cmd/vet: decide how to handle type checking failure

[ianlancetaylor]

This issue is to decide how cmd/vet should handle type checking failures.

In https://golang.org/cl/7393052 (pre-1.1) type checking was added to cmd/vet. A type checking error would issue a warning and cause vet to exit with status 1.

Issue #4895 complained about this behavior, because it meant that the type checker had to be able to find imported packages. In that issue Russ said "It would definitely be nice for 'go vet' to be able to run in degraded mode. I ran it on a large uncompiled tree over the weekend and was pretty annoyed by all the import messages."

https://golang.org/cl/7399051 (also pre-1.1) addressed this problem by only issuing the warning, and setting the exit status, if vet were run with the -v option.

People rarely ran vet with the -v option, because there was no way to do so when using go vet; it could only be done by using go tool vet.

https://golang.org/cl/40112 (pre-1.9) changed cmd/go to pass more flags to vet. In particular, after that change, go vet -v invoked cmd/vet with the -v option.

This led to issue #21188 complaining that go vet -v now caused unexpected errors when a package was not available (for example, the special "C" package used by cgo). This was essentially a repeat of #4895, only with the -v option that is now easier to use.

https://golang.org/cl/52851 (also pre-1.9) changed cmd/vet so that a type checking error just prints a message with -v, rather than issuing a warning and thus changing the exit status to 1.

How do we want to handle this going forward?

• Should a type checker error cause vet to exit with status 1?
• Should vet change its exit status depending on whether -v is used?
• Does it matter if the type checker error is specifically a missing import?
• Does it matter if the missing import is "C"?

CC @robpike

[mvdan]

Thanks for the thorough explanation of the history of the issue, Ian.

Does it matter if the missing import is "C"?

Something to add is #6774 - eventually, the "C" import will no longer be an error. Perhaps that's an answer to your question. If the issue won't be fixed in time for 1.10 it might not be the right answer, though.

Should vet change its exit status depending on whether -v is used?

I strongly believe this is a bad idea, because it's counter-intuitive. It's also worth noting that since -v was useless in go vet up until 1.9, the flag has never actually changed the exit status.

[robpike]

@robpike cc'ing myself.

[robpike]

I suggested we check in the CL that restores prior behavior, however awkwardly, and then address this in 1.10. It does not need to be a release blocker.

[rsc]

https://go-review.googlesource.com/c/go/+/74356 is for #18084 but the idea was to also hook it up to #16086, so that when run as "go vet", vet can expect that it will have complete type information and no magic import "C" to worry about. Then it should be a fatal error in that mode for vet to fail to typecheck.

So the decision here is "go vet" will be fixed so that it can always typecheck, and then failure to typecheck will be a fatal error. Given that decision, this is a duplicate of #16086.

[gopherbot]

Change https://golang.org/cl/121455 mentions this issue: doc: document new vet behaviour for typechecking failures