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/AliasAnalysis.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/TargetTransformInfo.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/GlobalVariable.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/Metadata.h"
36 #include "llvm/IR/Module.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/IR/Type.h"
39 #include "llvm/Support/CFG.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/ConstantRange.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/NoFolder.h"
44 #include "llvm/Support/PatternMatch.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51 using namespace PatternMatch;
53 static cl::opt<unsigned>
54 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
55 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
58 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
59 cl::desc("Duplicate return instructions into unconditional branches"));
62 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
63 cl::desc("Sink common instructions down to the end block"));
66 HoistCondStores("simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
67 cl::desc("Hoist conditional stores if an unconditional store preceeds"));
70 ParallelAndOr("simplifycfg-parallel-and-or", cl::Hidden, cl::init(true),
71 cl::desc("Use parallel-and-or mode for branch conditions"));
73 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
74 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
75 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
76 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
79 /// ValueEqualityComparisonCase - Represents a case of a switch.
80 struct ValueEqualityComparisonCase {
84 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
85 : Value(Value), Dest(Dest) {}
87 bool operator<(ValueEqualityComparisonCase RHS) const {
88 // Comparing pointers is ok as we only rely on the order for uniquing.
89 return Value < RHS.Value;
92 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
95 class SimplifyCFGOpt {
96 const TargetTransformInfo &TTI;
97 const DataLayout *const TD;
100 Value *isValueEqualityComparison(TerminatorInst *TI);
101 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
102 std::vector<ValueEqualityComparisonCase> &Cases);
103 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
105 IRBuilder<> &Builder);
106 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
107 IRBuilder<> &Builder);
109 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
110 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
111 bool SimplifyUnreachable(UnreachableInst *UI);
112 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
113 bool SimplifyIndirectBr(IndirectBrInst *IBI);
114 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
115 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
116 /// \brief Use parallel-and or parallel-or to generate conditions for
117 /// conditional branches.
118 bool SimplifyParallelAndOr(BasicBlock *BB, IRBuilder<> &Builder, Pass *P = 0);
119 /// \brief If \param BB is the merge block of an if-region, attempt to merge
120 /// the if-region with an adjacent if-region upstream if two if-regions
121 /// contain identical instructions.
122 bool MergeIfRegion(BasicBlock *BB, IRBuilder<> &Builder, Pass *P = 0);
123 /// \brief Compare a pair of blocks: \p Block1 and \p Block2, which
124 /// are from two if-regions whose entry blocks are \p Head1 and \p
125 /// Head2. \returns true if \p Block1 and \p Block2 contain identical
126 /// instructions, and have no memory reference alias with \p Head2.
127 /// This is used as a legality check for merging if-regions.
128 bool CompareIfRegionBlock(BasicBlock *Head1, BasicBlock *Head2,
129 BasicBlock *Block1, BasicBlock *Block2);
132 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *TD,
134 : TTI(TTI), TD(TD), AA(AA) {}
135 bool run(BasicBlock *BB);
139 /// SafeToMergeTerminators - Return true if it is safe to merge these two
140 /// terminator instructions together.
142 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
143 if (SI1 == SI2) return false; // Can't merge with self!
145 // It is not safe to merge these two switch instructions if they have a common
146 // successor, and if that successor has a PHI node, and if *that* PHI node has
147 // conflicting incoming values from the two switch blocks.
148 BasicBlock *SI1BB = SI1->getParent();
149 BasicBlock *SI2BB = SI2->getParent();
150 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
152 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
153 if (SI1Succs.count(*I))
154 for (BasicBlock::iterator BBI = (*I)->begin();
155 isa<PHINode>(BBI); ++BBI) {
156 PHINode *PN = cast<PHINode>(BBI);
157 if (PN->getIncomingValueForBlock(SI1BB) !=
158 PN->getIncomingValueForBlock(SI2BB))
165 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
166 /// to merge these two terminator instructions together, where SI1 is an
167 /// unconditional branch. PhiNodes will store all PHI nodes in common
170 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
173 SmallVectorImpl<PHINode*> &PhiNodes) {
174 if (SI1 == SI2) return false; // Can't merge with self!
175 assert(SI1->isUnconditional() && SI2->isConditional());
177 // We fold the unconditional branch if we can easily update all PHI nodes in
178 // common successors:
179 // 1> We have a constant incoming value for the conditional branch;
180 // 2> We have "Cond" as the incoming value for the unconditional branch;
181 // 3> SI2->getCondition() and Cond have same operands.
182 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
183 if (!Ci2) return false;
184 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
185 Cond->getOperand(1) == Ci2->getOperand(1)) &&
186 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
187 Cond->getOperand(1) == Ci2->getOperand(0)))
190 BasicBlock *SI1BB = SI1->getParent();
191 BasicBlock *SI2BB = SI2->getParent();
192 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
193 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
194 if (SI1Succs.count(*I))
195 for (BasicBlock::iterator BBI = (*I)->begin();
196 isa<PHINode>(BBI); ++BBI) {
197 PHINode *PN = cast<PHINode>(BBI);
198 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
199 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
201 PhiNodes.push_back(PN);
206 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
207 /// now be entries in it from the 'NewPred' block. The values that will be
208 /// flowing into the PHI nodes will be the same as those coming in from
209 /// ExistPred, an existing predecessor of Succ.
210 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
211 BasicBlock *ExistPred) {
212 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
215 for (BasicBlock::iterator I = Succ->begin();
216 (PN = dyn_cast<PHINode>(I)); ++I)
217 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
221 /// GetIfCondition - Given a basic block (BB) with two predecessors,
222 /// check to see if the merge at this block is due
223 /// to an "if condition". If so, return the boolean condition that determines
224 /// which entry into BB will be taken. Also, return by references the block
225 /// that will be entered from if the condition is true, and the block that will
226 /// be entered if the condition is false.
228 /// This does no checking to see if the true/false blocks have large or unsavory
229 /// instructions in them.
230 static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue,
231 BasicBlock *&IfFalse) {
232 PHINode *SomePHI = dyn_cast<PHINode>(BB->begin());
233 BasicBlock *Pred1 = NULL;
234 BasicBlock *Pred2 = NULL;
237 if (SomePHI->getNumIncomingValues() != 2)
239 Pred1 = SomePHI->getIncomingBlock(0);
240 Pred2 = SomePHI->getIncomingBlock(1);
242 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
243 if (PI == PE) // No predecessor
246 if (PI == PE) // Only one predecessor
249 if (PI != PE) // More than two predecessors
253 // We can only handle branches. Other control flow will be lowered to
254 // branches if possible anyway.
255 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator());
256 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator());
257 if (Pred1Br == 0 || Pred2Br == 0)
260 // Eliminate code duplication by ensuring that Pred1Br is conditional if
262 if (Pred2Br->isConditional()) {
263 // If both branches are conditional, we don't have an "if statement". In
264 // reality, we could transform this case, but since the condition will be
265 // required anyway, we stand no chance of eliminating it, so the xform is
266 // probably not profitable.
267 if (Pred1Br->isConditional())
270 std::swap(Pred1, Pred2);
271 std::swap(Pred1Br, Pred2Br);
274 if (Pred1Br->isConditional()) {
275 // The only thing we have to watch out for here is to make sure that Pred2
276 // doesn't have incoming edges from other blocks. If it does, the condition
277 // doesn't dominate BB.
278 if (Pred2->getSinglePredecessor() == 0)
281 // If we found a conditional branch predecessor, make sure that it branches
282 // to BB and Pred2Br. If it doesn't, this isn't an "if statement".
283 if (Pred1Br->getSuccessor(0) == BB &&
284 Pred1Br->getSuccessor(1) == Pred2) {
287 } else if (Pred1Br->getSuccessor(0) == Pred2 &&
288 Pred1Br->getSuccessor(1) == BB) {
292 // We know that one arm of the conditional goes to BB, so the other must
293 // go somewhere unrelated, and this must not be an "if statement".
297 return Pred1Br->getCondition();
300 // Ok, if we got here, both predecessors end with an unconditional branch to
301 // BB. Don't panic! If both blocks only have a single (identical)
302 // predecessor, and THAT is a conditional branch, then we're all ok!
303 BasicBlock *CommonPred = Pred1->getSinglePredecessor();
304 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor())
307 // Otherwise, if this is a conditional branch, then we can use it!
308 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator());
309 if (BI == 0) return 0;
311 assert(BI->isConditional() && "Two successors but not conditional?");
312 if (BI->getSuccessor(0) == Pred1) {
319 return BI->getCondition();
322 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
323 /// given instruction, which is assumed to be safe to speculate. 1 means
324 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
325 static unsigned ComputeSpeculationCost(const User *I) {
326 assert(isSafeToSpeculativelyExecute(I) &&
327 "Instruction is not safe to speculatively execute!");
328 switch (Operator::getOpcode(I)) {
330 // In doubt, be conservative.
332 case Instruction::GetElementPtr:
333 // GEPs are cheap if all indices are constant.
334 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
337 case Instruction::Load:
338 case Instruction::Add:
339 case Instruction::Sub:
340 case Instruction::And:
341 case Instruction::Or:
342 case Instruction::Xor:
343 case Instruction::Shl:
344 case Instruction::LShr:
345 case Instruction::AShr:
346 case Instruction::ICmp:
347 case Instruction::Trunc:
348 case Instruction::ZExt:
349 case Instruction::SExt:
350 return 1; // These are all cheap.
352 case Instruction::Call:
353 case Instruction::Select:
358 /// DominatesMergePoint - If we have a merge point of an "if condition" as
359 /// accepted above, return true if the specified value dominates the block. We
360 /// don't handle the true generality of domination here, just a special case
361 /// which works well enough for us.
363 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
364 /// see if V (which must be an instruction) and its recursive operands
365 /// that do not dominate BB have a combined cost lower than CostRemaining and
366 /// are non-trapping. If both are true, the instruction is inserted into the
367 /// set and true is returned.
369 /// The cost for most non-trapping instructions is defined as 1 except for
370 /// Select whose cost is 2.
372 /// After this function returns, CostRemaining is decreased by the cost of
373 /// V plus its non-dominating operands. If that cost is greater than
374 /// CostRemaining, false is returned and CostRemaining is undefined.
375 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
376 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
377 unsigned &CostRemaining) {
378 Instruction *I = dyn_cast<Instruction>(V);
380 // Non-instructions all dominate instructions, but not all constantexprs
381 // can be executed unconditionally.
382 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
387 BasicBlock *PBB = I->getParent();
389 // We don't want to allow weird loops that might have the "if condition" in
390 // the bottom of this block.
391 if (PBB == BB) return false;
393 // If this instruction is defined in a block that contains an unconditional
394 // branch to BB, then it must be in the 'conditional' part of the "if
395 // statement". If not, it definitely dominates the region.
396 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
397 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB)
400 // If we aren't allowing aggressive promotion anymore, then don't consider
401 // instructions in the 'if region'.
402 if (AggressiveInsts == 0) return false;
404 // If we have seen this instruction before, don't count it again.
405 if (AggressiveInsts->count(I)) return true;
407 // Okay, it looks like the instruction IS in the "condition". Check to
408 // see if it's a cheap instruction to unconditionally compute, and if it
409 // only uses stuff defined outside of the condition. If so, hoist it out.
410 if (!isSafeToSpeculativelyExecute(I))
413 unsigned Cost = ComputeSpeculationCost(I);
415 if (Cost > CostRemaining)
418 CostRemaining -= Cost;
420 // Okay, we can only really hoist these out if their operands do
421 // not take us over the cost threshold.
422 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
423 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining))
425 // Okay, it's safe to do this! Remember this instruction.
426 AggressiveInsts->insert(I);
430 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
431 /// and PointerNullValue. Return NULL if value is not a constant int.
432 static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) {
433 // Normal constant int.
434 ConstantInt *CI = dyn_cast<ConstantInt>(V);
435 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy())
438 // This is some kind of pointer constant. Turn it into a pointer-sized
439 // ConstantInt if possible.
440 IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType()));
442 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
443 if (isa<ConstantPointerNull>(V))
444 return ConstantInt::get(PtrTy, 0);
446 // IntToPtr const int.
447 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
448 if (CE->getOpcode() == Instruction::IntToPtr)
449 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
450 // The constant is very likely to have the right type already.
451 if (CI->getType() == PtrTy)
454 return cast<ConstantInt>
455 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
460 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
461 /// collection of icmp eq/ne instructions that compare a value against a
462 /// constant, return the value being compared, and stick the constant into the
465 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
466 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) {
467 Instruction *I = dyn_cast<Instruction>(V);
468 if (I == 0) return 0;
470 // If this is an icmp against a constant, handle this as one of the cases.
471 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
472 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) {
476 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
477 // (x & ~2^x) == y --> x == y || x == y|2^x
478 // This undoes a transformation done by instcombine to fuse 2 compares.
479 if (match(ICI->getOperand(0),
480 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
481 APInt Not = ~RHSC->getValue();
482 if (Not.isPowerOf2()) {
485 ConstantInt::get(C->getContext(), C->getValue() | Not));
493 return I->getOperand(0);
496 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
499 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
501 // Shift the range if the compare is fed by an add. This is the range
502 // compare idiom as emitted by instcombine.
504 match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)));
506 Span = Span.subtract(RHSC->getValue());
508 // If this is an and/!= check then we want to optimize "x ugt 2" into
511 Span = Span.inverse();
513 // If there are a ton of values, we don't want to make a ginormous switch.
514 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
517 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
518 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
520 return hasAdd ? RHSVal : I->getOperand(0);
525 // Otherwise, we can only handle an | or &, depending on isEQ.
526 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
529 unsigned NumValsBeforeLHS = Vals.size();
530 unsigned UsedICmpsBeforeLHS = UsedICmps;
531 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD,
533 unsigned NumVals = Vals.size();
534 unsigned UsedICmpsBeforeRHS = UsedICmps;
535 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
539 Vals.resize(NumVals);
540 UsedICmps = UsedICmpsBeforeRHS;
543 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
544 // set it and return success.
545 if (Extra == 0 || Extra == I->getOperand(1)) {
546 Extra = I->getOperand(1);
550 Vals.resize(NumValsBeforeLHS);
551 UsedICmps = UsedICmpsBeforeLHS;
555 // If the LHS can't be folded in, but Extra is available and RHS can, try to
557 if (Extra == 0 || Extra == I->getOperand(0)) {
558 Value *OldExtra = Extra;
559 Extra = I->getOperand(0);
560 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
563 assert(Vals.size() == NumValsBeforeLHS);
570 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
571 Instruction *Cond = 0;
572 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
573 Cond = dyn_cast<Instruction>(SI->getCondition());
574 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
575 if (BI->isConditional())
576 Cond = dyn_cast<Instruction>(BI->getCondition());
577 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
578 Cond = dyn_cast<Instruction>(IBI->getAddress());
581 TI->eraseFromParent();
582 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
585 /// isValueEqualityComparison - Return true if the specified terminator checks
586 /// to see if a value is equal to constant integer value.
587 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
589 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
590 // Do not permit merging of large switch instructions into their
591 // predecessors unless there is only one predecessor.
592 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
593 pred_end(SI->getParent())) <= 128)
594 CV = SI->getCondition();
595 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
596 if (BI->isConditional() && BI->getCondition()->hasOneUse())
597 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
598 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), TD))
599 CV = ICI->getOperand(0);
601 // Unwrap any lossless ptrtoint cast.
602 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext()))
603 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV))
604 CV = PTII->getOperand(0);
608 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
609 /// decode all of the 'cases' that it represents and return the 'default' block.
610 BasicBlock *SimplifyCFGOpt::
611 GetValueEqualityComparisonCases(TerminatorInst *TI,
612 std::vector<ValueEqualityComparisonCase>
614 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
615 Cases.reserve(SI->getNumCases());
616 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
617 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
618 i.getCaseSuccessor()));
619 return SI->getDefaultDest();
622 BranchInst *BI = cast<BranchInst>(TI);
623 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
624 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
625 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
628 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
632 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
633 /// in the list that match the specified block.
634 static void EliminateBlockCases(BasicBlock *BB,
635 std::vector<ValueEqualityComparisonCase> &Cases) {
636 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
639 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
642 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
643 std::vector<ValueEqualityComparisonCase > &C2) {
644 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
646 // Make V1 be smaller than V2.
647 if (V1->size() > V2->size())
650 if (V1->size() == 0) return false;
651 if (V1->size() == 1) {
653 ConstantInt *TheVal = (*V1)[0].Value;
654 for (unsigned i = 0, e = V2->size(); i != e; ++i)
655 if (TheVal == (*V2)[i].Value)
659 // Otherwise, just sort both lists and compare element by element.
660 array_pod_sort(V1->begin(), V1->end());
661 array_pod_sort(V2->begin(), V2->end());
662 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
663 while (i1 != e1 && i2 != e2) {
664 if ((*V1)[i1].Value == (*V2)[i2].Value)
666 if ((*V1)[i1].Value < (*V2)[i2].Value)
674 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
675 /// terminator instruction and its block is known to only have a single
676 /// predecessor block, check to see if that predecessor is also a value
677 /// comparison with the same value, and if that comparison determines the
678 /// outcome of this comparison. If so, simplify TI. This does a very limited
679 /// form of jump threading.
680 bool SimplifyCFGOpt::
681 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
683 IRBuilder<> &Builder) {
684 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
685 if (!PredVal) return false; // Not a value comparison in predecessor.
687 Value *ThisVal = isValueEqualityComparison(TI);
688 assert(ThisVal && "This isn't a value comparison!!");
689 if (ThisVal != PredVal) return false; // Different predicates.
691 // TODO: Preserve branch weight metadata, similarly to how
692 // FoldValueComparisonIntoPredecessors preserves it.
694 // Find out information about when control will move from Pred to TI's block.
695 std::vector<ValueEqualityComparisonCase> PredCases;
696 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
698 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
700 // Find information about how control leaves this block.
701 std::vector<ValueEqualityComparisonCase> ThisCases;
702 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
703 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
705 // If TI's block is the default block from Pred's comparison, potentially
706 // simplify TI based on this knowledge.
707 if (PredDef == TI->getParent()) {
708 // If we are here, we know that the value is none of those cases listed in
709 // PredCases. If there are any cases in ThisCases that are in PredCases, we
711 if (!ValuesOverlap(PredCases, ThisCases))
714 if (isa<BranchInst>(TI)) {
715 // Okay, one of the successors of this condbr is dead. Convert it to a
717 assert(ThisCases.size() == 1 && "Branch can only have one case!");
718 // Insert the new branch.
719 Instruction *NI = Builder.CreateBr(ThisDef);
722 // Remove PHI node entries for the dead edge.
723 ThisCases[0].Dest->removePredecessor(TI->getParent());
725 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
726 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
728 EraseTerminatorInstAndDCECond(TI);
732 SwitchInst *SI = cast<SwitchInst>(TI);
733 // Okay, TI has cases that are statically dead, prune them away.
734 SmallPtrSet<Constant*, 16> DeadCases;
735 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
736 DeadCases.insert(PredCases[i].Value);
738 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
739 << "Through successor TI: " << *TI);
741 // Collect branch weights into a vector.
742 SmallVector<uint32_t, 8> Weights;
743 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
744 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
746 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
748 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
750 Weights.push_back(CI->getValue().getZExtValue());
752 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
754 if (DeadCases.count(i.getCaseValue())) {
756 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
759 i.getCaseSuccessor()->removePredecessor(TI->getParent());
763 if (HasWeight && Weights.size() >= 2)
764 SI->setMetadata(LLVMContext::MD_prof,
765 MDBuilder(SI->getParent()->getContext()).
766 createBranchWeights(Weights));
768 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
772 // Otherwise, TI's block must correspond to some matched value. Find out
773 // which value (or set of values) this is.
774 ConstantInt *TIV = 0;
775 BasicBlock *TIBB = TI->getParent();
776 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
777 if (PredCases[i].Dest == TIBB) {
779 return false; // Cannot handle multiple values coming to this block.
780 TIV = PredCases[i].Value;
782 assert(TIV && "No edge from pred to succ?");
784 // Okay, we found the one constant that our value can be if we get into TI's
785 // BB. Find out which successor will unconditionally be branched to.
786 BasicBlock *TheRealDest = 0;
787 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
788 if (ThisCases[i].Value == TIV) {
789 TheRealDest = ThisCases[i].Dest;
793 // If not handled by any explicit cases, it is handled by the default case.
794 if (TheRealDest == 0) TheRealDest = ThisDef;
796 // Remove PHI node entries for dead edges.
797 BasicBlock *CheckEdge = TheRealDest;
798 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
799 if (*SI != CheckEdge)
800 (*SI)->removePredecessor(TIBB);
804 // Insert the new branch.
805 Instruction *NI = Builder.CreateBr(TheRealDest);
808 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
809 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
811 EraseTerminatorInstAndDCECond(TI);
816 /// ConstantIntOrdering - This class implements a stable ordering of constant
817 /// integers that does not depend on their address. This is important for
818 /// applications that sort ConstantInt's to ensure uniqueness.
819 struct ConstantIntOrdering {
820 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
821 return LHS->getValue().ult(RHS->getValue());
826 static int ConstantIntSortPredicate(const void *P1, const void *P2) {
827 const ConstantInt *LHS = *(const ConstantInt*const*)P1;
828 const ConstantInt *RHS = *(const ConstantInt*const*)P2;
829 if (LHS->getValue().ult(RHS->getValue()))
831 if (LHS->getValue() == RHS->getValue())
836 static inline bool HasBranchWeights(const Instruction* I) {
837 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
838 if (ProfMD && ProfMD->getOperand(0))
839 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
840 return MDS->getString().equals("branch_weights");
845 /// Get Weights of a given TerminatorInst, the default weight is at the front
846 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
848 static void GetBranchWeights(TerminatorInst *TI,
849 SmallVectorImpl<uint64_t> &Weights) {
850 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
852 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
853 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i));
855 Weights.push_back(CI->getValue().getZExtValue());
858 // If TI is a conditional eq, the default case is the false case,
859 // and the corresponding branch-weight data is at index 2. We swap the
860 // default weight to be the first entry.
861 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
862 assert(Weights.size() == 2);
863 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
864 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
865 std::swap(Weights.front(), Weights.back());
869 /// Sees if any of the weights are too big for a uint32_t, and halves all the
870 /// weights if any are.
871 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
873 for (unsigned i = 0; i < Weights.size(); ++i)
874 if (Weights[i] > UINT_MAX) {
882 for (unsigned i = 0; i < Weights.size(); ++i)
886 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
887 /// equality comparison instruction (either a switch or a branch on "X == c").
888 /// See if any of the predecessors of the terminator block are value comparisons
889 /// on the same value. If so, and if safe to do so, fold them together.
890 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
891 IRBuilder<> &Builder) {
892 BasicBlock *BB = TI->getParent();
893 Value *CV = isValueEqualityComparison(TI); // CondVal
894 assert(CV && "Not a comparison?");
895 bool Changed = false;
897 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
898 while (!Preds.empty()) {
899 BasicBlock *Pred = Preds.pop_back_val();
901 // See if the predecessor is a comparison with the same value.
902 TerminatorInst *PTI = Pred->getTerminator();
903 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
905 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
906 // Figure out which 'cases' to copy from SI to PSI.
907 std::vector<ValueEqualityComparisonCase> BBCases;
908 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
910 std::vector<ValueEqualityComparisonCase> PredCases;
911 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
913 // Based on whether the default edge from PTI goes to BB or not, fill in
914 // PredCases and PredDefault with the new switch cases we would like to
916 SmallVector<BasicBlock*, 8> NewSuccessors;
918 // Update the branch weight metadata along the way
919 SmallVector<uint64_t, 8> Weights;
920 bool PredHasWeights = HasBranchWeights(PTI);
921 bool SuccHasWeights = HasBranchWeights(TI);
923 if (PredHasWeights) {
924 GetBranchWeights(PTI, Weights);
925 // branch-weight metadata is inconsistent here.
926 if (Weights.size() != 1 + PredCases.size())
927 PredHasWeights = SuccHasWeights = false;
928 } else if (SuccHasWeights)
929 // If there are no predecessor weights but there are successor weights,
930 // populate Weights with 1, which will later be scaled to the sum of
931 // successor's weights
932 Weights.assign(1 + PredCases.size(), 1);
934 SmallVector<uint64_t, 8> SuccWeights;
935 if (SuccHasWeights) {
936 GetBranchWeights(TI, SuccWeights);
937 // branch-weight metadata is inconsistent here.
938 if (SuccWeights.size() != 1 + BBCases.size())
939 PredHasWeights = SuccHasWeights = false;
940 } else if (PredHasWeights)
941 SuccWeights.assign(1 + BBCases.size(), 1);
943 if (PredDefault == BB) {
944 // If this is the default destination from PTI, only the edges in TI
945 // that don't occur in PTI, or that branch to BB will be activated.
946 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
947 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
948 if (PredCases[i].Dest != BB)
949 PTIHandled.insert(PredCases[i].Value);
951 // The default destination is BB, we don't need explicit targets.
952 std::swap(PredCases[i], PredCases.back());
954 if (PredHasWeights || SuccHasWeights) {
955 // Increase weight for the default case.
956 Weights[0] += Weights[i+1];
957 std::swap(Weights[i+1], Weights.back());
961 PredCases.pop_back();
965 // Reconstruct the new switch statement we will be building.
966 if (PredDefault != BBDefault) {
967 PredDefault->removePredecessor(Pred);
968 PredDefault = BBDefault;
969 NewSuccessors.push_back(BBDefault);
972 unsigned CasesFromPred = Weights.size();
973 uint64_t ValidTotalSuccWeight = 0;
974 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
975 if (!PTIHandled.count(BBCases[i].Value) &&
976 BBCases[i].Dest != BBDefault) {
977 PredCases.push_back(BBCases[i]);
978 NewSuccessors.push_back(BBCases[i].Dest);
979 if (SuccHasWeights || PredHasWeights) {
980 // The default weight is at index 0, so weight for the ith case
981 // should be at index i+1. Scale the cases from successor by
982 // PredDefaultWeight (Weights[0]).
983 Weights.push_back(Weights[0] * SuccWeights[i+1]);
984 ValidTotalSuccWeight += SuccWeights[i+1];
988 if (SuccHasWeights || PredHasWeights) {
989 ValidTotalSuccWeight += SuccWeights[0];
990 // Scale the cases from predecessor by ValidTotalSuccWeight.
991 for (unsigned i = 1; i < CasesFromPred; ++i)
992 Weights[i] *= ValidTotalSuccWeight;
993 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
994 Weights[0] *= SuccWeights[0];
997 // If this is not the default destination from PSI, only the edges
998 // in SI that occur in PSI with a destination of BB will be
1000 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
1001 std::map<ConstantInt*, uint64_t> WeightsForHandled;
1002 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1003 if (PredCases[i].Dest == BB) {
1004 PTIHandled.insert(PredCases[i].Value);
1006 if (PredHasWeights || SuccHasWeights) {
1007 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
1008 std::swap(Weights[i+1], Weights.back());
1012 std::swap(PredCases[i], PredCases.back());
1013 PredCases.pop_back();
1017 // Okay, now we know which constants were sent to BB from the
1018 // predecessor. Figure out where they will all go now.
1019 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
1020 if (PTIHandled.count(BBCases[i].Value)) {
1021 // If this is one we are capable of getting...
1022 if (PredHasWeights || SuccHasWeights)
1023 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
1024 PredCases.push_back(BBCases[i]);
1025 NewSuccessors.push_back(BBCases[i].Dest);
1026 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
1029 // If there are any constants vectored to BB that TI doesn't handle,
1030 // they must go to the default destination of TI.
1031 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
1033 E = PTIHandled.end(); I != E; ++I) {
1034 if (PredHasWeights || SuccHasWeights)
1035 Weights.push_back(WeightsForHandled[*I]);
1036 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
1037 NewSuccessors.push_back(BBDefault);
1041 // Okay, at this point, we know which new successor Pred will get. Make
1042 // sure we update the number of entries in the PHI nodes for these
1044 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
1045 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
1047 Builder.SetInsertPoint(PTI);
1048 // Convert pointer to int before we switch.
1049 if (CV->getType()->isPointerTy()) {
1050 assert(TD && "Cannot switch on pointer without DataLayout");
1051 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()),
1055 // Now that the successors are updated, create the new Switch instruction.
1056 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
1058 NewSI->setDebugLoc(PTI->getDebugLoc());
1059 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1060 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
1062 if (PredHasWeights || SuccHasWeights) {
1063 // Halve the weights if any of them cannot fit in an uint32_t
1064 FitWeights(Weights);
1066 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1068 NewSI->setMetadata(LLVMContext::MD_prof,
1069 MDBuilder(BB->getContext()).
1070 createBranchWeights(MDWeights));
1073 EraseTerminatorInstAndDCECond(PTI);
1075 // Okay, last check. If BB is still a successor of PSI, then we must
1076 // have an infinite loop case. If so, add an infinitely looping block
1077 // to handle the case to preserve the behavior of the code.
1078 BasicBlock *InfLoopBlock = 0;
1079 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1080 if (NewSI->getSuccessor(i) == BB) {
1081 if (InfLoopBlock == 0) {
1082 // Insert it at the end of the function, because it's either code,
1083 // or it won't matter if it's hot. :)
1084 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1085 "infloop", BB->getParent());
1086 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1088 NewSI->setSuccessor(i, InfLoopBlock);
1097 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1098 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1099 // would need to do this), we can't hoist the invoke, as there is nowhere
1100 // to put the select in this case.
1101 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1102 Instruction *I1, Instruction *I2) {
1103 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1105 for (BasicBlock::iterator BBI = SI->begin();
1106 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1107 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1108 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1109 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1117 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1118 /// BB2, hoist any common code in the two blocks up into the branch block. The
1119 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1120 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1121 // This does very trivial matching, with limited scanning, to find identical
1122 // instructions in the two blocks. In particular, we don't want to get into
1123 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1124 // such, we currently just scan for obviously identical instructions in an
1126 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1127 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1129 BasicBlock::iterator BB1_Itr = BB1->begin();
1130 BasicBlock::iterator BB2_Itr = BB2->begin();
1132 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1133 // Skip debug info if it is not identical.
1134 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1135 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1136 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1137 while (isa<DbgInfoIntrinsic>(I1))
1139 while (isa<DbgInfoIntrinsic>(I2))
1142 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1143 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1146 BasicBlock *BIParent = BI->getParent();
1148 bool Changed = false;
1150 // If we are hoisting the terminator instruction, don't move one (making a
1151 // broken BB), instead clone it, and remove BI.
1152 if (isa<TerminatorInst>(I1))
1153 goto HoistTerminator;
1155 // For a normal instruction, we just move one to right before the branch,
1156 // then replace all uses of the other with the first. Finally, we remove
1157 // the now redundant second instruction.
1158 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1159 if (!I2->use_empty())
1160 I2->replaceAllUsesWith(I1);
1161 I1->intersectOptionalDataWith(I2);
1162 I2->eraseFromParent();
1167 // Skip debug info if it is not identical.
1168 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1169 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1170 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1171 while (isa<DbgInfoIntrinsic>(I1))
1173 while (isa<DbgInfoIntrinsic>(I2))
1176 } while (I1->isIdenticalToWhenDefined(I2));
1181 // It may not be possible to hoist an invoke.
1182 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1185 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1187 for (BasicBlock::iterator BBI = SI->begin();
1188 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1189 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1190 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1194 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1196 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1201 // Okay, it is safe to hoist the terminator.
1202 Instruction *NT = I1->clone();
1203 BIParent->getInstList().insert(BI, NT);
1204 if (!NT->getType()->isVoidTy()) {
1205 I1->replaceAllUsesWith(NT);
1206 I2->replaceAllUsesWith(NT);
1210 IRBuilder<true, NoFolder> Builder(NT);
1211 // Hoisting one of the terminators from our successor is a great thing.
1212 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1213 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1214 // nodes, so we insert select instruction to compute the final result.
1215 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1216 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1218 for (BasicBlock::iterator BBI = SI->begin();
1219 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1220 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1221 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1222 if (BB1V == BB2V) continue;
1224 // These values do not agree. Insert a select instruction before NT
1225 // that determines the right value.
1226 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1228 SI = cast<SelectInst>
1229 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1230 BB1V->getName()+"."+BB2V->getName()));
1232 // Make the PHI node use the select for all incoming values for BB1/BB2
1233 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1234 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1235 PN->setIncomingValue(i, SI);
1239 // Update any PHI nodes in our new successors.
1240 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1241 AddPredecessorToBlock(*SI, BIParent, BB1);
1243 EraseTerminatorInstAndDCECond(BI);
1247 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1248 /// check whether BBEnd has only two predecessors and the other predecessor
1249 /// ends with an unconditional branch. If it is true, sink any common code
1250 /// in the two predecessors to BBEnd.
1251 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1252 assert(BI1->isUnconditional());
1253 BasicBlock *BB1 = BI1->getParent();
1254 BasicBlock *BBEnd = BI1->getSuccessor(0);
1256 // Check that BBEnd has two predecessors and the other predecessor ends with
1257 // an unconditional branch.
1258 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1259 BasicBlock *Pred0 = *PI++;
1260 if (PI == PE) // Only one predecessor.
1262 BasicBlock *Pred1 = *PI++;
1263 if (PI != PE) // More than two predecessors.
1265 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1266 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1267 if (!BI2 || !BI2->isUnconditional())
1270 // Gather the PHI nodes in BBEnd.
1271 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1272 Instruction *FirstNonPhiInBBEnd = 0;
1273 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1275 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1276 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1277 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1278 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1280 FirstNonPhiInBBEnd = &*I;
1284 if (!FirstNonPhiInBBEnd)
1288 // This does very trivial matching, with limited scanning, to find identical
1289 // instructions in the two blocks. We scan backward for obviously identical
1290 // instructions in an identical order.
1291 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1292 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1293 RE2 = BB2->getInstList().rend();
1295 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1298 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1301 // Skip the unconditional branches.
1305 bool Changed = false;
1306 while (RI1 != RE1 && RI2 != RE2) {
1308 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1311 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1315 Instruction *I1 = &*RI1, *I2 = &*RI2;
1316 // I1 and I2 should have a single use in the same PHI node, and they
1317 // perform the same operation.
1318 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1319 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1320 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1321 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1322 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1323 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1324 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1325 !I1->hasOneUse() || !I2->hasOneUse() ||
1326 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1327 MapValueFromBB1ToBB2[I1].first != I2)
1330 // Check whether we should swap the operands of ICmpInst.
1331 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1332 bool SwapOpnds = false;
1333 if (ICmp1 && ICmp2 &&
1334 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1335 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1336 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1337 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1338 ICmp2->swapOperands();
1341 if (!I1->isSameOperationAs(I2)) {
1343 ICmp2->swapOperands();
1347 // The operands should be either the same or they need to be generated
1348 // with a PHI node after sinking. We only handle the case where there is
1349 // a single pair of different operands.
1350 Value *DifferentOp1 = 0, *DifferentOp2 = 0;
1351 unsigned Op1Idx = 0;
1352 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1353 if (I1->getOperand(I) == I2->getOperand(I))
1355 // Early exit if we have more-than one pair of different operands or
1356 // the different operand is already in MapValueFromBB1ToBB2.
1357 // Early exit if we need a PHI node to replace a constant.
1359 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1360 MapValueFromBB1ToBB2.end() ||
1361 isa<Constant>(I1->getOperand(I)) ||
1362 isa<Constant>(I2->getOperand(I))) {
1363 // If we can't sink the instructions, undo the swapping.
1365 ICmp2->swapOperands();
1368 DifferentOp1 = I1->getOperand(I);
1370 DifferentOp2 = I2->getOperand(I);
1373 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1374 // remove (I1, I2) from MapValueFromBB1ToBB2.
1376 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1377 DifferentOp1->getName() + ".sink",
1379 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1380 // I1 should use NewPN instead of DifferentOp1.
1381 I1->setOperand(Op1Idx, NewPN);
1382 NewPN->addIncoming(DifferentOp1, BB1);
1383 NewPN->addIncoming(DifferentOp2, BB2);
1384 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1386 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1387 MapValueFromBB1ToBB2.erase(I1);
1389 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1390 DEBUG(dbgs() << " " << *I2 << "\n";);
1391 // We need to update RE1 and RE2 if we are going to sink the first
1392 // instruction in the basic block down.
1393 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1394 // Sink the instruction.
1395 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1396 if (!OldPN->use_empty())
1397 OldPN->replaceAllUsesWith(I1);
1398 OldPN->eraseFromParent();
1400 if (!I2->use_empty())
1401 I2->replaceAllUsesWith(I1);
1402 I1->intersectOptionalDataWith(I2);
1403 I2->eraseFromParent();
1406 RE1 = BB1->getInstList().rend();
1408 RE2 = BB2->getInstList().rend();
1409 FirstNonPhiInBBEnd = I1;
1416 /// \brief Determine if we can hoist sink a sole store instruction out of a
1417 /// conditional block.
1419 /// We are looking for code like the following:
1421 /// store i32 %add, i32* %arrayidx2
1422 /// ... // No other stores or function calls (we could be calling a memory
1423 /// ... // function).
1424 /// %cmp = icmp ult %x, %y
1425 /// br i1 %cmp, label %EndBB, label %ThenBB
1427 /// store i32 %add5, i32* %arrayidx2
1431 /// We are going to transform this into:
1433 /// store i32 %add, i32* %arrayidx2
1435 /// %cmp = icmp ult %x, %y
1436 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1437 /// store i32 %add.add5, i32* %arrayidx2
1440 /// \return The pointer to the value of the previous store if the store can be
1441 /// hoisted into the predecessor block. 0 otherwise.
1442 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1443 BasicBlock *StoreBB, BasicBlock *EndBB) {
1444 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1448 // Volatile or atomic.
1449 if (!StoreToHoist->isSimple())
1452 Value *StorePtr = StoreToHoist->getPointerOperand();
1454 // Look for a store to the same pointer in BrBB.
1455 unsigned MaxNumInstToLookAt = 10;
1456 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1457 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1458 Instruction *CurI = &*RI;
1460 // Could be calling an instruction that effects memory like free().
1461 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1464 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1465 // Found the previous store make sure it stores to the same location.
1466 if (SI && SI->getPointerOperand() == StorePtr)
1467 // Found the previous store, return its value operand.
1468 return SI->getValueOperand();
1470 return 0; // Unknown store.
1476 /// \brief Speculate a conditional basic block flattening the CFG.
1478 /// Note that this is a very risky transform currently. Speculating
1479 /// instructions like this is most often not desirable. Instead, there is an MI
1480 /// pass which can do it with full awareness of the resource constraints.
1481 /// However, some cases are "obvious" and we should do directly. An example of
1482 /// this is speculating a single, reasonably cheap instruction.
1484 /// There is only one distinct advantage to flattening the CFG at the IR level:
1485 /// it makes very common but simplistic optimizations such as are common in
1486 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1487 /// modeling their effects with easier to reason about SSA value graphs.
1490 /// An illustration of this transform is turning this IR:
1493 /// %cmp = icmp ult %x, %y
1494 /// br i1 %cmp, label %EndBB, label %ThenBB
1496 /// %sub = sub %x, %y
1499 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1506 /// %cmp = icmp ult %x, %y
1507 /// %sub = sub %x, %y
1508 /// %cond = select i1 %cmp, 0, %sub
1512 /// \returns true if the conditional block is removed.
1513 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) {
1514 // Be conservative for now. FP select instruction can often be expensive.
1515 Value *BrCond = BI->getCondition();
1516 if (isa<FCmpInst>(BrCond))
1519 BasicBlock *BB = BI->getParent();
1520 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1522 // If ThenBB is actually on the false edge of the conditional branch, remember
1523 // to swap the select operands later.
1524 bool Invert = false;
1525 if (ThenBB != BI->getSuccessor(0)) {
1526 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1529 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1531 // Keep a count of how many times instructions are used within CondBB when
1532 // they are candidates for sinking into CondBB. Specifically:
1533 // - They are defined in BB, and
1534 // - They have no side effects, and
1535 // - All of their uses are in CondBB.
1536 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1538 unsigned SpeculationCost = 0;
1539 Value *SpeculatedStoreValue = 0;
1540 StoreInst *SpeculatedStore = 0;
1541 for (BasicBlock::iterator BBI = ThenBB->begin(),
1542 BBE = llvm::prior(ThenBB->end());
1543 BBI != BBE; ++BBI) {
1544 Instruction *I = BBI;
1546 if (isa<DbgInfoIntrinsic>(I))
1549 // Only speculatively execution a single instruction (not counting the
1550 // terminator) for now.
1552 if (SpeculationCost > 1)
1555 // Don't hoist the instruction if it's unsafe or expensive.
1556 if (!isSafeToSpeculativelyExecute(I) &&
1557 !(HoistCondStores &&
1558 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1561 if (!SpeculatedStoreValue &&
1562 ComputeSpeculationCost(I) > PHINodeFoldingThreshold)
1565 // Store the store speculation candidate.
1566 if (SpeculatedStoreValue)
1567 SpeculatedStore = cast<StoreInst>(I);
1569 // Do not hoist the instruction if any of its operands are defined but not
1570 // used in BB. The transformation will prevent the operand from
1571 // being sunk into the use block.
1572 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1574 Instruction *OpI = dyn_cast<Instruction>(*i);
1575 if (!OpI || OpI->getParent() != BB ||
1576 OpI->mayHaveSideEffects())
1577 continue; // Not a candidate for sinking.
1579 ++SinkCandidateUseCounts[OpI];
1583 // Consider any sink candidates which are only used in CondBB as costs for
1584 // speculation. Note, while we iterate over a DenseMap here, we are summing
1585 // and so iteration order isn't significant.
1586 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1587 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1589 if (I->first->getNumUses() == I->second) {
1591 if (SpeculationCost > 1)
1595 // Check that the PHI nodes can be converted to selects.
1596 bool HaveRewritablePHIs = false;
1597 for (BasicBlock::iterator I = EndBB->begin();
1598 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1599 Value *OrigV = PN->getIncomingValueForBlock(BB);
1600 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1602 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1603 // Skip PHIs which are trivial.
1607 HaveRewritablePHIs = true;
1608 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1609 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1610 if (!OrigCE && !ThenCE)
1611 continue; // Known safe and cheap.
1613 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1614 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1616 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE) : 0;
1617 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE) : 0;
1618 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1621 // Account for the cost of an unfolded ConstantExpr which could end up
1622 // getting expanded into Instructions.
1623 // FIXME: This doesn't account for how many operations are combined in the
1624 // constant expression.
1626 if (SpeculationCost > 1)
1630 // If there are no PHIs to process, bail early. This helps ensure idempotence
1632 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1635 // If we get here, we can hoist the instruction and if-convert.
1636 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1638 // Insert a select of the value of the speculated store.
1639 if (SpeculatedStoreValue) {
1640 IRBuilder<true, NoFolder> Builder(BI);
1641 Value *TrueV = SpeculatedStore->getValueOperand();
1642 Value *FalseV = SpeculatedStoreValue;
1644 std::swap(TrueV, FalseV);
1645 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1646 "." + FalseV->getName());
1647 SpeculatedStore->setOperand(0, S);
1650 // Hoist the instructions.
1651 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1652 llvm::prior(ThenBB->end()));
1654 // Insert selects and rewrite the PHI operands.
1655 IRBuilder<true, NoFolder> Builder(BI);
1656 for (BasicBlock::iterator I = EndBB->begin();
1657 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1658 unsigned OrigI = PN->getBasicBlockIndex(BB);
1659 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1660 Value *OrigV = PN->getIncomingValue(OrigI);
1661 Value *ThenV = PN->getIncomingValue(ThenI);
1663 // Skip PHIs which are trivial.
1667 // Create a select whose true value is the speculatively executed value and
1668 // false value is the preexisting value. Swap them if the branch
1669 // destinations were inverted.
1670 Value *TrueV = ThenV, *FalseV = OrigV;
1672 std::swap(TrueV, FalseV);
1673 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1674 TrueV->getName() + "." + FalseV->getName());
1675 PN->setIncomingValue(OrigI, V);
1676 PN->setIncomingValue(ThenI, V);
1683 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1684 /// across this block.
1685 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1686 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1689 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1690 if (isa<DbgInfoIntrinsic>(BBI))
1692 if (Size > 10) return false; // Don't clone large BB's.
1695 // We can only support instructions that do not define values that are
1696 // live outside of the current basic block.
1697 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1699 Instruction *U = cast<Instruction>(*UI);
1700 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1703 // Looks ok, continue checking.
1709 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1710 /// that is defined in the same block as the branch and if any PHI entries are
1711 /// constants, thread edges corresponding to that entry to be branches to their
1712 /// ultimate destination.
1713 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) {
1714 BasicBlock *BB = BI->getParent();
1715 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1716 // NOTE: we currently cannot transform this case if the PHI node is used
1717 // outside of the block.
1718 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1721 // Degenerate case of a single entry PHI.
1722 if (PN->getNumIncomingValues() == 1) {
1723 FoldSingleEntryPHINodes(PN->getParent());
1727 // Now we know that this block has multiple preds and two succs.
1728 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1730 // Okay, this is a simple enough basic block. See if any phi values are
1732 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1733 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1734 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1736 // Okay, we now know that all edges from PredBB should be revectored to
1737 // branch to RealDest.
1738 BasicBlock *PredBB = PN->getIncomingBlock(i);
1739 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1741 if (RealDest == BB) continue; // Skip self loops.
1742 // Skip if the predecessor's terminator is an indirect branch.
1743 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1745 // The dest block might have PHI nodes, other predecessors and other
1746 // difficult cases. Instead of being smart about this, just insert a new
1747 // block that jumps to the destination block, effectively splitting
1748 // the edge we are about to create.
1749 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1750 RealDest->getName()+".critedge",
1751 RealDest->getParent(), RealDest);
1752 BranchInst::Create(RealDest, EdgeBB);
1754 // Update PHI nodes.
1755 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1757 // BB may have instructions that are being threaded over. Clone these
1758 // instructions into EdgeBB. We know that there will be no uses of the
1759 // cloned instructions outside of EdgeBB.
1760 BasicBlock::iterator InsertPt = EdgeBB->begin();
1761 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1762 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1763 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1764 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1767 // Clone the instruction.
1768 Instruction *N = BBI->clone();
1769 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1771 // Update operands due to translation.
1772 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1774 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1775 if (PI != TranslateMap.end())
1779 // Check for trivial simplification.
1780 if (Value *V = SimplifyInstruction(N, TD)) {
1781 TranslateMap[BBI] = V;
1782 delete N; // Instruction folded away, don't need actual inst
1784 // Insert the new instruction into its new home.
1785 EdgeBB->getInstList().insert(InsertPt, N);
1786 if (!BBI->use_empty())
1787 TranslateMap[BBI] = N;
1791 // Loop over all of the edges from PredBB to BB, changing them to branch
1792 // to EdgeBB instead.
1793 TerminatorInst *PredBBTI = PredBB->getTerminator();
1794 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1795 if (PredBBTI->getSuccessor(i) == BB) {
1796 BB->removePredecessor(PredBB);
1797 PredBBTI->setSuccessor(i, EdgeBB);
1800 // Recurse, simplifying any other constants.
1801 return FoldCondBranchOnPHI(BI, TD) | true;
1807 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1808 /// PHI node, see if we can eliminate it.
1809 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) {
1810 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1811 // statement", which has a very simple dominance structure. Basically, we
1812 // are trying to find the condition that is being branched on, which
1813 // subsequently causes this merge to happen. We really want control
1814 // dependence information for this check, but simplifycfg can't keep it up
1815 // to date, and this catches most of the cases we care about anyway.
1816 BasicBlock *BB = PN->getParent();
1817 BasicBlock *IfTrue, *IfFalse;
1818 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1820 // Don't bother if the branch will be constant folded trivially.
1821 isa<ConstantInt>(IfCond))
1824 // Okay, we found that we can merge this two-entry phi node into a select.
1825 // Doing so would require us to fold *all* two entry phi nodes in this block.
1826 // At some point this becomes non-profitable (particularly if the target
1827 // doesn't support cmov's). Only do this transformation if there are two or
1828 // fewer PHI nodes in this block.
1829 unsigned NumPhis = 0;
1830 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1834 // Loop over the PHI's seeing if we can promote them all to select
1835 // instructions. While we are at it, keep track of the instructions
1836 // that need to be moved to the dominating block.
1837 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1838 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1839 MaxCostVal1 = PHINodeFoldingThreshold;
1841 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1842 PHINode *PN = cast<PHINode>(II++);
1843 if (Value *V = SimplifyInstruction(PN, TD)) {
1844 PN->replaceAllUsesWith(V);
1845 PN->eraseFromParent();
1849 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1851 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1856 // If we folded the first phi, PN dangles at this point. Refresh it. If
1857 // we ran out of PHIs then we simplified them all.
1858 PN = dyn_cast<PHINode>(BB->begin());
1859 if (PN == 0) return true;
1861 // Don't fold i1 branches on PHIs which contain binary operators. These can
1862 // often be turned into switches and other things.
1863 if (PN->getType()->isIntegerTy(1) &&
1864 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1865 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1866 isa<BinaryOperator>(IfCond)))
1869 // If we all PHI nodes are promotable, check to make sure that all
1870 // instructions in the predecessor blocks can be promoted as well. If
1871 // not, we won't be able to get rid of the control flow, so it's not
1872 // worth promoting to select instructions.
1873 BasicBlock *DomBlock = 0;
1874 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1875 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1876 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1879 DomBlock = *pred_begin(IfBlock1);
1880 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1881 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1882 // This is not an aggressive instruction that we can promote.
1883 // Because of this, we won't be able to get rid of the control
1884 // flow, so the xform is not worth it.
1889 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1892 DomBlock = *pred_begin(IfBlock2);
1893 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1894 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1895 // This is not an aggressive instruction that we can promote.
1896 // Because of this, we won't be able to get rid of the control
1897 // flow, so the xform is not worth it.
1902 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1903 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1905 // If we can still promote the PHI nodes after this gauntlet of tests,
1906 // do all of the PHI's now.
1907 Instruction *InsertPt = DomBlock->getTerminator();
1908 IRBuilder<true, NoFolder> Builder(InsertPt);
1910 // Move all 'aggressive' instructions, which are defined in the
1911 // conditional parts of the if's up to the dominating block.
1913 DomBlock->getInstList().splice(InsertPt,
1914 IfBlock1->getInstList(), IfBlock1->begin(),
1915 IfBlock1->getTerminator());
1917 DomBlock->getInstList().splice(InsertPt,
1918 IfBlock2->getInstList(), IfBlock2->begin(),
1919 IfBlock2->getTerminator());
1921 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1922 // Change the PHI node into a select instruction.
1923 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1924 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1927 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1928 PN->replaceAllUsesWith(NV);
1930 PN->eraseFromParent();
1933 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1934 // has been flattened. Change DomBlock to jump directly to our new block to
1935 // avoid other simplifycfg's kicking in on the diamond.
1936 TerminatorInst *OldTI = DomBlock->getTerminator();
1937 Builder.SetInsertPoint(OldTI);
1938 Builder.CreateBr(BB);
1939 OldTI->eraseFromParent();
1943 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1944 /// to two returning blocks, try to merge them together into one return,
1945 /// introducing a select if the return values disagree.
1946 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1947 IRBuilder<> &Builder) {
1948 assert(BI->isConditional() && "Must be a conditional branch");
1949 BasicBlock *TrueSucc = BI->getSuccessor(0);
1950 BasicBlock *FalseSucc = BI->getSuccessor(1);
1951 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1952 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1954 // Check to ensure both blocks are empty (just a return) or optionally empty
1955 // with PHI nodes. If there are other instructions, merging would cause extra
1956 // computation on one path or the other.
1957 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1959 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1962 Builder.SetInsertPoint(BI);
1963 // Okay, we found a branch that is going to two return nodes. If
1964 // there is no return value for this function, just change the
1965 // branch into a return.
1966 if (FalseRet->getNumOperands() == 0) {
1967 TrueSucc->removePredecessor(BI->getParent());
1968 FalseSucc->removePredecessor(BI->getParent());
1969 Builder.CreateRetVoid();
1970 EraseTerminatorInstAndDCECond(BI);
1974 // Otherwise, figure out what the true and false return values are
1975 // so we can insert a new select instruction.
1976 Value *TrueValue = TrueRet->getReturnValue();
1977 Value *FalseValue = FalseRet->getReturnValue();
1979 // Unwrap any PHI nodes in the return blocks.
1980 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1981 if (TVPN->getParent() == TrueSucc)
1982 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1983 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1984 if (FVPN->getParent() == FalseSucc)
1985 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1987 // In order for this transformation to be safe, we must be able to
1988 // unconditionally execute both operands to the return. This is
1989 // normally the case, but we could have a potentially-trapping
1990 // constant expression that prevents this transformation from being
1992 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1995 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1999 // Okay, we collected all the mapped values and checked them for sanity, and
2000 // defined to really do this transformation. First, update the CFG.
2001 TrueSucc->removePredecessor(BI->getParent());
2002 FalseSucc->removePredecessor(BI->getParent());
2004 // Insert select instructions where needed.
2005 Value *BrCond = BI->getCondition();
2007 // Insert a select if the results differ.
2008 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
2009 } else if (isa<UndefValue>(TrueValue)) {
2010 TrueValue = FalseValue;
2012 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
2013 FalseValue, "retval");
2017 Value *RI = !TrueValue ?
2018 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
2022 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
2023 << "\n " << *BI << "NewRet = " << *RI
2024 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
2026 EraseTerminatorInstAndDCECond(BI);
2031 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
2032 /// probabilities of the branch taking each edge. Fills in the two APInt
2033 /// parameters and return true, or returns false if no or invalid metadata was
2035 static bool ExtractBranchMetadata(BranchInst *BI,
2036 uint64_t &ProbTrue, uint64_t &ProbFalse) {
2037 assert(BI->isConditional() &&
2038 "Looking for probabilities on unconditional branch?");
2039 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
2040 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
2041 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
2042 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
2043 if (!CITrue || !CIFalse) return false;
2044 ProbTrue = CITrue->getValue().getZExtValue();
2045 ProbFalse = CIFalse->getValue().getZExtValue();
2049 /// checkCSEInPredecessor - Return true if the given instruction is available
2050 /// in its predecessor block. If yes, the instruction will be removed.
2052 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2053 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2055 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
2056 Instruction *PBI = &*I;
2057 // Check whether Inst and PBI generate the same value.
2058 if (Inst->isIdenticalTo(PBI)) {
2059 Inst->replaceAllUsesWith(PBI);
2060 Inst->eraseFromParent();
2067 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
2068 /// predecessor branches to us and one of our successors, fold the block into
2069 /// the predecessor and use logical operations to pick the right destination.
2070 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
2071 BasicBlock *BB = BI->getParent();
2073 Instruction *Cond = 0;
2074 if (BI->isConditional())
2075 Cond = dyn_cast<Instruction>(BI->getCondition());
2077 // For unconditional branch, check for a simple CFG pattern, where
2078 // BB has a single predecessor and BB's successor is also its predecessor's
2079 // successor. If such pattern exisits, check for CSE between BB and its
2081 if (BasicBlock *PB = BB->getSinglePredecessor())
2082 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2083 if (PBI->isConditional() &&
2084 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2085 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2086 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2088 Instruction *Curr = I++;
2089 if (isa<CmpInst>(Curr)) {
2093 // Quit if we can't remove this instruction.
2094 if (!checkCSEInPredecessor(Curr, PB))
2103 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2104 Cond->getParent() != BB || !Cond->hasOneUse())
2107 // Only allow this if the condition is a simple instruction that can be
2108 // executed unconditionally. It must be in the same block as the branch, and
2109 // must be at the front of the block.
2110 BasicBlock::iterator FrontIt = BB->front();
2112 // Ignore dbg intrinsics.
2113 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2115 // Allow a single instruction to be hoisted in addition to the compare
2116 // that feeds the branch. We later ensure that any values that _it_ uses
2117 // were also live in the predecessor, so that we don't unnecessarily create
2118 // register pressure or inhibit out-of-order execution.
2119 Instruction *BonusInst = 0;
2120 if (&*FrontIt != Cond &&
2121 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
2122 isSafeToSpeculativelyExecute(FrontIt)) {
2123 BonusInst = &*FrontIt;
2126 // Ignore dbg intrinsics.
2127 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2130 // Only a single bonus inst is allowed.
2131 if (&*FrontIt != Cond)
2134 // Make sure the instruction after the condition is the cond branch.
2135 BasicBlock::iterator CondIt = Cond; ++CondIt;
2137 // Ingore dbg intrinsics.
2138 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2143 // Cond is known to be a compare or binary operator. Check to make sure that
2144 // neither operand is a potentially-trapping constant expression.
2145 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2148 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2152 // Finally, don't infinitely unroll conditional loops.
2153 BasicBlock *TrueDest = BI->getSuccessor(0);
2154 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
2155 if (TrueDest == BB || FalseDest == BB)
2158 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2159 BasicBlock *PredBlock = *PI;
2160 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2162 // Check that we have two conditional branches. If there is a PHI node in
2163 // the common successor, verify that the same value flows in from both
2165 SmallVector<PHINode*, 4> PHIs;
2166 if (PBI == 0 || PBI->isUnconditional() ||
2167 (BI->isConditional() &&
2168 !SafeToMergeTerminators(BI, PBI)) ||
2169 (!BI->isConditional() &&
2170 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2173 // Determine if the two branches share a common destination.
2174 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2175 bool InvertPredCond = false;
2177 if (BI->isConditional()) {
2178 if (PBI->getSuccessor(0) == TrueDest)
2179 Opc = Instruction::Or;
2180 else if (PBI->getSuccessor(1) == FalseDest)
2181 Opc = Instruction::And;
2182 else if (PBI->getSuccessor(0) == FalseDest)
2183 Opc = Instruction::And, InvertPredCond = true;
2184 else if (PBI->getSuccessor(1) == TrueDest)
2185 Opc = Instruction::Or, InvertPredCond = true;
2189 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2193 // Ensure that any values used in the bonus instruction are also used
2194 // by the terminator of the predecessor. This means that those values
2195 // must already have been resolved, so we won't be inhibiting the
2196 // out-of-order core by speculating them earlier.
2198 // Collect the values used by the bonus inst
2199 SmallPtrSet<Value*, 4> UsedValues;
2200 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2201 OE = BonusInst->op_end(); OI != OE; ++OI) {
2203 if (!isa<Constant>(V))
2204 UsedValues.insert(V);
2207 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2208 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2210 // Walk up to four levels back up the use-def chain of the predecessor's
2211 // terminator to see if all those values were used. The choice of four
2212 // levels is arbitrary, to provide a compile-time-cost bound.
2213 while (!Worklist.empty()) {
2214 std::pair<Value*, unsigned> Pair = Worklist.back();
2215 Worklist.pop_back();
2217 if (Pair.second >= 4) continue;
2218 UsedValues.erase(Pair.first);
2219 if (UsedValues.empty()) break;
2221 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2222 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2224 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2228 if (!UsedValues.empty()) return false;
2231 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2232 IRBuilder<> Builder(PBI);
2234 // If we need to invert the condition in the pred block to match, do so now.
2235 if (InvertPredCond) {
2236 Value *NewCond = PBI->getCondition();
2238 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2239 CmpInst *CI = cast<CmpInst>(NewCond);
2240 CI->setPredicate(CI->getInversePredicate());
2242 NewCond = Builder.CreateNot(NewCond,
2243 PBI->getCondition()->getName()+".not");
2246 PBI->setCondition(NewCond);
2247 PBI->swapSuccessors();
2250 // If we have a bonus inst, clone it into the predecessor block.
2251 Instruction *NewBonus = 0;
2253 NewBonus = BonusInst->clone();
2254 PredBlock->getInstList().insert(PBI, NewBonus);
2255 NewBonus->takeName(BonusInst);
2256 BonusInst->setName(BonusInst->getName()+".old");
2259 // Clone Cond into the predecessor basic block, and or/and the
2260 // two conditions together.
2261 Instruction *New = Cond->clone();
2262 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2263 PredBlock->getInstList().insert(PBI, New);
2264 New->takeName(Cond);
2265 Cond->setName(New->getName()+".old");
2267 if (BI->isConditional()) {
2268 Instruction *NewCond =
2269 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2271 PBI->setCondition(NewCond);
2273 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2274 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2276 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2278 SmallVector<uint64_t, 8> NewWeights;
2280 if (PBI->getSuccessor(0) == BB) {
2281 if (PredHasWeights && SuccHasWeights) {
2282 // PBI: br i1 %x, BB, FalseDest
2283 // BI: br i1 %y, TrueDest, FalseDest
2284 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2285 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2286 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2287 // TrueWeight for PBI * FalseWeight for BI.
2288 // We assume that total weights of a BranchInst can fit into 32 bits.
2289 // Therefore, we will not have overflow using 64-bit arithmetic.
2290 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2291 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2293 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2294 PBI->setSuccessor(0, TrueDest);
2296 if (PBI->getSuccessor(1) == BB) {
2297 if (PredHasWeights && SuccHasWeights) {
2298 // PBI: br i1 %x, TrueDest, BB
2299 // BI: br i1 %y, TrueDest, FalseDest
2300 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2301 // FalseWeight for PBI * TrueWeight for BI.
2302 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2303 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2304 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2305 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2307 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2308 PBI->setSuccessor(1, FalseDest);
2310 if (NewWeights.size() == 2) {
2311 // Halve the weights if any of them cannot fit in an uint32_t
2312 FitWeights(NewWeights);
2314 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2315 PBI->setMetadata(LLVMContext::MD_prof,
2316 MDBuilder(BI->getContext()).
2317 createBranchWeights(MDWeights));
2319 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2321 // Update PHI nodes in the common successors.
2322 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2323 ConstantInt *PBI_C = cast<ConstantInt>(
2324 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2325 assert(PBI_C->getType()->isIntegerTy(1));
2326 Instruction *MergedCond = 0;
2327 if (PBI->getSuccessor(0) == TrueDest) {
2328 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2329 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2330 // is false: !PBI_Cond and BI_Value
2331 Instruction *NotCond =
2332 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2335 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2340 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2341 PBI->getCondition(), MergedCond,
2344 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2345 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2346 // is false: PBI_Cond and BI_Value
2348 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2349 PBI->getCondition(), New,
2351 if (PBI_C->isOne()) {
2352 Instruction *NotCond =
2353 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2356 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2357 NotCond, MergedCond,
2362 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2365 // Change PBI from Conditional to Unconditional.
2366 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2367 EraseTerminatorInstAndDCECond(PBI);
2371 // TODO: If BB is reachable from all paths through PredBlock, then we
2372 // could replace PBI's branch probabilities with BI's.
2374 // Copy any debug value intrinsics into the end of PredBlock.
2375 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2376 if (isa<DbgInfoIntrinsic>(*I))
2377 I->clone()->insertBefore(PBI);
2384 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2385 /// predecessor of another block, this function tries to simplify it. We know
2386 /// that PBI and BI are both conditional branches, and BI is in one of the
2387 /// successor blocks of PBI - PBI branches to BI.
2388 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2389 assert(PBI->isConditional() && BI->isConditional());
2390 BasicBlock *BB = BI->getParent();
2392 // If this block ends with a branch instruction, and if there is a
2393 // predecessor that ends on a branch of the same condition, make
2394 // this conditional branch redundant.
2395 if (PBI->getCondition() == BI->getCondition() &&
2396 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2397 // Okay, the outcome of this conditional branch is statically
2398 // knowable. If this block had a single pred, handle specially.
2399 if (BB->getSinglePredecessor()) {
2400 // Turn this into a branch on constant.
2401 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2402 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2404 return true; // Nuke the branch on constant.
2407 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2408 // in the constant and simplify the block result. Subsequent passes of
2409 // simplifycfg will thread the block.
2410 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2411 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2412 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2413 std::distance(PB, PE),
2414 BI->getCondition()->getName() + ".pr",
2416 // Okay, we're going to insert the PHI node. Since PBI is not the only
2417 // predecessor, compute the PHI'd conditional value for all of the preds.
2418 // Any predecessor where the condition is not computable we keep symbolic.
2419 for (pred_iterator PI = PB; PI != PE; ++PI) {
2420 BasicBlock *P = *PI;
2421 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2422 PBI != BI && PBI->isConditional() &&
2423 PBI->getCondition() == BI->getCondition() &&
2424 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2425 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2426 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2429 NewPN->addIncoming(BI->getCondition(), P);
2433 BI->setCondition(NewPN);
2438 // If this is a conditional branch in an empty block, and if any
2439 // predecessors is a conditional branch to one of our destinations,
2440 // fold the conditions into logical ops and one cond br.
2441 BasicBlock::iterator BBI = BB->begin();
2442 // Ignore dbg intrinsics.
2443 while (isa<DbgInfoIntrinsic>(BBI))
2449 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2454 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2456 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2457 PBIOp = 0, BIOp = 1;
2458 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2459 PBIOp = 1, BIOp = 0;
2460 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2465 // Check to make sure that the other destination of this branch
2466 // isn't BB itself. If so, this is an infinite loop that will
2467 // keep getting unwound.
2468 if (PBI->getSuccessor(PBIOp) == BB)
2471 // Do not perform this transformation if it would require
2472 // insertion of a large number of select instructions. For targets
2473 // without predication/cmovs, this is a big pessimization.
2474 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2476 unsigned NumPhis = 0;
2477 for (BasicBlock::iterator II = CommonDest->begin();
2478 isa<PHINode>(II); ++II, ++NumPhis)
2479 if (NumPhis > 2) // Disable this xform.
2482 // Finally, if everything is ok, fold the branches to logical ops.
2483 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2485 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2486 << "AND: " << *BI->getParent());
2489 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2490 // branch in it, where one edge (OtherDest) goes back to itself but the other
2491 // exits. We don't *know* that the program avoids the infinite loop
2492 // (even though that seems likely). If we do this xform naively, we'll end up
2493 // recursively unpeeling the loop. Since we know that (after the xform is
2494 // done) that the block *is* infinite if reached, we just make it an obviously
2495 // infinite loop with no cond branch.
2496 if (OtherDest == BB) {
2497 // Insert it at the end of the function, because it's either code,
2498 // or it won't matter if it's hot. :)
2499 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2500 "infloop", BB->getParent());
2501 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2502 OtherDest = InfLoopBlock;
2505 DEBUG(dbgs() << *PBI->getParent()->getParent());
2507 // BI may have other predecessors. Because of this, we leave
2508 // it alone, but modify PBI.
2510 // Make sure we get to CommonDest on True&True directions.
2511 Value *PBICond = PBI->getCondition();
2512 IRBuilder<true, NoFolder> Builder(PBI);
2514 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2516 Value *BICond = BI->getCondition();
2518 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2520 // Merge the conditions.
2521 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2523 // Modify PBI to branch on the new condition to the new dests.
2524 PBI->setCondition(Cond);
2525 PBI->setSuccessor(0, CommonDest);
2526 PBI->setSuccessor(1, OtherDest);
2528 // Update branch weight for PBI.
2529 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2530 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2532 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2534 if (PredHasWeights && SuccHasWeights) {
2535 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2536 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2537 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2538 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2539 // The weight to CommonDest should be PredCommon * SuccTotal +
2540 // PredOther * SuccCommon.
2541 // The weight to OtherDest should be PredOther * SuccOther.
2542 SmallVector<uint64_t, 2> NewWeights;
2543 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2544 PredOther * SuccCommon);
2545 NewWeights.push_back(PredOther * SuccOther);
2546 // Halve the weights if any of them cannot fit in an uint32_t
2547 FitWeights(NewWeights);
2549 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2550 PBI->setMetadata(LLVMContext::MD_prof,
2551 MDBuilder(BI->getContext()).
2552 createBranchWeights(MDWeights));
2555 // OtherDest may have phi nodes. If so, add an entry from PBI's
2556 // block that are identical to the entries for BI's block.
2557 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2559 // We know that the CommonDest already had an edge from PBI to
2560 // it. If it has PHIs though, the PHIs may have different
2561 // entries for BB and PBI's BB. If so, insert a select to make
2564 for (BasicBlock::iterator II = CommonDest->begin();
2565 (PN = dyn_cast<PHINode>(II)); ++II) {
2566 Value *BIV = PN->getIncomingValueForBlock(BB);
2567 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2568 Value *PBIV = PN->getIncomingValue(PBBIdx);
2570 // Insert a select in PBI to pick the right value.
2571 Value *NV = cast<SelectInst>
2572 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2573 PN->setIncomingValue(PBBIdx, NV);
2577 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2578 DEBUG(dbgs() << *PBI->getParent()->getParent());
2580 // This basic block is probably dead. We know it has at least
2581 // one fewer predecessor.
2585 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2586 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2587 // Takes care of updating the successors and removing the old terminator.
2588 // Also makes sure not to introduce new successors by assuming that edges to
2589 // non-successor TrueBBs and FalseBBs aren't reachable.
2590 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2591 BasicBlock *TrueBB, BasicBlock *FalseBB,
2592 uint32_t TrueWeight,
2593 uint32_t FalseWeight){
2594 // Remove any superfluous successor edges from the CFG.
2595 // First, figure out which successors to preserve.
2596 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2598 BasicBlock *KeepEdge1 = TrueBB;
2599 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2601 // Then remove the rest.
2602 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2603 BasicBlock *Succ = OldTerm->getSuccessor(I);
2604 // Make sure only to keep exactly one copy of each edge.
2605 if (Succ == KeepEdge1)
2607 else if (Succ == KeepEdge2)
2610 Succ->removePredecessor(OldTerm->getParent());
2613 IRBuilder<> Builder(OldTerm);
2614 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2616 // Insert an appropriate new terminator.
2617 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2618 if (TrueBB == FalseBB)
2619 // We were only looking for one successor, and it was present.
2620 // Create an unconditional branch to it.
2621 Builder.CreateBr(TrueBB);
2623 // We found both of the successors we were looking for.
2624 // Create a conditional branch sharing the condition of the select.
2625 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2626 if (TrueWeight != FalseWeight)
2627 NewBI->setMetadata(LLVMContext::MD_prof,
2628 MDBuilder(OldTerm->getContext()).
2629 createBranchWeights(TrueWeight, FalseWeight));
2631 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2632 // Neither of the selected blocks were successors, so this
2633 // terminator must be unreachable.
2634 new UnreachableInst(OldTerm->getContext(), OldTerm);
2636 // One of the selected values was a successor, but the other wasn't.
2637 // Insert an unconditional branch to the one that was found;
2638 // the edge to the one that wasn't must be unreachable.
2640 // Only TrueBB was found.
2641 Builder.CreateBr(TrueBB);
2643 // Only FalseBB was found.
2644 Builder.CreateBr(FalseBB);
2647 EraseTerminatorInstAndDCECond(OldTerm);
2651 // SimplifySwitchOnSelect - Replaces
2652 // (switch (select cond, X, Y)) on constant X, Y
2653 // with a branch - conditional if X and Y lead to distinct BBs,
2654 // unconditional otherwise.
2655 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2656 // Check for constant integer values in the select.
2657 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2658 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2659 if (!TrueVal || !FalseVal)
2662 // Find the relevant condition and destinations.
2663 Value *Condition = Select->getCondition();
2664 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2665 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2667 // Get weight for TrueBB and FalseBB.
2668 uint32_t TrueWeight = 0, FalseWeight = 0;
2669 SmallVector<uint64_t, 8> Weights;
2670 bool HasWeights = HasBranchWeights(SI);
2672 GetBranchWeights(SI, Weights);
2673 if (Weights.size() == 1 + SI->getNumCases()) {
2674 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2675 getSuccessorIndex()];
2676 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2677 getSuccessorIndex()];
2681 // Perform the actual simplification.
2682 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2683 TrueWeight, FalseWeight);
2686 // SimplifyIndirectBrOnSelect - Replaces
2687 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2688 // blockaddress(@fn, BlockB)))
2690 // (br cond, BlockA, BlockB).
2691 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2692 // Check that both operands of the select are block addresses.
2693 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2694 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2698 // Extract the actual blocks.
2699 BasicBlock *TrueBB = TBA->getBasicBlock();
2700 BasicBlock *FalseBB = FBA->getBasicBlock();
2702 // Perform the actual simplification.
2703 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2707 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2708 /// instruction (a seteq/setne with a constant) as the only instruction in a
2709 /// block that ends with an uncond branch. We are looking for a very specific
2710 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2711 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2712 /// default value goes to an uncond block with a seteq in it, we get something
2715 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2717 /// %tmp = icmp eq i8 %A, 92
2720 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2722 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2723 /// the PHI, merging the third icmp into the switch.
2724 static bool TryToSimplifyUncondBranchWithICmpInIt(
2725 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2726 const DataLayout *TD) {
2727 BasicBlock *BB = ICI->getParent();
2729 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2731 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2733 Value *V = ICI->getOperand(0);
2734 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2736 // The pattern we're looking for is where our only predecessor is a switch on
2737 // 'V' and this block is the default case for the switch. In this case we can
2738 // fold the compared value into the switch to simplify things.
2739 BasicBlock *Pred = BB->getSinglePredecessor();
2740 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2742 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2743 if (SI->getCondition() != V)
2746 // If BB is reachable on a non-default case, then we simply know the value of
2747 // V in this block. Substitute it and constant fold the icmp instruction
2749 if (SI->getDefaultDest() != BB) {
2750 ConstantInt *VVal = SI->findCaseDest(BB);
2751 assert(VVal && "Should have a unique destination value");
2752 ICI->setOperand(0, VVal);
2754 if (Value *V = SimplifyInstruction(ICI, TD)) {
2755 ICI->replaceAllUsesWith(V);
2756 ICI->eraseFromParent();
2758 // BB is now empty, so it is likely to simplify away.
2759 return SimplifyCFG(BB, TTI, TD) | true;
2762 // Ok, the block is reachable from the default dest. If the constant we're
2763 // comparing exists in one of the other edges, then we can constant fold ICI
2765 if (SI->findCaseValue(Cst) != SI->case_default()) {
2767 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2768 V = ConstantInt::getFalse(BB->getContext());
2770 V = ConstantInt::getTrue(BB->getContext());
2772 ICI->replaceAllUsesWith(V);
2773 ICI->eraseFromParent();
2774 // BB is now empty, so it is likely to simplify away.
2775 return SimplifyCFG(BB, TTI, TD) | true;
2778 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2780 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2781 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2782 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2783 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2786 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2788 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2789 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2791 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2792 std::swap(DefaultCst, NewCst);
2794 // Replace ICI (which is used by the PHI for the default value) with true or
2795 // false depending on if it is EQ or NE.
2796 ICI->replaceAllUsesWith(DefaultCst);
2797 ICI->eraseFromParent();
2799 // Okay, the switch goes to this block on a default value. Add an edge from
2800 // the switch to the merge point on the compared value.
2801 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2802 BB->getParent(), BB);
2803 SmallVector<uint64_t, 8> Weights;
2804 bool HasWeights = HasBranchWeights(SI);
2806 GetBranchWeights(SI, Weights);
2807 if (Weights.size() == 1 + SI->getNumCases()) {
2808 // Split weight for default case to case for "Cst".
2809 Weights[0] = (Weights[0]+1) >> 1;
2810 Weights.push_back(Weights[0]);
2812 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2813 SI->setMetadata(LLVMContext::MD_prof,
2814 MDBuilder(SI->getContext()).
2815 createBranchWeights(MDWeights));
2818 SI->addCase(Cst, NewBB);
2820 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2821 Builder.SetInsertPoint(NewBB);
2822 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2823 Builder.CreateBr(SuccBlock);
2824 PHIUse->addIncoming(NewCst, NewBB);
2828 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2829 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2830 /// fold it into a switch instruction if so.
2831 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD,
2832 IRBuilder<> &Builder) {
2833 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2834 if (Cond == 0) return false;
2837 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2838 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2839 // 'setne's and'ed together, collect them.
2841 std::vector<ConstantInt*> Values;
2842 bool TrueWhenEqual = true;
2843 Value *ExtraCase = 0;
2844 unsigned UsedICmps = 0;
2846 if (Cond->getOpcode() == Instruction::Or) {
2847 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2849 } else if (Cond->getOpcode() == Instruction::And) {
2850 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2852 TrueWhenEqual = false;
2855 // If we didn't have a multiply compared value, fail.
2856 if (CompVal == 0) return false;
2858 // Avoid turning single icmps into a switch.
2862 // There might be duplicate constants in the list, which the switch
2863 // instruction can't handle, remove them now.
2864 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2865 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2867 // If Extra was used, we require at least two switch values to do the
2868 // transformation. A switch with one value is just an cond branch.
2869 if (ExtraCase && Values.size() < 2) return false;
2871 // TODO: Preserve branch weight metadata, similarly to how
2872 // FoldValueComparisonIntoPredecessors preserves it.
2874 // Figure out which block is which destination.
2875 BasicBlock *DefaultBB = BI->getSuccessor(1);
2876 BasicBlock *EdgeBB = BI->getSuccessor(0);
2877 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2879 BasicBlock *BB = BI->getParent();
2881 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2882 << " cases into SWITCH. BB is:\n" << *BB);
2884 // If there are any extra values that couldn't be folded into the switch
2885 // then we evaluate them with an explicit branch first. Split the block
2886 // right before the condbr to handle it.
2888 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2889 // Remove the uncond branch added to the old block.
2890 TerminatorInst *OldTI = BB->getTerminator();
2891 Builder.SetInsertPoint(OldTI);
2894 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2896 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2898 OldTI->eraseFromParent();
2900 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2901 // for the edge we just added.
2902 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2904 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2905 << "\nEXTRABB = " << *BB);
2909 Builder.SetInsertPoint(BI);
2910 // Convert pointer to int before we switch.
2911 if (CompVal->getType()->isPointerTy()) {
2912 assert(TD && "Cannot switch on pointer without DataLayout");
2913 CompVal = Builder.CreatePtrToInt(CompVal,
2914 TD->getIntPtrType(CompVal->getContext()),
2918 // Create the new switch instruction now.
2919 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2921 // Add all of the 'cases' to the switch instruction.
2922 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2923 New->addCase(Values[i], EdgeBB);
2925 // We added edges from PI to the EdgeBB. As such, if there were any
2926 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2927 // the number of edges added.
2928 for (BasicBlock::iterator BBI = EdgeBB->begin();
2929 isa<PHINode>(BBI); ++BBI) {
2930 PHINode *PN = cast<PHINode>(BBI);
2931 Value *InVal = PN->getIncomingValueForBlock(BB);
2932 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2933 PN->addIncoming(InVal, BB);
2936 // Erase the old branch instruction.
2937 EraseTerminatorInstAndDCECond(BI);
2939 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2943 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2944 // If this is a trivial landing pad that just continues unwinding the caught
2945 // exception then zap the landing pad, turning its invokes into calls.
2946 BasicBlock *BB = RI->getParent();
2947 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2948 if (RI->getValue() != LPInst)
2949 // Not a landing pad, or the resume is not unwinding the exception that
2950 // caused control to branch here.
2953 // Check that there are no other instructions except for debug intrinsics.
2954 BasicBlock::iterator I = LPInst, E = RI;
2956 if (!isa<DbgInfoIntrinsic>(I))
2959 // Turn all invokes that unwind here into calls and delete the basic block.
2960 bool InvokeRequiresTableEntry = false;
2961 bool Changed = false;
2962 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2963 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2965 if (II->hasFnAttr(Attribute::UWTable)) {
2966 // Don't remove an `invoke' instruction if the ABI requires an entry into
2968 InvokeRequiresTableEntry = true;
2972 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2974 // Insert a call instruction before the invoke.
2975 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2977 Call->setCallingConv(II->getCallingConv());
2978 Call->setAttributes(II->getAttributes());
2979 Call->setDebugLoc(II->getDebugLoc());
2981 // Anything that used the value produced by the invoke instruction now uses
2982 // the value produced by the call instruction. Note that we do this even
2983 // for void functions and calls with no uses so that the callgraph edge is
2985 II->replaceAllUsesWith(Call);
2986 BB->removePredecessor(II->getParent());
2988 // Insert a branch to the normal destination right before the invoke.
2989 BranchInst::Create(II->getNormalDest(), II);
2991 // Finally, delete the invoke instruction!
2992 II->eraseFromParent();
2996 if (!InvokeRequiresTableEntry)
2997 // The landingpad is now unreachable. Zap it.
2998 BB->eraseFromParent();
3003 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
3004 BasicBlock *BB = RI->getParent();
3005 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
3007 // Find predecessors that end with branches.
3008 SmallVector<BasicBlock*, 8> UncondBranchPreds;
3009 SmallVector<BranchInst*, 8> CondBranchPreds;
3010 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
3011 BasicBlock *P = *PI;
3012 TerminatorInst *PTI = P->getTerminator();
3013 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
3014 if (BI->isUnconditional())
3015 UncondBranchPreds.push_back(P);
3017 CondBranchPreds.push_back(BI);
3021 // If we found some, do the transformation!
3022 if (!UncondBranchPreds.empty() && DupRet) {
3023 while (!UncondBranchPreds.empty()) {
3024 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
3025 DEBUG(dbgs() << "FOLDING: " << *BB
3026 << "INTO UNCOND BRANCH PRED: " << *Pred);
3027 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
3030 // If we eliminated all predecessors of the block, delete the block now.
3031 if (pred_begin(BB) == pred_end(BB))
3032 // We know there are no successors, so just nuke the block.
3033 BB->eraseFromParent();
3038 // Check out all of the conditional branches going to this return
3039 // instruction. If any of them just select between returns, change the
3040 // branch itself into a select/return pair.
3041 while (!CondBranchPreds.empty()) {
3042 BranchInst *BI = CondBranchPreds.pop_back_val();
3044 // Check to see if the non-BB successor is also a return block.
3045 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
3046 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
3047 SimplifyCondBranchToTwoReturns(BI, Builder))
3053 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
3054 BasicBlock *BB = UI->getParent();
3056 bool Changed = false;
3058 // If there are any instructions immediately before the unreachable that can
3059 // be removed, do so.
3060 while (UI != BB->begin()) {
3061 BasicBlock::iterator BBI = UI;
3063 // Do not delete instructions that can have side effects which might cause
3064 // the unreachable to not be reachable; specifically, calls and volatile
3065 // operations may have this effect.
3066 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
3068 if (BBI->mayHaveSideEffects()) {
3069 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
3070 if (SI->isVolatile())
3072 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3073 if (LI->isVolatile())
3075 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3076 if (RMWI->isVolatile())
3078 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3079 if (CXI->isVolatile())
3081 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3082 !isa<LandingPadInst>(BBI)) {
3085 // Note that deleting LandingPad's here is in fact okay, although it
3086 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3087 // all the predecessors of this block will be the unwind edges of Invokes,
3088 // and we can therefore guarantee this block will be erased.
3091 // Delete this instruction (any uses are guaranteed to be dead)
3092 if (!BBI->use_empty())
3093 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3094 BBI->eraseFromParent();
3098 // If the unreachable instruction is the first in the block, take a gander
3099 // at all of the predecessors of this instruction, and simplify them.
3100 if (&BB->front() != UI) return Changed;
3102 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3103 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3104 TerminatorInst *TI = Preds[i]->getTerminator();
3105 IRBuilder<> Builder(TI);
3106 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3107 if (BI->isUnconditional()) {
3108 if (BI->getSuccessor(0) == BB) {
3109 new UnreachableInst(TI->getContext(), TI);
3110 TI->eraseFromParent();
3114 if (BI->getSuccessor(0) == BB) {
3115 Builder.CreateBr(BI->getSuccessor(1));
3116 EraseTerminatorInstAndDCECond(BI);
3117 } else if (BI->getSuccessor(1) == BB) {
3118 Builder.CreateBr(BI->getSuccessor(0));
3119 EraseTerminatorInstAndDCECond(BI);
3123 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3124 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3126 if (i.getCaseSuccessor() == BB) {
3127 BB->removePredecessor(SI->getParent());
3132 // If the default value is unreachable, figure out the most popular
3133 // destination and make it the default.
3134 if (SI->getDefaultDest() == BB) {
3135 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3136 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3138 std::pair<unsigned, unsigned> &entry =
3139 Popularity[i.getCaseSuccessor()];
3140 if (entry.first == 0) {
3142 entry.second = i.getCaseIndex();
3148 // Find the most popular block.
3149 unsigned MaxPop = 0;
3150 unsigned MaxIndex = 0;
3151 BasicBlock *MaxBlock = 0;
3152 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3153 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3154 if (I->second.first > MaxPop ||
3155 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3156 MaxPop = I->second.first;
3157 MaxIndex = I->second.second;
3158 MaxBlock = I->first;
3162 // Make this the new default, allowing us to delete any explicit
3164 SI->setDefaultDest(MaxBlock);
3167 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3169 if (isa<PHINode>(MaxBlock->begin()))
3170 for (unsigned i = 0; i != MaxPop-1; ++i)
3171 MaxBlock->removePredecessor(SI->getParent());
3173 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3175 if (i.getCaseSuccessor() == MaxBlock) {
3181 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3182 if (II->getUnwindDest() == BB) {
3183 // Convert the invoke to a call instruction. This would be a good
3184 // place to note that the call does not throw though.
3185 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3186 II->removeFromParent(); // Take out of symbol table
3188 // Insert the call now...
3189 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3190 Builder.SetInsertPoint(BI);
3191 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3192 Args, II->getName());
3193 CI->setCallingConv(II->getCallingConv());
3194 CI->setAttributes(II->getAttributes());
3195 // If the invoke produced a value, the call does now instead.
3196 II->replaceAllUsesWith(CI);
3203 // If this block is now dead, remove it.
3204 if (pred_begin(BB) == pred_end(BB) &&
3205 BB != &BB->getParent()->getEntryBlock()) {
3206 // We know there are no successors, so just nuke the block.
3207 BB->eraseFromParent();
3214 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3215 /// integer range comparison into a sub, an icmp and a branch.
3216 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3217 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3219 // Make sure all cases point to the same destination and gather the values.
3220 SmallVector<ConstantInt *, 16> Cases;
3221 SwitchInst::CaseIt I = SI->case_begin();
3222 Cases.push_back(I.getCaseValue());
3223 SwitchInst::CaseIt PrevI = I++;
3224 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3225 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3227 Cases.push_back(I.getCaseValue());
3229 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3231 // Sort the case values, then check if they form a range we can transform.
3232 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3233 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3234 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3238 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3239 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3241 Value *Sub = SI->getCondition();
3242 if (!Offset->isNullValue())
3243 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3245 // If NumCases overflowed, then all possible values jump to the successor.
3246 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3247 Cmp = ConstantInt::getTrue(SI->getContext());
3249 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3250 BranchInst *NewBI = Builder.CreateCondBr(
3251 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3253 // Update weight for the newly-created conditional branch.
3254 SmallVector<uint64_t, 8> Weights;
3255 bool HasWeights = HasBranchWeights(SI);
3257 GetBranchWeights(SI, Weights);
3258 if (Weights.size() == 1 + SI->getNumCases()) {
3259 // Combine all weights for the cases to be the true weight of NewBI.
3260 // We assume that the sum of all weights for a Terminator can fit into 32
3262 uint32_t NewTrueWeight = 0;
3263 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3264 NewTrueWeight += (uint32_t)Weights[I];
3265 NewBI->setMetadata(LLVMContext::MD_prof,
3266 MDBuilder(SI->getContext()).
3267 createBranchWeights(NewTrueWeight,
3268 (uint32_t)Weights[0]));
3272 // Prune obsolete incoming values off the successor's PHI nodes.
3273 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3274 isa<PHINode>(BBI); ++BBI) {
3275 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3276 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3278 SI->eraseFromParent();
3283 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3284 /// and use it to remove dead cases.
3285 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3286 Value *Cond = SI->getCondition();
3287 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
3288 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3289 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3291 // Gather dead cases.
3292 SmallVector<ConstantInt*, 8> DeadCases;
3293 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3294 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3295 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3296 DeadCases.push_back(I.getCaseValue());
3297 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3298 << I.getCaseValue() << "' is dead.\n");
3302 SmallVector<uint64_t, 8> Weights;
3303 bool HasWeight = HasBranchWeights(SI);
3305 GetBranchWeights(SI, Weights);
3306 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3309 // Remove dead cases from the switch.
3310 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3311 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3312 assert(Case != SI->case_default() &&
3313 "Case was not found. Probably mistake in DeadCases forming.");
3315 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3319 // Prune unused values from PHI nodes.
3320 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3321 SI->removeCase(Case);
3324 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3325 SI->setMetadata(LLVMContext::MD_prof,
3326 MDBuilder(SI->getParent()->getContext()).
3327 createBranchWeights(MDWeights));
3330 return !DeadCases.empty();
3333 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3334 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3335 /// by an unconditional branch), look at the phi node for BB in the successor
3336 /// block and see if the incoming value is equal to CaseValue. If so, return
3337 /// the phi node, and set PhiIndex to BB's index in the phi node.
3338 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3341 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3342 return NULL; // BB must be empty to be a candidate for simplification.
3343 if (!BB->getSinglePredecessor())
3344 return NULL; // BB must be dominated by the switch.
3346 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3347 if (!Branch || !Branch->isUnconditional())
3348 return NULL; // Terminator must be unconditional branch.
3350 BasicBlock *Succ = Branch->getSuccessor(0);
3352 BasicBlock::iterator I = Succ->begin();
3353 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3354 int Idx = PHI->getBasicBlockIndex(BB);
3355 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3357 Value *InValue = PHI->getIncomingValue(Idx);
3358 if (InValue != CaseValue) continue;
3367 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3368 /// instruction to a phi node dominated by the switch, if that would mean that
3369 /// some of the destination blocks of the switch can be folded away.
3370 /// Returns true if a change is made.
3371 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3372 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3373 ForwardingNodesMap ForwardingNodes;
3375 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3376 ConstantInt *CaseValue = I.getCaseValue();
3377 BasicBlock *CaseDest = I.getCaseSuccessor();
3380 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3384 ForwardingNodes[PHI].push_back(PhiIndex);
3387 bool Changed = false;
3389 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3390 E = ForwardingNodes.end(); I != E; ++I) {
3391 PHINode *Phi = I->first;
3392 SmallVectorImpl<int> &Indexes = I->second;
3394 if (Indexes.size() < 2) continue;
3396 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3397 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3404 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3405 /// initializing an array of constants like C.
3406 static bool ValidLookupTableConstant(Constant *C) {
3407 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3408 return CE->isGEPWithNoNotionalOverIndexing();
3410 return isa<ConstantFP>(C) ||
3411 isa<ConstantInt>(C) ||
3412 isa<ConstantPointerNull>(C) ||
3413 isa<GlobalValue>(C) ||
3417 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3418 /// its constant value in ConstantPool, returning 0 if it's not there.
3419 static Constant *LookupConstant(Value *V,
3420 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3421 if (Constant *C = dyn_cast<Constant>(V))
3423 return ConstantPool.lookup(V);
3426 /// ConstantFold - Try to fold instruction I into a constant. This works for
3427 /// simple instructions such as binary operations where both operands are
3428 /// constant or can be replaced by constants from the ConstantPool. Returns the
3429 /// resulting constant on success, 0 otherwise.
3430 static Constant *ConstantFold(Instruction *I,
3431 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3432 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3433 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool);
3436 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool);
3439 return ConstantExpr::get(BO->getOpcode(), A, B);
3442 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3443 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3446 Constant *B = LookupConstant(I->getOperand(1), ConstantPool);
3449 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B);
3452 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3453 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3456 if (A->isAllOnesValue())
3457 return LookupConstant(Select->getTrueValue(), ConstantPool);
3458 if (A->isNullValue())
3459 return LookupConstant(Select->getFalseValue(), ConstantPool);
3463 if (CastInst *Cast = dyn_cast<CastInst>(I)) {
3464 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3467 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy());
3473 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3474 /// at the common destination basic block, *CommonDest, for one of the case
3475 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3476 /// case), of a switch instruction SI.
3478 GetCaseResults(SwitchInst *SI,
3479 ConstantInt *CaseVal,
3480 BasicBlock *CaseDest,
3481 BasicBlock **CommonDest,
3482 SmallVectorImpl<std::pair<PHINode*,Constant*> > &Res) {
3483 // The block from which we enter the common destination.
3484 BasicBlock *Pred = SI->getParent();
3486 // If CaseDest is empty except for some side-effect free instructions through
3487 // which we can constant-propagate the CaseVal, continue to its successor.
3488 SmallDenseMap<Value*, Constant*> ConstantPool;
3489 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3490 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3492 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3493 // If the terminator is a simple branch, continue to the next block.
3494 if (T->getNumSuccessors() != 1)
3497 CaseDest = T->getSuccessor(0);
3498 } else if (isa<DbgInfoIntrinsic>(I)) {
3499 // Skip debug intrinsic.
3501 } else if (Constant *C = ConstantFold(I, ConstantPool)) {
3502 // Instruction is side-effect free and constant.
3503 ConstantPool.insert(std::make_pair(I, C));
3509 // If we did not have a CommonDest before, use the current one.
3511 *CommonDest = CaseDest;
3512 // If the destination isn't the common one, abort.
3513 if (CaseDest != *CommonDest)
3516 // Get the values for this case from phi nodes in the destination block.
3517 BasicBlock::iterator I = (*CommonDest)->begin();
3518 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3519 int Idx = PHI->getBasicBlockIndex(Pred);
3523 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3528 // Note: If the constant comes from constant-propagating the case value
3529 // through the CaseDest basic block, it will be safe to remove the
3530 // instructions in that block. They cannot be used (except in the phi nodes
3531 // we visit) outside CaseDest, because that block does not dominate its
3532 // successor. If it did, we would not be in this phi node.
3534 // Be conservative about which kinds of constants we support.
3535 if (!ValidLookupTableConstant(ConstVal))
3538 Res.push_back(std::make_pair(PHI, ConstVal));
3545 /// SwitchLookupTable - This class represents a lookup table that can be used
3546 /// to replace a switch.
3547 class SwitchLookupTable {
3549 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3550 /// with the contents of Values, using DefaultValue to fill any holes in the
3552 SwitchLookupTable(Module &M,
3554 ConstantInt *Offset,
3555 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3556 Constant *DefaultValue,
3557 const DataLayout *TD);
3559 /// BuildLookup - Build instructions with Builder to retrieve the value at
3560 /// the position given by Index in the lookup table.
3561 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3563 /// WouldFitInRegister - Return true if a table with TableSize elements of
3564 /// type ElementType would fit in a target-legal register.
3565 static bool WouldFitInRegister(const DataLayout *TD,
3567 const Type *ElementType);
3570 // Depending on the contents of the table, it can be represented in
3573 // For tables where each element contains the same value, we just have to
3574 // store that single value and return it for each lookup.
3577 // For small tables with integer elements, we can pack them into a bitmap
3578 // that fits into a target-legal register. Values are retrieved by
3579 // shift and mask operations.
3582 // The table is stored as an array of values. Values are retrieved by load
3583 // instructions from the table.
3587 // For SingleValueKind, this is the single value.
3588 Constant *SingleValue;
3590 // For BitMapKind, this is the bitmap.
3591 ConstantInt *BitMap;
3592 IntegerType *BitMapElementTy;
3594 // For ArrayKind, this is the array.
3595 GlobalVariable *Array;
3599 SwitchLookupTable::SwitchLookupTable(Module &M,
3601 ConstantInt *Offset,
3602 const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
3603 Constant *DefaultValue,
3604 const DataLayout *TD)
3605 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3606 assert(Values.size() && "Can't build lookup table without values!");
3607 assert(TableSize >= Values.size() && "Can't fit values in table!");
3609 // If all values in the table are equal, this is that value.
3610 SingleValue = Values.begin()->second;
3612 // Build up the table contents.
3613 SmallVector<Constant*, 64> TableContents(TableSize);
3614 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3615 ConstantInt *CaseVal = Values[I].first;
3616 Constant *CaseRes = Values[I].second;
3617 assert(CaseRes->getType() == DefaultValue->getType());
3619 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3621 TableContents[Idx] = CaseRes;
3623 if (CaseRes != SingleValue)
3627 // Fill in any holes in the table with the default result.
3628 if (Values.size() < TableSize) {
3629 for (uint64_t I = 0; I < TableSize; ++I) {
3630 if (!TableContents[I])
3631 TableContents[I] = DefaultValue;
3634 if (DefaultValue != SingleValue)
3638 // If each element in the table contains the same value, we only need to store
3639 // that single value.
3641 Kind = SingleValueKind;
3645 // If the type is integer and the table fits in a register, build a bitmap.
3646 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) {
3647 IntegerType *IT = cast<IntegerType>(DefaultValue->getType());
3648 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3649 for (uint64_t I = TableSize; I > 0; --I) {
3650 TableInt <<= IT->getBitWidth();
3651 // Insert values into the bitmap. Undef values are set to zero.
3652 if (!isa<UndefValue>(TableContents[I - 1])) {
3653 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3654 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3657 BitMap = ConstantInt::get(M.getContext(), TableInt);
3658 BitMapElementTy = IT;
3664 // Store the table in an array.
3665 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize);
3666 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3668 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3669 GlobalVariable::PrivateLinkage,
3672 Array->setUnnamedAddr(true);
3676 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3678 case SingleValueKind:
3681 // Type of the bitmap (e.g. i59).
3682 IntegerType *MapTy = BitMap->getType();
3684 // Cast Index to the same type as the bitmap.
3685 // Note: The Index is <= the number of elements in the table, so
3686 // truncating it to the width of the bitmask is safe.
3687 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3689 // Multiply the shift amount by the element width.
3690 ShiftAmt = Builder.CreateMul(ShiftAmt,
3691 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3695 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3696 "switch.downshift");
3698 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3702 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3703 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3705 return Builder.CreateLoad(GEP, "switch.load");
3708 llvm_unreachable("Unknown lookup table kind!");
3711 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD,
3713 const Type *ElementType) {
3716 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3719 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3720 // are <= 15, we could try to narrow the type.
3722 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3723 if (TableSize >= UINT_MAX/IT->getBitWidth())
3725 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth());
3728 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3729 /// for this switch, based on the number of cases, size of the table and the
3730 /// types of the results.
3731 static bool ShouldBuildLookupTable(SwitchInst *SI,
3733 const TargetTransformInfo &TTI,
3734 const DataLayout *TD,
3735 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3736 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3737 return false; // TableSize overflowed, or mul below might overflow.
3739 bool AllTablesFitInRegister = true;
3740 bool HasIllegalType = false;
3741 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3742 E = ResultTypes.end(); I != E; ++I) {
3743 Type *Ty = I->second;
3745 // Saturate this flag to true.
3746 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3748 // Saturate this flag to false.
3749 AllTablesFitInRegister = AllTablesFitInRegister &&
3750 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty);
3752 // If both flags saturate, we're done. NOTE: This *only* works with
3753 // saturating flags, and all flags have to saturate first due to the
3754 // non-deterministic behavior of iterating over a dense map.
3755 if (HasIllegalType && !AllTablesFitInRegister)
3759 // If each table would fit in a register, we should build it anyway.
3760 if (AllTablesFitInRegister)
3763 // Don't build a table that doesn't fit in-register if it has illegal types.
3767 // The table density should be at least 40%. This is the same criterion as for
3768 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3769 // FIXME: Find the best cut-off.
3770 return SI->getNumCases() * 10 >= TableSize * 4;
3773 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3774 /// phi nodes in a common successor block with different constant values,
3775 /// replace the switch with lookup tables.
3776 static bool SwitchToLookupTable(SwitchInst *SI,
3777 IRBuilder<> &Builder,
3778 const TargetTransformInfo &TTI,
3779 const DataLayout* TD) {
3780 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3782 // Only build lookup table when we have a target that supports it.
3783 if (!TTI.shouldBuildLookupTables())
3786 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3787 // split off a dense part and build a lookup table for that.
3789 // FIXME: This creates arrays of GEPs to constant strings, which means each
3790 // GEP needs a runtime relocation in PIC code. We should just build one big
3791 // string and lookup indices into that.
3793 // Ignore the switch if the number of cases is too small.
3794 // This is similar to the check when building jump tables in
3795 // SelectionDAGBuilder::handleJTSwitchCase.
3796 // FIXME: Determine the best cut-off.
3797 if (SI->getNumCases() < 4)
3800 // Figure out the corresponding result for each case value and phi node in the
3801 // common destination, as well as the the min and max case values.
3802 assert(SI->case_begin() != SI->case_end());
3803 SwitchInst::CaseIt CI = SI->case_begin();
3804 ConstantInt *MinCaseVal = CI.getCaseValue();
3805 ConstantInt *MaxCaseVal = CI.getCaseValue();
3807 BasicBlock *CommonDest = 0;
3808 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3809 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3810 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3811 SmallDenseMap<PHINode*, Type*> ResultTypes;
3812 SmallVector<PHINode*, 4> PHIs;
3814 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3815 ConstantInt *CaseVal = CI.getCaseValue();
3816 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3817 MinCaseVal = CaseVal;
3818 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3819 MaxCaseVal = CaseVal;
3821 // Resulting value at phi nodes for this case value.
3822 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3824 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3828 // Append the result from this case to the list for each phi.
3829 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3830 if (!ResultLists.count(I->first))
3831 PHIs.push_back(I->first);
3832 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3836 // Get the resulting values for the default case.
3837 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3838 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3839 DefaultResultsList))
3841 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3842 PHINode *PHI = DefaultResultsList[I].first;
3843 Constant *Result = DefaultResultsList[I].second;
3844 DefaultResults[PHI] = Result;
3845 ResultTypes[PHI] = Result->getType();
3848 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3849 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3850 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes))
3853 // Create the BB that does the lookups.
3854 Module &Mod = *CommonDest->getParent()->getParent();
3855 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3857 CommonDest->getParent(),
3860 // Check whether the condition value is within the case range, and branch to
3862 Builder.SetInsertPoint(SI);
3863 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3865 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3866 MinCaseVal->getType(), TableSize));
3867 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3869 // Populate the BB that does the lookups.
3870 Builder.SetInsertPoint(LookupBB);
3871 bool ReturnedEarly = false;
3872 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3873 PHINode *PHI = PHIs[I];
3875 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3876 DefaultResults[PHI], TD);
3878 Value *Result = Table.BuildLookup(TableIndex, Builder);
3880 // If the result is used to return immediately from the function, we want to
3881 // do that right here.
3882 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) &&
3883 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) {
3884 Builder.CreateRet(Result);
3885 ReturnedEarly = true;
3889 PHI->addIncoming(Result, LookupBB);
3893 Builder.CreateBr(CommonDest);
3895 // Remove the switch.
3896 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3897 BasicBlock *Succ = SI->getSuccessor(i);
3898 if (Succ == SI->getDefaultDest()) continue;
3899 Succ->removePredecessor(SI->getParent());
3901 SI->eraseFromParent();
3907 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3908 BasicBlock *BB = SI->getParent();
3910 if (isValueEqualityComparison(SI)) {
3911 // If we only have one predecessor, and if it is a branch on this value,
3912 // see if that predecessor totally determines the outcome of this switch.
3913 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3914 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3915 return SimplifyCFG(BB, TTI, TD) | true;
3917 Value *Cond = SI->getCondition();
3918 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3919 if (SimplifySwitchOnSelect(SI, Select))
3920 return SimplifyCFG(BB, TTI, TD) | true;
3922 // If the block only contains the switch, see if we can fold the block
3923 // away into any preds.
3924 BasicBlock::iterator BBI = BB->begin();
3925 // Ignore dbg intrinsics.
3926 while (isa<DbgInfoIntrinsic>(BBI))
3929 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3930 return SimplifyCFG(BB, TTI, TD) | true;
3933 // Try to transform the switch into an icmp and a branch.
3934 if (TurnSwitchRangeIntoICmp(SI, Builder))
3935 return SimplifyCFG(BB, TTI, TD) | true;
3937 // Remove unreachable cases.
3938 if (EliminateDeadSwitchCases(SI))
3939 return SimplifyCFG(BB, TTI, TD) | true;
3941 if (ForwardSwitchConditionToPHI(SI))
3942 return SimplifyCFG(BB, TTI, TD) | true;
3944 if (SwitchToLookupTable(SI, Builder, TTI, TD))
3945 return SimplifyCFG(BB, TTI, TD) | true;
3950 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3951 BasicBlock *BB = IBI->getParent();
3952 bool Changed = false;
3954 // Eliminate redundant destinations.
3955 SmallPtrSet<Value *, 8> Succs;
3956 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3957 BasicBlock *Dest = IBI->getDestination(i);
3958 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3959 Dest->removePredecessor(BB);
3960 IBI->removeDestination(i);
3966 if (IBI->getNumDestinations() == 0) {
3967 // If the indirectbr has no successors, change it to unreachable.
3968 new UnreachableInst(IBI->getContext(), IBI);
3969 EraseTerminatorInstAndDCECond(IBI);
3973 if (IBI->getNumDestinations() == 1) {
3974 // If the indirectbr has one successor, change it to a direct branch.
3975 BranchInst::Create(IBI->getDestination(0), IBI);
3976 EraseTerminatorInstAndDCECond(IBI);
3980 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3981 if (SimplifyIndirectBrOnSelect(IBI, SI))
3982 return SimplifyCFG(BB, TTI, TD) | true;
3987 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3988 BasicBlock *BB = BI->getParent();
3990 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3993 // If the Terminator is the only non-phi instruction, simplify the block.
3994 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3995 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3996 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3999 // If the only instruction in the block is a seteq/setne comparison
4000 // against a constant, try to simplify the block.
4001 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
4002 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
4003 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
4005 if (I->isTerminator() &&
4006 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD))
4010 // If this basic block is ONLY a compare and a branch, and if a predecessor
4011 // branches to us and our successor, fold the comparison into the
4012 // predecessor and use logical operations to update the incoming value
4013 // for PHI nodes in common successor.
4014 if (FoldBranchToCommonDest(BI))
4015 return SimplifyCFG(BB, TTI, TD) | true;
4020 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
4021 BasicBlock *BB = BI->getParent();
4023 // Conditional branch
4024 if (isValueEqualityComparison(BI)) {
4025 // If we only have one predecessor, and if it is a branch on this value,
4026 // see if that predecessor totally determines the outcome of this
4028 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
4029 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
4030 return SimplifyCFG(BB, TTI, TD) | true;
4032 // This block must be empty, except for the setcond inst, if it exists.
4033 // Ignore dbg intrinsics.
4034 BasicBlock::iterator I = BB->begin();
4035 // Ignore dbg intrinsics.
4036 while (isa<DbgInfoIntrinsic>(I))
4039 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4040 return SimplifyCFG(BB, TTI, TD) | true;
4041 } else if (&*I == cast<Instruction>(BI->getCondition())){
4043 // Ignore dbg intrinsics.
4044 while (isa<DbgInfoIntrinsic>(I))
4046 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4047 return SimplifyCFG(BB, TTI, TD) | true;
4051 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4052 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
4055 // If this basic block is ONLY a compare and a branch, and if a predecessor
4056 // branches to us and one of our successors, fold the comparison into the
4057 // predecessor and use logical operations to pick the right destination.
4058 if (FoldBranchToCommonDest(BI))
4059 return SimplifyCFG(BB, TTI, TD) | true;
4061 // We have a conditional branch to two blocks that are only reachable
4062 // from BI. We know that the condbr dominates the two blocks, so see if
4063 // there is any identical code in the "then" and "else" blocks. If so, we
4064 // can hoist it up to the branching block.
4065 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
4066 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
4067 if (HoistThenElseCodeToIf(BI))
4068 return SimplifyCFG(BB, TTI, TD) | true;
4070 // If Successor #1 has multiple preds, we may be able to conditionally
4071 // execute Successor #0 if it branches to successor #1.
4072 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4073 if (Succ0TI->getNumSuccessors() == 1 &&
4074 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4075 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
4076 return SimplifyCFG(BB, TTI, TD) | true;
4078 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
4079 // If Successor #0 has multiple preds, we may be able to conditionally
4080 // execute Successor #1 if it branches to successor #0.
4081 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4082 if (Succ1TI->getNumSuccessors() == 1 &&
4083 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4084 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
4085 return SimplifyCFG(BB, TTI, TD) | true;
4088 // If this is a branch on a phi node in the current block, thread control
4089 // through this block if any PHI node entries are constants.
4090 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4091 if (PN->getParent() == BI->getParent())
4092 if (FoldCondBranchOnPHI(BI, TD))
4093 return SimplifyCFG(BB, TTI, TD) | true;
4095 // Scan predecessor blocks for conditional branches.
4096 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4097 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4098 if (PBI != BI && PBI->isConditional())
4099 if (SimplifyCondBranchToCondBranch(PBI, BI))
4100 return SimplifyCFG(BB, TTI, TD) | true;
4105 /// If \param [in] BB has more than one predecessor that is a conditional
4106 /// branch, attempt to use parallel and/or for the branch condition. \returns
4107 /// true on success.
4111 /// %cmp10 = fcmp une float %tmp1, %tmp2
4112 /// br i1 %cmp1, label %if.then, label %lor.rhs
4116 /// %cmp11 = fcmp une float %tmp3, %tmp4
4117 /// br i1 %cmp11, label %if.then, label %ifend
4119 /// if.end: // the merge block
4122 /// if.then: // has two predecessors, both of them contains conditional branch.
4124 /// br label %if.end;
4128 /// %cmp10 = fcmp une float %tmp1, %tmp2
4130 /// %cmp11 = fcmp une float %tmp3, %tmp4
4131 /// %cmp12 = or i1 %cmp10, %cmp11 // parallel-or mode.
4132 /// br i1 %cmp12, label %if.then, label %ifend
4139 /// br label %if.end;
4141 /// Current implementation handles two cases.
4142 /// Case 1: \param BB is on the else-path.
4148 /// BB3 \ | where, BB1, BB2 contain conditional branches.
4149 /// \ | / BB3 contains unconditional branch.
4150 /// \ | / BB4 corresponds to \param BB which is also the merge.
4154 /// Corresponding source code:
4156 /// if (a == b && c == d)
4157 /// statement; // BB3
4159 /// Case 2: \param BB BB is on the then-path.
4164 /// \ / | where BB1, BB2 contain conditional branches.
4165 /// BB => BB3 | BB3 contains unconditiona branch and corresponds
4166 /// \ / to \param BB. BB4 is the merge.
4169 /// Corresponding source code:
4171 /// if (a == b || c == d)
4172 /// statement; // BB3
4174 /// In both cases, \param BB is the common successor of conditional branches.
4175 /// In Case 1, \param BB (BB4) has an unconditional branch (BB3) as
4176 /// its predecessor. In Case 2, \param BB (BB3) only has conditional branches
4177 /// as its predecessors.
4179 bool SimplifyCFGOpt::SimplifyParallelAndOr(BasicBlock *BB, IRBuilder<> &Builder,
4181 PHINode *PHI = dyn_cast<PHINode>(BB->begin());
4183 return false; // For simplicity, avoid cases containing PHI nodes.
4185 BasicBlock *LastCondBlock = NULL;
4186 BasicBlock *FirstCondBlock = NULL;
4187 BasicBlock *UnCondBlock = NULL;
4190 // Check predecessors of \param BB.
4191 SmallPtrSet<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
4192 for (SmallPtrSetIterator<BasicBlock*> PI = Preds.begin(), PE = Preds.end();
4194 BasicBlock *Pred = *PI;
4195 BranchInst *PBI = dyn_cast<BranchInst>(Pred->getTerminator());
4197 // All predecessors should terminate with a branch.
4201 BasicBlock *PP = Pred->getSinglePredecessor();
4203 if (PBI->isUnconditional()) {
4204 // Case 1: Pred (BB3) is an unconditional block, it should
4205 // have a single predecessor (BB2) that is also a predecessor
4206 // of \param BB (BB4) and should not have address-taken.
4207 // There should exist only one such unconditional
4208 // branch among the predecessors.
4209 if (UnCondBlock || !PP || (Preds.count(PP) == 0) ||
4210 Pred->hasAddressTaken())
4217 // Only conditional branches are allowed beyond this point.
4218 assert(PBI->isConditional());
4220 // Condition's unique use should be the branch instruction.
4221 Value *PC = PBI->getCondition();
4222 if (!PC || !PC->hasOneUse())
4225 if (PP && Preds.count(PP)) {
4226 // These are internal condition blocks to be merged from, e.g.,
4227 // BB2 in both cases.
4228 // Should not be address-taken.
4229 if (Pred->hasAddressTaken())
4232 // Instructions in the internal condition blocks should be safe
4234 for (BasicBlock::iterator BI = Pred->begin(), BE = PBI; BI != BE;) {
4235 Instruction *CI = BI++;
4236 if (isa<PHINode>(CI) ||
4237 !isSafeToSpeculativelyExecute(CI))
4241 // This is the condition block to be merged into, e.g. BB1 in
4245 FirstCondBlock = Pred;
4248 // Find whether BB is uniformly on the true (or false) path
4249 // for all of its predecessors.
4250 BasicBlock *PS1 = PBI->getSuccessor(0);
4251 BasicBlock *PS2 = PBI->getSuccessor(1);
4252 BasicBlock *PS = (PS1 == BB) ? PS2 : PS1;
4253 int CIdx = (PS1 == BB) ? 0 : 1;
4257 else if (CIdx != Idx)
4260 // PS is the successor which is not BB. Check successors to identify
4261 // the last conditional branch.
4262 if (Preds.count(PS) == 0) {
4264 // BB must have an unique successor.
4265 TerminatorInst *TBB = BB->getTerminator();
4266 if (TBB->getNumSuccessors() != 1)
4269 BasicBlock *SBB = TBB->getSuccessor(0);
4270 PHI = dyn_cast<PHINode>(SBB->begin());
4274 // PS (BB4) should be BB's successor.
4277 LastCondBlock = Pred;
4279 BranchInst *BPS = dyn_cast<BranchInst>(PS->getTerminator());
4280 if (BPS && BPS->isUnconditional()) {
4281 // Case 1: PS(BB3) should be an unconditional branch.
4282 LastCondBlock = Pred;
4287 if (!FirstCondBlock || !LastCondBlock || (FirstCondBlock == LastCondBlock))
4290 // Do the transformation.
4292 bool Iteration = true;
4293 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
4294 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
4295 BranchInst *PBI = dyn_cast<BranchInst>(FirstCondBlock->getTerminator());
4296 Value *PC = PBI->getCondition();
4298 CB = PBI->getSuccessor(1 - Idx);
4299 // Delete the conditional branch.
4300 FirstCondBlock->getInstList().pop_back();
4301 FirstCondBlock->getInstList().splice(FirstCondBlock->end(), CB->getInstList());
4302 PBI = cast<BranchInst>(FirstCondBlock->getTerminator());
4303 Value *CC = PBI->getCondition();
4304 // Merge conditions.
4305 Builder.SetInsertPoint(PBI);
4308 // Case 2, use parallel or.
4309 NC = Builder.CreateOr(PC, CC);
4311 // Case 1, use parallel and.
4312 NC = Builder.CreateAnd(PC, CC);
4314 PBI->replaceUsesOfWith(CC, NC);
4316 if (CB == LastCondBlock)
4318 // Remove internal conditional branches.
4319 CB->dropAllReferences();
4320 // make CB unreachable and let downstream to delete the block.
4321 new UnreachableInst(CB->getContext(), CB);
4322 } while (Iteration);
4324 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
4325 DEBUG(dbgs() << "Use parallel and/or in:\n" << *FirstCondBlock);
4329 /// Compare blocks from two if-regions, where \param Head1 is the entry of the
4330 /// 1st if-region. \param Head2 is the entry of the 2nd if-region. \param
4331 /// Block1 is a block in the 1st if-region to compare. \param Block2 is a block
4332 // in the 2nd if-region to compare. \returns true if \param Block1 and \param
4333 /// Block2 have identical instructions and do not have memory reference alias
4334 /// with \param Head2.
4336 bool SimplifyCFGOpt::CompareIfRegionBlock(BasicBlock *Head1, BasicBlock *Head2,
4337 BasicBlock *Block1, BasicBlock *Block2) {
4338 TerminatorInst *PTI2 = Head2->getTerminator();
4339 Instruction *PBI2 = Head2->begin();
4341 bool eq1 = (Block1 == Head1);
4342 bool eq2 = (Block2 == Head2);
4344 // An empty then-path or else-path.
4345 return (eq1 == eq2);
4348 // Check whether instructions in Block1 and Block2 are identical
4349 // and do not alias with instructions in Head2.
4350 BasicBlock::iterator iter1 = Block1->begin();
4351 BasicBlock::iterator end1 = Block1->getTerminator();
4352 BasicBlock::iterator iter2 = Block2->begin();
4353 BasicBlock::iterator end2 = Block2->getTerminator();
4356 if (iter1 == end1) {
4362 if (!iter1->isIdenticalTo(iter2))
4365 // Illegal to remove instructions with side effects except
4366 // non-volatile stores.
4367 if (iter1->mayHaveSideEffects()) {
4368 Instruction *CurI = &*iter1;
4369 StoreInst *SI = dyn_cast<StoreInst>(CurI);
4370 if (!SI || SI->isVolatile())
4374 // For simplicity and speed, data dependency check can be
4375 // avoided if read from memory doesn't exist.
4376 if (iter1->mayReadFromMemory())
4379 if (iter1->mayWriteToMemory()) {
4380 for (BasicBlock::iterator BI = PBI2, BE = PTI2; BI != BE; ++BI) {
4381 if (BI->mayReadFromMemory() || BI->mayWriteToMemory()) {
4382 // Check alias with Head2.
4383 if (!AA || AA->alias(iter1, BI))
4395 /// Check whether \param BB is the merge block of a if-region. If yes, check
4396 /// whether there exists an adjacent if-region upstream, the two if-regions
4397 /// contain identical instuctions and can be legally merged. \returns true if
4398 /// the two if-regions are merged.
4410 bool SimplifyCFGOpt::MergeIfRegion(BasicBlock *BB, IRBuilder<> &Builder,
4412 BasicBlock *IfTrue2, *IfFalse2;
4413 Value *IfCond2 = GetIfCondition(BB, IfTrue2, IfFalse2);
4414 Instruction *CInst2 = dyn_cast_or_null<Instruction>(IfCond2);
4418 BasicBlock *SecondEntryBlock = CInst2->getParent();
4419 if (SecondEntryBlock->hasAddressTaken())
4422 BasicBlock *IfTrue1, *IfFalse1;
4423 Value *IfCond1 = GetIfCondition(SecondEntryBlock, IfTrue1, IfFalse1);
4424 Instruction *CInst1 = dyn_cast_or_null<Instruction>(IfCond1);
4428 BasicBlock *FirstEntryBlock = CInst1->getParent();
4430 // Either then-path or else-path should be empty.
4431 if ((IfTrue1 != FirstEntryBlock) && (IfFalse1 != FirstEntryBlock))
4433 if ((IfTrue2 != SecondEntryBlock) && (IfFalse2 != SecondEntryBlock))
4436 TerminatorInst *PTI2 = SecondEntryBlock->getTerminator();
4437 Instruction *PBI2 = SecondEntryBlock->begin();
4439 if (!CompareIfRegionBlock(FirstEntryBlock, SecondEntryBlock, IfTrue1, IfTrue2))
4442 if (!CompareIfRegionBlock(FirstEntryBlock, SecondEntryBlock, IfFalse1, IfFalse2))
4445 // Check whether \param SecondEntryBlock has side-effect and is safe to speculate.
4446 for (BasicBlock::iterator BI = PBI2, BE = PTI2; BI != BE; ++BI) {
4447 Instruction *CI = BI;
4448 if (isa<PHINode>(CI) || CI->mayHaveSideEffects() ||
4449 !isSafeToSpeculativelyExecute(CI))
4453 // Merge \param SecondEntryBlock into \param FirstEntryBlock.
4454 FirstEntryBlock->getInstList().pop_back();
4455 FirstEntryBlock->getInstList().splice(FirstEntryBlock->end(), SecondEntryBlock->getInstList());
4456 BranchInst *PBI = dyn_cast<BranchInst>(FirstEntryBlock->getTerminator());
4457 Value *CC = PBI->getCondition();
4458 BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
4459 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
4460 Builder.SetInsertPoint(PBI);
4461 Value *NC = Builder.CreateOr(CInst1, CC);
4462 PBI->replaceUsesOfWith(CC, NC);
4464 Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
4467 if (IfTrue1 != FirstEntryBlock) {
4468 IfTrue1->dropAllReferences();
4469 IfTrue1->eraseFromParent();
4473 if (IfFalse1 != FirstEntryBlock) {
4474 IfFalse1->dropAllReferences();
4475 IfFalse1->eraseFromParent();
4478 // Remove \param SecondEntryBlock
4479 SecondEntryBlock->dropAllReferences();
4480 SecondEntryBlock->eraseFromParent();
4481 DEBUG(dbgs() << "If conditions merged into:\n" << *FirstEntryBlock);
4485 /// Check if passing a value to an instruction will cause undefined behavior.
4486 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4487 Constant *C = dyn_cast<Constant>(V);
4494 if (C->isNullValue()) {
4495 // Only look at the first use, avoid hurting compile time with long uselists
4496 User *Use = *I->use_begin();
4498 // Now make sure that there are no instructions in between that can alter
4499 // control flow (eg. calls)
4500 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4501 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4504 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4505 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4506 if (GEP->getPointerOperand() == I)
4507 return passingValueIsAlwaysUndefined(V, GEP);
4509 // Look through bitcasts.
4510 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4511 return passingValueIsAlwaysUndefined(V, BC);
4513 // Load from null is undefined.
4514 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4515 if (!LI->isVolatile())
4516 return LI->getPointerAddressSpace() == 0;
4518 // Store to null is undefined.
4519 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4520 if (!SI->isVolatile())
4521 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4526 /// If BB has an incoming value that will always trigger undefined behavior
4527 /// (eg. null pointer dereference), remove the branch leading here.
4528 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4529 for (BasicBlock::iterator i = BB->begin();
4530 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4531 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4532 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4533 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4534 IRBuilder<> Builder(T);
4535 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4536 BB->removePredecessor(PHI->getIncomingBlock(i));
4537 // Turn uncoditional branches into unreachables and remove the dead
4538 // destination from conditional branches.
4539 if (BI->isUnconditional())
4540 Builder.CreateUnreachable();
4542 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4543 BI->getSuccessor(0));
4544 BI->eraseFromParent();
4547 // TODO: SwitchInst.
4553 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4554 bool Changed = false;
4556 assert(BB && BB->getParent() && "Block not embedded in function!");
4557 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4559 // Remove basic blocks that have no predecessors (except the entry block)...
4560 // or that just have themself as a predecessor. These are unreachable.
4561 if ((pred_begin(BB) == pred_end(BB) &&
4562 BB != &BB->getParent()->getEntryBlock()) ||
4563 BB->getSinglePredecessor() == BB) {
4564 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4565 DeleteDeadBlock(BB);
4569 // Check to see if we can constant propagate this terminator instruction
4571 Changed |= ConstantFoldTerminator(BB, true);
4573 // Check for and eliminate duplicate PHI nodes in this block.
4574 Changed |= EliminateDuplicatePHINodes(BB);
4576 // Check for and remove branches that will always cause undefined behavior.
4577 Changed |= removeUndefIntroducingPredecessor(BB);
4579 // Merge basic blocks into their predecessor if there is only one distinct
4580 // pred, and if there is only one distinct successor of the predecessor, and
4581 // if there are no PHI nodes.
4583 if (MergeBlockIntoPredecessor(BB))
4586 IRBuilder<> Builder(BB);
4587 // Whether to optimize conditional branches.
4588 bool OptCB = (ParallelAndOr && AA && TTI.hasBranchDivergence());
4590 if (OptCB && SimplifyParallelAndOr(BB, Builder))
4593 // If there is a trivial two-entry PHI node in this basic block, and we can
4594 // eliminate it, do so now.
4595 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4596 if (PN->getNumIncomingValues() == 2)
4597 Changed |= FoldTwoEntryPHINode(PN, TD);
4599 Builder.SetInsertPoint(BB->getTerminator());
4600 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4601 if (BI->isUnconditional()) {
4602 if (SimplifyUncondBranch(BI, Builder)) return true;
4604 if (SimplifyCondBranch(BI, Builder)) return true;
4606 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4607 if (SimplifyReturn(RI, Builder)) return true;
4608 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4609 if (SimplifyResume(RI, Builder)) return true;
4610 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4611 if (SimplifySwitch(SI, Builder)) return true;
4612 } else if (UnreachableInst *UI =
4613 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4614 if (SimplifyUnreachable(UI)) return true;
4615 } else if (IndirectBrInst *IBI =
4616 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4617 if (SimplifyIndirectBr(IBI)) return true;
4620 if (OptCB && MergeIfRegion(BB, Builder))
4626 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4627 /// example, it adjusts branches to branches to eliminate the extra hop, it
4628 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4629 /// of the CFG. It returns true if a modification was made.
4631 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4632 const DataLayout *TD, AliasAnalysis *AA) {
4633 return SimplifyCFGOpt(TTI, TD, AA).run(BB);