Re: [Qemu-devel] [PATCH 28/32] ppc: Avoid double translation for lvx/lvxl/stvx/stvxl

[Richard Henderson]

On 07/27/2016 03:51 AM, Benjamin Herrenschmidt wrote:
> -    tcg_gen_andi_tl(EA, EA, ~0xf);                                            \
> -    /* We only need to swap high and low halves. gen_qemu_ld64 does necessary \
> -       64-bit byteswap already. */                                            \
> -    if (ctx->le_mode) {                                                       \
> -        gen_qemu_ld64(ctx, cpu_avrl[rD(ctx->opcode)], EA);                    \
> -        tcg_gen_addi_tl(EA, EA, 8);                                           \
> -        gen_qemu_ld64(ctx, cpu_avrh[rD(ctx->opcode)], EA);                    \
> -    } else {                                                                  \
> -        gen_qemu_ld64(ctx, cpu_avrh[rD(ctx->opcode)], EA);                    \
> -        tcg_gen_addi_tl(EA, EA, 8);                                           \
> -        gen_qemu_ld64(ctx, cpu_avrl[rD(ctx->opcode)], EA);                    \
> -    }                                                                         \
> +    gen_helper_lvx(cpu_env, t0, EA);                                          \

This, I'm not so keen on.

(1) The helper, since it writes to registers controlled by tcg, must be 
described to clobber all registers.  Which will noticeably increase memory 
traffic to ENV.  For instance, you won't be able to hold the guest register 
holding the address in a host register across the call.

(2) We're going to have to teach tcg about 16-byte data types soon anyway, for 
the proper emulation of 16-byte atomic operations.


r~