On Mon, Nov 4, 2013 at 12:34 AM, Bill Schmidt
<wschm...@linux.vnet.ibm.com> wrote:
> Hi,
>
> This patch fixes the widening multiply high/low operations to work
> correctly in the presence of the first patch of this series, which
> reverses the meanings of multiply even/odd instructions. Here we
> reorder the input operands to the vector merge low/high instructions.
>
> The general rule is that vmrghh(x,y) [BE] = vmrglh(y,x) [LE], and so on;
> that is, we need to reverse the usage of merge high and merge low, and
> also swap their inputs, to obtain the same semantics. In this case we
> are only swapping the inputs, because the reversed usage of high and low
> has already been done for us in the generic handling code for
> VEC_WIDEN_MULT_LO_EXPR.
>
> Bootstrapped and tested with the rest of the patch set on
> powerpc64{,le}-unknown-linux-gnu, with no regressions. Is this ok for
> trunk?
>
> Thanks,
> Bill
>
>
> 2013-11-03 Bill Schmidt <wschm...@linux.vnet.ibm.com>
>
> * config/rs6000/altivec.md (vec_widen_umult_hi_v16qi): Swap
> arguments to merge instruction for little endian.
> (vec_widen_umult_lo_v16qi): Likewise.
> (vec_widen_smult_hi_v16qi): Likewise.
> (vec_widen_smult_lo_v16qi): Likewise.
> (vec_widen_umult_hi_v8hi): Likewise.
> (vec_widen_umult_lo_v8hi): Likewise.
> (vec_widen_smult_hi_v8hi): Likewise.
> (vec_widen_smult_lo_v8hi): Likewise.
This patch is okay.
I agree with Richard's suggestion to address the other parts of the
patch by restructuring the patterns.
Thanks, David
