1316729 - gtest death tests don't work on Windows

Status: RESOLVED FIXED

Description

[Eric Rahm [:erahm] (out until Jan 25th)]

In bug 1316437 we noticed that death tests don't work on Windows.

Judging from test logs it looked like all gtests would run, it would hit the death test, spawn a new process, all gtests would run, it would hit the death test, spawn a new process, ad infinium.

I'm not sure if it has to do with our hacks to gtest or if it's a bug that's been fixed upstream. To reproduce you can run the windows equivalent of:

> GTEST_ALSO_RUN_DISABLED_TESTS=0 ./mach gtest STLWrapper*

Comment 1

[Ted Mielczarek [:ted] [:ted.mielczarek]]

This is almost certainly because Windows lacks fork(), so gtest probably does some fiddling under the hood to re-launch the current process. I don't know that we'll be able to fix this unless gtest offers hooks for customizing that behavior.

Comment 2

[Eric Rahm [:erahm] (out until Jan 25th)]

(In reply to Ted Mielczarek [:ted.mielczarek] from comment #1)
> This is almost certainly because Windows lacks fork(), so gtest probably
> does some fiddling under the hood to re-launch the current process. I don't
> know that we'll be able to fix this unless gtest offers hooks for
> customizing that behavior.

In theory it's supposed to launch a new process and run the test, so the question is why it's not doing that. I would guess our gtest main hook isn't reading command-line args.

Comment 3

[Eric Rahm [:erahm] (out until Jan 25th)]

(In reply to Eric Rahm [:erahm] from comment #2)
> (In reply to Ted Mielczarek [:ted.mielczarek] from comment #1)
> > This is almost certainly because Windows lacks fork(), so gtest probably
> > does some fiddling under the hood to re-launch the current process. I don't
> > know that we'll be able to fix this unless gtest offers hooks for
> > customizing that behavior.
> 
> In theory it's supposed to launch a new process and run the test, so the
> question is why it's not doing that. I would guess our gtest main hook isn't
> reading command-line args.

That definitely appears to be the problem, we implement our own main and initialize without passing in args [1], vs the standard gtest main which does [2]. The Windows death test spawns a new process and needs to pass it args indicating this is a child death test [3], but of course those aren't parsed because of [1].

[1] http://searchfox.org/mozilla-central/rev/4b6cab91f93c73ae591dafaea40fd5704b41810e/testing/gtest/mozilla/GTestRunner.cpp#82
[2] http://searchfox.org/mozilla-central/rev/4b6cab91f93c73ae591dafaea40fd5704b41810e/testing/gtest/gtest/src/gtest_main.cc#37
[3] http://searchfox.org/mozilla-central/rev/4b6cab91f93c73ae591dafaea40fd5704b41810e/testing/gtest/gtest/src/gtest-death-test.cc#639-643

Comment 4

[Eric Rahm [:erahm] (out until Jan 25th)]

https://treeherder.mozilla.org/#/jobs?repo=try&revision=096c1b8440c3

Comment 5

[Eric Rahm [:erahm] (out until Jan 25th)]

https://treeherder.mozilla.org/#/jobs?repo=try&revision=4d200c3a1d76

Comment 6

[Eric Rahm [:erahm] (out until Jan 25th)]

Created attachment 8810492 [details] [diff] [review]
Pass command-line args to gtest

MozReview-Commit-ID: 8Xs52E9NnCq

Comment 7

[Benjamin Smedberg]

Comment on attachment 8810492 [details] [diff] [review]
Pass command-line args to gtest

stealing review

Comment 8

[Eric Rahm [:erahm] (out until Jan 25th)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/8df58dfdea13c9c79975abb4e0c6514b1b4da5f6
Bug 1316729 - Pass command-line args to gtest. r=ted

Comment 9

[Wes Kocher (:KWierso)]

https://hg.mozilla.org/mozilla-central/rev/8df58dfdea13