The difference between XTheadVector and RVV1.0 is same as the other patchs: 1. Different mask reg layout. 2. Different tail/masked elements process policy. 3. Simpler acceleration judgment logic.
Signed-off-by: Huang Tao <[email protected]> --- target/riscv/helper.h | 25 ++++ .../riscv/insn_trans/trans_xtheadvector.c.inc | 61 ++++++++-- target/riscv/xtheadvector_helper.c | 115 ++++++++++++++++++ 3 files changed, 192 insertions(+), 9 deletions(-) diff --git a/target/riscv/helper.h b/target/riscv/helper.h index 6599b2f2f5..77251af8c9 100644 --- a/target/riscv/helper.h +++ b/target/riscv/helper.h @@ -1600,3 +1600,28 @@ DEF_HELPER_6(th_vxor_vx_b, void, ptr, ptr, tl, ptr, env, i32) DEF_HELPER_6(th_vxor_vx_h, void, ptr, ptr, tl, ptr, env, i32) DEF_HELPER_6(th_vxor_vx_w, void, ptr, ptr, tl, ptr, env, i32) DEF_HELPER_6(th_vxor_vx_d, void, ptr, ptr, tl, ptr, env, i32) + +DEF_HELPER_6(th_vsll_vv_b, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsll_vv_h, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsll_vv_w, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsll_vv_d, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsrl_vv_b, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsrl_vv_h, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsrl_vv_w, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsrl_vv_d, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsra_vv_b, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsra_vv_h, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsra_vv_w, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsra_vv_d, void, ptr, ptr, ptr, ptr, env, i32) +DEF_HELPER_6(th_vsll_vx_b, void, ptr, ptr, tl, ptr, env, i32) +DEF_HELPER_6(th_vsll_vx_h, void, ptr, ptr, tl, ptr, env, i32) +DEF_HELPER_6(th_vsll_vx_w, void, ptr, ptr, tl, ptr, env, i32) +DEF_HELPER_6(th_vsll_vx_d, void, ptr, ptr, tl, ptr, env, i32) +DEF_HELPER_6(th_vsrl_vx_b, void, ptr, ptr, tl, ptr, env, i32) +DEF_HELPER_6(th_vsrl_vx_h, void, ptr, ptr, tl, ptr, env, i32) +DEF_HELPER_6(th_vsrl_vx_w, void, ptr, ptr, tl, ptr, env, i32) +DEF_HELPER_6(th_vsrl_vx_d, void, ptr, ptr, tl, ptr, env, i32) +DEF_HELPER_6(th_vsra_vx_b, void, ptr, ptr, tl, ptr, env, i32) +DEF_HELPER_6(th_vsra_vx_h, void, ptr, ptr, tl, ptr, env, i32) +DEF_HELPER_6(th_vsra_vx_w, void, ptr, ptr, tl, ptr, env, i32) +DEF_HELPER_6(th_vsra_vx_d, void, ptr, ptr, tl, ptr, env, i32) diff --git a/target/riscv/insn_trans/trans_xtheadvector.c.inc b/target/riscv/insn_trans/trans_xtheadvector.c.inc index 2b7b2cfe20..d72320699c 100644 --- a/target/riscv/insn_trans/trans_xtheadvector.c.inc +++ b/target/riscv/insn_trans/trans_xtheadvector.c.inc @@ -1325,21 +1325,64 @@ GEN_OPIVI_GVEC_TRANS_TH(th_vand_vi, IMM_SX, th_vand_vx, andi) GEN_OPIVI_GVEC_TRANS_TH(th_vor_vi, IMM_SX, th_vor_vx, ori) GEN_OPIVI_GVEC_TRANS_TH(th_vxor_vi, IMM_SX, th_vxor_vx, xori) +/* Vector Single-Width Bit Shift Instructions */ +GEN_OPIVV_GVEC_TRANS_TH(th_vsll_vv, shlv) +GEN_OPIVV_GVEC_TRANS_TH(th_vsrl_vv, shrv) +GEN_OPIVV_GVEC_TRANS_TH(th_vsra_vv, sarv) + +#define GVecGen2sFn32_Th GVecGen2sFn32 + +/* + * This function is almost the copy of do_opivx_gvec_shift, except: + * 1) XTheadVector simplifies the judgment logic of whether + * to accelerate or not for its lack of fractional LMUL and + * VTA. + */ +static inline bool +do_opivx_gvec_shift_th(DisasContext *s, arg_rmrr *a, GVecGen2sFn32_Th *gvec_fn, + gen_helper_opivx_th *fn) +{ + if (a->vm && s->vl_eq_vlmax) { + TCGv_i32 src1 = tcg_temp_new_i32(); + + tcg_gen_trunc_tl_i32(src1, get_gpr(s, a->rs1, EXT_NONE)); + tcg_gen_extract_i32(src1, src1, 0, s->sew + 3); + gvec_fn(s->sew, vreg_ofs(s, a->rd), vreg_ofs(s, a->rs2), + src1, MAXSZ(s), MAXSZ(s)); + + finalize_rvv_inst(s); + return true; + } + return opivx_trans_th(a->rd, a->rs1, a->rs2, a->vm, fn, s); +} + +#define GEN_OPIVX_GVEC_SHIFT_TRANS_TH(NAME, SUF) \ +static bool trans_##NAME(DisasContext *s, arg_rmrr *a) \ +{ \ + static gen_helper_opivx * const fns[4] = { \ + gen_helper_##NAME##_b, gen_helper_##NAME##_h, \ + gen_helper_##NAME##_w, gen_helper_##NAME##_d, \ + }; \ + if (!opivx_check_th(s, a)) { \ + return false; \ + } \ + return do_opivx_gvec_shift_th(s, a, tcg_gen_gvec_##SUF, fns[s->sew]); \ +} + +GEN_OPIVX_GVEC_SHIFT_TRANS_TH(th_vsll_vx, shls) +GEN_OPIVX_GVEC_SHIFT_TRANS_TH(th_vsrl_vx, shrs) +GEN_OPIVX_GVEC_SHIFT_TRANS_TH(th_vsra_vx, sars) + +GEN_OPIVI_GVEC_TRANS_TH(th_vsll_vi, IMM_TRUNC_SEW, th_vsll_vx, shli) +GEN_OPIVI_GVEC_TRANS_TH(th_vsrl_vi, IMM_TRUNC_SEW, th_vsrl_vx, shri) +GEN_OPIVI_GVEC_TRANS_TH(th_vsra_vi, IMM_TRUNC_SEW, th_vsra_vx, sari) + #define TH_TRANS_STUB(NAME) \ static bool trans_##NAME(DisasContext *s, arg_##NAME *a) \ { \ return require_xtheadvector(s); \ } -TH_TRANS_STUB(th_vsll_vv) -TH_TRANS_STUB(th_vsll_vx) -TH_TRANS_STUB(th_vsll_vi) -TH_TRANS_STUB(th_vsrl_vv) -TH_TRANS_STUB(th_vsrl_vx) -TH_TRANS_STUB(th_vsrl_vi) -TH_TRANS_STUB(th_vsra_vv) -TH_TRANS_STUB(th_vsra_vx) -TH_TRANS_STUB(th_vsra_vi) TH_TRANS_STUB(th_vnsrl_vv) TH_TRANS_STUB(th_vnsrl_vx) TH_TRANS_STUB(th_vnsrl_vi) diff --git a/target/riscv/xtheadvector_helper.c b/target/riscv/xtheadvector_helper.c index 85fa69dd82..d3f10ad873 100644 --- a/target/riscv/xtheadvector_helper.c +++ b/target/riscv/xtheadvector_helper.c @@ -1274,3 +1274,118 @@ GEN_TH_VX(th_vxor_vx_b, 1, 1, clearb_th) GEN_TH_VX(th_vxor_vx_h, 2, 2, clearh_th) GEN_TH_VX(th_vxor_vx_w, 4, 4, clearl_th) GEN_TH_VX(th_vxor_vx_d, 8, 8, clearq_th) + +/* Vector Single-Width Bit Shift Instructions */ +#define TH_SLL(N, M) (N << (M)) +#define TH_SRL(N, M) (N >> (M)) + +/* + * generate the helpers for shift instructions with two vector operators + * + * GEN_TH_SHIFT_VV and GEN_TH_SHIFT_VX are almost the copy of + * GEN_VEXT_SHIFT_VV and GEN_VEXT_SHIFT_VX, except: + * 1) different mask layout + * 2) different data encoding + * 3) different masked/tail elements process policy + */ +#define GEN_TH_SHIFT_VV(NAME, TS1, TS2, HS1, HS2, OP, MASK, CLEAR_FN) \ +void HELPER(NAME)(void *vd, void *v0, void *vs1, \ + void *vs2, CPURISCVState *env, uint32_t desc) \ +{ \ + uint32_t mlen = th_mlen(desc); \ + uint32_t vm = th_vm(desc); \ + uint32_t vl = env->vl; \ + uint32_t esz = sizeof(TS1); \ + uint32_t vlmax = th_maxsz(desc) / esz; \ + uint32_t i; \ + \ + VSTART_CHECK_EARLY_EXIT(env); \ + for (i = env->vstart; i < vl; i++) { \ + if (!vm && !th_elem_mask(v0, mlen, i)) { \ + continue; \ + } \ + TS1 s1 = *((TS1 *)vs1 + HS1(i)); \ + TS2 s2 = *((TS2 *)vs2 + HS2(i)); \ + *((TS1 *)vd + HS1(i)) = OP(s2, s1 & MASK); \ + } \ + env->vstart = 0; \ + CLEAR_FN(vd, vl, vl * esz, vlmax * esz); \ +} + +GEN_TH_SHIFT_VV(th_vsll_vv_b, uint8_t, uint8_t, H1, H1, TH_SLL, + 0x7, clearb_th) +GEN_TH_SHIFT_VV(th_vsll_vv_h, uint16_t, uint16_t, H2, H2, TH_SLL, + 0xf, clearh_th) +GEN_TH_SHIFT_VV(th_vsll_vv_w, uint32_t, uint32_t, H4, H4, TH_SLL, + 0x1f, clearl_th) +GEN_TH_SHIFT_VV(th_vsll_vv_d, uint64_t, uint64_t, H8, H8, TH_SLL, + 0x3f, clearq_th) + +GEN_TH_SHIFT_VV(th_vsrl_vv_b, uint8_t, uint8_t, H1, H1, TH_SRL, + 0x7, clearb_th) +GEN_TH_SHIFT_VV(th_vsrl_vv_h, uint16_t, uint16_t, H2, H2, TH_SRL, + 0xf, clearh_th) +GEN_TH_SHIFT_VV(th_vsrl_vv_w, uint32_t, uint32_t, H4, H4, TH_SRL, + 0x1f, clearl_th) +GEN_TH_SHIFT_VV(th_vsrl_vv_d, uint64_t, uint64_t, H8, H8, TH_SRL, + 0x3f, clearq_th) + +GEN_TH_SHIFT_VV(th_vsra_vv_b, uint8_t, int8_t, H1, H1, TH_SRL, + 0x7, clearb_th) +GEN_TH_SHIFT_VV(th_vsra_vv_h, uint16_t, int16_t, H2, H2, TH_SRL, + 0xf, clearh_th) +GEN_TH_SHIFT_VV(th_vsra_vv_w, uint32_t, int32_t, H4, H4, TH_SRL, + 0x1f, clearl_th) +GEN_TH_SHIFT_VV(th_vsra_vv_d, uint64_t, int64_t, H8, H8, TH_SRL, + 0x3f, clearq_th) + +/* generate the helpers for shift instructions with one vector and one scalar */ +#define GEN_TH_SHIFT_VX(NAME, TD, TS2, HD, HS2, OP, MASK, CLEAR_FN) \ +void HELPER(NAME)(void *vd, void *v0, target_ulong s1, \ + void *vs2, CPURISCVState *env, uint32_t desc) \ +{ \ + uint32_t mlen = th_mlen(desc); \ + uint32_t vm = th_vm(desc); \ + uint32_t vl = env->vl; \ + uint32_t esz = sizeof(TD); \ + uint32_t vlmax = th_maxsz(desc) / esz; \ + uint32_t i; \ + \ + VSTART_CHECK_EARLY_EXIT(env); \ + for (i = env->vstart; i < vl; i++) { \ + if (!vm && !th_elem_mask(v0, mlen, i)) { \ + continue; \ + } \ + TS2 s2 = *((TS2 *)vs2 + HS2(i)); \ + *((TD *)vd + HD(i)) = OP(s2, s1 & MASK); \ + } \ + env->vstart = 0; \ + CLEAR_FN(vd, vl, vl * esz, vlmax * esz); \ +} + +GEN_TH_SHIFT_VX(th_vsll_vx_b, uint8_t, int8_t, H1, H1, TH_SLL, + 0x7, clearb_th) +GEN_TH_SHIFT_VX(th_vsll_vx_h, uint16_t, int16_t, H2, H2, TH_SLL, + 0xf, clearh_th) +GEN_TH_SHIFT_VX(th_vsll_vx_w, uint32_t, int32_t, H4, H4, TH_SLL, + 0x1f, clearl_th) +GEN_TH_SHIFT_VX(th_vsll_vx_d, uint64_t, int64_t, H8, H8, TH_SLL, + 0x3f, clearq_th) + +GEN_TH_SHIFT_VX(th_vsrl_vx_b, uint8_t, uint8_t, H1, H1, TH_SRL, + 0x7, clearb_th) +GEN_TH_SHIFT_VX(th_vsrl_vx_h, uint16_t, uint16_t, H2, H2, TH_SRL, + 0xf, clearh_th) +GEN_TH_SHIFT_VX(th_vsrl_vx_w, uint32_t, uint32_t, H4, H4, TH_SRL, + 0x1f, clearl_th) +GEN_TH_SHIFT_VX(th_vsrl_vx_d, uint64_t, uint64_t, H8, H8, TH_SRL, + 0x3f, clearq_th) + +GEN_TH_SHIFT_VX(th_vsra_vx_b, int8_t, int8_t, H1, H1, TH_SRL, + 0x7, clearb_th) +GEN_TH_SHIFT_VX(th_vsra_vx_h, int16_t, int16_t, H2, H2, TH_SRL, + 0xf, clearh_th) +GEN_TH_SHIFT_VX(th_vsra_vx_w, int32_t, int32_t, H4, H4, TH_SRL, + 0x1f, clearl_th) +GEN_TH_SHIFT_VX(th_vsra_vx_d, int64_t, int64_t, H8, H8, TH_SRL, + 0x3f, clearq_th) -- 2.44.0
