| 1 |
280 |
jeremybenn |
/* Lower vector operations to scalar operations.
|
| 2 |
|
|
Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009
|
| 3 |
|
|
Free Software Foundation, Inc.
|
| 4 |
|
|
|
| 5 |
|
|
This file is part of GCC.
|
| 6 |
|
|
|
| 7 |
|
|
GCC is free software; you can redistribute it and/or modify it
|
| 8 |
|
|
under the terms of the GNU General Public License as published by the
|
| 9 |
|
|
Free Software Foundation; either version 3, or (at your option) any
|
| 10 |
|
|
later version.
|
| 11 |
|
|
|
| 12 |
|
|
GCC is distributed in the hope that it will be useful, but WITHOUT
|
| 13 |
|
|
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
| 14 |
|
|
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
| 15 |
|
|
for more details.
|
| 16 |
|
|
|
| 17 |
|
|
You should have received a copy of the GNU General Public License
|
| 18 |
|
|
along with GCC; see the file COPYING3. If not see
|
| 19 |
|
|
<http://www.gnu.org/licenses/>. */
|
| 20 |
|
|
|
| 21 |
|
|
#include "config.h"
|
| 22 |
|
|
#include "system.h"
|
| 23 |
|
|
#include "coretypes.h"
|
| 24 |
|
|
#include "tree.h"
|
| 25 |
|
|
#include "tm.h"
|
| 26 |
|
|
#include "rtl.h"
|
| 27 |
|
|
#include "expr.h"
|
| 28 |
|
|
#include "insn-codes.h"
|
| 29 |
|
|
#include "diagnostic.h"
|
| 30 |
|
|
#include "optabs.h"
|
| 31 |
|
|
#include "machmode.h"
|
| 32 |
|
|
#include "langhooks.h"
|
| 33 |
|
|
#include "tree-flow.h"
|
| 34 |
|
|
#include "gimple.h"
|
| 35 |
|
|
#include "tree-iterator.h"
|
| 36 |
|
|
#include "tree-pass.h"
|
| 37 |
|
|
#include "flags.h"
|
| 38 |
|
|
#include "ggc.h"
|
| 39 |
|
|
|
| 40 |
|
|
|
| 41 |
|
|
/* Build a constant of type TYPE, made of VALUE's bits replicated
|
| 42 |
|
|
every TYPE_SIZE (INNER_TYPE) bits to fit TYPE's precision. */
|
| 43 |
|
|
static tree
|
| 44 |
|
|
build_replicated_const (tree type, tree inner_type, HOST_WIDE_INT value)
|
| 45 |
|
|
{
|
| 46 |
|
|
int width = tree_low_cst (TYPE_SIZE (inner_type), 1);
|
| 47 |
|
|
int n = HOST_BITS_PER_WIDE_INT / width;
|
| 48 |
|
|
unsigned HOST_WIDE_INT low, high, mask;
|
| 49 |
|
|
tree ret;
|
| 50 |
|
|
|
| 51 |
|
|
gcc_assert (n);
|
| 52 |
|
|
|
| 53 |
|
|
if (width == HOST_BITS_PER_WIDE_INT)
|
| 54 |
|
|
low = value;
|
| 55 |
|
|
else
|
| 56 |
|
|
{
|
| 57 |
|
|
mask = ((HOST_WIDE_INT)1 << width) - 1;
|
| 58 |
|
|
low = (unsigned HOST_WIDE_INT) ~0 / mask * (value & mask);
|
| 59 |
|
|
}
|
| 60 |
|
|
|
| 61 |
|
|
if (TYPE_PRECISION (type) < HOST_BITS_PER_WIDE_INT)
|
| 62 |
|
|
low &= ((HOST_WIDE_INT)1 << TYPE_PRECISION (type)) - 1, high = 0;
|
| 63 |
|
|
else if (TYPE_PRECISION (type) == HOST_BITS_PER_WIDE_INT)
|
| 64 |
|
|
high = 0;
|
| 65 |
|
|
else if (TYPE_PRECISION (type) == 2 * HOST_BITS_PER_WIDE_INT)
|
| 66 |
|
|
high = low;
|
| 67 |
|
|
else
|
| 68 |
|
|
gcc_unreachable ();
|
| 69 |
|
|
|
| 70 |
|
|
ret = build_int_cst_wide (type, low, high);
|
| 71 |
|
|
return ret;
|
| 72 |
|
|
}
|
| 73 |
|
|
|
| 74 |
|
|
static GTY(()) tree vector_inner_type;
|
| 75 |
|
|
static GTY(()) tree vector_last_type;
|
| 76 |
|
|
static GTY(()) int vector_last_nunits;
|
| 77 |
|
|
|
| 78 |
|
|
/* Return a suitable vector types made of SUBPARTS units each of mode
|
| 79 |
|
|
"word_mode" (the global variable). */
|
| 80 |
|
|
static tree
|
| 81 |
|
|
build_word_mode_vector_type (int nunits)
|
| 82 |
|
|
{
|
| 83 |
|
|
if (!vector_inner_type)
|
| 84 |
|
|
vector_inner_type = lang_hooks.types.type_for_mode (word_mode, 1);
|
| 85 |
|
|
else if (vector_last_nunits == nunits)
|
| 86 |
|
|
{
|
| 87 |
|
|
gcc_assert (TREE_CODE (vector_last_type) == VECTOR_TYPE);
|
| 88 |
|
|
return vector_last_type;
|
| 89 |
|
|
}
|
| 90 |
|
|
|
| 91 |
|
|
/* We build a new type, but we canonicalize it nevertheless,
|
| 92 |
|
|
because it still saves some memory. */
|
| 93 |
|
|
vector_last_nunits = nunits;
|
| 94 |
|
|
vector_last_type = type_hash_canon (nunits,
|
| 95 |
|
|
build_vector_type (vector_inner_type,
|
| 96 |
|
|
nunits));
|
| 97 |
|
|
return vector_last_type;
|
| 98 |
|
|
}
|
| 99 |
|
|
|
| 100 |
|
|
typedef tree (*elem_op_func) (gimple_stmt_iterator *,
|
| 101 |
|
|
tree, tree, tree, tree, tree, enum tree_code);
|
| 102 |
|
|
|
| 103 |
|
|
static inline tree
|
| 104 |
|
|
tree_vec_extract (gimple_stmt_iterator *gsi, tree type,
|
| 105 |
|
|
tree t, tree bitsize, tree bitpos)
|
| 106 |
|
|
{
|
| 107 |
|
|
if (bitpos)
|
| 108 |
|
|
return gimplify_build3 (gsi, BIT_FIELD_REF, type, t, bitsize, bitpos);
|
| 109 |
|
|
else
|
| 110 |
|
|
return gimplify_build1 (gsi, VIEW_CONVERT_EXPR, type, t);
|
| 111 |
|
|
}
|
| 112 |
|
|
|
| 113 |
|
|
static tree
|
| 114 |
|
|
do_unop (gimple_stmt_iterator *gsi, tree inner_type, tree a,
|
| 115 |
|
|
tree b ATTRIBUTE_UNUSED, tree bitpos, tree bitsize,
|
| 116 |
|
|
enum tree_code code)
|
| 117 |
|
|
{
|
| 118 |
|
|
a = tree_vec_extract (gsi, inner_type, a, bitsize, bitpos);
|
| 119 |
|
|
return gimplify_build1 (gsi, code, inner_type, a);
|
| 120 |
|
|
}
|
| 121 |
|
|
|
| 122 |
|
|
static tree
|
| 123 |
|
|
do_binop (gimple_stmt_iterator *gsi, tree inner_type, tree a, tree b,
|
| 124 |
|
|
tree bitpos, tree bitsize, enum tree_code code)
|
| 125 |
|
|
{
|
| 126 |
|
|
a = tree_vec_extract (gsi, inner_type, a, bitsize, bitpos);
|
| 127 |
|
|
b = tree_vec_extract (gsi, inner_type, b, bitsize, bitpos);
|
| 128 |
|
|
return gimplify_build2 (gsi, code, inner_type, a, b);
|
| 129 |
|
|
}
|
| 130 |
|
|
|
| 131 |
|
|
/* Expand vector addition to scalars. This does bit twiddling
|
| 132 |
|
|
in order to increase parallelism:
|
| 133 |
|
|
|
| 134 |
|
|
a + b = (((int) a & 0x7f7f7f7f) + ((int) b & 0x7f7f7f7f)) ^
|
| 135 |
|
|
(a ^ b) & 0x80808080
|
| 136 |
|
|
|
| 137 |
|
|
a - b = (((int) a | 0x80808080) - ((int) b & 0x7f7f7f7f)) ^
|
| 138 |
|
|
(a ^ ~b) & 0x80808080
|
| 139 |
|
|
|
| 140 |
|
|
-b = (0x80808080 - ((int) b & 0x7f7f7f7f)) ^ (~b & 0x80808080)
|
| 141 |
|
|
|
| 142 |
|
|
This optimization should be done only if 4 vector items or more
|
| 143 |
|
|
fit into a word. */
|
| 144 |
|
|
static tree
|
| 145 |
|
|
do_plus_minus (gimple_stmt_iterator *gsi, tree word_type, tree a, tree b,
|
| 146 |
|
|
tree bitpos ATTRIBUTE_UNUSED, tree bitsize ATTRIBUTE_UNUSED,
|
| 147 |
|
|
enum tree_code code)
|
| 148 |
|
|
{
|
| 149 |
|
|
tree inner_type = TREE_TYPE (TREE_TYPE (a));
|
| 150 |
|
|
unsigned HOST_WIDE_INT max;
|
| 151 |
|
|
tree low_bits, high_bits, a_low, b_low, result_low, signs;
|
| 152 |
|
|
|
| 153 |
|
|
max = GET_MODE_MASK (TYPE_MODE (inner_type));
|
| 154 |
|
|
low_bits = build_replicated_const (word_type, inner_type, max >> 1);
|
| 155 |
|
|
high_bits = build_replicated_const (word_type, inner_type, max & ~(max >> 1));
|
| 156 |
|
|
|
| 157 |
|
|
a = tree_vec_extract (gsi, word_type, a, bitsize, bitpos);
|
| 158 |
|
|
b = tree_vec_extract (gsi, word_type, b, bitsize, bitpos);
|
| 159 |
|
|
|
| 160 |
|
|
signs = gimplify_build2 (gsi, BIT_XOR_EXPR, word_type, a, b);
|
| 161 |
|
|
b_low = gimplify_build2 (gsi, BIT_AND_EXPR, word_type, b, low_bits);
|
| 162 |
|
|
if (code == PLUS_EXPR)
|
| 163 |
|
|
a_low = gimplify_build2 (gsi, BIT_AND_EXPR, word_type, a, low_bits);
|
| 164 |
|
|
else
|
| 165 |
|
|
{
|
| 166 |
|
|
a_low = gimplify_build2 (gsi, BIT_IOR_EXPR, word_type, a, high_bits);
|
| 167 |
|
|
signs = gimplify_build1 (gsi, BIT_NOT_EXPR, word_type, signs);
|
| 168 |
|
|
}
|
| 169 |
|
|
|
| 170 |
|
|
signs = gimplify_build2 (gsi, BIT_AND_EXPR, word_type, signs, high_bits);
|
| 171 |
|
|
result_low = gimplify_build2 (gsi, code, word_type, a_low, b_low);
|
| 172 |
|
|
return gimplify_build2 (gsi, BIT_XOR_EXPR, word_type, result_low, signs);
|
| 173 |
|
|
}
|
| 174 |
|
|
|
| 175 |
|
|
static tree
|
| 176 |
|
|
do_negate (gimple_stmt_iterator *gsi, tree word_type, tree b,
|
| 177 |
|
|
tree unused ATTRIBUTE_UNUSED, tree bitpos ATTRIBUTE_UNUSED,
|
| 178 |
|
|
tree bitsize ATTRIBUTE_UNUSED,
|
| 179 |
|
|
enum tree_code code ATTRIBUTE_UNUSED)
|
| 180 |
|
|
{
|
| 181 |
|
|
tree inner_type = TREE_TYPE (TREE_TYPE (b));
|
| 182 |
|
|
HOST_WIDE_INT max;
|
| 183 |
|
|
tree low_bits, high_bits, b_low, result_low, signs;
|
| 184 |
|
|
|
| 185 |
|
|
max = GET_MODE_MASK (TYPE_MODE (inner_type));
|
| 186 |
|
|
low_bits = build_replicated_const (word_type, inner_type, max >> 1);
|
| 187 |
|
|
high_bits = build_replicated_const (word_type, inner_type, max & ~(max >> 1));
|
| 188 |
|
|
|
| 189 |
|
|
b = tree_vec_extract (gsi, word_type, b, bitsize, bitpos);
|
| 190 |
|
|
|
| 191 |
|
|
b_low = gimplify_build2 (gsi, BIT_AND_EXPR, word_type, b, low_bits);
|
| 192 |
|
|
signs = gimplify_build1 (gsi, BIT_NOT_EXPR, word_type, b);
|
| 193 |
|
|
signs = gimplify_build2 (gsi, BIT_AND_EXPR, word_type, signs, high_bits);
|
| 194 |
|
|
result_low = gimplify_build2 (gsi, MINUS_EXPR, word_type, high_bits, b_low);
|
| 195 |
|
|
return gimplify_build2 (gsi, BIT_XOR_EXPR, word_type, result_low, signs);
|
| 196 |
|
|
}
|
| 197 |
|
|
|
| 198 |
|
|
/* Expand a vector operation to scalars, by using many operations
|
| 199 |
|
|
whose type is the vector type's inner type. */
|
| 200 |
|
|
static tree
|
| 201 |
|
|
expand_vector_piecewise (gimple_stmt_iterator *gsi, elem_op_func f,
|
| 202 |
|
|
tree type, tree inner_type,
|
| 203 |
|
|
tree a, tree b, enum tree_code code)
|
| 204 |
|
|
{
|
| 205 |
|
|
VEC(constructor_elt,gc) *v;
|
| 206 |
|
|
tree part_width = TYPE_SIZE (inner_type);
|
| 207 |
|
|
tree index = bitsize_int (0);
|
| 208 |
|
|
int nunits = TYPE_VECTOR_SUBPARTS (type);
|
| 209 |
|
|
int delta = tree_low_cst (part_width, 1)
|
| 210 |
|
|
/ tree_low_cst (TYPE_SIZE (TREE_TYPE (type)), 1);
|
| 211 |
|
|
int i;
|
| 212 |
|
|
|
| 213 |
|
|
v = VEC_alloc(constructor_elt, gc, (nunits + delta - 1) / delta);
|
| 214 |
|
|
for (i = 0; i < nunits;
|
| 215 |
|
|
i += delta, index = int_const_binop (PLUS_EXPR, index, part_width, 0))
|
| 216 |
|
|
{
|
| 217 |
|
|
tree result = f (gsi, inner_type, a, b, index, part_width, code);
|
| 218 |
|
|
constructor_elt *ce = VEC_quick_push (constructor_elt, v, NULL);
|
| 219 |
|
|
ce->index = NULL_TREE;
|
| 220 |
|
|
ce->value = result;
|
| 221 |
|
|
}
|
| 222 |
|
|
|
| 223 |
|
|
return build_constructor (type, v);
|
| 224 |
|
|
}
|
| 225 |
|
|
|
| 226 |
|
|
/* Expand a vector operation to scalars with the freedom to use
|
| 227 |
|
|
a scalar integer type, or to use a different size for the items
|
| 228 |
|
|
in the vector type. */
|
| 229 |
|
|
static tree
|
| 230 |
|
|
expand_vector_parallel (gimple_stmt_iterator *gsi, elem_op_func f, tree type,
|
| 231 |
|
|
tree a, tree b,
|
| 232 |
|
|
enum tree_code code)
|
| 233 |
|
|
{
|
| 234 |
|
|
tree result, compute_type;
|
| 235 |
|
|
enum machine_mode mode;
|
| 236 |
|
|
int n_words = tree_low_cst (TYPE_SIZE_UNIT (type), 1) / UNITS_PER_WORD;
|
| 237 |
|
|
|
| 238 |
|
|
/* We have three strategies. If the type is already correct, just do
|
| 239 |
|
|
the operation an element at a time. Else, if the vector is wider than
|
| 240 |
|
|
one word, do it a word at a time; finally, if the vector is smaller
|
| 241 |
|
|
than one word, do it as a scalar. */
|
| 242 |
|
|
if (TYPE_MODE (TREE_TYPE (type)) == word_mode)
|
| 243 |
|
|
return expand_vector_piecewise (gsi, f,
|
| 244 |
|
|
type, TREE_TYPE (type),
|
| 245 |
|
|
a, b, code);
|
| 246 |
|
|
else if (n_words > 1)
|
| 247 |
|
|
{
|
| 248 |
|
|
tree word_type = build_word_mode_vector_type (n_words);
|
| 249 |
|
|
result = expand_vector_piecewise (gsi, f,
|
| 250 |
|
|
word_type, TREE_TYPE (word_type),
|
| 251 |
|
|
a, b, code);
|
| 252 |
|
|
result = force_gimple_operand_gsi (gsi, result, true, NULL, true,
|
| 253 |
|
|
GSI_SAME_STMT);
|
| 254 |
|
|
}
|
| 255 |
|
|
else
|
| 256 |
|
|
{
|
| 257 |
|
|
/* Use a single scalar operation with a mode no wider than word_mode. */
|
| 258 |
|
|
mode = mode_for_size (tree_low_cst (TYPE_SIZE (type), 1), MODE_INT, 0);
|
| 259 |
|
|
compute_type = lang_hooks.types.type_for_mode (mode, 1);
|
| 260 |
|
|
result = f (gsi, compute_type, a, b, NULL_TREE, NULL_TREE, code);
|
| 261 |
|
|
}
|
| 262 |
|
|
|
| 263 |
|
|
return result;
|
| 264 |
|
|
}
|
| 265 |
|
|
|
| 266 |
|
|
/* Expand a vector operation to scalars; for integer types we can use
|
| 267 |
|
|
special bit twiddling tricks to do the sums a word at a time, using
|
| 268 |
|
|
function F_PARALLEL instead of F. These tricks are done only if
|
| 269 |
|
|
they can process at least four items, that is, only if the vector
|
| 270 |
|
|
holds at least four items and if a word can hold four items. */
|
| 271 |
|
|
static tree
|
| 272 |
|
|
expand_vector_addition (gimple_stmt_iterator *gsi,
|
| 273 |
|
|
elem_op_func f, elem_op_func f_parallel,
|
| 274 |
|
|
tree type, tree a, tree b, enum tree_code code)
|
| 275 |
|
|
{
|
| 276 |
|
|
int parts_per_word = UNITS_PER_WORD
|
| 277 |
|
|
/ tree_low_cst (TYPE_SIZE_UNIT (TREE_TYPE (type)), 1);
|
| 278 |
|
|
|
| 279 |
|
|
if (INTEGRAL_TYPE_P (TREE_TYPE (type))
|
| 280 |
|
|
&& parts_per_word >= 4
|
| 281 |
|
|
&& TYPE_VECTOR_SUBPARTS (type) >= 4)
|
| 282 |
|
|
return expand_vector_parallel (gsi, f_parallel,
|
| 283 |
|
|
type, a, b, code);
|
| 284 |
|
|
else
|
| 285 |
|
|
return expand_vector_piecewise (gsi, f,
|
| 286 |
|
|
type, TREE_TYPE (type),
|
| 287 |
|
|
a, b, code);
|
| 288 |
|
|
}
|
| 289 |
|
|
|
| 290 |
|
|
static tree
|
| 291 |
|
|
expand_vector_operation (gimple_stmt_iterator *gsi, tree type, tree compute_type,
|
| 292 |
|
|
gimple assign, enum tree_code code)
|
| 293 |
|
|
{
|
| 294 |
|
|
enum machine_mode compute_mode = TYPE_MODE (compute_type);
|
| 295 |
|
|
|
| 296 |
|
|
/* If the compute mode is not a vector mode (hence we are not decomposing
|
| 297 |
|
|
a BLKmode vector to smaller, hardware-supported vectors), we may want
|
| 298 |
|
|
to expand the operations in parallel. */
|
| 299 |
|
|
if (GET_MODE_CLASS (compute_mode) != MODE_VECTOR_INT
|
| 300 |
|
|
&& GET_MODE_CLASS (compute_mode) != MODE_VECTOR_FLOAT
|
| 301 |
|
|
&& GET_MODE_CLASS (compute_mode) != MODE_VECTOR_FRACT
|
| 302 |
|
|
&& GET_MODE_CLASS (compute_mode) != MODE_VECTOR_UFRACT
|
| 303 |
|
|
&& GET_MODE_CLASS (compute_mode) != MODE_VECTOR_ACCUM
|
| 304 |
|
|
&& GET_MODE_CLASS (compute_mode) != MODE_VECTOR_UACCUM)
|
| 305 |
|
|
switch (code)
|
| 306 |
|
|
{
|
| 307 |
|
|
case PLUS_EXPR:
|
| 308 |
|
|
case MINUS_EXPR:
|
| 309 |
|
|
if (!TYPE_OVERFLOW_TRAPS (type))
|
| 310 |
|
|
return expand_vector_addition (gsi, do_binop, do_plus_minus, type,
|
| 311 |
|
|
gimple_assign_rhs1 (assign),
|
| 312 |
|
|
gimple_assign_rhs2 (assign), code);
|
| 313 |
|
|
break;
|
| 314 |
|
|
|
| 315 |
|
|
case NEGATE_EXPR:
|
| 316 |
|
|
if (!TYPE_OVERFLOW_TRAPS (type))
|
| 317 |
|
|
return expand_vector_addition (gsi, do_unop, do_negate, type,
|
| 318 |
|
|
gimple_assign_rhs1 (assign),
|
| 319 |
|
|
NULL_TREE, code);
|
| 320 |
|
|
break;
|
| 321 |
|
|
|
| 322 |
|
|
case BIT_AND_EXPR:
|
| 323 |
|
|
case BIT_IOR_EXPR:
|
| 324 |
|
|
case BIT_XOR_EXPR:
|
| 325 |
|
|
return expand_vector_parallel (gsi, do_binop, type,
|
| 326 |
|
|
gimple_assign_rhs1 (assign),
|
| 327 |
|
|
gimple_assign_rhs2 (assign), code);
|
| 328 |
|
|
|
| 329 |
|
|
case BIT_NOT_EXPR:
|
| 330 |
|
|
return expand_vector_parallel (gsi, do_unop, type,
|
| 331 |
|
|
gimple_assign_rhs1 (assign),
|
| 332 |
|
|
NULL_TREE, code);
|
| 333 |
|
|
|
| 334 |
|
|
default:
|
| 335 |
|
|
break;
|
| 336 |
|
|
}
|
| 337 |
|
|
|
| 338 |
|
|
if (TREE_CODE_CLASS (code) == tcc_unary)
|
| 339 |
|
|
return expand_vector_piecewise (gsi, do_unop, type, compute_type,
|
| 340 |
|
|
gimple_assign_rhs1 (assign),
|
| 341 |
|
|
NULL_TREE, code);
|
| 342 |
|
|
else
|
| 343 |
|
|
return expand_vector_piecewise (gsi, do_binop, type, compute_type,
|
| 344 |
|
|
gimple_assign_rhs1 (assign),
|
| 345 |
|
|
gimple_assign_rhs2 (assign), code);
|
| 346 |
|
|
}
|
| 347 |
|
|
|
| 348 |
|
|
/* Return a type for the widest vector mode whose components are of mode
|
| 349 |
|
|
INNER_MODE, or NULL_TREE if none is found.
|
| 350 |
|
|
SATP is true for saturating fixed-point types. */
|
| 351 |
|
|
|
| 352 |
|
|
static tree
|
| 353 |
|
|
type_for_widest_vector_mode (enum machine_mode inner_mode, optab op, int satp)
|
| 354 |
|
|
{
|
| 355 |
|
|
enum machine_mode best_mode = VOIDmode, mode;
|
| 356 |
|
|
int best_nunits = 0;
|
| 357 |
|
|
|
| 358 |
|
|
if (SCALAR_FLOAT_MODE_P (inner_mode))
|
| 359 |
|
|
mode = MIN_MODE_VECTOR_FLOAT;
|
| 360 |
|
|
else if (SCALAR_FRACT_MODE_P (inner_mode))
|
| 361 |
|
|
mode = MIN_MODE_VECTOR_FRACT;
|
| 362 |
|
|
else if (SCALAR_UFRACT_MODE_P (inner_mode))
|
| 363 |
|
|
mode = MIN_MODE_VECTOR_UFRACT;
|
| 364 |
|
|
else if (SCALAR_ACCUM_MODE_P (inner_mode))
|
| 365 |
|
|
mode = MIN_MODE_VECTOR_ACCUM;
|
| 366 |
|
|
else if (SCALAR_UACCUM_MODE_P (inner_mode))
|
| 367 |
|
|
mode = MIN_MODE_VECTOR_UACCUM;
|
| 368 |
|
|
else
|
| 369 |
|
|
mode = MIN_MODE_VECTOR_INT;
|
| 370 |
|
|
|
| 371 |
|
|
for (; mode != VOIDmode; mode = GET_MODE_WIDER_MODE (mode))
|
| 372 |
|
|
if (GET_MODE_INNER (mode) == inner_mode
|
| 373 |
|
|
&& GET_MODE_NUNITS (mode) > best_nunits
|
| 374 |
|
|
&& optab_handler (op, mode)->insn_code != CODE_FOR_nothing)
|
| 375 |
|
|
best_mode = mode, best_nunits = GET_MODE_NUNITS (mode);
|
| 376 |
|
|
|
| 377 |
|
|
if (best_mode == VOIDmode)
|
| 378 |
|
|
return NULL_TREE;
|
| 379 |
|
|
else
|
| 380 |
|
|
{
|
| 381 |
|
|
/* For fixed-point modes, we need to pass satp as the 2nd parameter. */
|
| 382 |
|
|
if (ALL_FIXED_POINT_MODE_P (best_mode))
|
| 383 |
|
|
return lang_hooks.types.type_for_mode (best_mode, satp);
|
| 384 |
|
|
|
| 385 |
|
|
return lang_hooks.types.type_for_mode (best_mode, 1);
|
| 386 |
|
|
}
|
| 387 |
|
|
}
|
| 388 |
|
|
|
| 389 |
|
|
/* Process one statement. If we identify a vector operation, expand it. */
|
| 390 |
|
|
|
| 391 |
|
|
static void
|
| 392 |
|
|
expand_vector_operations_1 (gimple_stmt_iterator *gsi)
|
| 393 |
|
|
{
|
| 394 |
|
|
gimple stmt = gsi_stmt (*gsi);
|
| 395 |
|
|
tree lhs, rhs1, rhs2 = NULL, type, compute_type;
|
| 396 |
|
|
enum tree_code code;
|
| 397 |
|
|
enum machine_mode compute_mode;
|
| 398 |
|
|
optab op;
|
| 399 |
|
|
enum gimple_rhs_class rhs_class;
|
| 400 |
|
|
tree new_rhs;
|
| 401 |
|
|
|
| 402 |
|
|
if (gimple_code (stmt) != GIMPLE_ASSIGN)
|
| 403 |
|
|
return;
|
| 404 |
|
|
|
| 405 |
|
|
code = gimple_assign_rhs_code (stmt);
|
| 406 |
|
|
rhs_class = get_gimple_rhs_class (code);
|
| 407 |
|
|
|
| 408 |
|
|
if (rhs_class != GIMPLE_UNARY_RHS && rhs_class != GIMPLE_BINARY_RHS)
|
| 409 |
|
|
return;
|
| 410 |
|
|
|
| 411 |
|
|
lhs = gimple_assign_lhs (stmt);
|
| 412 |
|
|
rhs1 = gimple_assign_rhs1 (stmt);
|
| 413 |
|
|
type = gimple_expr_type (stmt);
|
| 414 |
|
|
if (rhs_class == GIMPLE_BINARY_RHS)
|
| 415 |
|
|
rhs2 = gimple_assign_rhs2 (stmt);
|
| 416 |
|
|
|
| 417 |
|
|
if (TREE_CODE (type) != VECTOR_TYPE)
|
| 418 |
|
|
return;
|
| 419 |
|
|
|
| 420 |
|
|
if (code == NOP_EXPR
|
| 421 |
|
|
|| code == FLOAT_EXPR
|
| 422 |
|
|
|| code == FIX_TRUNC_EXPR
|
| 423 |
|
|
|| code == VIEW_CONVERT_EXPR)
|
| 424 |
|
|
return;
|
| 425 |
|
|
|
| 426 |
|
|
gcc_assert (code != CONVERT_EXPR);
|
| 427 |
|
|
|
| 428 |
|
|
/* The signedness is determined from input argument. */
|
| 429 |
|
|
if (code == VEC_UNPACK_FLOAT_HI_EXPR
|
| 430 |
|
|
|| code == VEC_UNPACK_FLOAT_LO_EXPR)
|
| 431 |
|
|
type = TREE_TYPE (rhs1);
|
| 432 |
|
|
|
| 433 |
|
|
/* Choose between vector shift/rotate by vector and vector shift/rotate by
|
| 434 |
|
|
scalar */
|
| 435 |
|
|
if (code == LSHIFT_EXPR
|
| 436 |
|
|
|| code == RSHIFT_EXPR
|
| 437 |
|
|
|| code == LROTATE_EXPR
|
| 438 |
|
|
|| code == RROTATE_EXPR)
|
| 439 |
|
|
{
|
| 440 |
|
|
/* If the 2nd argument is vector, we need a vector/vector shift */
|
| 441 |
|
|
if (VECTOR_MODE_P (TYPE_MODE (TREE_TYPE (rhs2))))
|
| 442 |
|
|
op = optab_for_tree_code (code, type, optab_vector);
|
| 443 |
|
|
else
|
| 444 |
|
|
{
|
| 445 |
|
|
/* Try for a vector/scalar shift, and if we don't have one, see if we
|
| 446 |
|
|
have a vector/vector shift */
|
| 447 |
|
|
op = optab_for_tree_code (code, type, optab_scalar);
|
| 448 |
|
|
if (!op
|
| 449 |
|
|
|| (op->handlers[(int) TYPE_MODE (type)].insn_code
|
| 450 |
|
|
== CODE_FOR_nothing))
|
| 451 |
|
|
op = optab_for_tree_code (code, type, optab_vector);
|
| 452 |
|
|
}
|
| 453 |
|
|
}
|
| 454 |
|
|
else
|
| 455 |
|
|
op = optab_for_tree_code (code, type, optab_default);
|
| 456 |
|
|
|
| 457 |
|
|
/* For widening/narrowing vector operations, the relevant type is of the
|
| 458 |
|
|
arguments, not the widened result. VEC_UNPACK_FLOAT_*_EXPR is
|
| 459 |
|
|
calculated in the same way above. */
|
| 460 |
|
|
if (code == WIDEN_SUM_EXPR
|
| 461 |
|
|
|| code == VEC_WIDEN_MULT_HI_EXPR
|
| 462 |
|
|
|| code == VEC_WIDEN_MULT_LO_EXPR
|
| 463 |
|
|
|| code == VEC_UNPACK_HI_EXPR
|
| 464 |
|
|
|| code == VEC_UNPACK_LO_EXPR
|
| 465 |
|
|
|| code == VEC_PACK_TRUNC_EXPR
|
| 466 |
|
|
|| code == VEC_PACK_SAT_EXPR
|
| 467 |
|
|
|| code == VEC_PACK_FIX_TRUNC_EXPR)
|
| 468 |
|
|
type = TREE_TYPE (rhs1);
|
| 469 |
|
|
|
| 470 |
|
|
/* Optabs will try converting a negation into a subtraction, so
|
| 471 |
|
|
look for it as well. TODO: negation of floating-point vectors
|
| 472 |
|
|
might be turned into an exclusive OR toggling the sign bit. */
|
| 473 |
|
|
if (op == NULL
|
| 474 |
|
|
&& code == NEGATE_EXPR
|
| 475 |
|
|
&& INTEGRAL_TYPE_P (TREE_TYPE (type)))
|
| 476 |
|
|
op = optab_for_tree_code (MINUS_EXPR, type, optab_default);
|
| 477 |
|
|
|
| 478 |
|
|
/* For very wide vectors, try using a smaller vector mode. */
|
| 479 |
|
|
compute_type = type;
|
| 480 |
|
|
if (TYPE_MODE (type) == BLKmode && op)
|
| 481 |
|
|
{
|
| 482 |
|
|
tree vector_compute_type
|
| 483 |
|
|
= type_for_widest_vector_mode (TYPE_MODE (TREE_TYPE (type)), op,
|
| 484 |
|
|
TYPE_SATURATING (TREE_TYPE (type)));
|
| 485 |
|
|
if (vector_compute_type != NULL_TREE
|
| 486 |
|
|
&& (TYPE_VECTOR_SUBPARTS (vector_compute_type)
|
| 487 |
|
|
< TYPE_VECTOR_SUBPARTS (compute_type)))
|
| 488 |
|
|
compute_type = vector_compute_type;
|
| 489 |
|
|
}
|
| 490 |
|
|
|
| 491 |
|
|
/* If we are breaking a BLKmode vector into smaller pieces,
|
| 492 |
|
|
type_for_widest_vector_mode has already looked into the optab,
|
| 493 |
|
|
so skip these checks. */
|
| 494 |
|
|
if (compute_type == type)
|
| 495 |
|
|
{
|
| 496 |
|
|
compute_mode = TYPE_MODE (compute_type);
|
| 497 |
|
|
if ((GET_MODE_CLASS (compute_mode) == MODE_VECTOR_INT
|
| 498 |
|
|
|| GET_MODE_CLASS (compute_mode) == MODE_VECTOR_FLOAT
|
| 499 |
|
|
|| GET_MODE_CLASS (compute_mode) == MODE_VECTOR_FRACT
|
| 500 |
|
|
|| GET_MODE_CLASS (compute_mode) == MODE_VECTOR_UFRACT
|
| 501 |
|
|
|| GET_MODE_CLASS (compute_mode) == MODE_VECTOR_ACCUM
|
| 502 |
|
|
|| GET_MODE_CLASS (compute_mode) == MODE_VECTOR_UACCUM)
|
| 503 |
|
|
&& op != NULL
|
| 504 |
|
|
&& optab_handler (op, compute_mode)->insn_code != CODE_FOR_nothing)
|
| 505 |
|
|
return;
|
| 506 |
|
|
else
|
| 507 |
|
|
/* There is no operation in hardware, so fall back to scalars. */
|
| 508 |
|
|
compute_type = TREE_TYPE (type);
|
| 509 |
|
|
}
|
| 510 |
|
|
|
| 511 |
|
|
gcc_assert (code != VEC_LSHIFT_EXPR && code != VEC_RSHIFT_EXPR);
|
| 512 |
|
|
new_rhs = expand_vector_operation (gsi, type, compute_type, stmt, code);
|
| 513 |
|
|
if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (new_rhs)))
|
| 514 |
|
|
new_rhs = gimplify_build1 (gsi, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
|
| 515 |
|
|
new_rhs);
|
| 516 |
|
|
|
| 517 |
|
|
/* NOTE: We should avoid using gimple_assign_set_rhs_from_tree. One
|
| 518 |
|
|
way to do it is change expand_vector_operation and its callees to
|
| 519 |
|
|
return a tree_code, RHS1 and RHS2 instead of a tree. */
|
| 520 |
|
|
gimple_assign_set_rhs_from_tree (gsi, new_rhs);
|
| 521 |
|
|
|
| 522 |
|
|
gimple_set_modified (gsi_stmt (*gsi), true);
|
| 523 |
|
|
}
|
| 524 |
|
|
|
| 525 |
|
|
/* Use this to lower vector operations introduced by the vectorizer,
|
| 526 |
|
|
if it may need the bit-twiddling tricks implemented in this file. */
|
| 527 |
|
|
|
| 528 |
|
|
static bool
|
| 529 |
|
|
gate_expand_vector_operations (void)
|
| 530 |
|
|
{
|
| 531 |
|
|
return flag_tree_vectorize != 0;
|
| 532 |
|
|
}
|
| 533 |
|
|
|
| 534 |
|
|
static unsigned int
|
| 535 |
|
|
expand_vector_operations (void)
|
| 536 |
|
|
{
|
| 537 |
|
|
gimple_stmt_iterator gsi;
|
| 538 |
|
|
basic_block bb;
|
| 539 |
|
|
|
| 540 |
|
|
FOR_EACH_BB (bb)
|
| 541 |
|
|
{
|
| 542 |
|
|
for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
|
| 543 |
|
|
{
|
| 544 |
|
|
expand_vector_operations_1 (&gsi);
|
| 545 |
|
|
update_stmt_if_modified (gsi_stmt (gsi));
|
| 546 |
|
|
}
|
| 547 |
|
|
}
|
| 548 |
|
|
return 0;
|
| 549 |
|
|
}
|
| 550 |
|
|
|
| 551 |
|
|
struct gimple_opt_pass pass_lower_vector =
|
| 552 |
|
|
{
|
| 553 |
|
|
{
|
| 554 |
|
|
GIMPLE_PASS,
|
| 555 |
|
|
"veclower", /* name */
|
| 556 |
|
|
0, /* gate */
|
| 557 |
|
|
expand_vector_operations, /* execute */
|
| 558 |
|
|
NULL, /* sub */
|
| 559 |
|
|
NULL, /* next */
|
| 560 |
|
|
0, /* static_pass_number */
|
| 561 |
|
|
TV_NONE, /* tv_id */
|
| 562 |
|
|
PROP_cfg, /* properties_required */
|
| 563 |
|
|
0, /* properties_provided */
|
| 564 |
|
|
0, /* properties_destroyed */
|
| 565 |
|
|
0, /* todo_flags_start */
|
| 566 |
|
|
TODO_dump_func | TODO_ggc_collect
|
| 567 |
|
|
| TODO_verify_stmts /* todo_flags_finish */
|
| 568 |
|
|
}
|
| 569 |
|
|
};
|
| 570 |
|
|
|
| 571 |
|
|
struct gimple_opt_pass pass_lower_vector_ssa =
|
| 572 |
|
|
{
|
| 573 |
|
|
{
|
| 574 |
|
|
GIMPLE_PASS,
|
| 575 |
|
|
"veclower2", /* name */
|
| 576 |
|
|
gate_expand_vector_operations, /* gate */
|
| 577 |
|
|
expand_vector_operations, /* execute */
|
| 578 |
|
|
NULL, /* sub */
|
| 579 |
|
|
NULL, /* next */
|
| 580 |
|
|
0, /* static_pass_number */
|
| 581 |
|
|
TV_NONE, /* tv_id */
|
| 582 |
|
|
PROP_cfg, /* properties_required */
|
| 583 |
|
|
0, /* properties_provided */
|
| 584 |
|
|
0, /* properties_destroyed */
|
| 585 |
|
|
0, /* todo_flags_start */
|
| 586 |
|
|
TODO_dump_func | TODO_update_ssa /* todo_flags_finish */
|
| 587 |
|
|
| TODO_verify_ssa
|
| 588 |
|
|
| TODO_verify_stmts | TODO_verify_flow
|
| 589 |
|
|
}
|
| 590 |
|
|
};
|
| 591 |
|
|
|
| 592 |
|
|
#include "gt-tree-vect-generic.h"
|