https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98674
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |marxin at gcc dot gnu.org
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #3)
> While GCC 9 vectorizes this case (and thus the vectorization failure is a
> regression) dependence analysis isn't presented with the problematical access
> but instead we see
>
> _1 = *p_22;
> _5 = (unsigned short) _1;
> bswapdst_10 = _5 r>> 8;
> _8 = (short int) bswapdst_10;
> *p_22 = _8;
>
> where the problematical access is created by the bswap pass which
> detects
>
> 16 bit bswap implementation found at: _8 = _4 | _7;
>
> creates the replacement load but then fails half-way, not emitting a
> bswap!? That's a bug worth fixing IMHO (either do all or none of the
> transform).
Oh, so it produces
load_dst_11 = MEM[(short int *)p_22];
bswapdst_10 = load_dst_11 r>> 8;
_8 = (short int) bswapdst_10;
*p_22 = _8;
it's just pattern recog that turns the rotate back to a shift sequence. So
indeed dependence analysis is what should be fixed.
