Re: Frontend pass assumptions

[pocmatos at gmail dot com (Paulo J. Matos)]

Richard Guenther <richard.guenther@gmail.com> writes:

> On Fri, Sep 10, 2010 at 4:06 PM, Paulo J. Matos <pocmatos@gmail.com> wrote:
>>
>> Or I could move the pass to be executed before the CFG is created. Would
>> there be any issues with this?
>
> Yes.  Profile information is not available at that point.
>

Thanks. Makes sense if I intend to use the profile information in the
future, which I currently don't. Back to fixing it! :)

>> Regarding passes execution order. It seems on one hand that passes
>> execution order would be determined by GCC during runtime depending on
>> the properties required and provided by each pass. But on the other
>> passes.c seem to schedule the passes explicitly through NEXT_PASS().
>> Could you just briefly tell me how passes are actually scheduled?
>
> They are scheduled statically specified by the schedule initialized
> in passes.c.
>

I see, and then the properties provided and required are checked for. If
they don't agree with the current position of the pass in the scheduling
table, the thing breaks.

-- 
PMatos