[Qemu-devel] Re: Compile files only once: some planning

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Mar 24, 2010 at 10:24:12PM +0200, Blue Swirl wrote:
> On 3/24/10, Michael S. Tsirkin <mst@redhat.com> wrote:
> > On Wed, Mar 24, 2010 at 10:05:10PM +0200, Blue Swirl wrote:
> >  > On 3/24/10, Michael S. Tsirkin <mst@redhat.com> wrote:
> >  > > On Tue, Mar 23, 2010 at 11:43:51PM +0200, Blue Swirl wrote:
> >  > >  > rtl8139.c, e1000.c: need to convert ldl/stl to cpu_physical_memory_read/write.
> >  > >
> >  > >
> >  > > I don't see how it would help. These still get target_phys_addr_t which
> >  > >  is per-target. Further, a ton of devices do
> >  > >  cpu_register_physical_memory/qemu_register_coalesced_mmio.
> >  > >  These are also per target.
> >  >
> >  > I don't know what I was eating yesterday: there are no references to
> >  > ldl/stl in either rtl8139 or e1000. In fact, the conversion is simple
> >  > for the device itself, just add a property "be". The attached patch
> >  > performs this part.
> >  >
> >  > But now there is a bigger problem, how to pass the property to the
> >  > device. It's not fair to require the user to remember to set it.
> >
> >
> > I still don't fully understand how come e1000 cares about
> >  target endianness.
> 
> It shouldn't. Maybe the real fix is to remove the byte swapping.

Presumably it's there for a reason?

> >  > >  A simple solution would be to change all of cpu_XX functions to
> >  > >  get a 64 bit address. This is a lot of churn, if we do this
> >  > >  anyway we should also pass length to callbacks, this way
> >  > >  rwhandler will get very small or go away completely.
> >  >
> >  > It's not too much effort to keep the target_phys_addr_t type.
> >
> >
> > I don't understand - target_phys_addr_t is different for different
> >  targets to we will need to recompile the code for each target.
> >  What am I missing?
> 
> target_phys_addr_t is 64 bit on a 64 bit host, on a 32 bit host it's
> size will be either 64 or 32 bits depending on the target. So the
> files are compiled once on 64 bit host, twice on 32 bit host if both
> 32 and 64 bits targets are selected.

How about just making it 64 bit unconditionally?
How much do we save by using a 32 bit address value?

-- 
MST