]>
Commit | Line | Data |
---|---|---|
4a61a337 | 1 | /* Analysis Utilities for Loop Vectorization. |
ce084dfc | 2 | Copyright (C) 2006, 2007, 2008, 2009, 2010 Free Software Foundation, Inc. |
4a61a337 | 3 | Contributed by Dorit Nuzman <dorit@il.ibm.com> |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it under | |
8 | the terms of the GNU General Public License as published by the Free | |
8c4c00c1 | 9 | Software Foundation; either version 3, or (at your option) any later |
4a61a337 | 10 | version. |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
13 | 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 | |
8c4c00c1 | 18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
4a61a337 | 20 | |
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "tm.h" | |
25 | #include "ggc.h" | |
26 | #include "tree.h" | |
4a61a337 | 27 | #include "target.h" |
28 | #include "basic-block.h" | |
ce084dfc | 29 | #include "gimple-pretty-print.h" |
4a61a337 | 30 | #include "tree-flow.h" |
31 | #include "tree-dump.h" | |
4a61a337 | 32 | #include "cfgloop.h" |
33 | #include "expr.h" | |
34 | #include "optabs.h" | |
35 | #include "params.h" | |
36 | #include "tree-data-ref.h" | |
37 | #include "tree-vectorizer.h" | |
38 | #include "recog.h" | |
0b205f4c | 39 | #include "diagnostic-core.h" |
4a61a337 | 40 | |
334ec2d8 | 41 | /* Function prototypes */ |
48e1416a | 42 | static void vect_pattern_recog_1 |
75a70cf9 | 43 | (gimple (* ) (gimple, tree *, tree *), gimple_stmt_iterator); |
44 | static bool widened_name_p (tree, gimple, tree *, gimple *); | |
4a61a337 | 45 | |
46 | /* Pattern recognition functions */ | |
75a70cf9 | 47 | static gimple vect_recog_widen_sum_pattern (gimple, tree *, tree *); |
48 | static gimple vect_recog_widen_mult_pattern (gimple, tree *, tree *); | |
49 | static gimple vect_recog_dot_prod_pattern (gimple, tree *, tree *); | |
50 | static gimple vect_recog_pow_pattern (gimple, tree *, tree *); | |
4a61a337 | 51 | static vect_recog_func_ptr vect_vect_recog_func_ptrs[NUM_PATTERNS] = { |
52 | vect_recog_widen_mult_pattern, | |
53 | vect_recog_widen_sum_pattern, | |
c37bea13 | 54 | vect_recog_dot_prod_pattern, |
55 | vect_recog_pow_pattern}; | |
4a61a337 | 56 | |
57 | ||
58 | /* Function widened_name_p | |
59 | ||
60 | Check whether NAME, an ssa-name used in USE_STMT, | |
61 | is a result of a type-promotion, such that: | |
62 | DEF_STMT: NAME = NOP (name0) | |
48e1416a | 63 | where the type of name0 (HALF_TYPE) is smaller than the type of NAME. |
4a61a337 | 64 | */ |
65 | ||
66 | static bool | |
75a70cf9 | 67 | widened_name_p (tree name, gimple use_stmt, tree *half_type, gimple *def_stmt) |
4a61a337 | 68 | { |
69 | tree dummy; | |
75a70cf9 | 70 | gimple dummy_gimple; |
4a61a337 | 71 | loop_vec_info loop_vinfo; |
72 | stmt_vec_info stmt_vinfo; | |
4a61a337 | 73 | tree type = TREE_TYPE (name); |
74 | tree oprnd0; | |
75 | enum vect_def_type dt; | |
76 | tree def; | |
77 | ||
78 | stmt_vinfo = vinfo_for_stmt (use_stmt); | |
79 | loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
80 | ||
37545e54 | 81 | if (!vect_is_simple_use (name, loop_vinfo, NULL, def_stmt, &def, &dt)) |
4a61a337 | 82 | return false; |
83 | ||
f083cd24 | 84 | if (dt != vect_internal_def |
85 | && dt != vect_external_def && dt != vect_constant_def) | |
4a61a337 | 86 | return false; |
87 | ||
88 | if (! *def_stmt) | |
89 | return false; | |
90 | ||
75a70cf9 | 91 | if (!is_gimple_assign (*def_stmt)) |
4a61a337 | 92 | return false; |
93 | ||
75a70cf9 | 94 | if (gimple_assign_rhs_code (*def_stmt) != NOP_EXPR) |
4a61a337 | 95 | return false; |
96 | ||
75a70cf9 | 97 | oprnd0 = gimple_assign_rhs1 (*def_stmt); |
4a61a337 | 98 | |
99 | *half_type = TREE_TYPE (oprnd0); | |
100 | if (!INTEGRAL_TYPE_P (type) || !INTEGRAL_TYPE_P (*half_type) | |
101 | || (TYPE_UNSIGNED (type) != TYPE_UNSIGNED (*half_type)) | |
102 | || (TYPE_PRECISION (type) < (TYPE_PRECISION (*half_type) * 2))) | |
103 | return false; | |
104 | ||
48e1416a | 105 | if (!vect_is_simple_use (oprnd0, loop_vinfo, NULL, &dummy_gimple, &dummy, |
37545e54 | 106 | &dt)) |
4a61a337 | 107 | return false; |
108 | ||
4a61a337 | 109 | return true; |
110 | } | |
111 | ||
75a70cf9 | 112 | /* Helper to return a new temporary for pattern of TYPE for STMT. If STMT |
113 | is NULL, the caller must set SSA_NAME_DEF_STMT for the returned SSA var. */ | |
114 | ||
115 | static tree | |
116 | vect_recog_temp_ssa_var (tree type, gimple stmt) | |
117 | { | |
118 | tree var = create_tmp_var (type, "patt"); | |
119 | ||
120 | add_referenced_var (var); | |
121 | var = make_ssa_name (var, stmt); | |
122 | return var; | |
123 | } | |
4a61a337 | 124 | |
125 | /* Function vect_recog_dot_prod_pattern | |
126 | ||
127 | Try to find the following pattern: | |
128 | ||
129 | type x_t, y_t; | |
130 | TYPE1 prod; | |
131 | TYPE2 sum = init; | |
132 | loop: | |
133 | sum_0 = phi <init, sum_1> | |
134 | S1 x_t = ... | |
135 | S2 y_t = ... | |
136 | S3 x_T = (TYPE1) x_t; | |
137 | S4 y_T = (TYPE1) y_t; | |
138 | S5 prod = x_T * y_T; | |
139 | [S6 prod = (TYPE2) prod; #optional] | |
140 | S7 sum_1 = prod + sum_0; | |
141 | ||
48e1416a | 142 | where 'TYPE1' is exactly double the size of type 'type', and 'TYPE2' is the |
143 | same size of 'TYPE1' or bigger. This is a special case of a reduction | |
4a61a337 | 144 | computation. |
48e1416a | 145 | |
4a61a337 | 146 | Input: |
147 | ||
148 | * LAST_STMT: A stmt from which the pattern search begins. In the example, | |
149 | when this function is called with S7, the pattern {S3,S4,S5,S6,S7} will be | |
150 | detected. | |
151 | ||
152 | Output: | |
153 | ||
154 | * TYPE_IN: The type of the input arguments to the pattern. | |
155 | ||
156 | * TYPE_OUT: The type of the output of this pattern. | |
157 | ||
158 | * Return value: A new stmt that will be used to replace the sequence of | |
159 | stmts that constitute the pattern. In this case it will be: | |
160 | WIDEN_DOT_PRODUCT <x_t, y_t, sum_0> | |
221e9a92 | 161 | |
162 | Note: The dot-prod idiom is a widening reduction pattern that is | |
163 | vectorized without preserving all the intermediate results. It | |
164 | produces only N/2 (widened) results (by summing up pairs of | |
165 | intermediate results) rather than all N results. Therefore, we | |
166 | cannot allow this pattern when we want to get all the results and in | |
167 | the correct order (as is the case when this computation is in an | |
168 | inner-loop nested in an outer-loop that us being vectorized). */ | |
4a61a337 | 169 | |
75a70cf9 | 170 | static gimple |
171 | vect_recog_dot_prod_pattern (gimple last_stmt, tree *type_in, tree *type_out) | |
4a61a337 | 172 | { |
75a70cf9 | 173 | gimple stmt; |
4a61a337 | 174 | tree oprnd0, oprnd1; |
175 | tree oprnd00, oprnd01; | |
176 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt); | |
177 | tree type, half_type; | |
75a70cf9 | 178 | gimple pattern_stmt; |
4a61a337 | 179 | tree prod_type; |
221e9a92 | 180 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
181 | struct loop *loop = LOOP_VINFO_LOOP (loop_info); | |
75a70cf9 | 182 | tree var, rhs; |
4a61a337 | 183 | |
75a70cf9 | 184 | if (!is_gimple_assign (last_stmt)) |
4a61a337 | 185 | return NULL; |
186 | ||
75a70cf9 | 187 | type = gimple_expr_type (last_stmt); |
4a61a337 | 188 | |
48e1416a | 189 | /* Look for the following pattern |
4a61a337 | 190 | DX = (TYPE1) X; |
191 | DY = (TYPE1) Y; | |
48e1416a | 192 | DPROD = DX * DY; |
4a61a337 | 193 | DDPROD = (TYPE2) DPROD; |
194 | sum_1 = DDPROD + sum_0; | |
48e1416a | 195 | In which |
4a61a337 | 196 | - DX is double the size of X |
197 | - DY is double the size of Y | |
198 | - DX, DY, DPROD all have the same type | |
199 | - sum is the same size of DPROD or bigger | |
200 | - sum has been recognized as a reduction variable. | |
201 | ||
202 | This is equivalent to: | |
203 | DPROD = X w* Y; #widen mult | |
204 | sum_1 = DPROD w+ sum_0; #widen summation | |
205 | or | |
206 | DPROD = X w* Y; #widen mult | |
207 | sum_1 = DPROD + sum_0; #summation | |
208 | */ | |
209 | ||
210 | /* Starting from LAST_STMT, follow the defs of its uses in search | |
211 | of the above pattern. */ | |
212 | ||
75a70cf9 | 213 | if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR) |
4a61a337 | 214 | return NULL; |
215 | ||
216 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) | |
217 | { | |
218 | /* Has been detected as widening-summation? */ | |
219 | ||
220 | stmt = STMT_VINFO_RELATED_STMT (stmt_vinfo); | |
75a70cf9 | 221 | type = gimple_expr_type (stmt); |
222 | if (gimple_assign_rhs_code (stmt) != WIDEN_SUM_EXPR) | |
4a61a337 | 223 | return NULL; |
75a70cf9 | 224 | oprnd0 = gimple_assign_rhs1 (stmt); |
225 | oprnd1 = gimple_assign_rhs2 (stmt); | |
4a61a337 | 226 | half_type = TREE_TYPE (oprnd0); |
227 | } | |
228 | else | |
229 | { | |
75a70cf9 | 230 | gimple def_stmt; |
4a61a337 | 231 | |
232 | if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def) | |
233 | return NULL; | |
75a70cf9 | 234 | oprnd0 = gimple_assign_rhs1 (last_stmt); |
235 | oprnd1 = gimple_assign_rhs2 (last_stmt); | |
1ea6a73c | 236 | if (!types_compatible_p (TREE_TYPE (oprnd0), type) |
237 | || !types_compatible_p (TREE_TYPE (oprnd1), type)) | |
4a61a337 | 238 | return NULL; |
239 | stmt = last_stmt; | |
240 | ||
241 | if (widened_name_p (oprnd0, stmt, &half_type, &def_stmt)) | |
242 | { | |
243 | stmt = def_stmt; | |
75a70cf9 | 244 | oprnd0 = gimple_assign_rhs1 (stmt); |
4a61a337 | 245 | } |
246 | else | |
247 | half_type = type; | |
248 | } | |
249 | ||
250 | /* So far so good. Since last_stmt was detected as a (summation) reduction, | |
251 | we know that oprnd1 is the reduction variable (defined by a loop-header | |
252 | phi), and oprnd0 is an ssa-name defined by a stmt in the loop body. | |
253 | Left to check that oprnd0 is defined by a (widen_)mult_expr */ | |
254 | ||
255 | prod_type = half_type; | |
256 | stmt = SSA_NAME_DEF_STMT (oprnd0); | |
4af22cd9 | 257 | |
258 | /* It could not be the dot_prod pattern if the stmt is outside the loop. */ | |
4ba74662 | 259 | if (!gimple_bb (stmt) || !flow_bb_inside_loop_p (loop, gimple_bb (stmt))) |
4af22cd9 | 260 | return NULL; |
261 | ||
48e1416a | 262 | /* FORNOW. Can continue analyzing the def-use chain when this stmt in a phi |
e3e3ee4b | 263 | inside the loop (in case we are analyzing an outer-loop). */ |
75a70cf9 | 264 | if (!is_gimple_assign (stmt)) |
48e1416a | 265 | return NULL; |
4a61a337 | 266 | stmt_vinfo = vinfo_for_stmt (stmt); |
267 | gcc_assert (stmt_vinfo); | |
f083cd24 | 268 | if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_internal_def) |
23d9a167 | 269 | return NULL; |
75a70cf9 | 270 | if (gimple_assign_rhs_code (stmt) != MULT_EXPR) |
4a61a337 | 271 | return NULL; |
272 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) | |
273 | { | |
274 | /* Has been detected as a widening multiplication? */ | |
275 | ||
276 | stmt = STMT_VINFO_RELATED_STMT (stmt_vinfo); | |
75a70cf9 | 277 | if (gimple_assign_rhs_code (stmt) != WIDEN_MULT_EXPR) |
4a61a337 | 278 | return NULL; |
279 | stmt_vinfo = vinfo_for_stmt (stmt); | |
280 | gcc_assert (stmt_vinfo); | |
f083cd24 | 281 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_internal_def); |
75a70cf9 | 282 | oprnd00 = gimple_assign_rhs1 (stmt); |
283 | oprnd01 = gimple_assign_rhs2 (stmt); | |
4a61a337 | 284 | } |
285 | else | |
286 | { | |
287 | tree half_type0, half_type1; | |
75a70cf9 | 288 | gimple def_stmt; |
4a61a337 | 289 | tree oprnd0, oprnd1; |
290 | ||
75a70cf9 | 291 | oprnd0 = gimple_assign_rhs1 (stmt); |
292 | oprnd1 = gimple_assign_rhs2 (stmt); | |
1ea6a73c | 293 | if (!types_compatible_p (TREE_TYPE (oprnd0), prod_type) |
294 | || !types_compatible_p (TREE_TYPE (oprnd1), prod_type)) | |
4a61a337 | 295 | return NULL; |
296 | if (!widened_name_p (oprnd0, stmt, &half_type0, &def_stmt)) | |
297 | return NULL; | |
75a70cf9 | 298 | oprnd00 = gimple_assign_rhs1 (def_stmt); |
4a61a337 | 299 | if (!widened_name_p (oprnd1, stmt, &half_type1, &def_stmt)) |
300 | return NULL; | |
75a70cf9 | 301 | oprnd01 = gimple_assign_rhs1 (def_stmt); |
1ea6a73c | 302 | if (!types_compatible_p (half_type0, half_type1)) |
4a61a337 | 303 | return NULL; |
304 | if (TYPE_PRECISION (prod_type) != TYPE_PRECISION (half_type0) * 2) | |
305 | return NULL; | |
306 | } | |
307 | ||
308 | half_type = TREE_TYPE (oprnd00); | |
309 | *type_in = half_type; | |
310 | *type_out = type; | |
48e1416a | 311 | |
4a61a337 | 312 | /* Pattern detected. Create a stmt to be used to replace the pattern: */ |
75a70cf9 | 313 | var = vect_recog_temp_ssa_var (type, NULL); |
314 | rhs = build3 (DOT_PROD_EXPR, type, oprnd00, oprnd01, oprnd1), | |
315 | pattern_stmt = gimple_build_assign (var, rhs); | |
48e1416a | 316 | |
4a61a337 | 317 | if (vect_print_dump_info (REPORT_DETAILS)) |
318 | { | |
319 | fprintf (vect_dump, "vect_recog_dot_prod_pattern: detected: "); | |
75a70cf9 | 320 | print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); |
4a61a337 | 321 | } |
221e9a92 | 322 | |
323 | /* We don't allow changing the order of the computation in the inner-loop | |
324 | when doing outer-loop vectorization. */ | |
ade2ac53 | 325 | gcc_assert (!nested_in_vect_loop_p (loop, last_stmt)); |
221e9a92 | 326 | |
75a70cf9 | 327 | return pattern_stmt; |
4a61a337 | 328 | } |
48e1416a | 329 | |
4a61a337 | 330 | /* Function vect_recog_widen_mult_pattern |
331 | ||
332 | Try to find the following pattern: | |
333 | ||
334 | type a_t, b_t; | |
335 | TYPE a_T, b_T, prod_T; | |
336 | ||
337 | S1 a_t = ; | |
338 | S2 b_t = ; | |
339 | S3 a_T = (TYPE) a_t; | |
340 | S4 b_T = (TYPE) b_t; | |
341 | S5 prod_T = a_T * b_T; | |
342 | ||
343 | where type 'TYPE' is at least double the size of type 'type'. | |
344 | ||
345 | Input: | |
346 | ||
347 | * LAST_STMT: A stmt from which the pattern search begins. In the example, | |
348 | when this function is called with S5, the pattern {S3,S4,S5} is be detected. | |
349 | ||
350 | Output: | |
351 | ||
352 | * TYPE_IN: The type of the input arguments to the pattern. | |
353 | ||
354 | * TYPE_OUT: The type of the output of this pattern. | |
355 | ||
356 | * Return value: A new stmt that will be used to replace the sequence of | |
357 | stmts that constitute the pattern. In this case it will be: | |
358 | WIDEN_MULT <a_t, b_t> | |
359 | */ | |
360 | ||
75a70cf9 | 361 | static gimple |
48e1416a | 362 | vect_recog_widen_mult_pattern (gimple last_stmt, |
363 | tree *type_in, | |
c6c91d61 | 364 | tree *type_out) |
4a61a337 | 365 | { |
75a70cf9 | 366 | gimple def_stmt0, def_stmt1; |
c6c91d61 | 367 | tree oprnd0, oprnd1; |
368 | tree type, half_type0, half_type1; | |
75a70cf9 | 369 | gimple pattern_stmt; |
b334cbba | 370 | tree vectype, vectype_out; |
c6c91d61 | 371 | tree dummy; |
75a70cf9 | 372 | tree var; |
c6c91d61 | 373 | enum tree_code dummy_code; |
862bb3cd | 374 | int dummy_int; |
375 | VEC (tree, heap) *dummy_vec; | |
c6c91d61 | 376 | |
75a70cf9 | 377 | if (!is_gimple_assign (last_stmt)) |
c6c91d61 | 378 | return NULL; |
379 | ||
75a70cf9 | 380 | type = gimple_expr_type (last_stmt); |
c6c91d61 | 381 | |
382 | /* Starting from LAST_STMT, follow the defs of its uses in search | |
383 | of the above pattern. */ | |
384 | ||
75a70cf9 | 385 | if (gimple_assign_rhs_code (last_stmt) != MULT_EXPR) |
c6c91d61 | 386 | return NULL; |
387 | ||
75a70cf9 | 388 | oprnd0 = gimple_assign_rhs1 (last_stmt); |
389 | oprnd1 = gimple_assign_rhs2 (last_stmt); | |
1ea6a73c | 390 | if (!types_compatible_p (TREE_TYPE (oprnd0), type) |
391 | || !types_compatible_p (TREE_TYPE (oprnd1), type)) | |
c6c91d61 | 392 | return NULL; |
393 | ||
394 | /* Check argument 0 */ | |
395 | if (!widened_name_p (oprnd0, last_stmt, &half_type0, &def_stmt0)) | |
396 | return NULL; | |
75a70cf9 | 397 | oprnd0 = gimple_assign_rhs1 (def_stmt0); |
c6c91d61 | 398 | |
399 | /* Check argument 1 */ | |
400 | if (!widened_name_p (oprnd1, last_stmt, &half_type1, &def_stmt1)) | |
401 | return NULL; | |
75a70cf9 | 402 | oprnd1 = gimple_assign_rhs1 (def_stmt1); |
c6c91d61 | 403 | |
1ea6a73c | 404 | if (!types_compatible_p (half_type0, half_type1)) |
c6c91d61 | 405 | return NULL; |
406 | ||
407 | /* Pattern detected. */ | |
408 | if (vect_print_dump_info (REPORT_DETAILS)) | |
409 | fprintf (vect_dump, "vect_recog_widen_mult_pattern: detected: "); | |
410 | ||
411 | /* Check target support */ | |
412 | vectype = get_vectype_for_scalar_type (half_type0); | |
b334cbba | 413 | vectype_out = get_vectype_for_scalar_type (type); |
f031fa03 | 414 | if (!vectype |
1548bce8 | 415 | || !vectype_out |
b334cbba | 416 | || !supportable_widening_operation (WIDEN_MULT_EXPR, last_stmt, |
417 | vectype_out, vectype, | |
75a70cf9 | 418 | &dummy, &dummy, &dummy_code, |
862bb3cd | 419 | &dummy_code, &dummy_int, &dummy_vec)) |
c6c91d61 | 420 | return NULL; |
421 | ||
422 | *type_in = vectype; | |
b334cbba | 423 | *type_out = vectype_out; |
c6c91d61 | 424 | |
425 | /* Pattern supported. Create a stmt to be used to replace the pattern: */ | |
75a70cf9 | 426 | var = vect_recog_temp_ssa_var (type, NULL); |
427 | pattern_stmt = gimple_build_assign_with_ops (WIDEN_MULT_EXPR, var, oprnd0, | |
428 | oprnd1); | |
429 | SSA_NAME_DEF_STMT (var) = pattern_stmt; | |
430 | ||
c6c91d61 | 431 | if (vect_print_dump_info (REPORT_DETAILS)) |
75a70cf9 | 432 | print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); |
433 | ||
434 | return pattern_stmt; | |
4a61a337 | 435 | } |
436 | ||
437 | ||
c37bea13 | 438 | /* Function vect_recog_pow_pattern |
439 | ||
440 | Try to find the following pattern: | |
441 | ||
442 | x = POW (y, N); | |
443 | ||
444 | with POW being one of pow, powf, powi, powif and N being | |
445 | either 2 or 0.5. | |
446 | ||
447 | Input: | |
448 | ||
449 | * LAST_STMT: A stmt from which the pattern search begins. | |
450 | ||
451 | Output: | |
452 | ||
453 | * TYPE_IN: The type of the input arguments to the pattern. | |
454 | ||
455 | * TYPE_OUT: The type of the output of this pattern. | |
456 | ||
457 | * Return value: A new stmt that will be used to replace the sequence of | |
458 | stmts that constitute the pattern. In this case it will be: | |
75a70cf9 | 459 | x = x * x |
c37bea13 | 460 | or |
75a70cf9 | 461 | x = sqrt (x) |
c37bea13 | 462 | */ |
463 | ||
75a70cf9 | 464 | static gimple |
465 | vect_recog_pow_pattern (gimple last_stmt, tree *type_in, tree *type_out) | |
c37bea13 | 466 | { |
75a70cf9 | 467 | tree fn, base, exp = NULL; |
468 | gimple stmt; | |
469 | tree var; | |
c37bea13 | 470 | |
75a70cf9 | 471 | if (!is_gimple_call (last_stmt) || gimple_call_lhs (last_stmt) == NULL) |
c37bea13 | 472 | return NULL; |
473 | ||
75a70cf9 | 474 | fn = gimple_call_fndecl (last_stmt); |
c37bea13 | 475 | switch (DECL_FUNCTION_CODE (fn)) |
476 | { | |
477 | case BUILT_IN_POWIF: | |
478 | case BUILT_IN_POWI: | |
479 | case BUILT_IN_POWF: | |
480 | case BUILT_IN_POW: | |
75a70cf9 | 481 | base = gimple_call_arg (last_stmt, 0); |
482 | exp = gimple_call_arg (last_stmt, 1); | |
c37bea13 | 483 | if (TREE_CODE (exp) != REAL_CST |
484 | && TREE_CODE (exp) != INTEGER_CST) | |
75a70cf9 | 485 | return NULL; |
c37bea13 | 486 | break; |
487 | ||
75a70cf9 | 488 | default: |
489 | return NULL; | |
c37bea13 | 490 | } |
491 | ||
492 | /* We now have a pow or powi builtin function call with a constant | |
493 | exponent. */ | |
494 | ||
c37bea13 | 495 | *type_out = NULL_TREE; |
496 | ||
497 | /* Catch squaring. */ | |
498 | if ((host_integerp (exp, 0) | |
499 | && tree_low_cst (exp, 0) == 2) | |
500 | || (TREE_CODE (exp) == REAL_CST | |
501 | && REAL_VALUES_EQUAL (TREE_REAL_CST (exp), dconst2))) | |
4957645d | 502 | { |
503 | *type_in = TREE_TYPE (base); | |
75a70cf9 | 504 | |
505 | var = vect_recog_temp_ssa_var (TREE_TYPE (base), NULL); | |
506 | stmt = gimple_build_assign_with_ops (MULT_EXPR, var, base, base); | |
507 | SSA_NAME_DEF_STMT (var) = stmt; | |
508 | return stmt; | |
4957645d | 509 | } |
c37bea13 | 510 | |
511 | /* Catch square root. */ | |
512 | if (TREE_CODE (exp) == REAL_CST | |
513 | && REAL_VALUES_EQUAL (TREE_REAL_CST (exp), dconsthalf)) | |
514 | { | |
515 | tree newfn = mathfn_built_in (TREE_TYPE (base), BUILT_IN_SQRT); | |
4957645d | 516 | *type_in = get_vectype_for_scalar_type (TREE_TYPE (base)); |
517 | if (*type_in) | |
518 | { | |
75a70cf9 | 519 | gimple stmt = gimple_build_call (newfn, 1, base); |
520 | if (vectorizable_function (stmt, *type_in, *type_in) | |
521 | != NULL_TREE) | |
522 | { | |
523 | var = vect_recog_temp_ssa_var (TREE_TYPE (base), stmt); | |
48e1416a | 524 | gimple_call_set_lhs (stmt, var); |
75a70cf9 | 525 | return stmt; |
526 | } | |
4957645d | 527 | } |
c37bea13 | 528 | } |
529 | ||
75a70cf9 | 530 | return NULL; |
c37bea13 | 531 | } |
532 | ||
533 | ||
4a61a337 | 534 | /* Function vect_recog_widen_sum_pattern |
535 | ||
536 | Try to find the following pattern: | |
537 | ||
48e1416a | 538 | type x_t; |
4a61a337 | 539 | TYPE x_T, sum = init; |
540 | loop: | |
541 | sum_0 = phi <init, sum_1> | |
542 | S1 x_t = *p; | |
543 | S2 x_T = (TYPE) x_t; | |
544 | S3 sum_1 = x_T + sum_0; | |
545 | ||
48e1416a | 546 | where type 'TYPE' is at least double the size of type 'type', i.e - we're |
4a61a337 | 547 | summing elements of type 'type' into an accumulator of type 'TYPE'. This is |
9ca2c29a | 548 | a special case of a reduction computation. |
4a61a337 | 549 | |
550 | Input: | |
551 | ||
552 | * LAST_STMT: A stmt from which the pattern search begins. In the example, | |
553 | when this function is called with S3, the pattern {S2,S3} will be detected. | |
48e1416a | 554 | |
4a61a337 | 555 | Output: |
48e1416a | 556 | |
4a61a337 | 557 | * TYPE_IN: The type of the input arguments to the pattern. |
558 | ||
559 | * TYPE_OUT: The type of the output of this pattern. | |
560 | ||
561 | * Return value: A new stmt that will be used to replace the sequence of | |
562 | stmts that constitute the pattern. In this case it will be: | |
563 | WIDEN_SUM <x_t, sum_0> | |
221e9a92 | 564 | |
48e1416a | 565 | Note: The widening-sum idiom is a widening reduction pattern that is |
221e9a92 | 566 | vectorized without preserving all the intermediate results. It |
48e1416a | 567 | produces only N/2 (widened) results (by summing up pairs of |
568 | intermediate results) rather than all N results. Therefore, we | |
569 | cannot allow this pattern when we want to get all the results and in | |
570 | the correct order (as is the case when this computation is in an | |
221e9a92 | 571 | inner-loop nested in an outer-loop that us being vectorized). */ |
4a61a337 | 572 | |
75a70cf9 | 573 | static gimple |
574 | vect_recog_widen_sum_pattern (gimple last_stmt, tree *type_in, tree *type_out) | |
4a61a337 | 575 | { |
75a70cf9 | 576 | gimple stmt; |
4a61a337 | 577 | tree oprnd0, oprnd1; |
578 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt); | |
579 | tree type, half_type; | |
75a70cf9 | 580 | gimple pattern_stmt; |
221e9a92 | 581 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
582 | struct loop *loop = LOOP_VINFO_LOOP (loop_info); | |
75a70cf9 | 583 | tree var; |
4a61a337 | 584 | |
75a70cf9 | 585 | if (!is_gimple_assign (last_stmt)) |
4a61a337 | 586 | return NULL; |
587 | ||
75a70cf9 | 588 | type = gimple_expr_type (last_stmt); |
4a61a337 | 589 | |
590 | /* Look for the following pattern | |
591 | DX = (TYPE) X; | |
592 | sum_1 = DX + sum_0; | |
593 | In which DX is at least double the size of X, and sum_1 has been | |
594 | recognized as a reduction variable. | |
595 | */ | |
596 | ||
597 | /* Starting from LAST_STMT, follow the defs of its uses in search | |
598 | of the above pattern. */ | |
599 | ||
75a70cf9 | 600 | if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR) |
4a61a337 | 601 | return NULL; |
602 | ||
603 | if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def) | |
604 | return NULL; | |
605 | ||
75a70cf9 | 606 | oprnd0 = gimple_assign_rhs1 (last_stmt); |
607 | oprnd1 = gimple_assign_rhs2 (last_stmt); | |
1ea6a73c | 608 | if (!types_compatible_p (TREE_TYPE (oprnd0), type) |
609 | || !types_compatible_p (TREE_TYPE (oprnd1), type)) | |
4a61a337 | 610 | return NULL; |
611 | ||
612 | /* So far so good. Since last_stmt was detected as a (summation) reduction, | |
613 | we know that oprnd1 is the reduction variable (defined by a loop-header | |
614 | phi), and oprnd0 is an ssa-name defined by a stmt in the loop body. | |
615 | Left to check that oprnd0 is defined by a cast from type 'type' to type | |
616 | 'TYPE'. */ | |
617 | ||
618 | if (!widened_name_p (oprnd0, last_stmt, &half_type, &stmt)) | |
619 | return NULL; | |
620 | ||
75a70cf9 | 621 | oprnd0 = gimple_assign_rhs1 (stmt); |
4a61a337 | 622 | *type_in = half_type; |
623 | *type_out = type; | |
624 | ||
625 | /* Pattern detected. Create a stmt to be used to replace the pattern: */ | |
75a70cf9 | 626 | var = vect_recog_temp_ssa_var (type, NULL); |
627 | pattern_stmt = gimple_build_assign_with_ops (WIDEN_SUM_EXPR, var, | |
628 | oprnd0, oprnd1); | |
629 | SSA_NAME_DEF_STMT (var) = pattern_stmt; | |
630 | ||
4a61a337 | 631 | if (vect_print_dump_info (REPORT_DETAILS)) |
632 | { | |
633 | fprintf (vect_dump, "vect_recog_widen_sum_pattern: detected: "); | |
75a70cf9 | 634 | print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); |
4a61a337 | 635 | } |
221e9a92 | 636 | |
637 | /* We don't allow changing the order of the computation in the inner-loop | |
638 | when doing outer-loop vectorization. */ | |
ade2ac53 | 639 | gcc_assert (!nested_in_vect_loop_p (loop, last_stmt)); |
221e9a92 | 640 | |
75a70cf9 | 641 | return pattern_stmt; |
4a61a337 | 642 | } |
643 | ||
644 | ||
48e1416a | 645 | /* Function vect_pattern_recog_1 |
4a61a337 | 646 | |
647 | Input: | |
648 | PATTERN_RECOG_FUNC: A pointer to a function that detects a certain | |
649 | computation pattern. | |
650 | STMT: A stmt from which the pattern search should start. | |
651 | ||
652 | If PATTERN_RECOG_FUNC successfully detected the pattern, it creates an | |
48e1416a | 653 | expression that computes the same functionality and can be used to |
654 | replace the sequence of stmts that are involved in the pattern. | |
4a61a337 | 655 | |
656 | Output: | |
48e1416a | 657 | This function checks if the expression returned by PATTERN_RECOG_FUNC is |
658 | supported in vector form by the target. We use 'TYPE_IN' to obtain the | |
659 | relevant vector type. If 'TYPE_IN' is already a vector type, then this | |
4a61a337 | 660 | indicates that target support had already been checked by PATTERN_RECOG_FUNC. |
661 | If 'TYPE_OUT' is also returned by PATTERN_RECOG_FUNC, we check that it fits | |
662 | to the available target pattern. | |
663 | ||
48e1416a | 664 | This function also does some bookkeeping, as explained in the documentation |
4a61a337 | 665 | for vect_recog_pattern. */ |
666 | ||
667 | static void | |
668 | vect_pattern_recog_1 ( | |
75a70cf9 | 669 | gimple (* vect_recog_func) (gimple, tree *, tree *), |
670 | gimple_stmt_iterator si) | |
4a61a337 | 671 | { |
75a70cf9 | 672 | gimple stmt = gsi_stmt (si), pattern_stmt; |
4a61a337 | 673 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
674 | stmt_vec_info pattern_stmt_info; | |
675 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4a61a337 | 676 | tree pattern_vectype; |
677 | tree type_in, type_out; | |
4a61a337 | 678 | enum tree_code code; |
eefa05c8 | 679 | int i; |
680 | gimple next; | |
4a61a337 | 681 | |
75a70cf9 | 682 | pattern_stmt = (* vect_recog_func) (stmt, &type_in, &type_out); |
683 | if (!pattern_stmt) | |
48e1416a | 684 | return; |
685 | ||
686 | if (VECTOR_MODE_P (TYPE_MODE (type_in))) | |
687 | { | |
688 | /* No need to check target support (already checked by the pattern | |
689 | recognition function). */ | |
b334cbba | 690 | if (type_out) |
691 | gcc_assert (VECTOR_MODE_P (TYPE_MODE (type_out))); | |
692 | pattern_vectype = type_out ? type_out : type_in; | |
4a61a337 | 693 | } |
694 | else | |
695 | { | |
8458f4ca | 696 | enum machine_mode vec_mode; |
4a61a337 | 697 | enum insn_code icode; |
698 | optab optab; | |
699 | ||
700 | /* Check target support */ | |
b334cbba | 701 | type_in = get_vectype_for_scalar_type (type_in); |
702 | if (!type_in) | |
703 | return; | |
704 | if (type_out) | |
705 | type_out = get_vectype_for_scalar_type (type_out); | |
706 | else | |
707 | type_out = type_in; | |
ed67497f | 708 | if (!type_out) |
709 | return; | |
b334cbba | 710 | pattern_vectype = type_out; |
f031fa03 | 711 | |
75a70cf9 | 712 | if (is_gimple_assign (pattern_stmt)) |
713 | code = gimple_assign_rhs_code (pattern_stmt); | |
714 | else | |
715 | { | |
716 | gcc_assert (is_gimple_call (pattern_stmt)); | |
717 | code = CALL_EXPR; | |
718 | } | |
719 | ||
b334cbba | 720 | optab = optab_for_tree_code (code, type_in, optab_default); |
721 | vec_mode = TYPE_MODE (type_in); | |
4a61a337 | 722 | if (!optab |
d6bf3b14 | 723 | || (icode = optab_handler (optab, vec_mode)) == CODE_FOR_nothing |
b334cbba | 724 | || (insn_data[icode].operand[0].mode != TYPE_MODE (type_out))) |
4a61a337 | 725 | return; |
726 | } | |
727 | ||
728 | /* Found a vectorizable pattern. */ | |
729 | if (vect_print_dump_info (REPORT_DETAILS)) | |
730 | { | |
48e1416a | 731 | fprintf (vect_dump, "pattern recognized: "); |
75a70cf9 | 732 | print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); |
4a61a337 | 733 | } |
48e1416a | 734 | |
75a70cf9 | 735 | /* Mark the stmts that are involved in the pattern. */ |
736 | gsi_insert_before (&si, pattern_stmt, GSI_SAME_STMT); | |
737 | set_vinfo_for_stmt (pattern_stmt, | |
37545e54 | 738 | new_stmt_vec_info (pattern_stmt, loop_vinfo, NULL)); |
75a70cf9 | 739 | pattern_stmt_info = vinfo_for_stmt (pattern_stmt); |
48e1416a | 740 | |
4a61a337 | 741 | STMT_VINFO_RELATED_STMT (pattern_stmt_info) = stmt; |
742 | STMT_VINFO_DEF_TYPE (pattern_stmt_info) = STMT_VINFO_DEF_TYPE (stmt_info); | |
743 | STMT_VINFO_VECTYPE (pattern_stmt_info) = pattern_vectype; | |
744 | STMT_VINFO_IN_PATTERN_P (stmt_info) = true; | |
75a70cf9 | 745 | STMT_VINFO_RELATED_STMT (stmt_info) = pattern_stmt; |
4a61a337 | 746 | |
eefa05c8 | 747 | /* Patterns cannot be vectorized using SLP, because they change the order of |
748 | computation. */ | |
48148244 | 749 | FOR_EACH_VEC_ELT (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo), i, next) |
eefa05c8 | 750 | if (next == stmt) |
751 | VEC_ordered_remove (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo), i); | |
4a61a337 | 752 | } |
753 | ||
754 | ||
755 | /* Function vect_pattern_recog | |
756 | ||
757 | Input: | |
758 | LOOP_VINFO - a struct_loop_info of a loop in which we want to look for | |
759 | computation idioms. | |
760 | ||
761 | Output - for each computation idiom that is detected we insert a new stmt | |
762 | that provides the same functionality and that can be vectorized. We | |
763 | also record some information in the struct_stmt_info of the relevant | |
764 | stmts, as explained below: | |
765 | ||
766 | At the entry to this function we have the following stmts, with the | |
767 | following initial value in the STMT_VINFO fields: | |
768 | ||
769 | stmt in_pattern_p related_stmt vec_stmt | |
770 | S1: a_i = .... - - - | |
771 | S2: a_2 = ..use(a_i).. - - - | |
772 | S3: a_1 = ..use(a_2).. - - - | |
773 | S4: a_0 = ..use(a_1).. - - - | |
774 | S5: ... = ..use(a_0).. - - - | |
775 | ||
776 | Say the sequence {S1,S2,S3,S4} was detected as a pattern that can be | |
777 | represented by a single stmt. We then: | |
778 | - create a new stmt S6 that will replace the pattern. | |
779 | - insert the new stmt S6 before the last stmt in the pattern | |
780 | - fill in the STMT_VINFO fields as follows: | |
781 | ||
782 | in_pattern_p related_stmt vec_stmt | |
48e1416a | 783 | S1: a_i = .... - - - |
4a61a337 | 784 | S2: a_2 = ..use(a_i).. - - - |
785 | S3: a_1 = ..use(a_2).. - - - | |
786 | > S6: a_new = .... - S4 - | |
787 | S4: a_0 = ..use(a_1).. true S6 - | |
788 | S5: ... = ..use(a_0).. - - - | |
789 | ||
790 | (the last stmt in the pattern (S4) and the new pattern stmt (S6) point | |
791 | to each other through the RELATED_STMT field). | |
792 | ||
793 | S6 will be marked as relevant in vect_mark_stmts_to_be_vectorized instead | |
794 | of S4 because it will replace all its uses. Stmts {S1,S2,S3} will | |
795 | remain irrelevant unless used by stmts other than S4. | |
796 | ||
797 | If vectorization succeeds, vect_transform_stmt will skip over {S1,S2,S3} | |
334ec2d8 | 798 | (because they are marked as irrelevant). It will vectorize S6, and record |
48e1416a | 799 | a pointer to the new vector stmt VS6 both from S6 (as usual), and also |
4a61a337 | 800 | from S4. We do that so that when we get to vectorizing stmts that use the |
801 | def of S4 (like S5 that uses a_0), we'll know where to take the relevant | |
802 | vector-def from. S4 will be skipped, and S5 will be vectorized as usual: | |
803 | ||
804 | in_pattern_p related_stmt vec_stmt | |
805 | S1: a_i = .... - - - | |
806 | S2: a_2 = ..use(a_i).. - - - | |
807 | S3: a_1 = ..use(a_2).. - - - | |
808 | > VS6: va_new = .... - - - | |
809 | S6: a_new = .... - S4 VS6 | |
810 | S4: a_0 = ..use(a_1).. true S6 VS6 | |
811 | > VS5: ... = ..vuse(va_new).. - - - | |
812 | S5: ... = ..use(a_0).. - - - | |
813 | ||
814 | DCE could then get rid of {S1,S2,S3,S4,S5,S6} (if their defs are not used | |
815 | elsewhere), and we'll end up with: | |
816 | ||
48e1416a | 817 | VS6: va_new = .... |
4a61a337 | 818 | VS5: ... = ..vuse(va_new).. |
819 | ||
820 | If vectorization does not succeed, DCE will clean S6 away (its def is | |
821 | not used), and we'll end up with the original sequence. | |
822 | */ | |
823 | ||
824 | void | |
825 | vect_pattern_recog (loop_vec_info loop_vinfo) | |
826 | { | |
827 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
828 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
829 | unsigned int nbbs = loop->num_nodes; | |
75a70cf9 | 830 | gimple_stmt_iterator si; |
4a61a337 | 831 | unsigned int i, j; |
75a70cf9 | 832 | gimple (* vect_recog_func_ptr) (gimple, tree *, tree *); |
4a61a337 | 833 | |
834 | if (vect_print_dump_info (REPORT_DETAILS)) | |
835 | fprintf (vect_dump, "=== vect_pattern_recog ==="); | |
836 | ||
837 | /* Scan through the loop stmts, applying the pattern recognition | |
838 | functions starting at each stmt visited: */ | |
839 | for (i = 0; i < nbbs; i++) | |
840 | { | |
841 | basic_block bb = bbs[i]; | |
75a70cf9 | 842 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) |
4a61a337 | 843 | { |
4a61a337 | 844 | /* Scan over all generic vect_recog_xxx_pattern functions. */ |
845 | for (j = 0; j < NUM_PATTERNS; j++) | |
846 | { | |
847 | vect_recog_func_ptr = vect_vect_recog_func_ptrs[j]; | |
848 | vect_pattern_recog_1 (vect_recog_func_ptr, si); | |
849 | } | |
850 | } | |
851 | } | |
852 | } |