[Bug middle-end/48377] [4.6/4.7 regression] miscompilation at -O3

["rguenther at suse dot de" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48377

--- Comment #33 from rguenther at suse dot de <rguenther at suse dot de> 2011-04-07 13:57:09 UTC ---
On Thu, 7 Apr 2011, jakub at gcc dot gnu.org wrote:

> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48377
> 
> --- Comment #32 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-04-07 13:43:47 UTC ---
> Given that without vectorization that code surely works well on all the weird
> ABIs, such ADJUST_FIELD_ALIGN is used only on targets that either aren't strict
> alignment targets, or aren't strict alignment targets for the particular types.
> You know, for most CPUs strict alignment isn't a binary thing, many targets
> including i?86/x86_64 are only partially strict alignment targets.
> Even for targets where say int needs to be accessed aligned often e.g. double
> can be just 4 byte aligned and insns that read/write doubles handle it.
> 
> We really shouldn't change __alignof__ of types/INDIRECT_REFs/etc., those are
> all ABI changes.  The problem with vectorization is that often the vectors have
> strict alignment, while non-vector ops are not.

Sure, I agree with all of the above.  Still, if we need target hooks for
this kind of stuff then something is really rotten in the middle-end.

Richard.