[v5,11/36] tcg/tci: Add TCG_TARGET_CALL_{RET,ARG}_I128

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                | 19 +++++++++++++++++
 tcg/tci.c                | 44 ++++++++++++++++++++--------------------
 tcg/tci/tcg-target.c.inc | 10 ++++-----
 4 files changed, 49 insertions(+), 27 deletions(-)

Richard Henderson <richard.henderson@linaro.org> writes:

> 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                | 19 +++++++++++++++++
>  tcg/tci.c                | 44 ++++++++++++++++++++--------------------
>  tcg/tci/tcg-target.c.inc | 10 ++++-----
>  4 files changed, 49 insertions(+), 27 deletions(-)
>
> diff --git a/tcg/tci/tcg-target.h b/tcg/tci/tcg-target.h
> index 1414ab4d5b..7140a76a73 100644
> --- a/tcg/tci/tcg-target.h
> +++ b/tcg/tci/tcg-target.h
> @@ -160,10 +160,13 @@ typedef enum {
>  #if TCG_TARGET_REG_BITS == 32
>  # define TCG_TARGET_CALL_ARG_I32        TCG_CALL_ARG_EVEN
>  # define TCG_TARGET_CALL_ARG_I64        TCG_CALL_ARG_EVEN
> +# define TCG_TARGET_CALL_ARG_I128       TCG_CALL_ARG_EVEN
>  #else
>  # define TCG_TARGET_CALL_ARG_I32        TCG_CALL_ARG_NORMAL
>  # define TCG_TARGET_CALL_ARG_I64        TCG_CALL_ARG_NORMAL
> +# define TCG_TARGET_CALL_ARG_I128       TCG_CALL_ARG_NORMAL
>  #endif
> +#define TCG_TARGET_CALL_RET_I128        TCG_CALL_RET_NORMAL
>  
>  #define HAVE_TCG_QEMU_TB_EXEC
>  #define TCG_TARGET_NEED_POOL_LABELS
> diff --git a/tcg/tcg.c b/tcg/tcg.c
> index 084e3c3a54..4c43fd28ba 100644
> --- a/tcg/tcg.c
> +++ b/tcg/tcg.c
> @@ -570,6 +570,22 @@ static GHashTable *helper_table;
>  #ifdef CONFIG_TCG_INTERPRETER
>  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] = {
> +        &ffi_type_uint64,
> +        &ffi_type_uint64,
> +        NULL
> +    };
> +    static ffi_type ffi_type_i128 = {
> +        .size = 16,
> +        .alignment = __alignof__(Int128),
> +        .type = FFI_TYPE_STRUCT,
> +        .elements = ffi_type_i128_elements,
> +    };
> +
>      switch (argmask) {
>      case dh_typecode_void:
>          return &ffi_type_void;
> @@ -583,6 +599,8 @@ static ffi_type *typecode_to_ffi(int argmask)
>          return &ffi_type_sint64;
>      case dh_typecode_ptr:
>          return &ffi_type_pointer;
> +    case dh_typecode_i128:
> +        return &ffi_type_i128;
>      }
>      g_assert_not_reached();
>  }
> @@ -613,6 +631,7 @@ static void init_ffi_layouts(void)
>          /* Ignoring the return type, find the last non-zero field. */
>          nargs = 32 - clz32(typemask >> 3);
>          nargs = DIV_ROUND_UP(nargs, 3);
> +        assert(nargs <= MAX_CALL_IARGS);
>  
>          ca = g_malloc0(sizeof(*ca) + nargs * sizeof(ffi_type *));
>          ca->cif.rtype = typecode_to_ffi(typemask & 7);
> diff --git a/tcg/tci.c b/tcg/tci.c
> index eeccdde8bc..022fe9d0f8 100644
> --- a/tcg/tci.c
> +++ b/tcg/tci.c
> @@ -470,12 +470,9 @@ uintptr_t QEMU_DISABLE_CFI tcg_qemu_tb_exec(CPUArchState *env,
>      tcg_target_ulong regs[TCG_TARGET_NB_REGS];
>      uint64_t stack[(TCG_STATIC_CALL_ARGS_SIZE + TCG_STATIC_FRAME_SIZE)
>                     / sizeof(uint64_t)];
> -    void *call_slots[TCG_STATIC_CALL_ARGS_SIZE / sizeof(uint64_t)];
>  
>      regs[TCG_AREG0] = (tcg_target_ulong)env;
>      regs[TCG_REG_CALL_STACK] = (uintptr_t)stack;
> -    /* Other call_slots entries initialized at first use (see below). */
> -    call_slots[0] = NULL;
>      tci_assert(tb_ptr);
>  
>      for (;;) {
> @@ -498,26 +495,26 @@ 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_CALL_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;
> +                for (i = s = 0; i < n; ++i) {

Why the pre-increment for ++i here? It doesn't make a difference surely?

> +                    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);
>              }
>  
>              switch (len) {
> @@ -542,6 +539,9 @@ uintptr_t QEMU_DISABLE_CFI tcg_qemu_tb_exec(CPUArchState *env,
>                   */
>                  memcpy(&regs[TCG_REG_R0], stack, 8);
>                  break;
> +            case 3: /* Int128 */
> +                memcpy(&regs[TCG_REG_R0], stack, 16);
> +                break;
>              default:
>                  g_assert_not_reached();
>              }
> diff --git a/tcg/tci/tcg-target.c.inc b/tcg/tci/tcg-target.c.inc
> index e3b0ff303f..c1d34d7bd1 100644
> --- a/tcg/tci/tcg-target.c.inc
> +++ b/tcg/tci/tcg-target.c.inc
> @@ -203,7 +203,7 @@ static const int tcg_target_call_iarg_regs[] = { };
>  static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind kind, int slot)
>  {
>      tcg_debug_assert(kind == TCG_CALL_RET_NORMAL);
> -    tcg_debug_assert(slot >= 0 && slot < 64 / TCG_TARGET_REG_BITS);
> +    tcg_debug_assert(slot >= 0 && slot < 128 / TCG_TARGET_REG_BITS);
>      return TCG_REG_R0 + slot;
>  }
>  
> @@ -573,11 +573,11 @@ static void tcg_out_call(TCGContext *s, const tcg_insn_unit *func,
>  
>      if (cif->rtype == &ffi_type_void) {
>          which = 0;
> -    } else if (cif->rtype->size == 4) {
> -        which = 1;
>      } else {
> -        tcg_debug_assert(cif->rtype->size == 8);
> -        which = 2;
> +        tcg_debug_assert(cif->rtype->size == 4 ||
> +                         cif->rtype->size == 8 ||
> +                         cif->rtype->size == 16);
> +        which = ctz32(cif->rtype->size) - 1;
>      }
>      new_pool_l2(s, 20, s->code_ptr, 0, (uintptr_t)func, (uintptr_t)cif);
>      insn = deposit32(insn, 0, 8, INDEX_op_call);

Otherwise:

Reviewed-by: Alex Bennée <alex.bennee@linaro.org>

On 1/27/23 04:00, Alex Bennée wrote:
>> +                for (i = s = 0; i < n; ++i) {
> 
> Why the pre-increment for ++i here? It doesn't make a difference surely?

Why not?  It doesn't make a difference.

(Semi-reflexive preference driven by c++, where post-inc operator might (or might not) be 
less efficient, needing to return a copy of the class with the "old" value.  But if you 
don't know, or don't care about the old value, use the pre-inc form.)


r~