https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96135
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2020-07-10
Keywords| |missed-optimization
Target Milestone|--- |9.4
Ever confirmed|0 |1
Priority|P3 |P2
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
The bswap pass doesn't handle these patterns at all (it doesn't look at
stores).
What does handle this case is store-merging which - on trunk - figures
bswap in f() but not in g() likely because of the BIT_FIELD_REF ->
cast folding which makes the stores appear inhomogenous:
_3 = VIEW_CONVERT_EXPR<long long int>(x_2(D));
_4 = BIT_FIELD_REF <x_2(D), 8, 56>;
v.c[0] = _4;
...
_11 = (char) _3;
v.c[7] = _11;
store-merging already handles the cast vs. BIT_FIELD_REF case for f() but
it appearantly doesn't consider to look through a VIEW_CONVERT.
With -O3 we vectorize this in an inconvenient way and fully elide the store
so store-merging isn't the correct pass to handle this:
_3 = BIT_FIELD_REF <i_2(D), 8, 56>;
_4 = BIT_FIELD_REF <i_2(D), 8, 48>;
_5 = BIT_FIELD_REF <i_2(D), 8, 40>;
_6 = BIT_FIELD_REF <i_2(D), 8, 32>;
_7 = BIT_FIELD_REF <i_2(D), 8, 24>;
_8 = BIT_FIELD_REF <i_2(D), 8, 16>;
_9 = BIT_FIELD_REF <i_2(D), 8, 8>;
_10 = (char) i_2(D);
_21 = {_3, _4, _5, _6, _7, _8, _9, _10};
_18 = VIEW_CONVERT_EXPR<long long int>(_21);
v ={v} {CLOBBER};
return _18;
the vectorizer is also confused about BIT_FIELD_REF vs. cast here
(I repeatedly thought of removing that simplification ... but the
user could have written it as well :/). And it would look for
a vector function argument but that's something that could be fixed.
The above is all for GCC 10. GCC 8 possibly was lucky and did not have
that BIT_FIELD_REF -> cast simplification.