Re: [Qemu-devel] [PATCH] CONFIG_MMU_MAP powerpc host support

[Piotras]

Hi!

The problem can be avoided by replacing some of asm with C. 
The gcc seems to generate equally-efficient code in this case. 
I don't have a ppc machine to test it, but something like this may 
work:

RES_TYPE glue(glue(ld, USUFFIX), MEMSUFFIX)(void *ptr)
{
    target_ulong addr;
    register target_phys_addr_t add asm ("r3");
    RES_TYPE val;

    addr = (target_ulong)ptr;
    add = env->mmu_map[CPU_MEM_INDEX].add_read[addr >> TARGET_PAGE_BITS];
    asm volatile (
#if (DATA_SIZE == 1)
        "lbzx   %0, %1, %2\n"
#elif (DATA_SIZE == 2) && defined(BSWAP_NEEDED)
        "lhbrx  %0, %1, %2\n"
#elif (DATA_SIZE == 2)
        "lhzx   %0, %1, %2\n"
#elif (DATA_SIZE == 4) && defined(BSWAP_NEEDED)
        "lwbrx  %0, %1, %2\n"
#elif (DATA_SIZE == 4)
        "lwzx   %0, %1, %2\n"
#else
#error unsupported size
#endif
        : "=r" (val)
        : "r" (addr),
          "r" (add)
        : "memory");

    return val;
}

You will have to modify lds and st as well (make sure to replace 
"add_read" with "add_write" for st). It is critical to use a specific 
register as %2 (Magnus decided to use r3).


Regards,

Piotrek

On Tue, 28 Dec 2004 12:18:57 +0100, Pierre d'Herbemont <address@hidden> wrote:
> 
> [...]
> 
> > Actually I'm a bit unsure about the displacement syntax but
> > I know that register names are accepted by GNU binutils and
> > cctools because I've written quite a lot of AltiVec code
> > which works on both Linux and MacOSX without any tricks.
> 
> Hum, seems to be a good trick... Google search returns the "-mregnames"
> option. But it seems that it doesn't deal with ha16() or lo16(), so
> we'll have to deal with it. However I think it is still better than
> duplicating the PowerPC code.
> 
> Pierre.