Re: [Qemu-devel] [PATCH 7/7] tcg-i386: Perform tail call to qemu_ret_ld*_mmu

[Paolo Bonzini]

Il 27/08/2013 23:46, Richard Henderson ha scritto:
> This does require the fast path always load to the function return
> value register, but apparently the loaded value usually needs to be
> spilled back to its memory slot anyway so the change in register
> does not really change much.

Even for something like

    mov (%rdi), %rax
    add (%r8), %rax

?  Memory operands should avoid the need to spill anything.

Is this change really an advantage considering the additional icache
footprint of the new helpers?

Paolo

> Signed-off-by: Richard Henderson <address@hidden>
> ---
>  tcg/i386/tcg-target.c | 107 
> ++++++++++++++++++--------------------------------
>  1 file changed, 39 insertions(+), 68 deletions(-)
> 
> diff --git a/tcg/i386/tcg-target.c b/tcg/i386/tcg-target.c
> index 5aee0fa..b1d05b8 100644
> --- a/tcg/i386/tcg-target.c
> +++ b/tcg/i386/tcg-target.c
> @@ -1025,11 +1025,20 @@ static void tcg_out_jmp(TCGContext *s, 
> tcg_target_long dest)
>  /* helper signature: helper_ret_ld_mmu(CPUState *env, target_ulong addr,
>   *                                     int mmu_idx, uintptr_t ra)
>   */
> -static const void * const qemu_ld_helpers[4] = {
> +static const void * const qemu_ld_helpers[8] = {
>      helper_ret_ldub_mmu,
>      helper_ret_lduw_mmu,
>      helper_ret_ldul_mmu,
>      helper_ret_ldq_mmu,
> +
> +    helper_ret_ldsb_mmu,
> +    helper_ret_ldsw_mmu,
> +#if TCG_TARGET_REG_BITS == 64
> +    helper_ret_ldsl_mmu,
> +#else
> +    helper_ret_ldul_mmu,
> +#endif
> +    helper_ret_ldq_mmu,
>  };
>  
>  /* helper signature: helper_ret_st_mmu(CPUState *env, target_ulong addr,
> @@ -1473,9 +1482,8 @@ static void add_qemu_ldst_label(TCGContext *s,
>  static void tcg_out_qemu_ld_slow_path(TCGContext *s, TCGLabelQemuLdst *l)
>  {
>      int opc = l->opc;
> -    int s_bits = opc & 3;
> -    TCGReg data_reg;
>      uint8_t **label_ptr = &l->label_ptr[0];
> +    TCGReg retaddr;
>  
>      /* resolve label address */
>      *(uint32_t *)label_ptr[0] = (uint32_t)(s->code_ptr - label_ptr[0] - 4);
> @@ -1500,58 +1508,21 @@ static void tcg_out_qemu_ld_slow_path(TCGContext *s, 
> TCGLabelQemuLdst *l)
>          tcg_out_sti(s, TCG_TYPE_I32, TCG_REG_ESP, ofs, l->mem_index);
>          ofs += 4;
>  
> -        tcg_out_sti(s, TCG_TYPE_I32, TCG_REG_ESP, ofs, (uintptr_t)l->raddr);
> +        retaddr = TCG_REG_EAX;
> +        tcg_out_movi(s, TCG_TYPE_I32, retaddr, (uintptr_t)l->raddr);
> +        tcg_out_st(s, TCG_TYPE_I32, retaddr, TCG_REG_ESP, ofs);
>      } else {
>          tcg_out_mov(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[0], 
> TCG_AREG0);
>          /* The second argument is already loaded with addrlo.  */
>          tcg_out_movi(s, TCG_TYPE_I32, tcg_target_call_iarg_regs[2],
>                       l->mem_index);
> -        tcg_out_movi(s, TCG_TYPE_PTR, tcg_target_call_iarg_regs[3],
> -                     (uintptr_t)l->raddr);
> -    }
> -
> -    tcg_out_calli(s, (tcg_target_long)qemu_ld_helpers[s_bits]);
> -
> -    data_reg = l->datalo_reg;
> -    switch(opc) {
> -    case 0 | 4:
> -        tcg_out_ext8s(s, data_reg, TCG_REG_EAX, P_REXW);
> -        break;
> -    case 1 | 4:
> -        tcg_out_ext16s(s, data_reg, TCG_REG_EAX, P_REXW);
> -        break;
> -    case 0:
> -        tcg_out_ext8u(s, data_reg, TCG_REG_EAX);
> -        break;
> -    case 1:
> -        tcg_out_ext16u(s, data_reg, TCG_REG_EAX);
> -        break;
> -    case 2:
> -        tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
> -        break;
> -#if TCG_TARGET_REG_BITS == 64
> -    case 2 | 4:
> -        tcg_out_ext32s(s, data_reg, TCG_REG_EAX);
> -        break;
> -#endif
> -    case 3:
> -        if (TCG_TARGET_REG_BITS == 64) {
> -            tcg_out_mov(s, TCG_TYPE_I64, data_reg, TCG_REG_RAX);
> -        } else if (data_reg == TCG_REG_EDX) {
> -            /* xchg %edx, %eax */
> -            tcg_out_opc(s, OPC_XCHG_ax_r32 + TCG_REG_EDX, 0, 0, 0);
> -            tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EAX);
> -        } else {
> -            tcg_out_mov(s, TCG_TYPE_I32, data_reg, TCG_REG_EAX);
> -            tcg_out_mov(s, TCG_TYPE_I32, l->datahi_reg, TCG_REG_EDX);
> -        }
> -        break;
> -    default:
> -        tcg_abort();
> +        retaddr = tcg_target_call_iarg_regs[3];
> +        tcg_out_movi(s, TCG_TYPE_PTR, retaddr, (uintptr_t)l->raddr);
>      }
>  
> -    /* Jump to the code corresponding to next IR of qemu_st */
> -    tcg_out_jmp(s, (tcg_target_long)l->raddr);
> +    /* "Tail call" to the helper, with the return address back inline.  */
> +    tcg_out_push(s, retaddr);
> +    tcg_out_jmp(s, (tcg_target_long)qemu_ld_helpers[opc]);
>  }
>  
>  /*
> @@ -2125,38 +2096,38 @@ static const TCGTargetOpDef x86_op_defs[] = {
>  #endif
>  
>  #if TCG_TARGET_REG_BITS == 64
> -    { INDEX_op_qemu_ld8u, { "r", "L" } },
> -    { INDEX_op_qemu_ld8s, { "r", "L" } },
> -    { INDEX_op_qemu_ld16u, { "r", "L" } },
> -    { INDEX_op_qemu_ld16s, { "r", "L" } },
> -    { INDEX_op_qemu_ld32, { "r", "L" } },
> -    { INDEX_op_qemu_ld32u, { "r", "L" } },
> -    { INDEX_op_qemu_ld32s, { "r", "L" } },
> -    { INDEX_op_qemu_ld64, { "r", "L" } },
> +    { INDEX_op_qemu_ld8u, { "a", "L" } },
> +    { INDEX_op_qemu_ld8s, { "a", "L" } },
> +    { INDEX_op_qemu_ld16u, { "a", "L" } },
> +    { INDEX_op_qemu_ld16s, { "a", "L" } },
> +    { INDEX_op_qemu_ld32, { "a", "L" } },
> +    { INDEX_op_qemu_ld32u, { "a", "L" } },
> +    { INDEX_op_qemu_ld32s, { "a", "L" } },
> +    { INDEX_op_qemu_ld64, { "a", "L" } },
>  
>      { INDEX_op_qemu_st8, { "L", "L" } },
>      { INDEX_op_qemu_st16, { "L", "L" } },
>      { INDEX_op_qemu_st32, { "L", "L" } },
>      { INDEX_op_qemu_st64, { "L", "L" } },
>  #elif TARGET_LONG_BITS <= TCG_TARGET_REG_BITS
> -    { INDEX_op_qemu_ld8u, { "r", "L" } },
> -    { INDEX_op_qemu_ld8s, { "r", "L" } },
> -    { INDEX_op_qemu_ld16u, { "r", "L" } },
> -    { INDEX_op_qemu_ld16s, { "r", "L" } },
> -    { INDEX_op_qemu_ld32, { "r", "L" } },
> -    { INDEX_op_qemu_ld64, { "r", "r", "L" } },
> +    { INDEX_op_qemu_ld8u, { "a", "L" } },
> +    { INDEX_op_qemu_ld8s, { "a", "L" } },
> +    { INDEX_op_qemu_ld16u, { "a", "L" } },
> +    { INDEX_op_qemu_ld16s, { "a", "L" } },
> +    { INDEX_op_qemu_ld32, { "a", "L" } },
> +    { INDEX_op_qemu_ld64, { "a", "d", "L" } },
>  
>      { INDEX_op_qemu_st8, { "cb", "L" } },
>      { INDEX_op_qemu_st16, { "L", "L" } },
>      { INDEX_op_qemu_st32, { "L", "L" } },
>      { INDEX_op_qemu_st64, { "L", "L", "L" } },
>  #else
> -    { INDEX_op_qemu_ld8u, { "r", "L", "L" } },
> -    { INDEX_op_qemu_ld8s, { "r", "L", "L" } },
> -    { INDEX_op_qemu_ld16u, { "r", "L", "L" } },
> -    { INDEX_op_qemu_ld16s, { "r", "L", "L" } },
> -    { INDEX_op_qemu_ld32, { "r", "L", "L" } },
> -    { INDEX_op_qemu_ld64, { "r", "r", "L", "L" } },
> +    { INDEX_op_qemu_ld8u, { "a", "L", "L" } },
> +    { INDEX_op_qemu_ld8s, { "a", "L", "L" } },
> +    { INDEX_op_qemu_ld16u, { "a", "L", "L" } },
> +    { INDEX_op_qemu_ld16s, { "a", "L", "L" } },
> +    { INDEX_op_qemu_ld32, { "a", "L", "L" } },
> +    { INDEX_op_qemu_ld64, { "a", "d", "L", "L" } },
>  
>      { INDEX_op_qemu_st8, { "cb", "L", "L" } },
>      { INDEX_op_qemu_st16, { "L", "L", "L" } },
>