llvm dev - Feb 2017 - using getTargetInsertSubreg in DAG

I'm not sure what I'm missing and was hoping someone could point it out
to
me.

We'd like to use a 16 bit intrinsic to do some work on a f32. The problem
with the code is that we're getting a f32 returned from the intrinsic when
it should be a i16, so we are unable to pattern match. We're trying to
follow along the two examples in code code in SparcISelLowering.cpp.

given f(x, y), it should extract the upper bits of y and x, copy the bits
from y specified in the bitmask to x and then reinsert the upper bits into
x and return the f32.

Code:

  SDLoc DL(Op);
  EVT VT = Op.getValueType();

  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);

  SDValue ConstMask = DAG.getConstant(0x8000, SDLoc(Op.getNode()),
MVT::i16);
  SDValue y_MSB =  DAG.getTargetExtractSubreg(TARGET::sub_even, DL,
MVT::i16, RHS);
  SDValue x_MSB = DAG.getTargetExtractSubreg(TARGET::sub_even, DL,
MVT::i16, LHS);
  SDValue x_LSB = DAG.getTargetExtractSubreg(TARGET::sub_odd, DL, MVT::i16,
LHS);

  x_MSB = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL,
                          VT, DAG.getTargetConstant(Intrinsic::myIntrinsic,
                          DL, MVT::i16), y_MSB, ConstMask);

  SDValue result = SDValue(DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF,
                                                DL, MVT::f32), 0);

  result = DAG.getTargetInsertSubreg(TARGET::sub_even, DL, MVT::f32,
result, x_MSB);
  result = DAG.getTargetInsertSubreg(TARGET::sub_odd, DL, MVT::f32, result,
x_LSB);

  return result;
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170213/e94de60e/attachment.html>

I found the bug on my end, a simple mistake with the typing of the
intrinsic. Sorry for the spam.



On Mon, Feb 13, 2017 at 4:37 PM, Ryan Taylor <ryta1203 at gmail.com>
wrote:
> I'm not sure what I'm missing and was hoping someone could point it
out to
> me.
>
> We'd like to use a 16 bit intrinsic to do some work on a f32. The
problem
> with the code is that we're getting a f32 returned from the intrinsic
when
> it should be a i16, so we are unable to pattern match. We're trying to
> follow along the two examples in code code in SparcISelLowering.cpp.
>
> given f(x, y), it should extract the upper bits of y and x, copy the bits
> from y specified in the bitmask to x and then reinsert the upper bits into
> x and return the f32.
>
> Code:
>
>   SDLoc DL(Op);
>   EVT VT = Op.getValueType();
>
>   SDValue LHS = Op.getOperand(0);
>   SDValue RHS = Op.getOperand(1);
>
>   SDValue ConstMask = DAG.getConstant(0x8000, SDLoc(Op.getNode()),
> MVT::i16);
>   SDValue y_MSB =  DAG.getTargetExtractSubreg(TARGET::sub_even, DL,
> MVT::i16, RHS);
>   SDValue x_MSB = DAG.getTargetExtractSubreg(TARGET::sub_even, DL,
> MVT::i16, LHS);
>   SDValue x_LSB = DAG.getTargetExtractSubreg(TARGET::sub_odd, DL,
> MVT::i16, LHS);
>
>   x_MSB = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL,
>                           VT, DAG.getTargetConstant(
> Intrinsic::myIntrinsic,
>                           DL, MVT::i16), y_MSB, ConstMask);
>
>   SDValue result = SDValue(DAG.getMachineNode(TargetOpcode::IMPLICIT_DEF,
>                                                 DL, MVT::f32), 0);
>
>   result = DAG.getTargetInsertSubreg(TARGET::sub_even, DL, MVT::f32,
> result, x_MSB);
>   result = DAG.getTargetInsertSubreg(TARGET::sub_odd, DL, MVT::f32,
> result, x_LSB);
>
>   return result;
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170213/12220e54/attachment.html>