[OpenBIOS] [PATCH, RFC 6/6] ppc: Fix stack setup

[Andreas Färber]

Am 11.10.2010 um 11:16 schrieb Alexander Graf:

> Andreas Färber wrote:
>> Am 11.10.2010 um 04:53 schrieb Segher Boessenkool:
>>
>>>> +    mfpvr    r14                /* check if ppc64 */
>>>> +    li        r15, 4
>>>> +    srw        r14, r14, r15    /* >> 4 */
>>>
>>> srwi 14,14,4
>>>
>>> but you actually want to shift 16, not 4.
>>
>> Oops!
>>
>>>> +    cmplwi    r14, 0x33
>>>> +    blt        1f
>>>> +    cmplwi    r14, 0x7033
>>>> +    bge        1f
>>>
>>> This is too simplistic.
>>
>> Are you saying the original is_ppc64() function I transformed into
>> assembler code is wrong, too?
>>
>> static int is_ppc64(void)
>> {
>>    unsigned int pvr;
>>    asm volatile("mfspr %0, 0x11f" : "=r" (pvr) );
>>
>>    return ((pvr >= 0x330000) && (pvr < 0x70330000));
>> }
>
> In here we don't know if SF=1 because we forcefully disable it always.

Ah! You mean this?

li r0, 0
mtmsr r0

That would mean I don't need a 64-bit RTAS unless AIX sets SF=1. :)

>>> Here's something that detects
>>> whether the CPU is currently in 64-bit mode, without
>>> using 64-bit-only instructions:
>>>
>>> lis 3,0x4000
>>> add 3,3,3
>>> addc 3,3,3
>>> adde 3,3,3
>>> cmplwi 3,0
>>> beq 64bit
>>> bne 32bit
>>
>> Thanks! I'll give it a try.
>>
>>> Anyway, can't you always align to 128 bytes, and create
>>> one frame for the 64-bit ABI?  Or do you really want to
>>> save those few bytes...
>>
>> As for "those few bytes", I guess I mixed up bits and hexadecimal
>> digits once again. An example:
>>
>> -m 1024
>>
>> RAM       0..0x40000000
>> OpenBIOS   @ 0x3ff00000
>> Hash table @ 0x3fef0000 vs. 0x3fe00000
>>
>> We'd be wasting 960 KiB. If that's acceptable, it'll simplify the  
>> code
>> a little of course.
>
> I'd personally say go for simplicity.

On second thoughts, unfortunately I don't think 32-bit CPUs can handle  
clrrdi, so we'd still need the distinction here since we're  
subtracting from RAM size.

Or better, we'd need a distinction at the original setup of r1, so  
that it fits into the low 32 or 31 bits in the first place. Currently  
everything over -m 2048 would overlap with ISA and at some point PCI  
and fw_cfg. Memory would probably have to become non-contiguous then,  
breaking some more assumptions...

Andreas