Re: [PATCH] Fix PR80222

[Richard Biener <rguenther@suse.de>]

On May 13, 2017 10:49:31 AM GMT+02:00, Eric Botcazou <ebotcazou@adacore.com> wrote:
>> Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to
>trunk.
>> 
>> Richard.
>> 
>> 2017-03-28  Richard Biener  <rguenther@suse.de>
>> 
>> 	PR middle-end/80222
>> 	* gimple-fold.c (gimple_fold_indirect_ref): Do not touch
>> 	TYPE_REF_CAN_ALIAS_ALL references.
>> 	* fold-const.c (fold_indirect_ref_1): Likewise.
>
>You just backported it onto the 6 branch and the fold_indirect_ref_1
>hunk 
>alone introduced regressions in the C testsuite on the branch for
>strict-
>alignment architectures, e.g. SPARC.  Reduced testcase:
>
>typedef long long V
>__attribute__ ((vector_size (2 * sizeof (long long)), may_alias));
>
>typedef struct S { V b; } P __attribute__((aligned (1)));
>
>__attribute__((noinline, noclone)) void
>bar (P *p)
>{
>  p->b[1] = 5;
>}
>
>Before the patch at -O0:
>
>bar:
>        save    %sp, -96, %sp
>        st      %i0, [%fp+68]
>        ld      [%fp+68], %g1
>        ldub    [%g1+12], %g2
>        and     %g2, 0, %g2
>        stb     %g2, [%g1+12]
>        ldub    [%g1+13], %g2
>        and     %g2, 0, %g2
>        stb     %g2, [%g1+13]
>        ldub    [%g1+14], %g2
>        and     %g2, 0, %g2
>        stb     %g2, [%g1+14]
>        ldub    [%g1+15], %g2
>        and     %g2, 0, %g2
>        or      %g2, 5, %g2
>        stb     %g2, [%g1+15]
>        ldub    [%g1+8], %g2
>        and     %g2, 0, %g2
>        stb     %g2, [%g1+8]
>        ldub    [%g1+9], %g2
>        and     %g2, 0, %g2
>        stb     %g2, [%g1+9]
>        ldub    [%g1+10], %g2
>        and     %g2, 0, %g2
>        stb     %g2, [%g1+10]
>        ldub    [%g1+11], %g2
>        and     %g2, 0, %g2
>        stb     %g2, [%g1+11]
>        nop
>        return  %i7+8
>         nop
>
>
>After the patch:
>
>bar:
>        save    %sp, -96, %sp
>        st      %i0, [%fp+68]
>        ld      [%fp+68], %g1
>        add     %g1, 8, %g1
>        mov     0, %g2
>        mov     5, %g3
>        std     %g2, [%g1]
>        nop
>        return  %i7+8
>         nop
>
>that is to say, the compiler now generates an unaligned store.

Does this happen on the GCC7 branch as well?  The patch just guards an indirect ref folding (I refrained from trying to make it correct given I think it's premature optimization).

I'll try to investigate on Monday if you don't beat me to it.  Feel free to revert the backport in the meantime.

Note I think you can trigger the same bug with some source changes independent of the patch which means the GENERIC must be somehow invalid.

Thanks,
Richard.