[Qemu-devel] Performances on Mac OS X

[Qemu-devel] Performances on Mac OS X

[Laurent Amon]

Hello,

Just to let you know that I have been benchmarking with linux-test/nbench on my
Titanium G4/400 and that I have been able to sqeeze a few percents (some 5%) of
performance by compiling with -mcpu=7400 wirth GCC 3.3. This is the only flag
that seems useful.
I have also tried -mpowerpc-gpopt, -mpowerpc-gfxopt, -falign-loops,
-funroll-loops, -fnew-ra, -ftracer.

When compiling with -finline-functions, the generated i386-softmmu/op.h is bad,
and if you do not use it do compile dyngen, there is no improvement.
-Os actually decreases performance and I have not been able to generate a
profiled  executable with -fprofile-arcs.

Owners of PPC7450 or of G5 should try the relevant switch.

I was wondering whether the implementation of qemu is tied to gcc or if it might
be interesting to try to compile using IBM's compiler, which is able to generate
Altivec code...

Lga.

Re: [Qemu-devel] Performances on Mac OS X

[Falk Hueffner]

Laurent Amon <amon@stanfordalumni.org> writes:

> Just to let you know that I have been benchmarking with linux-test/nbench on my
> Titanium G4/400 and that I have been able to sqeeze a few percents (some 5%) of
> performance by compiling with -mcpu=7400 wirth GCC 3.3. This is the only flag
> that seems useful.
> I have also tried -mpowerpc-gpopt, -mpowerpc-gfxopt, -falign-loops,
> -funroll-loops, -fnew-ra, -ftracer.

Well, the instruction stubs don't contain loops and are short enough
to be trivial to assign registers. So this is not really surprising.

> I was wondering whether the implementation of qemu is tied to gcc or
> if it might be interesting to try to compile using IBM's compiler,
> which is able to generate Altivec code...

It might be possible with some work, but Altivec will not be useful
for the vast majority of instructions, basically only for MMX and
similar things, which AFAIK aren't yet supported in qemu.

-- 
	Falk

RE: [Qemu-devel] Performances on Mac OS X

[Natalia Portillo]

GCC supports Altivec, just see the help. 

> -----Mensaje original-----
> De: qemu-devel-bounces+iosglpgc=teleline.es@nongnu.org 
> [mailto:qemu-devel-bounces+iosglpgc=teleline.es@nongnu.org] 
> En nombre de Laurent Amon
> Enviado el: lunes, 02 de agosto de 2004 13:08
> Para: qemu-devel@nongnu.org
> Asunto: [Qemu-devel] Performances on Mac OS X
> 
> 
> Hello,
> 
> Just to let you know that I have been benchmarking with 
> linux-test/nbench on my Titanium G4/400 and that I have been 
> able to sqeeze a few percents (some 5%) of performance by 
> compiling with -mcpu=7400 wirth GCC 3.3. This is the only 
> flag that seems useful.
> I have also tried -mpowerpc-gpopt, -mpowerpc-gfxopt, 
> -falign-loops, -funroll-loops, -fnew-ra, -ftracer.
> 
> When compiling with -finline-functions, the generated 
> i386-softmmu/op.h is bad, and if you do not use it do compile 
> dyngen, there is no improvement.
> -Os actually decreases performance and I have not been able 
> to generate a profiled  executable with -fprofile-arcs.
> 
> Owners of PPC7450 or of G5 should try the relevant switch.
> 
> I was wondering whether the implementation of qemu is tied to 
> gcc or if it might be interesting to try to compile using 
> IBM's compiler, which is able to generate Altivec code...
> 
> Lga.
> 
> 
> 
> _______________________________________________
> Qemu-devel mailing list
> Qemu-devel@nongnu.org
> http://lists.nongnu.org/mailman/listinfo/qemu-devel

RE: [Qemu-devel] Performances on Mac OS X

[Laurent Amon]

Actually, GCC 3.3 supports Altivec extensions (like vector keywords) but does
not generate Altivec instructions on regular code, so you need to write
specific code.
The IBM compiler optimizes regular code using the Altivec.

Lga.

Selon Natalia Portillo <claunia@claunia.com>:

> GCC supports Altivec, just see the help.