https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116398
--- Comment #25 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
And some further attempts in the 1500..5000 range:
$ for i in 1500 2000 2500 3000 3500 4000 5000; do echo COMBLIM=$i time
./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128 -march=z196
-fpreprocessed -w; COMBLIM=$i time ./cc1plus -quiet -nostdinc -O2 pr101523.ii
-mlong-double-128 -march=z196 -fpreprocessed -w; done
COMBLIM=1500 time ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w
14.12user 0.15system 0:14.32elapsed 99%CPU (0avgtext+0avgdata
316544maxresident)k
0inputs+824outputs (0major+106337minor)pagefaults 0swaps
COMBLIM=2000 time ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w
14.44user 0.12system 0:14.63elapsed 99%CPU (0avgtext+0avgdata
318616maxresident)k
0inputs+824outputs (0major+104871minor)pagefaults 0swaps
COMBLIM=2500 time ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w
14.61user 0.13system 0:14.79elapsed 99%CPU (0avgtext+0avgdata
318880maxresident)k
0inputs+848outputs (0major+106934minor)pagefaults 0swaps
COMBLIM=3000 time ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w
15.04user 0.15system 0:15.26elapsed 99%CPU (0avgtext+0avgdata
320672maxresident)k
0inputs+856outputs (0major+106836minor)pagefaults 0swaps
COMBLIM=3500 time ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w
15.54user 0.13system 0:15.73elapsed 99%CPU (0avgtext+0avgdata
318856maxresident)k
0inputs+840outputs (0major+107193minor)pagefaults 0swaps
COMBLIM=4000 time ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w
15.53user 0.12system 0:15.72elapsed 99%CPU (0avgtext+0avgdata
321640maxresident)k
0inputs+856outputs (0major+107824minor)pagefaults 0swaps
COMBLIM=5000 time ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w
17.20user 0.15system 0:17.41elapsed 99%CPU (0avgtext+0avgdata
306276maxresident)k
0inputs+872outputs (0major+107194minor)pagefaults 0swaps
$ for i in 1500 2000 2500 3000 3500 4000 5000; do rm pr101523.ii.*; echo
COMBLIM=$i ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w -fdump-rtl-combine-stats; COMBLIM=$i ./cc1plus
-quiet -nostdinc -O2 pr101523.ii -mlong-double-128 -march=z196 -fpreprocessed
-w -fdump-rtl-combine-stats; awk
'/^attempts:/{attempts+=$2}/^merges:/{merges+=$2}/^extras:/{extras+=$2}/^successes/{successes+=$2}/^two-insn
combine:/{two+=$3}/^three-insn combine:/{three+=$3}/^four-insn
combine:/{four+=$3}/^insn-with-note combine/{note+=$3}END{printf "attempts:
%d\nmerges: %d\nextras: %d\nsuccesses: %d\ntwo-insn combine: %d\nthree-insn
combine: %d\nfour-insn combine: %d\ninsn-with-note combine: %d\n", attempts,
merges, extras, successes, two, three, four, note}' pr101523.ii.*combine*; done
COMBLIM=1500 ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w -fdump-rtl-combine-stats
attempts: 50696
merges: 49743
extras: 14793
successes: 4383
two-insn combine: 4382
three-insn combine: 1
four-insn combine: 0
insn-with-note combine: 0
COMBLIM=2000 ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w -fdump-rtl-combine-stats
attempts: 59541
merges: 58588
extras: 17422
successes: 5229
two-insn combine: 5228
three-insn combine: 1
four-insn combine: 0
insn-with-note combine: 0
COMBLIM=2500 ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w -fdump-rtl-combine-stats
attempts: 63403
merges: 62450
extras: 18572
successes: 5606
two-insn combine: 5605
three-insn combine: 1
four-insn combine: 0
insn-with-note combine: 0
COMBLIM=3000 ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w -fdump-rtl-combine-stats
attempts: 63558
merges: 62605
extras: 18727
successes: 5762
two-insn combine: 5761
three-insn combine: 1
four-insn combine: 0
insn-with-note combine: 0
COMBLIM=3500 ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w -fdump-rtl-combine-stats
attempts: 64203
merges: 63250
extras: 19372
successes: 6416
two-insn combine: 6415
three-insn combine: 1
four-insn combine: 0
insn-with-note combine: 0
COMBLIM=4000 ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w -fdump-rtl-combine-stats
attempts: 70572
merges: 69616
extras: 21019
successes: 6745
two-insn combine: 6744
three-insn combine: 1
four-insn combine: 0
insn-with-note combine: 0
COMBLIM=5000 ./cc1plus -quiet -nostdinc -O2 pr101523.ii -mlong-double-128
-march=z196 -fpreprocessed -w -fdump-rtl-combine-stats
attempts: 99213
merges: 98257
extras: 28052
successes: 7632
two-insn combine: 7631
three-insn combine: 1
four-insn combine: 0
insn-with-note combine: 0
So, I think even defaulting to say 3000 could work pretty well.
It would be nice to find out what is the largest number seen say during
x86_64/i686/aarch64/powerpc64le bootstraps/regtests, so that the chosen default
limit doesn't trigger too often.
And ideally make it a param.
