[llvm-dev] RFC: A cross platform way of using shell commands in lit tests

Tue Aug 16 11:51:25 PDT 2016

No wonder your power supply blew up if you're doing crazy stuff like that!

Although it's more work, honestly I like the idea of running as much as we
can from within lit to avoid excess process creation.  We should probably
see how far that can get is and then take a look at the remaining
dependencies.

On Tue, Aug 16, 2016 at 11:41 AM Reid Kleckner <rnk at google.com> wrote:

> I selected the pants-on-head crazy button of "put all utilities on PATH",
> and it works from cmd too. We could probably teach lit to find these
> utilities by searching relative to 'git', which is presumably already on
> PATH.
>
> On Tue, Aug 16, 2016 at 11:39 AM, Zachary Turner <zturner at google.com>
> wrote:
>
>> That's only if you use git bash, not cmd.
>>
>> On Tue, Aug 16, 2016 at 11:38 AM Reid Kleckner <rnk at google.com> wrote:
>>
>>> One other thing to note is that the dependency problem is going to go
>>> away very soon just a result of moving to git. The normal Windows install
>>> of git comes with all the utilities you need to run the LLVM tests anyway.
>>> I haven't bothered to install gnuwin32 on my new machine.
>>>
>>> I think we do want to rewrite a few key utilities, like 'rm', to make
>>> them reliable on Windows, but for most things like 'grep' and 'sed', we
>>> should use the standard things. There's no reason to want to have our own
>>> subtly different versions of these tools.
>>>
>>> If we do go all the way to avoid process creation, then it becomes
>>> interesting to have our own version of these things, and at that point,
>>> maybe having a fork or checked-in copy of busybox is the way to go.
>>>
>>> On Tue, Aug 16, 2016 at 11:23 AM, Zachary Turner <zturner at google.com>
>>> wrote:
>>>
>>>> Unfortunately the proposal here doesn't get rid of process creation, it
>>>> just changes the process being created from a GnuWin32 one to one that we
>>>> write ourselves.  Happy to entertain suggestions for how to get rid of
>>>> process creation entirely, but it doesn't seem simple to me.
>>>>
>>>> On Tue, Aug 16, 2016 at 11:20 AM Yaron Keren <yaron.keren at gmail.com>
>>>> wrote:
>>>>
>>>>> You can use msys2 rather than gnuwin32.
>>>>> Nevertheless it would be great to get rid of this dependency, as well
>>>>> as speedup the regressions tests. Process creation is slower on Windows
>>>>> than Linux and so the tests run much slower.
>>>>>
>>>>>
>>>>> 2016-08-16 21:14 GMT+03:00 Zachary Turner via llvm-dev <
>>>>> llvm-dev at lists.llvm.org>:
>>>>>
>>>>>> That's more like a workaround though.  The real problem is that file
>>>>>> deletion is racy on Windows.  I can't count the number of times I've
>>>>>> encountered this error, only to find out that adding a retry fixed the
>>>>>> problem.  Sure, if you can remove opened files then there's no problem to
>>>>>> begin with, but even if you can't, you can still remove the file by
>>>>>> retrying again 100ms later or so.  It seems silly to disable a test which
>>>>>> could be testing useful functionality because of some quirk of an OS.
>>>>>>
>>>>>> If you had your own robust rm command, it can retry a few times,
>>>>>> allowing the test to run everywhere.
>>>>>>
>>>>>> On Tue, Aug 16, 2016 at 11:07 AM Reid Kleckner via llvm-dev <
>>>>>> llvm-dev at lists.llvm.org> wrote:
>>>>>>
>>>>>>> On Tue, Aug 16, 2016 at 10:09 AM, Greg Bedwell via llvm-dev <
>>>>>>> llvm-dev at lists.llvm.org> wrote:
>>>>>>>
>>>>>>>> Anecdotally, most of the times I've seen people add "REQUIRES:
>>>>>>>>> shell" on
>>>>>>>>> a test they really meant "disable on windows". I think/hope that's
>>>>>>>>> usually temporary while they investigate problems, but it might
>>>>>>>>> explain
>>>>>>>>> why so many tests say they require shell when they don't.
>>>>>>>>>
>>>>>>>>>
>>>>>>>> In at least one case I can think of
>>>>>>>> ("clang/test/Format/style-on-command-line.cpp") the "REQUIRES: shell" is
>>>>>>>> there to work around annoying intermittent failures on Windows where 'rm'
>>>>>>>> was failing with permission denied errors just occasionally enough to be a
>>>>>>>> problem.  I've definitely seen this happen in a number of other places
>>>>>>>> previously although I can't think whether this affects any other lit tests
>>>>>>>> off-hand, but I know that personally I've had to implement 'robust rm'
>>>>>>>> (basically, keep on trying until the error goes away on its own) on Windows
>>>>>>>> in a number of different systems to work around this sort of problem
>>>>>>>> before.  Replacing gnuwin32 with something where we could easily implement
>>>>>>>> something like 'robust <cmd>' versions for Windows when we encounter these
>>>>>>>> sorts of issues would be fantastic, so I'm definitely in favour of removing
>>>>>>>> the dependency.
>>>>>>>>
>>>>>>>
>>>>>>> You can use "REQUIRES: can-remove-opened-file" to express this
>>>>>>> requirement more precisely.
>>>>>>> _______________________________________________
>>>>>>> LLVM Developers mailing list
>>>>>>> llvm-dev at lists.llvm.org
>>>>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> LLVM Developers mailing list
>>>>>> llvm-dev at lists.llvm.org
>>>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>>>>
>>>>>>
>>>>>
>>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160816/bce475d0/attachment.html>