
1655   EraseTerminatorInstAndDCECond(BI);
1656 
1657   return true;
1658 }
1659 
1660 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1661 /// probabilities of the branch taking each edge. Fills in the two APInt
1662 /// parameters and return true, or returns false if no or invalid metadata was
1663 /// found.
ExtractBranchMetadata(BranchInst * BI,APInt & ProbTrue,APInt & ProbFalse)1664 static bool ExtractBranchMetadata(BranchInst *BI,
1665                                   APInt &ProbTrue, APInt &ProbFalse) {
1666   assert(BI->isConditional() &&
1667          "Looking for probabilities on unconditional branch?");
1668   MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1669   if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1670   ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1671   ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1672   if (!CITrue || !CIFalse) return false;
1673   ProbTrue = CITrue->getValue();
1674   ProbFalse = CIFalse->getValue();
1675   assert(ProbTrue.getBitWidth() == 32 && ProbFalse.getBitWidth() == 32 &&
1676          "Branch probability metadata must be 32-bit integers");
1677   return true;
1678 }
1679 
1680 /// MultiplyAndLosePrecision - Multiplies A and B, then returns the result. In
1681 /// the event of overflow, logically-shifts all four inputs right until the
1682 /// multiply fits.
MultiplyAndLosePrecision(APInt & A,APInt & B,APInt & C,APInt & D,unsigned & BitsLost)1683 static APInt MultiplyAndLosePrecision(APInt &A, APInt &B, APInt &C, APInt &D,
1684                                       unsigned &BitsLost) {
1685   BitsLost = 0;
1686   bool Overflow = false;
1687   APInt Result = A.umul_ov(B, Overflow);
1688   if (Overflow) {
1689     APInt MaxB = APInt::getMaxValue(A.getBitWidth()).udiv(A);
1690     do {
1691       B = B.lshr(1);
1692       ++BitsLost;
1693     } while (B.ugt(MaxB));
1694     A = A.lshr(BitsLost);
1695     C = C.lshr(BitsLost);
1696     D = D.lshr(BitsLost);
1697     Result = A * B;
1698   }
1699   return Result;
1700 }
1701 
1702 /// checkCSEInPredecessor - Return true if the given instruction is available
1703 /// in its predecessor block. If yes, the instruction will be removed.
1704 ///
checkCSEInPredecessor(Instruction * Inst,BasicBlock * PB)1705 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1706   if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1707     return false;
1708   for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1709     Instruction *PBI = &*I;
1710     // Check whether Inst and PBI generate the same value.
1711     if (Inst->isIdenticalTo(PBI)) {
1712       Inst->replaceAllUsesWith(PBI);
1713       Inst->eraseFromParent();
1714       return true;
1715     }
1716   }
1717   return false;
1718 }
1719 
1720 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1721 /// predecessor branches to us and one of our successors, fold the block into
1722 /// the predecessor and use logical operations to pick the right destination.
FoldBranchToCommonDest(BranchInst * BI)1723 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1724   BasicBlock *BB = BI->getParent();
1725 
1726   Instruction *Cond = 0;
1727   if (BI->isConditional())
1728     Cond = dyn_cast<Instruction>(BI->getCondition());
1729   else {
1730     // For unconditional branch, check for a simple CFG pattern, where
1731     // BB has a single predecessor and BB's successor is also its predecessor's
1732     // successor. If such pattern exisits, check for CSE between BB and its
1733     // predecessor.
1734     if (BasicBlock *PB = BB->getSinglePredecessor())
1735       if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1736         if (PBI->isConditional() &&
1737             (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1738              BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1739           for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1740                I != E; ) {
1741             Instruction *Curr = I++;
1742             if (isa<CmpInst>(Curr)) {
1743               Cond = Curr;
1744               break;
1745             }
1746             // Quit if we can't remove this instruction.
1747             if (!checkCSEInPredecessor(Curr, PB))
1748               return false;
1749           }
1750         }
1751 
1752     if (Cond == 0)
1753       return false;
1754   }
1755 
1756   if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1757     Cond->getParent() != BB || !Cond->hasOneUse())
1758   return false;
1759 
1760   // Only allow this if the condition is a simple instruction that can be
1761   // executed unconditionally.  It must be in the same block as the branch, and
1762   // must be at the front of the block.
1763   BasicBlock::iterator FrontIt = BB->front();
1764 
1765   // Ignore dbg intrinsics.
1766   while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1767 
1768   // Allow a single instruction to be hoisted in addition to the compare
1769   // that feeds the branch.  We later ensure that any values that _it_ uses
1770   // were also live in the predecessor, so that we don't unnecessarily create
1771   // register pressure or inhibit out-of-order execution.
1772   Instruction *BonusInst = 0;
1773   if (&*FrontIt != Cond &&
1774       FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
1775       isSafeToSpeculativelyExecute(FrontIt)) {
1776     BonusInst = &*FrontIt;
1777     ++FrontIt;
1778 
1779     // Ignore dbg intrinsics.
1780     while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1781   }
1782 
1783   // Only a single bonus inst is allowed.
1784   if (&*FrontIt != Cond)
1785     return false;
1786 
1787   // Make sure the instruction after the condition is the cond branch.
1788   BasicBlock::iterator CondIt = Cond; ++CondIt;
1789 
1790   // Ingore dbg intrinsics.
1791   while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
1792 
1793   if (&*CondIt != BI)
1794     return false;
1795 
1796   // Cond is known to be a compare or binary operator.  Check to make sure that
1797   // neither operand is a potentially-trapping constant expression.
1798   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
1799     if (CE->canTrap())
1800       return false;
1801   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
1802     if (CE->canTrap())
1803       return false;
1804 
1805   // Finally, don't infinitely unroll conditional loops.
1806   BasicBlock *TrueDest  = BI->getSuccessor(0);
1807   BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
1808   if (TrueDest == BB || FalseDest == BB)
1809     return false;
1810 
1811   for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1812     BasicBlock *PredBlock = *PI;
1813     BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
1814 
1815     // Check that we have two conditional branches.  If there is a PHI node in
1816     // the common successor, verify that the same value flows in from both
1817     // blocks.
1818     SmallVector<PHINode*, 4> PHIs;
1819     if (PBI == 0 || PBI->isUnconditional() ||
1820         (BI->isConditional() &&
1821          !SafeToMergeTerminators(BI, PBI)) ||
1822         (!BI->isConditional() &&
1823          !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
1824       continue;
1825 
1826     // Determine if the two branches share a common destination.
1827     Instruction::BinaryOps Opc;
1828     bool InvertPredCond = false;
1829 
1830     if (BI->isConditional()) {
1831       if (PBI->getSuccessor(0) == TrueDest)
1832         Opc = Instruction::Or;
1833       else if (PBI->getSuccessor(1) == FalseDest)
1834         Opc = Instruction::And;
1835       else if (PBI->getSuccessor(0) == FalseDest)
1836         Opc = Instruction::And, InvertPredCond = true;
1837       else if (PBI->getSuccessor(1) == TrueDest)
1838         Opc = Instruction::Or, InvertPredCond = true;
1839       else
1840         continue;
1841     } else {
1842       if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
1843         continue;
1844     }
1845 
1846     // Ensure that any values used in the bonus instruction are also used
1847     // by the terminator of the predecessor.  This means that those values
1848     // must already have been resolved, so we won't be inhibiting the
1849     // out-of-order core by speculating them earlier.
1850     if (BonusInst) {
1851       // Collect the values used by the bonus inst
1852       SmallPtrSet<Value*, 4> UsedValues;
1853       for (Instruction::op_iterator OI = BonusInst->op_begin(),
1854            OE = BonusInst->op_end(); OI != OE; ++OI) {
1855         Value *V = *OI;
1856         if (!isa<Constant>(V))
1857           UsedValues.insert(V);
1858       }
1859 
1860       SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
1861       Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
1862 
1863       // Walk up to four levels back up the use-def chain of the predecessor's
1864       // terminator to see if all those values were used.  The choice of four
1865       // levels is arbitrary, to provide a compile-time-cost bound.
1866       while (!Worklist.empty()) {
1867         std::pair<Value*, unsigned> Pair = Worklist.back();
1868         Worklist.pop_back();
1869 
1870         if (Pair.second >= 4) continue;
1871         UsedValues.erase(Pair.first);
1872         if (UsedValues.empty()) break;
1873 
1874         if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
1875           for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
1876                OI != OE; ++OI)
1877             Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
1878         }
1879       }
1880 
1881       if (!UsedValues.empty()) return false;
1882     }
1883 
1884     DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
1885     IRBuilder<> Builder(PBI);
1886 
1887     // If we need to invert the condition in the pred block to match, do so now.
1888     if (InvertPredCond) {
1889       Value *NewCond = PBI->getCondition();
1890 
1891       if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
1892         CmpInst *CI = cast<CmpInst>(NewCond);
1893         CI->setPredicate(CI->getInversePredicate());
1894       } else {
1895         NewCond = Builder.CreateNot(NewCond,
1896                                     PBI->getCondition()->getName()+".not");
1897       }
1898 
1899       PBI->setCondition(NewCond);
1900       PBI->swapSuccessors();
1901     }
1902 
1903     // If we have a bonus inst, clone it into the predecessor block.
1904     Instruction *NewBonus = 0;
1905     if (BonusInst) {
1906       NewBonus = BonusInst->clone();
1907       PredBlock->getInstList().insert(PBI, NewBonus);
1908       NewBonus->takeName(BonusInst);
1909       BonusInst->setName(BonusInst->getName()+".old");
1910     }
1911 
1912     // Clone Cond into the predecessor basic block, and or/and the
1913     // two conditions together.
1914     Instruction *New = Cond->clone();
1915     if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
1916     PredBlock->getInstList().insert(PBI, New);
1917     New->takeName(Cond);
1918     Cond->setName(New->getName()+".old");
1919 
1920     if (BI->isConditional()) {
1921       Instruction *NewCond =
1922         cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
1923                                             New, "or.cond"));
1924       PBI->setCondition(NewCond);
1925 
1926       if (PBI->getSuccessor(0) == BB) {
1927         AddPredecessorToBlock(TrueDest, PredBlock, BB);
1928         PBI->setSuccessor(0, TrueDest);
1929       }
1930       if (PBI->getSuccessor(1) == BB) {
1931         AddPredecessorToBlock(FalseDest, PredBlock, BB);
1932         PBI->setSuccessor(1, FalseDest);
1933       }
1934     } else {
1935       // Update PHI nodes in the common successors.
1936       for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
1937         ConstantInt *PBI_C = cast<ConstantInt>(
1938           PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
1939         assert(PBI_C->getType()->isIntegerTy(1));
1940         Instruction *MergedCond = 0;
1941         if (PBI->getSuccessor(0) == TrueDest) {
1942           // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
1943           // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
1944           //       is false: !PBI_Cond and BI_Value
1945           Instruction *NotCond =
1946             cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
1947                                 "not.cond"));
1948           MergedCond =
1949             cast<Instruction>(Builder.CreateBinOp(Instruction::And,
1950                                 NotCond, New,
1951                                 "and.cond"));
1952           if (PBI_C->isOne())
1953             MergedCond =
1954               cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
1955                                   PBI->getCondition(), MergedCond,
1956                                   "or.cond"));
1957         } else {
1958           // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
1959           // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
1960           //       is false: PBI_Cond and BI_Value
1961           MergedCond =
1962             cast<Instruction>(Builder.CreateBinOp(Instruction::And,
1963                                 PBI->getCondition(), New,
1964                                 "and.cond"));
1965           if (PBI_C->isOne()) {
1966             Instruction *NotCond =
1967               cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
1968                                   "not.cond"));
1969             MergedCond =
1970               cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
1971                                   NotCond, MergedCond,
1972                                   "or.cond"));
1973           }
1974         }
1975         // Update PHI Node.
1976         PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
1977                                   MergedCond);
1978       }
1979       // Change PBI from Conditional to Unconditional.
1980       BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
1981       EraseTerminatorInstAndDCECond(PBI);
1982       PBI = New_PBI;
1983     }
1984 
1985     // TODO: If BB is reachable from all paths through PredBlock, then we
1986     // could replace PBI's branch probabilities with BI's.
1987 
1988     // Merge probability data into PredBlock's branch.
1989     APInt A, B, C, D;
1990     if (PBI->isConditional() && BI->isConditional() &&
1991         ExtractBranchMetadata(PBI, C, D) && ExtractBranchMetadata(BI, A, B)) {
1992       // Given IR which does:
1993       //   bbA:
1994       //     br i1 %x, label %bbB, label %bbC
1995       //   bbB:
1996       //     br i1 %y, label %bbD, label %bbC
1997       // Let's call the probability that we take the edge from %bbA to %bbB
1998       // 'a', from %bbA to %bbC, 'b', from %bbB to %bbD 'c' and from %bbB to
1999       // %bbC probability 'd'.
2000       //
2001       // We transform the IR into:
2002       //   bbA:
2003       //     br i1 %z, label %bbD, label %bbC
2004       // where the probability of going to %bbD is (a*c) and going to bbC is
2005       // (b+a*d).
2006       //
2007       // Probabilities aren't stored as ratios directly. Using branch weights,
2008       // we get:
2009       // (a*c)% = A*C, (b+(a*d))% = A*D+B*C+B*D.
2010 
2011       // In the event of overflow, we want to drop the LSB of the input
2012       // probabilities.
2013       unsigned BitsLost;
2014 
2015       // Ignore overflow result on ProbTrue.
2016       APInt ProbTrue = MultiplyAndLosePrecision(A, C, B, D, BitsLost);
2017 
2018       APInt Tmp1 = MultiplyAndLosePrecision(B, D, A, C, BitsLost);
2019       if (BitsLost) {
2020         ProbTrue = ProbTrue.lshr(BitsLost*2);
2021       }
2022 
2023       APInt Tmp2 = MultiplyAndLosePrecision(A, D, C, B, BitsLost);
2024       if (BitsLost) {
2025         ProbTrue = ProbTrue.lshr(BitsLost*2);
2026         Tmp1 = Tmp1.lshr(BitsLost*2);
2027       }
2028 
2029       APInt Tmp3 = MultiplyAndLosePrecision(B, C, A, D, BitsLost);
2030       if (BitsLost) {
2031         ProbTrue = ProbTrue.lshr(BitsLost*2);
2032         Tmp1 = Tmp1.lshr(BitsLost*2);
2033         Tmp2 = Tmp2.lshr(BitsLost*2);
2034       }
2035 
2036       bool Overflow1 = false, Overflow2 = false;
2037       APInt Tmp4 = Tmp2.uadd_ov(Tmp3, Overflow1);
2038       APInt ProbFalse = Tmp4.uadd_ov(Tmp1, Overflow2);
2039 
2040       if (Overflow1 || Overflow2) {
2041         ProbTrue = ProbTrue.lshr(1);
2042         Tmp1 = Tmp1.lshr(1);
2043         Tmp2 = Tmp2.lshr(1);
2044         Tmp3 = Tmp3.lshr(1);
2045         Tmp4 = Tmp2 + Tmp3;
2046         ProbFalse = Tmp4 + Tmp1;
2047       }
2048 
2049       // The sum of branch weights must fit in 32-bits.
2050       if (ProbTrue.isNegative() && ProbFalse.isNegative()) {
2051         ProbTrue = ProbTrue.lshr(1);
2052         ProbFalse = ProbFalse.lshr(1);
2053       }
2054 
2055       if (ProbTrue != ProbFalse) {
2056         // Normalize the result.
2057         APInt GCD = APIntOps::GreatestCommonDivisor(ProbTrue, ProbFalse);
2058         ProbTrue = ProbTrue.udiv(GCD);
2059         ProbFalse = ProbFalse.udiv(GCD);
2060 
2061         MDBuilder MDB(BI->getContext());
2062         MDNode *N = MDB.createBranchWeights(ProbTrue.getZExtValue(),
2063                                             ProbFalse.getZExtValue());
2064         PBI->setMetadata(LLVMContext::MD_prof, N);
2065       } else {
2066         PBI->setMetadata(LLVMContext::MD_prof, NULL);
2067       }
2068     } else {
2069       PBI->setMetadata(LLVMContext::MD_prof, NULL);
2070     }
2071 
2072     // Copy any debug value intrinsics into the end of PredBlock.
2073     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2074       if (isa<DbgInfoIntrinsic>(*I))
2075         I->clone()->insertBefore(PBI);
2076 
2077     return true;
2078   }
2079   return false;
2080 }
2081 
2082 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2083 /// predecessor of another block, this function tries to simplify it.  We know
2084 /// that PBI and BI are both conditional branches, and BI is in one of the
2085 /// successor blocks of PBI - PBI branches to BI.
SimplifyCondBranchToCondBranch(BranchInst * PBI,BranchInst * BI)2086 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2087   assert(PBI->isConditional() && BI->isConditional());
2088   BasicBlock *BB = BI->getParent();
2089 
2090   // If this block ends with a branch instruction, and if there is a
2091   // predecessor that ends on a branch of the same condition, make
2092   // this conditional branch redundant.
2093   if (PBI->getCondition() == BI->getCondition() &&
2094       PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2095     // Okay, the outcome of this conditional branch is statically
2096     // knowable.  If this block had a single pred, handle specially.
2097     if (BB->getSinglePredecessor()) {
2098       // Turn this into a branch on constant.
2099       bool CondIsTrue = PBI->getSuccessor(0) == BB;
2100       BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2101                                         CondIsTrue));
2102       return true;  // Nuke the branch on constant.
2103     }
2104 
2105     // Otherwise, if there are multiple predecessors, insert a PHI that merges
2106     // in the constant and simplify the block result.  Subsequent passes of
2107     // simplifycfg will thread the block.
2108     if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2109       pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2110       PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2111                                        std::distance(PB, PE),
2112                                        BI->getCondition()->getName() + ".pr",
2113                                        BB->begin());
2114       // Okay, we're going to insert the PHI node.  Since PBI is not the only
2115       // predecessor, compute the PHI'd conditional value for all of the preds.
2116       // Any predecessor where the condition is not computable we keep symbolic.
2117       for (pred_iterator PI = PB; PI != PE; ++PI) {
2118         BasicBlock *P = *PI;
2119         if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2120             PBI != BI && PBI->isConditional() &&
2121             PBI->getCondition() == BI->getCondition() &&
2122             PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2123           bool CondIsTrue = PBI->getSuccessor(0) == BB;
2124           NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2125                                               CondIsTrue), P);
2126         } else {
2127           NewPN->addIncoming(BI->getCondition(), P);
2128         }
2129       }
2130 
2131       BI->setCondition(NewPN);
2132       return true;
2133     }
2134   }
2135 
2136   // If this is a conditional branch in an empty block, and if any
2137   // predecessors is a conditional branch to one of our destinations,
2138   // fold the conditions into logical ops and one cond br.
2139   BasicBlock::iterator BBI = BB->begin();
2140   // Ignore dbg intrinsics.
2141   while (isa<DbgInfoIntrinsic>(BBI))
2142     ++BBI;
2143   if (&*BBI != BI)
2144     return false;
2145 
2146 
2147   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2148     if (CE->canTrap())
2149       return false;
2150 
2151   int PBIOp, BIOp;
2152   if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2153     PBIOp = BIOp = 0;
2154   else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2155     PBIOp = 0, BIOp = 1;
2156   else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2157     PBIOp = 1, BIOp = 0;
2158   else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2159     PBIOp = BIOp = 1;
2160   else
2161     return false;
2162 
2163   // Check to make sure that the other destination of this branch
2164   // isn't BB itself.  If so, this is an infinite loop that will
2165   // keep getting unwound.
2166   if (PBI->getSuccessor(PBIOp) == BB)
2167     return false;
2168 
2169   // Do not perform this transformation if it would require
2170   // insertion of a large number of select instructions. For targets
2171   // without predication/cmovs, this is a big pessimization.
2172   BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2173 
2174   unsigned NumPhis = 0;
2175   for (BasicBlock::iterator II = CommonDest->begin();
2176        isa<PHINode>(II); ++II, ++NumPhis)
2177     if (NumPhis > 2) // Disable this xform.
2178       return false;
2179 
2180   // Finally, if everything is ok, fold the branches to logical ops.
2181   BasicBlock *OtherDest  = BI->getSuccessor(BIOp ^ 1);
2182 
2183   DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2184                << "AND: " << *BI->getParent());
2185 
2186 
2187   // If OtherDest *is* BB, then BB is a basic block with a single conditional
2188   // branch in it, where one edge (OtherDest) goes back to itself but the other
2189   // exits.  We don't *know* that the program avoids the infinite loop
2190   // (even though that seems likely).  If we do this xform naively, we'll end up
2191   // recursively unpeeling the loop.  Since we know that (after the xform is
2192   // done) that the block *is* infinite if reached, we just make it an obviously
2193   // infinite loop with no cond branch.
2194   if (OtherDest == BB) {
2195     // Insert it at the end of the function, because it's either code,
2196     // or it won't matter if it's hot. :)
2197     BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2198                                                   "infloop", BB->getParent());
2199     BranchInst::Create(InfLoopBlock, InfLoopBlock);
2200     OtherDest = InfLoopBlock;
2201   }
2202 
2203   DEBUG(dbgs() << *PBI->getParent()->getParent());
2204 
2205   // BI may have other predecessors.  Because of this, we leave
2206   // it alone, but modify PBI.
2207 
2208   // Make sure we get to CommonDest on True&True directions.
2209   Value *PBICond = PBI->getCondition();
2210   IRBuilder<true, NoFolder> Builder(PBI);
2211   if (PBIOp)
2212     PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2213 
2214   Value *BICond = BI->getCondition();
2215   if (BIOp)
2216     BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2217 
2218   // Merge the conditions.
2219   Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2220 
2221   // Modify PBI to branch on the new condition to the new dests.
2222   PBI->setCondition(Cond);
2223   PBI->setSuccessor(0, CommonDest);
2224   PBI->setSuccessor(1, OtherDest);
2225 
2226   // OtherDest may have phi nodes.  If so, add an entry from PBI's
2227   // block that are identical to the entries for BI's block.
2228   AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2229 
2230   // We know that the CommonDest already had an edge from PBI to
2231   // it.  If it has PHIs though, the PHIs may have different
2232   // entries for BB and PBI's BB.  If so, insert a select to make
2233   // them agree.
2234   PHINode *PN;
2235   for (BasicBlock::iterator II = CommonDest->begin();
2236        (PN = dyn_cast<PHINode>(II)); ++II) {
2237     Value *BIV = PN->getIncomingValueForBlock(BB);
2238     unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2239     Value *PBIV = PN->getIncomingValue(PBBIdx);
2240     if (BIV != PBIV) {
2241       // Insert a select in PBI to pick the right value.
2242       Value *NV = cast<SelectInst>
2243         (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2244       PN->setIncomingValue(PBBIdx, NV);
2245     }
2246   }
2247 
2248   DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2249   DEBUG(dbgs() << *PBI->getParent()->getParent());
2250 
2251   // This basic block is probably dead.  We know it has at least
2252   // one fewer predecessor.
2253   return true;
2254 }
2255 
2256 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2257 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2258 // Takes care of updating the successors and removing the old terminator.
2259 // Also makes sure not to introduce new successors by assuming that edges to
2260 // non-successor TrueBBs and FalseBBs aren't reachable.
SimplifyTerminatorOnSelect(TerminatorInst * OldTerm,Value * Cond,BasicBlock * TrueBB,BasicBlock * FalseBB)2261 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2262                                        BasicBlock *TrueBB, BasicBlock *FalseBB){
2263   // Remove any superfluous successor edges from the CFG.
2264   // First, figure out which successors to preserve.
2265   // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2266   // successor.
2267   BasicBlock *KeepEdge1 = TrueBB;
2268   BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2269 
2270   // Then remove the rest.
2271   for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2272     BasicBlock *Succ = OldTerm->getSuccessor(I);
2273     // Make sure only to keep exactly one copy of each edge.
2274     if (Succ == KeepEdge1)
2275       KeepEdge1 = 0;
2276     else if (Succ == KeepEdge2)
2277       KeepEdge2 = 0;
2278     else
2279       Succ->removePredecessor(OldTerm->getParent());
2280   }
2281 
2282   IRBuilder<> Builder(OldTerm);
2283   Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2284 
2285   // Insert an appropriate new terminator.
2286   if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2287     if (TrueBB == FalseBB)
2288       // We were only looking for one successor, and it was present.
2289       // Create an unconditional branch to it.
2290       Builder.CreateBr(TrueBB);
2291     else
2292       // We found both of the successors we were looking for.
2293       // Create a conditional branch sharing the condition of the select.
2294       Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2295   } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2296     // Neither of the selected blocks were successors, so this
2297     // terminator must be unreachable.
2298     new UnreachableInst(OldTerm->getContext(), OldTerm);
2299   } else {
2300     // One of the selected values was a successor, but the other wasn't.
2301     // Insert an unconditional branch to the one that was found;
2302     // the edge to the one that wasn't must be unreachable.
2303     if (KeepEdge1 == 0)
2304       // Only TrueBB was found.
2305       Builder.CreateBr(TrueBB);
2306     else
2307       // Only FalseBB was found.
2308       Builder.CreateBr(FalseBB);
2309   }
2310 
2311   EraseTerminatorInstAndDCECond(OldTerm);
2312   return true;
2313 }
2314 
2315 // SimplifySwitchOnSelect - Replaces
2316 //   (switch (select cond, X, Y)) on constant X, Y
2317 // with a branch - conditional if X and Y lead to distinct BBs,
2318 // unconditional otherwise.
SimplifySwitchOnSelect(SwitchInst * SI,SelectInst * Select)2319 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2320   // Check for constant integer values in the select.
2321   ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2322   ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2323   if (!TrueVal || !FalseVal)
2324     return false;
2325 
2326   // Find the relevant condition and destinations.
2327   Value *Condition = Select->getCondition();
2328   BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2329   BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2330 
2331   // Perform the actual simplification.
2332   return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB);
2333 }
2334 
2335 // SimplifyIndirectBrOnSelect - Replaces
2336 //   (indirectbr (select cond, blockaddress(@fn, BlockA),
2337 //                             blockaddress(@fn, BlockB)))
2338 // with
2339 //   (br cond, BlockA, BlockB).
SimplifyIndirectBrOnSelect(IndirectBrInst * IBI,SelectInst * SI)2340 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2341   // Check that both operands of the select are block addresses.
2342   BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2343   BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2344   if (!TBA || !FBA)
2345     return false;
2346 
2347   // Extract the actual blocks.
2348   BasicBlock *TrueBB = TBA->getBasicBlock();
2349   BasicBlock *FalseBB = FBA->getBasicBlock();
2350 
2351   // Perform the actual simplification.
2352   return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB);
2353 }
2354 
2355 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2356 /// instruction (a seteq/setne with a constant) as the only instruction in a
2357 /// block that ends with an uncond branch.  We are looking for a very specific
2358 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified.  In
2359 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2360 /// default value goes to an uncond block with a seteq in it, we get something
2361 /// like:
2362 ///
2363 ///   switch i8 %A, label %DEFAULT [ i8 1, label %end    i8 2, label %end ]
2364 /// DEFAULT:
2365 ///   %tmp = icmp eq i8 %A, 92
2366 ///   br label %end
2367 /// end:
2368 ///   ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2369 ///
2370 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2371 /// the PHI, merging the third icmp into the switch.
TryToSimplifyUncondBranchWithICmpInIt(ICmpInst * ICI,const TargetData * TD,IRBuilder<> & Builder)2372 static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
2373                                                   const TargetData *TD,
2374                                                   IRBuilder<> &Builder) {
2375   BasicBlock *BB = ICI->getParent();
2376 
2377   // If the block has any PHIs in it or the icmp has multiple uses, it is too
2378   // complex.
2379   if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2380 
2381   Value *V = ICI->getOperand(0);
2382   ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2383 
2384   // The pattern we're looking for is where our only predecessor is a switch on
2385   // 'V' and this block is the default case for the switch.  In this case we can
2386   // fold the compared value into the switch to simplify things.
2387   BasicBlock *Pred = BB->getSinglePredecessor();
2388   if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2389 
2390   SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2391   if (SI->getCondition() != V)
2392     return false;
2393 
2394   // If BB is reachable on a non-default case, then we simply know the value of
2395   // V in this block.  Substitute it and constant fold the icmp instruction
2396   // away.
2397   if (SI->getDefaultDest() != BB) {
2398     ConstantInt *VVal = SI->findCaseDest(BB);
2399     assert(VVal && "Should have a unique destination value");
2400     ICI->setOperand(0, VVal);
2401 
2402     if (Value *V = SimplifyInstruction(ICI, TD)) {
2403       ICI->replaceAllUsesWith(V);
2404       ICI->eraseFromParent();
2405     }
2406     // BB is now empty, so it is likely to simplify away.
2407     return SimplifyCFG(BB) | true;
2408   }
2409 
2410   // Ok, the block is reachable from the default dest.  If the constant we're
2411   // comparing exists in one of the other edges, then we can constant fold ICI
2412   // and zap it.
2413   if (SI->findCaseValue(Cst) != SI->case_default()) {
2414     Value *V;
2415     if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2416       V = ConstantInt::getFalse(BB->getContext());
2417     else
2418       V = ConstantInt::getTrue(BB->getContext());
2419 
2420     ICI->replaceAllUsesWith(V);
2421     ICI->eraseFromParent();
2422     // BB is now empty, so it is likely to simplify away.
2423     return SimplifyCFG(BB) | true;
2424   }
2425 
2426   // The use of the icmp has to be in the 'end' block, by the only PHI node in
2427   // the block.
2428   BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2429   PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2430   if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2431       isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2432     return false;
2433 
2434   // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2435   // true in the PHI.
2436   Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2437   Constant *NewCst     = ConstantInt::getFalse(BB->getContext());
2438 
2439   if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2440     std::swap(DefaultCst, NewCst);
2441 
2442   // Replace ICI (which is used by the PHI for the default value) with true or
2443   // false depending on if it is EQ or NE.
2444   ICI->replaceAllUsesWith(DefaultCst);
2445   ICI->eraseFromParent();
2446 
2447   // Okay, the switch goes to this block on a default value.  Add an edge from
2448   // the switch to the merge point on the compared value.
2449   BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2450                                          BB->getParent(), BB);
2451   SI->addCase(Cst, NewBB);
2452 
2453   // NewBB branches to the phi block, add the uncond branch and the phi entry.
2454   Builder.SetInsertPoint(NewBB);
2455   Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2456   Builder.CreateBr(SuccBlock);
2457   PHIUse->addIncoming(NewCst, NewBB);
2458   return true;
2459 }
2460 
2461 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2462 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2463 /// fold it into a switch instruction if so.
SimplifyBranchOnICmpChain(BranchInst * BI,const TargetData * TD,IRBuilder<> & Builder)2464 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD,
2465                                       IRBuilder<> &Builder) {
2466   Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2467   if (Cond == 0) return false;
2468 
2469 
2470   // Change br (X == 0 | X == 1), T, F into a switch instruction.
2471   // If this is a bunch of seteq's or'd together, or if it's a bunch of
2472   // 'setne's and'ed together, collect them.
2473   Value *CompVal = 0;
2474   std::vector<ConstantInt*> Values;
2475   bool TrueWhenEqual = true;
2476   Value *ExtraCase = 0;
2477   unsigned UsedICmps = 0;
2478 
2479   if (Cond->getOpcode() == Instruction::Or) {
2480     CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2481                                      UsedICmps);
2482   } else if (Cond->getOpcode() == Instruction::And) {
2483     CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2484                                      UsedICmps);
2485     TrueWhenEqual = false;
2486   }
2487 
2488   // If we didn't have a multiply compared value, fail.
2489   if (CompVal == 0) return false;
2490 
2491   // Avoid turning single icmps into a switch.
2492   if (UsedICmps <= 1)
2493     return false;
2494 
2495   // There might be duplicate constants in the list, which the switch
2496   // instruction can't handle, remove them now.
2497   array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2498   Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2499 
2500   // If Extra was used, we require at least two switch values to do the
2501   // transformation.  A switch with one value is just an cond branch.
2502   if (ExtraCase && Values.size() < 2) return false;
2503 
2504   // TODO: Preserve branch weight metadata, similarly to how
2505   // FoldValueComparisonIntoPredecessors preserves it.
2506 
2507   // Figure out which block is which destination.
2508   BasicBlock *DefaultBB = BI->getSuccessor(1);
2509   BasicBlock *EdgeBB    = BI->getSuccessor(0);
2510   if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2511 
2512   BasicBlock *BB = BI->getParent();
2513 
2514   DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2515                << " cases into SWITCH.  BB is:\n" << *BB);
2516 
2517   // If there are any extra values that couldn't be folded into the switch
2518   // then we evaluate them with an explicit branch first.  Split the block
2519   // right before the condbr to handle it.
2520   if (ExtraCase) {
2521     BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2522     // Remove the uncond branch added to the old block.
2523     TerminatorInst *OldTI = BB->getTerminator();
2524     Builder.SetInsertPoint(OldTI);
2525 
2526     if (TrueWhenEqual)
2527       Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2528     else
2529       Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2530 
2531     OldTI->eraseFromParent();
2532 
2533     // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2534     // for the edge we just added.
2535     AddPredecessorToBlock(EdgeBB, BB, NewBB);
2536 
2537     DEBUG(dbgs() << "  ** 'icmp' chain unhandled condition: " << *ExtraCase
2538           << "\nEXTRABB = " << *BB);
2539     BB = NewBB;
2540   }
2541 
2542   Builder.SetInsertPoint(BI);
2543   // Convert pointer to int before we switch.
2544   if (CompVal->getType()->isPointerTy()) {
2545     assert(TD && "Cannot switch on pointer without TargetData");
2546     CompVal = Builder.CreatePtrToInt(CompVal,
2547                                      TD->getIntPtrType(CompVal->getContext()),
2548                                      "magicptr");
2549   }
2550 
2551   // Create the new switch instruction now.
2552   SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2553 
2554   // Add all of the 'cases' to the switch instruction.
2555   for (unsigned i = 0, e = Values.size(); i != e; ++i)
2556     New->addCase(Values[i], EdgeBB);
2557 
2558   // We added edges from PI to the EdgeBB.  As such, if there were any
2559   // PHI nodes in EdgeBB, they need entries to be added corresponding to
2560   // the number of edges added.
2561   for (BasicBlock::iterator BBI = EdgeBB->begin();
2562        isa<PHINode>(BBI); ++BBI) {
2563     PHINode *PN = cast<PHINode>(BBI);
2564     Value *InVal = PN->getIncomingValueForBlock(BB);
2565     for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2566       PN->addIncoming(InVal, BB);
2567   }
2568 
2569   // Erase the old branch instruction.
2570   EraseTerminatorInstAndDCECond(BI);
2571 
2572   DEBUG(dbgs() << "  ** 'icmp' chain result is:\n" << *BB << '\n');
2573   return true;
2574 }
2575 
SimplifyResume(ResumeInst * RI,IRBuilder<> & Builder)2576 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2577   // If this is a trivial landing pad that just continues unwinding the caught
2578   // exception then zap the landing pad, turning its invokes into calls.
2579   BasicBlock *BB = RI->getParent();
2580   LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2581   if (RI->getValue() != LPInst)
2582     // Not a landing pad, or the resume is not unwinding the exception that
2583     // caused control to branch here.
2584     return false;
2585 
2586   // Check that there are no other instructions except for debug intrinsics.
2587   BasicBlock::iterator I = LPInst, E = RI;
2588   while (++I != E)
2589     if (!isa<DbgInfoIntrinsic>(I))
2590       return false;
2591 
2592   // Turn all invokes that unwind here into calls and delete the basic block.
2593   for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2594     InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2595     SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2596     // Insert a call instruction before the invoke.
2597     CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2598     Call->takeName(II);
2599     Call->setCallingConv(II->getCallingConv());
2600     Call->setAttributes(II->getAttributes());
2601     Call->setDebugLoc(II->getDebugLoc());
2602 
2603     // Anything that used the value produced by the invoke instruction now uses
2604     // the value produced by the call instruction.  Note that we do this even
2605     // for void functions and calls with no uses so that the callgraph edge is
2606     // updated.
2607     II->replaceAllUsesWith(Call);
2608     BB->removePredecessor(II->getParent());
2609 
2610     // Insert a branch to the normal destination right before the invoke.
2611     BranchInst::Create(II->getNormalDest(), II);
2612 
2613     // Finally, delete the invoke instruction!
2614     II->eraseFromParent();
2615   }
2616 
2617   // The landingpad is now unreachable.  Zap it.
2618   BB->eraseFromParent();
2619   return true;
2620 }
2621 
SimplifyReturn(ReturnInst * RI,IRBuilder<> & Builder)2622 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2623   BasicBlock *BB = RI->getParent();
2624   if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2625 
2626   // Find predecessors that end with branches.
2627   SmallVector<BasicBlock*, 8> UncondBranchPreds;
2628   SmallVector<BranchInst*, 8> CondBranchPreds;
2629   for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2630     BasicBlock *P = *PI;
2631     TerminatorInst *PTI = P->getTerminator();
2632     if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2633       if (BI->isUnconditional())
2634         UncondBranchPreds.push_back(P);
2635       else
2636         CondBranchPreds.push_back(BI);
2637     }
2638   }
2639 
2640   // If we found some, do the transformation!
2641   if (!UncondBranchPreds.empty() && DupRet) {
2642     while (!UncondBranchPreds.empty()) {
2643       BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2644       DEBUG(dbgs() << "FOLDING: " << *BB
2645             << "INTO UNCOND BRANCH PRED: " << *Pred);
2646       (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2647     }
2648 
2649     // If we eliminated all predecessors of the block, delete the block now.
2650     if (pred_begin(BB) == pred_end(BB))
2651       // We know there are no successors, so just nuke the block.
2652       BB->eraseFromParent();
2653 
2654     return true;
2655   }
2656 
2657   // Check out all of the conditional branches going to this return
2658   // instruction.  If any of them just select between returns, change the
2659   // branch itself into a select/return pair.
2660   while (!CondBranchPreds.empty()) {
2661     BranchInst *BI = CondBranchPreds.pop_back_val();
2662 
2663     // Check to see if the non-BB successor is also a return block.
2664     if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2665         isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2666         SimplifyCondBranchToTwoReturns(BI, Builder))
2667       return true;
2668   }
2669   return false;
2670 }
2671 
SimplifyUnreachable(UnreachableInst * UI)2672 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2673   BasicBlock *BB = UI->getParent();
2674 
2675   bool Changed = false;
2676 
2677   // If there are any instructions immediately before the unreachable that can
2678   // be removed, do so.
2679   while (UI != BB->begin()) {
2680     BasicBlock::iterator BBI = UI;
2681     --BBI;
2682     // Do not delete instructions that can have side effects which might cause
2683     // the unreachable to not be reachable; specifically, calls and volatile
2684     // operations may have this effect.
2685     if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2686 
2687     if (BBI->mayHaveSideEffects()) {
2688       if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2689         if (SI->isVolatile())
2690           break;
2691       } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2692         if (LI->isVolatile())
2693           break;
2694       } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2695         if (RMWI->isVolatile())
2696           break;
2697       } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2698         if (CXI->isVolatile())
2699           break;
2700       } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2701                  !isa<LandingPadInst>(BBI)) {
2702         break;
2703       }
2704       // Note that deleting LandingPad's here is in fact okay, although it
2705       // involves a bit of subtle reasoning. If this inst is a LandingPad,
2706       // all the predecessors of this block will be the unwind edges of Invokes,
2707       // and we can therefore guarantee this block will be erased.
2708     }
2709 
2710     // Delete this instruction (any uses are guaranteed to be dead)
2711     if (!BBI->use_empty())
2712       BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2713     BBI->eraseFromParent();
2714     Changed = true;
2715   }
2716 
2717   // If the unreachable instruction is the first in the block, take a gander
2718   // at all of the predecessors of this instruction, and simplify them.
2719   if (&BB->front() != UI) return Changed;
2720 
2721   SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2722   for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
2723     TerminatorInst *TI = Preds[i]->getTerminator();
2724     IRBuilder<> Builder(TI);
2725     if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2726       if (BI->isUnconditional()) {
2727         if (BI->getSuccessor(0) == BB) {
2728           new UnreachableInst(TI->getContext(), TI);
2729           TI->eraseFromParent();
2730           Changed = true;
2731         }
2732       } else {
2733         if (BI->getSuccessor(0) == BB) {
2734           Builder.CreateBr(BI->getSuccessor(1));
2735           EraseTerminatorInstAndDCECond(BI);
2736         } else if (BI->getSuccessor(1) == BB) {
2737           Builder.CreateBr(BI->getSuccessor(0));
2738           EraseTerminatorInstAndDCECond(BI);
2739           Changed = true;
2740         }
2741       }
2742     } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2743       for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2744            i != e; ++i)
2745         if (i.getCaseSuccessor() == BB) {
2746           BB->removePredecessor(SI->getParent());
2747           SI->removeCase(i);
2748           --i; --e;
2749           Changed = true;
2750         }
2751       // If the default value is unreachable, figure out the most popular
2752       // destination and make it the default.
2753       if (SI->getDefaultDest() == BB) {
2754         std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
2755         for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2756              i != e; ++i) {
2757           std::pair<unsigned, unsigned> &entry =
2758               Popularity[i.getCaseSuccessor()];
2759           if (entry.first == 0) {
2760             entry.first = 1;
2761             entry.second = i.getCaseIndex();
2762           } else {
2763             entry.first++;
2764           }
2765         }
2766 
2767         // Find the most popular block.
2768         unsigned MaxPop = 0;
2769         unsigned MaxIndex = 0;
2770         BasicBlock *MaxBlock = 0;
2771         for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
2772              I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
2773           if (I->second.first > MaxPop ||
2774               (I->second.first == MaxPop && MaxIndex > I->second.second)) {
2775             MaxPop = I->second.first;
2776             MaxIndex = I->second.second;
2777             MaxBlock = I->first;
2778           }
2779         }
2780         if (MaxBlock) {
2781           // Make this the new default, allowing us to delete any explicit
2782           // edges to it.
2783           SI->setDefaultDest(MaxBlock);
2784           Changed = true;
2785 
2786           // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
2787           // it.
2788           if (isa<PHINode>(MaxBlock->begin()))
2789             for (unsigned i = 0; i != MaxPop-1; ++i)
2790               MaxBlock->removePredecessor(SI->getParent());
2791 
2792           for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2793                i != e; ++i)
2794             if (i.getCaseSuccessor() == MaxBlock) {
2795               SI->removeCase(i);
2796               --i; --e;
2797             }
2798         }
2799       }
2800     } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
2801       if (II->getUnwindDest() == BB) {
2802         // Convert the invoke to a call instruction.  This would be a good
2803         // place to note that the call does not throw though.
2804         BranchInst *BI = Builder.CreateBr(II->getNormalDest());
2805         II->removeFromParent();   // Take out of symbol table
2806 
2807         // Insert the call now...
2808         SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
2809         Builder.SetInsertPoint(BI);
2810         CallInst *CI = Builder.CreateCall(II->getCalledValue(),
2811                                           Args, II->getName());
2812         CI->setCallingConv(II->getCallingConv());
2813         CI->setAttributes(II->getAttributes());
2814         // If the invoke produced a value, the call does now instead.
2815         II->replaceAllUsesWith(CI);
2816         delete II;
2817         Changed = true;
2818       }
2819     }
2820   }
2821 
2822   // If this block is now dead, remove it.
2823   if (pred_begin(BB) == pred_end(BB) &&
2824       BB != &BB->getParent()->getEntryBlock()) {
2825     // We know there are no successors, so just nuke the block.
2826     BB->eraseFromParent();
2827     return true;
2828   }
2829 
2830   return Changed;
2831 }
2832 
2833 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
2834 /// integer range comparison into a sub, an icmp and a branch.
TurnSwitchRangeIntoICmp(SwitchInst * SI,IRBuilder<> & Builder)2835 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
2836   assert(SI->getNumCases() > 1 && "Degenerate switch?");
2837 
2838   // Make sure all cases point to the same destination and gather the values.
2839   SmallVector<ConstantInt *, 16> Cases;
2840   SwitchInst::CaseIt I = SI->case_begin();
2841   Cases.push_back(I.getCaseValue());
2842   SwitchInst::CaseIt PrevI = I++;
2843   for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
2844     if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
2845       return false;
2846     Cases.push_back(I.getCaseValue());
2847   }
2848   assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
2849 
2850   // Sort the case values, then check if they form a range we can transform.
2851   array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
2852   for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
2853     if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
2854       return false;
2855   }
2856 
2857   Constant *Offset = ConstantExpr::getNeg(Cases.back());
2858   Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
2859 
2860   Value *Sub = SI->getCondition();
2861   if (!Offset->isNullValue())
2862     Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
2863   Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
2864   Builder.CreateCondBr(
2865       Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
2866 
2867   // Prune obsolete incoming values off the successor's PHI nodes.
2868   for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
2869        isa<PHINode>(BBI); ++BBI) {
2870     for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
2871       cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
2872   }
2873   SI->eraseFromParent();
2874 
2875   return true;
2876 }
2877 
2878 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
2879 /// and use it to remove dead cases.
EliminateDeadSwitchCases(SwitchInst * SI)2880 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
2881   Value *Cond = SI->getCondition();
2882   unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
2883   APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
2884   ComputeMaskedBits(Cond, KnownZero, KnownOne);
2885 
2886   // Gather dead cases.
2887   SmallVector<ConstantInt*, 8> DeadCases;
2888   for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
2889     if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
2890         (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
2891       DeadCases.push_back(I.getCaseValue());
2892       DEBUG(dbgs() << "SimplifyCFG: switch case '"
2893                    << I.getCaseValue() << "' is dead.\n");
2894     }
2895   }
2896 
2897   // Remove dead cases from the switch.
2898   for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
2899     SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
2900     assert(Case != SI->case_default() &&
2901            "Case was not found. Probably mistake in DeadCases forming.");
2902     // Prune unused values from PHI nodes.
2903     Case.getCaseSuccessor()->removePredecessor(SI->getParent());
2904     SI->removeCase(Case);
2905   }
2906 
2907   return !DeadCases.empty();
2908 }
2909 
2910 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
2911 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
2912 /// by an unconditional branch), look at the phi node for BB in the successor
2913 /// block and see if the incoming value is equal to CaseValue. If so, return
2914 /// the phi node, and set PhiIndex to BB's index in the phi node.
FindPHIForConditionForwarding(ConstantInt * CaseValue,BasicBlock * BB,int * PhiIndex)2915 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
2916                                               BasicBlock *BB,
2917                                               int *PhiIndex) {
2918   if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
2919     return NULL; // BB must be empty to be a candidate for simplification.
2920   if (!BB->getSinglePredecessor())
2921     return NULL; // BB must be dominated by the switch.
2922 
2923   BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
2924   if (!Branch || !Branch->isUnconditional())
2925     return NULL; // Terminator must be unconditional branch.
2926 
2927   BasicBlock *Succ = Branch->getSuccessor(0);
2928 
2929   BasicBlock::iterator I = Succ->begin();
2930   while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
2931     int Idx = PHI->getBasicBlockIndex(BB);
2932     assert(Idx >= 0 && "PHI has no entry for predecessor?");
2933 
2934     Value *InValue = PHI->getIncomingValue(Idx);
2935     if (InValue != CaseValue) continue;
2936 
2937     *PhiIndex = Idx;
2938     return PHI;
2939   }
2940 
2941   return NULL;
2942 }
2943 
2944 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
2945 /// instruction to a phi node dominated by the switch, if that would mean that
2946 /// some of the destination blocks of the switch can be folded away.
2947 /// Returns true if a change is made.
ForwardSwitchConditionToPHI(SwitchInst * SI)2948 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
2949   typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
2950   ForwardingNodesMap ForwardingNodes;
2951 
2952   for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
2953     ConstantInt *CaseValue = I.getCaseValue();
2954     BasicBlock *CaseDest = I.getCaseSuccessor();
2955 
2956     int PhiIndex;
2957     PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
2958                                                  &PhiIndex);
2959     if (!PHI) continue;
2960 
2961     ForwardingNodes[PHI].push_back(PhiIndex);
2962   }
2963 
2964   bool Changed = false;
2965 
2966   for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
2967        E = ForwardingNodes.end(); I != E; ++I) {
2968     PHINode *Phi = I->first;
2969     SmallVector<int,4> &Indexes = I->second;
2970 
2971     if (Indexes.size() < 2) continue;
2972 
2973     for (size_t I = 0, E = Indexes.size(); I != E; ++I)
2974       Phi->setIncomingValue(Indexes[I], SI->getCondition());
2975     Changed = true;
2976   }
2977 
2978   return Changed;
2979 }
2980 
2981 /// ValidLookupTableConstant - Return true if the backend will be able to handle
2982 /// initializing an array of constants like C.
ValidLookupTableConstant(Constant * C)2983 static bool ValidLookupTableConstant(Constant *C) {
2984   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2985     return CE->isGEPWithNoNotionalOverIndexing();
2986 
2987   return isa<ConstantFP>(C) ||
2988       isa<ConstantInt>(C) ||
2989       isa<ConstantPointerNull>(C) ||
2990       isa<GlobalValue>(C) ||
2991       isa<UndefValue>(C);
2992 }
2993 
2994 /// GetCaseResulsts - Try to determine the resulting constant values in phi
2995 /// nodes at the common destination basic block for one of the case
2996 /// destinations of a switch instruction.
GetCaseResults(SwitchInst * SI,BasicBlock * CaseDest,BasicBlock ** CommonDest,SmallVector<std::pair<PHINode *,Constant * >,4> & Res)2997 static bool GetCaseResults(SwitchInst *SI,
2998                            BasicBlock *CaseDest,
2999                            BasicBlock **CommonDest,
3000                            SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) {
3001   // The block from which we enter the common destination.
3002   BasicBlock *Pred = SI->getParent();
3003 
3004   // If CaseDest is empty, continue to its successor.
3005   if (CaseDest->getFirstNonPHIOrDbg() == CaseDest->getTerminator() &&
3006       !isa<PHINode>(CaseDest->begin())) {
3007 
3008     TerminatorInst *Terminator = CaseDest->getTerminator();
3009     if (Terminator->getNumSuccessors() != 1)
3010       return false;
3011 
3012     Pred = CaseDest;
3013     CaseDest = Terminator->getSuccessor(0);
3014   }
3015 
3016   // If we did not have a CommonDest before, use the current one.
3017   if (!*CommonDest)
3018     *CommonDest = CaseDest;
3019   // If the destination isn't the common one, abort.
3020   if (CaseDest != *CommonDest)
3021     return false;
3022 
3023   // Get the values for this case from phi nodes in the destination block.
3024   BasicBlock::iterator I = (*CommonDest)->begin();
3025   while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3026     int Idx = PHI->getBasicBlockIndex(Pred);
3027     if (Idx == -1)
3028       continue;
3029 
3030     Constant *ConstVal = dyn_cast<Constant>(PHI->getIncomingValue(Idx));
3031     if (!ConstVal)
3032       return false;
3033 
3034     // Be conservative about which kinds of constants we support.
3035     if (!ValidLookupTableConstant(ConstVal))
3036       return false;
3037 
3038     Res.push_back(std::make_pair(PHI, ConstVal));
3039   }
3040 
3041   return true;
3042 }
3043 
3044 /// BuildLookupTable - Build a lookup table with the contents of Results, using
3045 /// DefaultResult to fill the holes in the table. If the table ends up
3046 /// containing the same result in each element, set *SingleResult to that value
3047 /// and return NULL.
BuildLookupTable(Module & M,uint64_t TableSize,ConstantInt * Offset,const SmallVector<std::pair<ConstantInt *,Constant * >,4> & Results,Constant * DefaultResult,Constant ** SingleResult)3048 static GlobalVariable *BuildLookupTable(Module &M,
3049                                         uint64_t TableSize,
3050                                         ConstantInt *Offset,
3051               const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Results,
3052                                         Constant *DefaultResult,
3053                                         Constant **SingleResult) {
3054   assert(Results.size() && "Need values to build lookup table");
3055   assert(TableSize >= Results.size() && "Table needs to hold all values");
3056 
3057   // If all values in the table are equal, this is that value.
3058   Constant *SameResult = Results.begin()->second;
3059 
3060   // Build up the table contents.
3061   std::vector<Constant*> TableContents(TableSize);
3062   for (size_t I = 0, E = Results.size(); I != E; ++I) {
3063     ConstantInt *CaseVal = Results[I].first;
3064     Constant *CaseRes = Results[I].second;
3065 
3066     uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
3067     TableContents[Idx] = CaseRes;
3068 
3069     if (CaseRes != SameResult)
3070       SameResult = NULL;
3071   }
3072 
3073   // Fill in any holes in the table with the default result.
3074   if (Results.size() < TableSize) {
3075     for (unsigned i = 0; i < TableSize; ++i) {
3076       if (!TableContents[i])
3077         TableContents[i] = DefaultResult;
3078     }
3079 
3080     if (DefaultResult != SameResult)
3081       SameResult = NULL;
3082   }
3083 
3084   // Same result was used in the entire table; just return that.
3085   if (SameResult) {
3086     *SingleResult = SameResult;
3087     return NULL;
3088   }
3089 
3090   ArrayType *ArrayTy = ArrayType::get(DefaultResult->getType(), TableSize);
3091   Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3092 
3093   GlobalVariable *GV = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3094                                           GlobalVariable::PrivateLinkage,
3095                                           Initializer,
3096                                           "switch.table");
3097   GV->setUnnamedAddr(true);
3098   return GV;
3099 }
3100 
3101 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3102 /// phi nodes in a common successor block with different constant values,
3103 /// replace the switch with lookup tables.
SwitchToLookupTable(SwitchInst * SI,IRBuilder<> & Builder)3104 static bool SwitchToLookupTable(SwitchInst *SI,
3105                                 IRBuilder<> &Builder) {
3106   assert(SI->getNumCases() > 1 && "Degenerate switch?");
3107   // FIXME: Handle unreachable cases.
3108 
3109   // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3110   // split off a dense part and build a lookup table for that.
3111 
3112   // FIXME: If the results are all integers and the lookup table would fit in a
3113   // target-legal register, we should store them as a bitmap and use shift/mask
3114   // to look up the result.
3115 
3116   // FIXME: This creates arrays of GEPs to constant strings, which means each
3117   // GEP needs a runtime relocation in PIC code. We should just build one big
3118   // string and lookup indices into that.
3119 
3120   // Ignore the switch if the number of cases are too small.
3121   // This is similar to the check when building jump tables in
3122   // SelectionDAGBuilder::handleJTSwitchCase.
3123   // FIXME: Determine the best cut-off.
3124   if (SI->getNumCases() < 4)
3125     return false;
3126 
3127   // Figure out the corresponding result for each case value and phi node in the
3128   // common destination, as well as the the min and max case values.
3129   assert(SI->case_begin() != SI->case_end());
3130   SwitchInst::CaseIt CI = SI->case_begin();
3131   ConstantInt *MinCaseVal = CI.getCaseValue();
3132   ConstantInt *MaxCaseVal = CI.getCaseValue();
3133 
3134   BasicBlock *CommonDest = NULL;
3135   typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3136   SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3137   SmallDenseMap<PHINode*, Constant*> DefaultResults;
3138   SmallDenseMap<PHINode*, Type*> ResultTypes;
3139   SmallVector<PHINode*, 4> PHIs;
3140 
3141   for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3142     ConstantInt *CaseVal = CI.getCaseValue();
3143     if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3144       MinCaseVal = CaseVal;
3145     if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3146       MaxCaseVal = CaseVal;
3147 
3148     // Resulting value at phi nodes for this case value.
3149     typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3150     ResultsTy Results;
3151     if (!GetCaseResults(SI, CI.getCaseSuccessor(), &CommonDest, Results))
3152       return false;
3153 
3154     // Append the result from this case to the list for each phi.
3155     for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3156       if (!ResultLists.count(I->first))
3157         PHIs.push_back(I->first);
3158       ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3159     }
3160   }
3161 
3162   // Get the resulting values for the default case.
3163   SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3164   if (!GetCaseResults(SI, SI->getDefaultDest(), &CommonDest, DefaultResultsList))
3165     return false;
3166   for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3167     PHINode *PHI = DefaultResultsList[I].first;
3168     Constant *Result = DefaultResultsList[I].second;
3169     DefaultResults[PHI] = Result;
3170     ResultTypes[PHI] = Result->getType();
3171   }
3172 
3173   APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3174   // The table density should be at lest 40%. This is the same criterion as for
3175   // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3176   // FIXME: Find the best cut-off.
3177   // Be careful to avoid overlow in the density computation.
3178   if (RangeSpread.zextOrSelf(64).ugt(UINT64_MAX / 4 - 1))
3179     return false;
3180   uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3181   if (SI->getNumCases() * 10 < TableSize * 4)
3182     return false;
3183 
3184   // Build the lookup tables.
3185   SmallDenseMap<PHINode*, GlobalVariable*> LookupTables;
3186   SmallDenseMap<PHINode*, Constant*> SingleResults;
3187 
3188   Module &Mod = *CommonDest->getParent()->getParent();
3189   for (SmallVector<PHINode*, 4>::iterator I = PHIs.begin(), E = PHIs.end();
3190        I != E; ++I) {
3191     PHINode *PHI = *I;
3192 
3193     Constant *SingleResult = NULL;
3194     LookupTables[PHI] = BuildLookupTable(Mod, TableSize, MinCaseVal,
3195                                          ResultLists[PHI], DefaultResults[PHI],
3196                                          &SingleResult);
3197     SingleResults[PHI] = SingleResult;
3198   }
3199 
3200   // Create the BB that does the lookups.
3201   BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3202                                             "switch.lookup",
3203                                             CommonDest->getParent(),
3204                                             CommonDest);
3205 
3206   // Check whether the condition value is within the case range, and branch to
3207   // the new BB.
3208   Builder.SetInsertPoint(SI);
3209   Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3210                                         "switch.tableidx");
3211   Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3212       MinCaseVal->getType(), TableSize));
3213   Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3214 
3215   // Populate the BB that does the lookups.
3216   Builder.SetInsertPoint(LookupBB);
3217   bool ReturnedEarly = false;
3218   for (SmallVector<PHINode*, 4>::iterator I = PHIs.begin(), E = PHIs.end();
3219        I != E; ++I) {
3220     PHINode *PHI = *I;
3221     // There was a single result for this phi; just use that.
3222     if (Constant *SingleResult = SingleResults[PHI]) {
3223       PHI->addIncoming(SingleResult, LookupBB);
3224       continue;
3225     }
3226 
3227     Value *GEPIndices[] = { Builder.getInt32(0), TableIndex };
3228     Value *GEP = Builder.CreateInBoundsGEP(LookupTables[PHI], GEPIndices,
3229                                            "switch.gep");
3230     Value *Result = Builder.CreateLoad(GEP, "switch.load");
3231 
3232     // If the result is only going to be used to return from the function,
3233     // we want to do that right here.
3234     if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin())) {
3235       if (CommonDest->getFirstNonPHIOrDbg() == CommonDest->getTerminator()) {
3236         Builder.CreateRet(Result);
3237         ReturnedEarly = true;
3238       }
3239     }
3240 
3241     if (!ReturnedEarly)
3242       PHI->addIncoming(Result, LookupBB);
3243   }
3244 
3245   if (!ReturnedEarly)
3246     Builder.CreateBr(CommonDest);
3247 
3248   // Remove the switch.
3249   for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3250     BasicBlock *Succ = SI->getSuccessor(i);
3251     if (Succ == SI->getDefaultDest()) continue;
3252     Succ->removePredecessor(SI->getParent());
3253   }
3254   SI->eraseFromParent();
3255 
3256   ++NumLookupTables;
3257   return true;
3258 }
3259 
SimplifySwitch(SwitchInst * SI,IRBuilder<> & Builder)3260 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3261   // If this switch is too complex to want to look at, ignore it.
3262   if (!isValueEqualityComparison(SI))
3263     return false;
3264 
3265   BasicBlock *BB = SI->getParent();
3266 
3267   // If we only have one predecessor, and if it is a branch on this value,
3268   // see if that predecessor totally determines the outcome of this switch.
3269   if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3270     if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3271       return SimplifyCFG(BB) | true;
3272 
3273   Value *Cond = SI->getCondition();
3274   if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3275     if (SimplifySwitchOnSelect(SI, Select))
3276       return SimplifyCFG(BB) | true;
3277 
3278   // If the block only contains the switch, see if we can fold the block
3279   // away into any preds.
3280   BasicBlock::iterator BBI = BB->begin();
3281   // Ignore dbg intrinsics.
3282   while (isa<DbgInfoIntrinsic>(BBI))
3283     ++BBI;
3284   if (SI == &*BBI)
3285     if (FoldValueComparisonIntoPredecessors(SI, Builder))
3286       return SimplifyCFG(BB) | true;
3287 
3288   // Try to transform the switch into an icmp and a branch.
3289   if (TurnSwitchRangeIntoICmp(SI, Builder))
3290     return SimplifyCFG(BB) | true;
3291 
3292   // Remove unreachable cases.
3293   if (EliminateDeadSwitchCases(SI))
3294     return SimplifyCFG(BB) | true;
3295 
3296   if (ForwardSwitchConditionToPHI(SI))
3297     return SimplifyCFG(BB) | true;
3298 
3299   if (SwitchToLookupTable(SI, Builder))
3300     return SimplifyCFG(BB) | true;
3301 
3302   return false;
3303 }
3304 
SimplifyIndirectBr(IndirectBrInst * IBI)3305 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3306   BasicBlock *BB = IBI->getParent();
3307   bool Changed = false;
3308 
3309   // Eliminate redundant destinations.
3310   SmallPtrSet<Value *, 8> Succs;
3311   for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3312     BasicBlock *Dest = IBI->getDestination(i);
3313     if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3314       Dest->removePredecessor(BB);
3315       IBI->removeDestination(i);
3316       --i; --e;
3317       Changed = true;
3318     }
3319   }
3320 
3321   if (IBI->getNumDestinations() == 0) {
3322     // If the indirectbr has no successors, change it to unreachable.
3323     new UnreachableInst(IBI->getContext(), IBI);
3324     EraseTerminatorInstAndDCECond(IBI);
3325     return true;
3326   }
3327 
3328   if (IBI->getNumDestinations() == 1) {
3329     // If the indirectbr has one successor, change it to a direct branch.
3330     BranchInst::Create(IBI->getDestination(0), IBI);
3331     EraseTerminatorInstAndDCECond(IBI);
3332     return true;
3333   }
3334 
3335   if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3336     if (SimplifyIndirectBrOnSelect(IBI, SI))
3337       return SimplifyCFG(BB) | true;
3338   }
3339   return Changed;
3340 }
3341 
SimplifyUncondBranch(BranchInst * BI,IRBuilder<> & Builder)3342 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3343   BasicBlock *BB = BI->getParent();
3344 
3345   // If the Terminator is the only non-phi instruction, simplify the block.
3346   BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3347   if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3348       TryToSimplifyUncondBranchFromEmptyBlock(BB))
3349     return true;
3350 
3351   // If the only instruction in the block is a seteq/setne comparison
3352   // against a constant, try to simplify the block.
3353   if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3354     if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3355       for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3356         ;
3357       if (I->isTerminator() &&
3358           TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder))
3359         return true;
3360     }
3361 
3362   // If this basic block is ONLY a compare and a branch, and if a predecessor
3363   // branches to us and our successor, fold the comparison into the
3364   // predecessor and use logical operations to update the incoming value
3365   // for PHI nodes in common successor.
3366   if (FoldBranchToCommonDest(BI))
3367     return SimplifyCFG(BB) | true;
3368   return false;
3369 }
3370 
3371 
SimplifyCondBranch(BranchInst * BI,IRBuilder<> & Builder)3372 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3373   BasicBlock *BB = BI->getParent();
3374 
3375   // Conditional branch
3376   if (isValueEqualityComparison(BI)) {
3377     // If we only have one predecessor, and if it is a branch on this value,
3378     // see if that predecessor totally determines the outcome of this
3379     // switch.
3380     if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3381       if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3382         return SimplifyCFG(BB) | true;
3383 
3384     // This block must be empty, except for the setcond inst, if it exists.
3385     // Ignore dbg intrinsics.
3386     BasicBlock::iterator I = BB->begin();
3387     // Ignore dbg intrinsics.
3388     while (isa<DbgInfoIntrinsic>(I))
3389       ++I;
3390     if (&*I == BI) {
3391       if (FoldValueComparisonIntoPredecessors(BI, Builder))
3392         return SimplifyCFG(BB) | true;
3393     } else if (&*I == cast<Instruction>(BI->getCondition())){
3394       ++I;
3395       // Ignore dbg intrinsics.
3396       while (isa<DbgInfoIntrinsic>(I))
3397         ++I;
3398       if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3399         return SimplifyCFG(BB) | true;
3400     }
3401   }
3402 
3403   // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3404   if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3405     return true;
3406 
3407   // If this basic block is ONLY a compare and a branch, and if a predecessor
3408   // branches to us and one of our successors, fold the comparison into the
3409   // predecessor and use logical operations to pick the right destination.
3410   if (FoldBranchToCommonDest(BI))
3411     return SimplifyCFG(BB) | true;
3412 
3413   // We have a conditional branch to two blocks that are only reachable
3414   // from BI.  We know that the condbr dominates the two blocks, so see if
3415   // there is any identical code in the "then" and "else" blocks.  If so, we
3416   // can hoist it up to the branching block.
3417   if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3418     if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3419       if (HoistThenElseCodeToIf(BI))
3420         return SimplifyCFG(BB) | true;
3421     } else {
3422       // If Successor #1 has multiple preds, we may be able to conditionally
3423       // execute Successor #0 if it branches to successor #1.
3424       TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3425       if (Succ0TI->getNumSuccessors() == 1 &&
3426           Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3427         if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
3428           return SimplifyCFG(BB) | true;
3429     }
3430   } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3431     // If Successor #0 has multiple preds, we may be able to conditionally
3432     // execute Successor #1 if it branches to successor #0.
3433     TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3434     if (Succ1TI->getNumSuccessors() == 1 &&
3435         Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
3436       if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
3437         return SimplifyCFG(BB) | true;
3438   }
3439 
3440   // If this is a branch on a phi node in the current block, thread control
3441   // through this block if any PHI node entries are constants.
3442   if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
3443     if (PN->getParent() == BI->getParent())
3444       if (FoldCondBranchOnPHI(BI, TD))
3445         return SimplifyCFG(BB) | true;
3446 
3447   // Scan predecessor blocks for conditional branches.
3448   for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
3449     if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
3450       if (PBI != BI && PBI->isConditional())
3451         if (SimplifyCondBranchToCondBranch(PBI, BI))
3452           return SimplifyCFG(BB) | true;
3453 
3454   return false;
3455 }
3456 
3457 /// Check if passing a value to an instruction will cause undefined behavior.
passingValueIsAlwaysUndefined(Value * V,Instruction * I)3458 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
3459   Constant *C = dyn_cast<Constant>(V);
3460   if (!C)
3461     return false;
3462 
3463   if (!I->hasOneUse()) // Only look at single-use instructions, for compile time
3464     return false;
3465 
3466   if (C->isNullValue()) {
3467     Instruction *Use = I->use_back();
3468 
3469     // Now make sure that there are no instructions in between that can alter
3470     // control flow (eg. calls)
3471     for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
3472       if (i == I->getParent()->end() || i->mayHaveSideEffects())
3473         return false;
3474 
3475     // Look through GEPs. A load from a GEP derived from NULL is still undefined
3476     if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
3477       if (GEP->getPointerOperand() == I)
3478         return passingValueIsAlwaysUndefined(V, GEP);
3479 
3480     // Look through bitcasts.
3481     if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
3482       return passingValueIsAlwaysUndefined(V, BC);
3483 
3484     // Load from null is undefined.
3485     if (LoadInst *LI = dyn_cast<LoadInst>(Use))
3486       return LI->getPointerAddressSpace() == 0;
3487 
3488     // Store to null is undefined.
3489     if (StoreInst *SI = dyn_cast<StoreInst>(Use))
3490       return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
3491   }
3492   return false;
3493 }
3494 
3495 /// If BB has an incoming value that will always trigger undefined behavior
3496 /// (eg. null pointer dereference), remove the branch leading here.
removeUndefIntroducingPredecessor(BasicBlock * BB)3497 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
3498   for (BasicBlock::iterator i = BB->begin();
3499        PHINode *PHI = dyn_cast<PHINode>(i); ++i)
3500     for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
3501       if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
3502         TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
3503         IRBuilder<> Builder(T);
3504         if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
3505           BB->removePredecessor(PHI->getIncomingBlock(i));
3506           // Turn uncoditional branches into unreachables and remove the dead
3507           // destination from conditional branches.
3508           if (BI->isUnconditional())
3509             Builder.CreateUnreachable();
3510           else
3511             Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
3512                                                          BI->getSuccessor(0));
3513           BI->eraseFromParent();
3514           return true;
3515         }
3516         // TODO: SwitchInst.
3517       }
3518 
3519   return false;
3520 }
3521 
run(BasicBlock * BB)3522 bool SimplifyCFGOpt::run(BasicBlock *BB) {
3523   bool Changed = false;
3524 
3525   assert(BB && BB->getParent() && "Block not embedded in function!");
3526   assert(BB->getTerminator() && "Degenerate basic block encountered!");
3527 
3528   // Remove basic blocks that have no predecessors (except the entry block)...
3529   // or that just have themself as a predecessor.  These are unreachable.
3530   if ((pred_begin(BB) == pred_end(BB) &&
3531        BB != &BB->getParent()->getEntryBlock()) ||
3532       BB->getSinglePredecessor() == BB) {
3533     DEBUG(dbgs() << "Removing BB: \n" << *BB);
3534     DeleteDeadBlock(BB);
3535     return true;
3536   }
3537 
3538   // Check to see if we can constant propagate this terminator instruction
3539   // away...
3540   Changed |= ConstantFoldTerminator(BB, true);
3541 
3542   // Check for and eliminate duplicate PHI nodes in this block.
3543   Changed |= EliminateDuplicatePHINodes(BB);
3544 
3545   // Check for and remove branches that will always cause undefined behavior.
3546   Changed |= removeUndefIntroducingPredecessor(BB);
3547 
3548   // Merge basic blocks into their predecessor if there is only one distinct
3549   // pred, and if there is only one distinct successor of the predecessor, and
3550   // if there are no PHI nodes.
3551   //
3552   if (MergeBlockIntoPredecessor(BB))
3553     return true;
3554 
3555   IRBuilder<> Builder(BB);
3556 
3557   // If there is a trivial two-entry PHI node in this basic block, and we can
3558   // eliminate it, do so now.
3559   if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
3560     if (PN->getNumIncomingValues() == 2)
3561       Changed |= FoldTwoEntryPHINode(PN, TD);
3562 
3563   Builder.SetInsertPoint(BB->getTerminator());
3564   if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
3565     if (BI->isUnconditional()) {
3566       if (SimplifyUncondBranch(BI, Builder)) return true;
3567     } else {
3568       if (SimplifyCondBranch(BI, Builder)) return true;
3569     }
3570   } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
3571     if (SimplifyReturn(RI, Builder)) return true;
3572   } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
3573     if (SimplifyResume(RI, Builder)) return true;
3574   } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
3575     if (SimplifySwitch(SI, Builder)) return true;
3576   } else if (UnreachableInst *UI =
3577                dyn_cast<UnreachableInst>(BB->getTerminator())) {
3578     if (SimplifyUnreachable(UI)) return true;
3579   } else if (IndirectBrInst *IBI =
3580                dyn_cast<IndirectBrInst>(BB->getTerminator())) {
3581     if (SimplifyIndirectBr(IBI)) return true;
3582   }
3583 
3584   return Changed;
3585 }
3586 
3587 /// SimplifyCFG - This function is used to do simplification of a CFG.  For
3588 /// example, it adjusts branches to branches to eliminate the extra hop, it
3589 /// eliminates unreachable basic blocks, and does other "peephole" optimization
3590 /// of the CFG.  It returns true if a modification was made.
3591 ///
SimplifyCFG(BasicBlock * BB,const TargetData * TD)3592 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) {
3593   return SimplifyCFGOpt(TD).run(BB);
3594 }
3595