gcc/
* Makefile.in (TEXI_GCC_FILES): Add arm-neon-intrinsics.texi.
* config.gcc (arm*-*-*): Add arm_neon.h to extra headers.
(with_fpu): Allow --with-fpu=neon.
* config/arm/aof.h (ADDITIONAL_REGISTER_NAMES): Add Q0-Q15.
* config/arm/aout.h (ADDITIONAL_REGISTER_NAMES): Add Q0-Q15.
* config/arm/arm-modes.def (EI, OI, CI, XI): New modes.
* config/arm/arm-protos.h (neon_immediate_valid_for_move)
(neon_immediate_valid_for_logic, neon_output_logic_immediate)
(neon_pairwise_reduce, neon_expand_vector_init, neon_reinterpret)
(neon_emit_pair_result_insn, neon_disambiguate_copy)
(neon_vector_mem_operand, neon_struct_mem_operand, output_move_quad)
(output_move_neon): Add prototypes.
* config/arm/arm.c (FL_NEON): New flag for NEON processor capability.
(all_fpus): Add FPUTYPE_NEON.
(fp_model_for_fpu): Add NEON field.
(arm_return_in_memory): Return vectors <= 16 bytes in ARM registers.
(arm_arg_partial_bytes): Allow NEON vectors to be passed partially
in registers.
(arm_legitimate_address_p): Don't support fancy addressing for NEON
structure moves.
(thumb2_legitimate_address_p): Likewise.
(neon_valid_immediate): Recognize and prepare constants suitable for
NEON instructions.
(neon_immediate_valid_for_move): New function. Recognize and prepare
immediates for NEON move instructions.
(neon_immediate_valid_for_logic): New function. Recognize and
prepare immediates for NEON logic instructions.
(neon_output_logic_immediate): New function. Create asm string
suitable for outputting immediate logic instructions.
(neon_pairwise_reduce): New function. Implement reduction using
pairwise operations.
(neon_expand_vector_init): New function. Expand a (possibly
non-constant) vector initialization.
(neon_vector_mem_operand): New function. Memory operands supported
for quad-word loads/stores to/from ARM or NEON registers. Don't
allow base+offset addressing for core regs.
(neon_struct_mem_operand): New function. Valid mems for NEON
structure moves.
(coproc_secondary_reload_class): Enable NEON registers to be loaded
from neon_vector_mem_operand addresses without a secondary register.
(add_minipool_forward_ref): Handle >8-byte minipool entries.
(add_minipool_backward_ref): Likewise.
(dump_minipool): Likewise.
(push_minipool_fix): Likewise.
(output_move_quad): New function. Output quad-word moves, loads and
stores using ARM registers.
(output_move_vfp): Add support for vectors in VFP (NEON) D
registers.
(output_move_neon): Output a NEON load/store to/from a quadword
register.
(arm_print_operand): Implement new codes:
- 'c' for unadorned integers (without a # sign).
- 'J', 'K' for reg+2/reg+3, reg+3/reg+2 in little/big-endian
mode.
- 'e', 'f' for the low and high D parts of a NEON Q register.
- 'q' outputs a NEON Q register.
- 'h' outputs ranges of D registers for VLDM/VSTM etc.
- 'T' prints NEON opcode features from a coded bitmask.
- 'F' is similar to T, but signed/unsigned codes both print as
'i'.
- 't' is similar to T, but 'u' is printed instead of 'p'.
- 'O' prints 'r' if NEON instruction should perform rounding (as
specified by bitmask), else prints nothing.
- '#' is a punctuation character to stop operand numbers from
running together with following digits in the assembler
strings for instructions (when using mode attributes).
(arm_assemble_integer): Handle extra NEON vector modes. Permute
constant vectors in big-endian mode, where necessary.
(arm_hard_regno_mode_ok): Allow vectors in VFP/NEON registers.
Handle EI, OI, CI, XI modes.
(ashlv4hi3, ashlv2si3, lshrv4hi3, lshrv2si3, ashrv4hi3)
(ashrv2si3): Rename IWMMXT2_BUILTINs to...
(ashlv4hi3_iwmmxt, ashlv2si3_iwmmxt, lshrv4hi3_iwmmxt)
(lshrv2si3_iwmmxt, ashrv4hi3_iwmmxt, ashrv2si3_iwmmxt): New names.
(neon_builtin_type_bits): Add enumeration, one bit for each vector
type.
(v8qi_UP, v4hi_UP, v2si_UP, v2sf_UP, di_UP, v16qi_UP, v8hi_UP)
(v4si_UP, v4sf_UP, v2di_UP, ti_UP, ei_UP, oi_UP, UP): Define macros
to turn v8qi, etc. into bits defined above.
(neon_itype): New enumeration. Classifications of NEON builtins.
(neon_builtin_datum): Define struct. Contains information about
a single builtin (with multiple modes).
(CF): Define helper macro for...
(VAR1...VAR10): Define builtins with a type, name and 1-10 different
modes.
(neon_builtin_data): New array. Define information about builtins
for use during initialization/expansion.
(arm_init_neon_builtins): New function.
(arm_init_builtins): Call arm_init_neon_builtins if TARGET_NEON is
true.
(neon_builtin_compare): New function.
(locate_neon_builtin_icode): New function. Find an insn code for a
builtin given a function code for that builtin. Also return type of
builtin (NEON_BINOP, NEON_UNOP etc.).
(builtin_arg): New enumeration. Types of arguments for builtins.
(arm_expand_neon_args): New function. Expand a generic NEON builtin.
Takes a variable argument list of builtin_arg types, terminated by
NEON_ARG_STOP.
(arm_expand_neon_builtin): New function. Expand a NEON builtin.
(neon_reinterpret): New function. Expand NEON reinterpret intrinsic.
(neon_emit_pair_result_insn): New function. Support returning pairs
of vectors via a pointer.
(neon_disambiguate_copy): New function. Set up operands for a
multi-word copy such that registers do not get clobbered.
(arm_expand_builtin): Call arm_expand_neon_builtin if fcode >=
ARM_BUILTIN_NEON_BASE.
(arm_file_start): Set float-abi attribute for NEON.
(arm_vector_mode_supported_p): Enable NEON vector modes.
(arm_mangle_map_entry): New.
(arm_mangle_map): New.
(arm_mangle_vector_type): New.
* config/arm/arm.h (TARGET_CPU_CPP_BUILTINS): Define __ARM_NEON__
when appropriate.
(TARGET_NEON): New macro. Target supports NEON.
(fputype): Add FPUTYPE_NEON.
(UNITS_PER_SIMD_WORD): Define. Allow quad-word registers to be used
for vectorization based on command-line arg.
(NEON_REGNO_OK_FOR_NREGS): Define.
(VALID_NEON_DREG_MODE, VALID_NEON_QREG_MODE)
(VALID_NEON_STRUCT_MODE): Define.
(PRINT_OPERAND_PUNCT_VALID_P): '#' is valid punctuation.
(arm_builtins): Add ARM_BUILTIN_NEON_BASE.
* config/arm/arm.md (VUNSPEC_POOL_16): Insert constant for unspec.
(consttable_16): Add pattern for outputting 16-byte minipool
entries.
(movv2si, movv4hi, movv8qi): Remove blank expanders (redefined in
vec-common.md).
(vec-common.md, neon.md): Include md files.
* config/arm/arm.opt (mvectorize-with-neon-quad): Add option.
* config/arm/constraints.md (constraint "Dn", "Dl", "DL"): Define.
(memory_constraint "Ut", "Un", "Us"): Define.
* config/arm/iwmmxt.md (VMMX, VSHFT): New mode macros.
(MMX_char): New mode attribute.
(addv8qi3, addv4hi3, addv2si3): Remove. Replace with...
(*add<mode>3_iwmmxt): New insn pattern.
(subv8qi3, subv4hi3, subv2si3): Remove. Replace with...
(*sub<mode>3_iwmmxt): New insn pattern.
(mulv4hi3): Rename to...
(*mulv4hi3_iwmmxt): This.
(smaxv8qi3, smaxv4hi3, smaxv2si3, umaxv8qi3, umaxv4hi3)
(umaxv2si3, sminv8qi3, sminv4hi3, sminv2si3, uminv8qi3)
(uminv4hi3, uminv2si3): Remove. Replace with...
(*smax<mode>3_iwmmxt, *umax<mode>3_iwmmxt, *smin<mode>3_iwmmxt)
(*umin<mode>3_iwmmxt): These.
(ashrv4hi3, ashrv2si3, ashrdi3_iwmmxt): Replace with...
(ashr<mode>3_iwmmxt): This new pattern.
(lshrv4hi3, lshrv2si3, lshrdi3_iwmmxt): Replace with...
(lshr<mode>3_iwmmxt): This new pattern.
(ashlv4hi3, ashlv2si3, ashldi3_iwmmxt): Replace with...
(ashl<mode>3_iwmmxt): This new pattern.
* config/arm/neon-docgen.ml: New file. Generate documentation for
intrinsics.
* config/arm/neon-gen.ml: New file. Generate arm_neon.h header.
* config/arm/arm_neon.h: New (autogenerated).
* config/arm/neon-testgen.ml: New file. Generate NEON tests
automatically.
* config/arm/neon.md: New file. Define NEON instructions.
* config/arm/neon.ml: New file. Abstract description of NEON
instructions, used to generate arm_neon.h header, documentation and tests.
* config/arm/t-arm (MD_INCLUDES): Add vec-common.md, neon.md.
* vec-common.md: New file. Shared parts for iWMMXt and NEON vector
support.
* doc/extend.texi (ARM Built-in Functions): Rename and remove
extraneous comma.
(ARM NEON Intrinsics): New subsection.
* doc/arm-neon-intrinsics.texi: New (autogenerated).
gcc/testsuite/
* gcc.dg/vect/vect.exp: Check is-effective-target arm_neon_hw.
* gcc.dg/vect/tree-vect.h: Check for NEON SIMD support.
* lib/gcc-dg.exp (cleanup-saved-temps): Fix comment.
* lib/target-supports.exp (check_effective_target_arm_neon_ok)
(check_effective_target_arm_neon_hw): New.
* gcc.target/arm/neon/neon.exp: New file.
* gcc.target/arm/neon/polytypes.c: New file.
* gcc.target/arm/neon/v*.c (1870 files): New (autogenerated).
Co-Authored-By: Joseph Myers <joseph@codesourcery.com>
Co-Authored-By: Mark Shinwell <shinwell@codesourcery.com>
Co-Authored-By: Paul Brook <paul@codesourcery.com>
From-SVN: r126911
Daniel Jacobowitz
Julian Brown