llvm dev - Sep 2016 - [RFC] Require PRs for XFAILing tests

> On Sep 29, 2016, at 7:52 AM, David Blaikie <dblaikie at gmail.com>
wrote:
> 
> 
> 
> On Wed, Sep 28, 2016 at 11:58 AM Robinson, Paul via llvm-dev <llvm-dev
at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote:
> On 28 September 2016 at 10:08, Chris Bieneman via llvm-dev
> <llvm-dev at lists.llvm.org <mailto:llvm-dev at
lists.llvm.org>> wrote:
> > I cannot think of any situation where a universally failing test
> > should be in-tree unless it is a bug that someone is expecting to fix.
> 
> It seems moderately common to mark something XFAIL temporarily to get
> the bots green while then going ahead to fix the issue.  Your proposal
> would add extra overhead to that flow by requiring a PR as well.  This
> has value when it turns out that fix can't happen in the short term for
> any reason.  I don't have a feel for how common that is, although
I'm
> sure it does happen.
> I think the overhead is worth the added value, but then I'm a process
> kind of guy.
> 
> 
> On 28 September 2016 at 10:28, Renato Golin via llvm-dev
> <llvm-dev at lists.llvm.org <mailto:llvm-dev at
lists.llvm.org>> wrote:
> > We already have an unwritten rule to create PRs for XFAILs, and we
> > normally don't XFAIL lightly (I don't, at least). But creating
one PR
> > for every existing XFAIL may end up as a long list of never looked
> > PRs. :)
> 
> As opposed to the other ~9000 open PRs?  At least they would be tracked.
> 
> I'd be inclined to agree (or at least voice the same concern) as Renato
here - as a project we don't really have very good bug hygiene, so adding
more bug filing doesn't seem to me like it'd give us much value.
I haven’t done a full audit, but we have 257 XFAILs in LLVM.

* 44 of those are vg_leak failures on TableGen tests which should be UNSUPPORTED
because we allow TableGen to leak for performance the same way we allow clang to
leak.
* 15 of them are vg_leak in the OCaml bindings. Someone familiar with OCaml
should chime in on it, but I suspect those too should be UNSUPPORTED
* 125 of them are universal failure (XFAIL: *). Many of these have been marked
this way for years. I suspect that if we take the time to go through these we
will likely find that many of these test cases either should be tracked by bugs,
or should be removed from the tree

From there, many of the test cases are XFAIL on features where they really
should be UNSUPPORTED. I suspect that if we do a full audit of the XFAILs we
would likely find <100 which should actually be XFAIL, and tracking those
seems valuable to me.
> 
> Auditing existing XFAILs can be done today without filing bugs for them.
Yes, we can audit them today. Making bugs required for them makes it easier to
audit them because there will (in theory) be a bug describing the justification
for the XFAIL and what the underlying issue is. Digging up the reason why an
XFAIL was added in 2009 is a little bit challenging today if there isn’t a PR
associated with it (or a really good comment or commit message).

-Chris
> 
> And we still always have the option to (& in many cases do) file bugs
for XFAILs to discuss them, etc.
> 
> But I don't feel strongly about it either way, so happy to leave the
folks who do to make the call/do the work.
> 
> - Dave
>  
> --paulr
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
<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/20160929/653ac085/attachment.html>

On Thu, Sep 29, 2016 at 10:29 AM, Chris Bieneman via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> I haven’t done a full audit, but we have 257 XFAILs in LLVM.
>
> * 44 of those are vg_leak failures on TableGen tests which should be
> UNSUPPORTED because we allow TableGen to leak for performance the same way
> we allow clang to leak.
> * 15 of them are vg_leak in the OCaml bindings. Someone familiar with
> OCaml should chime in on it, but I suspect those too should be UNSUPPORTED
>
Do we still support running the test suite with Valgrind? I don't think we
run it continuously anymore. Maybe these are just noise now and we should
remove them. Users can still run Valgrind on LLVM, but LLVM's test suite is
an implementation detail, and we don't have to gaurantee that it is leak
free.

> * 125 of them are universal failure (XFAIL: *). Many of these have been
> marked this way for years. I suspect that if we take the time to go through
> these we will likely find that many of these test cases either should be
> tracked by bugs, or should be removed from the tree
>
> From there, many of the test cases are XFAIL on features where they really
> should be UNSUPPORTED. I suspect that if we do a full audit of the XFAILs
> we would likely find <100 which should actually be XFAIL, and tracking
> those seems valuable to me.
>
Bummer.
> Auditing existing XFAILs can be done today without filing bugs for them.
>
>
> Yes, we can audit them today. Making bugs required for them makes it
> easier to audit them because there will (in theory) be a bug describing the
> justification for the XFAIL and what the underlying issue is. Digging up
> the reason why an XFAIL was added in 2009 is a little bit challenging today
> if there isn’t a PR associated with it (or a really good comment or commit
> message).
>
Can we require a comment for every XFAIL? That seems like a clear guideline
that can be enforced by code review, and it's less heavyweight.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160929/85f5c593/attachment.html>

> On Sep 29, 2016, at 10:38 AM, Reid Kleckner <rnk at google.com>
wrote:
> 
> On Thu, Sep 29, 2016 at 10:29 AM, Chris Bieneman via llvm-dev <llvm-dev
at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote:
> I haven’t done a full audit, but we have 257 XFAILs in LLVM.
> 
> * 44 of those are vg_leak failures on TableGen tests which should be
UNSUPPORTED because we allow TableGen to leak for performance the same way we
allow clang to leak.
> * 15 of them are vg_leak in the OCaml bindings. Someone familiar with OCaml
should chime in on it, but I suspect those too should be UNSUPPORTED
> 
> Do we still support running the test suite with Valgrind? I don't think
we run it continuously anymore. Maybe these are just noise now and we should
remove them. Users can still run Valgrind on LLVM, but LLVM's test suite is
an implementation detail, and we don't have to gaurantee that it is leak
free.
>  
> * 125 of them are universal failure (XFAIL: *). Many of these have been
marked this way for years. I suspect that if we take the time to go through
these we will likely find that many of these test cases either should be tracked
by bugs, or should be removed from the tree
> 
> From there, many of the test cases are XFAIL on features where they really
should be UNSUPPORTED. I suspect that if we do a full audit of the XFAILs we
would likely find <100 which should actually be XFAIL, and tracking those
seems valuable to me.
> 
> Bummer. 
>> Auditing existing XFAILs can be done today without filing bugs for
them.
> 
> Yes, we can audit them today. Making bugs required for them makes it easier
to audit them because there will (in theory) be a bug describing the
justification for the XFAIL and what the underlying issue is. Digging up the
reason why an XFAIL was added in 2009 is a little bit challenging today if there
isn’t a PR associated with it (or a really good comment or commit message).
> 
> Can we require a comment for every XFAIL? That seems like a clear guideline
that can be enforced by code review, and it's less heavyweight.
Requiring comments for XFAIL is much harder to enforce via tools, and would need
to be enforced with code review. One advantage I see to this process is that the
tooling support is a relatively simple patch to LIT.

-Chris
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160929/ca8f27e7/attachment.html>