[htdig-dev] Re: Efficiency Improvement for String.cc

[Lachlan A. <lac...@ip...>]

On Sunday 02 February 2003 11:35, Neal Richter wrote:

> very few of our
> inline functions end up getting inlined.  I've done a fair bit of
> work with gprof and looking at what functions are actually inlined
> in the assembly.  At least with -O2 I'm not seeing some of them
> making it.

What version of  gcc  are you using?  I've just compiled using 3.1,=20
and it seems to be inlining.  For example, the copy constructor gives=20
the output below, which clearly shows line 229 of  htString.h  being=20
inlined.

=09.align 2
=09.p2align 4,,15
=2Eglobl _ZN6StringC2ERKS_
=09.type=09_ZN6StringC2ERKS_,@function
_ZN6StringC2ERKS_:
=2ELFB7:
=09.loc 1 67 0
=09pushl=09%ebp
=2ELCFI34:
=09movl=09%esp, %ebp
=2ELCFI35:
=09pushl=09%ebx
=2ELCFI36:
=09pushl=09%edx
=09call=09.L31
=2EL31:
=09popl=09%ebx
=09addl=09$_GLOBAL_OFFSET_TABLE_+[.-.L31], %ebx
=09movl=098(%ebp), %edx
=09movl=0912(%ebp), %ecx
=2ELBB10:
=09movl=09_ZTV6String@GOT(%ebx), %eax
=09addl=09$8, %eax
=09.loc 1 68 0
=09movl=09$0, 4(%edx)
=09.loc 1 67 0
=09movl=09%eax, (%edx)
=09.file 38 "htString.h"
=09.loc 38 229 0
=2ELBB11:
=2ELBB12:
=09movl=094(%ecx), %eax
=09.loc 1 68 0
=2ELBE12:
=2ELBE11:
=09movl=09$0, 8(%edx)
=09.loc 1 69 0
=09movl=09$0, 12(%edx)
=09.loc 38 229 0
=09testl=09%eax, %eax
=09jle=09.L25
=09pushl=09%eax
=09pushl=09%eax
=09movl=0912(%ecx), %eax
=09pushl=09%eax
=09pushl=09%edx
=2ELCFI37:
=09call=09_ZN6String4copyEPKcii@PLT
=09addl=09$16, %esp
=09.loc 1 73 0
=2EL25:
=2ELBE10:
=09movl=09-4(%ebp), %ebx
=09movl=09%ebp, %esp
=09popl=09%ebp
=09ret
=2ELFE7:
=2ELfe7:
=09.size=09_ZN6StringC2ERKS_,.Lfe7-_ZN6StringC2ERKS_

This isn't from DDD, but using the command

/bin/sh ../libtool --mode=3Dcompile g++ -DHAVE_CONFIG_H -I. -I.=20
-I../include=20
-DDEFAULT_CONFIG_FILE=3D\"/home/lha/devel/htdig/install//conf/htdig.conf\=
"           =20
-I../include -I../htlib          -I../htnet -I../htcommon               =20
-I../htword             -I../db -I../db                   =20
-I/usr/include  -g -O2 -Wall -fno-rtti -fno-exceptions -S String.cc

which is what my makefile generates, but with '-c' replaced by '-S'.

>   My next set of changes are going to be enhancements to WordDB, &
> WordKey.

These sound like good improvements.  However when I do a dig of my=20
local computer, it seems rather disk bound, with CPU usage hovering=20
around 50%, of which a lot is system time.  I haven't yet checked how=20
much of this is reading the input docs and how much is database=20
access, but my guess is that implementing your more compact database=20
format will give more bang for your development buck.  It has the=20
added advantages that it uses less disk space, and that it speeds up=20
searching (assuming that is also disk bound).  However, I am really=20
keen to get 3.2.0b5 out and I personally won't be working on=20
optimisations until then.  Thoughts?

>   I need to get more familiar with how gcc does some of these
> optimizations.. but at the same time I'm inclined not to rely on it
> too much.

Agreed.  We should eliminate unnecessary function calls, but *some*=20
will still be needed, and we might as well get those inlined.

Cheers,
Lachlan