On Thu, 2 Feb 2023 at 15:13, Philipp Tomsich <philipp.tomsich@xxxxxxxx> wrote:
>
> On Thu, 2 Feb 2023 at 13:42, Lawrence Hunter
> <lawrence.hunter@xxxxxxxxxxxxxxx> wrote:
> >
> > From: Dickon Hood <dickon.hood@xxxxxxxxxxxxxxx>
> >
> > Add an initial implementation of the vrol.* and vror.* instructions,
> > with mappings between the RISC-V instructions and their internal TCG
> > accelerated implmentations.
> >
> > There are some missing ror helpers, so I've bodged it by converting them
> > to rols.
> >
> > Co-authored-by: Kiran Ostrolenk <kiran.ostrolenk@xxxxxxxxxxxxxxx>
> > Signed-off-by: Kiran Ostrolenk <kiran.ostrolenk@xxxxxxxxxxxxxxx>
> > Signed-off-by: Dickon Hood <dickon.hood@xxxxxxxxxxxxxxx>
> > ---
> > target/riscv/helper.h | 20 ++++++++
> > target/riscv/insn32.decode | 6 +++
> > target/riscv/insn_trans/trans_rvzvkb.c.inc | 20 ++++++++
> > target/riscv/vcrypto_helper.c | 58 ++++++++++++++++++++++
> > target/riscv/vector_helper.c | 45 -----------------
> > target/riscv/vector_internals.h | 52 +++++++++++++++++++
> > 6 files changed, 156 insertions(+), 45 deletions(-)
> >
> > diff --git a/target/riscv/helper.h b/target/riscv/helper.h
> > index 32f1179e29..e5b6b3360f 100644
> > --- a/target/riscv/helper.h
> > +++ b/target/riscv/helper.h
> > @@ -1142,3 +1142,23 @@ DEF_HELPER_6(vclmul_vv, void, ptr, ptr, ptr, ptr, env, i32)
> > DEF_HELPER_6(vclmul_vx, void, ptr, ptr, tl, ptr, env, i32)
> > DEF_HELPER_6(vclmulh_vv, void, ptr, ptr, ptr, ptr, env, i32)
> > DEF_HELPER_6(vclmulh_vx, void, ptr, ptr, tl, ptr, env, i32)
> > +
> > +DEF_HELPER_6(vror_vv_b, void, ptr, ptr, ptr, ptr, env, i32)
> > +DEF_HELPER_6(vror_vv_h, void, ptr, ptr, ptr, ptr, env, i32)
> > +DEF_HELPER_6(vror_vv_w, void, ptr, ptr, ptr, ptr, env, i32)
> > +DEF_HELPER_6(vror_vv_d, void, ptr, ptr, ptr, ptr, env, i32)
> > +
> > +DEF_HELPER_6(vror_vx_b, void, ptr, ptr, tl, ptr, env, i32)
> > +DEF_HELPER_6(vror_vx_h, void, ptr, ptr, tl, ptr, env, i32)
> > +DEF_HELPER_6(vror_vx_w, void, ptr, ptr, tl, ptr, env, i32)
> > +DEF_HELPER_6(vror_vx_d, void, ptr, ptr, tl, ptr, env, i32)
> > +
> > +DEF_HELPER_6(vrol_vv_b, void, ptr, ptr, ptr, ptr, env, i32)
> > +DEF_HELPER_6(vrol_vv_h, void, ptr, ptr, ptr, ptr, env, i32)
> > +DEF_HELPER_6(vrol_vv_w, void, ptr, ptr, ptr, ptr, env, i32)
> > +DEF_HELPER_6(vrol_vv_d, void, ptr, ptr, ptr, ptr, env, i32)
> > +
> > +DEF_HELPER_6(vrol_vx_b, void, ptr, ptr, tl, ptr, env, i32)
> > +DEF_HELPER_6(vrol_vx_h, void, ptr, ptr, tl, ptr, env, i32)
> > +DEF_HELPER_6(vrol_vx_w, void, ptr, ptr, tl, ptr, env, i32)
> > +DEF_HELPER_6(vrol_vx_d, void, ptr, ptr, tl, ptr, env, i32)
> > diff --git a/target/riscv/insn32.decode b/target/riscv/insn32.decode
> > index b4d88dd1cb..725f907ad1 100644
> > --- a/target/riscv/insn32.decode
> > +++ b/target/riscv/insn32.decode
> > @@ -896,3 +896,9 @@ vclmul_vv 001100 . ..... ..... 010 ..... 1010111 @r_vm
> > vclmul_vx 001100 . ..... ..... 110 ..... 1010111 @r_vm
> > vclmulh_vv 001101 . ..... ..... 010 ..... 1010111 @r_vm
> > vclmulh_vx 001101 . ..... ..... 110 ..... 1010111 @r_vm
> > +vrol_vv 010101 . ..... ..... 000 ..... 1010111 @r_vm
> > +vrol_vx 010101 . ..... ..... 100 ..... 1010111 @r_vm
> > +vror_vv 010100 . ..... ..... 000 ..... 1010111 @r_vm
> > +vror_vx 010100 . ..... ..... 100 ..... 1010111 @r_vm
> > +vror_vi 010100 . ..... ..... 011 ..... 1010111 @r_vm
> > +vror_vi2 010101 . ..... ..... 011 ..... 1010111 @r_vm
>
> We have only a single vror_vi instruction... and it has a 6bit immediate.
> There's no need to deviate from the spec. Just write it as follows:
> %imm_z6 26:1 15:5
> @r2_zimm6 ..... . vm:1 ..... ..... ... ..... ....... &rmrr %rs2
> rs1=%imm_z6 %rd
> vror_vi 01010. . ..... ..... 011 ..... 1010111 @r2_zimm6
>
> > diff --git a/target/riscv/insn_trans/trans_rvzvkb.c.inc b/target/riscv/insn_trans/trans_rvzvkb.c.inc
> > index 533141e559..d2a7a92d42 100644
> > --- a/target/riscv/insn_trans/trans_rvzvkb.c.inc
> > +++ b/target/riscv/insn_trans/trans_rvzvkb.c.inc
> > @@ -95,3 +95,23 @@ static bool vclmul_vx_check(DisasContext *s, arg_rmrr *a)
> >
> > GEN_VX_MASKED_TRANS(vclmul_vx, vclmul_vx_check)
> > GEN_VX_MASKED_TRANS(vclmulh_vx, vclmul_vx_check)
> > +
> > +GEN_OPIVV_TRANS(vror_vv, zvkb_vv_check)
> > +GEN_OPIVX_TRANS(vror_vx, zvkb_vx_check)
> > +GEN_OPIVV_TRANS(vrol_vv, zvkb_vv_check)
> > +GEN_OPIVX_TRANS(vrol_vx, zvkb_vx_check)
> > +
> > +GEN_OPIVI_TRANS(vror_vi, IMM_TRUNC_SEW, vror_vx, zvkb_vx_check)
>
> Please introduce a IMM_ZIMM6 that integrates with the extract_imm as follows:
> case IMM_ZIMM6:
> return extract64(imm, 0, 6);
>
> > +
> > +/*
> > + * Immediates are 5b long, and we need six for the rotate-immediate. The
> > + * decision has been taken to remove the vrol.vi instruction -- you can
> > + * emulate it with a ror, after all -- and use the bottom bit of the funct6
> > + * part of the opcode to encode the extra bit. I've chosen to implement it
> > + * like this because it's easy and reasonably clean.
> > + */
> > +static bool trans_vror_vi2(DisasContext *s, arg_rmrr *a)
> > +{
> > + a->rs1 += 32;
> > + return trans_vror_vi(s, a);
> > +}
>
> As discussed above, please handle this in a single trans_vror_vi:
> GEN_OPIVI_GVEC_TRANS_CHECK(vror_vi, IMM_ZIMM6, vror_vx, rotri, zvkb_check_vi)
On the second pass over these patches, here's how we can use gvec
support for both vror and vrol:
/* Synthesize a rotate-right from a negate(shift-amount) + rotate-left */
static void tcg_gen_gvec_rotrs(unsigned vece, uint32_t dofs, uint32_t aofs,
TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
{
TCGv_i32 tmp = tcg_temp_new_i32();
tcg_gen_neg_i32(tmp, shift);
tcg_gen_gvec_rotls(vece, dofs, aofs, tmp, oprsz, maxsz);
tcg_temp_free_i32(tmp);
}
/* vrol.v[vx] */
GEN_OPIVV_GVEC_TRANS_CHECK(vrol_vv, rotlv, zvkb_check_vv)
GEN_OPIVX_GVEC_SHIFT_TRANS_CHECK(vrol_vx, rotls, zvkb_check_vx)
/* vror.v[vxi] */
GEN_OPIVV_GVEC_TRANS_CHECK(vror_vv, rotrv, zvkb_check_vv)
GEN_OPIVX_GVEC_SHIFT_TRANS_CHECK(vror_vx, rotrs, zvkb_check_vx)
GEN_OPIVI_GVEC_TRANS_CHECK(vror_vi, IMM_ZIMM6, vror_vx, rotri, zvkb_check_vi)
> > diff --git a/target/riscv/vcrypto_helper.c b/target/riscv/vcrypto_helper.c
> > index 46e2e510c5..7ec75c5589 100644
> > --- a/target/riscv/vcrypto_helper.c
> > +++ b/target/riscv/vcrypto_helper.c
> > @@ -57,3 +57,61 @@ GEN_VEXT_VV(vclmul_vv, 8)
> > GEN_VEXT_VX(vclmul_vx, 8)
> > GEN_VEXT_VV(vclmulh_vv, 8)
> > GEN_VEXT_VX(vclmulh_vx, 8)
> > +
> > +/*
> > + * Looks a mess, but produces reasonable (aarch32) code on clang:
> > + * https://godbolt.org/z/jchjsTda8
> > + */
> > +#define DO_ROR(x, n) \
> > + ((x >> (n & ((sizeof(x) << 3) - 1))) | \
> > + (x << ((sizeof(x) << 3) - (n & ((sizeof(x) << 3) - 1)))))
> > +#define DO_ROL(x, n) \
> > + ((x << (n & ((sizeof(x) << 3) - 1))) | \
> > + (x >> ((sizeof(x) << 3) - (n & ((sizeof(x) << 3) - 1)))))
> > +
> > +RVVCALL(OPIVV2, vror_vv_b, OP_UUU_B, H1, H1, H1, DO_ROR)
> > +RVVCALL(OPIVV2, vror_vv_h, OP_UUU_H, H2, H2, H2, DO_ROR)
> > +RVVCALL(OPIVV2, vror_vv_w, OP_UUU_W, H4, H4, H4, DO_ROR)
> > +RVVCALL(OPIVV2, vror_vv_d, OP_UUU_D, H8, H8, H8, DO_ROR)
>
> Indeed, this is a mess: we already have ror8/16/32/64 available.
> Why not the following?
>
> /* vror.vv */
> GEN_VEXT_SHIFT_VV(vror_vv_b, uint8_t, uint8_t, H1, H1, ror8, 0x7)
> GEN_VEXT_SHIFT_VV(vror_vv_h, uint16_t, uint16_t, H2, H2, ror16, 0xf)
> GEN_VEXT_SHIFT_VV(vror_vv_w, uint32_t, uint32_t, H4, H4, ror32, 0x1f)
> GEN_VEXT_SHIFT_VV(vror_vv_d, uint64_t, uint64_t, H8, H8, ror64, 0x3f)
>
> > +GEN_VEXT_VV(vror_vv_b, 1)
> > +GEN_VEXT_VV(vror_vv_h, 2)
> > +GEN_VEXT_VV(vror_vv_w, 4)
> > +GEN_VEXT_VV(vror_vv_d, 8)
> > +
> > +/*
> > + * There's a missing tcg_gen_gvec_rotrs() helper function.
> > + */
> > +#define GEN_VEXT_VX_RTOL(NAME, ESZ) \
> > +void HELPER(NAME)(void *vd, void *v0, target_ulong s1, void *vs2, \
> > + CPURISCVState *env, uint32_t desc) \
> > +{ \
> > + do_vext_vx(vd, v0, (ESZ << 3) - s1, vs2, env, desc, do_##NAME, ESZ); \
> > +}
> > +
> > +/* DO_ROL because GEN_VEXT_VX_RTOL() converts from R to L */
> > +RVVCALL(OPIVX2, vror_vx_b, OP_UUU_B, H1, H1, DO_ROL)
> > +RVVCALL(OPIVX2, vror_vx_h, OP_UUU_H, H2, H2, DO_ROL)
> > +RVVCALL(OPIVX2, vror_vx_w, OP_UUU_W, H4, H4, DO_ROL)
> > +RVVCALL(OPIVX2, vror_vx_d, OP_UUU_D, H8, H8, DO_ROL)
>
> Same applies as above:
> /* vror.vx */
> GEN_VEXT_SHIFT_VX(vror_vx_b, uint8_t, uint8_t, H1, H1, ror8, 0x7)
> GEN_VEXT_SHIFT_VX(vror_vx_h, uint16_t, uint16_t, H2, H2, ror16, 0xf)
> GEN_VEXT_SHIFT_VX(vror_vx_w, uint32_t, uint32_t, H4, H4, ror32, 0x1f)
> GEN_VEXT_SHIFT_VX(vror_vx_d, uint64_t, uint64_t, H8, H8, ror64, 0x3f)
>
> > +GEN_VEXT_VX_RTOL(vror_vx_b, 1)
> > +GEN_VEXT_VX_RTOL(vror_vx_h, 2)
> > +GEN_VEXT_VX_RTOL(vror_vx_w, 4)
> > +GEN_VEXT_VX_RTOL(vror_vx_d, 8)
> > +
> > +RVVCALL(OPIVV2, vrol_vv_b, OP_UUU_B, H1, H1, H1, DO_ROL)
> > +RVVCALL(OPIVV2, vrol_vv_h, OP_UUU_H, H2, H2, H2, DO_ROL)
> > +RVVCALL(OPIVV2, vrol_vv_w, OP_UUU_W, H4, H4, H4, DO_ROL)
> > +RVVCALL(OPIVV2, vrol_vv_d, OP_UUU_D, H8, H8, H8, DO_ROL)
> > +GEN_VEXT_VV(vrol_vv_b, 1)
> > +GEN_VEXT_VV(vrol_vv_h, 2)
> > +GEN_VEXT_VV(vrol_vv_w, 4)
> > +GEN_VEXT_VV(vrol_vv_d, 8)
> > +
> > +RVVCALL(OPIVX2, vrol_vx_b, OP_UUU_B, H1, H1, DO_ROL)
> > +RVVCALL(OPIVX2, vrol_vx_h, OP_UUU_H, H2, H2, DO_ROL)
> > +RVVCALL(OPIVX2, vrol_vx_w, OP_UUU_W, H4, H4, DO_ROL)
> > +RVVCALL(OPIVX2, vrol_vx_d, OP_UUU_D, H8, H8, DO_ROL)
> > +GEN_VEXT_VX(vrol_vx_b, 1)
> > +GEN_VEXT_VX(vrol_vx_h, 2)
> > +GEN_VEXT_VX(vrol_vx_w, 4)
> > +GEN_VEXT_VX(vrol_vx_d, 8)
> > diff --git a/target/riscv/vector_helper.c b/target/riscv/vector_helper.c
> > index ab470092f6..ff7b03cbe3 100644
> > --- a/target/riscv/vector_helper.c
> > +++ b/target/riscv/vector_helper.c
> > @@ -76,26 +76,6 @@ target_ulong HELPER(vsetvl)(CPURISCVState *env, target_ulong s1,
> > return vl;
> > }
> >
> > -/*
> > - * Note that vector data is stored in host-endian 64-bit chunks,
> > - * so addressing units smaller than that needs a host-endian fixup.
> > - */
> > -#if HOST_BIG_ENDIAN
> > -#define H1(x) ((x) ^ 7)
> > -#define H1_2(x) ((x) ^ 6)
> > -#define H1_4(x) ((x) ^ 4)
> > -#define H2(x) ((x) ^ 3)
> > -#define H4(x) ((x) ^ 1)
> > -#define H8(x) ((x))
> > -#else
> > -#define H1(x) (x)
> > -#define H1_2(x) (x)
> > -#define H1_4(x) (x)
> > -#define H2(x) (x)
> > -#define H4(x) (x)
> > -#define H8(x) (x)
> > -#endif
> > -
> > /*
> > * Get the maximum number of elements can be operated.
> > *
> > @@ -683,18 +663,11 @@ GEN_VEXT_ST_WHOLE(vs8r_v, int8_t, ste_b)
> > *** Vector Integer Arithmetic Instructions
> > */
> >
> > -/* expand macro args before macro */
> > -#define RVVCALL(macro, ...) macro(__VA_ARGS__)
> > -
> > /* (TD, T1, T2, TX1, TX2) */
> > #define OP_SSS_B int8_t, int8_t, int8_t, int8_t, int8_t
> > #define OP_SSS_H int16_t, int16_t, int16_t, int16_t, int16_t
> > #define OP_SSS_W int32_t, int32_t, int32_t, int32_t, int32_t
> > #define OP_SSS_D int64_t, int64_t, int64_t, int64_t, int64_t
> > -#define OP_UUU_B uint8_t, uint8_t, uint8_t, uint8_t, uint8_t
> > -#define OP_UUU_H uint16_t, uint16_t, uint16_t, uint16_t, uint16_t
> > -#define OP_UUU_W uint32_t, uint32_t, uint32_t, uint32_t, uint32_t
> > -#define OP_UUU_D uint64_t, uint64_t, uint64_t, uint64_t, uint64_t
> > #define OP_SUS_B int8_t, uint8_t, int8_t, uint8_t, int8_t
> > #define OP_SUS_H int16_t, uint16_t, int16_t, uint16_t, int16_t
> > #define OP_SUS_W int32_t, uint32_t, int32_t, uint32_t, int32_t
> > @@ -718,14 +691,6 @@ GEN_VEXT_ST_WHOLE(vs8r_v, int8_t, ste_b)
> > #define NOP_UUU_H uint16_t, uint16_t, uint32_t, uint16_t, uint32_t
> > #define NOP_UUU_W uint32_t, uint32_t, uint64_t, uint32_t, uint64_t
> >
> > -
> > -#define OPIVV2(NAME, TD, T1, T2, TX1, TX2, HD, HS1, HS2, OP) \
> > -static void do_##NAME(void *vd, void *vs1, void *vs2, int i) \
> > -{ \
> > - TX1 s1 = *((T1 *)vs1 + HS1(i)); \
> > - TX2 s2 = *((T2 *)vs2 + HS2(i)); \
> > - *((TD *)vd + HD(i)) = OP(s2, s1); \
> > -}
> > #define DO_SUB(N, M) (N - M)
> > #define DO_RSUB(N, M) (M - N)
> >
> > @@ -747,16 +712,6 @@ GEN_VEXT_VV(vsub_vv_h, 2)
> > GEN_VEXT_VV(vsub_vv_w, 4)
> > GEN_VEXT_VV(vsub_vv_d, 8)
> >
> > -/*
> > - * (T1)s1 gives the real operator type.
> > - * (TX1)(T1)s1 expands the operator type of widen or narrow operations.
> > - */
> > -#define OPIVX2(NAME, TD, T1, T2, TX1, TX2, HD, HS2, OP) \
> > -static void do_##NAME(void *vd, target_long s1, void *vs2, int i) \
> > -{ \
> > - TX2 s2 = *((T2 *)vs2 + HS2(i)); \
> > - *((TD *)vd + HD(i)) = OP(s2, (TX1)(T1)s1); \
> > -}
> >
> > RVVCALL(OPIVX2, vadd_vx_b, OP_SSS_B, H1, H1, DO_ADD)
> > RVVCALL(OPIVX2, vadd_vx_h, OP_SSS_H, H2, H2, DO_ADD)
> > diff --git a/target/riscv/vector_internals.h b/target/riscv/vector_internals.h
> > index 49529d2379..a0fbac7bf3 100644
> > --- a/target/riscv/vector_internals.h
> > +++ b/target/riscv/vector_internals.h
> > @@ -28,6 +28,26 @@ static inline uint32_t vext_nf(uint32_t desc)
> > return FIELD_EX32(simd_data(desc), VDATA, NF);
> > }
> >
> > +/*
> > + * Note that vector data is stored in host-endian 64-bit chunks,
> > + * so addressing units smaller than that needs a host-endian fixup.
> > + */
> > +#if HOST_BIG_ENDIAN
> > +#define H1(x) ((x) ^ 7)
> > +#define H1_2(x) ((x) ^ 6)
> > +#define H1_4(x) ((x) ^ 4)
> > +#define H2(x) ((x) ^ 3)
> > +#define H4(x) ((x) ^ 1)
> > +#define H8(x) ((x))
> > +#else
> > +#define H1(x) (x)
> > +#define H1_2(x) (x)
> > +#define H1_4(x) (x)
> > +#define H2(x) (x)
> > +#define H4(x) (x)
> > +#define H8(x) (x)
> > +#endif
> > +
> > /*
> > * Encode LMUL to lmul as following:
> > * LMUL vlmul lmul
> > @@ -96,9 +116,30 @@ static inline uint32_t vext_get_total_elems(CPURISCVState *env, uint32_t desc,
> > void vext_set_elems_1s(void *base, uint32_t is_agnostic, uint32_t cnt,
> > uint32_t tot);
> >
> > +/*
> > + *** Vector Integer Arithmetic Instructions
> > + */
> > +
> > +/* expand macro args before macro */
> > +#define RVVCALL(macro, ...) macro(__VA_ARGS__)
> > +
> > +/* (TD, T1, T2, TX1, TX2) */
> > +#define OP_UUU_B uint8_t, uint8_t, uint8_t, uint8_t, uint8_t
> > +#define OP_UUU_H uint16_t, uint16_t, uint16_t, uint16_t, uint16_t
> > +#define OP_UUU_W uint32_t, uint32_t, uint32_t, uint32_t, uint32_t
> > +#define OP_UUU_D uint64_t, uint64_t, uint64_t, uint64_t, uint64_t
> > +
> > /* operation of two vector elements */
> > typedef void opivv2_fn(void *vd, void *vs1, void *vs2, int i);
> >
> > +#define OPIVV2(NAME, TD, T1, T2, TX1, TX2, HD, HS1, HS2, OP) \
> > +static void do_##NAME(void *vd, void *vs1, void *vs2, int i) \
> > +{ \
> > + TX1 s1 = *((T1 *)vs1 + HS1(i)); \
> > + TX2 s2 = *((T2 *)vs2 + HS2(i)); \
> > + *((TD *)vd + HD(i)) = OP(s2, s1); \
> > +}
> > +
> > void do_vext_vv(void *vd, void *v0, void *vs1, void *vs2,
> > CPURISCVState *env, uint32_t desc,
> > opivv2_fn *fn, uint32_t esz);
> > @@ -115,6 +156,17 @@ void HELPER(NAME)(void *vd, void *v0, void *vs1, \
> >
> > typedef void opivx2_fn(void *vd, target_long s1, void *vs2, int i);
> >
> > +/*
> > + * (T1)s1 gives the real operator type.
> > + * (TX1)(T1)s1 expands the operator type of widen or narrow operations.
> > + */
> > +#define OPIVX2(NAME, TD, T1, T2, TX1, TX2, HD, HS2, OP) \
> > +static void do_##NAME(void *vd, target_long s1, void *vs2, int i) \
> > +{ \
> > + TX2 s2 = *((T2 *)vs2 + HS2(i)); \
> > + *((TD *)vd + HD(i)) = OP(s2, (TX1)(T1)s1); \
> > +}
> > +
> > void do_vext_vx(void *vd, void *v0, target_long s1, void *vs2,
> > CPURISCVState *env, uint32_t desc,
> > opivx2_fn fn, uint32_t esz);
>
> Again: refactoring needs to go into a separate patch.
>
> > --
> > 2.39.1
> >
