From: Filipe C. <fi...@gm...> - 2008-03-29 18:57:05
|
On 29 Mar, 2008, at 13:08, Julian Seward wrote: > > What about something really large, like konqueror, mozilla, OOo? > > Do you have some more comprehensive before/after numbers? > > * what are the before/after code expansion ratios for apps as a whole, > as shown by these lines (run with -v) ? > > --8732-- transtab: new 3,095 (68,098 -> 1,067,412; ratio 156:10) > > * what are the actual measured performance effects? Have you made > friends yet with 'make perf' ? > > J Also, the make perf suite: Old valgrind: -- Running tests in perf ---------------------------------------------- bigcode1 valgrind :0.23s no: 2.4s (10.6x, -----) me: 4.7s (20.3x, -----) bigcode2 valgrind :0.28s no: 6.3s (22.6x, -----) me:13.7s (49.0x, -----) bz2 valgrind :0.99s no: 2.4s ( 2.4x, -----) me:10.7s (10.8x, -----) fbench valgrind :0.56s no: 0.7s ( 1.3x, -----) me: 7.1s (12.7x, -----) ffbench valgrind :0.40s no: 0.9s ( 2.3x, -----) me: 5.2s (13.0x, -----) heap valgrind :0.28s no: 0.8s ( 2.7x, -----) me:15.7s (56.0x, -----) sarp valgrind :0.03s no: 0.2s ( 6.0x, -----) me: 3.6s (119.0x, -----) tinycc valgrind :0.31s no: 2.1s ( 6.7x, -----) me:19.2s (62.0x, -----) -- Finished tests in perf ---------------------------------------------- real 2m13.018s user 1m39.190s sys 0m32.710s Patched valgrind: -- Running tests in perf ---------------------------------------------- bigcode1 new-valgrind:0.24s no: 2.9s (12.2x, -----) me: 6.1s (25.4x, -----) bigcode2 new-valgrind:0.27s no: 7.8s (28.7x, -----) me:16.1s (59.6x, -----) bz2 new-valgrind:0.96s no: 4.8s ( 5.0x, -----) me:13.4s (14.0x, -----) fbench new-valgrind:0.57s no: 1.2s ( 2.0x, -----) me: 6.4s (11.3x, -----) ffbench new-valgrind:0.40s no: 0.9s ( 2.3x, -----) me: 5.2s (13.0x, -----) heap new-valgrind:0.27s no: 0.7s ( 2.7x, -----) me:14.5s (53.6x, -----) sarp new-valgrind:0.04s no: 0.2s ( 4.2x, -----) me: 3.6s (89.5x, -----) tinycc new-valgrind:0.31s no: 1.5s ( 4.9x, -----) me:17.9s (57.6x, -----) -- Finished tests in perf ---------------------------------------------- real 2m11.275s user 1m46.779s sys 0m23.545s This is with a x86_64 machine with vmware. Is there a way to compile the tests in 32-bit mode? Or must I do some hacking in the Makefiles? I'll revert the optimization patch so we could see if the problem is in the regalloc bugfixes or in the optimization. - Filipe Cabecinhas |