1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "simplifycfg"
15 #include "llvm/Transforms/Utils/Local.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/GlobalVariable.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/MDBuilder.h"
34 #include "llvm/IR/Metadata.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/IR/Type.h"
38 #include "llvm/Support/CFG.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/ConstantRange.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/NoFolder.h"
43 #include "llvm/Support/PatternMatch.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 using namespace PatternMatch;
52 static cl::opt<unsigned>
53 PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
54 cl::desc("Control the amount of phi node folding to perform (default = 1)"));
57 DupRet("simplifycfg-dup-ret", cl::Hidden, cl::init(false),
58 cl::desc("Duplicate return instructions into unconditional branches"));
61 SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
62 cl::desc("Sink common instructions down to the end block"));
65 HoistCondStores("simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
66 cl::desc("Hoist conditional stores if an unconditional store preceeds"));
68 STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
69 STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
70 STATISTIC(NumSinkCommons, "Number of common instructions sunk down to the end block");
71 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
74 /// ValueEqualityComparisonCase - Represents a case of a switch.
75 struct ValueEqualityComparisonCase {
79 ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
80 : Value(Value), Dest(Dest) {}
82 bool operator<(ValueEqualityComparisonCase RHS) const {
83 // Comparing pointers is ok as we only rely on the order for uniquing.
84 return Value < RHS.Value;
87 bool operator==(BasicBlock *RHSDest) const { return Dest == RHSDest; }
90 class SimplifyCFGOpt {
91 const TargetTransformInfo &TTI;
92 const DataLayout *const TD;
94 Value *isValueEqualityComparison(TerminatorInst *TI);
95 BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
96 std::vector<ValueEqualityComparisonCase> &Cases);
97 bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
99 IRBuilder<> &Builder);
100 bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
101 IRBuilder<> &Builder);
103 bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
104 bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
105 bool SimplifyUnreachable(UnreachableInst *UI);
106 bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
107 bool SimplifyIndirectBr(IndirectBrInst *IBI);
108 bool SimplifyUncondBranch(BranchInst *BI, IRBuilder <> &Builder);
109 bool SimplifyCondBranch(BranchInst *BI, IRBuilder <>&Builder);
112 SimplifyCFGOpt(const TargetTransformInfo &TTI, const DataLayout *TD)
113 : TTI(TTI), TD(TD) {}
114 bool run(BasicBlock *BB);
118 /// SafeToMergeTerminators - Return true if it is safe to merge these two
119 /// terminator instructions together.
121 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
122 if (SI1 == SI2) return false; // Can't merge with self!
124 // It is not safe to merge these two switch instructions if they have a common
125 // successor, and if that successor has a PHI node, and if *that* PHI node has
126 // conflicting incoming values from the two switch blocks.
127 BasicBlock *SI1BB = SI1->getParent();
128 BasicBlock *SI2BB = SI2->getParent();
129 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
131 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
132 if (SI1Succs.count(*I))
133 for (BasicBlock::iterator BBI = (*I)->begin();
134 isa<PHINode>(BBI); ++BBI) {
135 PHINode *PN = cast<PHINode>(BBI);
136 if (PN->getIncomingValueForBlock(SI1BB) !=
137 PN->getIncomingValueForBlock(SI2BB))
144 /// isProfitableToFoldUnconditional - Return true if it is safe and profitable
145 /// to merge these two terminator instructions together, where SI1 is an
146 /// unconditional branch. PhiNodes will store all PHI nodes in common
149 static bool isProfitableToFoldUnconditional(BranchInst *SI1,
152 SmallVectorImpl<PHINode*> &PhiNodes) {
153 if (SI1 == SI2) return false; // Can't merge with self!
154 assert(SI1->isUnconditional() && SI2->isConditional());
156 // We fold the unconditional branch if we can easily update all PHI nodes in
157 // common successors:
158 // 1> We have a constant incoming value for the conditional branch;
159 // 2> We have "Cond" as the incoming value for the unconditional branch;
160 // 3> SI2->getCondition() and Cond have same operands.
161 CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
162 if (!Ci2) return false;
163 if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
164 Cond->getOperand(1) == Ci2->getOperand(1)) &&
165 !(Cond->getOperand(0) == Ci2->getOperand(1) &&
166 Cond->getOperand(1) == Ci2->getOperand(0)))
169 BasicBlock *SI1BB = SI1->getParent();
170 BasicBlock *SI2BB = SI2->getParent();
171 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
172 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
173 if (SI1Succs.count(*I))
174 for (BasicBlock::iterator BBI = (*I)->begin();
175 isa<PHINode>(BBI); ++BBI) {
176 PHINode *PN = cast<PHINode>(BBI);
177 if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
178 !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
180 PhiNodes.push_back(PN);
185 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
186 /// now be entries in it from the 'NewPred' block. The values that will be
187 /// flowing into the PHI nodes will be the same as those coming in from
188 /// ExistPred, an existing predecessor of Succ.
189 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
190 BasicBlock *ExistPred) {
191 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
194 for (BasicBlock::iterator I = Succ->begin();
195 (PN = dyn_cast<PHINode>(I)); ++I)
196 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
200 /// GetIfCondition - Given a basic block (BB) with two predecessors (and at
201 /// least one PHI node in it), check to see if the merge at this block is due
202 /// to an "if condition". If so, return the boolean condition that determines
203 /// which entry into BB will be taken. Also, return by references the block
204 /// that will be entered from if the condition is true, and the block that will
205 /// be entered if the condition is false.
207 /// This does no checking to see if the true/false blocks have large or unsavory
208 /// instructions in them.
209 static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue,
210 BasicBlock *&IfFalse) {
211 PHINode *SomePHI = cast<PHINode>(BB->begin());
212 assert(SomePHI->getNumIncomingValues() == 2 &&
213 "Function can only handle blocks with 2 predecessors!");
214 BasicBlock *Pred1 = SomePHI->getIncomingBlock(0);
215 BasicBlock *Pred2 = SomePHI->getIncomingBlock(1);
217 // We can only handle branches. Other control flow will be lowered to
218 // branches if possible anyway.
219 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator());
220 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator());
221 if (Pred1Br == 0 || Pred2Br == 0)
224 // Eliminate code duplication by ensuring that Pred1Br is conditional if
226 if (Pred2Br->isConditional()) {
227 // If both branches are conditional, we don't have an "if statement". In
228 // reality, we could transform this case, but since the condition will be
229 // required anyway, we stand no chance of eliminating it, so the xform is
230 // probably not profitable.
231 if (Pred1Br->isConditional())
234 std::swap(Pred1, Pred2);
235 std::swap(Pred1Br, Pred2Br);
238 if (Pred1Br->isConditional()) {
239 // The only thing we have to watch out for here is to make sure that Pred2
240 // doesn't have incoming edges from other blocks. If it does, the condition
241 // doesn't dominate BB.
242 if (Pred2->getSinglePredecessor() == 0)
245 // If we found a conditional branch predecessor, make sure that it branches
246 // to BB and Pred2Br. If it doesn't, this isn't an "if statement".
247 if (Pred1Br->getSuccessor(0) == BB &&
248 Pred1Br->getSuccessor(1) == Pred2) {
251 } else if (Pred1Br->getSuccessor(0) == Pred2 &&
252 Pred1Br->getSuccessor(1) == BB) {
256 // We know that one arm of the conditional goes to BB, so the other must
257 // go somewhere unrelated, and this must not be an "if statement".
261 return Pred1Br->getCondition();
264 // Ok, if we got here, both predecessors end with an unconditional branch to
265 // BB. Don't panic! If both blocks only have a single (identical)
266 // predecessor, and THAT is a conditional branch, then we're all ok!
267 BasicBlock *CommonPred = Pred1->getSinglePredecessor();
268 if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor())
271 // Otherwise, if this is a conditional branch, then we can use it!
272 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator());
273 if (BI == 0) return 0;
275 assert(BI->isConditional() && "Two successors but not conditional?");
276 if (BI->getSuccessor(0) == Pred1) {
283 return BI->getCondition();
286 /// ComputeSpeculationCost - Compute an abstract "cost" of speculating the
287 /// given instruction, which is assumed to be safe to speculate. 1 means
288 /// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
289 static unsigned ComputeSpeculationCost(const User *I) {
290 assert(isSafeToSpeculativelyExecute(I) &&
291 "Instruction is not safe to speculatively execute!");
292 switch (Operator::getOpcode(I)) {
294 // In doubt, be conservative.
296 case Instruction::GetElementPtr:
297 // GEPs are cheap if all indices are constant.
298 if (!cast<GEPOperator>(I)->hasAllConstantIndices())
301 case Instruction::Load:
302 case Instruction::Add:
303 case Instruction::Sub:
304 case Instruction::And:
305 case Instruction::Or:
306 case Instruction::Xor:
307 case Instruction::Shl:
308 case Instruction::LShr:
309 case Instruction::AShr:
310 case Instruction::ICmp:
311 case Instruction::Trunc:
312 case Instruction::ZExt:
313 case Instruction::SExt:
314 return 1; // These are all cheap.
316 case Instruction::Call:
317 case Instruction::Select:
322 /// DominatesMergePoint - If we have a merge point of an "if condition" as
323 /// accepted above, return true if the specified value dominates the block. We
324 /// don't handle the true generality of domination here, just a special case
325 /// which works well enough for us.
327 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
328 /// see if V (which must be an instruction) and its recursive operands
329 /// that do not dominate BB have a combined cost lower than CostRemaining and
330 /// are non-trapping. If both are true, the instruction is inserted into the
331 /// set and true is returned.
333 /// The cost for most non-trapping instructions is defined as 1 except for
334 /// Select whose cost is 2.
336 /// After this function returns, CostRemaining is decreased by the cost of
337 /// V plus its non-dominating operands. If that cost is greater than
338 /// CostRemaining, false is returned and CostRemaining is undefined.
339 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
340 SmallPtrSet<Instruction*, 4> *AggressiveInsts,
341 unsigned &CostRemaining) {
342 Instruction *I = dyn_cast<Instruction>(V);
344 // Non-instructions all dominate instructions, but not all constantexprs
345 // can be executed unconditionally.
346 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
351 BasicBlock *PBB = I->getParent();
353 // We don't want to allow weird loops that might have the "if condition" in
354 // the bottom of this block.
355 if (PBB == BB) return false;
357 // If this instruction is defined in a block that contains an unconditional
358 // branch to BB, then it must be in the 'conditional' part of the "if
359 // statement". If not, it definitely dominates the region.
360 BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator());
361 if (BI == 0 || BI->isConditional() || BI->getSuccessor(0) != BB)
364 // If we aren't allowing aggressive promotion anymore, then don't consider
365 // instructions in the 'if region'.
366 if (AggressiveInsts == 0) return false;
368 // If we have seen this instruction before, don't count it again.
369 if (AggressiveInsts->count(I)) return true;
371 // Okay, it looks like the instruction IS in the "condition". Check to
372 // see if it's a cheap instruction to unconditionally compute, and if it
373 // only uses stuff defined outside of the condition. If so, hoist it out.
374 if (!isSafeToSpeculativelyExecute(I))
377 unsigned Cost = ComputeSpeculationCost(I);
379 if (Cost > CostRemaining)
382 CostRemaining -= Cost;
384 // Okay, we can only really hoist these out if their operands do
385 // not take us over the cost threshold.
386 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
387 if (!DominatesMergePoint(*i, BB, AggressiveInsts, CostRemaining))
389 // Okay, it's safe to do this! Remember this instruction.
390 AggressiveInsts->insert(I);
394 /// GetConstantInt - Extract ConstantInt from value, looking through IntToPtr
395 /// and PointerNullValue. Return NULL if value is not a constant int.
396 static ConstantInt *GetConstantInt(Value *V, const DataLayout *TD) {
397 // Normal constant int.
398 ConstantInt *CI = dyn_cast<ConstantInt>(V);
399 if (CI || !TD || !isa<Constant>(V) || !V->getType()->isPointerTy())
402 // This is some kind of pointer constant. Turn it into a pointer-sized
403 // ConstantInt if possible.
404 IntegerType *PtrTy = cast<IntegerType>(TD->getIntPtrType(V->getType()));
406 // Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
407 if (isa<ConstantPointerNull>(V))
408 return ConstantInt::get(PtrTy, 0);
410 // IntToPtr const int.
411 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
412 if (CE->getOpcode() == Instruction::IntToPtr)
413 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(0))) {
414 // The constant is very likely to have the right type already.
415 if (CI->getType() == PtrTy)
418 return cast<ConstantInt>
419 (ConstantExpr::getIntegerCast(CI, PtrTy, /*isSigned=*/false));
424 /// GatherConstantCompares - Given a potentially 'or'd or 'and'd together
425 /// collection of icmp eq/ne instructions that compare a value against a
426 /// constant, return the value being compared, and stick the constant into the
429 GatherConstantCompares(Value *V, std::vector<ConstantInt*> &Vals, Value *&Extra,
430 const DataLayout *TD, bool isEQ, unsigned &UsedICmps) {
431 Instruction *I = dyn_cast<Instruction>(V);
432 if (I == 0) return 0;
434 // If this is an icmp against a constant, handle this as one of the cases.
435 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
436 if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) {
440 if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
441 // (x & ~2^x) == y --> x == y || x == y|2^x
442 // This undoes a transformation done by instcombine to fuse 2 compares.
443 if (match(ICI->getOperand(0),
444 m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
445 APInt Not = ~RHSC->getValue();
446 if (Not.isPowerOf2()) {
449 ConstantInt::get(C->getContext(), C->getValue() | Not));
457 return I->getOperand(0);
460 // If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
463 ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
465 // Shift the range if the compare is fed by an add. This is the range
466 // compare idiom as emitted by instcombine.
468 match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)));
470 Span = Span.subtract(RHSC->getValue());
472 // If this is an and/!= check then we want to optimize "x ugt 2" into
475 Span = Span.inverse();
477 // If there are a ton of values, we don't want to make a ginormous switch.
478 if (Span.getSetSize().ugt(8) || Span.isEmptySet())
481 for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
482 Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
484 return hasAdd ? RHSVal : I->getOperand(0);
489 // Otherwise, we can only handle an | or &, depending on isEQ.
490 if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
493 unsigned NumValsBeforeLHS = Vals.size();
494 unsigned UsedICmpsBeforeLHS = UsedICmps;
495 if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD,
497 unsigned NumVals = Vals.size();
498 unsigned UsedICmpsBeforeRHS = UsedICmps;
499 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
503 Vals.resize(NumVals);
504 UsedICmps = UsedICmpsBeforeRHS;
507 // The RHS of the or/and can't be folded in and we haven't used "Extra" yet,
508 // set it and return success.
509 if (Extra == 0 || Extra == I->getOperand(1)) {
510 Extra = I->getOperand(1);
514 Vals.resize(NumValsBeforeLHS);
515 UsedICmps = UsedICmpsBeforeLHS;
519 // If the LHS can't be folded in, but Extra is available and RHS can, try to
521 if (Extra == 0 || Extra == I->getOperand(0)) {
522 Value *OldExtra = Extra;
523 Extra = I->getOperand(0);
524 if (Value *RHS = GatherConstantCompares(I->getOperand(1), Vals, Extra, TD,
527 assert(Vals.size() == NumValsBeforeLHS);
534 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
535 Instruction *Cond = 0;
536 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
537 Cond = dyn_cast<Instruction>(SI->getCondition());
538 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
539 if (BI->isConditional())
540 Cond = dyn_cast<Instruction>(BI->getCondition());
541 } else if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(TI)) {
542 Cond = dyn_cast<Instruction>(IBI->getAddress());
545 TI->eraseFromParent();
546 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
549 /// isValueEqualityComparison - Return true if the specified terminator checks
550 /// to see if a value is equal to constant integer value.
551 Value *SimplifyCFGOpt::isValueEqualityComparison(TerminatorInst *TI) {
553 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
554 // Do not permit merging of large switch instructions into their
555 // predecessors unless there is only one predecessor.
556 if (SI->getNumSuccessors()*std::distance(pred_begin(SI->getParent()),
557 pred_end(SI->getParent())) <= 128)
558 CV = SI->getCondition();
559 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
560 if (BI->isConditional() && BI->getCondition()->hasOneUse())
561 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
562 if (ICI->isEquality() && GetConstantInt(ICI->getOperand(1), TD))
563 CV = ICI->getOperand(0);
565 // Unwrap any lossless ptrtoint cast.
566 if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext()))
567 if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV))
568 CV = PTII->getOperand(0);
572 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
573 /// decode all of the 'cases' that it represents and return the 'default' block.
574 BasicBlock *SimplifyCFGOpt::
575 GetValueEqualityComparisonCases(TerminatorInst *TI,
576 std::vector<ValueEqualityComparisonCase>
578 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
579 Cases.reserve(SI->getNumCases());
580 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
581 Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
582 i.getCaseSuccessor()));
583 return SI->getDefaultDest();
586 BranchInst *BI = cast<BranchInst>(TI);
587 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
588 BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
589 Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
592 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
596 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
597 /// in the list that match the specified block.
598 static void EliminateBlockCases(BasicBlock *BB,
599 std::vector<ValueEqualityComparisonCase> &Cases) {
600 Cases.erase(std::remove(Cases.begin(), Cases.end(), BB), Cases.end());
603 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
606 ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
607 std::vector<ValueEqualityComparisonCase > &C2) {
608 std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
610 // Make V1 be smaller than V2.
611 if (V1->size() > V2->size())
614 if (V1->size() == 0) return false;
615 if (V1->size() == 1) {
617 ConstantInt *TheVal = (*V1)[0].Value;
618 for (unsigned i = 0, e = V2->size(); i != e; ++i)
619 if (TheVal == (*V2)[i].Value)
623 // Otherwise, just sort both lists and compare element by element.
624 array_pod_sort(V1->begin(), V1->end());
625 array_pod_sort(V2->begin(), V2->end());
626 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
627 while (i1 != e1 && i2 != e2) {
628 if ((*V1)[i1].Value == (*V2)[i2].Value)
630 if ((*V1)[i1].Value < (*V2)[i2].Value)
638 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
639 /// terminator instruction and its block is known to only have a single
640 /// predecessor block, check to see if that predecessor is also a value
641 /// comparison with the same value, and if that comparison determines the
642 /// outcome of this comparison. If so, simplify TI. This does a very limited
643 /// form of jump threading.
644 bool SimplifyCFGOpt::
645 SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
647 IRBuilder<> &Builder) {
648 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
649 if (!PredVal) return false; // Not a value comparison in predecessor.
651 Value *ThisVal = isValueEqualityComparison(TI);
652 assert(ThisVal && "This isn't a value comparison!!");
653 if (ThisVal != PredVal) return false; // Different predicates.
655 // TODO: Preserve branch weight metadata, similarly to how
656 // FoldValueComparisonIntoPredecessors preserves it.
658 // Find out information about when control will move from Pred to TI's block.
659 std::vector<ValueEqualityComparisonCase> PredCases;
660 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
662 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
664 // Find information about how control leaves this block.
665 std::vector<ValueEqualityComparisonCase> ThisCases;
666 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
667 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
669 // If TI's block is the default block from Pred's comparison, potentially
670 // simplify TI based on this knowledge.
671 if (PredDef == TI->getParent()) {
672 // If we are here, we know that the value is none of those cases listed in
673 // PredCases. If there are any cases in ThisCases that are in PredCases, we
675 if (!ValuesOverlap(PredCases, ThisCases))
678 if (isa<BranchInst>(TI)) {
679 // Okay, one of the successors of this condbr is dead. Convert it to a
681 assert(ThisCases.size() == 1 && "Branch can only have one case!");
682 // Insert the new branch.
683 Instruction *NI = Builder.CreateBr(ThisDef);
686 // Remove PHI node entries for the dead edge.
687 ThisCases[0].Dest->removePredecessor(TI->getParent());
689 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
690 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
692 EraseTerminatorInstAndDCECond(TI);
696 SwitchInst *SI = cast<SwitchInst>(TI);
697 // Okay, TI has cases that are statically dead, prune them away.
698 SmallPtrSet<Constant*, 16> DeadCases;
699 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
700 DeadCases.insert(PredCases[i].Value);
702 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
703 << "Through successor TI: " << *TI);
705 // Collect branch weights into a vector.
706 SmallVector<uint32_t, 8> Weights;
707 MDNode* MD = SI->getMetadata(LLVMContext::MD_prof);
708 bool HasWeight = MD && (MD->getNumOperands() == 2 + SI->getNumCases());
710 for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
712 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(MD_i));
714 Weights.push_back(CI->getValue().getZExtValue());
716 for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
718 if (DeadCases.count(i.getCaseValue())) {
720 std::swap(Weights[i.getCaseIndex()+1], Weights.back());
723 i.getCaseSuccessor()->removePredecessor(TI->getParent());
727 if (HasWeight && Weights.size() >= 2)
728 SI->setMetadata(LLVMContext::MD_prof,
729 MDBuilder(SI->getParent()->getContext()).
730 createBranchWeights(Weights));
732 DEBUG(dbgs() << "Leaving: " << *TI << "\n");
736 // Otherwise, TI's block must correspond to some matched value. Find out
737 // which value (or set of values) this is.
738 ConstantInt *TIV = 0;
739 BasicBlock *TIBB = TI->getParent();
740 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
741 if (PredCases[i].Dest == TIBB) {
743 return false; // Cannot handle multiple values coming to this block.
744 TIV = PredCases[i].Value;
746 assert(TIV && "No edge from pred to succ?");
748 // Okay, we found the one constant that our value can be if we get into TI's
749 // BB. Find out which successor will unconditionally be branched to.
750 BasicBlock *TheRealDest = 0;
751 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
752 if (ThisCases[i].Value == TIV) {
753 TheRealDest = ThisCases[i].Dest;
757 // If not handled by any explicit cases, it is handled by the default case.
758 if (TheRealDest == 0) TheRealDest = ThisDef;
760 // Remove PHI node entries for dead edges.
761 BasicBlock *CheckEdge = TheRealDest;
762 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
763 if (*SI != CheckEdge)
764 (*SI)->removePredecessor(TIBB);
768 // Insert the new branch.
769 Instruction *NI = Builder.CreateBr(TheRealDest);
772 DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
773 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
775 EraseTerminatorInstAndDCECond(TI);
780 /// ConstantIntOrdering - This class implements a stable ordering of constant
781 /// integers that does not depend on their address. This is important for
782 /// applications that sort ConstantInt's to ensure uniqueness.
783 struct ConstantIntOrdering {
784 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
785 return LHS->getValue().ult(RHS->getValue());
790 static int ConstantIntSortPredicate(const void *P1, const void *P2) {
791 const ConstantInt *LHS = *(const ConstantInt*const*)P1;
792 const ConstantInt *RHS = *(const ConstantInt*const*)P2;
793 if (LHS->getValue().ult(RHS->getValue()))
795 if (LHS->getValue() == RHS->getValue())
800 static inline bool HasBranchWeights(const Instruction* I) {
801 MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
802 if (ProfMD && ProfMD->getOperand(0))
803 if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
804 return MDS->getString().equals("branch_weights");
809 /// Get Weights of a given TerminatorInst, the default weight is at the front
810 /// of the vector. If TI is a conditional eq, we need to swap the branch-weight
812 static void GetBranchWeights(TerminatorInst *TI,
813 SmallVectorImpl<uint64_t> &Weights) {
814 MDNode* MD = TI->getMetadata(LLVMContext::MD_prof);
816 for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
817 ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i));
819 Weights.push_back(CI->getValue().getZExtValue());
822 // If TI is a conditional eq, the default case is the false case,
823 // and the corresponding branch-weight data is at index 2. We swap the
824 // default weight to be the first entry.
825 if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
826 assert(Weights.size() == 2);
827 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
828 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
829 std::swap(Weights.front(), Weights.back());
833 /// Sees if any of the weights are too big for a uint32_t, and halves all the
834 /// weights if any are.
835 static void FitWeights(MutableArrayRef<uint64_t> Weights) {
837 for (unsigned i = 0; i < Weights.size(); ++i)
838 if (Weights[i] > UINT_MAX) {
846 for (unsigned i = 0; i < Weights.size(); ++i)
850 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
851 /// equality comparison instruction (either a switch or a branch on "X == c").
852 /// See if any of the predecessors of the terminator block are value comparisons
853 /// on the same value. If so, and if safe to do so, fold them together.
854 bool SimplifyCFGOpt::FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
855 IRBuilder<> &Builder) {
856 BasicBlock *BB = TI->getParent();
857 Value *CV = isValueEqualityComparison(TI); // CondVal
858 assert(CV && "Not a comparison?");
859 bool Changed = false;
861 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
862 while (!Preds.empty()) {
863 BasicBlock *Pred = Preds.pop_back_val();
865 // See if the predecessor is a comparison with the same value.
866 TerminatorInst *PTI = Pred->getTerminator();
867 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
869 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
870 // Figure out which 'cases' to copy from SI to PSI.
871 std::vector<ValueEqualityComparisonCase> BBCases;
872 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
874 std::vector<ValueEqualityComparisonCase> PredCases;
875 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
877 // Based on whether the default edge from PTI goes to BB or not, fill in
878 // PredCases and PredDefault with the new switch cases we would like to
880 SmallVector<BasicBlock*, 8> NewSuccessors;
882 // Update the branch weight metadata along the way
883 SmallVector<uint64_t, 8> Weights;
884 bool PredHasWeights = HasBranchWeights(PTI);
885 bool SuccHasWeights = HasBranchWeights(TI);
887 if (PredHasWeights) {
888 GetBranchWeights(PTI, Weights);
889 // branch-weight metadata is inconsistent here.
890 if (Weights.size() != 1 + PredCases.size())
891 PredHasWeights = SuccHasWeights = false;
892 } else if (SuccHasWeights)
893 // If there are no predecessor weights but there are successor weights,
894 // populate Weights with 1, which will later be scaled to the sum of
895 // successor's weights
896 Weights.assign(1 + PredCases.size(), 1);
898 SmallVector<uint64_t, 8> SuccWeights;
899 if (SuccHasWeights) {
900 GetBranchWeights(TI, SuccWeights);
901 // branch-weight metadata is inconsistent here.
902 if (SuccWeights.size() != 1 + BBCases.size())
903 PredHasWeights = SuccHasWeights = false;
904 } else if (PredHasWeights)
905 SuccWeights.assign(1 + BBCases.size(), 1);
907 if (PredDefault == BB) {
908 // If this is the default destination from PTI, only the edges in TI
909 // that don't occur in PTI, or that branch to BB will be activated.
910 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
911 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
912 if (PredCases[i].Dest != BB)
913 PTIHandled.insert(PredCases[i].Value);
915 // The default destination is BB, we don't need explicit targets.
916 std::swap(PredCases[i], PredCases.back());
918 if (PredHasWeights || SuccHasWeights) {
919 // Increase weight for the default case.
920 Weights[0] += Weights[i+1];
921 std::swap(Weights[i+1], Weights.back());
925 PredCases.pop_back();
929 // Reconstruct the new switch statement we will be building.
930 if (PredDefault != BBDefault) {
931 PredDefault->removePredecessor(Pred);
932 PredDefault = BBDefault;
933 NewSuccessors.push_back(BBDefault);
936 unsigned CasesFromPred = Weights.size();
937 uint64_t ValidTotalSuccWeight = 0;
938 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
939 if (!PTIHandled.count(BBCases[i].Value) &&
940 BBCases[i].Dest != BBDefault) {
941 PredCases.push_back(BBCases[i]);
942 NewSuccessors.push_back(BBCases[i].Dest);
943 if (SuccHasWeights || PredHasWeights) {
944 // The default weight is at index 0, so weight for the ith case
945 // should be at index i+1. Scale the cases from successor by
946 // PredDefaultWeight (Weights[0]).
947 Weights.push_back(Weights[0] * SuccWeights[i+1]);
948 ValidTotalSuccWeight += SuccWeights[i+1];
952 if (SuccHasWeights || PredHasWeights) {
953 ValidTotalSuccWeight += SuccWeights[0];
954 // Scale the cases from predecessor by ValidTotalSuccWeight.
955 for (unsigned i = 1; i < CasesFromPred; ++i)
956 Weights[i] *= ValidTotalSuccWeight;
957 // Scale the default weight by SuccDefaultWeight (SuccWeights[0]).
958 Weights[0] *= SuccWeights[0];
961 // If this is not the default destination from PSI, only the edges
962 // in SI that occur in PSI with a destination of BB will be
964 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
965 std::map<ConstantInt*, uint64_t> WeightsForHandled;
966 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
967 if (PredCases[i].Dest == BB) {
968 PTIHandled.insert(PredCases[i].Value);
970 if (PredHasWeights || SuccHasWeights) {
971 WeightsForHandled[PredCases[i].Value] = Weights[i+1];
972 std::swap(Weights[i+1], Weights.back());
976 std::swap(PredCases[i], PredCases.back());
977 PredCases.pop_back();
981 // Okay, now we know which constants were sent to BB from the
982 // predecessor. Figure out where they will all go now.
983 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
984 if (PTIHandled.count(BBCases[i].Value)) {
985 // If this is one we are capable of getting...
986 if (PredHasWeights || SuccHasWeights)
987 Weights.push_back(WeightsForHandled[BBCases[i].Value]);
988 PredCases.push_back(BBCases[i]);
989 NewSuccessors.push_back(BBCases[i].Dest);
990 PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
993 // If there are any constants vectored to BB that TI doesn't handle,
994 // they must go to the default destination of TI.
995 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
997 E = PTIHandled.end(); I != E; ++I) {
998 if (PredHasWeights || SuccHasWeights)
999 Weights.push_back(WeightsForHandled[*I]);
1000 PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
1001 NewSuccessors.push_back(BBDefault);
1005 // Okay, at this point, we know which new successor Pred will get. Make
1006 // sure we update the number of entries in the PHI nodes for these
1008 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
1009 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
1011 Builder.SetInsertPoint(PTI);
1012 // Convert pointer to int before we switch.
1013 if (CV->getType()->isPointerTy()) {
1014 assert(TD && "Cannot switch on pointer without DataLayout");
1015 CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()),
1019 // Now that the successors are updated, create the new Switch instruction.
1020 SwitchInst *NewSI = Builder.CreateSwitch(CV, PredDefault,
1022 NewSI->setDebugLoc(PTI->getDebugLoc());
1023 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
1024 NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
1026 if (PredHasWeights || SuccHasWeights) {
1027 // Halve the weights if any of them cannot fit in an uint32_t
1028 FitWeights(Weights);
1030 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
1032 NewSI->setMetadata(LLVMContext::MD_prof,
1033 MDBuilder(BB->getContext()).
1034 createBranchWeights(MDWeights));
1037 EraseTerminatorInstAndDCECond(PTI);
1039 // Okay, last check. If BB is still a successor of PSI, then we must
1040 // have an infinite loop case. If so, add an infinitely looping block
1041 // to handle the case to preserve the behavior of the code.
1042 BasicBlock *InfLoopBlock = 0;
1043 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
1044 if (NewSI->getSuccessor(i) == BB) {
1045 if (InfLoopBlock == 0) {
1046 // Insert it at the end of the function, because it's either code,
1047 // or it won't matter if it's hot. :)
1048 InfLoopBlock = BasicBlock::Create(BB->getContext(),
1049 "infloop", BB->getParent());
1050 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1052 NewSI->setSuccessor(i, InfLoopBlock);
1061 // isSafeToHoistInvoke - If we would need to insert a select that uses the
1062 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
1063 // would need to do this), we can't hoist the invoke, as there is nowhere
1064 // to put the select in this case.
1065 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
1066 Instruction *I1, Instruction *I2) {
1067 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1069 for (BasicBlock::iterator BBI = SI->begin();
1070 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1071 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1072 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1073 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
1081 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
1082 /// BB2, hoist any common code in the two blocks up into the branch block. The
1083 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
1084 static bool HoistThenElseCodeToIf(BranchInst *BI) {
1085 // This does very trivial matching, with limited scanning, to find identical
1086 // instructions in the two blocks. In particular, we don't want to get into
1087 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
1088 // such, we currently just scan for obviously identical instructions in an
1090 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
1091 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
1093 BasicBlock::iterator BB1_Itr = BB1->begin();
1094 BasicBlock::iterator BB2_Itr = BB2->begin();
1096 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
1097 // Skip debug info if it is not identical.
1098 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1099 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1100 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1101 while (isa<DbgInfoIntrinsic>(I1))
1103 while (isa<DbgInfoIntrinsic>(I2))
1106 if (isa<PHINode>(I1) || !I1->isIdenticalToWhenDefined(I2) ||
1107 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
1110 BasicBlock *BIParent = BI->getParent();
1112 bool Changed = false;
1114 // If we are hoisting the terminator instruction, don't move one (making a
1115 // broken BB), instead clone it, and remove BI.
1116 if (isa<TerminatorInst>(I1))
1117 goto HoistTerminator;
1119 // For a normal instruction, we just move one to right before the branch,
1120 // then replace all uses of the other with the first. Finally, we remove
1121 // the now redundant second instruction.
1122 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
1123 if (!I2->use_empty())
1124 I2->replaceAllUsesWith(I1);
1125 I1->intersectOptionalDataWith(I2);
1126 I2->eraseFromParent();
1131 // Skip debug info if it is not identical.
1132 DbgInfoIntrinsic *DBI1 = dyn_cast<DbgInfoIntrinsic>(I1);
1133 DbgInfoIntrinsic *DBI2 = dyn_cast<DbgInfoIntrinsic>(I2);
1134 if (!DBI1 || !DBI2 || !DBI1->isIdenticalToWhenDefined(DBI2)) {
1135 while (isa<DbgInfoIntrinsic>(I1))
1137 while (isa<DbgInfoIntrinsic>(I2))
1140 } while (I1->isIdenticalToWhenDefined(I2));
1145 // It may not be possible to hoist an invoke.
1146 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
1149 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1151 for (BasicBlock::iterator BBI = SI->begin();
1152 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1153 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1154 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1158 if (isa<ConstantExpr>(BB1V) && !isSafeToSpeculativelyExecute(BB1V))
1160 if (isa<ConstantExpr>(BB2V) && !isSafeToSpeculativelyExecute(BB2V))
1165 // Okay, it is safe to hoist the terminator.
1166 Instruction *NT = I1->clone();
1167 BIParent->getInstList().insert(BI, NT);
1168 if (!NT->getType()->isVoidTy()) {
1169 I1->replaceAllUsesWith(NT);
1170 I2->replaceAllUsesWith(NT);
1174 IRBuilder<true, NoFolder> Builder(NT);
1175 // Hoisting one of the terminators from our successor is a great thing.
1176 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
1177 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
1178 // nodes, so we insert select instruction to compute the final result.
1179 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
1180 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
1182 for (BasicBlock::iterator BBI = SI->begin();
1183 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1184 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1185 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1186 if (BB1V == BB2V) continue;
1188 // These values do not agree. Insert a select instruction before NT
1189 // that determines the right value.
1190 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
1192 SI = cast<SelectInst>
1193 (Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
1194 BB1V->getName()+"."+BB2V->getName()));
1196 // Make the PHI node use the select for all incoming values for BB1/BB2
1197 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
1198 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
1199 PN->setIncomingValue(i, SI);
1203 // Update any PHI nodes in our new successors.
1204 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
1205 AddPredecessorToBlock(*SI, BIParent, BB1);
1207 EraseTerminatorInstAndDCECond(BI);
1211 /// SinkThenElseCodeToEnd - Given an unconditional branch that goes to BBEnd,
1212 /// check whether BBEnd has only two predecessors and the other predecessor
1213 /// ends with an unconditional branch. If it is true, sink any common code
1214 /// in the two predecessors to BBEnd.
1215 static bool SinkThenElseCodeToEnd(BranchInst *BI1) {
1216 assert(BI1->isUnconditional());
1217 BasicBlock *BB1 = BI1->getParent();
1218 BasicBlock *BBEnd = BI1->getSuccessor(0);
1220 // Check that BBEnd has two predecessors and the other predecessor ends with
1221 // an unconditional branch.
1222 pred_iterator PI = pred_begin(BBEnd), PE = pred_end(BBEnd);
1223 BasicBlock *Pred0 = *PI++;
1224 if (PI == PE) // Only one predecessor.
1226 BasicBlock *Pred1 = *PI++;
1227 if (PI != PE) // More than two predecessors.
1229 BasicBlock *BB2 = (Pred0 == BB1) ? Pred1 : Pred0;
1230 BranchInst *BI2 = dyn_cast<BranchInst>(BB2->getTerminator());
1231 if (!BI2 || !BI2->isUnconditional())
1234 // Gather the PHI nodes in BBEnd.
1235 std::map<Value*, std::pair<Value*, PHINode*> > MapValueFromBB1ToBB2;
1236 Instruction *FirstNonPhiInBBEnd = 0;
1237 for (BasicBlock::iterator I = BBEnd->begin(), E = BBEnd->end();
1239 if (PHINode *PN = dyn_cast<PHINode>(I)) {
1240 Value *BB1V = PN->getIncomingValueForBlock(BB1);
1241 Value *BB2V = PN->getIncomingValueForBlock(BB2);
1242 MapValueFromBB1ToBB2[BB1V] = std::make_pair(BB2V, PN);
1244 FirstNonPhiInBBEnd = &*I;
1248 if (!FirstNonPhiInBBEnd)
1252 // This does very trivial matching, with limited scanning, to find identical
1253 // instructions in the two blocks. We scan backward for obviously identical
1254 // instructions in an identical order.
1255 BasicBlock::InstListType::reverse_iterator RI1 = BB1->getInstList().rbegin(),
1256 RE1 = BB1->getInstList().rend(), RI2 = BB2->getInstList().rbegin(),
1257 RE2 = BB2->getInstList().rend();
1259 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1262 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1265 // Skip the unconditional branches.
1269 bool Changed = false;
1270 while (RI1 != RE1 && RI2 != RE2) {
1272 while (RI1 != RE1 && isa<DbgInfoIntrinsic>(&*RI1)) ++RI1;
1275 while (RI2 != RE2 && isa<DbgInfoIntrinsic>(&*RI2)) ++RI2;
1279 Instruction *I1 = &*RI1, *I2 = &*RI2;
1280 // I1 and I2 should have a single use in the same PHI node, and they
1281 // perform the same operation.
1282 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
1283 if (isa<PHINode>(I1) || isa<PHINode>(I2) ||
1284 isa<TerminatorInst>(I1) || isa<TerminatorInst>(I2) ||
1285 isa<LandingPadInst>(I1) || isa<LandingPadInst>(I2) ||
1286 isa<AllocaInst>(I1) || isa<AllocaInst>(I2) ||
1287 I1->mayHaveSideEffects() || I2->mayHaveSideEffects() ||
1288 I1->mayReadOrWriteMemory() || I2->mayReadOrWriteMemory() ||
1289 !I1->hasOneUse() || !I2->hasOneUse() ||
1290 MapValueFromBB1ToBB2.find(I1) == MapValueFromBB1ToBB2.end() ||
1291 MapValueFromBB1ToBB2[I1].first != I2)
1294 // Check whether we should swap the operands of ICmpInst.
1295 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(I1), *ICmp2 = dyn_cast<ICmpInst>(I2);
1296 bool SwapOpnds = false;
1297 if (ICmp1 && ICmp2 &&
1298 ICmp1->getOperand(0) != ICmp2->getOperand(0) &&
1299 ICmp1->getOperand(1) != ICmp2->getOperand(1) &&
1300 (ICmp1->getOperand(0) == ICmp2->getOperand(1) ||
1301 ICmp1->getOperand(1) == ICmp2->getOperand(0))) {
1302 ICmp2->swapOperands();
1305 if (!I1->isSameOperationAs(I2)) {
1307 ICmp2->swapOperands();
1311 // The operands should be either the same or they need to be generated
1312 // with a PHI node after sinking. We only handle the case where there is
1313 // a single pair of different operands.
1314 Value *DifferentOp1 = 0, *DifferentOp2 = 0;
1315 unsigned Op1Idx = 0;
1316 for (unsigned I = 0, E = I1->getNumOperands(); I != E; ++I) {
1317 if (I1->getOperand(I) == I2->getOperand(I))
1319 // Early exit if we have more-than one pair of different operands or
1320 // the different operand is already in MapValueFromBB1ToBB2.
1321 // Early exit if we need a PHI node to replace a constant.
1323 MapValueFromBB1ToBB2.find(I1->getOperand(I)) !=
1324 MapValueFromBB1ToBB2.end() ||
1325 isa<Constant>(I1->getOperand(I)) ||
1326 isa<Constant>(I2->getOperand(I))) {
1327 // If we can't sink the instructions, undo the swapping.
1329 ICmp2->swapOperands();
1332 DifferentOp1 = I1->getOperand(I);
1334 DifferentOp2 = I2->getOperand(I);
1337 // We insert the pair of different operands to MapValueFromBB1ToBB2 and
1338 // remove (I1, I2) from MapValueFromBB1ToBB2.
1340 PHINode *NewPN = PHINode::Create(DifferentOp1->getType(), 2,
1341 DifferentOp1->getName() + ".sink",
1343 MapValueFromBB1ToBB2[DifferentOp1] = std::make_pair(DifferentOp2, NewPN);
1344 // I1 should use NewPN instead of DifferentOp1.
1345 I1->setOperand(Op1Idx, NewPN);
1346 NewPN->addIncoming(DifferentOp1, BB1);
1347 NewPN->addIncoming(DifferentOp2, BB2);
1348 DEBUG(dbgs() << "Create PHI node " << *NewPN << "\n";);
1350 PHINode *OldPN = MapValueFromBB1ToBB2[I1].second;
1351 MapValueFromBB1ToBB2.erase(I1);
1353 DEBUG(dbgs() << "SINK common instructions " << *I1 << "\n";);
1354 DEBUG(dbgs() << " " << *I2 << "\n";);
1355 // We need to update RE1 and RE2 if we are going to sink the first
1356 // instruction in the basic block down.
1357 bool UpdateRE1 = (I1 == BB1->begin()), UpdateRE2 = (I2 == BB2->begin());
1358 // Sink the instruction.
1359 BBEnd->getInstList().splice(FirstNonPhiInBBEnd, BB1->getInstList(), I1);
1360 if (!OldPN->use_empty())
1361 OldPN->replaceAllUsesWith(I1);
1362 OldPN->eraseFromParent();
1364 if (!I2->use_empty())
1365 I2->replaceAllUsesWith(I1);
1366 I1->intersectOptionalDataWith(I2);
1367 I2->eraseFromParent();
1370 RE1 = BB1->getInstList().rend();
1372 RE2 = BB2->getInstList().rend();
1373 FirstNonPhiInBBEnd = I1;
1380 /// \brief Determine if we can hoist sink a sole store instruction out of a
1381 /// conditional block.
1383 /// We are looking for code like the following:
1385 /// store i32 %add, i32* %arrayidx2
1386 /// ... // No other stores or function calls (we could be calling a memory
1387 /// ... // function).
1388 /// %cmp = icmp ult %x, %y
1389 /// br i1 %cmp, label %EndBB, label %ThenBB
1391 /// store i32 %add5, i32* %arrayidx2
1395 /// We are going to transform this into:
1397 /// store i32 %add, i32* %arrayidx2
1399 /// %cmp = icmp ult %x, %y
1400 /// %add.add5 = select i1 %cmp, i32 %add, %add5
1401 /// store i32 %add.add5, i32* %arrayidx2
1404 /// \return The pointer to the value of the previous store if the store can be
1405 /// hoisted into the predecessor block. 0 otherwise.
1406 static Value *isSafeToSpeculateStore(Instruction *I, BasicBlock *BrBB,
1407 BasicBlock *StoreBB, BasicBlock *EndBB) {
1408 StoreInst *StoreToHoist = dyn_cast<StoreInst>(I);
1412 // Volatile or atomic.
1413 if (!StoreToHoist->isSimple())
1416 Value *StorePtr = StoreToHoist->getPointerOperand();
1418 // Look for a store to the same pointer in BrBB.
1419 unsigned MaxNumInstToLookAt = 10;
1420 for (BasicBlock::reverse_iterator RI = BrBB->rbegin(),
1421 RE = BrBB->rend(); RI != RE && (--MaxNumInstToLookAt); ++RI) {
1422 Instruction *CurI = &*RI;
1424 // Could be calling an instruction that effects memory like free().
1425 if (CurI->mayHaveSideEffects() && !isa<StoreInst>(CurI))
1428 StoreInst *SI = dyn_cast<StoreInst>(CurI);
1429 // Found the previous store make sure it stores to the same location.
1430 if (SI && SI->getPointerOperand() == StorePtr)
1431 // Found the previous store, return its value operand.
1432 return SI->getValueOperand();
1434 return 0; // Unknown store.
1440 /// \brief Speculate a conditional basic block flattening the CFG.
1442 /// Note that this is a very risky transform currently. Speculating
1443 /// instructions like this is most often not desirable. Instead, there is an MI
1444 /// pass which can do it with full awareness of the resource constraints.
1445 /// However, some cases are "obvious" and we should do directly. An example of
1446 /// this is speculating a single, reasonably cheap instruction.
1448 /// There is only one distinct advantage to flattening the CFG at the IR level:
1449 /// it makes very common but simplistic optimizations such as are common in
1450 /// instcombine and the DAG combiner more powerful by removing CFG edges and
1451 /// modeling their effects with easier to reason about SSA value graphs.
1454 /// An illustration of this transform is turning this IR:
1457 /// %cmp = icmp ult %x, %y
1458 /// br i1 %cmp, label %EndBB, label %ThenBB
1460 /// %sub = sub %x, %y
1463 /// %phi = phi [ %sub, %ThenBB ], [ 0, %EndBB ]
1470 /// %cmp = icmp ult %x, %y
1471 /// %sub = sub %x, %y
1472 /// %cond = select i1 %cmp, 0, %sub
1476 /// \returns true if the conditional block is removed.
1477 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *ThenBB) {
1478 // Be conservative for now. FP select instruction can often be expensive.
1479 Value *BrCond = BI->getCondition();
1480 if (isa<FCmpInst>(BrCond))
1483 BasicBlock *BB = BI->getParent();
1484 BasicBlock *EndBB = ThenBB->getTerminator()->getSuccessor(0);
1486 // If ThenBB is actually on the false edge of the conditional branch, remember
1487 // to swap the select operands later.
1488 bool Invert = false;
1489 if (ThenBB != BI->getSuccessor(0)) {
1490 assert(ThenBB == BI->getSuccessor(1) && "No edge from 'if' block?");
1493 assert(EndBB == BI->getSuccessor(!Invert) && "No edge from to end block");
1495 // Keep a count of how many times instructions are used within CondBB when
1496 // they are candidates for sinking into CondBB. Specifically:
1497 // - They are defined in BB, and
1498 // - They have no side effects, and
1499 // - All of their uses are in CondBB.
1500 SmallDenseMap<Instruction *, unsigned, 4> SinkCandidateUseCounts;
1502 unsigned SpeculationCost = 0;
1503 Value *SpeculatedStoreValue = 0;
1504 StoreInst *SpeculatedStore = 0;
1505 for (BasicBlock::iterator BBI = ThenBB->begin(),
1506 BBE = llvm::prior(ThenBB->end());
1507 BBI != BBE; ++BBI) {
1508 Instruction *I = BBI;
1510 if (isa<DbgInfoIntrinsic>(I))
1513 // Only speculatively execution a single instruction (not counting the
1514 // terminator) for now.
1516 if (SpeculationCost > 1)
1519 // Don't hoist the instruction if it's unsafe or expensive.
1520 if (!isSafeToSpeculativelyExecute(I) &&
1521 !(HoistCondStores &&
1522 (SpeculatedStoreValue = isSafeToSpeculateStore(I, BB, ThenBB,
1525 if (!SpeculatedStoreValue &&
1526 ComputeSpeculationCost(I) > PHINodeFoldingThreshold)
1529 // Store the store speculation candidate.
1530 if (SpeculatedStoreValue)
1531 SpeculatedStore = cast<StoreInst>(I);
1533 // Do not hoist the instruction if any of its operands are defined but not
1534 // used in BB. The transformation will prevent the operand from
1535 // being sunk into the use block.
1536 for (User::op_iterator i = I->op_begin(), e = I->op_end();
1538 Instruction *OpI = dyn_cast<Instruction>(*i);
1539 if (!OpI || OpI->getParent() != BB ||
1540 OpI->mayHaveSideEffects())
1541 continue; // Not a candidate for sinking.
1543 ++SinkCandidateUseCounts[OpI];
1547 // Consider any sink candidates which are only used in CondBB as costs for
1548 // speculation. Note, while we iterate over a DenseMap here, we are summing
1549 // and so iteration order isn't significant.
1550 for (SmallDenseMap<Instruction *, unsigned, 4>::iterator I =
1551 SinkCandidateUseCounts.begin(), E = SinkCandidateUseCounts.end();
1553 if (I->first->getNumUses() == I->second) {
1555 if (SpeculationCost > 1)
1559 // Check that the PHI nodes can be converted to selects.
1560 bool HaveRewritablePHIs = false;
1561 for (BasicBlock::iterator I = EndBB->begin();
1562 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1563 Value *OrigV = PN->getIncomingValueForBlock(BB);
1564 Value *ThenV = PN->getIncomingValueForBlock(ThenBB);
1566 // FIXME: Try to remove some of the duplication with HoistThenElseCodeToIf.
1567 // Skip PHIs which are trivial.
1571 HaveRewritablePHIs = true;
1572 ConstantExpr *OrigCE = dyn_cast<ConstantExpr>(OrigV);
1573 ConstantExpr *ThenCE = dyn_cast<ConstantExpr>(ThenV);
1574 if (!OrigCE && !ThenCE)
1575 continue; // Known safe and cheap.
1577 if ((ThenCE && !isSafeToSpeculativelyExecute(ThenCE)) ||
1578 (OrigCE && !isSafeToSpeculativelyExecute(OrigCE)))
1580 unsigned OrigCost = OrigCE ? ComputeSpeculationCost(OrigCE) : 0;
1581 unsigned ThenCost = ThenCE ? ComputeSpeculationCost(ThenCE) : 0;
1582 if (OrigCost + ThenCost > 2 * PHINodeFoldingThreshold)
1585 // Account for the cost of an unfolded ConstantExpr which could end up
1586 // getting expanded into Instructions.
1587 // FIXME: This doesn't account for how many operations are combined in the
1588 // constant expression.
1590 if (SpeculationCost > 1)
1594 // If there are no PHIs to process, bail early. This helps ensure idempotence
1596 if (!HaveRewritablePHIs && !(HoistCondStores && SpeculatedStoreValue))
1599 // If we get here, we can hoist the instruction and if-convert.
1600 DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *ThenBB << "\n";);
1602 // Insert a select of the value of the speculated store.
1603 if (SpeculatedStoreValue) {
1604 IRBuilder<true, NoFolder> Builder(BI);
1605 Value *TrueV = SpeculatedStore->getValueOperand();
1606 Value *FalseV = SpeculatedStoreValue;
1608 std::swap(TrueV, FalseV);
1609 Value *S = Builder.CreateSelect(BrCond, TrueV, FalseV, TrueV->getName() +
1610 "." + FalseV->getName());
1611 SpeculatedStore->setOperand(0, S);
1614 // Hoist the instructions.
1615 BB->getInstList().splice(BI, ThenBB->getInstList(), ThenBB->begin(),
1616 llvm::prior(ThenBB->end()));
1618 // Insert selects and rewrite the PHI operands.
1619 IRBuilder<true, NoFolder> Builder(BI);
1620 for (BasicBlock::iterator I = EndBB->begin();
1621 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1622 unsigned OrigI = PN->getBasicBlockIndex(BB);
1623 unsigned ThenI = PN->getBasicBlockIndex(ThenBB);
1624 Value *OrigV = PN->getIncomingValue(OrigI);
1625 Value *ThenV = PN->getIncomingValue(ThenI);
1627 // Skip PHIs which are trivial.
1631 // Create a select whose true value is the speculatively executed value and
1632 // false value is the preexisting value. Swap them if the branch
1633 // destinations were inverted.
1634 Value *TrueV = ThenV, *FalseV = OrigV;
1636 std::swap(TrueV, FalseV);
1637 Value *V = Builder.CreateSelect(BrCond, TrueV, FalseV,
1638 TrueV->getName() + "." + FalseV->getName());
1639 PN->setIncomingValue(OrigI, V);
1640 PN->setIncomingValue(ThenI, V);
1647 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1648 /// across this block.
1649 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1650 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1653 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1654 if (isa<DbgInfoIntrinsic>(BBI))
1656 if (Size > 10) return false; // Don't clone large BB's.
1659 // We can only support instructions that do not define values that are
1660 // live outside of the current basic block.
1661 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1663 Instruction *U = cast<Instruction>(*UI);
1664 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1667 // Looks ok, continue checking.
1673 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1674 /// that is defined in the same block as the branch and if any PHI entries are
1675 /// constants, thread edges corresponding to that entry to be branches to their
1676 /// ultimate destination.
1677 static bool FoldCondBranchOnPHI(BranchInst *BI, const DataLayout *TD) {
1678 BasicBlock *BB = BI->getParent();
1679 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1680 // NOTE: we currently cannot transform this case if the PHI node is used
1681 // outside of the block.
1682 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1685 // Degenerate case of a single entry PHI.
1686 if (PN->getNumIncomingValues() == 1) {
1687 FoldSingleEntryPHINodes(PN->getParent());
1691 // Now we know that this block has multiple preds and two succs.
1692 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1694 // Okay, this is a simple enough basic block. See if any phi values are
1696 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1697 ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
1698 if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
1700 // Okay, we now know that all edges from PredBB should be revectored to
1701 // branch to RealDest.
1702 BasicBlock *PredBB = PN->getIncomingBlock(i);
1703 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1705 if (RealDest == BB) continue; // Skip self loops.
1706 // Skip if the predecessor's terminator is an indirect branch.
1707 if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
1709 // The dest block might have PHI nodes, other predecessors and other
1710 // difficult cases. Instead of being smart about this, just insert a new
1711 // block that jumps to the destination block, effectively splitting
1712 // the edge we are about to create.
1713 BasicBlock *EdgeBB = BasicBlock::Create(BB->getContext(),
1714 RealDest->getName()+".critedge",
1715 RealDest->getParent(), RealDest);
1716 BranchInst::Create(RealDest, EdgeBB);
1718 // Update PHI nodes.
1719 AddPredecessorToBlock(RealDest, EdgeBB, BB);
1721 // BB may have instructions that are being threaded over. Clone these
1722 // instructions into EdgeBB. We know that there will be no uses of the
1723 // cloned instructions outside of EdgeBB.
1724 BasicBlock::iterator InsertPt = EdgeBB->begin();
1725 DenseMap<Value*, Value*> TranslateMap; // Track translated values.
1726 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1727 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1728 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1731 // Clone the instruction.
1732 Instruction *N = BBI->clone();
1733 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1735 // Update operands due to translation.
1736 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1738 DenseMap<Value*, Value*>::iterator PI = TranslateMap.find(*i);
1739 if (PI != TranslateMap.end())
1743 // Check for trivial simplification.
1744 if (Value *V = SimplifyInstruction(N, TD)) {
1745 TranslateMap[BBI] = V;
1746 delete N; // Instruction folded away, don't need actual inst
1748 // Insert the new instruction into its new home.
1749 EdgeBB->getInstList().insert(InsertPt, N);
1750 if (!BBI->use_empty())
1751 TranslateMap[BBI] = N;
1755 // Loop over all of the edges from PredBB to BB, changing them to branch
1756 // to EdgeBB instead.
1757 TerminatorInst *PredBBTI = PredBB->getTerminator();
1758 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1759 if (PredBBTI->getSuccessor(i) == BB) {
1760 BB->removePredecessor(PredBB);
1761 PredBBTI->setSuccessor(i, EdgeBB);
1764 // Recurse, simplifying any other constants.
1765 return FoldCondBranchOnPHI(BI, TD) | true;
1771 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1772 /// PHI node, see if we can eliminate it.
1773 static bool FoldTwoEntryPHINode(PHINode *PN, const DataLayout *TD) {
1774 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1775 // statement", which has a very simple dominance structure. Basically, we
1776 // are trying to find the condition that is being branched on, which
1777 // subsequently causes this merge to happen. We really want control
1778 // dependence information for this check, but simplifycfg can't keep it up
1779 // to date, and this catches most of the cases we care about anyway.
1780 BasicBlock *BB = PN->getParent();
1781 BasicBlock *IfTrue, *IfFalse;
1782 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1784 // Don't bother if the branch will be constant folded trivially.
1785 isa<ConstantInt>(IfCond))
1788 // Okay, we found that we can merge this two-entry phi node into a select.
1789 // Doing so would require us to fold *all* two entry phi nodes in this block.
1790 // At some point this becomes non-profitable (particularly if the target
1791 // doesn't support cmov's). Only do this transformation if there are two or
1792 // fewer PHI nodes in this block.
1793 unsigned NumPhis = 0;
1794 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1798 // Loop over the PHI's seeing if we can promote them all to select
1799 // instructions. While we are at it, keep track of the instructions
1800 // that need to be moved to the dominating block.
1801 SmallPtrSet<Instruction*, 4> AggressiveInsts;
1802 unsigned MaxCostVal0 = PHINodeFoldingThreshold,
1803 MaxCostVal1 = PHINodeFoldingThreshold;
1805 for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
1806 PHINode *PN = cast<PHINode>(II++);
1807 if (Value *V = SimplifyInstruction(PN, TD)) {
1808 PN->replaceAllUsesWith(V);
1809 PN->eraseFromParent();
1813 if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
1815 !DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
1820 // If we folded the first phi, PN dangles at this point. Refresh it. If
1821 // we ran out of PHIs then we simplified them all.
1822 PN = dyn_cast<PHINode>(BB->begin());
1823 if (PN == 0) return true;
1825 // Don't fold i1 branches on PHIs which contain binary operators. These can
1826 // often be turned into switches and other things.
1827 if (PN->getType()->isIntegerTy(1) &&
1828 (isa<BinaryOperator>(PN->getIncomingValue(0)) ||
1829 isa<BinaryOperator>(PN->getIncomingValue(1)) ||
1830 isa<BinaryOperator>(IfCond)))
1833 // If we all PHI nodes are promotable, check to make sure that all
1834 // instructions in the predecessor blocks can be promoted as well. If
1835 // not, we won't be able to get rid of the control flow, so it's not
1836 // worth promoting to select instructions.
1837 BasicBlock *DomBlock = 0;
1838 BasicBlock *IfBlock1 = PN->getIncomingBlock(0);
1839 BasicBlock *IfBlock2 = PN->getIncomingBlock(1);
1840 if (cast<BranchInst>(IfBlock1->getTerminator())->isConditional()) {
1843 DomBlock = *pred_begin(IfBlock1);
1844 for (BasicBlock::iterator I = IfBlock1->begin();!isa<TerminatorInst>(I);++I)
1845 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1846 // This is not an aggressive instruction that we can promote.
1847 // Because of this, we won't be able to get rid of the control
1848 // flow, so the xform is not worth it.
1853 if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
1856 DomBlock = *pred_begin(IfBlock2);
1857 for (BasicBlock::iterator I = IfBlock2->begin();!isa<TerminatorInst>(I);++I)
1858 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1859 // This is not an aggressive instruction that we can promote.
1860 // Because of this, we won't be able to get rid of the control
1861 // flow, so the xform is not worth it.
1866 DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1867 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1869 // If we can still promote the PHI nodes after this gauntlet of tests,
1870 // do all of the PHI's now.
1871 Instruction *InsertPt = DomBlock->getTerminator();
1872 IRBuilder<true, NoFolder> Builder(InsertPt);
1874 // Move all 'aggressive' instructions, which are defined in the
1875 // conditional parts of the if's up to the dominating block.
1877 DomBlock->getInstList().splice(InsertPt,
1878 IfBlock1->getInstList(), IfBlock1->begin(),
1879 IfBlock1->getTerminator());
1881 DomBlock->getInstList().splice(InsertPt,
1882 IfBlock2->getInstList(), IfBlock2->begin(),
1883 IfBlock2->getTerminator());
1885 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1886 // Change the PHI node into a select instruction.
1887 Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1888 Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1891 cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
1892 PN->replaceAllUsesWith(NV);
1894 PN->eraseFromParent();
1897 // At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
1898 // has been flattened. Change DomBlock to jump directly to our new block to
1899 // avoid other simplifycfg's kicking in on the diamond.
1900 TerminatorInst *OldTI = DomBlock->getTerminator();
1901 Builder.SetInsertPoint(OldTI);
1902 Builder.CreateBr(BB);
1903 OldTI->eraseFromParent();
1907 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1908 /// to two returning blocks, try to merge them together into one return,
1909 /// introducing a select if the return values disagree.
1910 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
1911 IRBuilder<> &Builder) {
1912 assert(BI->isConditional() && "Must be a conditional branch");
1913 BasicBlock *TrueSucc = BI->getSuccessor(0);
1914 BasicBlock *FalseSucc = BI->getSuccessor(1);
1915 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1916 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1918 // Check to ensure both blocks are empty (just a return) or optionally empty
1919 // with PHI nodes. If there are other instructions, merging would cause extra
1920 // computation on one path or the other.
1921 if (!TrueSucc->getFirstNonPHIOrDbg()->isTerminator())
1923 if (!FalseSucc->getFirstNonPHIOrDbg()->isTerminator())
1926 Builder.SetInsertPoint(BI);
1927 // Okay, we found a branch that is going to two return nodes. If
1928 // there is no return value for this function, just change the
1929 // branch into a return.
1930 if (FalseRet->getNumOperands() == 0) {
1931 TrueSucc->removePredecessor(BI->getParent());
1932 FalseSucc->removePredecessor(BI->getParent());
1933 Builder.CreateRetVoid();
1934 EraseTerminatorInstAndDCECond(BI);
1938 // Otherwise, figure out what the true and false return values are
1939 // so we can insert a new select instruction.
1940 Value *TrueValue = TrueRet->getReturnValue();
1941 Value *FalseValue = FalseRet->getReturnValue();
1943 // Unwrap any PHI nodes in the return blocks.
1944 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1945 if (TVPN->getParent() == TrueSucc)
1946 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1947 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1948 if (FVPN->getParent() == FalseSucc)
1949 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1951 // In order for this transformation to be safe, we must be able to
1952 // unconditionally execute both operands to the return. This is
1953 // normally the case, but we could have a potentially-trapping
1954 // constant expression that prevents this transformation from being
1956 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1959 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1963 // Okay, we collected all the mapped values and checked them for sanity, and
1964 // defined to really do this transformation. First, update the CFG.
1965 TrueSucc->removePredecessor(BI->getParent());
1966 FalseSucc->removePredecessor(BI->getParent());
1968 // Insert select instructions where needed.
1969 Value *BrCond = BI->getCondition();
1971 // Insert a select if the results differ.
1972 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1973 } else if (isa<UndefValue>(TrueValue)) {
1974 TrueValue = FalseValue;
1976 TrueValue = Builder.CreateSelect(BrCond, TrueValue,
1977 FalseValue, "retval");
1981 Value *RI = !TrueValue ?
1982 Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
1986 DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1987 << "\n " << *BI << "NewRet = " << *RI
1988 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
1990 EraseTerminatorInstAndDCECond(BI);
1995 /// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
1996 /// probabilities of the branch taking each edge. Fills in the two APInt
1997 /// parameters and return true, or returns false if no or invalid metadata was
1999 static bool ExtractBranchMetadata(BranchInst *BI,
2000 uint64_t &ProbTrue, uint64_t &ProbFalse) {
2001 assert(BI->isConditional() &&
2002 "Looking for probabilities on unconditional branch?");
2003 MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
2004 if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
2005 ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
2006 ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
2007 if (!CITrue || !CIFalse) return false;
2008 ProbTrue = CITrue->getValue().getZExtValue();
2009 ProbFalse = CIFalse->getValue().getZExtValue();
2013 /// checkCSEInPredecessor - Return true if the given instruction is available
2014 /// in its predecessor block. If yes, the instruction will be removed.
2016 static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
2017 if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
2019 for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
2020 Instruction *PBI = &*I;
2021 // Check whether Inst and PBI generate the same value.
2022 if (Inst->isIdenticalTo(PBI)) {
2023 Inst->replaceAllUsesWith(PBI);
2024 Inst->eraseFromParent();
2031 /// FoldBranchToCommonDest - If this basic block is simple enough, and if a
2032 /// predecessor branches to us and one of our successors, fold the block into
2033 /// the predecessor and use logical operations to pick the right destination.
2034 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
2035 BasicBlock *BB = BI->getParent();
2037 Instruction *Cond = 0;
2038 if (BI->isConditional())
2039 Cond = dyn_cast<Instruction>(BI->getCondition());
2041 // For unconditional branch, check for a simple CFG pattern, where
2042 // BB has a single predecessor and BB's successor is also its predecessor's
2043 // successor. If such pattern exisits, check for CSE between BB and its
2045 if (BasicBlock *PB = BB->getSinglePredecessor())
2046 if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
2047 if (PBI->isConditional() &&
2048 (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
2049 BI->getSuccessor(0) == PBI->getSuccessor(1))) {
2050 for (BasicBlock::iterator I = BB->begin(), E = BB->end();
2052 Instruction *Curr = I++;
2053 if (isa<CmpInst>(Curr)) {
2057 // Quit if we can't remove this instruction.
2058 if (!checkCSEInPredecessor(Curr, PB))
2067 if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
2068 Cond->getParent() != BB || !Cond->hasOneUse())
2071 // Only allow this if the condition is a simple instruction that can be
2072 // executed unconditionally. It must be in the same block as the branch, and
2073 // must be at the front of the block.
2074 BasicBlock::iterator FrontIt = BB->front();
2076 // Ignore dbg intrinsics.
2077 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2079 // Allow a single instruction to be hoisted in addition to the compare
2080 // that feeds the branch. We later ensure that any values that _it_ uses
2081 // were also live in the predecessor, so that we don't unnecessarily create
2082 // register pressure or inhibit out-of-order execution.
2083 Instruction *BonusInst = 0;
2084 if (&*FrontIt != Cond &&
2085 FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
2086 isSafeToSpeculativelyExecute(FrontIt)) {
2087 BonusInst = &*FrontIt;
2090 // Ignore dbg intrinsics.
2091 while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
2094 // Only a single bonus inst is allowed.
2095 if (&*FrontIt != Cond)
2098 // Make sure the instruction after the condition is the cond branch.
2099 BasicBlock::iterator CondIt = Cond; ++CondIt;
2101 // Ingore dbg intrinsics.
2102 while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
2107 // Cond is known to be a compare or binary operator. Check to make sure that
2108 // neither operand is a potentially-trapping constant expression.
2109 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
2112 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
2116 // Finally, don't infinitely unroll conditional loops.
2117 BasicBlock *TrueDest = BI->getSuccessor(0);
2118 BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
2119 if (TrueDest == BB || FalseDest == BB)
2122 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2123 BasicBlock *PredBlock = *PI;
2124 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
2126 // Check that we have two conditional branches. If there is a PHI node in
2127 // the common successor, verify that the same value flows in from both
2129 SmallVector<PHINode*, 4> PHIs;
2130 if (PBI == 0 || PBI->isUnconditional() ||
2131 (BI->isConditional() &&
2132 !SafeToMergeTerminators(BI, PBI)) ||
2133 (!BI->isConditional() &&
2134 !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
2137 // Determine if the two branches share a common destination.
2138 Instruction::BinaryOps Opc = Instruction::BinaryOpsEnd;
2139 bool InvertPredCond = false;
2141 if (BI->isConditional()) {
2142 if (PBI->getSuccessor(0) == TrueDest)
2143 Opc = Instruction::Or;
2144 else if (PBI->getSuccessor(1) == FalseDest)
2145 Opc = Instruction::And;
2146 else if (PBI->getSuccessor(0) == FalseDest)
2147 Opc = Instruction::And, InvertPredCond = true;
2148 else if (PBI->getSuccessor(1) == TrueDest)
2149 Opc = Instruction::Or, InvertPredCond = true;
2153 if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
2157 // Ensure that any values used in the bonus instruction are also used
2158 // by the terminator of the predecessor. This means that those values
2159 // must already have been resolved, so we won't be inhibiting the
2160 // out-of-order core by speculating them earlier.
2162 // Collect the values used by the bonus inst
2163 SmallPtrSet<Value*, 4> UsedValues;
2164 for (Instruction::op_iterator OI = BonusInst->op_begin(),
2165 OE = BonusInst->op_end(); OI != OE; ++OI) {
2167 if (!isa<Constant>(V))
2168 UsedValues.insert(V);
2171 SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
2172 Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
2174 // Walk up to four levels back up the use-def chain of the predecessor's
2175 // terminator to see if all those values were used. The choice of four
2176 // levels is arbitrary, to provide a compile-time-cost bound.
2177 while (!Worklist.empty()) {
2178 std::pair<Value*, unsigned> Pair = Worklist.back();
2179 Worklist.pop_back();
2181 if (Pair.second >= 4) continue;
2182 UsedValues.erase(Pair.first);
2183 if (UsedValues.empty()) break;
2185 if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
2186 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
2188 Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
2192 if (!UsedValues.empty()) return false;
2195 DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
2196 IRBuilder<> Builder(PBI);
2198 // If we need to invert the condition in the pred block to match, do so now.
2199 if (InvertPredCond) {
2200 Value *NewCond = PBI->getCondition();
2202 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
2203 CmpInst *CI = cast<CmpInst>(NewCond);
2204 CI->setPredicate(CI->getInversePredicate());
2206 NewCond = Builder.CreateNot(NewCond,
2207 PBI->getCondition()->getName()+".not");
2210 PBI->setCondition(NewCond);
2211 PBI->swapSuccessors();
2214 // If we have a bonus inst, clone it into the predecessor block.
2215 Instruction *NewBonus = 0;
2217 NewBonus = BonusInst->clone();
2218 PredBlock->getInstList().insert(PBI, NewBonus);
2219 NewBonus->takeName(BonusInst);
2220 BonusInst->setName(BonusInst->getName()+".old");
2223 // Clone Cond into the predecessor basic block, and or/and the
2224 // two conditions together.
2225 Instruction *New = Cond->clone();
2226 if (BonusInst) New->replaceUsesOfWith(BonusInst, NewBonus);
2227 PredBlock->getInstList().insert(PBI, New);
2228 New->takeName(Cond);
2229 Cond->setName(New->getName()+".old");
2231 if (BI->isConditional()) {
2232 Instruction *NewCond =
2233 cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
2235 PBI->setCondition(NewCond);
2237 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2238 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2240 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2242 SmallVector<uint64_t, 8> NewWeights;
2244 if (PBI->getSuccessor(0) == BB) {
2245 if (PredHasWeights && SuccHasWeights) {
2246 // PBI: br i1 %x, BB, FalseDest
2247 // BI: br i1 %y, TrueDest, FalseDest
2248 //TrueWeight is TrueWeight for PBI * TrueWeight for BI.
2249 NewWeights.push_back(PredTrueWeight * SuccTrueWeight);
2250 //FalseWeight is FalseWeight for PBI * TotalWeight for BI +
2251 // TrueWeight for PBI * FalseWeight for BI.
2252 // We assume that total weights of a BranchInst can fit into 32 bits.
2253 // Therefore, we will not have overflow using 64-bit arithmetic.
2254 NewWeights.push_back(PredFalseWeight * (SuccFalseWeight +
2255 SuccTrueWeight) + PredTrueWeight * SuccFalseWeight);
2257 AddPredecessorToBlock(TrueDest, PredBlock, BB);
2258 PBI->setSuccessor(0, TrueDest);
2260 if (PBI->getSuccessor(1) == BB) {
2261 if (PredHasWeights && SuccHasWeights) {
2262 // PBI: br i1 %x, TrueDest, BB
2263 // BI: br i1 %y, TrueDest, FalseDest
2264 //TrueWeight is TrueWeight for PBI * TotalWeight for BI +
2265 // FalseWeight for PBI * TrueWeight for BI.
2266 NewWeights.push_back(PredTrueWeight * (SuccFalseWeight +
2267 SuccTrueWeight) + PredFalseWeight * SuccTrueWeight);
2268 //FalseWeight is FalseWeight for PBI * FalseWeight for BI.
2269 NewWeights.push_back(PredFalseWeight * SuccFalseWeight);
2271 AddPredecessorToBlock(FalseDest, PredBlock, BB);
2272 PBI->setSuccessor(1, FalseDest);
2274 if (NewWeights.size() == 2) {
2275 // Halve the weights if any of them cannot fit in an uint32_t
2276 FitWeights(NewWeights);
2278 SmallVector<uint32_t, 8> MDWeights(NewWeights.begin(),NewWeights.end());
2279 PBI->setMetadata(LLVMContext::MD_prof,
2280 MDBuilder(BI->getContext()).
2281 createBranchWeights(MDWeights));
2283 PBI->setMetadata(LLVMContext::MD_prof, NULL);
2285 // Update PHI nodes in the common successors.
2286 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
2287 ConstantInt *PBI_C = cast<ConstantInt>(
2288 PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
2289 assert(PBI_C->getType()->isIntegerTy(1));
2290 Instruction *MergedCond = 0;
2291 if (PBI->getSuccessor(0) == TrueDest) {
2292 // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
2293 // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
2294 // is false: !PBI_Cond and BI_Value
2295 Instruction *NotCond =
2296 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2299 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2304 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2305 PBI->getCondition(), MergedCond,
2308 // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
2309 // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
2310 // is false: PBI_Cond and BI_Value
2312 cast<Instruction>(Builder.CreateBinOp(Instruction::And,
2313 PBI->getCondition(), New,
2315 if (PBI_C->isOne()) {
2316 Instruction *NotCond =
2317 cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
2320 cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
2321 NotCond, MergedCond,
2326 PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
2329 // Change PBI from Conditional to Unconditional.
2330 BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
2331 EraseTerminatorInstAndDCECond(PBI);
2335 // TODO: If BB is reachable from all paths through PredBlock, then we
2336 // could replace PBI's branch probabilities with BI's.
2338 // Copy any debug value intrinsics into the end of PredBlock.
2339 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
2340 if (isa<DbgInfoIntrinsic>(*I))
2341 I->clone()->insertBefore(PBI);
2348 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
2349 /// predecessor of another block, this function tries to simplify it. We know
2350 /// that PBI and BI are both conditional branches, and BI is in one of the
2351 /// successor blocks of PBI - PBI branches to BI.
2352 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
2353 assert(PBI->isConditional() && BI->isConditional());
2354 BasicBlock *BB = BI->getParent();
2356 // If this block ends with a branch instruction, and if there is a
2357 // predecessor that ends on a branch of the same condition, make
2358 // this conditional branch redundant.
2359 if (PBI->getCondition() == BI->getCondition() &&
2360 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2361 // Okay, the outcome of this conditional branch is statically
2362 // knowable. If this block had a single pred, handle specially.
2363 if (BB->getSinglePredecessor()) {
2364 // Turn this into a branch on constant.
2365 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2366 BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2368 return true; // Nuke the branch on constant.
2371 // Otherwise, if there are multiple predecessors, insert a PHI that merges
2372 // in the constant and simplify the block result. Subsequent passes of
2373 // simplifycfg will thread the block.
2374 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
2375 pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
2376 PHINode *NewPN = PHINode::Create(Type::getInt1Ty(BB->getContext()),
2377 std::distance(PB, PE),
2378 BI->getCondition()->getName() + ".pr",
2380 // Okay, we're going to insert the PHI node. Since PBI is not the only
2381 // predecessor, compute the PHI'd conditional value for all of the preds.
2382 // Any predecessor where the condition is not computable we keep symbolic.
2383 for (pred_iterator PI = PB; PI != PE; ++PI) {
2384 BasicBlock *P = *PI;
2385 if ((PBI = dyn_cast<BranchInst>(P->getTerminator())) &&
2386 PBI != BI && PBI->isConditional() &&
2387 PBI->getCondition() == BI->getCondition() &&
2388 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
2389 bool CondIsTrue = PBI->getSuccessor(0) == BB;
2390 NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
2393 NewPN->addIncoming(BI->getCondition(), P);
2397 BI->setCondition(NewPN);
2402 // If this is a conditional branch in an empty block, and if any
2403 // predecessors is a conditional branch to one of our destinations,
2404 // fold the conditions into logical ops and one cond br.
2405 BasicBlock::iterator BBI = BB->begin();
2406 // Ignore dbg intrinsics.
2407 while (isa<DbgInfoIntrinsic>(BBI))
2413 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
2418 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
2420 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
2421 PBIOp = 0, BIOp = 1;
2422 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
2423 PBIOp = 1, BIOp = 0;
2424 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
2429 // Check to make sure that the other destination of this branch
2430 // isn't BB itself. If so, this is an infinite loop that will
2431 // keep getting unwound.
2432 if (PBI->getSuccessor(PBIOp) == BB)
2435 // Do not perform this transformation if it would require
2436 // insertion of a large number of select instructions. For targets
2437 // without predication/cmovs, this is a big pessimization.
2438 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
2440 unsigned NumPhis = 0;
2441 for (BasicBlock::iterator II = CommonDest->begin();
2442 isa<PHINode>(II); ++II, ++NumPhis)
2443 if (NumPhis > 2) // Disable this xform.
2446 // Finally, if everything is ok, fold the branches to logical ops.
2447 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
2449 DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
2450 << "AND: " << *BI->getParent());
2453 // If OtherDest *is* BB, then BB is a basic block with a single conditional
2454 // branch in it, where one edge (OtherDest) goes back to itself but the other
2455 // exits. We don't *know* that the program avoids the infinite loop
2456 // (even though that seems likely). If we do this xform naively, we'll end up
2457 // recursively unpeeling the loop. Since we know that (after the xform is
2458 // done) that the block *is* infinite if reached, we just make it an obviously
2459 // infinite loop with no cond branch.
2460 if (OtherDest == BB) {
2461 // Insert it at the end of the function, because it's either code,
2462 // or it won't matter if it's hot. :)
2463 BasicBlock *InfLoopBlock = BasicBlock::Create(BB->getContext(),
2464 "infloop", BB->getParent());
2465 BranchInst::Create(InfLoopBlock, InfLoopBlock);
2466 OtherDest = InfLoopBlock;
2469 DEBUG(dbgs() << *PBI->getParent()->getParent());
2471 // BI may have other predecessors. Because of this, we leave
2472 // it alone, but modify PBI.
2474 // Make sure we get to CommonDest on True&True directions.
2475 Value *PBICond = PBI->getCondition();
2476 IRBuilder<true, NoFolder> Builder(PBI);
2478 PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
2480 Value *BICond = BI->getCondition();
2482 BICond = Builder.CreateNot(BICond, BICond->getName()+".not");
2484 // Merge the conditions.
2485 Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
2487 // Modify PBI to branch on the new condition to the new dests.
2488 PBI->setCondition(Cond);
2489 PBI->setSuccessor(0, CommonDest);
2490 PBI->setSuccessor(1, OtherDest);
2492 // Update branch weight for PBI.
2493 uint64_t PredTrueWeight, PredFalseWeight, SuccTrueWeight, SuccFalseWeight;
2494 bool PredHasWeights = ExtractBranchMetadata(PBI, PredTrueWeight,
2496 bool SuccHasWeights = ExtractBranchMetadata(BI, SuccTrueWeight,
2498 if (PredHasWeights && SuccHasWeights) {
2499 uint64_t PredCommon = PBIOp ? PredFalseWeight : PredTrueWeight;
2500 uint64_t PredOther = PBIOp ?PredTrueWeight : PredFalseWeight;
2501 uint64_t SuccCommon = BIOp ? SuccFalseWeight : SuccTrueWeight;
2502 uint64_t SuccOther = BIOp ? SuccTrueWeight : SuccFalseWeight;
2503 // The weight to CommonDest should be PredCommon * SuccTotal +
2504 // PredOther * SuccCommon.
2505 // The weight to OtherDest should be PredOther * SuccOther.
2506 SmallVector<uint64_t, 2> NewWeights;
2507 NewWeights.push_back(PredCommon * (SuccCommon + SuccOther) +
2508 PredOther * SuccCommon);
2509 NewWeights.push_back(PredOther * SuccOther);
2510 // Halve the weights if any of them cannot fit in an uint32_t
2511 FitWeights(NewWeights);
2513 SmallVector<uint32_t, 2> MDWeights(NewWeights.begin(),NewWeights.end());
2514 PBI->setMetadata(LLVMContext::MD_prof,
2515 MDBuilder(BI->getContext()).
2516 createBranchWeights(MDWeights));
2519 // OtherDest may have phi nodes. If so, add an entry from PBI's
2520 // block that are identical to the entries for BI's block.
2521 AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
2523 // We know that the CommonDest already had an edge from PBI to
2524 // it. If it has PHIs though, the PHIs may have different
2525 // entries for BB and PBI's BB. If so, insert a select to make
2528 for (BasicBlock::iterator II = CommonDest->begin();
2529 (PN = dyn_cast<PHINode>(II)); ++II) {
2530 Value *BIV = PN->getIncomingValueForBlock(BB);
2531 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
2532 Value *PBIV = PN->getIncomingValue(PBBIdx);
2534 // Insert a select in PBI to pick the right value.
2535 Value *NV = cast<SelectInst>
2536 (Builder.CreateSelect(PBICond, PBIV, BIV, PBIV->getName()+".mux"));
2537 PN->setIncomingValue(PBBIdx, NV);
2541 DEBUG(dbgs() << "INTO: " << *PBI->getParent());
2542 DEBUG(dbgs() << *PBI->getParent()->getParent());
2544 // This basic block is probably dead. We know it has at least
2545 // one fewer predecessor.
2549 // SimplifyTerminatorOnSelect - Simplifies a terminator by replacing it with a
2550 // branch to TrueBB if Cond is true or to FalseBB if Cond is false.
2551 // Takes care of updating the successors and removing the old terminator.
2552 // Also makes sure not to introduce new successors by assuming that edges to
2553 // non-successor TrueBBs and FalseBBs aren't reachable.
2554 static bool SimplifyTerminatorOnSelect(TerminatorInst *OldTerm, Value *Cond,
2555 BasicBlock *TrueBB, BasicBlock *FalseBB,
2556 uint32_t TrueWeight,
2557 uint32_t FalseWeight){
2558 // Remove any superfluous successor edges from the CFG.
2559 // First, figure out which successors to preserve.
2560 // If TrueBB and FalseBB are equal, only try to preserve one copy of that
2562 BasicBlock *KeepEdge1 = TrueBB;
2563 BasicBlock *KeepEdge2 = TrueBB != FalseBB ? FalseBB : 0;
2565 // Then remove the rest.
2566 for (unsigned I = 0, E = OldTerm->getNumSuccessors(); I != E; ++I) {
2567 BasicBlock *Succ = OldTerm->getSuccessor(I);
2568 // Make sure only to keep exactly one copy of each edge.
2569 if (Succ == KeepEdge1)
2571 else if (Succ == KeepEdge2)
2574 Succ->removePredecessor(OldTerm->getParent());
2577 IRBuilder<> Builder(OldTerm);
2578 Builder.SetCurrentDebugLocation(OldTerm->getDebugLoc());
2580 // Insert an appropriate new terminator.
2581 if ((KeepEdge1 == 0) && (KeepEdge2 == 0)) {
2582 if (TrueBB == FalseBB)
2583 // We were only looking for one successor, and it was present.
2584 // Create an unconditional branch to it.
2585 Builder.CreateBr(TrueBB);
2587 // We found both of the successors we were looking for.
2588 // Create a conditional branch sharing the condition of the select.
2589 BranchInst *NewBI = Builder.CreateCondBr(Cond, TrueBB, FalseBB);
2590 if (TrueWeight != FalseWeight)
2591 NewBI->setMetadata(LLVMContext::MD_prof,
2592 MDBuilder(OldTerm->getContext()).
2593 createBranchWeights(TrueWeight, FalseWeight));
2595 } else if (KeepEdge1 && (KeepEdge2 || TrueBB == FalseBB)) {
2596 // Neither of the selected blocks were successors, so this
2597 // terminator must be unreachable.
2598 new UnreachableInst(OldTerm->getContext(), OldTerm);
2600 // One of the selected values was a successor, but the other wasn't.
2601 // Insert an unconditional branch to the one that was found;
2602 // the edge to the one that wasn't must be unreachable.
2604 // Only TrueBB was found.
2605 Builder.CreateBr(TrueBB);
2607 // Only FalseBB was found.
2608 Builder.CreateBr(FalseBB);
2611 EraseTerminatorInstAndDCECond(OldTerm);
2615 // SimplifySwitchOnSelect - Replaces
2616 // (switch (select cond, X, Y)) on constant X, Y
2617 // with a branch - conditional if X and Y lead to distinct BBs,
2618 // unconditional otherwise.
2619 static bool SimplifySwitchOnSelect(SwitchInst *SI, SelectInst *Select) {
2620 // Check for constant integer values in the select.
2621 ConstantInt *TrueVal = dyn_cast<ConstantInt>(Select->getTrueValue());
2622 ConstantInt *FalseVal = dyn_cast<ConstantInt>(Select->getFalseValue());
2623 if (!TrueVal || !FalseVal)
2626 // Find the relevant condition and destinations.
2627 Value *Condition = Select->getCondition();
2628 BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
2629 BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
2631 // Get weight for TrueBB and FalseBB.
2632 uint32_t TrueWeight = 0, FalseWeight = 0;
2633 SmallVector<uint64_t, 8> Weights;
2634 bool HasWeights = HasBranchWeights(SI);
2636 GetBranchWeights(SI, Weights);
2637 if (Weights.size() == 1 + SI->getNumCases()) {
2638 TrueWeight = (uint32_t)Weights[SI->findCaseValue(TrueVal).
2639 getSuccessorIndex()];
2640 FalseWeight = (uint32_t)Weights[SI->findCaseValue(FalseVal).
2641 getSuccessorIndex()];
2645 // Perform the actual simplification.
2646 return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB,
2647 TrueWeight, FalseWeight);
2650 // SimplifyIndirectBrOnSelect - Replaces
2651 // (indirectbr (select cond, blockaddress(@fn, BlockA),
2652 // blockaddress(@fn, BlockB)))
2654 // (br cond, BlockA, BlockB).
2655 static bool SimplifyIndirectBrOnSelect(IndirectBrInst *IBI, SelectInst *SI) {
2656 // Check that both operands of the select are block addresses.
2657 BlockAddress *TBA = dyn_cast<BlockAddress>(SI->getTrueValue());
2658 BlockAddress *FBA = dyn_cast<BlockAddress>(SI->getFalseValue());
2662 // Extract the actual blocks.
2663 BasicBlock *TrueBB = TBA->getBasicBlock();
2664 BasicBlock *FalseBB = FBA->getBasicBlock();
2666 // Perform the actual simplification.
2667 return SimplifyTerminatorOnSelect(IBI, SI->getCondition(), TrueBB, FalseBB,
2671 /// TryToSimplifyUncondBranchWithICmpInIt - This is called when we find an icmp
2672 /// instruction (a seteq/setne with a constant) as the only instruction in a
2673 /// block that ends with an uncond branch. We are looking for a very specific
2674 /// pattern that occurs when "A == 1 || A == 2 || A == 3" gets simplified. In
2675 /// this case, we merge the first two "or's of icmp" into a switch, but then the
2676 /// default value goes to an uncond block with a seteq in it, we get something
2679 /// switch i8 %A, label %DEFAULT [ i8 1, label %end i8 2, label %end ]
2681 /// %tmp = icmp eq i8 %A, 92
2684 /// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
2686 /// We prefer to split the edge to 'end' so that there is a true/false entry to
2687 /// the PHI, merging the third icmp into the switch.
2688 static bool TryToSimplifyUncondBranchWithICmpInIt(
2689 ICmpInst *ICI, IRBuilder<> &Builder, const TargetTransformInfo &TTI,
2690 const DataLayout *TD) {
2691 BasicBlock *BB = ICI->getParent();
2693 // If the block has any PHIs in it or the icmp has multiple uses, it is too
2695 if (isa<PHINode>(BB->begin()) || !ICI->hasOneUse()) return false;
2697 Value *V = ICI->getOperand(0);
2698 ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
2700 // The pattern we're looking for is where our only predecessor is a switch on
2701 // 'V' and this block is the default case for the switch. In this case we can
2702 // fold the compared value into the switch to simplify things.
2703 BasicBlock *Pred = BB->getSinglePredecessor();
2704 if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
2706 SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
2707 if (SI->getCondition() != V)
2710 // If BB is reachable on a non-default case, then we simply know the value of
2711 // V in this block. Substitute it and constant fold the icmp instruction
2713 if (SI->getDefaultDest() != BB) {
2714 ConstantInt *VVal = SI->findCaseDest(BB);
2715 assert(VVal && "Should have a unique destination value");
2716 ICI->setOperand(0, VVal);
2718 if (Value *V = SimplifyInstruction(ICI, TD)) {
2719 ICI->replaceAllUsesWith(V);
2720 ICI->eraseFromParent();
2722 // BB is now empty, so it is likely to simplify away.
2723 return SimplifyCFG(BB, TTI, TD) | true;
2726 // Ok, the block is reachable from the default dest. If the constant we're
2727 // comparing exists in one of the other edges, then we can constant fold ICI
2729 if (SI->findCaseValue(Cst) != SI->case_default()) {
2731 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2732 V = ConstantInt::getFalse(BB->getContext());
2734 V = ConstantInt::getTrue(BB->getContext());
2736 ICI->replaceAllUsesWith(V);
2737 ICI->eraseFromParent();
2738 // BB is now empty, so it is likely to simplify away.
2739 return SimplifyCFG(BB, TTI, TD) | true;
2742 // The use of the icmp has to be in the 'end' block, by the only PHI node in
2744 BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
2745 PHINode *PHIUse = dyn_cast<PHINode>(ICI->use_back());
2746 if (PHIUse == 0 || PHIUse != &SuccBlock->front() ||
2747 isa<PHINode>(++BasicBlock::iterator(PHIUse)))
2750 // If the icmp is a SETEQ, then the default dest gets false, the new edge gets
2752 Constant *DefaultCst = ConstantInt::getTrue(BB->getContext());
2753 Constant *NewCst = ConstantInt::getFalse(BB->getContext());
2755 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
2756 std::swap(DefaultCst, NewCst);
2758 // Replace ICI (which is used by the PHI for the default value) with true or
2759 // false depending on if it is EQ or NE.
2760 ICI->replaceAllUsesWith(DefaultCst);
2761 ICI->eraseFromParent();
2763 // Okay, the switch goes to this block on a default value. Add an edge from
2764 // the switch to the merge point on the compared value.
2765 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
2766 BB->getParent(), BB);
2767 SmallVector<uint64_t, 8> Weights;
2768 bool HasWeights = HasBranchWeights(SI);
2770 GetBranchWeights(SI, Weights);
2771 if (Weights.size() == 1 + SI->getNumCases()) {
2772 // Split weight for default case to case for "Cst".
2773 Weights[0] = (Weights[0]+1) >> 1;
2774 Weights.push_back(Weights[0]);
2776 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
2777 SI->setMetadata(LLVMContext::MD_prof,
2778 MDBuilder(SI->getContext()).
2779 createBranchWeights(MDWeights));
2782 SI->addCase(Cst, NewBB);
2784 // NewBB branches to the phi block, add the uncond branch and the phi entry.
2785 Builder.SetInsertPoint(NewBB);
2786 Builder.SetCurrentDebugLocation(SI->getDebugLoc());
2787 Builder.CreateBr(SuccBlock);
2788 PHIUse->addIncoming(NewCst, NewBB);
2792 /// SimplifyBranchOnICmpChain - The specified branch is a conditional branch.
2793 /// Check to see if it is branching on an or/and chain of icmp instructions, and
2794 /// fold it into a switch instruction if so.
2795 static bool SimplifyBranchOnICmpChain(BranchInst *BI, const DataLayout *TD,
2796 IRBuilder<> &Builder) {
2797 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
2798 if (Cond == 0) return false;
2801 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2802 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2803 // 'setne's and'ed together, collect them.
2805 std::vector<ConstantInt*> Values;
2806 bool TrueWhenEqual = true;
2807 Value *ExtraCase = 0;
2808 unsigned UsedICmps = 0;
2810 if (Cond->getOpcode() == Instruction::Or) {
2811 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
2813 } else if (Cond->getOpcode() == Instruction::And) {
2814 CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, false,
2816 TrueWhenEqual = false;
2819 // If we didn't have a multiply compared value, fail.
2820 if (CompVal == 0) return false;
2822 // Avoid turning single icmps into a switch.
2826 // There might be duplicate constants in the list, which the switch
2827 // instruction can't handle, remove them now.
2828 array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
2829 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2831 // If Extra was used, we require at least two switch values to do the
2832 // transformation. A switch with one value is just an cond branch.
2833 if (ExtraCase && Values.size() < 2) return false;
2835 // TODO: Preserve branch weight metadata, similarly to how
2836 // FoldValueComparisonIntoPredecessors preserves it.
2838 // Figure out which block is which destination.
2839 BasicBlock *DefaultBB = BI->getSuccessor(1);
2840 BasicBlock *EdgeBB = BI->getSuccessor(0);
2841 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2843 BasicBlock *BB = BI->getParent();
2845 DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
2846 << " cases into SWITCH. BB is:\n" << *BB);
2848 // If there are any extra values that couldn't be folded into the switch
2849 // then we evaluate them with an explicit branch first. Split the block
2850 // right before the condbr to handle it.
2852 BasicBlock *NewBB = BB->splitBasicBlock(BI, "switch.early.test");
2853 // Remove the uncond branch added to the old block.
2854 TerminatorInst *OldTI = BB->getTerminator();
2855 Builder.SetInsertPoint(OldTI);
2858 Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
2860 Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
2862 OldTI->eraseFromParent();
2864 // If there are PHI nodes in EdgeBB, then we need to add a new entry to them
2865 // for the edge we just added.
2866 AddPredecessorToBlock(EdgeBB, BB, NewBB);
2868 DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
2869 << "\nEXTRABB = " << *BB);
2873 Builder.SetInsertPoint(BI);
2874 // Convert pointer to int before we switch.
2875 if (CompVal->getType()->isPointerTy()) {
2876 assert(TD && "Cannot switch on pointer without DataLayout");
2877 CompVal = Builder.CreatePtrToInt(CompVal,
2878 TD->getIntPtrType(CompVal->getContext()),
2882 // Create the new switch instruction now.
2883 SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
2885 // Add all of the 'cases' to the switch instruction.
2886 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2887 New->addCase(Values[i], EdgeBB);
2889 // We added edges from PI to the EdgeBB. As such, if there were any
2890 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2891 // the number of edges added.
2892 for (BasicBlock::iterator BBI = EdgeBB->begin();
2893 isa<PHINode>(BBI); ++BBI) {
2894 PHINode *PN = cast<PHINode>(BBI);
2895 Value *InVal = PN->getIncomingValueForBlock(BB);
2896 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2897 PN->addIncoming(InVal, BB);
2900 // Erase the old branch instruction.
2901 EraseTerminatorInstAndDCECond(BI);
2903 DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
2907 bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
2908 // If this is a trivial landing pad that just continues unwinding the caught
2909 // exception then zap the landing pad, turning its invokes into calls.
2910 BasicBlock *BB = RI->getParent();
2911 LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
2912 if (RI->getValue() != LPInst)
2913 // Not a landing pad, or the resume is not unwinding the exception that
2914 // caused control to branch here.
2917 // Check that there are no other instructions except for debug intrinsics.
2918 BasicBlock::iterator I = LPInst, E = RI;
2920 if (!isa<DbgInfoIntrinsic>(I))
2923 // Turn all invokes that unwind here into calls and delete the basic block.
2924 bool InvokeRequiresTableEntry = false;
2925 bool Changed = false;
2926 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
2927 InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
2929 if (II->hasFnAttr(Attribute::UWTable)) {
2930 // Don't remove an `invoke' instruction if the ABI requires an entry into
2932 InvokeRequiresTableEntry = true;
2936 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
2938 // Insert a call instruction before the invoke.
2939 CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
2941 Call->setCallingConv(II->getCallingConv());
2942 Call->setAttributes(II->getAttributes());
2943 Call->setDebugLoc(II->getDebugLoc());
2945 // Anything that used the value produced by the invoke instruction now uses
2946 // the value produced by the call instruction. Note that we do this even
2947 // for void functions and calls with no uses so that the callgraph edge is
2949 II->replaceAllUsesWith(Call);
2950 BB->removePredecessor(II->getParent());
2952 // Insert a branch to the normal destination right before the invoke.
2953 BranchInst::Create(II->getNormalDest(), II);
2955 // Finally, delete the invoke instruction!
2956 II->eraseFromParent();
2960 if (!InvokeRequiresTableEntry)
2961 // The landingpad is now unreachable. Zap it.
2962 BB->eraseFromParent();
2967 bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
2968 BasicBlock *BB = RI->getParent();
2969 if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
2971 // Find predecessors that end with branches.
2972 SmallVector<BasicBlock*, 8> UncondBranchPreds;
2973 SmallVector<BranchInst*, 8> CondBranchPreds;
2974 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
2975 BasicBlock *P = *PI;
2976 TerminatorInst *PTI = P->getTerminator();
2977 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
2978 if (BI->isUnconditional())
2979 UncondBranchPreds.push_back(P);
2981 CondBranchPreds.push_back(BI);
2985 // If we found some, do the transformation!
2986 if (!UncondBranchPreds.empty() && DupRet) {
2987 while (!UncondBranchPreds.empty()) {
2988 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
2989 DEBUG(dbgs() << "FOLDING: " << *BB
2990 << "INTO UNCOND BRANCH PRED: " << *Pred);
2991 (void)FoldReturnIntoUncondBranch(RI, BB, Pred);
2994 // If we eliminated all predecessors of the block, delete the block now.
2995 if (pred_begin(BB) == pred_end(BB))
2996 // We know there are no successors, so just nuke the block.
2997 BB->eraseFromParent();
3002 // Check out all of the conditional branches going to this return
3003 // instruction. If any of them just select between returns, change the
3004 // branch itself into a select/return pair.
3005 while (!CondBranchPreds.empty()) {
3006 BranchInst *BI = CondBranchPreds.pop_back_val();
3008 // Check to see if the non-BB successor is also a return block.
3009 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
3010 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
3011 SimplifyCondBranchToTwoReturns(BI, Builder))
3017 bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
3018 BasicBlock *BB = UI->getParent();
3020 bool Changed = false;
3022 // If there are any instructions immediately before the unreachable that can
3023 // be removed, do so.
3024 while (UI != BB->begin()) {
3025 BasicBlock::iterator BBI = UI;
3027 // Do not delete instructions that can have side effects which might cause
3028 // the unreachable to not be reachable; specifically, calls and volatile
3029 // operations may have this effect.
3030 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
3032 if (BBI->mayHaveSideEffects()) {
3033 if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
3034 if (SI->isVolatile())
3036 } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
3037 if (LI->isVolatile())
3039 } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
3040 if (RMWI->isVolatile())
3042 } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
3043 if (CXI->isVolatile())
3045 } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
3046 !isa<LandingPadInst>(BBI)) {
3049 // Note that deleting LandingPad's here is in fact okay, although it
3050 // involves a bit of subtle reasoning. If this inst is a LandingPad,
3051 // all the predecessors of this block will be the unwind edges of Invokes,
3052 // and we can therefore guarantee this block will be erased.
3055 // Delete this instruction (any uses are guaranteed to be dead)
3056 if (!BBI->use_empty())
3057 BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
3058 BBI->eraseFromParent();
3062 // If the unreachable instruction is the first in the block, take a gander
3063 // at all of the predecessors of this instruction, and simplify them.
3064 if (&BB->front() != UI) return Changed;
3066 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
3067 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
3068 TerminatorInst *TI = Preds[i]->getTerminator();
3069 IRBuilder<> Builder(TI);
3070 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
3071 if (BI->isUnconditional()) {
3072 if (BI->getSuccessor(0) == BB) {
3073 new UnreachableInst(TI->getContext(), TI);
3074 TI->eraseFromParent();
3078 if (BI->getSuccessor(0) == BB) {
3079 Builder.CreateBr(BI->getSuccessor(1));
3080 EraseTerminatorInstAndDCECond(BI);
3081 } else if (BI->getSuccessor(1) == BB) {
3082 Builder.CreateBr(BI->getSuccessor(0));
3083 EraseTerminatorInstAndDCECond(BI);
3087 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
3088 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3090 if (i.getCaseSuccessor() == BB) {
3091 BB->removePredecessor(SI->getParent());
3096 // If the default value is unreachable, figure out the most popular
3097 // destination and make it the default.
3098 if (SI->getDefaultDest() == BB) {
3099 std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
3100 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3102 std::pair<unsigned, unsigned> &entry =
3103 Popularity[i.getCaseSuccessor()];
3104 if (entry.first == 0) {
3106 entry.second = i.getCaseIndex();
3112 // Find the most popular block.
3113 unsigned MaxPop = 0;
3114 unsigned MaxIndex = 0;
3115 BasicBlock *MaxBlock = 0;
3116 for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
3117 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
3118 if (I->second.first > MaxPop ||
3119 (I->second.first == MaxPop && MaxIndex > I->second.second)) {
3120 MaxPop = I->second.first;
3121 MaxIndex = I->second.second;
3122 MaxBlock = I->first;
3126 // Make this the new default, allowing us to delete any explicit
3128 SI->setDefaultDest(MaxBlock);
3131 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
3133 if (isa<PHINode>(MaxBlock->begin()))
3134 for (unsigned i = 0; i != MaxPop-1; ++i)
3135 MaxBlock->removePredecessor(SI->getParent());
3137 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
3139 if (i.getCaseSuccessor() == MaxBlock) {
3145 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
3146 if (II->getUnwindDest() == BB) {
3147 // Convert the invoke to a call instruction. This would be a good
3148 // place to note that the call does not throw though.
3149 BranchInst *BI = Builder.CreateBr(II->getNormalDest());
3150 II->removeFromParent(); // Take out of symbol table
3152 // Insert the call now...
3153 SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
3154 Builder.SetInsertPoint(BI);
3155 CallInst *CI = Builder.CreateCall(II->getCalledValue(),
3156 Args, II->getName());
3157 CI->setCallingConv(II->getCallingConv());
3158 CI->setAttributes(II->getAttributes());
3159 // If the invoke produced a value, the call does now instead.
3160 II->replaceAllUsesWith(CI);
3167 // If this block is now dead, remove it.
3168 if (pred_begin(BB) == pred_end(BB) &&
3169 BB != &BB->getParent()->getEntryBlock()) {
3170 // We know there are no successors, so just nuke the block.
3171 BB->eraseFromParent();
3178 /// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
3179 /// integer range comparison into a sub, an icmp and a branch.
3180 static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
3181 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3183 // Make sure all cases point to the same destination and gather the values.
3184 SmallVector<ConstantInt *, 16> Cases;
3185 SwitchInst::CaseIt I = SI->case_begin();
3186 Cases.push_back(I.getCaseValue());
3187 SwitchInst::CaseIt PrevI = I++;
3188 for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
3189 if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
3191 Cases.push_back(I.getCaseValue());
3193 assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
3195 // Sort the case values, then check if they form a range we can transform.
3196 array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
3197 for (unsigned I = 1, E = Cases.size(); I != E; ++I) {
3198 if (Cases[I-1]->getValue() != Cases[I]->getValue()+1)
3202 Constant *Offset = ConstantExpr::getNeg(Cases.back());
3203 Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
3205 Value *Sub = SI->getCondition();
3206 if (!Offset->isNullValue())
3207 Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
3209 // If NumCases overflowed, then all possible values jump to the successor.
3210 if (NumCases->isNullValue() && SI->getNumCases() != 0)
3211 Cmp = ConstantInt::getTrue(SI->getContext());
3213 Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
3214 BranchInst *NewBI = Builder.CreateCondBr(
3215 Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
3217 // Update weight for the newly-created conditional branch.
3218 SmallVector<uint64_t, 8> Weights;
3219 bool HasWeights = HasBranchWeights(SI);
3221 GetBranchWeights(SI, Weights);
3222 if (Weights.size() == 1 + SI->getNumCases()) {
3223 // Combine all weights for the cases to be the true weight of NewBI.
3224 // We assume that the sum of all weights for a Terminator can fit into 32
3226 uint32_t NewTrueWeight = 0;
3227 for (unsigned I = 1, E = Weights.size(); I != E; ++I)
3228 NewTrueWeight += (uint32_t)Weights[I];
3229 NewBI->setMetadata(LLVMContext::MD_prof,
3230 MDBuilder(SI->getContext()).
3231 createBranchWeights(NewTrueWeight,
3232 (uint32_t)Weights[0]));
3236 // Prune obsolete incoming values off the successor's PHI nodes.
3237 for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
3238 isa<PHINode>(BBI); ++BBI) {
3239 for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
3240 cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
3242 SI->eraseFromParent();
3247 /// EliminateDeadSwitchCases - Compute masked bits for the condition of a switch
3248 /// and use it to remove dead cases.
3249 static bool EliminateDeadSwitchCases(SwitchInst *SI) {
3250 Value *Cond = SI->getCondition();
3251 unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
3252 APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
3253 ComputeMaskedBits(Cond, KnownZero, KnownOne);
3255 // Gather dead cases.
3256 SmallVector<ConstantInt*, 8> DeadCases;
3257 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3258 if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
3259 (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
3260 DeadCases.push_back(I.getCaseValue());
3261 DEBUG(dbgs() << "SimplifyCFG: switch case '"
3262 << I.getCaseValue() << "' is dead.\n");
3266 SmallVector<uint64_t, 8> Weights;
3267 bool HasWeight = HasBranchWeights(SI);
3269 GetBranchWeights(SI, Weights);
3270 HasWeight = (Weights.size() == 1 + SI->getNumCases());
3273 // Remove dead cases from the switch.
3274 for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
3275 SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
3276 assert(Case != SI->case_default() &&
3277 "Case was not found. Probably mistake in DeadCases forming.");
3279 std::swap(Weights[Case.getCaseIndex()+1], Weights.back());
3283 // Prune unused values from PHI nodes.
3284 Case.getCaseSuccessor()->removePredecessor(SI->getParent());
3285 SI->removeCase(Case);
3288 SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
3289 SI->setMetadata(LLVMContext::MD_prof,
3290 MDBuilder(SI->getParent()->getContext()).
3291 createBranchWeights(MDWeights));
3294 return !DeadCases.empty();
3297 /// FindPHIForConditionForwarding - If BB would be eligible for simplification
3298 /// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
3299 /// by an unconditional branch), look at the phi node for BB in the successor
3300 /// block and see if the incoming value is equal to CaseValue. If so, return
3301 /// the phi node, and set PhiIndex to BB's index in the phi node.
3302 static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
3305 if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
3306 return NULL; // BB must be empty to be a candidate for simplification.
3307 if (!BB->getSinglePredecessor())
3308 return NULL; // BB must be dominated by the switch.
3310 BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
3311 if (!Branch || !Branch->isUnconditional())
3312 return NULL; // Terminator must be unconditional branch.
3314 BasicBlock *Succ = Branch->getSuccessor(0);
3316 BasicBlock::iterator I = Succ->begin();
3317 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3318 int Idx = PHI->getBasicBlockIndex(BB);
3319 assert(Idx >= 0 && "PHI has no entry for predecessor?");
3321 Value *InValue = PHI->getIncomingValue(Idx);
3322 if (InValue != CaseValue) continue;
3331 /// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
3332 /// instruction to a phi node dominated by the switch, if that would mean that
3333 /// some of the destination blocks of the switch can be folded away.
3334 /// Returns true if a change is made.
3335 static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
3336 typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
3337 ForwardingNodesMap ForwardingNodes;
3339 for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
3340 ConstantInt *CaseValue = I.getCaseValue();
3341 BasicBlock *CaseDest = I.getCaseSuccessor();
3344 PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
3348 ForwardingNodes[PHI].push_back(PhiIndex);
3351 bool Changed = false;
3353 for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
3354 E = ForwardingNodes.end(); I != E; ++I) {
3355 PHINode *Phi = I->first;
3356 SmallVector<int,4> &Indexes = I->second;
3358 if (Indexes.size() < 2) continue;
3360 for (size_t I = 0, E = Indexes.size(); I != E; ++I)
3361 Phi->setIncomingValue(Indexes[I], SI->getCondition());
3368 /// ValidLookupTableConstant - Return true if the backend will be able to handle
3369 /// initializing an array of constants like C.
3370 static bool ValidLookupTableConstant(Constant *C) {
3371 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
3372 return CE->isGEPWithNoNotionalOverIndexing();
3374 return isa<ConstantFP>(C) ||
3375 isa<ConstantInt>(C) ||
3376 isa<ConstantPointerNull>(C) ||
3377 isa<GlobalValue>(C) ||
3381 /// LookupConstant - If V is a Constant, return it. Otherwise, try to look up
3382 /// its constant value in ConstantPool, returning 0 if it's not there.
3383 static Constant *LookupConstant(Value *V,
3384 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3385 if (Constant *C = dyn_cast<Constant>(V))
3387 return ConstantPool.lookup(V);
3390 /// ConstantFold - Try to fold instruction I into a constant. This works for
3391 /// simple instructions such as binary operations where both operands are
3392 /// constant or can be replaced by constants from the ConstantPool. Returns the
3393 /// resulting constant on success, 0 otherwise.
3394 static Constant *ConstantFold(Instruction *I,
3395 const SmallDenseMap<Value*, Constant*>& ConstantPool) {
3396 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
3397 Constant *A = LookupConstant(BO->getOperand(0), ConstantPool);
3400 Constant *B = LookupConstant(BO->getOperand(1), ConstantPool);
3403 return ConstantExpr::get(BO->getOpcode(), A, B);
3406 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
3407 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3410 Constant *B = LookupConstant(I->getOperand(1), ConstantPool);
3413 return ConstantExpr::getCompare(Cmp->getPredicate(), A, B);
3416 if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
3417 Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
3420 if (A->isAllOnesValue())
3421 return LookupConstant(Select->getTrueValue(), ConstantPool);
3422 if (A->isNullValue())
3423 return LookupConstant(Select->getFalseValue(), ConstantPool);
3427 if (CastInst *Cast = dyn_cast<CastInst>(I)) {
3428 Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
3431 return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy());
3437 /// GetCaseResults - Try to determine the resulting constant values in phi nodes
3438 /// at the common destination basic block, *CommonDest, for one of the case
3439 /// destionations CaseDest corresponding to value CaseVal (0 for the default
3440 /// case), of a switch instruction SI.
3441 static bool GetCaseResults(SwitchInst *SI,
3442 ConstantInt *CaseVal,
3443 BasicBlock *CaseDest,
3444 BasicBlock **CommonDest,
3445 SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) {
3446 // The block from which we enter the common destination.
3447 BasicBlock *Pred = SI->getParent();
3449 // If CaseDest is empty except for some side-effect free instructions through
3450 // which we can constant-propagate the CaseVal, continue to its successor.
3451 SmallDenseMap<Value*, Constant*> ConstantPool;
3452 ConstantPool.insert(std::make_pair(SI->getCondition(), CaseVal));
3453 for (BasicBlock::iterator I = CaseDest->begin(), E = CaseDest->end(); I != E;
3455 if (TerminatorInst *T = dyn_cast<TerminatorInst>(I)) {
3456 // If the terminator is a simple branch, continue to the next block.
3457 if (T->getNumSuccessors() != 1)
3460 CaseDest = T->getSuccessor(0);
3461 } else if (isa<DbgInfoIntrinsic>(I)) {
3462 // Skip debug intrinsic.
3464 } else if (Constant *C = ConstantFold(I, ConstantPool)) {
3465 // Instruction is side-effect free and constant.
3466 ConstantPool.insert(std::make_pair(I, C));
3472 // If we did not have a CommonDest before, use the current one.
3474 *CommonDest = CaseDest;
3475 // If the destination isn't the common one, abort.
3476 if (CaseDest != *CommonDest)
3479 // Get the values for this case from phi nodes in the destination block.
3480 BasicBlock::iterator I = (*CommonDest)->begin();
3481 while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
3482 int Idx = PHI->getBasicBlockIndex(Pred);
3486 Constant *ConstVal = LookupConstant(PHI->getIncomingValue(Idx),
3491 // Note: If the constant comes from constant-propagating the case value
3492 // through the CaseDest basic block, it will be safe to remove the
3493 // instructions in that block. They cannot be used (except in the phi nodes
3494 // we visit) outside CaseDest, because that block does not dominate its
3495 // successor. If it did, we would not be in this phi node.
3497 // Be conservative about which kinds of constants we support.
3498 if (!ValidLookupTableConstant(ConstVal))
3501 Res.push_back(std::make_pair(PHI, ConstVal));
3508 /// SwitchLookupTable - This class represents a lookup table that can be used
3509 /// to replace a switch.
3510 class SwitchLookupTable {
3512 /// SwitchLookupTable - Create a lookup table to use as a switch replacement
3513 /// with the contents of Values, using DefaultValue to fill any holes in the
3515 SwitchLookupTable(Module &M,
3517 ConstantInt *Offset,
3518 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3519 Constant *DefaultValue,
3520 const DataLayout *TD);
3522 /// BuildLookup - Build instructions with Builder to retrieve the value at
3523 /// the position given by Index in the lookup table.
3524 Value *BuildLookup(Value *Index, IRBuilder<> &Builder);
3526 /// WouldFitInRegister - Return true if a table with TableSize elements of
3527 /// type ElementType would fit in a target-legal register.
3528 static bool WouldFitInRegister(const DataLayout *TD,
3530 const Type *ElementType);
3533 // Depending on the contents of the table, it can be represented in
3536 // For tables where each element contains the same value, we just have to
3537 // store that single value and return it for each lookup.
3540 // For small tables with integer elements, we can pack them into a bitmap
3541 // that fits into a target-legal register. Values are retrieved by
3542 // shift and mask operations.
3545 // The table is stored as an array of values. Values are retrieved by load
3546 // instructions from the table.
3550 // For SingleValueKind, this is the single value.
3551 Constant *SingleValue;
3553 // For BitMapKind, this is the bitmap.
3554 ConstantInt *BitMap;
3555 IntegerType *BitMapElementTy;
3557 // For ArrayKind, this is the array.
3558 GlobalVariable *Array;
3562 SwitchLookupTable::SwitchLookupTable(Module &M,
3564 ConstantInt *Offset,
3565 const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
3566 Constant *DefaultValue,
3567 const DataLayout *TD)
3568 : SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
3569 assert(Values.size() && "Can't build lookup table without values!");
3570 assert(TableSize >= Values.size() && "Can't fit values in table!");
3572 // If all values in the table are equal, this is that value.
3573 SingleValue = Values.begin()->second;
3575 // Build up the table contents.
3576 SmallVector<Constant*, 64> TableContents(TableSize);
3577 for (size_t I = 0, E = Values.size(); I != E; ++I) {
3578 ConstantInt *CaseVal = Values[I].first;
3579 Constant *CaseRes = Values[I].second;
3580 assert(CaseRes->getType() == DefaultValue->getType());
3582 uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
3584 TableContents[Idx] = CaseRes;
3586 if (CaseRes != SingleValue)
3590 // Fill in any holes in the table with the default result.
3591 if (Values.size() < TableSize) {
3592 for (uint64_t I = 0; I < TableSize; ++I) {
3593 if (!TableContents[I])
3594 TableContents[I] = DefaultValue;
3597 if (DefaultValue != SingleValue)
3601 // If each element in the table contains the same value, we only need to store
3602 // that single value.
3604 Kind = SingleValueKind;
3608 // If the type is integer and the table fits in a register, build a bitmap.
3609 if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) {
3610 IntegerType *IT = cast<IntegerType>(DefaultValue->getType());
3611 APInt TableInt(TableSize * IT->getBitWidth(), 0);
3612 for (uint64_t I = TableSize; I > 0; --I) {
3613 TableInt <<= IT->getBitWidth();
3614 // Insert values into the bitmap. Undef values are set to zero.
3615 if (!isa<UndefValue>(TableContents[I - 1])) {
3616 ConstantInt *Val = cast<ConstantInt>(TableContents[I - 1]);
3617 TableInt |= Val->getValue().zext(TableInt.getBitWidth());
3620 BitMap = ConstantInt::get(M.getContext(), TableInt);
3621 BitMapElementTy = IT;
3627 // Store the table in an array.
3628 ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize);
3629 Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
3631 Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
3632 GlobalVariable::PrivateLinkage,
3635 Array->setUnnamedAddr(true);
3639 Value *SwitchLookupTable::BuildLookup(Value *Index, IRBuilder<> &Builder) {
3641 case SingleValueKind:
3644 // Type of the bitmap (e.g. i59).
3645 IntegerType *MapTy = BitMap->getType();
3647 // Cast Index to the same type as the bitmap.
3648 // Note: The Index is <= the number of elements in the table, so
3649 // truncating it to the width of the bitmask is safe.
3650 Value *ShiftAmt = Builder.CreateZExtOrTrunc(Index, MapTy, "switch.cast");
3652 // Multiply the shift amount by the element width.
3653 ShiftAmt = Builder.CreateMul(ShiftAmt,
3654 ConstantInt::get(MapTy, BitMapElementTy->getBitWidth()),
3658 Value *DownShifted = Builder.CreateLShr(BitMap, ShiftAmt,
3659 "switch.downshift");
3661 return Builder.CreateTrunc(DownShifted, BitMapElementTy,
3665 Value *GEPIndices[] = { Builder.getInt32(0), Index };
3666 Value *GEP = Builder.CreateInBoundsGEP(Array, GEPIndices,
3668 return Builder.CreateLoad(GEP, "switch.load");
3671 llvm_unreachable("Unknown lookup table kind!");
3674 bool SwitchLookupTable::WouldFitInRegister(const DataLayout *TD,
3676 const Type *ElementType) {
3679 const IntegerType *IT = dyn_cast<IntegerType>(ElementType);
3682 // FIXME: If the type is wider than it needs to be, e.g. i8 but all values
3683 // are <= 15, we could try to narrow the type.
3685 // Avoid overflow, fitsInLegalInteger uses unsigned int for the width.
3686 if (TableSize >= UINT_MAX/IT->getBitWidth())
3688 return TD->fitsInLegalInteger(TableSize * IT->getBitWidth());
3691 /// ShouldBuildLookupTable - Determine whether a lookup table should be built
3692 /// for this switch, based on the number of cases, size of the table and the
3693 /// types of the results.
3694 static bool ShouldBuildLookupTable(SwitchInst *SI,
3696 const TargetTransformInfo &TTI,
3697 const DataLayout *TD,
3698 const SmallDenseMap<PHINode*, Type*>& ResultTypes) {
3699 if (SI->getNumCases() > TableSize || TableSize >= UINT64_MAX / 10)
3700 return false; // TableSize overflowed, or mul below might overflow.
3702 bool AllTablesFitInRegister = true;
3703 bool HasIllegalType = false;
3704 for (SmallDenseMap<PHINode*, Type*>::const_iterator I = ResultTypes.begin(),
3705 E = ResultTypes.end(); I != E; ++I) {
3706 Type *Ty = I->second;
3708 // Saturate this flag to true.
3709 HasIllegalType = HasIllegalType || !TTI.isTypeLegal(Ty);
3711 // Saturate this flag to false.
3712 AllTablesFitInRegister = AllTablesFitInRegister &&
3713 SwitchLookupTable::WouldFitInRegister(TD, TableSize, Ty);
3715 // If both flags saturate, we're done. NOTE: This *only* works with
3716 // saturating flags, and all flags have to saturate first due to the
3717 // non-deterministic behavior of iterating over a dense map.
3718 if (HasIllegalType && !AllTablesFitInRegister)
3722 // If each table would fit in a register, we should build it anyway.
3723 if (AllTablesFitInRegister)
3726 // Don't build a table that doesn't fit in-register if it has illegal types.
3730 // The table density should be at least 40%. This is the same criterion as for
3731 // jump tables, see SelectionDAGBuilder::handleJTSwitchCase.
3732 // FIXME: Find the best cut-off.
3733 return SI->getNumCases() * 10 >= TableSize * 4;
3736 /// SwitchToLookupTable - If the switch is only used to initialize one or more
3737 /// phi nodes in a common successor block with different constant values,
3738 /// replace the switch with lookup tables.
3739 static bool SwitchToLookupTable(SwitchInst *SI,
3740 IRBuilder<> &Builder,
3741 const TargetTransformInfo &TTI,
3742 const DataLayout* TD) {
3743 assert(SI->getNumCases() > 1 && "Degenerate switch?");
3745 // Only build lookup table when we have a target that supports it.
3746 if (!TTI.shouldBuildLookupTables())
3749 // FIXME: If the switch is too sparse for a lookup table, perhaps we could
3750 // split off a dense part and build a lookup table for that.
3752 // FIXME: This creates arrays of GEPs to constant strings, which means each
3753 // GEP needs a runtime relocation in PIC code. We should just build one big
3754 // string and lookup indices into that.
3756 // Ignore the switch if the number of cases is too small.
3757 // This is similar to the check when building jump tables in
3758 // SelectionDAGBuilder::handleJTSwitchCase.
3759 // FIXME: Determine the best cut-off.
3760 if (SI->getNumCases() < 4)
3763 // Figure out the corresponding result for each case value and phi node in the
3764 // common destination, as well as the the min and max case values.
3765 assert(SI->case_begin() != SI->case_end());
3766 SwitchInst::CaseIt CI = SI->case_begin();
3767 ConstantInt *MinCaseVal = CI.getCaseValue();
3768 ConstantInt *MaxCaseVal = CI.getCaseValue();
3770 BasicBlock *CommonDest = 0;
3771 typedef SmallVector<std::pair<ConstantInt*, Constant*>, 4> ResultListTy;
3772 SmallDenseMap<PHINode*, ResultListTy> ResultLists;
3773 SmallDenseMap<PHINode*, Constant*> DefaultResults;
3774 SmallDenseMap<PHINode*, Type*> ResultTypes;
3775 SmallVector<PHINode*, 4> PHIs;
3777 for (SwitchInst::CaseIt E = SI->case_end(); CI != E; ++CI) {
3778 ConstantInt *CaseVal = CI.getCaseValue();
3779 if (CaseVal->getValue().slt(MinCaseVal->getValue()))
3780 MinCaseVal = CaseVal;
3781 if (CaseVal->getValue().sgt(MaxCaseVal->getValue()))
3782 MaxCaseVal = CaseVal;
3784 // Resulting value at phi nodes for this case value.
3785 typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
3787 if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
3791 // Append the result from this case to the list for each phi.
3792 for (ResultsTy::iterator I = Results.begin(), E = Results.end(); I!=E; ++I) {
3793 if (!ResultLists.count(I->first))
3794 PHIs.push_back(I->first);
3795 ResultLists[I->first].push_back(std::make_pair(CaseVal, I->second));
3799 // Get the resulting values for the default case.
3800 SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
3801 if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
3802 DefaultResultsList))
3804 for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
3805 PHINode *PHI = DefaultResultsList[I].first;
3806 Constant *Result = DefaultResultsList[I].second;
3807 DefaultResults[PHI] = Result;
3808 ResultTypes[PHI] = Result->getType();
3811 APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
3812 uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
3813 if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes))
3816 // Create the BB that does the lookups.
3817 Module &Mod = *CommonDest->getParent()->getParent();
3818 BasicBlock *LookupBB = BasicBlock::Create(Mod.getContext(),
3820 CommonDest->getParent(),
3823 // Check whether the condition value is within the case range, and branch to
3825 Builder.SetInsertPoint(SI);
3826 Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
3828 Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
3829 MinCaseVal->getType(), TableSize));
3830 Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
3832 // Populate the BB that does the lookups.
3833 Builder.SetInsertPoint(LookupBB);
3834 bool ReturnedEarly = false;
3835 for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
3836 PHINode *PHI = PHIs[I];
3838 SwitchLookupTable Table(Mod, TableSize, MinCaseVal, ResultLists[PHI],
3839 DefaultResults[PHI], TD);
3841 Value *Result = Table.BuildLookup(TableIndex, Builder);
3843 // If the result is used to return immediately from the function, we want to
3844 // do that right here.
3845 if (PHI->hasOneUse() && isa<ReturnInst>(*PHI->use_begin()) &&
3846 *PHI->use_begin() == CommonDest->getFirstNonPHIOrDbg()) {
3847 Builder.CreateRet(Result);
3848 ReturnedEarly = true;
3852 PHI->addIncoming(Result, LookupBB);
3856 Builder.CreateBr(CommonDest);
3858 // Remove the switch.
3859 for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
3860 BasicBlock *Succ = SI->getSuccessor(i);
3861 if (Succ == SI->getDefaultDest()) continue;
3862 Succ->removePredecessor(SI->getParent());
3864 SI->eraseFromParent();
3870 bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
3871 BasicBlock *BB = SI->getParent();
3873 if (isValueEqualityComparison(SI)) {
3874 // If we only have one predecessor, and if it is a branch on this value,
3875 // see if that predecessor totally determines the outcome of this switch.
3876 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3877 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred, Builder))
3878 return SimplifyCFG(BB, TTI, TD) | true;
3880 Value *Cond = SI->getCondition();
3881 if (SelectInst *Select = dyn_cast<SelectInst>(Cond))
3882 if (SimplifySwitchOnSelect(SI, Select))
3883 return SimplifyCFG(BB, TTI, TD) | true;
3885 // If the block only contains the switch, see if we can fold the block
3886 // away into any preds.
3887 BasicBlock::iterator BBI = BB->begin();
3888 // Ignore dbg intrinsics.
3889 while (isa<DbgInfoIntrinsic>(BBI))
3892 if (FoldValueComparisonIntoPredecessors(SI, Builder))
3893 return SimplifyCFG(BB, TTI, TD) | true;
3896 // Try to transform the switch into an icmp and a branch.
3897 if (TurnSwitchRangeIntoICmp(SI, Builder))
3898 return SimplifyCFG(BB, TTI, TD) | true;
3900 // Remove unreachable cases.
3901 if (EliminateDeadSwitchCases(SI))
3902 return SimplifyCFG(BB, TTI, TD) | true;
3904 if (ForwardSwitchConditionToPHI(SI))
3905 return SimplifyCFG(BB, TTI, TD) | true;
3907 if (SwitchToLookupTable(SI, Builder, TTI, TD))
3908 return SimplifyCFG(BB, TTI, TD) | true;
3913 bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
3914 BasicBlock *BB = IBI->getParent();
3915 bool Changed = false;
3917 // Eliminate redundant destinations.
3918 SmallPtrSet<Value *, 8> Succs;
3919 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
3920 BasicBlock *Dest = IBI->getDestination(i);
3921 if (!Dest->hasAddressTaken() || !Succs.insert(Dest)) {
3922 Dest->removePredecessor(BB);
3923 IBI->removeDestination(i);
3929 if (IBI->getNumDestinations() == 0) {
3930 // If the indirectbr has no successors, change it to unreachable.
3931 new UnreachableInst(IBI->getContext(), IBI);
3932 EraseTerminatorInstAndDCECond(IBI);
3936 if (IBI->getNumDestinations() == 1) {
3937 // If the indirectbr has one successor, change it to a direct branch.
3938 BranchInst::Create(IBI->getDestination(0), IBI);
3939 EraseTerminatorInstAndDCECond(IBI);
3943 if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
3944 if (SimplifyIndirectBrOnSelect(IBI, SI))
3945 return SimplifyCFG(BB, TTI, TD) | true;
3950 bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
3951 BasicBlock *BB = BI->getParent();
3953 if (SinkCommon && SinkThenElseCodeToEnd(BI))
3956 // If the Terminator is the only non-phi instruction, simplify the block.
3957 BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
3958 if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
3959 TryToSimplifyUncondBranchFromEmptyBlock(BB))
3962 // If the only instruction in the block is a seteq/setne comparison
3963 // against a constant, try to simplify the block.
3964 if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
3965 if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
3966 for (++I; isa<DbgInfoIntrinsic>(I); ++I)
3968 if (I->isTerminator() &&
3969 TryToSimplifyUncondBranchWithICmpInIt(ICI, Builder, TTI, TD))
3973 // If this basic block is ONLY a compare and a branch, and if a predecessor
3974 // branches to us and our successor, fold the comparison into the
3975 // predecessor and use logical operations to update the incoming value
3976 // for PHI nodes in common successor.
3977 if (FoldBranchToCommonDest(BI))
3978 return SimplifyCFG(BB, TTI, TD) | true;
3983 bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
3984 BasicBlock *BB = BI->getParent();
3986 // Conditional branch
3987 if (isValueEqualityComparison(BI)) {
3988 // If we only have one predecessor, and if it is a branch on this value,
3989 // see if that predecessor totally determines the outcome of this
3991 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
3992 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
3993 return SimplifyCFG(BB, TTI, TD) | true;
3995 // This block must be empty, except for the setcond inst, if it exists.
3996 // Ignore dbg intrinsics.
3997 BasicBlock::iterator I = BB->begin();
3998 // Ignore dbg intrinsics.
3999 while (isa<DbgInfoIntrinsic>(I))
4002 if (FoldValueComparisonIntoPredecessors(BI, Builder))
4003 return SimplifyCFG(BB, TTI, TD) | true;
4004 } else if (&*I == cast<Instruction>(BI->getCondition())){
4006 // Ignore dbg intrinsics.
4007 while (isa<DbgInfoIntrinsic>(I))
4009 if (&*I == BI && FoldValueComparisonIntoPredecessors(BI, Builder))
4010 return SimplifyCFG(BB, TTI, TD) | true;
4014 // Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
4015 if (SimplifyBranchOnICmpChain(BI, TD, Builder))
4018 // If this basic block is ONLY a compare and a branch, and if a predecessor
4019 // branches to us and one of our successors, fold the comparison into the
4020 // predecessor and use logical operations to pick the right destination.
4021 if (FoldBranchToCommonDest(BI))
4022 return SimplifyCFG(BB, TTI, TD) | true;
4024 // We have a conditional branch to two blocks that are only reachable
4025 // from BI. We know that the condbr dominates the two blocks, so see if
4026 // there is any identical code in the "then" and "else" blocks. If so, we
4027 // can hoist it up to the branching block.
4028 if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
4029 if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
4030 if (HoistThenElseCodeToIf(BI))
4031 return SimplifyCFG(BB, TTI, TD) | true;
4033 // If Successor #1 has multiple preds, we may be able to conditionally
4034 // execute Successor #0 if it branches to successor #1.
4035 TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
4036 if (Succ0TI->getNumSuccessors() == 1 &&
4037 Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
4038 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
4039 return SimplifyCFG(BB, TTI, TD) | true;
4041 } else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
4042 // If Successor #0 has multiple preds, we may be able to conditionally
4043 // execute Successor #1 if it branches to successor #0.
4044 TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
4045 if (Succ1TI->getNumSuccessors() == 1 &&
4046 Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
4047 if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
4048 return SimplifyCFG(BB, TTI, TD) | true;
4051 // If this is a branch on a phi node in the current block, thread control
4052 // through this block if any PHI node entries are constants.
4053 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
4054 if (PN->getParent() == BI->getParent())
4055 if (FoldCondBranchOnPHI(BI, TD))
4056 return SimplifyCFG(BB, TTI, TD) | true;
4058 // Scan predecessor blocks for conditional branches.
4059 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
4060 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
4061 if (PBI != BI && PBI->isConditional())
4062 if (SimplifyCondBranchToCondBranch(PBI, BI))
4063 return SimplifyCFG(BB, TTI, TD) | true;
4068 /// Check if passing a value to an instruction will cause undefined behavior.
4069 static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
4070 Constant *C = dyn_cast<Constant>(V);
4077 if (C->isNullValue()) {
4078 // Only look at the first use, avoid hurting compile time with long uselists
4079 User *Use = *I->use_begin();
4081 // Now make sure that there are no instructions in between that can alter
4082 // control flow (eg. calls)
4083 for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
4084 if (i == I->getParent()->end() || i->mayHaveSideEffects())
4087 // Look through GEPs. A load from a GEP derived from NULL is still undefined
4088 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
4089 if (GEP->getPointerOperand() == I)
4090 return passingValueIsAlwaysUndefined(V, GEP);
4092 // Look through bitcasts.
4093 if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
4094 return passingValueIsAlwaysUndefined(V, BC);
4096 // Load from null is undefined.
4097 if (LoadInst *LI = dyn_cast<LoadInst>(Use))
4098 if (!LI->isVolatile())
4099 return LI->getPointerAddressSpace() == 0;
4101 // Store to null is undefined.
4102 if (StoreInst *SI = dyn_cast<StoreInst>(Use))
4103 if (!SI->isVolatile())
4104 return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
4109 /// If BB has an incoming value that will always trigger undefined behavior
4110 /// (eg. null pointer dereference), remove the branch leading here.
4111 static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
4112 for (BasicBlock::iterator i = BB->begin();
4113 PHINode *PHI = dyn_cast<PHINode>(i); ++i)
4114 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
4115 if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
4116 TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
4117 IRBuilder<> Builder(T);
4118 if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
4119 BB->removePredecessor(PHI->getIncomingBlock(i));
4120 // Turn uncoditional branches into unreachables and remove the dead
4121 // destination from conditional branches.
4122 if (BI->isUnconditional())
4123 Builder.CreateUnreachable();
4125 Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
4126 BI->getSuccessor(0));
4127 BI->eraseFromParent();
4130 // TODO: SwitchInst.
4136 bool SimplifyCFGOpt::run(BasicBlock *BB) {
4137 bool Changed = false;
4139 assert(BB && BB->getParent() && "Block not embedded in function!");
4140 assert(BB->getTerminator() && "Degenerate basic block encountered!");
4142 // Remove basic blocks that have no predecessors (except the entry block)...
4143 // or that just have themself as a predecessor. These are unreachable.
4144 if ((pred_begin(BB) == pred_end(BB) &&
4145 BB != &BB->getParent()->getEntryBlock()) ||
4146 BB->getSinglePredecessor() == BB) {
4147 DEBUG(dbgs() << "Removing BB: \n" << *BB);
4148 DeleteDeadBlock(BB);
4152 // Check to see if we can constant propagate this terminator instruction
4154 Changed |= ConstantFoldTerminator(BB, true);
4156 // Check for and eliminate duplicate PHI nodes in this block.
4157 Changed |= EliminateDuplicatePHINodes(BB);
4159 // Check for and remove branches that will always cause undefined behavior.
4160 Changed |= removeUndefIntroducingPredecessor(BB);
4162 // Merge basic blocks into their predecessor if there is only one distinct
4163 // pred, and if there is only one distinct successor of the predecessor, and
4164 // if there are no PHI nodes.
4166 if (MergeBlockIntoPredecessor(BB))
4169 IRBuilder<> Builder(BB);
4171 // If there is a trivial two-entry PHI node in this basic block, and we can
4172 // eliminate it, do so now.
4173 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
4174 if (PN->getNumIncomingValues() == 2)
4175 Changed |= FoldTwoEntryPHINode(PN, TD);
4177 Builder.SetInsertPoint(BB->getTerminator());
4178 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
4179 if (BI->isUnconditional()) {
4180 if (SimplifyUncondBranch(BI, Builder)) return true;
4182 if (SimplifyCondBranch(BI, Builder)) return true;
4184 } else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
4185 if (SimplifyReturn(RI, Builder)) return true;
4186 } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
4187 if (SimplifyResume(RI, Builder)) return true;
4188 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
4189 if (SimplifySwitch(SI, Builder)) return true;
4190 } else if (UnreachableInst *UI =
4191 dyn_cast<UnreachableInst>(BB->getTerminator())) {
4192 if (SimplifyUnreachable(UI)) return true;
4193 } else if (IndirectBrInst *IBI =
4194 dyn_cast<IndirectBrInst>(BB->getTerminator())) {
4195 if (SimplifyIndirectBr(IBI)) return true;
4201 /// SimplifyCFG - This function is used to do simplification of a CFG. For
4202 /// example, it adjusts branches to branches to eliminate the extra hop, it
4203 /// eliminates unreachable basic blocks, and does other "peephole" optimization
4204 /// of the CFG. It returns true if a modification was made.
4206 bool llvm::SimplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI,
4207 const DataLayout *TD) {
4208 return SimplifyCFGOpt(TTI, TD).run(BB);