cmd/go: TestScript/build_trimpath_goroot fails during the release process

[heschi]

When run under x/build/cmd/release, the test fails in all longtest configurations:

--- FAIL: TestScript (0.01s)
    --- FAIL: TestScript/build_trimpath_goroot (0.04s)
        script_test.go:260: 
            # Regression test for https://go.dev/issue/51461 and https://go.dev/issue/51483.
            #
            # When built with -trimpath, runtime.GOROOT() returned the bogus string "go"
            # if GOROOT was not set explicitly in the environment.
            # It should instead return the empty string, since we know that we don't
            # have a valid path to return.
            #
            # TODO(#51483): when runtime.GOROOT() returns the empty string,
            # go/build should default to 'go env GOROOT' instead. (0.044s)
            > env GOROOT=
            > env GOROOT_FINAL=
            > [trimpath] ! go env GOROOT
            > [trimpath] stderr '^go: cannot find GOROOT directory: ''go'' binary is trimmed and GOROOT is not set$'
            > [trimpath] stop
            > [short] skip
            > go run .
            [stderr]
            go: cannot find GOROOT directory: /usr/local/go
            [exit status 2]
            FAIL: testdata/script/build_trimpath_goroot.txt:21: unexpected command failure
            
FAIL
FAIL    cmd/go  244.852s

To reproduce, run something like:
release -target linux-amd64-longtest -user $USER -version go1.19beta1 -rev 957cc90c9b350a1d8ef90d992298dfe33cc13980

[heschi]

Worth noting that my redesigned release process in https://go.dev/cl/398694 doesn't exhibit the problem. I'd be interested to know if this indicates an improvement or a bug in the new process.

[bcmills]

I can reproduce this fairly trivially by running:

GOROOT_FINAL=/bogus GOROOT=$(go env GOROOT) ../bin/go test cmd/go -run=TestScript/build_trimpath_goroot

#51504 is peripherally related, although this particular case is definitely a bug in the test (because it reproduces even if I set GOROOT explicitly to the correct value).

[bcmills]

I'd be interested to know if this indicates an improvement or a bug in the new process.

I'm going with “probably bug”, because the fact that the test doesn't fail seems to imply that GOROOT_FINAL isn't set to its final value while the tests are being built and run.

[heschi]

Ah, yeah, that's correct -- I thought we wouldn't need to set GOROOT_FINAL while running the tests, but that was at least somewhat while I was hoping to use run.bash. OK, will fix.

[gopherbot]

Change https://go.dev/cl/399156 mentions this issue: cmd/go: fix TestScript/build_trimpath_goroot when built with a mismatched GOROOT_FINAL