[PATCH] Improve alias information for arrays

[Eric Botcazou]

>   #   FRAME.1_2 = V_MAY_DEF <FRAME.1_1>;
>   FRAME.1.zoo_data_base[1]{lb: 1 sz: 7} = "Animal ";
>   D.633_3 = 1;
>   D.641_4 = &FRAME.1.zoo_data_base[1]{lb: 1 sz: 7}[D.633_3 ...]{lb: 1 sz:
> 1};
>   D.634_5 = D.641_4;
>   D.635_6 = (character[1:6] *) D.634_5;
>   #   VUSE <TMT.7_8>;
>   D.636_7 = __builtin_memcmp (D.635_6, "Animal", 6);
>
> we say D.641 points to ANYTHING and so give it a TMT, but at the
> point of initialization we have a V_MAY_DEF, which looks bogus.

I think we should have a VUSE for FRAME.1 on the call to __builtin_memcmp 
because its address precisely escapes at this call point.  Look at the 
constraint computed for

D.641_4 = &FRAME.1.zoo_data_base[1]{lb: 1 sz: 7}[D.633_3 ...]{lb: 1 sz:

It should reference FRAME.1 and not ANYTHING.

> Is there any chance btw. for Ada no longer disabling struct-aliasing?

I don't really know why it has been disabled (there is no comment).  But given 
that we have made no progress wrt to our problems with SRA, I'd doubt it.

-- 
Eric Botcazou