cmd/compile: Fannkuch11 on AMD64 slow down 6% after removing assembler backend instruction reordering

[cherrymui]

CL https://go-review.googlesource.com/c/36205/ deletes assembler backend instruction reordering. This makes Fannkuch-11 on AMD64 slow down 6%. Figure out why and improve the compiler.

[gopherbot]

CL https://golang.org/cl/38431 mentions this issue.

[josharian]

Tip (6897030) looks like a net win vs 1.8.1. Code size is smaller:

• 1.8.1: "".fannkuch t=1 size=844 args=0x10 locals=0x80
• tip: "".fannkuch STEXT size=829 args=0x10 locals=0x80

Execution time is faster, at least on my machine:

name          old time/op  new time/op  delta
Fannkuch11-8   2.74s ± 2%   2.68s ± 2%  -2.32%  (p=0.000 n=20+19)

So whatever regression occurred has at the least been compensated for, if not fixed by CL 38431.

I do see one notable regression, though, which is in the number of panicindex calls. I've filed #20332 for that.

I'm going to close this as fixed. @cherrymui or @randall77, please reopen if you disagree.

[navytux]

@josharian, sad to bring bad news but here is how it looks on my machine:

$ benchstat go18.txt gotip.txt
name          old time/op  new time/op  delta
Fannkuch11-4   2.68s ± 0%   2.94s ± 0%  +9.88%  (p=0.000 n=20+19)

in other words it is ~ 10% regression.

$ go18 version
go version go1.8.1 linux/amd64

$ go version
go version devel +ee69c21747 Thu May 11 18:28:40 2017 +0000 linux/amd64

Linux deco 4.9.0-2-amd64 #1 SMP Debian 4.9.18-1 (2017-03-30) x86_64 GNU/Linux
Intel(R) Core(TM) i7-6600U CPU @ 2.60GHz

[josharian]

@navytux will you re-measure including CL 43291?

[navytux]

let me try

[navytux]

Strangely CL 43291 make things only worse:

$ benchstat gotip.txt gotip+cl43291.txt 
name          old time/op  new time/op  delta
Fannkuch11-4   2.94s ± 0%   3.01s ± 0%  +2.29%  (p=0.000 n=19+19)

and I've doble-checked it is reproducible via recompiling (make.bash) and benchmarking tip then tip+CL then tip again then tip+CL again.

[josharian]

OK, re-opening.

One other things I see in tip vs 1.8.1 is that in 1.8.1, all the panicindex calls are at the end of the function, so jumps to them are always unlikely (forward jumps). In tip, they're scattered throughout. Not sure whether that's responsible, but I'll see if I can fix it and check whether it helps.