GCC inline parameters (PR 10160 testcase)

[Jan Hubicka]

> > Guess what?  This joke on memory abuse goes away with "-O3
> > -fno-unit-at-a-time".
> 
> First, -funit-at-a-time is not in 3.3, so this is a bit off-topic
> relative to Eric's comment.
> 
> Second, Jan's analysis just points out that something is wrong with the
> inlining strategy, not the particular heuristic to which you point.
> 
> Third, if -O3 enables -funit-at-a-time, then whoever did that has
> introduced a regression, and should fix the problem -- perhaps by taking
> -funit-at-a-time out of -O3.
I will look into this.  Of course a bit of rearanging of the testcase
produces a testcase that does not need -funit-at-a-time.
With -funit-at-a-time we can actually estimate the overall growth of
whole compilation unit via inlining and I think that is something we
should limit to avoid such a explossions. (or perhaps do that per
function basis but I guess the higher granuality we get the better, but
then we won't work around problems like O(n^2) scheduler.

I currently have to finish 3 papers so I will dig into that ASAP
afterwards (ie my reply to the bugreport didn't mean that I am not going to
look into this just wanted to explain that it is more a generic problem
of inlining)

Honza
> 
> Yours,
> 
> -- 
> Mark Mitchell <mark@codesourcery.com>
> CodeSourcery, LLC