Re: [Syllable-kernel] SMP: -O1 vs. -O2

[Kristian V. D. V. <va...@li...>]

On Sunday 17 October 2004 7:33 pm, Jake Hamby wrote:
> Kristian Van Der Vliet wrote:
>> So far the only thing which has made any difference is to compile with
>> -O1 instead of -O2, so it must be an over-agresive optimisation going on
>> somewhere.  It just doesn't appear to be in the atomic code.
>
> Can you figure out which file(s) need to be compiled with -O1 for the
> kernel to work?  You should be able to do a binary search:  compile half
> of the files with -O1 and keep dividing by two until you've isolated the
> files which are responsible.  Then post your results and I'll look at
> the assembly output to see if anything pops out.  Hopefully the problem
> is isolated to one or two specific source files.

Attached is the result of tonights work.  I've identified smp.c as the culprit 
(Makes sense, at least)  I was working with the stock 0.5.4 source which did 
not include your previous patches for -O2, so I've used -O3 and -O1, although 
I will confirm tommorrow with -O1 and -O2 just to be 100% sure.

The zip contains the following files:

smp.o.O1		- smp.c compiled with the following flags:

-D__KERNEL__ -D__BUILD_KERNEL__ -no-fPIC -c -g -O1 -fno-strict-aliasing -Wall 
-march=i586 -mcpu=i586

smp.o.O3		- smp.c compiled with the following flags:

-D__KERNEL__ -D__BUILD_KERNEL__ -no-fPIC -c -g -O3 -fno-strict-aliasing -Wall 
-march=i586 -mcpu=i586

smp.s.O1		- smp.c compiled with the following flags:

-D__KERNEL__ -D__BUILD_KERNEL__ -no-fPIC -s -g -O1 -fno-strict-aliasing -Wall 
-march=i586 -mcpu=i586

smp.s.O3		- smp.c compiled with the following flags:

-D__KERNEL__ -D__BUILD_KERNEL__ -no-fPIC -s -g -O3 -fno-strict-aliasing -Wall 
-march=i586 -mcpu=i586

smp.O1.dump	- Output from objdump -ds smp.o.O1
smp.O3.dump	- Output from objsump -ds smp.o.O3

If you need anything else please give me a shout and I'll get see what I can 
do.

-- 
Vanders
http://syllable.sourceforge.net/
http://www.liqwyd.com