On Tue, Apr 12, 2011 at 3:59 PM, Richard Sandiford
<richard.sandif...@linaro.org> wrote:
> This patch adds vec_load_lanes and vec_store_lanes optabs for instructions
> like NEON's vldN and vstN. The optabs are defined this way because the
> vectors must be allocated to a block of consecutive registers.
>
> Tested on x86_64-linux-gnu and arm-linux-gnueabi. OK to install?
>
> Richard
>
>
> gcc/
> * doc/md.texi (vec_load_lanes, vec_store_lanes): Document.
> * optabs.h (COI_vec_load_lanes, COI_vec_store_lanes): New
> convert_optab_index values.
> (vec_load_lanes_optab, vec_store_lanes_optab): New convert optabs.
> * genopinit.c (optabs): Initialize the new optabs.
> * internal-fn.def (LOAD_LANES, STORE_LANES): New internal functions.
> * internal-fn.c (get_multi_vector_move, expand_LOAD_LANES)
> (expand_STORE_LANES): New functions.
> * tree.h (build_simple_array_type): Declare.
> * tree.c (build_simple_array_type): New function.
> * tree-vectorizer.h (vect_model_store_cost): Add a bool argument.
> (vect_model_load_cost): Likewise.
> (vect_store_lanes_supported, vect_load_lanes_supported)
> (vect_record_strided_load_vectors): Declare.
> * tree-vect-data-refs.c (vect_lanes_optab_supported_p)
> (vect_store_lanes_supported, vect_load_lanes_supported): New functions.
> (vect_transform_strided_load): Split out statement recording into...
> (vect_record_strided_load_vectors): ...this new function.
> * tree-vect-stmts.c (create_vector_array, read_vector_array)
> (write_vector_array, create_array_ref): New functions.
> (vect_model_store_cost): Add store_lanes_p argument.
> (vect_model_load_cost): Add load_lanes_p argument.
> (vectorizable_store): Try to use store-lanes functions for
> interleaved stores.
> (vectorizable_load): Likewise load-lanes and loads.
> * tree-vect-slp.c (vect_get_and_check_slp_defs)
> (vect_build_slp_tree):
>
> Index: gcc/doc/md.texi
> ===================================================================
> --- gcc/doc/md.texi 2011-04-12 12:16:46.000000000 +0100
> +++ gcc/doc/md.texi 2011-04-12 14:48:28.000000000 +0100
> @@ -3846,6 +3846,48 @@ into consecutive memory locations. Oper
> consecutive memory locations, operand 1 is the first register, and
> operand 2 is a constant: the number of consecutive registers.
>
> +@cindex @code{vec_load_lanes@var{m}@var{n}} instruction pattern
> +@item @samp{vec_load_lanes@var{m}@var{n}}
> +Perform an interleaved load of several vectors from memory operand 1
> +into register operand 0. Both operands have mode @var{m}. The register
> +operand is viewed as holding consecutive vectors of mode @var{n},
> +while the memory operand is a flat array that contains the same number
> +of elements. The operation is equivalent to:
> +
> +@smallexample
> +int c = GET_MODE_SIZE (@var{m}) / GET_MODE_SIZE (@var{n});
> +for (j = 0; j < GET_MODE_NUNITS (@var{n}); j++)
> + for (i = 0; i < c; i++)
> + operand0[i][j] = operand1[j * c + i];
> +@end smallexample
> +
> +For example, @samp{vec_load_lanestiv4hi} loads 8 16-bit values
> +from memory into a register of mode @samp{TI}@. The register
> +contains two consecutive vectors of mode @samp{V4HI}@.
So vec_load_lanestiv2qi would load ... ? c == 8 here. Intuitively
such operation would have adjacent blocks of siv2qi memory. But
maybe you want to constrain the mode size to GET_MODE_SIZE (@var{n})
* GET_MODE_NUNITS (@var{n})? In which case the mode m is
redundant? You could specify that we load NUNITS adjacent vectors into
an integer mode of appropriate size.
> +This pattern can only be used if:
> +@smallexample
> +TARGET_ARRAY_MODE_SUPPORTED_P (@var{n}, @var{c})
> +@end smallexample
> +is true. GCC assumes that, if a target supports this kind of
> +instruction for some mode @var{n}, it also supports unaligned
> +loads for vectors of mode @var{n}.
> +
> +@cindex @code{vec_store_lanes@var{m}@var{n}} instruction pattern
> +@item @samp{vec_store_lanes@var{m}@var{n}}
> +Equivalent to @samp{vec_load_lanes@var{m}@var{n}}, with the memory
> +and register operands reversed. That is, the instruction is
> +equivalent to:
> +
> +@smallexample
> +int c = GET_MODE_SIZE (@var{m}) / GET_MODE_SIZE (@var{n});
> +for (j = 0; j < GET_MODE_NUNITS (@var{n}); j++)
> + for (i = 0; i < c; i++)
> + operand0[j * c + i] = operand1[i][j];
> +@end smallexample
> +
> +for a memory operand 0 and register operand 1.
> +
> @cindex @code{vec_set@var{m}} instruction pattern
> @item @samp{vec_set@var{m}}
> Set given field in the vector value. Operand 0 is the vector to modify,
> Index: gcc/optabs.h
> ===================================================================
> --- gcc/optabs.h 2011-04-12 12:16:46.000000000 +0100
> +++ gcc/optabs.h 2011-04-12 14:48:28.000000000 +0100
> @@ -578,6 +578,9 @@ enum convert_optab_index
> COI_satfract,
> COI_satfractuns,
>
> + COI_vec_load_lanes,
> + COI_vec_store_lanes,
> +
Um, they are not really conversion optabs. Any reason they can't
use the direct_optab table and path? What are the two modes
usually? I don't see how you specify the kind of permutation that
is performed on the load - so, why not go the targetm.expand_builtin
path instead (well, targetm.expand_internal_fn, of course - or rather
targetm.expand_gimple_call which we need anyway for expanding
directly from gimple calls at some point).
> COI_MAX
> };
>
> @@ -598,6 +601,8 @@ #define fract_optab (&convert_optab_tabl
> #define fractuns_optab (&convert_optab_table[COI_fractuns])
> #define satfract_optab (&convert_optab_table[COI_satfract])
> #define satfractuns_optab (&convert_optab_table[COI_satfractuns])
> +#define vec_load_lanes_optab (&convert_optab_table[COI_vec_load_lanes])
> +#define vec_store_lanes_optab (&convert_optab_table[COI_vec_store_lanes])
>
> /* Contains the optab used for each rtx code. */
> extern optab code_to_optab[NUM_RTX_CODE + 1];
> Index: gcc/genopinit.c
> ===================================================================
> --- gcc/genopinit.c 2011-04-12 12:16:46.000000000 +0100
> +++ gcc/genopinit.c 2011-04-12 14:48:28.000000000 +0100
> @@ -74,6 +74,8 @@ static const char * const optabs[] =
> "set_convert_optab_handler (fractuns_optab, $B, $A,
> CODE_FOR_$(fractuns$Q$a$I$b2$))",
> "set_convert_optab_handler (satfract_optab, $B, $A,
> CODE_FOR_$(satfract$a$Q$b2$))",
> "set_convert_optab_handler (satfractuns_optab, $B, $A,
> CODE_FOR_$(satfractuns$I$a$Q$b2$))",
> + "set_convert_optab_handler (vec_load_lanes_optab, $A, $B,
> CODE_FOR_$(vec_load_lanes$a$b$))",
> + "set_convert_optab_handler (vec_store_lanes_optab, $A, $B,
> CODE_FOR_$(vec_store_lanes$a$b$))",
> "set_optab_handler (add_optab, $A, CODE_FOR_$(add$P$a3$))",
> "set_optab_handler (addv_optab, $A, CODE_FOR_$(add$F$a3$)),\n\
> set_optab_handler (add_optab, $A, CODE_FOR_$(add$F$a3$))",
> Index: gcc/internal-fn.def
> ===================================================================
> --- gcc/internal-fn.def 2011-04-12 14:10:42.000000000 +0100
> +++ gcc/internal-fn.def 2011-04-12 14:48:28.000000000 +0100
> @@ -32,3 +32,6 @@ along with GCC; see the file COPYING3.
>
> where NAME is the name of the function and FLAGS is a set of
> ECF_* flags. */
> +
> +DEF_INTERNAL_FN (LOAD_LANES, ECF_CONST | ECF_LEAF)
> +DEF_INTERNAL_FN (STORE_LANES, ECF_CONST | ECF_LEAF)
> Index: gcc/internal-fn.c
> ===================================================================
> --- gcc/internal-fn.c 2011-04-12 14:10:42.000000000 +0100
> +++ gcc/internal-fn.c 2011-04-12 14:48:28.000000000 +0100
> @@ -41,6 +41,69 @@ #define DEF_INTERNAL_FN(CODE, FLAGS) FLA
> 0
> };
>
> +/* ARRAY_TYPE is an array of vector modes. Return the associated insn
> + for load-lanes-style optab OPTAB. The insn must exist. */
> +
> +static enum insn_code
> +get_multi_vector_move (tree array_type, convert_optab optab)
> +{
> + enum insn_code icode;
> + enum machine_mode imode;
> + enum machine_mode vmode;
> +
> + gcc_assert (TREE_CODE (array_type) == ARRAY_TYPE);
> + imode = TYPE_MODE (array_type);
> + vmode = TYPE_MODE (TREE_TYPE (array_type));
> +
> + icode = convert_optab_handler (optab, imode, vmode);
> + gcc_assert (icode != CODE_FOR_nothing);
> + return icode;
> +}
> +
> +/* Expand: LHS = LOAD_LANES (ARGS[0]). */
> +
> +static void
> +expand_LOAD_LANES (tree lhs, tree *args)
> +{
> + struct expand_operand ops[2];
> + tree type;
> + rtx target, mem;
> +
> + type = TREE_TYPE (lhs);
> +
> + target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE);
> + mem = expand_normal (args[0]);
> +
> + gcc_assert (MEM_P (mem));
> + PUT_MODE (mem, TYPE_MODE (type));
> +
> + create_output_operand (&ops[0], target, TYPE_MODE (type));
> + create_fixed_operand (&ops[1], mem);
> + expand_insn (get_multi_vector_move (type, vec_load_lanes_optab), 2, ops);
> +}
> +
> +/* Expand: LHS = STORE_LANES (ARGS[0]). */
> +
> +static void
> +expand_STORE_LANES (tree lhs, tree *args)
> +{
> + struct expand_operand ops[2];
> + tree type;
> + rtx target, rhs;
> +
> + type = TREE_TYPE (args[0]);
> +
> + target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE);
> + rhs = expand_normal (args[0]);
> +
> + gcc_assert (MEM_P (target));
> + PUT_MODE (target, TYPE_MODE (type));
> +
> + create_fixed_operand (&ops[0], target);
> + create_input_operand (&ops[1], rhs, TYPE_MODE (type));
> + expand_insn (get_multi_vector_move (type, vec_store_lanes_optab), 2, ops);
> +}
> +
> /* Routines to expand each internal function, indexed by function number.
> Each routine has the prototype:
>
> Index: gcc/tree.h
> ===================================================================
> --- gcc/tree.h 2011-04-12 12:16:46.000000000 +0100
> +++ gcc/tree.h 2011-04-12 14:48:28.000000000 +0100
> @@ -4198,6 +4198,7 @@ extern tree build_type_no_quals (tree);
> extern tree build_index_type (tree);
> extern tree build_array_type (tree, tree);
> extern tree build_nonshared_array_type (tree, tree);
> +extern tree build_simple_array_type (tree, unsigned HOST_WIDE_INT);
> extern tree build_function_type (tree, tree);
> extern tree build_function_type_list (tree, ...);
> extern tree build_function_type_skip_args (tree, bitmap);
> Index: gcc/tree.c
> ===================================================================
> --- gcc/tree.c 2011-04-12 12:16:46.000000000 +0100
> +++ gcc/tree.c 2011-04-12 14:48:28.000000000 +0100
> @@ -7385,6 +7385,15 @@ build_nonshared_array_type (tree elt_typ
> return build_array_type_1 (elt_type, index_type, false);
> }
>
> +/* Return a representation of ELT_TYPE[NELTS], using indices of type
> + sizetype. */
> +
> +tree
> +build_simple_array_type (tree elt_type, unsigned HOST_WIDE_INT nelts)
build_array_type_nelts
The rest looks ok to me.
Richard.
