> > Honza reports high CPU usage from lookup_external_ref at LTRANS time. > This is likely because build_abbrev_table calls it on all DIEs > rather than just type DIEs which we populate the ref table with. > So we end up calling it for all the DIEs refering to early debug, > possibly many distinct ones with the same die.symbol (but usually > different die_offset). That will end up with loads of hashtable > collisions. > > The following patch restricts the build_abbrev_table lookups to > type DIEs. > > We might consider optimizing the refs to decrease the number of > relocations the link editor has to process but I guess it's not > worth it and the rest of the optimize_external_refs isn't prepared > for the LTO case anyways. > > Bootstrapped and tested on x86_64-unknown-linux-gnu. > > Honza has yet to confirm the patch helps, once he does I plan to > install it.
Before patch I got about 5% of lookup_external_ref while building cc1 and after patch the function is off the profile :) Honza > > Richard. > > 2018-10-31 Richard Biener <rguent...@suse.de> > > * dwarf2out.c (build_abbrev_table): Guard lookup_external_ref call > with is_type_die. > > diff --git a/gcc/dwarf2out.c b/gcc/dwarf2out.c > index 30bbfee9052..8b478aa265f 100644 > --- a/gcc/dwarf2out.c > +++ b/gcc/dwarf2out.c > @@ -9023,8 +9023,9 @@ build_abbrev_table (dw_die_ref die, > external_ref_hash_type *extern_map) > struct external_ref *ref_p; > gcc_assert (AT_ref (a)->comdat_type_p || AT_ref (a)->die_id.die_symbol); > > - ref_p = lookup_external_ref (extern_map, c); > - if (ref_p->stub && ref_p->stub != die) > + if (is_type_die (c) > + && (ref_p = lookup_external_ref (extern_map, c)) > + && ref_p->stub && ref_p->stub != die) > change_AT_die_ref (a, ref_p->stub); > else > /* We aren't changing this reference, so mark it external. */
