What was the behavior with Go 1.12? (Is this a regression?)

I see this issue as a regression, go1.12 did not fail. So I could run tests with additional custom (application relevant) flags.

I think this is likely happening due to changed init precedence, that test framework init ran after app-level init.

Looks like it's probably due to CL 173722, which was intended to address #21051.

CC @bradfitz

OK, so if that is an intended behavior I can fix my code by adding a go1.13 build tagged init() with testing.Init() call inside.

I don't think that a breaking change was intended. This seems like something that should be fixed before Go 1.13 is released.

We knew that fbc6a97 was going to break some people. It's arguably on the fence on whether it's inside the go1compat umbrella (it's half API, half tooling) and we guessed it would affect very few people. (generally more advanced people who could read the docs/release notes & work around)

That said, I haven't looked at this particular issue yet. I'll let @cespare have a look first.

It seems possible to fix this by putting the call to testing.Init in an earlier init function injected into the test variant of the package, rather than in testing.MainStart.

If I understand correctly, the init functions within a package run in the order in which the source files were presented to the compiler, right?

Fun. I'll take a look today.

What we (or at least I) intended with #21051 / CL 173722 was:

• Anyone using go test shouldn't observe any difference or need to explicitly call testing.Init.
• People using testing outside of go test (notably, testing.Benchmark within a package main) would need to use testing.Init if they want to use test flags.

So @vearutop this means that you shouldn't need to use testing.Init in your example code. In any case, it certainly shouldn't fail with such a mysterious error.

As @bcmills alluded to, the problem is that code which calls flag.Parse during initialization won't get the test flags. The initialization of the user's test package ("ptest"/"pxtest" in the go tool internal parlance) happens before the package main test runner initialization ("testmain") so there's nothing we can do in the generated testmain code to fix this.

The workaround that makes sense to me, which is what I think @bcmills is suggesting, is that we can insert a bit of code into the ptest/pxtest packages before we build them:

var _ = func() bool {
	testing.Init()
	return true
}()

(Putting this in a variable declaration expression means that it happens before flag.Parse is called in a variable initialization as well as an init function. Calling flag.Parse as part of variable declaration is ill-advised, of course, but it does work today.)

I think this is pretty straightforward. I'll send a CL soon.

If I understand correctly, the init functions within a package run in the order in which the source files were presented to the compiler, right?

Yeah, and same for variable declarations, modulo initialization dependencies. I'll ensure that the synthesized file is passed to the compiler first.

Change https://golang.org/cl/176098 mentions this issue: cmd/go: move automatic testing.Init call into generated test code

@cespare I'm afraid this is still broken.

Given somewhere/init.go:

package somewhere

import (
	"flag"
)

var StopOnFailure bool

func init() {
	flag.BoolVar(&StopOnFailure, "stop-on-failure", false,
		"Stop processing on first failed scenario.")

	flag.Parse()
}

And mytest/my_test.go:

package my_test

import (
	"fmt"
	"testing"

	"path/to/somewhere"
)

func TestSomething(t *testing.T) {
	_ = somewhere.StopOnFailure
}

func BenchmarkSomething(b *testing.B) {
	b.ReportAllocs()
	for i := 0; i < b.N; i++ {
		_ = fmt.Sprintf("%d", i)
	}
}

gotip (go version devel +1e3ffb0c90 Tue May 14 22:30:48 2019 +0000 darwin/amd64) fails:

gotip test ./mytest/...
flag provided but not defined: -test.testlogfile
Usage of /var/folders/lb/j0gz8jln36z4frqkrzmkdb0c0000gp/T/go-build623057355/b001/mytest.test:
  -stop-on-failure
    	Stop processing on first failed scenario.
FAIL path/to/mytest	0.005s
FAIL

gotip test -bench=. ./mytest/...
flag provided but not defined: -test.bench
Usage of /var/folders/lb/j0gz8jln36z4frqkrzmkdb0c0000gp/T/go-build638505735/b001/mytest.test:
  -stop-on-failure
    	Stop processing on first failed scenario.
exit status 2
FAIL	 path/to/mytest	0.005s
FAIL

@vearutop thanks for bringing it to my attention.

It turns out that this particular repro code relies on undefined behavior: it assumes that the testing package is initialized before the somewhere package. There's nothing in the spec that says it will be. And indeed, it so happens that on tip there are some initialization changes that already broke this code before any of my testing changes landed; see https://golang.org/cl/161337, https://golang.org/cl/170318, and related discussion over on #31636. If I run the repro with Go at 8515d9c, which is this commit where CL 170318 went in and before any of my testing changes, I get a failure:

~/apps/godev/bin/go test
flag provided but not defined: -test.timeout
Usage of /tmp/go-build059416027/b001/aaa.test:
  -stop-on-failure
        Stop processing on first failed scenario.
exit status 2
FAIL    aaa     0.001s

(And if I run with -race, it sometimes fails and sometimes succeeds.)

However, this all kind of doesn't matter because you could fix your repro code by making somewhere import testing, perhaps just by adding _ "testing" to the imports. Then you're guaranteed that testing is initialized first, and so it works up until my changes break it. So this is a problem.

To state it more generally, the "testinginit" workaround I did over in https://golang.org/cl/176098 ensures that testing.Init is called before the package under test is initialized. It does not ensure that testing.Init is called before that package's dependencies are initialized, so if they call flag.Parse during initialization they may not see the testing flags.

I think that a solution is to move the testinginit workaround from the package under test into the testing package. That is, go test would insert this file into the testing package:

package testing
func init() { Init() }

(This would be very simple with a test build tag, but we decided in #21360 not to do that.)

@bcmills thoughts?

Not sure to understand:

And indeed, it so happens that in Go 1.12 there were some initialization changes that already broke this code

With go1.12.5 that example case never fails for me.

Sorry, I meant to say that there are some changes on tip (that will be in Go 1.13). I updated my comment.

And also I tried adding _ "testing" to the imports:

package somewhere

import (
	"flag"
	_ "testing"
)

var StopOnFailure bool

func init() {
	flag.BoolVar(&StopOnFailure, "stop-on-failure", false,
		"Stop processing on first failed scenario.")

	flag.Parse()
}

but it did not help, same failure on gotip.

Right, I'm saying that without the _ "testing" import your repro code is buggy due to depending on undefined behavior, and that bugginess is exposed by the init ordering changes unrelated to my testing changes that are on tip.

By adding the _ "testing" import, your repro code no longer depends on undefined behavior, but it was broken by my change.

So:

• You need to fix your code if it depends on testing being initialized first without actually importing it
• We need to fix Go so that any package (not just the package under test) sees that the test flags have been registered when running as a test

It's hard to understand why code would call flag.Init in an init function that is not in the main package. That will ignore flags defined in other packages that are initialized later. I think I would want to see a clear explanation as to why we need to cater to that case at all, since it seems broken by design.

Hello,
I know I drop late to this, we also stumbled into this issue and while we kind of worked around it, I feel that what we do now to address this is a bit more convoluted than it used/needs to be.
So the essence of the problem is the following question:

• what if multiple packages of the application need for the init()-ialization part parameters that are to be extracted from the command line via Parse?. We initially implemented that with a special singleton object that was created in a configuration package that was imported by every other package needing to access some command line parameters. The configuration package did the flag.Parse in its init() function and everything worked just fine, until this change.
  Now we had to re-implement the initialization to be explicitly called via a chain of events that follow back to an initial call in the application's main function that actually triggers the flag.Parse() at a later time just to make it compatible with the new rules.
  So what I'm saying is that this change renders unusable any package init()-ialization that would be dependent on cli parameters. I don't think this should be the case, this seems an artificial limitation to me.
  I do understand the reasoning for this change, but maybe something else can be made to address this scenario as well? Maybe change in the flag.Parse() implementation to allow multiple interleaved calls to flag.Var and flag.Parse without generating errors?

@GreatSnoopy this change shipped two releases ago, over a year ago - I think if you have further suggestions or bugs, you should file a new issue.

Sorry for that, not trying to reopen the issue, just to ask a question that is closely related to this particular issue and I thought this would be the most appropriate place. I will create a new issue.

Calling flag.Parse as part of variable declaration is ill-advised

Why ?

@Wang-Kai This issue was closed a year and a half ago. Closed issues are not a good place for discussions. Thanks.