Excessive exception size cost ...

Hi guys,

On Wed, 2012-02-29 at 16:29 +0000, Michael Meeks wrote:
>       So - I did some more analysis, and re-compiled all of LibreOffice both
> with and without the attached patch - analysing all 287 shared libraries
> we get:

        So - because of the expert skepticism of my estimate of where the
wasteage is: ie. exception unwind tables, I re-ran my relocstats.pl tool
(which I've checked in here):

http://cgit.freedesktop.org/libreoffice/contrib/dev-tools/tree/scripts/

        over the code both with and without those extra bad_alloc exceptions we
spit out on every string allocation (which are used to do nothing more
useful than abort ;-).

        Here are the numbers from:

$ relocstat.pl --data-profile *.so # in program/

Section size breakdown # with no string exceptions
 code            74126kb - 45%
 exceptions      46003kb - 28%
 data            20913kb - 13%
 linking         15432kb - 9.3%
 data relocs     8462kb - 5.1%
...
 Total: 170285850 bytes

Section size breakdown # with string exceptions thrown everywhere
 code            75233kb - 44%
 exceptions      48131kb - 28%
 data            20913kb - 12%
 linking         15445kb - 9.1%
data relocs     8462kb -  5%
...
 Total: 173612058 bytes

        So - there is the 1.9% size saving ~3.3Mb saved (which is a lower bound
- we can do better by being more complete).

        Perhaps what is more frightening, is the sheer weight of the exception
information: we have fourty-eight (48) Mb of exception unwind
information vs. 75Mb of code; that is cf.

http://cgit.freedesktop.org/libreoffice/contrib/dev-tools/tree/scripts/relocstat.pl#n547

        from just two section types: .gcc_except_table and .eh_frame.

        It appears that for every 10 bytes of .text (ie. code) we create 6+
bytes of exception unwind information.

        Given that we then that in (100 - epsilon)% of the generated cases
don't do anything at all useful with the results beyond the crash
handler, this seems rather a high cost to pay.

        It is also a markedly higher proportion than mozilla:

Section size breakdown
 code            15923kb - 60%
 exceptions      4870kb - 18%
 data            3226kb - 12%
 data relocs     1682kb - 6.3%
 linking          460kb - 1.7%
...

        This is compiled with openSUSE 12.1's default: gcc 4.6.2, and operating
on stripped binaries.

        We can also see that of the two potential causes of bloat removal of
not doing this:

        a) not in-lining:

                if (error_return) throw ::std::bad_alloc();
vs.
        b) not generating huge unwind tables for methods doing pure
           string operations
        
        The win breaks down thus:

        code change:      -1107kb
        exception change  -2128kb

        So 2/3rds of the reduction is simply shrinking the unwind tables.

        While it shouldn't affect the ultimate numbers, I am using the
--enable-merged-libs feature here (as is obvious from the
autogen.lastrun I posted).

        So - there we are: exceptions hurt, they hurt really a lot size-wise,
and they provide us with very little real value since we just abort when
they are thrown in ~all cases. Alternatively, perhaps they are truly a
productivity tool - they certainly let us generate more bytes of output
more quickly ;-)

        Feedback appreciated, my relocstat.pl tool is untouched for several
years, possibly it is not adding up right - please do check it out.

        ATB,

                Michael.

-- 
michael.meeks@suse.com  <><, Pseudo Engineer, itinerant idiot