#define DEBUG_TYPE "simplifycfg"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/GlobalVariable.h"
+#include "llvm/IRBuilder.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
+#include "llvm/LLVMContext.h"
+#include "llvm/MDBuilder.h"
+#include "llvm/Metadata.h"
+#include "llvm/Operator.h"
#include "llvm/Type.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/GlobalVariable.h"
-#include "llvm/Analysis/InstructionSimplify.h"
-#include "llvm/Analysis/ValueTracking.h"
-#include "llvm/Target/TargetData.h"
-#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/ADT/STLExtras.h"
+#include "llvm/Analysis/InstructionSimplify.h"
+#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ConstantRange.h"
#include "llvm/Support/Debug.h"
-#include "llvm/Support/IRBuilder.h"
+#include "llvm/Support/NoFolder.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetData.h"
+#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include <algorithm>
#include <set>
#include <map>
STATISTIC(NumSpeculations, "Number of speculative executed instructions");
namespace {
+ /// ValueEqualityComparisonCase - Represents a case of a switch.
+ struct ValueEqualityComparisonCase {
+ ConstantInt *Value;
+ BasicBlock *Dest;
+
+ ValueEqualityComparisonCase(ConstantInt *Value, BasicBlock *Dest)
+ : Value(Value), Dest(Dest) {}
+
+ bool operator<(ValueEqualityComparisonCase RHS) const {
+ // Comparing pointers is ok as we only rely on the order for uniquing.
+ return Value < RHS.Value;
+ }
+ };
+
class SimplifyCFGOpt {
const TargetData *const TD;
Value *isValueEqualityComparison(TerminatorInst *TI);
BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
- std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases);
+ std::vector<ValueEqualityComparisonCase> &Cases);
bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
BasicBlock *Pred,
IRBuilder<> &Builder);
IRBuilder<> &Builder);
bool SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder);
- bool SimplifyUnwind(UnwindInst *UI, IRBuilder<> &Builder);
+ bool SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder);
bool SimplifyUnreachable(UnreachableInst *UI);
bool SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder);
bool SimplifyIndirectBr(IndirectBrInst *IBI);
///
static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
if (SI1 == SI2) return false; // Can't merge with self!
-
+
// It is not safe to merge these two switch instructions if they have a common
// successor, and if that successor has a PHI node, and if *that* PHI node has
// conflicting incoming values from the two switch blocks.
BasicBlock *SI1BB = SI1->getParent();
BasicBlock *SI2BB = SI2->getParent();
SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
-
+
for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
if (SI1Succs.count(*I))
for (BasicBlock::iterator BBI = (*I)->begin();
PN->getIncomingValueForBlock(SI2BB))
return false;
}
-
+
+ return true;
+}
+
+/// isProfitableToFoldUnconditional - Return true if it is safe and profitable
+/// to merge these two terminator instructions together, where SI1 is an
+/// unconditional branch. PhiNodes will store all PHI nodes in common
+/// successors.
+///
+static bool isProfitableToFoldUnconditional(BranchInst *SI1,
+ BranchInst *SI2,
+ Instruction *Cond,
+ SmallVectorImpl<PHINode*> &PhiNodes) {
+ if (SI1 == SI2) return false; // Can't merge with self!
+ assert(SI1->isUnconditional() && SI2->isConditional());
+
+ // We fold the unconditional branch if we can easily update all PHI nodes in
+ // common successors:
+ // 1> We have a constant incoming value for the conditional branch;
+ // 2> We have "Cond" as the incoming value for the unconditional branch;
+ // 3> SI2->getCondition() and Cond have same operands.
+ CmpInst *Ci2 = dyn_cast<CmpInst>(SI2->getCondition());
+ if (!Ci2) return false;
+ if (!(Cond->getOperand(0) == Ci2->getOperand(0) &&
+ Cond->getOperand(1) == Ci2->getOperand(1)) &&
+ !(Cond->getOperand(0) == Ci2->getOperand(1) &&
+ Cond->getOperand(1) == Ci2->getOperand(0)))
+ return false;
+
+ BasicBlock *SI1BB = SI1->getParent();
+ BasicBlock *SI2BB = SI2->getParent();
+ SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
+ for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
+ if (SI1Succs.count(*I))
+ for (BasicBlock::iterator BBI = (*I)->begin();
+ isa<PHINode>(BBI); ++BBI) {
+ PHINode *PN = cast<PHINode>(BBI);
+ if (PN->getIncomingValueForBlock(SI1BB) != Cond ||
+ !isa<ConstantInt>(PN->getIncomingValueForBlock(SI2BB)))
+ return false;
+ PhiNodes.push_back(PN);
+ }
return true;
}
static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
BasicBlock *ExistPred) {
if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
-
+
PHINode *PN;
for (BasicBlock::iterator I = Succ->begin();
(PN = dyn_cast<PHINode>(I)); ++I)
// doesn't dominate BB.
if (Pred2->getSinglePredecessor() == 0)
return 0;
-
+
// If we found a conditional branch predecessor, make sure that it branches
// to BB and Pred2Br. If it doesn't, this isn't an "if statement".
if (Pred1Br->getSuccessor(0) == BB &&
// Otherwise, if this is a conditional branch, then we can use it!
BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator());
if (BI == 0) return 0;
-
+
assert(BI->isConditional() && "Two successors but not conditional?");
if (BI->getSuccessor(0) == Pred1) {
IfTrue = Pred1;
return BI->getCondition();
}
+/// ComputeSpeculuationCost - Compute an abstract "cost" of speculating the
+/// given instruction, which is assumed to be safe to speculate. 1 means
+/// cheap, 2 means less cheap, and UINT_MAX means prohibitively expensive.
+static unsigned ComputeSpeculationCost(const User *I) {
+ assert(isSafeToSpeculativelyExecute(I) &&
+ "Instruction is not safe to speculatively execute!");
+ switch (Operator::getOpcode(I)) {
+ default:
+ // In doubt, be conservative.
+ return UINT_MAX;
+ case Instruction::GetElementPtr:
+ // GEPs are cheap if all indices are constant.
+ if (!cast<GEPOperator>(I)->hasAllConstantIndices())
+ return UINT_MAX;
+ return 1;
+ case Instruction::Load:
+ case Instruction::Add:
+ case Instruction::Sub:
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ case Instruction::Shl:
+ case Instruction::LShr:
+ case Instruction::AShr:
+ case Instruction::ICmp:
+ case Instruction::Trunc:
+ case Instruction::ZExt:
+ case Instruction::SExt:
+ return 1; // These are all cheap.
+
+ case Instruction::Call:
+ case Instruction::Select:
+ return 2;
+ }
+}
+
/// DominatesMergePoint - If we have a merge point of an "if condition" as
/// accepted above, return true if the specified value dominates the block. We
/// don't handle the true generality of domination here, just a special case
// If we aren't allowing aggressive promotion anymore, then don't consider
// instructions in the 'if region'.
if (AggressiveInsts == 0) return false;
-
+
// If we have seen this instruction before, don't count it again.
if (AggressiveInsts->count(I)) return true;
// Okay, it looks like the instruction IS in the "condition". Check to
// see if it's a cheap instruction to unconditionally compute, and if it
// only uses stuff defined outside of the condition. If so, hoist it out.
- if (!I->isSafeToSpeculativelyExecute())
+ if (!isSafeToSpeculativelyExecute(I))
return false;
- unsigned Cost = 0;
-
- switch (I->getOpcode()) {
- default: return false; // Cannot hoist this out safely.
- case Instruction::Load:
- // We have to check to make sure there are no instructions before the
- // load in its basic block, as we are going to hoist the load out to its
- // predecessor.
- if (PBB->getFirstNonPHIOrDbg() != I)
- return false;
- Cost = 1;
- break;
- case Instruction::GetElementPtr:
- // GEPs are cheap if all indices are constant.
- if (!cast<GetElementPtrInst>(I)->hasAllConstantIndices())
- return false;
- Cost = 1;
- break;
- case Instruction::Add:
- case Instruction::Sub:
- case Instruction::And:
- case Instruction::Or:
- case Instruction::Xor:
- case Instruction::Shl:
- case Instruction::LShr:
- case Instruction::AShr:
- case Instruction::ICmp:
- case Instruction::Trunc:
- case Instruction::ZExt:
- case Instruction::SExt:
- Cost = 1;
- break; // These are all cheap and non-trapping instructions.
-
- case Instruction::Select:
- Cost = 2;
- break;
- }
+ unsigned Cost = ComputeSpeculationCost(I);
if (Cost > CostRemaining)
return false;
// This is some kind of pointer constant. Turn it into a pointer-sized
// ConstantInt if possible.
- const IntegerType *PtrTy = TD->getIntPtrType(V->getContext());
+ IntegerType *PtrTy = TD->getIntPtrType(V->getContext());
// Null pointer means 0, see SelectionDAGBuilder::getValue(const Value*).
if (isa<ConstantPointerNull>(V))
const TargetData *TD, bool isEQ, unsigned &UsedICmps) {
Instruction *I = dyn_cast<Instruction>(V);
if (I == 0) return 0;
-
+
// If this is an icmp against a constant, handle this as one of the cases.
if (ICmpInst *ICI = dyn_cast<ICmpInst>(I)) {
if (ConstantInt *C = GetConstantInt(I->getOperand(1), TD)) {
Vals.push_back(C);
return I->getOperand(0);
}
-
+
// If we have "x ult 3" comparison, for example, then we can add 0,1,2 to
// the set.
ConstantRange Span =
ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
-
+
// If this is an and/!= check then we want to optimize "x ugt 2" into
// x != 0 && x != 1.
if (!isEQ)
Span = Span.inverse();
-
+
// If there are a ton of values, we don't want to make a ginormous switch.
- if (Span.getSetSize().ugt(8) || Span.isEmptySet() ||
- // We don't handle wrapped sets yet.
- Span.isWrappedSet())
+ if (Span.getSetSize().ugt(8) || Span.isEmptySet())
return 0;
-
+
for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
UsedICmps++;
}
return 0;
}
-
+
// Otherwise, we can only handle an | or &, depending on isEQ.
if (I->getOpcode() != (isEQ ? Instruction::Or : Instruction::And))
return 0;
-
+
unsigned NumValsBeforeLHS = Vals.size();
unsigned UsedICmpsBeforeLHS = UsedICmps;
if (Value *LHS = GatherConstantCompares(I->getOperand(0), Vals, Extra, TD,
Extra = I->getOperand(1);
return LHS;
}
-
+
Vals.resize(NumValsBeforeLHS);
UsedICmps = UsedICmpsBeforeLHS;
return 0;
}
-
+
// If the LHS can't be folded in, but Extra is available and RHS can, try to
// use LHS as Extra.
if (Extra == 0 || Extra == I->getOperand(0)) {
assert(Vals.size() == NumValsBeforeLHS);
Extra = OldExtra;
}
-
+
return 0;
}
-
+
static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
- Instruction* Cond = 0;
+ Instruction *Cond = 0;
if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
Cond = dyn_cast<Instruction>(SI->getCondition());
} else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
/// decode all of the 'cases' that it represents and return the 'default' block.
BasicBlock *SimplifyCFGOpt::
GetValueEqualityComparisonCases(TerminatorInst *TI,
- std::vector<std::pair<ConstantInt*,
- BasicBlock*> > &Cases) {
+ std::vector<ValueEqualityComparisonCase>
+ &Cases) {
if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
Cases.reserve(SI->getNumCases());
- for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i)
- Cases.push_back(std::make_pair(SI->getCaseValue(i), SI->getSuccessor(i)));
+ for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
+ Cases.push_back(ValueEqualityComparisonCase(i.getCaseValue(),
+ i.getCaseSuccessor()));
return SI->getDefaultDest();
}
BranchInst *BI = cast<BranchInst>(TI);
ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
- Cases.push_back(std::make_pair(GetConstantInt(ICI->getOperand(1), TD),
- BI->getSuccessor(ICI->getPredicate() ==
- ICmpInst::ICMP_NE)));
+ BasicBlock *Succ = BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_NE);
+ Cases.push_back(ValueEqualityComparisonCase(GetConstantInt(ICI->getOperand(1),
+ TD),
+ Succ));
return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
}
/// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
/// in the list that match the specified block.
static void EliminateBlockCases(BasicBlock *BB,
- std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases) {
+ std::vector<ValueEqualityComparisonCase> &Cases) {
for (unsigned i = 0, e = Cases.size(); i != e; ++i)
- if (Cases[i].second == BB) {
+ if (Cases[i].Dest == BB) {
Cases.erase(Cases.begin()+i);
--i; --e;
}
/// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
/// well.
static bool
-ValuesOverlap(std::vector<std::pair<ConstantInt*, BasicBlock*> > &C1,
- std::vector<std::pair<ConstantInt*, BasicBlock*> > &C2) {
- std::vector<std::pair<ConstantInt*, BasicBlock*> > *V1 = &C1, *V2 = &C2;
+ValuesOverlap(std::vector<ValueEqualityComparisonCase> &C1,
+ std::vector<ValueEqualityComparisonCase > &C2) {
+ std::vector<ValueEqualityComparisonCase> *V1 = &C1, *V2 = &C2;
// Make V1 be smaller than V2.
if (V1->size() > V2->size())
if (V1->size() == 0) return false;
if (V1->size() == 1) {
// Just scan V2.
- ConstantInt *TheVal = (*V1)[0].first;
+ ConstantInt *TheVal = (*V1)[0].Value;
for (unsigned i = 0, e = V2->size(); i != e; ++i)
- if (TheVal == (*V2)[i].first)
+ if (TheVal == (*V2)[i].Value)
return true;
}
array_pod_sort(V2->begin(), V2->end());
unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
while (i1 != e1 && i2 != e2) {
- if ((*V1)[i1].first == (*V2)[i2].first)
+ if ((*V1)[i1].Value == (*V2)[i2].Value)
return true;
- if ((*V1)[i1].first < (*V2)[i2].first)
+ if ((*V1)[i1].Value < (*V2)[i2].Value)
++i1;
else
++i2;
assert(ThisVal && "This isn't a value comparison!!");
if (ThisVal != PredVal) return false; // Different predicates.
+ // TODO: Preserve branch weight metadata, similarly to how
+ // FoldValueComparisonIntoPredecessors preserves it.
+
// Find out information about when control will move from Pred to TI's block.
- std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases;
+ std::vector<ValueEqualityComparisonCase> PredCases;
BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
PredCases);
EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
// Find information about how control leaves this block.
- std::vector<std::pair<ConstantInt*, BasicBlock*> > ThisCases;
+ std::vector<ValueEqualityComparisonCase> ThisCases;
BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
// can simplify TI.
if (!ValuesOverlap(PredCases, ThisCases))
return false;
-
+
if (isa<BranchInst>(TI)) {
// Okay, one of the successors of this condbr is dead. Convert it to a
// uncond br.
(void) NI;
// Remove PHI node entries for the dead edge.
- ThisCases[0].second->removePredecessor(TI->getParent());
+ ThisCases[0].Dest->removePredecessor(TI->getParent());
DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
<< "Through successor TI: " << *TI << "Leaving: " << *NI << "\n");
EraseTerminatorInstAndDCECond(TI);
return true;
}
-
+
SwitchInst *SI = cast<SwitchInst>(TI);
// Okay, TI has cases that are statically dead, prune them away.
SmallPtrSet<Constant*, 16> DeadCases;
for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
- DeadCases.insert(PredCases[i].first);
+ DeadCases.insert(PredCases[i].Value);
DEBUG(dbgs() << "Threading pred instr: " << *Pred->getTerminator()
<< "Through successor TI: " << *TI);
- for (unsigned i = SI->getNumCases()-1; i != 0; --i)
- if (DeadCases.count(SI->getCaseValue(i))) {
- SI->getSuccessor(i)->removePredecessor(TI->getParent());
+ for (SwitchInst::CaseIt i = SI->case_end(), e = SI->case_begin(); i != e;) {
+ --i;
+ if (DeadCases.count(i.getCaseValue())) {
+ i.getCaseSuccessor()->removePredecessor(TI->getParent());
SI->removeCase(i);
}
+ }
DEBUG(dbgs() << "Leaving: " << *TI << "\n");
return true;
}
-
+
// Otherwise, TI's block must correspond to some matched value. Find out
// which value (or set of values) this is.
ConstantInt *TIV = 0;
BasicBlock *TIBB = TI->getParent();
for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
- if (PredCases[i].second == TIBB) {
+ if (PredCases[i].Dest == TIBB) {
if (TIV != 0)
return false; // Cannot handle multiple values coming to this block.
- TIV = PredCases[i].first;
+ TIV = PredCases[i].Value;
}
assert(TIV && "No edge from pred to succ?");
// BB. Find out which successor will unconditionally be branched to.
BasicBlock *TheRealDest = 0;
for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
- if (ThisCases[i].first == TIV) {
- TheRealDest = ThisCases[i].second;
+ if (ThisCases[i].Value == TIV) {
+ TheRealDest = ThisCases[i].Dest;
break;
}
return -1;
}
+static inline bool HasBranchWeights(const Instruction* I) {
+ MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
+ if (ProfMD && ProfMD->getOperand(0))
+ if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0)))
+ return MDS->getString().equals("branch_weights");
+
+ return false;
+}
+
+/// Tries to get a branch weight for the given instruction, returns NULL if it
+/// can't. Pos starts at 0.
+static ConstantInt* GetWeight(Instruction* I, int Pos) {
+ MDNode* ProfMD = I->getMetadata(LLVMContext::MD_prof);
+ if (ProfMD && ProfMD->getOperand(0)) {
+ if (MDString* MDS = dyn_cast<MDString>(ProfMD->getOperand(0))) {
+ if (MDS->getString().equals("branch_weights")) {
+ assert(ProfMD->getNumOperands() >= 3);
+ return dyn_cast<ConstantInt>(ProfMD->getOperand(1 + Pos));
+ }
+ }
+ }
+
+ return 0;
+}
+
+/// Scale the given weights based on the successor TI's metadata. Scaling is
+/// done by multiplying every weight by the sum of the successor's weights.
+static void ScaleWeights(Instruction* STI, MutableArrayRef<uint64_t> Weights) {
+ // Sum the successor's weights
+ assert(HasBranchWeights(STI));
+ unsigned Scale = 0;
+ MDNode* ProfMD = STI->getMetadata(LLVMContext::MD_prof);
+ for (unsigned i = 1; i < ProfMD->getNumOperands(); ++i) {
+ ConstantInt* CI = dyn_cast<ConstantInt>(ProfMD->getOperand(i));
+ assert(CI);
+ Scale += CI->getValue().getZExtValue();
+ }
+
+ // Skip default, as it's replaced during the folding
+ for (unsigned i = 1; i < Weights.size(); ++i) {
+ Weights[i] *= Scale;
+ }
+}
+
+/// Sees if any of the weights are too big for a uint32_t, and halves all the
+/// weights if any are.
+static void FitWeights(MutableArrayRef<uint64_t> Weights) {
+ bool Halve = false;
+ for (unsigned i = 0; i < Weights.size(); ++i)
+ if (Weights[i] > UINT_MAX) {
+ Halve = true;
+ break;
+ }
+
+ if (! Halve)
+ return;
+
+ for (unsigned i = 0; i < Weights.size(); ++i)
+ Weights[i] /= 2;
+}
+
/// FoldValueComparisonIntoPredecessors - The specified terminator is a value
/// equality comparison instruction (either a switch or a branch on "X == c").
/// See if any of the predecessors of the terminator block are value comparisons
if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
// Figure out which 'cases' to copy from SI to PSI.
- std::vector<std::pair<ConstantInt*, BasicBlock*> > BBCases;
+ std::vector<ValueEqualityComparisonCase> BBCases;
BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
- std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases;
+ std::vector<ValueEqualityComparisonCase> PredCases;
BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
// Based on whether the default edge from PTI goes to BB or not, fill in
// build.
SmallVector<BasicBlock*, 8> NewSuccessors;
+ // Update the branch weight metadata along the way
+ SmallVector<uint64_t, 8> Weights;
+ uint64_t PredDefaultWeight = 0;
+ bool PredHasWeights = HasBranchWeights(PTI);
+ bool SuccHasWeights = HasBranchWeights(TI);
+
+ if (PredHasWeights) {
+ MDNode* MD = PTI->getMetadata(LLVMContext::MD_prof);
+ assert(MD);
+ for (unsigned i = 1, e = MD->getNumOperands(); i < e; ++i) {
+ ConstantInt* CI = dyn_cast<ConstantInt>(MD->getOperand(i));
+ assert(CI);
+ Weights.push_back(CI->getValue().getZExtValue());
+ }
+
+ // If the predecessor is a conditional eq, then swap the default weight
+ // to be the first entry.
+ if (BranchInst* BI = dyn_cast<BranchInst>(PTI)) {
+ assert(Weights.size() == 2);
+ ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
+
+ if (ICI->getPredicate() == ICmpInst::ICMP_EQ) {
+ std::swap(Weights.front(), Weights.back());
+ }
+ }
+
+ PredDefaultWeight = Weights.front();
+ } else if (SuccHasWeights) {
+ // If there are no predecessor weights but there are successor weights,
+ // populate Weights with 1, which will later be scaled to the sum of
+ // successor's weights
+ Weights.assign(1 + PredCases.size(), 1);
+ PredDefaultWeight = 1;
+ }
+
+ uint64_t SuccDefaultWeight = 0;
+ if (SuccHasWeights) {
+ int Index = 0;
+ if (BranchInst* BI = dyn_cast<BranchInst>(TI)) {
+ ICmpInst* ICI = dyn_cast<ICmpInst>(BI->getCondition());
+ assert(ICI);
+
+ if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
+ Index = 1;
+ }
+
+ SuccDefaultWeight = GetWeight(TI, Index)->getValue().getZExtValue();
+ }
+
if (PredDefault == BB) {
// If this is the default destination from PTI, only the edges in TI
// that don't occur in PTI, or that branch to BB will be activated.
std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
- if (PredCases[i].second != BB)
- PTIHandled.insert(PredCases[i].first);
+ if (PredCases[i].Dest != BB)
+ PTIHandled.insert(PredCases[i].Value);
else {
// The default destination is BB, we don't need explicit targets.
std::swap(PredCases[i], PredCases.back());
+
+ if (PredHasWeights) {
+ std::swap(Weights[i+1], Weights.back());
+ Weights.pop_back();
+ }
+
PredCases.pop_back();
--i; --e;
}
PredDefault = BBDefault;
NewSuccessors.push_back(BBDefault);
}
+
+ if (SuccHasWeights) {
+ ScaleWeights(TI, Weights);
+ Weights.front() *= SuccDefaultWeight;
+ } else if (PredHasWeights) {
+ Weights.front() /= (1 + BBCases.size());
+ }
+
for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
- if (!PTIHandled.count(BBCases[i].first) &&
- BBCases[i].second != BBDefault) {
+ if (!PTIHandled.count(BBCases[i].Value) &&
+ BBCases[i].Dest != BBDefault) {
PredCases.push_back(BBCases[i]);
- NewSuccessors.push_back(BBCases[i].second);
+ NewSuccessors.push_back(BBCases[i].Dest);
+ if (SuccHasWeights) {
+ Weights.push_back(PredDefaultWeight *
+ GetWeight(TI, i)->getValue().getZExtValue());
+ } else if (PredHasWeights) {
+ // Split the old default's weight amongst the children
+ assert(PredDefaultWeight != 0);
+ Weights.push_back(PredDefaultWeight / (1 + BBCases.size()));
+ }
}
} else {
+ // FIXME: preserve branch weight metadata, similarly to the 'then'
+ // above. For now, drop it.
+ PredHasWeights = false;
+ SuccHasWeights = false;
+
// If this is not the default destination from PSI, only the edges
// in SI that occur in PSI with a destination of BB will be
// activated.
std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
- if (PredCases[i].second == BB) {
- PTIHandled.insert(PredCases[i].first);
+ if (PredCases[i].Dest == BB) {
+ PTIHandled.insert(PredCases[i].Value);
std::swap(PredCases[i], PredCases.back());
PredCases.pop_back();
--i; --e;
// Okay, now we know which constants were sent to BB from the
// predecessor. Figure out where they will all go now.
for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
- if (PTIHandled.count(BBCases[i].first)) {
+ if (PTIHandled.count(BBCases[i].Value)) {
// If this is one we are capable of getting...
PredCases.push_back(BBCases[i]);
- NewSuccessors.push_back(BBCases[i].second);
- PTIHandled.erase(BBCases[i].first);// This constant is taken care of
+ NewSuccessors.push_back(BBCases[i].Dest);
+ PTIHandled.erase(BBCases[i].Value);// This constant is taken care of
}
// If there are any constants vectored to BB that TI doesn't handle,
// they must go to the default destination of TI.
- for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
+ for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
PTIHandled.begin(),
E = PTIHandled.end(); I != E; ++I) {
- PredCases.push_back(std::make_pair(*I, BBDefault));
+ PredCases.push_back(ValueEqualityComparisonCase(*I, BBDefault));
NewSuccessors.push_back(BBDefault);
}
}
PredCases.size());
NewSI->setDebugLoc(PTI->getDebugLoc());
for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
- NewSI->addCase(PredCases[i].first, PredCases[i].second);
+ NewSI->addCase(PredCases[i].Value, PredCases[i].Dest);
+
+ if (PredHasWeights || SuccHasWeights) {
+ // Halve the weights if any of them cannot fit in an uint32_t
+ FitWeights(Weights);
+
+ SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
+
+ NewSI->setMetadata(LLVMContext::MD_prof,
+ MDBuilder(BB->getContext()).
+ createBranchWeights(MDWeights));
+ }
EraseTerminatorInstAndDCECond(PTI);
/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
/// BB2, hoist any common code in the two blocks up into the branch block. The
/// caller of this function guarantees that BI's block dominates BB1 and BB2.
-static bool HoistThenElseCodeToIf(BranchInst *BI, IRBuilder<> &Builder) {
+static bool HoistThenElseCodeToIf(BranchInst *BI) {
// This does very trivial matching, with limited scanning, to find identical
// instructions in the two blocks. In particular, we don't want to get into
// O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
NT->takeName(I1);
}
- Builder.SetInsertPoint(NT);
+ IRBuilder<true, NoFolder> Builder(NT);
// Hoisting one of the terminators from our successor is a great thing.
// Unfortunately, the successors of the if/else blocks may have PHI nodes in
// them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
Value *BB1V = PN->getIncomingValueForBlock(BB1);
Value *BB2V = PN->getIncomingValueForBlock(BB2);
if (BB1V == BB2V) continue;
-
+
// These values do not agree. Insert a select instruction before NT
// that determines the right value.
SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
- if (SI == 0)
+ if (SI == 0)
SI = cast<SelectInst>
(Builder.CreateSelect(BI->getCondition(), BB1V, BB2V,
BB1V->getName()+"."+BB2V->getName()));
/// and an BB2 and the only successor of BB1 is BB2, hoist simple code
/// (for now, restricted to a single instruction that's side effect free) from
/// the BB1 into the branch block to speculatively execute it.
-static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1,
- IRBuilder<> &Builder) {
+///
+/// Turn
+/// BB:
+/// %t1 = icmp
+/// br i1 %t1, label %BB1, label %BB2
+/// BB1:
+/// %t3 = add %t2, c
+/// br label BB2
+/// BB2:
+/// =>
+/// BB:
+/// %t1 = icmp
+/// %t4 = add %t2, c
+/// %t3 = select i1 %t1, %t2, %t3
+static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) {
// Only speculatively execution a single instruction (not counting the
// terminator) for now.
Instruction *HInst = NULL;
return false;
HInst = I;
}
- if (!HInst)
- return false;
+
+ BasicBlock *BIParent = BI->getParent();
+
+ // Check the instruction to be hoisted, if there is one.
+ if (HInst) {
+ // Don't hoist the instruction if it's unsafe or expensive.
+ if (!isSafeToSpeculativelyExecute(HInst))
+ return false;
+ if (ComputeSpeculationCost(HInst) > PHINodeFoldingThreshold)
+ return false;
+
+ // Do not hoist the instruction if any of its operands are defined but not
+ // used in this BB. The transformation will prevent the operand from
+ // being sunk into the use block.
+ for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end();
+ i != e; ++i) {
+ Instruction *OpI = dyn_cast<Instruction>(*i);
+ if (OpI && OpI->getParent() == BIParent &&
+ !OpI->mayHaveSideEffects() &&
+ !OpI->isUsedInBasicBlock(BIParent))
+ return false;
+ }
+ }
// Be conservative for now. FP select instruction can often be expensive.
Value *BrCond = BI->getCondition();
Invert = true;
}
- // Turn
- // BB:
- // %t1 = icmp
- // br i1 %t1, label %BB1, label %BB2
- // BB1:
- // %t3 = add %t2, c
- // br label BB2
- // BB2:
- // =>
- // BB:
- // %t1 = icmp
- // %t4 = add %t2, c
- // %t3 = select i1 %t1, %t2, %t3
- switch (HInst->getOpcode()) {
- default: return false; // Not safe / profitable to hoist.
- case Instruction::Add:
- case Instruction::Sub:
- // Not worth doing for vector ops.
- if (HInst->getType()->isVectorTy())
- return false;
- break;
- case Instruction::And:
- case Instruction::Or:
- case Instruction::Xor:
- case Instruction::Shl:
- case Instruction::LShr:
- case Instruction::AShr:
- // Don't mess with vector operations.
- if (HInst->getType()->isVectorTy())
- return false;
- break; // These are all cheap and non-trapping instructions.
- }
-
- // If the instruction is obviously dead, don't try to predicate it.
- if (HInst->use_empty()) {
- HInst->eraseFromParent();
- return true;
- }
-
- // Can we speculatively execute the instruction? And what is the value
- // if the condition is false? Consider the phi uses, if the incoming value
- // from the "if" block are all the same V, then V is the value of the
- // select if the condition is false.
- BasicBlock *BIParent = BI->getParent();
- SmallVector<PHINode*, 4> PHIUses;
- Value *FalseV = NULL;
-
+ // Collect interesting PHIs, and scan for hazards.
+ SmallSetVector<std::pair<Value *, Value *>, 4> PHIs;
BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0);
- for (Value::use_iterator UI = HInst->use_begin(), E = HInst->use_end();
- UI != E; ++UI) {
- // Ignore any user that is not a PHI node in BB2. These can only occur in
- // unreachable blocks, because they would not be dominated by the instr.
- PHINode *PN = dyn_cast<PHINode>(*UI);
- if (!PN || PN->getParent() != BB2)
- return false;
- PHIUses.push_back(PN);
-
- Value *PHIV = PN->getIncomingValueForBlock(BIParent);
- if (!FalseV)
- FalseV = PHIV;
- else if (FalseV != PHIV)
- return false; // Inconsistent value when condition is false.
- }
-
- assert(FalseV && "Must have at least one user, and it must be a PHI");
-
- // Do not hoist the instruction if any of its operands are defined but not
- // used in this BB. The transformation will prevent the operand from
- // being sunk into the use block.
- for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end();
- i != e; ++i) {
- Instruction *OpI = dyn_cast<Instruction>(*i);
- if (OpI && OpI->getParent() == BIParent &&
- !OpI->isUsedInBasicBlock(BIParent))
- return false;
- }
+ for (BasicBlock::iterator I = BB2->begin();
+ PHINode *PN = dyn_cast<PHINode>(I); ++I) {
+ Value *BB1V = PN->getIncomingValueForBlock(BB1);
+ Value *BIParentV = PN->getIncomingValueForBlock(BIParent);
- // If we get here, we can hoist the instruction. Try to place it
- // before the icmp instruction preceding the conditional branch.
- BasicBlock::iterator InsertPos = BI;
- if (InsertPos != BIParent->begin())
- --InsertPos;
- // Skip debug info between condition and branch.
- while (InsertPos != BIParent->begin() && isa<DbgInfoIntrinsic>(InsertPos))
- --InsertPos;
- if (InsertPos == BrCond && !isa<PHINode>(BrCond)) {
- SmallPtrSet<Instruction *, 4> BB1Insns;
- for(BasicBlock::iterator BB1I = BB1->begin(), BB1E = BB1->end();
- BB1I != BB1E; ++BB1I)
- BB1Insns.insert(BB1I);
- for(Value::use_iterator UI = BrCond->use_begin(), UE = BrCond->use_end();
- UI != UE; ++UI) {
- Instruction *Use = cast<Instruction>(*UI);
- if (!BB1Insns.count(Use)) continue;
-
- // If BrCond uses the instruction that place it just before
- // branch instruction.
- InsertPos = BI;
- break;
+ // Skip PHIs which are trivial.
+ if (BB1V == BIParentV)
+ continue;
+
+ // Check for saftey.
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BB1V)) {
+ // An unfolded ConstantExpr could end up getting expanded into
+ // Instructions. Don't speculate this and another instruction at
+ // the same time.
+ if (HInst)
+ return false;
+ if (!isSafeToSpeculativelyExecute(CE))
+ return false;
+ if (ComputeSpeculationCost(CE) > PHINodeFoldingThreshold)
+ return false;
}
- } else
- InsertPos = BI;
- BIParent->getInstList().splice(InsertPos, BB1->getInstList(), HInst);
- // Create a select whose true value is the speculatively executed value and
- // false value is the previously determined FalseV.
- Builder.SetInsertPoint(BI);
- SelectInst *SI;
- if (Invert)
- SI = cast<SelectInst>
- (Builder.CreateSelect(BrCond, FalseV, HInst,
- FalseV->getName() + "." + HInst->getName()));
- else
- SI = cast<SelectInst>
- (Builder.CreateSelect(BrCond, HInst, FalseV,
- HInst->getName() + "." + FalseV->getName()));
+ // Ok, we may insert a select for this PHI.
+ PHIs.insert(std::make_pair(BB1V, BIParentV));
+ }
+
+ // If there are no PHIs to process, bail early. This helps ensure idempotence
+ // as well.
+ if (PHIs.empty())
+ return false;
- // Make the PHI node use the select for all incoming values for "then" and
- // "if" blocks.
- for (unsigned i = 0, e = PHIUses.size(); i != e; ++i) {
- PHINode *PN = PHIUses[i];
- for (unsigned j = 0, ee = PN->getNumIncomingValues(); j != ee; ++j)
- if (PN->getIncomingBlock(j) == BB1 || PN->getIncomingBlock(j) == BIParent)
- PN->setIncomingValue(j, SI);
+ // If we get here, we can hoist the instruction and if-convert.
+ DEBUG(dbgs() << "SPECULATIVELY EXECUTING BB" << *BB1 << "\n";);
+
+ // Hoist the instruction.
+ if (HInst)
+ BIParent->getInstList().splice(BI, BB1->getInstList(), HInst);
+
+ // Insert selects and rewrite the PHI operands.
+ IRBuilder<true, NoFolder> Builder(BI);
+ for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
+ Value *TrueV = PHIs[i].first;
+ Value *FalseV = PHIs[i].second;
+
+ // Create a select whose true value is the speculatively executed value and
+ // false value is the previously determined FalseV.
+ SelectInst *SI;
+ if (Invert)
+ SI = cast<SelectInst>
+ (Builder.CreateSelect(BrCond, FalseV, TrueV,
+ FalseV->getName() + "." + TrueV->getName()));
+ else
+ SI = cast<SelectInst>
+ (Builder.CreateSelect(BrCond, TrueV, FalseV,
+ TrueV->getName() + "." + FalseV->getName()));
+
+ // Make the PHI node use the select for all incoming values for "then" and
+ // "if" blocks.
+ for (BasicBlock::iterator I = BB2->begin();
+ PHINode *PN = dyn_cast<PHINode>(I); ++I) {
+ unsigned BB1I = PN->getBasicBlockIndex(BB1);
+ unsigned BIParentI = PN->getBasicBlockIndex(BIParent);
+ Value *BB1V = PN->getIncomingValue(BB1I);
+ Value *BIParentV = PN->getIncomingValue(BIParentI);
+ if (TrueV == BB1V && FalseV == BIParentV) {
+ PN->setIncomingValue(BB1I, SI);
+ PN->setIncomingValue(BIParentI, SI);
+ }
+ }
}
++NumSpeculations;
static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
BranchInst *BI = cast<BranchInst>(BB->getTerminator());
unsigned Size = 0;
-
+
for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
if (isa<DbgInfoIntrinsic>(BBI))
continue;
if (Size > 10) return false; // Don't clone large BB's.
++Size;
-
+
// We can only support instructions that do not define values that are
// live outside of the current basic block.
for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
Instruction *U = cast<Instruction>(*UI);
if (U->getParent() != BB || isa<PHINode>(U)) return false;
}
-
+
// Looks ok, continue checking.
}
// outside of the block.
if (!PN || PN->getParent() != BB || !PN->hasOneUse())
return false;
-
+
// Degenerate case of a single entry PHI.
if (PN->getNumIncomingValues() == 1) {
FoldSingleEntryPHINodes(PN->getParent());
- return true;
+ return true;
}
// Now we know that this block has multiple preds and two succs.
if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
-
+
// Okay, this is a simple enough basic block. See if any phi values are
// constants.
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
ConstantInt *CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i));
if (CB == 0 || !CB->getType()->isIntegerTy(1)) continue;
-
+
// Okay, we now know that all edges from PredBB should be revectored to
// branch to RealDest.
BasicBlock *PredBB = PN->getIncomingBlock(i);
BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
-
+
if (RealDest == BB) continue; // Skip self loops.
-
+ // Skip if the predecessor's terminator is an indirect branch.
+ if (isa<IndirectBrInst>(PredBB->getTerminator())) continue;
+
// The dest block might have PHI nodes, other predecessors and other
// difficult cases. Instead of being smart about this, just insert a new
// block that jumps to the destination block, effectively splitting
RealDest->getName()+".critedge",
RealDest->getParent(), RealDest);
BranchInst::Create(RealDest, EdgeBB);
-
+
// Update PHI nodes.
AddPredecessorToBlock(RealDest, EdgeBB, BB);
// Clone the instruction.
Instruction *N = BBI->clone();
if (BBI->hasName()) N->setName(BBI->getName()+".c");
-
+
// Update operands due to translation.
for (User::op_iterator i = N->op_begin(), e = N->op_end();
i != e; ++i) {
if (PI != TranslateMap.end())
*i = PI->second;
}
-
+
// Check for trivial simplification.
if (Value *V = SimplifyInstruction(N, TD)) {
TranslateMap[BBI] = V;
BB->removePredecessor(PredBB);
PredBBTI->setSuccessor(i, EdgeBB);
}
-
+
// Recurse, simplifying any other constants.
return FoldCondBranchOnPHI(BI, TD) | true;
}
/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
/// PHI node, see if we can eliminate it.
-static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD,
- IRBuilder<> &Builder) {
+static bool FoldTwoEntryPHINode(PHINode *PN, const TargetData *TD) {
// Ok, this is a two entry PHI node. Check to see if this is a simple "if
// statement", which has a very simple dominance structure. Basically, we
// are trying to find the condition that is being branched on, which
// Don't bother if the branch will be constant folded trivially.
isa<ConstantInt>(IfCond))
return false;
-
+
// Okay, we found that we can merge this two-entry phi node into a select.
// Doing so would require us to fold *all* two entry phi nodes in this block.
// At some point this becomes non-profitable (particularly if the target
for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
if (NumPhis > 2)
return false;
-
+
// Loop over the PHI's seeing if we can promote them all to select
// instructions. While we are at it, keep track of the instructions
// that need to be moved to the dominating block.
SmallPtrSet<Instruction*, 4> AggressiveInsts;
unsigned MaxCostVal0 = PHINodeFoldingThreshold,
MaxCostVal1 = PHINodeFoldingThreshold;
-
+
for (BasicBlock::iterator II = BB->begin(); isa<PHINode>(II);) {
PHINode *PN = cast<PHINode>(II++);
if (Value *V = SimplifyInstruction(PN, TD)) {
PN->eraseFromParent();
continue;
}
-
+
if (!DominatesMergePoint(PN->getIncomingValue(0), BB, &AggressiveInsts,
MaxCostVal0) ||
!DominatesMergePoint(PN->getIncomingValue(1), BB, &AggressiveInsts,
MaxCostVal1))
return false;
}
-
- // If we folded the the first phi, PN dangles at this point. Refresh it. If
+
+ // If we folded the first phi, PN dangles at this point. Refresh it. If
// we ran out of PHIs then we simplified them all.
PN = dyn_cast<PHINode>(BB->begin());
if (PN == 0) return true;
-
+
// Don't fold i1 branches on PHIs which contain binary operators. These can
// often be turned into switches and other things.
if (PN->getType()->isIntegerTy(1) &&
isa<BinaryOperator>(PN->getIncomingValue(1)) ||
isa<BinaryOperator>(IfCond)))
return false;
-
+
// If we all PHI nodes are promotable, check to make sure that all
// instructions in the predecessor blocks can be promoted as well. If
// not, we won't be able to get rid of the control flow, so it's not
return false;
}
}
-
+
if (cast<BranchInst>(IfBlock2->getTerminator())->isConditional()) {
IfBlock2 = 0;
} else {
return false;
}
}
-
+
DEBUG(dbgs() << "FOUND IF CONDITION! " << *IfCond << " T: "
<< IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
-
+
// If we can still promote the PHI nodes after this gauntlet of tests,
// do all of the PHI's now.
Instruction *InsertPt = DomBlock->getTerminator();
- Builder.SetInsertPoint(InsertPt);
-
+ IRBuilder<true, NoFolder> Builder(InsertPt);
+
// Move all 'aggressive' instructions, which are defined in the
// conditional parts of the if's up to the dominating block.
if (IfBlock1)
DomBlock->getInstList().splice(InsertPt,
IfBlock2->getInstList(), IfBlock2->begin(),
IfBlock2->getTerminator());
-
+
while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
// Change the PHI node into a select instruction.
Value *TrueVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
Value *FalseVal = PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
-
- SelectInst *NV =
+
+ SelectInst *NV =
cast<SelectInst>(Builder.CreateSelect(IfCond, TrueVal, FalseVal, ""));
PN->replaceAllUsesWith(NV);
NV->takeName(PN);
PN->eraseFromParent();
}
-
+
// At this point, IfBlock1 and IfBlock2 are both empty, so our if statement
// has been flattened. Change DomBlock to jump directly to our new block to
// avoid other simplifycfg's kicking in on the diamond.
/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
/// to two returning blocks, try to merge them together into one return,
/// introducing a select if the return values disagree.
-static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
+static bool SimplifyCondBranchToTwoReturns(BranchInst *BI,
IRBuilder<> &Builder) {
assert(BI->isConditional() && "Must be a conditional branch");
BasicBlock *TrueSucc = BI->getSuccessor(0);
BasicBlock *FalseSucc = BI->getSuccessor(1);
ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
-
+
// Check to ensure both blocks are empty (just a return) or optionally empty
// with PHI nodes. If there are other instructions, merging would cause extra
// computation on one path or the other.
EraseTerminatorInstAndDCECond(BI);
return true;
}
-
+
// Otherwise, figure out what the true and false return values are
// so we can insert a new select instruction.
Value *TrueValue = TrueRet->getReturnValue();
Value *FalseValue = FalseRet->getReturnValue();
-
+
// Unwrap any PHI nodes in the return blocks.
if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
if (TVPN->getParent() == TrueSucc)
if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
if (FVPN->getParent() == FalseSucc)
FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
-
+
// In order for this transformation to be safe, we must be able to
// unconditionally execute both operands to the return. This is
// normally the case, but we could have a potentially-trapping
if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
if (FCV->canTrap())
return false;
-
+
// Okay, we collected all the mapped values and checked them for sanity, and
// defined to really do this transformation. First, update the CFG.
TrueSucc->removePredecessor(BI->getParent());
FalseSucc->removePredecessor(BI->getParent());
-
+
// Insert select instructions where needed.
Value *BrCond = BI->getCondition();
if (TrueValue) {
}
}
- Value *RI = !TrueValue ?
+ Value *RI = !TrueValue ?
Builder.CreateRetVoid() : Builder.CreateRet(TrueValue);
(void) RI;
-
+
DEBUG(dbgs() << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
<< "\n " << *BI << "NewRet = " << *RI
<< "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc);
-
+
EraseTerminatorInstAndDCECond(BI);
return true;
}
+/// ExtractBranchMetadata - Given a conditional BranchInstruction, retrieve the
+/// probabilities of the branch taking each edge. Fills in the two APInt
+/// parameters and return true, or returns false if no or invalid metadata was
+/// found.
+static bool ExtractBranchMetadata(BranchInst *BI,
+ APInt &ProbTrue, APInt &ProbFalse) {
+ assert(BI->isConditional() &&
+ "Looking for probabilities on unconditional branch?");
+ MDNode *ProfileData = BI->getMetadata(LLVMContext::MD_prof);
+ if (!ProfileData || ProfileData->getNumOperands() != 3) return false;
+ ConstantInt *CITrue = dyn_cast<ConstantInt>(ProfileData->getOperand(1));
+ ConstantInt *CIFalse = dyn_cast<ConstantInt>(ProfileData->getOperand(2));
+ if (!CITrue || !CIFalse) return false;
+ ProbTrue = CITrue->getValue();
+ ProbFalse = CIFalse->getValue();
+ assert(ProbTrue.getBitWidth() == 32 && ProbFalse.getBitWidth() == 32 &&
+ "Branch probability metadata must be 32-bit integers");
+ return true;
+}
+
+/// MultiplyAndLosePrecision - Multiplies A and B, then returns the result. In
+/// the event of overflow, logically-shifts all four inputs right until the
+/// multiply fits.
+static APInt MultiplyAndLosePrecision(APInt &A, APInt &B, APInt &C, APInt &D,
+ unsigned &BitsLost) {
+ BitsLost = 0;
+ bool Overflow = false;
+ APInt Result = A.umul_ov(B, Overflow);
+ if (Overflow) {
+ APInt MaxB = APInt::getMaxValue(A.getBitWidth()).udiv(A);
+ do {
+ B = B.lshr(1);
+ ++BitsLost;
+ } while (B.ugt(MaxB));
+ A = A.lshr(BitsLost);
+ C = C.lshr(BitsLost);
+ D = D.lshr(BitsLost);
+ Result = A * B;
+ }
+ return Result;
+}
+
+/// checkCSEInPredecessor - Return true if the given instruction is available
+/// in its predecessor block. If yes, the instruction will be removed.
+///
+static bool checkCSEInPredecessor(Instruction *Inst, BasicBlock *PB) {
+ if (!isa<BinaryOperator>(Inst) && !isa<CmpInst>(Inst))
+ return false;
+ for (BasicBlock::iterator I = PB->begin(), E = PB->end(); I != E; I++) {
+ Instruction *PBI = &*I;
+ // Check whether Inst and PBI generate the same value.
+ if (Inst->isIdenticalTo(PBI)) {
+ Inst->replaceAllUsesWith(PBI);
+ Inst->eraseFromParent();
+ return true;
+ }
+ }
+ return false;
+}
+
/// FoldBranchToCommonDest - If this basic block is simple enough, and if a
/// predecessor branches to us and one of our successors, fold the block into
/// the predecessor and use logical operations to pick the right destination.
bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
BasicBlock *BB = BI->getParent();
- Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
+ Instruction *Cond = 0;
+ if (BI->isConditional())
+ Cond = dyn_cast<Instruction>(BI->getCondition());
+ else {
+ // For unconditional branch, check for a simple CFG pattern, where
+ // BB has a single predecessor and BB's successor is also its predecessor's
+ // successor. If such pattern exisits, check for CSE between BB and its
+ // predecessor.
+ if (BasicBlock *PB = BB->getSinglePredecessor())
+ if (BranchInst *PBI = dyn_cast<BranchInst>(PB->getTerminator()))
+ if (PBI->isConditional() &&
+ (BI->getSuccessor(0) == PBI->getSuccessor(0) ||
+ BI->getSuccessor(0) == PBI->getSuccessor(1))) {
+ for (BasicBlock::iterator I = BB->begin(), E = BB->end();
+ I != E; ) {
+ Instruction *Curr = I++;
+ if (isa<CmpInst>(Curr)) {
+ Cond = Curr;
+ break;
+ }
+ // Quit if we can't remove this instruction.
+ if (!checkCSEInPredecessor(Curr, PB))
+ return false;
+ }
+ }
+
+ if (Cond == 0)
+ return false;
+ }
+
if (Cond == 0 || (!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
Cond->getParent() != BB || !Cond->hasOneUse())
return false;
// Ignore dbg intrinsics.
while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
-
+
// Allow a single instruction to be hoisted in addition to the compare
// that feeds the branch. We later ensure that any values that _it_ uses
// were also live in the predecessor, so that we don't unnecessarily create
Instruction *BonusInst = 0;
if (&*FrontIt != Cond &&
FrontIt->hasOneUse() && *FrontIt->use_begin() == Cond &&
- FrontIt->isSafeToSpeculativelyExecute()) {
+ isSafeToSpeculativelyExecute(FrontIt)) {
BonusInst = &*FrontIt;
++FrontIt;
-
+
// Ignore dbg intrinsics.
while (isa<DbgInfoIntrinsic>(FrontIt)) ++FrontIt;
}
// Only a single bonus inst is allowed.
if (&*FrontIt != Cond)
return false;
-
+
// Make sure the instruction after the condition is the cond branch.
BasicBlock::iterator CondIt = Cond; ++CondIt;
// Ingore dbg intrinsics.
while (isa<DbgInfoIntrinsic>(CondIt)) ++CondIt;
-
+
if (&*CondIt != BI)
return false;
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
if (CE->canTrap())
return false;
-
+
// Finally, don't infinitely unroll conditional loops.
BasicBlock *TrueDest = BI->getSuccessor(0);
- BasicBlock *FalseDest = BI->getSuccessor(1);
+ BasicBlock *FalseDest = (BI->isConditional()) ? BI->getSuccessor(1) : 0;
if (TrueDest == BB || FalseDest == BB)
return false;
for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
BasicBlock *PredBlock = *PI;
BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
-
+
// Check that we have two conditional branches. If there is a PHI node in
// the common successor, verify that the same value flows in from both
// blocks.
- if (PBI == 0 || PBI->isUnconditional() || !SafeToMergeTerminators(BI, PBI))
+ SmallVector<PHINode*, 4> PHIs;
+ if (PBI == 0 || PBI->isUnconditional() ||
+ (BI->isConditional() &&
+ !SafeToMergeTerminators(BI, PBI)) ||
+ (!BI->isConditional() &&
+ !isProfitableToFoldUnconditional(BI, PBI, Cond, PHIs)))
continue;
-
+
// Determine if the two branches share a common destination.
Instruction::BinaryOps Opc;
bool InvertPredCond = false;
-
- if (PBI->getSuccessor(0) == TrueDest)
- Opc = Instruction::Or;
- else if (PBI->getSuccessor(1) == FalseDest)
- Opc = Instruction::And;
- else if (PBI->getSuccessor(0) == FalseDest)
- Opc = Instruction::And, InvertPredCond = true;
- else if (PBI->getSuccessor(1) == TrueDest)
- Opc = Instruction::Or, InvertPredCond = true;
- else
- continue;
+
+ if (BI->isConditional()) {
+ if (PBI->getSuccessor(0) == TrueDest)
+ Opc = Instruction::Or;
+ else if (PBI->getSuccessor(1) == FalseDest)
+ Opc = Instruction::And;
+ else if (PBI->getSuccessor(0) == FalseDest)
+ Opc = Instruction::And, InvertPredCond = true;
+ else if (PBI->getSuccessor(1) == TrueDest)
+ Opc = Instruction::Or, InvertPredCond = true;
+ else
+ continue;
+ } else {
+ if (PBI->getSuccessor(0) != TrueDest && PBI->getSuccessor(1) != TrueDest)
+ continue;
+ }
// Ensure that any values used in the bonus instruction are also used
// by the terminator of the predecessor. This means that those values
- // must already have been resolved, so we won't be inhibiting the
+ // must already have been resolved, so we won't be inhibiting the
// out-of-order core by speculating them earlier.
if (BonusInst) {
// Collect the values used by the bonus inst
SmallPtrSet<Value*, 4> UsedValues;
for (Instruction::op_iterator OI = BonusInst->op_begin(),
OE = BonusInst->op_end(); OI != OE; ++OI) {
- Value* V = *OI;
+ Value *V = *OI;
if (!isa<Constant>(V))
UsedValues.insert(V);
}
SmallVector<std::pair<Value*, unsigned>, 4> Worklist;
Worklist.push_back(std::make_pair(PBI->getOperand(0), 0));
-
+
// Walk up to four levels back up the use-def chain of the predecessor's
// terminator to see if all those values were used. The choice of four
// levels is arbitrary, to provide a compile-time-cost bound.
while (!Worklist.empty()) {
std::pair<Value*, unsigned> Pair = Worklist.back();
Worklist.pop_back();
-
+
if (Pair.second >= 4) continue;
UsedValues.erase(Pair.first);
if (UsedValues.empty()) break;
-
+
if (Instruction *I = dyn_cast<Instruction>(Pair.first)) {
for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
OI != OE; ++OI)
Worklist.push_back(std::make_pair(OI->get(), Pair.second+1));
- }
+ }
}
-
+
if (!UsedValues.empty()) return false;
}
DEBUG(dbgs() << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB);
- IRBuilder<> Builder(PBI);
+ IRBuilder<> Builder(PBI);
// If we need to invert the condition in the pred block to match, do so now.
if (InvertPredCond) {
Value *NewCond = PBI->getCondition();
-
+
if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) {
CmpInst *CI = cast<CmpInst>(NewCond);
CI->setPredicate(CI->getInversePredicate());
} else {
- NewCond = Builder.CreateNot(NewCond,
+ NewCond = Builder.CreateNot(NewCond,
PBI->getCondition()->getName()+".not");
}
-
+
PBI->setCondition(NewCond);
- BasicBlock *OldTrue = PBI->getSuccessor(0);
- BasicBlock *OldFalse = PBI->getSuccessor(1);
- PBI->setSuccessor(0, OldFalse);
- PBI->setSuccessor(1, OldTrue);
+ PBI->swapSuccessors();
}
-
+
// If we have a bonus inst, clone it into the predecessor block.
Instruction *NewBonus = 0;
if (BonusInst) {
NewBonus->takeName(BonusInst);
BonusInst->setName(BonusInst->getName()+".old");
}
-
+
// Clone Cond into the predecessor basic block, and or/and the
// two conditions together.
Instruction *New = Cond->clone();
PredBlock->getInstList().insert(PBI, New);
New->takeName(Cond);
Cond->setName(New->getName()+".old");
-
- Instruction *NewCond =
- cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
+
+ if (BI->isConditional()) {
+ Instruction *NewCond =
+ cast<Instruction>(Builder.CreateBinOp(Opc, PBI->getCondition(),
New, "or.cond"));
- PBI->setCondition(NewCond);
- if (PBI->getSuccessor(0) == BB) {
- AddPredecessorToBlock(TrueDest, PredBlock, BB);
- PBI->setSuccessor(0, TrueDest);
+ PBI->setCondition(NewCond);
+
+ if (PBI->getSuccessor(0) == BB) {
+ AddPredecessorToBlock(TrueDest, PredBlock, BB);
+ PBI->setSuccessor(0, TrueDest);
+ }
+ if (PBI->getSuccessor(1) == BB) {
+ AddPredecessorToBlock(FalseDest, PredBlock, BB);
+ PBI->setSuccessor(1, FalseDest);
+ }
+ } else {
+ // Update PHI nodes in the common successors.
+ for (unsigned i = 0, e = PHIs.size(); i != e; ++i) {
+ ConstantInt *PBI_C = cast<ConstantInt>(
+ PHIs[i]->getIncomingValueForBlock(PBI->getParent()));
+ assert(PBI_C->getType()->isIntegerTy(1));
+ Instruction *MergedCond = 0;
+ if (PBI->getSuccessor(0) == TrueDest) {
+ // Create (PBI_Cond and PBI_C) or (!PBI_Cond and BI_Value)
+ // PBI_C is true: PBI_Cond or (!PBI_Cond and BI_Value)
+ // is false: !PBI_Cond and BI_Value
+ Instruction *NotCond =
+ cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
+ "not.cond"));
+ MergedCond =
+ cast<Instruction>(Builder.CreateBinOp(Instruction::And,
+ NotCond, New,
+ "and.cond"));
+ if (PBI_C->isOne())
+ MergedCond =
+ cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
+ PBI->getCondition(), MergedCond,
+ "or.cond"));
+ } else {
+ // Create (PBI_Cond and BI_Value) or (!PBI_Cond and PBI_C)
+ // PBI_C is true: (PBI_Cond and BI_Value) or (!PBI_Cond)
+ // is false: PBI_Cond and BI_Value
+ MergedCond =
+ cast<Instruction>(Builder.CreateBinOp(Instruction::And,
+ PBI->getCondition(), New,
+ "and.cond"));
+ if (PBI_C->isOne()) {
+ Instruction *NotCond =
+ cast<Instruction>(Builder.CreateNot(PBI->getCondition(),
+ "not.cond"));
+ MergedCond =
+ cast<Instruction>(Builder.CreateBinOp(Instruction::Or,
+ NotCond, MergedCond,
+ "or.cond"));
+ }
+ }
+ // Update PHI Node.
+ PHIs[i]->setIncomingValue(PHIs[i]->getBasicBlockIndex(PBI->getParent()),
+ MergedCond);
+ }
+ // Change PBI from Conditional to Unconditional.
+ BranchInst *New_PBI = BranchInst::Create(TrueDest, PBI);
+ EraseTerminatorInstAndDCECond(PBI);
+ PBI = New_PBI;
}
- if (PBI->getSuccessor(1) == BB) {
- AddPredecessorToBlock(FalseDest, PredBlock, BB);
- PBI->setSuccessor(1, FalseDest);
+
+ // TODO: If BB is reachable from all paths through PredBlock, then we
+ // could replace PBI's branch probabilities with BI's.
+
+ // Merge probability data into PredBlock's branch.
+ APInt A, B, C, D;
+ if (PBI->isConditional() && BI->isConditional() &&
+ ExtractBranchMetadata(PBI, C, D) && ExtractBranchMetadata(BI, A, B)) {
+ // Given IR which does:
+ // bbA:
+ // br i1 %x, label %bbB, label %bbC
+ // bbB:
+ // br i1 %y, label %bbD, label %bbC
+ // Let's call the probability that we take the edge from %bbA to %bbB
+ // 'a', from %bbA to %bbC, 'b', from %bbB to %bbD 'c' and from %bbB to
+ // %bbC probability 'd'.
+ //
+ // We transform the IR into:
+ // bbA:
+ // br i1 %z, label %bbD, label %bbC
+ // where the probability of going to %bbD is (a*c) and going to bbC is
+ // (b+a*d).
+ //
+ // Probabilities aren't stored as ratios directly. Using branch weights,
+ // we get:
+ // (a*c)% = A*C, (b+(a*d))% = A*D+B*C+B*D.
+
+ // In the event of overflow, we want to drop the LSB of the input
+ // probabilities.
+ unsigned BitsLost;
+
+ // Ignore overflow result on ProbTrue.
+ APInt ProbTrue = MultiplyAndLosePrecision(A, C, B, D, BitsLost);
+
+ APInt Tmp1 = MultiplyAndLosePrecision(B, D, A, C, BitsLost);
+ if (BitsLost) {
+ ProbTrue = ProbTrue.lshr(BitsLost*2);
+ }
+
+ APInt Tmp2 = MultiplyAndLosePrecision(A, D, C, B, BitsLost);
+ if (BitsLost) {
+ ProbTrue = ProbTrue.lshr(BitsLost*2);
+ Tmp1 = Tmp1.lshr(BitsLost*2);
+ }
+
+ APInt Tmp3 = MultiplyAndLosePrecision(B, C, A, D, BitsLost);
+ if (BitsLost) {
+ ProbTrue = ProbTrue.lshr(BitsLost*2);
+ Tmp1 = Tmp1.lshr(BitsLost*2);
+ Tmp2 = Tmp2.lshr(BitsLost*2);
+ }
+
+ bool Overflow1 = false, Overflow2 = false;
+ APInt Tmp4 = Tmp2.uadd_ov(Tmp3, Overflow1);
+ APInt ProbFalse = Tmp4.uadd_ov(Tmp1, Overflow2);
+
+ if (Overflow1 || Overflow2) {
+ ProbTrue = ProbTrue.lshr(1);
+ Tmp1 = Tmp1.lshr(1);
+ Tmp2 = Tmp2.lshr(1);
+ Tmp3 = Tmp3.lshr(1);
+ Tmp4 = Tmp2 + Tmp3;
+ ProbFalse = Tmp4 + Tmp1;
+ }
+
+ // The sum of branch weights must fit in 32-bits.
+ if (ProbTrue.isNegative() && ProbFalse.isNegative()) {
+ ProbTrue = ProbTrue.lshr(1);
+ ProbFalse = ProbFalse.lshr(1);
+ }
+
+ if (ProbTrue != ProbFalse) {
+ // Normalize the result.
+ APInt GCD = APIntOps::GreatestCommonDivisor(ProbTrue, ProbFalse);
+ ProbTrue = ProbTrue.udiv(GCD);
+ ProbFalse = ProbFalse.udiv(GCD);
+
+ MDBuilder MDB(BI->getContext());
+ MDNode *N = MDB.createBranchWeights(ProbTrue.getZExtValue(),
+ ProbFalse.getZExtValue());
+ PBI->setMetadata(LLVMContext::MD_prof, N);
+ } else {
+ PBI->setMetadata(LLVMContext::MD_prof, NULL);
+ }
+ } else {
+ PBI->setMetadata(LLVMContext::MD_prof, NULL);
}
// Copy any debug value intrinsics into the end of PredBlock.
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
if (isa<DbgInfoIntrinsic>(*I))
I->clone()->insertBefore(PBI);
-
+
return true;
}
return false;
BasicBlock *BB = BI->getParent();
// If this block ends with a branch instruction, and if there is a
- // predecessor that ends on a branch of the same condition, make
+ // predecessor that ends on a branch of the same condition, make
// this conditional branch redundant.
if (PBI->getCondition() == BI->getCondition() &&
PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
if (BB->getSinglePredecessor()) {
// Turn this into a branch on constant.
bool CondIsTrue = PBI->getSuccessor(0) == BB;
- BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
+ BI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
CondIsTrue));
return true; // Nuke the branch on constant.
}
-
+
// Otherwise, if there are multiple predecessors, insert a PHI that merges
// in the constant and simplify the block result. Subsequent passes of
// simplifycfg will thread the block.
PBI->getCondition() == BI->getCondition() &&
PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
bool CondIsTrue = PBI->getSuccessor(0) == BB;
- NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
+ NewPN->addIncoming(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
CondIsTrue), P);
} else {
NewPN->addIncoming(BI->getCondition(), P);
}
}
-
+
BI->setCondition(NewPN);
return true;
}
}
-
+
// If this is a conditional branch in an empty block, and if any
// predecessors is a conditional branch to one of our destinations,
// fold the conditions into logical ops and one cond br.
if (&*BBI != BI)
return false;
-
+
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
if (CE->canTrap())
return false;
-
+
int PBIOp, BIOp;
if (PBI->getSuccessor(0) == BI->getSuccessor(0))
PBIOp = BIOp = 0;
PBIOp = BIOp = 1;
else
return false;
-
+
// Check to make sure that the other destination of this branch
// isn't BB itself. If so, this is an infinite loop that will
// keep getting unwound.
if (PBI->getSuccessor(PBIOp) == BB)
return false;
-
- // Do not perform this transformation if it would require
+
+ // Do not perform this transformation if it would require
// insertion of a large number of select instructions. For targets
// without predication/cmovs, this is a big pessimization.
BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
-
+
unsigned NumPhis = 0;
for (BasicBlock::iterator II = CommonDest->begin();
isa<PHINode>(II); ++II, ++NumPhis)
if (NumPhis > 2) // Disable this xform.
return false;
-
+
// Finally, if everything is ok, fold the branches to logical ops.
BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
-
+
DEBUG(dbgs() << "FOLDING BRs:" << *PBI->getParent()
<< "AND: " << *BI->getParent());
-
-
+
+
// If OtherDest *is* BB, then BB is a basic block with a single conditional
// branch in it, where one edge (OtherDest) goes back to itself but the other
// exits. We don't *know* that the program avoids the infinite loop
"infloop", BB->getParent());
BranchInst::Create(InfLoopBlock, InfLoopBlock);
OtherDest = InfLoopBlock;
- }
-
+ }
+
DEBUG(dbgs() << *PBI->getParent()->getParent());
// BI may have other predecessors. Because of this, we leave
// it alone, but modify PBI.
-
+
// Make sure we get to CommonDest on True&True directions.
Value *PBICond = PBI->getCondition();
- IRBuilder<> Builder(PBI);
+ IRBuilder<true, NoFolder> Builder(PBI);
if (PBIOp)
PBICond = Builder.CreateNot(PBICond, PBICond->getName()+".not");
// Merge the conditions.
Value *Cond = Builder.CreateOr(PBICond, BICond, "brmerge");
-
+
// Modify PBI to branch on the new condition to the new dests.
PBI->setCondition(Cond);
PBI->setSuccessor(0, CommonDest);
PBI->setSuccessor(1, OtherDest);
-
+
// OtherDest may have phi nodes. If so, add an entry from PBI's
// block that are identical to the entries for BI's block.
AddPredecessorToBlock(OtherDest, PBI->getParent(), BB);
-
+
// We know that the CommonDest already had an edge from PBI to
// it. If it has PHIs though, the PHIs may have different
// entries for BB and PBI's BB. If so, insert a select to make
PN->setIncomingValue(PBBIdx, NV);
}
}
-
+
DEBUG(dbgs() << "INTO: " << *PBI->getParent());
DEBUG(dbgs() << *PBI->getParent()->getParent());
-
+
// This basic block is probably dead. We know it has at least
// one fewer predecessor.
return true;
// Find the relevant condition and destinations.
Value *Condition = Select->getCondition();
- BasicBlock *TrueBB = SI->getSuccessor(SI->findCaseValue(TrueVal));
- BasicBlock *FalseBB = SI->getSuccessor(SI->findCaseValue(FalseVal));
+ BasicBlock *TrueBB = SI->findCaseValue(TrueVal).getCaseSuccessor();
+ BasicBlock *FalseBB = SI->findCaseValue(FalseVal).getCaseSuccessor();
// Perform the actual simplification.
return SimplifyTerminatorOnSelect(SI, Condition, TrueBB, FalseBB);
/// br label %end
/// end:
/// ... = phi i1 [ true, %entry ], [ %tmp, %DEFAULT ], [ true, %entry ]
-///
+///
/// We prefer to split the edge to 'end' so that there is a true/false entry to
/// the PHI, merging the third icmp into the switch.
static bool TryToSimplifyUncondBranchWithICmpInIt(ICmpInst *ICI,
Value *V = ICI->getOperand(0);
ConstantInt *Cst = cast<ConstantInt>(ICI->getOperand(1));
-
+
// The pattern we're looking for is where our only predecessor is a switch on
// 'V' and this block is the default case for the switch. In this case we can
// fold the compared value into the switch to simplify things.
BasicBlock *Pred = BB->getSinglePredecessor();
if (Pred == 0 || !isa<SwitchInst>(Pred->getTerminator())) return false;
-
+
SwitchInst *SI = cast<SwitchInst>(Pred->getTerminator());
if (SI->getCondition() != V)
return false;
-
+
// If BB is reachable on a non-default case, then we simply know the value of
// V in this block. Substitute it and constant fold the icmp instruction
// away.
ConstantInt *VVal = SI->findCaseDest(BB);
assert(VVal && "Should have a unique destination value");
ICI->setOperand(0, VVal);
-
+
if (Value *V = SimplifyInstruction(ICI, TD)) {
ICI->replaceAllUsesWith(V);
ICI->eraseFromParent();
// BB is now empty, so it is likely to simplify away.
return SimplifyCFG(BB) | true;
}
-
+
// Ok, the block is reachable from the default dest. If the constant we're
// comparing exists in one of the other edges, then we can constant fold ICI
// and zap it.
- if (SI->findCaseValue(Cst) != 0) {
+ if (SI->findCaseValue(Cst) != SI->case_default()) {
Value *V;
if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
V = ConstantInt::getFalse(BB->getContext());
else
V = ConstantInt::getTrue(BB->getContext());
-
+
ICI->replaceAllUsesWith(V);
ICI->eraseFromParent();
// BB is now empty, so it is likely to simplify away.
return SimplifyCFG(BB) | true;
}
-
+
// The use of the icmp has to be in the 'end' block, by the only PHI node in
// the block.
BasicBlock *SuccBlock = BB->getTerminator()->getSuccessor(0);
BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "switch.edge",
BB->getParent(), BB);
SI->addCase(Cst, NewBB);
-
+
// NewBB branches to the phi block, add the uncond branch and the phi entry.
Builder.SetInsertPoint(NewBB);
Builder.SetCurrentDebugLocation(SI->getDebugLoc());
IRBuilder<> &Builder) {
Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
if (Cond == 0) return false;
-
-
+
+
// Change br (X == 0 | X == 1), T, F into a switch instruction.
// If this is a bunch of seteq's or'd together, or if it's a bunch of
// 'setne's and'ed together, collect them.
bool TrueWhenEqual = true;
Value *ExtraCase = 0;
unsigned UsedICmps = 0;
-
+
if (Cond->getOpcode() == Instruction::Or) {
CompVal = GatherConstantCompares(Cond, Values, ExtraCase, TD, true,
UsedICmps);
UsedICmps);
TrueWhenEqual = false;
}
-
+
// If we didn't have a multiply compared value, fail.
if (CompVal == 0) return false;
// instruction can't handle, remove them now.
array_pod_sort(Values.begin(), Values.end(), ConstantIntSortPredicate);
Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
-
+
// If Extra was used, we require at least two switch values to do the
// transformation. A switch with one value is just an cond branch.
if (ExtraCase && Values.size() < 2) return false;
-
+
+ // TODO: Preserve branch weight metadata, similarly to how
+ // FoldValueComparisonIntoPredecessors preserves it.
+
// Figure out which block is which destination.
BasicBlock *DefaultBB = BI->getSuccessor(1);
BasicBlock *EdgeBB = BI->getSuccessor(0);
if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
-
+
BasicBlock *BB = BI->getParent();
-
+
DEBUG(dbgs() << "Converting 'icmp' chain with " << Values.size()
<< " cases into SWITCH. BB is:\n" << *BB);
-
+
// If there are any extra values that couldn't be folded into the switch
// then we evaluate them with an explicit branch first. Split the block
// right before the condbr to handle it.
Builder.CreateCondBr(ExtraCase, EdgeBB, NewBB);
else
Builder.CreateCondBr(ExtraCase, NewBB, EdgeBB);
-
+
OldTI->eraseFromParent();
-
+
// If there are PHI nodes in EdgeBB, then we need to add a new entry to them
// for the edge we just added.
AddPredecessorToBlock(EdgeBB, BB, NewBB);
-
+
DEBUG(dbgs() << " ** 'icmp' chain unhandled condition: " << *ExtraCase
<< "\nEXTRABB = " << *BB);
BB = NewBB;
TD->getIntPtrType(CompVal->getContext()),
"magicptr");
}
-
+
// Create the new switch instruction now.
SwitchInst *New = Builder.CreateSwitch(CompVal, DefaultBB, Values.size());
// Add all of the 'cases' to the switch instruction.
for (unsigned i = 0, e = Values.size(); i != e; ++i)
New->addCase(Values[i], EdgeBB);
-
+
// We added edges from PI to the EdgeBB. As such, if there were any
// PHI nodes in EdgeBB, they need entries to be added corresponding to
// the number of edges added.
for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
PN->addIncoming(InVal, BB);
}
-
+
// Erase the old branch instruction.
EraseTerminatorInstAndDCECond(BI);
-
+
DEBUG(dbgs() << " ** 'icmp' chain result is:\n" << *BB << '\n');
return true;
}
+bool SimplifyCFGOpt::SimplifyResume(ResumeInst *RI, IRBuilder<> &Builder) {
+ // If this is a trivial landing pad that just continues unwinding the caught
+ // exception then zap the landing pad, turning its invokes into calls.
+ BasicBlock *BB = RI->getParent();
+ LandingPadInst *LPInst = dyn_cast<LandingPadInst>(BB->getFirstNonPHI());
+ if (RI->getValue() != LPInst)
+ // Not a landing pad, or the resume is not unwinding the exception that
+ // caused control to branch here.
+ return false;
+
+ // Check that there are no other instructions except for debug intrinsics.
+ BasicBlock::iterator I = LPInst, E = RI;
+ while (++I != E)
+ if (!isa<DbgInfoIntrinsic>(I))
+ return false;
+
+ // Turn all invokes that unwind here into calls and delete the basic block.
+ for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE;) {
+ InvokeInst *II = cast<InvokeInst>((*PI++)->getTerminator());
+ SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
+ // Insert a call instruction before the invoke.
+ CallInst *Call = CallInst::Create(II->getCalledValue(), Args, "", II);
+ Call->takeName(II);
+ Call->setCallingConv(II->getCallingConv());
+ Call->setAttributes(II->getAttributes());
+ Call->setDebugLoc(II->getDebugLoc());
+
+ // Anything that used the value produced by the invoke instruction now uses
+ // the value produced by the call instruction. Note that we do this even
+ // for void functions and calls with no uses so that the callgraph edge is
+ // updated.
+ II->replaceAllUsesWith(Call);
+ BB->removePredecessor(II->getParent());
+
+ // Insert a branch to the normal destination right before the invoke.
+ BranchInst::Create(II->getNormalDest(), II);
+
+ // Finally, delete the invoke instruction!
+ II->eraseFromParent();
+ }
+
+ // The landingpad is now unreachable. Zap it.
+ BB->eraseFromParent();
+ return true;
+}
+
bool SimplifyCFGOpt::SimplifyReturn(ReturnInst *RI, IRBuilder<> &Builder) {
BasicBlock *BB = RI->getParent();
if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
-
+
// Find predecessors that end with branches.
SmallVector<BasicBlock*, 8> UncondBranchPreds;
SmallVector<BranchInst*, 8> CondBranchPreds;
CondBranchPreds.push_back(BI);
}
}
-
+
// If we found some, do the transformation!
if (!UncondBranchPreds.empty() && DupRet) {
while (!UncondBranchPreds.empty()) {
<< "INTO UNCOND BRANCH PRED: " << *Pred);
(void)FoldReturnIntoUncondBranch(RI, BB, Pred);
}
-
+
// If we eliminated all predecessors of the block, delete the block now.
if (pred_begin(BB) == pred_end(BB))
// We know there are no successors, so just nuke the block.
BB->eraseFromParent();
-
+
return true;
}
-
+
// Check out all of the conditional branches going to this return
// instruction. If any of them just select between returns, change the
// branch itself into a select/return pair.
while (!CondBranchPreds.empty()) {
BranchInst *BI = CondBranchPreds.pop_back_val();
-
+
// Check to see if the non-BB successor is also a return block.
if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
return false;
}
-bool SimplifyCFGOpt::SimplifyUnwind(UnwindInst *UI, IRBuilder<> &Builder) {
- // Check to see if the first instruction in this block is just an unwind.
- // If so, replace any invoke instructions which use this as an exception
- // destination with call instructions.
+bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
BasicBlock *BB = UI->getParent();
- if (!BB->getFirstNonPHIOrDbg()->isTerminator()) return false;
bool Changed = false;
- SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
- while (!Preds.empty()) {
- BasicBlock *Pred = Preds.back();
- InvokeInst *II = dyn_cast<InvokeInst>(Pred->getTerminator());
- if (II && II->getUnwindDest() == BB) {
- // Insert a new branch instruction before the invoke, because this
- // is now a fall through.
- Builder.SetInsertPoint(II);
- BranchInst *BI = Builder.CreateBr(II->getNormalDest());
- Pred->getInstList().remove(II); // Take out of symbol table
-
- // Insert the call now.
- SmallVector<Value*,8> Args(II->op_begin(), II->op_end()-3);
- Builder.SetInsertPoint(BI);
- CallInst *CI = Builder.CreateCall(II->getCalledValue(),
- Args.begin(), Args.end(),
- II->getName());
- CI->setCallingConv(II->getCallingConv());
- CI->setAttributes(II->getAttributes());
- // If the invoke produced a value, the Call now does instead.
- II->replaceAllUsesWith(CI);
- delete II;
- Changed = true;
- }
-
- Preds.pop_back();
- }
-
- // If this block is now dead (and isn't the entry block), remove it.
- if (pred_begin(BB) == pred_end(BB) &&
- BB != &BB->getParent()->getEntryBlock()) {
- // We know there are no successors, so just nuke the block.
- BB->eraseFromParent();
- return true;
- }
-
- return Changed;
-}
-bool SimplifyCFGOpt::SimplifyUnreachable(UnreachableInst *UI) {
- BasicBlock *BB = UI->getParent();
-
- bool Changed = false;
-
// If there are any instructions immediately before the unreachable that can
// be removed, do so.
while (UI != BB->begin()) {
BasicBlock::iterator BBI = UI;
--BBI;
- // Do not delete instructions that can have side effects, like calls
- // (which may never return) and volatile loads and stores.
+ // Do not delete instructions that can have side effects which might cause
+ // the unreachable to not be reachable; specifically, calls and volatile
+ // operations may have this effect.
if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
-
- if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
- if (SI->isVolatile())
- break;
-
- if (LoadInst *LI = dyn_cast<LoadInst>(BBI))
- if (LI->isVolatile())
+
+ if (BBI->mayHaveSideEffects()) {
+ if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
+ if (SI->isVolatile())
+ break;
+ } else if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
+ if (LI->isVolatile())
+ break;
+ } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(BBI)) {
+ if (RMWI->isVolatile())
+ break;
+ } else if (AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(BBI)) {
+ if (CXI->isVolatile())
+ break;
+ } else if (!isa<FenceInst>(BBI) && !isa<VAArgInst>(BBI) &&
+ !isa<LandingPadInst>(BBI)) {
break;
-
+ }
+ // Note that deleting LandingPad's here is in fact okay, although it
+ // involves a bit of subtle reasoning. If this inst is a LandingPad,
+ // all the predecessors of this block will be the unwind edges of Invokes,
+ // and we can therefore guarantee this block will be erased.
+ }
+
// Delete this instruction (any uses are guaranteed to be dead)
if (!BBI->use_empty())
BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
BBI->eraseFromParent();
Changed = true;
}
-
+
// If the unreachable instruction is the first in the block, take a gander
// at all of the predecessors of this instruction, and simplify them.
if (&BB->front() != UI) return Changed;
-
+
SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
TerminatorInst *TI = Preds[i]->getTerminator();
-
+ IRBuilder<> Builder(TI);
if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
if (BI->isUnconditional()) {
if (BI->getSuccessor(0) == BB) {
}
} else {
if (BI->getSuccessor(0) == BB) {
- BranchInst::Create(BI->getSuccessor(1), BI);
+ Builder.CreateBr(BI->getSuccessor(1));
EraseTerminatorInstAndDCECond(BI);
} else if (BI->getSuccessor(1) == BB) {
- BranchInst::Create(BI->getSuccessor(0), BI);
+ Builder.CreateBr(BI->getSuccessor(0));
EraseTerminatorInstAndDCECond(BI);
Changed = true;
}
}
} else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
- for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i)
- if (SI->getSuccessor(i) == BB) {
+ for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
+ i != e; ++i)
+ if (i.getCaseSuccessor() == BB) {
BB->removePredecessor(SI->getParent());
SI->removeCase(i);
--i; --e;
}
// If the default value is unreachable, figure out the most popular
// destination and make it the default.
- if (SI->getSuccessor(0) == BB) {
+ if (SI->getDefaultDest() == BB) {
std::map<BasicBlock*, std::pair<unsigned, unsigned> > Popularity;
- for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
- std::pair<unsigned, unsigned>& entry =
- Popularity[SI->getSuccessor(i)];
+ for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
+ i != e; ++i) {
+ std::pair<unsigned, unsigned> &entry =
+ Popularity[i.getCaseSuccessor()];
if (entry.first == 0) {
entry.first = 1;
- entry.second = i;
+ entry.second = i.getCaseIndex();
} else {
entry.first++;
}
BasicBlock *MaxBlock = 0;
for (std::map<BasicBlock*, std::pair<unsigned, unsigned> >::iterator
I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
- if (I->second.first > MaxPop ||
+ if (I->second.first > MaxPop ||
(I->second.first == MaxPop && MaxIndex > I->second.second)) {
MaxPop = I->second.first;
MaxIndex = I->second.second;
if (MaxBlock) {
// Make this the new default, allowing us to delete any explicit
// edges to it.
- SI->setSuccessor(0, MaxBlock);
+ SI->setDefaultDest(MaxBlock);
Changed = true;
-
+
// If MaxBlock has phinodes in it, remove MaxPop-1 entries from
// it.
if (isa<PHINode>(MaxBlock->begin()))
for (unsigned i = 0; i != MaxPop-1; ++i)
MaxBlock->removePredecessor(SI->getParent());
-
- for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i)
- if (SI->getSuccessor(i) == MaxBlock) {
+
+ for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
+ i != e; ++i)
+ if (i.getCaseSuccessor() == MaxBlock) {
SI->removeCase(i);
--i; --e;
}
if (II->getUnwindDest() == BB) {
// Convert the invoke to a call instruction. This would be a good
// place to note that the call does not throw though.
- BranchInst *BI = BranchInst::Create(II->getNormalDest(), II);
+ BranchInst *BI = Builder.CreateBr(II->getNormalDest());
II->removeFromParent(); // Take out of symbol table
-
+
// Insert the call now...
SmallVector<Value*, 8> Args(II->op_begin(), II->op_end()-3);
- CallInst *CI = CallInst::Create(II->getCalledValue(),
- Args.begin(), Args.end(),
- II->getName(), BI);
+ Builder.SetInsertPoint(BI);
+ CallInst *CI = Builder.CreateCall(II->getCalledValue(),
+ Args, II->getName());
CI->setCallingConv(II->getCallingConv());
CI->setAttributes(II->getAttributes());
// If the invoke produced a value, the call does now instead.
}
}
}
-
+
// If this block is now dead, remove it.
if (pred_begin(BB) == pred_end(BB) &&
BB != &BB->getParent()->getEntryBlock()) {
/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
/// integer range comparison into a sub, an icmp and a branch.
static bool TurnSwitchRangeIntoICmp(SwitchInst *SI, IRBuilder<> &Builder) {
- assert(SI->getNumCases() > 2 && "Degenerate switch?");
+ assert(SI->getNumCases() > 1 && "Degenerate switch?");
// Make sure all cases point to the same destination and gather the values.
SmallVector<ConstantInt *, 16> Cases;
- Cases.push_back(SI->getCaseValue(1));
- for (unsigned I = 2, E = SI->getNumCases(); I != E; ++I) {
- if (SI->getSuccessor(I-1) != SI->getSuccessor(I))
+ SwitchInst::CaseIt I = SI->case_begin();
+ Cases.push_back(I.getCaseValue());
+ SwitchInst::CaseIt PrevI = I++;
+ for (SwitchInst::CaseIt E = SI->case_end(); I != E; PrevI = I++) {
+ if (PrevI.getCaseSuccessor() != I.getCaseSuccessor())
return false;
- Cases.push_back(SI->getCaseValue(I));
+ Cases.push_back(I.getCaseValue());
}
- assert(Cases.size() == SI->getNumCases()-1 && "Not all cases gathered");
+ assert(Cases.size() == SI->getNumCases() && "Not all cases gathered");
// Sort the case values, then check if they form a range we can transform.
array_pod_sort(Cases.begin(), Cases.end(), ConstantIntSortPredicate);
}
Constant *Offset = ConstantExpr::getNeg(Cases.back());
- Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases()-1);
+ Constant *NumCases = ConstantInt::get(Offset->getType(), SI->getNumCases());
Value *Sub = SI->getCondition();
if (!Offset->isNullValue())
- Sub = BinaryOperator::CreateAdd(Sub, Offset, Sub->getName()+".off", SI);
- Value *Cmp = new ICmpInst(SI, ICmpInst::ICMP_ULT, Sub, NumCases, "switch");
- Builder.CreateCondBr(Cmp, SI->getSuccessor(1), SI->getDefaultDest());
+ Sub = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
+ Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
+ Builder.CreateCondBr(
+ Cmp, SI->case_begin().getCaseSuccessor(), SI->getDefaultDest());
// Prune obsolete incoming values off the successor's PHI nodes.
- for (BasicBlock::iterator BBI = SI->getSuccessor(1)->begin();
+ for (BasicBlock::iterator BBI = SI->case_begin().getCaseSuccessor()->begin();
isa<PHINode>(BBI); ++BBI) {
- for (unsigned I = 0, E = SI->getNumCases()-2; I != E; ++I)
+ for (unsigned I = 0, E = SI->getNumCases()-1; I != E; ++I)
cast<PHINode>(BBI)->removeIncomingValue(SI->getParent());
}
SI->eraseFromParent();
Value *Cond = SI->getCondition();
unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
- ComputeMaskedBits(Cond, APInt::getAllOnesValue(Bits), KnownZero, KnownOne);
+ ComputeMaskedBits(Cond, KnownZero, KnownOne);
// Gather dead cases.
SmallVector<ConstantInt*, 8> DeadCases;
- for (unsigned I = 1, E = SI->getNumCases(); I != E; ++I) {
- if ((SI->getCaseValue(I)->getValue() & KnownZero) != 0 ||
- (SI->getCaseValue(I)->getValue() & KnownOne) != KnownOne) {
- DeadCases.push_back(SI->getCaseValue(I));
+ for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
+ if ((I.getCaseValue()->getValue() & KnownZero) != 0 ||
+ (I.getCaseValue()->getValue() & KnownOne) != KnownOne) {
+ DeadCases.push_back(I.getCaseValue());
DEBUG(dbgs() << "SimplifyCFG: switch case '"
- << SI->getCaseValue(I)->getValue() << "' is dead.\n");
+ << I.getCaseValue() << "' is dead.\n");
}
}
// Remove dead cases from the switch.
for (unsigned I = 0, E = DeadCases.size(); I != E; ++I) {
- unsigned Case = SI->findCaseValue(DeadCases[I]);
+ SwitchInst::CaseIt Case = SI->findCaseValue(DeadCases[I]);
+ assert(Case != SI->case_default() &&
+ "Case was not found. Probably mistake in DeadCases forming.");
// Prune unused values from PHI nodes.
- SI->getSuccessor(Case)->removePredecessor(SI->getParent());
+ Case.getCaseSuccessor()->removePredecessor(SI->getParent());
SI->removeCase(Case);
}
return !DeadCases.empty();
}
+/// FindPHIForConditionForwarding - If BB would be eligible for simplification
+/// by TryToSimplifyUncondBranchFromEmptyBlock (i.e. it is empty and terminated
+/// by an unconditional branch), look at the phi node for BB in the successor
+/// block and see if the incoming value is equal to CaseValue. If so, return
+/// the phi node, and set PhiIndex to BB's index in the phi node.
+static PHINode *FindPHIForConditionForwarding(ConstantInt *CaseValue,
+ BasicBlock *BB,
+ int *PhiIndex) {
+ if (BB->getFirstNonPHIOrDbg() != BB->getTerminator())
+ return NULL; // BB must be empty to be a candidate for simplification.
+ if (!BB->getSinglePredecessor())
+ return NULL; // BB must be dominated by the switch.
+
+ BranchInst *Branch = dyn_cast<BranchInst>(BB->getTerminator());
+ if (!Branch || !Branch->isUnconditional())
+ return NULL; // Terminator must be unconditional branch.
+
+ BasicBlock *Succ = Branch->getSuccessor(0);
+
+ BasicBlock::iterator I = Succ->begin();
+ while (PHINode *PHI = dyn_cast<PHINode>(I++)) {
+ int Idx = PHI->getBasicBlockIndex(BB);
+ assert(Idx >= 0 && "PHI has no entry for predecessor?");
+
+ Value *InValue = PHI->getIncomingValue(Idx);
+ if (InValue != CaseValue) continue;
+
+ *PhiIndex = Idx;
+ return PHI;
+ }
+
+ return NULL;
+}
+
+/// ForwardSwitchConditionToPHI - Try to forward the condition of a switch
+/// instruction to a phi node dominated by the switch, if that would mean that
+/// some of the destination blocks of the switch can be folded away.
+/// Returns true if a change is made.
+static bool ForwardSwitchConditionToPHI(SwitchInst *SI) {
+ typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
+ ForwardingNodesMap ForwardingNodes;
+
+ for (SwitchInst::CaseIt I = SI->case_begin(), E = SI->case_end(); I != E; ++I) {
+ ConstantInt *CaseValue = I.getCaseValue();
+ BasicBlock *CaseDest = I.getCaseSuccessor();
+
+ int PhiIndex;
+ PHINode *PHI = FindPHIForConditionForwarding(CaseValue, CaseDest,
+ &PhiIndex);
+ if (!PHI) continue;
+
+ ForwardingNodes[PHI].push_back(PhiIndex);
+ }
+
+ bool Changed = false;
+
+ for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
+ E = ForwardingNodes.end(); I != E; ++I) {
+ PHINode *Phi = I->first;
+ SmallVector<int,4> &Indexes = I->second;
+
+ if (Indexes.size() < 2) continue;
+
+ for (size_t I = 0, E = Indexes.size(); I != E; ++I)
+ Phi->setIncomingValue(Indexes[I], SI->getCondition());
+ Changed = true;
+ }
+
+ return Changed;
+}
+
bool SimplifyCFGOpt::SimplifySwitch(SwitchInst *SI, IRBuilder<> &Builder) {
// If this switch is too complex to want to look at, ignore it.
if (!isValueEqualityComparison(SI))
if (EliminateDeadSwitchCases(SI))
return SimplifyCFG(BB) | true;
+ if (ForwardSwitchConditionToPHI(SI))
+ return SimplifyCFG(BB) | true;
+
return false;
}
bool SimplifyCFGOpt::SimplifyIndirectBr(IndirectBrInst *IBI) {
BasicBlock *BB = IBI->getParent();
bool Changed = false;
-
+
// Eliminate redundant destinations.
SmallPtrSet<Value *, 8> Succs;
for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
--i; --e;
Changed = true;
}
- }
+ }
if (IBI->getNumDestinations() == 0) {
// If the indirectbr has no successors, change it to unreachable.
EraseTerminatorInstAndDCECond(IBI);
return true;
}
-
+
if (IBI->getNumDestinations() == 1) {
// If the indirectbr has one successor, change it to a direct branch.
BranchInst::Create(IBI->getDestination(0), IBI);
EraseTerminatorInstAndDCECond(IBI);
return true;
}
-
+
if (SelectInst *SI = dyn_cast<SelectInst>(IBI->getAddress())) {
if (SimplifyIndirectBrOnSelect(IBI, SI))
return SimplifyCFG(BB) | true;
bool SimplifyCFGOpt::SimplifyUncondBranch(BranchInst *BI, IRBuilder<> &Builder){
BasicBlock *BB = BI->getParent();
-
+
// If the Terminator is the only non-phi instruction, simplify the block.
- BasicBlock::iterator I = BB->getFirstNonPHIOrDbg();
+ BasicBlock::iterator I = BB->getFirstNonPHIOrDbgOrLifetime();
if (I->isTerminator() && BB != &BB->getParent()->getEntryBlock() &&
TryToSimplifyUncondBranchFromEmptyBlock(BB))
return true;
-
+
// If the only instruction in the block is a seteq/setne comparison
// against a constant, try to simplify the block.
if (ICmpInst *ICI = dyn_cast<ICmpInst>(I))
if (ICI->isEquality() && isa<ConstantInt>(ICI->getOperand(1))) {
for (++I; isa<DbgInfoIntrinsic>(I); ++I)
;
- if (I->isTerminator()
- && TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder))
+ if (I->isTerminator() &&
+ TryToSimplifyUncondBranchWithICmpInIt(ICI, TD, Builder))
return true;
}
-
+
+ // If this basic block is ONLY a compare and a branch, and if a predecessor
+ // branches to us and our successor, fold the comparison into the
+ // predecessor and use logical operations to update the incoming value
+ // for PHI nodes in common successor.
+ if (FoldBranchToCommonDest(BI))
+ return SimplifyCFG(BB) | true;
return false;
}
bool SimplifyCFGOpt::SimplifyCondBranch(BranchInst *BI, IRBuilder<> &Builder) {
BasicBlock *BB = BI->getParent();
-
+
// Conditional branch
if (isValueEqualityComparison(BI)) {
// If we only have one predecessor, and if it is a branch on this value,
if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred, Builder))
return SimplifyCFG(BB) | true;
-
+
// This block must be empty, except for the setcond inst, if it exists.
// Ignore dbg intrinsics.
BasicBlock::iterator I = BB->begin();
return SimplifyCFG(BB) | true;
}
}
-
+
// Try to turn "br (X == 0 | X == 1), T, F" into a switch instruction.
if (SimplifyBranchOnICmpChain(BI, TD, Builder))
return true;
-
+
+ // If this basic block is ONLY a compare and a branch, and if a predecessor
+ // branches to us and one of our successors, fold the comparison into the
+ // predecessor and use logical operations to pick the right destination.
+ if (FoldBranchToCommonDest(BI))
+ return SimplifyCFG(BB) | true;
+
// We have a conditional branch to two blocks that are only reachable
// from BI. We know that the condbr dominates the two blocks, so see if
// there is any identical code in the "then" and "else" blocks. If so, we
// can hoist it up to the branching block.
if (BI->getSuccessor(0)->getSinglePredecessor() != 0) {
if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
- if (HoistThenElseCodeToIf(BI, Builder))
+ if (HoistThenElseCodeToIf(BI))
return SimplifyCFG(BB) | true;
} else {
// If Successor #1 has multiple preds, we may be able to conditionally
TerminatorInst *Succ0TI = BI->getSuccessor(0)->getTerminator();
if (Succ0TI->getNumSuccessors() == 1 &&
Succ0TI->getSuccessor(0) == BI->getSuccessor(1))
- if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0), Builder))
+ if (SpeculativelyExecuteBB(BI, BI->getSuccessor(0)))
return SimplifyCFG(BB) | true;
}
} else if (BI->getSuccessor(1)->getSinglePredecessor() != 0) {
TerminatorInst *Succ1TI = BI->getSuccessor(1)->getTerminator();
if (Succ1TI->getNumSuccessors() == 1 &&
Succ1TI->getSuccessor(0) == BI->getSuccessor(0))
- if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1), Builder))
+ if (SpeculativelyExecuteBB(BI, BI->getSuccessor(1)))
return SimplifyCFG(BB) | true;
}
-
+
// If this is a branch on a phi node in the current block, thread control
// through this block if any PHI node entries are constants.
if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
if (PN->getParent() == BI->getParent())
if (FoldCondBranchOnPHI(BI, TD))
return SimplifyCFG(BB) | true;
-
- // If this basic block is ONLY a setcc and a branch, and if a predecessor
- // branches to us and one of our successors, fold the setcc into the
- // predecessor and use logical operations to pick the right destination.
- if (FoldBranchToCommonDest(BI))
- return SimplifyCFG(BB) | true;
-
+
// Scan predecessor blocks for conditional branches.
for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
return false;
}
+/// Check if passing a value to an instruction will cause undefined behavior.
+static bool passingValueIsAlwaysUndefined(Value *V, Instruction *I) {
+ Constant *C = dyn_cast<Constant>(V);
+ if (!C)
+ return false;
+
+ if (!I->hasOneUse()) // Only look at single-use instructions, for compile time
+ return false;
+
+ if (C->isNullValue()) {
+ Instruction *Use = I->use_back();
+
+ // Now make sure that there are no instructions in between that can alter
+ // control flow (eg. calls)
+ for (BasicBlock::iterator i = ++BasicBlock::iterator(I); &*i != Use; ++i)
+ if (i == I->getParent()->end() || i->mayHaveSideEffects())
+ return false;
+
+ // Look through GEPs. A load from a GEP derived from NULL is still undefined
+ if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Use))
+ if (GEP->getPointerOperand() == I)
+ return passingValueIsAlwaysUndefined(V, GEP);
+
+ // Look through bitcasts.
+ if (BitCastInst *BC = dyn_cast<BitCastInst>(Use))
+ return passingValueIsAlwaysUndefined(V, BC);
+
+ // Load from null is undefined.
+ if (LoadInst *LI = dyn_cast<LoadInst>(Use))
+ return LI->getPointerAddressSpace() == 0;
+
+ // Store to null is undefined.
+ if (StoreInst *SI = dyn_cast<StoreInst>(Use))
+ return SI->getPointerAddressSpace() == 0 && SI->getPointerOperand() == I;
+ }
+ return false;
+}
+
+/// If BB has an incoming value that will always trigger undefined behavior
+/// (eg. null pointer dereference), remove the branch leading here.
+static bool removeUndefIntroducingPredecessor(BasicBlock *BB) {
+ for (BasicBlock::iterator i = BB->begin();
+ PHINode *PHI = dyn_cast<PHINode>(i); ++i)
+ for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
+ if (passingValueIsAlwaysUndefined(PHI->getIncomingValue(i), PHI)) {
+ TerminatorInst *T = PHI->getIncomingBlock(i)->getTerminator();
+ IRBuilder<> Builder(T);
+ if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
+ BB->removePredecessor(PHI->getIncomingBlock(i));
+ // Turn uncoditional branches into unreachables and remove the dead
+ // destination from conditional branches.
+ if (BI->isUnconditional())
+ Builder.CreateUnreachable();
+ else
+ Builder.CreateBr(BI->getSuccessor(0) == BB ? BI->getSuccessor(1) :
+ BI->getSuccessor(0));
+ BI->eraseFromParent();
+ return true;
+ }
+ // TODO: SwitchInst.
+ }
+
+ return false;
+}
+
bool SimplifyCFGOpt::run(BasicBlock *BB) {
bool Changed = false;
// Check to see if we can constant propagate this terminator instruction
// away...
- Changed |= ConstantFoldTerminator(BB);
+ Changed |= ConstantFoldTerminator(BB, true);
// Check for and eliminate duplicate PHI nodes in this block.
Changed |= EliminateDuplicatePHINodes(BB);
+ // Check for and remove branches that will always cause undefined behavior.
+ Changed |= removeUndefIntroducingPredecessor(BB);
+
// Merge basic blocks into their predecessor if there is only one distinct
// pred, and if there is only one distinct successor of the predecessor, and
// if there are no PHI nodes.
//
if (MergeBlockIntoPredecessor(BB))
return true;
-
+
IRBuilder<> Builder(BB);
// If there is a trivial two-entry PHI node in this basic block, and we can
// eliminate it, do so now.
if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
if (PN->getNumIncomingValues() == 2)
- Changed |= FoldTwoEntryPHINode(PN, TD, Builder);
+ Changed |= FoldTwoEntryPHINode(PN, TD);
Builder.SetInsertPoint(BB->getTerminator());
if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
}
} else if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
if (SimplifyReturn(RI, Builder)) return true;
+ } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
+ if (SimplifyResume(RI, Builder)) return true;
} else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
if (SimplifySwitch(SI, Builder)) return true;
} else if (UnreachableInst *UI =
dyn_cast<UnreachableInst>(BB->getTerminator())) {
if (SimplifyUnreachable(UI)) return true;
- } else if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
- if (SimplifyUnwind(UI, Builder)) return true;
} else if (IndirectBrInst *IBI =
dyn_cast<IndirectBrInst>(BB->getTerminator())) {
if (SimplifyIndirectBr(IBI)) return true;
projects / oota-llvm.git / blobdiff