[Bug rtl-optimization/78972] New: [5/6/7 Regression] poor x86 simd instruction scheduling
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78972 Bug ID: 78972 Summary: [5/6/7 Regression] poor x86 simd instruction scheduling Product: gcc Version: 7.0 Status: UNCONFIRMED Severity: normal Priority: P3 Component: rtl-optimization Assignee: unassigned at gcc dot gnu.org Reporter: liquidsun at gmail dot com Target Milestone: --- Created attachment 40441 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=40441&action=edit example.c
[Bug target/78972] [5/6/7 Regression] poor x86 simd instruction scheduling
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78972 --- Comment #1 from Andrew M. --- gcc versions >= 5 started dropping all of the additions down to the bottom of the function instead of keeping a running total. Optimization appears to follow 4.x.x up to tree-reassoc1 where >= 5 uses slightly different addition scheduling. This stays the same until rtl-expand, where _all_ of the additions get deferred to the bottom of the function, requiring a massive stack frame and a large performance hit. No version of 4.x.x I tried had this problem, so it looks like it was introduced in 5.
[Bug target/78972] [5/6/7 Regression] poor x86 simd instruction scheduling
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78972 --- Comment #2 from Andrew M. --- Created attachment 40442 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=40442&action=edit generated code for gcc-4.9.4 example.c -O1
[Bug target/78972] [5/6/7 Regression] poor x86 simd instruction scheduling
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78972 --- Comment #3 from Andrew M. --- Created attachment 40443 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=40443&action=edit generated code for gcc-6.3 example.c -O1
[Bug rtl-optimization/78972] [5/6/7 Regression] poor x86 simd instruction scheduling
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78972 Andrew M. changed: What|Removed |Added Component|target |rtl-optimization --- Comment #4 from Andrew M. --- SSE2/AVX/AVX2 are all generated similarly poorly. Non-SIMD versions do not appear to be affected. I don't know enough (anything) about gcc to investigate any further at this point
[Bug target/78972] [5/6/7 Regression] poor x86 simd instruction scheduling
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78972 --- Comment #8 from Andrew M. --- (In reply to Andrew Pinski from comment #7) > One thing to try is -fno-tree-ter. Stack sizes for -fno-tree-ter: 4.9.4: 272 bytes 5.1-5.4: 288 bytes 6.1-6.3: 560 bytes 7: 560 bytes Performance improves a lot with -fno-tree-ter with 5.x going back to 4.9 levels, and 6.x being somewhere inbetween 4.9 and 6.x without -f-no-tree-ter.
