[RFA 4.7/4.6] Re: [PATCH v2] Target-specific limits on vector alignment

[Richard Guenther]

On Tue, Aug 7, 2012 at 4:56 PM, Ulrich Weigand <uweigand@de.ibm.com> wrote:
> Richard Guenther wrote:
>> On Fri, Jul 27, 2012 at 5:24 PM, Ulrich Weigand <uweigand@de.ibm.com> wrote:
>> > ChangeLog:
>> >
>> >         * target.def (vector_alignment): New target hook.
>> >         * doc/tm.texi.in (TARGET_VECTOR_ALIGNMENT): Document new hook.
>> >         * doc/tm.texi: Regenerate.
>> >         * targhooks.c (default_vector_alignment): New function.
>> >         * targhooks.h (default_vector_alignment): Add prototype.
>> >         * stor-layout.c (layout_type): Use targetm.vector_alignment.
>> >         * config/arm/arm.c (arm_vector_alignment): New function.
>> >         (TARGET_VECTOR_ALIGNMENT): Define.
>> >
>> >         * tree-vect-data-refs.c (vect_update_misalignment_for_peel): Use
>> >         vector type alignment instead of size.
>> >         * tree-vect-loop-manip.c (vect_do_peeling_for_loop_bound): Use
>> >         element type size directly instead of computing it from alignment.
>> >         Fix variable naming and comment.
>> >
>> > testsuite/ChangeLog:
>> >
>> >         * lib/target-supports.exp
>> >         (check_effective_target_vect_natural_alignment): New function.
>> >         * gcc.dg/align-2.c: Only run on targets with natural alignment
>> >         of vector types.
>> >         * gcc.dg/vect/slp-25.c: Adjust tests for targets without natural
>> >         alignment of vector types.
>
>
>> > OK for mainline?
>>
>> Ok.
>
> Would it be OK to backport this to 4.7 and possibly 4.6?
>
> This patch represents a change in the ABI on ARM (the alignment could
> potentially affect struct member offsets, for example), which could
> conceivably cause incompatibilities with code compiled with older
> versions of GCC.  We had originally decided we nevertheless want to
> implement this change on ARM, because:
>
> - GCC is now in compliance with the platform ABI document
> - the old code could actually be buggy since GCC *assumed* 16-byte
>   alignment that wasn't actually guaranteed
> - code actually affected by this change ought to be rare (code using
>   NEON vector types is rare in the first place, code using *structure
>   members* of such types is even rarer, and code using such structures
>   in cross-module APIs seems to be extremely rare)
>
> Nevertheless, given we do want to make this change, it would be best
> to roll it out as quickly as possible, to minimize the time period
> where people might use old (not yet fixed) compilers to generate
> non-ABI-compliant binaries.  Thus, we think it would be good for
> the change to be applied to all still open release branches as well.
>
> (Note that while the patch contains changes to common code, those
> should be no-ops for all targets that do not implement the new hook.)

I'll defer the decision to the target maintainers.  But please double-check
for any changes in the vectorizer parts when backporting to 4.6.

Thanks,
Richard.

> Bye,
> Ulrich
>
> --
>   Dr. Ulrich Weigand
>   GNU Toolchain for Linux on System z and Cell BE
>   Ulrich.Weigand@de.ibm.com
>