#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Support/ConstantRange.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/NoFolder.h"
+#include "llvm/Support/PatternMatch.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include <algorithm>
#include <map>
#include <set>
using namespace llvm;
+using namespace PatternMatch;
static cl::opt<unsigned>
PHINodeFoldingThreshold("phi-node-folding-threshold", cl::Hidden, cl::init(1),
SinkCommon("simplifycfg-sink-common", cl::Hidden, cl::init(true),
cl::desc("Sink common instructions down to the end block"));
-static cl::opt<bool>
-HoistCondStores("simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
- cl::desc("Hoist conditional stores if an unconditional store preceeds"));
+static cl::opt<bool> HoistCondStores(
+ "simplifycfg-hoist-cond-stores", cl::Hidden, cl::init(true),
+ cl::desc("Hoist conditional stores if an unconditional store precedes"));
STATISTIC(NumBitMaps, "Number of switch instructions turned into bitmaps");
STATISTIC(NumLookupTables, "Number of switch instructions turned into lookup tables");
class SimplifyCFGOpt {
const TargetTransformInfo &TTI;
const DataLayout *const TD;
-
Value *isValueEqualityComparison(TerminatorInst *TI);
BasicBlock *GetValueEqualityComparisonCases(TerminatorInst *TI,
std::vector<ValueEqualityComparisonCase> &Cases);
PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
}
-
-/// GetIfCondition - Given a basic block (BB) with two predecessors (and at
-/// least one PHI node in it), check to see if the merge at this block is due
-/// to an "if condition". If so, return the boolean condition that determines
-/// which entry into BB will be taken. Also, return by references the block
-/// that will be entered from if the condition is true, and the block that will
-/// be entered if the condition is false.
-///
-/// This does no checking to see if the true/false blocks have large or unsavory
-/// instructions in them.
-static Value *GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue,
- BasicBlock *&IfFalse) {
- PHINode *SomePHI = cast<PHINode>(BB->begin());
- assert(SomePHI->getNumIncomingValues() == 2 &&
- "Function can only handle blocks with 2 predecessors!");
- BasicBlock *Pred1 = SomePHI->getIncomingBlock(0);
- BasicBlock *Pred2 = SomePHI->getIncomingBlock(1);
-
- // We can only handle branches. Other control flow will be lowered to
- // branches if possible anyway.
- BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator());
- BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator());
- if (Pred1Br == 0 || Pred2Br == 0)
- return 0;
-
- // Eliminate code duplication by ensuring that Pred1Br is conditional if
- // either are.
- if (Pred2Br->isConditional()) {
- // If both branches are conditional, we don't have an "if statement". In
- // reality, we could transform this case, but since the condition will be
- // required anyway, we stand no chance of eliminating it, so the xform is
- // probably not profitable.
- if (Pred1Br->isConditional())
- return 0;
-
- std::swap(Pred1, Pred2);
- std::swap(Pred1Br, Pred2Br);
- }
-
- if (Pred1Br->isConditional()) {
- // The only thing we have to watch out for here is to make sure that Pred2
- // doesn't have incoming edges from other blocks. If it does, the condition
- // 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 &&
- Pred1Br->getSuccessor(1) == Pred2) {
- IfTrue = Pred1;
- IfFalse = Pred2;
- } else if (Pred1Br->getSuccessor(0) == Pred2 &&
- Pred1Br->getSuccessor(1) == BB) {
- IfTrue = Pred2;
- IfFalse = Pred1;
- } else {
- // We know that one arm of the conditional goes to BB, so the other must
- // go somewhere unrelated, and this must not be an "if statement".
- return 0;
- }
-
- return Pred1Br->getCondition();
- }
-
- // Ok, if we got here, both predecessors end with an unconditional branch to
- // BB. Don't panic! If both blocks only have a single (identical)
- // predecessor, and THAT is a conditional branch, then we're all ok!
- BasicBlock *CommonPred = Pred1->getSinglePredecessor();
- if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor())
- return 0;
-
- // 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;
- IfFalse = Pred2;
- } else {
- IfTrue = Pred2;
- IfFalse = Pred1;
- }
- return BI->getCondition();
-}
-
/// ComputeSpeculationCost - 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.
// 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)) {
+ Value *RHSVal;
+ ConstantInt *RHSC;
+
if (ICI->getPredicate() == (isEQ ? ICmpInst::ICMP_EQ:ICmpInst::ICMP_NE)) {
+ // (x & ~2^x) == y --> x == y || x == y|2^x
+ // This undoes a transformation done by instcombine to fuse 2 compares.
+ if (match(ICI->getOperand(0),
+ m_And(m_Value(RHSVal), m_ConstantInt(RHSC)))) {
+ APInt Not = ~RHSC->getValue();
+ if (Not.isPowerOf2()) {
+ Vals.push_back(C);
+ Vals.push_back(
+ ConstantInt::get(C->getContext(), C->getValue() | Not));
+ UsedICmps++;
+ return RHSVal;
+ }
+ }
+
UsedICmps++;
Vals.push_back(C);
return I->getOperand(0);
ConstantRange Span =
ConstantRange::makeICmpRegion(ICI->getPredicate(), C->getValue());
+ // Shift the range if the compare is fed by an add. This is the range
+ // compare idiom as emitted by instcombine.
+ bool hasAdd =
+ match(I->getOperand(0), m_Add(m_Value(RHSVal), m_ConstantInt(RHSC)));
+ if (hasAdd)
+ Span = Span.subtract(RHSC->getValue());
+
// If this is an and/!= check then we want to optimize "x ugt 2" into
// x != 0 && x != 1.
if (!isEQ)
for (APInt Tmp = Span.getLower(); Tmp != Span.getUpper(); ++Tmp)
Vals.push_back(ConstantInt::get(V->getContext(), Tmp));
UsedICmps++;
- return I->getOperand(0);
+ return hasAdd ? RHSVal : I->getOperand(0);
}
return 0;
}
CV = ICI->getOperand(0);
// Unwrap any lossless ptrtoint cast.
- if (TD && CV && CV->getType() == TD->getIntPtrType(CV->getContext()))
- if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV))
- CV = PTII->getOperand(0);
+ if (TD && CV) {
+ if (PtrToIntInst *PTII = dyn_cast<PtrToIntInst>(CV)) {
+ Value *Ptr = PTII->getPointerOperand();
+ if (PTII->getType() == TD->getIntPtrType(Ptr->getType()))
+ CV = Ptr;
+ }
+ }
return CV;
}
};
}
-static int ConstantIntSortPredicate(const void *P1, const void *P2) {
- const ConstantInt *LHS = *(const ConstantInt*const*)P1;
- const ConstantInt *RHS = *(const ConstantInt*const*)P2;
+static int ConstantIntSortPredicate(ConstantInt *const *P1,
+ ConstantInt *const *P2) {
+ const ConstantInt *LHS = *P1;
+ const ConstantInt *RHS = *P2;
if (LHS->getValue().ult(RHS->getValue()))
return 1;
if (LHS->getValue() == RHS->getValue())
// Convert pointer to int before we switch.
if (CV->getType()->isPointerTy()) {
assert(TD && "Cannot switch on pointer without DataLayout");
- CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getContext()),
+ CV = Builder.CreatePtrToInt(CV, TD->getIntPtrType(CV->getType()),
"magicptr");
}
return true;
}
+/// \returns True if this block contains a CallInst with the NoDuplicate
+/// attribute.
+static bool HasNoDuplicateCall(const BasicBlock *BB) {
+ for (BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
+ const CallInst *CI = dyn_cast<CallInst>(I);
+ if (!CI)
+ continue;
+ if (CI->cannotDuplicate())
+ return true;
+ }
+ return false;
+}
+
/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
/// across this block.
static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
// Now we know that this block has multiple preds and two succs.
if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
+ if (HasNoDuplicateCall(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) {
// 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
- // out-of-order core by speculating them earlier.
- if (BonusInst) {
+ // out-of-order core by speculating them earlier. We also allow
+ // instructions that are used by the terminator's condition because it
+ // exposes more merging opportunities.
+ bool UsedByBranch = (BonusInst && BonusInst->hasOneUse() &&
+ *BonusInst->use_begin() == Cond);
+
+ if (BonusInst && !UsedByBranch) {
// 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;
- if (!isa<Constant>(V))
+ if (!isa<Constant>(V) && !isa<Argument>(V))
UsedValues.insert(V);
}
if (CompVal->getType()->isPointerTy()) {
assert(TD && "Cannot switch on pointer without DataLayout");
CompVal = Builder.CreatePtrToInt(CompVal,
- TD->getIntPtrType(CompVal->getContext()),
+ TD->getIntPtrType(CompVal->getType()),
"magicptr");
}
/// and use it to remove dead cases.
static bool EliminateDeadSwitchCases(SwitchInst *SI) {
Value *Cond = SI->getCondition();
- unsigned Bits = cast<IntegerType>(Cond->getType())->getBitWidth();
+ unsigned Bits = Cond->getType()->getIntegerBitWidth();
APInt KnownZero(Bits, 0), KnownOne(Bits, 0);
ComputeMaskedBits(Cond, KnownZero, KnownOne);
Case.getCaseSuccessor()->removePredecessor(SI->getParent());
SI->removeCase(Case);
}
- if (HasWeight) {
+ if (HasWeight && Weights.size() >= 2) {
SmallVector<uint32_t, 8> MDWeights(Weights.begin(), Weights.end());
SI->setMetadata(LLVMContext::MD_prof,
MDBuilder(SI->getParent()->getContext()).
for (ForwardingNodesMap::iterator I = ForwardingNodes.begin(),
E = ForwardingNodes.end(); I != E; ++I) {
PHINode *Phi = I->first;
- SmallVector<int,4> &Indexes = I->second;
+ SmallVectorImpl<int> &Indexes = I->second;
if (Indexes.size() < 2) continue;
/// simple instructions such as binary operations where both operands are
/// constant or can be replaced by constants from the ConstantPool. Returns the
/// resulting constant on success, 0 otherwise.
-static Constant *ConstantFold(Instruction *I,
- const SmallDenseMap<Value*, Constant*>& ConstantPool) {
- if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
- Constant *A = LookupConstant(BO->getOperand(0), ConstantPool);
- if (!A)
- return 0;
- Constant *B = LookupConstant(BO->getOperand(1), ConstantPool);
- if (!B)
- return 0;
- return ConstantExpr::get(BO->getOpcode(), A, B);
- }
-
- if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
- Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
- if (!A)
- return 0;
- Constant *B = LookupConstant(I->getOperand(1), ConstantPool);
- if (!B)
- return 0;
- return ConstantExpr::getCompare(Cmp->getPredicate(), A, B);
- }
-
+static Constant *
+ConstantFold(Instruction *I,
+ const SmallDenseMap<Value *, Constant *> &ConstantPool,
+ const DataLayout *DL) {
if (SelectInst *Select = dyn_cast<SelectInst>(I)) {
Constant *A = LookupConstant(Select->getCondition(), ConstantPool);
if (!A)
return 0;
}
- if (CastInst *Cast = dyn_cast<CastInst>(I)) {
- Constant *A = LookupConstant(I->getOperand(0), ConstantPool);
- if (!A)
+ SmallVector<Constant *, 4> COps;
+ for (unsigned N = 0, E = I->getNumOperands(); N != E; ++N) {
+ if (Constant *A = LookupConstant(I->getOperand(N), ConstantPool))
+ COps.push_back(A);
+ else
return 0;
- return ConstantExpr::getCast(Cast->getOpcode(), A, Cast->getDestTy());
}
- return 0;
+ if (CmpInst *Cmp = dyn_cast<CmpInst>(I))
+ return ConstantFoldCompareInstOperands(Cmp->getPredicate(), COps[0],
+ COps[1], DL);
+
+ return ConstantFoldInstOperands(I->getOpcode(), I->getType(), COps, DL);
}
/// GetCaseResults - Try to determine the resulting constant values in phi nodes
/// at the common destination basic block, *CommonDest, for one of the case
/// destionations CaseDest corresponding to value CaseVal (0 for the default
/// case), of a switch instruction SI.
-static bool GetCaseResults(SwitchInst *SI,
- ConstantInt *CaseVal,
- BasicBlock *CaseDest,
- BasicBlock **CommonDest,
- SmallVector<std::pair<PHINode*,Constant*>, 4> &Res) {
+static bool
+GetCaseResults(SwitchInst *SI,
+ ConstantInt *CaseVal,
+ BasicBlock *CaseDest,
+ BasicBlock **CommonDest,
+ SmallVectorImpl<std::pair<PHINode *, Constant *> > &Res,
+ const DataLayout *DL) {
// The block from which we enter the common destination.
BasicBlock *Pred = SI->getParent();
} else if (isa<DbgInfoIntrinsic>(I)) {
// Skip debug intrinsic.
continue;
- } else if (Constant *C = ConstantFold(I, ConstantPool)) {
+ } else if (Constant *C = ConstantFold(I, ConstantPool, DL)) {
// Instruction is side-effect free and constant.
ConstantPool.insert(std::make_pair(I, C));
} else {
Res.push_back(std::make_pair(PHI, ConstVal));
}
- return true;
+ return Res.size() > 0;
}
namespace {
SwitchLookupTable(Module &M,
uint64_t TableSize,
ConstantInt *Offset,
- const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
+ const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
Constant *DefaultValue,
const DataLayout *TD);
SwitchLookupTable::SwitchLookupTable(Module &M,
uint64_t TableSize,
ConstantInt *Offset,
- const SmallVector<std::pair<ConstantInt*, Constant*>, 4>& Values,
+ const SmallVectorImpl<std::pair<ConstantInt*, Constant*> >& Values,
Constant *DefaultValue,
const DataLayout *TD)
: SingleValue(0), BitMap(0), BitMapElementTy(0), Array(0) {
// If all values in the table are equal, this is that value.
SingleValue = Values.begin()->second;
+ Type *ValueType = Values.begin()->second->getType();
+
// Build up the table contents.
SmallVector<Constant*, 64> TableContents(TableSize);
for (size_t I = 0, E = Values.size(); I != E; ++I) {
ConstantInt *CaseVal = Values[I].first;
Constant *CaseRes = Values[I].second;
- assert(CaseRes->getType() == DefaultValue->getType());
+ assert(CaseRes->getType() == ValueType);
uint64_t Idx = (CaseVal->getValue() - Offset->getValue())
.getLimitedValue();
// Fill in any holes in the table with the default result.
if (Values.size() < TableSize) {
+ assert(DefaultValue && "Need a default value to fill the lookup table holes.");
+ assert(DefaultValue->getType() == ValueType);
for (uint64_t I = 0; I < TableSize; ++I) {
if (!TableContents[I])
TableContents[I] = DefaultValue;
}
// If the type is integer and the table fits in a register, build a bitmap.
- if (WouldFitInRegister(TD, TableSize, DefaultValue->getType())) {
- IntegerType *IT = cast<IntegerType>(DefaultValue->getType());
+ if (WouldFitInRegister(TD, TableSize, ValueType)) {
+ IntegerType *IT = cast<IntegerType>(ValueType);
APInt TableInt(TableSize * IT->getBitWidth(), 0);
for (uint64_t I = TableSize; I > 0; --I) {
TableInt <<= IT->getBitWidth();
}
// Store the table in an array.
- ArrayType *ArrayTy = ArrayType::get(DefaultValue->getType(), TableSize);
+ ArrayType *ArrayTy = ArrayType::get(ValueType, TableSize);
Constant *Initializer = ConstantArray::get(ArrayTy, TableContents);
Array = new GlobalVariable(M, ArrayTy, /*constant=*/ true,
// GEP needs a runtime relocation in PIC code. We should just build one big
// string and lookup indices into that.
- // Ignore the switch if the number of cases is too small.
- // This is similar to the check when building jump tables in
- // SelectionDAGBuilder::handleJTSwitchCase.
- // FIXME: Determine the best cut-off.
- if (SI->getNumCases() < 4)
+ // Ignore switches with less than three cases. Lookup tables will not make them
+ // faster, so we don't analyze them.
+ if (SI->getNumCases() < 3)
return false;
// Figure out the corresponding result for each case value and phi node in the
typedef SmallVector<std::pair<PHINode*, Constant*>, 4> ResultsTy;
ResultsTy Results;
if (!GetCaseResults(SI, CaseVal, CI.getCaseSuccessor(), &CommonDest,
- Results))
+ Results, TD))
return false;
// Append the result from this case to the list for each phi.
}
}
- // Get the resulting values for the default case.
+ // Keep track of the result types.
+ for (size_t I = 0, E = PHIs.size(); I != E; ++I) {
+ PHINode *PHI = PHIs[I];
+ ResultTypes[PHI] = ResultLists[PHI][0].second->getType();
+ }
+
+ uint64_t NumResults = ResultLists[PHIs[0]].size();
+ APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
+ uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
+ bool TableHasHoles = (NumResults < TableSize);
+
+ // If the table has holes, we need a constant result for the default case.
SmallVector<std::pair<PHINode*, Constant*>, 4> DefaultResultsList;
- if (!GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
- DefaultResultsList))
+ if (TableHasHoles && !GetCaseResults(SI, 0, SI->getDefaultDest(), &CommonDest,
+ DefaultResultsList, TD))
return false;
+
for (size_t I = 0, E = DefaultResultsList.size(); I != E; ++I) {
PHINode *PHI = DefaultResultsList[I].first;
Constant *Result = DefaultResultsList[I].second;
DefaultResults[PHI] = Result;
- ResultTypes[PHI] = Result->getType();
}
- APInt RangeSpread = MaxCaseVal->getValue() - MinCaseVal->getValue();
- uint64_t TableSize = RangeSpread.getLimitedValue() + 1;
if (!ShouldBuildLookupTable(SI, TableSize, TTI, TD, ResultTypes))
return false;
CommonDest->getParent(),
CommonDest);
- // Check whether the condition value is within the case range, and branch to
- // the new BB.
+ // Compute the table index value.
Builder.SetInsertPoint(SI);
Value *TableIndex = Builder.CreateSub(SI->getCondition(), MinCaseVal,
"switch.tableidx");
- Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
- MinCaseVal->getType(), TableSize));
- Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
+
+ // Compute the maximum table size representable by the integer type we are
+ // switching upon.
+ unsigned CaseSize = MinCaseVal->getType()->getPrimitiveSizeInBits();
+ uint64_t MaxTableSize = CaseSize > 63 ? UINT64_MAX : 1ULL << CaseSize;
+ assert(MaxTableSize >= TableSize &&
+ "It is impossible for a switch to have more entries than the max "
+ "representable value of its input integer type's size.");
+
+ // If we have a fully covered lookup table, unconditionally branch to the
+ // lookup table BB. Otherwise, check if the condition value is within the case
+ // range. If it is so, branch to the new BB. Otherwise branch to SI's default
+ // destination.
+ const bool GeneratingCoveredLookupTable = MaxTableSize == TableSize;
+ if (GeneratingCoveredLookupTable) {
+ Builder.CreateBr(LookupBB);
+ SI->getDefaultDest()->removePredecessor(SI->getParent());
+ } else {
+ Value *Cmp = Builder.CreateICmpULT(TableIndex, ConstantInt::get(
+ MinCaseVal->getType(), TableSize));
+ Builder.CreateCondBr(Cmp, LookupBB, SI->getDefaultDest());
+ }
// Populate the BB that does the lookups.
Builder.SetInsertPoint(LookupBB);
Builder.CreateBr(CommonDest);
// Remove the switch.
- for (unsigned i = 0; i < SI->getNumSuccessors(); ++i) {
+ for (unsigned i = 0, e = SI->getNumSuccessors(); i < e; ++i) {
BasicBlock *Succ = SI->getSuccessor(i);
- if (Succ == SI->getDefaultDest()) continue;
+
+ if (Succ == SI->getDefaultDest())
+ continue;
Succ->removePredecessor(SI->getParent());
}
SI->eraseFromParent();