Allow SelectionDAG to create EXTRACT_VECTOR_ELT nodes with

non-constant indices. Only a few of the peephole checks require
a constant index.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@54764 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Dan Gohman 2008-08-13 21:51:37 +00:00
parent 03857b29d8
commit 6ab6422f25
2 changed files with 20 additions and 11 deletions

View File

@ -2302,15 +2302,14 @@ SDValue SelectionDAG::getNode(unsigned Opcode, MVT VT,
break;
}
case ISD::EXTRACT_VECTOR_ELT:
assert(N2C && "Bad EXTRACT_VECTOR_ELT!");
// EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF.
if (N1.getOpcode() == ISD::UNDEF)
return getNode(ISD::UNDEF, VT);
// EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is
// expanding copies of large vectors from registers.
if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
if (N2C &&
N1.getOpcode() == ISD::CONCAT_VECTORS &&
N1.getNumOperands() > 0) {
unsigned Factor =
N1.getOperand(0).getValueType().getVectorNumElements();
@ -2321,18 +2320,17 @@ SDValue SelectionDAG::getNode(unsigned Opcode, MVT VT,
// EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is
// expanding large vector constants.
if (N1.getOpcode() == ISD::BUILD_VECTOR)
if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR)
return N1.getOperand(N2C->getValue());
// EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector
// operations are lowered to scalars.
if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT)
if (ConstantSDNode *IEC = dyn_cast<ConstantSDNode>(N1.getOperand(2))) {
if (IEC == N2C)
return N1.getOperand(1);
else
return getNode(ISD::EXTRACT_VECTOR_ELT, VT, N1.getOperand(0), N2);
}
if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) {
if (N1.getOperand(2) == N2)
return N1.getOperand(1);
else
return getNode(ISD::EXTRACT_VECTOR_ELT, VT, N1.getOperand(0), N2);
}
break;
case ISD::EXTRACT_ELEMENT:
assert(N2C && (unsigned)N2C->getValue() < 2 && "Bad EXTRACT_ELEMENT!");

View File

@ -0,0 +1,11 @@
; RUN: llvm-as < %s | llc -march=x86-64 | grep movss | count 2
; PR2676
define float @foo(<4 x float> %p, i32 %t) {
%z = extractelement <4 x float> %p, i32 %t
ret float %z
}
define <4 x float> @bar(<4 x float> %p, float %f, i32 %t) {
%z = insertelement <4 x float> %p, float %f, i32 %t
ret <4 x float> %z
}