os,os/exec: should Environ record environment variables in the test log?

[bcmills]

Since CL 81895, os.Getenv and os.LookupEnv both record the keys in the test log:
https://cs.opensource.google/go/go/+/master:src/os/env.go;l=98-115;drc=b2faff18ce28edad98303d2c3134dec1331fd7b5

However, os.Environ does not:
https://cs.opensource.google/go/go/+/master:src/os/env.go;l=137-141;drc=b2faff18ce28edad98303d2c3134dec1331fd7b5

This means that, for example, a program that iterates over os.Environ() looking for keys has different caching behavior than one that calls LookupEnv on those exact same keys. This difference is especially pronounced in tests that execute subprocesses, because the variables accessed by the subprocess (which may be very significant, especially in the case of libraries like golang.org/x/tools/go/packages!) do not end up recorded in the cache key.

(For example, see the odd-looking Getenv loop in testenv.GoToolPath!)

It's not obvious to me whether the difference is an intentional decision or an oversight. Either way, I think we should consider recording all keys in the test log whenever os.Environ or execenv.Default is called.

(CC @ianlancetaylor)

[bcmills]

This question is motivated by:

• cmd/go: add known variables to go command to the test cache key hash #32285
• cmd/go,testing: GOWORK value does not invalidate test cache #60052

[ianlancetaylor]

I note that that was actually an unresolved comment on https://go.dev/cl/81895: https://go-review.googlesource.com/c/go/+/81895/comment/486513ce_5f69647f/ .

The main problem with adding all of os.Environ to the test hash is that it means that when users run tests that call os.Environ, those tests will generally not be hashed. For example, my current environment has at least the following variables with unpredictable values: SESSION_MANAGER, SSH_AUTH_SOCK, GNOME_TERMINAL_SCREEN. Caching would apply if I ran a test twice in a row, but not if I run it, log out, log in again, and run it again.