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 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
672 if (DeadCases.count(i.getCaseValue())) {
673 i.getCaseSuccessor()->removePredecessor(TI->getParent());
678 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
682 // Otherwise, TI's block must correspond to some matched value. Find out
683 // which value (or set of values) this is.
684 ConstantInt *TIV = 0;
685 BasicBlock *TIBB = TI->getParent();
686 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
687 if (PredCases[i].Dest == TIBB) {
689 return false; // Cannot handle multiple values coming to this block.
690 TIV = PredCases[i].Value;
692 assert(TIV && "No edge from pred to succ?");
694 // Okay, we found the one constant that our value can be if we get into TI's
695 // BB. Find out which successor will unconditionally be branched to.
696 BasicBlock *TheRealDest = 0;
697 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
698 if (ThisCases[i].Value == TIV) {
699 TheRealDest = ThisCases[i].Dest;
703 // If not handled by any explicit cases, it is handled by the default case.
704 if (TheRealDest == 0) TheRealDest = ThisDef;
706 // Remove PHI node entries for dead edges.
707 BasicBlock *CheckEdge = TheRealDest;
708 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
709 if (*SI != CheckEdge)
710 (*SI)->removePredecessor(TIBB);
714 // Insert the new branch.
715 Instruction *NI = Builder.CreateBr(TheRealDest);
718 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
719 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
721 EraseTerminatorInstAndDCECond(TI);
726 /// ConstantIntOrdering - This class implements a stable ordering of constant
727 /// integers that does not depend on their address. This is important for
728 /// applications that sort ConstantInt's to ensure uniqueness.
729 struct ConstantIntOrdering {
730 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
731 return LHS->getValue().ult(RHS->getValue());
736 static int ConstantIntSortPredicate(const void *P1, const void *P2) {
737 const ConstantInt *LHS = *(const ConstantInt*const*)P1;
738 const ConstantInt *RHS = *(const ConstantInt*const*)P2;
739 if (LHS->getValue().ult(RHS->getValue()))
741 if (LHS->getValue() == RHS->getValue())
746 static inline bool HasBranchWeights(const Instruction* I) {
747 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
748 if (ProfMD && ProfMD->getOperand(0))
749 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
750 return MDS->getString().equals("branch_weights");
755 /// Get Weights of a given TerminatorInst, the default weight is at the front
756 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
758 static void GetBranchWeights(TerminatorInst *TI,
759 SmallVectorImpl<uint64_t> &Weights) {
760 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
762 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
763 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i));
765 Weights.push_back(CI->getValue().getZExtValue());
768 // If TI is a conditional eq, the default case is the false case,
769 // and the corresponding branch-weight data is at index 2. We swap the
770 // default weight to be the first entry.
771 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
772 assert(Weights.size() == 2);
773 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
774 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
775 std::swap(Weights.front(), Weights.back());
779 /// Sees if any of the weights are too big for a uint32_t, and halves all the
780 /// weights if any are.
781 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
783 for (unsigned i = 0; i < Weights.size(); ++i)
784 if (Weights[i] > UINT_MAX) {
792 for (unsigned i = 0; i < Weights.size(); ++i)
796 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
797 /// equality comparison instruction (either a switch or a branch on "X == c").
798 /// See if any of the predecessors of the terminator block are value comparisons
799 /// on the same value. If so, and if safe to do so, fold them together.
800 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
801 IRBuilder<> &Builder) {
802 BasicBlock *BB = TI->getParent();
803 Value *CV = isValueEqualityComparison(TI); // CondVal
804 assert(CV && "Not a comparison?");
805 bool Changed = false;
807 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
808 while (!Preds.empty()) {
809 BasicBlock *Pred = Preds.pop_back_val();
811 // See if the predecessor is a comparison with the same value.
812 TerminatorInst *PTI = Pred->getTerminator();
813 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
815 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
816 // Figure out which 'cases' to copy from SI to PSI.
817 std::vector<ValueEqualityComparisonCase> BBCases;
818 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
820 std::vector<ValueEqualityComparisonCase> PredCases;
821 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
823 // Based on whether the default edge from PTI goes to BB or not, fill in
824 // PredCases and PredDefault with the new switch cases we would like to
826 SmallVector<BasicBlock*, 8> NewSuccessors;
828 // Update the branch weight metadata along the way
829 SmallVector<uint64_t, 8> Weights;
830 bool PredHasWeights = HasBranchWeights(PTI);
831 bool SuccHasWeights = HasBranchWeights(TI);
833 if (PredHasWeights) {
834 GetBranchWeights(PTI, Weights);
835 // branch-weight metadata is inconsistant here.
836 if (Weights.size() != 1 + PredCases.size())
837 PredHasWeights = SuccHasWeights = false;
838 } else if (SuccHasWeights)
839 // If there are no predecessor weights but there are successor weights,
840 // populate Weights with 1, which will later be scaled to the sum of
841 // successor's weights
842 Weights.assign(1 + PredCases.size(), 1);
844 SmallVector<uint64_t, 8> SuccWeights;
845 if (SuccHasWeights) {
846 GetBranchWeights(TI, SuccWeights);
847 // branch-weight metadata is inconsistant here.
848 if (SuccWeights.size() != 1 + BBCases.size())
849 PredHasWeights = SuccHasWeights = false;
850 } else if (PredHasWeights)
851 SuccWeights.assign(1 + BBCases.size(), 1);
853 if (PredDefault == BB) {
854 // If this is the default destination from PTI, only the edges in TI
855 // that don't occur in PTI, or that branch to BB will be activated.
856 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
857 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
858 if (PredCases[i].Dest != BB)
859 PTIHandled.insert(PredCases[i].Value);
861 // The default destination is BB, we don't need explicit targets.
862 std::swap(PredCases[i], PredCases.back());
864 if (PredHasWeights || SuccHasWeights) {
865 // Increase weight for the default case.
866 Weights[0] += Weights[i+1];
867 std::swap(Weights[i+1], Weights.back());
871 PredCases.pop_back();
875 // Reconstruct the new switch statement we will be building.
876 if (PredDefault != BBDefault) {
877 PredDefault->removePredecessor(Pred);
878 PredDefault = BBDefault;
879 NewSuccessors.push_back(BBDefault);
882 unsigned CasesFromPred = Weights.size();
883 uint64_t ValidTotalSuccWeight = 0;
884 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
885 if (!PTIHandled.count(BBCases[i].Value) &&
886 BBCases[i].Dest != BBDefault) {
887 PredCases.push_back(BBCases[i]);
888 NewSuccessors.push_back(BBCases[i].Dest);
889 if (SuccHasWeights || PredHasWeights) {
890 // The default weight is at index 0, so weight for the ith case
891 // should be at index i+1. Scale the cases from successor by
892 // PredDefaultWeight (Weights[0]).
893 Weights.push_back(Weights[0] * SuccWeights[i+1]);
894 ValidTotalSuccWeight += SuccWeights[i+1];
898 if (SuccHasWeights || PredHasWeights) {
899 ValidTotalSuccWeight += SuccWeights[0];
900 // Scale the cases from predecessor by ValidTotalSuccWeight.
901 for (unsigned i = 1; i < CasesFromPred; ++i)
902 Weights[i] *= ValidTotalSuccWeight;
903 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
904 Weights[0] *= SuccWeights[0];
907 // If this is not the default destination from PSI, only the edges
908 // in SI that occur in PSI with a destination of BB will be
910 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
911 std::map<ConstantInt*, uint64_t> WeightsForHandled;
912 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
913 if (PredCases[i].Dest == BB) {
914 PTIHandled.insert(PredCases[i].Value);
916 if (PredHasWeights || SuccHasWeights) {
917 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
918 std::swap(Weights[i+1], Weights.back());
922 std::swap(PredCases[i], PredCases.back());
923 PredCases.pop_back();
927 // Okay, now we know which constants were sent to BB from the
928 // predecessor. Figure out where they will all go now.
929 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
930 if (PTIHandled.count(BBCases[i].Value)) {
931 // If this is one we are capable of getting...
932 if (PredHasWeights || SuccHasWeights)
933 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
934 PredCases.push_back(BBCases[i]);
935 NewSuccessors.push_back(BBCases[i].Dest);
936 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
939 // If there are any constants vectored to BB that TI doesn't handle,
940 // they must go to the default destination of TI.
941 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
943 E = PTIHandled.end(); I != E; ++I) {
944 if (PredHasWeights || SuccHasWeights)
945 Weights.push_back(WeightsForHandled[*I]);
946 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
947 NewSuccessors.push_back(BBDefault);
951 // Okay, at this point, we know which new successor Pred will get. Make
952 // sure we update the number of entries in the PHI nodes for these
954 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
955 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
957 Builder.SetInsertPoint(PTI);
958 // Convert pointer to int before we switch.
959 if (CV->getType()->isPointerTy()) {
960 assert(TD && "Cannot switch on pointer without TargetData");
961 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()),
965 // Now that the successors are updated, create the new Switch instruction.
966 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
968 NewSI->setDebugLoc(PTI->getDebugLoc());
969 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
970 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
972 if (PredHasWeights || SuccHasWeights) {
973 // Halve the weights if any of them cannot fit in an uint32_t
976 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
978 NewSI->setMetadata(LLVMContext::MD_prof,
979 MDBuilder(BB->getContext()).
980 createBranchWeights(MDWeights));
983 EraseTerminatorInstAndDCECond(PTI);
985 // Okay, last check. If BB is still a successor of PSI, then we must
986 // have an infinite loop case. If so, add an infinitely looping block
987 // to handle the case to preserve the behavior of the code.
988 BasicBlock *InfLoopBlock = 0;
989 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
990 if (NewSI->getSuccessor(i) == BB) {
991 if (InfLoopBlock == 0) {
992 // Insert it at the end of the function, because it's either code,
993 // or it won't matter if it's hot. :)
994 InfLoopBlock = BasicBlock::Create(BB->getContext(),
995 "infloop", BB->getParent());
996 BranchInst::Create(InfLoopBlock, InfLoopBlock);
998 NewSI->setSuccessor(i, InfLoopBlock);
1007 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1008 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1009 // would need to do this), we can't hoist the invoke, as there is nowhere
1010 // to put the select in this case.
1011 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1012 Instruction *I1, Instruction *I2) {
1013 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1015 for (BasicBlock::iterator BBI = SI->begin();
1016 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1017 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1018 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1019 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1027 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1028 /// BB2, hoist any common code in the two blocks up into the branch block. The
1029 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1030 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1031 // This does very trivial matching, with limited scanning, to find identical
1032 // instructions in the two blocks. In particular, we don't want to get into
1033 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1034 // such, we currently just scan for obviously identical instructions in an
1036 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1037 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1039 BasicBlock::iterator BB1_Itr = BB1->begin();
1040 BasicBlock::iterator BB2_Itr = BB2->begin();
1042 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1043 // Skip debug info if it is not identical.
1044 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1045 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1046 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1047 while (isa<DbgInfoIntrinsic>(I1))
1049 while (isa<DbgInfoIntrinsic>(I2))
1052 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1053 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1056 // If we get here, we can hoist at least one instruction.
1057 BasicBlock *BIParent = BI->getParent();
1060 // If we are hoisting the terminator instruction, don't move one (making a
1061 // broken BB), instead clone it, and remove BI.
1062 if (isa<TerminatorInst>(I1))
1063 goto HoistTerminator;
1065 // For a normal instruction, we just move one to right before the branch,
1066 // then replace all uses of the other with the first. Finally, we remove
1067 // the now redundant second instruction.
1068 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1069 if (!I2->use_empty())
1070 I2->replaceAllUsesWith(I1);
1071 I1->intersectOptionalDataWith(I2);
1072 I2->eraseFromParent();
1076 // Skip debug info if it is not identical.
1077 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1078 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1079 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1080 while (isa<DbgInfoIntrinsic>(I1))
1082 while (isa<DbgInfoIntrinsic>(I2))
1085 } while (I1->isIdenticalToWhenDefined(I2));
1090 // It may not be possible to hoist an invoke.
1091 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1094 // Okay, it is safe to hoist the terminator.
1095 Instruction *NT = I1->clone();
1096 BIParent->getInstList().insert(BI, NT);
1097 if (!NT->getType()->isVoidTy()) {
1098 I1->replaceAllUsesWith(NT);
1099 I2->replaceAllUsesWith(NT);
1103 IRBuilder<true, NoFolder> Builder(NT);
1104 // Hoisting one of the terminators from our successor is a great thing.
1105 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1106 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1107 // nodes, so we insert select instruction to compute the final result.
1108 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1109 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1111 for (BasicBlock::iterator BBI = SI->begin();
1112 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1113 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1114 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1115 if (BB1V == BB2V) continue;
1117 // These values do not agree. Insert a select instruction before NT
1118 // that determines the right value.
1119 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1121 SI = cast<SelectInst>
1122 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1123 BB1V->getName()+"."+BB2V->getName()));
1125 // Make the PHI node use the select for all incoming values for BB1/BB2
1126 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1127 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1128 PN->setIncomingValue(i, SI);
1132 // Update any PHI nodes in our new successors.
1133 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1134 AddPredecessorToBlock(*SI, BIParent, BB1);
1136 EraseTerminatorInstAndDCECond(BI);
1140 /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1
1141 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code
1142 /// (for now, restricted to a single instruction that's side effect free) from
1143 /// the BB1 into the branch block to speculatively execute it.
1148 /// br i1 %t1, label %BB1, label %BB2
1150 /// %t3 = add %t2, c
1156 /// %t4 = add %t2, c
1157 /// %t3 = select i1 %t1, %t2, %t3
1158 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) {
1159 // Only speculatively execution a single instruction (not counting the
1160 // terminator) for now.
1161 Instruction *HInst = NULL;
1162 Instruction *Term = BB1->getTerminator();
1163 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end();
1164 BBI != BBE; ++BBI) {
1165 Instruction *I = BBI;
1167 if (isa<DbgInfoIntrinsic>(I)) continue;
1168 if (I == Term) break;
1175 BasicBlock *BIParent = BI->getParent();
1177 // Check the instruction to be hoisted, if there is one.
1179 // Don't hoist the instruction if it's unsafe or expensive.
1180 if (!isSafeToSpeculativelyExecute(HInst))
1182 if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold)
1185 // Do not hoist the instruction if any of its operands are defined but not
1186 // used in this BB. The transformation will prevent the operand from
1187 // being sunk into the use block.
1188 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end();
1190 Instruction *OpI = dyn_cast<Instruction>(*i);
1191 if (OpI && OpI->getParent() == BIParent &&
1192 !OpI->mayHaveSideEffects() &&
1193 !OpI->isUsedInBasicBlock(BIParent))
1198 // Be conservative for now. FP select instruction can often be expensive.
1199 Value *BrCond = BI->getCondition();
1200 if (isa<FCmpInst>(BrCond))
1203 // If BB1 is actually on the false edge of the conditional branch, remember
1204 // to swap the select operands later.
1205 bool Invert = false;
1206 if (BB1 != BI->getSuccessor(0)) {
1207 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?");
1211 // Collect interesting PHIs, and scan for hazards.
1212 SmallSetVector<std::pair<Value *, Value *>, 4> PHIs;
1213 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0);
1214 for (BasicBlock::iterator I = BB2->begin();
1215 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1216 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1217 Value *BIParentV = PN->getIncomingValueForBlock(BIParent);
1219 // Skip PHIs which are trivial.
1220 if (BB1V == BIParentV)
1223 // Check for saftey.
1224 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) {
1225 // An unfolded ConstantExpr could end up getting expanded into
1226 // Instructions. Don't speculate this and another instruction at
1230 if (!isSafeToSpeculativelyExecute(CE))
1232 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold)
1236 // Ok, we may insert a select for this PHI.
1237 PHIs.insert(std::make_pair(BB1V, BIParentV));
1240 // If there are no PHIs to process, bail early. This helps ensure idempotence
1245 // If we get here, we can hoist the instruction and if-convert.
1246 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";);
1248 // Hoist the instruction.
1250 BIParent->getInstList().splice(BI, BB1->getInstList(), HInst);
1252 // Insert selects and rewrite the PHI operands.
1253 IRBuilder<true, NoFolder> Builder(BI);
1254 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
1255 Value *TrueV = PHIs[i].first;
1256 Value *FalseV = PHIs[i].second;
1258 // Create a select whose true value is the speculatively executed value and
1259 // false value is the previously determined FalseV.
1262 SI = cast<SelectInst>
1263 (Builder.CreateSelect(BrCond, FalseV, TrueV,
1264 FalseV->getName() + "." + TrueV->getName()));
1266 SI = cast<SelectInst>
1267 (Builder.CreateSelect(BrCond, TrueV, FalseV,
1268 TrueV->getName() + "." + FalseV->getName()));
1270 // Make the PHI node use the select for all incoming values for "then" and
1272 for (BasicBlock::iterator I = BB2->begin();
1273 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1274 unsigned BB1I = PN->getBasicBlockIndex(BB1);
1275 unsigned BIParentI = PN->getBasicBlockIndex(BIParent);
1276 Value *BB1V = PN->getIncomingValue(BB1I);
1277 Value *BIParentV = PN->getIncomingValue(BIParentI);
1278 if (TrueV == BB1V && FalseV == BIParentV) {
1279 PN->setIncomingValue(BB1I, SI);
1280 PN->setIncomingValue(BIParentI, SI);
1289 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1290 /// across this block.
1291 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1292 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1295 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1296 if (isa<DbgInfoIntrinsic>(BBI))
1298 if (Size > 10) return false; // Don't clone large BB's.
1301 // We can only support instructions that do not define values that are
1302 // live outside of the current basic block.
1303 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1305 Instruction *U = cast<Instruction>(*UI);
1306 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1309 // Looks ok, continue checking.
1315 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1316 /// that is defined in the same block as the branch and if any PHI entries are
1317 /// constants, thread edges corresponding to that entry to be branches to their
1318 /// ultimate destination.
1319 static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) {
1320 BasicBlock *BB = BI->getParent();
1321 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1322 // NOTE: we currently cannot transform this case if the PHI node is used
1323 // outside of the block.
1324 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1327 // Degenerate case of a single entry PHI.
1328 if (PN->getNumIncomingValues() == 1) {
1329 FoldSingleEntryPHINodes(PN->getParent());
1333 // Now we know that this block has multiple preds and two succs.
1334 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1336 // Okay, this is a simple enough basic block. See if any phi values are
1338 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1339 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1340 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1342 // Okay, we now know that all edges from PredBB should be revectored to
1343 // branch to RealDest.
1344 BasicBlock *PredBB = PN->getIncomingBlock(i);
1345 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1347 if (RealDest == BB) continue; // Skip self loops.
1348 // Skip if the predecessor's terminator is an indirect branch.
1349 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1351 // The dest block might have PHI nodes, other predecessors and other
1352 // difficult cases. Instead of being smart about this, just insert a new
1353 // block that jumps to the destination block, effectively splitting
1354 // the edge we are about to create.
1355 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1356 RealDest->getName()+".critedge",
1357 RealDest->getParent(), RealDest);
1358 BranchInst::Create(RealDest, EdgeBB);
1360 // Update PHI nodes.
1361 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1363 // BB may have instructions that are being threaded over. Clone these
1364 // instructions into EdgeBB. We know that there will be no uses of the
1365 // cloned instructions outside of EdgeBB.
1366 BasicBlock::iterator InsertPt = EdgeBB->begin();
1367 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1368 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1369 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1370 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1373 // Clone the instruction.
1374 Instruction *N = BBI->clone();
1375 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1377 // Update operands due to translation.
1378 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1380 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1381 if (PI != TranslateMap.end())
1385 // Check for trivial simplification.
1386 if (Value *V = SimplifyInstruction(N, TD)) {
1387 TranslateMap[BBI] = V;
1388 delete N; // Instruction folded away, don't need actual inst
1390 // Insert the new instruction into its new home.
1391 EdgeBB->getInstList().insert(InsertPt, N);
1392 if (!BBI->use_empty())
1393 TranslateMap[BBI] = N;
1397 // Loop over all of the edges from PredBB to BB, changing them to branch
1398 // to EdgeBB instead.
1399 TerminatorInst *PredBBTI = PredBB->getTerminator();
1400 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1401 if (PredBBTI->getSuccessor(i) == BB) {
1402 BB->removePredecessor(PredBB);
1403 PredBBTI->setSuccessor(i, EdgeBB);
1406 // Recurse, simplifying any other constants.
1407 return FoldCondBranchOnPHI(BI, TD) | true;
1413 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1414 /// PHI node, see if we can eliminate it.
1415 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) {
1416 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1417 // statement", which has a very simple dominance structure. Basically, we
1418 // are trying to find the condition that is being branched on, which
1419 // subsequently causes this merge to happen. We really want control
1420 // dependence information for this check, but simplifycfg can't keep it up
1421 // to date, and this catches most of the cases we care about anyway.
1422 BasicBlock *BB = PN->getParent();
1423 BasicBlock *IfTrue, *IfFalse;
1424 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1426 // Don't bother if the branch will be constant folded trivially.
1427 isa<ConstantInt>(IfCond))
1430 // Okay, we found that we can merge this two-entry phi node into a select.
1431 // Doing so would require us to fold *all* two entry phi nodes in this block.
1432 // At some point this becomes non-profitable (particularly if the target
1433 // doesn't support cmov's). Only do this transformation if there are two or
1434 // fewer PHI nodes in this block.
1435 unsigned NumPhis = 0;
1436 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1440 // Loop over the PHI's seeing if we can promote them all to select
1441 // instructions. While we are at it, keep track of the instructions
1442 // that need to be moved to the dominating block.
1443 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1444 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1445 MaxCostVal1 = PHINodeFoldingThreshold;
1447 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1448 PHINode *PN = cast<PHINode>(II++);
1449 if (Value *V = SimplifyInstruction(PN, TD)) {
1450 PN->replaceAllUsesWith(V);
1451 PN->eraseFromParent();
1455 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1457 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1462 // If we folded the first phi, PN dangles at this point. Refresh it. If
1463 // we ran out of PHIs then we simplified them all.
1464 PN = dyn_cast<PHINode>(BB->begin());
1465 if (PN == 0) return true;
1467 // Don't fold i1 branches on PHIs which contain binary operators. These can
1468 // often be turned into switches and other things.
1469 if (PN->getType()->isIntegerTy(1) &&
1470 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1471 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1472 isa<BinaryOperator>(IfCond)))
1475 // If we all PHI nodes are promotable, check to make sure that all
1476 // instructions in the predecessor blocks can be promoted as well. If
1477 // not, we won't be able to get rid of the control flow, so it's not
1478 // worth promoting to select instructions.
1479 BasicBlock *DomBlock = 0;
1480 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1481 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1482 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1485 DomBlock = *pred_begin(IfBlock1);
1486 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1487 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1488 // This is not an aggressive instruction that we can promote.
1489 // Because of this, we won't be able to get rid of the control
1490 // flow, so the xform is not worth it.
1495 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1498 DomBlock = *pred_begin(IfBlock2);
1499 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1500 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1501 // This is not an aggressive instruction that we can promote.
1502 // Because of this, we won't be able to get rid of the control
1503 // flow, so the xform is not worth it.
1508 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1509 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1511 // If we can still promote the PHI nodes after this gauntlet of tests,
1512 // do all of the PHI's now.
1513 Instruction *InsertPt = DomBlock->getTerminator();
1514 IRBuilder<true, NoFolder> Builder(InsertPt);
1516 // Move all 'aggressive' instructions, which are defined in the
1517 // conditional parts of the if's up to the dominating block.
1519 DomBlock->getInstList().splice(InsertPt,
1520 IfBlock1->getInstList(), IfBlock1->begin(),
1521 IfBlock1->getTerminator());
1523 DomBlock->getInstList().splice(InsertPt,
1524 IfBlock2->getInstList(), IfBlock2->begin(),
1525 IfBlock2->getTerminator());
1527 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1528 // Change the PHI node into a select instruction.
1529 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1530 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1533 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1534 PN->replaceAllUsesWith(NV);
1536 PN->eraseFromParent();
1539 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1540 // has been flattened. Change DomBlock to jump directly to our new block to
1541 // avoid other simplifycfg's kicking in on the diamond.
1542 TerminatorInst *OldTI = DomBlock->getTerminator();
1543 Builder.SetInsertPoint(OldTI);
1544 Builder.CreateBr(BB);
1545 OldTI->eraseFromParent();
1549 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1550 /// to two returning blocks, try to merge them together into one return,
1551 /// introducing a select if the return values disagree.
1552 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1553 IRBuilder<> &Builder) {
1554 assert(BI->isConditional() && "Must be a conditional branch");
1555 BasicBlock *TrueSucc = BI->getSuccessor(0);
1556 BasicBlock *FalseSucc = BI->getSuccessor(1);
1557 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1558 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1560 // Check to ensure both blocks are empty (just a return) or optionally empty
1561 // with PHI nodes. If there are other instructions, merging would cause extra
1562 // computation on one path or the other.
1563 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1565 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1568 Builder.SetInsertPoint(BI);
1569 // Okay, we found a branch that is going to two return nodes. If
1570 // there is no return value for this function, just change the
1571 // branch into a return.
1572 if (FalseRet->getNumOperands() == 0) {
1573 TrueSucc->removePredecessor(BI->getParent());
1574 FalseSucc->removePredecessor(BI->getParent());
1575 Builder.CreateRetVoid();
1576 EraseTerminatorInstAndDCECond(BI);
1580 // Otherwise, figure out what the true and false return values are
1581 // so we can insert a new select instruction.
1582 Value *TrueValue = TrueRet->getReturnValue();
1583 Value *FalseValue = FalseRet->getReturnValue();
1585 // Unwrap any PHI nodes in the return blocks.
1586 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1587 if (TVPN->getParent() == TrueSucc)
1588 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1589 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1590 if (FVPN->getParent() == FalseSucc)
1591 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1593 // In order for this transformation to be safe, we must be able to
1594 // unconditionally execute both operands to the return. This is
1595 // normally the case, but we could have a potentially-trapping
1596 // constant expression that prevents this transformation from being
1598 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1601 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1605 // Okay, we collected all the mapped values and checked them for sanity, and
1606 // defined to really do this transformation. First, update the CFG.
1607 TrueSucc->removePredecessor(BI->getParent());
1608 FalseSucc->removePredecessor(BI->getParent());
1610 // Insert select instructions where needed.
1611 Value *BrCond = BI->getCondition();
1613 // Insert a select if the results differ.
1614 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1615 } else if (isa<UndefValue>(TrueValue)) {
1616 TrueValue = FalseValue;
1618 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1619 FalseValue, "retval");
1623 Value *RI = !TrueValue ?
1624 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1628 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1629 << "\n " << *BI << "NewRet = " << *RI
1630 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1632 EraseTerminatorInstAndDCECond(BI);
1637 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1638 /// probabilities of the branch taking each edge. Fills in the two APInt
1639 /// parameters and return true, or returns false if no or invalid metadata was
1641 static bool ExtractBranchMetadata(BranchInst *BI,
1642 APInt &ProbTrue, APInt &ProbFalse) {
1643 assert(BI->isConditional() &&
1644 "Looking for probabilities on unconditional branch?");
1645 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1646 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1647 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1648 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1649 if (!CITrue || !CIFalse) return false;
1650 ProbTrue = CITrue->getValue();
1651 ProbFalse = CIFalse->getValue();
1652 assert(ProbTrue.getBitWidth() == 32 && ProbFalse.getBitWidth() == 32 &&
1653 "Branch probability metadata must be 32-bit integers");
1657 /// MultiplyAndLosePrecision - Multiplies A and B, then returns the result. In
1658 /// the event of overflow, logically-shifts all four inputs right until the
1660 static APInt MultiplyAndLosePrecision(APInt &A, APInt &B, APInt &C, APInt &D,
1661 unsigned &BitsLost) {
1663 bool Overflow = false;
1664 APInt Result = A.umul_ov(B, Overflow);
1666 APInt MaxB = APInt::getMaxValue(A.getBitWidth()).udiv(A);
1670 } while (B.ugt(MaxB));
1671 A = A.lshr(BitsLost);
1672 C = C.lshr(BitsLost);
1673 D = D.lshr(BitsLost);
1679 /// checkCSEInPredecessor - Return true if the given instruction is available
1680 /// in its predecessor block. If yes, the instruction will be removed.
1682 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1683 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1685 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1686 Instruction *PBI = &*I;
1687 // Check whether Inst and PBI generate the same value.
1688 if (Inst->isIdenticalTo(PBI)) {
1689 Inst->replaceAllUsesWith(PBI);
1690 Inst->eraseFromParent();
1697 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1698 /// predecessor branches to us and one of our successors, fold the block into
1699 /// the predecessor and use logical operations to pick the right destination.
1700 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1701 BasicBlock *BB = BI->getParent();
1703 Instruction *Cond = 0;
1704 if (BI->isConditional())
1705 Cond = dyn_cast<Instruction>(BI->getCondition());
1707 // For unconditional branch, check for a simple CFG pattern, where
1708 // BB has a single predecessor and BB's successor is also its predecessor's
1709 // successor. If such pattern exisits, check for CSE between BB and its
1711 if (BasicBlock *PB = BB->getSinglePredecessor())
1712 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1713 if (PBI->isConditional() &&
1714 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1715 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1716 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1718 Instruction *Curr = I++;
1719 if (isa<CmpInst>(Curr)) {
1723 // Quit if we can't remove this instruction.
1724 if (!checkCSEInPredecessor(Curr, PB))
1733 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1734 Cond->getParent() != BB || !Cond->hasOneUse())
1737 // Only allow this if the condition is a simple instruction that can be
1738 // executed unconditionally. It must be in the same block as the branch, and
1739 // must be at the front of the block.
1740 BasicBlock::iterator FrontIt = BB->front();
1742 // Ignore dbg intrinsics.
1743 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1745 // Allow a single instruction to be hoisted in addition to the compare
1746 // that feeds the branch. We later ensure that any values that _it_ uses
1747 // were also live in the predecessor, so that we don't unnecessarily create
1748 // register pressure or inhibit out-of-order execution.
1749 Instruction *BonusInst = 0;
1750 if (&*FrontIt != Cond &&
1751 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
1752 isSafeToSpeculativelyExecute(FrontIt)) {
1753 BonusInst = &*FrontIt;
1756 // Ignore dbg intrinsics.
1757 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1760 // Only a single bonus inst is allowed.
1761 if (&*FrontIt != Cond)
1764 // Make sure the instruction after the condition is the cond branch.
1765 BasicBlock::iterator CondIt = Cond; ++CondIt;
1767 // Ingore dbg intrinsics.
1768 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
1773 // Cond is known to be a compare or binary operator. Check to make sure that
1774 // neither operand is a potentially-trapping constant expression.
1775 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
1778 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
1782 // Finally, don't infinitely unroll conditional loops.
1783 BasicBlock *TrueDest = BI->getSuccessor(0);
1784 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
1785 if (TrueDest == BB || FalseDest == BB)
1788 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1789 BasicBlock *PredBlock = *PI;
1790 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
1792 // Check that we have two conditional branches. If there is a PHI node in
1793 // the common successor, verify that the same value flows in from both
1795 SmallVector<PHINode*, 4> PHIs;
1796 if (PBI == 0 || PBI->isUnconditional() ||
1797 (BI->isConditional() &&
1798 !SafeToMergeTerminators(BI, PBI)) ||
1799 (!BI->isConditional() &&
1800 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
1803 // Determine if the two branches share a common destination.
1804 Instruction::BinaryOps Opc;
1805 bool InvertPredCond = false;
1807 if (BI->isConditional()) {
1808 if (PBI->getSuccessor(0) == TrueDest)
1809 Opc = Instruction::Or;
1810 else if (PBI->getSuccessor(1) == FalseDest)
1811 Opc = Instruction::And;
1812 else if (PBI->getSuccessor(0) == FalseDest)
1813 Opc = Instruction::And, InvertPredCond = true;
1814 else if (PBI->getSuccessor(1) == TrueDest)
1815 Opc = Instruction::Or, InvertPredCond = true;
1819 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
1823 // Ensure that any values used in the bonus instruction are also used
1824 // by the terminator of the predecessor. This means that those values
1825 // must already have been resolved, so we won't be inhibiting the
1826 // out-of-order core by speculating them earlier.
1828 // Collect the values used by the bonus inst
1829 SmallPtrSet<Value*, 4> UsedValues;
1830 for (Instruction::op_iterator OI = BonusInst->op_begin(),
1831 OE = BonusInst->op_end(); OI != OE; ++OI) {
1833 if (!isa<Constant>(V))
1834 UsedValues.insert(V);
1837 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
1838 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
1840 // Walk up to four levels back up the use-def chain of the predecessor's
1841 // terminator to see if all those values were used. The choice of four
1842 // levels is arbitrary, to provide a compile-time-cost bound.
1843 while (!Worklist.empty()) {
1844 std::pair<Value*, unsigned> Pair = Worklist.back();
1845 Worklist.pop_back();
1847 if (Pair.second >= 4) continue;
1848 UsedValues.erase(Pair.first);
1849 if (UsedValues.empty()) break;
1851 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
1852 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
1854 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
1858 if (!UsedValues.empty()) return false;
1861 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
1862 IRBuilder<> Builder(PBI);
1864 // If we need to invert the condition in the pred block to match, do so now.
1865 if (InvertPredCond) {
1866 Value *NewCond = PBI->getCondition();
1868 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
1869 CmpInst *CI = cast<CmpInst>(NewCond);
1870 CI->setPredicate(CI->getInversePredicate());
1872 NewCond = Builder.CreateNot(NewCond,
1873 PBI->getCondition()->getName()+".not");
1876 PBI->setCondition(NewCond);
1877 PBI->swapSuccessors();
1880 // If we have a bonus inst, clone it into the predecessor block.
1881 Instruction *NewBonus = 0;
1883 NewBonus = BonusInst->clone();
1884 PredBlock->getInstList().insert(PBI, NewBonus);
1885 NewBonus->takeName(BonusInst);
1886 BonusInst->setName(BonusInst->getName()+".old");
1889 // Clone Cond into the predecessor basic block, and or/and the
1890 // two conditions together.
1891 Instruction *New = Cond->clone();
1892 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
1893 PredBlock->getInstList().insert(PBI, New);
1894 New->takeName(Cond);
1895 Cond->setName(New->getName()+".old");
1897 if (BI->isConditional()) {
1898 Instruction *NewCond =
1899 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
1901 PBI->setCondition(NewCond);
1903 if (PBI->getSuccessor(0) == BB) {
1904 AddPredecessorToBlock(TrueDest, PredBlock, BB);
1905 PBI->setSuccessor(0, TrueDest);
1907 if (PBI->getSuccessor(1) == BB) {
1908 AddPredecessorToBlock(FalseDest, PredBlock, BB);
1909 PBI->setSuccessor(1, FalseDest);
1912 // Update PHI nodes in the common successors.
1913 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
1914 ConstantInt *PBI_C = cast<ConstantInt>(
1915 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
1916 assert(PBI_C->getType()->isIntegerTy(1));
1917 Instruction *MergedCond = 0;
1918 if (PBI->getSuccessor(0) == TrueDest) {
1919 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
1920 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
1921 // is false: !PBI_Cond and BI_Value
1922 Instruction *NotCond =
1923 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
1926 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
1931 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
1932 PBI->getCondition(), MergedCond,
1935 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
1936 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
1937 // is false: PBI_Cond and BI_Value
1939 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
1940 PBI->getCondition(), New,
1942 if (PBI_C->isOne()) {
1943 Instruction *NotCond =
1944 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
1947 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
1948 NotCond, MergedCond,
1953 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
1956 // Change PBI from Conditional to Unconditional.
1957 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
1958 EraseTerminatorInstAndDCECond(PBI);
1962 // TODO: If BB is reachable from all paths through PredBlock, then we
1963 // could replace PBI's branch probabilities with BI's.
1965 // Merge probability data into PredBlock's branch.
1967 if (PBI->isConditional() && BI->isConditional() &&
1968 ExtractBranchMetadata(PBI, C, D) && ExtractBranchMetadata(BI, A, B)) {
1969 // Given IR which does:
1971 // br i1 %x, label %bbB, label %bbC
1973 // br i1 %y, label %bbD, label %bbC
1974 // Let's call the probability that we take the edge from %bbA to %bbB
1975 // 'a', from %bbA to %bbC, 'b', from %bbB to %bbD 'c' and from %bbB to
1976 // %bbC probability 'd'.
1978 // We transform the IR into:
1980 // br i1 %z, label %bbD, label %bbC
1981 // where the probability of going to %bbD is (a*c) and going to bbC is
1984 // Probabilities aren't stored as ratios directly. Using branch weights,
1986 // (a*c)% = A*C, (b+(a*d))% = A*D+B*C+B*D.
1988 // In the event of overflow, we want to drop the LSB of the input
1992 // Ignore overflow result on ProbTrue.
1993 APInt ProbTrue = MultiplyAndLosePrecision(A, C, B, D, BitsLost);
1995 APInt Tmp1 = MultiplyAndLosePrecision(B, D, A, C, BitsLost);
1997 ProbTrue = ProbTrue.lshr(BitsLost*2);
2000 APInt Tmp2 = MultiplyAndLosePrecision(A, D, C, B, BitsLost);
2002 ProbTrue = ProbTrue.lshr(BitsLost*2);
2003 Tmp1 = Tmp1.lshr(BitsLost*2);
2006 APInt Tmp3 = MultiplyAndLosePrecision(B, C, A, D, BitsLost);
2008 ProbTrue = ProbTrue.lshr(BitsLost*2);
2009 Tmp1 = Tmp1.lshr(BitsLost*2);
2010 Tmp2 = Tmp2.lshr(BitsLost*2);
2013 bool Overflow1 = false, Overflow2 = false;
2014 APInt Tmp4 = Tmp2.uadd_ov(Tmp3, Overflow1);
2015 APInt ProbFalse = Tmp4.uadd_ov(Tmp1, Overflow2);
2017 if (Overflow1 || Overflow2) {
2018 ProbTrue = ProbTrue.lshr(1);
2019 Tmp1 = Tmp1.lshr(1);
2020 Tmp2 = Tmp2.lshr(1);
2021 Tmp3 = Tmp3.lshr(1);
2023 ProbFalse = Tmp4 + Tmp1;
2026 // The sum of branch weights must fit in 32-bits.
2027 if (ProbTrue.isNegative() && ProbFalse.isNegative()) {
2028 ProbTrue = ProbTrue.lshr(1);
2029 ProbFalse = ProbFalse.lshr(1);
2032 if (ProbTrue != ProbFalse) {
2033 // Normalize the result.
2034 APInt GCD = APIntOps::GreatestCommonDivisor(ProbTrue, ProbFalse);
2035 ProbTrue = ProbTrue.udiv(GCD);
2036 ProbFalse = ProbFalse.udiv(GCD);
2038 MDBuilder MDB(BI->getContext());
2039 MDNode *N = MDB.createBranchWeights(ProbTrue.getZExtValue(),
2040 ProbFalse.getZExtValue());
2041 PBI->setMetadata(LLVMContext::MD_prof, N);
2043 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2046 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2049 // Copy any debug value intrinsics into the end of PredBlock.
2050 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2051 if (isa<DbgInfoIntrinsic>(*I))
2052 I->clone()->insertBefore(PBI);
2059 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2060 /// predecessor of another block, this function tries to simplify it. We know
2061 /// that PBI and BI are both conditional branches, and BI is in one of the
2062 /// successor blocks of PBI - PBI branches to BI.
2063 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2064 assert(PBI->isConditional() && BI->isConditional());
2065 BasicBlock *BB = BI->getParent();
2067 // If this block ends with a branch instruction, and if there is a
2068 // predecessor that ends on a branch of the same condition, make
2069 // this conditional branch redundant.
2070 if (PBI->getCondition() == BI->getCondition() &&
2071 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2072 // Okay, the outcome of this conditional branch is statically
2073 // knowable. If this block had a single pred, handle specially.
2074 if (BB->getSinglePredecessor()) {
2075 // Turn this into a branch on constant.
2076 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2077 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2079 return true; // Nuke the branch on constant.
2082 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2083 // in the constant and simplify the block result. Subsequent passes of
2084 // simplifycfg will thread the block.
2085 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2086 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2087 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2088 std::distance(PB, PE),
2089 BI->getCondition()->getName() + ".pr",
2091 // Okay, we're going to insert the PHI node. Since PBI is not the only
2092 // predecessor, compute the PHI'd conditional value for all of the preds.
2093 // Any predecessor where the condition is not computable we keep symbolic.
2094 for (pred_iterator PI = PB; PI != PE; ++PI) {
2095 BasicBlock *P = *PI;
2096 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2097 PBI != BI && PBI->isConditional() &&
2098 PBI->getCondition() == BI->getCondition() &&
2099 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2100 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2101 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2104 NewPN->addIncoming(BI->getCondition(), P);
2108 BI->setCondition(NewPN);
2113 // If this is a conditional branch in an empty block, and if any
2114 // predecessors is a conditional branch to one of our destinations,
2115 // fold the conditions into logical ops and one cond br.
2116 BasicBlock::iterator BBI = BB->begin();
2117 // Ignore dbg intrinsics.
2118 while (isa<DbgInfoIntrinsic>(BBI))
2124 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2129 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2131 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2132 PBIOp = 0, BIOp = 1;
2133 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2134 PBIOp = 1, BIOp = 0;
2135 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2140 // Check to make sure that the other destination of this branch
2141 // isn't BB itself. If so, this is an infinite loop that will
2142 // keep getting unwound.
2143 if (PBI->getSuccessor(PBIOp) == BB)
2146 // Do not perform this transformation if it would require
2147 // insertion of a large number of select instructions. For targets
2148 // without predication/cmovs, this is a big pessimization.
2149 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2151 unsigned NumPhis = 0;
2152 for (BasicBlock::iterator II = CommonDest->begin();
2153 isa<PHINode>(II); ++II, ++NumPhis)
2154 if (NumPhis > 2) // Disable this xform.
2157 // Finally, if everything is ok, fold the branches to logical ops.
2158 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2160 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2161 << "AND: " << *BI->getParent());
2164 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2165 // branch in it, where one edge (OtherDest) goes back to itself but the other
2166 // exits. We don't *know* that the program avoids the infinite loop
2167 // (even though that seems likely). If we do this xform naively, we'll end up
2168 // recursively unpeeling the loop. Since we know that (after the xform is
2169 // done) that the block *is* infinite if reached, we just make it an obviously
2170 // infinite loop with no cond branch.
2171 if (OtherDest == BB) {
2172 // Insert it at the end of the function, because it's either code,
2173 // or it won't matter if it's hot. :)
2174 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2175 "infloop", BB->getParent());
2176 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2177 OtherDest = InfLoopBlock;
2180 DEBUG(dbgs() << *PBI->getParent()->getParent());
2182 // BI may have other predecessors. Because of this, we leave
2183 // it alone, but modify PBI.
2185 // Make sure we get to CommonDest on True&True directions.
2186 Value *PBICond = PBI->getCondition();
2187 IRBuilder<true, NoFolder> Builder(PBI);
2189 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2191 Value *BICond = BI->getCondition();
2193 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2195 // Merge the conditions.
2196 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2198 // Modify PBI to branch on the new condition to the new dests.
2199 PBI->setCondition(Cond);
2200 PBI->setSuccessor(0, CommonDest);
2201 PBI->setSuccessor(1, OtherDest);
2203 // OtherDest may have phi nodes. If so, add an entry from PBI's
2204 // block that are identical to the entries for BI's block.
2205 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2207 // We know that the CommonDest already had an edge from PBI to
2208 // it. If it has PHIs though, the PHIs may have different
2209 // entries for BB and PBI's BB. If so, insert a select to make
2212 for (BasicBlock::iterator II = CommonDest->begin();
2213 (PN = dyn_cast<PHINode>(II)); ++II) {
2214 Value *BIV = PN->getIncomingValueForBlock(BB);
2215 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2216 Value *PBIV = PN->getIncomingValue(PBBIdx);
2218 // Insert a select in PBI to pick the right value.
2219 Value *NV = cast<SelectInst>
2220 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2221 PN->setIncomingValue(PBBIdx, NV);
2225 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2226 DEBUG(dbgs() << *PBI->getParent()->getParent());
2228 // This basic block is probably dead. We know it has at least
2229 // one fewer predecessor.
2233 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2234 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2235 // Takes care of updating the successors and removing the old terminator.
2236 // Also makes sure not to introduce new successors by assuming that edges to
2237 // non-successor TrueBBs and FalseBBs aren't reachable.
2238 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2239 BasicBlock *TrueBB, BasicBlock *FalseBB){
2240 // Remove any superfluous successor edges from the CFG.
2241 // First, figure out which successors to preserve.
2242 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2244 BasicBlock *KeepEdge1 = TrueBB;
2245 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2247 // Then remove the rest.
2248 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2249 BasicBlock *Succ = OldTerm->getSuccessor(I);
2250 // Make sure only to keep exactly one copy of each edge.
2251 if (Succ == KeepEdge1)
2253 else if (Succ == KeepEdge2)
2256 Succ->removePredecessor(OldTerm->getParent());
2259 IRBuilder<> Builder(OldTerm);
2260 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2262 // Insert an appropriate new terminator.
2263 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2264 if (TrueBB == FalseBB)
2265 // We were only looking for one successor, and it was present.
2266 // Create an unconditional branch to it.
2267 Builder.CreateBr(TrueBB);
2269 // We found both of the successors we were looking for.
2270 // Create a conditional branch sharing the condition of the select.
2271 Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2272 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2273 // Neither of the selected blocks were successors, so this
2274 // terminator must be unreachable.
2275 new UnreachableInst(OldTerm->getContext(), OldTerm);
2277 // One of the selected values was a successor, but the other wasn't.
2278 // Insert an unconditional branch to the one that was found;
2279 // the edge to the one that wasn't must be unreachable.
2281 // Only TrueBB was found.
2282 Builder.CreateBr(TrueBB);
2284 // Only FalseBB was found.
2285 Builder.CreateBr(FalseBB);
2288 EraseTerminatorInstAndDCECond(OldTerm);
2292 // SimplifySwitchOnSelect - Replaces
2293 // (switch (select cond, X, Y)) on constant X, Y
2294 // with a branch - conditional if X and Y lead to distinct BBs,
2295 // unconditional otherwise.
2296 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2297 // Check for constant integer values in the select.
2298 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2299 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2300 if (!TrueVal || !FalseVal)
2303 // Find the relevant condition and destinations.
2304 Value *Condition = Select->getCondition();
2305 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2306 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2308 // Perform the actual simplification.
2309 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB);
2312 // SimplifyIndirectBrOnSelect - Replaces
2313 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2314 // blockaddress(@fn, BlockB)))
2316 // (br cond, BlockA, BlockB).
2317 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2318 // Check that both operands of the select are block addresses.
2319 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2320 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2324 // Extract the actual blocks.
2325 BasicBlock *TrueBB = TBA->getBasicBlock();
2326 BasicBlock *FalseBB = FBA->getBasicBlock();
2328 // Perform the actual simplification.
2329 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB);
2332 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2333 /// instruction (a seteq/setne with a constant) as the only instruction in a
2334 /// block that ends with an uncond branch. We are looking for a very specific
2335 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2336 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2337 /// default value goes to an uncond block with a seteq in it, we get something
2340 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2342 /// %tmp = icmp eq i8 %A, 92
2345 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2347 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2348 /// the PHI, merging the third icmp into the switch.
2349 static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
2350 const TargetData *TD,
2351 IRBuilder<> &Builder) {
2352 BasicBlock *BB = ICI->getParent();
2354 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2356 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2358 Value *V = ICI->getOperand(0);
2359 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2361 // The pattern we're looking for is where our only predecessor is a switch on
2362 // 'V' and this block is the default case for the switch. In this case we can
2363 // fold the compared value into the switch to simplify things.
2364 BasicBlock *Pred = BB->getSinglePredecessor();
2365 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2367 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2368 if (SI->getCondition() != V)
2371 // If BB is reachable on a non-default case, then we simply know the value of
2372 // V in this block. Substitute it and constant fold the icmp instruction
2374 if (SI->getDefaultDest() != BB) {
2375 ConstantInt *VVal = SI->findCaseDest(BB);
2376 assert(VVal && "Should have a unique destination value");
2377 ICI->setOperand(0, VVal);
2379 if (Value *V = SimplifyInstruction(ICI, TD)) {
2380 ICI->replaceAllUsesWith(V);
2381 ICI->eraseFromParent();
2383 // BB is now empty, so it is likely to simplify away.
2384 return SimplifyCFG(BB) | true;
2387 // Ok, the block is reachable from the default dest. If the constant we're
2388 // comparing exists in one of the other edges, then we can constant fold ICI
2390 if (SI->findCaseValue(Cst) != SI->case_default()) {
2392 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2393 V = ConstantInt::getFalse(BB->getContext());
2395 V = ConstantInt::getTrue(BB->getContext());
2397 ICI->replaceAllUsesWith(V);
2398 ICI->eraseFromParent();
2399 // BB is now empty, so it is likely to simplify away.
2400 return SimplifyCFG(BB) | true;
2403 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2405 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2406 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2407 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2408 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2411 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2413 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2414 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2416 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2417 std::swap(DefaultCst, NewCst);
2419 // Replace ICI (which is used by the PHI for the default value) with true or
2420 // false depending on if it is EQ or NE.
2421 ICI->replaceAllUsesWith(DefaultCst);
2422 ICI->eraseFromParent();
2424 // Okay, the switch goes to this block on a default value. Add an edge from
2425 // the switch to the merge point on the compared value.
2426 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2427 BB->getParent(), BB);
2428 SI->addCase(Cst, NewBB);
2430 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2431 Builder.SetInsertPoint(NewBB);
2432 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2433 Builder.CreateBr(SuccBlock);
2434 PHIUse->addIncoming(NewCst, NewBB);
2438 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2439 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2440 /// fold it into a switch instruction if so.
2441 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD,
2442 IRBuilder<> &Builder) {
2443 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2444 if (Cond == 0) return false;
2447 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2448 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2449 // 'setne's and'ed together, collect them.
2451 std::vector<ConstantInt*> Values;
2452 bool TrueWhenEqual = true;
2453 Value *ExtraCase = 0;
2454 unsigned UsedICmps = 0;
2456 if (Cond->getOpcode() == Instruction::Or) {
2457 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2459 } else if (Cond->getOpcode() == Instruction::And) {
2460 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2462 TrueWhenEqual = false;
2465 // If we didn't have a multiply compared value, fail.
2466 if (CompVal == 0) return false;
2468 // Avoid turning single icmps into a switch.
2472 // There might be duplicate constants in the list, which the switch
2473 // instruction can't handle, remove them now.
2474 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2475 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2477 // If Extra was used, we require at least two switch values to do the
2478 // transformation. A switch with one value is just an cond branch.
2479 if (ExtraCase && Values.size() < 2) return false;
2481 // TODO: Preserve branch weight metadata, similarly to how
2482 // FoldValueComparisonIntoPredecessors preserves it.
2484 // Figure out which block is which destination.
2485 BasicBlock *DefaultBB = BI->getSuccessor(1);
2486 BasicBlock *EdgeBB = BI->getSuccessor(0);
2487 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2489 BasicBlock *BB = BI->getParent();
2491 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2492 << " cases into SWITCH. BB is:\n" << *BB);
2494 // If there are any extra values that couldn't be folded into the switch
2495 // then we evaluate them with an explicit branch first. Split the block
2496 // right before the condbr to handle it.
2498 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2499 // Remove the uncond branch added to the old block.
2500 TerminatorInst *OldTI = BB->getTerminator();
2501 Builder.SetInsertPoint(OldTI);
2504 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2506 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2508 OldTI->eraseFromParent();
2510 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2511 // for the edge we just added.
2512 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2514 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2515 << "\nEXTRABB = " << *BB);
2519 Builder.SetInsertPoint(BI);
2520 // Convert pointer to int before we switch.
2521 if (CompVal->getType()->isPointerTy()) {
2522 assert(TD && "Cannot switch on pointer without TargetData");
2523 CompVal = Builder.CreatePtrToInt(CompVal,
2524 TD->getIntPtrType(CompVal->getContext()),
2528 // Create the new switch instruction now.
2529 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2531 // Add all of the 'cases' to the switch instruction.
2532 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2533 New->addCase(Values[i], EdgeBB);
2535 // We added edges from PI to the EdgeBB. As such, if there were any
2536 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2537 // the number of edges added.
2538 for (BasicBlock::iterator BBI = EdgeBB->begin();
2539 isa<PHINode>(BBI); ++BBI) {
2540 PHINode *PN = cast<PHINode>(BBI);
2541 Value *InVal = PN->getIncomingValueForBlock(BB);
2542 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2543 PN->addIncoming(InVal, BB);
2546 // Erase the old branch instruction.
2547 EraseTerminatorInstAndDCECond(BI);
2549 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2553 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2554 // If this is a trivial landing pad that just continues unwinding the caught
2555 // exception then zap the landing pad, turning its invokes into calls.
2556 BasicBlock *BB = RI->getParent();
2557 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2558 if (RI->getValue() != LPInst)
2559 // Not a landing pad, or the resume is not unwinding the exception that
2560 // caused control to branch here.
2563 // Check that there are no other instructions except for debug intrinsics.
2564 BasicBlock::iterator I = LPInst, E = RI;
2566 if (!isa<DbgInfoIntrinsic>(I))
2569 // Turn all invokes that unwind here into calls and delete the basic block.
2570 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2571 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2572 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2573 // Insert a call instruction before the invoke.
2574 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2576 Call->setCallingConv(II->getCallingConv());
2577 Call->setAttributes(II->getAttributes());
2578 Call->setDebugLoc(II->getDebugLoc());
2580 // Anything that used the value produced by the invoke instruction now uses
2581 // the value produced by the call instruction. Note that we do this even
2582 // for void functions and calls with no uses so that the callgraph edge is
2584 II->replaceAllUsesWith(Call);
2585 BB->removePredecessor(II->getParent());
2587 // Insert a branch to the normal destination right before the invoke.
2588 BranchInst::Create(II->getNormalDest(), II);
2590 // Finally, delete the invoke instruction!
2591 II->eraseFromParent();
2594 // The landingpad is now unreachable. Zap it.
2595 BB->eraseFromParent();
2599 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2600 BasicBlock *BB = RI->getParent();
2601 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2603 // Find predecessors that end with branches.
2604 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2605 SmallVector<BranchInst*, 8> CondBranchPreds;
2606 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2607 BasicBlock *P = *PI;
2608 TerminatorInst *PTI = P->getTerminator();
2609 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2610 if (BI->isUnconditional())
2611 UncondBranchPreds.push_back(P);
2613 CondBranchPreds.push_back(BI);
2617 // If we found some, do the transformation!
2618 if (!UncondBranchPreds.empty() && DupRet) {
2619 while (!UncondBranchPreds.empty()) {
2620 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2621 DEBUG(dbgs() << "FOLDING: " << *BB
2622 << "INTO UNCOND BRANCH PRED: " << *Pred);
2623 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2626 // If we eliminated all predecessors of the block, delete the block now.
2627 if (pred_begin(BB) == pred_end(BB))
2628 // We know there are no successors, so just nuke the block.
2629 BB->eraseFromParent();
2634 // Check out all of the conditional branches going to this return
2635 // instruction. If any of them just select between returns, change the
2636 // branch itself into a select/return pair.
2637 while (!CondBranchPreds.empty()) {
2638 BranchInst *BI = CondBranchPreds.pop_back_val();
2640 // Check to see if the non-BB successor is also a return block.
2641 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2642 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2643 SimplifyCondBranchToTwoReturns(BI, Builder))
2649 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2650 BasicBlock *BB = UI->getParent();
2652 bool Changed = false;
2654 // If there are any instructions immediately before the unreachable that can
2655 // be removed, do so.
2656 while (UI != BB->begin()) {
2657 BasicBlock::iterator BBI = UI;
2659 // Do not delete instructions that can have side effects which might cause
2660 // the unreachable to not be reachable; specifically, calls and volatile
2661 // operations may have this effect.
2662 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2664 if (BBI->mayHaveSideEffects()) {
2665 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2666 if (SI->isVolatile())
2668 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2669 if (LI->isVolatile())
2671 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2672 if (RMWI->isVolatile())
2674 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2675 if (CXI->isVolatile())
2677 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2678 !isa<LandingPadInst>(BBI)) {
2681 // Note that deleting LandingPad's here is in fact okay, although it
2682 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2683 // all the predecessors of this block will be the unwind edges of Invokes,
2684 // and we can therefore guarantee this block will be erased.
2687 // Delete this instruction (any uses are guaranteed to be dead)
2688 if (!BBI->use_empty())
2689 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2690 BBI->eraseFromParent();
2694 // If the unreachable instruction is the first in the block, take a gander
2695 // at all of the predecessors of this instruction, and simplify them.
2696 if (&BB->front() != UI) return Changed;
2698 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2699 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
2700 TerminatorInst *TI = Preds[i]->getTerminator();
2701 IRBuilder<> Builder(TI);
2702 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2703 if (BI->isUnconditional()) {
2704 if (BI->getSuccessor(0) == BB) {
2705 new UnreachableInst(TI->getContext(), TI);
2706 TI->eraseFromParent();
2710 if (BI->getSuccessor(0) == BB) {
2711 Builder.CreateBr(BI->getSuccessor(1));
2712 EraseTerminatorInstAndDCECond(BI);
2713 } else if (BI->getSuccessor(1) == BB) {
2714 Builder.CreateBr(BI->getSuccessor(0));
2715 EraseTerminatorInstAndDCECond(BI);
2719 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2720 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2722 if (i.getCaseSuccessor() == BB) {
2723 BB->removePredecessor(SI->getParent());
2728 // If the default value is unreachable, figure out the most popular
2729 // destination and make it the default.
2730 if (SI->getDefaultDest() == BB) {
2731 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
2732 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2734 std::pair<unsigned, unsigned> &entry =
2735 Popularity[i.getCaseSuccessor()];
2736 if (entry.first == 0) {
2738 entry.second = i.getCaseIndex();
2744 // Find the most popular block.
2745 unsigned MaxPop = 0;
2746 unsigned MaxIndex = 0;
2747 BasicBlock *MaxBlock = 0;
2748 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
2749 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
2750 if (I->second.first > MaxPop ||
2751 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
2752 MaxPop = I->second.first;
2753 MaxIndex = I->second.second;
2754 MaxBlock = I->first;
2758 // Make this the new default, allowing us to delete any explicit
2760 SI->setDefaultDest(MaxBlock);
2763 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
2765 if (isa<PHINode>(MaxBlock->begin()))
2766 for (unsigned i = 0; i != MaxPop-1; ++i)
2767 MaxBlock->removePredecessor(SI->getParent());
2769 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2771 if (i.getCaseSuccessor() == MaxBlock) {
2777 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
2778 if (II->getUnwindDest() == BB) {
2779 // Convert the invoke to a call instruction. This would be a good
2780 // place to note that the call does not throw though.
2781 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
2782 II->removeFromParent(); // Take out of symbol table
2784 // Insert the call now...
2785 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
2786 Builder.SetInsertPoint(BI);
2787 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
2788 Args, II->getName());
2789 CI->setCallingConv(II->getCallingConv());
2790 CI->setAttributes(II->getAttributes());
2791 // If the invoke produced a value, the call does now instead.
2792 II->replaceAllUsesWith(CI);
2799 // If this block is now dead, remove it.
2800 if (pred_begin(BB) == pred_end(BB) &&
2801 BB != &BB->getParent()->getEntryBlock()) {
2802 // We know there are no successors, so just nuke the block.
2803 BB->eraseFromParent();
2810 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
2811 /// integer range comparison into a sub, an icmp and a branch.
2812 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
2813 assert(SI->getNumCases() > 1 && "Degenerate switch?");
2815 // Make sure all cases point to the same destination and gather the values.
2816 SmallVector<ConstantInt *, 16> Cases;
2817 SwitchInst::CaseIt I = SI->case_begin();
2818 Cases.push_back(I.getCaseValue());
2819 SwitchInst::CaseIt PrevI = I++;
2820 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
2821 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
2823 Cases.push_back(I.getCaseValue());
2825 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
2827 // Sort the case values, then check if they form a range we can transform.
2828 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
2829 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
2830 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
2834 Constant *Offset = ConstantExpr::getNeg(Cases.back());
2835 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
2837 Value *Sub = SI->getCondition();
2838 if (!Offset->isNullValue())
2839 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
2840 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
2841 Builder.CreateCondBr(
2842 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
2844 // Prune obsolete incoming values off the successor's PHI nodes.
2845 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
2846 isa<PHINode>(BBI); ++BBI) {
2847 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
2848 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
2850 SI->eraseFromParent();
2855 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
2856 /// and use it to remove dead cases.
2857 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
2858 Value *Cond = SI->getCondition();
2859 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
2860 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
2861 ComputeMaskedBits(Cond, KnownZero, KnownOne);
2863 // Gather dead cases.
2864 SmallVector<ConstantInt*, 8> DeadCases;
2865 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
2866 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
2867 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
2868 DeadCases.push_back(I.getCaseValue());
2869 DEBUG(dbgs() << "SimplifyCFG: switch case '"
2870 << I.getCaseValue() << "' is dead.\n");
2874 // Remove dead cases from the switch.
2875 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
2876 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
2877 assert(Case != SI->case_default() &&
2878 "Case was not found. Probably mistake in DeadCases forming.");
2879 // Prune unused values from PHI nodes.
2880 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
2881 SI->removeCase(Case);
2884 return !DeadCases.empty();
2887 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
2888 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
2889 /// by an unconditional branch), look at the phi node for BB in the successor
2890 /// block and see if the incoming value is equal to CaseValue. If so, return
2891 /// the phi node, and set PhiIndex to BB's index in the phi node.
2892 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
2895 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
2896 return NULL; // BB must be empty to be a candidate for simplification.
2897 if (!BB->getSinglePredecessor())
2898 return NULL; // BB must be dominated by the switch.
2900 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
2901 if (!Branch || !Branch->isUnconditional())
2902 return NULL; // Terminator must be unconditional branch.
2904 BasicBlock *Succ = Branch->getSuccessor(0);
2906 BasicBlock::iterator I = Succ->begin();
2907 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
2908 int Idx = PHI->getBasicBlockIndex(BB);
2909 assert(Idx >= 0 && "PHI has no entry for predecessor?");
2911 Value *InValue = PHI->getIncomingValue(Idx);
2912 if (InValue != CaseValue) continue;
2921 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
2922 /// instruction to a phi node dominated by the switch, if that would mean that
2923 /// some of the destination blocks of the switch can be folded away.
2924 /// Returns true if a change is made.
2925 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
2926 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
2927 ForwardingNodesMap ForwardingNodes;
2929 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
2930 ConstantInt *CaseValue = I.getCaseValue();
2931 BasicBlock *CaseDest = I.getCaseSuccessor();
2934 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
2938 ForwardingNodes[PHI].push_back(PhiIndex);
2941 bool Changed = false;
2943 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
2944 E = ForwardingNodes.end(); I != E; ++I) {
2945 PHINode *Phi = I->first;
2946 SmallVector<int,4> &Indexes = I->second;
2948 if (Indexes.size() < 2) continue;
2950 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
2951 Phi->setIncomingValue(Indexes[I], SI->getCondition());
2958 /// ValidLookupTableConstant - Return true if the backend will be able to handle
2959 /// initializing an array of constants like C.
2960 static bool ValidLookupTableConstant(Constant *C) {
2961 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2962 return CE->isGEPWithNoNotionalOverIndexing();
2964 return isa<ConstantFP>(C) ||
2965 isa<ConstantInt>(C) ||
2966 isa<ConstantPointerNull>(C) ||
2967 isa<GlobalValue>(C) ||
2971 /// GetCaseResulsts - Try to determine the resulting constant values in phi
2972 /// nodes at the common destination basic block for one of the case
2973 /// destinations of a switch instruction.
2974 static bool GetCaseResults(SwitchInst *SI,
2975 BasicBlock *CaseDest,
2976 BasicBlock **CommonDest,
2977 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) {
2978 // The block from which we enter the common destination.
2979 BasicBlock *Pred = SI->getParent();
2981 // If CaseDest is empty, continue to its successor.
2982 if (CaseDest->getFirstNonPHIOrDbg() == CaseDest->getTerminator() &&
2983 !isa<PHINode>(CaseDest->begin())) {
2985 TerminatorInst *Terminator = CaseDest->getTerminator();
2986 if (Terminator->getNumSuccessors() != 1)
2990 CaseDest = Terminator->getSuccessor(0);
2993 // If we did not have a CommonDest before, use the current one.
2995 *CommonDest = CaseDest;
2996 // If the destination isn't the common one, abort.
2997 if (CaseDest != *CommonDest)
3000 // Get the values for this case from phi nodes in the destination block.
3001 BasicBlock::iterator I = (*CommonDest)->begin();
3002 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3003 int Idx = PHI->getBasicBlockIndex(Pred);
3007 Constant *ConstVal = dyn_cast<Constant>(PHI->getIncomingValue(Idx));
3011 // Be conservative about which kinds of constants we support.
3012 if (!ValidLookupTableConstant(ConstVal))
3015 Res.push_back(std::make_pair(PHI, ConstVal));
3021 /// BuildLookupTable - Build a lookup table with the contents of Results, using
3022 /// DefaultResult to fill the holes in the table. If the table ends up
3023 /// containing the same result in each element, set *SingleResult to that value
3024 /// and return NULL.
3025 static GlobalVariable *BuildLookupTable(Module &M,
3027 ConstantInt *Offset,
3028 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Results,
3029 Constant *DefaultResult,
3030 Constant **SingleResult) {
3031 assert(Results.size() && "Need values to build lookup table");
3032 assert(TableSize >= Results.size() && "Table needs to hold all values");
3034 // If all values in the table are equal, this is that value.
3035 Constant *SameResult = Results.begin()->second;
3037 // Build up the table contents.
3038 std::vector<Constant*> TableContents(TableSize);
3039 for (size_t I = 0, E = Results.size(); I != E; ++I) {
3040 ConstantInt *CaseVal = Results[I].first;
3041 Constant *CaseRes = Results[I].second;
3043 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
3044 TableContents[Idx] = CaseRes;
3046 if (CaseRes != SameResult)
3050 // Fill in any holes in the table with the default result.
3051 if (Results.size() < TableSize) {
3052 for (unsigned i = 0; i < TableSize; ++i) {
3053 if (!TableContents[i])
3054 TableContents[i] = DefaultResult;
3057 if (DefaultResult != SameResult)
3061 // Same result was used in the entire table; just return that.
3063 *SingleResult = SameResult;
3067 ArrayType *ArrayTy = ArrayType::get(DefaultResult->getType(), TableSize);
3068 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3070 GlobalVariable *GV = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3071 GlobalVariable::PrivateLinkage,
3074 GV->setUnnamedAddr(true);
3078 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3079 /// phi nodes in a common successor block with different constant values,
3080 /// replace the switch with lookup tables.
3081 static bool SwitchToLookupTable(SwitchInst *SI,
3082 IRBuilder<> &Builder) {
3083 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3084 // FIXME: Handle unreachable cases.
3086 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3087 // split off a dense part and build a lookup table for that.
3089 // FIXME: If the results are all integers and the lookup table would fit in a
3090 // target-legal register, we should store them as a bitmap and use shift/mask
3091 // to look up the result.
3093 // FIXME: This creates arrays of GEPs to constant strings, which means each
3094 // GEP needs a runtime relocation in PIC code. We should just build one big
3095 // string and lookup indices into that.
3097 // Ignore the switch if the number of cases are too small.
3098 // This is similar to the check when building jump tables in
3099 // SelectionDAGBuilder::handleJTSwitchCase.
3100 // FIXME: Determine the best cut-off.
3101 if (SI->getNumCases() < 4)
3104 // Figure out the corresponding result for each case value and phi node in the
3105 // common destination, as well as the the min and max case values.
3106 assert(SI->case_begin() != SI->case_end());
3107 SwitchInst::CaseIt CI = SI->case_begin();
3108 ConstantInt *MinCaseVal = CI.getCaseValue();
3109 ConstantInt *MaxCaseVal = CI.getCaseValue();
3111 BasicBlock *CommonDest = NULL;
3112 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3113 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3114 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3115 SmallDenseMap<PHINode*, Type*> ResultTypes;
3116 SmallVector<PHINode*, 4> PHIs;
3118 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3119 ConstantInt *CaseVal = CI.getCaseValue();
3120 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3121 MinCaseVal = CaseVal;
3122 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3123 MaxCaseVal = CaseVal;
3125 // Resulting value at phi nodes for this case value.
3126 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3128 if (!GetCaseResults(SI, CI.getCaseSuccessor(), &CommonDest, Results))
3131 // Append the result from this case to the list for each phi.
3132 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3133 if (!ResultLists.count(I->first))
3134 PHIs.push_back(I->first);
3135 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3139 // Get the resulting values for the default case.
3140 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3141 if (!GetCaseResults(SI, SI->getDefaultDest(), &CommonDest, DefaultResultsList))
3143 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3144 PHINode *PHI = DefaultResultsList[I].first;
3145 Constant *Result = DefaultResultsList[I].second;
3146 DefaultResults[PHI] = Result;
3147 ResultTypes[PHI] = Result->getType();
3150 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3151 // The table density should be at lest 40%. This is the same criterion as for
3152 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3153 // FIXME: Find the best cut-off.
3154 // Be careful to avoid overlow in the density computation.
3155 if (RangeSpread.zextOrSelf(64).ugt(UINT64_MAX / 4 - 1))
3157 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3158 if (SI->getNumCases() * 10 < TableSize * 4)
3161 // Build the lookup tables.
3162 SmallDenseMap<PHINode*, GlobalVariable*> LookupTables;
3163 SmallDenseMap<PHINode*, Constant*> SingleResults;
3165 Module &Mod = *CommonDest->getParent()->getParent();
3166 for (SmallVector<PHINode*, 4>::iterator I = PHIs.begin(), E = PHIs.end();
3170 Constant *SingleResult = NULL;
3171 LookupTables[PHI] = BuildLookupTable(Mod, TableSize, MinCaseVal,
3172 ResultLists[PHI], DefaultResults[PHI],
3174 SingleResults[PHI] = SingleResult;
3177 // Create the BB that does the lookups.
3178 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3180 CommonDest->getParent(),
3183 // Check whether the condition value is within the case range, and branch to
3185 Builder.SetInsertPoint(SI);
3186 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3188 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3189 MinCaseVal->getType(), TableSize));
3190 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3192 // Populate the BB that does the lookups.
3193 Builder.SetInsertPoint(LookupBB);
3194 bool ReturnedEarly = false;
3195 for (SmallVector<PHINode*, 4>::iterator I = PHIs.begin(), E = PHIs.end();
3198 // There was a single result for this phi; just use that.
3199 if (Constant *SingleResult = SingleResults[PHI]) {
3200 PHI->addIncoming(SingleResult, LookupBB);
3204 Value *GEPIndices[] = { Builder.getInt32(0), TableIndex };
3205 Value *GEP = Builder.CreateInBoundsGEP(LookupTables[PHI], GEPIndices,
3207 Value *Result = Builder.CreateLoad(GEP, "switch.load");
3209 // If the result is only going to be used to return from the function,
3210 // we want to do that right here.
3211 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin())) {
3212 if (CommonDest->getFirstNonPHIOrDbg() == CommonDest->getTerminator()) {
3213 Builder.CreateRet(Result);
3214 ReturnedEarly = true;
3219 PHI->addIncoming(Result, LookupBB);
3223 Builder.CreateBr(CommonDest);
3225 // Remove the switch.
3226 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3227 BasicBlock *Succ = SI->getSuccessor(i);
3228 if (Succ == SI->getDefaultDest()) continue;
3229 Succ->removePredecessor(SI->getParent());
3231 SI->eraseFromParent();
3237 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3238 // If this switch is too complex to want to look at, ignore it.
3239 if (!isValueEqualityComparison(SI))
3242 BasicBlock *BB = SI->getParent();
3244 // If we only have one predecessor, and if it is a branch on this value,
3245 // see if that predecessor totally determines the outcome of this switch.
3246 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3247 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3248 return SimplifyCFG(BB) | true;
3250 Value *Cond = SI->getCondition();
3251 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3252 if (SimplifySwitchOnSelect(SI, Select))
3253 return SimplifyCFG(BB) | true;
3255 // If the block only contains the switch, see if we can fold the block
3256 // away into any preds.
3257 BasicBlock::iterator BBI = BB->begin();
3258 // Ignore dbg intrinsics.
3259 while (isa<DbgInfoIntrinsic>(BBI))
3262 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3263 return SimplifyCFG(BB) | true;
3265 // Try to transform the switch into an icmp and a branch.
3266 if (TurnSwitchRangeIntoICmp(SI, Builder))
3267 return SimplifyCFG(BB) | true;
3269 // Remove unreachable cases.
3270 if (EliminateDeadSwitchCases(SI))
3271 return SimplifyCFG(BB) | true;
3273 if (ForwardSwitchConditionToPHI(SI))
3274 return SimplifyCFG(BB) | true;
3276 if (SwitchToLookupTable(SI, Builder))
3277 return SimplifyCFG(BB) | true;
3282 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3283 BasicBlock *BB = IBI->getParent();
3284 bool Changed = false;
3286 // Eliminate redundant destinations.
3287 SmallPtrSet<Value *, 8> Succs;
3288 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3289 BasicBlock *Dest = IBI->getDestination(i);
3290 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3291 Dest->removePredecessor(BB);
3292 IBI->removeDestination(i);
3298 if (IBI->getNumDestinations() == 0) {
3299 // If the indirectbr has no successors, change it to unreachable.
3300 new UnreachableInst(IBI->getContext(), IBI);
3301 EraseTerminatorInstAndDCECond(IBI);
3305 if (IBI->getNumDestinations() == 1) {
3306 // If the indirectbr has one successor, change it to a direct branch.
3307 BranchInst::Create(IBI->getDestination(0), IBI);
3308 EraseTerminatorInstAndDCECond(IBI);
3312 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3313 if (SimplifyIndirectBrOnSelect(IBI, SI))
3314 return SimplifyCFG(BB) | true;
3319 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3320 BasicBlock *BB = BI->getParent();
3322 // If the Terminator is the only non-phi instruction, simplify the block.
3323 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3324 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3325 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3328 // If the only instruction in the block is a seteq/setne comparison
3329 // against a constant, try to simplify the block.
3330 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3331 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3332 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3334 if (I->isTerminator() &&
3335 TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder))
3339 // If this basic block is ONLY a compare and a branch, and if a predecessor
3340 // branches to us and our successor, fold the comparison into the
3341 // predecessor and use logical operations to update the incoming value
3342 // for PHI nodes in common successor.
3343 if (FoldBranchToCommonDest(BI))
3344 return SimplifyCFG(BB) | true;
3349 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3350 BasicBlock *BB = BI->getParent();
3352 // Conditional branch
3353 if (isValueEqualityComparison(BI)) {
3354 // If we only have one predecessor, and if it is a branch on this value,
3355 // see if that predecessor totally determines the outcome of this
3357 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3358 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3359 return SimplifyCFG(BB) | true;
3361 // This block must be empty, except for the setcond inst, if it exists.
3362 // Ignore dbg intrinsics.
3363 BasicBlock::iterator I = BB->begin();
3364 // Ignore dbg intrinsics.
3365 while (isa<DbgInfoIntrinsic>(I))
3368 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3369 return SimplifyCFG(BB) | true;
3370 } else if (&*I == cast<Instruction>(BI->getCondition())){
3372 // Ignore dbg intrinsics.
3373 while (isa<DbgInfoIntrinsic>(I))
3375 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3376 return SimplifyCFG(BB) | true;
3380 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3381 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3384 // If this basic block is ONLY a compare and a branch, and if a predecessor
3385 // branches to us and one of our successors, fold the comparison into the
3386 // predecessor and use logical operations to pick the right destination.
3387 if (FoldBranchToCommonDest(BI))
3388 return SimplifyCFG(BB) | true;
3390 // We have a conditional branch to two blocks that are only reachable
3391 // from BI. We know that the condbr dominates the two blocks, so see if
3392 // there is any identical code in the "then" and "else" blocks. If so, we
3393 // can hoist it up to the branching block.
3394 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3395 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3396 if (HoistThenElseCodeToIf(BI))
3397 return SimplifyCFG(BB) | true;
3399 // If Successor #1 has multiple preds, we may be able to conditionally
3400 // execute Successor #0 if it branches to successor #1.
3401 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3402 if (Succ0TI->getNumSuccessors() == 1 &&
3403 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3404 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
3405 return SimplifyCFG(BB) | true;
3407 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3408 // If Successor #0 has multiple preds, we may be able to conditionally
3409 // execute Successor #1 if it branches to successor #0.
3410 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3411 if (Succ1TI->getNumSuccessors() == 1 &&
3412 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
3413 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
3414 return SimplifyCFG(BB) | true;
3417 // If this is a branch on a phi node in the current block, thread control
3418 // through this block if any PHI node entries are constants.
3419 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
3420 if (PN->getParent() == BI->getParent())
3421 if (FoldCondBranchOnPHI(BI, TD))
3422 return SimplifyCFG(BB) | true;
3424 // Scan predecessor blocks for conditional branches.
3425 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
3426 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
3427 if (PBI != BI && PBI->isConditional())
3428 if (SimplifyCondBranchToCondBranch(PBI, BI))
3429 return SimplifyCFG(BB) | true;
3434 /// Check if passing a value to an instruction will cause undefined behavior.
3435 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
3436 Constant *C = dyn_cast<Constant>(V);
3440 if (!I->hasOneUse()) // Only look at single-use instructions, for compile time
3443 if (C->isNullValue()) {
3444 Instruction *Use = I->use_back();
3446 // Now make sure that there are no instructions in between that can alter
3447 // control flow (eg. calls)
3448 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
3449 if (i == I->getParent()->end() || i->mayHaveSideEffects())
3452 // Look through GEPs. A load from a GEP derived from NULL is still undefined
3453 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
3454 if (GEP->getPointerOperand() == I)
3455 return passingValueIsAlwaysUndefined(V, GEP);
3457 // Look through bitcasts.
3458 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
3459 return passingValueIsAlwaysUndefined(V, BC);
3461 // Load from null is undefined.
3462 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
3463 return LI->getPointerAddressSpace() == 0;
3465 // Store to null is undefined.
3466 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
3467 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
3472 /// If BB has an incoming value that will always trigger undefined behavior
3473 /// (eg. null pointer dereference), remove the branch leading here.
3474 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
3475 for (BasicBlock::iterator i = BB->begin();
3476 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
3477 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
3478 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
3479 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
3480 IRBuilder<> Builder(T);
3481 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
3482 BB->removePredecessor(PHI->getIncomingBlock(i));
3483 // Turn uncoditional branches into unreachables and remove the dead
3484 // destination from conditional branches.
3485 if (BI->isUnconditional())
3486 Builder.CreateUnreachable();
3488 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
3489 BI->getSuccessor(0));
3490 BI->eraseFromParent();
3493 // TODO: SwitchInst.
3499 bool SimplifyCFGOpt::run(BasicBlock *BB) {
3500 bool Changed = false;
3502 assert(BB && BB->getParent() && "Block not embedded in function!");
3503 assert(BB->getTerminator() && "Degenerate basic block encountered!");
3505 // Remove basic blocks that have no predecessors (except the entry block)...
3506 // or that just have themself as a predecessor. These are unreachable.
3507 if ((pred_begin(BB) == pred_end(BB) &&
3508 BB != &BB->getParent()->getEntryBlock()) ||
3509 BB->getSinglePredecessor() == BB) {
3510 DEBUG(dbgs() << "Removing BB: \n" << *BB);
3511 DeleteDeadBlock(BB);
3515 // Check to see if we can constant propagate this terminator instruction
3517 Changed |= ConstantFoldTerminator(BB, true);
3519 // Check for and eliminate duplicate PHI nodes in this block.
3520 Changed |= EliminateDuplicatePHINodes(BB);
3522 // Check for and remove branches that will always cause undefined behavior.
3523 Changed |= removeUndefIntroducingPredecessor(BB);
3525 // Merge basic blocks into their predecessor if there is only one distinct
3526 // pred, and if there is only one distinct successor of the predecessor, and
3527 // if there are no PHI nodes.
3529 if (MergeBlockIntoPredecessor(BB))
3532 IRBuilder<> Builder(BB);
3534 // If there is a trivial two-entry PHI node in this basic block, and we can
3535 // eliminate it, do so now.
3536 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
3537 if (PN->getNumIncomingValues() == 2)
3538 Changed |= FoldTwoEntryPHINode(PN, TD);
3540 Builder.SetInsertPoint(BB->getTerminator());
3541 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
3542 if (BI->isUnconditional()) {
3543 if (SimplifyUncondBranch(BI, Builder)) return true;
3545 if (SimplifyCondBranch(BI, Builder)) return true;
3547 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
3548 if (SimplifyReturn(RI, Builder)) return true;
3549 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
3550 if (SimplifyResume(RI, Builder)) return true;
3551 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
3552 if (SimplifySwitch(SI, Builder)) return true;
3553 } else if (UnreachableInst *UI =
3554 dyn_cast<UnreachableInst>(BB->getTerminator())) {
3555 if (SimplifyUnreachable(UI)) return true;
3556 } else if (IndirectBrInst *IBI =
3557 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
3558 if (SimplifyIndirectBr(IBI)) return true;
3564 /// SimplifyCFG - This function is used to do simplification of a CFG. For
3565 /// example, it adjusts branches to branches to eliminate the extra hop, it
3566 /// eliminates unreachable basic blocks, and does other "peephole" optimization
3567 /// of the CFG. It returns true if a modification was made.
3569 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) {
3570 return SimplifyCFGOpt(TD).run(BB);