Hi Yongbok,
On 11/05/2015 14:15, Yongbok Kim wrote:
> Hi
> I have implemented this to have a flag which isn't that nice.
>
> The thing is that the fact misaligned accesses of MSA LD/ST should be allowed
> in R5 cores
> while all other instructions are not allowed.
> Therefore it is required which types of instruction is triggering the
> misaligned accesses.
>
> Initially I tried to fetch the instructions from the
> mips_cpu_do_unaligned_access() callback,
> but if in certain case that the LD/ST address and PC are having same TLB
> indexes it goes wrong.
>
> I also tried to increase mmu_idx to avoid this problem but that requires
> anyway a flag as it is not
> able to pass mmu_idx to cpu_{ld,st}XX_XXX(). (cpu_{ld,st}XX_XXX() are calling
> cpu_mmu_index() to get mmu_idx).
>
> I could use host address directly via {ld,st}xx_p() but then mmio will be
> left alone to be solved.
> Perhaps another flag for the only case of R5 + MSA + MMIO.
>
> I might able to change all the generic load/store macros such as
> cpu_ldst_template.h and
> softmmu_template.h to pass the misalignment information.
> However that would be a huge work impacting all the architectures.
>
> Do you have any other thought or suggestion for this? Or this flag would be
> the necessary evil?
I haven't reviewed this patch yet, but have you considered using always
byte-by-byte accesses for misaligned MSA loads/stores? The flag wouldn't
be required and also I suspect that we would benefit from the fast path.
Thanks,
Leon
