Re: [Qemu-devel] [PATCH 2/8] target/arm: Use generic vector infrastructure for aa64 add/sub/logic

["Alex Bennée" <alex.bennee@linaro.org>]

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

> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>  target/arm/translate-a64.c | 137 ++++++++++++++++++++++++++++-----------------
>  1 file changed, 87 insertions(+), 50 deletions(-)
>
> diff --git a/target/arm/translate-a64.c b/target/arm/translate-a64.c
> index 2200e25be0..025354f983 100644
> --- a/target/arm/translate-a64.c
> +++ b/target/arm/translate-a64.c
> @@ -21,6 +21,7 @@
>  #include "cpu.h"
>  #include "exec/exec-all.h"
>  #include "tcg-op.h"
> +#include "tcg-op-gvec.h"
>  #include "qemu/log.h"
>  #include "arm_ldst.h"
>  #include "translate.h"
> @@ -82,6 +83,7 @@ typedef void NeonGenTwoDoubleOPFn(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_ptr);
>  typedef void NeonGenOneOpFn(TCGv_i64, TCGv_i64);
>  typedef void CryptoTwoOpEnvFn(TCGv_ptr, TCGv_i32, TCGv_i32);
>  typedef void CryptoThreeOpEnvFn(TCGv_ptr, TCGv_i32, TCGv_i32, TCGv_i32);
> +typedef void GVecGenTwoFn(uint32_t, uint32_t, uint32_t, uint32_t, uint32_t);
>
>  /* initialize TCG globals.  */
>  void a64_translate_init(void)
> @@ -537,6 +539,21 @@ static inline int vec_reg_offset(DisasContext *s, int regno,
>      return offs;
>  }
>
> +/* Return the offset info CPUARMState of the "whole" vector register Qn.  */
> +static inline int vec_full_reg_offset(DisasContext *s, int regno)
> +{
> +    assert_fp_access_checked(s);
> +    return offsetof(CPUARMState, vfp.regs[regno * 2]);
> +}
> +
> +/* Return the byte size of the "whole" vector register, VL / 8.  */
> +static inline int vec_full_reg_size(DisasContext *s)
> +{
> +    /* FIXME SVE: We should put the composite ZCR_EL* value into tb->flags.
> +       In the meantime this is just the AdvSIMD length of 128.  */
> +    return 128 / 8;
> +}
> +
>  /* Return the offset into CPUARMState of a slice (from
>   * the least significant end) of FP register Qn (ie
>   * Dn, Sn, Hn or Bn).
> @@ -9042,11 +9059,38 @@ static void disas_simd_3same_logic(DisasContext *s, uint32_t insn)
>      bool is_q = extract32(insn, 30, 1);
>      TCGv_i64 tcg_op1, tcg_op2, tcg_res[2];
>      int pass;
> +    GVecGenTwoFn *gvec_op;
>
>      if (!fp_access_check(s)) {
>          return;
>      }
>
> +    switch (size + 4 * is_u) {

Hmm I find this switch a little too magical. I mean I can see that the
encoding abuses size for the final opcode when I look at the manual but
it reads badly.

> +    case 0: /* AND */
> +        gvec_op = tcg_gen_gvec_and8;
> +        goto do_gvec;
> +    case 1: /* BIC */
> +        gvec_op = tcg_gen_gvec_andc8;
> +        goto do_gvec;
> +    case 2: /* ORR */
> +        gvec_op = tcg_gen_gvec_or8;
> +        goto do_gvec;
> +    case 3: /* ORN */
> +        gvec_op = tcg_gen_gvec_orc8;
> +        goto do_gvec;
> +    case 4: /* EOR */
> +        gvec_op = tcg_gen_gvec_xor8;
> +        goto do_gvec;
> +    do_gvec:
> +        gvec_op(vec_full_reg_offset(s, rd),
> +                vec_full_reg_offset(s, rn),
> +                vec_full_reg_offset(s, rm),
> +                is_q ? 16 : 8, vec_full_reg_size(s));
> +        return;

No default case (although I guess we just fall through). What's wrong
with just having a !is_u test with gvec_op = tbl[size] and skipping all
the goto stuff?

> +    }
> +
> +    /* Note that we've now eliminated all !is_u.  */
> +
>      tcg_op1 = tcg_temp_new_i64();
>      tcg_op2 = tcg_temp_new_i64();
>      tcg_res[0] = tcg_temp_new_i64();
> @@ -9056,47 +9100,27 @@ static void disas_simd_3same_logic(DisasContext *s, uint32_t insn)
>          read_vec_element(s, tcg_op1, rn, pass, MO_64);
>          read_vec_element(s, tcg_op2, rm, pass, MO_64);
>
> -        if (!is_u) {
> -            switch (size) {
> -            case 0: /* AND */
> -                tcg_gen_and_i64(tcg_res[pass], tcg_op1, tcg_op2);
> -                break;
> -            case 1: /* BIC */
> -                tcg_gen_andc_i64(tcg_res[pass], tcg_op1, tcg_op2);
> -                break;
> -            case 2: /* ORR */
> -                tcg_gen_or_i64(tcg_res[pass], tcg_op1, tcg_op2);
> -                break;
> -            case 3: /* ORN */
> -                tcg_gen_orc_i64(tcg_res[pass], tcg_op1, tcg_op2);
> -                break;
> -            }
> -        } else {
> -            if (size != 0) {
> -                /* B* ops need res loaded to operate on */
> -                read_vec_element(s, tcg_res[pass], rd, pass, MO_64);
> -            }
> +        /* B* ops need res loaded to operate on */
> +        read_vec_element(s, tcg_res[pass], rd, pass, MO_64);
>
> -            switch (size) {
> -            case 0: /* EOR */
> -                tcg_gen_xor_i64(tcg_res[pass], tcg_op1, tcg_op2);
> -                break;
> -            case 1: /* BSL bitwise select */
> -                tcg_gen_xor_i64(tcg_op1, tcg_op1, tcg_op2);
> -                tcg_gen_and_i64(tcg_op1, tcg_op1, tcg_res[pass]);
> -                tcg_gen_xor_i64(tcg_res[pass], tcg_op2, tcg_op1);
> -                break;
> -            case 2: /* BIT, bitwise insert if true */
> -                tcg_gen_xor_i64(tcg_op1, tcg_op1, tcg_res[pass]);
> -                tcg_gen_and_i64(tcg_op1, tcg_op1, tcg_op2);
> -                tcg_gen_xor_i64(tcg_res[pass], tcg_res[pass], tcg_op1);
> -                break;
> -            case 3: /* BIF, bitwise insert if false */
> -                tcg_gen_xor_i64(tcg_op1, tcg_op1, tcg_res[pass]);
> -                tcg_gen_andc_i64(tcg_op1, tcg_op1, tcg_op2);
> -                tcg_gen_xor_i64(tcg_res[pass], tcg_res[pass], tcg_op1);
> -                break;
> -            }
> +        switch (size) {
> +        case 1: /* BSL bitwise select */
> +            tcg_gen_xor_i64(tcg_op1, tcg_op1, tcg_op2);
> +            tcg_gen_and_i64(tcg_op1, tcg_op1, tcg_res[pass]);
> +            tcg_gen_xor_i64(tcg_res[pass], tcg_op2, tcg_op1);
> +            break;
> +        case 2: /* BIT, bitwise insert if true */
> +            tcg_gen_xor_i64(tcg_op1, tcg_op1, tcg_res[pass]);
> +            tcg_gen_and_i64(tcg_op1, tcg_op1, tcg_op2);
> +            tcg_gen_xor_i64(tcg_res[pass], tcg_res[pass], tcg_op1);
> +            break;
> +        case 3: /* BIF, bitwise insert if false */
> +            tcg_gen_xor_i64(tcg_op1, tcg_op1, tcg_res[pass]);
> +            tcg_gen_andc_i64(tcg_op1, tcg_op1, tcg_op2);
> +            tcg_gen_xor_i64(tcg_res[pass], tcg_res[pass], tcg_op1);
> +            break;
> +        default:
> +            g_assert_not_reached();
>          }
>      }
>
> @@ -9370,6 +9394,7 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
>      int rn = extract32(insn, 5, 5);
>      int rd = extract32(insn, 0, 5);
>      int pass;
> +    GVecGenTwoFn *gvec_op;
>
>      switch (opcode) {
>      case 0x13: /* MUL, PMUL */
> @@ -9409,6 +9434,28 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
>          return;
>      }
>
> +    switch (opcode) {
> +    case 0x10: /* ADD, SUB */
> +        {
> +            static GVecGenTwoFn * const fns[4][2] = {
> +                { tcg_gen_gvec_add8, tcg_gen_gvec_sub8 },
> +                { tcg_gen_gvec_add16, tcg_gen_gvec_sub16 },
> +                { tcg_gen_gvec_add32, tcg_gen_gvec_sub32 },
> +                { tcg_gen_gvec_add64, tcg_gen_gvec_sub64 },
> +            };
> +            gvec_op = fns[size][u];
> +            goto do_gvec;
> +        }
> +        break;
> +
> +    do_gvec:
> +        gvec_op(vec_full_reg_offset(s, rd),
> +                vec_full_reg_offset(s, rn),
> +                vec_full_reg_offset(s, rm),
> +                is_q ? 16 : 8, vec_full_reg_size(s));
> +        return;
> +    }
> +
>      if (size == 3) {
>          assert(is_q);
>          for (pass = 0; pass < 2; pass++) {
> @@ -9581,16 +9628,6 @@ static void disas_simd_3same_int(DisasContext *s, uint32_t insn)
>                  genfn = fns[size][u];
>                  break;
>              }
> -            case 0x10: /* ADD, SUB */
> -            {
> -                static NeonGenTwoOpFn * const fns[3][2] = {
> -                    { gen_helper_neon_add_u8, gen_helper_neon_sub_u8 },
> -                    { gen_helper_neon_add_u16, gen_helper_neon_sub_u16 },
> -                    { tcg_gen_add_i32, tcg_gen_sub_i32 },
> -                };
> -                genfn = fns[size][u];
> -                break;
> -            }
>              case 0x11: /* CMTST, CMEQ */
>              {
>                  static NeonGenTwoOpFn * const fns[3][2] = {

Other than the comments on the switch the rest looks good to me.

--
Alex Bennée