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);
834 GetBranchWeights(PTI, Weights);
835 else if (SuccHasWeights)
836 // If there are no predecessor weights but there are successor weights,
837 // populate Weights with 1, which will later be scaled to the sum of
838 // successor's weights
839 Weights.assign(1 + PredCases.size(), 1);
841 SmallVector<uint64_t, 8> SuccWeights;
843 GetBranchWeights(TI, SuccWeights);
844 else if (PredHasWeights)
845 SuccWeights.assign(1 + BBCases.size(), 1);
847 if (PredDefault == BB) {
848 // If this is the default destination from PTI, only the edges in TI
849 // that don't occur in PTI, or that branch to BB will be activated.
850 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
851 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
852 if (PredCases[i].Dest != BB)
853 PTIHandled.insert(PredCases[i].Value);
855 // The default destination is BB, we don't need explicit targets.
856 std::swap(PredCases[i], PredCases.back());
858 if (PredHasWeights || SuccHasWeights) {
859 // Increase weight for the default case.
860 Weights[0] += Weights[i+1];
861 std::swap(Weights[i+1], Weights.back());
865 PredCases.pop_back();
869 // Reconstruct the new switch statement we will be building.
870 if (PredDefault != BBDefault) {
871 PredDefault->removePredecessor(Pred);
872 PredDefault = BBDefault;
873 NewSuccessors.push_back(BBDefault);
876 unsigned CasesFromPred = Weights.size();
877 uint64_t ValidTotalSuccWeight = 0;
878 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
879 if (!PTIHandled.count(BBCases[i].Value) &&
880 BBCases[i].Dest != BBDefault) {
881 PredCases.push_back(BBCases[i]);
882 NewSuccessors.push_back(BBCases[i].Dest);
883 if (SuccHasWeights || PredHasWeights) {
884 // The default weight is at index 0, so weight for the ith case
885 // should be at index i+1. Scale the cases from successor by
886 // PredDefaultWeight (Weights[0]).
887 Weights.push_back(Weights[0] * SuccWeights[i+1]);
888 ValidTotalSuccWeight += SuccWeights[i+1];
892 if (SuccHasWeights || PredHasWeights) {
893 ValidTotalSuccWeight += SuccWeights[0];
894 // Scale the cases from predecessor by ValidTotalSuccWeight.
895 for (unsigned i = 1; i < CasesFromPred; ++i)
896 Weights[i] *= ValidTotalSuccWeight;
897 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
898 Weights[0] *= SuccWeights[0];
901 // FIXME: preserve branch weight metadata, similarly to the 'then'
902 // above. For now, drop it.
903 PredHasWeights = false;
904 SuccHasWeights = false;
906 // If this is not the default destination from PSI, only the edges
907 // in SI that occur in PSI with a destination of BB will be
909 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
910 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
911 if (PredCases[i].Dest == BB) {
912 PTIHandled.insert(PredCases[i].Value);
913 std::swap(PredCases[i], PredCases.back());
914 PredCases.pop_back();
918 // Okay, now we know which constants were sent to BB from the
919 // predecessor. Figure out where they will all go now.
920 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
921 if (PTIHandled.count(BBCases[i].Value)) {
922 // If this is one we are capable of getting...
923 PredCases.push_back(BBCases[i]);
924 NewSuccessors.push_back(BBCases[i].Dest);
925 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
928 // If there are any constants vectored to BB that TI doesn't handle,
929 // they must go to the default destination of TI.
930 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
932 E = PTIHandled.end(); I != E; ++I) {
933 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
934 NewSuccessors.push_back(BBDefault);
938 // Okay, at this point, we know which new successor Pred will get. Make
939 // sure we update the number of entries in the PHI nodes for these
941 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
942 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
944 Builder.SetInsertPoint(PTI);
945 // Convert pointer to int before we switch.
946 if (CV->getType()->isPointerTy()) {
947 assert(TD && "Cannot switch on pointer without TargetData");
948 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()),
952 // Now that the successors are updated, create the new Switch instruction.
953 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
955 NewSI->setDebugLoc(PTI->getDebugLoc());
956 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
957 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
959 if (PredHasWeights || SuccHasWeights) {
960 // Halve the weights if any of them cannot fit in an uint32_t
963 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
965 NewSI->setMetadata(LLVMContext::MD_prof,
966 MDBuilder(BB->getContext()).
967 createBranchWeights(MDWeights));
970 EraseTerminatorInstAndDCECond(PTI);
972 // Okay, last check. If BB is still a successor of PSI, then we must
973 // have an infinite loop case. If so, add an infinitely looping block
974 // to handle the case to preserve the behavior of the code.
975 BasicBlock *InfLoopBlock = 0;
976 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
977 if (NewSI->getSuccessor(i) == BB) {
978 if (InfLoopBlock == 0) {
979 // Insert it at the end of the function, because it's either code,
980 // or it won't matter if it's hot. :)
981 InfLoopBlock = BasicBlock::Create(BB->getContext(),
982 "infloop", BB->getParent());
983 BranchInst::Create(InfLoopBlock, InfLoopBlock);
985 NewSI->setSuccessor(i, InfLoopBlock);
994 // isSafeToHoistInvoke - If we would need to insert a select that uses the
995 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
996 // would need to do this), we can't hoist the invoke, as there is nowhere
997 // to put the select in this case.
998 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
999 Instruction *I1, Instruction *I2) {
1000 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1002 for (BasicBlock::iterator BBI = SI->begin();
1003 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1004 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1005 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1006 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1014 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1015 /// BB2, hoist any common code in the two blocks up into the branch block. The
1016 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1017 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1018 // This does very trivial matching, with limited scanning, to find identical
1019 // instructions in the two blocks. In particular, we don't want to get into
1020 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1021 // such, we currently just scan for obviously identical instructions in an
1023 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1024 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1026 BasicBlock::iterator BB1_Itr = BB1->begin();
1027 BasicBlock::iterator BB2_Itr = BB2->begin();
1029 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1030 // Skip debug info if it is not identical.
1031 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1032 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1033 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1034 while (isa<DbgInfoIntrinsic>(I1))
1036 while (isa<DbgInfoIntrinsic>(I2))
1039 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1040 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1043 // If we get here, we can hoist at least one instruction.
1044 BasicBlock *BIParent = BI->getParent();
1047 // If we are hoisting the terminator instruction, don't move one (making a
1048 // broken BB), instead clone it, and remove BI.
1049 if (isa<TerminatorInst>(I1))
1050 goto HoistTerminator;
1052 // For a normal instruction, we just move one to right before the branch,
1053 // then replace all uses of the other with the first. Finally, we remove
1054 // the now redundant second instruction.
1055 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1056 if (!I2->use_empty())
1057 I2->replaceAllUsesWith(I1);
1058 I1->intersectOptionalDataWith(I2);
1059 I2->eraseFromParent();
1063 // Skip debug info if it is not identical.
1064 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1065 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1066 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1067 while (isa<DbgInfoIntrinsic>(I1))
1069 while (isa<DbgInfoIntrinsic>(I2))
1072 } while (I1->isIdenticalToWhenDefined(I2));
1077 // It may not be possible to hoist an invoke.
1078 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1081 // Okay, it is safe to hoist the terminator.
1082 Instruction *NT = I1->clone();
1083 BIParent->getInstList().insert(BI, NT);
1084 if (!NT->getType()->isVoidTy()) {
1085 I1->replaceAllUsesWith(NT);
1086 I2->replaceAllUsesWith(NT);
1090 IRBuilder<true, NoFolder> Builder(NT);
1091 // Hoisting one of the terminators from our successor is a great thing.
1092 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1093 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1094 // nodes, so we insert select instruction to compute the final result.
1095 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1096 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1098 for (BasicBlock::iterator BBI = SI->begin();
1099 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1100 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1101 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1102 if (BB1V == BB2V) continue;
1104 // These values do not agree. Insert a select instruction before NT
1105 // that determines the right value.
1106 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1108 SI = cast<SelectInst>
1109 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1110 BB1V->getName()+"."+BB2V->getName()));
1112 // Make the PHI node use the select for all incoming values for BB1/BB2
1113 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1114 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1115 PN->setIncomingValue(i, SI);
1119 // Update any PHI nodes in our new successors.
1120 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1121 AddPredecessorToBlock(*SI, BIParent, BB1);
1123 EraseTerminatorInstAndDCECond(BI);
1127 /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1
1128 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code
1129 /// (for now, restricted to a single instruction that's side effect free) from
1130 /// the BB1 into the branch block to speculatively execute it.
1135 /// br i1 %t1, label %BB1, label %BB2
1137 /// %t3 = add %t2, c
1143 /// %t4 = add %t2, c
1144 /// %t3 = select i1 %t1, %t2, %t3
1145 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) {
1146 // Only speculatively execution a single instruction (not counting the
1147 // terminator) for now.
1148 Instruction *HInst = NULL;
1149 Instruction *Term = BB1->getTerminator();
1150 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end();
1151 BBI != BBE; ++BBI) {
1152 Instruction *I = BBI;
1154 if (isa<DbgInfoIntrinsic>(I)) continue;
1155 if (I == Term) break;
1162 BasicBlock *BIParent = BI->getParent();
1164 // Check the instruction to be hoisted, if there is one.
1166 // Don't hoist the instruction if it's unsafe or expensive.
1167 if (!isSafeToSpeculativelyExecute(HInst))
1169 if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold)
1172 // Do not hoist the instruction if any of its operands are defined but not
1173 // used in this BB. The transformation will prevent the operand from
1174 // being sunk into the use block.
1175 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end();
1177 Instruction *OpI = dyn_cast<Instruction>(*i);
1178 if (OpI && OpI->getParent() == BIParent &&
1179 !OpI->mayHaveSideEffects() &&
1180 !OpI->isUsedInBasicBlock(BIParent))
1185 // Be conservative for now. FP select instruction can often be expensive.
1186 Value *BrCond = BI->getCondition();
1187 if (isa<FCmpInst>(BrCond))
1190 // If BB1 is actually on the false edge of the conditional branch, remember
1191 // to swap the select operands later.
1192 bool Invert = false;
1193 if (BB1 != BI->getSuccessor(0)) {
1194 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?");
1198 // Collect interesting PHIs, and scan for hazards.
1199 SmallSetVector<std::pair<Value *, Value *>, 4> PHIs;
1200 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0);
1201 for (BasicBlock::iterator I = BB2->begin();
1202 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1203 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1204 Value *BIParentV = PN->getIncomingValueForBlock(BIParent);
1206 // Skip PHIs which are trivial.
1207 if (BB1V == BIParentV)
1210 // Check for saftey.
1211 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) {
1212 // An unfolded ConstantExpr could end up getting expanded into
1213 // Instructions. Don't speculate this and another instruction at
1217 if (!isSafeToSpeculativelyExecute(CE))
1219 if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold)
1223 // Ok, we may insert a select for this PHI.
1224 PHIs.insert(std::make_pair(BB1V, BIParentV));
1227 // If there are no PHIs to process, bail early. This helps ensure idempotence
1232 // If we get here, we can hoist the instruction and if-convert.
1233 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";);
1235 // Hoist the instruction.
1237 BIParent->getInstList().splice(BI, BB1->getInstList(), HInst);
1239 // Insert selects and rewrite the PHI operands.
1240 IRBuilder<true, NoFolder> Builder(BI);
1241 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
1242 Value *TrueV = PHIs[i].first;
1243 Value *FalseV = PHIs[i].second;
1245 // Create a select whose true value is the speculatively executed value and
1246 // false value is the previously determined FalseV.
1249 SI = cast<SelectInst>
1250 (Builder.CreateSelect(BrCond, FalseV, TrueV,
1251 FalseV->getName() + "." + TrueV->getName()));
1253 SI = cast<SelectInst>
1254 (Builder.CreateSelect(BrCond, TrueV, FalseV,
1255 TrueV->getName() + "." + FalseV->getName()));
1257 // Make the PHI node use the select for all incoming values for "then" and
1259 for (BasicBlock::iterator I = BB2->begin();
1260 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1261 unsigned BB1I = PN->getBasicBlockIndex(BB1);
1262 unsigned BIParentI = PN->getBasicBlockIndex(BIParent);
1263 Value *BB1V = PN->getIncomingValue(BB1I);
1264 Value *BIParentV = PN->getIncomingValue(BIParentI);
1265 if (TrueV == BB1V && FalseV == BIParentV) {
1266 PN->setIncomingValue(BB1I, SI);
1267 PN->setIncomingValue(BIParentI, SI);
1276 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1277 /// across this block.
1278 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1279 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1282 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1283 if (isa<DbgInfoIntrinsic>(BBI))
1285 if (Size > 10) return false; // Don't clone large BB's.
1288 // We can only support instructions that do not define values that are
1289 // live outside of the current basic block.
1290 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1292 Instruction *U = cast<Instruction>(*UI);
1293 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1296 // Looks ok, continue checking.
1302 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1303 /// that is defined in the same block as the branch and if any PHI entries are
1304 /// constants, thread edges corresponding to that entry to be branches to their
1305 /// ultimate destination.
1306 static bool FoldCondBranchOnPHI(BranchInst *BI, const TargetData *TD) {
1307 BasicBlock *BB = BI->getParent();
1308 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1309 // NOTE: we currently cannot transform this case if the PHI node is used
1310 // outside of the block.
1311 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1314 // Degenerate case of a single entry PHI.
1315 if (PN->getNumIncomingValues() == 1) {
1316 FoldSingleEntryPHINodes(PN->getParent());
1320 // Now we know that this block has multiple preds and two succs.
1321 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1323 // Okay, this is a simple enough basic block. See if any phi values are
1325 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1326 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1327 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1329 // Okay, we now know that all edges from PredBB should be revectored to
1330 // branch to RealDest.
1331 BasicBlock *PredBB = PN->getIncomingBlock(i);
1332 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1334 if (RealDest == BB) continue; // Skip self loops.
1335 // Skip if the predecessor's terminator is an indirect branch.
1336 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1338 // The dest block might have PHI nodes, other predecessors and other
1339 // difficult cases. Instead of being smart about this, just insert a new
1340 // block that jumps to the destination block, effectively splitting
1341 // the edge we are about to create.
1342 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1343 RealDest->getName()+".critedge",
1344 RealDest->getParent(), RealDest);
1345 BranchInst::Create(RealDest, EdgeBB);
1347 // Update PHI nodes.
1348 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1350 // BB may have instructions that are being threaded over. Clone these
1351 // instructions into EdgeBB. We know that there will be no uses of the
1352 // cloned instructions outside of EdgeBB.
1353 BasicBlock::iterator InsertPt = EdgeBB->begin();
1354 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1355 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1356 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1357 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1360 // Clone the instruction.
1361 Instruction *N = BBI->clone();
1362 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1364 // Update operands due to translation.
1365 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1367 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1368 if (PI != TranslateMap.end())
1372 // Check for trivial simplification.
1373 if (Value *V = SimplifyInstruction(N, TD)) {
1374 TranslateMap[BBI] = V;
1375 delete N; // Instruction folded away, don't need actual inst
1377 // Insert the new instruction into its new home.
1378 EdgeBB->getInstList().insert(InsertPt, N);
1379 if (!BBI->use_empty())
1380 TranslateMap[BBI] = N;
1384 // Loop over all of the edges from PredBB to BB, changing them to branch
1385 // to EdgeBB instead.
1386 TerminatorInst *PredBBTI = PredBB->getTerminator();
1387 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1388 if (PredBBTI->getSuccessor(i) == BB) {
1389 BB->removePredecessor(PredBB);
1390 PredBBTI->setSuccessor(i, EdgeBB);
1393 // Recurse, simplifying any other constants.
1394 return FoldCondBranchOnPHI(BI, TD) | true;
1400 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1401 /// PHI node, see if we can eliminate it.
1402 static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) {
1403 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1404 // statement", which has a very simple dominance structure. Basically, we
1405 // are trying to find the condition that is being branched on, which
1406 // subsequently causes this merge to happen. We really want control
1407 // dependence information for this check, but simplifycfg can't keep it up
1408 // to date, and this catches most of the cases we care about anyway.
1409 BasicBlock *BB = PN->getParent();
1410 BasicBlock *IfTrue, *IfFalse;
1411 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1413 // Don't bother if the branch will be constant folded trivially.
1414 isa<ConstantInt>(IfCond))
1417 // Okay, we found that we can merge this two-entry phi node into a select.
1418 // Doing so would require us to fold *all* two entry phi nodes in this block.
1419 // At some point this becomes non-profitable (particularly if the target
1420 // doesn't support cmov's). Only do this transformation if there are two or
1421 // fewer PHI nodes in this block.
1422 unsigned NumPhis = 0;
1423 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1427 // Loop over the PHI's seeing if we can promote them all to select
1428 // instructions. While we are at it, keep track of the instructions
1429 // that need to be moved to the dominating block.
1430 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1431 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1432 MaxCostVal1 = PHINodeFoldingThreshold;
1434 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1435 PHINode *PN = cast<PHINode>(II++);
1436 if (Value *V = SimplifyInstruction(PN, TD)) {
1437 PN->replaceAllUsesWith(V);
1438 PN->eraseFromParent();
1442 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1444 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1449 // If we folded the first phi, PN dangles at this point. Refresh it. If
1450 // we ran out of PHIs then we simplified them all.
1451 PN = dyn_cast<PHINode>(BB->begin());
1452 if (PN == 0) return true;
1454 // Don't fold i1 branches on PHIs which contain binary operators. These can
1455 // often be turned into switches and other things.
1456 if (PN->getType()->isIntegerTy(1) &&
1457 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1458 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1459 isa<BinaryOperator>(IfCond)))
1462 // If we all PHI nodes are promotable, check to make sure that all
1463 // instructions in the predecessor blocks can be promoted as well. If
1464 // not, we won't be able to get rid of the control flow, so it's not
1465 // worth promoting to select instructions.
1466 BasicBlock *DomBlock = 0;
1467 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1468 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1469 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1472 DomBlock = *pred_begin(IfBlock1);
1473 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1474 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1475 // This is not an aggressive instruction that we can promote.
1476 // Because of this, we won't be able to get rid of the control
1477 // flow, so the xform is not worth it.
1482 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1485 DomBlock = *pred_begin(IfBlock2);
1486 for (BasicBlock::iterator I = IfBlock2->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 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1496 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1498 // If we can still promote the PHI nodes after this gauntlet of tests,
1499 // do all of the PHI's now.
1500 Instruction *InsertPt = DomBlock->getTerminator();
1501 IRBuilder<true, NoFolder> Builder(InsertPt);
1503 // Move all 'aggressive' instructions, which are defined in the
1504 // conditional parts of the if's up to the dominating block.
1506 DomBlock->getInstList().splice(InsertPt,
1507 IfBlock1->getInstList(), IfBlock1->begin(),
1508 IfBlock1->getTerminator());
1510 DomBlock->getInstList().splice(InsertPt,
1511 IfBlock2->getInstList(), IfBlock2->begin(),
1512 IfBlock2->getTerminator());
1514 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1515 // Change the PHI node into a select instruction.
1516 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1517 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1520 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1521 PN->replaceAllUsesWith(NV);
1523 PN->eraseFromParent();
1526 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1527 // has been flattened. Change DomBlock to jump directly to our new block to
1528 // avoid other simplifycfg's kicking in on the diamond.
1529 TerminatorInst *OldTI = DomBlock->getTerminator();
1530 Builder.SetInsertPoint(OldTI);
1531 Builder.CreateBr(BB);
1532 OldTI->eraseFromParent();
1536 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1537 /// to two returning blocks, try to merge them together into one return,
1538 /// introducing a select if the return values disagree.
1539 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1540 IRBuilder<> &Builder) {
1541 assert(BI->isConditional() && "Must be a conditional branch");
1542 BasicBlock *TrueSucc = BI->getSuccessor(0);
1543 BasicBlock *FalseSucc = BI->getSuccessor(1);
1544 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1545 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1547 // Check to ensure both blocks are empty (just a return) or optionally empty
1548 // with PHI nodes. If there are other instructions, merging would cause extra
1549 // computation on one path or the other.
1550 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1552 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1555 Builder.SetInsertPoint(BI);
1556 // Okay, we found a branch that is going to two return nodes. If
1557 // there is no return value for this function, just change the
1558 // branch into a return.
1559 if (FalseRet->getNumOperands() == 0) {
1560 TrueSucc->removePredecessor(BI->getParent());
1561 FalseSucc->removePredecessor(BI->getParent());
1562 Builder.CreateRetVoid();
1563 EraseTerminatorInstAndDCECond(BI);
1567 // Otherwise, figure out what the true and false return values are
1568 // so we can insert a new select instruction.
1569 Value *TrueValue = TrueRet->getReturnValue();
1570 Value *FalseValue = FalseRet->getReturnValue();
1572 // Unwrap any PHI nodes in the return blocks.
1573 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1574 if (TVPN->getParent() == TrueSucc)
1575 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1576 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1577 if (FVPN->getParent() == FalseSucc)
1578 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1580 // In order for this transformation to be safe, we must be able to
1581 // unconditionally execute both operands to the return. This is
1582 // normally the case, but we could have a potentially-trapping
1583 // constant expression that prevents this transformation from being
1585 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1588 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1592 // Okay, we collected all the mapped values and checked them for sanity, and
1593 // defined to really do this transformation. First, update the CFG.
1594 TrueSucc->removePredecessor(BI->getParent());
1595 FalseSucc->removePredecessor(BI->getParent());
1597 // Insert select instructions where needed.
1598 Value *BrCond = BI->getCondition();
1600 // Insert a select if the results differ.
1601 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1602 } else if (isa<UndefValue>(TrueValue)) {
1603 TrueValue = FalseValue;
1605 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1606 FalseValue, "retval");
1610 Value *RI = !TrueValue ?
1611 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1615 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1616 << "\n " << *BI << "NewRet = " << *RI
1617 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1619 EraseTerminatorInstAndDCECond(BI);
1624 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1625 /// probabilities of the branch taking each edge. Fills in the two APInt
1626 /// parameters and return true, or returns false if no or invalid metadata was
1628 static bool ExtractBranchMetadata(BranchInst *BI,
1629 APInt &ProbTrue, APInt &ProbFalse) {
1630 assert(BI->isConditional() &&
1631 "Looking for probabilities on unconditional branch?");
1632 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1633 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1634 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1635 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1636 if (!CITrue || !CIFalse) return false;
1637 ProbTrue = CITrue->getValue();
1638 ProbFalse = CIFalse->getValue();
1639 assert(ProbTrue.getBitWidth() == 32 && ProbFalse.getBitWidth() == 32 &&
1640 "Branch probability metadata must be 32-bit integers");
1644 /// MultiplyAndLosePrecision - Multiplies A and B, then returns the result. In
1645 /// the event of overflow, logically-shifts all four inputs right until the
1647 static APInt MultiplyAndLosePrecision(APInt &A, APInt &B, APInt &C, APInt &D,
1648 unsigned &BitsLost) {
1650 bool Overflow = false;
1651 APInt Result = A.umul_ov(B, Overflow);
1653 APInt MaxB = APInt::getMaxValue(A.getBitWidth()).udiv(A);
1657 } while (B.ugt(MaxB));
1658 A = A.lshr(BitsLost);
1659 C = C.lshr(BitsLost);
1660 D = D.lshr(BitsLost);
1666 /// checkCSEInPredecessor - Return true if the given instruction is available
1667 /// in its predecessor block. If yes, the instruction will be removed.
1669 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1670 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1672 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1673 Instruction *PBI = &*I;
1674 // Check whether Inst and PBI generate the same value.
1675 if (Inst->isIdenticalTo(PBI)) {
1676 Inst->replaceAllUsesWith(PBI);
1677 Inst->eraseFromParent();
1684 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1685 /// predecessor branches to us and one of our successors, fold the block into
1686 /// the predecessor and use logical operations to pick the right destination.
1687 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1688 BasicBlock *BB = BI->getParent();
1690 Instruction *Cond = 0;
1691 if (BI->isConditional())
1692 Cond = dyn_cast<Instruction>(BI->getCondition());
1694 // For unconditional branch, check for a simple CFG pattern, where
1695 // BB has a single predecessor and BB's successor is also its predecessor's
1696 // successor. If such pattern exisits, check for CSE between BB and its
1698 if (BasicBlock *PB = BB->getSinglePredecessor())
1699 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1700 if (PBI->isConditional() &&
1701 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1702 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1703 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1705 Instruction *Curr = I++;
1706 if (isa<CmpInst>(Curr)) {
1710 // Quit if we can't remove this instruction.
1711 if (!checkCSEInPredecessor(Curr, PB))
1720 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1721 Cond->getParent() != BB || !Cond->hasOneUse())
1724 // Only allow this if the condition is a simple instruction that can be
1725 // executed unconditionally. It must be in the same block as the branch, and
1726 // must be at the front of the block.
1727 BasicBlock::iterator FrontIt = BB->front();
1729 // Ignore dbg intrinsics.
1730 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1732 // Allow a single instruction to be hoisted in addition to the compare
1733 // that feeds the branch. We later ensure that any values that _it_ uses
1734 // were also live in the predecessor, so that we don't unnecessarily create
1735 // register pressure or inhibit out-of-order execution.
1736 Instruction *BonusInst = 0;
1737 if (&*FrontIt != Cond &&
1738 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
1739 isSafeToSpeculativelyExecute(FrontIt)) {
1740 BonusInst = &*FrontIt;
1743 // Ignore dbg intrinsics.
1744 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1747 // Only a single bonus inst is allowed.
1748 if (&*FrontIt != Cond)
1751 // Make sure the instruction after the condition is the cond branch.
1752 BasicBlock::iterator CondIt = Cond; ++CondIt;
1754 // Ingore dbg intrinsics.
1755 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
1760 // Cond is known to be a compare or binary operator. Check to make sure that
1761 // neither operand is a potentially-trapping constant expression.
1762 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
1765 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
1769 // Finally, don't infinitely unroll conditional loops.
1770 BasicBlock *TrueDest = BI->getSuccessor(0);
1771 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
1772 if (TrueDest == BB || FalseDest == BB)
1775 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1776 BasicBlock *PredBlock = *PI;
1777 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
1779 // Check that we have two conditional branches. If there is a PHI node in
1780 // the common successor, verify that the same value flows in from both
1782 SmallVector<PHINode*, 4> PHIs;
1783 if (PBI == 0 || PBI->isUnconditional() ||
1784 (BI->isConditional() &&
1785 !SafeToMergeTerminators(BI, PBI)) ||
1786 (!BI->isConditional() &&
1787 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
1790 // Determine if the two branches share a common destination.
1791 Instruction::BinaryOps Opc;
1792 bool InvertPredCond = false;
1794 if (BI->isConditional()) {
1795 if (PBI->getSuccessor(0) == TrueDest)
1796 Opc = Instruction::Or;
1797 else if (PBI->getSuccessor(1) == FalseDest)
1798 Opc = Instruction::And;
1799 else if (PBI->getSuccessor(0) == FalseDest)
1800 Opc = Instruction::And, InvertPredCond = true;
1801 else if (PBI->getSuccessor(1) == TrueDest)
1802 Opc = Instruction::Or, InvertPredCond = true;
1806 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
1810 // Ensure that any values used in the bonus instruction are also used
1811 // by the terminator of the predecessor. This means that those values
1812 // must already have been resolved, so we won't be inhibiting the
1813 // out-of-order core by speculating them earlier.
1815 // Collect the values used by the bonus inst
1816 SmallPtrSet<Value*, 4> UsedValues;
1817 for (Instruction::op_iterator OI = BonusInst->op_begin(),
1818 OE = BonusInst->op_end(); OI != OE; ++OI) {
1820 if (!isa<Constant>(V))
1821 UsedValues.insert(V);
1824 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
1825 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
1827 // Walk up to four levels back up the use-def chain of the predecessor's
1828 // terminator to see if all those values were used. The choice of four
1829 // levels is arbitrary, to provide a compile-time-cost bound.
1830 while (!Worklist.empty()) {
1831 std::pair<Value*, unsigned> Pair = Worklist.back();
1832 Worklist.pop_back();
1834 if (Pair.second >= 4) continue;
1835 UsedValues.erase(Pair.first);
1836 if (UsedValues.empty()) break;
1838 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
1839 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
1841 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
1845 if (!UsedValues.empty()) return false;
1848 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
1849 IRBuilder<> Builder(PBI);
1851 // If we need to invert the condition in the pred block to match, do so now.
1852 if (InvertPredCond) {
1853 Value *NewCond = PBI->getCondition();
1855 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
1856 CmpInst *CI = cast<CmpInst>(NewCond);
1857 CI->setPredicate(CI->getInversePredicate());
1859 NewCond = Builder.CreateNot(NewCond,
1860 PBI->getCondition()->getName()+".not");
1863 PBI->setCondition(NewCond);
1864 PBI->swapSuccessors();
1867 // If we have a bonus inst, clone it into the predecessor block.
1868 Instruction *NewBonus = 0;
1870 NewBonus = BonusInst->clone();
1871 PredBlock->getInstList().insert(PBI, NewBonus);
1872 NewBonus->takeName(BonusInst);
1873 BonusInst->setName(BonusInst->getName()+".old");
1876 // Clone Cond into the predecessor basic block, and or/and the
1877 // two conditions together.
1878 Instruction *New = Cond->clone();
1879 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
1880 PredBlock->getInstList().insert(PBI, New);
1881 New->takeName(Cond);
1882 Cond->setName(New->getName()+".old");
1884 if (BI->isConditional()) {
1885 Instruction *NewCond =
1886 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
1888 PBI->setCondition(NewCond);
1890 if (PBI->getSuccessor(0) == BB) {
1891 AddPredecessorToBlock(TrueDest, PredBlock, BB);
1892 PBI->setSuccessor(0, TrueDest);
1894 if (PBI->getSuccessor(1) == BB) {
1895 AddPredecessorToBlock(FalseDest, PredBlock, BB);
1896 PBI->setSuccessor(1, FalseDest);
1899 // Update PHI nodes in the common successors.
1900 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
1901 ConstantInt *PBI_C = cast<ConstantInt>(
1902 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
1903 assert(PBI_C->getType()->isIntegerTy(1));
1904 Instruction *MergedCond = 0;
1905 if (PBI->getSuccessor(0) == TrueDest) {
1906 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
1907 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
1908 // is false: !PBI_Cond and BI_Value
1909 Instruction *NotCond =
1910 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
1913 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
1918 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
1919 PBI->getCondition(), MergedCond,
1922 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
1923 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
1924 // is false: PBI_Cond and BI_Value
1926 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
1927 PBI->getCondition(), New,
1929 if (PBI_C->isOne()) {
1930 Instruction *NotCond =
1931 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
1934 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
1935 NotCond, MergedCond,
1940 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
1943 // Change PBI from Conditional to Unconditional.
1944 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
1945 EraseTerminatorInstAndDCECond(PBI);
1949 // TODO: If BB is reachable from all paths through PredBlock, then we
1950 // could replace PBI's branch probabilities with BI's.
1952 // Merge probability data into PredBlock's branch.
1954 if (PBI->isConditional() && BI->isConditional() &&
1955 ExtractBranchMetadata(PBI, C, D) && ExtractBranchMetadata(BI, A, B)) {
1956 // Given IR which does:
1958 // br i1 %x, label %bbB, label %bbC
1960 // br i1 %y, label %bbD, label %bbC
1961 // Let's call the probability that we take the edge from %bbA to %bbB
1962 // 'a', from %bbA to %bbC, 'b', from %bbB to %bbD 'c' and from %bbB to
1963 // %bbC probability 'd'.
1965 // We transform the IR into:
1967 // br i1 %z, label %bbD, label %bbC
1968 // where the probability of going to %bbD is (a*c) and going to bbC is
1971 // Probabilities aren't stored as ratios directly. Using branch weights,
1973 // (a*c)% = A*C, (b+(a*d))% = A*D+B*C+B*D.
1975 // In the event of overflow, we want to drop the LSB of the input
1979 // Ignore overflow result on ProbTrue.
1980 APInt ProbTrue = MultiplyAndLosePrecision(A, C, B, D, BitsLost);
1982 APInt Tmp1 = MultiplyAndLosePrecision(B, D, A, C, BitsLost);
1984 ProbTrue = ProbTrue.lshr(BitsLost*2);
1987 APInt Tmp2 = MultiplyAndLosePrecision(A, D, C, B, BitsLost);
1989 ProbTrue = ProbTrue.lshr(BitsLost*2);
1990 Tmp1 = Tmp1.lshr(BitsLost*2);
1993 APInt Tmp3 = MultiplyAndLosePrecision(B, C, A, D, BitsLost);
1995 ProbTrue = ProbTrue.lshr(BitsLost*2);
1996 Tmp1 = Tmp1.lshr(BitsLost*2);
1997 Tmp2 = Tmp2.lshr(BitsLost*2);
2000 bool Overflow1 = false, Overflow2 = false;
2001 APInt Tmp4 = Tmp2.uadd_ov(Tmp3, Overflow1);
2002 APInt ProbFalse = Tmp4.uadd_ov(Tmp1, Overflow2);
2004 if (Overflow1 || Overflow2) {
2005 ProbTrue = ProbTrue.lshr(1);
2006 Tmp1 = Tmp1.lshr(1);
2007 Tmp2 = Tmp2.lshr(1);
2008 Tmp3 = Tmp3.lshr(1);
2010 ProbFalse = Tmp4 + Tmp1;
2013 // The sum of branch weights must fit in 32-bits.
2014 if (ProbTrue.isNegative() && ProbFalse.isNegative()) {
2015 ProbTrue = ProbTrue.lshr(1);
2016 ProbFalse = ProbFalse.lshr(1);
2019 if (ProbTrue != ProbFalse) {
2020 // Normalize the result.
2021 APInt GCD = APIntOps::GreatestCommonDivisor(ProbTrue, ProbFalse);
2022 ProbTrue = ProbTrue.udiv(GCD);
2023 ProbFalse = ProbFalse.udiv(GCD);
2025 MDBuilder MDB(BI->getContext());
2026 MDNode *N = MDB.createBranchWeights(ProbTrue.getZExtValue(),
2027 ProbFalse.getZExtValue());
2028 PBI->setMetadata(LLVMContext::MD_prof, N);
2030 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2033 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2036 // Copy any debug value intrinsics into the end of PredBlock.
2037 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2038 if (isa<DbgInfoIntrinsic>(*I))
2039 I->clone()->insertBefore(PBI);
2046 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2047 /// predecessor of another block, this function tries to simplify it. We know
2048 /// that PBI and BI are both conditional branches, and BI is in one of the
2049 /// successor blocks of PBI - PBI branches to BI.
2050 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2051 assert(PBI->isConditional() && BI->isConditional());
2052 BasicBlock *BB = BI->getParent();
2054 // If this block ends with a branch instruction, and if there is a
2055 // predecessor that ends on a branch of the same condition, make
2056 // this conditional branch redundant.
2057 if (PBI->getCondition() == BI->getCondition() &&
2058 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2059 // Okay, the outcome of this conditional branch is statically
2060 // knowable. If this block had a single pred, handle specially.
2061 if (BB->getSinglePredecessor()) {
2062 // Turn this into a branch on constant.
2063 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2064 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2066 return true; // Nuke the branch on constant.
2069 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2070 // in the constant and simplify the block result. Subsequent passes of
2071 // simplifycfg will thread the block.
2072 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2073 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2074 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2075 std::distance(PB, PE),
2076 BI->getCondition()->getName() + ".pr",
2078 // Okay, we're going to insert the PHI node. Since PBI is not the only
2079 // predecessor, compute the PHI'd conditional value for all of the preds.
2080 // Any predecessor where the condition is not computable we keep symbolic.
2081 for (pred_iterator PI = PB; PI != PE; ++PI) {
2082 BasicBlock *P = *PI;
2083 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2084 PBI != BI && PBI->isConditional() &&
2085 PBI->getCondition() == BI->getCondition() &&
2086 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2087 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2088 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2091 NewPN->addIncoming(BI->getCondition(), P);
2095 BI->setCondition(NewPN);
2100 // If this is a conditional branch in an empty block, and if any
2101 // predecessors is a conditional branch to one of our destinations,
2102 // fold the conditions into logical ops and one cond br.
2103 BasicBlock::iterator BBI = BB->begin();
2104 // Ignore dbg intrinsics.
2105 while (isa<DbgInfoIntrinsic>(BBI))
2111 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2116 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2118 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2119 PBIOp = 0, BIOp = 1;
2120 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2121 PBIOp = 1, BIOp = 0;
2122 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2127 // Check to make sure that the other destination of this branch
2128 // isn't BB itself. If so, this is an infinite loop that will
2129 // keep getting unwound.
2130 if (PBI->getSuccessor(PBIOp) == BB)
2133 // Do not perform this transformation if it would require
2134 // insertion of a large number of select instructions. For targets
2135 // without predication/cmovs, this is a big pessimization.
2136 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2138 unsigned NumPhis = 0;
2139 for (BasicBlock::iterator II = CommonDest->begin();
2140 isa<PHINode>(II); ++II, ++NumPhis)
2141 if (NumPhis > 2) // Disable this xform.
2144 // Finally, if everything is ok, fold the branches to logical ops.
2145 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2147 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2148 << "AND: " << *BI->getParent());
2151 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2152 // branch in it, where one edge (OtherDest) goes back to itself but the other
2153 // exits. We don't *know* that the program avoids the infinite loop
2154 // (even though that seems likely). If we do this xform naively, we'll end up
2155 // recursively unpeeling the loop. Since we know that (after the xform is
2156 // done) that the block *is* infinite if reached, we just make it an obviously
2157 // infinite loop with no cond branch.
2158 if (OtherDest == BB) {
2159 // Insert it at the end of the function, because it's either code,
2160 // or it won't matter if it's hot. :)
2161 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2162 "infloop", BB->getParent());
2163 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2164 OtherDest = InfLoopBlock;
2167 DEBUG(dbgs() << *PBI->getParent()->getParent());
2169 // BI may have other predecessors. Because of this, we leave
2170 // it alone, but modify PBI.
2172 // Make sure we get to CommonDest on True&True directions.
2173 Value *PBICond = PBI->getCondition();
2174 IRBuilder<true, NoFolder> Builder(PBI);
2176 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2178 Value *BICond = BI->getCondition();
2180 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2182 // Merge the conditions.
2183 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2185 // Modify PBI to branch on the new condition to the new dests.
2186 PBI->setCondition(Cond);
2187 PBI->setSuccessor(0, CommonDest);
2188 PBI->setSuccessor(1, OtherDest);
2190 // OtherDest may have phi nodes. If so, add an entry from PBI's
2191 // block that are identical to the entries for BI's block.
2192 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2194 // We know that the CommonDest already had an edge from PBI to
2195 // it. If it has PHIs though, the PHIs may have different
2196 // entries for BB and PBI's BB. If so, insert a select to make
2199 for (BasicBlock::iterator II = CommonDest->begin();
2200 (PN = dyn_cast<PHINode>(II)); ++II) {
2201 Value *BIV = PN->getIncomingValueForBlock(BB);
2202 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2203 Value *PBIV = PN->getIncomingValue(PBBIdx);
2205 // Insert a select in PBI to pick the right value.
2206 Value *NV = cast<SelectInst>
2207 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2208 PN->setIncomingValue(PBBIdx, NV);
2212 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2213 DEBUG(dbgs() << *PBI->getParent()->getParent());
2215 // This basic block is probably dead. We know it has at least
2216 // one fewer predecessor.
2220 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2221 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2222 // Takes care of updating the successors and removing the old terminator.
2223 // Also makes sure not to introduce new successors by assuming that edges to
2224 // non-successor TrueBBs and FalseBBs aren't reachable.
2225 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2226 BasicBlock *TrueBB, BasicBlock *FalseBB){
2227 // Remove any superfluous successor edges from the CFG.
2228 // First, figure out which successors to preserve.
2229 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2231 BasicBlock *KeepEdge1 = TrueBB;
2232 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2234 // Then remove the rest.
2235 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2236 BasicBlock *Succ = OldTerm->getSuccessor(I);
2237 // Make sure only to keep exactly one copy of each edge.
2238 if (Succ == KeepEdge1)
2240 else if (Succ == KeepEdge2)
2243 Succ->removePredecessor(OldTerm->getParent());
2246 IRBuilder<> Builder(OldTerm);
2247 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2249 // Insert an appropriate new terminator.
2250 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2251 if (TrueBB == FalseBB)
2252 // We were only looking for one successor, and it was present.
2253 // Create an unconditional branch to it.
2254 Builder.CreateBr(TrueBB);
2256 // We found both of the successors we were looking for.
2257 // Create a conditional branch sharing the condition of the select.
2258 Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2259 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2260 // Neither of the selected blocks were successors, so this
2261 // terminator must be unreachable.
2262 new UnreachableInst(OldTerm->getContext(), OldTerm);
2264 // One of the selected values was a successor, but the other wasn't.
2265 // Insert an unconditional branch to the one that was found;
2266 // the edge to the one that wasn't must be unreachable.
2268 // Only TrueBB was found.
2269 Builder.CreateBr(TrueBB);
2271 // Only FalseBB was found.
2272 Builder.CreateBr(FalseBB);
2275 EraseTerminatorInstAndDCECond(OldTerm);
2279 // SimplifySwitchOnSelect - Replaces
2280 // (switch (select cond, X, Y)) on constant X, Y
2281 // with a branch - conditional if X and Y lead to distinct BBs,
2282 // unconditional otherwise.
2283 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2284 // Check for constant integer values in the select.
2285 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2286 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2287 if (!TrueVal || !FalseVal)
2290 // Find the relevant condition and destinations.
2291 Value *Condition = Select->getCondition();
2292 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2293 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2295 // Perform the actual simplification.
2296 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB);
2299 // SimplifyIndirectBrOnSelect - Replaces
2300 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2301 // blockaddress(@fn, BlockB)))
2303 // (br cond, BlockA, BlockB).
2304 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2305 // Check that both operands of the select are block addresses.
2306 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2307 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2311 // Extract the actual blocks.
2312 BasicBlock *TrueBB = TBA->getBasicBlock();
2313 BasicBlock *FalseBB = FBA->getBasicBlock();
2315 // Perform the actual simplification.
2316 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB);
2319 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2320 /// instruction (a seteq/setne with a constant) as the only instruction in a
2321 /// block that ends with an uncond branch. We are looking for a very specific
2322 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2323 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2324 /// default value goes to an uncond block with a seteq in it, we get something
2327 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2329 /// %tmp = icmp eq i8 %A, 92
2332 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2334 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2335 /// the PHI, merging the third icmp into the switch.
2336 static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
2337 const TargetData *TD,
2338 IRBuilder<> &Builder) {
2339 BasicBlock *BB = ICI->getParent();
2341 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2343 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2345 Value *V = ICI->getOperand(0);
2346 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2348 // The pattern we're looking for is where our only predecessor is a switch on
2349 // 'V' and this block is the default case for the switch. In this case we can
2350 // fold the compared value into the switch to simplify things.
2351 BasicBlock *Pred = BB->getSinglePredecessor();
2352 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2354 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2355 if (SI->getCondition() != V)
2358 // If BB is reachable on a non-default case, then we simply know the value of
2359 // V in this block. Substitute it and constant fold the icmp instruction
2361 if (SI->getDefaultDest() != BB) {
2362 ConstantInt *VVal = SI->findCaseDest(BB);
2363 assert(VVal && "Should have a unique destination value");
2364 ICI->setOperand(0, VVal);
2366 if (Value *V = SimplifyInstruction(ICI, TD)) {
2367 ICI->replaceAllUsesWith(V);
2368 ICI->eraseFromParent();
2370 // BB is now empty, so it is likely to simplify away.
2371 return SimplifyCFG(BB) | true;
2374 // Ok, the block is reachable from the default dest. If the constant we're
2375 // comparing exists in one of the other edges, then we can constant fold ICI
2377 if (SI->findCaseValue(Cst) != SI->case_default()) {
2379 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2380 V = ConstantInt::getFalse(BB->getContext());
2382 V = ConstantInt::getTrue(BB->getContext());
2384 ICI->replaceAllUsesWith(V);
2385 ICI->eraseFromParent();
2386 // BB is now empty, so it is likely to simplify away.
2387 return SimplifyCFG(BB) | true;
2390 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2392 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2393 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2394 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2395 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2398 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2400 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2401 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2403 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2404 std::swap(DefaultCst, NewCst);
2406 // Replace ICI (which is used by the PHI for the default value) with true or
2407 // false depending on if it is EQ or NE.
2408 ICI->replaceAllUsesWith(DefaultCst);
2409 ICI->eraseFromParent();
2411 // Okay, the switch goes to this block on a default value. Add an edge from
2412 // the switch to the merge point on the compared value.
2413 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2414 BB->getParent(), BB);
2415 SI->addCase(Cst, NewBB);
2417 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2418 Builder.SetInsertPoint(NewBB);
2419 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2420 Builder.CreateBr(SuccBlock);
2421 PHIUse->addIncoming(NewCst, NewBB);
2425 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2426 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2427 /// fold it into a switch instruction if so.
2428 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const TargetData *TD,
2429 IRBuilder<> &Builder) {
2430 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2431 if (Cond == 0) return false;
2434 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2435 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2436 // 'setne's and'ed together, collect them.
2438 std::vector<ConstantInt*> Values;
2439 bool TrueWhenEqual = true;
2440 Value *ExtraCase = 0;
2441 unsigned UsedICmps = 0;
2443 if (Cond->getOpcode() == Instruction::Or) {
2444 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2446 } else if (Cond->getOpcode() == Instruction::And) {
2447 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2449 TrueWhenEqual = false;
2452 // If we didn't have a multiply compared value, fail.
2453 if (CompVal == 0) return false;
2455 // Avoid turning single icmps into a switch.
2459 // There might be duplicate constants in the list, which the switch
2460 // instruction can't handle, remove them now.
2461 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2462 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2464 // If Extra was used, we require at least two switch values to do the
2465 // transformation. A switch with one value is just an cond branch.
2466 if (ExtraCase && Values.size() < 2) return false;
2468 // TODO: Preserve branch weight metadata, similarly to how
2469 // FoldValueComparisonIntoPredecessors preserves it.
2471 // Figure out which block is which destination.
2472 BasicBlock *DefaultBB = BI->getSuccessor(1);
2473 BasicBlock *EdgeBB = BI->getSuccessor(0);
2474 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2476 BasicBlock *BB = BI->getParent();
2478 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2479 << " cases into SWITCH. BB is:\n" << *BB);
2481 // If there are any extra values that couldn't be folded into the switch
2482 // then we evaluate them with an explicit branch first. Split the block
2483 // right before the condbr to handle it.
2485 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2486 // Remove the uncond branch added to the old block.
2487 TerminatorInst *OldTI = BB->getTerminator();
2488 Builder.SetInsertPoint(OldTI);
2491 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2493 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2495 OldTI->eraseFromParent();
2497 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2498 // for the edge we just added.
2499 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2501 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2502 << "\nEXTRABB = " << *BB);
2506 Builder.SetInsertPoint(BI);
2507 // Convert pointer to int before we switch.
2508 if (CompVal->getType()->isPointerTy()) {
2509 assert(TD && "Cannot switch on pointer without TargetData");
2510 CompVal = Builder.CreatePtrToInt(CompVal,
2511 TD->getIntPtrType(CompVal->getContext()),
2515 // Create the new switch instruction now.
2516 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2518 // Add all of the 'cases' to the switch instruction.
2519 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2520 New->addCase(Values[i], EdgeBB);
2522 // We added edges from PI to the EdgeBB. As such, if there were any
2523 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2524 // the number of edges added.
2525 for (BasicBlock::iterator BBI = EdgeBB->begin();
2526 isa<PHINode>(BBI); ++BBI) {
2527 PHINode *PN = cast<PHINode>(BBI);
2528 Value *InVal = PN->getIncomingValueForBlock(BB);
2529 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2530 PN->addIncoming(InVal, BB);
2533 // Erase the old branch instruction.
2534 EraseTerminatorInstAndDCECond(BI);
2536 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2540 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2541 // If this is a trivial landing pad that just continues unwinding the caught
2542 // exception then zap the landing pad, turning its invokes into calls.
2543 BasicBlock *BB = RI->getParent();
2544 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2545 if (RI->getValue() != LPInst)
2546 // Not a landing pad, or the resume is not unwinding the exception that
2547 // caused control to branch here.
2550 // Check that there are no other instructions except for debug intrinsics.
2551 BasicBlock::iterator I = LPInst, E = RI;
2553 if (!isa<DbgInfoIntrinsic>(I))
2556 // Turn all invokes that unwind here into calls and delete the basic block.
2557 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2558 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2559 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2560 // Insert a call instruction before the invoke.
2561 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2563 Call->setCallingConv(II->getCallingConv());
2564 Call->setAttributes(II->getAttributes());
2565 Call->setDebugLoc(II->getDebugLoc());
2567 // Anything that used the value produced by the invoke instruction now uses
2568 // the value produced by the call instruction. Note that we do this even
2569 // for void functions and calls with no uses so that the callgraph edge is
2571 II->replaceAllUsesWith(Call);
2572 BB->removePredecessor(II->getParent());
2574 // Insert a branch to the normal destination right before the invoke.
2575 BranchInst::Create(II->getNormalDest(), II);
2577 // Finally, delete the invoke instruction!
2578 II->eraseFromParent();
2581 // The landingpad is now unreachable. Zap it.
2582 BB->eraseFromParent();
2586 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2587 BasicBlock *BB = RI->getParent();
2588 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2590 // Find predecessors that end with branches.
2591 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2592 SmallVector<BranchInst*, 8> CondBranchPreds;
2593 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2594 BasicBlock *P = *PI;
2595 TerminatorInst *PTI = P->getTerminator();
2596 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2597 if (BI->isUnconditional())
2598 UncondBranchPreds.push_back(P);
2600 CondBranchPreds.push_back(BI);
2604 // If we found some, do the transformation!
2605 if (!UncondBranchPreds.empty() && DupRet) {
2606 while (!UncondBranchPreds.empty()) {
2607 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2608 DEBUG(dbgs() << "FOLDING: " << *BB
2609 << "INTO UNCOND BRANCH PRED: " << *Pred);
2610 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2613 // If we eliminated all predecessors of the block, delete the block now.
2614 if (pred_begin(BB) == pred_end(BB))
2615 // We know there are no successors, so just nuke the block.
2616 BB->eraseFromParent();
2621 // Check out all of the conditional branches going to this return
2622 // instruction. If any of them just select between returns, change the
2623 // branch itself into a select/return pair.
2624 while (!CondBranchPreds.empty()) {
2625 BranchInst *BI = CondBranchPreds.pop_back_val();
2627 // Check to see if the non-BB successor is also a return block.
2628 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2629 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2630 SimplifyCondBranchToTwoReturns(BI, Builder))
2636 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2637 BasicBlock *BB = UI->getParent();
2639 bool Changed = false;
2641 // If there are any instructions immediately before the unreachable that can
2642 // be removed, do so.
2643 while (UI != BB->begin()) {
2644 BasicBlock::iterator BBI = UI;
2646 // Do not delete instructions that can have side effects which might cause
2647 // the unreachable to not be reachable; specifically, calls and volatile
2648 // operations may have this effect.
2649 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2651 if (BBI->mayHaveSideEffects()) {
2652 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2653 if (SI->isVolatile())
2655 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2656 if (LI->isVolatile())
2658 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2659 if (RMWI->isVolatile())
2661 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2662 if (CXI->isVolatile())
2664 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2665 !isa<LandingPadInst>(BBI)) {
2668 // Note that deleting LandingPad's here is in fact okay, although it
2669 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2670 // all the predecessors of this block will be the unwind edges of Invokes,
2671 // and we can therefore guarantee this block will be erased.
2674 // Delete this instruction (any uses are guaranteed to be dead)
2675 if (!BBI->use_empty())
2676 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2677 BBI->eraseFromParent();
2681 // If the unreachable instruction is the first in the block, take a gander
2682 // at all of the predecessors of this instruction, and simplify them.
2683 if (&BB->front() != UI) return Changed;
2685 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2686 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
2687 TerminatorInst *TI = Preds[i]->getTerminator();
2688 IRBuilder<> Builder(TI);
2689 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2690 if (BI->isUnconditional()) {
2691 if (BI->getSuccessor(0) == BB) {
2692 new UnreachableInst(TI->getContext(), TI);
2693 TI->eraseFromParent();
2697 if (BI->getSuccessor(0) == BB) {
2698 Builder.CreateBr(BI->getSuccessor(1));
2699 EraseTerminatorInstAndDCECond(BI);
2700 } else if (BI->getSuccessor(1) == BB) {
2701 Builder.CreateBr(BI->getSuccessor(0));
2702 EraseTerminatorInstAndDCECond(BI);
2706 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2707 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2709 if (i.getCaseSuccessor() == BB) {
2710 BB->removePredecessor(SI->getParent());
2715 // If the default value is unreachable, figure out the most popular
2716 // destination and make it the default.
2717 if (SI->getDefaultDest() == BB) {
2718 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
2719 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2721 std::pair<unsigned, unsigned> &entry =
2722 Popularity[i.getCaseSuccessor()];
2723 if (entry.first == 0) {
2725 entry.second = i.getCaseIndex();
2731 // Find the most popular block.
2732 unsigned MaxPop = 0;
2733 unsigned MaxIndex = 0;
2734 BasicBlock *MaxBlock = 0;
2735 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
2736 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
2737 if (I->second.first > MaxPop ||
2738 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
2739 MaxPop = I->second.first;
2740 MaxIndex = I->second.second;
2741 MaxBlock = I->first;
2745 // Make this the new default, allowing us to delete any explicit
2747 SI->setDefaultDest(MaxBlock);
2750 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
2752 if (isa<PHINode>(MaxBlock->begin()))
2753 for (unsigned i = 0; i != MaxPop-1; ++i)
2754 MaxBlock->removePredecessor(SI->getParent());
2756 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2758 if (i.getCaseSuccessor() == MaxBlock) {
2764 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
2765 if (II->getUnwindDest() == BB) {
2766 // Convert the invoke to a call instruction. This would be a good
2767 // place to note that the call does not throw though.
2768 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
2769 II->removeFromParent(); // Take out of symbol table
2771 // Insert the call now...
2772 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
2773 Builder.SetInsertPoint(BI);
2774 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
2775 Args, II->getName());
2776 CI->setCallingConv(II->getCallingConv());
2777 CI->setAttributes(II->getAttributes());
2778 // If the invoke produced a value, the call does now instead.
2779 II->replaceAllUsesWith(CI);
2786 // If this block is now dead, remove it.
2787 if (pred_begin(BB) == pred_end(BB) &&
2788 BB != &BB->getParent()->getEntryBlock()) {
2789 // We know there are no successors, so just nuke the block.
2790 BB->eraseFromParent();
2797 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
2798 /// integer range comparison into a sub, an icmp and a branch.
2799 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
2800 assert(SI->getNumCases() > 1 && "Degenerate switch?");
2802 // Make sure all cases point to the same destination and gather the values.
2803 SmallVector<ConstantInt *, 16> Cases;
2804 SwitchInst::CaseIt I = SI->case_begin();
2805 Cases.push_back(I.getCaseValue());
2806 SwitchInst::CaseIt PrevI = I++;
2807 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
2808 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
2810 Cases.push_back(I.getCaseValue());
2812 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
2814 // Sort the case values, then check if they form a range we can transform.
2815 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
2816 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
2817 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
2821 Constant *Offset = ConstantExpr::getNeg(Cases.back());
2822 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
2824 Value *Sub = SI->getCondition();
2825 if (!Offset->isNullValue())
2826 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
2827 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
2828 Builder.CreateCondBr(
2829 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
2831 // Prune obsolete incoming values off the successor's PHI nodes.
2832 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
2833 isa<PHINode>(BBI); ++BBI) {
2834 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
2835 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
2837 SI->eraseFromParent();
2842 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
2843 /// and use it to remove dead cases.
2844 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
2845 Value *Cond = SI->getCondition();
2846 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
2847 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
2848 ComputeMaskedBits(Cond, KnownZero, KnownOne);
2850 // Gather dead cases.
2851 SmallVector<ConstantInt*, 8> DeadCases;
2852 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
2853 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
2854 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
2855 DeadCases.push_back(I.getCaseValue());
2856 DEBUG(dbgs() << "SimplifyCFG: switch case '"
2857 << I.getCaseValue() << "' is dead.\n");
2861 // Remove dead cases from the switch.
2862 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
2863 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
2864 assert(Case != SI->case_default() &&
2865 "Case was not found. Probably mistake in DeadCases forming.");
2866 // Prune unused values from PHI nodes.
2867 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
2868 SI->removeCase(Case);
2871 return !DeadCases.empty();
2874 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
2875 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
2876 /// by an unconditional branch), look at the phi node for BB in the successor
2877 /// block and see if the incoming value is equal to CaseValue. If so, return
2878 /// the phi node, and set PhiIndex to BB's index in the phi node.
2879 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
2882 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
2883 return NULL; // BB must be empty to be a candidate for simplification.
2884 if (!BB->getSinglePredecessor())
2885 return NULL; // BB must be dominated by the switch.
2887 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
2888 if (!Branch || !Branch->isUnconditional())
2889 return NULL; // Terminator must be unconditional branch.
2891 BasicBlock *Succ = Branch->getSuccessor(0);
2893 BasicBlock::iterator I = Succ->begin();
2894 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
2895 int Idx = PHI->getBasicBlockIndex(BB);
2896 assert(Idx >= 0 && "PHI has no entry for predecessor?");
2898 Value *InValue = PHI->getIncomingValue(Idx);
2899 if (InValue != CaseValue) continue;
2908 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
2909 /// instruction to a phi node dominated by the switch, if that would mean that
2910 /// some of the destination blocks of the switch can be folded away.
2911 /// Returns true if a change is made.
2912 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
2913 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
2914 ForwardingNodesMap ForwardingNodes;
2916 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
2917 ConstantInt *CaseValue = I.getCaseValue();
2918 BasicBlock *CaseDest = I.getCaseSuccessor();
2921 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
2925 ForwardingNodes[PHI].push_back(PhiIndex);
2928 bool Changed = false;
2930 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
2931 E = ForwardingNodes.end(); I != E; ++I) {
2932 PHINode *Phi = I->first;
2933 SmallVector<int,4> &Indexes = I->second;
2935 if (Indexes.size() < 2) continue;
2937 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
2938 Phi->setIncomingValue(Indexes[I], SI->getCondition());
2945 /// ValidLookupTableConstant - Return true if the backend will be able to handle
2946 /// initializing an array of constants like C.
2947 static bool ValidLookupTableConstant(Constant *C) {
2948 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
2949 return CE->isGEPWithNoNotionalOverIndexing();
2951 return isa<ConstantFP>(C) ||
2952 isa<ConstantInt>(C) ||
2953 isa<ConstantPointerNull>(C) ||
2954 isa<GlobalValue>(C) ||
2958 /// GetCaseResulsts - Try to determine the resulting constant values in phi
2959 /// nodes at the common destination basic block for one of the case
2960 /// destinations of a switch instruction.
2961 static bool GetCaseResults(SwitchInst *SI,
2962 BasicBlock *CaseDest,
2963 BasicBlock **CommonDest,
2964 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) {
2965 // The block from which we enter the common destination.
2966 BasicBlock *Pred = SI->getParent();
2968 // If CaseDest is empty, continue to its successor.
2969 if (CaseDest->getFirstNonPHIOrDbg() == CaseDest->getTerminator() &&
2970 !isa<PHINode>(CaseDest->begin())) {
2972 TerminatorInst *Terminator = CaseDest->getTerminator();
2973 if (Terminator->getNumSuccessors() != 1)
2977 CaseDest = Terminator->getSuccessor(0);
2980 // If we did not have a CommonDest before, use the current one.
2982 *CommonDest = CaseDest;
2983 // If the destination isn't the common one, abort.
2984 if (CaseDest != *CommonDest)
2987 // Get the values for this case from phi nodes in the destination block.
2988 BasicBlock::iterator I = (*CommonDest)->begin();
2989 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
2990 int Idx = PHI->getBasicBlockIndex(Pred);
2994 Constant *ConstVal = dyn_cast<Constant>(PHI->getIncomingValue(Idx));
2998 // Be conservative about which kinds of constants we support.
2999 if (!ValidLookupTableConstant(ConstVal))
3002 Res.push_back(std::make_pair(PHI, ConstVal));
3008 /// BuildLookupTable - Build a lookup table with the contents of Results, using
3009 /// DefaultResult to fill the holes in the table. If the table ends up
3010 /// containing the same result in each element, set *SingleResult to that value
3011 /// and return NULL.
3012 static GlobalVariable *BuildLookupTable(Module &M,
3014 ConstantInt *Offset,
3015 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Results,
3016 Constant *DefaultResult,
3017 Constant **SingleResult) {
3018 assert(Results.size() && "Need values to build lookup table");
3019 assert(TableSize >= Results.size() && "Table needs to hold all values");
3021 // If all values in the table are equal, this is that value.
3022 Constant *SameResult = Results.begin()->second;
3024 // Build up the table contents.
3025 std::vector<Constant*> TableContents(TableSize);
3026 for (size_t I = 0, E = Results.size(); I != E; ++I) {
3027 ConstantInt *CaseVal = Results[I].first;
3028 Constant *CaseRes = Results[I].second;
3030 uint64_t Idx = (CaseVal->getValue() - Offset->getValue()).getLimitedValue();
3031 TableContents[Idx] = CaseRes;
3033 if (CaseRes != SameResult)
3037 // Fill in any holes in the table with the default result.
3038 if (Results.size() < TableSize) {
3039 for (unsigned i = 0; i < TableSize; ++i) {
3040 if (!TableContents[i])
3041 TableContents[i] = DefaultResult;
3044 if (DefaultResult != SameResult)
3048 // Same result was used in the entire table; just return that.
3050 *SingleResult = SameResult;
3054 ArrayType *ArrayTy = ArrayType::get(DefaultResult->getType(), TableSize);
3055 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3057 GlobalVariable *GV = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3058 GlobalVariable::PrivateLinkage,
3061 GV->setUnnamedAddr(true);
3065 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3066 /// phi nodes in a common successor block with different constant values,
3067 /// replace the switch with lookup tables.
3068 static bool SwitchToLookupTable(SwitchInst *SI,
3069 IRBuilder<> &Builder) {
3070 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3071 // FIXME: Handle unreachable cases.
3073 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3074 // split off a dense part and build a lookup table for that.
3076 // FIXME: If the results are all integers and the lookup table would fit in a
3077 // target-legal register, we should store them as a bitmap and use shift/mask
3078 // to look up the result.
3080 // FIXME: This creates arrays of GEPs to constant strings, which means each
3081 // GEP needs a runtime relocation in PIC code. We should just build one big
3082 // string and lookup indices into that.
3084 // Ignore the switch if the number of cases are too small.
3085 // This is similar to the check when building jump tables in
3086 // SelectionDAGBuilder::handleJTSwitchCase.
3087 // FIXME: Determine the best cut-off.
3088 if (SI->getNumCases() < 4)
3091 // Figure out the corresponding result for each case value and phi node in the
3092 // common destination, as well as the the min and max case values.
3093 assert(SI->case_begin() != SI->case_end());
3094 SwitchInst::CaseIt CI = SI->case_begin();
3095 ConstantInt *MinCaseVal = CI.getCaseValue();
3096 ConstantInt *MaxCaseVal = CI.getCaseValue();
3098 BasicBlock *CommonDest = NULL;
3099 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3100 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3101 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3102 SmallDenseMap<PHINode*, Type*> ResultTypes;
3103 SmallVector<PHINode*, 4> PHIs;
3105 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3106 ConstantInt *CaseVal = CI.getCaseValue();
3107 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3108 MinCaseVal = CaseVal;
3109 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3110 MaxCaseVal = CaseVal;
3112 // Resulting value at phi nodes for this case value.
3113 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3115 if (!GetCaseResults(SI, CI.getCaseSuccessor(), &CommonDest, Results))
3118 // Append the result from this case to the list for each phi.
3119 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3120 if (!ResultLists.count(I->first))
3121 PHIs.push_back(I->first);
3122 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3126 // Get the resulting values for the default case.
3127 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3128 if (!GetCaseResults(SI, SI->getDefaultDest(), &CommonDest, DefaultResultsList))
3130 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3131 PHINode *PHI = DefaultResultsList[I].first;
3132 Constant *Result = DefaultResultsList[I].second;
3133 DefaultResults[PHI] = Result;
3134 ResultTypes[PHI] = Result->getType();
3137 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3138 // The table density should be at lest 40%. This is the same criterion as for
3139 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3140 // FIXME: Find the best cut-off.
3141 // Be careful to avoid overlow in the density computation.
3142 if (RangeSpread.zextOrSelf(64).ugt(UINT64_MAX / 4 - 1))
3144 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3145 if (SI->getNumCases() * 10 < TableSize * 4)
3148 // Build the lookup tables.
3149 SmallDenseMap<PHINode*, GlobalVariable*> LookupTables;
3150 SmallDenseMap<PHINode*, Constant*> SingleResults;
3152 Module &Mod = *CommonDest->getParent()->getParent();
3153 for (SmallVector<PHINode*, 4>::iterator I = PHIs.begin(), E = PHIs.end();
3157 Constant *SingleResult = NULL;
3158 LookupTables[PHI] = BuildLookupTable(Mod, TableSize, MinCaseVal,
3159 ResultLists[PHI], DefaultResults[PHI],
3161 SingleResults[PHI] = SingleResult;
3164 // Create the BB that does the lookups.
3165 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3167 CommonDest->getParent(),
3170 // Check whether the condition value is within the case range, and branch to
3172 Builder.SetInsertPoint(SI);
3173 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3175 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3176 MinCaseVal->getType(), TableSize));
3177 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3179 // Populate the BB that does the lookups.
3180 Builder.SetInsertPoint(LookupBB);
3181 bool ReturnedEarly = false;
3182 for (SmallVector<PHINode*, 4>::iterator I = PHIs.begin(), E = PHIs.end();
3185 // There was a single result for this phi; just use that.
3186 if (Constant *SingleResult = SingleResults[PHI]) {
3187 PHI->addIncoming(SingleResult, LookupBB);
3191 Value *GEPIndices[] = { Builder.getInt32(0), TableIndex };
3192 Value *GEP = Builder.CreateInBoundsGEP(LookupTables[PHI], GEPIndices,
3194 Value *Result = Builder.CreateLoad(GEP, "switch.load");
3196 // If the result is only going to be used to return from the function,
3197 // we want to do that right here.
3198 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin())) {
3199 if (CommonDest->getFirstNonPHIOrDbg() == CommonDest->getTerminator()) {
3200 Builder.CreateRet(Result);
3201 ReturnedEarly = true;
3206 PHI->addIncoming(Result, LookupBB);
3210 Builder.CreateBr(CommonDest);
3212 // Remove the switch.
3213 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3214 BasicBlock *Succ = SI->getSuccessor(i);
3215 if (Succ == SI->getDefaultDest()) continue;
3216 Succ->removePredecessor(SI->getParent());
3218 SI->eraseFromParent();
3224 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3225 // If this switch is too complex to want to look at, ignore it.
3226 if (!isValueEqualityComparison(SI))
3229 BasicBlock *BB = SI->getParent();
3231 // If we only have one predecessor, and if it is a branch on this value,
3232 // see if that predecessor totally determines the outcome of this switch.
3233 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3234 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3235 return SimplifyCFG(BB) | true;
3237 Value *Cond = SI->getCondition();
3238 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3239 if (SimplifySwitchOnSelect(SI, Select))
3240 return SimplifyCFG(BB) | true;
3242 // If the block only contains the switch, see if we can fold the block
3243 // away into any preds.
3244 BasicBlock::iterator BBI = BB->begin();
3245 // Ignore dbg intrinsics.
3246 while (isa<DbgInfoIntrinsic>(BBI))
3249 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3250 return SimplifyCFG(BB) | true;
3252 // Try to transform the switch into an icmp and a branch.
3253 if (TurnSwitchRangeIntoICmp(SI, Builder))
3254 return SimplifyCFG(BB) | true;
3256 // Remove unreachable cases.
3257 if (EliminateDeadSwitchCases(SI))
3258 return SimplifyCFG(BB) | true;
3260 if (ForwardSwitchConditionToPHI(SI))
3261 return SimplifyCFG(BB) | true;
3263 if (SwitchToLookupTable(SI, Builder))
3264 return SimplifyCFG(BB) | true;
3269 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3270 BasicBlock *BB = IBI->getParent();
3271 bool Changed = false;
3273 // Eliminate redundant destinations.
3274 SmallPtrSet<Value *, 8> Succs;
3275 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3276 BasicBlock *Dest = IBI->getDestination(i);
3277 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3278 Dest->removePredecessor(BB);
3279 IBI->removeDestination(i);
3285 if (IBI->getNumDestinations() == 0) {
3286 // If the indirectbr has no successors, change it to unreachable.
3287 new UnreachableInst(IBI->getContext(), IBI);
3288 EraseTerminatorInstAndDCECond(IBI);
3292 if (IBI->getNumDestinations() == 1) {
3293 // If the indirectbr has one successor, change it to a direct branch.
3294 BranchInst::Create(IBI->getDestination(0), IBI);
3295 EraseTerminatorInstAndDCECond(IBI);
3299 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3300 if (SimplifyIndirectBrOnSelect(IBI, SI))
3301 return SimplifyCFG(BB) | true;
3306 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3307 BasicBlock *BB = BI->getParent();
3309 // If the Terminator is the only non-phi instruction, simplify the block.
3310 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3311 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3312 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3315 // If the only instruction in the block is a seteq/setne comparison
3316 // against a constant, try to simplify the block.
3317 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3318 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3319 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3321 if (I->isTerminator() &&
3322 TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder))
3326 // If this basic block is ONLY a compare and a branch, and if a predecessor
3327 // branches to us and our successor, fold the comparison into the
3328 // predecessor and use logical operations to update the incoming value
3329 // for PHI nodes in common successor.
3330 if (FoldBranchToCommonDest(BI))
3331 return SimplifyCFG(BB) | true;
3336 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3337 BasicBlock *BB = BI->getParent();
3339 // Conditional branch
3340 if (isValueEqualityComparison(BI)) {
3341 // If we only have one predecessor, and if it is a branch on this value,
3342 // see if that predecessor totally determines the outcome of this
3344 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3345 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3346 return SimplifyCFG(BB) | true;
3348 // This block must be empty, except for the setcond inst, if it exists.
3349 // Ignore dbg intrinsics.
3350 BasicBlock::iterator I = BB->begin();
3351 // Ignore dbg intrinsics.
3352 while (isa<DbgInfoIntrinsic>(I))
3355 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3356 return SimplifyCFG(BB) | true;
3357 } else if (&*I == cast<Instruction>(BI->getCondition())){
3359 // Ignore dbg intrinsics.
3360 while (isa<DbgInfoIntrinsic>(I))
3362 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3363 return SimplifyCFG(BB) | true;
3367 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3368 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3371 // If this basic block is ONLY a compare and a branch, and if a predecessor
3372 // branches to us and one of our successors, fold the comparison into the
3373 // predecessor and use logical operations to pick the right destination.
3374 if (FoldBranchToCommonDest(BI))
3375 return SimplifyCFG(BB) | true;
3377 // We have a conditional branch to two blocks that are only reachable
3378 // from BI. We know that the condbr dominates the two blocks, so see if
3379 // there is any identical code in the "then" and "else" blocks. If so, we
3380 // can hoist it up to the branching block.
3381 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3382 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3383 if (HoistThenElseCodeToIf(BI))
3384 return SimplifyCFG(BB) | true;
3386 // If Successor #1 has multiple preds, we may be able to conditionally
3387 // execute Successor #0 if it branches to successor #1.
3388 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3389 if (Succ0TI->getNumSuccessors() == 1 &&
3390 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3391 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
3392 return SimplifyCFG(BB) | true;
3394 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3395 // If Successor #0 has multiple preds, we may be able to conditionally
3396 // execute Successor #1 if it branches to successor #0.
3397 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3398 if (Succ1TI->getNumSuccessors() == 1 &&
3399 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
3400 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
3401 return SimplifyCFG(BB) | true;
3404 // If this is a branch on a phi node in the current block, thread control
3405 // through this block if any PHI node entries are constants.
3406 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
3407 if (PN->getParent() == BI->getParent())
3408 if (FoldCondBranchOnPHI(BI, TD))
3409 return SimplifyCFG(BB) | true;
3411 // Scan predecessor blocks for conditional branches.
3412 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
3413 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
3414 if (PBI != BI && PBI->isConditional())
3415 if (SimplifyCondBranchToCondBranch(PBI, BI))
3416 return SimplifyCFG(BB) | true;
3421 /// Check if passing a value to an instruction will cause undefined behavior.
3422 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
3423 Constant *C = dyn_cast<Constant>(V);
3427 if (!I->hasOneUse()) // Only look at single-use instructions, for compile time
3430 if (C->isNullValue()) {
3431 Instruction *Use = I->use_back();
3433 // Now make sure that there are no instructions in between that can alter
3434 // control flow (eg. calls)
3435 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
3436 if (i == I->getParent()->end() || i->mayHaveSideEffects())
3439 // Look through GEPs. A load from a GEP derived from NULL is still undefined
3440 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
3441 if (GEP->getPointerOperand() == I)
3442 return passingValueIsAlwaysUndefined(V, GEP);
3444 // Look through bitcasts.
3445 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
3446 return passingValueIsAlwaysUndefined(V, BC);
3448 // Load from null is undefined.
3449 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
3450 return LI->getPointerAddressSpace() == 0;
3452 // Store to null is undefined.
3453 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
3454 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
3459 /// If BB has an incoming value that will always trigger undefined behavior
3460 /// (eg. null pointer dereference), remove the branch leading here.
3461 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
3462 for (BasicBlock::iterator i = BB->begin();
3463 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
3464 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
3465 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
3466 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
3467 IRBuilder<> Builder(T);
3468 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
3469 BB->removePredecessor(PHI->getIncomingBlock(i));
3470 // Turn uncoditional branches into unreachables and remove the dead
3471 // destination from conditional branches.
3472 if (BI->isUnconditional())
3473 Builder.CreateUnreachable();
3475 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
3476 BI->getSuccessor(0));
3477 BI->eraseFromParent();
3480 // TODO: SwitchInst.
3486 bool SimplifyCFGOpt::run(BasicBlock *BB) {
3487 bool Changed = false;
3489 assert(BB && BB->getParent() && "Block not embedded in function!");
3490 assert(BB->getTerminator() && "Degenerate basic block encountered!");
3492 // Remove basic blocks that have no predecessors (except the entry block)...
3493 // or that just have themself as a predecessor. These are unreachable.
3494 if ((pred_begin(BB) == pred_end(BB) &&
3495 BB != &BB->getParent()->getEntryBlock()) ||
3496 BB->getSinglePredecessor() == BB) {
3497 DEBUG(dbgs() << "Removing BB: \n" << *BB);
3498 DeleteDeadBlock(BB);
3502 // Check to see if we can constant propagate this terminator instruction
3504 Changed |= ConstantFoldTerminator(BB, true);
3506 // Check for and eliminate duplicate PHI nodes in this block.
3507 Changed |= EliminateDuplicatePHINodes(BB);
3509 // Check for and remove branches that will always cause undefined behavior.
3510 Changed |= removeUndefIntroducingPredecessor(BB);
3512 // Merge basic blocks into their predecessor if there is only one distinct
3513 // pred, and if there is only one distinct successor of the predecessor, and
3514 // if there are no PHI nodes.
3516 if (MergeBlockIntoPredecessor(BB))
3519 IRBuilder<> Builder(BB);
3521 // If there is a trivial two-entry PHI node in this basic block, and we can
3522 // eliminate it, do so now.
3523 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
3524 if (PN->getNumIncomingValues() == 2)
3525 Changed |= FoldTwoEntryPHINode(PN, TD);
3527 Builder.SetInsertPoint(BB->getTerminator());
3528 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
3529 if (BI->isUnconditional()) {
3530 if (SimplifyUncondBranch(BI, Builder)) return true;
3532 if (SimplifyCondBranch(BI, Builder)) return true;
3534 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
3535 if (SimplifyReturn(RI, Builder)) return true;
3536 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
3537 if (SimplifyResume(RI, Builder)) return true;
3538 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
3539 if (SimplifySwitch(SI, Builder)) return true;
3540 } else if (UnreachableInst *UI =
3541 dyn_cast<UnreachableInst>(BB->getTerminator())) {
3542 if (SimplifyUnreachable(UI)) return true;
3543 } else if (IndirectBrInst *IBI =
3544 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
3545 if (SimplifyIndirectBr(IBI)) return true;
3551 /// SimplifyCFG - This function is used to do simplification of a CFG. For
3552 /// example, it adjusts branches to branches to eliminate the extra hop, it
3553 /// eliminates unreachable basic blocks, and does other "peephole" optimization
3554 /// of the CFG. It returns true if a modification was made.
3556 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetData *TD) {
3557 return SimplifyCFGOpt(TD).run(BB);