Makefile.in (TEXI_GCC_FILES): Add arm-neon-intrinsics.texi.

    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

1 files changed, 60 insertions, 4 deletions

diff --git a/gcc/config/arm/arm.h b/gcc/config/arm/arm.h
index b9c6e85..6c4d95e 100644
--- a/gcc/config/arm/arm.h
+++ b/gcc/config/arm/arm.h
@@ -65,6 +65,9 @@ extern char arm_arch_name[];
 	if (TARGET_VFP)					\
 	  builtin_define ("__VFP_FP__");		\
 							\
+	if (TARGET_NEON)				\
+	  builtin_define ("__ARM_NEON__");		\
+							\
 	/* Add a define for interworking.		\
 	   Needed when building libgcc.a.  */		\
 	if (arm_cpp_interwork)				\
@@ -206,10 +209,23 @@ extern GTY(()) rtx aof_pic_label;
 /* 32-bit Thumb-2 code.  */
 #define TARGET_THUMB2			(TARGET_THUMB && arm_arch_thumb2)
 
+/* The following two macros concern the ability to execute coprocessor
+   instructions for VFPv3 or NEON.  TARGET_VFP3 is currently only ever
+   tested when we know we are generating for VFP hardware; we need to
+   be more careful with TARGET_NEON as noted below.  */
+
 /* FPU is VFPv3 (with twice the number of D registers).  Setting the FPU to
    Neon automatically enables VFPv3 too.  */
 #define TARGET_VFP3 (arm_fp_model == ARM_FP_MODEL_VFP \
-		     && (arm_fpu_arch == FPUTYPE_VFP3))
+		     && (arm_fpu_arch == FPUTYPE_VFP3 \
+			 || arm_fpu_arch == FPUTYPE_NEON))
+/* FPU supports Neon instructions.  The setting of this macro gets
+   revealed via __ARM_NEON__ so we add extra guards upon TARGET_32BIT
+   and TARGET_HARD_FLOAT to ensure that NEON instructions are
+   available.  */
+#define TARGET_NEON (TARGET_32BIT && TARGET_HARD_FLOAT \
+		     && arm_fp_model == ARM_FP_MODEL_VFP \
+		     && arm_fpu_arch == FPUTYPE_NEON)
 
 /* "DSP" multiply instructions, eg. SMULxy.  */
 #define TARGET_DSP_MULTIPLY \
@@ -282,7 +298,9 @@ enum fputype
   /* VFP.  */
   FPUTYPE_VFP,
   /* VFPv3.  */
-  FPUTYPE_VFP3
+  FPUTYPE_VFP3,
+  /* Neon.  */
+  FPUTYPE_NEON
 };
 
 /* Recast the floating point class to be the floating point attribute.  */
@@ -483,6 +501,12 @@ extern int arm_arch_hwdiv;
 
 #define UNITS_PER_WORD	4
 
+/* Use the option -mvectorize-with-neon-quad to override the use of doubleword
+   registers when autovectorizing for Neon, at least until multiple vector
+   widths are supported properly by the middle-end.  */
+#define UNITS_PER_SIMD_WORD \
+  (TARGET_NEON ? (TARGET_NEON_VECTORIZE_QUAD ? 16 : 8) : UNITS_PER_WORD)
+
 /* True if natural alignment is used for doubleword types.  */
 #define ARM_DOUBLEWORD_ALIGN	TARGET_AAPCS_BASED
 
@@ -941,6 +965,18 @@ extern int arm_structure_size_boundary;
 #define VFP_REGNO_OK_FOR_DOUBLE(REGNUM) \
   ((((REGNUM) - FIRST_VFP_REGNUM) & 1) == 0)
 
+/* Neon Quad values must start at a multiple of four registers.  */
+#define NEON_REGNO_OK_FOR_QUAD(REGNUM) \
+  ((((REGNUM) - FIRST_VFP_REGNUM) & 3) == 0)
+
+/* Neon structures of vectors must be in even register pairs and there
+   must be enough registers available.  Because of various patterns
+   requiring quad registers, we require them to start at a multiple of
+   four.  */
+#define NEON_REGNO_OK_FOR_NREGS(REGNUM, N) \
+  ((((REGNUM) - FIRST_VFP_REGNUM) & 3) == 0 \
+   && (LAST_VFP_REGNUM - (REGNUM) >= 2 * (N) - 1))
+
 /* The number of hard registers is 16 ARM + 8 FPA + 1 CC + 1 SFP + 1 AFP.  */
 /* + 16 Cirrus registers take us up to 43.  */
 /* Intel Wireless MMX Technology registers add 16 + 4 more.  */
@@ -994,6 +1030,21 @@ extern int arm_structure_size_boundary;
 #define VALID_IWMMXT_REG_MODE(MODE) \
  (arm_vector_mode_supported_p (MODE) || (MODE) == DImode)
 
+/* Modes valid for Neon D registers.  */
+#define VALID_NEON_DREG_MODE(MODE) \
+  ((MODE) == V2SImode || (MODE) == V4HImode || (MODE) == V8QImode \
+   || (MODE) == V2SFmode || (MODE) == DImode)
+
+/* Modes valid for Neon Q registers.  */
+#define VALID_NEON_QREG_MODE(MODE) \
+  ((MODE) == V4SImode || (MODE) == V8HImode || (MODE) == V16QImode \
+   || (MODE) == V4SFmode || (MODE) == V2DImode)
+
+/* Structure modes valid for Neon registers.  */
+#define VALID_NEON_STRUCT_MODE(MODE) \
+  ((MODE) == TImode || (MODE) == EImode || (MODE) == OImode \
+   || (MODE) == CImode || (MODE) == XImode)
+
 /* The order in which register should be allocated.  It is good to use ip
    since no saving is required (though calls clobber it) and it never contains
    function parameters.  It is quite good to use lr since other calls may
@@ -2409,7 +2460,7 @@ extern int making_const_table;
 
 #define PRINT_OPERAND_PUNCT_VALID_P(CODE)	\
   (CODE == '@' || CODE == '|' || CODE == '.'	\
-   || CODE == '(' || CODE == ')'		\
+   || CODE == '(' || CODE == ')' || CODE == '#'	\
    || (TARGET_32BIT && (CODE == '?'))		\
    || (TARGET_THUMB2 && (CODE == '!'))		\
    || (TARGET_THUMB && (CODE == '_')))
@@ -2581,6 +2632,9 @@ extern int making_const_table;
    : arm_gen_return_addr_mask ())
 
 
+/* Neon defines builtins from ARM_BUILTIN_MAX upwards, though they don't have
+   symbolic names defined here (which would require too much duplication).
+   FIXME?  */
 enum arm_builtins
 {
   ARM_BUILTIN_GETWCX,
@@ -2745,7 +2799,9 @@ enum arm_builtins
 
   ARM_BUILTIN_THREAD_POINTER,
 
-  ARM_BUILTIN_MAX
+  ARM_BUILTIN_NEON_BASE,
+
+  ARM_BUILTIN_MAX = ARM_BUILTIN_NEON_BASE  /* FIXME: Wrong!  */
 };
 
 /* Do not emit .note.GNU-stack by default.  */