1 |
280 |
jeremybenn |
/* Analysis Utilities for Loop Vectorization.
|
2 |
|
|
Copyright (C) 2006, 2007, 2008, 2009 Free Software Foundation, Inc.
|
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"
|
29 |
|
|
#include "diagnostic.h"
|
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"
|
39 |
|
|
#include "toplev.h"
|
40 |
|
|
|
41 |
|
|
/* Function prototypes */
|
42 |
|
|
static void vect_pattern_recog_1
|
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)
|
63 |
|
|
where the type of name0 (HALF_TYPE) is smaller than the type of NAME.
|
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 |
|
|
|
81 |
|
|
if (!vect_is_simple_use (name, loop_vinfo, NULL, def_stmt, &def, &dt))
|
82 |
|
|
return false;
|
83 |
|
|
|
84 |
|
|
if (dt != vect_internal_def
|
85 |
|
|
&& dt != vect_external_def && dt != vect_constant_def)
|
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 |
|
|
|
105 |
|
|
if (!vect_is_simple_use (oprnd0, loop_vinfo, NULL, &dummy_gimple, &dummy,
|
106 |
|
|
&dt))
|
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 |
|
|
|
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
|
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>
|
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 |
|
|
|
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 |
|
|
|
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);
|
236 |
|
|
if (!types_compatible_p (TREE_TYPE (oprnd0), type)
|
237 |
|
|
|| !types_compatible_p (TREE_TYPE (oprnd1), type))
|
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);
|
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). */
|
259 |
|
|
if (!is_gimple_assign (stmt))
|
260 |
|
|
return NULL;
|
261 |
|
|
stmt_vinfo = vinfo_for_stmt (stmt);
|
262 |
|
|
gcc_assert (stmt_vinfo);
|
263 |
|
|
if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_internal_def)
|
264 |
|
|
return NULL;
|
265 |
|
|
if (gimple_assign_rhs_code (stmt) != MULT_EXPR)
|
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);
|
272 |
|
|
if (gimple_assign_rhs_code (stmt) != WIDEN_MULT_EXPR)
|
273 |
|
|
return NULL;
|
274 |
|
|
stmt_vinfo = vinfo_for_stmt (stmt);
|
275 |
|
|
gcc_assert (stmt_vinfo);
|
276 |
|
|
gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_internal_def);
|
277 |
|
|
oprnd00 = gimple_assign_rhs1 (stmt);
|
278 |
|
|
oprnd01 = gimple_assign_rhs2 (stmt);
|
279 |
|
|
}
|
280 |
|
|
else
|
281 |
|
|
{
|
282 |
|
|
tree half_type0, half_type1;
|
283 |
|
|
gimple def_stmt;
|
284 |
|
|
tree oprnd0, oprnd1;
|
285 |
|
|
|
286 |
|
|
oprnd0 = gimple_assign_rhs1 (stmt);
|
287 |
|
|
oprnd1 = gimple_assign_rhs2 (stmt);
|
288 |
|
|
if (!types_compatible_p (TREE_TYPE (oprnd0), prod_type)
|
289 |
|
|
|| !types_compatible_p (TREE_TYPE (oprnd1), prod_type))
|
290 |
|
|
return NULL;
|
291 |
|
|
if (!widened_name_p (oprnd0, stmt, &half_type0, &def_stmt))
|
292 |
|
|
return NULL;
|
293 |
|
|
oprnd00 = gimple_assign_rhs1 (def_stmt);
|
294 |
|
|
if (!widened_name_p (oprnd1, stmt, &half_type1, &def_stmt))
|
295 |
|
|
return NULL;
|
296 |
|
|
oprnd01 = gimple_assign_rhs1 (def_stmt);
|
297 |
|
|
if (!types_compatible_p (half_type0, half_type1))
|
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: */
|
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 |
|
|
|
312 |
|
|
if (vect_print_dump_info (REPORT_DETAILS))
|
313 |
|
|
{
|
314 |
|
|
fprintf (vect_dump, "vect_recog_dot_prod_pattern: detected: ");
|
315 |
|
|
print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM);
|
316 |
|
|
}
|
317 |
|
|
|
318 |
|
|
/* We don't allow changing the order of the computation in the inner-loop
|
319 |
|
|
when doing outer-loop vectorization. */
|
320 |
|
|
gcc_assert (!nested_in_vect_loop_p (loop, last_stmt));
|
321 |
|
|
|
322 |
|
|
return pattern_stmt;
|
323 |
|
|
}
|
324 |
|
|
|
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 |
|
|
|
356 |
|
|
static gimple
|
357 |
|
|
vect_recog_widen_mult_pattern (gimple last_stmt,
|
358 |
|
|
tree *type_in,
|
359 |
|
|
tree *type_out)
|
360 |
|
|
{
|
361 |
|
|
gimple def_stmt0, def_stmt1;
|
362 |
|
|
tree oprnd0, oprnd1;
|
363 |
|
|
tree type, half_type0, half_type1;
|
364 |
|
|
gimple pattern_stmt;
|
365 |
|
|
tree vectype;
|
366 |
|
|
tree dummy;
|
367 |
|
|
tree var;
|
368 |
|
|
enum tree_code dummy_code;
|
369 |
|
|
int dummy_int;
|
370 |
|
|
VEC (tree, heap) *dummy_vec;
|
371 |
|
|
|
372 |
|
|
if (!is_gimple_assign (last_stmt))
|
373 |
|
|
return NULL;
|
374 |
|
|
|
375 |
|
|
type = gimple_expr_type (last_stmt);
|
376 |
|
|
|
377 |
|
|
/* Starting from LAST_STMT, follow the defs of its uses in search
|
378 |
|
|
of the above pattern. */
|
379 |
|
|
|
380 |
|
|
if (gimple_assign_rhs_code (last_stmt) != MULT_EXPR)
|
381 |
|
|
return NULL;
|
382 |
|
|
|
383 |
|
|
oprnd0 = gimple_assign_rhs1 (last_stmt);
|
384 |
|
|
oprnd1 = gimple_assign_rhs2 (last_stmt);
|
385 |
|
|
if (!types_compatible_p (TREE_TYPE (oprnd0), type)
|
386 |
|
|
|| !types_compatible_p (TREE_TYPE (oprnd1), type))
|
387 |
|
|
return NULL;
|
388 |
|
|
|
389 |
|
|
/* Check argument 0 */
|
390 |
|
|
if (!widened_name_p (oprnd0, last_stmt, &half_type0, &def_stmt0))
|
391 |
|
|
return NULL;
|
392 |
|
|
oprnd0 = gimple_assign_rhs1 (def_stmt0);
|
393 |
|
|
|
394 |
|
|
/* Check argument 1 */
|
395 |
|
|
if (!widened_name_p (oprnd1, last_stmt, &half_type1, &def_stmt1))
|
396 |
|
|
return NULL;
|
397 |
|
|
oprnd1 = gimple_assign_rhs1 (def_stmt1);
|
398 |
|
|
|
399 |
|
|
if (!types_compatible_p (half_type0, half_type1))
|
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);
|
408 |
|
|
if (!vectype
|
409 |
|
|
|| !supportable_widening_operation (WIDEN_MULT_EXPR, last_stmt, vectype,
|
410 |
|
|
&dummy, &dummy, &dummy_code,
|
411 |
|
|
&dummy_code, &dummy_int, &dummy_vec))
|
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: */
|
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 |
|
|
|
423 |
|
|
if (vect_print_dump_info (REPORT_DETAILS))
|
424 |
|
|
print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM);
|
425 |
|
|
|
426 |
|
|
return pattern_stmt;
|
427 |
|
|
}
|
428 |
|
|
|
429 |
|
|
|
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:
|
451 |
|
|
x = x * x
|
452 |
|
|
or
|
453 |
|
|
x = sqrt (x)
|
454 |
|
|
*/
|
455 |
|
|
|
456 |
|
|
static gimple
|
457 |
|
|
vect_recog_pow_pattern (gimple last_stmt, tree *type_in, tree *type_out)
|
458 |
|
|
{
|
459 |
|
|
tree fn, base, exp = NULL;
|
460 |
|
|
gimple stmt;
|
461 |
|
|
tree var;
|
462 |
|
|
|
463 |
|
|
if (!is_gimple_call (last_stmt) || gimple_call_lhs (last_stmt) == NULL)
|
464 |
|
|
return NULL;
|
465 |
|
|
|
466 |
|
|
fn = gimple_call_fndecl (last_stmt);
|
467 |
|
|
switch (DECL_FUNCTION_CODE (fn))
|
468 |
|
|
{
|
469 |
|
|
case BUILT_IN_POWIF:
|
470 |
|
|
case BUILT_IN_POWI:
|
471 |
|
|
case BUILT_IN_POWF:
|
472 |
|
|
case BUILT_IN_POW:
|
473 |
|
|
base = gimple_call_arg (last_stmt, 0);
|
474 |
|
|
exp = gimple_call_arg (last_stmt, 1);
|
475 |
|
|
if (TREE_CODE (exp) != REAL_CST
|
476 |
|
|
&& TREE_CODE (exp) != INTEGER_CST)
|
477 |
|
|
return NULL;
|
478 |
|
|
break;
|
479 |
|
|
|
480 |
|
|
default:
|
481 |
|
|
return NULL;
|
482 |
|
|
}
|
483 |
|
|
|
484 |
|
|
/* We now have a pow or powi builtin function call with a constant
|
485 |
|
|
exponent. */
|
486 |
|
|
|
487 |
|
|
*type_out = NULL_TREE;
|
488 |
|
|
|
489 |
|
|
/* Catch squaring. */
|
490 |
|
|
if ((host_integerp (exp, 0)
|
491 |
|
|
&& tree_low_cst (exp, 0) == 2)
|
492 |
|
|
|| (TREE_CODE (exp) == REAL_CST
|
493 |
|
|
&& REAL_VALUES_EQUAL (TREE_REAL_CST (exp), dconst2)))
|
494 |
|
|
{
|
495 |
|
|
*type_in = TREE_TYPE (base);
|
496 |
|
|
|
497 |
|
|
var = vect_recog_temp_ssa_var (TREE_TYPE (base), NULL);
|
498 |
|
|
stmt = gimple_build_assign_with_ops (MULT_EXPR, var, base, base);
|
499 |
|
|
SSA_NAME_DEF_STMT (var) = stmt;
|
500 |
|
|
return stmt;
|
501 |
|
|
}
|
502 |
|
|
|
503 |
|
|
/* Catch square root. */
|
504 |
|
|
if (TREE_CODE (exp) == REAL_CST
|
505 |
|
|
&& REAL_VALUES_EQUAL (TREE_REAL_CST (exp), dconsthalf))
|
506 |
|
|
{
|
507 |
|
|
tree newfn = mathfn_built_in (TREE_TYPE (base), BUILT_IN_SQRT);
|
508 |
|
|
*type_in = get_vectype_for_scalar_type (TREE_TYPE (base));
|
509 |
|
|
if (*type_in)
|
510 |
|
|
{
|
511 |
|
|
gimple stmt = gimple_build_call (newfn, 1, base);
|
512 |
|
|
if (vectorizable_function (stmt, *type_in, *type_in)
|
513 |
|
|
!= NULL_TREE)
|
514 |
|
|
{
|
515 |
|
|
var = vect_recog_temp_ssa_var (TREE_TYPE (base), stmt);
|
516 |
|
|
gimple_call_set_lhs (stmt, var);
|
517 |
|
|
return stmt;
|
518 |
|
|
}
|
519 |
|
|
}
|
520 |
|
|
}
|
521 |
|
|
|
522 |
|
|
return NULL;
|
523 |
|
|
}
|
524 |
|
|
|
525 |
|
|
|
526 |
|
|
/* Function vect_recog_widen_sum_pattern
|
527 |
|
|
|
528 |
|
|
Try to find the following pattern:
|
529 |
|
|
|
530 |
|
|
type x_t;
|
531 |
|
|
TYPE x_T, sum = init;
|
532 |
|
|
loop:
|
533 |
|
|
sum_0 = phi <init, sum_1>
|
534 |
|
|
S1 x_t = *p;
|
535 |
|
|
S2 x_T = (TYPE) x_t;
|
536 |
|
|
S3 sum_1 = x_T + sum_0;
|
537 |
|
|
|
538 |
|
|
where type 'TYPE' is at least double the size of type 'type', i.e - we're
|
539 |
|
|
summing elements of type 'type' into an accumulator of type 'TYPE'. This is
|
540 |
|
|
a special case of a reduction computation.
|
541 |
|
|
|
542 |
|
|
Input:
|
543 |
|
|
|
544 |
|
|
* LAST_STMT: A stmt from which the pattern search begins. In the example,
|
545 |
|
|
when this function is called with S3, the pattern {S2,S3} will be detected.
|
546 |
|
|
|
547 |
|
|
Output:
|
548 |
|
|
|
549 |
|
|
* TYPE_IN: The type of the input arguments to the pattern.
|
550 |
|
|
|
551 |
|
|
* TYPE_OUT: The type of the output of this pattern.
|
552 |
|
|
|
553 |
|
|
* Return value: A new stmt that will be used to replace the sequence of
|
554 |
|
|
stmts that constitute the pattern. In this case it will be:
|
555 |
|
|
WIDEN_SUM <x_t, sum_0>
|
556 |
|
|
|
557 |
|
|
Note: The widening-sum idiom is a widening reduction pattern that is
|
558 |
|
|
vectorized without preserving all the intermediate results. It
|
559 |
|
|
produces only N/2 (widened) results (by summing up pairs of
|
560 |
|
|
intermediate results) rather than all N results. Therefore, we
|
561 |
|
|
cannot allow this pattern when we want to get all the results and in
|
562 |
|
|
the correct order (as is the case when this computation is in an
|
563 |
|
|
inner-loop nested in an outer-loop that us being vectorized). */
|
564 |
|
|
|
565 |
|
|
static gimple
|
566 |
|
|
vect_recog_widen_sum_pattern (gimple last_stmt, tree *type_in, tree *type_out)
|
567 |
|
|
{
|
568 |
|
|
gimple stmt;
|
569 |
|
|
tree oprnd0, oprnd1;
|
570 |
|
|
stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt);
|
571 |
|
|
tree type, half_type;
|
572 |
|
|
gimple pattern_stmt;
|
573 |
|
|
loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
|
574 |
|
|
struct loop *loop = LOOP_VINFO_LOOP (loop_info);
|
575 |
|
|
tree var;
|
576 |
|
|
|
577 |
|
|
if (!is_gimple_assign (last_stmt))
|
578 |
|
|
return NULL;
|
579 |
|
|
|
580 |
|
|
type = gimple_expr_type (last_stmt);
|
581 |
|
|
|
582 |
|
|
/* Look for the following pattern
|
583 |
|
|
DX = (TYPE) X;
|
584 |
|
|
sum_1 = DX + sum_0;
|
585 |
|
|
In which DX is at least double the size of X, and sum_1 has been
|
586 |
|
|
recognized as a reduction variable.
|
587 |
|
|
*/
|
588 |
|
|
|
589 |
|
|
/* Starting from LAST_STMT, follow the defs of its uses in search
|
590 |
|
|
of the above pattern. */
|
591 |
|
|
|
592 |
|
|
if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR)
|
593 |
|
|
return NULL;
|
594 |
|
|
|
595 |
|
|
if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def)
|
596 |
|
|
return NULL;
|
597 |
|
|
|
598 |
|
|
oprnd0 = gimple_assign_rhs1 (last_stmt);
|
599 |
|
|
oprnd1 = gimple_assign_rhs2 (last_stmt);
|
600 |
|
|
if (!types_compatible_p (TREE_TYPE (oprnd0), type)
|
601 |
|
|
|| !types_compatible_p (TREE_TYPE (oprnd1), type))
|
602 |
|
|
return NULL;
|
603 |
|
|
|
604 |
|
|
/* So far so good. Since last_stmt was detected as a (summation) reduction,
|
605 |
|
|
we know that oprnd1 is the reduction variable (defined by a loop-header
|
606 |
|
|
phi), and oprnd0 is an ssa-name defined by a stmt in the loop body.
|
607 |
|
|
Left to check that oprnd0 is defined by a cast from type 'type' to type
|
608 |
|
|
'TYPE'. */
|
609 |
|
|
|
610 |
|
|
if (!widened_name_p (oprnd0, last_stmt, &half_type, &stmt))
|
611 |
|
|
return NULL;
|
612 |
|
|
|
613 |
|
|
oprnd0 = gimple_assign_rhs1 (stmt);
|
614 |
|
|
*type_in = half_type;
|
615 |
|
|
*type_out = type;
|
616 |
|
|
|
617 |
|
|
/* Pattern detected. Create a stmt to be used to replace the pattern: */
|
618 |
|
|
var = vect_recog_temp_ssa_var (type, NULL);
|
619 |
|
|
pattern_stmt = gimple_build_assign_with_ops (WIDEN_SUM_EXPR, var,
|
620 |
|
|
oprnd0, oprnd1);
|
621 |
|
|
SSA_NAME_DEF_STMT (var) = pattern_stmt;
|
622 |
|
|
|
623 |
|
|
if (vect_print_dump_info (REPORT_DETAILS))
|
624 |
|
|
{
|
625 |
|
|
fprintf (vect_dump, "vect_recog_widen_sum_pattern: detected: ");
|
626 |
|
|
print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM);
|
627 |
|
|
}
|
628 |
|
|
|
629 |
|
|
/* We don't allow changing the order of the computation in the inner-loop
|
630 |
|
|
when doing outer-loop vectorization. */
|
631 |
|
|
gcc_assert (!nested_in_vect_loop_p (loop, last_stmt));
|
632 |
|
|
|
633 |
|
|
return pattern_stmt;
|
634 |
|
|
}
|
635 |
|
|
|
636 |
|
|
|
637 |
|
|
/* Function vect_pattern_recog_1
|
638 |
|
|
|
639 |
|
|
Input:
|
640 |
|
|
PATTERN_RECOG_FUNC: A pointer to a function that detects a certain
|
641 |
|
|
computation pattern.
|
642 |
|
|
STMT: A stmt from which the pattern search should start.
|
643 |
|
|
|
644 |
|
|
If PATTERN_RECOG_FUNC successfully detected the pattern, it creates an
|
645 |
|
|
expression that computes the same functionality and can be used to
|
646 |
|
|
replace the sequence of stmts that are involved in the pattern.
|
647 |
|
|
|
648 |
|
|
Output:
|
649 |
|
|
This function checks if the expression returned by PATTERN_RECOG_FUNC is
|
650 |
|
|
supported in vector form by the target. We use 'TYPE_IN' to obtain the
|
651 |
|
|
relevant vector type. If 'TYPE_IN' is already a vector type, then this
|
652 |
|
|
indicates that target support had already been checked by PATTERN_RECOG_FUNC.
|
653 |
|
|
If 'TYPE_OUT' is also returned by PATTERN_RECOG_FUNC, we check that it fits
|
654 |
|
|
to the available target pattern.
|
655 |
|
|
|
656 |
|
|
This function also does some bookkeeping, as explained in the documentation
|
657 |
|
|
for vect_recog_pattern. */
|
658 |
|
|
|
659 |
|
|
static void
|
660 |
|
|
vect_pattern_recog_1 (
|
661 |
|
|
gimple (* vect_recog_func) (gimple, tree *, tree *),
|
662 |
|
|
gimple_stmt_iterator si)
|
663 |
|
|
{
|
664 |
|
|
gimple stmt = gsi_stmt (si), pattern_stmt;
|
665 |
|
|
stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
|
666 |
|
|
stmt_vec_info pattern_stmt_info;
|
667 |
|
|
loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
|
668 |
|
|
tree pattern_vectype;
|
669 |
|
|
tree type_in, type_out;
|
670 |
|
|
enum tree_code code;
|
671 |
|
|
|
672 |
|
|
pattern_stmt = (* vect_recog_func) (stmt, &type_in, &type_out);
|
673 |
|
|
if (!pattern_stmt)
|
674 |
|
|
return;
|
675 |
|
|
|
676 |
|
|
if (VECTOR_MODE_P (TYPE_MODE (type_in)))
|
677 |
|
|
{
|
678 |
|
|
/* No need to check target support (already checked by the pattern
|
679 |
|
|
recognition function). */
|
680 |
|
|
pattern_vectype = type_in;
|
681 |
|
|
}
|
682 |
|
|
else
|
683 |
|
|
{
|
684 |
|
|
enum machine_mode vec_mode;
|
685 |
|
|
enum insn_code icode;
|
686 |
|
|
optab optab;
|
687 |
|
|
|
688 |
|
|
/* Check target support */
|
689 |
|
|
pattern_vectype = get_vectype_for_scalar_type (type_in);
|
690 |
|
|
if (!pattern_vectype)
|
691 |
|
|
return;
|
692 |
|
|
|
693 |
|
|
if (is_gimple_assign (pattern_stmt))
|
694 |
|
|
code = gimple_assign_rhs_code (pattern_stmt);
|
695 |
|
|
else
|
696 |
|
|
{
|
697 |
|
|
gcc_assert (is_gimple_call (pattern_stmt));
|
698 |
|
|
code = CALL_EXPR;
|
699 |
|
|
}
|
700 |
|
|
|
701 |
|
|
optab = optab_for_tree_code (code, pattern_vectype, optab_default);
|
702 |
|
|
vec_mode = TYPE_MODE (pattern_vectype);
|
703 |
|
|
if (!optab
|
704 |
|
|
|| (icode = optab_handler (optab, vec_mode)->insn_code) ==
|
705 |
|
|
CODE_FOR_nothing
|
706 |
|
|
|| (type_out
|
707 |
|
|
&& (!get_vectype_for_scalar_type (type_out)
|
708 |
|
|
|| (insn_data[icode].operand[0].mode !=
|
709 |
|
|
TYPE_MODE (get_vectype_for_scalar_type (type_out))))))
|
710 |
|
|
return;
|
711 |
|
|
}
|
712 |
|
|
|
713 |
|
|
/* Found a vectorizable pattern. */
|
714 |
|
|
if (vect_print_dump_info (REPORT_DETAILS))
|
715 |
|
|
{
|
716 |
|
|
fprintf (vect_dump, "pattern recognized: ");
|
717 |
|
|
print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM);
|
718 |
|
|
}
|
719 |
|
|
|
720 |
|
|
/* Mark the stmts that are involved in the pattern. */
|
721 |
|
|
gsi_insert_before (&si, pattern_stmt, GSI_SAME_STMT);
|
722 |
|
|
set_vinfo_for_stmt (pattern_stmt,
|
723 |
|
|
new_stmt_vec_info (pattern_stmt, loop_vinfo, NULL));
|
724 |
|
|
pattern_stmt_info = vinfo_for_stmt (pattern_stmt);
|
725 |
|
|
|
726 |
|
|
STMT_VINFO_RELATED_STMT (pattern_stmt_info) = stmt;
|
727 |
|
|
STMT_VINFO_DEF_TYPE (pattern_stmt_info) = STMT_VINFO_DEF_TYPE (stmt_info);
|
728 |
|
|
STMT_VINFO_VECTYPE (pattern_stmt_info) = pattern_vectype;
|
729 |
|
|
STMT_VINFO_IN_PATTERN_P (stmt_info) = true;
|
730 |
|
|
STMT_VINFO_RELATED_STMT (stmt_info) = pattern_stmt;
|
731 |
|
|
|
732 |
|
|
return;
|
733 |
|
|
}
|
734 |
|
|
|
735 |
|
|
|
736 |
|
|
/* Function vect_pattern_recog
|
737 |
|
|
|
738 |
|
|
Input:
|
739 |
|
|
LOOP_VINFO - a struct_loop_info of a loop in which we want to look for
|
740 |
|
|
computation idioms.
|
741 |
|
|
|
742 |
|
|
Output - for each computation idiom that is detected we insert a new stmt
|
743 |
|
|
that provides the same functionality and that can be vectorized. We
|
744 |
|
|
also record some information in the struct_stmt_info of the relevant
|
745 |
|
|
stmts, as explained below:
|
746 |
|
|
|
747 |
|
|
At the entry to this function we have the following stmts, with the
|
748 |
|
|
following initial value in the STMT_VINFO fields:
|
749 |
|
|
|
750 |
|
|
stmt in_pattern_p related_stmt vec_stmt
|
751 |
|
|
S1: a_i = .... - - -
|
752 |
|
|
S2: a_2 = ..use(a_i).. - - -
|
753 |
|
|
S3: a_1 = ..use(a_2).. - - -
|
754 |
|
|
S4: a_0 = ..use(a_1).. - - -
|
755 |
|
|
S5: ... = ..use(a_0).. - - -
|
756 |
|
|
|
757 |
|
|
Say the sequence {S1,S2,S3,S4} was detected as a pattern that can be
|
758 |
|
|
represented by a single stmt. We then:
|
759 |
|
|
- create a new stmt S6 that will replace the pattern.
|
760 |
|
|
- insert the new stmt S6 before the last stmt in the pattern
|
761 |
|
|
- fill in the STMT_VINFO fields as follows:
|
762 |
|
|
|
763 |
|
|
in_pattern_p related_stmt vec_stmt
|
764 |
|
|
S1: a_i = .... - - -
|
765 |
|
|
S2: a_2 = ..use(a_i).. - - -
|
766 |
|
|
S3: a_1 = ..use(a_2).. - - -
|
767 |
|
|
> S6: a_new = .... - S4 -
|
768 |
|
|
S4: a_0 = ..use(a_1).. true S6 -
|
769 |
|
|
S5: ... = ..use(a_0).. - - -
|
770 |
|
|
|
771 |
|
|
(the last stmt in the pattern (S4) and the new pattern stmt (S6) point
|
772 |
|
|
to each other through the RELATED_STMT field).
|
773 |
|
|
|
774 |
|
|
S6 will be marked as relevant in vect_mark_stmts_to_be_vectorized instead
|
775 |
|
|
of S4 because it will replace all its uses. Stmts {S1,S2,S3} will
|
776 |
|
|
remain irrelevant unless used by stmts other than S4.
|
777 |
|
|
|
778 |
|
|
If vectorization succeeds, vect_transform_stmt will skip over {S1,S2,S3}
|
779 |
|
|
(because they are marked as irrelevant). It will vectorize S6, and record
|
780 |
|
|
a pointer to the new vector stmt VS6 both from S6 (as usual), and also
|
781 |
|
|
from S4. We do that so that when we get to vectorizing stmts that use the
|
782 |
|
|
def of S4 (like S5 that uses a_0), we'll know where to take the relevant
|
783 |
|
|
vector-def from. S4 will be skipped, and S5 will be vectorized as usual:
|
784 |
|
|
|
785 |
|
|
in_pattern_p related_stmt vec_stmt
|
786 |
|
|
S1: a_i = .... - - -
|
787 |
|
|
S2: a_2 = ..use(a_i).. - - -
|
788 |
|
|
S3: a_1 = ..use(a_2).. - - -
|
789 |
|
|
> VS6: va_new = .... - - -
|
790 |
|
|
S6: a_new = .... - S4 VS6
|
791 |
|
|
S4: a_0 = ..use(a_1).. true S6 VS6
|
792 |
|
|
> VS5: ... = ..vuse(va_new).. - - -
|
793 |
|
|
S5: ... = ..use(a_0).. - - -
|
794 |
|
|
|
795 |
|
|
DCE could then get rid of {S1,S2,S3,S4,S5,S6} (if their defs are not used
|
796 |
|
|
elsewhere), and we'll end up with:
|
797 |
|
|
|
798 |
|
|
VS6: va_new = ....
|
799 |
|
|
VS5: ... = ..vuse(va_new)..
|
800 |
|
|
|
801 |
|
|
If vectorization does not succeed, DCE will clean S6 away (its def is
|
802 |
|
|
not used), and we'll end up with the original sequence.
|
803 |
|
|
*/
|
804 |
|
|
|
805 |
|
|
void
|
806 |
|
|
vect_pattern_recog (loop_vec_info loop_vinfo)
|
807 |
|
|
{
|
808 |
|
|
struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
|
809 |
|
|
basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
|
810 |
|
|
unsigned int nbbs = loop->num_nodes;
|
811 |
|
|
gimple_stmt_iterator si;
|
812 |
|
|
unsigned int i, j;
|
813 |
|
|
gimple (* vect_recog_func_ptr) (gimple, tree *, tree *);
|
814 |
|
|
|
815 |
|
|
if (vect_print_dump_info (REPORT_DETAILS))
|
816 |
|
|
fprintf (vect_dump, "=== vect_pattern_recog ===");
|
817 |
|
|
|
818 |
|
|
/* Scan through the loop stmts, applying the pattern recognition
|
819 |
|
|
functions starting at each stmt visited: */
|
820 |
|
|
for (i = 0; i < nbbs; i++)
|
821 |
|
|
{
|
822 |
|
|
basic_block bb = bbs[i];
|
823 |
|
|
for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
|
824 |
|
|
{
|
825 |
|
|
/* Scan over all generic vect_recog_xxx_pattern functions. */
|
826 |
|
|
for (j = 0; j < NUM_PATTERNS; j++)
|
827 |
|
|
{
|
828 |
|
|
vect_recog_func_ptr = vect_vect_recog_func_ptrs[j];
|
829 |
|
|
vect_pattern_recog_1 (vect_recog_func_ptr, si);
|
830 |
|
|
}
|
831 |
|
|
}
|
832 |
|
|
}
|
833 |
|
|
}
|