cmd/go: do not cache tool output if tools print to stdout/stderr

[cherrymui]

Update, Oct 7 2020: see #27628 (comment) for most recent proposal in this issue.

What version of Go are you using (go version)?

tip (2e5c325), also Go 1.11

Does this issue reproduce with the latest release?

Yes

What operating system and processor architecture are you using (go env)?

darwin/amd64

What did you do?

$ go build -toolexec=/usr/bin/time hello.go
# command-line-arguments
        0.01 real         0.00 user         0.00 sys
# command-line-arguments
        0.12 real         0.11 user         0.02 sys
$ go build hello.go 
# command-line-arguments
        0.01 real         0.00 user         0.00 sys
# command-line-arguments
        0.12 real         0.11 user         0.02 sys
$

What did you expect to see?

The second invocation of go build doesn't have -toolexec, so it should not invoke the toolexec command (which I think it doesn't), nor reprint its output.

What did you see instead?

toolexec output is reprinted.

In fact, I think it probably should not cache at all if -toolexec is specified, since the external command that toolexec invokes may do anything, and (intentionally) not reproducible.

cc @dr2chase

[dr2chase]

Using a cooked command "wrap" =

#!/bin/bash
echo -n `basename "$1"` 1>&2 ; /usr/bin/time "$@"
date 1>&2 

it keeps reprinting the same date+time. This is definitely cached output.

go build -toolexec wrap .
go build .
rm hello
go build .

[mark-rushakoff]

#27207 is about -exec prevents test caching. Maybe there needs to be a more holistic review of -exec, -toolexec, and friends, and how they all relate to caching.

[bcmills]

I'm not sure whether we should cache the result of -toolexec (see #27207 (comment)), but if we do, the value of the -toolexec flag should be included in the cache key.

So there's definitely a problem here either way: we just need to decide whether the fix is to update the cache key or exclude the result from caching.

[ianlancetaylor]

I actually don't think those are the right choices. There is a plausible use of -toolexec to invoke a debugger or other interactive program, in which case we should never cache if -toolexec is used. And there is a plausible use of -toolexec to invoke a logger or other build annotator, in which we should completely ignore -toolexec and its argument for caching purposes. I can't think of an important use case in which we want to cache based on whether we are using the same -toolexec option as before.

I would probably pick "never cache if -toolexec is used" but I don't feel very strongly about it.

[cherrymui]

I would probably pick "never cache if -toolexec is used"

I vote for this.

[bcmills]

It looks like we have a consensus. (Moving from NeedsDecision to NeedsFix.)

[mvdan]

I agree that this bug needs fixing one way or another. The only way I find -toolexec useful right now is like -a -toolexec=whatever.

I have some thoughts about caching, though. It's unfortunate if we simply throw caching out the window, because it's entirely reasonable for toolexec programs to be deterministic given their input.

The way go build handles caching properly for go tool compile seems to be to first ask it compile -V=full, and use that as part of the cache key. This means the compile operations will need re-doing if the compiler version changes, but not otherwise.

Could we do the same for -toolexec as an opt-in? For example, if one does go build -toolexec=mytool, the Go tool would run a well defined version flag like mytool -toolversion, similar to the compiler's -V=full. If that fails, we do no caching at all with -toolexec=mytool. If it succeeds, its output is added to the cache key, and caching takes place.

I think this is the best of both worlds, because current non-deterministic toolexec programs would continue to use no caching at all, while other Go-specific complex tools could take advantage of the build cache.

We would have to define the bare minimum of information for -toolversion to report, though. For example, would a stringified https://golang.org/pkg/runtime/debug/#Module be enough, if available? Or perhaps a hash of the file contents in https://golang.org/pkg/os/#Executable?

[mvdan]

Thinking about my comment above a bit more, I think this issue should still be fixed by simply doing no caching when -toolexec is used. That would be strictly better than what we have right now.

Once this issue is fixed, I can then file the "opt-in caching for -toolexec" idea as a separate proposal.

[rsc]

Going to mark this as a proposal since the answer is unclear and it is bound up in #41145. Maybe they should be merged.

[bcmills]

It turns out that this also affects the cmd/go.TestScript/toolexec and cmd/go.TestScript/vet.

The vet test expects go vet to be silent for packages with no vet errors, but if the toolexec test ends up rebuilding shared packages in the standard library (e.g. because they are otherwise stale; see #46347), its output will pollute the vet test.

[janflyborg]

I have seen the same thing with printouts from a previous run with GODEBUG=gctrace=1, being printed, even if I start without that environment variable, later on (please see https://groups.google.com/g/golang-nuts/c/UP9hIrvqDj4 for a better explanation).

Is this related?

[gopherbot]

Change https://golang.org/cl/359614 mentions this issue: cmd/go: do not cache compiler or linker output

[bcmills]

I tried to make a run at this for Go 1.18, but ran into what appear to be some existing bugs in go test output caching that I think I'm going to have to detangle to get it into a testable state. 😞

[bcmills]

(Filed the above test-caching bug as #49267.)

[bcmills]

I don't think this is going to land for 1.21. (Too many existing caching bugs to work through to get it to a testable state.)

[bcmills]

Moving to Backlog. This will require diagnosing and fixing the existing caching bugs so that we can write an integration test for the new caching behavior, and that is a fairly large project.