Re: [PATCH] ARM Neon instruction set support

[Julian Brown <julian@codesourcery.com>]

[-- Attachment #1: Type: text/plain, Size: 14491 bytes --]

Hi,

Paul Brook wrote:
> On Saturday 01 April 2006 02:39, Julian Brown wrote:
> 
>>Hi,
>>
>>This patch provides support for ARM's new Advanced SIMD (Neon)
>>instruction set, and version 3 of the VFP instruction set. Code using
>>Neon instructions can be both assembled and disassembled. The full range
>>of SIMD data types is available, and should be correctly type-checked by
>>the assembler.
> 
> 
> As-is the patch looks generally ok to me, a couple of relatively minor points 
> below. Ok for the csl branch once those are resolved.
> 
> ARM have also defined all the existing VFP instructions in terms of the vector 
> mnemonics. eg. "faddd d0, d0, d0" becomes "vadd.f64 d0, d0, d0).
> We probably want to consider the implications of this before applying to 
> mainline. Hopefully this will be an incremental change that can be done 
> without rewriting everything, but it's best to check :-)

I've not considered in detail how that bit's going to be done yet, but 
I'll hold off on mainline submission until it's implemented, if that 
seems like a good idea. I think it should be possible without too much 
perturbation though, hopefully.

>>+   /* Undo polymorphism for Neon D and Q registers.  */
>>+   if (reg && type == REG_TYPE_NDQ)
>>+     type = (reg->type == REG_TYPE_NQ || reg->type == REG_TYPE_VFD)
>>+            ? reg->type : type;
> 
> 
> Probably clearer as a single if:
> 
> if (reg && type == REG_TYPE_NDQ
>     &&(reg->type == REG_TYPE_NQ || reg->type == REG_TYPE_VFD))
>   type = reg->type;

Done.

>>!   /* FIXME: Check if cortex-a8 supports these things!  */
>>!   {"cortex-a8",               ARM_ARCH_V7A,    ARM_FEATURE(0, FPU_VFP_V3
> 
> 
> Yes it does. You can remove the FIXME.

Done.

>>+   {"vfp3",              FPU_ARCH_VFP_V3},
> 
> 
> You also need to add some way of enabling NEON. I'm not sure if it's best to 
> have this as an independent architecture extension (like iWMMXt) or just a 
> different type of FPU. Probably the latter.

I've gone with the latter for this version of the patch (against the CSL 
branch this time), though I'm not entirely conviced it's the right way 
to go. The build attribute for Neon is seperate (like the iWMMXt one). 
I've made -mfpu=neon turn on vfp3 too.

> You also need to set the appropriate EABI build attributes to when NEON and 
> VFPv3 instructions are used.

I think I uncovered a bug here: in aeabi_set_public_attributes, 
fpu_arch_vfp_v[12] was used instead of fpu_vfp_ext_v[12], which meant 
that essentially the Tag_VFP_arch tag was never set to 1 (the *arch* 
versions set bits for each VFP version supported, i.e. vfp2 sets vfp1 
and vfp1xd, so when ARM_CPU_HAS_FEATURE asks about vfpv2 and only vfpv1 
instructions were used, the result is wrongly affirmative).

If that is indeed not the desired behaviour, I expect we want that bit 
on mainline too independent of the rest of the patch.

Tested with "make check" crossed to arm-none-eabi. I'll install it on 
the CSL branch shortly. (There might still be an issue with the build 
attribute for VFPv3, but I'm 99% sure it'll be OK. If it's not, I'll 
address that with a followup patch.)

Cheers,

Julian

ChangeLog (binutils):

     * readelf.c (arm_attr_tag_VFP_arch): Add VFPv3.

ChangeLog (gas):

2006-04-01  Julian Brown  <julian@codesourcery.com>

     * config/tc-arm.c (limits.h): Include.
     (fpu_arch_vfp_v3, fpu_vfp_ext_v3, fpu_neon_ext_v1)
     (fpu_vfp_v3_or_neon_ext): Declare constants.
     (neon_el_type): New enumeration of types for Neon vector
     elements.
     (neon_type_el): New struct. Define type and size of a vector
     element.
     (NEON_MAX_TYPE_ELS): Define constant. The maximum number of types
     per instruction.
     (neon_type): Define struct. The type of an instruction.
     (arm_it): Add 'vectype' for the current instruction.
     (isscalar, immisalign, regisimm, isquad): New predicates for
     operands.
     (vfp_sp_reg_pos): Rename to...
     (vfp_reg_pos): ...this, and add VFP_REG_Dd, VFP_REG_Dm, VFP_REG_Dn
     tags.
     (arm_reg_type): Add REG_TYPE_NQ (Neon Q register) and REG_TYPE_NDQ
     (Neon D or Q register).
     (reg_expected_msgs): Sync with above. Allow VFD to mean VFP or Neon
     D register.
     (GE_OPT_PREFIX_BIG): Define constant, for use in...
     (my_get_expression): Allow above constant as argument to accept
     64-bit constants with optional prefix.
     (arm_reg_parse): Add extra argument to return the specific type
     of register in when either a D or Q register (REG_TYPE_NDQ) is
     requested. Can be NULL.
     (parse_scalar): New function. Parse Neon scalar (vector reg and
     index).
     (parse_reg_list): Update for new arm_reg_parse args.
     (parse_vfp_reg_list): Allow parsing of Neon D/Q register lists.
     (parse_neon_el_struct_list): New function. Parse element/structure
     register lists for VLD<n>/VST<n> instructions.
     (s_arm_unwind_save_vfp): Update for new parse_vfp_reg_list args.
     (s_arm_unwind_save_mmxwr): Likewise.
     (s_arm_unwind_save_mmxwcg): Likewise.
     (s_arm_unwind_movsp): Likewise.
     (s_arm_unwind_setfp): Likewise.
     (parse_big_immediate): New function. Parse an immediate, which may
     be 64 bits wide. Put results in inst.operands[i].
     (parse_shift): Update for new arm_reg_parse args.
     (parse_address): Likewise. Add parsing of alignment specifiers.
     (parse_neon_mov): Parse the operands of a VMOV instruction.
     (operand_parse_code): Add OP_RND, OP_RNQ, OP_RNDQ, OP_RNSC,
     OP_NRDLST, OP_NSTRLST, OP_NILO, OP_RNDQ_I0, OP_RR_RNSC,
     OP_RNDQ_RNSC, OP_RND_RNSC, OP_VMOV, OP_RNDQ_IMVNb, OP_RNDQ_I63b,
     OP_I0, OP_I16z, OP_I32z, OP_I64, OP_I64z, OP_oI32b, OP_oRND,
     OP_oRNQ, OP_oRNDQ.
     (parse_operands): Handle new codes above.
     (encode_arm_vfp_sp_reg): Rename to...
     (encode_arm_vfp_reg): ...this. Handle D regs (0-31) too. Complain if
     selected VFP version only supports D0-D15.
     (do_vfp_sp_monadic, do_vfp_sp_dyadic, do_vfp_sp_compare_z)
     (do_vfp_dp_sp_cvt, do_vfp_reg_from_sp, do_vfp_reg2_from_sp2)
     (do_vfp_sp_from_reg, do_vfp_sp2_from_reg2, do_vfp_sp_ldst)
     (do_vfp_dp_ldst, vfp_sp_ldstm, vfp_dp_ldstm): Update for new
     encode_arm_vfp_reg name, and allow 32 D regs.
     (do_vfp_dp_rd_rm, do_vfp_dp_rn_rd, do_vfp_dp_rd_rn)
     (do_vfp_dp_rd_rn_rm, do_vfp_rm_rd_rn): New functions to encode VFP
     insns allowing 32 D regs.
     (do_vfp_sp_const, do_vfp_dp_const, vfp_conv, do_vfp_sp_conv_16)
     (do_vfp_dp_conv_16, do_vfp_sp_conv_32, do_vfp_dp_conv_32): Handle
     constant-load and conversion insns introduced with VFPv3.
     (neon_tab_entry): New struct.
     (NEON_ENC_TAB): Bit patterns for overloaded Neon instructions, and
     those which are the targets of pseudo-instructions.
     (neon_opc): Enumerate opcodes, use as indices into...
     (neon_enc_tab): ...this. Hold data from NEON_ENC_TAB.
     (NEON_ENC_INTEGER, NEON_ENC_ARMREG, NEON_ENC_POLY, NEON_ENC_FLOAT)
     (NEON_ENC_SCALAR, NEON_ENC_IMMED, NEON_ENC_INTERLV, NEON_ENC_LANE)
     (NEON_ENC_DUP): Define meaningful helper macros to look up values in
     neon_enc_tab.
     (neon_shape): Enumerate shapes (permitted register widths, etc.) for
     Neon instructions.
     (neon_type_mask): New. Compact type representation for type
     checking.
     (N_SU_ALL, N_SU_32, N_SU_16_64, N_SUF_32, N_I_ALL, N_IF_32): Common
     permitted type combinations.
     (N_IGNORE_TYPE): New macro.
     (neon_check_shape): New function. Check an instruction shape for
     multiple alternatives. Return the specific shape for the current
     instruction.
     (neon_modify_type_size): New function. Modify a vector type and
     size, depending on the bit mask in argument 1.
     (neon_type_promote): New function. Convert a given "key" type (of an
     operand) into the correct type for a different operand, based on a
     bit mask.
     (type_chk_of_el_type): New function. Convert a type and size into
     the compact representation used for type checking.
     (el_type_of_type_ckh): New function. Reverse of above (only when a
     single bit is set in the bit mask).
     (modify_types_allowed): New function. Alter a mask of allowed types
     based on a bit mask of modifications.
     (neon_check_type): New function. Check the type of the current
     instruction against the variable argument list. The "key" type of
     the instruction is returned.
     (neon_dp_fixup): New function. Fill in and modify instruction bits
     for a Neon data-processing instruction depending on whether we're in
     ARM mode or Thumb-2 mode.
     (neon_logbits): New function.
     (neon_three_same, neon_two_same, do_neon_dyadic_i_su)
     (do_neon_dyadic_i64_su, neon_imm_shift, do_neon_shl_imm)
     (do_neon_qshl_imm, neon_cmode_for_logic_imm)
     (neon_bits_same_in_bytes, neon_squash_bits, neon_is_quarter_float)
     (neon_qfloat_bits, neon_cmode_for_move_imm, neon_write_immbits)
     (neon_invert_size, do_neon_logic, do_neon_bitfield, neon_dyadic)
     (do_neon_dyadic_if_su, do_neon_dyadic_if_su_d, do_neon_dyadic_if_i)
     (do_neon_dyadic_if_i_d, do_neon_addsub_if_i, neon_exchange_operands)
     (neon_compare, do_neon_cmp, do_neon_cmp_inv, do_neon_ceq)
     (neon_scalar_for_mul, neon_mul_mac, do_neon_mac_maybe_scalar)
     (do_neon_tst, do_neon_mul, do_neon_qdmulh, do_neon_fcmp_absolute)
     (do_neon_fcmp_absolute_inv, do_neon_step, do_neon_abs_neg)
     (do_neon_sli, do_neon_sri, do_neon_qshlu_imm, do_neon_qmovn)
     (do_neon_qmovun, do_neon_rshift_sat_narrow)
     (do_neon_rshift_sat_narrow_u, do_neon_movn, do_neon_rshift_narrow)
     (do_neon_shll, neon_cvt_flavour, do_neon_cvt, neon_move_immediate)
     (do_neon_mvn, neon_mixed_length, do_neon_dyadic_long, do_neon_abal)
     (neon_mac_reg_scalar_long, do_neon_mac_maybe_scalar_long)
     (do_neon_dyadic_wide, do_neon_vmull, do_neon_ext, do_neon_rev)
     (do_neon_dup, do_neon_mov, do_neon_rshift_round_imm, do_neon_movl)
     (do_neon_trn, do_neon_zip_uzp, do_neon_sat_abs_neg)
     (do_neon_pair_long, do_neon_recip_est, do_neon_cls, do_neon_clz)
     (do_neon_cnt, do_neon_swp, do_neon_tbl_tbx, do_neon_ldm_stm)
     (do_neon_ldr_str, do_neon_ld_st_interleave, neon_alignment_bit)
     (do_neon_ld_st_lane, do_neon_ld_dup, do_neon_ldx_stx): New
     functions. Neon bit encoding and encoding helpers.
     (parse_neon_type): New function. Parse Neon type specifier.
     (opcode_lookup): Allow parsing of Neon type specifiers.
     (REGNUM2, REGSETH, REGSET2): New macros.
     (reg_names): Add new VFPv3 and Neon registers.
     (NUF, nUF, NCE, nCE): New macros for opcode table.
     (insns): More VFP registers allowed in fcpyd, fmdhr, fmdlr, fmrdh,
     fmrdl, fabsd, fnegd, fsqrtd, faddd, fsubd, fmuld, fdivd, fmacd,
     fmscd, fnmuld, fnmacd, fnmscd, fcmpd, fcmpzd, fcmped, fcmpezd,
     fmdrr, fmrrd. Add Neon instructions vaba, vhadd, vrhadd, vhsub,
     vqadd, vqsub, vrshl, vqrshl, vshl, vqshl{u}, vand, vbic, vorr, vorn,
     veor, vbsl, vbit, vbif, vabd, vmax, vmin, vcge, vcgt, vclt, vcle,
     vceq, vpmax, vpmin, vmla, vmls, vpadd, vadd, vsub, vtst, vmul,
     vqdmulh, vqrdmulh, vacge, vacgt, vaclt, vacle, vrecps, vrsqrts,
     vabs, vneg, v{r}shr,  v{r}sra, vsli, vsri, vqshrn, vq{r}shr{u}n,
     v{r}shrn, vshll, vcvt, vmov, vmvn, vabal, vabdl, vaddl, vsubl,
     vmlal, vmlsl, vaddw, vsubw, v{r}addhn, v{r}subhn, vqdmlal, vqdmlsl,
     vqdmull, vmull, vext, vrev64, vrev32, vrev16, vdup, vmovl, v{q}movn,
     vzip, vuzp, vqabs, vqneg, vpadal, vpaddl, vrecpe, vrsqrte, vcls,
     vclz, vcnt, vswp, vtrn, vtbl, vtbx, vldm, vstm, vldr, vstr,
     vld[1234], vst[1234], fconst[sd], f[us][lh]to[sd], fto[us][lh][sd].
     (tc_arm_regname_to_dw2regnum): Update for arm_reg_parse args.
     (arm_cpu_option_table): Add Neon and VFPv3 to Cortex-A8.
     (arm_option_cpu_value): Add vfp3 and neon.
     (aeabi_set_public_attributes): Support VFPv3 and NEON attributes.
     Fix VFPv1 attribute.


ChangeLog (gas/testsuite):

2006-04-01  Julian Brown  <julian@codesourcery.com>

     * gas/arm/copro.s: Avoid ldcl which encodes as a bad Neon
     instruction.
     * gas/arm/copro.d: Update accordingly.
     * gas/arm/neon-cond.s: New test. Conditional Neon opcodes in ARM
    mode.
     * gas/arm/neon-cond.d: Expected results of above.
     * gas/arm/neon-cov.s: New test. Coverage of Neon instructions.
     * gas/arm/neon-cov.d: Expected results of above.
     * gas/arm/neon-ldst-es.s: New test. Element and structure loads and
     stores.
     * gas/arm/neon-ldst-es.d: Expected results of above.
     * gas/arm/neon-ldst-rm.s: New test. Single and multiple register
     loads and stores.
     * gas/arm/neon-ldst-rm.d: Expected results of above.
     * gas/arm/neon-omit.s: New test. Omission of optional operands.
     * gas/arm/neon-omit.d: Expected results of above.
     * gas/arm/vfp1.d: Expect Neon syntax for some VFP instructions.
     * gas/arm/vfp1_t2.d: Likewise.
     * gas/arm/vfp1xD.d: Likewise.
     * gas/arm/vfp1xD_t2.d: Likewise.
     * gas/arm/vfp2.d: Likewise.
     * gas/arm/vfp2_t2.d: Likewise.
     * gas/arm/vfp3-32drs.s: New test. Extended D register range for VFP
     instructions.
     * gas/arm/vfp3-32drs.d: Expected results of above.
     * gas/arm/vfp3-const-conv.s: New test. VFPv3 constant-load and
     conversion instructions.
     * gas/arm/vfp3-const-conv.d: Expected results of above.

ChangeLog (include):

2006-04-01  Julian Brown  <julian@codesourcery.com>

     * opcode/arm.h (FPU_VFP_EXT_V3): Define constant.
     (FPU_NEON_EXT_V1): Likewise.
     (FPU_VFP_HARD): Update.
     (FPU_VFP_V3): Define macro.
     (FPU_ARCH_VFP_V3, FPU_ARCH_VFP_V3_PLUS_NEON_V1): Define macros.

ChangeLog (opcodes/arm-dis.c):

2006-04-01  Nathan Sidwell  <nathan@codesourcery.com>
             Julian Brown  <julian@codesourcery.com>

     * opcodes/arm-dis.c (coprocessor_opcodes): Add %A, %B, %k,
     convert %<code>[zy] into %[zy]<code>.  Expand meaning of
     %<bitfield>['`?].
     Add unified load/store instruction names.
     (neon_opcode_table): New.
     (arm_opcodes): Expand meaning of %<bitfield>['`?].
     (arm_decode_bitfield): New.
     (print_insn_coprocessor): Add pc argument. Add %A & %B
     specifiers. Use arm_decode_bitfield and adjust numeric
     specifiers.
     Adjust %z & %y.
     (print_insn_neon): New.
     (print_insn_arm): Adjust print_insn_coprocessor call. Call
     print_insn_neon. Use arm_decode_bitfield and adjust numeric
     specifiers.
     (print_insn_thumb32): Likewise.


[-- Attachment #2: neon-support-4.gz --]
[-- Type: application/x-gzip, Size: 65965 bytes --]