forked from OSchip/llvm-project
For targets with FABS/FNEG support, lower copysign to an integer load,
a select and FABS/FNEG. This speeds up a trivial (aka stupid) copysign benchmark I wrote from 6.73s to 2.64s, woo. llvm-svn: 26723
This commit is contained in:
parent
ec9d0bc3ec
commit
994d8e6bd4
|
@ -1798,7 +1798,30 @@ SDOperand SelectionDAGLegalize::LegalizeOp(SDOperand Op) {
|
|||
break;
|
||||
case TargetLowering::Legal: break;
|
||||
case TargetLowering::Expand:
|
||||
// Floating point mod -> fmod libcall.
|
||||
// If this target supports fabs/fneg natively, do this efficiently.
|
||||
if (TLI.isOperationLegal(ISD::FABS, Tmp1.getValueType()) &&
|
||||
TLI.isOperationLegal(ISD::FNEG, Tmp1.getValueType())) {
|
||||
// Get the sign bit of the RHS.
|
||||
MVT::ValueType IVT =
|
||||
Tmp2.getValueType() == MVT::f32 ? MVT::i32 : MVT::i64;
|
||||
SDOperand SignBit = DAG.getNode(ISD::BIT_CONVERT, IVT, Tmp2);
|
||||
SignBit = DAG.getSetCC(TLI.getSetCCResultTy(),
|
||||
SignBit, DAG.getConstant(0, IVT), ISD::SETLT);
|
||||
// Get the absolute value of the result.
|
||||
SDOperand AbsVal = DAG.getNode(ISD::FABS, Tmp1.getValueType(), Tmp1);
|
||||
// Select between the nabs and abs value based on the sign bit of
|
||||
// the input.
|
||||
Result = DAG.getNode(ISD::SELECT, AbsVal.getValueType(), SignBit,
|
||||
DAG.getNode(ISD::FNEG, AbsVal.getValueType(),
|
||||
AbsVal),
|
||||
AbsVal);
|
||||
Result = LegalizeOp(Result);
|
||||
break;
|
||||
}
|
||||
|
||||
// Otherwise, do bitwise ops!
|
||||
|
||||
// copysign -> copysignf/copysign libcall.
|
||||
const char *FnName;
|
||||
if (Node->getValueType(0) == MVT::f32) {
|
||||
FnName = "copysignf";
|
||||
|
|
Loading…
Reference in New Issue