Compilation performance comparison of GCC 3.4.2 and GCC 4.0.0 (041024) on MICO sources

[Richard Guenther]

On Mon, 25 Oct 2004 16:35:58 +0200, Steven Bosscher <stevenb@suse.de> wrote:
> On Monday 25 October 2004 14:01, Karel Gardas wrote:
> > As a first possible step, I would like to propose solving ir.cc
> > regressions, since if you do it, even -O1/-O2 compilation will be faster
> > in comparison with 3.4.2 (ir.cc is 580kB long generated file)
> 
> With the attached patchlet to overcome the problem that we basically
> assign tons of unrelated stuff to TV_EXPAND, I see this:
> 
>   tree find referenced vars:  15.39 (13%) usr   0.06 ( 2%) sys  15.89 (13%) wall
> 
> That seems too expensive for what the pass does.  Top of the oprofile:
> 
>  902280   10.4203  21489     8.9873  cc1plus                  cc1plus                  ggc_alloc_stat
>  700233    8.0869  23820     9.9622  cc1plus                  cc1plus                  cgraph_reset_static_var_maps

Maybe even using walk_tree_without_duplicates will help.  I can't see
why we don't do this.
Also we walk the tree recursively(!) in add_referenced_var through
find_vars_r.  But this
is just from a quick look.

Richard.