https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104665
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2022-02-23
Severity|normal |enhancement
Ever confirmed|0 |1
--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
A couple of reasons.
First is store merging happens too late.
Second reason is store merging does not work in the loop case.
Take:
#if 1
enum b : unsigned char{};
#else
typedef unsigned char b;
#endif
void serialize_le(b* __restrict dst, const unsigned* __restrict src)
{
// for (int i = 0; i < 128; ++i, ++src)
{
unsigned t = *src;
*dst++ = static_cast<b>((t >> 0) & 0xff);
*dst++ = static_cast<b>((t >> 8) & 0xff);
*dst++ = static_cast<b>((t >> 16) & 0xff);
*dst++ = static_cast<b>((t >> 24) & 0xff);
}
}
This gets optimized to one load followed by one store. But once you add the
loop, and use -fno-tree-vectorize (because GCC's vectorizer gets kicked in
which causes other issues), the stores are not merged into one.
Also store merging happens way after loop distrubution happens so ...
