encoding/json: 36.64% regression in StdlibJSON-72 sec/op on linux/arm64 at e62d3e6

[JunyangShao]

Discovered a regression in sec/op of 36.64% for benchmark StdlibJSON-72 at e62d3e6.

Looks like introduced by go.dev/cl/773962.
We will investigate this and update the issue.

@dsnet

Perf dashboard:
https://perf.golang.org/dashboard/?benchmark=StdlibJSON-72&unit=sec/op&platform=linux/arm64#commite62d3e6e897175a07aa44a7b2c7f99700072f22f

[dsnet]

Thanks for the report. I lack an arm64 machine. How can I reproduce this results? Where's the code the benchmarks themselves?

[prattmic]

You can use a gomote, but it looks like this appears on amd64 as well: https://perf.golang.org/dashboard/?benchmark=StdlibJSON&unit=&repository=go&branch=master&platform=all&days=30&end=2026-05-20T18%3A51 [1]

We really need to get the module surfaced on this dashboard.

StdlibJSON is BenchmarkStdlibJSON in github.com/aws/aws-sdk-go/private/protocol/json/jsonutil@v1.41.1

---

There is another benchmark called StandardJSON (not sure the difference), that shows a big performance improvement instead: https://perf.golang.org/dashboard/?benchmark=StandardJSON&unit=&repository=go&branch=master&platform=all&days=30&end=2026-05-20T18%3A51

It turns out StandardJSON is in go.uber.org/zap/zapcore@v1.19.1, so completely unrelated.

[1] Note: click the unit text "sec/op" at the top of a graph to go to the per-commit table.

[prattmic]

(To be clear, the benchmarks on our dashboard aren't as heavily vetted as I'd like, so "this is a bad benchmark" is a possible answer.)

[dsnet]

I didn't look too deeply. It's not bad benchmark per-say, but also not particular representative of real-world data. BenchmarkStdlibJSON is marshaling a lot of small values, which is something I know v2 performs worse at compared to v1 (since we more aggressively heap-allocate at the top-level). I have ideas for how to improve it to gain back equivalent v1 performance, but I don't think this blocks release in Go 1.27.

[JunyangShao]

Thanks! I tentatively assign this to you @dsnet 👍 .