Re: [lfs-dev] gold

[Ken Moffat]

On Tue, Feb 07, 2017 at 10:14:34PM -0600, DJ Lucas wrote:
> 
> 
> I could have sworn the gold testsuite was a mess, but I just ran it and all
> 243 tests pass. Commit coming shortly.
> 
Well it certainly looks messy compared to the other parts of
binutils, and for me every result (when it finally stops compiling)
is within separate 'Entering directory' and 'Leaving directory'
messages, all for binutils-2.27/build/gold/testsuite.

I had one failure (relro_test.sh) but I didn't save the
testsuite.log from gold.

ĸen
-- 
`I shall take my mountains', said Lu-Tze. `The climate will be good
for them.' -- Small Gods
-- 
http://lists.linuxfromscratch.org/listinfo/lfs-dev
FAQ: http://www.linuxfromscratch.org/faq/
Unsubscribe: See the above information page

Re: [lfs-dev] gold

[DJ Lucas]

On 02/07/2017 09:35 PM, Ken Moffat wrote:

On Mon, Feb 06, 2017 at 07:28:24PM -0600, DJ Lucas wrote:



On 02/04/2017 10:58 AM, Ken Moffat wrote:

On Sat, Feb 04, 2017 at 12:07:55AM -0600, DJ Lucas wrote:


Interesting, I had just suggested to Bruce (the other day on IRC) that we go 
ahead and do it. I've seen absolutely no failures with bfd as the default 
linker with both big DEs installed. I haven't gotten around to newer QT yet. I 
was hoping we'd see a 2.28 release well before now. Does QT pick up gold by 
default?


It seems not - I have both ld.gold, ld.bfd and ldd (bfd) in /usr/bin,
with a /usr/bin/gold directory containing an ld symlink.  I normally
put that first on the PATH, but using a normal PATH it used ld.bfd.



I'm still of the mind to provide everything that the package can provide
with two caveats. 1. We aren't increasing dependencies drastically. 2. No
harm comes from the change. This even at the expense of a couple of SBU --
0.1 for bison in Ch5 and ~3 for the gold test suite IIRC -- and regardless
of any perceived benefit. I might be in the minority here, so opinions are
welcome (not sure if Ken's are fully expressed above).

On the flip, nothing prevents us from burring the two additional switches in
the wiki if there are strong feelings the other way.

i[...]


As to usability, in the default configuration, you can use -fuse-gold to
CFLAGS/CXXFLAGS if you want to use it explicitly, but otherwise, as far as I
have been able to tell, it is ignored unless the package maintainer
explicitly checks for it (and presumably, the maintainer knows whether their
dependent libraries are correct or if they still depend on bfd's DL-like
workarounds).

With 2.27, however, the test suite is, um...messy, with a completely
different testing framework. 2.28 will hopefully be out the door before
package freeze, but I'd like to wait and see what things look like at its
release.



I initially delayed responding because I was hoping to run some
tests - got as far as running a 4.10-rc7 kernel on my Kaveri with a
couple of 'Compute Unit' patches which were on lkml and probably
heading for 4.11 - mixed results (kernel compile with -j4
considerably less-slow, but rerunning pass 1 of binutils needs more
tests : from time to time, that takes a *lot* longer on a
fully-complete system, this may have been one of htose time).  But
I'm about to step into a maze of perl modules (latest versions of all
the things used by the testsuites of the modules needed by biber) and
I'm not sure if I'll have time.  Plus other local requirements.

So - now that I have calmed down, no objections to putting gold into
the book, particularly if somewhere we explain how to force it (I
had not bothered to google deeply enough to find -fuse-gold).

You mention the wiki - isn't that only for BLFS ?  Documenting this
might be most useful for BLFS users, but the logical place to do
that is surely in LFS.

ĸen, hoping that 2.28 arrives in time for 8.0 and is less messy in
its test results - but because I hide most static libs so that I know
what uses them, I suspect it will take me ages to get a good version
of my scripts.



I could have sworn the gold testsuite was a mess, but I just ran it and 
all 243 tests pass. Commit coming shortly.


--DJ

--
http://lists.linuxfromscratch.org/listinfo/lfs-dev
FAQ: http://www.linuxfromscratch.org/faq/
Unsubscribe: See the above information page

Re: [lfs-dev] gold

[Ken Moffat]

On Mon, Feb 06, 2017 at 07:28:24PM -0600, DJ Lucas wrote:
> 
> 
> On 02/04/2017 10:58 AM, Ken Moffat wrote:
> > On Sat, Feb 04, 2017 at 12:07:55AM -0600, DJ Lucas wrote:
> > > 
> > > Interesting, I had just suggested to Bruce (the other day on IRC) that we 
> > > go ahead and do it. I've seen absolutely no failures with bfd as the 
> > > default linker with both big DEs installed. I haven't gotten around to 
> > > newer QT yet. I was hoping we'd see a 2.28 release well before now. Does 
> > > QT pick up gold by default?
> > > 
> > It seems not - I have both ld.gold, ld.bfd and ldd (bfd) in /usr/bin,
> > with a /usr/bin/gold directory containing an ld symlink.  I normally
> > put that first on the PATH, but using a normal PATH it used ld.bfd.
> > 
> 
> I'm still of the mind to provide everything that the package can provide
> with two caveats. 1. We aren't increasing dependencies drastically. 2. No
> harm comes from the change. This even at the expense of a couple of SBU --
> 0.1 for bison in Ch5 and ~3 for the gold test suite IIRC -- and regardless
> of any perceived benefit. I might be in the minority here, so opinions are
> welcome (not sure if Ken's are fully expressed above).
> 
> On the flip, nothing prevents us from burring the two additional switches in
> the wiki if there are strong feelings the other way.
i[...]
> 
> As to usability, in the default configuration, you can use -fuse-gold to
> CFLAGS/CXXFLAGS if you want to use it explicitly, but otherwise, as far as I
> have been able to tell, it is ignored unless the package maintainer
> explicitly checks for it (and presumably, the maintainer knows whether their
> dependent libraries are correct or if they still depend on bfd's DL-like
> workarounds).
> 
> With 2.27, however, the test suite is, um...messy, with a completely
> different testing framework. 2.28 will hopefully be out the door before
> package freeze, but I'd like to wait and see what things look like at its
> release.
> 

I initially delayed responding because I was hoping to run some
tests - got as far as running a 4.10-rc7 kernel on my Kaveri with a
couple of 'Compute Unit' patches which were on lkml and probably
heading for 4.11 - mixed results (kernel compile with -j4
considerably less-slow, but rerunning pass 1 of binutils needs more
tests : from time to time, that takes a *lot* longer on a
fully-complete system, this may have been one of htose time).  But
I'm about to step into a maze of perl modules (latest versions of all
the things used by the testsuites of the modules needed by biber) and
I'm not sure if I'll have time.  Plus other local requirements.

So - now that I have calmed down, no objections to putting gold into
the book, particularly if somewhere we explain how to force it (I
had not bothered to google deeply enough to find -fuse-gold).

You mention the wiki - isn't that only for BLFS ?  Documenting this
might be most useful for BLFS users, but the logical place to do
that is surely in LFS.

ĸen, hoping that 2.28 arrives in time for 8.0 and is less messy in
its test results - but because I hide most static libs so that I know
what uses them, I suspect it will take me ages to get a good version
of my scripts.
-- 
`I shall take my mountains', said Lu-Tze. `The climate will be good
for them.' -- Small Gods
-- 
http://lists.linuxfromscratch.org/listinfo/lfs-dev
FAQ: http://www.linuxfromscratch.org/faq/
Unsubscribe: See the above information page

Re: [lfs-dev] gold

[DJ Lucas]

On 02/04/2017 10:58 AM, Ken Moffat wrote:

On Sat, Feb 04, 2017 at 12:07:55AM -0600, DJ Lucas wrote:


Interesting, I had just suggested to Bruce (the other day on IRC) that we go 
ahead and do it. I've seen absolutely no failures with bfd as the default 
linker with both big DEs installed. I haven't gotten around to newer QT yet. I 
was hoping we'd see a 2.28 release well before now. Does QT pick up gold by 
default?


It seems not - I have both ld.gold, ld.bfd and ldd (bfd) in /usr/bin,
with a /usr/bin/gold directory containing an ld symlink.  I normally
put that first on the PATH, but using a normal PATH it used ld.bfd.



I'm still of the mind to provide everything that the package can provide 
with two caveats. 1. We aren't increasing dependencies drastically. 2. 
No harm comes from the change. This even at the expense of a couple of 
SBU -- 0.1 for bison in Ch5 and ~3 for the gold test suite IIRC -- and 
regardless of any perceived benefit. I might be in the minority here, so 
opinions are welcome (not sure if Ken's are fully expressed above).


On the flip, nothing prevents us from burring the two additional 
switches in the wiki if there are strong feelings the other way. While I 
think a *drastic* speed/size benefit has been sufficiently debunked 
(even if only with anecdotal evidence), that doesn't render it useless. 
Samba was a good example where the maintainers were tracking down issues 
in their code that bfd managed to gloss over.


As to usability, in the default configuration, you can use -fuse-gold to 
CFLAGS/CXXFLAGS if you want to use it explicitly, but otherwise, as far 
as I have been able to tell, it is ignored unless the package maintainer 
explicitly checks for it (and presumably, the maintainer knows whether 
their dependent libraries are correct or if they still depend on bfd's 
DL-like workarounds).


With 2.27, however, the test suite is, um...messy, with a completely 
different testing framework. 2.28 will hopefully be out the door before 
package freeze, but I'd like to wait and see what things look like at 
its release.


Some interesting linker optimizations coming down the pipe as well, this 
one https://sourceware.org/ml/binutils/2017-02/msg00032.html for 
instance has piqued my interest. That is a 3.97% reduction in size. Now 
granted, that's a really big executable, and not anywhere near typical.


--DJ

--
http://lists.linuxfromscratch.org/listinfo/lfs-dev
FAQ: http://www.linuxfromscratch.org/faq/
Unsubscribe: See the above information page

Re: [lfs-dev] gold

[Ken Moffat]

On Sat, Feb 04, 2017 at 04:58:38PM +, Ken Moffat wrote:
> On Sat, Feb 04, 2017 at 12:07:55AM -0600, DJ Lucas wrote:
> > 
> > Interesting, I had just suggested to Bruce (the other day on IRC) that we 
> > go ahead and do it. I've seen absolutely no failures with bfd as the 
> > default linker with both big DEs installed. I haven't gotten around to 
> > newer QT yet. I was hoping we'd see a 2.28 release well before now. Does QT 
> > pick up gold by default?
> > 
> It seems not - I have both ld.gold, ld.bfd and ldd (bfd) in /usr/bin,
> with a /usr/bin/gold directory containing an ld symlink.  I normally
> put that first on the PATH, but using a normal PATH it used ld.bfd.
> 

Hmm, I just took a look at my scripts to see where I mention gold,
and I discovered that I added -no-use-gold-linker to qt5 a while
ago, so presumably it has not worked for a while.  That may well be
in qtwebengine (I only really build qt for qupzilla on my normal
builds).

And that switch apparently no longer works.  Strange.

ĸen
-- 
`I shall take my mountains', said Lu-Tze. `The climate will be good
for them.' -- Small Gods
-- 
http://lists.linuxfromscratch.org/listinfo/lfs-dev
FAQ: http://www.linuxfromscratch.org/faq/
Unsubscribe: See the above information page

Re: [lfs-dev] gold

[Ken Moffat]

On Sat, Feb 04, 2017 at 12:07:55AM -0600, DJ Lucas wrote:
> 
> Interesting, I had just suggested to Bruce (the other day on IRC) that we go 
> ahead and do it. I've seen absolutely no failures with bfd as the default 
> linker with both big DEs installed. I haven't gotten around to newer QT yet. 
> I was hoping we'd see a 2.28 release well before now. Does QT pick up gold by 
> default?
> 
It seems not - I have both ld.gold, ld.bfd and ldd (bfd) in /usr/bin,
with a /usr/bin/gold directory containing an ld symlink.  I normally
put that first on the PATH, but using a normal PATH it used ld.bfd.

ĸen
-- 
`I shall take my mountains', said Lu-Tze. `The climate will be good
for them.' -- Small Gods
-- 
http://lists.linuxfromscratch.org/listinfo/lfs-dev
FAQ: http://www.linuxfromscratch.org/faq/
Unsubscribe: See the above information page

Re: [lfs-dev] gold

[DJ Lucas]

On February 3, 2017 9:51:29 PM CST, Ken Moffat  wrote:
>I'm increasingly starting to think that I'm going to stop building
>the gold linker (it will save me a little time in /tools by not
>building bison, and rather more time in the binutils tests in
>chroot).
>
>This annoyance is prompted because qt-5.8.0 (or perhaps qtwebengine)
>provokes a failure in gold.
>
>From previous posts, ISTR that DJ was unconvinced that gold actually
>improved link times in his tests - I've been hoping to retry a few
>things for myself, but so far I haven't had time and for the moment
>there are too many other problems to look at.
>
>This is really dispiriting - it was supposed to be a faster way of
>linking, but the reality doesn't seem to match that (too much breaks
>from time to time).

Interesting, I had just suggested to Bruce (the other day on IRC) that we go 
ahead and do it. I've seen absolutely no failures with bfd as the default 
linker with both big DEs installed. I haven't gotten around to newer QT yet. I 
was hoping we'd see a 2.28 release well before now. Does QT pick up gold by 
default?

--DJ

-- 
Sent from my Android device with K-9 Mail. Please excuse my brevity.

-- 
http://lists.linuxfromscratch.org/listinfo/lfs-dev
FAQ: http://www.linuxfromscratch.org/faq/
Unsubscribe: See the above information page