[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 Richard Guenther changed: What|Removed |Added Status|UNCONFIRMED |RESOLVED Resolution|

[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 --- Comment #11 from Richard Guenther 2011-01-20 10:36:32 UTC --- Author: rguenth Date: Thu Jan 20 10:36:29 2011 New Revision: 169051 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=169051 Log: 2011-01-20 Richard Guenther PR tree-o

[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 --- Comment #10 from Richard Guenther 2011-01-20 10:33:18 UTC --- Author: rguenth Date: Thu Jan 20 10:33:15 2011 New Revision: 169050 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=169050 Log: 2011-01-20 Richard Guenther PR tree-o

[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 --- Comment #9 from Martin Reinecke 2011-01-19 17:26:31 UTC --- (In reply to comment #8) > Can you check if the following patch solves your problem? Yes, this patch gets performance back to normal on the 4.5 branch and on trunk. Great! > The di

[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 Richard Guenther changed: What|Removed |Added CC||rguenth at gcc dot gnu.org --- Comment

[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 --- Comment #7 from Martin Reinecke 2011-01-19 14:16:18 UTC --- OK, I located the problematic commit, at least on the 4.5 branch: it's revision number 167492 (fix for PR tree-optimization/46806). Between revisions 167491 and 167492 the CPU time

[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 --- Comment #6 from Uros Bizjak 2011-01-06 07:38:11 UTC --- (In reply to comment #5) > Some loop performance is very sensitive to code sizes. This change > > -mulpd%xmm10, %xmm2 > +mulpd%xmm0, %xmm2 > > will impact loop size d

[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 --- Comment #5 from H.J. Lu 2011-01-05 20:09:11 UTC --- (In reply to comment #3) > > this could be the reason for slowdown. > > > $ gcc -lm testcase2.s > $ time ./a.out > > real0m4.239s > user0m4.234s > sys0m0.001s > > The im

[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 --- Comment #4 from Uros Bizjak 2011-01-05 19:48:58 UTC --- Applying the same medicine to original test gets us from: wall time for map2alm: 6.908527s to wall time for map2alm: 6.703142s where 4.5.1 wins with: wall time for map2alm: 6.651740

[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 --- Comment #3 from Uros Bizjak 2011-01-05 19:30:49 UTC --- > this could be the reason for slowdown. Hm, not really. But, by changing the generated assembly around loop entry: $ diff -u testcase2.s testcase2_.s --- testcase2.s2011-01-05 20

[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 --- Comment #2 from Uros Bizjak 2011-01-05 17:31:20 UTC --- The only difference in the hot loop is the usage of two regs in the address: 4.5.1: .L142: movapd%xmm0, (%rcx) mulpd%xmm6, %xmm2 addq$32, %rbx movapd%xm

[Bug other/47167] Performance regression in numerical code

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47167 --- Comment #1 from Martin Reinecke 2011-01-05 14:42:20 UTC --- Created attachment 22904 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22904 shorter test case More compact test case; the hot spot is marked with "CRITICAL LOOP". Compile wit