From a08582f41e94b5a7ee3e53ad1e719d24431aa675 Mon Sep 17 00:00:00 2001 From: Peter Maydell Date: Fri, 31 Jan 2014 14:47:36 +0000 Subject: target-arm: A64: Add SIMD three-different multiply accumulate insns Add support for the multiply-accumulate instructions from the SIMD three-different instructions group (C3.6.15): * skeleton decode of unallocated encodings and split of the group into its three sub-parts * framework for handling the 64x64->128 widening subpart * implementation of the multiply-accumulate instructions SMLAL, SMLAL2, UMLAL, UMLAL2, SMLSL, SMLSL2, UMLSL, UMLSL2, UMULL, UMULL2, SMULL, SMULL2 Signed-off-by: Peter Maydell Reviewed-by: Richard Henderson --- target-arm/translate-a64.c | 233 ++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 232 insertions(+), 1 deletion(-) (limited to 'target-arm') diff --git a/target-arm/translate-a64.c b/target-arm/translate-a64.c index c22bbb5cf3..e8479a52db 100644 --- a/target-arm/translate-a64.c +++ b/target-arm/translate-a64.c @@ -700,6 +700,9 @@ static void do_fp_ld(DisasContext *s, int destidx, TCGv_i64 tcg_addr, int size) * zero extend as we are filling a partial chunk of the vector register. * These functions don't support 128 bit loads/stores, which would be * normal load/store operations. + * + * The _i32 versions are useful when operating on 32 bit quantities + * (eg for floating point single or using Neon helper functions). */ /* Get value of an element within a vector register */ @@ -735,6 +738,32 @@ static void read_vec_element(DisasContext *s, TCGv_i64 tcg_dest, int srcidx, } } +static void read_vec_element_i32(DisasContext *s, TCGv_i32 tcg_dest, int srcidx, + int element, TCGMemOp memop) +{ + int vect_off = vec_reg_offset(srcidx, element, memop & MO_SIZE); + switch (memop) { + case MO_8: + tcg_gen_ld8u_i32(tcg_dest, cpu_env, vect_off); + break; + case MO_16: + tcg_gen_ld16u_i32(tcg_dest, cpu_env, vect_off); + break; + case MO_8|MO_SIGN: + tcg_gen_ld8s_i32(tcg_dest, cpu_env, vect_off); + break; + case MO_16|MO_SIGN: + tcg_gen_ld16s_i32(tcg_dest, cpu_env, vect_off); + break; + case MO_32: + case MO_32|MO_SIGN: + tcg_gen_ld_i32(tcg_dest, cpu_env, vect_off); + break; + default: + g_assert_not_reached(); + } +} + /* Set value of an element within a vector register */ static void write_vec_element(DisasContext *s, TCGv_i64 tcg_src, int destidx, int element, TCGMemOp memop) @@ -5518,6 +5547,150 @@ static void disas_simd_shift_imm(DisasContext *s, uint32_t insn) unsupported_encoding(s, insn); } +static void handle_3rd_widening(DisasContext *s, int is_q, int is_u, int size, + int opcode, int rd, int rn, int rm) +{ + /* 3-reg-different widening insns: 64 x 64 -> 128 */ + TCGv_i64 tcg_res[2]; + int pass, accop; + + tcg_res[0] = tcg_temp_new_i64(); + tcg_res[1] = tcg_temp_new_i64(); + + /* Does this op do an adding accumulate, a subtracting accumulate, + * or no accumulate at all? + */ + switch (opcode) { + case 5: + case 8: + case 9: + accop = 1; + break; + case 10: + case 11: + accop = -1; + break; + default: + accop = 0; + break; + } + + if (accop != 0) { + read_vec_element(s, tcg_res[0], rd, 0, MO_64); + read_vec_element(s, tcg_res[1], rd, 1, MO_64); + } + + /* size == 2 means two 32x32->64 operations; this is worth special + * casing because we can generally handle it inline. + */ + if (size == 2) { + for (pass = 0; pass < 2; pass++) { + TCGv_i64 tcg_op1 = tcg_temp_new_i64(); + TCGv_i64 tcg_op2 = tcg_temp_new_i64(); + TCGv_i64 tcg_passres; + TCGMemOp memop = MO_32 | (is_u ? 0 : MO_SIGN); + + int elt = pass + is_q * 2; + + read_vec_element(s, tcg_op1, rn, elt, memop); + read_vec_element(s, tcg_op2, rm, elt, memop); + + if (accop == 0) { + tcg_passres = tcg_res[pass]; + } else { + tcg_passres = tcg_temp_new_i64(); + } + + switch (opcode) { + case 8: /* SMLAL, SMLAL2, UMLAL, UMLAL2 */ + case 10: /* SMLSL, SMLSL2, UMLSL, UMLSL2 */ + case 12: /* UMULL, UMULL2, SMULL, SMULL2 */ + tcg_gen_mul_i64(tcg_passres, tcg_op1, tcg_op2); + break; + default: + g_assert_not_reached(); + } + + if (accop > 0) { + tcg_gen_add_i64(tcg_res[pass], tcg_res[pass], tcg_passres); + tcg_temp_free_i64(tcg_passres); + } else if (accop < 0) { + tcg_gen_sub_i64(tcg_res[pass], tcg_res[pass], tcg_passres); + tcg_temp_free_i64(tcg_passres); + } + + tcg_temp_free_i64(tcg_op1); + tcg_temp_free_i64(tcg_op2); + } + } else { + /* size 0 or 1, generally helper functions */ + for (pass = 0; pass < 2; pass++) { + TCGv_i32 tcg_op1 = tcg_temp_new_i32(); + TCGv_i32 tcg_op2 = tcg_temp_new_i32(); + TCGv_i64 tcg_passres; + int elt = pass + is_q * 2; + + read_vec_element_i32(s, tcg_op1, rn, elt, MO_32); + read_vec_element_i32(s, tcg_op2, rm, elt, MO_32); + + if (accop == 0) { + tcg_passres = tcg_res[pass]; + } else { + tcg_passres = tcg_temp_new_i64(); + } + + switch (opcode) { + case 8: /* SMLAL, SMLAL2, UMLAL, UMLAL2 */ + case 10: /* SMLSL, SMLSL2, UMLSL, UMLSL2 */ + case 12: /* UMULL, UMULL2, SMULL, SMULL2 */ + if (size == 0) { + if (is_u) { + gen_helper_neon_mull_u8(tcg_passres, tcg_op1, tcg_op2); + } else { + gen_helper_neon_mull_s8(tcg_passres, tcg_op1, tcg_op2); + } + } else { + if (is_u) { + gen_helper_neon_mull_u16(tcg_passres, tcg_op1, tcg_op2); + } else { + gen_helper_neon_mull_s16(tcg_passres, tcg_op1, tcg_op2); + } + } + break; + default: + g_assert_not_reached(); + } + tcg_temp_free_i32(tcg_op1); + tcg_temp_free_i32(tcg_op2); + + if (accop > 0) { + if (size == 0) { + gen_helper_neon_addl_u16(tcg_res[pass], tcg_res[pass], + tcg_passres); + } else { + gen_helper_neon_addl_u32(tcg_res[pass], tcg_res[pass], + tcg_passres); + } + tcg_temp_free_i64(tcg_passres); + } else if (accop < 0) { + if (size == 0) { + gen_helper_neon_subl_u16(tcg_res[pass], tcg_res[pass], + tcg_passres); + } else { + gen_helper_neon_subl_u32(tcg_res[pass], tcg_res[pass], + tcg_passres); + } + tcg_temp_free_i64(tcg_passres); + } + } + } + + write_vec_element(s, tcg_res[0], rd, 0, MO_64); + write_vec_element(s, tcg_res[1], rd, 1, MO_64); + tcg_temp_free_i64(tcg_res[0]); + tcg_temp_free_i64(tcg_res[1]); +} + /* C3.6.15 AdvSIMD three different * 31 30 29 28 24 23 22 21 20 16 15 12 11 10 9 5 4 0 * +---+---+---+-----------+------+---+------+--------+-----+------+------+ @@ -5526,7 +5699,65 @@ static void disas_simd_shift_imm(DisasContext *s, uint32_t insn) */ static void disas_simd_three_reg_diff(DisasContext *s, uint32_t insn) { - unsupported_encoding(s, insn); + /* Instructions in this group fall into three basic classes + * (in each case with the operation working on each element in + * the input vectors): + * (1) widening 64 x 64 -> 128 (with possibly Vd as an extra + * 128 bit input) + * (2) wide 64 x 128 -> 128 + * (3) narrowing 128 x 128 -> 64 + * Here we do initial decode, catch unallocated cases and + * dispatch to separate functions for each class. + */ + int is_q = extract32(insn, 30, 1); + int is_u = extract32(insn, 29, 1); + int size = extract32(insn, 22, 2); + int opcode = extract32(insn, 12, 4); + int rm = extract32(insn, 16, 5); + int rn = extract32(insn, 5, 5); + int rd = extract32(insn, 0, 5); + + switch (opcode) { + case 1: /* SADDW, SADDW2, UADDW, UADDW2 */ + case 3: /* SSUBW, SSUBW2, USUBW, USUBW2 */ + /* 64 x 128 -> 128 */ + unsupported_encoding(s, insn); + break; + case 4: /* ADDHN, ADDHN2, RADDHN, RADDHN2 */ + case 6: /* SUBHN, SUBHN2, RSUBHN, RSUBHN2 */ + /* 128 x 128 -> 64 */ + unsupported_encoding(s, insn); + break; + case 9: + case 11: + case 13: + case 14: + if (is_u) { + unallocated_encoding(s); + return; + } + /* fall through */ + case 0: + case 2: + case 5: + case 7: + unsupported_encoding(s, insn); + break; + case 8: + case 10: + case 12: + /* 64 x 64 -> 128 */ + if (size == 3) { + unallocated_encoding(s); + return; + } + handle_3rd_widening(s, is_q, is_u, size, opcode, rd, rn, rm); + break; + default: + /* opcode 15 not allocated */ + unallocated_encoding(s); + break; + } } /* C3.6.16 AdvSIMD three same -- cgit v1.2.1