149151 - Use -lazyload to distribute cost of loading objects thruought their lifecycle

Status: RESOLVED FIXED

(SeaMonkey :: Build Config, defect)

Description

[Nidheesh Dubey]

Monitoring process startup, I noticed libxpcom.so dlopen'ing
    libdemangle.so.1 via it's .init section.  Why can't this be a standard
    dependency?  or even better, couldn't this dlopen be deferred until
    the demangling library is needed?

    This is where I start talking about defferred loading.  If you have to
    load a lot of objects, is it possible to load them when you need them,
    rather that all up front incase you need them?

    After we've loaded the process and fired the initial .init's, ld.so.1
    reveals a whole bunch of unreferenced dependencies (as ldd did above),
    and more importantly some unused:

    00496: 1: file=libjsj.so  unused: does not satisfy any references.
    00496: 1: file=libplc4.so  unused: does not satisfy any refe
    00496: 1: file=libgtk-1.2.so.0  unused: does not satisfy any refere
    00496: 1: file=libgmodule-1.2.so.0  unused: does not satisfy any re
    00496: 1: file=libXext.so.0  unused: cyclic group [3] member: unreferen
    00496: 1: file=libX11.so.4  unused: cyclic group [3] member: unreferenc
    00496: 1: file=libm.so.1  unused: does not satisfy any references
    00496: 1: file=librt.so.1  unused: does not satisfy any references
    00496: 1: file=libmp.so.2  unused: does not satisfy any references
    00496: 1: file=libaio.so.1  unused: does not satisfy any references
    00496: 1: file=libmd5.so.1  unused: does not satisfy any references
    
    If you follow my advice in A, some of this might disappear.  But an
    additional technique is to use lazy-loading.  Build your objects
    with -zlazyload and then any dependencies will only be loaded when
    a symbolic reference is made to them.   This lazy-loading has the
    effect of sharing the cost of loading objects throughout the
    process lifecycle, rather than bearing all the costs up front.

Comment 1

[Margaret Chan]

Attachment: Patch with -lazyload and -zcombreloc option for solaris

Our linker expert suggests us to add these two options in the LDFLAGS.	This is
for improving performance

Comment 2

[Margaret Chan]

Hi [e-mail]:  Can you give me a r= for the above patch please?  I 
have put the flag in the place where it should only affect Solaris non-gcc 
(i.e., using Sun's compiler) builds.  Thanks, Margaret

Comment 3

[hacker formerly known as [e-mail]]

Comment on attachment 89156 [details] [diff] [review]
Patch with -lazyload and -zcombreloc option for solaris

r=cls

Comment 4

[Margaret Chan]

Thanks cls.  Didn't get to come back to this until now.  Who should I be asking
for sr= these days?   Margaret

Comment 5

[hacker formerly known as [e-mail]]

Per http://www.mozilla.org/hacking/reviewers.html, r=cls should be all you need
to check into the build system files.  You can ask leaf for sr= if you really
feel the need for a second review. :)

Comment 6

[Margaret Chan]

Great.  I am not going to bother leaf if I don't have to.  I just think that the
process requires that.  Glad to know that it doesn't.  I'll check it in soon. 
Thanks, Margaret

Comment 7

[Margaret Chan]

Tinderbox says that it is frozen for 1.1beta, I guess I'll wait till next week
for checkin then.  

Comment 8

[Roland Mainz]

I did some test builds on Solaris 2.7/SPARC with Sun Workshop 7/FCS, the patch
seems to cause no problems.

----

Margaret Chan wrote:
> Tinderbox says that it is frozen for 1.1beta, I guess I'll wait till next week
> for checkin then.

It's now frozen for 1.1final... ;-(

But I do not see problems with these flags  - what about sending an email to
[e-mail] and request approval for the "trunk" ?

Comment 9

[Margaret Chan]

Bummer...  It seems to be frozen forever...  Anyway, just send an email to
driver.  I'll see what they say.  Margaret

Comment 10

[Roland Mainz]

Margaret Chan wrote:
> Bummer...  It seems to be frozen forever...

Yeah. Grrrrrrr... xx@@@!!!... ;-(

> Anyway, just send an email to driver.

Erm... do you mean you will send an email or should I send the request ?

Comment 11

[Margaret Chan]

Oops.  Excuse my grammar.

> Anyway, just send an email to driver.

It should say "just sent an email to driver." since I have already done so.  I
Am currently waiting for a reply from them.  Margaret

Comment 12

[Asa Dotzler [:asa]]

Comment on attachment 89156 [details] [diff] [review]
Patch with -lazyload and -zcombreloc option for solaris

a=asa (on behalf of drivers) for checkin to 1.1

Comment 13

[Roland Mainz]

Margaret Chan:
Please read http://tinderbox.mozilla.org/showbuilds.cgi?tree=SeaMonkey :
-- snip --
"OPEN means that the tree is open only to drivers-approved checkins that have
been approved within the three days prior to the checkin."
-- snip --
or short: Driver approval is only valid if you check the code "in" ASAP; the
approval expires after three days...

... can you check the patch "in" before the a=asa (see comment #12) expires,
please ?

Comment 14

[Margaret Chan]

Thanks for the reminder, Roland.  I just did.  Marked it resolved fixed.