Re: sh4 pmap bug? (Re: port-dreamcast/34243)

[email protected] wrote:

> Right, but 5.2.6 says:
> 
>     "...  The case of a single process attaching the same shared
>     memory region two or more times within its own address space is
>     the same as described in the previous chapters (see Section
>     3.3.6)."

Ah, okay, but anyway such aliases could only happens
if shared pages have the different virtual indexes.
i.e. maybe we should fix ld.elf_so to consider virtual
indexes (colors) on mappings.

The SH4 programming manual also says:

>> When address translation information whereby a number of
>> 4-kbyte page UTLB entries are translated into the same
>> physical address is recorded in the UTLB, ensure that
>> the VPN [13:12] values are the same.

BTW, it also says:

>> When multiple items of address translation information use
>> the same physical memory to provide for future SuperH RISC
>> engine family expansion, ensure that the VPN [20:10] values
>> are the same.

It looks Renesas consider that they would release SH4 family with
virtual cache which has 2^9 indexes... (Umm, unlikely)
---
Izumi Tsutsui