Re: [Qemu-devel] [PATCH 5/6] target-ppc: add lxvb16x and lxvh8x

[Richard Henderson]

On 08/10/2016 02:51 PM, Nikunj A Dadhania wrote:
> I can fix the BE case using following but not sure if that will be
> correct !
>
>         tcg_gen_qemu_ld_i64(xth, EA, ctx->mem_idx, MO_Q);
>         gen_helper_bswap32x2(xth, xth);
>         tcg_gen_addi_tl(EA, EA, 8);
>         tcg_gen_qemu_ld_i64(xtl, EA, ctx->mem_idx, MO_Q);
>         gen_helper_bswap32x2(xtl, xtl);

This cannot be correct, because it assumes a host-dependent byte ordering.  You 
should be able to see different results depending on a BE or LE host.


> __vector uint32_t vrt32;
> uint32_t rb32[4] = {0x00010203, 0x20212223, 0x30313233, 0x40414243};
> asm("lxvw4x %x0, 0, %1 \n\t" \
>      : "=ws"(vrt32) : "r"(&rb32));
> printf("VRT32 = "); vec_put_u32(vrt32);
>
> Result On LE:
> VRT32 = 40414243 30313233 20212223 00010203
>
> Result On BE:
> VRT32 = 03020100 23222120 33323130 43424140

Did you really recompile the test program for BE and LE, or did you just use a 
different command-line switch to qemu with the same executable image?  I can't 
see how you could possibly get these results for BE.


r~