#define DEBUG_TYPE "instcombine"
#include "llvm/Transforms/Scalar.h"
#include "InstCombine.h"
-#include "llvm/IntrinsicInst.h"
+#include "llvm-c/Initialization.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/StringSwitch.h"
#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/MemoryBuiltins.h"
-#include "llvm/Target/TargetData.h"
-#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/IntrinsicInst.h"
#include "llvm/Support/CFG.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/PatternMatch.h"
-#include "llvm/ADT/SmallPtrSet.h"
-#include "llvm/ADT/Statistic.h"
-#include "llvm-c/Initialization.h"
+#include "llvm/Support/ValueHandle.h"
+#include "llvm/Target/TargetLibraryInfo.h"
+#include "llvm/Transforms/Utils/Local.h"
#include <algorithm>
#include <climits>
using namespace llvm;
STATISTIC(NumFactor , "Number of factorizations");
STATISTIC(NumReassoc , "Number of reassociations");
+static cl::opt<bool> UnsafeFPShrink("enable-double-float-shrink", cl::Hidden,
+ cl::init(false),
+ cl::desc("Enable unsafe double to float "
+ "shrinking for math lib calls"));
+
// Initialization Routines
void llvm::initializeInstCombine(PassRegistry &Registry) {
initializeInstCombinerPass(Registry);
}
char InstCombiner::ID = 0;
-INITIALIZE_PASS(InstCombiner, "instcombine",
+INITIALIZE_PASS_BEGIN(InstCombiner, "instcombine",
+ "Combine redundant instructions", false, false)
+INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
+INITIALIZE_PASS_END(InstCombiner, "instcombine",
"Combine redundant instructions", false, false)
void InstCombiner::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
+ AU.addRequired<TargetLibraryInfo>();
}
+Value *InstCombiner::EmitGEPOffset(User *GEP) {
+ return llvm::EmitGEPOffset(Builder, *getDataLayout(), GEP);
+}
+
/// ShouldChangeType - Return true if it is desirable to convert a computation
/// from 'From' to 'To'. We don't want to convert from a legal to an illegal
/// type for example, or from a smaller to a larger illegal type.
bool InstCombiner::ShouldChangeType(Type *From, Type *To) const {
assert(From->isIntegerTy() && To->isIntegerTy());
-
+
// If we don't have TD, we don't know if the source/dest are legal.
if (!TD) return false;
-
+
unsigned FromWidth = From->getPrimitiveSizeInBits();
unsigned ToWidth = To->getPrimitiveSizeInBits();
bool FromLegal = TD->isLegalInteger(FromWidth);
bool ToLegal = TD->isLegalInteger(ToWidth);
-
+
// If this is a legal integer from type, and the result would be an illegal
// type, don't do the transformation.
if (FromLegal && !ToLegal)
return false;
-
+
// Otherwise, if both are illegal, do not increase the size of the result. We
// do allow things like i160 -> i64, but not i64 -> i160.
if (!FromLegal && !ToLegal && ToWidth > FromWidth)
return false;
-
+
return true;
}
+// Return true, if No Signed Wrap should be maintained for I.
+// The No Signed Wrap flag can be kept if the operation "B (I.getOpcode) C",
+// where both B and C should be ConstantInts, results in a constant that does
+// not overflow. This function only handles the Add and Sub opcodes. For
+// all other opcodes, the function conservatively returns false.
+static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) {
+ OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(&I);
+ if (!OBO || !OBO->hasNoSignedWrap()) {
+ return false;
+ }
+
+ // We reason about Add and Sub Only.
+ Instruction::BinaryOps Opcode = I.getOpcode();
+ if (Opcode != Instruction::Add &&
+ Opcode != Instruction::Sub) {
+ return false;
+ }
+
+ ConstantInt *CB = dyn_cast<ConstantInt>(B);
+ ConstantInt *CC = dyn_cast<ConstantInt>(C);
+
+ if (!CB || !CC) {
+ return false;
+ }
+
+ const APInt &BVal = CB->getValue();
+ const APInt &CVal = CC->getValue();
+ bool Overflow = false;
+
+ if (Opcode == Instruction::Add) {
+ BVal.sadd_ov(CVal, Overflow);
+ } else {
+ BVal.ssub_ov(CVal, Overflow);
+ }
+
+ return !Overflow;
+}
+
+/// Conservatively clears subclassOptionalData after a reassociation or
+/// commutation. We preserve fast-math flags when applicable as they can be
+/// preserved.
+static void ClearSubclassDataAfterReassociation(BinaryOperator &I) {
+ FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
+ if (!FPMO) {
+ I.clearSubclassOptionalData();
+ return;
+ }
+
+ FastMathFlags FMF = I.getFastMathFlags();
+ I.clearSubclassOptionalData();
+ I.setFastMathFlags(FMF);
+}
/// SimplifyAssociativeOrCommutative - This performs a few simplifications for
/// operators which are associative or commutative:
I.setOperand(1, V);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
- I.clearSubclassOptionalData();
+ if (MaintainNoSignedWrap(I, B, C) &&
+ (!Op0 || (isa<BinaryOperator>(Op0) && Op0->hasNoSignedWrap()))) {
+ // Note: this is only valid because SimplifyBinOp doesn't look at
+ // the operands to Op0.
+ I.clearSubclassOptionalData();
+ I.setHasNoSignedWrap(true);
+ } else {
+ ClearSubclassDataAfterReassociation(I);
+ }
+
Changed = true;
++NumReassoc;
continue;
I.setOperand(1, C);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
- I.clearSubclassOptionalData();
+ ClearSubclassDataAfterReassociation(I);
Changed = true;
++NumReassoc;
continue;
I.setOperand(1, B);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
- I.clearSubclassOptionalData();
+ ClearSubclassDataAfterReassociation(I);
Changed = true;
++NumReassoc;
continue;
I.setOperand(1, V);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
- I.clearSubclassOptionalData();
+ ClearSubclassDataAfterReassociation(I);
Changed = true;
++NumReassoc;
continue;
Constant *C2 = cast<Constant>(Op1->getOperand(1));
Constant *Folded = ConstantExpr::get(Opcode, C1, C2);
- Instruction *New = BinaryOperator::Create(Opcode, A, B);
+ BinaryOperator *New = BinaryOperator::Create(Opcode, A, B);
+ if (isa<FPMathOperator>(New)) {
+ FastMathFlags Flags = I.getFastMathFlags();
+ Flags &= Op0->getFastMathFlags();
+ Flags &= Op1->getFastMathFlags();
+ New->setFastMathFlags(Flags);
+ }
InsertNewInstWith(New, I);
New->takeName(Op1);
I.setOperand(0, New);
I.setOperand(1, Folded);
// Conservatively clear the optional flags, since they may not be
// preserved by the reassociation.
- I.clearSubclassOptionalData();
+ ClearSubclassDataAfterReassociation(I);
+
Changed = true;
continue;
}
if (ConstantInt *C = dyn_cast<ConstantInt>(V))
return ConstantExpr::getNeg(C);
- if (ConstantVector *C = dyn_cast<ConstantVector>(V))
+ if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
if (C->getType()->getElementType()->isIntegerTy())
return ConstantExpr::getNeg(C);
// instruction if the LHS is a constant negative zero (which is the 'negate'
// form).
//
-Value *InstCombiner::dyn_castFNegVal(Value *V) const {
- if (BinaryOperator::isFNeg(V))
+Value *InstCombiner::dyn_castFNegVal(Value *V, bool IgnoreZeroSign) const {
+ if (BinaryOperator::isFNeg(V, IgnoreZeroSign))
return BinaryOperator::getFNegArgument(V);
// Constants can be considered to be negated values if they can be folded.
if (ConstantFP *C = dyn_cast<ConstantFP>(V))
return ConstantExpr::getFNeg(C);
- if (ConstantVector *C = dyn_cast<ConstantVector>(V))
+ if (ConstantDataVector *C = dyn_cast<ConstantDataVector>(V))
if (C->getType()->getElementType()->isFloatingPointTy())
return ConstantExpr::getFNeg(C);
Value *Op0 = SO, *Op1 = ConstOperand;
if (!ConstIsRHS)
std::swap(Op0, Op1);
-
- if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
- return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
+
+ if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I)) {
+ Value *RI = IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
SO->getName()+".op");
+ Instruction *FPInst = dyn_cast<Instruction>(RI);
+ if (FPInst && isa<FPMathOperator>(FPInst))
+ FPInst->copyFastMathFlags(BO);
+ return RI;
+ }
if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
SO->getName()+".cmp");
if (SrcTy && SrcTy->getNumElements() != DestTy->getNumElements())
return 0;
}
-
+
Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, this);
Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, this);
unsigned NumPHIValues = PN->getNumIncomingValues();
if (NumPHIValues == 0)
return 0;
-
+
// We normally only transform phis with a single use. However, if a PHI has
// multiple uses and they are all the same operation, we can fold *all* of the
// uses into the PHI.
}
// Otherwise, we can replace *all* users with the new PHI we form.
}
-
+
// Check to see if all of the operands of the PHI are simple constants
// (constantint/constantfp/undef). If there is one non-constant value,
// remember the BB it is in. If there is more than one or if *it* is a PHI,
if (isa<PHINode>(InVal)) return 0; // Itself a phi.
if (NonConstBB) return 0; // More than one non-const value.
-
+
NonConstBB = PN->getIncomingBlock(i);
// If the InVal is an invoke at the end of the pred block, then we can't
if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
if (II->getParent() == NonConstBB)
return 0;
-
+
// If the incoming non-constant value is in I's block, we will remove one
// instruction, but insert another equivalent one, leading to infinite
// instcombine.
if (NonConstBB == I.getParent())
return 0;
}
-
+
// If there is exactly one non-constant value, we can insert a copy of the
// operation in that block. However, if this is a critical edge, we would be
// inserting the computation one some other paths (e.g. inside a loop). Only
PHINode *NewPN = PHINode::Create(I.getType(), PN->getNumIncomingValues());
InsertNewInstBefore(NewPN, *PN);
NewPN->takeName(PN);
-
+
// If we are going to have to insert a new computation, do so right before the
// predecessors terminator.
if (NonConstBB)
Builder->SetInsertPoint(NonConstBB->getTerminator());
-
+
// Next, add all of the operands to the PHI.
if (SelectInst *SI = dyn_cast<SelectInst>(&I)) {
// We only currently try to fold the condition of a select when it is a phi,
Value *TrueVInPred = TrueV->DoPHITranslation(PhiTransBB, ThisBB);
Value *FalseVInPred = FalseV->DoPHITranslation(PhiTransBB, ThisBB);
Value *InV = 0;
- if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
+ // Beware of ConstantExpr: it may eventually evaluate to getNullValue,
+ // even if currently isNullValue gives false.
+ Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i));
+ if (InC && !isa<ConstantExpr>(InC))
InV = InC->isNullValue() ? FalseVInPred : TrueVInPred;
else
InV = Builder->CreateSelect(PN->getIncomingValue(i),
PN->getIncomingValue(i), C, "phitmp");
NewPN->addIncoming(InV, PN->getIncomingBlock(i));
}
- } else {
+ } else {
CastInst *CI = cast<CastInst>(&I);
Type *RetTy = CI->getType();
for (unsigned i = 0; i != NumPHIValues; ++i) {
Value *InV;
if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i)))
InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
- else
+ else
InV = Builder->CreateCast(CI->getOpcode(),
PN->getIncomingValue(i), I.getType(), "phitmp");
NewPN->addIncoming(InV, PN->getIncomingBlock(i));
}
}
-
+
for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
UI != E; ) {
Instruction *User = cast<Instruction>(*UI++);
return ReplaceInstUsesWith(I, NewPN);
}
-/// FindElementAtOffset - Given a type and a constant offset, determine whether
-/// or not there is a sequence of GEP indices into the type that will land us at
-/// the specified offset. If so, fill them into NewIndices and return the
-/// resultant element type, otherwise return null.
-Type *InstCombiner::FindElementAtOffset(Type *Ty, int64_t Offset,
- SmallVectorImpl<Value*> &NewIndices) {
- if (!TD) return 0;
- if (!Ty->isSized()) return 0;
-
+/// FindElementAtOffset - Given a pointer type and a constant offset, determine
+/// whether or not there is a sequence of GEP indices into the pointed type that
+/// will land us at the specified offset. If so, fill them into NewIndices and
+/// return the resultant element type, otherwise return null.
+Type *InstCombiner::FindElementAtOffset(Type *PtrTy, int64_t Offset,
+ SmallVectorImpl<Value*> &NewIndices) {
+ assert(PtrTy->isPtrOrPtrVectorTy());
+
+ if (!TD)
+ return 0;
+
+ Type *Ty = PtrTy->getPointerElementType();
+ if (!Ty->isSized())
+ return 0;
+
// Start with the index over the outer type. Note that the type size
// might be zero (even if the offset isn't zero) if the indexed type
// is something like [0 x {int, int}]
- Type *IntPtrTy = TD->getIntPtrType(Ty->getContext());
+ Type *IntPtrTy = TD->getIntPtrType(PtrTy);
int64_t FirstIdx = 0;
if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
FirstIdx = Offset/TySize;
Offset -= FirstIdx*TySize;
-
+
// Handle hosts where % returns negative instead of values [0..TySize).
if (Offset < 0) {
--FirstIdx;
}
assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
}
-
+
NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
-
+
// Index into the types. If we fail, set OrigBase to null.
while (Offset) {
// Indexing into tail padding between struct/array elements.
if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
return 0;
-
+
if (StructType *STy = dyn_cast<StructType>(Ty)) {
const StructLayout *SL = TD->getStructLayout(STy);
assert(Offset < (int64_t)SL->getSizeInBytes() &&
"Offset must stay within the indexed type");
-
+
unsigned Elt = SL->getElementContainingOffset(Offset);
NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
Elt));
-
+
Offset -= SL->getElementOffset(Elt);
Ty = STy->getElementType(Elt);
} else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
return 0;
}
}
-
+
return Ty;
}
+static bool shouldMergeGEPs(GEPOperator &GEP, GEPOperator &Src) {
+ // If this GEP has only 0 indices, it is the same pointer as
+ // Src. If Src is not a trivial GEP too, don't combine
+ // the indices.
+ if (GEP.hasAllZeroIndices() && !Src.hasAllZeroIndices() &&
+ !Src.hasOneUse())
+ return false;
+ return true;
+}
+
+/// Descale - Return a value X such that Val = X * Scale, or null if none. If
+/// the multiplication is known not to overflow then NoSignedWrap is set.
+Value *InstCombiner::Descale(Value *Val, APInt Scale, bool &NoSignedWrap) {
+ assert(isa<IntegerType>(Val->getType()) && "Can only descale integers!");
+ assert(cast<IntegerType>(Val->getType())->getBitWidth() ==
+ Scale.getBitWidth() && "Scale not compatible with value!");
+
+ // If Val is zero or Scale is one then Val = Val * Scale.
+ if (match(Val, m_Zero()) || Scale == 1) {
+ NoSignedWrap = true;
+ return Val;
+ }
+
+ // If Scale is zero then it does not divide Val.
+ if (Scale.isMinValue())
+ return 0;
+
+ // Look through chains of multiplications, searching for a constant that is
+ // divisible by Scale. For example, descaling X*(Y*(Z*4)) by a factor of 4
+ // will find the constant factor 4 and produce X*(Y*Z). Descaling X*(Y*8) by
+ // a factor of 4 will produce X*(Y*2). The principle of operation is to bore
+ // down from Val:
+ //
+ // Val = M1 * X || Analysis starts here and works down
+ // M1 = M2 * Y || Doesn't descend into terms with more
+ // M2 = Z * 4 \/ than one use
+ //
+ // Then to modify a term at the bottom:
+ //
+ // Val = M1 * X
+ // M1 = Z * Y || Replaced M2 with Z
+ //
+ // Then to work back up correcting nsw flags.
+
+ // Op - the term we are currently analyzing. Starts at Val then drills down.
+ // Replaced with its descaled value before exiting from the drill down loop.
+ Value *Op = Val;
+
+ // Parent - initially null, but after drilling down notes where Op came from.
+ // In the example above, Parent is (Val, 0) when Op is M1, because M1 is the
+ // 0'th operand of Val.
+ std::pair<Instruction*, unsigned> Parent;
+
+ // RequireNoSignedWrap - Set if the transform requires a descaling at deeper
+ // levels that doesn't overflow.
+ bool RequireNoSignedWrap = false;
+
+ // logScale - log base 2 of the scale. Negative if not a power of 2.
+ int32_t logScale = Scale.exactLogBase2();
+ for (;; Op = Parent.first->getOperand(Parent.second)) { // Drill down
+
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
+ // If Op is a constant divisible by Scale then descale to the quotient.
+ APInt Quotient(Scale), Remainder(Scale); // Init ensures right bitwidth.
+ APInt::sdivrem(CI->getValue(), Scale, Quotient, Remainder);
+ if (!Remainder.isMinValue())
+ // Not divisible by Scale.
+ return 0;
+ // Replace with the quotient in the parent.
+ Op = ConstantInt::get(CI->getType(), Quotient);
+ NoSignedWrap = true;
+ break;
+ }
+
+ if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) {
+
+ if (BO->getOpcode() == Instruction::Mul) {
+ // Multiplication.
+ NoSignedWrap = BO->hasNoSignedWrap();
+ if (RequireNoSignedWrap && !NoSignedWrap)
+ return 0;
+
+ // There are three cases for multiplication: multiplication by exactly
+ // the scale, multiplication by a constant different to the scale, and
+ // multiplication by something else.
+ Value *LHS = BO->getOperand(0);
+ Value *RHS = BO->getOperand(1);
+
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
+ // Multiplication by a constant.
+ if (CI->getValue() == Scale) {
+ // Multiplication by exactly the scale, replace the multiplication
+ // by its left-hand side in the parent.
+ Op = LHS;
+ break;
+ }
+
+ // Otherwise drill down into the constant.
+ if (!Op->hasOneUse())
+ return 0;
+
+ Parent = std::make_pair(BO, 1);
+ continue;
+ }
+
+ // Multiplication by something else. Drill down into the left-hand side
+ // since that's where the reassociate pass puts the good stuff.
+ if (!Op->hasOneUse())
+ return 0;
+
+ Parent = std::make_pair(BO, 0);
+ continue;
+ }
+
+ if (logScale > 0 && BO->getOpcode() == Instruction::Shl &&
+ isa<ConstantInt>(BO->getOperand(1))) {
+ // Multiplication by a power of 2.
+ NoSignedWrap = BO->hasNoSignedWrap();
+ if (RequireNoSignedWrap && !NoSignedWrap)
+ return 0;
+
+ Value *LHS = BO->getOperand(0);
+ int32_t Amt = cast<ConstantInt>(BO->getOperand(1))->
+ getLimitedValue(Scale.getBitWidth());
+ // Op = LHS << Amt.
+
+ if (Amt == logScale) {
+ // Multiplication by exactly the scale, replace the multiplication
+ // by its left-hand side in the parent.
+ Op = LHS;
+ break;
+ }
+ if (Amt < logScale || !Op->hasOneUse())
+ return 0;
+
+ // Multiplication by more than the scale. Reduce the multiplying amount
+ // by the scale in the parent.
+ Parent = std::make_pair(BO, 1);
+ Op = ConstantInt::get(BO->getType(), Amt - logScale);
+ break;
+ }
+ }
+
+ if (!Op->hasOneUse())
+ return 0;
+
+ if (CastInst *Cast = dyn_cast<CastInst>(Op)) {
+ if (Cast->getOpcode() == Instruction::SExt) {
+ // Op is sign-extended from a smaller type, descale in the smaller type.
+ unsigned SmallSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
+ APInt SmallScale = Scale.trunc(SmallSize);
+ // Suppose Op = sext X, and we descale X as Y * SmallScale. We want to
+ // descale Op as (sext Y) * Scale. In order to have
+ // sext (Y * SmallScale) = (sext Y) * Scale
+ // some conditions need to hold however: SmallScale must sign-extend to
+ // Scale and the multiplication Y * SmallScale should not overflow.
+ if (SmallScale.sext(Scale.getBitWidth()) != Scale)
+ // SmallScale does not sign-extend to Scale.
+ return 0;
+ assert(SmallScale.exactLogBase2() == logScale);
+ // Require that Y * SmallScale must not overflow.
+ RequireNoSignedWrap = true;
+
+ // Drill down through the cast.
+ Parent = std::make_pair(Cast, 0);
+ Scale = SmallScale;
+ continue;
+ }
+
+ if (Cast->getOpcode() == Instruction::Trunc) {
+ // Op is truncated from a larger type, descale in the larger type.
+ // Suppose Op = trunc X, and we descale X as Y * sext Scale. Then
+ // trunc (Y * sext Scale) = (trunc Y) * Scale
+ // always holds. However (trunc Y) * Scale may overflow even if
+ // trunc (Y * sext Scale) does not, so nsw flags need to be cleared
+ // from this point up in the expression (see later).
+ if (RequireNoSignedWrap)
+ return 0;
+
+ // Drill down through the cast.
+ unsigned LargeSize = Cast->getSrcTy()->getPrimitiveSizeInBits();
+ Parent = std::make_pair(Cast, 0);
+ Scale = Scale.sext(LargeSize);
+ if (logScale + 1 == (int32_t)Cast->getType()->getPrimitiveSizeInBits())
+ logScale = -1;
+ assert(Scale.exactLogBase2() == logScale);
+ continue;
+ }
+ }
+
+ // Unsupported expression, bail out.
+ return 0;
+ }
+
+ // We know that we can successfully descale, so from here on we can safely
+ // modify the IR. Op holds the descaled version of the deepest term in the
+ // expression. NoSignedWrap is 'true' if multiplying Op by Scale is known
+ // not to overflow.
+
+ if (!Parent.first)
+ // The expression only had one term.
+ return Op;
+
+ // Rewrite the parent using the descaled version of its operand.
+ assert(Parent.first->hasOneUse() && "Drilled down when more than one use!");
+ assert(Op != Parent.first->getOperand(Parent.second) &&
+ "Descaling was a no-op?");
+ Parent.first->setOperand(Parent.second, Op);
+ Worklist.Add(Parent.first);
+
+ // Now work back up the expression correcting nsw flags. The logic is based
+ // on the following observation: if X * Y is known not to overflow as a signed
+ // multiplication, and Y is replaced by a value Z with smaller absolute value,
+ // then X * Z will not overflow as a signed multiplication either. As we work
+ // our way up, having NoSignedWrap 'true' means that the descaled value at the
+ // current level has strictly smaller absolute value than the original.
+ Instruction *Ancestor = Parent.first;
+ do {
+ if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Ancestor)) {
+ // If the multiplication wasn't nsw then we can't say anything about the
+ // value of the descaled multiplication, and we have to clear nsw flags
+ // from this point on up.
+ bool OpNoSignedWrap = BO->hasNoSignedWrap();
+ NoSignedWrap &= OpNoSignedWrap;
+ if (NoSignedWrap != OpNoSignedWrap) {
+ BO->setHasNoSignedWrap(NoSignedWrap);
+ Worklist.Add(Ancestor);
+ }
+ } else if (Ancestor->getOpcode() == Instruction::Trunc) {
+ // The fact that the descaled input to the trunc has smaller absolute
+ // value than the original input doesn't tell us anything useful about
+ // the absolute values of the truncations.
+ NoSignedWrap = false;
+ }
+ assert((Ancestor->getOpcode() != Instruction::SExt || NoSignedWrap) &&
+ "Failed to keep proper track of nsw flags while drilling down?");
+
+ if (Ancestor == Val)
+ // Got to the top, all done!
+ return Val;
+
+ // Move up one level in the expression.
+ assert(Ancestor->hasOneUse() && "Drilled down when more than one use!");
+ Ancestor = Ancestor->use_back();
+ } while (1);
+}
Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
SmallVector<Value*, 8> Ops(GEP.op_begin(), GEP.op_end());
// by multiples of a zero size type with zero.
if (TD) {
bool MadeChange = false;
- Type *IntPtrTy = TD->getIntPtrType(GEP.getContext());
+ Type *IntPtrTy = TD->getIntPtrType(GEP.getPointerOperandType());
gep_type_iterator GTI = gep_type_begin(GEP);
for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
MadeChange = true;
}
- if ((*I)->getType() != IntPtrTy) {
+ Type *IndexTy = (*I)->getType();
+ if (IndexTy != IntPtrTy) {
// If we are using a wider index than needed for this platform, shrink
// it to what we need. If narrower, sign-extend it to what we need.
// This explicit cast can make subsequent optimizations more obvious.
// getelementptr instructions into a single instruction.
//
if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
-
- // If this GEP has only 0 indices, it is the same pointer as
- // Src. If Src is not a trivial GEP too, don't combine
- // the indices.
- if (GEP.hasAllZeroIndices() && !Src->hasAllZeroIndices() &&
- !Src->hasOneUse())
+ if (!shouldMergeGEPs(*cast<GEPOperator>(&GEP), *Src))
return 0;
- // Note that if our source is a gep chain itself that we wait for that
+ // Note that if our source is a gep chain itself then we wait for that
// chain to be resolved before we perform this transformation. This
// avoids us creating a TON of code in some cases.
- //
- if (GetElementPtrInst *SrcGEP =
- dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
- if (SrcGEP->getNumOperands() == 2)
+ if (GEPOperator *SrcGEP =
+ dyn_cast<GEPOperator>(Src->getOperand(0)))
+ if (SrcGEP->getNumOperands() == 2 && shouldMergeGEPs(*Src, *SrcGEP))
return 0; // Wait until our source is folded to completion.
SmallVector<Value*, 8> Indices;
if (!Indices.empty())
return (GEP.isInBounds() && Src->isInBounds()) ?
- GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices.begin(),
- Indices.end(), GEP.getName()) :
- GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
- Indices.end(), GEP.getName());
+ GetElementPtrInst::CreateInBounds(Src->getOperand(0), Indices,
+ GEP.getName()) :
+ GetElementPtrInst::Create(Src->getOperand(0), Indices, GEP.getName());
+ }
+
+ // Canonicalize (gep i8* X, -(ptrtoint Y)) to (sub (ptrtoint X), (ptrtoint Y))
+ // The GEP pattern is emitted by the SCEV expander for certain kinds of
+ // pointer arithmetic.
+ if (TD && GEP.getNumIndices() == 1 &&
+ match(GEP.getOperand(1), m_Neg(m_PtrToInt(m_Value())))) {
+ unsigned AS = GEP.getPointerAddressSpace();
+ if (GEP.getType() == Builder->getInt8PtrTy(AS) &&
+ GEP.getOperand(1)->getType()->getScalarSizeInBits() ==
+ TD->getPointerSizeInBits(AS)) {
+ Operator *Index = cast<Operator>(GEP.getOperand(1));
+ Value *PtrToInt = Builder->CreatePtrToInt(PtrOp, Index->getType());
+ Value *NewSub = Builder->CreateSub(PtrToInt, Index->getOperand(1));
+ return CastInst::Create(Instruction::IntToPtr, NewSub, GEP.getType());
+ }
}
// Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
Value *StrippedPtr = PtrOp->stripPointerCasts();
- PointerType *StrippedPtrTy =cast<PointerType>(StrippedPtr->getType());
+ PointerType *StrippedPtrTy = dyn_cast<PointerType>(StrippedPtr->getType());
+
+ // We do not handle pointer-vector geps here.
+ if (!StrippedPtrTy)
+ return 0;
+
if (StrippedPtr != PtrOp &&
StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) {
// -> GEP i8* X, ...
SmallVector<Value*, 8> Idx(GEP.idx_begin()+1, GEP.idx_end());
GetElementPtrInst *Res =
- GetElementPtrInst::Create(StrippedPtr, Idx.begin(),
- Idx.end(), GEP.getName());
+ GetElementPtrInst::Create(StrippedPtr, Idx, GEP.getName());
Res->setIsInBounds(GEP.isInBounds());
return Res;
}
-
+
if (ArrayType *XATy =
dyn_cast<ArrayType>(StrippedPtrTy->getElementType())){
// GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
// %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
// into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
Type *SrcElTy = StrippedPtrTy->getElementType();
- Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
+ Type *ResElTy = PtrOp->getType()->getPointerElementType();
if (TD && SrcElTy->isArrayTy() &&
- TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
+ TD->getTypeAllocSize(SrcElTy->getArrayElementType()) ==
TD->getTypeAllocSize(ResElTy)) {
- Value *Idx[2];
- Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
- Idx[1] = GEP.getOperand(1);
+ Type *IdxType = TD->getIntPtrType(GEP.getType());
+ Value *Idx[2] = { Constant::getNullValue(IdxType), GEP.getOperand(1) };
Value *NewGEP = GEP.isInBounds() ?
Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()) :
Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
// V and GEP are both pointer types --> BitCast
return new BitCastInst(NewGEP, GEP.getType());
}
-
+
// Transform things like:
+ // %V = mul i64 %N, 4
+ // %t = getelementptr i8* bitcast (i32* %arr to i8*), i32 %V
+ // into: %t1 = getelementptr i32* %arr, i32 %N; bitcast
+ if (TD && ResElTy->isSized() && SrcElTy->isSized()) {
+ // Check that changing the type amounts to dividing the index by a scale
+ // factor.
+ uint64_t ResSize = TD->getTypeAllocSize(ResElTy);
+ uint64_t SrcSize = TD->getTypeAllocSize(SrcElTy);
+ if (ResSize && SrcSize % ResSize == 0) {
+ Value *Idx = GEP.getOperand(1);
+ unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
+ uint64_t Scale = SrcSize / ResSize;
+
+ // Earlier transforms ensure that the index has type IntPtrType, which
+ // considerably simplifies the logic by eliminating implicit casts.
+ assert(Idx->getType() == TD->getIntPtrType(GEP.getType()) &&
+ "Index not cast to pointer width?");
+
+ bool NSW;
+ if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
+ // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
+ // If the multiplication NewIdx * Scale may overflow then the new
+ // GEP may not be "inbounds".
+ Value *NewGEP = GEP.isInBounds() && NSW ?
+ Builder->CreateInBoundsGEP(StrippedPtr, NewIdx, GEP.getName()) :
+ Builder->CreateGEP(StrippedPtr, NewIdx, GEP.getName());
+ // The NewGEP must be pointer typed, so must the old one -> BitCast
+ return new BitCastInst(NewGEP, GEP.getType());
+ }
+ }
+ }
+
+ // Similarly, transform things like:
// getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
// (where tmp = 8*tmp2) into:
// getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
-
- if (TD && SrcElTy->isArrayTy() && ResElTy->isIntegerTy(8)) {
- uint64_t ArrayEltSize =
- TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
-
- // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
- // allow either a mul, shift, or constant here.
- Value *NewIdx = 0;
- ConstantInt *Scale = 0;
- if (ArrayEltSize == 1) {
- NewIdx = GEP.getOperand(1);
- Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
- } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
- NewIdx = ConstantInt::get(CI->getType(), 1);
- Scale = CI;
- } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
- if (Inst->getOpcode() == Instruction::Shl &&
- isa<ConstantInt>(Inst->getOperand(1))) {
- ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
- uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
- Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
- 1ULL << ShAmtVal);
- NewIdx = Inst->getOperand(0);
- } else if (Inst->getOpcode() == Instruction::Mul &&
- isa<ConstantInt>(Inst->getOperand(1))) {
- Scale = cast<ConstantInt>(Inst->getOperand(1));
- NewIdx = Inst->getOperand(0);
+ if (TD && ResElTy->isSized() && SrcElTy->isSized() &&
+ SrcElTy->isArrayTy()) {
+ // Check that changing to the array element type amounts to dividing the
+ // index by a scale factor.
+ uint64_t ResSize = TD->getTypeAllocSize(ResElTy);
+ uint64_t ArrayEltSize
+ = TD->getTypeAllocSize(SrcElTy->getArrayElementType());
+ if (ResSize && ArrayEltSize % ResSize == 0) {
+ Value *Idx = GEP.getOperand(1);
+ unsigned BitWidth = Idx->getType()->getPrimitiveSizeInBits();
+ uint64_t Scale = ArrayEltSize / ResSize;
+
+ // Earlier transforms ensure that the index has type IntPtrType, which
+ // considerably simplifies the logic by eliminating implicit casts.
+ assert(Idx->getType() == TD->getIntPtrType(GEP.getType()) &&
+ "Index not cast to pointer width?");
+
+ bool NSW;
+ if (Value *NewIdx = Descale(Idx, APInt(BitWidth, Scale), NSW)) {
+ // Successfully decomposed Idx as NewIdx * Scale, form a new GEP.
+ // If the multiplication NewIdx * Scale may overflow then the new
+ // GEP may not be "inbounds".
+ Value *Off[2] = {
+ Constant::getNullValue(TD->getIntPtrType(GEP.getType())),
+ NewIdx
+ };
+
+ Value *NewGEP = GEP.isInBounds() && NSW ?
+ Builder->CreateInBoundsGEP(StrippedPtr, Off, GEP.getName()) :
+ Builder->CreateGEP(StrippedPtr, Off, GEP.getName());
+ // The NewGEP must be pointer typed, so must the old one -> BitCast
+ return new BitCastInst(NewGEP, GEP.getType());
}
}
-
- // If the index will be to exactly the right offset with the scale taken
- // out, perform the transformation. Note, we don't know whether Scale is
- // signed or not. We'll use unsigned version of division/modulo
- // operation after making sure Scale doesn't have the sign bit set.
- if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
- Scale->getZExtValue() % ArrayEltSize == 0) {
- Scale = ConstantInt::get(Scale->getType(),
- Scale->getZExtValue() / ArrayEltSize);
- if (Scale->getZExtValue() != 1) {
- Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
- false /*ZExt*/);
- NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
- }
-
- // Insert the new GEP instruction.
- Value *Idx[2];
- Idx[0] = Constant::getNullValue(Type::getInt32Ty(GEP.getContext()));
- Idx[1] = NewIdx;
- Value *NewGEP = GEP.isInBounds() ?
- Builder->CreateInBoundsGEP(StrippedPtr, Idx, GEP.getName()):
- Builder->CreateGEP(StrippedPtr, Idx, GEP.getName());
- // The NewGEP must be pointer typed, so must the old one -> BitCast
- return new BitCastInst(NewGEP, GEP.getType());
- }
}
}
}
+ if (!TD)
+ return 0;
+
/// See if we can simplify:
/// X = bitcast A* to B*
/// Y = gep X, <...constant indices...>
/// into a gep of the original struct. This is important for SROA and alias
/// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
- if (TD &&
- !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices() &&
+ Value *Operand = BCI->getOperand(0);
+ PointerType *OpType = cast<PointerType>(Operand->getType());
+ unsigned OffsetBits = TD->getPointerTypeSizeInBits(OpType);
+ APInt Offset(OffsetBits, 0);
+ if (!isa<BitCastInst>(Operand) &&
+ GEP.accumulateConstantOffset(*TD, Offset) &&
StrippedPtrTy->getAddressSpace() == GEP.getPointerAddressSpace()) {
- // Determine how much the GEP moves the pointer. We are guaranteed to get
- // a constant back from EmitGEPOffset.
- ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP));
- int64_t Offset = OffsetV->getSExtValue();
-
// If this GEP instruction doesn't move the pointer, just replace the GEP
// with a bitcast of the real input to the dest type.
- if (Offset == 0) {
+ if (!Offset) {
// If the bitcast is of an allocation, and the allocation will be
// converted to match the type of the cast, don't touch this.
- if (isa<AllocaInst>(BCI->getOperand(0)) ||
- isMalloc(BCI->getOperand(0))) {
+ if (isa<AllocaInst>(Operand) || isAllocationFn(Operand, TLI)) {
// See if the bitcast simplifies, if so, don't nuke this GEP yet.
if (Instruction *I = visitBitCast(*BCI)) {
if (I != BCI) {
return &GEP;
}
}
- return new BitCastInst(BCI->getOperand(0), GEP.getType());
+ return new BitCastInst(Operand, GEP.getType());
}
-
+
// Otherwise, if the offset is non-zero, we need to find out if there is a
// field at Offset in 'A's type. If so, we can pull the cast through the
// GEP.
SmallVector<Value*, 8> NewIndices;
- Type *InTy =
- cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
- if (FindElementAtOffset(InTy, Offset, NewIndices)) {
+ if (FindElementAtOffset(OpType, Offset.getSExtValue(), NewIndices)) {
Value *NGEP = GEP.isInBounds() ?
- Builder->CreateInBoundsGEP(BCI->getOperand(0), NewIndices) :
- Builder->CreateGEP(BCI->getOperand(0), NewIndices);
-
+ Builder->CreateInBoundsGEP(Operand, NewIndices) :
+ Builder->CreateGEP(Operand, NewIndices);
+
if (NGEP->getType() == GEP.getType())
return ReplaceInstUsesWith(GEP, NGEP);
NGEP->takeName(&GEP);
return new BitCastInst(NGEP, GEP.getType());
}
}
- }
-
+ }
+
return 0;
}
+static bool
+isAllocSiteRemovable(Instruction *AI, SmallVectorImpl<WeakVH> &Users,
+ const TargetLibraryInfo *TLI) {
+ SmallVector<Instruction*, 4> Worklist;
+ Worklist.push_back(AI);
+ do {
+ Instruction *PI = Worklist.pop_back_val();
+ for (Value::use_iterator UI = PI->use_begin(), UE = PI->use_end(); UI != UE;
+ ++UI) {
+ Instruction *I = cast<Instruction>(*UI);
+ switch (I->getOpcode()) {
+ default:
+ // Give up the moment we see something we can't handle.
+ return false;
-static bool IsOnlyNullComparedAndFreed(const Value &V) {
- for (Value::const_use_iterator UI = V.use_begin(), UE = V.use_end();
- UI != UE; ++UI) {
- const User *U = *UI;
- if (isFreeCall(U))
- continue;
- if (const ICmpInst *ICI = dyn_cast<ICmpInst>(U))
- if (ICI->isEquality() && isa<ConstantPointerNull>(ICI->getOperand(1)))
+ case Instruction::BitCast:
+ case Instruction::GetElementPtr:
+ Users.push_back(I);
+ Worklist.push_back(I);
continue;
- return false;
- }
+
+ case Instruction::ICmp: {
+ ICmpInst *ICI = cast<ICmpInst>(I);
+ // We can fold eq/ne comparisons with null to false/true, respectively.
+ if (!ICI->isEquality() || !isa<ConstantPointerNull>(ICI->getOperand(1)))
+ return false;
+ Users.push_back(I);
+ continue;
+ }
+
+ case Instruction::Call:
+ // Ignore no-op and store intrinsics.
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
+ switch (II->getIntrinsicID()) {
+ default:
+ return false;
+
+ case Intrinsic::memmove:
+ case Intrinsic::memcpy:
+ case Intrinsic::memset: {
+ MemIntrinsic *MI = cast<MemIntrinsic>(II);
+ if (MI->isVolatile() || MI->getRawDest() != PI)
+ return false;
+ }
+ // fall through
+ case Intrinsic::dbg_declare:
+ case Intrinsic::dbg_value:
+ case Intrinsic::invariant_start:
+ case Intrinsic::invariant_end:
+ case Intrinsic::lifetime_start:
+ case Intrinsic::lifetime_end:
+ case Intrinsic::objectsize:
+ Users.push_back(I);
+ continue;
+ }
+ }
+
+ if (isFreeCall(I, TLI)) {
+ Users.push_back(I);
+ continue;
+ }
+ return false;
+
+ case Instruction::Store: {
+ StoreInst *SI = cast<StoreInst>(I);
+ if (SI->isVolatile() || SI->getPointerOperand() != PI)
+ return false;
+ Users.push_back(I);
+ continue;
+ }
+ }
+ llvm_unreachable("missing a return?");
+ }
+ } while (!Worklist.empty());
return true;
}
-Instruction *InstCombiner::visitMalloc(Instruction &MI) {
+Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
// If we have a malloc call which is only used in any amount of comparisons
// to null and free calls, delete the calls and replace the comparisons with
// true or false as appropriate.
- if (IsOnlyNullComparedAndFreed(MI)) {
- for (Value::use_iterator UI = MI.use_begin(), UE = MI.use_end();
- UI != UE;) {
- // We can assume that every remaining use is a free call or an icmp eq/ne
- // to null, so the cast is safe.
- Instruction *I = cast<Instruction>(*UI);
-
- // Early increment here, as we're about to get rid of the user.
- ++UI;
-
- if (isFreeCall(I)) {
- EraseInstFromFunction(*cast<CallInst>(I));
- continue;
+ SmallVector<WeakVH, 64> Users;
+ if (isAllocSiteRemovable(&MI, Users, TLI)) {
+ for (unsigned i = 0, e = Users.size(); i != e; ++i) {
+ Instruction *I = cast_or_null<Instruction>(&*Users[i]);
+ if (!I) continue;
+
+ if (ICmpInst *C = dyn_cast<ICmpInst>(I)) {
+ ReplaceInstUsesWith(*C,
+ ConstantInt::get(Type::getInt1Ty(C->getContext()),
+ C->isFalseWhenEqual()));
+ } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
+ ReplaceInstUsesWith(*I, UndefValue::get(I->getType()));
+ } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
+ if (II->getIntrinsicID() == Intrinsic::objectsize) {
+ ConstantInt *CI = cast<ConstantInt>(II->getArgOperand(1));
+ uint64_t DontKnow = CI->isZero() ? -1ULL : 0;
+ ReplaceInstUsesWith(*I, ConstantInt::get(I->getType(), DontKnow));
+ }
}
- // Again, the cast is safe.
- ICmpInst *C = cast<ICmpInst>(I);
- ReplaceInstUsesWith(*C, ConstantInt::get(Type::getInt1Ty(C->getContext()),
- C->isFalseWhenEqual()));
- EraseInstFromFunction(*C);
+ EraseInstFromFunction(*I);
+ }
+
+ if (InvokeInst *II = dyn_cast<InvokeInst>(&MI)) {
+ // Replace invoke with a NOP intrinsic to maintain the original CFG
+ Module *M = II->getParent()->getParent()->getParent();
+ Function *F = Intrinsic::getDeclaration(M, Intrinsic::donothing);
+ InvokeInst::Create(F, II->getNormalDest(), II->getUnwindDest(),
+ None, "", II->getParent());
}
return EraseInstFromFunction(MI);
}
return 0;
}
+/// \brief Move the call to free before a NULL test.
+///
+/// Check if this free is accessed after its argument has been test
+/// against NULL (property 0).
+/// If yes, it is legal to move this call in its predecessor block.
+///
+/// The move is performed only if the block containing the call to free
+/// will be removed, i.e.:
+/// 1. it has only one predecessor P, and P has two successors
+/// 2. it contains the call and an unconditional branch
+/// 3. its successor is the same as its predecessor's successor
+///
+/// The profitability is out-of concern here and this function should
+/// be called only if the caller knows this transformation would be
+/// profitable (e.g., for code size).
+static Instruction *
+tryToMoveFreeBeforeNullTest(CallInst &FI) {
+ Value *Op = FI.getArgOperand(0);
+ BasicBlock *FreeInstrBB = FI.getParent();
+ BasicBlock *PredBB = FreeInstrBB->getSinglePredecessor();
+
+ // Validate part of constraint #1: Only one predecessor
+ // FIXME: We can extend the number of predecessor, but in that case, we
+ // would duplicate the call to free in each predecessor and it may
+ // not be profitable even for code size.
+ if (!PredBB)
+ return 0;
+
+ // Validate constraint #2: Does this block contains only the call to
+ // free and an unconditional branch?
+ // FIXME: We could check if we can speculate everything in the
+ // predecessor block
+ if (FreeInstrBB->size() != 2)
+ return 0;
+ BasicBlock *SuccBB;
+ if (!match(FreeInstrBB->getTerminator(), m_UnconditionalBr(SuccBB)))
+ return 0;
+
+ // Validate the rest of constraint #1 by matching on the pred branch.
+ TerminatorInst *TI = PredBB->getTerminator();
+ BasicBlock *TrueBB, *FalseBB;
+ ICmpInst::Predicate Pred;
+ if (!match(TI, m_Br(m_ICmp(Pred, m_Specific(Op), m_Zero()), TrueBB, FalseBB)))
+ return 0;
+ if (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE)
+ return 0;
+
+ // Validate constraint #3: Ensure the null case just falls through.
+ if (SuccBB != (Pred == ICmpInst::ICMP_EQ ? TrueBB : FalseBB))
+ return 0;
+ assert(FreeInstrBB == (Pred == ICmpInst::ICMP_EQ ? FalseBB : TrueBB) &&
+ "Broken CFG: missing edge from predecessor to successor");
+
+ FI.moveBefore(TI);
+ return &FI;
+}
Instruction *InstCombiner::visitFree(CallInst &FI) {
UndefValue::get(Type::getInt1PtrTy(FI.getContext())));
return EraseInstFromFunction(FI);
}
-
+
// If we have 'free null' delete the instruction. This can happen in stl code
// when lots of inlining happens.
if (isa<ConstantPointerNull>(Op))
return EraseInstFromFunction(FI);
+ // If we optimize for code size, try to move the call to free before the null
+ // test so that simplify cfg can remove the empty block and dead code
+ // elimination the branch. I.e., helps to turn something like:
+ // if (foo) free(foo);
+ // into
+ // free(foo);
+ if (MinimizeSize)
+ if (Instruction *I = tryToMoveFreeBeforeNullTest(FI))
+ return I;
+
return 0;
}
!isa<Constant>(X)) {
// Swap Destinations and condition...
BI.setCondition(X);
- BI.setSuccessor(0, FalseDest);
- BI.setSuccessor(1, TrueDest);
+ BI.swapSuccessors();
return &BI;
}
- // Cannonicalize fcmp_one -> fcmp_oeq
+ // Canonicalize fcmp_one -> fcmp_oeq
FCmpInst::Predicate FPred; Value *Y;
- if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
+ if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
TrueDest, FalseDest)) &&
BI.getCondition()->hasOneUse())
if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
FPred == FCmpInst::FCMP_OGE) {
FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
-
+
// Swap Destinations and condition.
- BI.setSuccessor(0, FalseDest);
- BI.setSuccessor(1, TrueDest);
+ BI.swapSuccessors();
Worklist.Add(Cond);
return &BI;
}
- // Cannonicalize icmp_ne -> icmp_eq
+ // Canonicalize icmp_ne -> icmp_eq
ICmpInst::Predicate IPred;
if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
TrueDest, FalseDest)) &&
ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
// Swap Destinations and condition.
- BI.setSuccessor(0, FalseDest);
- BI.setSuccessor(1, TrueDest);
+ BI.swapSuccessors();
Worklist.Add(Cond);
return &BI;
}
if (I->getOpcode() == Instruction::Add)
if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
// change 'switch (X+4) case 1:' into 'switch (X) case -3'
- for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
- SI.setOperand(i,
- ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
- AddRHS));
- SI.setOperand(0, I->getOperand(0));
+ // Skip the first item since that's the default case.
+ for (SwitchInst::CaseIt i = SI.case_begin(), e = SI.case_end();
+ i != e; ++i) {
+ ConstantInt* CaseVal = i.getCaseValue();
+ Constant* NewCaseVal = ConstantExpr::getSub(cast<Constant>(CaseVal),
+ AddRHS);
+ assert(isa<ConstantInt>(NewCaseVal) &&
+ "Result of expression should be constant");
+ i.setValue(cast<ConstantInt>(NewCaseVal));
+ }
+ SI.setCondition(I->getOperand(0));
Worklist.Add(I);
return &SI;
}
return ReplaceInstUsesWith(EV, Agg);
if (Constant *C = dyn_cast<Constant>(Agg)) {
- if (isa<UndefValue>(C))
- return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
-
- if (isa<ConstantAggregateZero>(C))
- return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
-
- if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
- // Extract the element indexed by the first index out of the constant
- Value *V = C->getOperand(*EV.idx_begin());
- if (EV.getNumIndices() > 1)
- // Extract the remaining indices out of the constant indexed by the
- // first index
- return ExtractValueInst::Create(V, EV.getIndices().slice(1));
- else
- return ReplaceInstUsesWith(EV, V);
+ if (Constant *C2 = C->getAggregateElement(*EV.idx_begin())) {
+ if (EV.getNumIndices() == 0)
+ return ReplaceInstUsesWith(EV, C2);
+ // Extract the remaining indices out of the constant indexed by the
+ // first index
+ return ExtractValueInst::Create(C2, EV.getIndices().slice(1));
}
return 0; // Can't handle other constants
- }
+ }
+
if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
// We're extracting from an insertvalue instruction, compare the indices
const unsigned *exti, *exte, *insi, *inse;
// %E = extractvalue { i32, { i32 } } %I, 1, 0
// with
// %E extractvalue { i32 } { i32 42 }, 0
- return ExtractValueInst::Create(IV->getInsertedValueOperand(),
+ return ExtractValueInst::Create(IV->getInsertedValueOperand(),
makeArrayRef(exti, exte));
}
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Agg)) {
EraseInstFromFunction(*II);
return BinaryOperator::CreateAdd(LHS, RHS);
}
-
+
// If the normal result of the add is dead, and the RHS is a constant,
// we can transform this into a range comparison.
// overflow = uadd a, -4 --> overflow = icmp ugt a, 3
// load from a GEP. This reduces the size of the load.
// FIXME: If a load is used only by extractvalue instructions then this
// could be done regardless of having multiple uses.
- if (!L->isVolatile() && L->hasOneUse()) {
+ if (L->isSimple() && L->hasOneUse()) {
// extractvalue has integer indices, getelementptr has Value*s. Convert.
SmallVector<Value*, 4> Indices;
// Prefix an i32 0 since we need the first element.
return 0;
}
+enum Personality_Type {
+ Unknown_Personality,
+ GNU_Ada_Personality,
+ GNU_CXX_Personality,
+ GNU_ObjC_Personality
+};
+
+/// RecognizePersonality - See if the given exception handling personality
+/// function is one that we understand. If so, return a description of it;
+/// otherwise return Unknown_Personality.
+static Personality_Type RecognizePersonality(Value *Pers) {
+ Function *F = dyn_cast<Function>(Pers->stripPointerCasts());
+ if (!F)
+ return Unknown_Personality;
+ return StringSwitch<Personality_Type>(F->getName())
+ .Case("__gnat_eh_personality", GNU_Ada_Personality)
+ .Case("__gxx_personality_v0", GNU_CXX_Personality)
+ .Case("__objc_personality_v0", GNU_ObjC_Personality)
+ .Default(Unknown_Personality);
+}
+
+/// isCatchAll - Return 'true' if the given typeinfo will match anything.
+static bool isCatchAll(Personality_Type Personality, Constant *TypeInfo) {
+ switch (Personality) {
+ case Unknown_Personality:
+ return false;
+ case GNU_Ada_Personality:
+ // While __gnat_all_others_value will match any Ada exception, it doesn't
+ // match foreign exceptions (or didn't, before gcc-4.7).
+ return false;
+ case GNU_CXX_Personality:
+ case GNU_ObjC_Personality:
+ return TypeInfo->isNullValue();
+ }
+ llvm_unreachable("Unknown personality!");
+}
+
+static bool shorter_filter(const Value *LHS, const Value *RHS) {
+ return
+ cast<ArrayType>(LHS->getType())->getNumElements()
+ <
+ cast<ArrayType>(RHS->getType())->getNumElements();
+}
+
+Instruction *InstCombiner::visitLandingPadInst(LandingPadInst &LI) {
+ // The logic here should be correct for any real-world personality function.
+ // However if that turns out not to be true, the offending logic can always
+ // be conditioned on the personality function, like the catch-all logic is.
+ Personality_Type Personality = RecognizePersonality(LI.getPersonalityFn());
+
+ // Simplify the list of clauses, eg by removing repeated catch clauses
+ // (these are often created by inlining).
+ bool MakeNewInstruction = false; // If true, recreate using the following:
+ SmallVector<Value *, 16> NewClauses; // - Clauses for the new instruction;
+ bool CleanupFlag = LI.isCleanup(); // - The new instruction is a cleanup.
+
+ SmallPtrSet<Value *, 16> AlreadyCaught; // Typeinfos known caught already.
+ for (unsigned i = 0, e = LI.getNumClauses(); i != e; ++i) {
+ bool isLastClause = i + 1 == e;
+ if (LI.isCatch(i)) {
+ // A catch clause.
+ Value *CatchClause = LI.getClause(i);
+ Constant *TypeInfo = cast<Constant>(CatchClause->stripPointerCasts());
+
+ // If we already saw this clause, there is no point in having a second
+ // copy of it.
+ if (AlreadyCaught.insert(TypeInfo)) {
+ // This catch clause was not already seen.
+ NewClauses.push_back(CatchClause);
+ } else {
+ // Repeated catch clause - drop the redundant copy.
+ MakeNewInstruction = true;
+ }
+
+ // If this is a catch-all then there is no point in keeping any following
+ // clauses or marking the landingpad as having a cleanup.
+ if (isCatchAll(Personality, TypeInfo)) {
+ if (!isLastClause)
+ MakeNewInstruction = true;
+ CleanupFlag = false;
+ break;
+ }
+ } else {
+ // A filter clause. If any of the filter elements were already caught
+ // then they can be dropped from the filter. It is tempting to try to
+ // exploit the filter further by saying that any typeinfo that does not
+ // occur in the filter can't be caught later (and thus can be dropped).
+ // However this would be wrong, since typeinfos can match without being
+ // equal (for example if one represents a C++ class, and the other some
+ // class derived from it).
+ assert(LI.isFilter(i) && "Unsupported landingpad clause!");
+ Value *FilterClause = LI.getClause(i);
+ ArrayType *FilterType = cast<ArrayType>(FilterClause->getType());
+ unsigned NumTypeInfos = FilterType->getNumElements();
+
+ // An empty filter catches everything, so there is no point in keeping any
+ // following clauses or marking the landingpad as having a cleanup. By
+ // dealing with this case here the following code is made a bit simpler.
+ if (!NumTypeInfos) {
+ NewClauses.push_back(FilterClause);
+ if (!isLastClause)
+ MakeNewInstruction = true;
+ CleanupFlag = false;
+ break;
+ }
+
+ bool MakeNewFilter = false; // If true, make a new filter.
+ SmallVector<Constant *, 16> NewFilterElts; // New elements.
+ if (isa<ConstantAggregateZero>(FilterClause)) {
+ // Not an empty filter - it contains at least one null typeinfo.
+ assert(NumTypeInfos > 0 && "Should have handled empty filter already!");
+ Constant *TypeInfo =
+ Constant::getNullValue(FilterType->getElementType());
+ // If this typeinfo is a catch-all then the filter can never match.
+ if (isCatchAll(Personality, TypeInfo)) {
+ // Throw the filter away.
+ MakeNewInstruction = true;
+ continue;
+ }
+
+ // There is no point in having multiple copies of this typeinfo, so
+ // discard all but the first copy if there is more than one.
+ NewFilterElts.push_back(TypeInfo);
+ if (NumTypeInfos > 1)
+ MakeNewFilter = true;
+ } else {
+ ConstantArray *Filter = cast<ConstantArray>(FilterClause);
+ SmallPtrSet<Value *, 16> SeenInFilter; // For uniquing the elements.
+ NewFilterElts.reserve(NumTypeInfos);
+
+ // Remove any filter elements that were already caught or that already
+ // occurred in the filter. While there, see if any of the elements are
+ // catch-alls. If so, the filter can be discarded.
+ bool SawCatchAll = false;
+ for (unsigned j = 0; j != NumTypeInfos; ++j) {
+ Value *Elt = Filter->getOperand(j);
+ Constant *TypeInfo = cast<Constant>(Elt->stripPointerCasts());
+ if (isCatchAll(Personality, TypeInfo)) {
+ // This element is a catch-all. Bail out, noting this fact.
+ SawCatchAll = true;
+ break;
+ }
+ if (AlreadyCaught.count(TypeInfo))
+ // Already caught by an earlier clause, so having it in the filter
+ // is pointless.
+ continue;
+ // There is no point in having multiple copies of the same typeinfo in
+ // a filter, so only add it if we didn't already.
+ if (SeenInFilter.insert(TypeInfo))
+ NewFilterElts.push_back(cast<Constant>(Elt));
+ }
+ // A filter containing a catch-all cannot match anything by definition.
+ if (SawCatchAll) {
+ // Throw the filter away.
+ MakeNewInstruction = true;
+ continue;
+ }
+
+ // If we dropped something from the filter, make a new one.
+ if (NewFilterElts.size() < NumTypeInfos)
+ MakeNewFilter = true;
+ }
+ if (MakeNewFilter) {
+ FilterType = ArrayType::get(FilterType->getElementType(),
+ NewFilterElts.size());
+ FilterClause = ConstantArray::get(FilterType, NewFilterElts);
+ MakeNewInstruction = true;
+ }
+
+ NewClauses.push_back(FilterClause);
+
+ // If the new filter is empty then it will catch everything so there is
+ // no point in keeping any following clauses or marking the landingpad
+ // as having a cleanup. The case of the original filter being empty was
+ // already handled above.
+ if (MakeNewFilter && !NewFilterElts.size()) {
+ assert(MakeNewInstruction && "New filter but not a new instruction!");
+ CleanupFlag = false;
+ break;
+ }
+ }
+ }
+
+ // If several filters occur in a row then reorder them so that the shortest
+ // filters come first (those with the smallest number of elements). This is
+ // advantageous because shorter filters are more likely to match, speeding up
+ // unwinding, but mostly because it increases the effectiveness of the other
+ // filter optimizations below.
+ for (unsigned i = 0, e = NewClauses.size(); i + 1 < e; ) {
+ unsigned j;
+ // Find the maximal 'j' s.t. the range [i, j) consists entirely of filters.
+ for (j = i; j != e; ++j)
+ if (!isa<ArrayType>(NewClauses[j]->getType()))
+ break;
+
+ // Check whether the filters are already sorted by length. We need to know
+ // if sorting them is actually going to do anything so that we only make a
+ // new landingpad instruction if it does.
+ for (unsigned k = i; k + 1 < j; ++k)
+ if (shorter_filter(NewClauses[k+1], NewClauses[k])) {
+ // Not sorted, so sort the filters now. Doing an unstable sort would be
+ // correct too but reordering filters pointlessly might confuse users.
+ std::stable_sort(NewClauses.begin() + i, NewClauses.begin() + j,
+ shorter_filter);
+ MakeNewInstruction = true;
+ break;
+ }
+
+ // Look for the next batch of filters.
+ i = j + 1;
+ }
+
+ // If typeinfos matched if and only if equal, then the elements of a filter L
+ // that occurs later than a filter F could be replaced by the intersection of
+ // the elements of F and L. In reality two typeinfos can match without being
+ // equal (for example if one represents a C++ class, and the other some class
+ // derived from it) so it would be wrong to perform this transform in general.
+ // However the transform is correct and useful if F is a subset of L. In that
+ // case L can be replaced by F, and thus removed altogether since repeating a
+ // filter is pointless. So here we look at all pairs of filters F and L where
+ // L follows F in the list of clauses, and remove L if every element of F is
+ // an element of L. This can occur when inlining C++ functions with exception
+ // specifications.
+ for (unsigned i = 0; i + 1 < NewClauses.size(); ++i) {
+ // Examine each filter in turn.
+ Value *Filter = NewClauses[i];
+ ArrayType *FTy = dyn_cast<ArrayType>(Filter->getType());
+ if (!FTy)
+ // Not a filter - skip it.
+ continue;
+ unsigned FElts = FTy->getNumElements();
+ // Examine each filter following this one. Doing this backwards means that
+ // we don't have to worry about filters disappearing under us when removed.
+ for (unsigned j = NewClauses.size() - 1; j != i; --j) {
+ Value *LFilter = NewClauses[j];
+ ArrayType *LTy = dyn_cast<ArrayType>(LFilter->getType());
+ if (!LTy)
+ // Not a filter - skip it.
+ continue;
+ // If Filter is a subset of LFilter, i.e. every element of Filter is also
+ // an element of LFilter, then discard LFilter.
+ SmallVectorImpl<Value *>::iterator J = NewClauses.begin() + j;
+ // If Filter is empty then it is a subset of LFilter.
+ if (!FElts) {
+ // Discard LFilter.
+ NewClauses.erase(J);
+ MakeNewInstruction = true;
+ // Move on to the next filter.
+ continue;
+ }
+ unsigned LElts = LTy->getNumElements();
+ // If Filter is longer than LFilter then it cannot be a subset of it.
+ if (FElts > LElts)
+ // Move on to the next filter.
+ continue;
+ // At this point we know that LFilter has at least one element.
+ if (isa<ConstantAggregateZero>(LFilter)) { // LFilter only contains zeros.
+ // Filter is a subset of LFilter iff Filter contains only zeros (as we
+ // already know that Filter is not longer than LFilter).
+ if (isa<ConstantAggregateZero>(Filter)) {
+ assert(FElts <= LElts && "Should have handled this case earlier!");
+ // Discard LFilter.
+ NewClauses.erase(J);
+ MakeNewInstruction = true;
+ }
+ // Move on to the next filter.
+ continue;
+ }
+ ConstantArray *LArray = cast<ConstantArray>(LFilter);
+ if (isa<ConstantAggregateZero>(Filter)) { // Filter only contains zeros.
+ // Since Filter is non-empty and contains only zeros, it is a subset of
+ // LFilter iff LFilter contains a zero.
+ assert(FElts > 0 && "Should have eliminated the empty filter earlier!");
+ for (unsigned l = 0; l != LElts; ++l)
+ if (LArray->getOperand(l)->isNullValue()) {
+ // LFilter contains a zero - discard it.
+ NewClauses.erase(J);
+ MakeNewInstruction = true;
+ break;
+ }
+ // Move on to the next filter.
+ continue;
+ }
+ // At this point we know that both filters are ConstantArrays. Loop over
+ // operands to see whether every element of Filter is also an element of
+ // LFilter. Since filters tend to be short this is probably faster than
+ // using a method that scales nicely.
+ ConstantArray *FArray = cast<ConstantArray>(Filter);
+ bool AllFound = true;
+ for (unsigned f = 0; f != FElts; ++f) {
+ Value *FTypeInfo = FArray->getOperand(f)->stripPointerCasts();
+ AllFound = false;
+ for (unsigned l = 0; l != LElts; ++l) {
+ Value *LTypeInfo = LArray->getOperand(l)->stripPointerCasts();
+ if (LTypeInfo == FTypeInfo) {
+ AllFound = true;
+ break;
+ }
+ }
+ if (!AllFound)
+ break;
+ }
+ if (AllFound) {
+ // Discard LFilter.
+ NewClauses.erase(J);
+ MakeNewInstruction = true;
+ }
+ // Move on to the next filter.
+ }
+ }
+
+ // If we changed any of the clauses, replace the old landingpad instruction
+ // with a new one.
+ if (MakeNewInstruction) {
+ LandingPadInst *NLI = LandingPadInst::Create(LI.getType(),
+ LI.getPersonalityFn(),
+ NewClauses.size());
+ for (unsigned i = 0, e = NewClauses.size(); i != e; ++i)
+ NLI->addClause(NewClauses[i]);
+ // A landing pad with no clauses must have the cleanup flag set. It is
+ // theoretically possible, though highly unlikely, that we eliminated all
+ // clauses. If so, force the cleanup flag to true.
+ if (NewClauses.empty())
+ CleanupFlag = true;
+ NLI->setCleanup(CleanupFlag);
+ return NLI;
+ }
+
+ // Even if none of the clauses changed, we may nonetheless have understood
+ // that the cleanup flag is pointless. Clear it if so.
+ if (LI.isCleanup() != CleanupFlag) {
+ assert(!CleanupFlag && "Adding a cleanup, not removing one?!");
+ LI.setCleanup(CleanupFlag);
+ return &LI;
+ }
+
+ return 0;
+}
+
assert(I->hasOneUse() && "Invariants didn't hold!");
// Cannot move control-flow-involving, volatile loads, vaarg, etc.
- if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
+ if (isa<PHINode>(I) || isa<LandingPadInst>(I) || I->mayHaveSideEffects() ||
+ isa<TerminatorInst>(I))
return false;
// Do not sink alloca instructions out of the entry block.
return false;
}
- BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
-
+ BasicBlock::iterator InsertPos = DestBlock->getFirstInsertionPt();
I->moveBefore(InsertPos);
++NumSunkInst;
return true;
/// many instructions are dead or constant). Additionally, if we find a branch
/// whose condition is a known constant, we only visit the reachable successors.
///
-static bool AddReachableCodeToWorklist(BasicBlock *BB,
+static bool AddReachableCodeToWorklist(BasicBlock *BB,
SmallPtrSet<BasicBlock*, 64> &Visited,
InstCombiner &IC,
- const TargetData *TD) {
+ const DataLayout *TD,
+ const TargetLibraryInfo *TLI) {
bool MadeIRChange = false;
SmallVector<BasicBlock*, 256> Worklist;
Worklist.push_back(BB);
do {
BB = Worklist.pop_back_val();
-
+
// We have now visited this block! If we've already been here, ignore it.
if (!Visited.insert(BB)) continue;
for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
Instruction *Inst = BBI++;
-
+
// DCE instruction if trivially dead.
- if (isInstructionTriviallyDead(Inst)) {
+ if (isInstructionTriviallyDead(Inst, TLI)) {
++NumDeadInst;
- DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
+ DEBUG(dbgs() << "IC: DCE: " << *Inst << '\n');
Inst->eraseFromParent();
continue;
}
-
+
// ConstantProp instruction if trivially constant.
if (!Inst->use_empty() && isa<Constant>(Inst->getOperand(0)))
- if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
- DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
+ if (Constant *C = ConstantFoldInstruction(Inst, TD, TLI)) {
+ DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: "
<< *Inst << '\n');
Inst->replaceAllUsesWith(C);
++NumConstProp;
Inst->eraseFromParent();
continue;
}
-
+
if (TD) {
// See if we can constant fold its operands.
for (User::op_iterator i = Inst->op_begin(), e = Inst->op_end();
Constant*& FoldRes = FoldedConstants[CE];
if (!FoldRes)
- FoldRes = ConstantFoldConstantExpression(CE, TD);
+ FoldRes = ConstantFoldConstantExpression(CE, TD, TLI);
if (!FoldRes)
FoldRes = CE;
} else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
// See if this is an explicit destination.
- for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
- if (SI->getCaseValue(i) == Cond) {
- BasicBlock *ReachableBB = SI->getSuccessor(i);
+ for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
+ i != e; ++i)
+ if (i.getCaseValue() == Cond) {
+ BasicBlock *ReachableBB = i.getCaseSuccessor();
Worklist.push_back(ReachableBB);
continue;
}
-
+
// Otherwise it is the default destination.
- Worklist.push_back(SI->getSuccessor(0));
+ Worklist.push_back(SI->getDefaultDest());
continue;
}
}
-
+
for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
Worklist.push_back(TI->getSuccessor(i));
} while (!Worklist.empty());
-
+
// Once we've found all of the instructions to add to instcombine's worklist,
// add them in reverse order. This way instcombine will visit from the top
// of the function down. This jives well with the way that it adds all uses
// some N^2 behavior in pathological cases.
IC.Worklist.AddInitialGroup(&InstrsForInstCombineWorklist[0],
InstrsForInstCombineWorklist.size());
-
+
return MadeIRChange;
}
bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
MadeIRChange = false;
-
- DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
- << F.getNameStr() << "\n");
+
+ DEBUG(dbgs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
+ << F.getName() << "\n");
{
// Do a depth-first traversal of the function, populate the worklist with
// the reachable instructions. Ignore blocks that are not reachable. Keep
// track of which blocks we visit.
SmallPtrSet<BasicBlock*, 64> Visited;
- MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
+ MadeIRChange |= AddReachableCodeToWorklist(F.begin(), Visited, *this, TD,
+ TLI);
// Do a quick scan over the function. If we find any blocks that are
// unreachable, remove any instructions inside of them. This prevents
// the instcombine code from having to deal with some bad special cases.
- for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
- if (!Visited.count(BB)) {
- Instruction *Term = BB->getTerminator();
- while (Term != BB->begin()) { // Remove instrs bottom-up
- BasicBlock::iterator I = Term; --I;
-
- DEBUG(errs() << "IC: DCE: " << *I << '\n');
- // A debug intrinsic shouldn't force another iteration if we weren't
- // going to do one without it.
- if (!isa<DbgInfoIntrinsic>(I)) {
- ++NumDeadInst;
- MadeIRChange = true;
- }
-
- // If I is not void type then replaceAllUsesWith undef.
- // This allows ValueHandlers and custom metadata to adjust itself.
- if (!I->getType()->isVoidTy())
- I->replaceAllUsesWith(UndefValue::get(I->getType()));
- I->eraseFromParent();
+ for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
+ if (Visited.count(BB)) continue;
+
+ // Delete the instructions backwards, as it has a reduced likelihood of
+ // having to update as many def-use and use-def chains.
+ Instruction *EndInst = BB->getTerminator(); // Last not to be deleted.
+ while (EndInst != BB->begin()) {
+ // Delete the next to last instruction.
+ BasicBlock::iterator I = EndInst;
+ Instruction *Inst = --I;
+ if (!Inst->use_empty())
+ Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
+ if (isa<LandingPadInst>(Inst)) {
+ EndInst = Inst;
+ continue;
+ }
+ if (!isa<DbgInfoIntrinsic>(Inst)) {
+ ++NumDeadInst;
+ MadeIRChange = true;
}
+ Inst->eraseFromParent();
}
+ }
}
while (!Worklist.isEmpty()) {
if (I == 0) continue; // skip null values.
// Check to see if we can DCE the instruction.
- if (isInstructionTriviallyDead(I)) {
- DEBUG(errs() << "IC: DCE: " << *I << '\n');
+ if (isInstructionTriviallyDead(I, TLI)) {
+ DEBUG(dbgs() << "IC: DCE: " << *I << '\n');
EraseInstFromFunction(*I);
++NumDeadInst;
MadeIRChange = true;
// Instruction isn't dead, see if we can constant propagate it.
if (!I->use_empty() && isa<Constant>(I->getOperand(0)))
- if (Constant *C = ConstantFoldInstruction(I, TD)) {
- DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
+ if (Constant *C = ConstantFoldInstruction(I, TD, TLI)) {
+ DEBUG(dbgs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
// Add operands to the worklist.
ReplaceInstUsesWith(*I, C);
BasicBlock *BB = I->getParent();
Instruction *UserInst = cast<Instruction>(I->use_back());
BasicBlock *UserParent;
-
+
// Get the block the use occurs in.
if (PHINode *PN = dyn_cast<PHINode>(UserInst))
UserParent = PN->getIncomingBlock(I->use_begin().getUse());
else
UserParent = UserInst->getParent();
-
+
if (UserParent != BB) {
bool UserIsSuccessor = false;
// See if the user is one of our successors.
// Now that we have an instruction, try combining it to simplify it.
Builder->SetInsertPoint(I->getParent(), I);
Builder->SetCurrentDebugLocation(I->getDebugLoc());
-
+
#ifndef NDEBUG
std::string OrigI;
#endif
DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
- DEBUG(errs() << "IC: Visiting: " << OrigI << '\n');
+ DEBUG(dbgs() << "IC: Visiting: " << OrigI << '\n');
if (Instruction *Result = visit(*I)) {
++NumCombined;
// Should we replace the old instruction with a new one?
if (Result != I) {
- DEBUG(errs() << "IC: Old = " << *I << '\n'
+ DEBUG(dbgs() << "IC: Old = " << *I << '\n'
<< " New = " << *Result << '\n');
if (!I->getDebugLoc().isUnknown())
// Everything uses the new instruction now.
I->replaceAllUsesWith(Result);
+ // Move the name to the new instruction first.
+ Result->takeName(I);
+
// Push the new instruction and any users onto the worklist.
Worklist.Add(Result);
Worklist.AddUsersToWorkList(*Result);
- // Move the name to the new instruction first.
- Result->takeName(I);
-
// Insert the new instruction into the basic block...
BasicBlock *InstParent = I->getParent();
BasicBlock::iterator InsertPos = I;
- if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
- while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
- ++InsertPos;
+ // If we replace a PHI with something that isn't a PHI, fix up the
+ // insertion point.
+ if (!isa<PHINode>(Result) && isa<PHINode>(InsertPos))
+ InsertPos = InstParent->getFirstInsertionPt();
InstParent->getInstList().insert(InsertPos, Result);
EraseInstFromFunction(*I);
} else {
#ifndef NDEBUG
- DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
+ DEBUG(dbgs() << "IC: Mod = " << OrigI << '\n'
<< " New = " << *I << '\n');
#endif
// If the instruction was modified, it's possible that it is now dead.
// if so, remove it.
- if (isInstructionTriviallyDead(I)) {
+ if (isInstructionTriviallyDead(I, TLI)) {
EraseInstFromFunction(*I);
} else {
Worklist.Add(I);
return MadeIRChange;
}
+namespace {
+class InstCombinerLibCallSimplifier : public LibCallSimplifier {
+ InstCombiner *IC;
+public:
+ InstCombinerLibCallSimplifier(const DataLayout *TD,
+ const TargetLibraryInfo *TLI,
+ InstCombiner *IC)
+ : LibCallSimplifier(TD, TLI, UnsafeFPShrink) {
+ this->IC = IC;
+ }
+
+ /// replaceAllUsesWith - override so that instruction replacement
+ /// can be defined in terms of the instruction combiner framework.
+ virtual void replaceAllUsesWith(Instruction *I, Value *With) const {
+ IC->ReplaceInstUsesWith(*I, With);
+ }
+};
+}
bool InstCombiner::runOnFunction(Function &F) {
- TD = getAnalysisIfAvailable<TargetData>();
+ if (skipOptnoneFunction(F))
+ return false;
+
+ TD = getAnalysisIfAvailable<DataLayout>();
+ TLI = &getAnalysis<TargetLibraryInfo>();
+ // Minimizing size?
+ MinimizeSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
+ Attribute::MinSize);
-
/// Builder - This is an IRBuilder that automatically inserts new
/// instructions into the worklist when they are created.
- IRBuilder<true, TargetFolder, InstCombineIRInserter>
+ IRBuilder<true, TargetFolder, InstCombineIRInserter>
TheBuilder(F.getContext(), TargetFolder(TD),
InstCombineIRInserter(Worklist));
Builder = &TheBuilder;
-
+
+ InstCombinerLibCallSimplifier TheSimplifier(TD, TLI, this);
+ Simplifier = &TheSimplifier;
+
bool EverMadeChange = false;
// Lower dbg.declare intrinsics otherwise their value may be clobbered
unsigned Iteration = 0;
while (DoOneIteration(F, Iteration++))
EverMadeChange = true;
-
+
Builder = 0;
return EverMadeChange;
}
projects / oota-llvm.git / blobdiff