> So with this the discriminator we assign might depend on whether > we have debug stmts or not. We output them only to debug info, so > it should in principle not cause compare-debug issues, right? And > we don't use discriminators to affect code generation (hopefully).
This is the reason of opts.cc change: > diff --git a/gcc/opts.cc b/gcc/opts.cc > index 6ca1ec7e865..60ad633b7ff 100644 > --- a/gcc/opts.cc > +++ b/gcc/opts.cc > @@ -1411,11 +1411,14 @@ finish_options (struct gcc_options *opts, struct > gcc_options *opts_set, > opts->x_debug_info_level = DINFO_LEVEL_NONE; > } > > + /* Also enable markers with -fauto-profile even when debug info is > disabled, > + so we assign same discriminators and can read back the profile info. */ > if (!opts_set->x_debug_nonbind_markers_p) > opts->x_debug_nonbind_markers_p > = (opts->x_optimize > - && opts->x_debug_info_level >= DINFO_LEVEL_NORMAL > - && (dwarf_debuginfo_p (opts) || codeview_debuginfo_p ()) > + && ((opts->x_debug_info_level >= DINFO_LEVEL_NORMAL > + && (dwarf_debuginfo_p (opts) || codeview_debuginfo_p ())) > + || opts->x_flag_auto_profile) > && !(opts->x_flag_selective_scheduling > || opts->x_flag_selective_scheduling2)); We only consume discriminators if we produce dwarf or if we do auto-profile and they indeed must agree. With -Wauto-profile you now get compiler complain if they does not. I enable debug stmt markers in both cases, so discriminators should be the same. I tested that on spec and it seems to work. As discussed on IRC I will look into possibility of enabling compare_debug for profiledbootstrap and autoprofiledbootstrap where it is currently off. We already have function to remove debug statements, so I guess we could remove them after auto-profile annotate pass at -O0, that I plan to look into incrementally now. My immediate plan is to fix the create_gcov consumer, so things can be finally properly tested. Honza
