Re: TODO_rebuild_alias and -O0

[Richard Guenther <richard dot guenther at gmail dot com>]

On Mon, Oct 8, 2012 at 6:04 PM, Ludovic Courtès
<ludovic.courtes@inria.fr> wrote:
> Richard Guenther <richard.guenther@gmail.com> skribis:
>
>> On Mon, Oct 8, 2012 at 3:32 PM, Ludovic Courtès <ludo@gnu.org> wrote:
>>> Richard Guenther <richard.guenther@gmail.com> skribis:
>>>
>>>> On Mon, Oct 8, 2012 at 1:42 PM, Ludovic Courtès <ludo@gnu.org> wrote:
>>>>> Richard Guenther <richard.guenther@gmail.com> skribis:
>>>>>
>>>>>> On Mon, Oct 8, 2012 at 11:58 AM, Ludovic Courtès <ludo@gnu.org> wrote:
>>>>>>> Richard Guenther <richard.guenther@gmail.com> skribis:
>>>>>>>
>>>>>>>> At -O0 no virtual operands are produced.  TODO_rebuild_alias only computes
>>>>>>>> points-to sets which are in itself not useful.
>>>>>>>>
>>>>>>>> What do you want to achieve with TODO_rebuild_alias?
>>>>>>>
>>>>>>> I basically want to use ‘ptr_derefs_may_alias_p’ in this particular pass.
>>>>>>
>>>>>> That should work.
>>>>>
>>>>> It actually does, except that ‘ptr_derefs_may_alias_p’ returns true for
>>>>> two SSA names in cases like this:
>>>>>
>>>>>   double *p, *q;
>>>>>   p = malloc (123);
>>>>>   q = malloc (234);
>>>>>
>>>>> (Where ‘malloc’ has the ‘malloc’ attribute.)
>>>>>
>>>>> For some reason, there’s no such false positive when using TODO_rebuild_alias.
>>>>
>>>> Well, "that should work" means it works non-conservatively with
>>>> TODO_rebuild_alias, of course.
>>>
>>> Right.  :-)  So how can I get maximum accuracy, while avoid the assertion
>>> failure on -O0?
>>
>> You have to analyze the assert more.  From the backtrace I cannot see
>> anything suspicious.
>
> The assert in ‘make_decl_rtl’ is:
>
>   gcc_assert (TREE_CODE (decl) != VAR_DECL
>               || TREE_STATIC (decl)
>               || TREE_PUBLIC (decl)
>               || DECL_EXTERNAL (decl)
>               || DECL_REGISTER (decl));
>
> Upon inspection, ‘decl’ is VAR_DECL, but has all 4 flags above cleared.
>
> In my example source file, the assertion is hit here:
>
>   numbers.c: In function ‘scm_c_make_polar’:
>   numbers.c:8728:10: internal compiler error: in make_decl_rtl, at varasm.c:1163
>
> That line in numbers.c reads this:
>
>   sincos (ang, &s, &c);
>
> And indeed, the dump of my pass (with -ftree-dump-all) reads this:
>
>   ;; Function scm_c_make_polar (scm_c_make_polar)
>
>   Now a gimple register: D.26884
>   No longer having address taken: s
>   No longer having address taken: c
>
>
>   Symbols to be put in SSA form
>
>   { s c D.26884 }
>
>
>   Incremental SSA update started at block: 0
>
>   Number of blocks in CFG: 8
>   Number of blocks to update: 7 ( 88%)
>
> So it looks like there’s additional state corresponding to these
> variables that needs updating?

Ah, no, but I suppose TODO_update_address_taken doesn't really work at -O0
so you need to fix that to do nothing at -O0.  Because s and c _do_ have their
address taken.

Richard.

> Thanks,
> Ludo’.