Re: [RFC PATCH v2 2/2] cputlb: implement load_helper_unaligned() for unaligned loads

[Richard Henderson <richard.henderson@linaro.org>]

On 6/9/21 7:10 AM, Philippe Mathieu-Daudé wrote:
> +    oi = make_memop_idx(MO_UB, mmu_idx);
> +    if (memop_big_endian(op)) {
> +        for (i = 0; i < size; ++i) {
> +            /* Big-endian load.  */
> +            uint8_t val8 = helper_ret_ldub_mmu(env, addr + i, oi, retaddr);
> +            val |= val8 << (((size - 1) * 8) - (i * 8));
> +        }
> +    } else {
> +        for (i = 0; i < size; ++i) {
> +            /* Little-endian load.  */
> +            uint8_t val8 = helper_ret_ldub_mmu(env, addr + i, oi, retaddr);
> +            val |= val8 << (i * 8);
> +        }
> +    }

This doesn't quite work.  You can't just call helper_ret_ldub_mmu, as the other 
option is full_ldub_code.  So, at present you've broken unaligned code loads.

We also need noinline markup for clang, like we do for helper_ret_stb_mmu. 
I've no proof of that, but it certainly makes sense to record how we expect the 
inline loop to be resolved.

Finally, you have to use uint64_t for val8, otherwise the shift fails for size 
== 8.

I'll fix these up and see how things go.


r~