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/ADT/DenseMap.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/GlobalVariable.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/MDBuilder.h"
34 #include "llvm/IR/Metadata.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/IR/Type.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/ConstantRange.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/NoFolder.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 static cl::opt<unsigned>
51 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
52 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
55 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
56 cl::desc("Duplicate return instructions into unconditional branches"));
59 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
60 cl::desc("Sink common instructions down to the end block"));
62 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
63 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
64 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
65 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
68 /// ValueEqualityComparisonCase - Represents a case of a switch.
69 struct ValueEqualityComparisonCase {
73 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
74 : Value(Value), Dest(Dest) {}
76 bool operator<(ValueEqualityComparisonCase RHS) const {
77 // Comparing pointers is ok as we only rely on the order for uniquing.
78 return Value < RHS.Value;
81 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
84 class SimplifyCFGOpt {
85 const TargetTransformInfo &TTI;
86 const DataLayout *const TD;
88 Value *isValueEqualityComparison(TerminatorInst *TI);
89 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
90 std::vector<ValueEqualityComparisonCase> &Cases);
91 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
93 IRBuilder<> &Builder);
94 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
95 IRBuilder<> &Builder);
97 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
98 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
99 bool SimplifyUnreachable(UnreachableInst *UI);
100 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
101 bool SimplifyIndirectBr(IndirectBrInst *IBI);
102 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
103 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
106 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *TD)
107 : TTI(TTI), TD(TD) {}
108 bool run(BasicBlock *BB);
112 /// SafeToMergeTerminators - Return true if it is safe to merge these two
113 /// terminator instructions together.
115 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
116 if (SI1 == SI2) return false; // Can't merge with self!
118 // It is not safe to merge these two switch instructions if they have a common
119 // successor, and if that successor has a PHI node, and if *that* PHI node has
120 // conflicting incoming values from the two switch blocks.
121 BasicBlock *SI1BB = SI1->getParent();
122 BasicBlock *SI2BB = SI2->getParent();
123 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
125 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
126 if (SI1Succs.count(*I))
127 for (BasicBlock::iterator BBI = (*I)->begin();
128 isa<PHINode>(BBI); ++BBI) {
129 PHINode *PN = cast<PHINode>(BBI);
130 if (PN->getIncomingValueForBlock(SI1BB) !=
131 PN->getIncomingValueForBlock(SI2BB))
138 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
139 /// to merge these two terminator instructions together, where SI1 is an
140 /// unconditional branch. PhiNodes will store all PHI nodes in common
143 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
146 SmallVectorImpl<PHINode*> &PhiNodes) {
147 if (SI1 == SI2) return false; // Can't merge with self!
148 assert(SI1->isUnconditional() && SI2->isConditional());
150 // We fold the unconditional branch if we can easily update all PHI nodes in
151 // common successors:
152 // 1> We have a constant incoming value for the conditional branch;
153 // 2> We have "Cond" as the incoming value for the unconditional branch;
154 // 3> SI2->getCondition() and Cond have same operands.
155 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
156 if (!Ci2) return false;
157 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
158 Cond->getOperand(1) == Ci2->getOperand(1)) &&
159 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
160 Cond->getOperand(1) == Ci2->getOperand(0)))
163 BasicBlock *SI1BB = SI1->getParent();
164 BasicBlock *SI2BB = SI2->getParent();
165 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
166 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
167 if (SI1Succs.count(*I))
168 for (BasicBlock::iterator BBI = (*I)->begin();
169 isa<PHINode>(BBI); ++BBI) {
170 PHINode *PN = cast<PHINode>(BBI);
171 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
172 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
174 PhiNodes.push_back(PN);
179 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
180 /// now be entries in it from the 'NewPred' block. The values that will be
181 /// flowing into the PHI nodes will be the same as those coming in from
182 /// ExistPred, an existing predecessor of Succ.
183 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
184 BasicBlock *ExistPred) {
185 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
188 for (BasicBlock::iterator I = Succ->begin();
189 (PN = dyn_cast<PHINode>(I)); ++I)
190 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
194 /// GetIfCondition - Given a basic block (BB) with two predecessors (and at
195 /// least one PHI node in it), check to see if the merge at this block is due
196 /// to an "if condition". If so, return the boolean condition that determines
197 /// which entry into BB will be taken. Also, return by references the block
198 /// that will be entered from if the condition is true, and the block that will
199 /// be entered if the condition is false.
201 /// This does no checking to see if the true/false blocks have large or unsavory
202 /// instructions in them.
203 static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue,
204 BasicBlock *&IfFalse) {
205 PHINode *SomePHI = cast<PHINode>(BB->begin());
206 assert(SomePHI->getNumIncomingValues() == 2 &&
207 "Function can only handle blocks with 2 predecessors!");
208 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0);
209 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1);
211 // We can only handle branches. Other control flow will be lowered to
212 // branches if possible anyway.
213 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator());
214 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator());
215 if (Pred1Br == 0 || Pred2Br == 0)
218 // Eliminate code duplication by ensuring that Pred1Br is conditional if
220 if (Pred2Br->isConditional()) {
221 // If both branches are conditional, we don't have an "if statement". In
222 // reality, we could transform this case, but since the condition will be
223 // required anyway, we stand no chance of eliminating it, so the xform is
224 // probably not profitable.
225 if (Pred1Br->isConditional())
228 std::swap(Pred1, Pred2);
229 std::swap(Pred1Br, Pred2Br);
232 if (Pred1Br->isConditional()) {
233 // The only thing we have to watch out for here is to make sure that Pred2
234 // doesn't have incoming edges from other blocks. If it does, the condition
235 // doesn't dominate BB.
236 if (Pred2->getSinglePredecessor() == 0)
239 // If we found a conditional branch predecessor, make sure that it branches
240 // to BB and Pred2Br. If it doesn't, this isn't an "if statement".
241 if (Pred1Br->getSuccessor(0) == BB &&
242 Pred1Br->getSuccessor(1) == Pred2) {
245 } else if (Pred1Br->getSuccessor(0) == Pred2 &&
246 Pred1Br->getSuccessor(1) == BB) {
250 // We know that one arm of the conditional goes to BB, so the other must
251 // go somewhere unrelated, and this must not be an "if statement".
255 return Pred1Br->getCondition();
258 // Ok, if we got here, both predecessors end with an unconditional branch to
259 // BB. Don't panic! If both blocks only have a single (identical)
260 // predecessor, and THAT is a conditional branch, then we're all ok!
261 BasicBlock *CommonPred = Pred1->getSinglePredecessor();
262 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor())
265 // Otherwise, if this is a conditional branch, then we can use it!
266 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator());
267 if (BI == 0) return 0;
269 assert(BI->isConditional() && "Two successors but not conditional?");
270 if (BI->getSuccessor(0) == Pred1) {
277 return BI->getCondition();
280 /// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the
281 /// given instruction, which is assumed to be safe to speculate. 1 means
282 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
283 static unsigned ComputeSpeculationCost(const User *I) {
284 assert(isSafeToSpeculativelyExecute(I) &&
285 "Instruction is not safe to speculatively execute!");
286 switch (Operator::getOpcode(I)) {
288 // In doubt, be conservative.
290 case Instruction::GetElementPtr:
291 // GEPs are cheap if all indices are constant.
292 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
295 case Instruction::Load:
296 case Instruction::Add:
297 case Instruction::Sub:
298 case Instruction::And:
299 case Instruction::Or:
300 case Instruction::Xor:
301 case Instruction::Shl:
302 case Instruction::LShr:
303 case Instruction::AShr:
304 case Instruction::ICmp:
305 case Instruction::Trunc:
306 case Instruction::ZExt:
307 case Instruction::SExt:
308 return 1; // These are all cheap.
310 case Instruction::Call:
311 case Instruction::Select:
316 /// DominatesMergePoint - If we have a merge point of an "if condition" as
317 /// accepted above, return true if the specified value dominates the block. We
318 /// don't handle the true generality of domination here, just a special case
319 /// which works well enough for us.
321 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
322 /// see if V (which must be an instruction) and its recursive operands
323 /// that do not dominate BB have a combined cost lower than CostRemaining and
324 /// are non-trapping. If both are true, the instruction is inserted into the
325 /// set and true is returned.
327 /// The cost for most non-trapping instructions is defined as 1 except for
328 /// Select whose cost is 2.
330 /// After this function returns, CostRemaining is decreased by the cost of
331 /// V plus its non-dominating operands. If that cost is greater than
332 /// CostRemaining, false is returned and CostRemaining is undefined.
333 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
334 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
335 unsigned &CostRemaining) {
336 Instruction *I = dyn_cast<Instruction>(V);
338 // Non-instructions all dominate instructions, but not all constantexprs
339 // can be executed unconditionally.
340 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
345 BasicBlock *PBB = I->getParent();
347 // We don't want to allow weird loops that might have the "if condition" in
348 // the bottom of this block.
349 if (PBB == BB) return false;
351 // If this instruction is defined in a block that contains an unconditional
352 // branch to BB, then it must be in the 'conditional' part of the "if
353 // statement". If not, it definitely dominates the region.
354 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
355 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB)
358 // If we aren't allowing aggressive promotion anymore, then don't consider
359 // instructions in the 'if region'.
360 if (AggressiveInsts == 0) return false;
362 // If we have seen this instruction before, don't count it again.
363 if (AggressiveInsts->count(I)) return true;
365 // Okay, it looks like the instruction IS in the "condition". Check to
366 // see if it's a cheap instruction to unconditionally compute, and if it
367 // only uses stuff defined outside of the condition. If so, hoist it out.
368 if (!isSafeToSpeculativelyExecute(I))
371 unsigned Cost = ComputeSpeculationCost(I);
373 if (Cost > CostRemaining)
376 CostRemaining -= Cost;
378 // Okay, we can only really hoist these out if their operands do
379 // not take us over the cost threshold.
380 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
381 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining))
383 // Okay, it's safe to do this! Remember this instruction.
384 AggressiveInsts->insert(I);
388 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
389 /// and PointerNullValue. Return NULL if value is not a constant int.
390 static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) {
391 // Normal constant int.
392 ConstantInt *CI = dyn_cast<ConstantInt>(V);
393 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy())
396 // This is some kind of pointer constant. Turn it into a pointer-sized
397 // ConstantInt if possible.
398 IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType()));
400 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
401 if (isa<ConstantPointerNull>(V))
402 return ConstantInt::get(PtrTy, 0);
404 // IntToPtr const int.
405 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
406 if (CE->getOpcode() == Instruction::IntToPtr)
407 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
408 // The constant is very likely to have the right type already.
409 if (CI->getType() == PtrTy)
412 return cast<ConstantInt>
413 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
418 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
419 /// collection of icmp eq/ne instructions that compare a value against a
420 /// constant, return the value being compared, and stick the constant into the
423 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
424 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) {
425 Instruction *I = dyn_cast<Instruction>(V);
426 if (I == 0) return 0;
428 // If this is an icmp against a constant, handle this as one of the cases.
429 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
430 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) {
431 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
434 return I->getOperand(0);
437 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
440 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
442 // If this is an and/!= check then we want to optimize "x ugt 2" into
445 Span = Span.inverse();
447 // If there are a ton of values, we don't want to make a ginormous switch.
448 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
451 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
452 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
454 return I->getOperand(0);
459 // Otherwise, we can only handle an | or &, depending on isEQ.
460 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
463 unsigned NumValsBeforeLHS = Vals.size();
464 unsigned UsedICmpsBeforeLHS = UsedICmps;
465 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD,
467 unsigned NumVals = Vals.size();
468 unsigned UsedICmpsBeforeRHS = UsedICmps;
469 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
473 Vals.resize(NumVals);
474 UsedICmps = UsedICmpsBeforeRHS;
477 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
478 // set it and return success.
479 if (Extra == 0 || Extra == I->getOperand(1)) {
480 Extra = I->getOperand(1);
484 Vals.resize(NumValsBeforeLHS);
485 UsedICmps = UsedICmpsBeforeLHS;
489 // If the LHS can't be folded in, but Extra is available and RHS can, try to
491 if (Extra == 0 || Extra == I->getOperand(0)) {
492 Value *OldExtra = Extra;
493 Extra = I->getOperand(0);
494 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
497 assert(Vals.size() == NumValsBeforeLHS);
504 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
505 Instruction *Cond = 0;
506 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
507 Cond = dyn_cast<Instruction>(SI->getCondition());
508 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
509 if (BI->isConditional())
510 Cond = dyn_cast<Instruction>(BI->getCondition());
511 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
512 Cond = dyn_cast<Instruction>(IBI->getAddress());
515 TI->eraseFromParent();
516 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
519 /// isValueEqualityComparison - Return true if the specified terminator checks
520 /// to see if a value is equal to constant integer value.
521 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
523 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
524 // Do not permit merging of large switch instructions into their
525 // predecessors unless there is only one predecessor.
526 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
527 pred_end(SI->getParent())) <= 128)
528 CV = SI->getCondition();
529 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
530 if (BI->isConditional() && BI->getCondition()->hasOneUse())
531 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
532 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ ||
533 ICI->getPredicate() == ICmpInst::ICMP_NE) &&
534 GetConstantInt(ICI->getOperand(1), TD))
535 CV = ICI->getOperand(0);
537 // Unwrap any lossless ptrtoint cast.
538 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext()))
539 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV))
540 CV = PTII->getOperand(0);
544 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
545 /// decode all of the 'cases' that it represents and return the 'default' block.
546 BasicBlock *SimplifyCFGOpt::
547 GetValueEqualityComparisonCases(TerminatorInst *TI,
548 std::vector<ValueEqualityComparisonCase>
550 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
551 Cases.reserve(SI->getNumCases());
552 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
553 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
554 i.getCaseSuccessor()));
555 return SI->getDefaultDest();
558 BranchInst *BI = cast<BranchInst>(TI);
559 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
560 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
561 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
564 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
568 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
569 /// in the list that match the specified block.
570 static void EliminateBlockCases(BasicBlock *BB,
571 std::vector<ValueEqualityComparisonCase> &Cases) {
572 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
575 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
578 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
579 std::vector<ValueEqualityComparisonCase > &C2) {
580 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
582 // Make V1 be smaller than V2.
583 if (V1->size() > V2->size())
586 if (V1->size() == 0) return false;
587 if (V1->size() == 1) {
589 ConstantInt *TheVal = (*V1)[0].Value;
590 for (unsigned i = 0, e = V2->size(); i != e; ++i)
591 if (TheVal == (*V2)[i].Value)
595 // Otherwise, just sort both lists and compare element by element.
596 array_pod_sort(V1->begin(), V1->end());
597 array_pod_sort(V2->begin(), V2->end());
598 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
599 while (i1 != e1 && i2 != e2) {
600 if ((*V1)[i1].Value == (*V2)[i2].Value)
602 if ((*V1)[i1].Value < (*V2)[i2].Value)
610 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
611 /// terminator instruction and its block is known to only have a single
612 /// predecessor block, check to see if that predecessor is also a value
613 /// comparison with the same value, and if that comparison determines the
614 /// outcome of this comparison. If so, simplify TI. This does a very limited
615 /// form of jump threading.
616 bool SimplifyCFGOpt::
617 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
619 IRBuilder<> &Builder) {
620 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
621 if (!PredVal) return false; // Not a value comparison in predecessor.
623 Value *ThisVal = isValueEqualityComparison(TI);
624 assert(ThisVal && "This isn't a value comparison!!");
625 if (ThisVal != PredVal) return false; // Different predicates.
627 // TODO: Preserve branch weight metadata, similarly to how
628 // FoldValueComparisonIntoPredecessors preserves it.
630 // Find out information about when control will move from Pred to TI's block.
631 std::vector<ValueEqualityComparisonCase> PredCases;
632 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
634 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
636 // Find information about how control leaves this block.
637 std::vector<ValueEqualityComparisonCase> ThisCases;
638 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
639 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
641 // If TI's block is the default block from Pred's comparison, potentially
642 // simplify TI based on this knowledge.
643 if (PredDef == TI->getParent()) {
644 // If we are here, we know that the value is none of those cases listed in
645 // PredCases. If there are any cases in ThisCases that are in PredCases, we
647 if (!ValuesOverlap(PredCases, ThisCases))
650 if (isa<BranchInst>(TI)) {
651 // Okay, one of the successors of this condbr is dead. Convert it to a
653 assert(ThisCases.size() == 1 && "Branch can only have one case!");
654 // Insert the new branch.
655 Instruction *NI = Builder.CreateBr(ThisDef);
658 // Remove PHI node entries for the dead edge.
659 ThisCases[0].Dest->removePredecessor(TI->getParent());
661 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
662 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
664 EraseTerminatorInstAndDCECond(TI);
668 SwitchInst *SI = cast<SwitchInst>(TI);
669 // Okay, TI has cases that are statically dead, prune them away.
670 SmallPtrSet<Constant*, 16> DeadCases;
671 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
672 DeadCases.insert(PredCases[i].Value);
674 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
675 << "Through successor TI: " << *TI);
677 // Collect branch weights into a vector.
678 SmallVector<uint32_t, 8> Weights;
679 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
680 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
682 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
684 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
686 Weights.push_back(CI->getValue().getZExtValue());
688 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
690 if (DeadCases.count(i.getCaseValue())) {
692 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
695 i.getCaseSuccessor()->removePredecessor(TI->getParent());
699 if (HasWeight && Weights.size() >= 2)
700 SI->setMetadata(LLVMContext::MD_prof,
701 MDBuilder(SI->getParent()->getContext()).
702 createBranchWeights(Weights));
704 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
708 // Otherwise, TI's block must correspond to some matched value. Find out
709 // which value (or set of values) this is.
710 ConstantInt *TIV = 0;
711 BasicBlock *TIBB = TI->getParent();
712 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
713 if (PredCases[i].Dest == TIBB) {
715 return false; // Cannot handle multiple values coming to this block.
716 TIV = PredCases[i].Value;
718 assert(TIV && "No edge from pred to succ?");
720 // Okay, we found the one constant that our value can be if we get into TI's
721 // BB. Find out which successor will unconditionally be branched to.
722 BasicBlock *TheRealDest = 0;
723 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
724 if (ThisCases[i].Value == TIV) {
725 TheRealDest = ThisCases[i].Dest;
729 // If not handled by any explicit cases, it is handled by the default case.
730 if (TheRealDest == 0) TheRealDest = ThisDef;
732 // Remove PHI node entries for dead edges.
733 BasicBlock *CheckEdge = TheRealDest;
734 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
735 if (*SI != CheckEdge)
736 (*SI)->removePredecessor(TIBB);
740 // Insert the new branch.
741 Instruction *NI = Builder.CreateBr(TheRealDest);
744 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
745 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
747 EraseTerminatorInstAndDCECond(TI);
752 /// ConstantIntOrdering - This class implements a stable ordering of constant
753 /// integers that does not depend on their address. This is important for
754 /// applications that sort ConstantInt's to ensure uniqueness.
755 struct ConstantIntOrdering {
756 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
757 return LHS->getValue().ult(RHS->getValue());
762 static int ConstantIntSortPredicate(const void *P1, const void *P2) {
763 const ConstantInt *LHS = *(const ConstantInt*const*)P1;
764 const ConstantInt *RHS = *(const ConstantInt*const*)P2;
765 if (LHS->getValue().ult(RHS->getValue()))
767 if (LHS->getValue() == RHS->getValue())
772 static inline bool HasBranchWeights(const Instruction* I) {
773 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
774 if (ProfMD && ProfMD->getOperand(0))
775 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
776 return MDS->getString().equals("branch_weights");
781 /// Get Weights of a given TerminatorInst, the default weight is at the front
782 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
784 static void GetBranchWeights(TerminatorInst *TI,
785 SmallVectorImpl<uint64_t> &Weights) {
786 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
788 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
789 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i));
791 Weights.push_back(CI->getValue().getZExtValue());
794 // If TI is a conditional eq, the default case is the false case,
795 // and the corresponding branch-weight data is at index 2. We swap the
796 // default weight to be the first entry.
797 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
798 assert(Weights.size() == 2);
799 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
800 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
801 std::swap(Weights.front(), Weights.back());
805 /// Sees if any of the weights are too big for a uint32_t, and halves all the
806 /// weights if any are.
807 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
809 for (unsigned i = 0; i < Weights.size(); ++i)
810 if (Weights[i] > UINT_MAX) {
818 for (unsigned i = 0; i < Weights.size(); ++i)
822 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
823 /// equality comparison instruction (either a switch or a branch on "X == c").
824 /// See if any of the predecessors of the terminator block are value comparisons
825 /// on the same value. If so, and if safe to do so, fold them together.
826 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
827 IRBuilder<> &Builder) {
828 BasicBlock *BB = TI->getParent();
829 Value *CV = isValueEqualityComparison(TI); // CondVal
830 assert(CV && "Not a comparison?");
831 bool Changed = false;
833 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
834 while (!Preds.empty()) {
835 BasicBlock *Pred = Preds.pop_back_val();
837 // See if the predecessor is a comparison with the same value.
838 TerminatorInst *PTI = Pred->getTerminator();
839 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
841 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
842 // Figure out which 'cases' to copy from SI to PSI.
843 std::vector<ValueEqualityComparisonCase> BBCases;
844 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
846 std::vector<ValueEqualityComparisonCase> PredCases;
847 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
849 // Based on whether the default edge from PTI goes to BB or not, fill in
850 // PredCases and PredDefault with the new switch cases we would like to
852 SmallVector<BasicBlock*, 8> NewSuccessors;
854 // Update the branch weight metadata along the way
855 SmallVector<uint64_t, 8> Weights;
856 bool PredHasWeights = HasBranchWeights(PTI);
857 bool SuccHasWeights = HasBranchWeights(TI);
859 if (PredHasWeights) {
860 GetBranchWeights(PTI, Weights);
861 // branch-weight metadata is inconsistent here.
862 if (Weights.size() != 1 + PredCases.size())
863 PredHasWeights = SuccHasWeights = false;
864 } else if (SuccHasWeights)
865 // If there are no predecessor weights but there are successor weights,
866 // populate Weights with 1, which will later be scaled to the sum of
867 // successor's weights
868 Weights.assign(1 + PredCases.size(), 1);
870 SmallVector<uint64_t, 8> SuccWeights;
871 if (SuccHasWeights) {
872 GetBranchWeights(TI, SuccWeights);
873 // branch-weight metadata is inconsistent here.
874 if (SuccWeights.size() != 1 + BBCases.size())
875 PredHasWeights = SuccHasWeights = false;
876 } else if (PredHasWeights)
877 SuccWeights.assign(1 + BBCases.size(), 1);
879 if (PredDefault == BB) {
880 // If this is the default destination from PTI, only the edges in TI
881 // that don't occur in PTI, or that branch to BB will be activated.
882 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
883 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
884 if (PredCases[i].Dest != BB)
885 PTIHandled.insert(PredCases[i].Value);
887 // The default destination is BB, we don't need explicit targets.
888 std::swap(PredCases[i], PredCases.back());
890 if (PredHasWeights || SuccHasWeights) {
891 // Increase weight for the default case.
892 Weights[0] += Weights[i+1];
893 std::swap(Weights[i+1], Weights.back());
897 PredCases.pop_back();
901 // Reconstruct the new switch statement we will be building.
902 if (PredDefault != BBDefault) {
903 PredDefault->removePredecessor(Pred);
904 PredDefault = BBDefault;
905 NewSuccessors.push_back(BBDefault);
908 unsigned CasesFromPred = Weights.size();
909 uint64_t ValidTotalSuccWeight = 0;
910 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
911 if (!PTIHandled.count(BBCases[i].Value) &&
912 BBCases[i].Dest != BBDefault) {
913 PredCases.push_back(BBCases[i]);
914 NewSuccessors.push_back(BBCases[i].Dest);
915 if (SuccHasWeights || PredHasWeights) {
916 // The default weight is at index 0, so weight for the ith case
917 // should be at index i+1. Scale the cases from successor by
918 // PredDefaultWeight (Weights[0]).
919 Weights.push_back(Weights[0] * SuccWeights[i+1]);
920 ValidTotalSuccWeight += SuccWeights[i+1];
924 if (SuccHasWeights || PredHasWeights) {
925 ValidTotalSuccWeight += SuccWeights[0];
926 // Scale the cases from predecessor by ValidTotalSuccWeight.
927 for (unsigned i = 1; i < CasesFromPred; ++i)
928 Weights[i] *= ValidTotalSuccWeight;
929 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
930 Weights[0] *= SuccWeights[0];
933 // If this is not the default destination from PSI, only the edges
934 // in SI that occur in PSI with a destination of BB will be
936 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
937 std::map<ConstantInt*, uint64_t> WeightsForHandled;
938 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
939 if (PredCases[i].Dest == BB) {
940 PTIHandled.insert(PredCases[i].Value);
942 if (PredHasWeights || SuccHasWeights) {
943 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
944 std::swap(Weights[i+1], Weights.back());
948 std::swap(PredCases[i], PredCases.back());
949 PredCases.pop_back();
953 // Okay, now we know which constants were sent to BB from the
954 // predecessor. Figure out where they will all go now.
955 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
956 if (PTIHandled.count(BBCases[i].Value)) {
957 // If this is one we are capable of getting...
958 if (PredHasWeights || SuccHasWeights)
959 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
960 PredCases.push_back(BBCases[i]);
961 NewSuccessors.push_back(BBCases[i].Dest);
962 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
965 // If there are any constants vectored to BB that TI doesn't handle,
966 // they must go to the default destination of TI.
967 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
969 E = PTIHandled.end(); I != E; ++I) {
970 if (PredHasWeights || SuccHasWeights)
971 Weights.push_back(WeightsForHandled[*I]);
972 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
973 NewSuccessors.push_back(BBDefault);
977 // Okay, at this point, we know which new successor Pred will get. Make
978 // sure we update the number of entries in the PHI nodes for these
980 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
981 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
983 Builder.SetInsertPoint(PTI);
984 // Convert pointer to int before we switch.
985 if (CV->getType()->isPointerTy()) {
986 assert(TD && "Cannot switch on pointer without DataLayout");
987 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()),
991 // Now that the successors are updated, create the new Switch instruction.
992 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
994 NewSI->setDebugLoc(PTI->getDebugLoc());
995 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
996 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
998 if (PredHasWeights || SuccHasWeights) {
999 // Halve the weights if any of them cannot fit in an uint32_t
1000 FitWeights(Weights);
1002 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1004 NewSI->setMetadata(LLVMContext::MD_prof,
1005 MDBuilder(BB->getContext()).
1006 createBranchWeights(MDWeights));
1009 EraseTerminatorInstAndDCECond(PTI);
1011 // Okay, last check. If BB is still a successor of PSI, then we must
1012 // have an infinite loop case. If so, add an infinitely looping block
1013 // to handle the case to preserve the behavior of the code.
1014 BasicBlock *InfLoopBlock = 0;
1015 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1016 if (NewSI->getSuccessor(i) == BB) {
1017 if (InfLoopBlock == 0) {
1018 // Insert it at the end of the function, because it's either code,
1019 // or it won't matter if it's hot. :)
1020 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1021 "infloop", BB->getParent());
1022 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1024 NewSI->setSuccessor(i, InfLoopBlock);
1033 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1034 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1035 // would need to do this), we can't hoist the invoke, as there is nowhere
1036 // to put the select in this case.
1037 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1038 Instruction *I1, Instruction *I2) {
1039 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1041 for (BasicBlock::iterator BBI = SI->begin();
1042 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1043 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1044 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1045 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1053 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1054 /// BB2, hoist any common code in the two blocks up into the branch block. The
1055 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1056 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1057 // This does very trivial matching, with limited scanning, to find identical
1058 // instructions in the two blocks. In particular, we don't want to get into
1059 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1060 // such, we currently just scan for obviously identical instructions in an
1062 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1063 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1065 BasicBlock::iterator BB1_Itr = BB1->begin();
1066 BasicBlock::iterator BB2_Itr = BB2->begin();
1068 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1069 // Skip debug info if it is not identical.
1070 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1071 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1072 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1073 while (isa<DbgInfoIntrinsic>(I1))
1075 while (isa<DbgInfoIntrinsic>(I2))
1078 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1079 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1082 // If we get here, we can hoist at least one instruction.
1083 BasicBlock *BIParent = BI->getParent();
1086 // If we are hoisting the terminator instruction, don't move one (making a
1087 // broken BB), instead clone it, and remove BI.
1088 if (isa<TerminatorInst>(I1))
1089 goto HoistTerminator;
1091 // For a normal instruction, we just move one to right before the branch,
1092 // then replace all uses of the other with the first. Finally, we remove
1093 // the now redundant second instruction.
1094 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1095 if (!I2->use_empty())
1096 I2->replaceAllUsesWith(I1);
1097 I1->intersectOptionalDataWith(I2);
1098 I2->eraseFromParent();
1102 // Skip debug info if it is not identical.
1103 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1104 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1105 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1106 while (isa<DbgInfoIntrinsic>(I1))
1108 while (isa<DbgInfoIntrinsic>(I2))
1111 } while (I1->isIdenticalToWhenDefined(I2));
1116 // It may not be possible to hoist an invoke.
1117 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1120 // Okay, it is safe to hoist the terminator.
1121 Instruction *NT = I1->clone();
1122 BIParent->getInstList().insert(BI, NT);
1123 if (!NT->getType()->isVoidTy()) {
1124 I1->replaceAllUsesWith(NT);
1125 I2->replaceAllUsesWith(NT);
1129 IRBuilder<true, NoFolder> Builder(NT);
1130 // Hoisting one of the terminators from our successor is a great thing.
1131 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1132 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1133 // nodes, so we insert select instruction to compute the final result.
1134 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1135 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1137 for (BasicBlock::iterator BBI = SI->begin();
1138 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1139 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1140 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1141 if (BB1V == BB2V) continue;
1143 // These values do not agree. Insert a select instruction before NT
1144 // that determines the right value.
1145 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1147 SI = cast<SelectInst>
1148 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1149 BB1V->getName()+"."+BB2V->getName()));
1151 // Make the PHI node use the select for all incoming values for BB1/BB2
1152 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1153 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1154 PN->setIncomingValue(i, SI);
1158 // Update any PHI nodes in our new successors.
1159 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1160 AddPredecessorToBlock(*SI, BIParent, BB1);
1162 EraseTerminatorInstAndDCECond(BI);
1166 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1167 /// check whether BBEnd has only two predecessors and the other predecessor
1168 /// ends with an unconditional branch. If it is true, sink any common code
1169 /// in the two predecessors to BBEnd.
1170 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1171 assert(BI1->isUnconditional());
1172 BasicBlock *BB1 = BI1->getParent();
1173 BasicBlock *BBEnd = BI1->getSuccessor(0);
1175 // Check that BBEnd has two predecessors and the other predecessor ends with
1176 // an unconditional branch.
1177 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1178 BasicBlock *Pred0 = *PI++;
1179 if (PI == PE) // Only one predecessor.
1181 BasicBlock *Pred1 = *PI++;
1182 if (PI != PE) // More than two predecessors.
1184 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1185 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1186 if (!BI2 || !BI2->isUnconditional())
1189 // Gather the PHI nodes in BBEnd.
1190 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1191 Instruction *FirstNonPhiInBBEnd = 0;
1192 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1194 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1195 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1196 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1197 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1199 FirstNonPhiInBBEnd = &*I;
1203 if (!FirstNonPhiInBBEnd)
1207 // This does very trivial matching, with limited scanning, to find identical
1208 // instructions in the two blocks. We scan backward for obviously identical
1209 // instructions in an identical order.
1210 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1211 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1212 RE2 = BB2->getInstList().rend();
1214 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1217 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1220 // Skip the unconditional branches.
1224 bool Changed = false;
1225 while (RI1 != RE1 && RI2 != RE2) {
1227 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1230 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1234 Instruction *I1 = &*RI1, *I2 = &*RI2;
1235 // I1 and I2 should have a single use in the same PHI node, and they
1236 // perform the same operation.
1237 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1238 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1239 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1240 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1241 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1242 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1243 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1244 !I1->hasOneUse() || !I2->hasOneUse() ||
1245 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1246 MapValueFromBB1ToBB2[I1].first != I2)
1249 // Check whether we should swap the operands of ICmpInst.
1250 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1251 bool SwapOpnds = false;
1252 if (ICmp1 && ICmp2 &&
1253 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1254 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1255 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1256 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1257 ICmp2->swapOperands();
1260 if (!I1->isSameOperationAs(I2)) {
1262 ICmp2->swapOperands();
1266 // The operands should be either the same or they need to be generated
1267 // with a PHI node after sinking. We only handle the case where there is
1268 // a single pair of different operands.
1269 Value *DifferentOp1 = 0, *DifferentOp2 = 0;
1270 unsigned Op1Idx = 0;
1271 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1272 if (I1->getOperand(I) == I2->getOperand(I))
1274 // Early exit if we have more-than one pair of different operands or
1275 // the different operand is already in MapValueFromBB1ToBB2.
1276 // Early exit if we need a PHI node to replace a constant.
1278 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1279 MapValueFromBB1ToBB2.end() ||
1280 isa<Constant>(I1->getOperand(I)) ||
1281 isa<Constant>(I2->getOperand(I))) {
1282 // If we can't sink the instructions, undo the swapping.
1284 ICmp2->swapOperands();
1287 DifferentOp1 = I1->getOperand(I);
1289 DifferentOp2 = I2->getOperand(I);
1292 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1293 // remove (I1, I2) from MapValueFromBB1ToBB2.
1295 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1296 DifferentOp1->getName() + ".sink",
1298 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1299 // I1 should use NewPN instead of DifferentOp1.
1300 I1->setOperand(Op1Idx, NewPN);
1301 NewPN->addIncoming(DifferentOp1, BB1);
1302 NewPN->addIncoming(DifferentOp2, BB2);
1303 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1305 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1306 MapValueFromBB1ToBB2.erase(I1);
1308 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1309 DEBUG(dbgs() << " " << *I2 << "\n";);
1310 // We need to update RE1 and RE2 if we are going to sink the first
1311 // instruction in the basic block down.
1312 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1313 // Sink the instruction.
1314 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1315 if (!OldPN->use_empty())
1316 OldPN->replaceAllUsesWith(I1);
1317 OldPN->eraseFromParent();
1319 if (!I2->use_empty())
1320 I2->replaceAllUsesWith(I1);
1321 I1->intersectOptionalDataWith(I2);
1322 I2->eraseFromParent();
1325 RE1 = BB1->getInstList().rend();
1327 RE2 = BB2->getInstList().rend();
1328 FirstNonPhiInBBEnd = I1;
1335 /// \brief Speculate a conditional basic block flattening the CFG.
1337 /// Note that this is a very risky transform currently. Speculating
1338 /// instructions like this is most often not desirable. Instead, there is an MI
1339 /// pass which can do it with full awareness of the resource constraints.
1340 /// However, some cases are "obvious" and we should do directly. An example of
1341 /// this is speculating a single, reasonably cheap instruction.
1343 /// There is only one distinct advantage to flattening the CFG at the IR level:
1344 /// it makes very common but simplistic optimizations such as are common in
1345 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1346 /// modeling their effects with easier to reason about SSA value graphs.
1349 /// An illustration of this transform is turning this IR:
1352 /// %cmp = icmp ult %x, %y
1353 /// br i1 %cmp, label %EndBB, label %ThenBB
1355 /// %sub = sub %x, %y
1358 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1365 /// %cmp = icmp ult %x, %y
1366 /// %sub = sub %x, %y
1367 /// %cond = select i1 %cmp, 0, %sub
1371 /// \returns true if the conditional block is removed.
1372 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB,
1373 const TargetTransformInfo &TTI) {
1374 // Be conservative for now. FP select instruction can often be expensive.
1375 Value *BrCond = BI->getCondition();
1376 if (isa<FCmpInst>(BrCond))
1379 BasicBlock *BB = BI->getParent();
1380 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1382 // If ThenBB is actually on the false edge of the conditional branch, remember
1383 // to swap the select operands later.
1384 bool Invert = false;
1385 if (ThenBB != BI->getSuccessor(0)) {
1386 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1389 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1391 // Keep a count of how many times instructions are used within CondBB when
1392 // they are candidates for sinking into CondBB. Specifically:
1393 // - They are defined in BB, and
1394 // - They have no side effects, and
1395 // - All of their uses are in CondBB.
1396 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1398 unsigned SpeculationCost = 0;
1399 for (BasicBlock::iterator BBI = ThenBB->begin(),
1400 BBE = llvm::prior(ThenBB->end());
1401 BBI != BBE; ++BBI) {
1402 Instruction *I = BBI;
1404 if (isa<DbgInfoIntrinsic>(I))
1407 // Only speculatively execution a single instruction (not counting the
1408 // terminator) for now.
1409 SpeculationCost += TTI.getUserCost(I);
1410 if (SpeculationCost > TargetTransformInfo::TCC_Basic)
1413 // Don't hoist the instruction if it's unsafe or expensive.
1414 if (!isSafeToSpeculativelyExecute(I))
1416 // FIXME: These should really be cost metrics, but our cost model doesn't
1417 // accurately model the expense of selects and floating point operations.
1418 // FIXME: Is it really safe to speculate floating point operations?
1419 // Signaling NaNs break with this, but we shouldn't care, right?
1420 if (isa<SelectInst>(I) || I->getType()->isFPOrFPVectorTy())
1422 // FIXME: The cost metric currently doesn't reason accurately about simple
1423 // versus complex GEPs, take a conservative approach here.
1424 if (GEPOperator *GEP = dyn_cast<GEPOperator>(I))
1425 if (!GEP->hasAllConstantIndices())
1428 // Do not hoist the instruction if any of its operands are defined but not
1429 // used in this BB. The transformation will prevent the operand from
1430 // being sunk into the use block.
1431 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1433 Instruction *OpI = dyn_cast<Instruction>(*i);
1434 if (!OpI || OpI->getParent() != BB ||
1435 OpI->mayHaveSideEffects())
1436 continue; // Not a candidate for sinking.
1438 ++SinkCandidateUseCounts[OpI];
1442 // Consider any sink candidates which are only used in CondBB as costs for
1443 // speculation. Note, while we iterate over a DenseMap here, we are summing
1444 // and so iteration order isn't significant.
1445 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1446 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1448 if (I->first->getNumUses() == I->second) {
1449 SpeculationCost += TTI.getUserCost(I->first);
1450 if (SpeculationCost > TargetTransformInfo::TCC_Basic)
1454 // Check that the PHI nodes can be converted to selects.
1455 bool HaveRewritablePHIs = false;
1456 for (BasicBlock::iterator I = EndBB->begin();
1457 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1458 Value *OrigV = PN->getIncomingValueForBlock(BB);
1459 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1461 // Skip PHIs which are trivial.
1465 HaveRewritablePHIs = true;
1466 ConstantExpr *CE = dyn_cast<ConstantExpr>(ThenV);
1468 continue; // Known safe and cheap.
1470 if (!isSafeToSpeculativelyExecute(CE))
1473 // Don't speculate into a select with a constant select expression operand.
1474 // FIXME: This should really be a cost metric, but our cost model doesn't
1475 // accurately model the expense of select.
1476 if (Operator::getOpcode(CE) == Instruction::Select)
1479 // Account for the cost of an unfolded ConstantExpr which could end up
1480 // getting expanded into Instructions.
1481 // FIXME: This doesn't account for how many operations are combined in the
1482 // constant expression. The cost functions in TTI don't yet correctly model
1483 // constant expression costs.
1484 SpeculationCost += TargetTransformInfo::TCC_Basic;
1485 if (SpeculationCost > TargetTransformInfo::TCC_Basic)
1489 // If there are no PHIs to process, bail early. This helps ensure idempotence
1491 if (!HaveRewritablePHIs)
1494 // If we get here, we can hoist the instruction and if-convert.
1495 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1497 // Hoist the instructions.
1498 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1499 llvm::prior(ThenBB->end()));
1501 // Insert selects and rewrite the PHI operands.
1502 IRBuilder<true, NoFolder> Builder(BI);
1503 for (BasicBlock::iterator I = EndBB->begin();
1504 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1505 unsigned OrigI = PN->getBasicBlockIndex(BB);
1506 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1507 Value *OrigV = PN->getIncomingValue(OrigI);
1508 Value *ThenV = PN->getIncomingValue(ThenI);
1510 // Skip PHIs which are trivial.
1514 // Create a select whose true value is the speculatively executed value and
1515 // false value is the preexisting value. Swap them if the branch
1516 // destinations were inverted.
1517 Value *TrueV = ThenV, *FalseV = OrigV;
1519 std::swap(TrueV, FalseV);
1520 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1521 TrueV->getName() + "." + FalseV->getName());
1522 PN->setIncomingValue(OrigI, V);
1523 PN->setIncomingValue(ThenI, V);
1530 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1531 /// across this block.
1532 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1533 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1536 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1537 if (isa<DbgInfoIntrinsic>(BBI))
1539 if (Size > 10) return false; // Don't clone large BB's.
1542 // We can only support instructions that do not define values that are
1543 // live outside of the current basic block.
1544 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1546 Instruction *U = cast<Instruction>(*UI);
1547 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1550 // Looks ok, continue checking.
1556 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1557 /// that is defined in the same block as the branch and if any PHI entries are
1558 /// constants, thread edges corresponding to that entry to be branches to their
1559 /// ultimate destination.
1560 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) {
1561 BasicBlock *BB = BI->getParent();
1562 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1563 // NOTE: we currently cannot transform this case if the PHI node is used
1564 // outside of the block.
1565 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1568 // Degenerate case of a single entry PHI.
1569 if (PN->getNumIncomingValues() == 1) {
1570 FoldSingleEntryPHINodes(PN->getParent());
1574 // Now we know that this block has multiple preds and two succs.
1575 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1577 // Okay, this is a simple enough basic block. See if any phi values are
1579 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1580 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1581 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1583 // Okay, we now know that all edges from PredBB should be revectored to
1584 // branch to RealDest.
1585 BasicBlock *PredBB = PN->getIncomingBlock(i);
1586 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1588 if (RealDest == BB) continue; // Skip self loops.
1589 // Skip if the predecessor's terminator is an indirect branch.
1590 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1592 // The dest block might have PHI nodes, other predecessors and other
1593 // difficult cases. Instead of being smart about this, just insert a new
1594 // block that jumps to the destination block, effectively splitting
1595 // the edge we are about to create.
1596 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1597 RealDest->getName()+".critedge",
1598 RealDest->getParent(), RealDest);
1599 BranchInst::Create(RealDest, EdgeBB);
1601 // Update PHI nodes.
1602 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1604 // BB may have instructions that are being threaded over. Clone these
1605 // instructions into EdgeBB. We know that there will be no uses of the
1606 // cloned instructions outside of EdgeBB.
1607 BasicBlock::iterator InsertPt = EdgeBB->begin();
1608 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1609 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1610 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1611 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1614 // Clone the instruction.
1615 Instruction *N = BBI->clone();
1616 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1618 // Update operands due to translation.
1619 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1621 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1622 if (PI != TranslateMap.end())
1626 // Check for trivial simplification.
1627 if (Value *V = SimplifyInstruction(N, TD)) {
1628 TranslateMap[BBI] = V;
1629 delete N; // Instruction folded away, don't need actual inst
1631 // Insert the new instruction into its new home.
1632 EdgeBB->getInstList().insert(InsertPt, N);
1633 if (!BBI->use_empty())
1634 TranslateMap[BBI] = N;
1638 // Loop over all of the edges from PredBB to BB, changing them to branch
1639 // to EdgeBB instead.
1640 TerminatorInst *PredBBTI = PredBB->getTerminator();
1641 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1642 if (PredBBTI->getSuccessor(i) == BB) {
1643 BB->removePredecessor(PredBB);
1644 PredBBTI->setSuccessor(i, EdgeBB);
1647 // Recurse, simplifying any other constants.
1648 return FoldCondBranchOnPHI(BI, TD) | true;
1654 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1655 /// PHI node, see if we can eliminate it.
1656 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) {
1657 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1658 // statement", which has a very simple dominance structure. Basically, we
1659 // are trying to find the condition that is being branched on, which
1660 // subsequently causes this merge to happen. We really want control
1661 // dependence information for this check, but simplifycfg can't keep it up
1662 // to date, and this catches most of the cases we care about anyway.
1663 BasicBlock *BB = PN->getParent();
1664 BasicBlock *IfTrue, *IfFalse;
1665 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1667 // Don't bother if the branch will be constant folded trivially.
1668 isa<ConstantInt>(IfCond))
1671 // Okay, we found that we can merge this two-entry phi node into a select.
1672 // Doing so would require us to fold *all* two entry phi nodes in this block.
1673 // At some point this becomes non-profitable (particularly if the target
1674 // doesn't support cmov's). Only do this transformation if there are two or
1675 // fewer PHI nodes in this block.
1676 unsigned NumPhis = 0;
1677 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1681 // Loop over the PHI's seeing if we can promote them all to select
1682 // instructions. While we are at it, keep track of the instructions
1683 // that need to be moved to the dominating block.
1684 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1685 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1686 MaxCostVal1 = PHINodeFoldingThreshold;
1688 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1689 PHINode *PN = cast<PHINode>(II++);
1690 if (Value *V = SimplifyInstruction(PN, TD)) {
1691 PN->replaceAllUsesWith(V);
1692 PN->eraseFromParent();
1696 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1698 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1703 // If we folded the first phi, PN dangles at this point. Refresh it. If
1704 // we ran out of PHIs then we simplified them all.
1705 PN = dyn_cast<PHINode>(BB->begin());
1706 if (PN == 0) return true;
1708 // Don't fold i1 branches on PHIs which contain binary operators. These can
1709 // often be turned into switches and other things.
1710 if (PN->getType()->isIntegerTy(1) &&
1711 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1712 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1713 isa<BinaryOperator>(IfCond)))
1716 // If we all PHI nodes are promotable, check to make sure that all
1717 // instructions in the predecessor blocks can be promoted as well. If
1718 // not, we won't be able to get rid of the control flow, so it's not
1719 // worth promoting to select instructions.
1720 BasicBlock *DomBlock = 0;
1721 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1722 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1723 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1726 DomBlock = *pred_begin(IfBlock1);
1727 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1728 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1729 // This is not an aggressive instruction that we can promote.
1730 // Because of this, we won't be able to get rid of the control
1731 // flow, so the xform is not worth it.
1736 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1739 DomBlock = *pred_begin(IfBlock2);
1740 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1741 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1742 // This is not an aggressive instruction that we can promote.
1743 // Because of this, we won't be able to get rid of the control
1744 // flow, so the xform is not worth it.
1749 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1750 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1752 // If we can still promote the PHI nodes after this gauntlet of tests,
1753 // do all of the PHI's now.
1754 Instruction *InsertPt = DomBlock->getTerminator();
1755 IRBuilder<true, NoFolder> Builder(InsertPt);
1757 // Move all 'aggressive' instructions, which are defined in the
1758 // conditional parts of the if's up to the dominating block.
1760 DomBlock->getInstList().splice(InsertPt,
1761 IfBlock1->getInstList(), IfBlock1->begin(),
1762 IfBlock1->getTerminator());
1764 DomBlock->getInstList().splice(InsertPt,
1765 IfBlock2->getInstList(), IfBlock2->begin(),
1766 IfBlock2->getTerminator());
1768 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1769 // Change the PHI node into a select instruction.
1770 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1771 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1774 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1775 PN->replaceAllUsesWith(NV);
1777 PN->eraseFromParent();
1780 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1781 // has been flattened. Change DomBlock to jump directly to our new block to
1782 // avoid other simplifycfg's kicking in on the diamond.
1783 TerminatorInst *OldTI = DomBlock->getTerminator();
1784 Builder.SetInsertPoint(OldTI);
1785 Builder.CreateBr(BB);
1786 OldTI->eraseFromParent();
1790 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1791 /// to two returning blocks, try to merge them together into one return,
1792 /// introducing a select if the return values disagree.
1793 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1794 IRBuilder<> &Builder) {
1795 assert(BI->isConditional() && "Must be a conditional branch");
1796 BasicBlock *TrueSucc = BI->getSuccessor(0);
1797 BasicBlock *FalseSucc = BI->getSuccessor(1);
1798 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1799 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1801 // Check to ensure both blocks are empty (just a return) or optionally empty
1802 // with PHI nodes. If there are other instructions, merging would cause extra
1803 // computation on one path or the other.
1804 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1806 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1809 Builder.SetInsertPoint(BI);
1810 // Okay, we found a branch that is going to two return nodes. If
1811 // there is no return value for this function, just change the
1812 // branch into a return.
1813 if (FalseRet->getNumOperands() == 0) {
1814 TrueSucc->removePredecessor(BI->getParent());
1815 FalseSucc->removePredecessor(BI->getParent());
1816 Builder.CreateRetVoid();
1817 EraseTerminatorInstAndDCECond(BI);
1821 // Otherwise, figure out what the true and false return values are
1822 // so we can insert a new select instruction.
1823 Value *TrueValue = TrueRet->getReturnValue();
1824 Value *FalseValue = FalseRet->getReturnValue();
1826 // Unwrap any PHI nodes in the return blocks.
1827 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1828 if (TVPN->getParent() == TrueSucc)
1829 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1830 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1831 if (FVPN->getParent() == FalseSucc)
1832 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1834 // In order for this transformation to be safe, we must be able to
1835 // unconditionally execute both operands to the return. This is
1836 // normally the case, but we could have a potentially-trapping
1837 // constant expression that prevents this transformation from being
1839 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1842 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1846 // Okay, we collected all the mapped values and checked them for sanity, and
1847 // defined to really do this transformation. First, update the CFG.
1848 TrueSucc->removePredecessor(BI->getParent());
1849 FalseSucc->removePredecessor(BI->getParent());
1851 // Insert select instructions where needed.
1852 Value *BrCond = BI->getCondition();
1854 // Insert a select if the results differ.
1855 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1856 } else if (isa<UndefValue>(TrueValue)) {
1857 TrueValue = FalseValue;
1859 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1860 FalseValue, "retval");
1864 Value *RI = !TrueValue ?
1865 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1869 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1870 << "\n " << *BI << "NewRet = " << *RI
1871 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1873 EraseTerminatorInstAndDCECond(BI);
1878 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1879 /// probabilities of the branch taking each edge. Fills in the two APInt
1880 /// parameters and return true, or returns false if no or invalid metadata was
1882 static bool ExtractBranchMetadata(BranchInst *BI,
1883 uint64_t &ProbTrue, uint64_t &ProbFalse) {
1884 assert(BI->isConditional() &&
1885 "Looking for probabilities on unconditional branch?");
1886 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
1887 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
1888 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
1889 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
1890 if (!CITrue || !CIFalse) return false;
1891 ProbTrue = CITrue->getValue().getZExtValue();
1892 ProbFalse = CIFalse->getValue().getZExtValue();
1896 /// checkCSEInPredecessor - Return true if the given instruction is available
1897 /// in its predecessor block. If yes, the instruction will be removed.
1899 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
1900 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
1902 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
1903 Instruction *PBI = &*I;
1904 // Check whether Inst and PBI generate the same value.
1905 if (Inst->isIdenticalTo(PBI)) {
1906 Inst->replaceAllUsesWith(PBI);
1907 Inst->eraseFromParent();
1914 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
1915 /// predecessor branches to us and one of our successors, fold the block into
1916 /// the predecessor and use logical operations to pick the right destination.
1917 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1918 BasicBlock *BB = BI->getParent();
1920 Instruction *Cond = 0;
1921 if (BI->isConditional())
1922 Cond = dyn_cast<Instruction>(BI->getCondition());
1924 // For unconditional branch, check for a simple CFG pattern, where
1925 // BB has a single predecessor and BB's successor is also its predecessor's
1926 // successor. If such pattern exisits, check for CSE between BB and its
1928 if (BasicBlock *PB = BB->getSinglePredecessor())
1929 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
1930 if (PBI->isConditional() &&
1931 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
1932 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
1933 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
1935 Instruction *Curr = I++;
1936 if (isa<CmpInst>(Curr)) {
1940 // Quit if we can't remove this instruction.
1941 if (!checkCSEInPredecessor(Curr, PB))
1950 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1951 Cond->getParent() != BB || !Cond->hasOneUse())
1954 // Only allow this if the condition is a simple instruction that can be
1955 // executed unconditionally. It must be in the same block as the branch, and
1956 // must be at the front of the block.
1957 BasicBlock::iterator FrontIt = BB->front();
1959 // Ignore dbg intrinsics.
1960 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1962 // Allow a single instruction to be hoisted in addition to the compare
1963 // that feeds the branch. We later ensure that any values that _it_ uses
1964 // were also live in the predecessor, so that we don't unnecessarily create
1965 // register pressure or inhibit out-of-order execution.
1966 Instruction *BonusInst = 0;
1967 if (&*FrontIt != Cond &&
1968 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
1969 isSafeToSpeculativelyExecute(FrontIt)) {
1970 BonusInst = &*FrontIt;
1973 // Ignore dbg intrinsics.
1974 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
1977 // Only a single bonus inst is allowed.
1978 if (&*FrontIt != Cond)
1981 // Make sure the instruction after the condition is the cond branch.
1982 BasicBlock::iterator CondIt = Cond; ++CondIt;
1984 // Ingore dbg intrinsics.
1985 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
1990 // Cond is known to be a compare or binary operator. Check to make sure that
1991 // neither operand is a potentially-trapping constant expression.
1992 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
1995 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
1999 // Finally, don't infinitely unroll conditional loops.
2000 BasicBlock *TrueDest = BI->getSuccessor(0);
2001 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
2002 if (TrueDest == BB || FalseDest == BB)
2005 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2006 BasicBlock *PredBlock = *PI;
2007 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2009 // Check that we have two conditional branches. If there is a PHI node in
2010 // the common successor, verify that the same value flows in from both
2012 SmallVector<PHINode*, 4> PHIs;
2013 if (PBI == 0 || PBI->isUnconditional() ||
2014 (BI->isConditional() &&
2015 !SafeToMergeTerminators(BI, PBI)) ||
2016 (!BI->isConditional() &&
2017 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2020 // Determine if the two branches share a common destination.
2021 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2022 bool InvertPredCond = false;
2024 if (BI->isConditional()) {
2025 if (PBI->getSuccessor(0) == TrueDest)
2026 Opc = Instruction::Or;
2027 else if (PBI->getSuccessor(1) == FalseDest)
2028 Opc = Instruction::And;
2029 else if (PBI->getSuccessor(0) == FalseDest)
2030 Opc = Instruction::And, InvertPredCond = true;
2031 else if (PBI->getSuccessor(1) == TrueDest)
2032 Opc = Instruction::Or, InvertPredCond = true;
2036 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2040 // Ensure that any values used in the bonus instruction are also used
2041 // by the terminator of the predecessor. This means that those values
2042 // must already have been resolved, so we won't be inhibiting the
2043 // out-of-order core by speculating them earlier.
2045 // Collect the values used by the bonus inst
2046 SmallPtrSet<Value*, 4> UsedValues;
2047 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2048 OE = BonusInst->op_end(); OI != OE; ++OI) {
2050 if (!isa<Constant>(V))
2051 UsedValues.insert(V);
2054 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2055 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2057 // Walk up to four levels back up the use-def chain of the predecessor's
2058 // terminator to see if all those values were used. The choice of four
2059 // levels is arbitrary, to provide a compile-time-cost bound.
2060 while (!Worklist.empty()) {
2061 std::pair<Value*, unsigned> Pair = Worklist.back();
2062 Worklist.pop_back();
2064 if (Pair.second >= 4) continue;
2065 UsedValues.erase(Pair.first);
2066 if (UsedValues.empty()) break;
2068 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2069 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2071 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2075 if (!UsedValues.empty()) return false;
2078 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2079 IRBuilder<> Builder(PBI);
2081 // If we need to invert the condition in the pred block to match, do so now.
2082 if (InvertPredCond) {
2083 Value *NewCond = PBI->getCondition();
2085 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2086 CmpInst *CI = cast<CmpInst>(NewCond);
2087 CI->setPredicate(CI->getInversePredicate());
2089 NewCond = Builder.CreateNot(NewCond,
2090 PBI->getCondition()->getName()+".not");
2093 PBI->setCondition(NewCond);
2094 PBI->swapSuccessors();
2097 // If we have a bonus inst, clone it into the predecessor block.
2098 Instruction *NewBonus = 0;
2100 NewBonus = BonusInst->clone();
2101 PredBlock->getInstList().insert(PBI, NewBonus);
2102 NewBonus->takeName(BonusInst);
2103 BonusInst->setName(BonusInst->getName()+".old");
2106 // Clone Cond into the predecessor basic block, and or/and the
2107 // two conditions together.
2108 Instruction *New = Cond->clone();
2109 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2110 PredBlock->getInstList().insert(PBI, New);
2111 New->takeName(Cond);
2112 Cond->setName(New->getName()+".old");
2114 if (BI->isConditional()) {
2115 Instruction *NewCond =
2116 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2118 PBI->setCondition(NewCond);
2120 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2121 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2123 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2125 SmallVector<uint64_t, 8> NewWeights;
2127 if (PBI->getSuccessor(0) == BB) {
2128 if (PredHasWeights && SuccHasWeights) {
2129 // PBI: br i1 %x, BB, FalseDest
2130 // BI: br i1 %y, TrueDest, FalseDest
2131 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2132 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2133 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2134 // TrueWeight for PBI * FalseWeight for BI.
2135 // We assume that total weights of a BranchInst can fit into 32 bits.
2136 // Therefore, we will not have overflow using 64-bit arithmetic.
2137 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2138 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2140 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2141 PBI->setSuccessor(0, TrueDest);
2143 if (PBI->getSuccessor(1) == BB) {
2144 if (PredHasWeights && SuccHasWeights) {
2145 // PBI: br i1 %x, TrueDest, BB
2146 // BI: br i1 %y, TrueDest, FalseDest
2147 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2148 // FalseWeight for PBI * TrueWeight for BI.
2149 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2150 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2151 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2152 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2154 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2155 PBI->setSuccessor(1, FalseDest);
2157 if (NewWeights.size() == 2) {
2158 // Halve the weights if any of them cannot fit in an uint32_t
2159 FitWeights(NewWeights);
2161 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2162 PBI->setMetadata(LLVMContext::MD_prof,
2163 MDBuilder(BI->getContext()).
2164 createBranchWeights(MDWeights));
2166 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2168 // Update PHI nodes in the common successors.
2169 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2170 ConstantInt *PBI_C = cast<ConstantInt>(
2171 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2172 assert(PBI_C->getType()->isIntegerTy(1));
2173 Instruction *MergedCond = 0;
2174 if (PBI->getSuccessor(0) == TrueDest) {
2175 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2176 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2177 // is false: !PBI_Cond and BI_Value
2178 Instruction *NotCond =
2179 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2182 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2187 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2188 PBI->getCondition(), MergedCond,
2191 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2192 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2193 // is false: PBI_Cond and BI_Value
2195 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2196 PBI->getCondition(), New,
2198 if (PBI_C->isOne()) {
2199 Instruction *NotCond =
2200 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2203 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2204 NotCond, MergedCond,
2209 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2212 // Change PBI from Conditional to Unconditional.
2213 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2214 EraseTerminatorInstAndDCECond(PBI);
2218 // TODO: If BB is reachable from all paths through PredBlock, then we
2219 // could replace PBI's branch probabilities with BI's.
2221 // Copy any debug value intrinsics into the end of PredBlock.
2222 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2223 if (isa<DbgInfoIntrinsic>(*I))
2224 I->clone()->insertBefore(PBI);
2231 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2232 /// predecessor of another block, this function tries to simplify it. We know
2233 /// that PBI and BI are both conditional branches, and BI is in one of the
2234 /// successor blocks of PBI - PBI branches to BI.
2235 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2236 assert(PBI->isConditional() && BI->isConditional());
2237 BasicBlock *BB = BI->getParent();
2239 // If this block ends with a branch instruction, and if there is a
2240 // predecessor that ends on a branch of the same condition, make
2241 // this conditional branch redundant.
2242 if (PBI->getCondition() == BI->getCondition() &&
2243 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2244 // Okay, the outcome of this conditional branch is statically
2245 // knowable. If this block had a single pred, handle specially.
2246 if (BB->getSinglePredecessor()) {
2247 // Turn this into a branch on constant.
2248 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2249 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2251 return true; // Nuke the branch on constant.
2254 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2255 // in the constant and simplify the block result. Subsequent passes of
2256 // simplifycfg will thread the block.
2257 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2258 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2259 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2260 std::distance(PB, PE),
2261 BI->getCondition()->getName() + ".pr",
2263 // Okay, we're going to insert the PHI node. Since PBI is not the only
2264 // predecessor, compute the PHI'd conditional value for all of the preds.
2265 // Any predecessor where the condition is not computable we keep symbolic.
2266 for (pred_iterator PI = PB; PI != PE; ++PI) {
2267 BasicBlock *P = *PI;
2268 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2269 PBI != BI && PBI->isConditional() &&
2270 PBI->getCondition() == BI->getCondition() &&
2271 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2272 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2273 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2276 NewPN->addIncoming(BI->getCondition(), P);
2280 BI->setCondition(NewPN);
2285 // If this is a conditional branch in an empty block, and if any
2286 // predecessors is a conditional branch to one of our destinations,
2287 // fold the conditions into logical ops and one cond br.
2288 BasicBlock::iterator BBI = BB->begin();
2289 // Ignore dbg intrinsics.
2290 while (isa<DbgInfoIntrinsic>(BBI))
2296 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2301 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2303 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2304 PBIOp = 0, BIOp = 1;
2305 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2306 PBIOp = 1, BIOp = 0;
2307 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2312 // Check to make sure that the other destination of this branch
2313 // isn't BB itself. If so, this is an infinite loop that will
2314 // keep getting unwound.
2315 if (PBI->getSuccessor(PBIOp) == BB)
2318 // Do not perform this transformation if it would require
2319 // insertion of a large number of select instructions. For targets
2320 // without predication/cmovs, this is a big pessimization.
2321 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2323 unsigned NumPhis = 0;
2324 for (BasicBlock::iterator II = CommonDest->begin();
2325 isa<PHINode>(II); ++II, ++NumPhis)
2326 if (NumPhis > 2) // Disable this xform.
2329 // Finally, if everything is ok, fold the branches to logical ops.
2330 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2332 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2333 << "AND: " << *BI->getParent());
2336 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2337 // branch in it, where one edge (OtherDest) goes back to itself but the other
2338 // exits. We don't *know* that the program avoids the infinite loop
2339 // (even though that seems likely). If we do this xform naively, we'll end up
2340 // recursively unpeeling the loop. Since we know that (after the xform is
2341 // done) that the block *is* infinite if reached, we just make it an obviously
2342 // infinite loop with no cond branch.
2343 if (OtherDest == BB) {
2344 // Insert it at the end of the function, because it's either code,
2345 // or it won't matter if it's hot. :)
2346 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2347 "infloop", BB->getParent());
2348 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2349 OtherDest = InfLoopBlock;
2352 DEBUG(dbgs() << *PBI->getParent()->getParent());
2354 // BI may have other predecessors. Because of this, we leave
2355 // it alone, but modify PBI.
2357 // Make sure we get to CommonDest on True&True directions.
2358 Value *PBICond = PBI->getCondition();
2359 IRBuilder<true, NoFolder> Builder(PBI);
2361 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2363 Value *BICond = BI->getCondition();
2365 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2367 // Merge the conditions.
2368 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2370 // Modify PBI to branch on the new condition to the new dests.
2371 PBI->setCondition(Cond);
2372 PBI->setSuccessor(0, CommonDest);
2373 PBI->setSuccessor(1, OtherDest);
2375 // Update branch weight for PBI.
2376 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2377 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2379 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2381 if (PredHasWeights && SuccHasWeights) {
2382 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2383 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2384 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2385 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2386 // The weight to CommonDest should be PredCommon * SuccTotal +
2387 // PredOther * SuccCommon.
2388 // The weight to OtherDest should be PredOther * SuccOther.
2389 SmallVector<uint64_t, 2> NewWeights;
2390 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2391 PredOther * SuccCommon);
2392 NewWeights.push_back(PredOther * SuccOther);
2393 // Halve the weights if any of them cannot fit in an uint32_t
2394 FitWeights(NewWeights);
2396 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2397 PBI->setMetadata(LLVMContext::MD_prof,
2398 MDBuilder(BI->getContext()).
2399 createBranchWeights(MDWeights));
2402 // OtherDest may have phi nodes. If so, add an entry from PBI's
2403 // block that are identical to the entries for BI's block.
2404 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2406 // We know that the CommonDest already had an edge from PBI to
2407 // it. If it has PHIs though, the PHIs may have different
2408 // entries for BB and PBI's BB. If so, insert a select to make
2411 for (BasicBlock::iterator II = CommonDest->begin();
2412 (PN = dyn_cast<PHINode>(II)); ++II) {
2413 Value *BIV = PN->getIncomingValueForBlock(BB);
2414 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2415 Value *PBIV = PN->getIncomingValue(PBBIdx);
2417 // Insert a select in PBI to pick the right value.
2418 Value *NV = cast<SelectInst>
2419 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2420 PN->setIncomingValue(PBBIdx, NV);
2424 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2425 DEBUG(dbgs() << *PBI->getParent()->getParent());
2427 // This basic block is probably dead. We know it has at least
2428 // one fewer predecessor.
2432 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2433 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2434 // Takes care of updating the successors and removing the old terminator.
2435 // Also makes sure not to introduce new successors by assuming that edges to
2436 // non-successor TrueBBs and FalseBBs aren't reachable.
2437 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2438 BasicBlock *TrueBB, BasicBlock *FalseBB,
2439 uint32_t TrueWeight,
2440 uint32_t FalseWeight){
2441 // Remove any superfluous successor edges from the CFG.
2442 // First, figure out which successors to preserve.
2443 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2445 BasicBlock *KeepEdge1 = TrueBB;
2446 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2448 // Then remove the rest.
2449 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2450 BasicBlock *Succ = OldTerm->getSuccessor(I);
2451 // Make sure only to keep exactly one copy of each edge.
2452 if (Succ == KeepEdge1)
2454 else if (Succ == KeepEdge2)
2457 Succ->removePredecessor(OldTerm->getParent());
2460 IRBuilder<> Builder(OldTerm);
2461 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2463 // Insert an appropriate new terminator.
2464 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2465 if (TrueBB == FalseBB)
2466 // We were only looking for one successor, and it was present.
2467 // Create an unconditional branch to it.
2468 Builder.CreateBr(TrueBB);
2470 // We found both of the successors we were looking for.
2471 // Create a conditional branch sharing the condition of the select.
2472 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2473 if (TrueWeight != FalseWeight)
2474 NewBI->setMetadata(LLVMContext::MD_prof,
2475 MDBuilder(OldTerm->getContext()).
2476 createBranchWeights(TrueWeight, FalseWeight));
2478 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2479 // Neither of the selected blocks were successors, so this
2480 // terminator must be unreachable.
2481 new UnreachableInst(OldTerm->getContext(), OldTerm);
2483 // One of the selected values was a successor, but the other wasn't.
2484 // Insert an unconditional branch to the one that was found;
2485 // the edge to the one that wasn't must be unreachable.
2487 // Only TrueBB was found.
2488 Builder.CreateBr(TrueBB);
2490 // Only FalseBB was found.
2491 Builder.CreateBr(FalseBB);
2494 EraseTerminatorInstAndDCECond(OldTerm);
2498 // SimplifySwitchOnSelect - Replaces
2499 // (switch (select cond, X, Y)) on constant X, Y
2500 // with a branch - conditional if X and Y lead to distinct BBs,
2501 // unconditional otherwise.
2502 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2503 // Check for constant integer values in the select.
2504 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2505 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2506 if (!TrueVal || !FalseVal)
2509 // Find the relevant condition and destinations.
2510 Value *Condition = Select->getCondition();
2511 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2512 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2514 // Get weight for TrueBB and FalseBB.
2515 uint32_t TrueWeight = 0, FalseWeight = 0;
2516 SmallVector<uint64_t, 8> Weights;
2517 bool HasWeights = HasBranchWeights(SI);
2519 GetBranchWeights(SI, Weights);
2520 if (Weights.size() == 1 + SI->getNumCases()) {
2521 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2522 getSuccessorIndex()];
2523 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2524 getSuccessorIndex()];
2528 // Perform the actual simplification.
2529 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2530 TrueWeight, FalseWeight);
2533 // SimplifyIndirectBrOnSelect - Replaces
2534 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2535 // blockaddress(@fn, BlockB)))
2537 // (br cond, BlockA, BlockB).
2538 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2539 // Check that both operands of the select are block addresses.
2540 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2541 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2545 // Extract the actual blocks.
2546 BasicBlock *TrueBB = TBA->getBasicBlock();
2547 BasicBlock *FalseBB = FBA->getBasicBlock();
2549 // Perform the actual simplification.
2550 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2554 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2555 /// instruction (a seteq/setne with a constant) as the only instruction in a
2556 /// block that ends with an uncond branch. We are looking for a very specific
2557 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2558 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2559 /// default value goes to an uncond block with a seteq in it, we get something
2562 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2564 /// %tmp = icmp eq i8 %A, 92
2567 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2569 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2570 /// the PHI, merging the third icmp into the switch.
2571 static bool TryToSimplifyUncondBranchWithICmpInIt(
2572 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2573 const DataLayout *TD) {
2574 BasicBlock *BB = ICI->getParent();
2576 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2578 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2580 Value *V = ICI->getOperand(0);
2581 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2583 // The pattern we're looking for is where our only predecessor is a switch on
2584 // 'V' and this block is the default case for the switch. In this case we can
2585 // fold the compared value into the switch to simplify things.
2586 BasicBlock *Pred = BB->getSinglePredecessor();
2587 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2589 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2590 if (SI->getCondition() != V)
2593 // If BB is reachable on a non-default case, then we simply know the value of
2594 // V in this block. Substitute it and constant fold the icmp instruction
2596 if (SI->getDefaultDest() != BB) {
2597 ConstantInt *VVal = SI->findCaseDest(BB);
2598 assert(VVal && "Should have a unique destination value");
2599 ICI->setOperand(0, VVal);
2601 if (Value *V = SimplifyInstruction(ICI, TD)) {
2602 ICI->replaceAllUsesWith(V);
2603 ICI->eraseFromParent();
2605 // BB is now empty, so it is likely to simplify away.
2606 return SimplifyCFG(BB, TTI, TD) | true;
2609 // Ok, the block is reachable from the default dest. If the constant we're
2610 // comparing exists in one of the other edges, then we can constant fold ICI
2612 if (SI->findCaseValue(Cst) != SI->case_default()) {
2614 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2615 V = ConstantInt::getFalse(BB->getContext());
2617 V = ConstantInt::getTrue(BB->getContext());
2619 ICI->replaceAllUsesWith(V);
2620 ICI->eraseFromParent();
2621 // BB is now empty, so it is likely to simplify away.
2622 return SimplifyCFG(BB, TTI, TD) | true;
2625 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2627 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2628 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2629 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2630 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2633 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2635 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2636 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2638 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2639 std::swap(DefaultCst, NewCst);
2641 // Replace ICI (which is used by the PHI for the default value) with true or
2642 // false depending on if it is EQ or NE.
2643 ICI->replaceAllUsesWith(DefaultCst);
2644 ICI->eraseFromParent();
2646 // Okay, the switch goes to this block on a default value. Add an edge from
2647 // the switch to the merge point on the compared value.
2648 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2649 BB->getParent(), BB);
2650 SmallVector<uint64_t, 8> Weights;
2651 bool HasWeights = HasBranchWeights(SI);
2653 GetBranchWeights(SI, Weights);
2654 if (Weights.size() == 1 + SI->getNumCases()) {
2655 // Split weight for default case to case for "Cst".
2656 Weights[0] = (Weights[0]+1) >> 1;
2657 Weights.push_back(Weights[0]);
2659 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2660 SI->setMetadata(LLVMContext::MD_prof,
2661 MDBuilder(SI->getContext()).
2662 createBranchWeights(MDWeights));
2665 SI->addCase(Cst, NewBB);
2667 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2668 Builder.SetInsertPoint(NewBB);
2669 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2670 Builder.CreateBr(SuccBlock);
2671 PHIUse->addIncoming(NewCst, NewBB);
2675 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2676 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2677 /// fold it into a switch instruction if so.
2678 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD,
2679 IRBuilder<> &Builder) {
2680 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2681 if (Cond == 0) return false;
2684 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2685 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2686 // 'setne's and'ed together, collect them.
2688 std::vector<ConstantInt*> Values;
2689 bool TrueWhenEqual = true;
2690 Value *ExtraCase = 0;
2691 unsigned UsedICmps = 0;
2693 if (Cond->getOpcode() == Instruction::Or) {
2694 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2696 } else if (Cond->getOpcode() == Instruction::And) {
2697 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2699 TrueWhenEqual = false;
2702 // If we didn't have a multiply compared value, fail.
2703 if (CompVal == 0) return false;
2705 // Avoid turning single icmps into a switch.
2709 // There might be duplicate constants in the list, which the switch
2710 // instruction can't handle, remove them now.
2711 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2712 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2714 // If Extra was used, we require at least two switch values to do the
2715 // transformation. A switch with one value is just an cond branch.
2716 if (ExtraCase && Values.size() < 2) return false;
2718 // TODO: Preserve branch weight metadata, similarly to how
2719 // FoldValueComparisonIntoPredecessors preserves it.
2721 // Figure out which block is which destination.
2722 BasicBlock *DefaultBB = BI->getSuccessor(1);
2723 BasicBlock *EdgeBB = BI->getSuccessor(0);
2724 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2726 BasicBlock *BB = BI->getParent();
2728 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2729 << " cases into SWITCH. BB is:\n" << *BB);
2731 // If there are any extra values that couldn't be folded into the switch
2732 // then we evaluate them with an explicit branch first. Split the block
2733 // right before the condbr to handle it.
2735 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2736 // Remove the uncond branch added to the old block.
2737 TerminatorInst *OldTI = BB->getTerminator();
2738 Builder.SetInsertPoint(OldTI);
2741 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2743 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2745 OldTI->eraseFromParent();
2747 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2748 // for the edge we just added.
2749 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2751 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2752 << "\nEXTRABB = " << *BB);
2756 Builder.SetInsertPoint(BI);
2757 // Convert pointer to int before we switch.
2758 if (CompVal->getType()->isPointerTy()) {
2759 assert(TD && "Cannot switch on pointer without DataLayout");
2760 CompVal = Builder.CreatePtrToInt(CompVal,
2761 TD->getIntPtrType(CompVal->getContext()),
2765 // Create the new switch instruction now.
2766 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2768 // Add all of the 'cases' to the switch instruction.
2769 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2770 New->addCase(Values[i], EdgeBB);
2772 // We added edges from PI to the EdgeBB. As such, if there were any
2773 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2774 // the number of edges added.
2775 for (BasicBlock::iterator BBI = EdgeBB->begin();
2776 isa<PHINode>(BBI); ++BBI) {
2777 PHINode *PN = cast<PHINode>(BBI);
2778 Value *InVal = PN->getIncomingValueForBlock(BB);
2779 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2780 PN->addIncoming(InVal, BB);
2783 // Erase the old branch instruction.
2784 EraseTerminatorInstAndDCECond(BI);
2786 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2790 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2791 // If this is a trivial landing pad that just continues unwinding the caught
2792 // exception then zap the landing pad, turning its invokes into calls.
2793 BasicBlock *BB = RI->getParent();
2794 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2795 if (RI->getValue() != LPInst)
2796 // Not a landing pad, or the resume is not unwinding the exception that
2797 // caused control to branch here.
2800 // Check that there are no other instructions except for debug intrinsics.
2801 BasicBlock::iterator I = LPInst, E = RI;
2803 if (!isa<DbgInfoIntrinsic>(I))
2806 // Turn all invokes that unwind here into calls and delete the basic block.
2807 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2808 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2809 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2810 // Insert a call instruction before the invoke.
2811 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2813 Call->setCallingConv(II->getCallingConv());
2814 Call->setAttributes(II->getAttributes());
2815 Call->setDebugLoc(II->getDebugLoc());
2817 // Anything that used the value produced by the invoke instruction now uses
2818 // the value produced by the call instruction. Note that we do this even
2819 // for void functions and calls with no uses so that the callgraph edge is
2821 II->replaceAllUsesWith(Call);
2822 BB->removePredecessor(II->getParent());
2824 // Insert a branch to the normal destination right before the invoke.
2825 BranchInst::Create(II->getNormalDest(), II);
2827 // Finally, delete the invoke instruction!
2828 II->eraseFromParent();
2831 // The landingpad is now unreachable. Zap it.
2832 BB->eraseFromParent();
2836 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2837 BasicBlock *BB = RI->getParent();
2838 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2840 // Find predecessors that end with branches.
2841 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2842 SmallVector<BranchInst*, 8> CondBranchPreds;
2843 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2844 BasicBlock *P = *PI;
2845 TerminatorInst *PTI = P->getTerminator();
2846 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2847 if (BI->isUnconditional())
2848 UncondBranchPreds.push_back(P);
2850 CondBranchPreds.push_back(BI);
2854 // If we found some, do the transformation!
2855 if (!UncondBranchPreds.empty() && DupRet) {
2856 while (!UncondBranchPreds.empty()) {
2857 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2858 DEBUG(dbgs() << "FOLDING: " << *BB
2859 << "INTO UNCOND BRANCH PRED: " << *Pred);
2860 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2863 // If we eliminated all predecessors of the block, delete the block now.
2864 if (pred_begin(BB) == pred_end(BB))
2865 // We know there are no successors, so just nuke the block.
2866 BB->eraseFromParent();
2871 // Check out all of the conditional branches going to this return
2872 // instruction. If any of them just select between returns, change the
2873 // branch itself into a select/return pair.
2874 while (!CondBranchPreds.empty()) {
2875 BranchInst *BI = CondBranchPreds.pop_back_val();
2877 // Check to see if the non-BB successor is also a return block.
2878 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
2879 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
2880 SimplifyCondBranchToTwoReturns(BI, Builder))
2886 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
2887 BasicBlock *BB = UI->getParent();
2889 bool Changed = false;
2891 // If there are any instructions immediately before the unreachable that can
2892 // be removed, do so.
2893 while (UI != BB->begin()) {
2894 BasicBlock::iterator BBI = UI;
2896 // Do not delete instructions that can have side effects which might cause
2897 // the unreachable to not be reachable; specifically, calls and volatile
2898 // operations may have this effect.
2899 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2901 if (BBI->mayHaveSideEffects()) {
2902 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
2903 if (SI->isVolatile())
2905 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
2906 if (LI->isVolatile())
2908 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
2909 if (RMWI->isVolatile())
2911 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
2912 if (CXI->isVolatile())
2914 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
2915 !isa<LandingPadInst>(BBI)) {
2918 // Note that deleting LandingPad's here is in fact okay, although it
2919 // involves a bit of subtle reasoning. If this inst is a LandingPad,
2920 // all the predecessors of this block will be the unwind edges of Invokes,
2921 // and we can therefore guarantee this block will be erased.
2924 // Delete this instruction (any uses are guaranteed to be dead)
2925 if (!BBI->use_empty())
2926 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
2927 BBI->eraseFromParent();
2931 // If the unreachable instruction is the first in the block, take a gander
2932 // at all of the predecessors of this instruction, and simplify them.
2933 if (&BB->front() != UI) return Changed;
2935 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2936 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
2937 TerminatorInst *TI = Preds[i]->getTerminator();
2938 IRBuilder<> Builder(TI);
2939 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2940 if (BI->isUnconditional()) {
2941 if (BI->getSuccessor(0) == BB) {
2942 new UnreachableInst(TI->getContext(), TI);
2943 TI->eraseFromParent();
2947 if (BI->getSuccessor(0) == BB) {
2948 Builder.CreateBr(BI->getSuccessor(1));
2949 EraseTerminatorInstAndDCECond(BI);
2950 } else if (BI->getSuccessor(1) == BB) {
2951 Builder.CreateBr(BI->getSuccessor(0));
2952 EraseTerminatorInstAndDCECond(BI);
2956 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2957 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2959 if (i.getCaseSuccessor() == BB) {
2960 BB->removePredecessor(SI->getParent());
2965 // If the default value is unreachable, figure out the most popular
2966 // destination and make it the default.
2967 if (SI->getDefaultDest() == BB) {
2968 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
2969 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2971 std::pair<unsigned, unsigned> &entry =
2972 Popularity[i.getCaseSuccessor()];
2973 if (entry.first == 0) {
2975 entry.second = i.getCaseIndex();
2981 // Find the most popular block.
2982 unsigned MaxPop = 0;
2983 unsigned MaxIndex = 0;
2984 BasicBlock *MaxBlock = 0;
2985 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
2986 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
2987 if (I->second.first > MaxPop ||
2988 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
2989 MaxPop = I->second.first;
2990 MaxIndex = I->second.second;
2991 MaxBlock = I->first;
2995 // Make this the new default, allowing us to delete any explicit
2997 SI->setDefaultDest(MaxBlock);
3000 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3002 if (isa<PHINode>(MaxBlock->begin()))
3003 for (unsigned i = 0; i != MaxPop-1; ++i)
3004 MaxBlock->removePredecessor(SI->getParent());
3006 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3008 if (i.getCaseSuccessor() == MaxBlock) {
3014 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3015 if (II->getUnwindDest() == BB) {
3016 // Convert the invoke to a call instruction. This would be a good
3017 // place to note that the call does not throw though.
3018 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3019 II->removeFromParent(); // Take out of symbol table
3021 // Insert the call now...
3022 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3023 Builder.SetInsertPoint(BI);
3024 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3025 Args, II->getName());
3026 CI->setCallingConv(II->getCallingConv());
3027 CI->setAttributes(II->getAttributes());
3028 // If the invoke produced a value, the call does now instead.
3029 II->replaceAllUsesWith(CI);
3036 // If this block is now dead, remove it.
3037 if (pred_begin(BB) == pred_end(BB) &&
3038 BB != &BB->getParent()->getEntryBlock()) {
3039 // We know there are no successors, so just nuke the block.
3040 BB->eraseFromParent();
3047 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3048 /// integer range comparison into a sub, an icmp and a branch.
3049 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3050 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3052 // Make sure all cases point to the same destination and gather the values.
3053 SmallVector<ConstantInt *, 16> Cases;
3054 SwitchInst::CaseIt I = SI->case_begin();
3055 Cases.push_back(I.getCaseValue());
3056 SwitchInst::CaseIt PrevI = I++;
3057 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3058 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3060 Cases.push_back(I.getCaseValue());
3062 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3064 // Sort the case values, then check if they form a range we can transform.
3065 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3066 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3067 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3071 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3072 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3074 Value *Sub = SI->getCondition();
3075 if (!Offset->isNullValue())
3076 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3077 Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3078 BranchInst *NewBI = Builder.CreateCondBr(
3079 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3081 // Update weight for the newly-created conditional branch.
3082 SmallVector<uint64_t, 8> Weights;
3083 bool HasWeights = HasBranchWeights(SI);
3085 GetBranchWeights(SI, Weights);
3086 if (Weights.size() == 1 + SI->getNumCases()) {
3087 // Combine all weights for the cases to be the true weight of NewBI.
3088 // We assume that the sum of all weights for a Terminator can fit into 32
3090 uint32_t NewTrueWeight = 0;
3091 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3092 NewTrueWeight += (uint32_t)Weights[I];
3093 NewBI->setMetadata(LLVMContext::MD_prof,
3094 MDBuilder(SI->getContext()).
3095 createBranchWeights(NewTrueWeight,
3096 (uint32_t)Weights[0]));
3100 // Prune obsolete incoming values off the successor's PHI nodes.
3101 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3102 isa<PHINode>(BBI); ++BBI) {
3103 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3104 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3106 SI->eraseFromParent();
3111 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3112 /// and use it to remove dead cases.
3113 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3114 Value *Cond = SI->getCondition();
3115 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
3116 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3117 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3119 // Gather dead cases.
3120 SmallVector<ConstantInt*, 8> DeadCases;
3121 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3122 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3123 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3124 DeadCases.push_back(I.getCaseValue());
3125 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3126 << I.getCaseValue() << "' is dead.\n");
3130 SmallVector<uint64_t, 8> Weights;
3131 bool HasWeight = HasBranchWeights(SI);
3133 GetBranchWeights(SI, Weights);
3134 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3137 // Remove dead cases from the switch.
3138 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3139 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3140 assert(Case != SI->case_default() &&
3141 "Case was not found. Probably mistake in DeadCases forming.");
3143 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3147 // Prune unused values from PHI nodes.
3148 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3149 SI->removeCase(Case);
3152 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3153 SI->setMetadata(LLVMContext::MD_prof,
3154 MDBuilder(SI->getParent()->getContext()).
3155 createBranchWeights(MDWeights));
3158 return !DeadCases.empty();
3161 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3162 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3163 /// by an unconditional branch), look at the phi node for BB in the successor
3164 /// block and see if the incoming value is equal to CaseValue. If so, return
3165 /// the phi node, and set PhiIndex to BB's index in the phi node.
3166 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3169 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3170 return NULL; // BB must be empty to be a candidate for simplification.
3171 if (!BB->getSinglePredecessor())
3172 return NULL; // BB must be dominated by the switch.
3174 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3175 if (!Branch || !Branch->isUnconditional())
3176 return NULL; // Terminator must be unconditional branch.
3178 BasicBlock *Succ = Branch->getSuccessor(0);
3180 BasicBlock::iterator I = Succ->begin();
3181 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3182 int Idx = PHI->getBasicBlockIndex(BB);
3183 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3185 Value *InValue = PHI->getIncomingValue(Idx);
3186 if (InValue != CaseValue) continue;
3195 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3196 /// instruction to a phi node dominated by the switch, if that would mean that
3197 /// some of the destination blocks of the switch can be folded away.
3198 /// Returns true if a change is made.
3199 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3200 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3201 ForwardingNodesMap ForwardingNodes;
3203 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3204 ConstantInt *CaseValue = I.getCaseValue();
3205 BasicBlock *CaseDest = I.getCaseSuccessor();
3208 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3212 ForwardingNodes[PHI].push_back(PhiIndex);
3215 bool Changed = false;
3217 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3218 E = ForwardingNodes.end(); I != E; ++I) {
3219 PHINode *Phi = I->first;
3220 SmallVector<int,4> &Indexes = I->second;
3222 if (Indexes.size() < 2) continue;
3224 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3225 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3232 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3233 /// initializing an array of constants like C.
3234 static bool ValidLookupTableConstant(Constant *C) {
3235 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3236 return CE->isGEPWithNoNotionalOverIndexing();
3238 return isa<ConstantFP>(C) ||
3239 isa<ConstantInt>(C) ||
3240 isa<ConstantPointerNull>(C) ||
3241 isa<GlobalValue>(C) ||
3245 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3246 /// its constant value in ConstantPool, returning 0 if it's not there.
3247 static Constant *LookupConstant(Value *V,
3248 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3249 if (Constant *C = dyn_cast<Constant>(V))
3251 return ConstantPool.lookup(V);
3254 /// ConstantFold - Try to fold instruction I into a constant. This works for
3255 /// simple instructions such as binary operations where both operands are
3256 /// constant or can be replaced by constants from the ConstantPool. Returns the
3257 /// resulting constant on success, 0 otherwise.
3258 static Constant *ConstantFold(Instruction *I,
3259 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3260 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3261 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool);
3264 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool);
3267 return ConstantExpr::get(BO->getOpcode(), A, B);
3270 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3271 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3274 Constant *B = LookupConstant(I->getOperand(1), ConstantPool);
3277 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B);
3280 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3281 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3284 if (A->isAllOnesValue())
3285 return LookupConstant(Select->getTrueValue(), ConstantPool);
3286 if (A->isNullValue())
3287 return LookupConstant(Select->getFalseValue(), ConstantPool);
3291 if (CastInst *Cast = dyn_cast<CastInst>(I)) {
3292 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3295 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy());
3301 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3302 /// at the common destination basic block, *CommonDest, for one of the case
3303 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3304 /// case), of a switch instruction SI.
3305 static bool GetCaseResults(SwitchInst *SI,
3306 ConstantInt *CaseVal,
3307 BasicBlock *CaseDest,
3308 BasicBlock **CommonDest,
3309 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) {
3310 // The block from which we enter the common destination.
3311 BasicBlock *Pred = SI->getParent();
3313 // If CaseDest is empty except for some side-effect free instructions through
3314 // which we can constant-propagate the CaseVal, continue to its successor.
3315 SmallDenseMap<Value*, Constant*> ConstantPool;
3316 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3317 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3319 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3320 // If the terminator is a simple branch, continue to the next block.
3321 if (T->getNumSuccessors() != 1)
3324 CaseDest = T->getSuccessor(0);
3325 } else if (isa<DbgInfoIntrinsic>(I)) {
3326 // Skip debug intrinsic.
3328 } else if (Constant *C = ConstantFold(I, ConstantPool)) {
3329 // Instruction is side-effect free and constant.
3330 ConstantPool.insert(std::make_pair(I, C));
3336 // If we did not have a CommonDest before, use the current one.
3338 *CommonDest = CaseDest;
3339 // If the destination isn't the common one, abort.
3340 if (CaseDest != *CommonDest)
3343 // Get the values for this case from phi nodes in the destination block.
3344 BasicBlock::iterator I = (*CommonDest)->begin();
3345 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3346 int Idx = PHI->getBasicBlockIndex(Pred);
3350 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3355 // Note: If the constant comes from constant-propagating the case value
3356 // through the CaseDest basic block, it will be safe to remove the
3357 // instructions in that block. They cannot be used (except in the phi nodes
3358 // we visit) outside CaseDest, because that block does not dominate its
3359 // successor. If it did, we would not be in this phi node.
3361 // Be conservative about which kinds of constants we support.
3362 if (!ValidLookupTableConstant(ConstVal))
3365 Res.push_back(std::make_pair(PHI, ConstVal));
3372 /// SwitchLookupTable - This class represents a lookup table that can be used
3373 /// to replace a switch.
3374 class SwitchLookupTable {
3376 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3377 /// with the contents of Values, using DefaultValue to fill any holes in the
3379 SwitchLookupTable(Module &M,
3381 ConstantInt *Offset,
3382 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3383 Constant *DefaultValue,
3384 const DataLayout *TD);
3386 /// BuildLookup - Build instructions with Builder to retrieve the value at
3387 /// the position given by Index in the lookup table.
3388 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3390 /// WouldFitInRegister - Return true if a table with TableSize elements of
3391 /// type ElementType would fit in a target-legal register.
3392 static bool WouldFitInRegister(const DataLayout *TD,
3394 const Type *ElementType);
3397 // Depending on the contents of the table, it can be represented in
3400 // For tables where each element contains the same value, we just have to
3401 // store that single value and return it for each lookup.
3404 // For small tables with integer elements, we can pack them into a bitmap
3405 // that fits into a target-legal register. Values are retrieved by
3406 // shift and mask operations.
3409 // The table is stored as an array of values. Values are retrieved by load
3410 // instructions from the table.
3414 // For SingleValueKind, this is the single value.
3415 Constant *SingleValue;
3417 // For BitMapKind, this is the bitmap.
3418 ConstantInt *BitMap;
3419 IntegerType *BitMapElementTy;
3421 // For ArrayKind, this is the array.
3422 GlobalVariable *Array;
3426 SwitchLookupTable::SwitchLookupTable(Module &M,
3428 ConstantInt *Offset,
3429 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3430 Constant *DefaultValue,
3431 const DataLayout *TD)
3432 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3433 assert(Values.size() && "Can't build lookup table without values!");
3434 assert(TableSize >= Values.size() && "Can't fit values in table!");
3436 // If all values in the table are equal, this is that value.
3437 SingleValue = Values.begin()->second;
3439 // Build up the table contents.
3440 SmallVector<Constant*, 64> TableContents(TableSize);
3441 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3442 ConstantInt *CaseVal = Values[I].first;
3443 Constant *CaseRes = Values[I].second;
3444 assert(CaseRes->getType() == DefaultValue->getType());
3446 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3448 TableContents[Idx] = CaseRes;
3450 if (CaseRes != SingleValue)
3454 // Fill in any holes in the table with the default result.
3455 if (Values.size() < TableSize) {
3456 for (uint64_t I = 0; I < TableSize; ++I) {
3457 if (!TableContents[I])
3458 TableContents[I] = DefaultValue;
3461 if (DefaultValue != SingleValue)
3465 // If each element in the table contains the same value, we only need to store
3466 // that single value.
3468 Kind = SingleValueKind;
3472 // If the type is integer and the table fits in a register, build a bitmap.
3473 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) {
3474 IntegerType *IT = cast<IntegerType>(DefaultValue->getType());
3475 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3476 for (uint64_t I = TableSize; I > 0; --I) {
3477 TableInt <<= IT->getBitWidth();
3478 // Insert values into the bitmap. Undef values are set to zero.
3479 if (!isa<UndefValue>(TableContents[I - 1])) {
3480 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3481 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3484 BitMap = ConstantInt::get(M.getContext(), TableInt);
3485 BitMapElementTy = IT;
3491 // Store the table in an array.
3492 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize);
3493 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3495 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3496 GlobalVariable::PrivateLinkage,
3499 Array->setUnnamedAddr(true);
3503 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3505 case SingleValueKind:
3508 // Type of the bitmap (e.g. i59).
3509 IntegerType *MapTy = BitMap->getType();
3511 // Cast Index to the same type as the bitmap.
3512 // Note: The Index is <= the number of elements in the table, so
3513 // truncating it to the width of the bitmask is safe.
3514 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3516 // Multiply the shift amount by the element width.
3517 ShiftAmt = Builder.CreateMul(ShiftAmt,
3518 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3522 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3523 "switch.downshift");
3525 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3529 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3530 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3532 return Builder.CreateLoad(GEP, "switch.load");
3535 llvm_unreachable("Unknown lookup table kind!");
3538 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD,
3540 const Type *ElementType) {
3543 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3546 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3547 // are <= 15, we could try to narrow the type.
3549 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3550 if (TableSize >= UINT_MAX/IT->getBitWidth())
3552 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth());
3555 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3556 /// for this switch, based on the number of caes, size of the table and the
3557 /// types of the results.
3558 static bool ShouldBuildLookupTable(SwitchInst *SI,
3560 const TargetTransformInfo &TTI,
3561 const DataLayout *TD,
3562 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3563 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3564 return false; // TableSize overflowed, or mul below might overflow.
3566 bool AllTablesFitInRegister = true;
3567 bool HasIllegalType = false;
3568 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3569 E = ResultTypes.end(); I != E; ++I) {
3570 Type *Ty = I->second;
3572 // Saturate this flag to true.
3573 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3575 // Saturate this flag to false.
3576 AllTablesFitInRegister = AllTablesFitInRegister &&
3577 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty);
3579 // If both flags saturate, we're done. NOTE: This *only* works with
3580 // saturating flags, and all flags have to saturate first due to the
3581 // non-deterministic behavior of iterating over a dense map.
3582 if (HasIllegalType && !AllTablesFitInRegister)
3586 // If each table would fit in a register, we should build it anyway.
3587 if (AllTablesFitInRegister)
3590 // Don't build a table that doesn't fit in-register if it has illegal types.
3594 // The table density should be at least 40%. This is the same criterion as for
3595 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3596 // FIXME: Find the best cut-off.
3597 return SI->getNumCases() * 10 >= TableSize * 4;
3600 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3601 /// phi nodes in a common successor block with different constant values,
3602 /// replace the switch with lookup tables.
3603 static bool SwitchToLookupTable(SwitchInst *SI,
3604 IRBuilder<> &Builder,
3605 const TargetTransformInfo &TTI,
3606 const DataLayout* TD) {
3607 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3609 // Only build lookup table when we have a target that supports it.
3610 if (!TTI.shouldBuildLookupTables())
3613 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3614 // split off a dense part and build a lookup table for that.
3616 // FIXME: This creates arrays of GEPs to constant strings, which means each
3617 // GEP needs a runtime relocation in PIC code. We should just build one big
3618 // string and lookup indices into that.
3620 // Ignore the switch if the number of cases is too small.
3621 // This is similar to the check when building jump tables in
3622 // SelectionDAGBuilder::handleJTSwitchCase.
3623 // FIXME: Determine the best cut-off.
3624 if (SI->getNumCases() < 4)
3627 // Figure out the corresponding result for each case value and phi node in the
3628 // common destination, as well as the the min and max case values.
3629 assert(SI->case_begin() != SI->case_end());
3630 SwitchInst::CaseIt CI = SI->case_begin();
3631 ConstantInt *MinCaseVal = CI.getCaseValue();
3632 ConstantInt *MaxCaseVal = CI.getCaseValue();
3634 BasicBlock *CommonDest = 0;
3635 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3636 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3637 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3638 SmallDenseMap<PHINode*, Type*> ResultTypes;
3639 SmallVector<PHINode*, 4> PHIs;
3641 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3642 ConstantInt *CaseVal = CI.getCaseValue();
3643 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3644 MinCaseVal = CaseVal;
3645 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3646 MaxCaseVal = CaseVal;
3648 // Resulting value at phi nodes for this case value.
3649 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3651 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3655 // Append the result from this case to the list for each phi.
3656 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3657 if (!ResultLists.count(I->first))
3658 PHIs.push_back(I->first);
3659 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3663 // Get the resulting values for the default case.
3664 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3665 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3666 DefaultResultsList))
3668 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3669 PHINode *PHI = DefaultResultsList[I].first;
3670 Constant *Result = DefaultResultsList[I].second;
3671 DefaultResults[PHI] = Result;
3672 ResultTypes[PHI] = Result->getType();
3675 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3676 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3677 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes))
3680 // Create the BB that does the lookups.
3681 Module &Mod = *CommonDest->getParent()->getParent();
3682 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3684 CommonDest->getParent(),
3687 // Check whether the condition value is within the case range, and branch to
3689 Builder.SetInsertPoint(SI);
3690 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3692 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3693 MinCaseVal->getType(), TableSize));
3694 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3696 // Populate the BB that does the lookups.
3697 Builder.SetInsertPoint(LookupBB);
3698 bool ReturnedEarly = false;
3699 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3700 PHINode *PHI = PHIs[I];
3702 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3703 DefaultResults[PHI], TD);
3705 Value *Result = Table.BuildLookup(TableIndex, Builder);
3707 // If the result is used to return immediately from the function, we want to
3708 // do that right here.
3709 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) &&
3710 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) {
3711 Builder.CreateRet(Result);
3712 ReturnedEarly = true;
3716 PHI->addIncoming(Result, LookupBB);
3720 Builder.CreateBr(CommonDest);
3722 // Remove the switch.
3723 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3724 BasicBlock *Succ = SI->getSuccessor(i);
3725 if (Succ == SI->getDefaultDest()) continue;
3726 Succ->removePredecessor(SI->getParent());
3728 SI->eraseFromParent();
3734 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3735 BasicBlock *BB = SI->getParent();
3737 if (isValueEqualityComparison(SI)) {
3738 // If we only have one predecessor, and if it is a branch on this value,
3739 // see if that predecessor totally determines the outcome of this switch.
3740 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3741 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3742 return SimplifyCFG(BB, TTI, TD) | true;
3744 Value *Cond = SI->getCondition();
3745 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3746 if (SimplifySwitchOnSelect(SI, Select))
3747 return SimplifyCFG(BB, TTI, TD) | true;
3749 // If the block only contains the switch, see if we can fold the block
3750 // away into any preds.
3751 BasicBlock::iterator BBI = BB->begin();
3752 // Ignore dbg intrinsics.
3753 while (isa<DbgInfoIntrinsic>(BBI))
3756 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3757 return SimplifyCFG(BB, TTI, TD) | true;
3760 // Try to transform the switch into an icmp and a branch.
3761 if (TurnSwitchRangeIntoICmp(SI, Builder))
3762 return SimplifyCFG(BB, TTI, TD) | true;
3764 // Remove unreachable cases.
3765 if (EliminateDeadSwitchCases(SI))
3766 return SimplifyCFG(BB, TTI, TD) | true;
3768 if (ForwardSwitchConditionToPHI(SI))
3769 return SimplifyCFG(BB, TTI, TD) | true;
3771 if (SwitchToLookupTable(SI, Builder, TTI, TD))
3772 return SimplifyCFG(BB, TTI, TD) | true;
3777 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3778 BasicBlock *BB = IBI->getParent();
3779 bool Changed = false;
3781 // Eliminate redundant destinations.
3782 SmallPtrSet<Value *, 8> Succs;
3783 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3784 BasicBlock *Dest = IBI->getDestination(i);
3785 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3786 Dest->removePredecessor(BB);
3787 IBI->removeDestination(i);
3793 if (IBI->getNumDestinations() == 0) {
3794 // If the indirectbr has no successors, change it to unreachable.
3795 new UnreachableInst(IBI->getContext(), IBI);
3796 EraseTerminatorInstAndDCECond(IBI);
3800 if (IBI->getNumDestinations() == 1) {
3801 // If the indirectbr has one successor, change it to a direct branch.
3802 BranchInst::Create(IBI->getDestination(0), IBI);
3803 EraseTerminatorInstAndDCECond(IBI);
3807 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3808 if (SimplifyIndirectBrOnSelect(IBI, SI))
3809 return SimplifyCFG(BB, TTI, TD) | true;
3814 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3815 BasicBlock *BB = BI->getParent();
3817 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3820 // If the Terminator is the only non-phi instruction, simplify the block.
3821 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3822 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3823 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3826 // If the only instruction in the block is a seteq/setne comparison
3827 // against a constant, try to simplify the block.
3828 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3829 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3830 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3832 if (I->isTerminator() &&
3833 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD))
3837 // If this basic block is ONLY a compare and a branch, and if a predecessor
3838 // branches to us and our successor, fold the comparison into the
3839 // predecessor and use logical operations to update the incoming value
3840 // for PHI nodes in common successor.
3841 if (FoldBranchToCommonDest(BI))
3842 return SimplifyCFG(BB, TTI, TD) | true;
3847 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3848 BasicBlock *BB = BI->getParent();
3850 // Conditional branch
3851 if (isValueEqualityComparison(BI)) {
3852 // If we only have one predecessor, and if it is a branch on this value,
3853 // see if that predecessor totally determines the outcome of this
3855 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3856 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3857 return SimplifyCFG(BB, TTI, TD) | true;
3859 // This block must be empty, except for the setcond inst, if it exists.
3860 // Ignore dbg intrinsics.
3861 BasicBlock::iterator I = BB->begin();
3862 // Ignore dbg intrinsics.
3863 while (isa<DbgInfoIntrinsic>(I))
3866 if (FoldValueComparisonIntoPredecessors(BI, Builder))
3867 return SimplifyCFG(BB, TTI, TD) | true;
3868 } else if (&*I == cast<Instruction>(BI->getCondition())){
3870 // Ignore dbg intrinsics.
3871 while (isa<DbgInfoIntrinsic>(I))
3873 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
3874 return SimplifyCFG(BB, TTI, TD) | true;
3878 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
3879 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
3882 // If this basic block is ONLY a compare and a branch, and if a predecessor
3883 // branches to us and one of our successors, fold the comparison into the
3884 // predecessor and use logical operations to pick the right destination.
3885 if (FoldBranchToCommonDest(BI))
3886 return SimplifyCFG(BB, TTI, TD) | true;
3888 // We have a conditional branch to two blocks that are only reachable
3889 // from BI. We know that the condbr dominates the two blocks, so see if
3890 // there is any identical code in the "then" and "else" blocks. If so, we
3891 // can hoist it up to the branching block.
3892 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
3893 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3894 if (HoistThenElseCodeToIf(BI))
3895 return SimplifyCFG(BB, TTI, TD) | true;
3897 // If Successor #1 has multiple preds, we may be able to conditionally
3898 // execute Successor #0 if it branches to successor #1.
3899 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
3900 if (Succ0TI->getNumSuccessors() == 1 &&
3901 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
3902 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), TTI))
3903 return SimplifyCFG(BB, TTI, TD) | true;
3905 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
3906 // If Successor #0 has multiple preds, we may be able to conditionally
3907 // execute Successor #1 if it branches to successor #0.
3908 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
3909 if (Succ1TI->getNumSuccessors() == 1 &&
3910 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
3911 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), TTI))
3912 return SimplifyCFG(BB, TTI, TD) | true;
3915 // If this is a branch on a phi node in the current block, thread control
3916 // through this block if any PHI node entries are constants.
3917 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
3918 if (PN->getParent() == BI->getParent())
3919 if (FoldCondBranchOnPHI(BI, TD))
3920 return SimplifyCFG(BB, TTI, TD) | true;
3922 // Scan predecessor blocks for conditional branches.
3923 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
3924 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
3925 if (PBI != BI && PBI->isConditional())
3926 if (SimplifyCondBranchToCondBranch(PBI, BI))
3927 return SimplifyCFG(BB, TTI, TD) | true;
3932 /// Check if passing a value to an instruction will cause undefined behavior.
3933 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
3934 Constant *C = dyn_cast<Constant>(V);
3941 if (C->isNullValue()) {
3942 // Only look at the first use, avoid hurting compile time with long uselists
3943 User *Use = *I->use_begin();
3945 // Now make sure that there are no instructions in between that can alter
3946 // control flow (eg. calls)
3947 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
3948 if (i == I->getParent()->end() || i->mayHaveSideEffects())
3951 // Look through GEPs. A load from a GEP derived from NULL is still undefined
3952 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
3953 if (GEP->getPointerOperand() == I)
3954 return passingValueIsAlwaysUndefined(V, GEP);
3956 // Look through bitcasts.
3957 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
3958 return passingValueIsAlwaysUndefined(V, BC);
3960 // Load from null is undefined.
3961 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
3962 return LI->getPointerAddressSpace() == 0;
3964 // Store to null is undefined.
3965 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
3966 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
3971 /// If BB has an incoming value that will always trigger undefined behavior
3972 /// (eg. null pointer dereference), remove the branch leading here.
3973 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
3974 for (BasicBlock::iterator i = BB->begin();
3975 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
3976 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
3977 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
3978 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
3979 IRBuilder<> Builder(T);
3980 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
3981 BB->removePredecessor(PHI->getIncomingBlock(i));
3982 // Turn uncoditional branches into unreachables and remove the dead
3983 // destination from conditional branches.
3984 if (BI->isUnconditional())
3985 Builder.CreateUnreachable();
3987 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
3988 BI->getSuccessor(0));
3989 BI->eraseFromParent();
3992 // TODO: SwitchInst.
3998 bool SimplifyCFGOpt::run(BasicBlock *BB) {
3999 bool Changed = false;
4001 assert(BB && BB->getParent() && "Block not embedded in function!");
4002 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4004 // Remove basic blocks that have no predecessors (except the entry block)...
4005 // or that just have themself as a predecessor. These are unreachable.
4006 if ((pred_begin(BB) == pred_end(BB) &&
4007 BB != &BB->getParent()->getEntryBlock()) ||
4008 BB->getSinglePredecessor() == BB) {
4009 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4010 DeleteDeadBlock(BB);
4014 // Check to see if we can constant propagate this terminator instruction
4016 Changed |= ConstantFoldTerminator(BB, true);
4018 // Check for and eliminate duplicate PHI nodes in this block.
4019 Changed |= EliminateDuplicatePHINodes(BB);
4021 // Check for and remove branches that will always cause undefined behavior.
4022 Changed |= removeUndefIntroducingPredecessor(BB);
4024 // Merge basic blocks into their predecessor if there is only one distinct
4025 // pred, and if there is only one distinct successor of the predecessor, and
4026 // if there are no PHI nodes.
4028 if (MergeBlockIntoPredecessor(BB))
4031 IRBuilder<> Builder(BB);
4033 // If there is a trivial two-entry PHI node in this basic block, and we can
4034 // eliminate it, do so now.
4035 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4036 if (PN->getNumIncomingValues() == 2)
4037 Changed |= FoldTwoEntryPHINode(PN, TD);
4039 Builder.SetInsertPoint(BB->getTerminator());
4040 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4041 if (BI->isUnconditional()) {
4042 if (SimplifyUncondBranch(BI, Builder)) return true;
4044 if (SimplifyCondBranch(BI, Builder)) return true;
4046 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4047 if (SimplifyReturn(RI, Builder)) return true;
4048 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4049 if (SimplifyResume(RI, Builder)) return true;
4050 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4051 if (SimplifySwitch(SI, Builder)) return true;
4052 } else if (UnreachableInst *UI =
4053 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4054 if (SimplifyUnreachable(UI)) return true;
4055 } else if (IndirectBrInst *IBI =
4056 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4057 if (SimplifyIndirectBr(IBI)) return true;
4063 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4064 /// example, it adjusts branches to branches to eliminate the extra hop, it
4065 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4066 /// of the CFG. It returns true if a modification was made.
4068 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4069 const DataLayout *TD) {
4070 return SimplifyCFGOpt(TTI, TD).run(BB);