c++-analyzer slow down ?

Hi,

There could be something wrong with my build system, but I get the
impression that doing a scan-build of a C++ project (like llvm/clang,
for example) has recently drastically increased (hours longer) in
duration. Could there be a good reason for this, or do I need to go
check my build system ?

Thanks,

John Smith.

John,

When did you start seeing this?

We've recently added experimental inter-procedural analyzes for C++, which would definitely explain the slowdown.

Cheers,
Anna.

Well I started a run of 'make all; make install; scan-build make' this
morning, and when I came home this evening it was still running. I
(thoughtlessly) terminated the process immediately, thinking something
was wrong with my system. The installed clang version says it's 'clang
version 3.2 (trunk 161033)'. The last run I did that didnt seem have
this problem (at least not this excessive) was with 'clang version 3.2
(trunk 160917) '

Does that make sense ?

Regards,

john Smith.

John,

When did you start seeing this?

We've recently added experimental inter-procedural analyzes for C++, which would definitely explain the slowdown.

Well I started a run of 'make all; make install; scan-build make' this
morning, and when I came home this evening it was still running. I
(thoughtlessly) terminated the process immediately, thinking something
was wrong with my system. The installed clang version says it's 'clang
version 3.2 (trunk 161033)'. The last run I did that didnt seem have
this problem (at least not this excessive) was with 'clang version 3.2
(trunk 160917) '

John,

IPA was added several weeks ago. There is nothing we know of that would introduce a substantial slowdown, added in the last week.

Can you retest with TOT? Maybe some other memory hungry task was running at the same time on the machine?

Thanks,
Anna.

Hrm. I started another scan-build of llvm/clang (with r161050) at
about 19:48, and it now is 02:56 and it's still analyzing. The compile
took about 2 hours on the same machine, which is about the same time
it took with other revisions (so thats pretty normal). Could it be
something llvm-clang project specific that makes the analysis take
longer ? Just thinking out loud here... It could still be something
with my build system, of course.

And im sorry, I dont know what 'TOT' is...
:frowning:

Regards,

John Smith.

"Tip of Tree", i.e, the current state of clang/llvm in subversion.

-- Marshall

Marshall Clow Idio Software <mailto:mclow.lists@gmail.com>

A.D. 1517: Martin Luther nails his 95 Theses to the church door and is promptly moderated down to (-1, Flamebait).
        -- Yu Suzuki

Well, the scan-build of svn r161050 using that same version finally
completed at about 08.00u. Since I started it at about about 19.48,
that means that it roughly took 12 hours to complete ? Versus roughly
2 hours for a 'normal' compile on the same system...

Oh, well....

John,

Could you test with an older clang version? You said that it is likely a regression from a week ago. Can we check for sure how long it takes with an older version of clang?

Thanks,
Anna.

Sure. I think it didnt last this long with r160917. Would that
revision be alright ? I could also go for r160630, which is older, but
im more certain that it didnt take this long with r160630 than I am
with r160917.

Regards,

John Smith.

John,

Could you test with an older clang version? You said that it is likely a regression from a week ago. Can we check for sure how long it takes with an older version of clang?

Thanks,
Anna.

Sure. I think it didnt last this long with r160917. Would that
revision be alright ? I could also go for r160630, which is older, but
im more certain that it didnt take this long with r160630 than I am
with r160917.

Either one should be fine.

Anna.

Well I did a r160917 scan-build of r160917, and that took roughly 12
hours. Guess I am wrong then, and that this isnt new. Perhaps I wasnt
paying enough attention to the duration before, and just thought that
it was faster.

Thanks for the time and effort to take it seriously, though. And my
apologies for the confusion it caused,

Regards,

John Smith

Well I did a r160917 scan-build of r160917, and that took roughly 12
hours. Guess I am wrong then, and that this isnt new. Perhaps I wasnt
paying enough attention to the duration before, and just thought that
it was faster.

Thanks for the time and effort to take it seriously, though. And my
apologies for the confusion it caused,

No problem! We will work on improving performance overall, however, it is possible that analysis would be 6 time slower than execution.

Cheers,
Anna.