@Jiebin thank you for your report. Before we can investigate you need to compete your error report by providing us with detailed information about what you did, what you saw, and what you expected to see. These steps need to be reproducible by someone else for us to investigate.

Thank you

@davecheney Thanks. And I have provided the info above. Please tell me if you need other info.

What is go-benchmark(build)? How would I get it and run it?

What is go-benchmark(build)? How would I get it and run it?

Go-benchmark is a benchmark. "Build" is an indicator for that benchmark. You can get it from the official website or run it from Phoronix.com.

@jiebinn please understand that we are not in a position to reproduce the phoronix benchmark setup. You're going to have to meet us half way by writing down in this PR the steps required to reproduce the issue.

@randall77 Here is the steps. You can replace the go-package and then run it on Ubuntu or Clear Linux.

1. cd ~
   wget https://phoronix-test-suite.com/releases/phoronix-test-suite-8.6.1.tar.gz
2. tar zxf ~/phoronix-test-suite-8.4.1.tar.gz
3. cd ~/phoronix-test-suite
   sudo ~/phoronix-test-suite/install-sh
4. phoronix-test-suite
   y
   n
5. phoronix-test-suite run go-benchmark
   1
   n
6. It will show the result of go-benchmark-"build".

Is it possible to extract the code that the phoronix benchmark is running? it will be building some source code at some revision. If you can identify which programs it builds, then we can reduce the scope of the problem to time go build

@davecheney Thanks. I will looking into how to run go-benchmark from the source code.

@jiebinn thank you, that will be the best way to resolve this. Once we've confirmed if there is a regression in the build benchmarks between Go 1.11 and Go 1.12 without phoronix-test-suite, then that can be addressed first. If there is no regression then the issue may be with how phoronix-test-suite invokes go build and that will probably have to be escalated to someone on the phoronix-test-suite side.

@davecheney I agree. I have found the regression on Ubuntu(11%) and Clear Linux(10%) by running it with phoronix-test-suite. Need to make sure if that's a problem of phoronix-test-suite.

@davecheney Here is the steps for go-benchmark(build).

1. wget http://www.phoronix-test-suite.com/benchmark-files/golang-benchmarks-121017.tar.gz
2. tar zxvf golang-benchmarks-121017.tar.gz
   go-benchmark-04122017/x/benchmarks/build
3. go build "build.go" and run it.

I don't see any regression from 1.11.2 to 1.12.2. In fact, 1.12.2 is a tad faster.

name   old time/op                 new time/op                 delta
Build                  31.9s ± 2%                  31.6s ± 1%  -1.16%  (p=0.023 n=10+10)

This is on linux/amd64.

In any case, lots of things have changed about the build system from 1.11 to 1.12, e.g. more aggressive inlining. I wouldn't expect build times to be stable. We really care about regressions in generated code speed, but it probably isn't worth chasing down a few percent in build speed.

Sorry, I lied. Just building build.go with different Go versions is not enough. You also have to change the version of the Go tool that the benchmark shells out to.
Now I get (from 1.11.2 to 1.12.2):

name   old time/op                 new time/op                 delta
Build                  21.5s ± 1%                  24.9s ± 1%  +16.19%  (p=0.008 n=5+5)
name   old binary-size             new binary-size             delta
Build                  13.0M ± 0%                  14.6M ± 0%  +12.15%  (p=0.008 n=5+5)

The larger binary size makes me think inlining is to blame for the extra time.

@randall77 are we still waiting for info? Add back the label if need be. Seems like this needs investigation.

I'm going to close this issue. It falls under the general heading of making compiling faster (see all the ToolSpeed tagged issues) and/or making binaries smaller (#6853).
It may be that inlining is too aggressive here also, that falls under #17566.