cmd/compile: Fannkuch benchmark performance regression

[kklobe]

Please answer these questions before submitting your issue. Thanks!

What version of Go are you using (go version)?

go version go1.11rc1 darwin/amd64

Does this issue reproduce with the latest release?

No.

What operating system and processor architecture are you using (go env)?

macOS 10.13.6 (darwin/amd64)

What did you do?

tested the benchmark at:

https://benchmarksgame-team.pages.debian.net/benchmarksgame/program/fannkuchredux-go-1.html

using 1.10.3 and 1.11rc1 with:

time go run fannkuchredux.go 12
time go1.11rc1 run fannkuchredux.go 12

results:
1.10.3: go run fannkuchredux.go 12 44.65s user 0.21s system 372% cpu 12.033 total
1.11rc1: go1.11rc1 run fannkuchredux.go 12 57.70s user 0.24s system 373% cpu 15.528 total

What did you expect to see?

1.11rc1 timings closer to 1.10.3

What did you see instead?

1.11rc1 timings slower than 1.10.3

[bradfitz]

/cc @randall77

[TocarIP]

Disabling branchelim seems to bring performance back to 1.10 level. Interestingly enough, on test/bench/go1/fannkuch disabling CMOV hurts performance by 20%

[navytux]

Possibly related issue about still unfixed Fannkuch11 performance regression from Go1.8 -> Go1.9 time: #18977.

For the reference: Fannkuch11 performance depends heavily on BCE: #24660.

[TocarIP]

I've looked at instruction-level profile and found that all conditional moves are out of the hot-spot. Main problem is that enabling branchelim pass caused different register allocation, which in turn put register spills/fills inside the inner loop.

[randall77]

Punting to 1.13, too late for anything major in 1.12.