#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/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/STLExtras.h"
#include "llvm/ADT/SetVector.h"
-#include "llvm/ADT/SmallVector.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);
bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI,
IRBuilder<> &Builder);
- bool SimplifyResume(ResumeInst *RI, 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);
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;
+}
+
/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
/// now be entries in it from the 'NewPred' block. The values that will be
/// flowing into the PHI nodes will be the same as those coming in from
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)
/// 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;
if (ThisVal != PredVal) return false; // Different predicates.
// 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.
(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");
// 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;
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;
}
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
// 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());
NewSuccessors.push_back(BBDefault);
}
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);
}
} else {
// 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,
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);
EraseTerminatorInstAndDCECond(PTI);
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;
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;
// 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;
// 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
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);
- }
- if (PBI->getSuccessor(1) == BB) {
- AddPredecessorToBlock(FalseDest, PredBlock, BB);
- PBI->setSuccessor(1, FalseDest);
+ 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;
}
// TODO: If BB is reachable from all paths through PredBlock, then we
// Merge probability data into PredBlock's branch.
APInt A, B, C, D;
- if (ExtractBranchMetadata(PBI, C, D) && ExtractBranchMetadata(BI, A, B)) {
+ 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
// we get:
// (a*c)% = A*C, (b+(a*d))% = A*D+B*C+B*D.
- bool Overflow1 = false, Overflow2 = false, Overflow3 = false;
- bool Overflow4 = false, Overflow5 = false, Overflow6 = false;
- APInt ProbTrue = A.umul_ov(C, Overflow1);
+ // In the event of overflow, we want to drop the LSB of the input
+ // probabilities.
+ unsigned BitsLost;
- APInt Tmp1 = A.umul_ov(D, Overflow2);
- APInt Tmp2 = B.umul_ov(C, Overflow3);
- APInt Tmp3 = B.umul_ov(D, Overflow4);
- APInt Tmp4 = Tmp1.uadd_ov(Tmp2, Overflow5);
- APInt ProbFalse = Tmp4.uadd_ov(Tmp3, Overflow6);
+ // Ignore overflow result on ProbTrue.
+ APInt ProbTrue = MultiplyAndLosePrecision(A, C, B, D, BitsLost);
- APInt GCD = APIntOps::GreatestCommonDivisor(ProbTrue, ProbFalse);
- ProbTrue = ProbTrue.udiv(GCD);
- ProbFalse = ProbFalse.udiv(GCD);
+ APInt Tmp1 = MultiplyAndLosePrecision(B, D, A, C, BitsLost);
+ if (BitsLost) {
+ ProbTrue = ProbTrue.lshr(BitsLost*2);
+ }
- if (Overflow1 || Overflow2 || Overflow3 || Overflow4 || Overflow5 ||
- Overflow6) {
- DEBUG(dbgs() << "Overflow recomputing branch weight on: " << *PBI
- << "when merging with: " << *BI);
- PBI->setMetadata(LLVMContext::MD_prof, NULL);
+ 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 {
- LLVMContext &Context = BI->getContext();
- Value *Ops[3];
- Ops[0] = BI->getMetadata(LLVMContext::MD_prof)->getOperand(0);
- Ops[1] = ConstantInt::get(Context, ProbTrue);
- Ops[2] = ConstantInt::get(Context, ProbFalse);
- PBI->setMetadata(LLVMContext::MD_prof, MDNode::get(Context, Ops));
+ PBI->setMetadata(LLVMContext::MD_prof, NULL);
}
} else {
PBI->setMetadata(LLVMContext::MD_prof, NULL);
// 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);
// 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());
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.
- 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, 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();
}
}
} 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) {
+ for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
+ i != e; ++i) {
std::pair<unsigned, unsigned> &entry =
- Popularity[SI->getSuccessor(i)];
+ Popularity[i.getCaseSuccessor()];
if (entry.first == 0) {
entry.first = 1;
- entry.second = i;
+ entry.second = i.getCaseIndex();
} else {
entry.first++;
}
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
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;
}
/// 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 = Builder.CreateAdd(Sub, Offset, Sub->getName()+".off");
Value *Cmp = Builder.CreateICmpULT(Sub, NumCases, "switch");
- Builder.CreateCondBr(Cmp, SI->getSuccessor(1), SI->getDefaultDest());
+ 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);
}
typedef DenseMap<PHINode*, SmallVector<int,4> > ForwardingNodesMap;
ForwardingNodesMap ForwardingNodes;
- for (unsigned I = 1; I < SI->getNumCases(); ++I) { // 0 is the default case.
- ConstantInt *CaseValue = SI->getCaseValue(I);
- BasicBlock *CaseDest = SI->getSuccessor(I);
+ 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,
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;
}
} else {
if (SimplifyCondBranch(BI, Builder)) return true;
}
- } else if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) {
- if (SimplifyResume(RI, Builder)) return true;
} 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;