Re: [Qemu-devel] [PATCH] target-i386: Fix mulx for identical target

[Richard Henderson]

On 11/17/2015 04:40 AM, Max Reitz wrote:
> -                    tcg_gen_mulu2_i64(cpu_regs[s->vex_v], cpu_regs[reg],
> -                                      cpu_T[0], cpu_regs[R_EDX]);
> +                    tcg_gen_op3_i64(INDEX_op_mul_i64, cpu_regs[s->vex_v],
> +                                    cpu_T[0], cpu_regs[R_EDX]);
> +                    if (TCG_TARGET_HAS_muluh_i64) {
> +                        tcg_gen_op3_i64(INDEX_op_muluh_i64, cpu_regs[reg],
> +                                        cpu_T[0], cpu_regs[R_EDX]);
> +                    } else {
> +                        gen_helper_muluh_i64(cpu_regs[reg],
> +                                             cpu_T[0], cpu_regs[R_EDX]);
> +                    }

This is not the correct fix.

I do see the operand write order present in the pseudo-code of the intel 
manual.  So that part of the diagnosis is fine.

But the correct solution is to put the outputs of tcg_gen_mulu2 into 
temporaries (cpu_T[0] and cpu_T[1]) and then afterward write back the results 
into the architectual registers in the correct order.


r~