From: Peter Maydell Date: Fri, 22 May 2020 14:55:16 +0000 (+0100) Subject: target/arm: Convert Neon narrowing shifts with op==8 to decodetree X-Git-Url: http://git.osdn.net/view?a=commitdiff_plain;h=712182d340e33c2ce86143f25fb2f04ae23d90de;p=qmiga%2Fqemu.git target/arm: Convert Neon narrowing shifts with op==8 to decodetree Convert the Neon narrowing shifts where op==8 to decodetree: * VSHRN * VRSHRN * VQSHRUN * VQRSHRUN Signed-off-by: Peter Maydell Reviewed-by: Richard Henderson Message-id: 20200522145520.6778-6-peter.maydell@linaro.org --- diff --git a/target/arm/neon-dp.decode b/target/arm/neon-dp.decode index 66c41a53e5..8161995aee 100644 --- a/target/arm/neon-dp.decode +++ b/target/arm/neon-dp.decode @@ -232,6 +232,17 @@ VMINNM_fp_3s 1111 001 1 0 . 1 . .... .... 1111 ... 1 .... @3same_fp @2reg_shl_b .... ... . . . 001 shift:3 .... .... 0 q:1 . . .... \ &2reg_shift vm=%vm_dp vd=%vd_dp size=0 +# Narrowing right shifts: here the Q bit is part of the opcode decode +@2reg_shrn_d .... ... . . . 1 ..... .... .... 0 . . . .... \ + &2reg_shift vm=%vm_dp vd=%vd_dp size=3 q=0 \ + shift=%neon_rshift_i5 +@2reg_shrn_s .... ... . . . 01 .... .... .... 0 . . . .... \ + &2reg_shift vm=%vm_dp vd=%vd_dp size=2 q=0 \ + shift=%neon_rshift_i4 +@2reg_shrn_h .... ... . . . 001 ... .... .... 0 . . . .... \ + &2reg_shift vm=%vm_dp vd=%vd_dp size=1 q=0 \ + shift=%neon_rshift_i3 + VSHR_S_2sh 1111 001 0 1 . ...... .... 0000 . . . 1 .... @2reg_shr_d VSHR_S_2sh 1111 001 0 1 . ...... .... 0000 . . . 1 .... @2reg_shr_s VSHR_S_2sh 1111 001 0 1 . ...... .... 0000 . . . 1 .... @2reg_shr_h @@ -301,3 +312,19 @@ VQSHL_U_64_2sh 1111 001 1 1 . ...... .... 0111 . . . 1 .... @2reg_shl_d VQSHL_U_2sh 1111 001 1 1 . ...... .... 0111 . . . 1 .... @2reg_shl_s VQSHL_U_2sh 1111 001 1 1 . ...... .... 0111 . . . 1 .... @2reg_shl_h VQSHL_U_2sh 1111 001 1 1 . ...... .... 0111 . . . 1 .... @2reg_shl_b + +VSHRN_64_2sh 1111 001 0 1 . ...... .... 1000 . 0 . 1 .... @2reg_shrn_d +VSHRN_32_2sh 1111 001 0 1 . ...... .... 1000 . 0 . 1 .... @2reg_shrn_s +VSHRN_16_2sh 1111 001 0 1 . ...... .... 1000 . 0 . 1 .... @2reg_shrn_h + +VRSHRN_64_2sh 1111 001 0 1 . ...... .... 1000 . 1 . 1 .... @2reg_shrn_d +VRSHRN_32_2sh 1111 001 0 1 . ...... .... 1000 . 1 . 1 .... @2reg_shrn_s +VRSHRN_16_2sh 1111 001 0 1 . ...... .... 1000 . 1 . 1 .... @2reg_shrn_h + +VQSHRUN_64_2sh 1111 001 1 1 . ...... .... 1000 . 0 . 1 .... @2reg_shrn_d +VQSHRUN_32_2sh 1111 001 1 1 . ...... .... 1000 . 0 . 1 .... @2reg_shrn_s +VQSHRUN_16_2sh 1111 001 1 1 . ...... .... 1000 . 0 . 1 .... @2reg_shrn_h + +VQRSHRUN_64_2sh 1111 001 1 1 . ...... .... 1000 . 1 . 1 .... @2reg_shrn_d +VQRSHRUN_32_2sh 1111 001 1 1 . ...... .... 1000 . 1 . 1 .... @2reg_shrn_s +VQRSHRUN_16_2sh 1111 001 1 1 . ...... .... 1000 . 1 . 1 .... @2reg_shrn_h diff --git a/target/arm/translate-neon.inc.c b/target/arm/translate-neon.inc.c index baa985b16c..fe3fb7f62f 100644 --- a/target/arm/translate-neon.inc.c +++ b/target/arm/translate-neon.inc.c @@ -1396,3 +1396,170 @@ static bool do_2shift_env_32(DisasContext *s, arg_2reg_shift *a, DO_2SHIFT_ENV(VQSHLU, qshlu_s) DO_2SHIFT_ENV(VQSHL_U, qshl_u) DO_2SHIFT_ENV(VQSHL_S, qshl_s) + +static bool do_2shift_narrow_64(DisasContext *s, arg_2reg_shift *a, + NeonGenTwo64OpFn *shiftfn, + NeonGenNarrowEnvFn *narrowfn) +{ + /* 2-reg-and-shift narrowing-shift operations, size == 3 case */ + TCGv_i64 constimm, rm1, rm2; + TCGv_i32 rd; + + if (!arm_dc_feature(s, ARM_FEATURE_NEON)) { + return false; + } + + /* UNDEF accesses to D16-D31 if they don't exist. */ + if (!dc_isar_feature(aa32_simd_r32, s) && + ((a->vd | a->vm) & 0x10)) { + return false; + } + + if (a->vm & 1) { + return false; + } + + if (!vfp_access_check(s)) { + return true; + } + + /* + * This is always a right shift, and the shiftfn is always a + * left-shift helper, which thus needs the negated shift count. + */ + constimm = tcg_const_i64(-a->shift); + rm1 = tcg_temp_new_i64(); + rm2 = tcg_temp_new_i64(); + + /* Load both inputs first to avoid potential overwrite if rm == rd */ + neon_load_reg64(rm1, a->vm); + neon_load_reg64(rm2, a->vm + 1); + + shiftfn(rm1, rm1, constimm); + rd = tcg_temp_new_i32(); + narrowfn(rd, cpu_env, rm1); + neon_store_reg(a->vd, 0, rd); + + shiftfn(rm2, rm2, constimm); + rd = tcg_temp_new_i32(); + narrowfn(rd, cpu_env, rm2); + neon_store_reg(a->vd, 1, rd); + + tcg_temp_free_i64(rm1); + tcg_temp_free_i64(rm2); + tcg_temp_free_i64(constimm); + + return true; +} + +static bool do_2shift_narrow_32(DisasContext *s, arg_2reg_shift *a, + NeonGenTwoOpFn *shiftfn, + NeonGenNarrowEnvFn *narrowfn) +{ + /* 2-reg-and-shift narrowing-shift operations, size < 3 case */ + TCGv_i32 constimm, rm1, rm2, rm3, rm4; + TCGv_i64 rtmp; + uint32_t imm; + + if (!arm_dc_feature(s, ARM_FEATURE_NEON)) { + return false; + } + + /* UNDEF accesses to D16-D31 if they don't exist. */ + if (!dc_isar_feature(aa32_simd_r32, s) && + ((a->vd | a->vm) & 0x10)) { + return false; + } + + if (a->vm & 1) { + return false; + } + + if (!vfp_access_check(s)) { + return true; + } + + /* + * This is always a right shift, and the shiftfn is always a + * left-shift helper, which thus needs the negated shift count + * duplicated into each lane of the immediate value. + */ + if (a->size == 1) { + imm = (uint16_t)(-a->shift); + imm |= imm << 16; + } else { + /* size == 2 */ + imm = -a->shift; + } + constimm = tcg_const_i32(imm); + + /* Load all inputs first to avoid potential overwrite */ + rm1 = neon_load_reg(a->vm, 0); + rm2 = neon_load_reg(a->vm, 1); + rm3 = neon_load_reg(a->vm + 1, 0); + rm4 = neon_load_reg(a->vm + 1, 1); + rtmp = tcg_temp_new_i64(); + + shiftfn(rm1, rm1, constimm); + shiftfn(rm2, rm2, constimm); + + tcg_gen_concat_i32_i64(rtmp, rm1, rm2); + tcg_temp_free_i32(rm2); + + narrowfn(rm1, cpu_env, rtmp); + neon_store_reg(a->vd, 0, rm1); + + shiftfn(rm3, rm3, constimm); + shiftfn(rm4, rm4, constimm); + tcg_temp_free_i32(constimm); + + tcg_gen_concat_i32_i64(rtmp, rm3, rm4); + tcg_temp_free_i32(rm4); + + narrowfn(rm3, cpu_env, rtmp); + tcg_temp_free_i64(rtmp); + neon_store_reg(a->vd, 1, rm3); + return true; +} + +#define DO_2SN_64(INSN, FUNC, NARROWFUNC) \ + static bool trans_##INSN##_2sh(DisasContext *s, arg_2reg_shift *a) \ + { \ + return do_2shift_narrow_64(s, a, FUNC, NARROWFUNC); \ + } +#define DO_2SN_32(INSN, FUNC, NARROWFUNC) \ + static bool trans_##INSN##_2sh(DisasContext *s, arg_2reg_shift *a) \ + { \ + return do_2shift_narrow_32(s, a, FUNC, NARROWFUNC); \ + } + +static void gen_neon_narrow_u32(TCGv_i32 dest, TCGv_ptr env, TCGv_i64 src) +{ + tcg_gen_extrl_i64_i32(dest, src); +} + +static void gen_neon_narrow_u16(TCGv_i32 dest, TCGv_ptr env, TCGv_i64 src) +{ + gen_helper_neon_narrow_u16(dest, src); +} + +static void gen_neon_narrow_u8(TCGv_i32 dest, TCGv_ptr env, TCGv_i64 src) +{ + gen_helper_neon_narrow_u8(dest, src); +} + +DO_2SN_64(VSHRN_64, gen_ushl_i64, gen_neon_narrow_u32) +DO_2SN_32(VSHRN_32, gen_ushl_i32, gen_neon_narrow_u16) +DO_2SN_32(VSHRN_16, gen_helper_neon_shl_u16, gen_neon_narrow_u8) + +DO_2SN_64(VRSHRN_64, gen_helper_neon_rshl_u64, gen_neon_narrow_u32) +DO_2SN_32(VRSHRN_32, gen_helper_neon_rshl_u32, gen_neon_narrow_u16) +DO_2SN_32(VRSHRN_16, gen_helper_neon_rshl_u16, gen_neon_narrow_u8) + +DO_2SN_64(VQSHRUN_64, gen_sshl_i64, gen_helper_neon_unarrow_sat32) +DO_2SN_32(VQSHRUN_32, gen_sshl_i32, gen_helper_neon_unarrow_sat16) +DO_2SN_32(VQSHRUN_16, gen_helper_neon_shl_s16, gen_helper_neon_unarrow_sat8) + +DO_2SN_64(VQRSHRUN_64, gen_helper_neon_rshl_s64, gen_helper_neon_unarrow_sat32) +DO_2SN_32(VQRSHRUN_32, gen_helper_neon_rshl_s32, gen_helper_neon_unarrow_sat16) +DO_2SN_32(VQRSHRUN_16, gen_helper_neon_rshl_s16, gen_helper_neon_unarrow_sat8) diff --git a/target/arm/translate.c b/target/arm/translate.c index c32a16085c..11330b9296 100644 --- a/target/arm/translate.c +++ b/target/arm/translate.c @@ -5280,6 +5280,7 @@ static int disas_neon_data_insn(DisasContext *s, uint32_t insn) case 5: /* VSHL, VSLI */ case 6: /* VQSHLU */ case 7: /* VQSHL */ + case 8: /* VSHRN, VRSHRN, VQSHRUN, VQRSHRUN */ return 1; /* handled by decodetree */ default: break;