This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: What is consensus?


On Thu, Apr 5, 2012 at 6:04 AM, Joseph S. Myers <joseph@codesourcery.com> wrote:
> On Thu, 5 Apr 2012, Carlos O'Donell wrote:
>
>> The more controversial point I make here is that we probably need to
>> take a hard stand on build/testsuite regressions and simply revert any
>> patches that break or regress the testsuite.
>
> The way the libm tests work it's unavoidable that a newly added test will
> typically fail on each architecture until the ulps baselines are updated,
> since they are architecture-specific. ?I don't think we should expect
> people to test on all architectures; we should state in such cases that
> ulps updates that do not add unduly large ulps entries (any ulps at all
> for functions that should be exact) are considered obvious. ?We should
> also be liberal about conditioning out tests for particular floating-point
> formats etc. if they show up bugs (always with a comment added pointing to
> the bug in Bugzilla), for those cases where they show up an
> architecture-specific problem that isn't being immediately fixed.

(a) libm.

I agree that libm is going to be problematic. Could you please add
some wording to Consensus regarding libm and feel free to include
vague wording about unduly large ulps?

(b) Testing on all architectures.

What might you propose for situations where a patch breaks the
testsuite for an architecture?

There are some architectures which are very importantly IMO, namely
x86 and x86_64 (and probably x32 soon). I think everyone should be
able to test on those, and any build breakage there or testsuite
regression should be grounds for immediate reverting to keep the tree
"green."

Comments? Should the wording be ammended to say "If you break x86 or
x86_64 be prepared for your commit to be reverted?"

Cheers,
Carlos.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]