On Tue, 18 Aug 2026, [email protected] wrote: > From: Kyrylo Tkachov <[email protected]> > > The parallel check targets start 128 runtest instances, or as many as > GCC_TEST_PARALLEL_SLOTS says. The -j level make was given is looked at only > to decide whether to go parallel at all, never for how parallel to be. > > That constant predates machines with more cores than it, and it is now what > limits the testsuite on them. On a 352 core machine the testsuite runs 128 > wide whatever -j says, and the rest of the machine is idle. > > Take the number from MFLAGS, which GNU make 4.0 and later set, but never go > below the old constant. The instances claim work in batches of ten tests > rather than taking a fixed share each, so an instance that draws a long batch > is the tail of the whole target, and instances beyond the token count are > there to pick up the slack. On x86_64 at -j12 check-gcc > takes a median of 772.8 s with 12 slots, 718.5 s with 24, 717.7 s with 48 and > 732.0 s with 128. Following -j alone is the slowest of the four. > So the floor is retained, and at 128 no -j below 128 changes behaviour at all. > > A bare -j and older make do not give a number, and there the floor is > retained as well. An explicit GCC_TEST_PARALLEL_SLOTS still wins, and the per > language check_$lang_parallelize caps still apply. > > On aarch64 at -j352, check-gcc falls by 24%, from a median of 118.1 s to > 89.3 s over four interleaved rounds, with identical results in every round. > It also runs far more repeatably: the spread across rounds falls from 10.6% > to 4.5%, and the slowest patched round is still 22% faster than the median > unpatched one. > > Ok for trunk?
OK. Please give others a chance to comment. Thanks, Richard. > Thanks, > Kyrill > > gcc/ChangeLog: > > * Makefile.in (check_p_min_slots, check_p_jobs, check_p_slots): New > variables. > (check_p_subdirs): Use check_p_slots. > > Signed-off-by: Kyrylo Tkachov <[email protected]> > --- > gcc/Makefile.in | 19 ++++++++++++++++++- > 1 file changed, 18 insertions(+), 1 deletion(-) > > diff --git a/gcc/Makefile.in b/gcc/Makefile.in > index 5c7cb98d22b..c718585e607 100644 > --- a/gcc/Makefile.in > +++ b/gcc/Makefile.in > @@ -4760,8 +4760,25 @@ check_p_tool=$(firstword $(subst _, ,$*)) > check_p_count=$(check_$(check_p_tool)_parallelize) > check_p_subno=$(word 2,$(subst _, ,$*)) > check_p_subdir=$(subst _,,$*) > +# Number of runtest instances to start. An explicit GCC_TEST_PARALLEL_SLOTS > +# wins. Otherwise start one per job token, but never fewer than > +# check_p_min_slots: the instances claim work in batches of ten tests, so a > +# batch that runs long is the tail of the whole target, and having more > +# instances than tokens lets the ones that are left pick up the slack. Above > +# that floor the token count is the useful number, since 128 instances on a > +# machine given -j256 leave half of it idle. GNU make 4.0 and later put the > +# level in MFLAGS. A bare -j, and older make, do not, and there the floor > +# stands on its own. > +check_p_min_slots=128 > +check_p_jobs=$(patsubst -j%,%,$(filter -j%,$(MFLAGS))) > +# max(check_p_jobs, check_p_min_slots). A number is <= the floor exactly > when > +# it appears among the first check_p_min_slots counting numbers. > +check_p_slots=$(strip $(if > $(GCC_TEST_PARALLEL_SLOTS),$(GCC_TEST_PARALLEL_SLOTS),\ > + $(if $(check_p_jobs),$(if $(filter $(check_p_jobs),\ > + $(wordlist 1,$(check_p_min_slots),$(one_to_9999))),\ > + $(check_p_min_slots),$(check_p_jobs)),$(check_p_min_slots)))) > check_p_subdirs=$(wordlist 1,$(check_p_count),$(wordlist 1, \ > - $(if > $(GCC_TEST_PARALLEL_SLOTS),$(GCC_TEST_PARALLEL_SLOTS),128), \ > + $(check_p_slots), \ > $(one_to_9999))) > > # For parallelized check-% targets, this decides whether parallelization > -- Richard Biener <[email protected]> SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg, Germany; GF: Jochen Jaser, Andrew McDonald, Abhinav Puri; (HRB 36809, AG Nuernberg)
