[OpenBIOS] [PATCH 4/4] ppc: avoid runtime relocations

[Segher Boessenkool]

>>>> Because the sizes of cells and pointers do not match on PPC64,
>>>
>>> So fix that.  Most OF implementations on 64-bit PowerPC use 64-bit
>>> cells; you pretty much need it for things like device access, esp.
>>> if running in real mode.
>>>
>>> The Forth standard does not define pointers that do not fit in a
>>> cell; OF-specific words like rl@ take a cell-sized "address" as
>>> well.
>>
>> I'm fairly sure that Andreas is aware of this (for example, one of
>> the patches I posted earlier for SPARC32 fixes the generation of
>> memory properties but will truncate the maximum address to 32-bits
>> on PPC64 since the code assumes that a single cell can hold a
>> virtual address).
>>
>> Andreas - are you happy for these two patchsets to be applied? We'd
>> like to hope they won't break PPC64 too badly until you can find
>> some time to continue work on it.
>
> My (our) requirement is that the guest-visible parts remain compatible
> with Apple OpenFirmware (the hardware we emulate today), and I'd
> prefer to share a common code base for ppc and ppc64.

You can do both with a 64-bit OF; you can provide a 32-bit client
interface.

Apple's OF on the G5 machines is a 32-bit OF, and they do a lot of
very much non-standard things (including a lot that are not correct).

It is possible that the Linux kernel will not work with certain
64-bit device tree things, because of bugs in the kernel.  I
recommend you fix such bugs, or avoid them.  But you really want
to have #address-cells = #size-cells = 2 in the root node; without
it, for example, you cannot have 4GB or more of memory.

> My understanding is that if we enlarge the internal cell size as
> suggested by Segher, encode-int and other Forth words will start
> widening their output, too, so that property values will start to
> differ, no?

No.  encode-int will encode a 32-bit integer, always.  All of the
device tree data is completely independent of word size and endianness,
etc.

"#address-cells" in the device tree actually means "number of 32-bit
integers per address".  It's historical...


Segher