Re: [PATCH 06/38] target/riscv: SIMD 16-bit Shift Instructions

[LIU Zhiwei <zhiwei_liu@c-sky.com>]

On 2021/3/17 3:54, Alistair Francis wrote:
> On Mon, Mar 15, 2021 at 10:40 PM LIU Zhiwei <zhiwei_liu@c-sky.com> wrote:
>>
>>
>> On 2021/3/16 5:25, Alistair Francis wrote:
>>> On Fri, Feb 12, 2021 at 10:16 AM LIU Zhiwei <zhiwei_liu@c-sky.com> wrote:
>>>> Signed-off-by: LIU Zhiwei <zhiwei_liu@c-sky.com>
>>>> ---
>>>>    target/riscv/helper.h                   |   9 ++
>>>>    target/riscv/insn32.decode              |  17 ++++
>>>>    target/riscv/insn_trans/trans_rvp.c.inc | 115 ++++++++++++++++++++++++
>>>>    target/riscv/packed_helper.c            | 104 +++++++++++++++++++++
>>>>    4 files changed, 245 insertions(+)
>>>>
>>>> diff --git a/target/riscv/helper.h b/target/riscv/helper.h
>>>> index a69a6b4e84..20bf400ac2 100644
>>>> --- a/target/riscv/helper.h
>>>> +++ b/target/riscv/helper.h
>>>> @@ -1184,3 +1184,12 @@ DEF_HELPER_3(rsub8, tl, env, tl, tl)
>>>>    DEF_HELPER_3(ursub8, tl, env, tl, tl)
>>>>    DEF_HELPER_3(ksub8, tl, env, tl, tl)
>>>>    DEF_HELPER_3(uksub8, tl, env, tl, tl)
>>>> +
>>>> +DEF_HELPER_3(sra16, tl, env, tl, tl)
>>>> +DEF_HELPER_3(sra16_u, tl, env, tl, tl)
>>>> +DEF_HELPER_3(srl16, tl, env, tl, tl)
>>>> +DEF_HELPER_3(srl16_u, tl, env, tl, tl)
>>>> +DEF_HELPER_3(sll16, tl, env, tl, tl)
>>>> +DEF_HELPER_3(ksll16, tl, env, tl, tl)
>>>> +DEF_HELPER_3(kslra16, tl, env, tl, tl)
>>>> +DEF_HELPER_3(kslra16_u, tl, env, tl, tl)
>>>> diff --git a/target/riscv/insn32.decode b/target/riscv/insn32.decode
>>>> index 358dd1fa10..6f053bfeb7 100644
>>>> --- a/target/riscv/insn32.decode
>>>> +++ b/target/riscv/insn32.decode
>>>> @@ -23,6 +23,7 @@
>>>>    %rd        7:5
>>>>
>>>>    %sh10    20:10
>>>> +%sh4    20:4
>>>>    %csr    20:12
>>>>    %rm     12:3
>>>>    %nf     29:3                     !function=ex_plus_1
>>>> @@ -59,6 +60,7 @@
>>>>    @j       ....................      ..... ....... &j      imm=%imm_j          %rd
>>>>
>>>>    @sh      ......  ...... .....  ... ..... ....... &shift  shamt=%sh10      %rs1 %rd
>>>> +@sh4     ......  ...... .....  ... ..... ....... &shift  shamt=%sh4      %rs1 %rd
>>>>    @csr     ............   .....  ... ..... .......               %csr     %rs1 %rd
>>>>
>>>>    @atom_ld ..... aq:1 rl:1 ..... ........ ..... ....... &atomic rs2=0     %rs1 %rd
>>>> @@ -635,3 +637,18 @@ rsub8      0000101  ..... ..... 000 ..... 1111111 @r
>>>>    ursub8     0010101  ..... ..... 000 ..... 1111111 @r
>>>>    ksub8      0001101  ..... ..... 000 ..... 1111111 @r
>>>>    uksub8     0011101  ..... ..... 000 ..... 1111111 @r
>>>> +
>>>> +sra16      0101000  ..... ..... 000 ..... 1111111 @r
>>>> +sra16_u    0110000  ..... ..... 000 ..... 1111111 @r
>>>> +srai16     0111000  0.... ..... 000 ..... 1111111 @sh4
>>>> +srai16_u   0111000  1.... ..... 000 ..... 1111111 @sh4
>>>> +srl16      0101001  ..... ..... 000 ..... 1111111 @r
>>>> +srl16_u    0110001  ..... ..... 000 ..... 1111111 @r
>>>> +srli16     0111001  0.... ..... 000 ..... 1111111 @sh4
>>>> +srli16_u   0111001  1.... ..... 000 ..... 1111111 @sh4
>>>> +sll16      0101010  ..... ..... 000 ..... 1111111 @r
>>>> +slli16     0111010  0.... ..... 000 ..... 1111111 @sh4
>>>> +ksll16     0110010  ..... ..... 000 ..... 1111111 @r
>>>> +kslli16    0111010  1.... ..... 000 ..... 1111111 @sh4
>>>> +kslra16    0101011  ..... ..... 000 ..... 1111111 @r
>>>> +kslra16_u  0110011  ..... ..... 000 ..... 1111111 @r
>>>> diff --git a/target/riscv/insn_trans/trans_rvp.c.inc b/target/riscv/insn_trans/trans_rvp.c.inc
>>>> index 109f560ec9..848edab7e5 100644
>>>> --- a/target/riscv/insn_trans/trans_rvp.c.inc
>>>> +++ b/target/riscv/insn_trans/trans_rvp.c.inc
>>>> @@ -238,3 +238,118 @@ GEN_RVP_R_OOL(rsub8);
>>>>    GEN_RVP_R_OOL(ursub8);
>>>>    GEN_RVP_R_OOL(ksub8);
>>>>    GEN_RVP_R_OOL(uksub8);
>>>> +
>>>> +/* 16-bit Shift Instructions */
>>>> +static bool rvp_shift_ool(DisasContext *ctx, arg_r *a,
>>>> +                          gen_helper_rvp_r *fn, target_ulong mask)
>>>> +{
>>>> +    TCGv src1, src2, dst;
>>>> +
>>>> +    src1 = tcg_temp_new();
>>>> +    src2 = tcg_temp_new();
>>>> +    dst = tcg_temp_new();
>>>> +
>>>> +    gen_get_gpr(src1, a->rs1);
>>>> +    gen_get_gpr(src2, a->rs2);
>>>> +    tcg_gen_andi_tl(src2, src2, mask);
>>>> +
>>>> +    fn(dst, cpu_env, src1, src2);
>>>> +    gen_set_gpr(a->rd, dst);
>>>> +
>>>> +    tcg_temp_free(src1);
>>>> +    tcg_temp_free(src2);
>>>> +    tcg_temp_free(dst);
>>>> +    return true;
>>>> +}
>>>> +
>>>> +typedef void GenGvecShift(unsigned, uint32_t, uint32_t, TCGv_i32,
>>>> +                          uint32_t, uint32_t);
>>>> +static inline bool
>>>> +rvp_shift(DisasContext *ctx, arg_r *a, uint8_t vece,
>>>> +          GenGvecShift *f64, gen_helper_rvp_r *fn,
>>>> +          uint8_t mask)
>>>> +{
>>>> +    if (!has_ext(ctx, RVP)) {
>>>> +        return false;
>>>> +    }
>>>> +
>>>> +#ifdef TARGET_RISCV64
>>> Hmm....
>>>
>>> I don't want to add any more #defines on the RISC-V xlen. We are
>>> trying to make the QEMU RISC-V implementation xlen independent.
>> I noticed the change, but was not quite clear about the benefit of it.
>>
>> Could you give a brief explanation?
> Yep, there are a few reasons for it.
>
> AFAIK every QEMU platform except RISC-V allows the 64-bit binary to
> run the 32-bit guests. So for example in the ARM QEMU builds I can use
> qemu-system-aarch64 to run 32-bit ARMv7 guests. This is a step towards
> allowing us to do the same with RISC-V.
Get it. By explicitly specifying a 32bit CPU option in command line, 
qemu-system-riscv64
can run a 32-bit guest application. Is it right?
>
> It's also a step towards allowing fixed XLEN CPUS to run. For example
> 4 64-bit application CPUs and a single 32-bit power management CPU can
> all run together.
>
> Also XLEN in RISC-V is configurable. A 64-bit Hypervisor can have
> 32-bit XLEN guests according to the spec. This is a step towards
> allowing that as well.
Really interesting points.

I have not used QEMU in this way. Are these features ready now?
>>> Can you use `riscv_cpu_is_32bit(env)` instead, here are everywhere
>>> else you add a #define TARGET... ?
>> Sure, I think there are two ways.
>>
>> 1) Get env from the current_cpu, then call riscv_cpu_is_32bit(env).
>>
>> It's some strange,  because I can't find current_cpu reference from many
>> archs.
>>
>> I don't know whether it has side effects.
>>
>> 2)  Add a similar function cpu_is_32bit(DisasContext *ctx).
> This is probably a better option, but I'm open to either way if you
> have a strong preference.
I will add a patch in this way in v2. Thanks a lot.

Zhiwei
>
> Alistair
>
>> In this way, the type of  misa field  in struct DisasContext should be
>> target_ulong.
>> Currently, the type of misa filed is uint32_t.
>>
>> Do you think which one is better? Thanks very much.
>>
>> Zhiwei
>>> Alistair
>>>
>>>> +    if (a->rd && a->rs1 && a->rs2) {
>>>> +        TCGv_i32 shift = tcg_temp_new_i32();
>>>> +        tcg_gen_extrl_i64_i32(shift, cpu_gpr[a->rs2]);
>>>> +        tcg_gen_andi_i32(shift, shift, mask);
>>>> +        f64(vece, offsetof(CPURISCVState, gpr[a->rd]),
>>>> +            offsetof(CPURISCVState, gpr[a->rs1]),
>>>> +            shift, 8, 8);
>>>> +        tcg_temp_free_i32(shift);
>>>> +        return true;
>>>> +    }
>>>> +#endif
>>>> +    return rvp_shift_ool(ctx, a, fn, mask);
>>>> +}
>>>> +
>>>> +#define GEN_RVP_SHIFT(NAME, GVEC, VECE)                     \
>>>> +static bool trans_##NAME(DisasContext *s, arg_r *a)         \
>>>> +{                                                           \
>>>> +    return rvp_shift(s, a, VECE, GVEC, gen_helper_##NAME,   \
>>>> +                     (8 << VECE) - 1);                      \
>>>> +}
>>>> +
>>>> +GEN_RVP_SHIFT(sra16, tcg_gen_gvec_sars, 1);
>>>> +GEN_RVP_SHIFT(srl16, tcg_gen_gvec_shrs, 1);
>>>> +GEN_RVP_SHIFT(sll16, tcg_gen_gvec_shls, 1);
>>>> +GEN_RVP_R_OOL(sra16_u);
>>>> +GEN_RVP_R_OOL(srl16_u);
>>>> +GEN_RVP_R_OOL(ksll16);
>>>> +GEN_RVP_R_OOL(kslra16);
>>>> +GEN_RVP_R_OOL(kslra16_u);
>>>> +
>>>> +static bool rvp_shifti_ool(DisasContext *ctx, arg_shift *a,
>>>> +                           gen_helper_rvp_r *fn)
>>>> +{
>>>> +    TCGv src1, dst, shift;
>>>> +
>>>> +    src1 = tcg_temp_new();
>>>> +    dst = tcg_temp_new();
>>>> +
>>>> +    gen_get_gpr(src1, a->rs1);
>>>> +    shift = tcg_const_tl(a->shamt);
>>>> +    fn(dst, cpu_env, src1, shift);
>>>> +    gen_set_gpr(a->rd, dst);
>>>> +
>>>> +    tcg_temp_free(src1);
>>>> +    tcg_temp_free(dst);
>>>> +    tcg_temp_free(shift);
>>>> +    return true;
>>>> +}
>>>> +
>>>> +static inline bool
>>>> +rvp_shifti(DisasContext *ctx, arg_shift *a,
>>>> +           void (* f64)(TCGv_i64, TCGv_i64, int64_t),
>>>> +           gen_helper_rvp_r *fn)
>>>> +{
>>>> +    if (!has_ext(ctx, RVP)) {
>>>> +        return false;
>>>> +    }
>>>> +
>>>> +#ifdef TARGET_RISCV64
>>>> +    if (a->rd && a->rs1 && f64) {
>>>> +        f64(cpu_gpr[a->rd], cpu_gpr[a->rs1], a->shamt);
>>>> +        return true;
>>>> +    }
>>>> +#endif
>>>> +    return rvp_shifti_ool(ctx, a, fn);
>>>> +}
>>>> +
>>>> +#define GEN_RVP_SHIFTI(NAME, OP, GVEC)                   \
>>>> +static bool trans_##NAME(DisasContext *s, arg_shift *a)  \
>>>> +{                                                        \
>>>> +    return rvp_shifti(s, a, GVEC, gen_helper_##OP);      \
>>>> +}
>>>> +
>>>> +GEN_RVP_SHIFTI(srai16, sra16, tcg_gen_vec_sar16i_i64);
>>>> +GEN_RVP_SHIFTI(srli16, srl16, tcg_gen_vec_shr16i_i64);
>>>> +GEN_RVP_SHIFTI(slli16, sll16, tcg_gen_vec_shl16i_i64);
>>>> +GEN_RVP_SHIFTI(srai16_u, sra16_u, NULL);
>>>> +GEN_RVP_SHIFTI(srli16_u, srl16_u, NULL);
>>>> +GEN_RVP_SHIFTI(kslli16, ksll16, NULL);
>>>> diff --git a/target/riscv/packed_helper.c b/target/riscv/packed_helper.c
>>>> index 62db072204..7e31c2fe46 100644
>>>> --- a/target/riscv/packed_helper.c
>>>> +++ b/target/riscv/packed_helper.c
>>>> @@ -425,3 +425,107 @@ static inline void do_uksub8(CPURISCVState *env, void *vd, void *va,
>>>>    }
>>>>
>>>>    RVPR(uksub8, 1, 1);
>>>> +
>>>> +/* 16-bit Shift Instructions */
>>>> +static inline void do_sra16(CPURISCVState *env, void *vd, void *va,
>>>> +                            void *vb, uint8_t i)
>>>> +{
>>>> +    int16_t *d = vd, *a = va;
>>>> +    uint8_t shift = *(uint8_t *)vb & 0xf;
>>>> +    d[i] = a[i] >> shift;
>>>> +}
>>>> +
>>>> +RVPR(sra16, 1, 2);
>>>> +
>>>> +static inline void do_srl16(CPURISCVState *env, void *vd, void *va,
>>>> +                            void *vb, uint8_t i)
>>>> +{
>>>> +    uint16_t *d = vd, *a = va;
>>>> +    uint8_t shift = *(uint8_t *)vb & 0xf;
>>>> +    d[i] = a[i] >> shift;
>>>> +}
>>>> +
>>>> +RVPR(srl16, 1, 2);
>>>> +
>>>> +static inline void do_sll16(CPURISCVState *env, void *vd, void *va,
>>>> +                            void *vb, uint8_t i)
>>>> +{
>>>> +    uint16_t *d = vd, *a = va;
>>>> +    uint8_t shift = *(uint8_t *)vb & 0xf;
>>>> +    d[i] = a[i] << shift;
>>>> +}
>>>> +
>>>> +RVPR(sll16, 1, 2);
>>>> +
>>>> +static inline void do_sra16_u(CPURISCVState *env, void *vd, void *va,
>>>> +                              void *vb, uint8_t i)
>>>> +{
>>>> +    int16_t *d = vd, *a = va;
>>>> +    uint8_t shift = *(uint8_t *)vb & 0xf;
>>>> +
>>>> +    d[i] = vssra16(env, 0, a[i], shift);
>>>> +}
>>>> +
>>>> +RVPR(sra16_u, 1, 2);
>>>> +
>>>> +static inline void do_srl16_u(CPURISCVState *env, void *vd, void *va,
>>>> +                              void *vb, uint8_t i)
>>>> +{
>>>> +    uint16_t *d = vd, *a = va;
>>>> +    uint8_t shift = *(uint8_t *)vb & 0xf;
>>>> +
>>>> +    d[i] = vssrl16(env, 0, a[i], shift);
>>>> +}
>>>> +
>>>> +RVPR(srl16_u, 1, 2);
>>>> +
>>>> +static inline void do_ksll16(CPURISCVState *env, void *vd, void *va,
>>>> +                             void *vb, uint8_t i)
>>>> +{
>>>> +    int16_t *d = vd, *a = va, result;
>>>> +    uint8_t shift = *(uint8_t *)vb & 0xf;
>>>> +
>>>> +    result = a[i] << shift;
>>>> +    if (shift > (clrsb32(a[i]) - 16)) {
>>>> +        env->vxsat = 0x1;
>>>> +        d[i] = (a[i] & INT16_MIN) ? INT16_MIN : INT16_MAX;
>>>> +    } else {
>>>> +        d[i] = result;
>>>> +    }
>>>> +}
>>>> +
>>>> +RVPR(ksll16, 1, 2);
>>>> +
>>>> +static inline void do_kslra16(CPURISCVState *env, void *vd, void *va,
>>>> +                              void *vb, uint8_t i)
>>>> +{
>>>> +    int16_t *d = vd, *a = va;
>>>> +    int32_t shift = sextract32((*(target_ulong *)vb), 0, 5);
>>>> +
>>>> +    if (shift >= 0) {
>>>> +        do_ksll16(env, vd, va, vb, i);
>>>> +    } else {
>>>> +        shift = -shift;
>>>> +        shift = (shift == 16) ? 15 : shift;
>>>> +        d[i] = a[i] >> shift;
>>>> +    }
>>>> +}
>>>> +
>>>> +RVPR(kslra16, 1, 2);
>>>> +
>>>> +static inline void do_kslra16_u(CPURISCVState *env, void *vd, void *va,
>>>> +                                void *vb, uint8_t i)
>>>> +{
>>>> +    int16_t *d = vd, *a = va;
>>>> +    int32_t shift = sextract32((*(uint32_t *)vb), 0, 5);
>>>> +
>>>> +    if (shift >= 0) {
>>>> +        do_ksll16(env, vd, va, vb, i);
>>>> +    } else {
>>>> +        shift = -shift;
>>>> +        shift = (shift == 16) ? 15 : shift;
>>>> +        d[i] = vssra16(env, 0, a[i], shift);
>>>> +    }
>>>> +}
>>>> +
>>>> +RVPR(kslra16_u, 1, 2);
>>>> --
>>>> 2.17.1
>>>>