1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
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"
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)"));
54 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
55 cl::desc("Duplicate return instructions into unconditional branches"));
57 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
58 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
61 /// ValueEqualityComparisonCase - Represents a case of a switch.
62 struct ValueEqualityComparisonCase {
66 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
67 : Value(Value), Dest(Dest) {}
69 bool operator<(ValueEqualityComparisonCase RHS) const {
70 // Comparing pointers is ok as we only rely on the order for uniquing.
71 return Value < RHS.Value;
75 class SimplifyCFGOpt {
76 const TargetData *const TD;
78 Value *isValueEqualityComparison(TerminatorInst *TI);
79 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
80 std::vector<ValueEqualityComparisonCase> &Cases);
81 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
83 IRBuilder<> &Builder);
84 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
85 IRBuilder<> &Builder);
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);
96 explicit SimplifyCFGOpt(const TargetData *td) : TD(td) {}
97 bool run(BasicBlock *BB);
101 /// SafeToMergeTerminators - Return true if it is safe to merge these two
102 /// terminator instructions together.
104 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
105 if (SI1 == SI2) return false; // Can't merge with self!
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));
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))
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
132 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
135 SmallVectorImpl<PHINode*> &PhiNodes) {
136 if (SI1 == SI2) return false; // Can't merge with self!
137 assert(SI1->isUnconditional() && SI2->isConditional());
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)))
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)))
163 PhiNodes.push_back(PN);
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.
172 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
173 BasicBlock *ExistPred) {
174 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
177 for (BasicBlock::iterator I = Succ->begin();
178 (PN = dyn_cast<PHINode>(I)); ++I)
179 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
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.
190 /// This does no checking to see if the true/false blocks have large or unsavory
191 /// instructions in them.
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);
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)
207 // Eliminate code duplication by ensuring that Pred1Br is conditional if
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())
217 std::swap(Pred1, Pred2);
218 std::swap(Pred1Br, Pred2Br);
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)
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) {
234 } else if (Pred1Br->getSuccessor(0) == Pred2 &&
235 Pred1Br->getSuccessor(1) == BB) {
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".
244 return Pred1Br->getCondition();
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())
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;
258 assert(BI->isConditional() && "Two successors but not conditional?");
259 if (BI->getSuccessor(0) == Pred1) {
266 return BI->getCondition();
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.
272 static unsigned ComputeSpeculationCost(const User *I) {
273 assert(isSafeToSpeculativelyExecute(I) &&
274 "Instruction is not safe to speculatively execute!");
275 switch (Operator::getOpcode(I)) {
277 // In doubt, be conservative.
279 case Instruction::GetElementPtr:
280 // GEPs are cheap if all indices are constant.
281 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
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.
299 case Instruction::Call:
300 case Instruction::Select:
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.
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.
316 /// The cost for most non-trapping instructions is defined as 1 except for
317 /// Select whose cost is 2.
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.
322 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
323 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
324 unsigned &CostRemaining) {
325 Instruction *I = dyn_cast<Instruction>(V);
327 // Non-instructions all dominate instructions, but not all constantexprs
328 // can be executed unconditionally.
329 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
334 BasicBlock *PBB = I->getParent();
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;
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)
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;
351 // If we have seen this instruction before, don't count it again.
352 if (AggressiveInsts->count(I)) return true;
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))
360 unsigned Cost = ComputeSpeculationCost(I);
362 if (Cost > CostRemaining)
365 CostRemaining -= Cost;
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))
372 // Okay, it's safe to do this! Remember this instruction.
373 AggressiveInsts->insert(I);
377 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
378 /// and PointerNullValue. Return NULL if value is not a constant int.
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())
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());
389 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
390 if (isa<ConstantPointerNull>(V))
391 return ConstantInt::get(PtrTy, 0);
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)
401 return cast<ConstantInt>
402 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
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
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;
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)) {
423 return I->getOperand(0);
426 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
429 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
431 // If this is an and/!= check then we want to optimize "x ugt 2" into
434 Span = Span.inverse();
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())
440 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
441 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
443 return I->getOperand(0);
448 // Otherwise, we can only handle an | or &, depending on isEQ.
449 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
452 unsigned NumValsBeforeLHS = Vals.size();
453 unsigned UsedICmpsBeforeLHS = UsedICmps;
454 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD,
456 unsigned NumVals = Vals.size();
457 unsigned UsedICmpsBeforeRHS = UsedICmps;
458 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
462 Vals.resize(NumVals);
463 UsedICmps = UsedICmpsBeforeRHS;
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);
473 Vals.resize(NumValsBeforeLHS);
474 UsedICmps = UsedICmpsBeforeLHS;
478 // If the LHS can't be folded in, but Extra is available and RHS can, try to
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,
486 assert(Vals.size() == NumValsBeforeLHS);
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());
504 TI->eraseFromParent();
505 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
508 /// isValueEqualityComparison - Return true if the specified terminator checks
509 /// to see if a value is equal to constant integer value.
510 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
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);
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);
533 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
534 /// decode all of the 'cases' that it represents and return the 'default' block.
535 BasicBlock *SimplifyCFGOpt::
536 GetValueEqualityComparisonCases(TerminatorInst *TI,
537 std::vector<ValueEqualityComparisonCase>
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();
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),
553 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
557 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
558 /// in the list that match the specified block.
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);
568 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
571 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
572 std::vector<ValueEqualityComparisonCase > &C2) {
573 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
575 // Make V1 be smaller than V2.
576 if (V1->size() > V2->size())
579 if (V1->size() == 0) return false;
580 if (V1->size() == 1) {
582 ConstantInt *TheVal = (*V1)[0].Value;
583 for (unsigned i = 0, e = V2->size(); i != e; ++i)
584 if (TheVal == (*V2)[i].Value)
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)
595 if ((*V1)[i1].Value < (*V2)[i2].Value)
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::
610 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
612 IRBuilder<> &Builder) {
613 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
614 if (!PredVal) return false; // Not a value comparison in predecessor.
616 Value *ThisVal = isValueEqualityComparison(TI);
617 assert(ThisVal && "This isn't a value comparison!!");
618 if (ThisVal != PredVal) return false; // Different predicates.
620 // TODO: Preserve branch weight metadata, similarly to how
621 // FoldValueComparisonIntoPredecessors preserves it.
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(),
627 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
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.
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
640 if (!ValuesOverlap(PredCases, ThisCases))
643 if (isa<BranchInst>(TI)) {
644 // Okay, one of the successors of this condbr is dead. Convert it to a
646 assert(ThisCases.size() == 1 && "Branch can only have one case!");
647 // Insert the new branch.
648 Instruction *NI = Builder.CreateBr(ThisDef);
651 // Remove PHI node entries for the dead edge.
652 ThisCases[0].Dest->removePredecessor(TI->getParent());
654 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
655 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
657 EraseTerminatorInstAndDCECond(TI);
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);
667 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
668 << "Through successor TI: " << *TI);
670 // Collect branch weights into a vector.
671 SmallVector<uint32_t, 8> Weights;
672 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
673 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
675 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
677 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
679 Weights.push_back(CI->getValue().getZExtValue());
681 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
683 if (DeadCases.count(i.getCaseValue())) {
685 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
688 i.getCaseSuccessor()->removePredecessor(TI->getParent());
693 SI->setMetadata(LLVMContext::MD_prof,
694 MDBuilder(SI->getParent()->getContext()).
695 createBranchWeights(Weights));
697 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
701 // Otherwise, TI's block must correspond to some matched value. Find out
702 // which value (or set of values) this is.
703 ConstantInt *TIV = 0;
704 BasicBlock *TIBB = TI->getParent();
705 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
706 if (PredCases[i].Dest == TIBB) {
708 return false; // Cannot handle multiple values coming to this block.
709 TIV = PredCases[i].Value;
711 assert(TIV && "No edge from pred to succ?");
713 // Okay, we found the one constant that our value can be if we get into TI's
714 // BB. Find out which successor will unconditionally be branched to.
715 BasicBlock *TheRealDest = 0;
716 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
717 if (ThisCases[i].Value == TIV) {
718 TheRealDest = ThisCases[i].Dest;
722 // If not handled by any explicit cases, it is handled by the default case.
723 if (TheRealDest == 0) TheRealDest = ThisDef;
725 // Remove PHI node entries for dead edges.
726 BasicBlock *CheckEdge = TheRealDest;
727 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
728 if (*SI != CheckEdge)
729 (*SI)->removePredecessor(TIBB);
733 // Insert the new branch.
734 Instruction *NI = Builder.CreateBr(TheRealDest);
737 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
738 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
740 EraseTerminatorInstAndDCECond(TI);
745 /// ConstantIntOrdering - This class implements a stable ordering of constant
746 /// integers that does not depend on their address. This is important for
747 /// applications that sort ConstantInt's to ensure uniqueness.
748 struct ConstantIntOrdering {
749 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
750 return LHS->getValue().ult(RHS->getValue());
755 static int ConstantIntSortPredicate(const void *P1, const void *P2) {
756 const ConstantInt *LHS = *(const ConstantInt*const*)P1;
757 const ConstantInt *RHS = *(const ConstantInt*const*)P2;
758 if (LHS->getValue().ult(RHS->getValue()))
760 if (LHS->getValue() == RHS->getValue())
765 static inline bool HasBranchWeights(const Instruction* I) {
766 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
767 if (ProfMD && ProfMD->getOperand(0))
768 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
769 return MDS->getString().equals("branch_weights");
774 /// Get Weights of a given TerminatorInst, the default weight is at the front
775 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
777 static void GetBranchWeights(TerminatorInst *TI,
778 SmallVectorImpl<uint64_t> &Weights) {
779 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
781 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
782 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i));
784 Weights.push_back(CI->getValue().getZExtValue());
787 // If TI is a conditional eq, the default case is the false case,
788 // and the corresponding branch-weight data is at index 2. We swap the
789 // default weight to be the first entry.
790 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
791 assert(Weights.size() == 2);
792 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
793 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
794 std::swap(Weights.front(), Weights.back());
798 /// Sees if any of the weights are too big for a uint32_t, and halves all the
799 /// weights if any are.
800 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
802 for (unsigned i = 0; i < Weights.size(); ++i)
803 if (Weights[i] > UINT_MAX) {
811 for (unsigned i = 0; i < Weights.size(); ++i)
815 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
816 /// equality comparison instruction (either a switch or a branch on "X == c").
817 /// See if any of the predecessors of the terminator block are value comparisons
818 /// on the same value. If so, and if safe to do so, fold them together.
819 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
820 IRBuilder<> &Builder) {
821 BasicBlock *BB = TI->getParent();
822 Value *CV = isValueEqualityComparison(TI); // CondVal
823 assert(CV && "Not a comparison?");
824 bool Changed = false;
826 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
827 while (!Preds.empty()) {
828 BasicBlock *Pred = Preds.pop_back_val();
830 // See if the predecessor is a comparison with the same value.
831 TerminatorInst *PTI = Pred->getTerminator();
832 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
834 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
835 // Figure out which 'cases' to copy from SI to PSI.
836 std::vector<ValueEqualityComparisonCase> BBCases;
837 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
839 std::vector<ValueEqualityComparisonCase> PredCases;
840 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
842 // Based on whether the default edge from PTI goes to BB or not, fill in
843 // PredCases and PredDefault with the new switch cases we would like to
845 SmallVector<BasicBlock*, 8> NewSuccessors;
847 // Update the branch weight metadata along the way
848 SmallVector<uint64_t, 8> Weights;
849 bool PredHasWeights = HasBranchWeights(PTI);
850 bool SuccHasWeights = HasBranchWeights(TI);
852 if (PredHasWeights) {
853 GetBranchWeights(PTI, Weights);
854 // branch-weight metadata is inconsistant here.
855 if (Weights.size() != 1 + PredCases.size())
856 PredHasWeights = SuccHasWeights = false;
857 } else if (SuccHasWeights)
858 // If there are no predecessor weights but there are successor weights,
859 // populate Weights with 1, which will later be scaled to the sum of
860 // successor's weights
861 Weights.assign(1 + PredCases.size(), 1);
863 SmallVector<uint64_t, 8> SuccWeights;
864 if (SuccHasWeights) {
865 GetBranchWeights(TI, SuccWeights);
866 // branch-weight metadata is inconsistant here.
867 if (SuccWeights.size() != 1 + BBCases.size())
868 PredHasWeights = SuccHasWeights = false;
869 } else if (PredHasWeights)
870 SuccWeights.assign(1 + BBCases.size(), 1);
872 if (PredDefault == BB) {
873 // If this is the default destination from PTI, only the edges in TI
874 // that don't occur in PTI, or that branch to BB will be activated.
875 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
876 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
877 if (PredCases[i].Dest != BB)
878 PTIHandled.insert(PredCases[i].Value);
880 // The default destination is BB, we don't need explicit targets.
881 std::swap(PredCases[i], PredCases.back());
883 if (PredHasWeights || SuccHasWeights) {
884 // Increase weight for the default case.
885 Weights[0] += Weights[i+1];
886 std::swap(Weights[i+1], Weights.back());
890 PredCases.pop_back();
894 // Reconstruct the new switch statement we will be building.
895 if (PredDefault != BBDefault) {
896 PredDefault->removePredecessor(Pred);
897 PredDefault = BBDefault;
898 NewSuccessors.push_back(BBDefault);
901 unsigned CasesFromPred = Weights.size();
902 uint64_t ValidTotalSuccWeight = 0;
903 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
904 if (!PTIHandled.count(BBCases[i].Value) &&
905 BBCases[i].Dest != BBDefault) {
906 PredCases.push_back(BBCases[i]);
907 NewSuccessors.push_back(BBCases[i].Dest);
908 if (SuccHasWeights || PredHasWeights) {
909 // The default weight is at index 0, so weight for the ith case
910 // should be at index i+1. Scale the cases from successor by
911 // PredDefaultWeight (Weights[0]).
912 Weights.push_back(Weights[0] * SuccWeights[i+1]);
913 ValidTotalSuccWeight += SuccWeights[i+1];
917 if (SuccHasWeights || PredHasWeights) {
918 ValidTotalSuccWeight += SuccWeights[0];
919 // Scale the cases from predecessor by ValidTotalSuccWeight.
920 for (unsigned i = 1; i < CasesFromPred; ++i)
921 Weights[i] *= ValidTotalSuccWeight;
922 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
923 Weights[0] *= SuccWeights[0];
926 // If this is not the default destination from PSI, only the edges
927 // in SI that occur in PSI with a destination of BB will be
929 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
930 std::map<ConstantInt*, uint64_t> WeightsForHandled;
931 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
932 if (PredCases[i].Dest == BB) {
933 PTIHandled.insert(PredCases[i].Value);
935 if (PredHasWeights || SuccHasWeights) {
936 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
937 std::swap(Weights[i+1], Weights.back());
941 std::swap(PredCases[i], PredCases.back());
942 PredCases.pop_back();
946 // Okay, now we know which constants were sent to BB from the
947 // predecessor. Figure out where they will all go now.
948 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
949 if (PTIHandled.count(BBCases[i].Value)) {
950 // If this is one we are capable of getting...
951 if (PredHasWeights || SuccHasWeights)
952 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
953 PredCases.push_back(BBCases[i]);
954 NewSuccessors.push_back(BBCases[i].Dest);
955 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
958 // If there are any constants vectored to BB that TI doesn't handle,
959 // they must go to the default destination of TI.
960 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
962 E = PTIHandled.end(); I != E; ++I) {
963 if (PredHasWeights || SuccHasWeights)
964 Weights.push_back(WeightsForHandled[*I]);
965 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
966 NewSuccessors.push_back(BBDefault);
970 // Okay, at this point, we know which new successor Pred will get. Make
971 // sure we update the number of entries in the PHI nodes for these
973 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
974 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
976 Builder.SetInsertPoint(PTI);
977 // Convert pointer to int before we switch.
978 if (CV->getType()->isPointerTy()) {
979 assert(TD && "Cannot switch on pointer without TargetData");
980 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()),
984 // Now that the successors are updated, create the new Switch instruction.
985 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
987 NewSI->setDebugLoc(PTI->getDebugLoc());
988 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
989 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
991 if (PredHasWeights || SuccHasWeights) {
992 // Halve the weights if any of them cannot fit in an uint32_t
995 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
997 NewSI->setMetadata(LLVMContext::MD_prof,
998 MDBuilder(BB->getContext()).
999 createBranchWeights(MDWeights));
1002 EraseTerminatorInstAndDCECond(PTI);
1004 // Okay, last check. If BB is still a successor of PSI, then we must
1005 // have an infinite loop case. If so, add an infinitely looping block
1006 // to handle the case to preserve the behavior of the code.
1007 BasicBlock *InfLoopBlock = 0;
1008 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1009 if (NewSI->getSuccessor(i) == BB) {
1010 if (InfLoopBlock == 0) {
1011 // Insert it at the end of the function, because it's either code,
1012 // or it won't matter if it's hot. :)
1013 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1014 "infloop", BB->getParent());
1015 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1017 NewSI->setSuccessor(i, InfLoopBlock);
1026 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1027 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1028 // would need to do this), we can't hoist the invoke, as there is nowhere
1029 // to put the select in this case.
1030 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1031 Instruction *I1, Instruction *I2) {
1032 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1034 for (BasicBlock::iterator BBI = SI->begin();
1035 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1036 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1037 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1038 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1046 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1047 /// BB2, hoist any common code in the two blocks up into the branch block. The
1048 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1049 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1050 // This does very trivial matching, with limited scanning, to find identical
1051 // instructions in the two blocks. In particular, we don't want to get into
1052 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1053 // such, we currently just scan for obviously identical instructions in an
1055 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1056 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1058 BasicBlock::iterator BB1_Itr = BB1->begin();
1059 BasicBlock::iterator BB2_Itr = BB2->begin();
1061 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1062 // Skip debug info if it is not identical.
1063 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1064 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1065 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1066 while (isa<DbgInfoIntrinsic>(I1))
1068 while (isa<DbgInfoIntrinsic>(I2))
1071 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1072 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1075 // If we get here, we can hoist at least one instruction.
1076 BasicBlock *BIParent = BI->getParent();
1079 // If we are hoisting the terminator instruction, don't move one (making a
1080 // broken BB), instead clone it, and remove BI.
1081 if (isa<TerminatorInst>(I1))
1082 goto HoistTerminator;
1084 // For a normal instruction, we just move one to right before the branch,
1085 // then replace all uses of the other with the first. Finally, we remove
1086 // the now redundant second instruction.
1087 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1088 if (!I2->use_empty())
1089 I2->replaceAllUsesWith(I1);
1090 I1->intersectOptionalDataWith(I2);
1091 I2->eraseFromParent();
1095 // Skip debug info if it is not identical.
1096 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1097 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1098 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1099 while (isa<DbgInfoIntrinsic>(I1))
1101 while (isa<DbgInfoIntrinsic>(I2))
1104 } while (I1->isIdenticalToWhenDefined(I2));
1109 // It may not be possible to hoist an invoke.
1110 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1113 // Okay, it is safe to hoist the terminator.
1114 Instruction *NT = I1->clone();
1115 BIParent->getInstList().insert(BI, NT);
1116 if (!NT->getType()->isVoidTy()) {
1117 I1->replaceAllUsesWith(NT);
1118 I2->replaceAllUsesWith(NT);
1122 IRBuilder<true, NoFolder> Builder(NT);
1123 // Hoisting one of the terminators from our successor is a great thing.
1124 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1125 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1126 // nodes, so we insert select instruction to compute the final result.
1127 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1128 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1130 for (BasicBlock::iterator BBI = SI->begin();
1131 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1132 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1133 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1134 if (BB1V == BB2V) continue;
1136 // These values do not agree. Insert a select instruction before NT
1137 // that determines the right value.
1138 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1140 SI = cast<SelectInst>
1141 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1142 BB1V->getName()+"."+BB2V->getName()));
1144 // Make the PHI node use the select for all incoming values for BB1/BB2
1145 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1146 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1147 PN->setIncomingValue(i, SI);
1151 // Update any PHI nodes in our new successors.
1152 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1153 AddPredecessorToBlock(*SI, BIParent, BB1);
1155 EraseTerminatorInstAndDCECond(BI);
1159 /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1
1160 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code
1161 /// (for now, restricted to a single instruction that's side effect free) from
1162 /// the BB1 into the branch block to speculatively execute it.
1167 /// br i1 %t1, label %BB1, label %BB2
1169 /// %t3 = add %t2, c
1175 /// %t4 = add %t2, c
1176 /// %t3 = select i1 %t1, %t2, %t3
1177 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) {
1178 // Only speculatively execution a single instruction (not counting the
1179 // terminator) for now.
1180 Instruction *HInst = NULL;
1181 Instruction *Term = BB1->getTerminator();
1182 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end();
1183 BBI != BBE; ++BBI) {
1184 Instruction *I = BBI;
1186 if (isa<DbgInfoIntrinsic>(I)) continue;
1187 if (I == Term) break;
1194 BasicBlock *BIParent = BI->getParent();
1196 // Check the instruction to be hoisted, if there is one.
1198 // Don't hoist the instruction if it's unsafe or expensive.
1199 if (!isSafeToSpeculativelyExecute(HInst))
1201 if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold)
1204 // Do not hoist the instruction if any of its operands are defined but not
1205 // used in this BB. The transformation will prevent the operand from
1206 // being sunk into the use block.
1207 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end();
1209 Instruction *OpI = dyn_cast<Instruction>(*i);
1210 if (OpI && OpI->getParent() == BIParent &&
1211 !OpI->mayHaveSideEffects() &&
1212 !OpI->isUsedInBasicBlock(BIParent))
1217 // Be conservative for now. FP select instruction can often be expensive.
1218 Value *BrCond = BI->getCondition();
1219 if (isa<FCmpInst>(BrCond))
1222 // If BB1 is actually on the false edge of the conditional branch, remember
1223 // to swap the select operands later.
1224 bool Invert = false;
1225 if (BB1 != BI->getSuccessor(0)) {
1226 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?");
1230 // Collect interesting PHIs, and scan for hazards.
1231 SmallSetVector<std::pair<Value *, Value *>, 4> PHIs;
1232 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0);
1233 for (BasicBlock::iterator I = BB2->begin();
1234 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1235 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1236 Value *BIParentV = PN->getIncomingValueForBlock(BIParent);
1238 // Skip PHIs which are trivial.
1239 if (BB1V == BIParentV)
1242 // Check for saftey.
1243 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) {
1244 // An unfolded ConstantExpr could end up getting expanded into
1245 // Instructions. Don't speculate this and another instruction at
1249 if (!isSafeToSpeculativelyExecute(CE))
1251 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold)
1255 // Ok, we may insert a select for this PHI.
1256 PHIs.insert(std::make_pair(BB1V, BIParentV));
1259 // If there are no PHIs to process, bail early. This helps ensure idempotence
1264 // If we get here, we can hoist the instruction and if-convert.
1265 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";);
1267 // Hoist the instruction.
1269 BIParent->getInstList().splice(BI, BB1->getInstList(), HInst);
1271 // Insert selects and rewrite the PHI operands.
1272 IRBuilder<true, NoFolder> Builder(BI);
1273 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
1274 Value *TrueV = PHIs[i].first;
1275 Value *FalseV = PHIs[i].second;
1277 // Create a select whose true value is the speculatively executed value and
1278 // false value is the previously determined FalseV.
1281 SI = cast<SelectInst>
1282 (Builder.CreateSelect(BrCond, FalseV, TrueV,
1283 FalseV->getName() + "." + TrueV->getName()));
1285 SI = cast<SelectInst>
1286 (Builder.CreateSelect(BrCond, TrueV, FalseV,
1287 TrueV->getName() + "." + FalseV->getName()));
1289 // Make the PHI node use the select for all incoming values for "then" and
1291 for (BasicBlock::iterator I = BB2->begin();
1292 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1293 unsigned BB1I = PN->getBasicBlockIndex(BB1);
1294 unsigned BIParentI = PN->getBasicBlockIndex(BIParent);
1295 Value *BB1V = PN->getIncomingValue(BB1I);
1296 Value *BIParentV = PN->getIncomingValue(BIParentI);
1297 if (TrueV == BB1V && FalseV == BIParentV) {
1298 PN->setIncomingValue(BB1I, SI);
1299 PN->setIncomingValue(BIParentI, SI);
1308 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1309 /// across this block.
1310 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1311 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1314 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1315 if (isa<DbgInfoIntrinsic>(BBI))
1317 if (Size > 10) return false; // Don't clone large BB's.
1320 // We can only support instructions that do not define values that are
1321 // live outside of the current basic block.
1322 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1324 Instruction *U = cast<Instruction>(*UI);
1325 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1328 // Looks ok, continue checking.
1334 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1335 /// that is defined in the same block as the branch and if any PHI entries are
1336 /// constants, thread edges corresponding to that entry to be branches to their
1337 /// ultimate destination.
1338 static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) {
1339 BasicBlock *BB = BI->getParent();
1340 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1341 // NOTE: we currently cannot transform this case if the PHI node is used
1342 // outside of the block.
1343 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1346 // Degenerate case of a single entry PHI.
1347 if (PN->getNumIncomingValues() == 1) {
1348 FoldSingleEntryPHINodes(PN->getParent());
1352 // Now we know that this block has multiple preds and two succs.
1353 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1355 // Okay, this is a simple enough basic block. See if any phi values are
1357 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1358 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1359 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1361 // Okay, we now know that all edges from PredBB should be revectored to
1362 // branch to RealDest.
1363 BasicBlock *PredBB = PN->getIncomingBlock(i);
1364 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1366 if (RealDest == BB) continue; // Skip self loops.
1367 // Skip if the predecessor's terminator is an indirect branch.
1368 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1370 // The dest block might have PHI nodes, other predecessors and other
1371 // difficult cases. Instead of being smart about this, just insert a new
1372 // block that jumps to the destination block, effectively splitting
1373 // the edge we are about to create.
1374 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1375 RealDest->getName()+".critedge",
1376 RealDest->getParent(), RealDest);
1377 BranchInst::Create(RealDest, EdgeBB);
1379 // Update PHI nodes.
1380 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1382 // BB may have instructions that are being threaded over. Clone these
1383 // instructions into EdgeBB. We know that there will be no uses of the
1384 // cloned instructions outside of EdgeBB.
1385 BasicBlock::iterator InsertPt = EdgeBB->begin();
1386 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1387 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1388 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1389 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1392 // Clone the instruction.
1393 Instruction *N = BBI->clone();
1394 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1396 // Update operands due to translation.
1397 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1399 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1400 if (PI != TranslateMap.end())
1404 // Check for trivial simplification.
1405 if (Value *V = SimplifyInstruction(N, TD)) {
1406 TranslateMap[BBI] = V;
1407 delete N; // Instruction folded away, don't need actual inst
1409 // Insert the new instruction into its new home.
1410 EdgeBB->getInstList().insert(InsertPt, N);
1411 if (!BBI->use_empty())
1412 TranslateMap[BBI] = N;
1416 // Loop over all of the edges from PredBB to BB, changing them to branch
1417 // to EdgeBB instead.
1418 TerminatorInst *PredBBTI = PredBB->getTerminator();
1419 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1420 if (PredBBTI->getSuccessor(i) == BB) {
1421 BB->removePredecessor(PredBB);
1422 PredBBTI->setSuccessor(i, EdgeBB);
1425 // Recurse, simplifying any other constants.
1426 return FoldCondBranchOnPHI(BI, TD) | true;
1432 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1433 /// PHI node, see if we can eliminate it.
1434 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) {
1435 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1436 // statement", which has a very simple dominance structure. Basically, we
1437 // are trying to find the condition that is being branched on, which
1438 // subsequently causes this merge to happen. We really want control
1439 // dependence information for this check, but simplifycfg can't keep it up
1440 // to date, and this catches most of the cases we care about anyway.
1441 BasicBlock *BB = PN->getParent();
1442 BasicBlock *IfTrue, *IfFalse;
1443 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1445 // Don't bother if the branch will be constant folded trivially.
1446 isa<ConstantInt>(IfCond))
1449 // Okay, we found that we can merge this two-entry phi node into a select.
1450 // Doing so would require us to fold *all* two entry phi nodes in this block.
1451 // At some point this becomes non-profitable (particularly if the target
1452 // doesn't support cmov's). Only do this transformation if there are two or
1453 // fewer PHI nodes in this block.
1454 unsigned NumPhis = 0;
1455 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1459 // Loop over the PHI's seeing if we can promote them all to select
1460 // instructions. While we are at it, keep track of the instructions
1461 // that need to be moved to the dominating block.
1462 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1463 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1464 MaxCostVal1 = PHINodeFoldingThreshold;
1466 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1467 PHINode *PN = cast<PHINode>(II++);
1468 if (Value *V = SimplifyInstruction(PN, TD)) {
1469 PN->replaceAllUsesWith(V);
1470 PN->eraseFromParent();
1474 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1476 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1481 // If we folded the first phi, PN dangles at this point. Refresh it. If
1482 // we ran out of PHIs then we simplified them all.
1483 PN = dyn_cast<PHINode>(BB->begin());
1484 if (PN == 0) return true;
1486 // Don't fold i1 branches on PHIs which contain binary operators. These can
1487 // often be turned into switches and other things.
1488 if (PN->getType()->isIntegerTy(1) &&
1489 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1490 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1491 isa<BinaryOperator>(IfCond)))
1494 // If we all PHI nodes are promotable, check to make sure that all
1495 // instructions in the predecessor blocks can be promoted as well. If
1496 // not, we won't be able to get rid of the control flow, so it's not
1497 // worth promoting to select instructions.
1498 BasicBlock *DomBlock = 0;
1499 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1500 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1501 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1504 DomBlock = *pred_begin(IfBlock1);
1505 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1506 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1507 // This is not an aggressive instruction that we can promote.
1508 // Because of this, we won't be able to get rid of the control
1509 // flow, so the xform is not worth it.
1514 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1517 DomBlock = *pred_begin(IfBlock2);
1518 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1519 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1520 // This is not an aggressive instruction that we can promote.
1521 // Because of this, we won't be able to get rid of the control
1522 // flow, so the xform is not worth it.
1527 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1528 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1530 // If we can still promote the PHI nodes after this gauntlet of tests,
1531 // do all of the PHI's now.
1532 Instruction *InsertPt = DomBlock->getTerminator();
1533 IRBuilder<true, NoFolder> Builder(InsertPt);
1535 // Move all 'aggressive' instructions, which are defined in the
1536 // conditional parts of the if's up to the dominating block.
1538 DomBlock->getInstList().splice(InsertPt,
1539 IfBlock1->getInstList(), IfBlock1->begin(),
1540 IfBlock1->getTerminator());
1542 DomBlock->getInstList().splice(InsertPt,
1543 IfBlock2->getInstList(), IfBlock2->begin(),
1544 IfBlock2->getTerminator());
1546 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1547 // Change the PHI node into a select instruction.
1548 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1549 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1552 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1553 PN->replaceAllUsesWith(NV);
1555 PN->eraseFromParent();
1558 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1559 // has been flattened. Change DomBlock to jump directly to our new block to
1560 // avoid other simplifycfg's kicking in on the diamond.
1561 TerminatorInst *OldTI = DomBlock->getTerminator();
1562 Builder.SetInsertPoint(OldTI);
1563 Builder.CreateBr(BB);
1564 OldTI->eraseFromParent();
1568 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1569 /// to two returning blocks, try to merge them together into one return,
1570 /// introducing a select if the return values disagree.
1571 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1572 IRBuilder<> &Builder) {
1573 assert(BI->isConditional() && "Must be a conditional branch");
1574 BasicBlock *TrueSucc = BI->getSuccessor(0);
1575 BasicBlock *FalseSucc = BI->getSuccessor(1);
1576 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1577 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1579 // Check to ensure both blocks are empty (just a return) or optionally empty
1580 // with PHI nodes. If there are other instructions, merging would cause extra
1581 // computation on one path or the other.
1582 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1584 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1587 Builder.SetInsertPoint(BI);
1588 // Okay, we found a branch that is going to two return nodes. If
1589 // there is no return value for this function, just change the
1590 // branch into a return.
1591 if (FalseRet->getNumOperands() == 0) {
1592 TrueSucc->removePredecessor(BI->getParent());
1593 FalseSucc->removePredecessor(BI->getParent());
1594 Builder.CreateRetVoid();
1595 EraseTerminatorInstAndDCECond(BI);
1599 // Otherwise, figure out what the true and false return values are
1600 // so we can insert a new select instruction.
1601 Value *TrueValue = TrueRet->getReturnValue();
1602 Value *FalseValue = FalseRet->getReturnValue();
1604 // Unwrap any PHI nodes in the return blocks.
1605 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1606 if (TVPN->getParent() == TrueSucc)
1607 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1608 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1609 if (FVPN->getParent() == FalseSucc)
1610 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1612 // In order for this transformation to be safe, we must be able to
1613 // unconditionally execute both operands to the return. This is
1614 // normally the case, but we could have a potentially-trapping
1615 // constant expression that prevents this transformation from being
1617 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1620 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1624 // Okay, we collected all the mapped values and checked them for sanity, and
1625 // defined to really do this transformation. First, update the CFG.
1626 TrueSucc->removePredecessor(BI->getParent());
1627 FalseSucc->removePredecessor(BI->getParent());
1629 // Insert select instructions where needed.
1630 Value *BrCond = BI->getCondition();
1632 // Insert a select if the results differ.
1633 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1634 } else if (isa<UndefValue>(TrueValue)) {
1635 TrueValue = FalseValue;
1637 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1638 FalseValue, "retval");
1642 Value *RI = !TrueValue ?
1643 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1647 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1648 << "\n " << *BI << "NewRet = " << *RI
1649 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1651 EraseTerminatorInstAndDCECond(BI);
1656 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1657 /// probabilities of the branch taking each edge. Fills in the two APInt
1658 /// parameters and return true, or returns false if no or invalid metadata was
1660 static bool ExtractBranchMetadata(BranchInst *BI,
1661 APInt &ProbTrue, APInt &ProbFalse) {
1662 assert(BI->isConditional() &&
1663 "Looking for probabilities on unconditional branch?");
1664 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1665 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1666 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1667 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1668 if (!CITrue || !CIFalse) return false;
1669 ProbTrue = CITrue->getValue();
1670 ProbFalse = CIFalse->getValue();
1671 assert(ProbTrue.getBitWidth() == 32 && ProbFalse.getBitWidth() == 32 &&
1672 "Branch probability metadata must be 32-bit integers");
1676 /// MultiplyAndLosePrecision - Multiplies A and B, then returns the result. In
1677 /// the event of overflow, logically-shifts all four inputs right until the
1679 static APInt MultiplyAndLosePrecision(APInt &A, APInt &B, APInt &C, APInt &D,
1680 unsigned &BitsLost) {
1682 bool Overflow = false;
1683 APInt Result = A.umul_ov(B, Overflow);
1685 APInt MaxB = APInt::getMaxValue(A.getBitWidth()).udiv(A);
1689 } while (B.ugt(MaxB));
1690 A = A.lshr(BitsLost);
1691 C = C.lshr(BitsLost);
1692 D = D.lshr(BitsLost);
1698 /// checkCSEInPredecessor - Return true if the given instruction is available
1699 /// in its predecessor block. If yes, the instruction will be removed.
1701 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1702 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1704 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1705 Instruction *PBI = &*I;
1706 // Check whether Inst and PBI generate the same value.
1707 if (Inst->isIdenticalTo(PBI)) {
1708 Inst->replaceAllUsesWith(PBI);
1709 Inst->eraseFromParent();
1716 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1717 /// predecessor branches to us and one of our successors, fold the block into
1718 /// the predecessor and use logical operations to pick the right destination.
1719 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1720 BasicBlock *BB = BI->getParent();
1722 Instruction *Cond = 0;
1723 if (BI->isConditional())
1724 Cond = dyn_cast<Instruction>(BI->getCondition());
1726 // For unconditional branch, check for a simple CFG pattern, where
1727 // BB has a single predecessor and BB's successor is also its predecessor's
1728 // successor. If such pattern exisits, check for CSE between BB and its
1730 if (BasicBlock *PB = BB->getSinglePredecessor())
1731 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1732 if (PBI->isConditional() &&
1733 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1734 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1735 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1737 Instruction *Curr = I++;
1738 if (isa<CmpInst>(Curr)) {
1742 // Quit if we can't remove this instruction.
1743 if (!checkCSEInPredecessor(Curr, PB))
1752 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1753 Cond->getParent() != BB || !Cond->hasOneUse())
1756 // Only allow this if the condition is a simple instruction that can be
1757 // executed unconditionally. It must be in the same block as the branch, and
1758 // must be at the front of the block.
1759 BasicBlock::iterator FrontIt = BB->front();
1761 // Ignore dbg intrinsics.
1762 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1764 // Allow a single instruction to be hoisted in addition to the compare
1765 // that feeds the branch. We later ensure that any values that _it_ uses
1766 // were also live in the predecessor, so that we don't unnecessarily create
1767 // register pressure or inhibit out-of-order execution.
1768 Instruction *BonusInst = 0;
1769 if (&*FrontIt != Cond &&
1770 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
1771 isSafeToSpeculativelyExecute(FrontIt)) {
1772 BonusInst = &*FrontIt;
1775 // Ignore dbg intrinsics.
1776 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1779 // Only a single bonus inst is allowed.
1780 if (&*FrontIt != Cond)
1783 // Make sure the instruction after the condition is the cond branch.
1784 BasicBlock::iterator CondIt = Cond; ++CondIt;
1786 // Ingore dbg intrinsics.
1787 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
1792 // Cond is known to be a compare or binary operator. Check to make sure that
1793 // neither operand is a potentially-trapping constant expression.
1794 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
1797 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
1801 // Finally, don't infinitely unroll conditional loops.
1802 BasicBlock *TrueDest = BI->getSuccessor(0);
1803 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
1804 if (TrueDest == BB || FalseDest == BB)
1807 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1808 BasicBlock *PredBlock = *PI;
1809 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
1811 // Check that we have two conditional branches. If there is a PHI node in
1812 // the common successor, verify that the same value flows in from both
1814 SmallVector<PHINode*, 4> PHIs;
1815 if (PBI == 0 || PBI->isUnconditional() ||
1816 (BI->isConditional() &&
1817 !SafeToMergeTerminators(BI, PBI)) ||
1818 (!BI->isConditional() &&
1819 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
1822 // Determine if the two branches share a common destination.
1823 Instruction::BinaryOps Opc;
1824 bool InvertPredCond = false;
1826 if (BI->isConditional()) {
1827 if (PBI->getSuccessor(0) == TrueDest)
1828 Opc = Instruction::Or;
1829 else if (PBI->getSuccessor(1) == FalseDest)
1830 Opc = Instruction::And;
1831 else if (PBI->getSuccessor(0) == FalseDest)
1832 Opc = Instruction::And, InvertPredCond = true;
1833 else if (PBI->getSuccessor(1) == TrueDest)
1834 Opc = Instruction::Or, InvertPredCond = true;
1838 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
1842 // Ensure that any values used in the bonus instruction are also used
1843 // by the terminator of the predecessor. This means that those values
1844 // must already have been resolved, so we won't be inhibiting the
1845 // out-of-order core by speculating them earlier.
1847 // Collect the values used by the bonus inst
1848 SmallPtrSet<Value*, 4> UsedValues;
1849 for (Instruction::op_iterator OI = BonusInst->op_begin(),
1850 OE = BonusInst->op_end(); OI != OE; ++OI) {
1852 if (!isa<Constant>(V))
1853 UsedValues.insert(V);
1856 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
1857 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
1859 // Walk up to four levels back up the use-def chain of the predecessor's
1860 // terminator to see if all those values were used. The choice of four
1861 // levels is arbitrary, to provide a compile-time-cost bound.
1862 while (!Worklist.empty()) {
1863 std::pair<Value*, unsigned> Pair = Worklist.back();
1864 Worklist.pop_back();
1866 if (Pair.second >= 4) continue;
1867 UsedValues.erase(Pair.first);
1868 if (UsedValues.empty()) break;
1870 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
1871 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
1873 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
1877 if (!UsedValues.empty()) return false;
1880 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
1881 IRBuilder<> Builder(PBI);
1883 // If we need to invert the condition in the pred block to match, do so now.
1884 if (InvertPredCond) {
1885 Value *NewCond = PBI->getCondition();
1887 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
1888 CmpInst *CI = cast<CmpInst>(NewCond);
1889 CI->setPredicate(CI->getInversePredicate());
1891 NewCond = Builder.CreateNot(NewCond,
1892 PBI->getCondition()->getName()+".not");
1895 PBI->setCondition(NewCond);
1896 PBI->swapSuccessors();
1899 // If we have a bonus inst, clone it into the predecessor block.
1900 Instruction *NewBonus = 0;
1902 NewBonus = BonusInst->clone();
1903 PredBlock->getInstList().insert(PBI, NewBonus);
1904 NewBonus->takeName(BonusInst);
1905 BonusInst->setName(BonusInst->getName()+".old");
1908 // Clone Cond into the predecessor basic block, and or/and the
1909 // two conditions together.
1910 Instruction *New = Cond->clone();
1911 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
1912 PredBlock->getInstList().insert(PBI, New);
1913 New->takeName(Cond);
1914 Cond->setName(New->getName()+".old");
1916 if (BI->isConditional()) {
1917 Instruction *NewCond =
1918 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
1920 PBI->setCondition(NewCond);
1922 if (PBI->getSuccessor(0) == BB) {
1923 AddPredecessorToBlock(TrueDest, PredBlock, BB);
1924 PBI->setSuccessor(0, TrueDest);
1926 if (PBI->getSuccessor(1) == BB) {
1927 AddPredecessorToBlock(FalseDest, PredBlock, BB);
1928 PBI->setSuccessor(1, FalseDest);
1931 // Update PHI nodes in the common successors.
1932 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
1933 ConstantInt *PBI_C = cast<ConstantInt>(
1934 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
1935 assert(PBI_C->getType()->isIntegerTy(1));
1936 Instruction *MergedCond = 0;
1937 if (PBI->getSuccessor(0) == TrueDest) {
1938 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
1939 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
1940 // is false: !PBI_Cond and BI_Value
1941 Instruction *NotCond =
1942 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
1945 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
1950 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
1951 PBI->getCondition(), MergedCond,
1954 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
1955 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
1956 // is false: PBI_Cond and BI_Value
1958 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
1959 PBI->getCondition(), New,
1961 if (PBI_C->isOne()) {
1962 Instruction *NotCond =
1963 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
1966 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
1967 NotCond, MergedCond,
1972 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
1975 // Change PBI from Conditional to Unconditional.
1976 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
1977 EraseTerminatorInstAndDCECond(PBI);
1981 // TODO: If BB is reachable from all paths through PredBlock, then we
1982 // could replace PBI's branch probabilities with BI's.
1984 // Merge probability data into PredBlock's branch.
1986 if (PBI->isConditional() && BI->isConditional() &&
1987 ExtractBranchMetadata(PBI, C, D) && ExtractBranchMetadata(BI, A, B)) {
1988 // Given IR which does:
1990 // br i1 %x, label %bbB, label %bbC
1992 // br i1 %y, label %bbD, label %bbC
1993 // Let's call the probability that we take the edge from %bbA to %bbB
1994 // 'a', from %bbA to %bbC, 'b', from %bbB to %bbD 'c' and from %bbB to
1995 // %bbC probability 'd'.
1997 // We transform the IR into:
1999 // br i1 %z, label %bbD, label %bbC
2000 // where the probability of going to %bbD is (a*c) and going to bbC is
2003 // Probabilities aren't stored as ratios directly. Using branch weights,
2005 // (a*c)% = A*C, (b+(a*d))% = A*D+B*C+B*D.
2007 // In the event of overflow, we want to drop the LSB of the input
2011 // Ignore overflow result on ProbTrue.
2012 APInt ProbTrue = MultiplyAndLosePrecision(A, C, B, D, BitsLost);
2014 APInt Tmp1 = MultiplyAndLosePrecision(B, D, A, C, BitsLost);
2016 ProbTrue = ProbTrue.lshr(BitsLost*2);
2019 APInt Tmp2 = MultiplyAndLosePrecision(A, D, C, B, BitsLost);
2021 ProbTrue = ProbTrue.lshr(BitsLost*2);
2022 Tmp1 = Tmp1.lshr(BitsLost*2);
2025 APInt Tmp3 = MultiplyAndLosePrecision(B, C, A, D, BitsLost);
2027 ProbTrue = ProbTrue.lshr(BitsLost*2);
2028 Tmp1 = Tmp1.lshr(BitsLost*2);
2029 Tmp2 = Tmp2.lshr(BitsLost*2);
2032 bool Overflow1 = false, Overflow2 = false;
2033 APInt Tmp4 = Tmp2.uadd_ov(Tmp3, Overflow1);
2034 APInt ProbFalse = Tmp4.uadd_ov(Tmp1, Overflow2);
2036 if (Overflow1 || Overflow2) {
2037 ProbTrue = ProbTrue.lshr(1);
2038 Tmp1 = Tmp1.lshr(1);
2039 Tmp2 = Tmp2.lshr(1);
2040 Tmp3 = Tmp3.lshr(1);
2042 ProbFalse = Tmp4 + Tmp1;
2045 // The sum of branch weights must fit in 32-bits.
2046 if (ProbTrue.isNegative() && ProbFalse.isNegative()) {
2047 ProbTrue = ProbTrue.lshr(1);
2048 ProbFalse = ProbFalse.lshr(1);
2051 if (ProbTrue != ProbFalse) {
2052 // Normalize the result.
2053 APInt GCD = APIntOps::GreatestCommonDivisor(ProbTrue, ProbFalse);
2054 ProbTrue = ProbTrue.udiv(GCD);
2055 ProbFalse = ProbFalse.udiv(GCD);
2057 MDBuilder MDB(BI->getContext());
2058 MDNode *N = MDB.createBranchWeights(ProbTrue.getZExtValue(),
2059 ProbFalse.getZExtValue());
2060 PBI->setMetadata(LLVMContext::MD_prof, N);
2062 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2065 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2068 // Copy any debug value intrinsics into the end of PredBlock.
2069 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2070 if (isa<DbgInfoIntrinsic>(*I))
2071 I->clone()->insertBefore(PBI);
2078 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2079 /// predecessor of another block, this function tries to simplify it. We know
2080 /// that PBI and BI are both conditional branches, and BI is in one of the
2081 /// successor blocks of PBI - PBI branches to BI.
2082 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2083 assert(PBI->isConditional() && BI->isConditional());
2084 BasicBlock *BB = BI->getParent();
2086 // If this block ends with a branch instruction, and if there is a
2087 // predecessor that ends on a branch of the same condition, make
2088 // this conditional branch redundant.
2089 if (PBI->getCondition() == BI->getCondition() &&
2090 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2091 // Okay, the outcome of this conditional branch is statically
2092 // knowable. If this block had a single pred, handle specially.
2093 if (BB->getSinglePredecessor()) {
2094 // Turn this into a branch on constant.
2095 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2096 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2098 return true; // Nuke the branch on constant.
2101 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2102 // in the constant and simplify the block result. Subsequent passes of
2103 // simplifycfg will thread the block.
2104 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2105 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2106 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2107 std::distance(PB, PE),
2108 BI->getCondition()->getName() + ".pr",
2110 // Okay, we're going to insert the PHI node. Since PBI is not the only
2111 // predecessor, compute the PHI'd conditional value for all of the preds.
2112 // Any predecessor where the condition is not computable we keep symbolic.
2113 for (pred_iterator PI = PB; PI != PE; ++PI) {
2114 BasicBlock *P = *PI;
2115 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2116 PBI != BI && PBI->isConditional() &&
2117 PBI->getCondition() == BI->getCondition() &&
2118 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2119 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2120 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2123 NewPN->addIncoming(BI->getCondition(), P);
2127 BI->setCondition(NewPN);
2132 // If this is a conditional branch in an empty block, and if any
2133 // predecessors is a conditional branch to one of our destinations,
2134 // fold the conditions into logical ops and one cond br.
2135 BasicBlock::iterator BBI = BB->begin();
2136 // Ignore dbg intrinsics.
2137 while (isa<DbgInfoIntrinsic>(BBI))
2143 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2148 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2150 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2151 PBIOp = 0, BIOp = 1;
2152 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2153 PBIOp = 1, BIOp = 0;
2154 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2159 // Check to make sure that the other destination of this branch
2160 // isn't BB itself. If so, this is an infinite loop that will
2161 // keep getting unwound.
2162 if (PBI->getSuccessor(PBIOp) == BB)
2165 // Do not perform this transformation if it would require
2166 // insertion of a large number of select instructions. For targets
2167 // without predication/cmovs, this is a big pessimization.
2168 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2170 unsigned NumPhis = 0;
2171 for (BasicBlock::iterator II = CommonDest->begin();
2172 isa<PHINode>(II); ++II, ++NumPhis)
2173 if (NumPhis > 2) // Disable this xform.
2176 // Finally, if everything is ok, fold the branches to logical ops.
2177 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2179 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2180 << "AND: " << *BI->getParent());
2183 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2184 // branch in it, where one edge (OtherDest) goes back to itself but the other
2185 // exits. We don't *know* that the program avoids the infinite loop
2186 // (even though that seems likely). If we do this xform naively, we'll end up
2187 // recursively unpeeling the loop. Since we know that (after the xform is
2188 // done) that the block *is* infinite if reached, we just make it an obviously
2189 // infinite loop with no cond branch.
2190 if (OtherDest == BB) {
2191 // Insert it at the end of the function, because it's either code,
2192 // or it won't matter if it's hot. :)
2193 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2194 "infloop", BB->getParent());
2195 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2196 OtherDest = InfLoopBlock;
2199 DEBUG(dbgs() << *PBI->getParent()->getParent());
2201 // BI may have other predecessors. Because of this, we leave
2202 // it alone, but modify PBI.
2204 // Make sure we get to CommonDest on True&True directions.
2205 Value *PBICond = PBI->getCondition();
2206 IRBuilder<true, NoFolder> Builder(PBI);
2208 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2210 Value *BICond = BI->getCondition();
2212 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2214 // Merge the conditions.
2215 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2217 // Modify PBI to branch on the new condition to the new dests.
2218 PBI->setCondition(Cond);
2219 PBI->setSuccessor(0, CommonDest);
2220 PBI->setSuccessor(1, OtherDest);
2222 // OtherDest may have phi nodes. If so, add an entry from PBI's
2223 // block that are identical to the entries for BI's block.
2224 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2226 // We know that the CommonDest already had an edge from PBI to
2227 // it. If it has PHIs though, the PHIs may have different
2228 // entries for BB and PBI's BB. If so, insert a select to make
2231 for (BasicBlock::iterator II = CommonDest->begin();
2232 (PN = dyn_cast<PHINode>(II)); ++II) {
2233 Value *BIV = PN->getIncomingValueForBlock(BB);
2234 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2235 Value *PBIV = PN->getIncomingValue(PBBIdx);
2237 // Insert a select in PBI to pick the right value.
2238 Value *NV = cast<SelectInst>
2239 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2240 PN->setIncomingValue(PBBIdx, NV);
2244 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2245 DEBUG(dbgs() << *PBI->getParent()->getParent());
2247 // This basic block is probably dead. We know it has at least
2248 // one fewer predecessor.
2252 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2253 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2254 // Takes care of updating the successors and removing the old terminator.
2255 // Also makes sure not to introduce new successors by assuming that edges to
2256 // non-successor TrueBBs and FalseBBs aren't reachable.
2257 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2258 BasicBlock *TrueBB, BasicBlock *FalseBB){
2259 // Remove any superfluous successor edges from the CFG.
2260 // First, figure out which successors to preserve.
2261 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2263 BasicBlock *KeepEdge1 = TrueBB;
2264 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2266 // Then remove the rest.
2267 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2268 BasicBlock *Succ = OldTerm->getSuccessor(I);
2269 // Make sure only to keep exactly one copy of each edge.
2270 if (Succ == KeepEdge1)
2272 else if (Succ == KeepEdge2)
2275 Succ->removePredecessor(OldTerm->getParent());
2278 IRBuilder<> Builder(OldTerm);
2279 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2281 // Insert an appropriate new terminator.
2282 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2283 if (TrueBB == FalseBB)
2284 // We were only looking for one successor, and it was present.
2285 // Create an unconditional branch to it.
2286 Builder.CreateBr(TrueBB);
2288 // We found both of the successors we were looking for.
2289 // Create a conditional branch sharing the condition of the select.
2290 Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2291 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2292 // Neither of the selected blocks were successors, so this
2293 // terminator must be unreachable.
2294 new UnreachableInst(OldTerm->getContext(), OldTerm);
2296 // One of the selected values was a successor, but the other wasn't.
2297 // Insert an unconditional branch to the one that was found;
2298 // the edge to the one that wasn't must be unreachable.
2300 // Only TrueBB was found.
2301 Builder.CreateBr(TrueBB);
2303 // Only FalseBB was found.
2304 Builder.CreateBr(FalseBB);
2307 EraseTerminatorInstAndDCECond(OldTerm);
2311 // SimplifySwitchOnSelect - Replaces
2312 // (switch (select cond, X, Y)) on constant X, Y
2313 // with a branch - conditional if X and Y lead to distinct BBs,
2314 // unconditional otherwise.
2315 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2316 // Check for constant integer values in the select.
2317 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2318 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2319 if (!TrueVal || !FalseVal)
2322 // Find the relevant condition and destinations.
2323 Value *Condition = Select->getCondition();
2324 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2325 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2327 // Perform the actual simplification.
2328 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB);
2331 // SimplifyIndirectBrOnSelect - Replaces
2332 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2333 // blockaddress(@fn, BlockB)))
2335 // (br cond, BlockA, BlockB).
2336 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2337 // Check that both operands of the select are block addresses.
2338 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2339 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2343 // Extract the actual blocks.
2344 BasicBlock *TrueBB = TBA->getBasicBlock();
2345 BasicBlock *FalseBB = FBA->getBasicBlock();
2347 // Perform the actual simplification.
2348 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB);
2351 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2352 /// instruction (a seteq/setne with a constant) as the only instruction in a
2353 /// block that ends with an uncond branch. We are looking for a very specific
2354 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2355 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2356 /// default value goes to an uncond block with a seteq in it, we get something
2359 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2361 /// %tmp = icmp eq i8 %A, 92
2364 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2366 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2367 /// the PHI, merging the third icmp into the switch.
2368 static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
2369 const TargetData *TD,
2370 IRBuilder<> &Builder) {
2371 BasicBlock *BB = ICI->getParent();
2373 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2375 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2377 Value *V = ICI->getOperand(0);
2378 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2380 // The pattern we're looking for is where our only predecessor is a switch on
2381 // 'V' and this block is the default case for the switch. In this case we can
2382 // fold the compared value into the switch to simplify things.
2383 BasicBlock *Pred = BB->getSinglePredecessor();
2384 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2386 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2387 if (SI->getCondition() != V)
2390 // If BB is reachable on a non-default case, then we simply know the value of
2391 // V in this block. Substitute it and constant fold the icmp instruction
2393 if (SI->getDefaultDest() != BB) {
2394 ConstantInt *VVal = SI->findCaseDest(BB);
2395 assert(VVal && "Should have a unique destination value");
2396 ICI->setOperand(0, VVal);
2398 if (Value *V = SimplifyInstruction(ICI, TD)) {
2399 ICI->replaceAllUsesWith(V);
2400 ICI->eraseFromParent();
2402 // BB is now empty, so it is likely to simplify away.
2403 return SimplifyCFG(BB) | true;
2406 // Ok, the block is reachable from the default dest. If the constant we're
2407 // comparing exists in one of the other edges, then we can constant fold ICI
2409 if (SI->findCaseValue(Cst) != SI->case_default()) {
2411 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2412 V = ConstantInt::getFalse(BB->getContext());
2414 V = ConstantInt::getTrue(BB->getContext());
2416 ICI->replaceAllUsesWith(V);
2417 ICI->eraseFromParent();
2418 // BB is now empty, so it is likely to simplify away.
2419 return SimplifyCFG(BB) | true;
2422 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2424 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2425 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2426 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2427 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2430 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2432 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2433 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2435 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2436 std::swap(DefaultCst, NewCst);
2438 // Replace ICI (which is used by the PHI for the default value) with true or
2439 // false depending on if it is EQ or NE.
2440 ICI->replaceAllUsesWith(DefaultCst);
2441 ICI->eraseFromParent();
2443 // Okay, the switch goes to this block on a default value. Add an edge from
2444 // the switch to the merge point on the compared value.
2445 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2446 BB->getParent(), BB);
2447 SI->addCase(Cst, NewBB);
2449 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2450 Builder.SetInsertPoint(NewBB);
2451 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2452 Builder.CreateBr(SuccBlock);
2453 PHIUse->addIncoming(NewCst, NewBB);
2457 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2458 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2459 /// fold it into a switch instruction if so.
2460 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD,
2461 IRBuilder<> &Builder) {
2462 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2463 if (Cond == 0) return false;
2466 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2467 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2468 // 'setne's and'ed together, collect them.
2470 std::vector<ConstantInt*> Values;
2471 bool TrueWhenEqual = true;
2472 Value *ExtraCase = 0;
2473 unsigned UsedICmps = 0;
2475 if (Cond->getOpcode() == Instruction::Or) {
2476 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2478 } else if (Cond->getOpcode() == Instruction::And) {
2479 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2481 TrueWhenEqual = false;
2484 // If we didn't have a multiply compared value, fail.
2485 if (CompVal == 0) return false;
2487 // Avoid turning single icmps into a switch.
2491 // There might be duplicate constants in the list, which the switch
2492 // instruction can't handle, remove them now.
2493 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2494 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2496 // If Extra was used, we require at least two switch values to do the
2497 // transformation. A switch with one value is just an cond branch.
2498 if (ExtraCase && Values.size() < 2) return false;
2500 // TODO: Preserve branch weight metadata, similarly to how
2501 // FoldValueComparisonIntoPredecessors preserves it.
2503 // Figure out which block is which destination.
2504 BasicBlock *DefaultBB = BI->getSuccessor(1);
2505 BasicBlock *EdgeBB = BI->getSuccessor(0);
2506 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2508 BasicBlock *BB = BI->getParent();
2510 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2511 << " cases into SWITCH. BB is:\n" << *BB);
2513 // If there are any extra values that couldn't be folded into the switch
2514 // then we evaluate them with an explicit branch first. Split the block
2515 // right before the condbr to handle it.
2517 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2518 // Remove the uncond branch added to the old block.
2519 TerminatorInst *OldTI = BB->getTerminator();
2520 Builder.SetInsertPoint(OldTI);
2523 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2525 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2527 OldTI->eraseFromParent();
2529 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2530 // for the edge we just added.
2531 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2533 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2534 << "\nEXTRABB = " << *BB);
2538 Builder.SetInsertPoint(BI);
2539 // Convert pointer to int before we switch.
2540 if (CompVal->getType()->isPointerTy()) {
2541 assert(TD && "Cannot switch on pointer without TargetData");
2542 CompVal = Builder.CreatePtrToInt(CompVal,
2543 TD->getIntPtrType(CompVal->getContext()),
2547 // Create the new switch instruction now.
2548 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2550 // Add all of the 'cases' to the switch instruction.
2551 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2552 New->addCase(Values[i], EdgeBB);
2554 // We added edges from PI to the EdgeBB. As such, if there were any
2555 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2556 // the number of edges added.
2557 for (BasicBlock::iterator BBI = EdgeBB->begin();
2558 isa<PHINode>(BBI); ++BBI) {
2559 PHINode *PN = cast<PHINode>(BBI);
2560 Value *InVal = PN->getIncomingValueForBlock(BB);
2561 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2562 PN->addIncoming(InVal, BB);
2565 // Erase the old branch instruction.
2566 EraseTerminatorInstAndDCECond(BI);
2568 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2572 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2573 // If this is a trivial landing pad that just continues unwinding the caught
2574 // exception then zap the landing pad, turning its invokes into calls.
2575 BasicBlock *BB = RI->getParent();
2576 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2577 if (RI->getValue() != LPInst)
2578 // Not a landing pad, or the resume is not unwinding the exception that
2579 // caused control to branch here.
2582 // Check that there are no other instructions except for debug intrinsics.
2583 BasicBlock::iterator I = LPInst, E = RI;
2585 if (!isa<DbgInfoIntrinsic>(I))
2588 // Turn all invokes that unwind here into calls and delete the basic block.
2589 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2590 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2591 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2592 // Insert a call instruction before the invoke.
2593 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2595 Call->setCallingConv(II->getCallingConv());
2596 Call->setAttributes(II->getAttributes());
2597 Call->setDebugLoc(II->getDebugLoc());
2599 // Anything that used the value produced by the invoke instruction now uses
2600 // the value produced by the call instruction. Note that we do this even
2601 // for void functions and calls with no uses so that the callgraph edge is
2603 II->replaceAllUsesWith(Call);
2604 BB->removePredecessor(II->getParent());
2606 // Insert a branch to the normal destination right before the invoke.
2607 BranchInst::Create(II->getNormalDest(), II);
2609 // Finally, delete the invoke instruction!
2610 II->eraseFromParent();
2613 // The landingpad is now unreachable. Zap it.
2614 BB->eraseFromParent();
2618 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2619 BasicBlock *BB = RI->getParent();
2620 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2622 // Find predecessors that end with branches.
2623 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2624 SmallVector<BranchInst*, 8> CondBranchPreds;
2625 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2626 BasicBlock *P = *PI;
2627 TerminatorInst *PTI = P->getTerminator();
2628 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2629 if (BI->isUnconditional())
2630 UncondBranchPreds.push_back(P);
2632 CondBranchPreds.push_back(BI);
2636 // If we found some, do the transformation!
2637 if (!UncondBranchPreds.empty() && DupRet) {
2638 while (!UncondBranchPreds.empty()) {
2639 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2640 DEBUG(dbgs() << "FOLDING: " << *BB
2641 << "INTO UNCOND BRANCH PRED: " << *Pred);
2642 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2645 // If we eliminated all predecessors of the block, delete the block now.
2646 if (pred_begin(BB) == pred_end(BB))
2647 // We know there are no successors, so just nuke the block.
2648 BB->eraseFromParent();
2653 // Check out all of the conditional branches going to this return
2654 // instruction. If any of them just select between returns, change the
2655 // branch itself into a select/return pair.
2656 while (!CondBranchPreds.empty()) {
2657 BranchInst *BI = CondBranchPreds.pop_back_val();
2659 // Check to see if the non-BB successor is also a return block.
2660 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2661 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2662 SimplifyCondBranchToTwoReturns(BI, Builder))
2668 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2669 BasicBlock *BB = UI->getParent();
2671 bool Changed = false;
2673 // If there are any instructions immediately before the unreachable that can
2674 // be removed, do so.
2675 while (UI != BB->begin()) {
2676 BasicBlock::iterator BBI = UI;
2678 // Do not delete instructions that can have side effects which might cause
2679 // the unreachable to not be reachable; specifically, calls and volatile
2680 // operations may have this effect.
2681 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2683 if (BBI->mayHaveSideEffects()) {
2684 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2685 if (SI->isVolatile())
2687 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2688 if (LI->isVolatile())
2690 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2691 if (RMWI->isVolatile())
2693 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2694 if (CXI->isVolatile())
2696 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2697 !isa<LandingPadInst>(BBI)) {
2700 // Note that deleting LandingPad's here is in fact okay, although it
2701 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2702 // all the predecessors of this block will be the unwind edges of Invokes,
2703 // and we can therefore guarantee this block will be erased.
2706 // Delete this instruction (any uses are guaranteed to be dead)
2707 if (!BBI->use_empty())
2708 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2709 BBI->eraseFromParent();
2713 // If the unreachable instruction is the first in the block, take a gander
2714 // at all of the predecessors of this instruction, and simplify them.
2715 if (&BB->front() != UI) return Changed;
2717 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2718 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
2719 TerminatorInst *TI = Preds[i]->getTerminator();
2720 IRBuilder<> Builder(TI);
2721 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2722 if (BI->isUnconditional()) {
2723 if (BI->getSuccessor(0) == BB) {
2724 new UnreachableInst(TI->getContext(), TI);
2725 TI->eraseFromParent();
2729 if (BI->getSuccessor(0) == BB) {
2730 Builder.CreateBr(BI->getSuccessor(1));
2731 EraseTerminatorInstAndDCECond(BI);
2732 } else if (BI->getSuccessor(1) == BB) {
2733 Builder.CreateBr(BI->getSuccessor(0));
2734 EraseTerminatorInstAndDCECond(BI);
2738 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2739 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2741 if (i.getCaseSuccessor() == BB) {
2742 BB->removePredecessor(SI->getParent());
2747 // If the default value is unreachable, figure out the most popular
2748 // destination and make it the default.
2749 if (SI->getDefaultDest() == BB) {
2750 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
2751 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2753 std::pair<unsigned, unsigned> &entry =
2754 Popularity[i.getCaseSuccessor()];
2755 if (entry.first == 0) {
2757 entry.second = i.getCaseIndex();
2763 // Find the most popular block.
2764 unsigned MaxPop = 0;
2765 unsigned MaxIndex = 0;
2766 BasicBlock *MaxBlock = 0;
2767 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
2768 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
2769 if (I->second.first > MaxPop ||
2770 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
2771 MaxPop = I->second.first;
2772 MaxIndex = I->second.second;
2773 MaxBlock = I->first;
2777 // Make this the new default, allowing us to delete any explicit
2779 SI->setDefaultDest(MaxBlock);
2782 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
2784 if (isa<PHINode>(MaxBlock->begin()))
2785 for (unsigned i = 0; i != MaxPop-1; ++i)
2786 MaxBlock->removePredecessor(SI->getParent());
2788 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2790 if (i.getCaseSuccessor() == MaxBlock) {
2796 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
2797 if (II->getUnwindDest() == BB) {
2798 // Convert the invoke to a call instruction. This would be a good
2799 // place to note that the call does not throw though.
2800 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
2801 II->removeFromParent(); // Take out of symbol table
2803 // Insert the call now...
2804 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
2805 Builder.SetInsertPoint(BI);
2806 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
2807 Args, II->getName());
2808 CI->setCallingConv(II->getCallingConv());
2809 CI->setAttributes(II->getAttributes());
2810 // If the invoke produced a value, the call does now instead.
2811 II->replaceAllUsesWith(CI);
2818 // If this block is now dead, remove it.
2819 if (pred_begin(BB) == pred_end(BB) &&
2820 BB != &BB->getParent()->getEntryBlock()) {
2821 // We know there are no successors, so just nuke the block.
2822 BB->eraseFromParent();
2829 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
2830 /// integer range comparison into a sub, an icmp and a branch.
2831 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
2832 assert(SI->getNumCases() > 1 && "Degenerate switch?");
2834 // Make sure all cases point to the same destination and gather the values.
2835 SmallVector<ConstantInt *, 16> Cases;
2836 SwitchInst::CaseIt I = SI->case_begin();
2837 Cases.push_back(I.getCaseValue());
2838 SwitchInst::CaseIt PrevI = I++;
2839 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
2840 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
2842 Cases.push_back(I.getCaseValue());
2844 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
2846 // Sort the case values, then check if they form a range we can transform.
2847 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
2848 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
2849 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
2853 Constant *Offset = ConstantExpr::getNeg(Cases.back());
2854 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
2856 Value *Sub = SI->getCondition();
2857 if (!Offset->isNullValue())
2858 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
2859 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
2860 Builder.CreateCondBr(
2861 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
2863 // Prune obsolete incoming values off the successor's PHI nodes.
2864 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
2865 isa<PHINode>(BBI); ++BBI) {
2866 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
2867 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
2869 SI->eraseFromParent();
2874 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
2875 /// and use it to remove dead cases.
2876 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
2877 Value *Cond = SI->getCondition();
2878 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
2879 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
2880 ComputeMaskedBits(Cond, KnownZero, KnownOne);
2882 // Gather dead cases.
2883 SmallVector<ConstantInt*, 8> DeadCases;
2884 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
2885 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
2886 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
2887 DeadCases.push_back(I.getCaseValue());
2888 DEBUG(dbgs() << "SimplifyCFG: switch case '"
2889 << I.getCaseValue() << "' is dead.\n");
2893 // Remove dead cases from the switch.
2894 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
2895 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
2896 assert(Case != SI->case_default() &&
2897 "Case was not found. Probably mistake in DeadCases forming.");
2898 // Prune unused values from PHI nodes.
2899 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
2900 SI->removeCase(Case);
2903 return !DeadCases.empty();
2906 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
2907 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
2908 /// by an unconditional branch), look at the phi node for BB in the successor
2909 /// block and see if the incoming value is equal to CaseValue. If so, return
2910 /// the phi node, and set PhiIndex to BB's index in the phi node.
2911 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
2914 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
2915 return NULL; // BB must be empty to be a candidate for simplification.
2916 if (!BB->getSinglePredecessor())
2917 return NULL; // BB must be dominated by the switch.
2919 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
2920 if (!Branch || !Branch->isUnconditional())
2921 return NULL; // Terminator must be unconditional branch.
2923 BasicBlock *Succ = Branch->getSuccessor(0);
2925 BasicBlock::iterator I = Succ->begin();
2926 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
2927 int Idx = PHI->getBasicBlockIndex(BB);
2928 assert(Idx >= 0 && "PHI has no entry for predecessor?");
2930 Value *InValue = PHI->getIncomingValue(Idx);
2931 if (InValue != CaseValue) continue;
2940 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
2941 /// instruction to a phi node dominated by the switch, if that would mean that
2942 /// some of the destination blocks of the switch can be folded away.
2943 /// Returns true if a change is made.
2944 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
2945 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
2946 ForwardingNodesMap ForwardingNodes;
2948 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
2949 ConstantInt *CaseValue = I.getCaseValue();
2950 BasicBlock *CaseDest = I.getCaseSuccessor();
2953 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
2957 ForwardingNodes[PHI].push_back(PhiIndex);
2960 bool Changed = false;
2962 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
2963 E = ForwardingNodes.end(); I != E; ++I) {
2964 PHINode *Phi = I->first;
2965 SmallVector<int,4> &Indexes = I->second;
2967 if (Indexes.size() < 2) continue;
2969 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
2970 Phi->setIncomingValue(Indexes[I], SI->getCondition());
2977 /// ValidLookupTableConstant - Return true if the backend will be able to handle
2978 /// initializing an array of constants like C.
2979 static bool ValidLookupTableConstant(Constant *C) {
2980 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2981 return CE->isGEPWithNoNotionalOverIndexing();
2983 return isa<ConstantFP>(C) ||
2984 isa<ConstantInt>(C) ||
2985 isa<ConstantPointerNull>(C) ||
2986 isa<GlobalValue>(C) ||
2990 /// GetCaseResulsts - Try to determine the resulting constant values in phi
2991 /// nodes at the common destination basic block for one of the case
2992 /// destinations of a switch instruction.
2993 static bool GetCaseResults(SwitchInst *SI,
2994 BasicBlock *CaseDest,
2995 BasicBlock **CommonDest,
2996 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) {
2997 // The block from which we enter the common destination.
2998 BasicBlock *Pred = SI->getParent();
3000 // If CaseDest is empty, continue to its successor.
3001 if (CaseDest->getFirstNonPHIOrDbg() == CaseDest->getTerminator() &&
3002 !isa<PHINode>(CaseDest->begin())) {
3004 TerminatorInst *Terminator = CaseDest->getTerminator();
3005 if (Terminator->getNumSuccessors() != 1)
3009 CaseDest = Terminator->getSuccessor(0);
3012 // If we did not have a CommonDest before, use the current one.
3014 *CommonDest = CaseDest;
3015 // If the destination isn't the common one, abort.
3016 if (CaseDest != *CommonDest)
3019 // Get the values for this case from phi nodes in the destination block.
3020 BasicBlock::iterator I = (*CommonDest)->begin();
3021 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3022 int Idx = PHI->getBasicBlockIndex(Pred);
3026 Constant *ConstVal = dyn_cast<Constant>(PHI->getIncomingValue(Idx));
3030 // Be conservative about which kinds of constants we support.
3031 if (!ValidLookupTableConstant(ConstVal))
3034 Res.push_back(std::make_pair(PHI, ConstVal));
3040 /// BuildLookupTable - Build a lookup table with the contents of Results, using
3041 /// DefaultResult to fill the holes in the table. If the table ends up
3042 /// containing the same result in each element, set *SingleResult to that value
3043 /// and return NULL.
3044 static GlobalVariable *BuildLookupTable(Module &M,
3046 ConstantInt *Offset,
3047 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Results,
3048 Constant *DefaultResult,
3049 Constant **SingleResult) {
3050 assert(Results.size() && "Need values to build lookup table");
3051 assert(TableSize >= Results.size() && "Table needs to hold all values");
3053 // If all values in the table are equal, this is that value.
3054 Constant *SameResult = Results.begin()->second;
3056 // Build up the table contents.
3057 std::vector<Constant*> TableContents(TableSize);
3058 for (size_t I = 0, E = Results.size(); I != E; ++I) {
3059 ConstantInt *CaseVal = Results[I].first;
3060 Constant *CaseRes = Results[I].second;
3062 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
3063 TableContents[Idx] = CaseRes;
3065 if (CaseRes != SameResult)
3069 // Fill in any holes in the table with the default result.
3070 if (Results.size() < TableSize) {
3071 for (unsigned i = 0; i < TableSize; ++i) {
3072 if (!TableContents[i])
3073 TableContents[i] = DefaultResult;
3076 if (DefaultResult != SameResult)
3080 // Same result was used in the entire table; just return that.
3082 *SingleResult = SameResult;
3086 ArrayType *ArrayTy = ArrayType::get(DefaultResult->getType(), TableSize);
3087 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3089 GlobalVariable *GV = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3090 GlobalVariable::PrivateLinkage,
3093 GV->setUnnamedAddr(true);
3097 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3098 /// phi nodes in a common successor block with different constant values,
3099 /// replace the switch with lookup tables.
3100 static bool SwitchToLookupTable(SwitchInst *SI,
3101 IRBuilder<> &Builder) {
3102 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3103 // FIXME: Handle unreachable cases.
3105 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3106 // split off a dense part and build a lookup table for that.
3108 // FIXME: If the results are all integers and the lookup table would fit in a
3109 // target-legal register, we should store them as a bitmap and use shift/mask
3110 // to look up the result.
3112 // FIXME: This creates arrays of GEPs to constant strings, which means each
3113 // GEP needs a runtime relocation in PIC code. We should just build one big
3114 // string and lookup indices into that.
3116 // Ignore the switch if the number of cases are too small.
3117 // This is similar to the check when building jump tables in
3118 // SelectionDAGBuilder::handleJTSwitchCase.
3119 // FIXME: Determine the best cut-off.
3120 if (SI->getNumCases() < 4)
3123 // Figure out the corresponding result for each case value and phi node in the
3124 // common destination, as well as the the min and max case values.
3125 assert(SI->case_begin() != SI->case_end());
3126 SwitchInst::CaseIt CI = SI->case_begin();
3127 ConstantInt *MinCaseVal = CI.getCaseValue();
3128 ConstantInt *MaxCaseVal = CI.getCaseValue();
3130 BasicBlock *CommonDest = NULL;
3131 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3132 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3133 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3134 SmallDenseMap<PHINode*, Type*> ResultTypes;
3135 SmallVector<PHINode*, 4> PHIs;
3137 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3138 ConstantInt *CaseVal = CI.getCaseValue();
3139 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3140 MinCaseVal = CaseVal;
3141 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3142 MaxCaseVal = CaseVal;
3144 // Resulting value at phi nodes for this case value.
3145 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3147 if (!GetCaseResults(SI, CI.getCaseSuccessor(), &CommonDest, Results))
3150 // Append the result from this case to the list for each phi.
3151 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3152 if (!ResultLists.count(I->first))
3153 PHIs.push_back(I->first);
3154 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3158 // Get the resulting values for the default case.
3159 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3160 if (!GetCaseResults(SI, SI->getDefaultDest(), &CommonDest, DefaultResultsList))
3162 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3163 PHINode *PHI = DefaultResultsList[I].first;
3164 Constant *Result = DefaultResultsList[I].second;
3165 DefaultResults[PHI] = Result;
3166 ResultTypes[PHI] = Result->getType();
3169 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3170 // The table density should be at lest 40%. This is the same criterion as for
3171 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3172 // FIXME: Find the best cut-off.
3173 // Be careful to avoid overlow in the density computation.
3174 if (RangeSpread.zextOrSelf(64).ugt(UINT64_MAX / 4 - 1))
3176 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3177 if (SI->getNumCases() * 10 < TableSize * 4)
3180 // Build the lookup tables.
3181 SmallDenseMap<PHINode*, GlobalVariable*> LookupTables;
3182 SmallDenseMap<PHINode*, Constant*> SingleResults;
3184 Module &Mod = *CommonDest->getParent()->getParent();
3185 for (SmallVector<PHINode*, 4>::iterator I = PHIs.begin(), E = PHIs.end();
3189 Constant *SingleResult = NULL;
3190 LookupTables[PHI] = BuildLookupTable(Mod, TableSize, MinCaseVal,
3191 ResultLists[PHI], DefaultResults[PHI],
3193 SingleResults[PHI] = SingleResult;
3196 // Create the BB that does the lookups.
3197 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3199 CommonDest->getParent(),
3202 // Check whether the condition value is within the case range, and branch to
3204 Builder.SetInsertPoint(SI);
3205 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3207 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3208 MinCaseVal->getType(), TableSize));
3209 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3211 // Populate the BB that does the lookups.
3212 Builder.SetInsertPoint(LookupBB);
3213 bool ReturnedEarly = false;
3214 for (SmallVector<PHINode*, 4>::iterator I = PHIs.begin(), E = PHIs.end();
3217 // There was a single result for this phi; just use that.
3218 if (Constant *SingleResult = SingleResults[PHI]) {
3219 PHI->addIncoming(SingleResult, LookupBB);
3223 Value *GEPIndices[] = { Builder.getInt32(0), TableIndex };
3224 Value *GEP = Builder.CreateInBoundsGEP(LookupTables[PHI], GEPIndices,
3226 Value *Result = Builder.CreateLoad(GEP, "switch.load");
3228 // If the result is only going to be used to return from the function,
3229 // we want to do that right here.
3230 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin())) {
3231 if (CommonDest->getFirstNonPHIOrDbg() == CommonDest->getTerminator()) {
3232 Builder.CreateRet(Result);
3233 ReturnedEarly = true;
3238 PHI->addIncoming(Result, LookupBB);
3242 Builder.CreateBr(CommonDest);
3244 // Remove the switch.
3245 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3246 BasicBlock *Succ = SI->getSuccessor(i);
3247 if (Succ == SI->getDefaultDest()) continue;
3248 Succ->removePredecessor(SI->getParent());
3250 SI->eraseFromParent();
3256 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3257 // If this switch is too complex to want to look at, ignore it.
3258 if (!isValueEqualityComparison(SI))
3261 BasicBlock *BB = SI->getParent();
3263 // If we only have one predecessor, and if it is a branch on this value,
3264 // see if that predecessor totally determines the outcome of this switch.
3265 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3266 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3267 return SimplifyCFG(BB) | true;
3269 Value *Cond = SI->getCondition();
3270 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3271 if (SimplifySwitchOnSelect(SI, Select))
3272 return SimplifyCFG(BB) | true;
3274 // If the block only contains the switch, see if we can fold the block
3275 // away into any preds.
3276 BasicBlock::iterator BBI = BB->begin();
3277 // Ignore dbg intrinsics.
3278 while (isa<DbgInfoIntrinsic>(BBI))
3281 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3282 return SimplifyCFG(BB) | true;
3284 // Try to transform the switch into an icmp and a branch.
3285 if (TurnSwitchRangeIntoICmp(SI, Builder))
3286 return SimplifyCFG(BB) | true;
3288 // Remove unreachable cases.
3289 if (EliminateDeadSwitchCases(SI))
3290 return SimplifyCFG(BB) | true;
3292 if (ForwardSwitchConditionToPHI(SI))
3293 return SimplifyCFG(BB) | true;
3295 if (SwitchToLookupTable(SI, Builder))
3296 return SimplifyCFG(BB) | true;
3301 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3302 BasicBlock *BB = IBI->getParent();
3303 bool Changed = false;
3305 // Eliminate redundant destinations.
3306 SmallPtrSet<Value *, 8> Succs;
3307 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3308 BasicBlock *Dest = IBI->getDestination(i);
3309 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3310 Dest->removePredecessor(BB);
3311 IBI->removeDestination(i);
3317 if (IBI->getNumDestinations() == 0) {
3318 // If the indirectbr has no successors, change it to unreachable.
3319 new UnreachableInst(IBI->getContext(), IBI);
3320 EraseTerminatorInstAndDCECond(IBI);
3324 if (IBI->getNumDestinations() == 1) {
3325 // If the indirectbr has one successor, change it to a direct branch.
3326 BranchInst::Create(IBI->getDestination(0), IBI);
3327 EraseTerminatorInstAndDCECond(IBI);
3331 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3332 if (SimplifyIndirectBrOnSelect(IBI, SI))
3333 return SimplifyCFG(BB) | true;
3338 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3339 BasicBlock *BB = BI->getParent();
3341 // If the Terminator is the only non-phi instruction, simplify the block.
3342 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3343 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3344 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3347 // If the only instruction in the block is a seteq/setne comparison
3348 // against a constant, try to simplify the block.
3349 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3350 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3351 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3353 if (I->isTerminator() &&
3354 TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder))
3358 // If this basic block is ONLY a compare and a branch, and if a predecessor
3359 // branches to us and our successor, fold the comparison into the
3360 // predecessor and use logical operations to update the incoming value
3361 // for PHI nodes in common successor.
3362 if (FoldBranchToCommonDest(BI))
3363 return SimplifyCFG(BB) | true;
3368 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3369 BasicBlock *BB = BI->getParent();
3371 // Conditional branch
3372 if (isValueEqualityComparison(BI)) {
3373 // If we only have one predecessor, and if it is a branch on this value,
3374 // see if that predecessor totally determines the outcome of this
3376 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3377 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3378 return SimplifyCFG(BB) | true;
3380 // This block must be empty, except for the setcond inst, if it exists.
3381 // Ignore dbg intrinsics.
3382 BasicBlock::iterator I = BB->begin();
3383 // Ignore dbg intrinsics.
3384 while (isa<DbgInfoIntrinsic>(I))
3387 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3388 return SimplifyCFG(BB) | true;
3389 } else if (&*I == cast<Instruction>(BI->getCondition())){
3391 // Ignore dbg intrinsics.
3392 while (isa<DbgInfoIntrinsic>(I))
3394 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3395 return SimplifyCFG(BB) | true;
3399 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3400 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3403 // If this basic block is ONLY a compare and a branch, and if a predecessor
3404 // branches to us and one of our successors, fold the comparison into the
3405 // predecessor and use logical operations to pick the right destination.
3406 if (FoldBranchToCommonDest(BI))
3407 return SimplifyCFG(BB) | true;
3409 // We have a conditional branch to two blocks that are only reachable
3410 // from BI. We know that the condbr dominates the two blocks, so see if
3411 // there is any identical code in the "then" and "else" blocks. If so, we
3412 // can hoist it up to the branching block.
3413 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3414 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3415 if (HoistThenElseCodeToIf(BI))
3416 return SimplifyCFG(BB) | true;
3418 // If Successor #1 has multiple preds, we may be able to conditionally
3419 // execute Successor #0 if it branches to successor #1.
3420 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3421 if (Succ0TI->getNumSuccessors() == 1 &&
3422 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3423 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
3424 return SimplifyCFG(BB) | true;
3426 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3427 // If Successor #0 has multiple preds, we may be able to conditionally
3428 // execute Successor #1 if it branches to successor #0.
3429 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3430 if (Succ1TI->getNumSuccessors() == 1 &&
3431 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
3432 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
3433 return SimplifyCFG(BB) | true;
3436 // If this is a branch on a phi node in the current block, thread control
3437 // through this block if any PHI node entries are constants.
3438 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
3439 if (PN->getParent() == BI->getParent())
3440 if (FoldCondBranchOnPHI(BI, TD))
3441 return SimplifyCFG(BB) | true;
3443 // Scan predecessor blocks for conditional branches.
3444 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
3445 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
3446 if (PBI != BI && PBI->isConditional())
3447 if (SimplifyCondBranchToCondBranch(PBI, BI))
3448 return SimplifyCFG(BB) | true;
3453 /// Check if passing a value to an instruction will cause undefined behavior.
3454 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
3455 Constant *C = dyn_cast<Constant>(V);
3459 if (!I->hasOneUse()) // Only look at single-use instructions, for compile time
3462 if (C->isNullValue()) {
3463 Instruction *Use = I->use_back();
3465 // Now make sure that there are no instructions in between that can alter
3466 // control flow (eg. calls)
3467 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
3468 if (i == I->getParent()->end() || i->mayHaveSideEffects())
3471 // Look through GEPs. A load from a GEP derived from NULL is still undefined
3472 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
3473 if (GEP->getPointerOperand() == I)
3474 return passingValueIsAlwaysUndefined(V, GEP);
3476 // Look through bitcasts.
3477 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
3478 return passingValueIsAlwaysUndefined(V, BC);
3480 // Load from null is undefined.
3481 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
3482 return LI->getPointerAddressSpace() == 0;
3484 // Store to null is undefined.
3485 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
3486 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
3491 /// If BB has an incoming value that will always trigger undefined behavior
3492 /// (eg. null pointer dereference), remove the branch leading here.
3493 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
3494 for (BasicBlock::iterator i = BB->begin();
3495 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
3496 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
3497 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
3498 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
3499 IRBuilder<> Builder(T);
3500 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
3501 BB->removePredecessor(PHI->getIncomingBlock(i));
3502 // Turn uncoditional branches into unreachables and remove the dead
3503 // destination from conditional branches.
3504 if (BI->isUnconditional())
3505 Builder.CreateUnreachable();
3507 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
3508 BI->getSuccessor(0));
3509 BI->eraseFromParent();
3512 // TODO: SwitchInst.
3518 bool SimplifyCFGOpt::run(BasicBlock *BB) {
3519 bool Changed = false;
3521 assert(BB && BB->getParent() && "Block not embedded in function!");
3522 assert(BB->getTerminator() && "Degenerate basic block encountered!");
3524 // Remove basic blocks that have no predecessors (except the entry block)...
3525 // or that just have themself as a predecessor. These are unreachable.
3526 if ((pred_begin(BB) == pred_end(BB) &&
3527 BB != &BB->getParent()->getEntryBlock()) ||
3528 BB->getSinglePredecessor() == BB) {
3529 DEBUG(dbgs() << "Removing BB: \n" << *BB);
3530 DeleteDeadBlock(BB);
3534 // Check to see if we can constant propagate this terminator instruction
3536 Changed |= ConstantFoldTerminator(BB, true);
3538 // Check for and eliminate duplicate PHI nodes in this block.
3539 Changed |= EliminateDuplicatePHINodes(BB);
3541 // Check for and remove branches that will always cause undefined behavior.
3542 Changed |= removeUndefIntroducingPredecessor(BB);
3544 // Merge basic blocks into their predecessor if there is only one distinct
3545 // pred, and if there is only one distinct successor of the predecessor, and
3546 // if there are no PHI nodes.
3548 if (MergeBlockIntoPredecessor(BB))
3551 IRBuilder<> Builder(BB);
3553 // If there is a trivial two-entry PHI node in this basic block, and we can
3554 // eliminate it, do so now.
3555 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
3556 if (PN->getNumIncomingValues() == 2)
3557 Changed |= FoldTwoEntryPHINode(PN, TD);
3559 Builder.SetInsertPoint(BB->getTerminator());
3560 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
3561 if (BI->isUnconditional()) {
3562 if (SimplifyUncondBranch(BI, Builder)) return true;
3564 if (SimplifyCondBranch(BI, Builder)) return true;
3566 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
3567 if (SimplifyReturn(RI, Builder)) return true;
3568 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
3569 if (SimplifyResume(RI, Builder)) return true;
3570 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
3571 if (SimplifySwitch(SI, Builder)) return true;
3572 } else if (UnreachableInst *UI =
3573 dyn_cast<UnreachableInst>(BB->getTerminator())) {
3574 if (SimplifyUnreachable(UI)) return true;
3575 } else if (IndirectBrInst *IBI =
3576 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
3577 if (SimplifyIndirectBr(IBI)) return true;
3583 /// SimplifyCFG - This function is used to do simplification of a CFG. For
3584 /// example, it adjusts branches to branches to eliminate the extra hop, it
3585 /// eliminates unreachable basic blocks, and does other "peephole" optimization
3586 /// of the CFG. It returns true if a modification was made.
3588 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) {
3589 return SimplifyCFGOpt(TD).run(BB);