Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-26 Thread Steve M. Robbins
On Sat, Feb 26, 2011 at 05:49:49PM +0100, Matthias Klose wrote:
 On 26.02.2011 04:42, Steve M. Robbins wrote:
 On Fri, Feb 25, 2011 at 03:57:28PM +0100, Matthias Klose wrote:
 On 25.02.2011 08:46, Steve M. Robbins wrote:
 
 Clearly one should be mindful of the effect on GCC -- that's why I
 asked the question on debian-gcc.  Do you have any specific concerns?
 
 Have any concerns been raised on the GCC mailing list?  I've googled
 and found only anecdotal positive reports:
 
http://www.listware.net/201003/gcc-gcc/99756-gmp-501-and-gcc-45.html
 
 
 
 Is there a GCC autobuilder suite that can do all these rebuilds?  I
 will upload there.
 
 I don't have such a setup.
 
 OK, but someone must have a similar setup.  People are occasionally
 rebuilding the archive to test new GCC versions.  Anyone on the
 debian-gcc list got an idea?
 
 does gcc still work when gmp5 is in the archive, and mpfr is not yet
 rebuilt against the new gmp5?

Sure: I've been running that way at home for a year.  Why wouldn't it
work?

Instead of asking cryptic questions, could you please spell out your
concerns in detail so that we could address them.

Thanks,
-Steve





signature.asc
Description: Digital signature


Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-26 Thread Matthias Klose

On 26.02.2011 18:08, Steve M. Robbins wrote:

On Sat, Feb 26, 2011 at 05:49:49PM +0100, Matthias Klose wrote:

On 26.02.2011 04:42, Steve M. Robbins wrote:

On Fri, Feb 25, 2011 at 03:57:28PM +0100, Matthias Klose wrote:

On 25.02.2011 08:46, Steve M. Robbins wrote:



Clearly one should be mindful of the effect on GCC -- that's why I
asked the question on debian-gcc.  Do you have any specific concerns?


Have any concerns been raised on the GCC mailing list?  I've googled
and found only anecdotal positive reports:

   http://www.listware.net/201003/gcc-gcc/99756-gmp-501-and-gcc-45.html




Is there a GCC autobuilder suite that can do all these rebuilds?  I
will upload there.


I don't have such a setup.


OK, but someone must have a similar setup.  People are occasionally
rebuilding the archive to test new GCC versions.  Anyone on the
debian-gcc list got an idea?


does gcc still work when gmp5 is in the archive, and mpfr is not yet
rebuilt against the new gmp5?


Sure: I've been running that way at home for a year.  Why wouldn't it
work?


I didn't ask about your *home*, but the *archive* (and the buildds).


Instead of asking cryptic questions, could you please spell out your
concerns in detail so that we could address them.


what is cryptic about the question?


--
To UNSUBSCRIBE, email to debian-gcc-requ...@lists.debian.org
with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Archive: http://lists.debian.org/4d69341e.4060...@debian.org



Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-26 Thread Steve M. Robbins
Dear Matthias,

On Sat, Feb 26, 2011 at 06:10:54PM +0100, Matthias Klose wrote:
 On 26.02.2011 18:08, Steve M. Robbins wrote:

 Instead of asking cryptic questions, could you please spell out your
 concerns in detail so that we could address them.
 
 what is cryptic about the question?

Thanks for your input.  However, I don't find this conversation
productive any longer.

Thanks,
-Steve


signature.asc
Description: Digital signature


Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-25 Thread Matthias Klose

On 25.02.2011 08:46, Steve M. Robbins wrote:

Matthias asks:


did you check, that all gcc versions do build with the new version
on all architectures, and that the gcc testsuite doesn't show
regressions with the new version? will gcc continue to work, while
re-building mpfr and mpclib with the new gmp?


What I have done is upload gmp to the experimental autobuilders.  The
GMP package build runs a comprehensive test suite that is passing on
all the architectures available to the experimental autobuilders.
This gives me some comfort that the code is reasonably sound.  In
addition, the fact that GMP 5 has been out for over a year gives me
some reason to believe that upstream sources have been adapted to
change in API.

Clearly one should be mindful of the effect on GCC -- that's why I
asked the question on debian-gcc.  Do you have any specific concerns?
Is there a GCC autobuilder suite that can do all these rebuilds?  I
will upload there.  However, to ask me to manually try all
combinations of architecture and GCC version is setting the bar too
high, IMHO.


I don't have such a setup. You should check that GCC continues to work with the 
new package, and doesn't show regressions, comparing with the test-summary.gz of 
an existing run. It's a bit hard to build things without a compiler. Note that 
we had exactly this scenario with an earlier PPL upload (or was it mpclib)?


  Matthias


--
To UNSUBSCRIBE, email to debian-gcc-requ...@lists.debian.org
with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Archive: http://lists.debian.org/4d67c358.1010...@debian.org



Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-25 Thread Steve M. Robbins
On Fri, Feb 25, 2011 at 03:57:28PM +0100, Matthias Klose wrote:
 On 25.02.2011 08:46, Steve M. Robbins wrote:

 Clearly one should be mindful of the effect on GCC -- that's why I
 asked the question on debian-gcc.  Do you have any specific concerns?

Have any concerns been raised on the GCC mailing list?  I've googled
and found only anecdotal positive reports:

  http://www.listware.net/201003/gcc-gcc/99756-gmp-501-and-gcc-45.html



 Is there a GCC autobuilder suite that can do all these rebuilds?  I
 will upload there.  
 
 I don't have such a setup.

OK, but someone must have a similar setup.  People are occasionally
rebuilding the archive to test new GCC versions.  Anyone on the
debian-gcc list got an idea?


-Steve


signature.asc
Description: Digital signature


Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-24 Thread Adam D. Barratt
On Sat, 2011-02-19 at 04:48 -0600, Steve M. Robbins wrote:
 On Sat, Feb 12, 2011 at 01:39:39PM +, Adam D. Barratt wrote:
  Have any of the reverse-dependencies been test-built against the new
  version?  Does the move to 5.0.1 imply any source changes being required
  for reverse-dependencies, or just rebuilds?  (I say just as there
  appear to be around 350 r-dependencies, including at least five from the
  GCC suite).
 
 I haven't done any test-builds.  Since the -dev package changed name,
 I presume that just rebuild won't work; rather, the sources have
 to edit their build-deps.

Out of interest, why is the -dev package versioned?

[...]
 Matthias also responded requesting:

 I see that both the runtime library and the -dev packages have
 different package names. But to be able to still use gmp3 for
 existing GCC versions, please change the source name too, such
 that gmp3 is still available in unstable after the upload of gmp5.
 
 Shall I go ahead and upload the source gmp5?  Then both gmp versions
 will co-exist in the repository and packages can choose to move to
 gmp5 at their leisure.

After some further investigation, it looks like this isn't feasible.
Neither gmp 4.3.2 nor 5.0.1 version their symbols and with both versions
in the archive simultaneously and co-installable, there's a reasonable
risk of a process ending up with both libraries loaded in to its address
space, which is generally not a good idea.

Regards,

Adam


-- 
To UNSUBSCRIBE, email to debian-gcc-requ...@lists.debian.org
with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Archive: 
http://lists.debian.org/1298578544.22974.292.ca...@hathi.jungle.funky-badger.org



Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-24 Thread Matthias Klose

On 24.02.2011 21:15, Adam D. Barratt wrote:

[...]

Matthias also responded requesting:

 I see that both the runtime library and the -dev packages have
 different package names. But to be able to still use gmp3 for
 existing GCC versions, please change the source name too, such
 that gmp3 is still available in unstable after the upload of gmp5.

Shall I go ahead and upload the source gmp5?  Then both gmp versions
will co-exist in the repository and packages can choose to move to
gmp5 at their leisure.


After some further investigation, it looks like this isn't feasible.
Neither gmp 4.3.2 nor 5.0.1 version their symbols and with both versions
in the archive simultaneously and co-installable, there's a reasonable
risk of a process ending up with both libraries loaded in to its address
space, which is generally not a good idea.


did you check, that all gcc versions do build with the new version on all 
architectures, and that the gcc testsuite doesn't show regressions with the new 
version? will gcc continue to work, while re-building mpfr and mpclib with the 
new gmp?


  Matthias


--
To UNSUBSCRIBE, email to debian-gcc-requ...@lists.debian.org
with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Archive: http://lists.debian.org/4d66ca6f.3010...@debian.org



Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-24 Thread Adam D. Barratt
On Thu, 2011-02-24 at 22:15 +0100, Matthias Klose wrote:
 On 24.02.2011 21:15, Adam D. Barratt wrote:
  [...]
  Matthias also responded requesting:
 
   I see that both the runtime library and the -dev packages have
   different package names. But to be able to still use gmp3 for
   existing GCC versions, please change the source name too, such
   that gmp3 is still available in unstable after the upload of gmp5.
 
  Shall I go ahead and upload the source gmp5?  Then both gmp versions
  will co-exist in the repository and packages can choose to move to
  gmp5 at their leisure.
 
  After some further investigation, it looks like this isn't feasible.
  Neither gmp 4.3.2 nor 5.0.1 version their symbols and with both versions
  in the archive simultaneously and co-installable, there's a reasonable
  risk of a process ending up with both libraries loaded in to its address
  space, which is generally not a good idea.
 
 did you check, that all gcc versions do build with the new version on all 
 architectures, and that the gcc testsuite doesn't show regressions with the 
 new 
 version? will gcc continue to work, while re-building mpfr and mpclib with 
 the 
 new gmp?

Steve's mail to this thread less than a week ago said he hadn't done any
test rebuilds at all; I suspect the chances that he's gone on to build
and test multiple GCC versions on multiple architectures in the meantime
are small.

Regards,

Adam


-- 
To UNSUBSCRIBE, email to debian-gcc-requ...@lists.debian.org
with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Archive: 
http://lists.debian.org/1298582383.22974.489.ca...@hathi.jungle.funky-badger.org



Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-24 Thread Steve M. Robbins
On Thu, Feb 24, 2011 at 08:15:44PM +, Adam D. Barratt wrote:
 On Sat, 2011-02-19 at 04:48 -0600, Steve M. Robbins wrote:
  On Sat, Feb 12, 2011 at 01:39:39PM +, Adam D. Barratt wrote:
   Have any of the reverse-dependencies been test-built against the new
   version?  Does the move to 5.0.1 imply any source changes being required
   for reverse-dependencies, or just rebuilds?  (I say just as there
   appear to be around 350 r-dependencies, including at least five from the
   GCC suite).
  
  I haven't done any test-builds.  Since the -dev package changed name,
  I presume that just rebuild won't work; rather, the sources have
  to edit their build-deps.
 
 Out of interest, why is the -dev package versioned?

I don't recall.  I believe it has been versioned since before I took
over maintenance.


  Shall I go ahead and upload the source gmp5?  Then both gmp versions
  will co-exist in the repository and packages can choose to move to
  gmp5 at their leisure.
 
 After some further investigation, it looks like this isn't feasible.
 Neither gmp 4.3.2 nor 5.0.1 version their symbols and with both versions
 in the archive simultaneously and co-installable, there's a reasonable
 risk of a process ending up with both libraries loaded in to its address
 space, which is generally not a good idea.

OK.  So with the recent change for ia64, the experimental build shows
no failures though 3 arches haven't been built (alpha, hppa, mips).
You previously said hppa doesn't have an autobuilder for experimental.
I believe the other two did build the previous version, but I can no
longer find the buildd page that shows all the build log history so I
can't verify that.

Matthias asks:

 did you check, that all gcc versions do build with the new version
 on all architectures, and that the gcc testsuite doesn't show
 regressions with the new version? will gcc continue to work, while
 re-building mpfr and mpclib with the new gmp?

What I have done is upload gmp to the experimental autobuilders.  The
GMP package build runs a comprehensive test suite that is passing on
all the architectures available to the experimental autobuilders.
This gives me some comfort that the code is reasonably sound.  In
addition, the fact that GMP 5 has been out for over a year gives me
some reason to believe that upstream sources have been adapted to
change in API.

Clearly one should be mindful of the effect on GCC -- that's why I
asked the question on debian-gcc.  Do you have any specific concerns?
Is there a GCC autobuilder suite that can do all these rebuilds?  I
will upload there.  However, to ask me to manually try all
combinations of architecture and GCC version is setting the bar too
high, IMHO.


So: what is the next step?

Thanks,
-Steve


signature.asc
Description: Digital signature


Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-19 Thread Steve M. Robbins
On Sat, Feb 12, 2011 at 01:39:39PM +, Adam D. Barratt wrote:
 On Sun, 2011-02-06 at 12:39 -0600, Steve M. Robbins wrote:

 Looking at the package names of the unstable and experimental versions,
 it looks like the main change is libgmp3c2 to libgmp3?  (There is also
 lib{32,64}gmp3 - lib{32,64}gmp10, but those libraries appear to only be
 used by gmp itself).

All the binary packages changed name:
libgmp3c2   -- libgmp10
libgmp3-dev -- libgmp10-dev
lib32gmp3   -- lib32gmp10
lib64gmp3   -- lib64gmp10
... etc ...

except the C++ bindings libgmpxx4ldbl (and its 32-bit  64-bit
variants).

In addition, I had to introduce libmp3, lib{32,64}mp3 because the mp
libraries SOVERSION remained at 3.  However, upstream subsequently
indicated these are fairly archaic [1] so I might just remove this
before uploading to unstable.

[1] http://gmplib.org/list-archives/gmp-devel/2010-November/001669.html


 Have any of the reverse-dependencies been test-built against the new
 version?  Does the move to 5.0.1 imply any source changes being required
 for reverse-dependencies, or just rebuilds?  (I say just as there
 appear to be around 350 r-dependencies, including at least five from the
 GCC suite).

I haven't done any test-builds.  Since the -dev package changed name,
I presume that just rebuild won't work; rather, the sources have
to edit their build-deps.


  Main question: should I go ahead and upload the new version when
  I get a free moment or do we need more investigation?
 
 At the very least I'd first like to confirm that the suggested change
 for ia64 works (rebuilding on the other architectures wouldn't hurt
 either)

I have uploaded gmp 5.0.1+dfsg-4 with ia64 compiling using -O2 and it
built, so this is confirmed.


 and better understand the scope of the changes, particularly in
 relation to e.g. gcc.

Matthias also responded requesting:

I see that both the runtime library and the -dev packages have
different package names. But to be able to still use gmp3 for
existing GCC versions, please change the source name too, such
that gmp3 is still available in unstable after the upload of gmp5.

Shall I go ahead and upload the source gmp5?  Then both gmp versions
will co-exist in the repository and packages can choose to move to
gmp5 at their leisure.

The -dev packages necessarily conflict.  In addition, the C++ bindings
package libgmpxx4ldbl did not change SOVERSION so any package that
depends on libgmpxx4ldbl alone without depending on libgmp3 or
libgmp10 could possibly be compiled against libgmp3 and run with
libgmp10, or vice-versa.  I'm not sure if that's a real problem or
not, but it seems potentially dangerous.  What would you suggest?

Thanks,
-Steve


signature.asc
Description: Digital signature


Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-14 Thread Matthias Klose

On 06.02.2011 19:39, Steve M. Robbins wrote:

Hi,

Now that squeeze is out, I'd like to move from GMP 4 to GMP 5.  The
latter was released upstream about a year ago and the gmp lists
aren't buzzing with outrageous bugs, so it appears stable enough.

I know GMP is used in gcc itself, so I'd appreciate some guidance
from the gcc team as well, since this is a major version change.


I see that both the runtime library and the -dev packages have different package 
names. But to be able to still use gmp3 for existing GCC versions, please change 
the source name too, such that gmp3 is still available in unstable after the 
upload of gmp5.


  Matthias


--
To UNSUBSCRIBE, email to debian-gcc-requ...@lists.debian.org
with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Archive: http://lists.debian.org/4d593ead.2080...@debian.org



Re: GMP transition: 4.3.2 to 5.0.1?

2011-02-12 Thread Adam D. Barratt
On Sun, 2011-02-06 at 12:39 -0600, Steve M. Robbins wrote:
 Now that squeeze is out, I'd like to move from GMP 4 to GMP 5.  The
 latter was released upstream about a year ago and the gmp lists
 aren't buzzing with outrageous bugs, so it appears stable enough.

Looking at the package names of the unstable and experimental versions,
it looks like the main change is libgmp3c2 to libgmp3?  (There is also
lib{32,64}gmp3 - lib{32,64}gmp10, but those libraries appear to only be
used by gmp itself).

 I know GMP is used in gcc itself, so I'd appreciate some guidance
 from the gcc team as well, since this is a major version change.
 
 I uploaded GMP 5 to experimental last year and it builds fine with two
 exceptions: hppa is BD-Uninstallable (?) and ia64 fails to build with
 an ICE.  The underlying cause is already known [1] and from reading
 the bug report I suspect I may be able to work around this by building
 one file with -O2 rather than -O3.  Other suggestions welcome.

hppa is in Needs-Build, not BD-Uninstallable; that's expected because
hppa doesn't have any autobuilders for experimental.

Have any of the reverse-dependencies been test-built against the new
version?  Does the move to 5.0.1 imply any source changes being required
for reverse-dependencies, or just rebuilds?  (I say just as there
appear to be around 350 r-dependencies, including at least five from the
GCC suite).

 Main question: should I go ahead and upload the new version when
 I get a free moment or do we need more investigation?

At the very least I'd first like to confirm that the suggested change
for ia64 works (rebuilding on the other architectures wouldn't hurt
either) and better understand the scope of the changes, particularly in
relation to e.g. gcc.

Regards,

Adam


-- 
To UNSUBSCRIBE, email to debian-gcc-requ...@lists.debian.org
with a subject of unsubscribe. Trouble? Contact listmas...@lists.debian.org
Archive: 
http://lists.debian.org/1297517979.27877.2727.ca...@hathi.jungle.funky-badger.org