On Thu, Mar 21, 2013 at 9:36 AM, Michael Zolotukhin
<michael.v.zolotuk...@gmail.com> wrote:
> Hi,
> I've found a little bit strange code in "mov<mode>_internal"
> RTL-pattern from config/i386/sse.md:
> case MODE_V2DF:
> if (TARGET_AVX
> && (misaligned_operand (operands[0], <MODE>mode)
> || misaligned_operand (operands[1], <MODE>mode)))
> return "vmovupd\t{%1, %0|%0, %1}";
> else
> return "%vmovapd\t{%1, %0|%0, %1}";
>
> That could lead to generation of vmovapd for misaligned operands. Was
> there any special intention behind that, or is that just a bug?
> Attached is a patch that removes TARGET_AVX from the condition, thus
> preventing generation of movapd for misaligned operands and enabling
> generation of movupd for them.
>
> If the patch is ok, could anyone please commit it? Of course, that
> could wait until 4.8 is released.
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu and i686-linux.
>
> Changelog:
> 2013-03-21 Michael Zolotukhin <michael.v.zolotuk...@gmail.com>
>
> * config/i386/sse.md (*mov<mode>_internal): Fix condition.
>
> --
Do you have a testcase to show there is a problem?
The misaligned case should only be generated from
ix86_avx256_split_vector_move_misalign.
--
H.J.
