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: [PATCH] Enable building glibc with gold.


On 01/09/2013 06:03 PM, Roland McGrath wrote:
>> On 01/09/2013 02:13 PM, Roland McGrath wrote:
>>> Meant to say as well, I don't think we should commit either my branch or
>>> another variant like your change until we have completely answered all the
>>> questions about regressions introduced by using gold.
>>
>> Why?
> 
> Because until that is done we are not sure that building with gold actually
> produces reliable usable results, obviously.
 
It is not a completely obvious conclusion and I didn't want to assume
what rationale you had for recommending against it. I respect your
opinion, but if you don't write it out I can't know what it is :-)

What I don't quite understand is why you are espousing what appears
to be a rather strict requirement for the addition of a new build
configuration. I don't know that it's healthy for the community to
adopt a "don't include it until it's perfect" policy.

Let us work by example...

Say the build matrix for glibc is a 2-dimensional thing with build
configurations on one axis and features on the other. As it stands
the matrix is not a sold sea of green-light functioning features.
The real matrix is a mix of greens, yellows, and reds depending
on the configuration e.g. compiler, binutils, flags, machine etc.

The addition of gold as a linker is a new row in this matrix, and
on x86_64, instead of the row being entirely green it has some
red blocks representing the testsuite failures.

The addition of this row in no way impacts the *sold* green rows
for both x86, and x86_64, when built with appropriate tools.
Regressions on those rows are very bad things and we will backout
code to remove.

At present it's cumbersome to test building glibc with gold, but
with either your patch, or mine (I'm not picky) *already* checked
in we could make it as easy a symlink flip (e.g. FC17 `alternatives')
and a rebuild with no need to switch branches / rebase or apply a
patch.

I think we should make it easier to build and test with gold.

Cheers,
Carlos.



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