Vorbis dev - Jul 2005 - [patch] libvorbis + gcc4

http://trac.xiph.org/cgi-bin/trac.cgi/ticket/583

looking at why gcc4 could increase vorbis so much, i first had a look at
the compiler options, and i saw -O20, as far as i know there is 
-O0, .. -O3 and -Os but no -O20. changing that to -O2 (patch attached)
one gets close to the result of gcc-3.4.

gcc3.4-O20.ogg  20423 

gcc4-O20.ogg    54623
gcc4-O3.ogg     54623

gcc4-O2.ogg     20423

looking at the hexdump of gcc4-02.ogg and gcc3.4-020.ogg, 
they differ slightly, things like:
-0000000 674f 5367 0200 0000 0000 0000 0000 0e8a
-0000010 25e5 0000 0000 5b97 0de7 1e01 7601 726f
+0000000 674f 5367 0200 0000 0000 0000 0000 fb5f
+0000010 0bb0 0000 0000 6e2e af20 1e01 7601 726f

luckily the wav files i get from oggdec do not differ.

the question is now which part of -O3(which -O20 is in the end) 
causes this increase in vorbis and is it a bug in vorbis or 
in gcc4?

j
-------------- next part --------------
A non-text attachment was scrubbed...
Name: vorbis-gcc4.patch
Type: text/x-patch
Size: 2348 bytes
Desc: not available
Url :
http://lists.xiph.org/pipermail/vorbis-dev/attachments/20050728/b4deab90/vorbis-gcc4.bin

>>>>> "j" == j  <j@v2v.cc> writes:
j> the question is now which part of -O3(which -O20 is in the end)
j> causes this increase in vorbis and is it a bug in vorbis or in
j> gcc4?

In gcc3 -O3 only adds -finline-functions and -frename-registers
to -O2.  As of gcc 4.0.0 it adds -finline-functions, -funswitch-
loops and -fgcse-after-reload.  

Since inline-functions wasn't causing the problem before gcc4,
it probably isn't now either.  I'd bet on -funswitch-loops.

This is what gcc.info has to say about that option:

,---- -funswitch-loops
|  Move branches with loop invariant conditions out of the loop,
|  with duplicates of the loop on both branches (modified according
|  to result of the condition).
`----

But give each possibility a test compile to see for sure.

Incidently, was there any performance difference between the
resulting binaries?

-JimC
-- 
James H. Cloos, Jr. <cloos@jhcloos.com>