diff options
| -rw-r--r-- | target/arm/cpu.h | 5 | ||||
| -rw-r--r-- | target/arm/translate-vfp.inc.c | 205 | ||||
| -rw-r--r-- | target/arm/translate.c | 14 | ||||
| -rw-r--r-- | target/arm/vfp.decode | 6 |
4 files changed, 224 insertions, 6 deletions
diff --git a/target/arm/cpu.h b/target/arm/cpu.h index dc50c86e31..9229862421 100644 --- a/target/arm/cpu.h +++ b/target/arm/cpu.h @@ -3377,6 +3377,11 @@ static inline bool isar_feature_aa32_fp_d32(const ARMISARegisters *id) return FIELD_EX64(id->mvfr0, MVFR0, SIMDREG) >= 2; } +static inline bool isar_feature_aa32_fpshvec(const ARMISARegisters *id) +{ + return FIELD_EX64(id->mvfr0, MVFR0, FPSHVEC) > 0; +} + /* * We always set the FP and SIMD FP16 fields to indicate identical * levels of support (assuming SIMD is implemented at all), so diff --git a/target/arm/translate-vfp.inc.c b/target/arm/translate-vfp.inc.c index 9729946d73..4f922dc840 100644 --- a/target/arm/translate-vfp.inc.c +++ b/target/arm/translate-vfp.inc.c @@ -1098,3 +1098,208 @@ static bool trans_VLDM_VSTM_dp(DisasContext *s, arg_VLDM_VSTM_dp *a) return true; } + +/* + * Types for callbacks for do_vfp_3op_sp() and do_vfp_3op_dp(). + * The callback should emit code to write a value to vd. If + * do_vfp_3op_{sp,dp}() was passed reads_vd then the TCGv vd + * will contain the old value of the relevant VFP register; + * otherwise it must be written to only. + */ +typedef void VFPGen3OpSPFn(TCGv_i32 vd, + TCGv_i32 vn, TCGv_i32 vm, TCGv_ptr fpst); +typedef void VFPGen3OpDPFn(TCGv_i64 vd, + TCGv_i64 vn, TCGv_i64 vm, TCGv_ptr fpst); + +/* + * Perform a 3-operand VFP data processing instruction. fn is the + * callback to do the actual operation; this function deals with the + * code to handle looping around for VFP vector processing. + */ +static bool do_vfp_3op_sp(DisasContext *s, VFPGen3OpSPFn *fn, + int vd, int vn, int vm, bool reads_vd) +{ + uint32_t delta_m = 0; + uint32_t delta_d = 0; + uint32_t bank_mask = 0; + int veclen = s->vec_len; + TCGv_i32 f0, f1, fd; + TCGv_ptr fpst; + + if (!dc_isar_feature(aa32_fpshvec, s) && + (veclen != 0 || s->vec_stride != 0)) { + return false; + } + + if (!vfp_access_check(s)) { + return true; + } + + if (veclen > 0) { + bank_mask = 0x18; + + /* Figure out what type of vector operation this is. */ + if ((vd & bank_mask) == 0) { + /* scalar */ + veclen = 0; + } else { + delta_d = s->vec_stride + 1; + + if ((vm & bank_mask) == 0) { + /* mixed scalar/vector */ + delta_m = 0; + } else { + /* vector */ + delta_m = delta_d; + } + } + } + + f0 = tcg_temp_new_i32(); + f1 = tcg_temp_new_i32(); + fd = tcg_temp_new_i32(); + fpst = get_fpstatus_ptr(0); + + neon_load_reg32(f0, vn); + neon_load_reg32(f1, vm); + + for (;;) { + if (reads_vd) { + neon_load_reg32(fd, vd); + } + fn(fd, f0, f1, fpst); + neon_store_reg32(fd, vd); + + if (veclen == 0) { + break; + } + + /* Set up the operands for the next iteration */ + veclen--; + vd = ((vd + delta_d) & (bank_mask - 1)) | (vd & bank_mask); + vn = ((vn + delta_d) & (bank_mask - 1)) | (vn & bank_mask); + neon_load_reg32(f0, vn); + if (delta_m) { + vm = ((vm + delta_m) & (bank_mask - 1)) | (vm & bank_mask); + neon_load_reg32(f1, vm); + } + } + + tcg_temp_free_i32(f0); + tcg_temp_free_i32(f1); + tcg_temp_free_i32(fd); + tcg_temp_free_ptr(fpst); + + return true; +} + +static bool do_vfp_3op_dp(DisasContext *s, VFPGen3OpDPFn *fn, + int vd, int vn, int vm, bool reads_vd) +{ + uint32_t delta_m = 0; + uint32_t delta_d = 0; + uint32_t bank_mask = 0; + int veclen = s->vec_len; + TCGv_i64 f0, f1, fd; + TCGv_ptr fpst; + + /* UNDEF accesses to D16-D31 if they don't exist */ + if (!dc_isar_feature(aa32_fp_d32, s) && ((vd | vn | vm) & 0x10)) { + return false; + } + + if (!dc_isar_feature(aa32_fpshvec, s) && + (veclen != 0 || s->vec_stride != 0)) { + return false; + } + + if (!vfp_access_check(s)) { + return true; + } + + if (veclen > 0) { + bank_mask = 0xc; + + /* Figure out what type of vector operation this is. */ + if ((vd & bank_mask) == 0) { + /* scalar */ + veclen = 0; + } else { + delta_d = (s->vec_stride >> 1) + 1; + + if ((vm & bank_mask) == 0) { + /* mixed scalar/vector */ + delta_m = 0; + } else { + /* vector */ + delta_m = delta_d; + } + } + } + + f0 = tcg_temp_new_i64(); + f1 = tcg_temp_new_i64(); + fd = tcg_temp_new_i64(); + fpst = get_fpstatus_ptr(0); + + neon_load_reg64(f0, vn); + neon_load_reg64(f1, vm); + + for (;;) { + if (reads_vd) { + neon_load_reg64(fd, vd); + } + fn(fd, f0, f1, fpst); + neon_store_reg64(fd, vd); + + if (veclen == 0) { + break; + } + /* Set up the operands for the next iteration */ + veclen--; + vd = ((vd + delta_d) & (bank_mask - 1)) | (vd & bank_mask); + vn = ((vn + delta_d) & (bank_mask - 1)) | (vn & bank_mask); + neon_load_reg64(f0, vn); + if (delta_m) { + vm = ((vm + delta_m) & (bank_mask - 1)) | (vm & bank_mask); + neon_load_reg64(f1, vm); + } + } + + tcg_temp_free_i64(f0); + tcg_temp_free_i64(f1); + tcg_temp_free_i64(fd); + tcg_temp_free_ptr(fpst); + + return true; +} + +static void gen_VMLA_sp(TCGv_i32 vd, TCGv_i32 vn, TCGv_i32 vm, TCGv_ptr fpst) +{ + /* Note that order of inputs to the add matters for NaNs */ + TCGv_i32 tmp = tcg_temp_new_i32(); + + gen_helper_vfp_muls(tmp, vn, vm, fpst); + gen_helper_vfp_adds(vd, vd, tmp, fpst); + tcg_temp_free_i32(tmp); +} + +static bool trans_VMLA_sp(DisasContext *s, arg_VMLA_sp *a) +{ + return do_vfp_3op_sp(s, gen_VMLA_sp, a->vd, a->vn, a->vm, true); +} + +static void gen_VMLA_dp(TCGv_i64 vd, TCGv_i64 vn, TCGv_i64 vm, TCGv_ptr fpst) +{ + /* Note that order of inputs to the add matters for NaNs */ + TCGv_i64 tmp = tcg_temp_new_i64(); + + gen_helper_vfp_muld(tmp, vn, vm, fpst); + gen_helper_vfp_addd(vd, vd, tmp, fpst); + tcg_temp_free_i64(tmp); +} + +static bool trans_VMLA_dp(DisasContext *s, arg_VMLA_sp *a) +{ + return do_vfp_3op_dp(s, gen_VMLA_dp, a->vd, a->vn, a->vm, true); +} diff --git a/target/arm/translate.c b/target/arm/translate.c index 4a462fe094..2215e83c86 100644 --- a/target/arm/translate.c +++ b/target/arm/translate.c @@ -3133,6 +3133,14 @@ static int disas_vfp_insn(DisasContext *s, uint32_t insn) op = ((insn >> 20) & 8) | ((insn >> 19) & 6) | ((insn >> 6) & 1); rn = VFP_SREG_N(insn); + switch (op) { + case 0: + /* Already handled by decodetree */ + return 1; + default: + break; + } + if (op == 15) { /* rn is opcode, encoded as per VFP_SREG_N. */ switch (rn) { @@ -3312,12 +3320,6 @@ static int disas_vfp_insn(DisasContext *s, uint32_t insn) for (;;) { /* Perform the calculation. */ switch (op) { - case 0: /* VMLA: fd + (fn * fm) */ - /* Note that order of inputs to the add matters for NaNs */ - gen_vfp_F1_mul(dp); - gen_mov_F0_vreg(dp, rd); - gen_vfp_add(dp); - break; case 1: /* VMLS: fd + -(fn * fm) */ gen_vfp_mul(dp); gen_vfp_F1_neg(dp); diff --git a/target/arm/vfp.decode b/target/arm/vfp.decode index 68c9ffcfd3..9530e17ae0 100644 --- a/target/arm/vfp.decode +++ b/target/arm/vfp.decode @@ -96,3 +96,9 @@ VLDM_VSTM_sp ---- 1101 0.1 l:1 rn:4 .... 1010 imm:8 \ vd=%vd_sp p=1 u=0 w=1 VLDM_VSTM_dp ---- 1101 0.1 l:1 rn:4 .... 1011 imm:8 \ vd=%vd_dp p=1 u=0 w=1 + +# 3-register VFP data-processing; bits [23,21:20,6] identify the operation. +VMLA_sp ---- 1110 0.00 .... .... 1010 .0.0 .... \ + vm=%vm_sp vn=%vn_sp vd=%vd_sp +VMLA_dp ---- 1110 0.00 .... .... 1011 .0.0 .... \ + vm=%vm_dp vn=%vn_dp vd=%vd_dp |