llvm dev - Jun 2013 - [LLVMdev] [cfe-dev] [3.3 Release] 3.3rc3 Now Available

Hi LLVM-ites!

We are doing a quick test of 3.3rc3. Our testers did a quick turn around and
created binaries which you can test. Please give them a go and let us know how
they work for you. They are here:

	http://llvm.org/pre-releases/3.3/rc3

We don't have a lot of time remaining in the release cycle, so please do
whatever you can to make sure they are solid. In particular, make sure that they
solve the PRs that were reported during the release
(http://llvm.org/bugs/show_bug.cgi?id=15999).

Share and enjoy!
-bw

Hi Bill,

I was running some benchmarks and was surprised that the startup
performance of your binaries is much worse than the stable binaries
for 3.2

I've used:
http://llvm.org/releases/3.2/clang+llvm-3.2-x86_64-linux-ubuntu-12.04.tar.gz
vs
http://llvm.org/pre-releases/3.3/rc3/clang+llvm-3.3rc3-Ubuntu-12.04.2-x86_64.tar.gz

On a simple hello.c file:
/ssd/bench$ TIMEFORMAT="%E"
/ssd/bench$ for I in `seq 1 5`; do \
  time /ssd/llvm3.2/bin/clang -c hello.c; \
  done
0.015
0.014
0.013
0.013
0.013

/ssd/bench$ for I in `seq 1 5`; do \
  time /ssd/llvm3.3rc/bin/clang -c hello.c; \
  done
0.992
0.978
0.982
0.981
0.976

Is this an expected behaviour (e.g. +Asserts or Debug build)
or a real regression?

P.s.
I'm not very familiar with the Clang release process, so I might be
missing something obvious...

--
Thanks,
Timur Iskhodzhanov

2013/6/5 Bill Wendling <wendling at apple.com>:
> Hi LLVM-ites!
>
> We are doing a quick test of 3.3rc3. Our testers did a quick turn around
and created binaries which you can test. Please give them a go and let us know
how they work for you. They are here:
>
>         http://llvm.org/pre-releases/3.3/rc3
>
> We don't have a lot of time remaining in the release cycle, so please
do whatever you can to make sure they are solid. In particular, make sure that
they solve the PRs that were reported during the release
(http://llvm.org/bugs/show_bug.cgi?id=15999).
>
> Share and enjoy!
> -bw
>
> _______________________________________________
> cfe-dev mailing list
> cfe-dev at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-dev

It's probably PR12517.

Looking at the clang binary, it's got a /home/ dir in RPATH:

$ objdump -p clang+llvm-3.3rc3-Ubuntu-12.04.2-x86_64/bin/clang | grep RPATH
  RPATH
$ORIGIN/../lib:/home/aadgrand/tmp/LLVM-3.3rc3/rc3/Phase3/Release+Asserts/llvmCore-3.3-rc3.obj/Release+Asserts/bin

This will slow things down if the system tries to automount /home/aadgrand.

Maybe we should merge r182559 to the release branch? Or at least we
should take care not to build it in a /home directory.

 - Hans

On Thu, Jun 6, 2013 at 4:19 PM, Timur Iskhodzhanov <timurrrr at
google.com> wrote:
> Hi Bill,
>
> I was running some benchmarks and was surprised that the startup
> performance of your binaries is much worse than the stable binaries
> for 3.2
>
> I've used:
>
http://llvm.org/releases/3.2/clang+llvm-3.2-x86_64-linux-ubuntu-12.04.tar.gz
> vs
>
http://llvm.org/pre-releases/3.3/rc3/clang+llvm-3.3rc3-Ubuntu-12.04.2-x86_64.tar.gz
>
> On a simple hello.c file:
> /ssd/bench$ TIMEFORMAT="%E"
> /ssd/bench$ for I in `seq 1 5`; do \
>   time /ssd/llvm3.2/bin/clang -c hello.c; \
>   done
> 0.015
> 0.014
> 0.013
> 0.013
> 0.013
>
> /ssd/bench$ for I in `seq 1 5`; do \
>   time /ssd/llvm3.3rc/bin/clang -c hello.c; \
>   done
> 0.992
> 0.978
> 0.982
> 0.981
> 0.976
>
> Is this an expected behaviour (e.g. +Asserts or Debug build)
> or a real regression?
>
> P.s.
> I'm not very familiar with the Clang release process, so I might be
> missing something obvious...
>
> --
> Thanks,
> Timur Iskhodzhanov
>
> 2013/6/5 Bill Wendling <wendling at apple.com>:
>> Hi LLVM-ites!
>>
>> We are doing a quick test of 3.3rc3. Our testers did a quick turn
around and created binaries which you can test. Please give them a go and let us
know how they work for you. They are here:
>>
>>         http://llvm.org/pre-releases/3.3/rc3
>>
>> We don't have a lot of time remaining in the release cycle, so
please do whatever you can to make sure they are solid. In particular, make sure
that they solve the PRs that were reported during the release
(http://llvm.org/bugs/show_bug.cgi?id=15999).
>>
>> Share and enjoy!
>> -bw