Re: building kernel
Thank you for answer, Falk.
Actually, I've started my experiments after I found out that
decompression of digital audio (mp3, ogg, etc) that are floation point
operations IIRC, are peformed on my 600ua much less efficiently, than,
say, on Celeron 566. For instance, xmms plaing 128kbps takes about 6%
of alpha and about 1% of C566.
Is this an expected performace?
Thanks,
Sacha
>>>>> On 29 Nov 2002 13:25:05 +0100
>>>>> "FH" == Falk Hueffner <falk.hueffner@student.uni-tuebingen.de> wrote:
FH>
FH> Alexander Kotelnikov <sacha@debian.org> writes:
>> Can anyone tell me, why linux kernel is built with -mno-fp-regs and
>> -Wa,-mev6, (see linux/arch/alpha/Makefile).
FH>
FH> The kernel must not use floating point registers, because they are not
FH> saved on context switch from user space to kernel space. Not using
FH> floating point math isn't enough, because gcc uses FP registers as
FH> register spilling space (which seems to me of doubtful usefulness,
FH> because it isn't faster than memory...)
FH>
FH> -Wa,-mev6 is needed so the assembler will accept EV6 opcodes. It won't
FH> affect the generated code, it just avoid the assembler barfing.
FH>
>> It is also interesting, when -mno-soft-float and -mfp-reg are essential
>> and not dangerous. Kernel built with -mno-soft-float -mfp-regs -Wa,mev6
>> oops'es at boot time,
FH>
FH> Well, that's not surprising...
FH>
>> without -Wa,mev6 system boots but many programs do not work.
FH>
FH> but this is weird... it really shouldn't affect the generated code.
FH>
>> BTW, what for kernel is built with -ffixed-8?
FH>
FH> Register 8 is hardwired to "current" (the running process context),
FH> and should therefore not be used by the compiler.
FH>
FH> --
FH> Falk
FH>
FH>
FH> --
FH> To UNSUBSCRIBE, email to debian-alpha-request@lists.debian.org
FH> with a subject of "unsubscribe". Trouble? Contact listmaster@lists.debian.org
FH>
--
Alexander Kotelnikov
Saint-Petersburg, Russia
Reply to: