Re: [PATCH] Fix PR tree-optimization/51315

[Richard Sandiford <rdsandiford at googlemail dot com>]

Eric Botcazou <ebotcazou@adacore.com> writes:
>> Sorry for the late notice, but this regressed memcpy-1.c for n32 and n64
>> on mips64-linux-gnu.  I realise memcpy-1.c isn't viewed as being a proper
>> SRA test, but in this case I think it really is showing a genuine problem.
>> We have:
>>
>>     struct a {int a,b,c;} a;
>>     int test(struct a a)
>>     {
>>     struct a nasty_local;
>>     __builtin_memcpy (&nasty_local,&a, sizeof(a));
>>     return nasty_local.a;
>>     }
>>
>> We apply LOCAL_ALIGNMENT to nasty_local during estimated_stack_frame_size,
>> so we have a VAR_DECL (nasty_local) with 64-bit alignment and a PARM_DECL
>> (a) with 32-bit alignment.  This fails the condition:
>>
>>       if (STRICT_ALIGNMENT
>> 	  && tree_non_aligned_mem_p (rhs, get_object_alignment (lhs)))
>>         lacc->grp_unscalarizable_region = 1;
>>
>> because LHS has 64-bit alignment but RHS has 32-bit alignment.
>
> Do you mean that the patch pessimizes this case?

Yeah.

> If so, yes, this is known, I ran into similar cases in Ada (we do this
> kind of local alignment promotion).

OK.  But passing small structures by value doesn't seem that rare --
especially in C++ -- and it doesn't feel right to disable SRA just because
the backend likes to increase the alignment of stack vars.  So...

> Index: tree-sra.c
> ===================================================================
> --- tree-sra.c  (revision 182780)
> +++ tree-sra.c  (working copy)
> @@ -1124,7 +1124,9 @@ build_accesses_from_assign (gimple stmt)
>      {
>        lacc->grp_assignment_write = 1;
>        if (STRICT_ALIGNMENT
> -         && tree_non_aligned_mem_p (rhs, get_object_alignment (lhs)))
> +         && tree_non_aligned_mem_p (rhs,
> +                                    MIN (TYPE_ALIGN (lacc->type),
> +                                         get_object_alignment (lhs))))
>          lacc->grp_unscalarizable_region = 1;
>      }
>
> @@ -1135,7 +1137,9 @@ build_accesses_from_assign (gimple stmt)
>           && !is_gimple_reg_type (racc->type))
>         bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
>        if (STRICT_ALIGNMENT
> -         && tree_non_aligned_mem_p (lhs, get_object_alignment (rhs)))
> +         && tree_non_aligned_mem_p (lhs,
> +                                    MIN (TYPE_ALIGN (racc->type),
> +                                         get_object_alignment (rhs))))
>          racc->grp_unscalarizable_region = 1;
>      }
>
> on the grounds that sub-accesses shouldn't be more aligned.  I think this 
> should be OK but, well...

...something like this sounds good, although you seem less than happy
with it :-)

(I suppose we shouldn't literally use MIN on get_object_alignment,
since it's a bit too expensive to evaluate twice.)

Richard