Re: [Qemu-devel] [PATCH v2 07/17] RISC-V: add vector extension atomic instructions

[Richard Henderson]

On 9/11/19 2:25 AM, liuzhiwei wrote:
> +            case 64:
> +                if (vector_elem_mask(env, vm, width, lmul, i)) {
> +                    int64_t tmp;
> +                    idx    = (target_long)env->vfp.vreg[src2].s64[j];
> +                    addr   = idx + env->gpr[rs1];
> +
> +#ifdef CONFIG_SOFTMMU
> +                    tmp = (int64_t)(int32_t)helper_atomic_xchgl_le(env, addr,
> +                        env->vfp.vreg[src3].s64[j],
> +                        make_memop_idx(memop & ~MO_SIGN, mem_idx));
> +#else
> +                    tmp = (int64_t)(int32_t)helper_atomic_xchgl_le(env, addr,
> +                        env->vfp.vreg[src3].s64[j]);
> +#endif
> +                    if (wd) {
> +                        env->vfp.vreg[src3].s64[j] = tmp;
> +                    }
> +                    env->vfp.vstart++;
> +                }
> +                break;

This will not link if !defined(CONFIG_ATOMIC64).

That's pretty rare these days, admittedly.  I think you'd need to compile for
ppc32 or mips32 (or riscv32!) host to see this.  You can force this condition
for i686 host with --extra-cflags='-march=i486', just to see if you've got it
right.

There should be two different versions of this helper: one that performs actual
atomic operations, as above, and a second that performs the same operation with
non-atomic operations.

The version of the helper that you call should be based on the translation time
setting of "tb_cflags(s->base.tb) & CF_PARALLEL":  If PARALLEL is set, call the
atomic helper otherwise the non-atomic helper.

If you arrive at a situation in which the host cannot handle any atomic
operation, then you must raise the EXCP_ATOMIC exception.  This will halt all
other cpus and run one instruction on this cpu while holding the exclusive lock.

If you cannot detect this condition any earlier than here at runtime, use
cpu_loop_exit_atomic(), but you must do so before altering any cpu state.
However, as per my comments for normal loads, you should be able to detect this
condition at translation time and call gen_helper_exit_atomic().


r~