Re: More memory reductions on PR c++/12850

[Jan Hubicka <hubicka at ucw dot cz>]

> > Yes, with SPARC cross.  Perhaps I do something silly but it didn't crash
> > for me.   (perhaps it is because I built sparc-linux, you do
> > sparc-solaris right?)
> 
> Yes, I build for sparc-sun-solaris2.* but according to
> 
> http://gcc.gnu.org/ml/gcc-testresults/2004-02/msg00254.html
> 
> the failures are also present on sparc64-unknown-linux-gnu and
> 
> http://gcc.gnu.org/ml/gcc-testresults/2004-02/msg00431.html
> 
> on sparc64-unknown-freebsd5.2 too.
> 
> [The 64 or not 64 doesn't play any role here.]

Aparently i got caught by using tree with merged tree-ssa cgrap code
that somehow manages to get across that bug. Sorry for that.
The patch has however uncovered latent problem in tree-optimize.c.
We have function that gets compiled, its body released and later the
DECL_ARGUMENTS are accessed by use_thunk that gets to the structure by
THUNK_TARGET.
Originally this dind't reproduce because bug in cgraph_finalize made
tree-optimizie to believe thta tthis is nested function and we don't
release the data in this case.

Mark, I have little understanding of the patch.  Do you think there is
better sollution than giving up the idea of freeing argument lists?
I will try do some measurements how much memory this consume.

Honza
> 
> 
> The ICE occurs on a sanity check.  Could it be the case that it requires 
> ENABLE_CHECKING to trigger and that you didn't configure with E_C on the 3.4 
> branch?
> 
> -- 
> Eric Botcazou