[llvm-dev] [RFC] Cross-project lit test suite

[Fangrui Song via llvm-dev]

On 2021-01-26, Mehdi AMINI wrote:
>On Tue, Jan 26, 2021 at 2:15 PM Adrian Prantl via llvm-dev <
>llvm-dev at lists.llvm.org> wrote:
>
>>
>>
>> > On Jan 26, 2021, at 1:38 PM, Fangrui Song <maskray at google.com> wrote:
>> >
>> > On 2021-01-26, Adrian Prantl via llvm-dev wrote:
>> >>
>> >>
>> >>> On Jan 26, 2021, at 11:26 AM, David Blaikie <dblaikie at gmail.com>
>> wrote:
>> >>>
>> >>>
>> >>>
>> >>> On Tue, Jan 26, 2021 at 11:08 AM Adrian Prantl <aprantl at apple.com
>> <mailto:aprantl at apple.com>> wrote:
>> >>>
>> >>>
>> >>>> On Jan 26, 2021, at 10:38 AM, David Blaikie <dblaikie at gmail.com
>> <mailto:dblaikie at gmail.com>> wrote:
>> >>>>
>> >>>>
>> >>>>
>> >>>> On Tue, Jan 26, 2021 at 4:50 AM James Henderson <
>> jh7370.2008 at my.bristol.ac.uk <mailto:jh7370.2008 at my.bristol.ac.uk>> wrote:
>> >>>> On Tue, 26 Jan 2021 at 00:28, David Blaikie <dblaikie at gmail.com
>> <mailto:dblaikie at gmail.com>> wrote:
>> >>>>
>> >>>>
>> >>>> On Mon, Jan 25, 2021 at 4:20 PM Mehdi AMINI <joker.eph at gmail.com
>> <mailto:joker.eph at gmail.com>> wrote:
>> >>>>
>> >>>>
>> >>>> On Mon, Jan 25, 2021 at 12:16 PM David Blaikie via llvm-dev <
>> llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote:
>> >>>> On Mon, Jan 25, 2021 at 3:36 AM James Henderson via llvm-dev
>> >>>> <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote:
>> >>>> >
>> >>>> > Dear all,
>> >>>> >
>> >>>> >
>> >>>> > Recently, I and a number of my colleagues have run into cases where
>> we would like the ability to write tests that involve components from
>> multiple LLVM projects, for example using both clang and LLD. Similarly, I
>> have seen a few instances recently where tests would ideally make use of
>> LLD but only to help generate input objects for testing an LLVM tool, such
>> as llvm-symbolizer (see for example https://reviews.llvm.org/D88988 <
>> https://reviews.llvm.org/D88988>). Currently, there is no location where
>> lit tests that use both clang and LLD can be put, whilst the
>> llvm-symbolizer cases I’ve hit are testing llvm-symbolizer (and not LLD),
>> so don’t really fit in the LLD test suite. I therefore have prototyped a
>> lit test suite that would be part of the monorepo, and which can support
>> tests that use elements from multiple projects - see
>> https://reviews.llvm.org/D95339 <https://reviews.llvm.org/D95339>. Tests
>> could be added to this suite as needed. The suite is modelled as an
>> additional top-level directory, and is enabled by enabling the
>> “cross-project-tests” project in CMake. I have initially added support for
>> both LLD and clang. At configuration time, the tests that require LLD or
>> clang will only be enabled when the respective projects are enabled, so
>> that developers continue to benefit from the subset of tests that are
>> applicable for the projects they are building. Note that I am not
>> especially familiar with CMake or lit, so this code may not be perfect, but
>> it should be sufficient to demonstrate what it can do.
>> >>>> >
>> >>>> >
>> >>>> >
>> >>>> > One could argue that these sorts of tests should belong in the
>> external (to the monorepo) test-suite, but this is a) quite distant from
>> the existing testing, and therefore easily forgotten, delaying potential
>> feedback for any breakages/resulting in potentially duplicate testing etc,
>> and b) is not as easy to set up and run (owing to the fact that it isn’t
>> part of the monorepo, isn’t connected to check-all etc), therefore making
>> it harder for developers to maintain the tests. Back in October 2019, there
>> was an extensive discussion on end-to-end testing and how to write them
>> (starting from
>> https://lists.llvm.org/pipermail/cfe-dev/2019-October/063509.html <
>> https://lists.llvm.org/pipermail/cfe-dev/2019-October/063509.html>). The
>> suggestion was that these tests would be lit-based and run as part of
>> check-all, and would not be inside the clang tree, although there was some
>> opposition. This concluded with a round table. Unfortunately, I am unaware
>> of what the conclusion of that round table conversation was, so it’s
>> possible that what I am proposing is redundant/being worked on by someone
>> else. Additionally, I don’t consider all classes of tests that the proposed
>> lit suite would be useful for to be “end-to-end” testing. For example,
>> llvm-symbolizer is usually used on linked output containing debug
>> information. Usually tests that consume objects that have debug data in
>> them rely on assembly that has been written by hand or generated by clang
>> prior to commit, with a limited set able to make use of yaml2obj to
>> generate the debug data instead. However, the output of these approaches is
>> typically not a fully linked output (yaml2obj output can be made to look
>> like one, but getting all the addresses to match up in a maintainable
>> manner makes this approach not particularly viable). Being able to use LLD
>> to link the object file produced would make the test significantly more
>> readable, much as using llvm-mc and assembly to generate test inputs is
>> more preferable to using prebuilt binaries. Such a test is ultimately not
>> really any more an end-to-end test than an llvm-symbolizer test that just
>> uses the object produced by the assembler directly.
>> >>>> >
>> >>>> >
>> >>>> >
>> >>>> > What do people think?
>> >>>>
>> >>>> Some concerns (the usual: Things should be tested in isolation, things
>> >>>> should be tested independently - but end to end tests have some value
>> >>>> too), but generally seems good.
>> >>>>
>> >>>> Indeed this is a usual concern: such tests shouldn't be seen as
>> replacing isolated lit tests ("unit tests").
>> >>>>
>> >>>> I completely agree. Indeed, the llvm-symbolizer test referenced in
>> the review I'd class as an isolated test - LLD was just being used to
>> generate the input for the test. The key here is that the thing being
>> tested was the llvm-symbolizer behaviour, and not the linker behaviour. As
>> mentioned, this isn't really different to how llvm-mc or llc might be used
>> to convert some input source (asm/IR etc) into something the tool under
>> test wants to be run on. Potentially, changes in those tools might break
>> things, but as long as the input is specific enough, this shouldn't happen
>> often.
>> >>>>
>> >>>> But I have another question about the cost of maintenance here: are
>> we gonna revert patches to either project when one of the integration tests
>> fails?
>> >>>>
>> >>>> Possibly, yeah. If they demonstrate a bug.
>> >>>>
>> >>>> That would be my intention - these tests should be classed as
>> first-class citizens as much as any other lit test. They're just unit tests
>> that use other components (LLD, clang etc) to generate inputs.
>> >>>>
>> >>>>
>> >>>> What about integration tests that require to be updated manually when
>> changing another component?
>> >>>>
>> >>>> If they need to be updated, because their failure isn't
>> representative of a bug, yes.
>> >>>>
>> >>>> Hopefully these sorts of failures would occur only infrequently. As
>> noted, my aim here isn't to provide a place in opensource LLVM to do
>> integration testing, but rather unit tests that just can't sit in the
>> corresponding lit area, so input variability should be minimal.
>> >>>>
>> >>>> That being said, the proposed suite could be used for integration
>> testing, if the community agreed such testing belonged in the monorepo -
>> indeed, I have plans for some downstream integration tests that would make
>> use of this if it lands - but that isn't the goal here.
>> >>>>
>> >>>> I find the cost of maintenance of end-to-end tests is often hard to
>> carry over, especially since they are only supplementing and not replacing
>> lit "unit-tests".
>> >>>>
>> >>>> One of the nice thing about end to end tests (as with all tests, if
>> designed carefully - eg: don't take some arbitrary code, compile it with
>> optimizations, and expect a very specific backtrace - optimizations might
>> lead to different line table/stack frame details (if some code was merged,
>> or moved, it might lose or gain a specific source file/line)) is that they
>> can be pretty resilient to implementation details, so less likely to need
>> updating due to changes in implementation details. If someone changes the
>> output format of llvm-symbolizer these would require updating and I think
>> it'd be reasonable to expect that to be updated/not left failing.
>> >>>>
>> >>>> Right, just as we would change other tests for tools where the output
>> format changed.
>> >>>>
>> >>>>
>> >>>> - Dave
>> >>>>
>> >>>>
>> >>>> Best,
>> >>>>
>> >>>> --
>> >>>> Mehdi
>> >>>>
>> >>>>
>> >>>>
>> >>>> Though perhaps debuginfo-tests (this presumably already supports the
>> >>>> multiple-subproject mechanical isssue you're discussing?) could be
>> >>>> generalized/renamed to be all our cross-project lit testing
>> >>>> (Essentially an in-monorepo, lit-only, high-reliability/not-flakey/etc
>> >>>> version of the test-suite).
>> >>>>
>> >>>>
>> >>>> The existing debug-info test area looks like it could work if it was
>> more generalised. It looks like we'd want the ability to have tests that
>> would still work without clang/lldb being built, and similarly which could
>> use LLD, but I don't think those are insurmountable issues - it would just
>> require taking some of the ideas from my existing prototype (or equivalent
>> from elsewhere) and merging them in.
>> >>>>
>> >>>> Sounds good to me - might want some buy-in from other debuginfo-tests
>> folks, though.
>> >>>
>> >>> I don't mind generalizing debuginfo-tests to support more use-cases,
>> but I have some very practical concerns: Because of the hardware
>> constraints on green-dragon (currently a single Intel Mac Mini running all
>> LLDB tests) I would like to avoid getting into a situation where we need to
>> build more than clang+lldb in order to run the debuginfo-tests. I also
>> don't want to have to run more than the debuginfo tests on that bot. So as
>> long as a separate "check-..." target is used and the bots keep working,
>> I'm fine with this.
>> >>>
>> >>> Build design question: How would you feel if it was the same check-
>> target, but it dynamically chose what to run based on which projects were
>> checked out/being built? I'm not sure if that's a good design or not, but
>> having a combinatorial set of explicit check-* names based on which
>> projects are enabled would seem unfortunaet.
>> >>
>> >> Generally I'm not a fan of automatic behavior: You could have a
>> misconfiguration that leaves out a target and the bot would still run the
>> same target and it would aways be green, but it wouldn't run any tests. But
>> I'm not going to stand in the way if such a set up makes most sense here.
>> >>
>> >>> (though, inversely - if some set of tests (like the existing debug
>> info ones) are in a subdirectory, and like the current
>> check-llvm-debuginfo-* etc where you can specify a target that's a
>> subdirectory, would mean you coudl run just the debuginfo ones - that
>> wouldn't be so bad, but maybe we'd eventually have tests under there that
>> might require lld for instance - and it'd still be nice for those tests to
>> degrade gracefully to "unsupported" if you weren't building lld (similarly,
>> if you aren't building clang, the existing debuginfo tests could degrade to
>> "unsupported"))
>> >>
>> >> That actually sounds nice. The unsupported category would show up in
>> the test result xml file and thus any change would show in the Jenkins
>> statistics.
>> >>
>> >> -- adrian
>> >
>> > I just tested a build. After `ninja lldb` (~4000 actions), `ninja lld`
>> requires just ~152 actions. `ninja llvm-symbolizer` requires 7 actions.
>> > Perhaps the additional actions are fine? (Linking lld may need some
>> resources, as it links in IR/CodeGen libraries for LLVM LTO.)
>> >
>> > There are currently some magic `IN_LIST` in CMakeLists.txt: compiler-rt,
>> lldb, mlir.
>> > compiler-rt & mlir are for one or two tiny targets.
>>
>> I guess this is fine — that particular bot is already building the lld
>> project.
>>

(My previous reply is about llvm-symbolizer. I think that is fine -
and thanks to Adrian for allowing it.

* lld/ELF and lld/COFF are very stable now.
* If we write specific llvm-symbolizer tests, the risk of lld changes causing
   trouble is low -- as long as we don't hard code the addresses.
* Symbolization code is amenable to section/segment address changes.

llvm-symbolizer on object files have comprehensive tests, but on
linkaged images the coverage is probably loose.)

>In this particular case maybe, but this is a valid general concern for the
>approach: this RFC is about cross-project tests in general and not only
>this specific case.
>
>In general the ability to not build and run the world to change a single
>component is quite valuable during development. Historically the testing of
>these components is very decoupled and it seems important to me to keep
>this property as much as possible.
>
>Here I can see how this does seem harmless when presented as "LLD was just
>being used to generate the input for the test. The key here is that the
>thing being tested was the llvm-symbolizer behaviour, and not the linker
>behaviour", but it is very easy to introduce coupling between these
>components, and increase the maintenance cost.
>
>I'd be nice to have more clear guidelines about what is / isn't OK there,
>and it isn't obvious to me how to draw the line or identify when is it
>suitable to introduce cross-components dependencies for the sake of testing.

I have a similar concern about whether to draw the line for such
cross-project testing. To make concrete examples:

* For lld/ELF changes, I generally just test check-lld-elf. The test
   passing gives me large confidence that my change is good. I commit
   most of my changes this way. When I know my change my affect other
   binary formats supported by lld, I may choose check-lld.
   Actually lldb/test/Shell has some `REQUIRES: lld` tests. Rarely I need to notice them.
   In rarer cases I need to test a stage-2 build with -fuse-ld=lld.
* For many clang changes, check-clang is sufficient. When changing
   public APIs, technically clang-tools-extra/lldb/flang/polly may all be
   affected. It seems that as of now the developer does not have to worry
   about lldb/flang/polly too much. The breakage is rare.
   (ce5e21868c22479df62ebd8884adc1bd7c964433 is an example I fix lldb
   after a removed API from clang)
* For many llvm changes, check-llvm is sufficient. Public API changes
   sometimes require clang changes (from my experience), but very rare.
   (I think) Developers can largely ignore polly if they don't use it.
   (And it looks that Polly has fewer eyes on it, so a breakage may last
   longer. Persoanlly I've only fixed that three times for others' LLVM
   commits).
* If the new top-level project requires tight tanglement with lld/ELF output,
   I'd surely be concerned. (For lld, given the positive experience with
   low maintenance cost on lldb/test/Shell tests, I think such tanglement
   is not too likely, but I can imagine some potential friction for the
   less stable ports of lld).
   But requiring `check-Y` in the workfolow of contributors who
   mostly only develop/care about X, I can see that it can lower productivity.