Re: [PATCH v2 33/36] tcg/tci: Add TCG_TARGET_CALL_{RET,ARG}_I128

[Philippe Mathieu-Daudé]

On 21/10/22 09:15, Richard Henderson wrote:
> Fill in the parameters for libffi for Int128.
> Adjust the interpreter to allow for 16-byte return values.
> Adjust tcg_out_call to record the return value length.
>
> Call parameters are no longer all the same size, so we
> cannot reuse the same call_slots array for every function.
> Compute it each time now, but only fill in slots required
> for the call we're about to make.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>   tcg/tci/tcg-target.h     |  3 +++
>   tcg/tcg.c                | 18 ++++++++++++++++
>   tcg/tci.c                | 45 ++++++++++++++++++++--------------------
>   tcg/tci/tcg-target.c.inc |  8 +++----
>   4 files changed, 48 insertions(+), 26 deletions(-)

>   static ffi_type *typecode_to_ffi(int argmask)
>   {
> +    /*
> +     * libffi does not support __int128_t, so we have forced Int128
> +     * to use the structure definition instead of the builtin type.
> +     */
> +    static ffi_type *ffi_type_i128_elements[3] = {

static const.

> +        &ffi_type_uint64,
> +        &ffi_type_uint64,
> +        NULL
> +    };
> +    static ffi_type ffi_type_i128 = {

static const.

> +        .size = 16,
> +        .alignment = __alignof__(Int128),
> +        .type = FFI_TYPE_STRUCT,
> +        .elements = ffi_type_i128_elements,
> +    };

> @@ -499,26 +496,27 @@ uintptr_t QEMU_DISABLE_CFI tcg_qemu_tb_exec(CPUArchState 
> *env,
>   
>           switch (opc) {
>           case INDEX_op_call:
> -            /*
> -             * Set up the ffi_avalue array once, delayed until now
> -             * because many TB's do not make any calls. In tcg_gen_callN,
> -             * we arranged for every real argument to be "left-aligned"
> -             * in each 64-bit slot.
> -             */
> -            if (unlikely(call_slots[0] == NULL)) {
> -                for (int i = 0; i < ARRAY_SIZE(call_slots); ++i) {
> -                    call_slots[i] = &stack[i];
> -                }
> -            }
> -
> -            tci_args_nl(insn, tb_ptr, &len, &ptr);
> -
> -            /* Helper functions may need to access the "return address" */
> -            tci_tb_ptr = (uintptr_t)tb_ptr;
> -
>               {
> -                void **pptr = ptr;
> -                ffi_call(pptr[1], pptr[0], stack, call_slots);
> +                void *call_slots[MAX_OPC_PARAM_IARGS];
> +                ffi_cif *cif;
> +                void *func;
> +                unsigned i, s, n;
> +
> +                tci_args_nl(insn, tb_ptr, &len, &ptr);
> +                func = ((void **)ptr)[0];
> +                cif = ((void **)ptr)[1];
> +
> +                n = cif->nargs;
> +                tci_assert(n <= MAX_OPC_PARAM_IARGS);
> +                for (i = s = 0; i < n; ++i) {
> +                    ffi_type *t = cif->arg_types[i];
> +                    call_slots[i] = &stack[s];
> +                    s += DIV_ROUND_UP(t->size, 8);
> +                }
> +
> +                /* Helper functions may need to access the "return address" */
> +                tci_tb_ptr = (uintptr_t)tb_ptr;
> +                ffi_call(cif, func, stack, call_slots);
>               }

Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>