windows exit() vs ExitProcess()

[Evan Martin]

Do any Windows experts care to comment on this patch?

https://github.com/doctorlove/ninja/commit/a4954a4f03d8e5127ee2f28691b67e5e1f5c74fd

I'm not sure I understand it.

[Rachel Blum]

This seems odd... I'm not aware of any issues with exit() vs.
ExitProcess() on Windows. I'd love to see a source for that claim.

Rachel

[Evan Martin]

Perhaps ExitProcess behaves more like _exit()? I know you're supposed
to use _exit in the presence of threads, but ninja doesn't use any
threads.

[Philip Craig]

A source for the claim of different behaviour. http://stackoverflow.com/questions/915312/different-ways-of-exiting-a-process-in-c

Xoreax Incredibuild injects threads into the process being parallelised (ninja, nmake), this is why we see a behaviour difference in ninja.

[Evan Martin]

On Wed, Jan 4, 2012 at 2:57 AM, Philip Craig <phi...@pobox.com> wrote:
> A source for the claim of different behaviour.
> http://stackoverflow.com/questions/915312/different-ways-of-exiting-a-process-in-c

Ah, so it is is really similar to _exit then, thanks!

> Xoreax Incredibuild injects threads into the process being parallelised
> (ninja, nmake), this is why we see a behaviour difference in ninja.

Wow, I didn't realize that. Do you know why?
(We used Incredibuild for Chrome but found it to be pretty unreliable,
which is too bad as it seemed pretty fancy. I think some of our
buildbot use it but few engineers do.)

[Rachel Blum]

> A source for the claim of different behaviour.
> http://stackoverflow.com/questions/915312/different-ways-of-exiting-a-process-in-c

I might misreading this, but it seems to say pretty clearly that
ExitProcess is _not_ recommended if you're running in an environment
that uses the CRT. (Which ninja is). Having said all that...

> Xoreax Incredibuild injects threads into the process being parallelised
> (ninja, nmake), this is why we see a behaviour difference in ninja.

Sigh. Since I haven't used Incredibuild for quite a while, I wasn't
aware it did that. But sure enough, they seem to have added "Automatic
Interception" for make-based builds.
(http://www.xoreax.com/xge_xoreax_grid_engine.htm#interfaces)

If you have a case where ninja reproducibly hangs due to calling
exit/_exit, I suggest also filing a bug with the Incredibuild guys -
the only way that can happen would be for them to do infinite waits in
atexit() functions they injected.

Rachel

[Scott Graham]

On Wed, Jan 4, 2012 at 7:44 AM, Evan Martin <mar...@danga.com> wrote:

On Wed, Jan 4, 2012 at 2:57 AM, Philip Craig <phi...@pobox.com> wrote:
> A source for the claim of different behaviour.
> http://stackoverflow.com/questions/915312/different-ways-of-exiting-a-process-in-c

Ah, so it is is really similar to _exit then, thanks!

I'm still a little skeptical on ExitProcess too (seems like a bug in IB if they really broke exit()), but probably could just use _exit everywhere to avoid the #if (http://msdn.microsoft.com/en-us/library/6wdz5232.aspx)