Add support for folding binary operators with vector zero operands.

llvm-svn: 43510
This commit is contained in:
Dan Gohman 2007-10-30 19:00:49 +00:00
parent de21d55e70
commit 9f39660c20
2 changed files with 64 additions and 19 deletions

View File

@ -450,14 +450,20 @@ Constant *llvm::ConstantFoldShuffleVectorInstruction(const Constant *V1,
/// EvalVectorOp - Given two vector constants and a function pointer, apply the
/// function pointer to each element pair, producing a new ConstantVector
/// constant.
/// constant. Either or both of V1 and V2 may be NULL, meaning a
/// ConstantAggregateZero operand.
static Constant *EvalVectorOp(const ConstantVector *V1,
const ConstantVector *V2,
const VectorType *VTy,
Constant *(*FP)(Constant*, Constant*)) {
std::vector<Constant*> Res;
for (unsigned i = 0, e = V1->getNumOperands(); i != e; ++i)
Res.push_back(FP(const_cast<Constant*>(V1->getOperand(i)),
const_cast<Constant*>(V2->getOperand(i))));
const Type *EltTy = VTy->getElementType();
for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
const Constant *C1 = V1 ? V1->getOperand(i) : Constant::getNullValue(EltTy);
const Constant *C2 = V2 ? V2->getOperand(i) : Constant::getNullValue(EltTy);
Res.push_back(FP(const_cast<Constant*>(C1),
const_cast<Constant*>(C2)));
}
return ConstantVector::get(Res);
}
@ -707,36 +713,40 @@ Constant *llvm::ConstantFoldBinaryInstruction(unsigned Opcode,
return ConstantFP::get(CFP1->getType(), C3V);
}
}
} else if (const ConstantVector *CP1 = dyn_cast<ConstantVector>(C1)) {
if (const ConstantVector *CP2 = dyn_cast<ConstantVector>(C2)) {
} else if (const VectorType *VTy = dyn_cast<VectorType>(C1->getType())) {
const ConstantVector *CP1 = dyn_cast<ConstantVector>(C1);
const ConstantVector *CP2 = dyn_cast<ConstantVector>(C2);
assert((CP1 != NULL || isa<ConstantAggregateZero>(C1)) &&
"Unexpected kind of vector constant!");
assert((CP2 != NULL || isa<ConstantAggregateZero>(C2)) &&
"Unexpected kind of vector constant!");
switch (Opcode) {
default:
break;
case Instruction::Add:
return EvalVectorOp(CP1, CP2, ConstantExpr::getAdd);
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getAdd);
case Instruction::Sub:
return EvalVectorOp(CP1, CP2, ConstantExpr::getSub);
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSub);
case Instruction::Mul:
return EvalVectorOp(CP1, CP2, ConstantExpr::getMul);
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getMul);
case Instruction::UDiv:
return EvalVectorOp(CP1, CP2, ConstantExpr::getUDiv);
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getUDiv);
case Instruction::SDiv:
return EvalVectorOp(CP1, CP2, ConstantExpr::getSDiv);
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSDiv);
case Instruction::FDiv:
return EvalVectorOp(CP1, CP2, ConstantExpr::getFDiv);
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getFDiv);
case Instruction::URem:
return EvalVectorOp(CP1, CP2, ConstantExpr::getURem);
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getURem);
case Instruction::SRem:
return EvalVectorOp(CP1, CP2, ConstantExpr::getSRem);
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getSRem);
case Instruction::FRem:
return EvalVectorOp(CP1, CP2, ConstantExpr::getFRem);
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getFRem);
case Instruction::And:
return EvalVectorOp(CP1, CP2, ConstantExpr::getAnd);
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getAnd);
case Instruction::Or:
return EvalVectorOp(CP1, CP2, ConstantExpr::getOr);
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getOr);
case Instruction::Xor:
return EvalVectorOp(CP1, CP2, ConstantExpr::getXor);
}
return EvalVectorOp(CP1, CP2, VTy, ConstantExpr::getXor);
}
}

View File

@ -0,0 +1,35 @@
; RUN: llvm-as < %s | opt -instcombine | llvm-dis | not grep zeroinitializer
define void @foo(i64 %A, i64 %B) {
bb8:
br label %bb30
bb30:
%s0 = phi i64 [ 0, %bb8 ], [ %r21, %bb30 ]
%l0 = phi i64 [ -2222, %bb8 ], [ %r23, %bb30 ]
%r2 = add i64 %s0, %B
%r3 = inttoptr i64 %r2 to <2 x double>*
%r4 = load <2 x double>* %r3, align 8
%r6 = bitcast <2 x double> %r4 to <2 x i64>
%r7 = bitcast <2 x double> zeroinitializer to <2 x i64>
%r8 = insertelement <2 x i64> undef, i64 9223372036854775807, i32 0
%r9 = insertelement <2 x i64> undef, i64 -9223372036854775808, i32 0
%r10 = insertelement <2 x i64> %r8, i64 9223372036854775807, i32 1
%r11 = insertelement <2 x i64> %r9, i64 -9223372036854775808, i32 1
%r12 = and <2 x i64> %r6, %r10
%r13 = and <2 x i64> %r7, %r11
%r14 = or <2 x i64> %r12, %r13
%r15 = bitcast <2 x i64> %r14 to <2 x double>
%r18 = add i64 %s0, %A
%r19 = inttoptr i64 %r18 to <2 x double>*
store <2 x double> %r15, <2 x double>* %r19, align 8
%r21 = add i64 16, %s0
%r23 = add i64 1, %l0
%r25 = icmp slt i64 %r23, 0
%r26 = zext i1 %r25 to i64
%r27 = icmp ne i64 %r26, 0
br i1 %r27, label %bb30, label %bb5
bb5:
ret void
}