https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119387
--- Comment #19 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #18)
> trunk is now faster than GCC 13, using about the same amount of memory
> without -g and only slightly more with -g. I consider the regression fixed
> on trunk sofar.
>
> It also hides the GC allocation issue somewhat. The profile is now dominated
> by the C++ FE:
>
> Samples: 130K of event 'cycles:P', Event count (approx.): 137748911889
>
> Overhead Samples Command Shared Object Symbol
>
> 17.11% 22269 cc1plus cc1plus [.]
> find_substitution(tree_node*)
Some obvious improvement to
/* Now check the list of available substitutions for this mangling
operation. */
if (!abbr || tags)
for (i = 0; i < size; ++i)
if (tree candidate = (*G.substitutions)[i])
{
/* NODE is a matched to a candidate if it's the same decl node or
if it's the same type. */
if (decl == candidate
|| (TYPE_P (candidate) && type && TYPE_P (node)
&& same_type_p (type, candidate))
|| NESTED_TEMPLATE_MATCH (node, candidate))
{
write_substitution (i);
return 1;
}
}
where all the time is spent would be to separate TYPE_P substitutions
from non-TYPE_P and/or "unswitch" the linear search over all candidates
based on type && TYPE_P (node). Just performing the unswitching
improves compile-time by ~3%.
