]>
Commit | Line | Data |
---|---|---|
20f06221 | 1 | /* Analysis Utilities for Loop Vectorization. |
ebfd146a | 2 | Copyright (C) 2006, 2007, 2008, 2009 Free Software Foundation, Inc. |
20f06221 DN |
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 | |
9dcd6f09 | 9 | Software Foundation; either version 3, or (at your option) any later |
20f06221 DN |
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 | |
9dcd6f09 NC |
18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
20f06221 DN |
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" | |
20f06221 DN |
27 | #include "target.h" |
28 | #include "basic-block.h" | |
29 | #include "diagnostic.h" | |
30 | #include "tree-flow.h" | |
31 | #include "tree-dump.h" | |
20f06221 DN |
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" | |
39 | #include "toplev.h" | |
40 | ||
c0220ea4 | 41 | /* Function prototypes */ |
20f06221 | 42 | static void vect_pattern_recog_1 |
726a989a RB |
43 | (gimple (* ) (gimple, tree *, tree *), gimple_stmt_iterator); |
44 | static bool widened_name_p (tree, gimple, tree *, gimple *); | |
20f06221 DN |
45 | |
46 | /* Pattern recognition functions */ | |
726a989a RB |
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 *); | |
20f06221 DN |
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, | |
0b2229b0 RG |
54 | vect_recog_dot_prod_pattern, |
55 | vect_recog_pow_pattern}; | |
20f06221 DN |
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) | |
63 | where the type of name0 (HALF_TYPE) is smaller than the type of NAME. | |
64 | */ | |
65 | ||
66 | static bool | |
726a989a | 67 | widened_name_p (tree name, gimple use_stmt, tree *half_type, gimple *def_stmt) |
20f06221 DN |
68 | { |
69 | tree dummy; | |
726a989a | 70 | gimple dummy_gimple; |
20f06221 DN |
71 | loop_vec_info loop_vinfo; |
72 | stmt_vec_info stmt_vinfo; | |
20f06221 DN |
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 | ||
a70d6342 | 81 | if (!vect_is_simple_use (name, loop_vinfo, NULL, def_stmt, &def, &dt)) |
20f06221 DN |
82 | return false; |
83 | ||
8644a673 IR |
84 | if (dt != vect_internal_def |
85 | && dt != vect_external_def && dt != vect_constant_def) | |
20f06221 DN |
86 | return false; |
87 | ||
88 | if (! *def_stmt) | |
89 | return false; | |
90 | ||
726a989a | 91 | if (!is_gimple_assign (*def_stmt)) |
20f06221 DN |
92 | return false; |
93 | ||
726a989a | 94 | if (gimple_assign_rhs_code (*def_stmt) != NOP_EXPR) |
20f06221 DN |
95 | return false; |
96 | ||
726a989a | 97 | oprnd0 = gimple_assign_rhs1 (*def_stmt); |
20f06221 DN |
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 | ||
a70d6342 IR |
105 | if (!vect_is_simple_use (oprnd0, loop_vinfo, NULL, &dummy_gimple, &dummy, |
106 | &dt)) | |
20f06221 DN |
107 | return false; |
108 | ||
20f06221 DN |
109 | return true; |
110 | } | |
111 | ||
726a989a RB |
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 | } | |
20f06221 DN |
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 | ||
142 | where 'TYPE1' is exactly double the size of type 'type', and 'TYPE2' is the | |
c0220ea4 | 143 | same size of 'TYPE1' or bigger. This is a special case of a reduction |
20f06221 DN |
144 | computation. |
145 | ||
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> | |
d29de1bf DN |
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). */ | |
20f06221 | 169 | |
726a989a RB |
170 | static gimple |
171 | vect_recog_dot_prod_pattern (gimple last_stmt, tree *type_in, tree *type_out) | |
20f06221 | 172 | { |
726a989a | 173 | gimple stmt; |
20f06221 DN |
174 | tree oprnd0, oprnd1; |
175 | tree oprnd00, oprnd01; | |
176 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt); | |
177 | tree type, half_type; | |
726a989a | 178 | gimple pattern_stmt; |
20f06221 | 179 | tree prod_type; |
d29de1bf DN |
180 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
181 | struct loop *loop = LOOP_VINFO_LOOP (loop_info); | |
726a989a | 182 | tree var, rhs; |
20f06221 | 183 | |
726a989a | 184 | if (!is_gimple_assign (last_stmt)) |
20f06221 DN |
185 | return NULL; |
186 | ||
726a989a | 187 | type = gimple_expr_type (last_stmt); |
20f06221 DN |
188 | |
189 | /* Look for the following pattern | |
190 | DX = (TYPE1) X; | |
191 | DY = (TYPE1) Y; | |
192 | DPROD = DX * DY; | |
193 | DDPROD = (TYPE2) DPROD; | |
194 | sum_1 = DDPROD + sum_0; | |
195 | In which | |
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 | ||
726a989a | 213 | if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR) |
20f06221 DN |
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); | |
726a989a RB |
221 | type = gimple_expr_type (stmt); |
222 | if (gimple_assign_rhs_code (stmt) != WIDEN_SUM_EXPR) | |
20f06221 | 223 | return NULL; |
726a989a RB |
224 | oprnd0 = gimple_assign_rhs1 (stmt); |
225 | oprnd1 = gimple_assign_rhs2 (stmt); | |
20f06221 DN |
226 | half_type = TREE_TYPE (oprnd0); |
227 | } | |
228 | else | |
229 | { | |
726a989a | 230 | gimple def_stmt; |
20f06221 DN |
231 | |
232 | if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def) | |
233 | return NULL; | |
726a989a RB |
234 | oprnd0 = gimple_assign_rhs1 (last_stmt); |
235 | oprnd1 = gimple_assign_rhs2 (last_stmt); | |
9600efe1 MM |
236 | if (!types_compatible_p (TREE_TYPE (oprnd0), type) |
237 | || !types_compatible_p (TREE_TYPE (oprnd1), type)) | |
20f06221 DN |
238 | return NULL; |
239 | stmt = last_stmt; | |
240 | ||
241 | if (widened_name_p (oprnd0, stmt, &half_type, &def_stmt)) | |
242 | { | |
243 | stmt = def_stmt; | |
726a989a | 244 | oprnd0 = gimple_assign_rhs1 (stmt); |
20f06221 DN |
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); | |
8665227f RG |
257 | /* FORNOW. Can continue analyzing the def-use chain when this stmt in a phi |
258 | inside the loop (in case we are analyzing an outer-loop). */ | |
726a989a | 259 | if (!is_gimple_assign (stmt)) |
8665227f | 260 | return NULL; |
20f06221 DN |
261 | stmt_vinfo = vinfo_for_stmt (stmt); |
262 | gcc_assert (stmt_vinfo); | |
8644a673 | 263 | if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_internal_def) |
b3130586 | 264 | return NULL; |
726a989a | 265 | if (gimple_assign_rhs_code (stmt) != MULT_EXPR) |
20f06221 DN |
266 | return NULL; |
267 | if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)) | |
268 | { | |
269 | /* Has been detected as a widening multiplication? */ | |
270 | ||
271 | stmt = STMT_VINFO_RELATED_STMT (stmt_vinfo); | |
726a989a | 272 | if (gimple_assign_rhs_code (stmt) != WIDEN_MULT_EXPR) |
20f06221 DN |
273 | return NULL; |
274 | stmt_vinfo = vinfo_for_stmt (stmt); | |
275 | gcc_assert (stmt_vinfo); | |
8644a673 | 276 | gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_internal_def); |
726a989a RB |
277 | oprnd00 = gimple_assign_rhs1 (stmt); |
278 | oprnd01 = gimple_assign_rhs2 (stmt); | |
20f06221 DN |
279 | } |
280 | else | |
281 | { | |
282 | tree half_type0, half_type1; | |
726a989a | 283 | gimple def_stmt; |
20f06221 DN |
284 | tree oprnd0, oprnd1; |
285 | ||
726a989a RB |
286 | oprnd0 = gimple_assign_rhs1 (stmt); |
287 | oprnd1 = gimple_assign_rhs2 (stmt); | |
9600efe1 MM |
288 | if (!types_compatible_p (TREE_TYPE (oprnd0), prod_type) |
289 | || !types_compatible_p (TREE_TYPE (oprnd1), prod_type)) | |
20f06221 DN |
290 | return NULL; |
291 | if (!widened_name_p (oprnd0, stmt, &half_type0, &def_stmt)) | |
292 | return NULL; | |
726a989a | 293 | oprnd00 = gimple_assign_rhs1 (def_stmt); |
20f06221 DN |
294 | if (!widened_name_p (oprnd1, stmt, &half_type1, &def_stmt)) |
295 | return NULL; | |
726a989a | 296 | oprnd01 = gimple_assign_rhs1 (def_stmt); |
9600efe1 | 297 | if (!types_compatible_p (half_type0, half_type1)) |
20f06221 DN |
298 | return NULL; |
299 | if (TYPE_PRECISION (prod_type) != TYPE_PRECISION (half_type0) * 2) | |
300 | return NULL; | |
301 | } | |
302 | ||
303 | half_type = TREE_TYPE (oprnd00); | |
304 | *type_in = half_type; | |
305 | *type_out = type; | |
306 | ||
307 | /* Pattern detected. Create a stmt to be used to replace the pattern: */ | |
726a989a RB |
308 | var = vect_recog_temp_ssa_var (type, NULL); |
309 | rhs = build3 (DOT_PROD_EXPR, type, oprnd00, oprnd01, oprnd1), | |
310 | pattern_stmt = gimple_build_assign (var, rhs); | |
311 | ||
20f06221 DN |
312 | if (vect_print_dump_info (REPORT_DETAILS)) |
313 | { | |
314 | fprintf (vect_dump, "vect_recog_dot_prod_pattern: detected: "); | |
726a989a | 315 | print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); |
20f06221 | 316 | } |
d29de1bf DN |
317 | |
318 | /* We don't allow changing the order of the computation in the inner-loop | |
319 | when doing outer-loop vectorization. */ | |
7c5222ff | 320 | gcc_assert (!nested_in_vect_loop_p (loop, last_stmt)); |
d29de1bf | 321 | |
726a989a | 322 | return pattern_stmt; |
20f06221 | 323 | } |
726a989a | 324 | |
20f06221 DN |
325 | /* Function vect_recog_widen_mult_pattern |
326 | ||
327 | Try to find the following pattern: | |
328 | ||
329 | type a_t, b_t; | |
330 | TYPE a_T, b_T, prod_T; | |
331 | ||
332 | S1 a_t = ; | |
333 | S2 b_t = ; | |
334 | S3 a_T = (TYPE) a_t; | |
335 | S4 b_T = (TYPE) b_t; | |
336 | S5 prod_T = a_T * b_T; | |
337 | ||
338 | where type 'TYPE' is at least double the size of type 'type'. | |
339 | ||
340 | Input: | |
341 | ||
342 | * LAST_STMT: A stmt from which the pattern search begins. In the example, | |
343 | when this function is called with S5, the pattern {S3,S4,S5} is be detected. | |
344 | ||
345 | Output: | |
346 | ||
347 | * TYPE_IN: The type of the input arguments to the pattern. | |
348 | ||
349 | * TYPE_OUT: The type of the output of this pattern. | |
350 | ||
351 | * Return value: A new stmt that will be used to replace the sequence of | |
352 | stmts that constitute the pattern. In this case it will be: | |
353 | WIDEN_MULT <a_t, b_t> | |
354 | */ | |
355 | ||
726a989a RB |
356 | static gimple |
357 | vect_recog_widen_mult_pattern (gimple last_stmt, | |
89d67cca DN |
358 | tree *type_in, |
359 | tree *type_out) | |
20f06221 | 360 | { |
726a989a | 361 | gimple def_stmt0, def_stmt1; |
89d67cca DN |
362 | tree oprnd0, oprnd1; |
363 | tree type, half_type0, half_type1; | |
726a989a | 364 | gimple pattern_stmt; |
89d67cca DN |
365 | tree vectype; |
366 | tree dummy; | |
726a989a | 367 | tree var; |
89d67cca | 368 | enum tree_code dummy_code; |
5d593372 IR |
369 | int dummy_int; |
370 | VEC (tree, heap) *dummy_vec; | |
89d67cca | 371 | |
726a989a | 372 | if (!is_gimple_assign (last_stmt)) |
89d67cca DN |
373 | return NULL; |
374 | ||
726a989a | 375 | type = gimple_expr_type (last_stmt); |
89d67cca DN |
376 | |
377 | /* Starting from LAST_STMT, follow the defs of its uses in search | |
378 | of the above pattern. */ | |
379 | ||
726a989a | 380 | if (gimple_assign_rhs_code (last_stmt) != MULT_EXPR) |
89d67cca DN |
381 | return NULL; |
382 | ||
726a989a RB |
383 | oprnd0 = gimple_assign_rhs1 (last_stmt); |
384 | oprnd1 = gimple_assign_rhs2 (last_stmt); | |
9600efe1 MM |
385 | if (!types_compatible_p (TREE_TYPE (oprnd0), type) |
386 | || !types_compatible_p (TREE_TYPE (oprnd1), type)) | |
89d67cca DN |
387 | return NULL; |
388 | ||
389 | /* Check argument 0 */ | |
390 | if (!widened_name_p (oprnd0, last_stmt, &half_type0, &def_stmt0)) | |
391 | return NULL; | |
726a989a | 392 | oprnd0 = gimple_assign_rhs1 (def_stmt0); |
89d67cca DN |
393 | |
394 | /* Check argument 1 */ | |
395 | if (!widened_name_p (oprnd1, last_stmt, &half_type1, &def_stmt1)) | |
396 | return NULL; | |
726a989a | 397 | oprnd1 = gimple_assign_rhs1 (def_stmt1); |
89d67cca | 398 | |
9600efe1 | 399 | if (!types_compatible_p (half_type0, half_type1)) |
89d67cca DN |
400 | return NULL; |
401 | ||
402 | /* Pattern detected. */ | |
403 | if (vect_print_dump_info (REPORT_DETAILS)) | |
404 | fprintf (vect_dump, "vect_recog_widen_mult_pattern: detected: "); | |
405 | ||
406 | /* Check target support */ | |
407 | vectype = get_vectype_for_scalar_type (half_type0); | |
03d3e953 IR |
408 | if (!vectype |
409 | || !supportable_widening_operation (WIDEN_MULT_EXPR, last_stmt, vectype, | |
726a989a | 410 | &dummy, &dummy, &dummy_code, |
5d593372 | 411 | &dummy_code, &dummy_int, &dummy_vec)) |
89d67cca DN |
412 | return NULL; |
413 | ||
414 | *type_in = vectype; | |
415 | *type_out = NULL_TREE; | |
416 | ||
417 | /* Pattern supported. Create a stmt to be used to replace the pattern: */ | |
726a989a RB |
418 | var = vect_recog_temp_ssa_var (type, NULL); |
419 | pattern_stmt = gimple_build_assign_with_ops (WIDEN_MULT_EXPR, var, oprnd0, | |
420 | oprnd1); | |
421 | SSA_NAME_DEF_STMT (var) = pattern_stmt; | |
422 | ||
89d67cca | 423 | if (vect_print_dump_info (REPORT_DETAILS)) |
726a989a RB |
424 | print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); |
425 | ||
426 | return pattern_stmt; | |
20f06221 DN |
427 | } |
428 | ||
429 | ||
0b2229b0 RG |
430 | /* Function vect_recog_pow_pattern |
431 | ||
432 | Try to find the following pattern: | |
433 | ||
434 | x = POW (y, N); | |
435 | ||
436 | with POW being one of pow, powf, powi, powif and N being | |
437 | either 2 or 0.5. | |
438 | ||
439 | Input: | |
440 | ||
441 | * LAST_STMT: A stmt from which the pattern search begins. | |
442 | ||
443 | Output: | |
444 | ||
445 | * TYPE_IN: The type of the input arguments to the pattern. | |
446 | ||
447 | * TYPE_OUT: The type of the output of this pattern. | |
448 | ||
449 | * Return value: A new stmt that will be used to replace the sequence of | |
450 | stmts that constitute the pattern. In this case it will be: | |
726a989a | 451 | x = x * x |
0b2229b0 | 452 | or |
726a989a | 453 | x = sqrt (x) |
0b2229b0 RG |
454 | */ |
455 | ||
726a989a RB |
456 | static gimple |
457 | vect_recog_pow_pattern (gimple last_stmt, tree *type_in, tree *type_out) | |
0b2229b0 | 458 | { |
0b2229b0 | 459 | tree type; |
726a989a RB |
460 | tree fn, base, exp = NULL; |
461 | gimple stmt; | |
462 | tree var; | |
0b2229b0 | 463 | |
726a989a | 464 | if (!is_gimple_call (last_stmt) || gimple_call_lhs (last_stmt) == NULL) |
0b2229b0 RG |
465 | return NULL; |
466 | ||
726a989a | 467 | type = gimple_expr_type (last_stmt); |
0b2229b0 | 468 | |
726a989a | 469 | fn = gimple_call_fndecl (last_stmt); |
0b2229b0 RG |
470 | switch (DECL_FUNCTION_CODE (fn)) |
471 | { | |
472 | case BUILT_IN_POWIF: | |
473 | case BUILT_IN_POWI: | |
474 | case BUILT_IN_POWF: | |
475 | case BUILT_IN_POW: | |
726a989a RB |
476 | base = gimple_call_arg (last_stmt, 0); |
477 | exp = gimple_call_arg (last_stmt, 1); | |
0b2229b0 RG |
478 | if (TREE_CODE (exp) != REAL_CST |
479 | && TREE_CODE (exp) != INTEGER_CST) | |
726a989a | 480 | return NULL; |
0b2229b0 RG |
481 | break; |
482 | ||
726a989a RB |
483 | default: |
484 | return NULL; | |
0b2229b0 RG |
485 | } |
486 | ||
487 | /* We now have a pow or powi builtin function call with a constant | |
488 | exponent. */ | |
489 | ||
0b2229b0 RG |
490 | *type_out = NULL_TREE; |
491 | ||
492 | /* Catch squaring. */ | |
493 | if ((host_integerp (exp, 0) | |
494 | && tree_low_cst (exp, 0) == 2) | |
495 | || (TREE_CODE (exp) == REAL_CST | |
496 | && REAL_VALUES_EQUAL (TREE_REAL_CST (exp), dconst2))) | |
c6b1b49b RG |
497 | { |
498 | *type_in = TREE_TYPE (base); | |
726a989a RB |
499 | |
500 | var = vect_recog_temp_ssa_var (TREE_TYPE (base), NULL); | |
501 | stmt = gimple_build_assign_with_ops (MULT_EXPR, var, base, base); | |
502 | SSA_NAME_DEF_STMT (var) = stmt; | |
503 | return stmt; | |
c6b1b49b | 504 | } |
0b2229b0 RG |
505 | |
506 | /* Catch square root. */ | |
507 | if (TREE_CODE (exp) == REAL_CST | |
508 | && REAL_VALUES_EQUAL (TREE_REAL_CST (exp), dconsthalf)) | |
509 | { | |
510 | tree newfn = mathfn_built_in (TREE_TYPE (base), BUILT_IN_SQRT); | |
c6b1b49b RG |
511 | *type_in = get_vectype_for_scalar_type (TREE_TYPE (base)); |
512 | if (*type_in) | |
513 | { | |
726a989a RB |
514 | gimple stmt = gimple_build_call (newfn, 1, base); |
515 | if (vectorizable_function (stmt, *type_in, *type_in) | |
516 | != NULL_TREE) | |
517 | { | |
518 | var = vect_recog_temp_ssa_var (TREE_TYPE (base), stmt); | |
519 | gimple_call_set_lhs (stmt, var); | |
520 | return stmt; | |
521 | } | |
c6b1b49b | 522 | } |
0b2229b0 RG |
523 | } |
524 | ||
726a989a | 525 | return NULL; |
0b2229b0 RG |
526 | } |
527 | ||
528 | ||
20f06221 DN |
529 | /* Function vect_recog_widen_sum_pattern |
530 | ||
531 | Try to find the following pattern: | |
532 | ||
533 | type x_t; | |
534 | TYPE x_T, sum = init; | |
535 | loop: | |
536 | sum_0 = phi <init, sum_1> | |
537 | S1 x_t = *p; | |
538 | S2 x_T = (TYPE) x_t; | |
539 | S3 sum_1 = x_T + sum_0; | |
540 | ||
541 | where type 'TYPE' is at least double the size of type 'type', i.e - we're | |
542 | summing elements of type 'type' into an accumulator of type 'TYPE'. This is | |
917f1b7e | 543 | a special case of a reduction computation. |
20f06221 DN |
544 | |
545 | Input: | |
546 | ||
547 | * LAST_STMT: A stmt from which the pattern search begins. In the example, | |
548 | when this function is called with S3, the pattern {S2,S3} will be detected. | |
549 | ||
550 | Output: | |
551 | ||
552 | * TYPE_IN: The type of the input arguments to the pattern. | |
553 | ||
554 | * TYPE_OUT: The type of the output of this pattern. | |
555 | ||
556 | * Return value: A new stmt that will be used to replace the sequence of | |
557 | stmts that constitute the pattern. In this case it will be: | |
558 | WIDEN_SUM <x_t, sum_0> | |
d29de1bf | 559 | |
84fbffb2 | 560 | Note: The widening-sum idiom is a widening reduction pattern that is |
d29de1bf DN |
561 | vectorized without preserving all the intermediate results. It |
562 | produces only N/2 (widened) results (by summing up pairs of | |
563 | intermediate results) rather than all N results. Therefore, we | |
564 | cannot allow this pattern when we want to get all the results and in | |
565 | the correct order (as is the case when this computation is in an | |
566 | inner-loop nested in an outer-loop that us being vectorized). */ | |
20f06221 | 567 | |
726a989a RB |
568 | static gimple |
569 | vect_recog_widen_sum_pattern (gimple last_stmt, tree *type_in, tree *type_out) | |
20f06221 | 570 | { |
726a989a | 571 | gimple stmt; |
20f06221 DN |
572 | tree oprnd0, oprnd1; |
573 | stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt); | |
574 | tree type, half_type; | |
726a989a | 575 | gimple pattern_stmt; |
d29de1bf DN |
576 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_vinfo); |
577 | struct loop *loop = LOOP_VINFO_LOOP (loop_info); | |
726a989a | 578 | tree var; |
20f06221 | 579 | |
726a989a | 580 | if (!is_gimple_assign (last_stmt)) |
20f06221 DN |
581 | return NULL; |
582 | ||
726a989a | 583 | type = gimple_expr_type (last_stmt); |
20f06221 DN |
584 | |
585 | /* Look for the following pattern | |
586 | DX = (TYPE) X; | |
587 | sum_1 = DX + sum_0; | |
588 | In which DX is at least double the size of X, and sum_1 has been | |
589 | recognized as a reduction variable. | |
590 | */ | |
591 | ||
592 | /* Starting from LAST_STMT, follow the defs of its uses in search | |
593 | of the above pattern. */ | |
594 | ||
726a989a | 595 | if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR) |
20f06221 DN |
596 | return NULL; |
597 | ||
598 | if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def) | |
599 | return NULL; | |
600 | ||
726a989a RB |
601 | oprnd0 = gimple_assign_rhs1 (last_stmt); |
602 | oprnd1 = gimple_assign_rhs2 (last_stmt); | |
9600efe1 MM |
603 | if (!types_compatible_p (TREE_TYPE (oprnd0), type) |
604 | || !types_compatible_p (TREE_TYPE (oprnd1), type)) | |
20f06221 DN |
605 | return NULL; |
606 | ||
607 | /* So far so good. Since last_stmt was detected as a (summation) reduction, | |
608 | we know that oprnd1 is the reduction variable (defined by a loop-header | |
609 | phi), and oprnd0 is an ssa-name defined by a stmt in the loop body. | |
610 | Left to check that oprnd0 is defined by a cast from type 'type' to type | |
611 | 'TYPE'. */ | |
612 | ||
613 | if (!widened_name_p (oprnd0, last_stmt, &half_type, &stmt)) | |
614 | return NULL; | |
615 | ||
726a989a | 616 | oprnd0 = gimple_assign_rhs1 (stmt); |
20f06221 DN |
617 | *type_in = half_type; |
618 | *type_out = type; | |
619 | ||
620 | /* Pattern detected. Create a stmt to be used to replace the pattern: */ | |
726a989a RB |
621 | var = vect_recog_temp_ssa_var (type, NULL); |
622 | pattern_stmt = gimple_build_assign_with_ops (WIDEN_SUM_EXPR, var, | |
623 | oprnd0, oprnd1); | |
624 | SSA_NAME_DEF_STMT (var) = pattern_stmt; | |
625 | ||
20f06221 DN |
626 | if (vect_print_dump_info (REPORT_DETAILS)) |
627 | { | |
628 | fprintf (vect_dump, "vect_recog_widen_sum_pattern: detected: "); | |
726a989a | 629 | print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); |
20f06221 | 630 | } |
d29de1bf DN |
631 | |
632 | /* We don't allow changing the order of the computation in the inner-loop | |
633 | when doing outer-loop vectorization. */ | |
7c5222ff | 634 | gcc_assert (!nested_in_vect_loop_p (loop, last_stmt)); |
d29de1bf | 635 | |
726a989a | 636 | return pattern_stmt; |
20f06221 DN |
637 | } |
638 | ||
639 | ||
640 | /* Function vect_pattern_recog_1 | |
641 | ||
642 | Input: | |
643 | PATTERN_RECOG_FUNC: A pointer to a function that detects a certain | |
644 | computation pattern. | |
645 | STMT: A stmt from which the pattern search should start. | |
646 | ||
647 | If PATTERN_RECOG_FUNC successfully detected the pattern, it creates an | |
648 | expression that computes the same functionality and can be used to | |
649 | replace the sequence of stmts that are involved in the pattern. | |
650 | ||
651 | Output: | |
652 | This function checks if the expression returned by PATTERN_RECOG_FUNC is | |
653 | supported in vector form by the target. We use 'TYPE_IN' to obtain the | |
654 | relevant vector type. If 'TYPE_IN' is already a vector type, then this | |
655 | indicates that target support had already been checked by PATTERN_RECOG_FUNC. | |
656 | If 'TYPE_OUT' is also returned by PATTERN_RECOG_FUNC, we check that it fits | |
657 | to the available target pattern. | |
658 | ||
c0220ea4 | 659 | This function also does some bookkeeping, as explained in the documentation |
20f06221 DN |
660 | for vect_recog_pattern. */ |
661 | ||
662 | static void | |
663 | vect_pattern_recog_1 ( | |
726a989a RB |
664 | gimple (* vect_recog_func) (gimple, tree *, tree *), |
665 | gimple_stmt_iterator si) | |
20f06221 | 666 | { |
726a989a | 667 | gimple stmt = gsi_stmt (si), pattern_stmt; |
20f06221 DN |
668 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
669 | stmt_vec_info pattern_stmt_info; | |
670 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
20f06221 DN |
671 | tree pattern_vectype; |
672 | tree type_in, type_out; | |
20f06221 | 673 | enum tree_code code; |
20f06221 | 674 | |
726a989a RB |
675 | pattern_stmt = (* vect_recog_func) (stmt, &type_in, &type_out); |
676 | if (!pattern_stmt) | |
20f06221 DN |
677 | return; |
678 | ||
679 | if (VECTOR_MODE_P (TYPE_MODE (type_in))) | |
680 | { | |
681 | /* No need to check target support (already checked by the pattern | |
682 | recognition function). */ | |
683 | pattern_vectype = type_in; | |
684 | } | |
685 | else | |
686 | { | |
32e8bb8e | 687 | enum machine_mode vec_mode; |
20f06221 DN |
688 | enum insn_code icode; |
689 | optab optab; | |
690 | ||
691 | /* Check target support */ | |
692 | pattern_vectype = get_vectype_for_scalar_type (type_in); | |
03d3e953 IR |
693 | if (!pattern_vectype) |
694 | return; | |
695 | ||
726a989a RB |
696 | if (is_gimple_assign (pattern_stmt)) |
697 | code = gimple_assign_rhs_code (pattern_stmt); | |
698 | else | |
699 | { | |
700 | gcc_assert (is_gimple_call (pattern_stmt)); | |
701 | code = CALL_EXPR; | |
702 | } | |
703 | ||
704 | optab = optab_for_tree_code (code, pattern_vectype, optab_default); | |
20f06221 DN |
705 | vec_mode = TYPE_MODE (pattern_vectype); |
706 | if (!optab | |
166cdb08 | 707 | || (icode = optab_handler (optab, vec_mode)->insn_code) == |
20f06221 DN |
708 | CODE_FOR_nothing |
709 | || (type_out | |
6887a8b9 SE |
710 | && (!get_vectype_for_scalar_type (type_out) |
711 | || (insn_data[icode].operand[0].mode != | |
712 | TYPE_MODE (get_vectype_for_scalar_type (type_out)))))) | |
20f06221 DN |
713 | return; |
714 | } | |
715 | ||
716 | /* Found a vectorizable pattern. */ | |
717 | if (vect_print_dump_info (REPORT_DETAILS)) | |
718 | { | |
719 | fprintf (vect_dump, "pattern recognized: "); | |
726a989a | 720 | print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); |
20f06221 DN |
721 | } |
722 | ||
726a989a RB |
723 | /* Mark the stmts that are involved in the pattern. */ |
724 | gsi_insert_before (&si, pattern_stmt, GSI_SAME_STMT); | |
725 | set_vinfo_for_stmt (pattern_stmt, | |
a70d6342 | 726 | new_stmt_vec_info (pattern_stmt, loop_vinfo, NULL)); |
726a989a | 727 | pattern_stmt_info = vinfo_for_stmt (pattern_stmt); |
20f06221 DN |
728 | |
729 | STMT_VINFO_RELATED_STMT (pattern_stmt_info) = stmt; | |
730 | STMT_VINFO_DEF_TYPE (pattern_stmt_info) = STMT_VINFO_DEF_TYPE (stmt_info); | |
731 | STMT_VINFO_VECTYPE (pattern_stmt_info) = pattern_vectype; | |
732 | STMT_VINFO_IN_PATTERN_P (stmt_info) = true; | |
726a989a | 733 | STMT_VINFO_RELATED_STMT (stmt_info) = pattern_stmt; |
20f06221 DN |
734 | |
735 | return; | |
736 | } | |
737 | ||
738 | ||
739 | /* Function vect_pattern_recog | |
740 | ||
741 | Input: | |
742 | LOOP_VINFO - a struct_loop_info of a loop in which we want to look for | |
743 | computation idioms. | |
744 | ||
745 | Output - for each computation idiom that is detected we insert a new stmt | |
746 | that provides the same functionality and that can be vectorized. We | |
747 | also record some information in the struct_stmt_info of the relevant | |
748 | stmts, as explained below: | |
749 | ||
750 | At the entry to this function we have the following stmts, with the | |
751 | following initial value in the STMT_VINFO fields: | |
752 | ||
753 | stmt in_pattern_p related_stmt vec_stmt | |
754 | S1: a_i = .... - - - | |
755 | S2: a_2 = ..use(a_i).. - - - | |
756 | S3: a_1 = ..use(a_2).. - - - | |
757 | S4: a_0 = ..use(a_1).. - - - | |
758 | S5: ... = ..use(a_0).. - - - | |
759 | ||
760 | Say the sequence {S1,S2,S3,S4} was detected as a pattern that can be | |
761 | represented by a single stmt. We then: | |
762 | - create a new stmt S6 that will replace the pattern. | |
763 | - insert the new stmt S6 before the last stmt in the pattern | |
764 | - fill in the STMT_VINFO fields as follows: | |
765 | ||
766 | in_pattern_p related_stmt vec_stmt | |
767 | S1: a_i = .... - - - | |
768 | S2: a_2 = ..use(a_i).. - - - | |
769 | S3: a_1 = ..use(a_2).. - - - | |
770 | > S6: a_new = .... - S4 - | |
771 | S4: a_0 = ..use(a_1).. true S6 - | |
772 | S5: ... = ..use(a_0).. - - - | |
773 | ||
774 | (the last stmt in the pattern (S4) and the new pattern stmt (S6) point | |
775 | to each other through the RELATED_STMT field). | |
776 | ||
777 | S6 will be marked as relevant in vect_mark_stmts_to_be_vectorized instead | |
778 | of S4 because it will replace all its uses. Stmts {S1,S2,S3} will | |
779 | remain irrelevant unless used by stmts other than S4. | |
780 | ||
781 | If vectorization succeeds, vect_transform_stmt will skip over {S1,S2,S3} | |
c0220ea4 | 782 | (because they are marked as irrelevant). It will vectorize S6, and record |
20f06221 DN |
783 | a pointer to the new vector stmt VS6 both from S6 (as usual), and also |
784 | from S4. We do that so that when we get to vectorizing stmts that use the | |
785 | def of S4 (like S5 that uses a_0), we'll know where to take the relevant | |
786 | vector-def from. S4 will be skipped, and S5 will be vectorized as usual: | |
787 | ||
788 | in_pattern_p related_stmt vec_stmt | |
789 | S1: a_i = .... - - - | |
790 | S2: a_2 = ..use(a_i).. - - - | |
791 | S3: a_1 = ..use(a_2).. - - - | |
792 | > VS6: va_new = .... - - - | |
793 | S6: a_new = .... - S4 VS6 | |
794 | S4: a_0 = ..use(a_1).. true S6 VS6 | |
795 | > VS5: ... = ..vuse(va_new).. - - - | |
796 | S5: ... = ..use(a_0).. - - - | |
797 | ||
798 | DCE could then get rid of {S1,S2,S3,S4,S5,S6} (if their defs are not used | |
799 | elsewhere), and we'll end up with: | |
800 | ||
801 | VS6: va_new = .... | |
802 | VS5: ... = ..vuse(va_new).. | |
803 | ||
804 | If vectorization does not succeed, DCE will clean S6 away (its def is | |
805 | not used), and we'll end up with the original sequence. | |
806 | */ | |
807 | ||
808 | void | |
809 | vect_pattern_recog (loop_vec_info loop_vinfo) | |
810 | { | |
811 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
812 | basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
813 | unsigned int nbbs = loop->num_nodes; | |
726a989a RB |
814 | gimple_stmt_iterator si; |
815 | gimple stmt; | |
20f06221 | 816 | unsigned int i, j; |
726a989a | 817 | gimple (* vect_recog_func_ptr) (gimple, tree *, tree *); |
20f06221 DN |
818 | |
819 | if (vect_print_dump_info (REPORT_DETAILS)) | |
820 | fprintf (vect_dump, "=== vect_pattern_recog ==="); | |
821 | ||
822 | /* Scan through the loop stmts, applying the pattern recognition | |
823 | functions starting at each stmt visited: */ | |
824 | for (i = 0; i < nbbs; i++) | |
825 | { | |
826 | basic_block bb = bbs[i]; | |
726a989a | 827 | for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) |
20f06221 | 828 | { |
726a989a | 829 | stmt = gsi_stmt (si); |
20f06221 DN |
830 | |
831 | /* Scan over all generic vect_recog_xxx_pattern functions. */ | |
832 | for (j = 0; j < NUM_PATTERNS; j++) | |
833 | { | |
834 | vect_recog_func_ptr = vect_vect_recog_func_ptrs[j]; | |
835 | vect_pattern_recog_1 (vect_recog_func_ptr, si); | |
836 | } | |
837 | } | |
838 | } | |
839 | } |