Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-vect-patterns.c @ 88:f214c1d5b862
merge 89
author | Nobuyasu Oshiro <dimolto@cr.ie.u-ryukyu.ac.jp> |
---|---|
date | Tue, 20 Dec 2011 18:53:46 +0900 |
parents | f6334be47118 |
children | 04ced10e8804 |
rev | line source |
---|---|
0 | 1 /* Analysis Utilities for Loop Vectorization. |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2 Copyright (C) 2006, 2007, 2008, 2009, 2010 Free Software Foundation, Inc. |
0 | 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 | |
9 Software Foundation; either version 3, or (at your option) any later | |
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 | |
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 "tm.h" | |
25 #include "ggc.h" | |
26 #include "tree.h" | |
27 #include "target.h" | |
28 #include "basic-block.h" | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
29 #include "gimple-pretty-print.h" |
0 | 30 #include "tree-flow.h" |
31 #include "tree-dump.h" | |
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" | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
39 #include "diagnostic-core.h" |
0 | 40 |
41 /* Function prototypes */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
42 static void vect_pattern_recog_1 |
0 | 43 (gimple (* ) (gimple, tree *, tree *), gimple_stmt_iterator); |
44 static bool widened_name_p (tree, gimple, tree *, gimple *); | |
45 | |
46 /* Pattern recognition functions */ | |
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 *); | |
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, | |
54 vect_recog_dot_prod_pattern, | |
55 vect_recog_pow_pattern}; | |
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) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
63 where the type of name0 (HALF_TYPE) is smaller than the type of NAME. |
0 | 64 */ |
65 | |
66 static bool | |
67 widened_name_p (tree name, gimple use_stmt, tree *half_type, gimple *def_stmt) | |
68 { | |
69 tree dummy; | |
70 gimple dummy_gimple; | |
71 loop_vec_info loop_vinfo; | |
72 stmt_vec_info stmt_vinfo; | |
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 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
81 if (!vect_is_simple_use (name, loop_vinfo, NULL, def_stmt, &def, &dt)) |
0 | 82 return false; |
83 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
84 if (dt != vect_internal_def |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
85 && dt != vect_external_def && dt != vect_constant_def) |
0 | 86 return false; |
87 | |
88 if (! *def_stmt) | |
89 return false; | |
90 | |
91 if (!is_gimple_assign (*def_stmt)) | |
92 return false; | |
93 | |
94 if (gimple_assign_rhs_code (*def_stmt) != NOP_EXPR) | |
95 return false; | |
96 | |
97 oprnd0 = gimple_assign_rhs1 (*def_stmt); | |
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 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
105 if (!vect_is_simple_use (oprnd0, loop_vinfo, NULL, &dummy_gimple, &dummy, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
106 &dt)) |
0 | 107 return false; |
108 | |
109 return true; | |
110 } | |
111 | |
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 } | |
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 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
142 where 'TYPE1' is exactly double the size of type 'type', and 'TYPE2' is the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
143 same size of 'TYPE1' or bigger. This is a special case of a reduction |
0 | 144 computation. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
145 |
0 | 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> | |
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). */ | |
169 | |
170 static gimple | |
171 vect_recog_dot_prod_pattern (gimple last_stmt, tree *type_in, tree *type_out) | |
172 { | |
173 gimple stmt; | |
174 tree oprnd0, oprnd1; | |
175 tree oprnd00, oprnd01; | |
176 stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt); | |
177 tree type, half_type; | |
178 gimple pattern_stmt; | |
179 tree prod_type; | |
180 loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
181 struct loop *loop = LOOP_VINFO_LOOP (loop_info); | |
182 tree var, rhs; | |
183 | |
184 if (!is_gimple_assign (last_stmt)) | |
185 return NULL; | |
186 | |
187 type = gimple_expr_type (last_stmt); | |
188 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
189 /* Look for the following pattern |
0 | 190 DX = (TYPE1) X; |
191 DY = (TYPE1) Y; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
192 DPROD = DX * DY; |
0 | 193 DDPROD = (TYPE2) DPROD; |
194 sum_1 = DDPROD + sum_0; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
195 In which |
0 | 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 | |
213 if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR) | |
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); | |
221 type = gimple_expr_type (stmt); | |
222 if (gimple_assign_rhs_code (stmt) != WIDEN_SUM_EXPR) | |
223 return NULL; | |
224 oprnd0 = gimple_assign_rhs1 (stmt); | |
225 oprnd1 = gimple_assign_rhs2 (stmt); | |
226 half_type = TREE_TYPE (oprnd0); | |
227 } | |
228 else | |
229 { | |
230 gimple def_stmt; | |
231 | |
232 if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def) | |
233 return NULL; | |
234 oprnd0 = gimple_assign_rhs1 (last_stmt); | |
235 oprnd1 = gimple_assign_rhs2 (last_stmt); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
236 if (!types_compatible_p (TREE_TYPE (oprnd0), type) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
237 || !types_compatible_p (TREE_TYPE (oprnd1), type)) |
0 | 238 return NULL; |
239 stmt = last_stmt; | |
240 | |
241 if (widened_name_p (oprnd0, stmt, &half_type, &def_stmt)) | |
242 { | |
243 stmt = def_stmt; | |
244 oprnd0 = gimple_assign_rhs1 (stmt); | |
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); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
257 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
258 /* It could not be the dot_prod pattern if the stmt is outside the loop. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
259 if (!gimple_bb (stmt) || !flow_bb_inside_loop_p (loop, gimple_bb (stmt))) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
260 return NULL; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
261 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
262 /* FORNOW. Can continue analyzing the def-use chain when this stmt in a phi |
0 | 263 inside the loop (in case we are analyzing an outer-loop). */ |
264 if (!is_gimple_assign (stmt)) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
265 return NULL; |
0 | 266 stmt_vinfo = vinfo_for_stmt (stmt); |
267 gcc_assert (stmt_vinfo); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
268 if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_internal_def) |
0 | 269 return NULL; |
270 if (gimple_assign_rhs_code (stmt) != MULT_EXPR) | |
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); | |
277 if (gimple_assign_rhs_code (stmt) != WIDEN_MULT_EXPR) | |
278 return NULL; | |
279 stmt_vinfo = vinfo_for_stmt (stmt); | |
280 gcc_assert (stmt_vinfo); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
281 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_internal_def); |
0 | 282 oprnd00 = gimple_assign_rhs1 (stmt); |
283 oprnd01 = gimple_assign_rhs2 (stmt); | |
284 } | |
285 else | |
286 { | |
287 tree half_type0, half_type1; | |
288 gimple def_stmt; | |
289 tree oprnd0, oprnd1; | |
290 | |
291 oprnd0 = gimple_assign_rhs1 (stmt); | |
292 oprnd1 = gimple_assign_rhs2 (stmt); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
293 if (!types_compatible_p (TREE_TYPE (oprnd0), prod_type) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
294 || !types_compatible_p (TREE_TYPE (oprnd1), prod_type)) |
0 | 295 return NULL; |
296 if (!widened_name_p (oprnd0, stmt, &half_type0, &def_stmt)) | |
297 return NULL; | |
298 oprnd00 = gimple_assign_rhs1 (def_stmt); | |
299 if (!widened_name_p (oprnd1, stmt, &half_type1, &def_stmt)) | |
300 return NULL; | |
301 oprnd01 = gimple_assign_rhs1 (def_stmt); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
302 if (!types_compatible_p (half_type0, half_type1)) |
0 | 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; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
311 |
0 | 312 /* Pattern detected. Create a stmt to be used to replace the pattern: */ |
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); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
316 |
0 | 317 if (vect_print_dump_info (REPORT_DETAILS)) |
318 { | |
319 fprintf (vect_dump, "vect_recog_dot_prod_pattern: detected: "); | |
320 print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); | |
321 } | |
322 | |
323 /* We don't allow changing the order of the computation in the inner-loop | |
324 when doing outer-loop vectorization. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
325 gcc_assert (!nested_in_vect_loop_p (loop, last_stmt)); |
0 | 326 |
327 return pattern_stmt; | |
328 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
329 |
0 | 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 | |
361 static gimple | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
362 vect_recog_widen_mult_pattern (gimple last_stmt, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
363 tree *type_in, |
0 | 364 tree *type_out) |
365 { | |
366 gimple def_stmt0, def_stmt1; | |
367 tree oprnd0, oprnd1; | |
368 tree type, half_type0, half_type1; | |
369 gimple pattern_stmt; | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
370 tree vectype, vectype_out; |
0 | 371 tree dummy; |
372 tree var; | |
373 enum tree_code dummy_code; | |
374 int dummy_int; | |
375 VEC (tree, heap) *dummy_vec; | |
376 | |
377 if (!is_gimple_assign (last_stmt)) | |
378 return NULL; | |
379 | |
380 type = gimple_expr_type (last_stmt); | |
381 | |
382 /* Starting from LAST_STMT, follow the defs of its uses in search | |
383 of the above pattern. */ | |
384 | |
385 if (gimple_assign_rhs_code (last_stmt) != MULT_EXPR) | |
386 return NULL; | |
387 | |
388 oprnd0 = gimple_assign_rhs1 (last_stmt); | |
389 oprnd1 = gimple_assign_rhs2 (last_stmt); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
390 if (!types_compatible_p (TREE_TYPE (oprnd0), type) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
391 || !types_compatible_p (TREE_TYPE (oprnd1), type)) |
0 | 392 return NULL; |
393 | |
394 /* Check argument 0 */ | |
395 if (!widened_name_p (oprnd0, last_stmt, &half_type0, &def_stmt0)) | |
396 return NULL; | |
397 oprnd0 = gimple_assign_rhs1 (def_stmt0); | |
398 | |
399 /* Check argument 1 */ | |
400 if (!widened_name_p (oprnd1, last_stmt, &half_type1, &def_stmt1)) | |
401 return NULL; | |
402 oprnd1 = gimple_assign_rhs1 (def_stmt1); | |
403 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
404 if (!types_compatible_p (half_type0, half_type1)) |
0 | 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); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
413 vectype_out = get_vectype_for_scalar_type (type); |
0 | 414 if (!vectype |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
415 || !vectype_out |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
416 || !supportable_widening_operation (WIDEN_MULT_EXPR, last_stmt, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
417 vectype_out, vectype, |
0 | 418 &dummy, &dummy, &dummy_code, |
419 &dummy_code, &dummy_int, &dummy_vec)) | |
420 return NULL; | |
421 | |
422 *type_in = vectype; | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
423 *type_out = vectype_out; |
0 | 424 |
425 /* Pattern supported. Create a stmt to be used to replace the pattern: */ | |
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 | |
431 if (vect_print_dump_info (REPORT_DETAILS)) | |
432 print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); | |
433 | |
434 return pattern_stmt; | |
435 } | |
436 | |
437 | |
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: | |
459 x = x * x | |
460 or | |
461 x = sqrt (x) | |
462 */ | |
463 | |
464 static gimple | |
465 vect_recog_pow_pattern (gimple last_stmt, tree *type_in, tree *type_out) | |
466 { | |
467 tree fn, base, exp = NULL; | |
468 gimple stmt; | |
469 tree var; | |
470 | |
471 if (!is_gimple_call (last_stmt) || gimple_call_lhs (last_stmt) == NULL) | |
472 return NULL; | |
473 | |
474 fn = gimple_call_fndecl (last_stmt); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
475 if (fn == NULL_TREE || DECL_BUILT_IN_CLASS (fn) != BUILT_IN_NORMAL) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
476 return NULL; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
477 |
0 | 478 switch (DECL_FUNCTION_CODE (fn)) |
479 { | |
480 case BUILT_IN_POWIF: | |
481 case BUILT_IN_POWI: | |
482 case BUILT_IN_POWF: | |
483 case BUILT_IN_POW: | |
484 base = gimple_call_arg (last_stmt, 0); | |
485 exp = gimple_call_arg (last_stmt, 1); | |
486 if (TREE_CODE (exp) != REAL_CST | |
487 && TREE_CODE (exp) != INTEGER_CST) | |
488 return NULL; | |
489 break; | |
490 | |
491 default: | |
492 return NULL; | |
493 } | |
494 | |
495 /* We now have a pow or powi builtin function call with a constant | |
496 exponent. */ | |
497 | |
498 *type_out = NULL_TREE; | |
499 | |
500 /* Catch squaring. */ | |
501 if ((host_integerp (exp, 0) | |
502 && tree_low_cst (exp, 0) == 2) | |
503 || (TREE_CODE (exp) == REAL_CST | |
504 && REAL_VALUES_EQUAL (TREE_REAL_CST (exp), dconst2))) | |
505 { | |
506 *type_in = TREE_TYPE (base); | |
507 | |
508 var = vect_recog_temp_ssa_var (TREE_TYPE (base), NULL); | |
509 stmt = gimple_build_assign_with_ops (MULT_EXPR, var, base, base); | |
510 SSA_NAME_DEF_STMT (var) = stmt; | |
511 return stmt; | |
512 } | |
513 | |
514 /* Catch square root. */ | |
515 if (TREE_CODE (exp) == REAL_CST | |
516 && REAL_VALUES_EQUAL (TREE_REAL_CST (exp), dconsthalf)) | |
517 { | |
518 tree newfn = mathfn_built_in (TREE_TYPE (base), BUILT_IN_SQRT); | |
519 *type_in = get_vectype_for_scalar_type (TREE_TYPE (base)); | |
520 if (*type_in) | |
521 { | |
522 gimple stmt = gimple_build_call (newfn, 1, base); | |
523 if (vectorizable_function (stmt, *type_in, *type_in) | |
524 != NULL_TREE) | |
525 { | |
526 var = vect_recog_temp_ssa_var (TREE_TYPE (base), stmt); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
527 gimple_call_set_lhs (stmt, var); |
0 | 528 return stmt; |
529 } | |
530 } | |
531 } | |
532 | |
533 return NULL; | |
534 } | |
535 | |
536 | |
537 /* Function vect_recog_widen_sum_pattern | |
538 | |
539 Try to find the following pattern: | |
540 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
541 type x_t; |
0 | 542 TYPE x_T, sum = init; |
543 loop: | |
544 sum_0 = phi <init, sum_1> | |
545 S1 x_t = *p; | |
546 S2 x_T = (TYPE) x_t; | |
547 S3 sum_1 = x_T + sum_0; | |
548 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
549 where type 'TYPE' is at least double the size of type 'type', i.e - we're |
0 | 550 summing elements of type 'type' into an accumulator of type 'TYPE'. This is |
551 a special case of a reduction computation. | |
552 | |
553 Input: | |
554 | |
555 * LAST_STMT: A stmt from which the pattern search begins. In the example, | |
556 when this function is called with S3, the pattern {S2,S3} will be detected. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
557 |
0 | 558 Output: |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
559 |
0 | 560 * TYPE_IN: The type of the input arguments to the pattern. |
561 | |
562 * TYPE_OUT: The type of the output of this pattern. | |
563 | |
564 * Return value: A new stmt that will be used to replace the sequence of | |
565 stmts that constitute the pattern. In this case it will be: | |
566 WIDEN_SUM <x_t, sum_0> | |
567 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
568 Note: The widening-sum idiom is a widening reduction pattern that is |
0 | 569 vectorized without preserving all the intermediate results. It |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
570 produces only N/2 (widened) results (by summing up pairs of |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
571 intermediate results) rather than all N results. Therefore, we |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
572 cannot allow this pattern when we want to get all the results and in |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
573 the correct order (as is the case when this computation is in an |
0 | 574 inner-loop nested in an outer-loop that us being vectorized). */ |
575 | |
576 static gimple | |
577 vect_recog_widen_sum_pattern (gimple last_stmt, tree *type_in, tree *type_out) | |
578 { | |
579 gimple stmt; | |
580 tree oprnd0, oprnd1; | |
581 stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt); | |
582 tree type, half_type; | |
583 gimple pattern_stmt; | |
584 loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_vinfo); | |
585 struct loop *loop = LOOP_VINFO_LOOP (loop_info); | |
586 tree var; | |
587 | |
588 if (!is_gimple_assign (last_stmt)) | |
589 return NULL; | |
590 | |
591 type = gimple_expr_type (last_stmt); | |
592 | |
593 /* Look for the following pattern | |
594 DX = (TYPE) X; | |
595 sum_1 = DX + sum_0; | |
596 In which DX is at least double the size of X, and sum_1 has been | |
597 recognized as a reduction variable. | |
598 */ | |
599 | |
600 /* Starting from LAST_STMT, follow the defs of its uses in search | |
601 of the above pattern. */ | |
602 | |
603 if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR) | |
604 return NULL; | |
605 | |
606 if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def) | |
607 return NULL; | |
608 | |
609 oprnd0 = gimple_assign_rhs1 (last_stmt); | |
610 oprnd1 = gimple_assign_rhs2 (last_stmt); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
611 if (!types_compatible_p (TREE_TYPE (oprnd0), type) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
612 || !types_compatible_p (TREE_TYPE (oprnd1), type)) |
0 | 613 return NULL; |
614 | |
615 /* So far so good. Since last_stmt was detected as a (summation) reduction, | |
616 we know that oprnd1 is the reduction variable (defined by a loop-header | |
617 phi), and oprnd0 is an ssa-name defined by a stmt in the loop body. | |
618 Left to check that oprnd0 is defined by a cast from type 'type' to type | |
619 'TYPE'. */ | |
620 | |
621 if (!widened_name_p (oprnd0, last_stmt, &half_type, &stmt)) | |
622 return NULL; | |
623 | |
624 oprnd0 = gimple_assign_rhs1 (stmt); | |
625 *type_in = half_type; | |
626 *type_out = type; | |
627 | |
628 /* Pattern detected. Create a stmt to be used to replace the pattern: */ | |
629 var = vect_recog_temp_ssa_var (type, NULL); | |
630 pattern_stmt = gimple_build_assign_with_ops (WIDEN_SUM_EXPR, var, | |
631 oprnd0, oprnd1); | |
632 SSA_NAME_DEF_STMT (var) = pattern_stmt; | |
633 | |
634 if (vect_print_dump_info (REPORT_DETAILS)) | |
635 { | |
636 fprintf (vect_dump, "vect_recog_widen_sum_pattern: detected: "); | |
637 print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); | |
638 } | |
639 | |
640 /* We don't allow changing the order of the computation in the inner-loop | |
641 when doing outer-loop vectorization. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
642 gcc_assert (!nested_in_vect_loop_p (loop, last_stmt)); |
0 | 643 |
644 return pattern_stmt; | |
645 } | |
646 | |
647 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
648 /* Function vect_pattern_recog_1 |
0 | 649 |
650 Input: | |
651 PATTERN_RECOG_FUNC: A pointer to a function that detects a certain | |
652 computation pattern. | |
653 STMT: A stmt from which the pattern search should start. | |
654 | |
655 If PATTERN_RECOG_FUNC successfully detected the pattern, it creates an | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
656 expression that computes the same functionality and can be used to |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
657 replace the sequence of stmts that are involved in the pattern. |
0 | 658 |
659 Output: | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
660 This function checks if the expression returned by PATTERN_RECOG_FUNC is |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
661 supported in vector form by the target. We use 'TYPE_IN' to obtain the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
662 relevant vector type. If 'TYPE_IN' is already a vector type, then this |
0 | 663 indicates that target support had already been checked by PATTERN_RECOG_FUNC. |
664 If 'TYPE_OUT' is also returned by PATTERN_RECOG_FUNC, we check that it fits | |
665 to the available target pattern. | |
666 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
667 This function also does some bookkeeping, as explained in the documentation |
0 | 668 for vect_recog_pattern. */ |
669 | |
670 static void | |
671 vect_pattern_recog_1 ( | |
672 gimple (* vect_recog_func) (gimple, tree *, tree *), | |
673 gimple_stmt_iterator si) | |
674 { | |
675 gimple stmt = gsi_stmt (si), pattern_stmt; | |
676 stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
677 stmt_vec_info pattern_stmt_info; | |
678 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
679 tree pattern_vectype; | |
680 tree type_in, type_out; | |
681 enum tree_code code; | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
682 int i; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
683 gimple next; |
0 | 684 |
685 pattern_stmt = (* vect_recog_func) (stmt, &type_in, &type_out); | |
686 if (!pattern_stmt) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
687 return; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
688 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
689 if (VECTOR_MODE_P (TYPE_MODE (type_in))) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
690 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
691 /* No need to check target support (already checked by the pattern |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
692 recognition function). */ |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
693 if (type_out) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
694 gcc_assert (VECTOR_MODE_P (TYPE_MODE (type_out))); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
695 pattern_vectype = type_out ? type_out : type_in; |
0 | 696 } |
697 else | |
698 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
699 enum machine_mode vec_mode; |
0 | 700 enum insn_code icode; |
701 optab optab; | |
702 | |
703 /* Check target support */ | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
704 type_in = get_vectype_for_scalar_type (type_in); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
705 if (!type_in) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
706 return; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
707 if (type_out) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
708 type_out = get_vectype_for_scalar_type (type_out); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
709 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
710 type_out = type_in; |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
711 if (!type_out) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
712 return; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
713 pattern_vectype = type_out; |
0 | 714 |
715 if (is_gimple_assign (pattern_stmt)) | |
716 code = gimple_assign_rhs_code (pattern_stmt); | |
717 else | |
718 { | |
719 gcc_assert (is_gimple_call (pattern_stmt)); | |
720 code = CALL_EXPR; | |
721 } | |
722 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
723 optab = optab_for_tree_code (code, type_in, optab_default); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
724 vec_mode = TYPE_MODE (type_in); |
0 | 725 if (!optab |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
726 || (icode = optab_handler (optab, vec_mode)) == CODE_FOR_nothing |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
727 || (insn_data[icode].operand[0].mode != TYPE_MODE (type_out))) |
0 | 728 return; |
729 } | |
730 | |
731 /* Found a vectorizable pattern. */ | |
732 if (vect_print_dump_info (REPORT_DETAILS)) | |
733 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
734 fprintf (vect_dump, "pattern recognized: "); |
0 | 735 print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM); |
736 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
737 |
0 | 738 /* Mark the stmts that are involved in the pattern. */ |
739 gsi_insert_before (&si, pattern_stmt, GSI_SAME_STMT); | |
740 set_vinfo_for_stmt (pattern_stmt, | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
741 new_stmt_vec_info (pattern_stmt, loop_vinfo, NULL)); |
0 | 742 pattern_stmt_info = vinfo_for_stmt (pattern_stmt); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
743 |
0 | 744 STMT_VINFO_RELATED_STMT (pattern_stmt_info) = stmt; |
745 STMT_VINFO_DEF_TYPE (pattern_stmt_info) = STMT_VINFO_DEF_TYPE (stmt_info); | |
746 STMT_VINFO_VECTYPE (pattern_stmt_info) = pattern_vectype; | |
747 STMT_VINFO_IN_PATTERN_P (stmt_info) = true; | |
748 STMT_VINFO_RELATED_STMT (stmt_info) = pattern_stmt; | |
749 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
750 /* Patterns cannot be vectorized using SLP, because they change the order of |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
751 computation. */ |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
752 FOR_EACH_VEC_ELT (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo), i, next) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
753 if (next == stmt) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
754 VEC_ordered_remove (gimple, LOOP_VINFO_REDUCTIONS (loop_vinfo), i); |
0 | 755 } |
756 | |
757 | |
758 /* Function vect_pattern_recog | |
759 | |
760 Input: | |
761 LOOP_VINFO - a struct_loop_info of a loop in which we want to look for | |
762 computation idioms. | |
763 | |
764 Output - for each computation idiom that is detected we insert a new stmt | |
765 that provides the same functionality and that can be vectorized. We | |
766 also record some information in the struct_stmt_info of the relevant | |
767 stmts, as explained below: | |
768 | |
769 At the entry to this function we have the following stmts, with the | |
770 following initial value in the STMT_VINFO fields: | |
771 | |
772 stmt in_pattern_p related_stmt vec_stmt | |
773 S1: a_i = .... - - - | |
774 S2: a_2 = ..use(a_i).. - - - | |
775 S3: a_1 = ..use(a_2).. - - - | |
776 S4: a_0 = ..use(a_1).. - - - | |
777 S5: ... = ..use(a_0).. - - - | |
778 | |
779 Say the sequence {S1,S2,S3,S4} was detected as a pattern that can be | |
780 represented by a single stmt. We then: | |
781 - create a new stmt S6 that will replace the pattern. | |
782 - insert the new stmt S6 before the last stmt in the pattern | |
783 - fill in the STMT_VINFO fields as follows: | |
784 | |
785 in_pattern_p related_stmt vec_stmt | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
786 S1: a_i = .... - - - |
0 | 787 S2: a_2 = ..use(a_i).. - - - |
788 S3: a_1 = ..use(a_2).. - - - | |
789 > S6: a_new = .... - S4 - | |
790 S4: a_0 = ..use(a_1).. true S6 - | |
791 S5: ... = ..use(a_0).. - - - | |
792 | |
793 (the last stmt in the pattern (S4) and the new pattern stmt (S6) point | |
794 to each other through the RELATED_STMT field). | |
795 | |
796 S6 will be marked as relevant in vect_mark_stmts_to_be_vectorized instead | |
797 of S4 because it will replace all its uses. Stmts {S1,S2,S3} will | |
798 remain irrelevant unless used by stmts other than S4. | |
799 | |
800 If vectorization succeeds, vect_transform_stmt will skip over {S1,S2,S3} | |
801 (because they are marked as irrelevant). It will vectorize S6, and record | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
802 a pointer to the new vector stmt VS6 both from S6 (as usual), and also |
0 | 803 from S4. We do that so that when we get to vectorizing stmts that use the |
804 def of S4 (like S5 that uses a_0), we'll know where to take the relevant | |
805 vector-def from. S4 will be skipped, and S5 will be vectorized as usual: | |
806 | |
807 in_pattern_p related_stmt vec_stmt | |
808 S1: a_i = .... - - - | |
809 S2: a_2 = ..use(a_i).. - - - | |
810 S3: a_1 = ..use(a_2).. - - - | |
811 > VS6: va_new = .... - - - | |
812 S6: a_new = .... - S4 VS6 | |
813 S4: a_0 = ..use(a_1).. true S6 VS6 | |
814 > VS5: ... = ..vuse(va_new).. - - - | |
815 S5: ... = ..use(a_0).. - - - | |
816 | |
817 DCE could then get rid of {S1,S2,S3,S4,S5,S6} (if their defs are not used | |
818 elsewhere), and we'll end up with: | |
819 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
820 VS6: va_new = .... |
0 | 821 VS5: ... = ..vuse(va_new).. |
822 | |
823 If vectorization does not succeed, DCE will clean S6 away (its def is | |
824 not used), and we'll end up with the original sequence. | |
825 */ | |
826 | |
827 void | |
828 vect_pattern_recog (loop_vec_info loop_vinfo) | |
829 { | |
830 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
831 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); | |
832 unsigned int nbbs = loop->num_nodes; | |
833 gimple_stmt_iterator si; | |
834 unsigned int i, j; | |
835 gimple (* vect_recog_func_ptr) (gimple, tree *, tree *); | |
836 | |
837 if (vect_print_dump_info (REPORT_DETAILS)) | |
838 fprintf (vect_dump, "=== vect_pattern_recog ==="); | |
839 | |
840 /* Scan through the loop stmts, applying the pattern recognition | |
841 functions starting at each stmt visited: */ | |
842 for (i = 0; i < nbbs; i++) | |
843 { | |
844 basic_block bb = bbs[i]; | |
845 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) | |
846 { | |
847 /* Scan over all generic vect_recog_xxx_pattern functions. */ | |
848 for (j = 0; j < NUM_PATTERNS; j++) | |
849 { | |
850 vect_recog_func_ptr = vect_vect_recog_func_ptrs[j]; | |
851 vect_pattern_recog_1 (vect_recog_func_ptr, si); | |
852 } | |
853 } | |
854 } | |
855 } |