Excepting MOVPRFX, which isn't a reduction. Presumably it is placed within the group because of its encoding.
Signed-off-by: Richard Henderson <[email protected]> --- target/arm/helper-sve.h | 44 +++++++++++++++++ target/arm/sve_helper.c | 116 +++++++++++++++++++++++++++++++++++++++------ target/arm/translate-sve.c | 64 +++++++++++++++++++++++++ target/arm/sve.def | 22 +++++++++ 4 files changed, 231 insertions(+), 15 deletions(-) diff --git a/target/arm/helper-sve.h b/target/arm/helper-sve.h index 964b15b104..937598d6f8 100644 --- a/target/arm/helper-sve.h +++ b/target/arm/helper-sve.h @@ -162,6 +162,50 @@ DEF_HELPER_FLAGS_5(sve_udiv_zpzz_s, TCG_CALL_NO_RWG, DEF_HELPER_FLAGS_5(sve_udiv_zpzz_d, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_orv_b, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_orv_h, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_orv_s, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_orv_d, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) + +DEF_HELPER_FLAGS_3(sve_eorv_b, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_eorv_h, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_eorv_s, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_eorv_d, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) + +DEF_HELPER_FLAGS_3(sve_andv_b, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_andv_h, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_andv_s, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_andv_d, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) + +DEF_HELPER_FLAGS_3(sve_saddv_b, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_saddv_h, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_saddv_s, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) + +DEF_HELPER_FLAGS_3(sve_uaddv_b, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_uaddv_h, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_uaddv_s, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_uaddv_d, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) + +DEF_HELPER_FLAGS_3(sve_smaxv_b, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_smaxv_h, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_smaxv_s, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_smaxv_d, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) + +DEF_HELPER_FLAGS_3(sve_umaxv_b, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_umaxv_h, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_umaxv_s, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_umaxv_d, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) + +DEF_HELPER_FLAGS_3(sve_sminv_b, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_sminv_h, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_sminv_s, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_sminv_d, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) + +DEF_HELPER_FLAGS_3(sve_uminv_b, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_uminv_h, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_uminv_s, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) +DEF_HELPER_FLAGS_3(sve_uminv_d, TCG_CALL_NO_RWG, i64, ptr, ptr, i32) + DEF_HELPER_FLAGS_5(sve_and_pred, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32) DEF_HELPER_FLAGS_5(sve_bic_pred, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32) DEF_HELPER_FLAGS_5(sve_eor_pred, TCG_CALL_NO_RWG, void, ptr, ptr, ptr, ptr, i32) diff --git a/target/arm/sve_helper.c b/target/arm/sve_helper.c index b617ea2c04..fca17440e7 100644 --- a/target/arm/sve_helper.c +++ b/target/arm/sve_helper.c @@ -196,11 +196,6 @@ DO_ZPZZ_D(sve_orr_zpzz_d, uint64_t, DO_ORR) DO_ZPZZ_D(sve_eor_zpzz_d, uint64_t, DO_EOR) DO_ZPZZ_D(sve_bic_zpzz_d, uint64_t, DO_BIC) -#undef DO_AND -#undef DO_ORR -#undef DO_EOR -#undef DO_BIC - #define DO_ADD(N, M) (N + M) #define DO_SUB(N, M) (N - M) @@ -216,9 +211,6 @@ DO_ZPZZ(sve_sub_zpzz_s, uint32_t, H1_4, DO_SUB) DO_ZPZZ_D(sve_add_zpzz_d, uint64_t, DO_ADD) DO_ZPZZ_D(sve_sub_zpzz_d, uint64_t, DO_SUB) -#undef DO_ADD -#undef DO_SUB - #define DO_MAX(N, M) ((N) >= (M) ? (N) : (M)) #define DO_MIN(N, M) ((N) >= (M) ? (M) : (N)) #define DO_ABD(N, M) ((N) >= (M) ? (N) - (M) : (M) - (N)) @@ -251,10 +243,6 @@ DO_ZPZZ_D(sve_umin_zpzz_d, uint64_t, DO_MIN) DO_ZPZZ_D(sve_sabd_zpzz_d, int64_t, DO_ABD) DO_ZPZZ_D(sve_uabd_zpzz_d, uint64_t, DO_ABD) -#undef DO_MAX -#undef DO_MIN -#undef DO_ABD - #define DO_MUL(N, M) (N * M) #define DO_DIV(N, M) (M ? N / M : 0) @@ -309,12 +297,110 @@ DO_ZPZZ_D(sve_umulh_zpzz_d, uint64_t, do_umulh_d) DO_ZPZZ_D(sve_sdiv_zpzz_d, int64_t, DO_DIV) DO_ZPZZ_D(sve_udiv_zpzz_d, uint64_t, DO_DIV) -#undef DO_MUL -#undef DO_DIV - #undef DO_ZPZZ #undef DO_ZPZZ_D +/* Two-operand reduction expander, controlled by a predicate. + * The difference between TYPERED and TYPERET has to do with + * sign-extension. E.g. for SMAX, TYPERED must be signed, + * but TYPERET must be unsigned so that e.g. a 32-bit value + * is not sign-extended to the ABI uint64_t return type. + */ +/* ??? If we were to vectorize this by hand the reduction ordering + * would change. For integer operands, this is perfectly fine. + */ +#define DO_VPZ(NAME, TYPEELT, TYPERED, TYPERET, H, INIT, OP) \ +uint64_t HELPER(NAME)(void *vn, void *vg, uint32_t desc) \ +{ \ + intptr_t iv, ib, opr_sz = simd_oprsz(desc); \ + TYPERED ret = INIT; \ + for (iv = ib = 0; iv < opr_sz; iv += 16, ib += 2) { \ + uint16_t pg = *(uint16_t *)(vg + H2(ib)); \ + intptr_t i = 0; \ + do { \ + TYPEELT nn = *(TYPEELT *)(vn + iv + H(i)); \ + ret = OP(ret, nn); \ + i += sizeof(TYPEELT), pg >>= sizeof(TYPEELT); \ + } while (pg); \ + } \ + return (TYPERET)ret; \ +} + +#define DO_VPZ_D(NAME, TYPEE, TYPER, INIT, OP) \ +uint64_t HELPER(NAME)(void *vn, void *vg, uint32_t desc) \ +{ \ + intptr_t i, opr_sz = simd_oprsz(desc) / 8; \ + TYPEE *n = vn; \ + uint8_t *pg = vg; \ + TYPER ret = INIT; \ + for (i = 0; i < opr_sz; i += 1) { \ + if (pg[H1(i)] & 1) { \ + TYPEE nn = n[i]; \ + ret = OP(ret, nn); \ + } \ + } \ + return ret; \ +} + +DO_VPZ(sve_orv_b, uint8_t, uint8_t, uint8_t, H1, 0, DO_ORR) +DO_VPZ(sve_orv_h, uint16_t, uint16_t, uint16_t, H1_2, 0, DO_ORR) +DO_VPZ(sve_orv_s, uint32_t, uint32_t, uint32_t, H1_4, 0, DO_ORR) +DO_VPZ_D(sve_orv_d, uint64_t, uint64_t, 0, DO_ORR) + +DO_VPZ(sve_eorv_b, uint8_t, uint8_t, uint8_t, H1, 0, DO_EOR) +DO_VPZ(sve_eorv_h, uint16_t, uint16_t, uint16_t, H1_2, 0, DO_EOR) +DO_VPZ(sve_eorv_s, uint32_t, uint32_t, uint32_t, H1_4, 0, DO_EOR) +DO_VPZ_D(sve_eorv_d, uint64_t, uint64_t, 0, DO_EOR) + +DO_VPZ(sve_andv_b, uint8_t, uint8_t, uint8_t, H1, -1, DO_AND) +DO_VPZ(sve_andv_h, uint16_t, uint16_t, uint16_t, H1_2, -1, DO_AND) +DO_VPZ(sve_andv_s, uint32_t, uint32_t, uint32_t, H1_4, -1, DO_AND) +DO_VPZ_D(sve_andv_d, uint64_t, uint64_t, -1, DO_AND) + +DO_VPZ(sve_saddv_b, int8_t, uint64_t, uint64_t, H1, 0, DO_ADD) +DO_VPZ(sve_saddv_h, int16_t, uint64_t, uint64_t, H1_2, 0, DO_ADD) +DO_VPZ(sve_saddv_s, int32_t, uint64_t, uint64_t, H1_4, 0, DO_ADD) + +DO_VPZ(sve_uaddv_b, uint8_t, uint64_t, uint64_t, H1, 0, DO_ADD) +DO_VPZ(sve_uaddv_h, uint16_t, uint64_t, uint64_t, H1_2, 0, DO_ADD) +DO_VPZ(sve_uaddv_s, uint32_t, uint64_t, uint64_t, H1_4, 0, DO_ADD) +DO_VPZ_D(sve_uaddv_d, uint64_t, uint64_t, 0, DO_ADD) + +DO_VPZ(sve_smaxv_b, int8_t, int8_t, uint8_t, H1, INT8_MIN, DO_MAX) +DO_VPZ(sve_smaxv_h, int16_t, int16_t, uint16_t, H1_2, INT16_MIN, DO_MAX) +DO_VPZ(sve_smaxv_s, int32_t, int32_t, uint32_t, H1_4, INT32_MIN, DO_MAX) +DO_VPZ_D(sve_smaxv_d, int64_t, int64_t, INT64_MIN, DO_MAX) + +DO_VPZ(sve_umaxv_b, uint8_t, uint8_t, uint8_t, H1, 0, DO_MAX) +DO_VPZ(sve_umaxv_h, uint16_t, uint16_t, uint16_t, H1_2, 0, DO_MAX) +DO_VPZ(sve_umaxv_s, uint32_t, uint32_t, uint32_t, H1_4, 0, DO_MAX) +DO_VPZ_D(sve_umaxv_d, uint64_t, uint64_t, 0, DO_MAX) + +DO_VPZ(sve_sminv_b, int8_t, int8_t, uint8_t, H1, INT8_MAX, DO_MIN) +DO_VPZ(sve_sminv_h, int16_t, int16_t, uint16_t, H1_2, INT16_MAX, DO_MIN) +DO_VPZ(sve_sminv_s, int32_t, int32_t, uint32_t, H1_4, INT32_MAX, DO_MIN) +DO_VPZ_D(sve_sminv_d, int64_t, int64_t, INT64_MAX, DO_MIN) + +DO_VPZ(sve_uminv_b, uint8_t, uint8_t, uint8_t, H1, -1, DO_MIN) +DO_VPZ(sve_uminv_h, uint16_t, uint16_t, uint16_t, H1_2, -1, DO_MIN) +DO_VPZ(sve_uminv_s, uint32_t, uint32_t, uint32_t, H1_4, -1, DO_MIN) +DO_VPZ_D(sve_uminv_d, uint64_t, uint64_t, -1, DO_MIN) + +#undef DO_VPZ +#undef DO_VPZ_D + +#undef DO_AND +#undef DO_ORR +#undef DO_EOR +#undef DO_BIC +#undef DO_ADD +#undef DO_SUB +#undef DO_MAX +#undef DO_MIN +#undef DO_ABD +#undef DO_MUL +#undef DO_DIV + void HELPER(sve_ldr)(CPUARMState *env, void *d, target_ulong addr, uint32_t len) { intptr_t i, len_align = QEMU_ALIGN_DOWN(len, 8); diff --git a/target/arm/translate-sve.c b/target/arm/translate-sve.c index d8b34020bb..4abc66ba5f 100644 --- a/target/arm/translate-sve.c +++ b/target/arm/translate-sve.c @@ -22,6 +22,7 @@ #include "exec/exec-all.h" #include "tcg-op.h" #include "tcg-op-gvec.h" +#include "tcg-gvec-desc.h" #include "qemu/log.h" #include "arm_ldst.h" #include "translate.h" @@ -201,6 +202,69 @@ void trans_UDIV_zpzz(DisasContext *s, arg_rprr_esz *a, uint32_t insn) #undef DO_ZPZZ +typedef void gen_helper_gvec_reduc(TCGv_i64, TCGv_ptr, TCGv_ptr, TCGv_i32); +static void do_vpz_ool(DisasContext *s, arg_rpr_esz *a, + gen_helper_gvec_reduc *fn) +{ + unsigned vsz = size_for_gvec(vec_full_reg_size(s)); + TCGv_i32 desc = tcg_const_i32(simd_desc(vsz, vsz, 0)); + TCGv_i64 temp = tcg_temp_new_i64(); + TCGv_ptr t_zn = tcg_temp_new_ptr(); + TCGv_ptr t_pg = tcg_temp_new_ptr(); + + tcg_gen_addi_ptr(t_zn, cpu_env, vec_full_reg_offset(s, a->rn)); + tcg_gen_addi_ptr(t_pg, cpu_env, pred_full_reg_offset(s, a->pg)); + fn(temp, t_zn, t_pg, desc); + tcg_temp_free_ptr(t_zn); + tcg_temp_free_ptr(t_pg); + tcg_temp_free_i32(desc); + + write_fp_dreg(s, a->rd, temp); + tcg_temp_free_i64(temp); +} + +#define DO_VPZ(NAME, name) \ +void trans_##NAME(DisasContext *s, arg_rpr_esz *a, uint32_t insn) \ +{ \ + static gen_helper_gvec_reduc * const fns[4] = { \ + gen_helper_sve_##name##_b, gen_helper_sve_##name##_h, \ + gen_helper_sve_##name##_s, gen_helper_sve_##name##_d, \ + }; \ + do_vpz_ool(s, a, fns[a->esz]); \ +} + +DO_VPZ(ORV, orv) +DO_VPZ(ANDV, andv) +DO_VPZ(EORV, eorv) + +DO_VPZ(UADDV, uaddv) +DO_VPZ(SMAXV, smaxv) +DO_VPZ(UMAXV, umaxv) +DO_VPZ(SMINV, sminv) +DO_VPZ(UMINV, uminv) + +void trans_SADDV(DisasContext *s, arg_rpr_esz *a, uint32_t insn) +{ + gen_helper_gvec_reduc *fn; + switch (a->esz) { + case 0: + fn = gen_helper_sve_saddv_b; + break; + case 1: + fn = gen_helper_sve_saddv_h; + break; + case 2: + fn = gen_helper_sve_saddv_s; + break; + default: + unallocated_encoding(s); + return; + }; + do_vpz_ool(s, a, fn); +} + +#undef DO_VPZ + static uint64_t pred_esz_mask[4] = { 0xffffffffffffffffull, 0x5555555555555555ull, 0x1111111111111111ull, 0x0101010101010101ull diff --git a/target/arm/sve.def b/target/arm/sve.def index 3bb4faaf89..c26b1377e8 100644 --- a/target/arm/sve.def +++ b/target/arm/sve.def @@ -35,6 +35,7 @@ &rri rd rn imm &rrr_esz rd rn rm esz +&rpr_esz rd pg rn esz &rprr_esz rd pg rn rm esz &pred_set rd pat esz i s @@ -52,6 +53,9 @@ @rdn_pg_rm_esz ........ esz:2 ... ... ... pg:3 rm:5 rd:5 &rprr_esz rn=%reg_movprfx @rdm_pg_rn_esz ........ esz:2 ... ... ... pg:3 rn:5 rd:5 &rprr_esz rm=%reg_movprfx +# One register operand, with governing predicate, vector element size +@rd_pg_rn_esz ........ esz:2 ... ... ... pg:3 rn:5 rd:5 &rpr_esz + # Basic Load/Store with 9-bit immediate offset @pd_rn_i9 ........ ........ ...... rn:5 . rd:4 &rri imm=%imm9_16_10 @rd_rn_i9 ........ ........ ...... rn:5 rd:5 &rri imm=%imm9_16_10 @@ -90,6 +94,24 @@ UDIV_zpzz 00000100 .. 010 101 000 ... ..... ..... @rdn_pg_rm_esz SDIV_zpzz 00000100 .. 010 110 000 ... ..... ..... @rdm_pg_rn_esz # SDIVR UDIV_zpzz 00000100 .. 010 111 000 ... ..... ..... @rdm_pg_rn_esz # UDIVR +### SVE Integer Reduction Group + +# SVE bitwise logical reduction (predicated) +ORV 00000100 .. 011 000 001 ... ..... ..... @rd_pg_rn_esz +EORV 00000100 .. 011 001 001 ... ..... ..... @rd_pg_rn_esz +ANDV 00000100 .. 011 010 001 ... ..... ..... @rd_pg_rn_esz + +# SVE integer add reduction (predicated) +UADDV 00000100 .. 000 001 001 ... ..... ..... @rd_pg_rn_esz +# Note that saddv requires size != 3, which is unallocated. +SADDV 00000100 .. 000 000 001 ... ..... ..... @rd_pg_rn_esz + +# SVE integer min/max reduction (predicated) +SMAXV 00000100 .. 001 000 001 ... ..... ..... @rd_pg_rn_esz +UMAXV 00000100 .. 001 001 001 ... ..... ..... @rd_pg_rn_esz +SMINV 00000100 .. 001 010 001 ... ..... ..... @rd_pg_rn_esz +UMINV 00000100 .. 001 011 001 ... ..... ..... @rd_pg_rn_esz + ### SVE Logical - Unpredicated Group # SVE bitwise logical operations (unpredicated) -- 2.14.3
