On 5/2/08, Adam Lackorzynski <adam@os.inf.tu-dresden.de> wrote:
>
>  On Thu May 01, 2008 at 18:02:46 +0300, Blue Swirl wrote:
>  > On 5/1/08, Adam Lackorzynski <adam@os.inf.tu-dresden.de> wrote:
>
> > >  For 64bit target T0 is 64bits so "=a" does not work and "=A" is needed.
>  > >  The strange thing is that I need to throw away the upper 32bits because
>  > >  otherwise it won't work. gen_func is defined to return just long but T0
>  > >  is unsigned long long, this seems inconsistent. The 'and' does not
>  > >  appear in 32bit targets so it does not harm there.
>  >
>  > This is because in this special case, T0 is not used as target CPU
>  > temporary, but instead to return next TB address. On i386 this is 32
>  > bits, so only EAX is needed. TCG does not touch EDX, so it contains
>  > garbage. This also means that moving EDX to high word of T0 and then
>  > throwing the high word away may be slightly wasteful.
>
>
> So I played a bit more with this by trying out the 'and' and the tmp
>  variable approaches. With the tmp variables the generated code looks ok
>  whereas with the 'and' approach it looks especially scary with gcc-4.3
>  (gcc-3.4 looks ok). I have two versions now, one condensed and ugly and
>  then one with separate parts for 32 and 64 targets. I think this one
>  should be prefered.

I made a new version that does not use T0 at all. Tested on i386 and
AMD64, both Sparc32 and Sparc64 work. AMD64 asm version does not seem
to be necessary.