[Qemu-devel] TCG for PowerPC host

[Qemu-devel] TCG for PowerPC host

[malc]

Hello,

I have just completed very quick and very ugly TCG port to PowerPC.
There's just enough code to make DOS boot and some demos run. Only
Linux/PPC32 host is supported and only for 32bit softmmu targets.

Fabrice mentioned that Sparc target should be used to test TCG
fully, should sparc*-linux-user (+ linux-user-test) be enough or
full system emulation is required?

If someone is interested in helping with TCG on PPC, please, drop
me a line, i'll try to brush the thing up and release it.

-- 
mailto:av1474@comtv.ru

Re: [Qemu-devel] TCG for PowerPC host

[Paul Brook]

> Fabrice mentioned that Sparc target should be used to test TCG
> fully, should sparc*-linux-user (+ linux-user-test) be enough or
> full system emulation is required?

You should definitely full system emulation. It exercises exceptions a lot 
more (most userspace applications never fault).

Paul

Re: [Qemu-devel] TCG for PowerPC host

[malc]

On Sun, 11 May 2008, Paul Brook wrote:

>> Fabrice mentioned that Sparc target should be used to test TCG
>> fully, should sparc*-linux-user (+ linux-user-test) be enough or
>> full system emulation is required?
>
> You should definitely full system emulation. It exercises exceptions a lot
> more (most userspace applications never fault).

While not beging sure what exceptions have to do with TCG i attempted it,
alas it seems at this stage that his is an exercise in futility.

  a. Approach taken by common TCG code when dealing with calling helper
     code is X86[_64] specific. Passing/returning long long quantities
     in SVR4PPC ABI doesn't map to this. I added a few temporary hacks,
     but at the end of the day they are just hacks.

  b. Even the last pre TCG snapshot of QEMU doesn't quite work with
     SPARC test iamge[1]. It makes a lot of progress (much more than
     specifically hacked TCG) but eventually the kernel crashes.

Also enabling CPU state logging (for both pre and TCG builds) yields
interesting  behaviour, namely - after a few guest instructions the
floating point part of the state is printed weirdly; more often than
not the values are incredibly huge, other times they are nans, in
reality env->fpr[0..N] are all bit zeroes. Given all that i do not
think that testing with sparc fullsystem emulation would get anyone
anywhere.

On the brighter side after fixing some issues i can boot i386 Linux
test image[2] and some DSL.

[1] http://fabrice.bellard.free.fr/qemu/sparc-test-0.2.tar.gz
[2] http://fabrice.bellard.free.fr/qemu/linux-0.2.img.bz2

-- 
mailto:av1474@comtv.ru