LIT-ify the test-suite

Sun Oct 25 06:33:32 PDT 2015

Hi all,

It's about a year later and I've finally found time to work on this and get
it up to scratch.

Because there's been a mailing list change in the meantime, the original
proposal and conversation was here:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20141020/241422.html

The major objection from the version I posted in March was the lack of
support for restarting failed builds where they left off. I agree now that
is a showstopper. So I've changed tack slightly, and instead of making
everything run through LIT, now LIT only runs the tests. The tests are now
built using CMake+Make/Ninja. So now the flow would be:

$ cd build
$ cmake -DCMAKE_C_COMPILER=/path/to/clang ..
$ make -j12
$ llvm-lit -sv .

timeit is still used for timing, and compile times + execution times appear
as metrics in the overall LIT report.

I think this is a useful change, as it reduces by far the amount of custom
code required. Now we just need some small amount of CMake glue and a small
lit.cfg. Also, everyone (should) already know how to use both of these
tools, so there is no new technology here.

The patch at http://reviews.llvm.org/D14046 has all of the glue within it,
as well as a python script that can parse the current Makefiles in
SingleSource/MultiSource and produce equivalent CMakeLists.txt. The
intention would be that this script would get run once and the resulting
CMakeLists.txt committed to the repository.

With this patch applied and the script run, the entire test-suite can be
built and tested successfully.

Features that are not implemented yet:
  * Cross-compilation
  * The --use-perf option (using timeit.sh instead of timeit.exe).
  * I can't think of any others, but I'm sure there are some.

What do people think? is this a good direction to go down? What features
are people using in their buildbots that the new system would miss?

Cheers,

James

---------- Forwarded message ---------
From: James Molloy <james at jamesmolloy.co.uk>
Date: Tue, 31 Mar 2015 at 16:21
Subject: Re: LIT-ify the test-suite
To: Daniel Dunbar <daniel at zuster.org>
Cc: LLVM Commits <llvm-commits at cs.uiuc.edu>

Hi Chris, Daniel,

Let's get this going again :)

I've taken on board your points and have put a work-in-progress patch up at
http://reviews.llvm.org/D8721 . The autogenerated lit.local.cfg's still
aren't perfect, and there are still some missing (legal issues have not
subsided I'm afraid!) but we're getting there.

Regarding Makefiles - after discussion at the LLVM conference in November
I'm still not really sure what the Makefiles provide over LIT in terms of
incremental development. They allow one particular source file to fail, and
then to remake and everything picks up at that source file, but I really
don't think that adds much value on top of LIT, which can do one test at a
time.

Granted, the multisource tests consist of multiple files so this is where
Makefiles win, but is that really a big enough pull? can we not just deal
with the test granularity?

Cheers,

James

On Mon, 27 Oct 2014 at 00:41 James Molloy <james at jamesmolloy.co.uk> wrote:

> Hi Daniel,
>
> Thanks for the comments!
>
> 1. I had trouble getting modules loadable from lit.local.cfgs. I'll try
> again, because I agree separating it out is a good thing.
>
> 2. I didn't do this because I had trouble getting the common module
> loaded. I also had trouble using a different format in different
> subdirectories - I didn't know if this was supposed to work. I can go back
> to this.
>
> 3. Sure.
>
> 4. The rationale for this was the lack of a need to compile a helper tool.
> It's also a fairly small python function. I didn't see any particular
> slowdown, I just felt it was a nicer design unless there is a performance
> reason otherwise.
>
> 5. Hmm. I think scraping Makefiles is pretty nasty, I'd tend more towards
> something like an equivalent of LLVMBuild.txt, which could be used to
> generate both lit.local.cfgs and Makefiles. What do you think about that?
>
> ----
>
> 1/2. My intention on sending my original email was to try and work out
> what people use the Makefiles for. They have a *lot* of features, and I
> really didn't know which of them people were using. It's been fairly clear
> from the responses that people still require incremental development, which
> really needs a dependency-based system like Make. I don't think putting all
> that complexity in LIT is a good idea at all, so at the moment I'm erring
> towards keeping both systems, for the different use cases. That way
> complexity can continue to be piled into the Makefiles without ruining the
> "Just Run It!" use case.
>
> What do you think?
>
> Cheers,
>
> James
>
> On 26 October 2014 18:39, Daniel Dunbar <daniel at zuster.org> wrote:
>
>> Hi James,
>>
>> Thanks for working on this!
>>
>> Comments on the specific patch:
>>
>> --
>>
>> 1. For tidyness, I think it would be good to move the test format
>> implementation into its own module (and just change lit.cfg to update
>> sys.path relative to __file__) instead of having it in the lit.cfg.
>>
>> 2. Instead of having the diamond inheritance in the test format, I would
>> just define two different test formats the SingleSrc format and the
>> MultiSrc format, and in the lit.local.cfg file for the top-level
>> directories, set the appropriate format. They can of course inherit from a
>> common base in the same module (see #1).
>>
>> 3. Re "FIXME: Push this into lit.formats" I disagree and think it should
>> just be local to the test-suite, it isn't a configuration any other project
>> is likely to use. There should be no harm in having it be independent
>> though, this is a use case lit should support.
>>
>> 4. I'm skeptical about moving fpcmp into the Python implementation -- why
>> not just use a subprocess here? The fpcmp implementation has had bugs over
>> the years, and actually its runtime performance does have an impact on the
>> test suite time (IIRC).
>>
>> 5. As for the lit.local.cfg issue -- one way to deal with this is to move
>> all of the per-test configuration information (hash outputs, command line
>> flags, etc.) into a well defined format. This could either be a JSON file
>> in the local directory, or it could be we just say the Makefile needs to
>> follow a very strict format for now, and have the Lit test format scrape
>> the information out of the local Makefile (and barf if it can't parse it).
>>
>> 6. Things like SMALL/LARGE problem size should probably always be pushed
>> into the C/C++ flags, not something individual lit.local.cfg files have to
>> check. Similarly, maybe "-I%S" should also be a default every test gets.
>>
>> --
>>
>> As for the general approach, I have also worked on this a couple times
>> over the years, and the blocker that has prevented me from ever checking
>> something in is:
>>
>> 1. How to manage the incremental development part of things? This
>> approach is great for automation and approachability, but for people who
>> are very familiar with the way the Makefiles work there are a lot of useful
>> workflows that aren't supported anymore (e.g., the ability to manually
>> rebuild different parts, the different ways to tweak the config, etc.).
>>
>> 2. Where does this leave the Makefile based config?
>>
>> Any thoughts on those?
>>
>>  - Daniel
>>
>> On Fri, Oct 24, 2014 at 7:20 AM, James Molloy <james at jamesmolloy.co.uk>
>> wrote:
>>
>>> Hi all,
>>>
>>> We've been trying to improve the LNT test driver to be a bit more clever
>>> with reruns, and have run into the problem that it is very tightly coupled
>>> to the Makefile system in the test-suite.
>>>
>>> The Makefile system was obviously designed some time ago before Clang
>>> was production ready, and it has a bunch of features to support running
>>> through LLC and suchlike that add serious complexity. It's nontrivial to
>>> just run a test and get a result!
>>>
>>> So I got to thinking, why don't we just use LIT, the really well written
>>> test driver that our regression tests use? That would allow us to simplify
>>> the interface, reduce the number of test drivers to 1, have pretty progress
>>> bars, export the results in JSON, shuffle the test order...
>>>
>>> So I attach an alpha version of a lit.cfg for the test-suite. Some
>>> caveats with this that make it not quite production-ready:
>>>   * It doesn't use perf for timing yet, it just uses python's time()
>>> interface.
>>>   * It doesn't yet report compile time, but that's actually a one line
>>> change.
>>>   * The biggest problem: A bunch of the test suites require
>>> lit.local.cfg files in them to tell lit how to build and run them. However,
>>> many of the test suite directories are under blanket non-LLVM licenses
>>> which makes for problems with our legal permissions.
>>>
>>> So attached are two patches: 0001 adds the lit.cfg - just apply it to
>>> the test-suite directory, and 0002 adds a selection of lit.local.cfgs - for
>>> all the subfolders that aren't covered by blanket licenses.
>>>
>>> Because it excludes some lit.local.cfgs, some tests won't run. However,
>>> all the unit tests do, so that's an easy proof of concept:
>>>
>>> $ cd test-suite
>>> $ CC=clang llvm-lit -sv ./MultiSource/UnitTests
>>>
>>> Please give it a try and let us know your thoughts, so I know how hard I
>>> should push this.
>>>
>>> Cheers,
>>>
>>> James
>>>
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20151025/caa5a544/attachment.html>