1 |
684 |
jeremybenn |
/* Lowering pass for OpenMP directives. Converts OpenMP directives
|
2 |
|
|
into explicit calls to the runtime library (libgomp) and data
|
3 |
|
|
marshalling to implement data sharing and copying clauses.
|
4 |
|
|
Contributed by Diego Novillo <dnovillo@redhat.com>
|
5 |
|
|
|
6 |
|
|
Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012
|
7 |
|
|
Free Software Foundation, Inc.
|
8 |
|
|
|
9 |
|
|
This file is part of GCC.
|
10 |
|
|
|
11 |
|
|
GCC is free software; you can redistribute it and/or modify it under
|
12 |
|
|
the terms of the GNU General Public License as published by the Free
|
13 |
|
|
Software Foundation; either version 3, or (at your option) any later
|
14 |
|
|
version.
|
15 |
|
|
|
16 |
|
|
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
|
17 |
|
|
WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
18 |
|
|
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
19 |
|
|
for more details.
|
20 |
|
|
|
21 |
|
|
You should have received a copy of the GNU General Public License
|
22 |
|
|
along with GCC; see the file COPYING3. If not see
|
23 |
|
|
<http://www.gnu.org/licenses/>. */
|
24 |
|
|
|
25 |
|
|
#include "config.h"
|
26 |
|
|
#include "system.h"
|
27 |
|
|
#include "coretypes.h"
|
28 |
|
|
#include "tm.h"
|
29 |
|
|
#include "tree.h"
|
30 |
|
|
#include "rtl.h"
|
31 |
|
|
#include "gimple.h"
|
32 |
|
|
#include "tree-iterator.h"
|
33 |
|
|
#include "tree-inline.h"
|
34 |
|
|
#include "langhooks.h"
|
35 |
|
|
#include "diagnostic-core.h"
|
36 |
|
|
#include "tree-flow.h"
|
37 |
|
|
#include "timevar.h"
|
38 |
|
|
#include "flags.h"
|
39 |
|
|
#include "function.h"
|
40 |
|
|
#include "expr.h"
|
41 |
|
|
#include "tree-pass.h"
|
42 |
|
|
#include "ggc.h"
|
43 |
|
|
#include "except.h"
|
44 |
|
|
#include "splay-tree.h"
|
45 |
|
|
#include "optabs.h"
|
46 |
|
|
#include "cfgloop.h"
|
47 |
|
|
|
48 |
|
|
|
49 |
|
|
/* Lowering of OpenMP parallel and workshare constructs proceeds in two
|
50 |
|
|
phases. The first phase scans the function looking for OMP statements
|
51 |
|
|
and then for variables that must be replaced to satisfy data sharing
|
52 |
|
|
clauses. The second phase expands code for the constructs, as well as
|
53 |
|
|
re-gimplifying things when variables have been replaced with complex
|
54 |
|
|
expressions.
|
55 |
|
|
|
56 |
|
|
Final code generation is done by pass_expand_omp. The flowgraph is
|
57 |
|
|
scanned for parallel regions which are then moved to a new
|
58 |
|
|
function, to be invoked by the thread library. */
|
59 |
|
|
|
60 |
|
|
/* Context structure. Used to store information about each parallel
|
61 |
|
|
directive in the code. */
|
62 |
|
|
|
63 |
|
|
typedef struct omp_context
|
64 |
|
|
{
|
65 |
|
|
/* This field must be at the beginning, as we do "inheritance": Some
|
66 |
|
|
callback functions for tree-inline.c (e.g., omp_copy_decl)
|
67 |
|
|
receive a copy_body_data pointer that is up-casted to an
|
68 |
|
|
omp_context pointer. */
|
69 |
|
|
copy_body_data cb;
|
70 |
|
|
|
71 |
|
|
/* The tree of contexts corresponding to the encountered constructs. */
|
72 |
|
|
struct omp_context *outer;
|
73 |
|
|
gimple stmt;
|
74 |
|
|
|
75 |
|
|
/* Map variables to fields in a structure that allows communication
|
76 |
|
|
between sending and receiving threads. */
|
77 |
|
|
splay_tree field_map;
|
78 |
|
|
tree record_type;
|
79 |
|
|
tree sender_decl;
|
80 |
|
|
tree receiver_decl;
|
81 |
|
|
|
82 |
|
|
/* These are used just by task contexts, if task firstprivate fn is
|
83 |
|
|
needed. srecord_type is used to communicate from the thread
|
84 |
|
|
that encountered the task construct to task firstprivate fn,
|
85 |
|
|
record_type is allocated by GOMP_task, initialized by task firstprivate
|
86 |
|
|
fn and passed to the task body fn. */
|
87 |
|
|
splay_tree sfield_map;
|
88 |
|
|
tree srecord_type;
|
89 |
|
|
|
90 |
|
|
/* A chain of variables to add to the top-level block surrounding the
|
91 |
|
|
construct. In the case of a parallel, this is in the child function. */
|
92 |
|
|
tree block_vars;
|
93 |
|
|
|
94 |
|
|
/* What to do with variables with implicitly determined sharing
|
95 |
|
|
attributes. */
|
96 |
|
|
enum omp_clause_default_kind default_kind;
|
97 |
|
|
|
98 |
|
|
/* Nesting depth of this context. Used to beautify error messages re
|
99 |
|
|
invalid gotos. The outermost ctx is depth 1, with depth 0 being
|
100 |
|
|
reserved for the main body of the function. */
|
101 |
|
|
int depth;
|
102 |
|
|
|
103 |
|
|
/* True if this parallel directive is nested within another. */
|
104 |
|
|
bool is_nested;
|
105 |
|
|
} omp_context;
|
106 |
|
|
|
107 |
|
|
|
108 |
|
|
struct omp_for_data_loop
|
109 |
|
|
{
|
110 |
|
|
tree v, n1, n2, step;
|
111 |
|
|
enum tree_code cond_code;
|
112 |
|
|
};
|
113 |
|
|
|
114 |
|
|
/* A structure describing the main elements of a parallel loop. */
|
115 |
|
|
|
116 |
|
|
struct omp_for_data
|
117 |
|
|
{
|
118 |
|
|
struct omp_for_data_loop loop;
|
119 |
|
|
tree chunk_size;
|
120 |
|
|
gimple for_stmt;
|
121 |
|
|
tree pre, iter_type;
|
122 |
|
|
int collapse;
|
123 |
|
|
bool have_nowait, have_ordered;
|
124 |
|
|
enum omp_clause_schedule_kind sched_kind;
|
125 |
|
|
struct omp_for_data_loop *loops;
|
126 |
|
|
};
|
127 |
|
|
|
128 |
|
|
|
129 |
|
|
static splay_tree all_contexts;
|
130 |
|
|
static int taskreg_nesting_level;
|
131 |
|
|
struct omp_region *root_omp_region;
|
132 |
|
|
static bitmap task_shared_vars;
|
133 |
|
|
|
134 |
|
|
static void scan_omp (gimple_seq, omp_context *);
|
135 |
|
|
static tree scan_omp_1_op (tree *, int *, void *);
|
136 |
|
|
|
137 |
|
|
#define WALK_SUBSTMTS \
|
138 |
|
|
case GIMPLE_BIND: \
|
139 |
|
|
case GIMPLE_TRY: \
|
140 |
|
|
case GIMPLE_CATCH: \
|
141 |
|
|
case GIMPLE_EH_FILTER: \
|
142 |
|
|
case GIMPLE_TRANSACTION: \
|
143 |
|
|
/* The sub-statements for these should be walked. */ \
|
144 |
|
|
*handled_ops_p = false; \
|
145 |
|
|
break;
|
146 |
|
|
|
147 |
|
|
/* Convenience function for calling scan_omp_1_op on tree operands. */
|
148 |
|
|
|
149 |
|
|
static inline tree
|
150 |
|
|
scan_omp_op (tree *tp, omp_context *ctx)
|
151 |
|
|
{
|
152 |
|
|
struct walk_stmt_info wi;
|
153 |
|
|
|
154 |
|
|
memset (&wi, 0, sizeof (wi));
|
155 |
|
|
wi.info = ctx;
|
156 |
|
|
wi.want_locations = true;
|
157 |
|
|
|
158 |
|
|
return walk_tree (tp, scan_omp_1_op, &wi, NULL);
|
159 |
|
|
}
|
160 |
|
|
|
161 |
|
|
static void lower_omp (gimple_seq, omp_context *);
|
162 |
|
|
static tree lookup_decl_in_outer_ctx (tree, omp_context *);
|
163 |
|
|
static tree maybe_lookup_decl_in_outer_ctx (tree, omp_context *);
|
164 |
|
|
|
165 |
|
|
/* Find an OpenMP clause of type KIND within CLAUSES. */
|
166 |
|
|
|
167 |
|
|
tree
|
168 |
|
|
find_omp_clause (tree clauses, enum omp_clause_code kind)
|
169 |
|
|
{
|
170 |
|
|
for (; clauses ; clauses = OMP_CLAUSE_CHAIN (clauses))
|
171 |
|
|
if (OMP_CLAUSE_CODE (clauses) == kind)
|
172 |
|
|
return clauses;
|
173 |
|
|
|
174 |
|
|
return NULL_TREE;
|
175 |
|
|
}
|
176 |
|
|
|
177 |
|
|
/* Return true if CTX is for an omp parallel. */
|
178 |
|
|
|
179 |
|
|
static inline bool
|
180 |
|
|
is_parallel_ctx (omp_context *ctx)
|
181 |
|
|
{
|
182 |
|
|
return gimple_code (ctx->stmt) == GIMPLE_OMP_PARALLEL;
|
183 |
|
|
}
|
184 |
|
|
|
185 |
|
|
|
186 |
|
|
/* Return true if CTX is for an omp task. */
|
187 |
|
|
|
188 |
|
|
static inline bool
|
189 |
|
|
is_task_ctx (omp_context *ctx)
|
190 |
|
|
{
|
191 |
|
|
return gimple_code (ctx->stmt) == GIMPLE_OMP_TASK;
|
192 |
|
|
}
|
193 |
|
|
|
194 |
|
|
|
195 |
|
|
/* Return true if CTX is for an omp parallel or omp task. */
|
196 |
|
|
|
197 |
|
|
static inline bool
|
198 |
|
|
is_taskreg_ctx (omp_context *ctx)
|
199 |
|
|
{
|
200 |
|
|
return gimple_code (ctx->stmt) == GIMPLE_OMP_PARALLEL
|
201 |
|
|
|| gimple_code (ctx->stmt) == GIMPLE_OMP_TASK;
|
202 |
|
|
}
|
203 |
|
|
|
204 |
|
|
|
205 |
|
|
/* Return true if REGION is a combined parallel+workshare region. */
|
206 |
|
|
|
207 |
|
|
static inline bool
|
208 |
|
|
is_combined_parallel (struct omp_region *region)
|
209 |
|
|
{
|
210 |
|
|
return region->is_combined_parallel;
|
211 |
|
|
}
|
212 |
|
|
|
213 |
|
|
|
214 |
|
|
/* Extract the header elements of parallel loop FOR_STMT and store
|
215 |
|
|
them into *FD. */
|
216 |
|
|
|
217 |
|
|
static void
|
218 |
|
|
extract_omp_for_data (gimple for_stmt, struct omp_for_data *fd,
|
219 |
|
|
struct omp_for_data_loop *loops)
|
220 |
|
|
{
|
221 |
|
|
tree t, var, *collapse_iter, *collapse_count;
|
222 |
|
|
tree count = NULL_TREE, iter_type = long_integer_type_node;
|
223 |
|
|
struct omp_for_data_loop *loop;
|
224 |
|
|
int i;
|
225 |
|
|
struct omp_for_data_loop dummy_loop;
|
226 |
|
|
location_t loc = gimple_location (for_stmt);
|
227 |
|
|
|
228 |
|
|
fd->for_stmt = for_stmt;
|
229 |
|
|
fd->pre = NULL;
|
230 |
|
|
fd->collapse = gimple_omp_for_collapse (for_stmt);
|
231 |
|
|
if (fd->collapse > 1)
|
232 |
|
|
fd->loops = loops;
|
233 |
|
|
else
|
234 |
|
|
fd->loops = &fd->loop;
|
235 |
|
|
|
236 |
|
|
fd->have_nowait = fd->have_ordered = false;
|
237 |
|
|
fd->sched_kind = OMP_CLAUSE_SCHEDULE_STATIC;
|
238 |
|
|
fd->chunk_size = NULL_TREE;
|
239 |
|
|
collapse_iter = NULL;
|
240 |
|
|
collapse_count = NULL;
|
241 |
|
|
|
242 |
|
|
for (t = gimple_omp_for_clauses (for_stmt); t ; t = OMP_CLAUSE_CHAIN (t))
|
243 |
|
|
switch (OMP_CLAUSE_CODE (t))
|
244 |
|
|
{
|
245 |
|
|
case OMP_CLAUSE_NOWAIT:
|
246 |
|
|
fd->have_nowait = true;
|
247 |
|
|
break;
|
248 |
|
|
case OMP_CLAUSE_ORDERED:
|
249 |
|
|
fd->have_ordered = true;
|
250 |
|
|
break;
|
251 |
|
|
case OMP_CLAUSE_SCHEDULE:
|
252 |
|
|
fd->sched_kind = OMP_CLAUSE_SCHEDULE_KIND (t);
|
253 |
|
|
fd->chunk_size = OMP_CLAUSE_SCHEDULE_CHUNK_EXPR (t);
|
254 |
|
|
break;
|
255 |
|
|
case OMP_CLAUSE_COLLAPSE:
|
256 |
|
|
if (fd->collapse > 1)
|
257 |
|
|
{
|
258 |
|
|
collapse_iter = &OMP_CLAUSE_COLLAPSE_ITERVAR (t);
|
259 |
|
|
collapse_count = &OMP_CLAUSE_COLLAPSE_COUNT (t);
|
260 |
|
|
}
|
261 |
|
|
default:
|
262 |
|
|
break;
|
263 |
|
|
}
|
264 |
|
|
|
265 |
|
|
/* FIXME: for now map schedule(auto) to schedule(static).
|
266 |
|
|
There should be analysis to determine whether all iterations
|
267 |
|
|
are approximately the same amount of work (then schedule(static)
|
268 |
|
|
is best) or if it varies (then schedule(dynamic,N) is better). */
|
269 |
|
|
if (fd->sched_kind == OMP_CLAUSE_SCHEDULE_AUTO)
|
270 |
|
|
{
|
271 |
|
|
fd->sched_kind = OMP_CLAUSE_SCHEDULE_STATIC;
|
272 |
|
|
gcc_assert (fd->chunk_size == NULL);
|
273 |
|
|
}
|
274 |
|
|
gcc_assert (fd->collapse == 1 || collapse_iter != NULL);
|
275 |
|
|
if (fd->sched_kind == OMP_CLAUSE_SCHEDULE_RUNTIME)
|
276 |
|
|
gcc_assert (fd->chunk_size == NULL);
|
277 |
|
|
else if (fd->chunk_size == NULL)
|
278 |
|
|
{
|
279 |
|
|
/* We only need to compute a default chunk size for ordered
|
280 |
|
|
static loops and dynamic loops. */
|
281 |
|
|
if (fd->sched_kind != OMP_CLAUSE_SCHEDULE_STATIC
|
282 |
|
|
|| fd->have_ordered
|
283 |
|
|
|| fd->collapse > 1)
|
284 |
|
|
fd->chunk_size = (fd->sched_kind == OMP_CLAUSE_SCHEDULE_STATIC)
|
285 |
|
|
? integer_zero_node : integer_one_node;
|
286 |
|
|
}
|
287 |
|
|
|
288 |
|
|
for (i = 0; i < fd->collapse; i++)
|
289 |
|
|
{
|
290 |
|
|
if (fd->collapse == 1)
|
291 |
|
|
loop = &fd->loop;
|
292 |
|
|
else if (loops != NULL)
|
293 |
|
|
loop = loops + i;
|
294 |
|
|
else
|
295 |
|
|
loop = &dummy_loop;
|
296 |
|
|
|
297 |
|
|
|
298 |
|
|
loop->v = gimple_omp_for_index (for_stmt, i);
|
299 |
|
|
gcc_assert (SSA_VAR_P (loop->v));
|
300 |
|
|
gcc_assert (TREE_CODE (TREE_TYPE (loop->v)) == INTEGER_TYPE
|
301 |
|
|
|| TREE_CODE (TREE_TYPE (loop->v)) == POINTER_TYPE);
|
302 |
|
|
var = TREE_CODE (loop->v) == SSA_NAME ? SSA_NAME_VAR (loop->v) : loop->v;
|
303 |
|
|
loop->n1 = gimple_omp_for_initial (for_stmt, i);
|
304 |
|
|
|
305 |
|
|
loop->cond_code = gimple_omp_for_cond (for_stmt, i);
|
306 |
|
|
loop->n2 = gimple_omp_for_final (for_stmt, i);
|
307 |
|
|
switch (loop->cond_code)
|
308 |
|
|
{
|
309 |
|
|
case LT_EXPR:
|
310 |
|
|
case GT_EXPR:
|
311 |
|
|
break;
|
312 |
|
|
case LE_EXPR:
|
313 |
|
|
if (POINTER_TYPE_P (TREE_TYPE (loop->n2)))
|
314 |
|
|
loop->n2 = fold_build_pointer_plus_hwi_loc (loc, loop->n2, 1);
|
315 |
|
|
else
|
316 |
|
|
loop->n2 = fold_build2_loc (loc,
|
317 |
|
|
PLUS_EXPR, TREE_TYPE (loop->n2), loop->n2,
|
318 |
|
|
build_int_cst (TREE_TYPE (loop->n2), 1));
|
319 |
|
|
loop->cond_code = LT_EXPR;
|
320 |
|
|
break;
|
321 |
|
|
case GE_EXPR:
|
322 |
|
|
if (POINTER_TYPE_P (TREE_TYPE (loop->n2)))
|
323 |
|
|
loop->n2 = fold_build_pointer_plus_hwi_loc (loc, loop->n2, -1);
|
324 |
|
|
else
|
325 |
|
|
loop->n2 = fold_build2_loc (loc,
|
326 |
|
|
MINUS_EXPR, TREE_TYPE (loop->n2), loop->n2,
|
327 |
|
|
build_int_cst (TREE_TYPE (loop->n2), 1));
|
328 |
|
|
loop->cond_code = GT_EXPR;
|
329 |
|
|
break;
|
330 |
|
|
default:
|
331 |
|
|
gcc_unreachable ();
|
332 |
|
|
}
|
333 |
|
|
|
334 |
|
|
t = gimple_omp_for_incr (for_stmt, i);
|
335 |
|
|
gcc_assert (TREE_OPERAND (t, 0) == var);
|
336 |
|
|
switch (TREE_CODE (t))
|
337 |
|
|
{
|
338 |
|
|
case PLUS_EXPR:
|
339 |
|
|
case POINTER_PLUS_EXPR:
|
340 |
|
|
loop->step = TREE_OPERAND (t, 1);
|
341 |
|
|
break;
|
342 |
|
|
case MINUS_EXPR:
|
343 |
|
|
loop->step = TREE_OPERAND (t, 1);
|
344 |
|
|
loop->step = fold_build1_loc (loc,
|
345 |
|
|
NEGATE_EXPR, TREE_TYPE (loop->step),
|
346 |
|
|
loop->step);
|
347 |
|
|
break;
|
348 |
|
|
default:
|
349 |
|
|
gcc_unreachable ();
|
350 |
|
|
}
|
351 |
|
|
|
352 |
|
|
if (iter_type != long_long_unsigned_type_node)
|
353 |
|
|
{
|
354 |
|
|
if (POINTER_TYPE_P (TREE_TYPE (loop->v)))
|
355 |
|
|
iter_type = long_long_unsigned_type_node;
|
356 |
|
|
else if (TYPE_UNSIGNED (TREE_TYPE (loop->v))
|
357 |
|
|
&& TYPE_PRECISION (TREE_TYPE (loop->v))
|
358 |
|
|
>= TYPE_PRECISION (iter_type))
|
359 |
|
|
{
|
360 |
|
|
tree n;
|
361 |
|
|
|
362 |
|
|
if (loop->cond_code == LT_EXPR)
|
363 |
|
|
n = fold_build2_loc (loc,
|
364 |
|
|
PLUS_EXPR, TREE_TYPE (loop->v),
|
365 |
|
|
loop->n2, loop->step);
|
366 |
|
|
else
|
367 |
|
|
n = loop->n1;
|
368 |
|
|
if (TREE_CODE (n) != INTEGER_CST
|
369 |
|
|
|| tree_int_cst_lt (TYPE_MAX_VALUE (iter_type), n))
|
370 |
|
|
iter_type = long_long_unsigned_type_node;
|
371 |
|
|
}
|
372 |
|
|
else if (TYPE_PRECISION (TREE_TYPE (loop->v))
|
373 |
|
|
> TYPE_PRECISION (iter_type))
|
374 |
|
|
{
|
375 |
|
|
tree n1, n2;
|
376 |
|
|
|
377 |
|
|
if (loop->cond_code == LT_EXPR)
|
378 |
|
|
{
|
379 |
|
|
n1 = loop->n1;
|
380 |
|
|
n2 = fold_build2_loc (loc,
|
381 |
|
|
PLUS_EXPR, TREE_TYPE (loop->v),
|
382 |
|
|
loop->n2, loop->step);
|
383 |
|
|
}
|
384 |
|
|
else
|
385 |
|
|
{
|
386 |
|
|
n1 = fold_build2_loc (loc,
|
387 |
|
|
MINUS_EXPR, TREE_TYPE (loop->v),
|
388 |
|
|
loop->n2, loop->step);
|
389 |
|
|
n2 = loop->n1;
|
390 |
|
|
}
|
391 |
|
|
if (TREE_CODE (n1) != INTEGER_CST
|
392 |
|
|
|| TREE_CODE (n2) != INTEGER_CST
|
393 |
|
|
|| !tree_int_cst_lt (TYPE_MIN_VALUE (iter_type), n1)
|
394 |
|
|
|| !tree_int_cst_lt (n2, TYPE_MAX_VALUE (iter_type)))
|
395 |
|
|
iter_type = long_long_unsigned_type_node;
|
396 |
|
|
}
|
397 |
|
|
}
|
398 |
|
|
|
399 |
|
|
if (collapse_count && *collapse_count == NULL)
|
400 |
|
|
{
|
401 |
|
|
if ((i == 0 || count != NULL_TREE)
|
402 |
|
|
&& TREE_CODE (TREE_TYPE (loop->v)) == INTEGER_TYPE
|
403 |
|
|
&& TREE_CONSTANT (loop->n1)
|
404 |
|
|
&& TREE_CONSTANT (loop->n2)
|
405 |
|
|
&& TREE_CODE (loop->step) == INTEGER_CST)
|
406 |
|
|
{
|
407 |
|
|
tree itype = TREE_TYPE (loop->v);
|
408 |
|
|
|
409 |
|
|
if (POINTER_TYPE_P (itype))
|
410 |
|
|
itype
|
411 |
|
|
= lang_hooks.types.type_for_size (TYPE_PRECISION (itype), 0);
|
412 |
|
|
t = build_int_cst (itype, (loop->cond_code == LT_EXPR ? -1 : 1));
|
413 |
|
|
t = fold_build2_loc (loc,
|
414 |
|
|
PLUS_EXPR, itype,
|
415 |
|
|
fold_convert_loc (loc, itype, loop->step), t);
|
416 |
|
|
t = fold_build2_loc (loc, PLUS_EXPR, itype, t,
|
417 |
|
|
fold_convert_loc (loc, itype, loop->n2));
|
418 |
|
|
t = fold_build2_loc (loc, MINUS_EXPR, itype, t,
|
419 |
|
|
fold_convert_loc (loc, itype, loop->n1));
|
420 |
|
|
if (TYPE_UNSIGNED (itype) && loop->cond_code == GT_EXPR)
|
421 |
|
|
t = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype,
|
422 |
|
|
fold_build1_loc (loc, NEGATE_EXPR, itype, t),
|
423 |
|
|
fold_build1_loc (loc, NEGATE_EXPR, itype,
|
424 |
|
|
fold_convert_loc (loc, itype,
|
425 |
|
|
loop->step)));
|
426 |
|
|
else
|
427 |
|
|
t = fold_build2_loc (loc, TRUNC_DIV_EXPR, itype, t,
|
428 |
|
|
fold_convert_loc (loc, itype, loop->step));
|
429 |
|
|
t = fold_convert_loc (loc, long_long_unsigned_type_node, t);
|
430 |
|
|
if (count != NULL_TREE)
|
431 |
|
|
count = fold_build2_loc (loc,
|
432 |
|
|
MULT_EXPR, long_long_unsigned_type_node,
|
433 |
|
|
count, t);
|
434 |
|
|
else
|
435 |
|
|
count = t;
|
436 |
|
|
if (TREE_CODE (count) != INTEGER_CST)
|
437 |
|
|
count = NULL_TREE;
|
438 |
|
|
}
|
439 |
|
|
else
|
440 |
|
|
count = NULL_TREE;
|
441 |
|
|
}
|
442 |
|
|
}
|
443 |
|
|
|
444 |
|
|
if (count)
|
445 |
|
|
{
|
446 |
|
|
if (!tree_int_cst_lt (count, TYPE_MAX_VALUE (long_integer_type_node)))
|
447 |
|
|
iter_type = long_long_unsigned_type_node;
|
448 |
|
|
else
|
449 |
|
|
iter_type = long_integer_type_node;
|
450 |
|
|
}
|
451 |
|
|
else if (collapse_iter && *collapse_iter != NULL)
|
452 |
|
|
iter_type = TREE_TYPE (*collapse_iter);
|
453 |
|
|
fd->iter_type = iter_type;
|
454 |
|
|
if (collapse_iter && *collapse_iter == NULL)
|
455 |
|
|
*collapse_iter = create_tmp_var (iter_type, ".iter");
|
456 |
|
|
if (collapse_count && *collapse_count == NULL)
|
457 |
|
|
{
|
458 |
|
|
if (count)
|
459 |
|
|
*collapse_count = fold_convert_loc (loc, iter_type, count);
|
460 |
|
|
else
|
461 |
|
|
*collapse_count = create_tmp_var (iter_type, ".count");
|
462 |
|
|
}
|
463 |
|
|
|
464 |
|
|
if (fd->collapse > 1)
|
465 |
|
|
{
|
466 |
|
|
fd->loop.v = *collapse_iter;
|
467 |
|
|
fd->loop.n1 = build_int_cst (TREE_TYPE (fd->loop.v), 0);
|
468 |
|
|
fd->loop.n2 = *collapse_count;
|
469 |
|
|
fd->loop.step = build_int_cst (TREE_TYPE (fd->loop.v), 1);
|
470 |
|
|
fd->loop.cond_code = LT_EXPR;
|
471 |
|
|
}
|
472 |
|
|
}
|
473 |
|
|
|
474 |
|
|
|
475 |
|
|
/* Given two blocks PAR_ENTRY_BB and WS_ENTRY_BB such that WS_ENTRY_BB
|
476 |
|
|
is the immediate dominator of PAR_ENTRY_BB, return true if there
|
477 |
|
|
are no data dependencies that would prevent expanding the parallel
|
478 |
|
|
directive at PAR_ENTRY_BB as a combined parallel+workshare region.
|
479 |
|
|
|
480 |
|
|
When expanding a combined parallel+workshare region, the call to
|
481 |
|
|
the child function may need additional arguments in the case of
|
482 |
|
|
GIMPLE_OMP_FOR regions. In some cases, these arguments are
|
483 |
|
|
computed out of variables passed in from the parent to the child
|
484 |
|
|
via 'struct .omp_data_s'. For instance:
|
485 |
|
|
|
486 |
|
|
#pragma omp parallel for schedule (guided, i * 4)
|
487 |
|
|
for (j ...)
|
488 |
|
|
|
489 |
|
|
Is lowered into:
|
490 |
|
|
|
491 |
|
|
# BLOCK 2 (PAR_ENTRY_BB)
|
492 |
|
|
.omp_data_o.i = i;
|
493 |
|
|
#pragma omp parallel [child fn: bar.omp_fn.0 ( ..., D.1598)
|
494 |
|
|
|
495 |
|
|
# BLOCK 3 (WS_ENTRY_BB)
|
496 |
|
|
.omp_data_i = &.omp_data_o;
|
497 |
|
|
D.1667 = .omp_data_i->i;
|
498 |
|
|
D.1598 = D.1667 * 4;
|
499 |
|
|
#pragma omp for schedule (guided, D.1598)
|
500 |
|
|
|
501 |
|
|
When we outline the parallel region, the call to the child function
|
502 |
|
|
'bar.omp_fn.0' will need the value D.1598 in its argument list, but
|
503 |
|
|
that value is computed *after* the call site. So, in principle we
|
504 |
|
|
cannot do the transformation.
|
505 |
|
|
|
506 |
|
|
To see whether the code in WS_ENTRY_BB blocks the combined
|
507 |
|
|
parallel+workshare call, we collect all the variables used in the
|
508 |
|
|
GIMPLE_OMP_FOR header check whether they appear on the LHS of any
|
509 |
|
|
statement in WS_ENTRY_BB. If so, then we cannot emit the combined
|
510 |
|
|
call.
|
511 |
|
|
|
512 |
|
|
FIXME. If we had the SSA form built at this point, we could merely
|
513 |
|
|
hoist the code in block 3 into block 2 and be done with it. But at
|
514 |
|
|
this point we don't have dataflow information and though we could
|
515 |
|
|
hack something up here, it is really not worth the aggravation. */
|
516 |
|
|
|
517 |
|
|
static bool
|
518 |
|
|
workshare_safe_to_combine_p (basic_block ws_entry_bb)
|
519 |
|
|
{
|
520 |
|
|
struct omp_for_data fd;
|
521 |
|
|
gimple ws_stmt = last_stmt (ws_entry_bb);
|
522 |
|
|
|
523 |
|
|
if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS)
|
524 |
|
|
return true;
|
525 |
|
|
|
526 |
|
|
gcc_assert (gimple_code (ws_stmt) == GIMPLE_OMP_FOR);
|
527 |
|
|
|
528 |
|
|
extract_omp_for_data (ws_stmt, &fd, NULL);
|
529 |
|
|
|
530 |
|
|
if (fd.collapse > 1 && TREE_CODE (fd.loop.n2) != INTEGER_CST)
|
531 |
|
|
return false;
|
532 |
|
|
if (fd.iter_type != long_integer_type_node)
|
533 |
|
|
return false;
|
534 |
|
|
|
535 |
|
|
/* FIXME. We give up too easily here. If any of these arguments
|
536 |
|
|
are not constants, they will likely involve variables that have
|
537 |
|
|
been mapped into fields of .omp_data_s for sharing with the child
|
538 |
|
|
function. With appropriate data flow, it would be possible to
|
539 |
|
|
see through this. */
|
540 |
|
|
if (!is_gimple_min_invariant (fd.loop.n1)
|
541 |
|
|
|| !is_gimple_min_invariant (fd.loop.n2)
|
542 |
|
|
|| !is_gimple_min_invariant (fd.loop.step)
|
543 |
|
|
|| (fd.chunk_size && !is_gimple_min_invariant (fd.chunk_size)))
|
544 |
|
|
return false;
|
545 |
|
|
|
546 |
|
|
return true;
|
547 |
|
|
}
|
548 |
|
|
|
549 |
|
|
|
550 |
|
|
/* Collect additional arguments needed to emit a combined
|
551 |
|
|
parallel+workshare call. WS_STMT is the workshare directive being
|
552 |
|
|
expanded. */
|
553 |
|
|
|
554 |
|
|
static VEC(tree,gc) *
|
555 |
|
|
get_ws_args_for (gimple ws_stmt)
|
556 |
|
|
{
|
557 |
|
|
tree t;
|
558 |
|
|
location_t loc = gimple_location (ws_stmt);
|
559 |
|
|
VEC(tree,gc) *ws_args;
|
560 |
|
|
|
561 |
|
|
if (gimple_code (ws_stmt) == GIMPLE_OMP_FOR)
|
562 |
|
|
{
|
563 |
|
|
struct omp_for_data fd;
|
564 |
|
|
|
565 |
|
|
extract_omp_for_data (ws_stmt, &fd, NULL);
|
566 |
|
|
|
567 |
|
|
ws_args = VEC_alloc (tree, gc, 3 + (fd.chunk_size != 0));
|
568 |
|
|
|
569 |
|
|
t = fold_convert_loc (loc, long_integer_type_node, fd.loop.n1);
|
570 |
|
|
VEC_quick_push (tree, ws_args, t);
|
571 |
|
|
|
572 |
|
|
t = fold_convert_loc (loc, long_integer_type_node, fd.loop.n2);
|
573 |
|
|
VEC_quick_push (tree, ws_args, t);
|
574 |
|
|
|
575 |
|
|
t = fold_convert_loc (loc, long_integer_type_node, fd.loop.step);
|
576 |
|
|
VEC_quick_push (tree, ws_args, t);
|
577 |
|
|
|
578 |
|
|
if (fd.chunk_size)
|
579 |
|
|
{
|
580 |
|
|
t = fold_convert_loc (loc, long_integer_type_node, fd.chunk_size);
|
581 |
|
|
VEC_quick_push (tree, ws_args, t);
|
582 |
|
|
}
|
583 |
|
|
|
584 |
|
|
return ws_args;
|
585 |
|
|
}
|
586 |
|
|
else if (gimple_code (ws_stmt) == GIMPLE_OMP_SECTIONS)
|
587 |
|
|
{
|
588 |
|
|
/* Number of sections is equal to the number of edges from the
|
589 |
|
|
GIMPLE_OMP_SECTIONS_SWITCH statement, except for the one to
|
590 |
|
|
the exit of the sections region. */
|
591 |
|
|
basic_block bb = single_succ (gimple_bb (ws_stmt));
|
592 |
|
|
t = build_int_cst (unsigned_type_node, EDGE_COUNT (bb->succs) - 1);
|
593 |
|
|
ws_args = VEC_alloc (tree, gc, 1);
|
594 |
|
|
VEC_quick_push (tree, ws_args, t);
|
595 |
|
|
return ws_args;
|
596 |
|
|
}
|
597 |
|
|
|
598 |
|
|
gcc_unreachable ();
|
599 |
|
|
}
|
600 |
|
|
|
601 |
|
|
|
602 |
|
|
/* Discover whether REGION is a combined parallel+workshare region. */
|
603 |
|
|
|
604 |
|
|
static void
|
605 |
|
|
determine_parallel_type (struct omp_region *region)
|
606 |
|
|
{
|
607 |
|
|
basic_block par_entry_bb, par_exit_bb;
|
608 |
|
|
basic_block ws_entry_bb, ws_exit_bb;
|
609 |
|
|
|
610 |
|
|
if (region == NULL || region->inner == NULL
|
611 |
|
|
|| region->exit == NULL || region->inner->exit == NULL
|
612 |
|
|
|| region->inner->cont == NULL)
|
613 |
|
|
return;
|
614 |
|
|
|
615 |
|
|
/* We only support parallel+for and parallel+sections. */
|
616 |
|
|
if (region->type != GIMPLE_OMP_PARALLEL
|
617 |
|
|
|| (region->inner->type != GIMPLE_OMP_FOR
|
618 |
|
|
&& region->inner->type != GIMPLE_OMP_SECTIONS))
|
619 |
|
|
return;
|
620 |
|
|
|
621 |
|
|
/* Check for perfect nesting PAR_ENTRY_BB -> WS_ENTRY_BB and
|
622 |
|
|
WS_EXIT_BB -> PAR_EXIT_BB. */
|
623 |
|
|
par_entry_bb = region->entry;
|
624 |
|
|
par_exit_bb = region->exit;
|
625 |
|
|
ws_entry_bb = region->inner->entry;
|
626 |
|
|
ws_exit_bb = region->inner->exit;
|
627 |
|
|
|
628 |
|
|
if (single_succ (par_entry_bb) == ws_entry_bb
|
629 |
|
|
&& single_succ (ws_exit_bb) == par_exit_bb
|
630 |
|
|
&& workshare_safe_to_combine_p (ws_entry_bb)
|
631 |
|
|
&& (gimple_omp_parallel_combined_p (last_stmt (par_entry_bb))
|
632 |
|
|
|| (last_and_only_stmt (ws_entry_bb)
|
633 |
|
|
&& last_and_only_stmt (par_exit_bb))))
|
634 |
|
|
{
|
635 |
|
|
gimple ws_stmt = last_stmt (ws_entry_bb);
|
636 |
|
|
|
637 |
|
|
if (region->inner->type == GIMPLE_OMP_FOR)
|
638 |
|
|
{
|
639 |
|
|
/* If this is a combined parallel loop, we need to determine
|
640 |
|
|
whether or not to use the combined library calls. There
|
641 |
|
|
are two cases where we do not apply the transformation:
|
642 |
|
|
static loops and any kind of ordered loop. In the first
|
643 |
|
|
case, we already open code the loop so there is no need
|
644 |
|
|
to do anything else. In the latter case, the combined
|
645 |
|
|
parallel loop call would still need extra synchronization
|
646 |
|
|
to implement ordered semantics, so there would not be any
|
647 |
|
|
gain in using the combined call. */
|
648 |
|
|
tree clauses = gimple_omp_for_clauses (ws_stmt);
|
649 |
|
|
tree c = find_omp_clause (clauses, OMP_CLAUSE_SCHEDULE);
|
650 |
|
|
if (c == NULL
|
651 |
|
|
|| OMP_CLAUSE_SCHEDULE_KIND (c) == OMP_CLAUSE_SCHEDULE_STATIC
|
652 |
|
|
|| find_omp_clause (clauses, OMP_CLAUSE_ORDERED))
|
653 |
|
|
{
|
654 |
|
|
region->is_combined_parallel = false;
|
655 |
|
|
region->inner->is_combined_parallel = false;
|
656 |
|
|
return;
|
657 |
|
|
}
|
658 |
|
|
}
|
659 |
|
|
|
660 |
|
|
region->is_combined_parallel = true;
|
661 |
|
|
region->inner->is_combined_parallel = true;
|
662 |
|
|
region->ws_args = get_ws_args_for (ws_stmt);
|
663 |
|
|
}
|
664 |
|
|
}
|
665 |
|
|
|
666 |
|
|
|
667 |
|
|
/* Return true if EXPR is variable sized. */
|
668 |
|
|
|
669 |
|
|
static inline bool
|
670 |
|
|
is_variable_sized (const_tree expr)
|
671 |
|
|
{
|
672 |
|
|
return !TREE_CONSTANT (TYPE_SIZE_UNIT (TREE_TYPE (expr)));
|
673 |
|
|
}
|
674 |
|
|
|
675 |
|
|
/* Return true if DECL is a reference type. */
|
676 |
|
|
|
677 |
|
|
static inline bool
|
678 |
|
|
is_reference (tree decl)
|
679 |
|
|
{
|
680 |
|
|
return lang_hooks.decls.omp_privatize_by_reference (decl);
|
681 |
|
|
}
|
682 |
|
|
|
683 |
|
|
/* Lookup variables in the decl or field splay trees. The "maybe" form
|
684 |
|
|
allows for the variable form to not have been entered, otherwise we
|
685 |
|
|
assert that the variable must have been entered. */
|
686 |
|
|
|
687 |
|
|
static inline tree
|
688 |
|
|
lookup_decl (tree var, omp_context *ctx)
|
689 |
|
|
{
|
690 |
|
|
tree *n;
|
691 |
|
|
n = (tree *) pointer_map_contains (ctx->cb.decl_map, var);
|
692 |
|
|
return *n;
|
693 |
|
|
}
|
694 |
|
|
|
695 |
|
|
static inline tree
|
696 |
|
|
maybe_lookup_decl (const_tree var, omp_context *ctx)
|
697 |
|
|
{
|
698 |
|
|
tree *n;
|
699 |
|
|
n = (tree *) pointer_map_contains (ctx->cb.decl_map, var);
|
700 |
|
|
return n ? *n : NULL_TREE;
|
701 |
|
|
}
|
702 |
|
|
|
703 |
|
|
static inline tree
|
704 |
|
|
lookup_field (tree var, omp_context *ctx)
|
705 |
|
|
{
|
706 |
|
|
splay_tree_node n;
|
707 |
|
|
n = splay_tree_lookup (ctx->field_map, (splay_tree_key) var);
|
708 |
|
|
return (tree) n->value;
|
709 |
|
|
}
|
710 |
|
|
|
711 |
|
|
static inline tree
|
712 |
|
|
lookup_sfield (tree var, omp_context *ctx)
|
713 |
|
|
{
|
714 |
|
|
splay_tree_node n;
|
715 |
|
|
n = splay_tree_lookup (ctx->sfield_map
|
716 |
|
|
? ctx->sfield_map : ctx->field_map,
|
717 |
|
|
(splay_tree_key) var);
|
718 |
|
|
return (tree) n->value;
|
719 |
|
|
}
|
720 |
|
|
|
721 |
|
|
static inline tree
|
722 |
|
|
maybe_lookup_field (tree var, omp_context *ctx)
|
723 |
|
|
{
|
724 |
|
|
splay_tree_node n;
|
725 |
|
|
n = splay_tree_lookup (ctx->field_map, (splay_tree_key) var);
|
726 |
|
|
return n ? (tree) n->value : NULL_TREE;
|
727 |
|
|
}
|
728 |
|
|
|
729 |
|
|
/* Return true if DECL should be copied by pointer. SHARED_CTX is
|
730 |
|
|
the parallel context if DECL is to be shared. */
|
731 |
|
|
|
732 |
|
|
static bool
|
733 |
|
|
use_pointer_for_field (tree decl, omp_context *shared_ctx)
|
734 |
|
|
{
|
735 |
|
|
if (AGGREGATE_TYPE_P (TREE_TYPE (decl)))
|
736 |
|
|
return true;
|
737 |
|
|
|
738 |
|
|
/* We can only use copy-in/copy-out semantics for shared variables
|
739 |
|
|
when we know the value is not accessible from an outer scope. */
|
740 |
|
|
if (shared_ctx)
|
741 |
|
|
{
|
742 |
|
|
/* ??? Trivially accessible from anywhere. But why would we even
|
743 |
|
|
be passing an address in this case? Should we simply assert
|
744 |
|
|
this to be false, or should we have a cleanup pass that removes
|
745 |
|
|
these from the list of mappings? */
|
746 |
|
|
if (TREE_STATIC (decl) || DECL_EXTERNAL (decl))
|
747 |
|
|
return true;
|
748 |
|
|
|
749 |
|
|
/* For variables with DECL_HAS_VALUE_EXPR_P set, we cannot tell
|
750 |
|
|
without analyzing the expression whether or not its location
|
751 |
|
|
is accessible to anyone else. In the case of nested parallel
|
752 |
|
|
regions it certainly may be. */
|
753 |
|
|
if (TREE_CODE (decl) != RESULT_DECL && DECL_HAS_VALUE_EXPR_P (decl))
|
754 |
|
|
return true;
|
755 |
|
|
|
756 |
|
|
/* Do not use copy-in/copy-out for variables that have their
|
757 |
|
|
address taken. */
|
758 |
|
|
if (TREE_ADDRESSABLE (decl))
|
759 |
|
|
return true;
|
760 |
|
|
|
761 |
|
|
/* Disallow copy-in/out in nested parallel if
|
762 |
|
|
decl is shared in outer parallel, otherwise
|
763 |
|
|
each thread could store the shared variable
|
764 |
|
|
in its own copy-in location, making the
|
765 |
|
|
variable no longer really shared. */
|
766 |
|
|
if (!TREE_READONLY (decl) && shared_ctx->is_nested)
|
767 |
|
|
{
|
768 |
|
|
omp_context *up;
|
769 |
|
|
|
770 |
|
|
for (up = shared_ctx->outer; up; up = up->outer)
|
771 |
|
|
if (is_taskreg_ctx (up) && maybe_lookup_decl (decl, up))
|
772 |
|
|
break;
|
773 |
|
|
|
774 |
|
|
if (up)
|
775 |
|
|
{
|
776 |
|
|
tree c;
|
777 |
|
|
|
778 |
|
|
for (c = gimple_omp_taskreg_clauses (up->stmt);
|
779 |
|
|
c; c = OMP_CLAUSE_CHAIN (c))
|
780 |
|
|
if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_SHARED
|
781 |
|
|
&& OMP_CLAUSE_DECL (c) == decl)
|
782 |
|
|
break;
|
783 |
|
|
|
784 |
|
|
if (c)
|
785 |
|
|
goto maybe_mark_addressable_and_ret;
|
786 |
|
|
}
|
787 |
|
|
}
|
788 |
|
|
|
789 |
|
|
/* For tasks avoid using copy-in/out, unless they are readonly
|
790 |
|
|
(in which case just copy-in is used). As tasks can be
|
791 |
|
|
deferred or executed in different thread, when GOMP_task
|
792 |
|
|
returns, the task hasn't necessarily terminated. */
|
793 |
|
|
if (!TREE_READONLY (decl) && is_task_ctx (shared_ctx))
|
794 |
|
|
{
|
795 |
|
|
tree outer;
|
796 |
|
|
maybe_mark_addressable_and_ret:
|
797 |
|
|
outer = maybe_lookup_decl_in_outer_ctx (decl, shared_ctx);
|
798 |
|
|
if (is_gimple_reg (outer))
|
799 |
|
|
{
|
800 |
|
|
/* Taking address of OUTER in lower_send_shared_vars
|
801 |
|
|
might need regimplification of everything that uses the
|
802 |
|
|
variable. */
|
803 |
|
|
if (!task_shared_vars)
|
804 |
|
|
task_shared_vars = BITMAP_ALLOC (NULL);
|
805 |
|
|
bitmap_set_bit (task_shared_vars, DECL_UID (outer));
|
806 |
|
|
TREE_ADDRESSABLE (outer) = 1;
|
807 |
|
|
}
|
808 |
|
|
return true;
|
809 |
|
|
}
|
810 |
|
|
}
|
811 |
|
|
|
812 |
|
|
return false;
|
813 |
|
|
}
|
814 |
|
|
|
815 |
|
|
/* Create a new VAR_DECL and copy information from VAR to it. */
|
816 |
|
|
|
817 |
|
|
tree
|
818 |
|
|
copy_var_decl (tree var, tree name, tree type)
|
819 |
|
|
{
|
820 |
|
|
tree copy = build_decl (DECL_SOURCE_LOCATION (var), VAR_DECL, name, type);
|
821 |
|
|
|
822 |
|
|
TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (var);
|
823 |
|
|
TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (var);
|
824 |
|
|
DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (var);
|
825 |
|
|
DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (var);
|
826 |
|
|
DECL_IGNORED_P (copy) = DECL_IGNORED_P (var);
|
827 |
|
|
DECL_CONTEXT (copy) = DECL_CONTEXT (var);
|
828 |
|
|
TREE_USED (copy) = 1;
|
829 |
|
|
DECL_SEEN_IN_BIND_EXPR_P (copy) = 1;
|
830 |
|
|
|
831 |
|
|
return copy;
|
832 |
|
|
}
|
833 |
|
|
|
834 |
|
|
/* Construct a new automatic decl similar to VAR. */
|
835 |
|
|
|
836 |
|
|
static tree
|
837 |
|
|
omp_copy_decl_2 (tree var, tree name, tree type, omp_context *ctx)
|
838 |
|
|
{
|
839 |
|
|
tree copy = copy_var_decl (var, name, type);
|
840 |
|
|
|
841 |
|
|
DECL_CONTEXT (copy) = current_function_decl;
|
842 |
|
|
DECL_CHAIN (copy) = ctx->block_vars;
|
843 |
|
|
ctx->block_vars = copy;
|
844 |
|
|
|
845 |
|
|
return copy;
|
846 |
|
|
}
|
847 |
|
|
|
848 |
|
|
static tree
|
849 |
|
|
omp_copy_decl_1 (tree var, omp_context *ctx)
|
850 |
|
|
{
|
851 |
|
|
return omp_copy_decl_2 (var, DECL_NAME (var), TREE_TYPE (var), ctx);
|
852 |
|
|
}
|
853 |
|
|
|
854 |
|
|
/* Build tree nodes to access the field for VAR on the receiver side. */
|
855 |
|
|
|
856 |
|
|
static tree
|
857 |
|
|
build_receiver_ref (tree var, bool by_ref, omp_context *ctx)
|
858 |
|
|
{
|
859 |
|
|
tree x, field = lookup_field (var, ctx);
|
860 |
|
|
|
861 |
|
|
/* If the receiver record type was remapped in the child function,
|
862 |
|
|
remap the field into the new record type. */
|
863 |
|
|
x = maybe_lookup_field (field, ctx);
|
864 |
|
|
if (x != NULL)
|
865 |
|
|
field = x;
|
866 |
|
|
|
867 |
|
|
x = build_simple_mem_ref (ctx->receiver_decl);
|
868 |
|
|
x = build3 (COMPONENT_REF, TREE_TYPE (field), x, field, NULL);
|
869 |
|
|
if (by_ref)
|
870 |
|
|
x = build_simple_mem_ref (x);
|
871 |
|
|
|
872 |
|
|
return x;
|
873 |
|
|
}
|
874 |
|
|
|
875 |
|
|
/* Build tree nodes to access VAR in the scope outer to CTX. In the case
|
876 |
|
|
of a parallel, this is a component reference; for workshare constructs
|
877 |
|
|
this is some variable. */
|
878 |
|
|
|
879 |
|
|
static tree
|
880 |
|
|
build_outer_var_ref (tree var, omp_context *ctx)
|
881 |
|
|
{
|
882 |
|
|
tree x;
|
883 |
|
|
|
884 |
|
|
if (is_global_var (maybe_lookup_decl_in_outer_ctx (var, ctx)))
|
885 |
|
|
x = var;
|
886 |
|
|
else if (is_variable_sized (var))
|
887 |
|
|
{
|
888 |
|
|
x = TREE_OPERAND (DECL_VALUE_EXPR (var), 0);
|
889 |
|
|
x = build_outer_var_ref (x, ctx);
|
890 |
|
|
x = build_simple_mem_ref (x);
|
891 |
|
|
}
|
892 |
|
|
else if (is_taskreg_ctx (ctx))
|
893 |
|
|
{
|
894 |
|
|
bool by_ref = use_pointer_for_field (var, NULL);
|
895 |
|
|
x = build_receiver_ref (var, by_ref, ctx);
|
896 |
|
|
}
|
897 |
|
|
else if (ctx->outer)
|
898 |
|
|
x = lookup_decl (var, ctx->outer);
|
899 |
|
|
else if (is_reference (var))
|
900 |
|
|
/* This can happen with orphaned constructs. If var is reference, it is
|
901 |
|
|
possible it is shared and as such valid. */
|
902 |
|
|
x = var;
|
903 |
|
|
else
|
904 |
|
|
gcc_unreachable ();
|
905 |
|
|
|
906 |
|
|
if (is_reference (var))
|
907 |
|
|
x = build_simple_mem_ref (x);
|
908 |
|
|
|
909 |
|
|
return x;
|
910 |
|
|
}
|
911 |
|
|
|
912 |
|
|
/* Build tree nodes to access the field for VAR on the sender side. */
|
913 |
|
|
|
914 |
|
|
static tree
|
915 |
|
|
build_sender_ref (tree var, omp_context *ctx)
|
916 |
|
|
{
|
917 |
|
|
tree field = lookup_sfield (var, ctx);
|
918 |
|
|
return build3 (COMPONENT_REF, TREE_TYPE (field),
|
919 |
|
|
ctx->sender_decl, field, NULL);
|
920 |
|
|
}
|
921 |
|
|
|
922 |
|
|
/* Add a new field for VAR inside the structure CTX->SENDER_DECL. */
|
923 |
|
|
|
924 |
|
|
static void
|
925 |
|
|
install_var_field (tree var, bool by_ref, int mask, omp_context *ctx)
|
926 |
|
|
{
|
927 |
|
|
tree field, type, sfield = NULL_TREE;
|
928 |
|
|
|
929 |
|
|
gcc_assert ((mask & 1) == 0
|
930 |
|
|
|| !splay_tree_lookup (ctx->field_map, (splay_tree_key) var));
|
931 |
|
|
gcc_assert ((mask & 2) == 0 || !ctx->sfield_map
|
932 |
|
|
|| !splay_tree_lookup (ctx->sfield_map, (splay_tree_key) var));
|
933 |
|
|
|
934 |
|
|
type = TREE_TYPE (var);
|
935 |
|
|
if (by_ref)
|
936 |
|
|
type = build_pointer_type (type);
|
937 |
|
|
else if ((mask & 3) == 1 && is_reference (var))
|
938 |
|
|
type = TREE_TYPE (type);
|
939 |
|
|
|
940 |
|
|
field = build_decl (DECL_SOURCE_LOCATION (var),
|
941 |
|
|
FIELD_DECL, DECL_NAME (var), type);
|
942 |
|
|
|
943 |
|
|
/* Remember what variable this field was created for. This does have a
|
944 |
|
|
side effect of making dwarf2out ignore this member, so for helpful
|
945 |
|
|
debugging we clear it later in delete_omp_context. */
|
946 |
|
|
DECL_ABSTRACT_ORIGIN (field) = var;
|
947 |
|
|
if (type == TREE_TYPE (var))
|
948 |
|
|
{
|
949 |
|
|
DECL_ALIGN (field) = DECL_ALIGN (var);
|
950 |
|
|
DECL_USER_ALIGN (field) = DECL_USER_ALIGN (var);
|
951 |
|
|
TREE_THIS_VOLATILE (field) = TREE_THIS_VOLATILE (var);
|
952 |
|
|
}
|
953 |
|
|
else
|
954 |
|
|
DECL_ALIGN (field) = TYPE_ALIGN (type);
|
955 |
|
|
|
956 |
|
|
if ((mask & 3) == 3)
|
957 |
|
|
{
|
958 |
|
|
insert_field_into_struct (ctx->record_type, field);
|
959 |
|
|
if (ctx->srecord_type)
|
960 |
|
|
{
|
961 |
|
|
sfield = build_decl (DECL_SOURCE_LOCATION (var),
|
962 |
|
|
FIELD_DECL, DECL_NAME (var), type);
|
963 |
|
|
DECL_ABSTRACT_ORIGIN (sfield) = var;
|
964 |
|
|
DECL_ALIGN (sfield) = DECL_ALIGN (field);
|
965 |
|
|
DECL_USER_ALIGN (sfield) = DECL_USER_ALIGN (field);
|
966 |
|
|
TREE_THIS_VOLATILE (sfield) = TREE_THIS_VOLATILE (field);
|
967 |
|
|
insert_field_into_struct (ctx->srecord_type, sfield);
|
968 |
|
|
}
|
969 |
|
|
}
|
970 |
|
|
else
|
971 |
|
|
{
|
972 |
|
|
if (ctx->srecord_type == NULL_TREE)
|
973 |
|
|
{
|
974 |
|
|
tree t;
|
975 |
|
|
|
976 |
|
|
ctx->srecord_type = lang_hooks.types.make_type (RECORD_TYPE);
|
977 |
|
|
ctx->sfield_map = splay_tree_new (splay_tree_compare_pointers, 0, 0);
|
978 |
|
|
for (t = TYPE_FIELDS (ctx->record_type); t ; t = TREE_CHAIN (t))
|
979 |
|
|
{
|
980 |
|
|
sfield = build_decl (DECL_SOURCE_LOCATION (var),
|
981 |
|
|
FIELD_DECL, DECL_NAME (t), TREE_TYPE (t));
|
982 |
|
|
DECL_ABSTRACT_ORIGIN (sfield) = DECL_ABSTRACT_ORIGIN (t);
|
983 |
|
|
insert_field_into_struct (ctx->srecord_type, sfield);
|
984 |
|
|
splay_tree_insert (ctx->sfield_map,
|
985 |
|
|
(splay_tree_key) DECL_ABSTRACT_ORIGIN (t),
|
986 |
|
|
(splay_tree_value) sfield);
|
987 |
|
|
}
|
988 |
|
|
}
|
989 |
|
|
sfield = field;
|
990 |
|
|
insert_field_into_struct ((mask & 1) ? ctx->record_type
|
991 |
|
|
: ctx->srecord_type, field);
|
992 |
|
|
}
|
993 |
|
|
|
994 |
|
|
if (mask & 1)
|
995 |
|
|
splay_tree_insert (ctx->field_map, (splay_tree_key) var,
|
996 |
|
|
(splay_tree_value) field);
|
997 |
|
|
if ((mask & 2) && ctx->sfield_map)
|
998 |
|
|
splay_tree_insert (ctx->sfield_map, (splay_tree_key) var,
|
999 |
|
|
(splay_tree_value) sfield);
|
1000 |
|
|
}
|
1001 |
|
|
|
1002 |
|
|
static tree
|
1003 |
|
|
install_var_local (tree var, omp_context *ctx)
|
1004 |
|
|
{
|
1005 |
|
|
tree new_var = omp_copy_decl_1 (var, ctx);
|
1006 |
|
|
insert_decl_map (&ctx->cb, var, new_var);
|
1007 |
|
|
return new_var;
|
1008 |
|
|
}
|
1009 |
|
|
|
1010 |
|
|
/* Adjust the replacement for DECL in CTX for the new context. This means
|
1011 |
|
|
copying the DECL_VALUE_EXPR, and fixing up the type. */
|
1012 |
|
|
|
1013 |
|
|
static void
|
1014 |
|
|
fixup_remapped_decl (tree decl, omp_context *ctx, bool private_debug)
|
1015 |
|
|
{
|
1016 |
|
|
tree new_decl, size;
|
1017 |
|
|
|
1018 |
|
|
new_decl = lookup_decl (decl, ctx);
|
1019 |
|
|
|
1020 |
|
|
TREE_TYPE (new_decl) = remap_type (TREE_TYPE (decl), &ctx->cb);
|
1021 |
|
|
|
1022 |
|
|
if ((!TREE_CONSTANT (DECL_SIZE (new_decl)) || private_debug)
|
1023 |
|
|
&& DECL_HAS_VALUE_EXPR_P (decl))
|
1024 |
|
|
{
|
1025 |
|
|
tree ve = DECL_VALUE_EXPR (decl);
|
1026 |
|
|
walk_tree (&ve, copy_tree_body_r, &ctx->cb, NULL);
|
1027 |
|
|
SET_DECL_VALUE_EXPR (new_decl, ve);
|
1028 |
|
|
DECL_HAS_VALUE_EXPR_P (new_decl) = 1;
|
1029 |
|
|
}
|
1030 |
|
|
|
1031 |
|
|
if (!TREE_CONSTANT (DECL_SIZE (new_decl)))
|
1032 |
|
|
{
|
1033 |
|
|
size = remap_decl (DECL_SIZE (decl), &ctx->cb);
|
1034 |
|
|
if (size == error_mark_node)
|
1035 |
|
|
size = TYPE_SIZE (TREE_TYPE (new_decl));
|
1036 |
|
|
DECL_SIZE (new_decl) = size;
|
1037 |
|
|
|
1038 |
|
|
size = remap_decl (DECL_SIZE_UNIT (decl), &ctx->cb);
|
1039 |
|
|
if (size == error_mark_node)
|
1040 |
|
|
size = TYPE_SIZE_UNIT (TREE_TYPE (new_decl));
|
1041 |
|
|
DECL_SIZE_UNIT (new_decl) = size;
|
1042 |
|
|
}
|
1043 |
|
|
}
|
1044 |
|
|
|
1045 |
|
|
/* The callback for remap_decl. Search all containing contexts for a
|
1046 |
|
|
mapping of the variable; this avoids having to duplicate the splay
|
1047 |
|
|
tree ahead of time. We know a mapping doesn't already exist in the
|
1048 |
|
|
given context. Create new mappings to implement default semantics. */
|
1049 |
|
|
|
1050 |
|
|
static tree
|
1051 |
|
|
omp_copy_decl (tree var, copy_body_data *cb)
|
1052 |
|
|
{
|
1053 |
|
|
omp_context *ctx = (omp_context *) cb;
|
1054 |
|
|
tree new_var;
|
1055 |
|
|
|
1056 |
|
|
if (TREE_CODE (var) == LABEL_DECL)
|
1057 |
|
|
{
|
1058 |
|
|
new_var = create_artificial_label (DECL_SOURCE_LOCATION (var));
|
1059 |
|
|
DECL_CONTEXT (new_var) = current_function_decl;
|
1060 |
|
|
insert_decl_map (&ctx->cb, var, new_var);
|
1061 |
|
|
return new_var;
|
1062 |
|
|
}
|
1063 |
|
|
|
1064 |
|
|
while (!is_taskreg_ctx (ctx))
|
1065 |
|
|
{
|
1066 |
|
|
ctx = ctx->outer;
|
1067 |
|
|
if (ctx == NULL)
|
1068 |
|
|
return var;
|
1069 |
|
|
new_var = maybe_lookup_decl (var, ctx);
|
1070 |
|
|
if (new_var)
|
1071 |
|
|
return new_var;
|
1072 |
|
|
}
|
1073 |
|
|
|
1074 |
|
|
if (is_global_var (var) || decl_function_context (var) != ctx->cb.src_fn)
|
1075 |
|
|
return var;
|
1076 |
|
|
|
1077 |
|
|
return error_mark_node;
|
1078 |
|
|
}
|
1079 |
|
|
|
1080 |
|
|
|
1081 |
|
|
/* Return the parallel region associated with STMT. */
|
1082 |
|
|
|
1083 |
|
|
/* Debugging dumps for parallel regions. */
|
1084 |
|
|
void dump_omp_region (FILE *, struct omp_region *, int);
|
1085 |
|
|
void debug_omp_region (struct omp_region *);
|
1086 |
|
|
void debug_all_omp_regions (void);
|
1087 |
|
|
|
1088 |
|
|
/* Dump the parallel region tree rooted at REGION. */
|
1089 |
|
|
|
1090 |
|
|
void
|
1091 |
|
|
dump_omp_region (FILE *file, struct omp_region *region, int indent)
|
1092 |
|
|
{
|
1093 |
|
|
fprintf (file, "%*sbb %d: %s\n", indent, "", region->entry->index,
|
1094 |
|
|
gimple_code_name[region->type]);
|
1095 |
|
|
|
1096 |
|
|
if (region->inner)
|
1097 |
|
|
dump_omp_region (file, region->inner, indent + 4);
|
1098 |
|
|
|
1099 |
|
|
if (region->cont)
|
1100 |
|
|
{
|
1101 |
|
|
fprintf (file, "%*sbb %d: GIMPLE_OMP_CONTINUE\n", indent, "",
|
1102 |
|
|
region->cont->index);
|
1103 |
|
|
}
|
1104 |
|
|
|
1105 |
|
|
if (region->exit)
|
1106 |
|
|
fprintf (file, "%*sbb %d: GIMPLE_OMP_RETURN\n", indent, "",
|
1107 |
|
|
region->exit->index);
|
1108 |
|
|
else
|
1109 |
|
|
fprintf (file, "%*s[no exit marker]\n", indent, "");
|
1110 |
|
|
|
1111 |
|
|
if (region->next)
|
1112 |
|
|
dump_omp_region (file, region->next, indent);
|
1113 |
|
|
}
|
1114 |
|
|
|
1115 |
|
|
DEBUG_FUNCTION void
|
1116 |
|
|
debug_omp_region (struct omp_region *region)
|
1117 |
|
|
{
|
1118 |
|
|
dump_omp_region (stderr, region, 0);
|
1119 |
|
|
}
|
1120 |
|
|
|
1121 |
|
|
DEBUG_FUNCTION void
|
1122 |
|
|
debug_all_omp_regions (void)
|
1123 |
|
|
{
|
1124 |
|
|
dump_omp_region (stderr, root_omp_region, 0);
|
1125 |
|
|
}
|
1126 |
|
|
|
1127 |
|
|
|
1128 |
|
|
/* Create a new parallel region starting at STMT inside region PARENT. */
|
1129 |
|
|
|
1130 |
|
|
struct omp_region *
|
1131 |
|
|
new_omp_region (basic_block bb, enum gimple_code type,
|
1132 |
|
|
struct omp_region *parent)
|
1133 |
|
|
{
|
1134 |
|
|
struct omp_region *region = XCNEW (struct omp_region);
|
1135 |
|
|
|
1136 |
|
|
region->outer = parent;
|
1137 |
|
|
region->entry = bb;
|
1138 |
|
|
region->type = type;
|
1139 |
|
|
|
1140 |
|
|
if (parent)
|
1141 |
|
|
{
|
1142 |
|
|
/* This is a nested region. Add it to the list of inner
|
1143 |
|
|
regions in PARENT. */
|
1144 |
|
|
region->next = parent->inner;
|
1145 |
|
|
parent->inner = region;
|
1146 |
|
|
}
|
1147 |
|
|
else
|
1148 |
|
|
{
|
1149 |
|
|
/* This is a toplevel region. Add it to the list of toplevel
|
1150 |
|
|
regions in ROOT_OMP_REGION. */
|
1151 |
|
|
region->next = root_omp_region;
|
1152 |
|
|
root_omp_region = region;
|
1153 |
|
|
}
|
1154 |
|
|
|
1155 |
|
|
return region;
|
1156 |
|
|
}
|
1157 |
|
|
|
1158 |
|
|
/* Release the memory associated with the region tree rooted at REGION. */
|
1159 |
|
|
|
1160 |
|
|
static void
|
1161 |
|
|
free_omp_region_1 (struct omp_region *region)
|
1162 |
|
|
{
|
1163 |
|
|
struct omp_region *i, *n;
|
1164 |
|
|
|
1165 |
|
|
for (i = region->inner; i ; i = n)
|
1166 |
|
|
{
|
1167 |
|
|
n = i->next;
|
1168 |
|
|
free_omp_region_1 (i);
|
1169 |
|
|
}
|
1170 |
|
|
|
1171 |
|
|
free (region);
|
1172 |
|
|
}
|
1173 |
|
|
|
1174 |
|
|
/* Release the memory for the entire omp region tree. */
|
1175 |
|
|
|
1176 |
|
|
void
|
1177 |
|
|
free_omp_regions (void)
|
1178 |
|
|
{
|
1179 |
|
|
struct omp_region *r, *n;
|
1180 |
|
|
for (r = root_omp_region; r ; r = n)
|
1181 |
|
|
{
|
1182 |
|
|
n = r->next;
|
1183 |
|
|
free_omp_region_1 (r);
|
1184 |
|
|
}
|
1185 |
|
|
root_omp_region = NULL;
|
1186 |
|
|
}
|
1187 |
|
|
|
1188 |
|
|
|
1189 |
|
|
/* Create a new context, with OUTER_CTX being the surrounding context. */
|
1190 |
|
|
|
1191 |
|
|
static omp_context *
|
1192 |
|
|
new_omp_context (gimple stmt, omp_context *outer_ctx)
|
1193 |
|
|
{
|
1194 |
|
|
omp_context *ctx = XCNEW (omp_context);
|
1195 |
|
|
|
1196 |
|
|
splay_tree_insert (all_contexts, (splay_tree_key) stmt,
|
1197 |
|
|
(splay_tree_value) ctx);
|
1198 |
|
|
ctx->stmt = stmt;
|
1199 |
|
|
|
1200 |
|
|
if (outer_ctx)
|
1201 |
|
|
{
|
1202 |
|
|
ctx->outer = outer_ctx;
|
1203 |
|
|
ctx->cb = outer_ctx->cb;
|
1204 |
|
|
ctx->cb.block = NULL;
|
1205 |
|
|
ctx->depth = outer_ctx->depth + 1;
|
1206 |
|
|
}
|
1207 |
|
|
else
|
1208 |
|
|
{
|
1209 |
|
|
ctx->cb.src_fn = current_function_decl;
|
1210 |
|
|
ctx->cb.dst_fn = current_function_decl;
|
1211 |
|
|
ctx->cb.src_node = cgraph_get_node (current_function_decl);
|
1212 |
|
|
gcc_checking_assert (ctx->cb.src_node);
|
1213 |
|
|
ctx->cb.dst_node = ctx->cb.src_node;
|
1214 |
|
|
ctx->cb.src_cfun = cfun;
|
1215 |
|
|
ctx->cb.copy_decl = omp_copy_decl;
|
1216 |
|
|
ctx->cb.eh_lp_nr = 0;
|
1217 |
|
|
ctx->cb.transform_call_graph_edges = CB_CGE_MOVE;
|
1218 |
|
|
ctx->depth = 1;
|
1219 |
|
|
}
|
1220 |
|
|
|
1221 |
|
|
ctx->cb.decl_map = pointer_map_create ();
|
1222 |
|
|
|
1223 |
|
|
return ctx;
|
1224 |
|
|
}
|
1225 |
|
|
|
1226 |
|
|
static gimple_seq maybe_catch_exception (gimple_seq);
|
1227 |
|
|
|
1228 |
|
|
/* Finalize task copyfn. */
|
1229 |
|
|
|
1230 |
|
|
static void
|
1231 |
|
|
finalize_task_copyfn (gimple task_stmt)
|
1232 |
|
|
{
|
1233 |
|
|
struct function *child_cfun;
|
1234 |
|
|
tree child_fn, old_fn;
|
1235 |
|
|
gimple_seq seq, new_seq;
|
1236 |
|
|
gimple bind;
|
1237 |
|
|
|
1238 |
|
|
child_fn = gimple_omp_task_copy_fn (task_stmt);
|
1239 |
|
|
if (child_fn == NULL_TREE)
|
1240 |
|
|
return;
|
1241 |
|
|
|
1242 |
|
|
child_cfun = DECL_STRUCT_FUNCTION (child_fn);
|
1243 |
|
|
|
1244 |
|
|
/* Inform the callgraph about the new function. */
|
1245 |
|
|
DECL_STRUCT_FUNCTION (child_fn)->curr_properties
|
1246 |
|
|
= cfun->curr_properties;
|
1247 |
|
|
|
1248 |
|
|
old_fn = current_function_decl;
|
1249 |
|
|
push_cfun (child_cfun);
|
1250 |
|
|
current_function_decl = child_fn;
|
1251 |
|
|
bind = gimplify_body (child_fn, false);
|
1252 |
|
|
seq = gimple_seq_alloc ();
|
1253 |
|
|
gimple_seq_add_stmt (&seq, bind);
|
1254 |
|
|
new_seq = maybe_catch_exception (seq);
|
1255 |
|
|
if (new_seq != seq)
|
1256 |
|
|
{
|
1257 |
|
|
bind = gimple_build_bind (NULL, new_seq, NULL);
|
1258 |
|
|
seq = gimple_seq_alloc ();
|
1259 |
|
|
gimple_seq_add_stmt (&seq, bind);
|
1260 |
|
|
}
|
1261 |
|
|
gimple_set_body (child_fn, seq);
|
1262 |
|
|
pop_cfun ();
|
1263 |
|
|
current_function_decl = old_fn;
|
1264 |
|
|
|
1265 |
|
|
cgraph_add_new_function (child_fn, false);
|
1266 |
|
|
}
|
1267 |
|
|
|
1268 |
|
|
/* Destroy a omp_context data structures. Called through the splay tree
|
1269 |
|
|
value delete callback. */
|
1270 |
|
|
|
1271 |
|
|
static void
|
1272 |
|
|
delete_omp_context (splay_tree_value value)
|
1273 |
|
|
{
|
1274 |
|
|
omp_context *ctx = (omp_context *) value;
|
1275 |
|
|
|
1276 |
|
|
pointer_map_destroy (ctx->cb.decl_map);
|
1277 |
|
|
|
1278 |
|
|
if (ctx->field_map)
|
1279 |
|
|
splay_tree_delete (ctx->field_map);
|
1280 |
|
|
if (ctx->sfield_map)
|
1281 |
|
|
splay_tree_delete (ctx->sfield_map);
|
1282 |
|
|
|
1283 |
|
|
/* We hijacked DECL_ABSTRACT_ORIGIN earlier. We need to clear it before
|
1284 |
|
|
it produces corrupt debug information. */
|
1285 |
|
|
if (ctx->record_type)
|
1286 |
|
|
{
|
1287 |
|
|
tree t;
|
1288 |
|
|
for (t = TYPE_FIELDS (ctx->record_type); t ; t = DECL_CHAIN (t))
|
1289 |
|
|
DECL_ABSTRACT_ORIGIN (t) = NULL;
|
1290 |
|
|
}
|
1291 |
|
|
if (ctx->srecord_type)
|
1292 |
|
|
{
|
1293 |
|
|
tree t;
|
1294 |
|
|
for (t = TYPE_FIELDS (ctx->srecord_type); t ; t = DECL_CHAIN (t))
|
1295 |
|
|
DECL_ABSTRACT_ORIGIN (t) = NULL;
|
1296 |
|
|
}
|
1297 |
|
|
|
1298 |
|
|
if (is_task_ctx (ctx))
|
1299 |
|
|
finalize_task_copyfn (ctx->stmt);
|
1300 |
|
|
|
1301 |
|
|
XDELETE (ctx);
|
1302 |
|
|
}
|
1303 |
|
|
|
1304 |
|
|
/* Fix up RECEIVER_DECL with a type that has been remapped to the child
|
1305 |
|
|
context. */
|
1306 |
|
|
|
1307 |
|
|
static void
|
1308 |
|
|
fixup_child_record_type (omp_context *ctx)
|
1309 |
|
|
{
|
1310 |
|
|
tree f, type = ctx->record_type;
|
1311 |
|
|
|
1312 |
|
|
/* ??? It isn't sufficient to just call remap_type here, because
|
1313 |
|
|
variably_modified_type_p doesn't work the way we expect for
|
1314 |
|
|
record types. Testing each field for whether it needs remapping
|
1315 |
|
|
and creating a new record by hand works, however. */
|
1316 |
|
|
for (f = TYPE_FIELDS (type); f ; f = DECL_CHAIN (f))
|
1317 |
|
|
if (variably_modified_type_p (TREE_TYPE (f), ctx->cb.src_fn))
|
1318 |
|
|
break;
|
1319 |
|
|
if (f)
|
1320 |
|
|
{
|
1321 |
|
|
tree name, new_fields = NULL;
|
1322 |
|
|
|
1323 |
|
|
type = lang_hooks.types.make_type (RECORD_TYPE);
|
1324 |
|
|
name = DECL_NAME (TYPE_NAME (ctx->record_type));
|
1325 |
|
|
name = build_decl (DECL_SOURCE_LOCATION (ctx->receiver_decl),
|
1326 |
|
|
TYPE_DECL, name, type);
|
1327 |
|
|
TYPE_NAME (type) = name;
|
1328 |
|
|
|
1329 |
|
|
for (f = TYPE_FIELDS (ctx->record_type); f ; f = DECL_CHAIN (f))
|
1330 |
|
|
{
|
1331 |
|
|
tree new_f = copy_node (f);
|
1332 |
|
|
DECL_CONTEXT (new_f) = type;
|
1333 |
|
|
TREE_TYPE (new_f) = remap_type (TREE_TYPE (f), &ctx->cb);
|
1334 |
|
|
DECL_CHAIN (new_f) = new_fields;
|
1335 |
|
|
walk_tree (&DECL_SIZE (new_f), copy_tree_body_r, &ctx->cb, NULL);
|
1336 |
|
|
walk_tree (&DECL_SIZE_UNIT (new_f), copy_tree_body_r,
|
1337 |
|
|
&ctx->cb, NULL);
|
1338 |
|
|
walk_tree (&DECL_FIELD_OFFSET (new_f), copy_tree_body_r,
|
1339 |
|
|
&ctx->cb, NULL);
|
1340 |
|
|
new_fields = new_f;
|
1341 |
|
|
|
1342 |
|
|
/* Arrange to be able to look up the receiver field
|
1343 |
|
|
given the sender field. */
|
1344 |
|
|
splay_tree_insert (ctx->field_map, (splay_tree_key) f,
|
1345 |
|
|
(splay_tree_value) new_f);
|
1346 |
|
|
}
|
1347 |
|
|
TYPE_FIELDS (type) = nreverse (new_fields);
|
1348 |
|
|
layout_type (type);
|
1349 |
|
|
}
|
1350 |
|
|
|
1351 |
|
|
TREE_TYPE (ctx->receiver_decl) = build_pointer_type (type);
|
1352 |
|
|
}
|
1353 |
|
|
|
1354 |
|
|
/* Instantiate decls as necessary in CTX to satisfy the data sharing
|
1355 |
|
|
specified by CLAUSES. */
|
1356 |
|
|
|
1357 |
|
|
static void
|
1358 |
|
|
scan_sharing_clauses (tree clauses, omp_context *ctx)
|
1359 |
|
|
{
|
1360 |
|
|
tree c, decl;
|
1361 |
|
|
bool scan_array_reductions = false;
|
1362 |
|
|
|
1363 |
|
|
for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c))
|
1364 |
|
|
{
|
1365 |
|
|
bool by_ref;
|
1366 |
|
|
|
1367 |
|
|
switch (OMP_CLAUSE_CODE (c))
|
1368 |
|
|
{
|
1369 |
|
|
case OMP_CLAUSE_PRIVATE:
|
1370 |
|
|
decl = OMP_CLAUSE_DECL (c);
|
1371 |
|
|
if (OMP_CLAUSE_PRIVATE_OUTER_REF (c))
|
1372 |
|
|
goto do_private;
|
1373 |
|
|
else if (!is_variable_sized (decl))
|
1374 |
|
|
install_var_local (decl, ctx);
|
1375 |
|
|
break;
|
1376 |
|
|
|
1377 |
|
|
case OMP_CLAUSE_SHARED:
|
1378 |
|
|
gcc_assert (is_taskreg_ctx (ctx));
|
1379 |
|
|
decl = OMP_CLAUSE_DECL (c);
|
1380 |
|
|
gcc_assert (!COMPLETE_TYPE_P (TREE_TYPE (decl))
|
1381 |
|
|
|| !is_variable_sized (decl));
|
1382 |
|
|
/* Global variables don't need to be copied,
|
1383 |
|
|
the receiver side will use them directly. */
|
1384 |
|
|
if (is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx)))
|
1385 |
|
|
break;
|
1386 |
|
|
by_ref = use_pointer_for_field (decl, ctx);
|
1387 |
|
|
if (! TREE_READONLY (decl)
|
1388 |
|
|
|| TREE_ADDRESSABLE (decl)
|
1389 |
|
|
|| by_ref
|
1390 |
|
|
|| is_reference (decl))
|
1391 |
|
|
{
|
1392 |
|
|
install_var_field (decl, by_ref, 3, ctx);
|
1393 |
|
|
install_var_local (decl, ctx);
|
1394 |
|
|
break;
|
1395 |
|
|
}
|
1396 |
|
|
/* We don't need to copy const scalar vars back. */
|
1397 |
|
|
OMP_CLAUSE_SET_CODE (c, OMP_CLAUSE_FIRSTPRIVATE);
|
1398 |
|
|
goto do_private;
|
1399 |
|
|
|
1400 |
|
|
case OMP_CLAUSE_LASTPRIVATE:
|
1401 |
|
|
/* Let the corresponding firstprivate clause create
|
1402 |
|
|
the variable. */
|
1403 |
|
|
if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c))
|
1404 |
|
|
break;
|
1405 |
|
|
/* FALLTHRU */
|
1406 |
|
|
|
1407 |
|
|
case OMP_CLAUSE_FIRSTPRIVATE:
|
1408 |
|
|
case OMP_CLAUSE_REDUCTION:
|
1409 |
|
|
decl = OMP_CLAUSE_DECL (c);
|
1410 |
|
|
do_private:
|
1411 |
|
|
if (is_variable_sized (decl))
|
1412 |
|
|
{
|
1413 |
|
|
if (is_task_ctx (ctx))
|
1414 |
|
|
install_var_field (decl, false, 1, ctx);
|
1415 |
|
|
break;
|
1416 |
|
|
}
|
1417 |
|
|
else if (is_taskreg_ctx (ctx))
|
1418 |
|
|
{
|
1419 |
|
|
bool global
|
1420 |
|
|
= is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx));
|
1421 |
|
|
by_ref = use_pointer_for_field (decl, NULL);
|
1422 |
|
|
|
1423 |
|
|
if (is_task_ctx (ctx)
|
1424 |
|
|
&& (global || by_ref || is_reference (decl)))
|
1425 |
|
|
{
|
1426 |
|
|
install_var_field (decl, false, 1, ctx);
|
1427 |
|
|
if (!global)
|
1428 |
|
|
install_var_field (decl, by_ref, 2, ctx);
|
1429 |
|
|
}
|
1430 |
|
|
else if (!global)
|
1431 |
|
|
install_var_field (decl, by_ref, 3, ctx);
|
1432 |
|
|
}
|
1433 |
|
|
install_var_local (decl, ctx);
|
1434 |
|
|
break;
|
1435 |
|
|
|
1436 |
|
|
case OMP_CLAUSE_COPYPRIVATE:
|
1437 |
|
|
case OMP_CLAUSE_COPYIN:
|
1438 |
|
|
decl = OMP_CLAUSE_DECL (c);
|
1439 |
|
|
by_ref = use_pointer_for_field (decl, NULL);
|
1440 |
|
|
install_var_field (decl, by_ref, 3, ctx);
|
1441 |
|
|
break;
|
1442 |
|
|
|
1443 |
|
|
case OMP_CLAUSE_DEFAULT:
|
1444 |
|
|
ctx->default_kind = OMP_CLAUSE_DEFAULT_KIND (c);
|
1445 |
|
|
break;
|
1446 |
|
|
|
1447 |
|
|
case OMP_CLAUSE_FINAL:
|
1448 |
|
|
case OMP_CLAUSE_IF:
|
1449 |
|
|
case OMP_CLAUSE_NUM_THREADS:
|
1450 |
|
|
case OMP_CLAUSE_SCHEDULE:
|
1451 |
|
|
if (ctx->outer)
|
1452 |
|
|
scan_omp_op (&OMP_CLAUSE_OPERAND (c, 0), ctx->outer);
|
1453 |
|
|
break;
|
1454 |
|
|
|
1455 |
|
|
case OMP_CLAUSE_NOWAIT:
|
1456 |
|
|
case OMP_CLAUSE_ORDERED:
|
1457 |
|
|
case OMP_CLAUSE_COLLAPSE:
|
1458 |
|
|
case OMP_CLAUSE_UNTIED:
|
1459 |
|
|
case OMP_CLAUSE_MERGEABLE:
|
1460 |
|
|
break;
|
1461 |
|
|
|
1462 |
|
|
default:
|
1463 |
|
|
gcc_unreachable ();
|
1464 |
|
|
}
|
1465 |
|
|
}
|
1466 |
|
|
|
1467 |
|
|
for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c))
|
1468 |
|
|
{
|
1469 |
|
|
switch (OMP_CLAUSE_CODE (c))
|
1470 |
|
|
{
|
1471 |
|
|
case OMP_CLAUSE_LASTPRIVATE:
|
1472 |
|
|
/* Let the corresponding firstprivate clause create
|
1473 |
|
|
the variable. */
|
1474 |
|
|
if (OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c))
|
1475 |
|
|
scan_array_reductions = true;
|
1476 |
|
|
if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c))
|
1477 |
|
|
break;
|
1478 |
|
|
/* FALLTHRU */
|
1479 |
|
|
|
1480 |
|
|
case OMP_CLAUSE_PRIVATE:
|
1481 |
|
|
case OMP_CLAUSE_FIRSTPRIVATE:
|
1482 |
|
|
case OMP_CLAUSE_REDUCTION:
|
1483 |
|
|
decl = OMP_CLAUSE_DECL (c);
|
1484 |
|
|
if (is_variable_sized (decl))
|
1485 |
|
|
install_var_local (decl, ctx);
|
1486 |
|
|
fixup_remapped_decl (decl, ctx,
|
1487 |
|
|
OMP_CLAUSE_CODE (c) == OMP_CLAUSE_PRIVATE
|
1488 |
|
|
&& OMP_CLAUSE_PRIVATE_DEBUG (c));
|
1489 |
|
|
if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION
|
1490 |
|
|
&& OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
|
1491 |
|
|
scan_array_reductions = true;
|
1492 |
|
|
break;
|
1493 |
|
|
|
1494 |
|
|
case OMP_CLAUSE_SHARED:
|
1495 |
|
|
decl = OMP_CLAUSE_DECL (c);
|
1496 |
|
|
if (! is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx)))
|
1497 |
|
|
fixup_remapped_decl (decl, ctx, false);
|
1498 |
|
|
break;
|
1499 |
|
|
|
1500 |
|
|
case OMP_CLAUSE_COPYPRIVATE:
|
1501 |
|
|
case OMP_CLAUSE_COPYIN:
|
1502 |
|
|
case OMP_CLAUSE_DEFAULT:
|
1503 |
|
|
case OMP_CLAUSE_IF:
|
1504 |
|
|
case OMP_CLAUSE_NUM_THREADS:
|
1505 |
|
|
case OMP_CLAUSE_SCHEDULE:
|
1506 |
|
|
case OMP_CLAUSE_NOWAIT:
|
1507 |
|
|
case OMP_CLAUSE_ORDERED:
|
1508 |
|
|
case OMP_CLAUSE_COLLAPSE:
|
1509 |
|
|
case OMP_CLAUSE_UNTIED:
|
1510 |
|
|
case OMP_CLAUSE_FINAL:
|
1511 |
|
|
case OMP_CLAUSE_MERGEABLE:
|
1512 |
|
|
break;
|
1513 |
|
|
|
1514 |
|
|
default:
|
1515 |
|
|
gcc_unreachable ();
|
1516 |
|
|
}
|
1517 |
|
|
}
|
1518 |
|
|
|
1519 |
|
|
if (scan_array_reductions)
|
1520 |
|
|
for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c))
|
1521 |
|
|
if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION
|
1522 |
|
|
&& OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
|
1523 |
|
|
{
|
1524 |
|
|
scan_omp (OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c), ctx);
|
1525 |
|
|
scan_omp (OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c), ctx);
|
1526 |
|
|
}
|
1527 |
|
|
else if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE
|
1528 |
|
|
&& OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c))
|
1529 |
|
|
scan_omp (OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c), ctx);
|
1530 |
|
|
}
|
1531 |
|
|
|
1532 |
|
|
/* Create a new name for omp child function. Returns an identifier. */
|
1533 |
|
|
|
1534 |
|
|
static GTY(()) unsigned int tmp_ompfn_id_num;
|
1535 |
|
|
|
1536 |
|
|
static tree
|
1537 |
|
|
create_omp_child_function_name (bool task_copy)
|
1538 |
|
|
{
|
1539 |
|
|
return (clone_function_name (current_function_decl,
|
1540 |
|
|
task_copy ? "_omp_cpyfn" : "_omp_fn"));
|
1541 |
|
|
}
|
1542 |
|
|
|
1543 |
|
|
/* Build a decl for the omp child function. It'll not contain a body
|
1544 |
|
|
yet, just the bare decl. */
|
1545 |
|
|
|
1546 |
|
|
static void
|
1547 |
|
|
create_omp_child_function (omp_context *ctx, bool task_copy)
|
1548 |
|
|
{
|
1549 |
|
|
tree decl, type, name, t;
|
1550 |
|
|
|
1551 |
|
|
name = create_omp_child_function_name (task_copy);
|
1552 |
|
|
if (task_copy)
|
1553 |
|
|
type = build_function_type_list (void_type_node, ptr_type_node,
|
1554 |
|
|
ptr_type_node, NULL_TREE);
|
1555 |
|
|
else
|
1556 |
|
|
type = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE);
|
1557 |
|
|
|
1558 |
|
|
decl = build_decl (gimple_location (ctx->stmt),
|
1559 |
|
|
FUNCTION_DECL, name, type);
|
1560 |
|
|
|
1561 |
|
|
if (!task_copy)
|
1562 |
|
|
ctx->cb.dst_fn = decl;
|
1563 |
|
|
else
|
1564 |
|
|
gimple_omp_task_set_copy_fn (ctx->stmt, decl);
|
1565 |
|
|
|
1566 |
|
|
TREE_STATIC (decl) = 1;
|
1567 |
|
|
TREE_USED (decl) = 1;
|
1568 |
|
|
DECL_ARTIFICIAL (decl) = 1;
|
1569 |
|
|
DECL_NAMELESS (decl) = 1;
|
1570 |
|
|
DECL_IGNORED_P (decl) = 0;
|
1571 |
|
|
TREE_PUBLIC (decl) = 0;
|
1572 |
|
|
DECL_UNINLINABLE (decl) = 1;
|
1573 |
|
|
DECL_EXTERNAL (decl) = 0;
|
1574 |
|
|
DECL_CONTEXT (decl) = NULL_TREE;
|
1575 |
|
|
DECL_INITIAL (decl) = make_node (BLOCK);
|
1576 |
|
|
|
1577 |
|
|
t = build_decl (DECL_SOURCE_LOCATION (decl),
|
1578 |
|
|
RESULT_DECL, NULL_TREE, void_type_node);
|
1579 |
|
|
DECL_ARTIFICIAL (t) = 1;
|
1580 |
|
|
DECL_IGNORED_P (t) = 1;
|
1581 |
|
|
DECL_CONTEXT (t) = decl;
|
1582 |
|
|
DECL_RESULT (decl) = t;
|
1583 |
|
|
|
1584 |
|
|
t = build_decl (DECL_SOURCE_LOCATION (decl),
|
1585 |
|
|
PARM_DECL, get_identifier (".omp_data_i"), ptr_type_node);
|
1586 |
|
|
DECL_ARTIFICIAL (t) = 1;
|
1587 |
|
|
DECL_NAMELESS (t) = 1;
|
1588 |
|
|
DECL_ARG_TYPE (t) = ptr_type_node;
|
1589 |
|
|
DECL_CONTEXT (t) = current_function_decl;
|
1590 |
|
|
TREE_USED (t) = 1;
|
1591 |
|
|
DECL_ARGUMENTS (decl) = t;
|
1592 |
|
|
if (!task_copy)
|
1593 |
|
|
ctx->receiver_decl = t;
|
1594 |
|
|
else
|
1595 |
|
|
{
|
1596 |
|
|
t = build_decl (DECL_SOURCE_LOCATION (decl),
|
1597 |
|
|
PARM_DECL, get_identifier (".omp_data_o"),
|
1598 |
|
|
ptr_type_node);
|
1599 |
|
|
DECL_ARTIFICIAL (t) = 1;
|
1600 |
|
|
DECL_NAMELESS (t) = 1;
|
1601 |
|
|
DECL_ARG_TYPE (t) = ptr_type_node;
|
1602 |
|
|
DECL_CONTEXT (t) = current_function_decl;
|
1603 |
|
|
TREE_USED (t) = 1;
|
1604 |
|
|
TREE_ADDRESSABLE (t) = 1;
|
1605 |
|
|
DECL_CHAIN (t) = DECL_ARGUMENTS (decl);
|
1606 |
|
|
DECL_ARGUMENTS (decl) = t;
|
1607 |
|
|
}
|
1608 |
|
|
|
1609 |
|
|
/* Allocate memory for the function structure. The call to
|
1610 |
|
|
allocate_struct_function clobbers CFUN, so we need to restore
|
1611 |
|
|
it afterward. */
|
1612 |
|
|
push_struct_function (decl);
|
1613 |
|
|
cfun->function_end_locus = gimple_location (ctx->stmt);
|
1614 |
|
|
pop_cfun ();
|
1615 |
|
|
}
|
1616 |
|
|
|
1617 |
|
|
|
1618 |
|
|
/* Scan an OpenMP parallel directive. */
|
1619 |
|
|
|
1620 |
|
|
static void
|
1621 |
|
|
scan_omp_parallel (gimple_stmt_iterator *gsi, omp_context *outer_ctx)
|
1622 |
|
|
{
|
1623 |
|
|
omp_context *ctx;
|
1624 |
|
|
tree name;
|
1625 |
|
|
gimple stmt = gsi_stmt (*gsi);
|
1626 |
|
|
|
1627 |
|
|
/* Ignore parallel directives with empty bodies, unless there
|
1628 |
|
|
are copyin clauses. */
|
1629 |
|
|
if (optimize > 0
|
1630 |
|
|
&& empty_body_p (gimple_omp_body (stmt))
|
1631 |
|
|
&& find_omp_clause (gimple_omp_parallel_clauses (stmt),
|
1632 |
|
|
OMP_CLAUSE_COPYIN) == NULL)
|
1633 |
|
|
{
|
1634 |
|
|
gsi_replace (gsi, gimple_build_nop (), false);
|
1635 |
|
|
return;
|
1636 |
|
|
}
|
1637 |
|
|
|
1638 |
|
|
ctx = new_omp_context (stmt, outer_ctx);
|
1639 |
|
|
if (taskreg_nesting_level > 1)
|
1640 |
|
|
ctx->is_nested = true;
|
1641 |
|
|
ctx->field_map = splay_tree_new (splay_tree_compare_pointers, 0, 0);
|
1642 |
|
|
ctx->default_kind = OMP_CLAUSE_DEFAULT_SHARED;
|
1643 |
|
|
ctx->record_type = lang_hooks.types.make_type (RECORD_TYPE);
|
1644 |
|
|
name = create_tmp_var_name (".omp_data_s");
|
1645 |
|
|
name = build_decl (gimple_location (stmt),
|
1646 |
|
|
TYPE_DECL, name, ctx->record_type);
|
1647 |
|
|
DECL_ARTIFICIAL (name) = 1;
|
1648 |
|
|
DECL_NAMELESS (name) = 1;
|
1649 |
|
|
TYPE_NAME (ctx->record_type) = name;
|
1650 |
|
|
create_omp_child_function (ctx, false);
|
1651 |
|
|
gimple_omp_parallel_set_child_fn (stmt, ctx->cb.dst_fn);
|
1652 |
|
|
|
1653 |
|
|
scan_sharing_clauses (gimple_omp_parallel_clauses (stmt), ctx);
|
1654 |
|
|
scan_omp (gimple_omp_body (stmt), ctx);
|
1655 |
|
|
|
1656 |
|
|
if (TYPE_FIELDS (ctx->record_type) == NULL)
|
1657 |
|
|
ctx->record_type = ctx->receiver_decl = NULL;
|
1658 |
|
|
else
|
1659 |
|
|
{
|
1660 |
|
|
layout_type (ctx->record_type);
|
1661 |
|
|
fixup_child_record_type (ctx);
|
1662 |
|
|
}
|
1663 |
|
|
}
|
1664 |
|
|
|
1665 |
|
|
/* Scan an OpenMP task directive. */
|
1666 |
|
|
|
1667 |
|
|
static void
|
1668 |
|
|
scan_omp_task (gimple_stmt_iterator *gsi, omp_context *outer_ctx)
|
1669 |
|
|
{
|
1670 |
|
|
omp_context *ctx;
|
1671 |
|
|
tree name, t;
|
1672 |
|
|
gimple stmt = gsi_stmt (*gsi);
|
1673 |
|
|
location_t loc = gimple_location (stmt);
|
1674 |
|
|
|
1675 |
|
|
/* Ignore task directives with empty bodies. */
|
1676 |
|
|
if (optimize > 0
|
1677 |
|
|
&& empty_body_p (gimple_omp_body (stmt)))
|
1678 |
|
|
{
|
1679 |
|
|
gsi_replace (gsi, gimple_build_nop (), false);
|
1680 |
|
|
return;
|
1681 |
|
|
}
|
1682 |
|
|
|
1683 |
|
|
ctx = new_omp_context (stmt, outer_ctx);
|
1684 |
|
|
if (taskreg_nesting_level > 1)
|
1685 |
|
|
ctx->is_nested = true;
|
1686 |
|
|
ctx->field_map = splay_tree_new (splay_tree_compare_pointers, 0, 0);
|
1687 |
|
|
ctx->default_kind = OMP_CLAUSE_DEFAULT_SHARED;
|
1688 |
|
|
ctx->record_type = lang_hooks.types.make_type (RECORD_TYPE);
|
1689 |
|
|
name = create_tmp_var_name (".omp_data_s");
|
1690 |
|
|
name = build_decl (gimple_location (stmt),
|
1691 |
|
|
TYPE_DECL, name, ctx->record_type);
|
1692 |
|
|
DECL_ARTIFICIAL (name) = 1;
|
1693 |
|
|
DECL_NAMELESS (name) = 1;
|
1694 |
|
|
TYPE_NAME (ctx->record_type) = name;
|
1695 |
|
|
create_omp_child_function (ctx, false);
|
1696 |
|
|
gimple_omp_task_set_child_fn (stmt, ctx->cb.dst_fn);
|
1697 |
|
|
|
1698 |
|
|
scan_sharing_clauses (gimple_omp_task_clauses (stmt), ctx);
|
1699 |
|
|
|
1700 |
|
|
if (ctx->srecord_type)
|
1701 |
|
|
{
|
1702 |
|
|
name = create_tmp_var_name (".omp_data_a");
|
1703 |
|
|
name = build_decl (gimple_location (stmt),
|
1704 |
|
|
TYPE_DECL, name, ctx->srecord_type);
|
1705 |
|
|
DECL_ARTIFICIAL (name) = 1;
|
1706 |
|
|
DECL_NAMELESS (name) = 1;
|
1707 |
|
|
TYPE_NAME (ctx->srecord_type) = name;
|
1708 |
|
|
create_omp_child_function (ctx, true);
|
1709 |
|
|
}
|
1710 |
|
|
|
1711 |
|
|
scan_omp (gimple_omp_body (stmt), ctx);
|
1712 |
|
|
|
1713 |
|
|
if (TYPE_FIELDS (ctx->record_type) == NULL)
|
1714 |
|
|
{
|
1715 |
|
|
ctx->record_type = ctx->receiver_decl = NULL;
|
1716 |
|
|
t = build_int_cst (long_integer_type_node, 0);
|
1717 |
|
|
gimple_omp_task_set_arg_size (stmt, t);
|
1718 |
|
|
t = build_int_cst (long_integer_type_node, 1);
|
1719 |
|
|
gimple_omp_task_set_arg_align (stmt, t);
|
1720 |
|
|
}
|
1721 |
|
|
else
|
1722 |
|
|
{
|
1723 |
|
|
tree *p, vla_fields = NULL_TREE, *q = &vla_fields;
|
1724 |
|
|
/* Move VLA fields to the end. */
|
1725 |
|
|
p = &TYPE_FIELDS (ctx->record_type);
|
1726 |
|
|
while (*p)
|
1727 |
|
|
if (!TYPE_SIZE_UNIT (TREE_TYPE (*p))
|
1728 |
|
|
|| ! TREE_CONSTANT (TYPE_SIZE_UNIT (TREE_TYPE (*p))))
|
1729 |
|
|
{
|
1730 |
|
|
*q = *p;
|
1731 |
|
|
*p = TREE_CHAIN (*p);
|
1732 |
|
|
TREE_CHAIN (*q) = NULL_TREE;
|
1733 |
|
|
q = &TREE_CHAIN (*q);
|
1734 |
|
|
}
|
1735 |
|
|
else
|
1736 |
|
|
p = &DECL_CHAIN (*p);
|
1737 |
|
|
*p = vla_fields;
|
1738 |
|
|
layout_type (ctx->record_type);
|
1739 |
|
|
fixup_child_record_type (ctx);
|
1740 |
|
|
if (ctx->srecord_type)
|
1741 |
|
|
layout_type (ctx->srecord_type);
|
1742 |
|
|
t = fold_convert_loc (loc, long_integer_type_node,
|
1743 |
|
|
TYPE_SIZE_UNIT (ctx->record_type));
|
1744 |
|
|
gimple_omp_task_set_arg_size (stmt, t);
|
1745 |
|
|
t = build_int_cst (long_integer_type_node,
|
1746 |
|
|
TYPE_ALIGN_UNIT (ctx->record_type));
|
1747 |
|
|
gimple_omp_task_set_arg_align (stmt, t);
|
1748 |
|
|
}
|
1749 |
|
|
}
|
1750 |
|
|
|
1751 |
|
|
|
1752 |
|
|
/* Scan an OpenMP loop directive. */
|
1753 |
|
|
|
1754 |
|
|
static void
|
1755 |
|
|
scan_omp_for (gimple stmt, omp_context *outer_ctx)
|
1756 |
|
|
{
|
1757 |
|
|
omp_context *ctx;
|
1758 |
|
|
size_t i;
|
1759 |
|
|
|
1760 |
|
|
ctx = new_omp_context (stmt, outer_ctx);
|
1761 |
|
|
|
1762 |
|
|
scan_sharing_clauses (gimple_omp_for_clauses (stmt), ctx);
|
1763 |
|
|
|
1764 |
|
|
scan_omp (gimple_omp_for_pre_body (stmt), ctx);
|
1765 |
|
|
for (i = 0; i < gimple_omp_for_collapse (stmt); i++)
|
1766 |
|
|
{
|
1767 |
|
|
scan_omp_op (gimple_omp_for_index_ptr (stmt, i), ctx);
|
1768 |
|
|
scan_omp_op (gimple_omp_for_initial_ptr (stmt, i), ctx);
|
1769 |
|
|
scan_omp_op (gimple_omp_for_final_ptr (stmt, i), ctx);
|
1770 |
|
|
scan_omp_op (gimple_omp_for_incr_ptr (stmt, i), ctx);
|
1771 |
|
|
}
|
1772 |
|
|
scan_omp (gimple_omp_body (stmt), ctx);
|
1773 |
|
|
}
|
1774 |
|
|
|
1775 |
|
|
/* Scan an OpenMP sections directive. */
|
1776 |
|
|
|
1777 |
|
|
static void
|
1778 |
|
|
scan_omp_sections (gimple stmt, omp_context *outer_ctx)
|
1779 |
|
|
{
|
1780 |
|
|
omp_context *ctx;
|
1781 |
|
|
|
1782 |
|
|
ctx = new_omp_context (stmt, outer_ctx);
|
1783 |
|
|
scan_sharing_clauses (gimple_omp_sections_clauses (stmt), ctx);
|
1784 |
|
|
scan_omp (gimple_omp_body (stmt), ctx);
|
1785 |
|
|
}
|
1786 |
|
|
|
1787 |
|
|
/* Scan an OpenMP single directive. */
|
1788 |
|
|
|
1789 |
|
|
static void
|
1790 |
|
|
scan_omp_single (gimple stmt, omp_context *outer_ctx)
|
1791 |
|
|
{
|
1792 |
|
|
omp_context *ctx;
|
1793 |
|
|
tree name;
|
1794 |
|
|
|
1795 |
|
|
ctx = new_omp_context (stmt, outer_ctx);
|
1796 |
|
|
ctx->field_map = splay_tree_new (splay_tree_compare_pointers, 0, 0);
|
1797 |
|
|
ctx->record_type = lang_hooks.types.make_type (RECORD_TYPE);
|
1798 |
|
|
name = create_tmp_var_name (".omp_copy_s");
|
1799 |
|
|
name = build_decl (gimple_location (stmt),
|
1800 |
|
|
TYPE_DECL, name, ctx->record_type);
|
1801 |
|
|
TYPE_NAME (ctx->record_type) = name;
|
1802 |
|
|
|
1803 |
|
|
scan_sharing_clauses (gimple_omp_single_clauses (stmt), ctx);
|
1804 |
|
|
scan_omp (gimple_omp_body (stmt), ctx);
|
1805 |
|
|
|
1806 |
|
|
if (TYPE_FIELDS (ctx->record_type) == NULL)
|
1807 |
|
|
ctx->record_type = NULL;
|
1808 |
|
|
else
|
1809 |
|
|
layout_type (ctx->record_type);
|
1810 |
|
|
}
|
1811 |
|
|
|
1812 |
|
|
|
1813 |
|
|
/* Check OpenMP nesting restrictions. */
|
1814 |
|
|
static void
|
1815 |
|
|
check_omp_nesting_restrictions (gimple stmt, omp_context *ctx)
|
1816 |
|
|
{
|
1817 |
|
|
switch (gimple_code (stmt))
|
1818 |
|
|
{
|
1819 |
|
|
case GIMPLE_OMP_FOR:
|
1820 |
|
|
case GIMPLE_OMP_SECTIONS:
|
1821 |
|
|
case GIMPLE_OMP_SINGLE:
|
1822 |
|
|
case GIMPLE_CALL:
|
1823 |
|
|
for (; ctx != NULL; ctx = ctx->outer)
|
1824 |
|
|
switch (gimple_code (ctx->stmt))
|
1825 |
|
|
{
|
1826 |
|
|
case GIMPLE_OMP_FOR:
|
1827 |
|
|
case GIMPLE_OMP_SECTIONS:
|
1828 |
|
|
case GIMPLE_OMP_SINGLE:
|
1829 |
|
|
case GIMPLE_OMP_ORDERED:
|
1830 |
|
|
case GIMPLE_OMP_MASTER:
|
1831 |
|
|
case GIMPLE_OMP_TASK:
|
1832 |
|
|
if (is_gimple_call (stmt))
|
1833 |
|
|
{
|
1834 |
|
|
warning (0, "barrier region may not be closely nested inside "
|
1835 |
|
|
"of work-sharing, critical, ordered, master or "
|
1836 |
|
|
"explicit task region");
|
1837 |
|
|
return;
|
1838 |
|
|
}
|
1839 |
|
|
warning (0, "work-sharing region may not be closely nested inside "
|
1840 |
|
|
"of work-sharing, critical, ordered, master or explicit "
|
1841 |
|
|
"task region");
|
1842 |
|
|
return;
|
1843 |
|
|
case GIMPLE_OMP_PARALLEL:
|
1844 |
|
|
return;
|
1845 |
|
|
default:
|
1846 |
|
|
break;
|
1847 |
|
|
}
|
1848 |
|
|
break;
|
1849 |
|
|
case GIMPLE_OMP_MASTER:
|
1850 |
|
|
for (; ctx != NULL; ctx = ctx->outer)
|
1851 |
|
|
switch (gimple_code (ctx->stmt))
|
1852 |
|
|
{
|
1853 |
|
|
case GIMPLE_OMP_FOR:
|
1854 |
|
|
case GIMPLE_OMP_SECTIONS:
|
1855 |
|
|
case GIMPLE_OMP_SINGLE:
|
1856 |
|
|
case GIMPLE_OMP_TASK:
|
1857 |
|
|
warning (0, "master region may not be closely nested inside "
|
1858 |
|
|
"of work-sharing or explicit task region");
|
1859 |
|
|
return;
|
1860 |
|
|
case GIMPLE_OMP_PARALLEL:
|
1861 |
|
|
return;
|
1862 |
|
|
default:
|
1863 |
|
|
break;
|
1864 |
|
|
}
|
1865 |
|
|
break;
|
1866 |
|
|
case GIMPLE_OMP_ORDERED:
|
1867 |
|
|
for (; ctx != NULL; ctx = ctx->outer)
|
1868 |
|
|
switch (gimple_code (ctx->stmt))
|
1869 |
|
|
{
|
1870 |
|
|
case GIMPLE_OMP_CRITICAL:
|
1871 |
|
|
case GIMPLE_OMP_TASK:
|
1872 |
|
|
warning (0, "ordered region may not be closely nested inside "
|
1873 |
|
|
"of critical or explicit task region");
|
1874 |
|
|
return;
|
1875 |
|
|
case GIMPLE_OMP_FOR:
|
1876 |
|
|
if (find_omp_clause (gimple_omp_for_clauses (ctx->stmt),
|
1877 |
|
|
OMP_CLAUSE_ORDERED) == NULL)
|
1878 |
|
|
warning (0, "ordered region must be closely nested inside "
|
1879 |
|
|
"a loop region with an ordered clause");
|
1880 |
|
|
return;
|
1881 |
|
|
case GIMPLE_OMP_PARALLEL:
|
1882 |
|
|
return;
|
1883 |
|
|
default:
|
1884 |
|
|
break;
|
1885 |
|
|
}
|
1886 |
|
|
break;
|
1887 |
|
|
case GIMPLE_OMP_CRITICAL:
|
1888 |
|
|
for (; ctx != NULL; ctx = ctx->outer)
|
1889 |
|
|
if (gimple_code (ctx->stmt) == GIMPLE_OMP_CRITICAL
|
1890 |
|
|
&& (gimple_omp_critical_name (stmt)
|
1891 |
|
|
== gimple_omp_critical_name (ctx->stmt)))
|
1892 |
|
|
{
|
1893 |
|
|
warning (0, "critical region may not be nested inside a critical "
|
1894 |
|
|
"region with the same name");
|
1895 |
|
|
return;
|
1896 |
|
|
}
|
1897 |
|
|
break;
|
1898 |
|
|
default:
|
1899 |
|
|
break;
|
1900 |
|
|
}
|
1901 |
|
|
}
|
1902 |
|
|
|
1903 |
|
|
|
1904 |
|
|
/* Helper function scan_omp.
|
1905 |
|
|
|
1906 |
|
|
Callback for walk_tree or operators in walk_gimple_stmt used to
|
1907 |
|
|
scan for OpenMP directives in TP. */
|
1908 |
|
|
|
1909 |
|
|
static tree
|
1910 |
|
|
scan_omp_1_op (tree *tp, int *walk_subtrees, void *data)
|
1911 |
|
|
{
|
1912 |
|
|
struct walk_stmt_info *wi = (struct walk_stmt_info *) data;
|
1913 |
|
|
omp_context *ctx = (omp_context *) wi->info;
|
1914 |
|
|
tree t = *tp;
|
1915 |
|
|
|
1916 |
|
|
switch (TREE_CODE (t))
|
1917 |
|
|
{
|
1918 |
|
|
case VAR_DECL:
|
1919 |
|
|
case PARM_DECL:
|
1920 |
|
|
case LABEL_DECL:
|
1921 |
|
|
case RESULT_DECL:
|
1922 |
|
|
if (ctx)
|
1923 |
|
|
*tp = remap_decl (t, &ctx->cb);
|
1924 |
|
|
break;
|
1925 |
|
|
|
1926 |
|
|
default:
|
1927 |
|
|
if (ctx && TYPE_P (t))
|
1928 |
|
|
*tp = remap_type (t, &ctx->cb);
|
1929 |
|
|
else if (!DECL_P (t))
|
1930 |
|
|
{
|
1931 |
|
|
*walk_subtrees = 1;
|
1932 |
|
|
if (ctx)
|
1933 |
|
|
{
|
1934 |
|
|
tree tem = remap_type (TREE_TYPE (t), &ctx->cb);
|
1935 |
|
|
if (tem != TREE_TYPE (t))
|
1936 |
|
|
{
|
1937 |
|
|
if (TREE_CODE (t) == INTEGER_CST)
|
1938 |
|
|
*tp = build_int_cst_wide (tem,
|
1939 |
|
|
TREE_INT_CST_LOW (t),
|
1940 |
|
|
TREE_INT_CST_HIGH (t));
|
1941 |
|
|
else
|
1942 |
|
|
TREE_TYPE (t) = tem;
|
1943 |
|
|
}
|
1944 |
|
|
}
|
1945 |
|
|
}
|
1946 |
|
|
break;
|
1947 |
|
|
}
|
1948 |
|
|
|
1949 |
|
|
return NULL_TREE;
|
1950 |
|
|
}
|
1951 |
|
|
|
1952 |
|
|
|
1953 |
|
|
/* Helper function for scan_omp.
|
1954 |
|
|
|
1955 |
|
|
Callback for walk_gimple_stmt used to scan for OpenMP directives in
|
1956 |
|
|
the current statement in GSI. */
|
1957 |
|
|
|
1958 |
|
|
static tree
|
1959 |
|
|
scan_omp_1_stmt (gimple_stmt_iterator *gsi, bool *handled_ops_p,
|
1960 |
|
|
struct walk_stmt_info *wi)
|
1961 |
|
|
{
|
1962 |
|
|
gimple stmt = gsi_stmt (*gsi);
|
1963 |
|
|
omp_context *ctx = (omp_context *) wi->info;
|
1964 |
|
|
|
1965 |
|
|
if (gimple_has_location (stmt))
|
1966 |
|
|
input_location = gimple_location (stmt);
|
1967 |
|
|
|
1968 |
|
|
/* Check the OpenMP nesting restrictions. */
|
1969 |
|
|
if (ctx != NULL)
|
1970 |
|
|
{
|
1971 |
|
|
if (is_gimple_omp (stmt))
|
1972 |
|
|
check_omp_nesting_restrictions (stmt, ctx);
|
1973 |
|
|
else if (is_gimple_call (stmt))
|
1974 |
|
|
{
|
1975 |
|
|
tree fndecl = gimple_call_fndecl (stmt);
|
1976 |
|
|
if (fndecl && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL
|
1977 |
|
|
&& DECL_FUNCTION_CODE (fndecl) == BUILT_IN_GOMP_BARRIER)
|
1978 |
|
|
check_omp_nesting_restrictions (stmt, ctx);
|
1979 |
|
|
}
|
1980 |
|
|
}
|
1981 |
|
|
|
1982 |
|
|
*handled_ops_p = true;
|
1983 |
|
|
|
1984 |
|
|
switch (gimple_code (stmt))
|
1985 |
|
|
{
|
1986 |
|
|
case GIMPLE_OMP_PARALLEL:
|
1987 |
|
|
taskreg_nesting_level++;
|
1988 |
|
|
scan_omp_parallel (gsi, ctx);
|
1989 |
|
|
taskreg_nesting_level--;
|
1990 |
|
|
break;
|
1991 |
|
|
|
1992 |
|
|
case GIMPLE_OMP_TASK:
|
1993 |
|
|
taskreg_nesting_level++;
|
1994 |
|
|
scan_omp_task (gsi, ctx);
|
1995 |
|
|
taskreg_nesting_level--;
|
1996 |
|
|
break;
|
1997 |
|
|
|
1998 |
|
|
case GIMPLE_OMP_FOR:
|
1999 |
|
|
scan_omp_for (stmt, ctx);
|
2000 |
|
|
break;
|
2001 |
|
|
|
2002 |
|
|
case GIMPLE_OMP_SECTIONS:
|
2003 |
|
|
scan_omp_sections (stmt, ctx);
|
2004 |
|
|
break;
|
2005 |
|
|
|
2006 |
|
|
case GIMPLE_OMP_SINGLE:
|
2007 |
|
|
scan_omp_single (stmt, ctx);
|
2008 |
|
|
break;
|
2009 |
|
|
|
2010 |
|
|
case GIMPLE_OMP_SECTION:
|
2011 |
|
|
case GIMPLE_OMP_MASTER:
|
2012 |
|
|
case GIMPLE_OMP_ORDERED:
|
2013 |
|
|
case GIMPLE_OMP_CRITICAL:
|
2014 |
|
|
ctx = new_omp_context (stmt, ctx);
|
2015 |
|
|
scan_omp (gimple_omp_body (stmt), ctx);
|
2016 |
|
|
break;
|
2017 |
|
|
|
2018 |
|
|
case GIMPLE_BIND:
|
2019 |
|
|
{
|
2020 |
|
|
tree var;
|
2021 |
|
|
|
2022 |
|
|
*handled_ops_p = false;
|
2023 |
|
|
if (ctx)
|
2024 |
|
|
for (var = gimple_bind_vars (stmt); var ; var = DECL_CHAIN (var))
|
2025 |
|
|
insert_decl_map (&ctx->cb, var, var);
|
2026 |
|
|
}
|
2027 |
|
|
break;
|
2028 |
|
|
default:
|
2029 |
|
|
*handled_ops_p = false;
|
2030 |
|
|
break;
|
2031 |
|
|
}
|
2032 |
|
|
|
2033 |
|
|
return NULL_TREE;
|
2034 |
|
|
}
|
2035 |
|
|
|
2036 |
|
|
|
2037 |
|
|
/* Scan all the statements starting at the current statement. CTX
|
2038 |
|
|
contains context information about the OpenMP directives and
|
2039 |
|
|
clauses found during the scan. */
|
2040 |
|
|
|
2041 |
|
|
static void
|
2042 |
|
|
scan_omp (gimple_seq body, omp_context *ctx)
|
2043 |
|
|
{
|
2044 |
|
|
location_t saved_location;
|
2045 |
|
|
struct walk_stmt_info wi;
|
2046 |
|
|
|
2047 |
|
|
memset (&wi, 0, sizeof (wi));
|
2048 |
|
|
wi.info = ctx;
|
2049 |
|
|
wi.want_locations = true;
|
2050 |
|
|
|
2051 |
|
|
saved_location = input_location;
|
2052 |
|
|
walk_gimple_seq (body, scan_omp_1_stmt, scan_omp_1_op, &wi);
|
2053 |
|
|
input_location = saved_location;
|
2054 |
|
|
}
|
2055 |
|
|
|
2056 |
|
|
/* Re-gimplification and code generation routines. */
|
2057 |
|
|
|
2058 |
|
|
/* Build a call to GOMP_barrier. */
|
2059 |
|
|
|
2060 |
|
|
static tree
|
2061 |
|
|
build_omp_barrier (void)
|
2062 |
|
|
{
|
2063 |
|
|
return build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_BARRIER), 0);
|
2064 |
|
|
}
|
2065 |
|
|
|
2066 |
|
|
/* If a context was created for STMT when it was scanned, return it. */
|
2067 |
|
|
|
2068 |
|
|
static omp_context *
|
2069 |
|
|
maybe_lookup_ctx (gimple stmt)
|
2070 |
|
|
{
|
2071 |
|
|
splay_tree_node n;
|
2072 |
|
|
n = splay_tree_lookup (all_contexts, (splay_tree_key) stmt);
|
2073 |
|
|
return n ? (omp_context *) n->value : NULL;
|
2074 |
|
|
}
|
2075 |
|
|
|
2076 |
|
|
|
2077 |
|
|
/* Find the mapping for DECL in CTX or the immediately enclosing
|
2078 |
|
|
context that has a mapping for DECL.
|
2079 |
|
|
|
2080 |
|
|
If CTX is a nested parallel directive, we may have to use the decl
|
2081 |
|
|
mappings created in CTX's parent context. Suppose that we have the
|
2082 |
|
|
following parallel nesting (variable UIDs showed for clarity):
|
2083 |
|
|
|
2084 |
|
|
iD.1562 = 0;
|
2085 |
|
|
#omp parallel shared(iD.1562) -> outer parallel
|
2086 |
|
|
iD.1562 = iD.1562 + 1;
|
2087 |
|
|
|
2088 |
|
|
#omp parallel shared (iD.1562) -> inner parallel
|
2089 |
|
|
iD.1562 = iD.1562 - 1;
|
2090 |
|
|
|
2091 |
|
|
Each parallel structure will create a distinct .omp_data_s structure
|
2092 |
|
|
for copying iD.1562 in/out of the directive:
|
2093 |
|
|
|
2094 |
|
|
outer parallel .omp_data_s.1.i -> iD.1562
|
2095 |
|
|
inner parallel .omp_data_s.2.i -> iD.1562
|
2096 |
|
|
|
2097 |
|
|
A shared variable mapping will produce a copy-out operation before
|
2098 |
|
|
the parallel directive and a copy-in operation after it. So, in
|
2099 |
|
|
this case we would have:
|
2100 |
|
|
|
2101 |
|
|
iD.1562 = 0;
|
2102 |
|
|
.omp_data_o.1.i = iD.1562;
|
2103 |
|
|
#omp parallel shared(iD.1562) -> outer parallel
|
2104 |
|
|
.omp_data_i.1 = &.omp_data_o.1
|
2105 |
|
|
.omp_data_i.1->i = .omp_data_i.1->i + 1;
|
2106 |
|
|
|
2107 |
|
|
.omp_data_o.2.i = iD.1562; -> **
|
2108 |
|
|
#omp parallel shared(iD.1562) -> inner parallel
|
2109 |
|
|
.omp_data_i.2 = &.omp_data_o.2
|
2110 |
|
|
.omp_data_i.2->i = .omp_data_i.2->i - 1;
|
2111 |
|
|
|
2112 |
|
|
|
2113 |
|
|
** This is a problem. The symbol iD.1562 cannot be referenced
|
2114 |
|
|
inside the body of the outer parallel region. But since we are
|
2115 |
|
|
emitting this copy operation while expanding the inner parallel
|
2116 |
|
|
directive, we need to access the CTX structure of the outer
|
2117 |
|
|
parallel directive to get the correct mapping:
|
2118 |
|
|
|
2119 |
|
|
.omp_data_o.2.i = .omp_data_i.1->i
|
2120 |
|
|
|
2121 |
|
|
Since there may be other workshare or parallel directives enclosing
|
2122 |
|
|
the parallel directive, it may be necessary to walk up the context
|
2123 |
|
|
parent chain. This is not a problem in general because nested
|
2124 |
|
|
parallelism happens only rarely. */
|
2125 |
|
|
|
2126 |
|
|
static tree
|
2127 |
|
|
lookup_decl_in_outer_ctx (tree decl, omp_context *ctx)
|
2128 |
|
|
{
|
2129 |
|
|
tree t;
|
2130 |
|
|
omp_context *up;
|
2131 |
|
|
|
2132 |
|
|
for (up = ctx->outer, t = NULL; up && t == NULL; up = up->outer)
|
2133 |
|
|
t = maybe_lookup_decl (decl, up);
|
2134 |
|
|
|
2135 |
|
|
gcc_assert (!ctx->is_nested || t || is_global_var (decl));
|
2136 |
|
|
|
2137 |
|
|
return t ? t : decl;
|
2138 |
|
|
}
|
2139 |
|
|
|
2140 |
|
|
|
2141 |
|
|
/* Similar to lookup_decl_in_outer_ctx, but return DECL if not found
|
2142 |
|
|
in outer contexts. */
|
2143 |
|
|
|
2144 |
|
|
static tree
|
2145 |
|
|
maybe_lookup_decl_in_outer_ctx (tree decl, omp_context *ctx)
|
2146 |
|
|
{
|
2147 |
|
|
tree t = NULL;
|
2148 |
|
|
omp_context *up;
|
2149 |
|
|
|
2150 |
|
|
for (up = ctx->outer, t = NULL; up && t == NULL; up = up->outer)
|
2151 |
|
|
t = maybe_lookup_decl (decl, up);
|
2152 |
|
|
|
2153 |
|
|
return t ? t : decl;
|
2154 |
|
|
}
|
2155 |
|
|
|
2156 |
|
|
|
2157 |
|
|
/* Construct the initialization value for reduction CLAUSE. */
|
2158 |
|
|
|
2159 |
|
|
tree
|
2160 |
|
|
omp_reduction_init (tree clause, tree type)
|
2161 |
|
|
{
|
2162 |
|
|
location_t loc = OMP_CLAUSE_LOCATION (clause);
|
2163 |
|
|
switch (OMP_CLAUSE_REDUCTION_CODE (clause))
|
2164 |
|
|
{
|
2165 |
|
|
case PLUS_EXPR:
|
2166 |
|
|
case MINUS_EXPR:
|
2167 |
|
|
case BIT_IOR_EXPR:
|
2168 |
|
|
case BIT_XOR_EXPR:
|
2169 |
|
|
case TRUTH_OR_EXPR:
|
2170 |
|
|
case TRUTH_ORIF_EXPR:
|
2171 |
|
|
case TRUTH_XOR_EXPR:
|
2172 |
|
|
case NE_EXPR:
|
2173 |
|
|
return build_zero_cst (type);
|
2174 |
|
|
|
2175 |
|
|
case MULT_EXPR:
|
2176 |
|
|
case TRUTH_AND_EXPR:
|
2177 |
|
|
case TRUTH_ANDIF_EXPR:
|
2178 |
|
|
case EQ_EXPR:
|
2179 |
|
|
return fold_convert_loc (loc, type, integer_one_node);
|
2180 |
|
|
|
2181 |
|
|
case BIT_AND_EXPR:
|
2182 |
|
|
return fold_convert_loc (loc, type, integer_minus_one_node);
|
2183 |
|
|
|
2184 |
|
|
case MAX_EXPR:
|
2185 |
|
|
if (SCALAR_FLOAT_TYPE_P (type))
|
2186 |
|
|
{
|
2187 |
|
|
REAL_VALUE_TYPE max, min;
|
2188 |
|
|
if (HONOR_INFINITIES (TYPE_MODE (type)))
|
2189 |
|
|
{
|
2190 |
|
|
real_inf (&max);
|
2191 |
|
|
real_arithmetic (&min, NEGATE_EXPR, &max, NULL);
|
2192 |
|
|
}
|
2193 |
|
|
else
|
2194 |
|
|
real_maxval (&min, 1, TYPE_MODE (type));
|
2195 |
|
|
return build_real (type, min);
|
2196 |
|
|
}
|
2197 |
|
|
else
|
2198 |
|
|
{
|
2199 |
|
|
gcc_assert (INTEGRAL_TYPE_P (type));
|
2200 |
|
|
return TYPE_MIN_VALUE (type);
|
2201 |
|
|
}
|
2202 |
|
|
|
2203 |
|
|
case MIN_EXPR:
|
2204 |
|
|
if (SCALAR_FLOAT_TYPE_P (type))
|
2205 |
|
|
{
|
2206 |
|
|
REAL_VALUE_TYPE max;
|
2207 |
|
|
if (HONOR_INFINITIES (TYPE_MODE (type)))
|
2208 |
|
|
real_inf (&max);
|
2209 |
|
|
else
|
2210 |
|
|
real_maxval (&max, 0, TYPE_MODE (type));
|
2211 |
|
|
return build_real (type, max);
|
2212 |
|
|
}
|
2213 |
|
|
else
|
2214 |
|
|
{
|
2215 |
|
|
gcc_assert (INTEGRAL_TYPE_P (type));
|
2216 |
|
|
return TYPE_MAX_VALUE (type);
|
2217 |
|
|
}
|
2218 |
|
|
|
2219 |
|
|
default:
|
2220 |
|
|
gcc_unreachable ();
|
2221 |
|
|
}
|
2222 |
|
|
}
|
2223 |
|
|
|
2224 |
|
|
/* Generate code to implement the input clauses, FIRSTPRIVATE and COPYIN,
|
2225 |
|
|
from the receiver (aka child) side and initializers for REFERENCE_TYPE
|
2226 |
|
|
private variables. Initialization statements go in ILIST, while calls
|
2227 |
|
|
to destructors go in DLIST. */
|
2228 |
|
|
|
2229 |
|
|
static void
|
2230 |
|
|
lower_rec_input_clauses (tree clauses, gimple_seq *ilist, gimple_seq *dlist,
|
2231 |
|
|
omp_context *ctx)
|
2232 |
|
|
{
|
2233 |
|
|
gimple_stmt_iterator diter;
|
2234 |
|
|
tree c, dtor, copyin_seq, x, ptr;
|
2235 |
|
|
bool copyin_by_ref = false;
|
2236 |
|
|
bool lastprivate_firstprivate = false;
|
2237 |
|
|
int pass;
|
2238 |
|
|
|
2239 |
|
|
*dlist = gimple_seq_alloc ();
|
2240 |
|
|
diter = gsi_start (*dlist);
|
2241 |
|
|
copyin_seq = NULL;
|
2242 |
|
|
|
2243 |
|
|
/* Do all the fixed sized types in the first pass, and the variable sized
|
2244 |
|
|
types in the second pass. This makes sure that the scalar arguments to
|
2245 |
|
|
the variable sized types are processed before we use them in the
|
2246 |
|
|
variable sized operations. */
|
2247 |
|
|
for (pass = 0; pass < 2; ++pass)
|
2248 |
|
|
{
|
2249 |
|
|
for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c))
|
2250 |
|
|
{
|
2251 |
|
|
enum omp_clause_code c_kind = OMP_CLAUSE_CODE (c);
|
2252 |
|
|
tree var, new_var;
|
2253 |
|
|
bool by_ref;
|
2254 |
|
|
location_t clause_loc = OMP_CLAUSE_LOCATION (c);
|
2255 |
|
|
|
2256 |
|
|
switch (c_kind)
|
2257 |
|
|
{
|
2258 |
|
|
case OMP_CLAUSE_PRIVATE:
|
2259 |
|
|
if (OMP_CLAUSE_PRIVATE_DEBUG (c))
|
2260 |
|
|
continue;
|
2261 |
|
|
break;
|
2262 |
|
|
case OMP_CLAUSE_SHARED:
|
2263 |
|
|
if (maybe_lookup_decl (OMP_CLAUSE_DECL (c), ctx) == NULL)
|
2264 |
|
|
{
|
2265 |
|
|
gcc_assert (is_global_var (OMP_CLAUSE_DECL (c)));
|
2266 |
|
|
continue;
|
2267 |
|
|
}
|
2268 |
|
|
case OMP_CLAUSE_FIRSTPRIVATE:
|
2269 |
|
|
case OMP_CLAUSE_COPYIN:
|
2270 |
|
|
case OMP_CLAUSE_REDUCTION:
|
2271 |
|
|
break;
|
2272 |
|
|
case OMP_CLAUSE_LASTPRIVATE:
|
2273 |
|
|
if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c))
|
2274 |
|
|
{
|
2275 |
|
|
lastprivate_firstprivate = true;
|
2276 |
|
|
if (pass != 0)
|
2277 |
|
|
continue;
|
2278 |
|
|
}
|
2279 |
|
|
break;
|
2280 |
|
|
default:
|
2281 |
|
|
continue;
|
2282 |
|
|
}
|
2283 |
|
|
|
2284 |
|
|
new_var = var = OMP_CLAUSE_DECL (c);
|
2285 |
|
|
if (c_kind != OMP_CLAUSE_COPYIN)
|
2286 |
|
|
new_var = lookup_decl (var, ctx);
|
2287 |
|
|
|
2288 |
|
|
if (c_kind == OMP_CLAUSE_SHARED || c_kind == OMP_CLAUSE_COPYIN)
|
2289 |
|
|
{
|
2290 |
|
|
if (pass != 0)
|
2291 |
|
|
continue;
|
2292 |
|
|
}
|
2293 |
|
|
else if (is_variable_sized (var))
|
2294 |
|
|
{
|
2295 |
|
|
/* For variable sized types, we need to allocate the
|
2296 |
|
|
actual storage here. Call alloca and store the
|
2297 |
|
|
result in the pointer decl that we created elsewhere. */
|
2298 |
|
|
if (pass == 0)
|
2299 |
|
|
continue;
|
2300 |
|
|
|
2301 |
|
|
if (c_kind != OMP_CLAUSE_FIRSTPRIVATE || !is_task_ctx (ctx))
|
2302 |
|
|
{
|
2303 |
|
|
gimple stmt;
|
2304 |
|
|
tree tmp, atmp;
|
2305 |
|
|
|
2306 |
|
|
ptr = DECL_VALUE_EXPR (new_var);
|
2307 |
|
|
gcc_assert (TREE_CODE (ptr) == INDIRECT_REF);
|
2308 |
|
|
ptr = TREE_OPERAND (ptr, 0);
|
2309 |
|
|
gcc_assert (DECL_P (ptr));
|
2310 |
|
|
x = TYPE_SIZE_UNIT (TREE_TYPE (new_var));
|
2311 |
|
|
|
2312 |
|
|
/* void *tmp = __builtin_alloca */
|
2313 |
|
|
atmp = builtin_decl_explicit (BUILT_IN_ALLOCA);
|
2314 |
|
|
stmt = gimple_build_call (atmp, 1, x);
|
2315 |
|
|
tmp = create_tmp_var_raw (ptr_type_node, NULL);
|
2316 |
|
|
gimple_add_tmp_var (tmp);
|
2317 |
|
|
gimple_call_set_lhs (stmt, tmp);
|
2318 |
|
|
|
2319 |
|
|
gimple_seq_add_stmt (ilist, stmt);
|
2320 |
|
|
|
2321 |
|
|
x = fold_convert_loc (clause_loc, TREE_TYPE (ptr), tmp);
|
2322 |
|
|
gimplify_assign (ptr, x, ilist);
|
2323 |
|
|
}
|
2324 |
|
|
}
|
2325 |
|
|
else if (is_reference (var))
|
2326 |
|
|
{
|
2327 |
|
|
/* For references that are being privatized for Fortran,
|
2328 |
|
|
allocate new backing storage for the new pointer
|
2329 |
|
|
variable. This allows us to avoid changing all the
|
2330 |
|
|
code that expects a pointer to something that expects
|
2331 |
|
|
a direct variable. Note that this doesn't apply to
|
2332 |
|
|
C++, since reference types are disallowed in data
|
2333 |
|
|
sharing clauses there, except for NRV optimized
|
2334 |
|
|
return values. */
|
2335 |
|
|
if (pass == 0)
|
2336 |
|
|
continue;
|
2337 |
|
|
|
2338 |
|
|
x = TYPE_SIZE_UNIT (TREE_TYPE (TREE_TYPE (new_var)));
|
2339 |
|
|
if (c_kind == OMP_CLAUSE_FIRSTPRIVATE && is_task_ctx (ctx))
|
2340 |
|
|
{
|
2341 |
|
|
x = build_receiver_ref (var, false, ctx);
|
2342 |
|
|
x = build_fold_addr_expr_loc (clause_loc, x);
|
2343 |
|
|
}
|
2344 |
|
|
else if (TREE_CONSTANT (x))
|
2345 |
|
|
{
|
2346 |
|
|
const char *name = NULL;
|
2347 |
|
|
if (DECL_NAME (var))
|
2348 |
|
|
name = IDENTIFIER_POINTER (DECL_NAME (new_var));
|
2349 |
|
|
|
2350 |
|
|
x = create_tmp_var_raw (TREE_TYPE (TREE_TYPE (new_var)),
|
2351 |
|
|
name);
|
2352 |
|
|
gimple_add_tmp_var (x);
|
2353 |
|
|
TREE_ADDRESSABLE (x) = 1;
|
2354 |
|
|
x = build_fold_addr_expr_loc (clause_loc, x);
|
2355 |
|
|
}
|
2356 |
|
|
else
|
2357 |
|
|
{
|
2358 |
|
|
tree atmp = builtin_decl_explicit (BUILT_IN_ALLOCA);
|
2359 |
|
|
x = build_call_expr_loc (clause_loc, atmp, 1, x);
|
2360 |
|
|
}
|
2361 |
|
|
|
2362 |
|
|
x = fold_convert_loc (clause_loc, TREE_TYPE (new_var), x);
|
2363 |
|
|
gimplify_assign (new_var, x, ilist);
|
2364 |
|
|
|
2365 |
|
|
new_var = build_simple_mem_ref_loc (clause_loc, new_var);
|
2366 |
|
|
}
|
2367 |
|
|
else if (c_kind == OMP_CLAUSE_REDUCTION
|
2368 |
|
|
&& OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
|
2369 |
|
|
{
|
2370 |
|
|
if (pass == 0)
|
2371 |
|
|
continue;
|
2372 |
|
|
}
|
2373 |
|
|
else if (pass != 0)
|
2374 |
|
|
continue;
|
2375 |
|
|
|
2376 |
|
|
switch (OMP_CLAUSE_CODE (c))
|
2377 |
|
|
{
|
2378 |
|
|
case OMP_CLAUSE_SHARED:
|
2379 |
|
|
/* Shared global vars are just accessed directly. */
|
2380 |
|
|
if (is_global_var (new_var))
|
2381 |
|
|
break;
|
2382 |
|
|
/* Set up the DECL_VALUE_EXPR for shared variables now. This
|
2383 |
|
|
needs to be delayed until after fixup_child_record_type so
|
2384 |
|
|
that we get the correct type during the dereference. */
|
2385 |
|
|
by_ref = use_pointer_for_field (var, ctx);
|
2386 |
|
|
x = build_receiver_ref (var, by_ref, ctx);
|
2387 |
|
|
SET_DECL_VALUE_EXPR (new_var, x);
|
2388 |
|
|
DECL_HAS_VALUE_EXPR_P (new_var) = 1;
|
2389 |
|
|
|
2390 |
|
|
/* ??? If VAR is not passed by reference, and the variable
|
2391 |
|
|
hasn't been initialized yet, then we'll get a warning for
|
2392 |
|
|
the store into the omp_data_s structure. Ideally, we'd be
|
2393 |
|
|
able to notice this and not store anything at all, but
|
2394 |
|
|
we're generating code too early. Suppress the warning. */
|
2395 |
|
|
if (!by_ref)
|
2396 |
|
|
TREE_NO_WARNING (var) = 1;
|
2397 |
|
|
break;
|
2398 |
|
|
|
2399 |
|
|
case OMP_CLAUSE_LASTPRIVATE:
|
2400 |
|
|
if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c))
|
2401 |
|
|
break;
|
2402 |
|
|
/* FALLTHRU */
|
2403 |
|
|
|
2404 |
|
|
case OMP_CLAUSE_PRIVATE:
|
2405 |
|
|
if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_PRIVATE)
|
2406 |
|
|
x = build_outer_var_ref (var, ctx);
|
2407 |
|
|
else if (OMP_CLAUSE_PRIVATE_OUTER_REF (c))
|
2408 |
|
|
{
|
2409 |
|
|
if (is_task_ctx (ctx))
|
2410 |
|
|
x = build_receiver_ref (var, false, ctx);
|
2411 |
|
|
else
|
2412 |
|
|
x = build_outer_var_ref (var, ctx);
|
2413 |
|
|
}
|
2414 |
|
|
else
|
2415 |
|
|
x = NULL;
|
2416 |
|
|
x = lang_hooks.decls.omp_clause_default_ctor (c, new_var, x);
|
2417 |
|
|
if (x)
|
2418 |
|
|
gimplify_and_add (x, ilist);
|
2419 |
|
|
/* FALLTHRU */
|
2420 |
|
|
|
2421 |
|
|
do_dtor:
|
2422 |
|
|
x = lang_hooks.decls.omp_clause_dtor (c, new_var);
|
2423 |
|
|
if (x)
|
2424 |
|
|
{
|
2425 |
|
|
gimple_seq tseq = NULL;
|
2426 |
|
|
|
2427 |
|
|
dtor = x;
|
2428 |
|
|
gimplify_stmt (&dtor, &tseq);
|
2429 |
|
|
gsi_insert_seq_before (&diter, tseq, GSI_SAME_STMT);
|
2430 |
|
|
}
|
2431 |
|
|
break;
|
2432 |
|
|
|
2433 |
|
|
case OMP_CLAUSE_FIRSTPRIVATE:
|
2434 |
|
|
if (is_task_ctx (ctx))
|
2435 |
|
|
{
|
2436 |
|
|
if (is_reference (var) || is_variable_sized (var))
|
2437 |
|
|
goto do_dtor;
|
2438 |
|
|
else if (is_global_var (maybe_lookup_decl_in_outer_ctx (var,
|
2439 |
|
|
ctx))
|
2440 |
|
|
|| use_pointer_for_field (var, NULL))
|
2441 |
|
|
{
|
2442 |
|
|
x = build_receiver_ref (var, false, ctx);
|
2443 |
|
|
SET_DECL_VALUE_EXPR (new_var, x);
|
2444 |
|
|
DECL_HAS_VALUE_EXPR_P (new_var) = 1;
|
2445 |
|
|
goto do_dtor;
|
2446 |
|
|
}
|
2447 |
|
|
}
|
2448 |
|
|
x = build_outer_var_ref (var, ctx);
|
2449 |
|
|
x = lang_hooks.decls.omp_clause_copy_ctor (c, new_var, x);
|
2450 |
|
|
gimplify_and_add (x, ilist);
|
2451 |
|
|
goto do_dtor;
|
2452 |
|
|
break;
|
2453 |
|
|
|
2454 |
|
|
case OMP_CLAUSE_COPYIN:
|
2455 |
|
|
by_ref = use_pointer_for_field (var, NULL);
|
2456 |
|
|
x = build_receiver_ref (var, by_ref, ctx);
|
2457 |
|
|
x = lang_hooks.decls.omp_clause_assign_op (c, new_var, x);
|
2458 |
|
|
append_to_statement_list (x, ©in_seq);
|
2459 |
|
|
copyin_by_ref |= by_ref;
|
2460 |
|
|
break;
|
2461 |
|
|
|
2462 |
|
|
case OMP_CLAUSE_REDUCTION:
|
2463 |
|
|
if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
|
2464 |
|
|
{
|
2465 |
|
|
tree placeholder = OMP_CLAUSE_REDUCTION_PLACEHOLDER (c);
|
2466 |
|
|
x = build_outer_var_ref (var, ctx);
|
2467 |
|
|
|
2468 |
|
|
if (is_reference (var))
|
2469 |
|
|
x = build_fold_addr_expr_loc (clause_loc, x);
|
2470 |
|
|
SET_DECL_VALUE_EXPR (placeholder, x);
|
2471 |
|
|
DECL_HAS_VALUE_EXPR_P (placeholder) = 1;
|
2472 |
|
|
lower_omp (OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c), ctx);
|
2473 |
|
|
gimple_seq_add_seq (ilist,
|
2474 |
|
|
OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c));
|
2475 |
|
|
OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c) = NULL;
|
2476 |
|
|
DECL_HAS_VALUE_EXPR_P (placeholder) = 0;
|
2477 |
|
|
}
|
2478 |
|
|
else
|
2479 |
|
|
{
|
2480 |
|
|
x = omp_reduction_init (c, TREE_TYPE (new_var));
|
2481 |
|
|
gcc_assert (TREE_CODE (TREE_TYPE (new_var)) != ARRAY_TYPE);
|
2482 |
|
|
gimplify_assign (new_var, x, ilist);
|
2483 |
|
|
}
|
2484 |
|
|
break;
|
2485 |
|
|
|
2486 |
|
|
default:
|
2487 |
|
|
gcc_unreachable ();
|
2488 |
|
|
}
|
2489 |
|
|
}
|
2490 |
|
|
}
|
2491 |
|
|
|
2492 |
|
|
/* The copyin sequence is not to be executed by the main thread, since
|
2493 |
|
|
that would result in self-copies. Perhaps not visible to scalars,
|
2494 |
|
|
but it certainly is to C++ operator=. */
|
2495 |
|
|
if (copyin_seq)
|
2496 |
|
|
{
|
2497 |
|
|
x = build_call_expr (builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM),
|
2498 |
|
|
0);
|
2499 |
|
|
x = build2 (NE_EXPR, boolean_type_node, x,
|
2500 |
|
|
build_int_cst (TREE_TYPE (x), 0));
|
2501 |
|
|
x = build3 (COND_EXPR, void_type_node, x, copyin_seq, NULL);
|
2502 |
|
|
gimplify_and_add (x, ilist);
|
2503 |
|
|
}
|
2504 |
|
|
|
2505 |
|
|
/* If any copyin variable is passed by reference, we must ensure the
|
2506 |
|
|
master thread doesn't modify it before it is copied over in all
|
2507 |
|
|
threads. Similarly for variables in both firstprivate and
|
2508 |
|
|
lastprivate clauses we need to ensure the lastprivate copying
|
2509 |
|
|
happens after firstprivate copying in all threads. */
|
2510 |
|
|
if (copyin_by_ref || lastprivate_firstprivate)
|
2511 |
|
|
gimplify_and_add (build_omp_barrier (), ilist);
|
2512 |
|
|
}
|
2513 |
|
|
|
2514 |
|
|
|
2515 |
|
|
/* Generate code to implement the LASTPRIVATE clauses. This is used for
|
2516 |
|
|
both parallel and workshare constructs. PREDICATE may be NULL if it's
|
2517 |
|
|
always true. */
|
2518 |
|
|
|
2519 |
|
|
static void
|
2520 |
|
|
lower_lastprivate_clauses (tree clauses, tree predicate, gimple_seq *stmt_list,
|
2521 |
|
|
omp_context *ctx)
|
2522 |
|
|
{
|
2523 |
|
|
tree x, c, label = NULL;
|
2524 |
|
|
bool par_clauses = false;
|
2525 |
|
|
|
2526 |
|
|
/* Early exit if there are no lastprivate clauses. */
|
2527 |
|
|
clauses = find_omp_clause (clauses, OMP_CLAUSE_LASTPRIVATE);
|
2528 |
|
|
if (clauses == NULL)
|
2529 |
|
|
{
|
2530 |
|
|
/* If this was a workshare clause, see if it had been combined
|
2531 |
|
|
with its parallel. In that case, look for the clauses on the
|
2532 |
|
|
parallel statement itself. */
|
2533 |
|
|
if (is_parallel_ctx (ctx))
|
2534 |
|
|
return;
|
2535 |
|
|
|
2536 |
|
|
ctx = ctx->outer;
|
2537 |
|
|
if (ctx == NULL || !is_parallel_ctx (ctx))
|
2538 |
|
|
return;
|
2539 |
|
|
|
2540 |
|
|
clauses = find_omp_clause (gimple_omp_parallel_clauses (ctx->stmt),
|
2541 |
|
|
OMP_CLAUSE_LASTPRIVATE);
|
2542 |
|
|
if (clauses == NULL)
|
2543 |
|
|
return;
|
2544 |
|
|
par_clauses = true;
|
2545 |
|
|
}
|
2546 |
|
|
|
2547 |
|
|
if (predicate)
|
2548 |
|
|
{
|
2549 |
|
|
gimple stmt;
|
2550 |
|
|
tree label_true, arm1, arm2;
|
2551 |
|
|
|
2552 |
|
|
label = create_artificial_label (UNKNOWN_LOCATION);
|
2553 |
|
|
label_true = create_artificial_label (UNKNOWN_LOCATION);
|
2554 |
|
|
arm1 = TREE_OPERAND (predicate, 0);
|
2555 |
|
|
arm2 = TREE_OPERAND (predicate, 1);
|
2556 |
|
|
gimplify_expr (&arm1, stmt_list, NULL, is_gimple_val, fb_rvalue);
|
2557 |
|
|
gimplify_expr (&arm2, stmt_list, NULL, is_gimple_val, fb_rvalue);
|
2558 |
|
|
stmt = gimple_build_cond (TREE_CODE (predicate), arm1, arm2,
|
2559 |
|
|
label_true, label);
|
2560 |
|
|
gimple_seq_add_stmt (stmt_list, stmt);
|
2561 |
|
|
gimple_seq_add_stmt (stmt_list, gimple_build_label (label_true));
|
2562 |
|
|
}
|
2563 |
|
|
|
2564 |
|
|
for (c = clauses; c ;)
|
2565 |
|
|
{
|
2566 |
|
|
tree var, new_var;
|
2567 |
|
|
location_t clause_loc = OMP_CLAUSE_LOCATION (c);
|
2568 |
|
|
|
2569 |
|
|
if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE)
|
2570 |
|
|
{
|
2571 |
|
|
var = OMP_CLAUSE_DECL (c);
|
2572 |
|
|
new_var = lookup_decl (var, ctx);
|
2573 |
|
|
|
2574 |
|
|
if (OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c))
|
2575 |
|
|
{
|
2576 |
|
|
lower_omp (OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c), ctx);
|
2577 |
|
|
gimple_seq_add_seq (stmt_list,
|
2578 |
|
|
OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c));
|
2579 |
|
|
}
|
2580 |
|
|
OMP_CLAUSE_LASTPRIVATE_GIMPLE_SEQ (c) = NULL;
|
2581 |
|
|
|
2582 |
|
|
x = build_outer_var_ref (var, ctx);
|
2583 |
|
|
if (is_reference (var))
|
2584 |
|
|
new_var = build_simple_mem_ref_loc (clause_loc, new_var);
|
2585 |
|
|
x = lang_hooks.decls.omp_clause_assign_op (c, x, new_var);
|
2586 |
|
|
gimplify_and_add (x, stmt_list);
|
2587 |
|
|
}
|
2588 |
|
|
c = OMP_CLAUSE_CHAIN (c);
|
2589 |
|
|
if (c == NULL && !par_clauses)
|
2590 |
|
|
{
|
2591 |
|
|
/* If this was a workshare clause, see if it had been combined
|
2592 |
|
|
with its parallel. In that case, continue looking for the
|
2593 |
|
|
clauses also on the parallel statement itself. */
|
2594 |
|
|
if (is_parallel_ctx (ctx))
|
2595 |
|
|
break;
|
2596 |
|
|
|
2597 |
|
|
ctx = ctx->outer;
|
2598 |
|
|
if (ctx == NULL || !is_parallel_ctx (ctx))
|
2599 |
|
|
break;
|
2600 |
|
|
|
2601 |
|
|
c = find_omp_clause (gimple_omp_parallel_clauses (ctx->stmt),
|
2602 |
|
|
OMP_CLAUSE_LASTPRIVATE);
|
2603 |
|
|
par_clauses = true;
|
2604 |
|
|
}
|
2605 |
|
|
}
|
2606 |
|
|
|
2607 |
|
|
if (label)
|
2608 |
|
|
gimple_seq_add_stmt (stmt_list, gimple_build_label (label));
|
2609 |
|
|
}
|
2610 |
|
|
|
2611 |
|
|
|
2612 |
|
|
/* Generate code to implement the REDUCTION clauses. */
|
2613 |
|
|
|
2614 |
|
|
static void
|
2615 |
|
|
lower_reduction_clauses (tree clauses, gimple_seq *stmt_seqp, omp_context *ctx)
|
2616 |
|
|
{
|
2617 |
|
|
gimple_seq sub_seq = NULL;
|
2618 |
|
|
gimple stmt;
|
2619 |
|
|
tree x, c;
|
2620 |
|
|
int count = 0;
|
2621 |
|
|
|
2622 |
|
|
/* First see if there is exactly one reduction clause. Use OMP_ATOMIC
|
2623 |
|
|
update in that case, otherwise use a lock. */
|
2624 |
|
|
for (c = clauses; c && count < 2; c = OMP_CLAUSE_CHAIN (c))
|
2625 |
|
|
if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION)
|
2626 |
|
|
{
|
2627 |
|
|
if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
|
2628 |
|
|
{
|
2629 |
|
|
/* Never use OMP_ATOMIC for array reductions. */
|
2630 |
|
|
count = -1;
|
2631 |
|
|
break;
|
2632 |
|
|
}
|
2633 |
|
|
count++;
|
2634 |
|
|
}
|
2635 |
|
|
|
2636 |
|
|
if (count == 0)
|
2637 |
|
|
return;
|
2638 |
|
|
|
2639 |
|
|
for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c))
|
2640 |
|
|
{
|
2641 |
|
|
tree var, ref, new_var;
|
2642 |
|
|
enum tree_code code;
|
2643 |
|
|
location_t clause_loc = OMP_CLAUSE_LOCATION (c);
|
2644 |
|
|
|
2645 |
|
|
if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_REDUCTION)
|
2646 |
|
|
continue;
|
2647 |
|
|
|
2648 |
|
|
var = OMP_CLAUSE_DECL (c);
|
2649 |
|
|
new_var = lookup_decl (var, ctx);
|
2650 |
|
|
if (is_reference (var))
|
2651 |
|
|
new_var = build_simple_mem_ref_loc (clause_loc, new_var);
|
2652 |
|
|
ref = build_outer_var_ref (var, ctx);
|
2653 |
|
|
code = OMP_CLAUSE_REDUCTION_CODE (c);
|
2654 |
|
|
|
2655 |
|
|
/* reduction(-:var) sums up the partial results, so it acts
|
2656 |
|
|
identically to reduction(+:var). */
|
2657 |
|
|
if (code == MINUS_EXPR)
|
2658 |
|
|
code = PLUS_EXPR;
|
2659 |
|
|
|
2660 |
|
|
if (count == 1)
|
2661 |
|
|
{
|
2662 |
|
|
tree addr = build_fold_addr_expr_loc (clause_loc, ref);
|
2663 |
|
|
|
2664 |
|
|
addr = save_expr (addr);
|
2665 |
|
|
ref = build1 (INDIRECT_REF, TREE_TYPE (TREE_TYPE (addr)), addr);
|
2666 |
|
|
x = fold_build2_loc (clause_loc, code, TREE_TYPE (ref), ref, new_var);
|
2667 |
|
|
x = build2 (OMP_ATOMIC, void_type_node, addr, x);
|
2668 |
|
|
gimplify_and_add (x, stmt_seqp);
|
2669 |
|
|
return;
|
2670 |
|
|
}
|
2671 |
|
|
|
2672 |
|
|
if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
|
2673 |
|
|
{
|
2674 |
|
|
tree placeholder = OMP_CLAUSE_REDUCTION_PLACEHOLDER (c);
|
2675 |
|
|
|
2676 |
|
|
if (is_reference (var))
|
2677 |
|
|
ref = build_fold_addr_expr_loc (clause_loc, ref);
|
2678 |
|
|
SET_DECL_VALUE_EXPR (placeholder, ref);
|
2679 |
|
|
DECL_HAS_VALUE_EXPR_P (placeholder) = 1;
|
2680 |
|
|
lower_omp (OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c), ctx);
|
2681 |
|
|
gimple_seq_add_seq (&sub_seq, OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c));
|
2682 |
|
|
OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c) = NULL;
|
2683 |
|
|
OMP_CLAUSE_REDUCTION_PLACEHOLDER (c) = NULL;
|
2684 |
|
|
}
|
2685 |
|
|
else
|
2686 |
|
|
{
|
2687 |
|
|
x = build2 (code, TREE_TYPE (ref), ref, new_var);
|
2688 |
|
|
ref = build_outer_var_ref (var, ctx);
|
2689 |
|
|
gimplify_assign (ref, x, &sub_seq);
|
2690 |
|
|
}
|
2691 |
|
|
}
|
2692 |
|
|
|
2693 |
|
|
stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_START),
|
2694 |
|
|
0);
|
2695 |
|
|
gimple_seq_add_stmt (stmt_seqp, stmt);
|
2696 |
|
|
|
2697 |
|
|
gimple_seq_add_seq (stmt_seqp, sub_seq);
|
2698 |
|
|
|
2699 |
|
|
stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END),
|
2700 |
|
|
0);
|
2701 |
|
|
gimple_seq_add_stmt (stmt_seqp, stmt);
|
2702 |
|
|
}
|
2703 |
|
|
|
2704 |
|
|
|
2705 |
|
|
/* Generate code to implement the COPYPRIVATE clauses. */
|
2706 |
|
|
|
2707 |
|
|
static void
|
2708 |
|
|
lower_copyprivate_clauses (tree clauses, gimple_seq *slist, gimple_seq *rlist,
|
2709 |
|
|
omp_context *ctx)
|
2710 |
|
|
{
|
2711 |
|
|
tree c;
|
2712 |
|
|
|
2713 |
|
|
for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c))
|
2714 |
|
|
{
|
2715 |
|
|
tree var, new_var, ref, x;
|
2716 |
|
|
bool by_ref;
|
2717 |
|
|
location_t clause_loc = OMP_CLAUSE_LOCATION (c);
|
2718 |
|
|
|
2719 |
|
|
if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_COPYPRIVATE)
|
2720 |
|
|
continue;
|
2721 |
|
|
|
2722 |
|
|
var = OMP_CLAUSE_DECL (c);
|
2723 |
|
|
by_ref = use_pointer_for_field (var, NULL);
|
2724 |
|
|
|
2725 |
|
|
ref = build_sender_ref (var, ctx);
|
2726 |
|
|
x = new_var = lookup_decl_in_outer_ctx (var, ctx);
|
2727 |
|
|
if (by_ref)
|
2728 |
|
|
{
|
2729 |
|
|
x = build_fold_addr_expr_loc (clause_loc, new_var);
|
2730 |
|
|
x = fold_convert_loc (clause_loc, TREE_TYPE (ref), x);
|
2731 |
|
|
}
|
2732 |
|
|
gimplify_assign (ref, x, slist);
|
2733 |
|
|
|
2734 |
|
|
ref = build_receiver_ref (var, false, ctx);
|
2735 |
|
|
if (by_ref)
|
2736 |
|
|
{
|
2737 |
|
|
ref = fold_convert_loc (clause_loc,
|
2738 |
|
|
build_pointer_type (TREE_TYPE (new_var)),
|
2739 |
|
|
ref);
|
2740 |
|
|
ref = build_fold_indirect_ref_loc (clause_loc, ref);
|
2741 |
|
|
}
|
2742 |
|
|
if (is_reference (var))
|
2743 |
|
|
{
|
2744 |
|
|
ref = fold_convert_loc (clause_loc, TREE_TYPE (new_var), ref);
|
2745 |
|
|
ref = build_simple_mem_ref_loc (clause_loc, ref);
|
2746 |
|
|
new_var = build_simple_mem_ref_loc (clause_loc, new_var);
|
2747 |
|
|
}
|
2748 |
|
|
x = lang_hooks.decls.omp_clause_assign_op (c, new_var, ref);
|
2749 |
|
|
gimplify_and_add (x, rlist);
|
2750 |
|
|
}
|
2751 |
|
|
}
|
2752 |
|
|
|
2753 |
|
|
|
2754 |
|
|
/* Generate code to implement the clauses, FIRSTPRIVATE, COPYIN, LASTPRIVATE,
|
2755 |
|
|
and REDUCTION from the sender (aka parent) side. */
|
2756 |
|
|
|
2757 |
|
|
static void
|
2758 |
|
|
lower_send_clauses (tree clauses, gimple_seq *ilist, gimple_seq *olist,
|
2759 |
|
|
omp_context *ctx)
|
2760 |
|
|
{
|
2761 |
|
|
tree c;
|
2762 |
|
|
|
2763 |
|
|
for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c))
|
2764 |
|
|
{
|
2765 |
|
|
tree val, ref, x, var;
|
2766 |
|
|
bool by_ref, do_in = false, do_out = false;
|
2767 |
|
|
location_t clause_loc = OMP_CLAUSE_LOCATION (c);
|
2768 |
|
|
|
2769 |
|
|
switch (OMP_CLAUSE_CODE (c))
|
2770 |
|
|
{
|
2771 |
|
|
case OMP_CLAUSE_PRIVATE:
|
2772 |
|
|
if (OMP_CLAUSE_PRIVATE_OUTER_REF (c))
|
2773 |
|
|
break;
|
2774 |
|
|
continue;
|
2775 |
|
|
case OMP_CLAUSE_FIRSTPRIVATE:
|
2776 |
|
|
case OMP_CLAUSE_COPYIN:
|
2777 |
|
|
case OMP_CLAUSE_LASTPRIVATE:
|
2778 |
|
|
case OMP_CLAUSE_REDUCTION:
|
2779 |
|
|
break;
|
2780 |
|
|
default:
|
2781 |
|
|
continue;
|
2782 |
|
|
}
|
2783 |
|
|
|
2784 |
|
|
val = OMP_CLAUSE_DECL (c);
|
2785 |
|
|
var = lookup_decl_in_outer_ctx (val, ctx);
|
2786 |
|
|
|
2787 |
|
|
if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_COPYIN
|
2788 |
|
|
&& is_global_var (var))
|
2789 |
|
|
continue;
|
2790 |
|
|
if (is_variable_sized (val))
|
2791 |
|
|
continue;
|
2792 |
|
|
by_ref = use_pointer_for_field (val, NULL);
|
2793 |
|
|
|
2794 |
|
|
switch (OMP_CLAUSE_CODE (c))
|
2795 |
|
|
{
|
2796 |
|
|
case OMP_CLAUSE_PRIVATE:
|
2797 |
|
|
case OMP_CLAUSE_FIRSTPRIVATE:
|
2798 |
|
|
case OMP_CLAUSE_COPYIN:
|
2799 |
|
|
do_in = true;
|
2800 |
|
|
break;
|
2801 |
|
|
|
2802 |
|
|
case OMP_CLAUSE_LASTPRIVATE:
|
2803 |
|
|
if (by_ref || is_reference (val))
|
2804 |
|
|
{
|
2805 |
|
|
if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c))
|
2806 |
|
|
continue;
|
2807 |
|
|
do_in = true;
|
2808 |
|
|
}
|
2809 |
|
|
else
|
2810 |
|
|
{
|
2811 |
|
|
do_out = true;
|
2812 |
|
|
if (lang_hooks.decls.omp_private_outer_ref (val))
|
2813 |
|
|
do_in = true;
|
2814 |
|
|
}
|
2815 |
|
|
break;
|
2816 |
|
|
|
2817 |
|
|
case OMP_CLAUSE_REDUCTION:
|
2818 |
|
|
do_in = true;
|
2819 |
|
|
do_out = !(by_ref || is_reference (val));
|
2820 |
|
|
break;
|
2821 |
|
|
|
2822 |
|
|
default:
|
2823 |
|
|
gcc_unreachable ();
|
2824 |
|
|
}
|
2825 |
|
|
|
2826 |
|
|
if (do_in)
|
2827 |
|
|
{
|
2828 |
|
|
ref = build_sender_ref (val, ctx);
|
2829 |
|
|
x = by_ref ? build_fold_addr_expr_loc (clause_loc, var) : var;
|
2830 |
|
|
gimplify_assign (ref, x, ilist);
|
2831 |
|
|
if (is_task_ctx (ctx))
|
2832 |
|
|
DECL_ABSTRACT_ORIGIN (TREE_OPERAND (ref, 1)) = NULL;
|
2833 |
|
|
}
|
2834 |
|
|
|
2835 |
|
|
if (do_out)
|
2836 |
|
|
{
|
2837 |
|
|
ref = build_sender_ref (val, ctx);
|
2838 |
|
|
gimplify_assign (var, ref, olist);
|
2839 |
|
|
}
|
2840 |
|
|
}
|
2841 |
|
|
}
|
2842 |
|
|
|
2843 |
|
|
/* Generate code to implement SHARED from the sender (aka parent)
|
2844 |
|
|
side. This is trickier, since GIMPLE_OMP_PARALLEL_CLAUSES doesn't
|
2845 |
|
|
list things that got automatically shared. */
|
2846 |
|
|
|
2847 |
|
|
static void
|
2848 |
|
|
lower_send_shared_vars (gimple_seq *ilist, gimple_seq *olist, omp_context *ctx)
|
2849 |
|
|
{
|
2850 |
|
|
tree var, ovar, nvar, f, x, record_type;
|
2851 |
|
|
|
2852 |
|
|
if (ctx->record_type == NULL)
|
2853 |
|
|
return;
|
2854 |
|
|
|
2855 |
|
|
record_type = ctx->srecord_type ? ctx->srecord_type : ctx->record_type;
|
2856 |
|
|
for (f = TYPE_FIELDS (record_type); f ; f = DECL_CHAIN (f))
|
2857 |
|
|
{
|
2858 |
|
|
ovar = DECL_ABSTRACT_ORIGIN (f);
|
2859 |
|
|
nvar = maybe_lookup_decl (ovar, ctx);
|
2860 |
|
|
if (!nvar || !DECL_HAS_VALUE_EXPR_P (nvar))
|
2861 |
|
|
continue;
|
2862 |
|
|
|
2863 |
|
|
/* If CTX is a nested parallel directive. Find the immediately
|
2864 |
|
|
enclosing parallel or workshare construct that contains a
|
2865 |
|
|
mapping for OVAR. */
|
2866 |
|
|
var = lookup_decl_in_outer_ctx (ovar, ctx);
|
2867 |
|
|
|
2868 |
|
|
if (use_pointer_for_field (ovar, ctx))
|
2869 |
|
|
{
|
2870 |
|
|
x = build_sender_ref (ovar, ctx);
|
2871 |
|
|
var = build_fold_addr_expr (var);
|
2872 |
|
|
gimplify_assign (x, var, ilist);
|
2873 |
|
|
}
|
2874 |
|
|
else
|
2875 |
|
|
{
|
2876 |
|
|
x = build_sender_ref (ovar, ctx);
|
2877 |
|
|
gimplify_assign (x, var, ilist);
|
2878 |
|
|
|
2879 |
|
|
if (!TREE_READONLY (var)
|
2880 |
|
|
/* We don't need to receive a new reference to a result
|
2881 |
|
|
or parm decl. In fact we may not store to it as we will
|
2882 |
|
|
invalidate any pending RSO and generate wrong gimple
|
2883 |
|
|
during inlining. */
|
2884 |
|
|
&& !((TREE_CODE (var) == RESULT_DECL
|
2885 |
|
|
|| TREE_CODE (var) == PARM_DECL)
|
2886 |
|
|
&& DECL_BY_REFERENCE (var)))
|
2887 |
|
|
{
|
2888 |
|
|
x = build_sender_ref (ovar, ctx);
|
2889 |
|
|
gimplify_assign (var, x, olist);
|
2890 |
|
|
}
|
2891 |
|
|
}
|
2892 |
|
|
}
|
2893 |
|
|
}
|
2894 |
|
|
|
2895 |
|
|
|
2896 |
|
|
/* A convenience function to build an empty GIMPLE_COND with just the
|
2897 |
|
|
condition. */
|
2898 |
|
|
|
2899 |
|
|
static gimple
|
2900 |
|
|
gimple_build_cond_empty (tree cond)
|
2901 |
|
|
{
|
2902 |
|
|
enum tree_code pred_code;
|
2903 |
|
|
tree lhs, rhs;
|
2904 |
|
|
|
2905 |
|
|
gimple_cond_get_ops_from_tree (cond, &pred_code, &lhs, &rhs);
|
2906 |
|
|
return gimple_build_cond (pred_code, lhs, rhs, NULL_TREE, NULL_TREE);
|
2907 |
|
|
}
|
2908 |
|
|
|
2909 |
|
|
|
2910 |
|
|
/* Build the function calls to GOMP_parallel_start etc to actually
|
2911 |
|
|
generate the parallel operation. REGION is the parallel region
|
2912 |
|
|
being expanded. BB is the block where to insert the code. WS_ARGS
|
2913 |
|
|
will be set if this is a call to a combined parallel+workshare
|
2914 |
|
|
construct, it contains the list of additional arguments needed by
|
2915 |
|
|
the workshare construct. */
|
2916 |
|
|
|
2917 |
|
|
static void
|
2918 |
|
|
expand_parallel_call (struct omp_region *region, basic_block bb,
|
2919 |
|
|
gimple entry_stmt, VEC(tree,gc) *ws_args)
|
2920 |
|
|
{
|
2921 |
|
|
tree t, t1, t2, val, cond, c, clauses;
|
2922 |
|
|
gimple_stmt_iterator gsi;
|
2923 |
|
|
gimple stmt;
|
2924 |
|
|
enum built_in_function start_ix;
|
2925 |
|
|
int start_ix2;
|
2926 |
|
|
location_t clause_loc;
|
2927 |
|
|
VEC(tree,gc) *args;
|
2928 |
|
|
|
2929 |
|
|
clauses = gimple_omp_parallel_clauses (entry_stmt);
|
2930 |
|
|
|
2931 |
|
|
/* Determine what flavor of GOMP_parallel_start we will be
|
2932 |
|
|
emitting. */
|
2933 |
|
|
start_ix = BUILT_IN_GOMP_PARALLEL_START;
|
2934 |
|
|
if (is_combined_parallel (region))
|
2935 |
|
|
{
|
2936 |
|
|
switch (region->inner->type)
|
2937 |
|
|
{
|
2938 |
|
|
case GIMPLE_OMP_FOR:
|
2939 |
|
|
gcc_assert (region->inner->sched_kind != OMP_CLAUSE_SCHEDULE_AUTO);
|
2940 |
|
|
start_ix2 = ((int)BUILT_IN_GOMP_PARALLEL_LOOP_STATIC_START
|
2941 |
|
|
+ (region->inner->sched_kind
|
2942 |
|
|
== OMP_CLAUSE_SCHEDULE_RUNTIME
|
2943 |
|
|
? 3 : region->inner->sched_kind));
|
2944 |
|
|
start_ix = (enum built_in_function)start_ix2;
|
2945 |
|
|
break;
|
2946 |
|
|
case GIMPLE_OMP_SECTIONS:
|
2947 |
|
|
start_ix = BUILT_IN_GOMP_PARALLEL_SECTIONS_START;
|
2948 |
|
|
break;
|
2949 |
|
|
default:
|
2950 |
|
|
gcc_unreachable ();
|
2951 |
|
|
}
|
2952 |
|
|
}
|
2953 |
|
|
|
2954 |
|
|
/* By default, the value of NUM_THREADS is zero (selected at run time)
|
2955 |
|
|
and there is no conditional. */
|
2956 |
|
|
cond = NULL_TREE;
|
2957 |
|
|
val = build_int_cst (unsigned_type_node, 0);
|
2958 |
|
|
|
2959 |
|
|
c = find_omp_clause (clauses, OMP_CLAUSE_IF);
|
2960 |
|
|
if (c)
|
2961 |
|
|
cond = OMP_CLAUSE_IF_EXPR (c);
|
2962 |
|
|
|
2963 |
|
|
c = find_omp_clause (clauses, OMP_CLAUSE_NUM_THREADS);
|
2964 |
|
|
if (c)
|
2965 |
|
|
{
|
2966 |
|
|
val = OMP_CLAUSE_NUM_THREADS_EXPR (c);
|
2967 |
|
|
clause_loc = OMP_CLAUSE_LOCATION (c);
|
2968 |
|
|
}
|
2969 |
|
|
else
|
2970 |
|
|
clause_loc = gimple_location (entry_stmt);
|
2971 |
|
|
|
2972 |
|
|
/* Ensure 'val' is of the correct type. */
|
2973 |
|
|
val = fold_convert_loc (clause_loc, unsigned_type_node, val);
|
2974 |
|
|
|
2975 |
|
|
/* If we found the clause 'if (cond)', build either
|
2976 |
|
|
(cond != 0) or (cond ? val : 1u). */
|
2977 |
|
|
if (cond)
|
2978 |
|
|
{
|
2979 |
|
|
gimple_stmt_iterator gsi;
|
2980 |
|
|
|
2981 |
|
|
cond = gimple_boolify (cond);
|
2982 |
|
|
|
2983 |
|
|
if (integer_zerop (val))
|
2984 |
|
|
val = fold_build2_loc (clause_loc,
|
2985 |
|
|
EQ_EXPR, unsigned_type_node, cond,
|
2986 |
|
|
build_int_cst (TREE_TYPE (cond), 0));
|
2987 |
|
|
else
|
2988 |
|
|
{
|
2989 |
|
|
basic_block cond_bb, then_bb, else_bb;
|
2990 |
|
|
edge e, e_then, e_else;
|
2991 |
|
|
tree tmp_then, tmp_else, tmp_join, tmp_var;
|
2992 |
|
|
|
2993 |
|
|
tmp_var = create_tmp_var (TREE_TYPE (val), NULL);
|
2994 |
|
|
if (gimple_in_ssa_p (cfun))
|
2995 |
|
|
{
|
2996 |
|
|
tmp_then = make_ssa_name (tmp_var, NULL);
|
2997 |
|
|
tmp_else = make_ssa_name (tmp_var, NULL);
|
2998 |
|
|
tmp_join = make_ssa_name (tmp_var, NULL);
|
2999 |
|
|
}
|
3000 |
|
|
else
|
3001 |
|
|
{
|
3002 |
|
|
tmp_then = tmp_var;
|
3003 |
|
|
tmp_else = tmp_var;
|
3004 |
|
|
tmp_join = tmp_var;
|
3005 |
|
|
}
|
3006 |
|
|
|
3007 |
|
|
e = split_block (bb, NULL);
|
3008 |
|
|
cond_bb = e->src;
|
3009 |
|
|
bb = e->dest;
|
3010 |
|
|
remove_edge (e);
|
3011 |
|
|
|
3012 |
|
|
then_bb = create_empty_bb (cond_bb);
|
3013 |
|
|
else_bb = create_empty_bb (then_bb);
|
3014 |
|
|
set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb);
|
3015 |
|
|
set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb);
|
3016 |
|
|
|
3017 |
|
|
stmt = gimple_build_cond_empty (cond);
|
3018 |
|
|
gsi = gsi_start_bb (cond_bb);
|
3019 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
3020 |
|
|
|
3021 |
|
|
gsi = gsi_start_bb (then_bb);
|
3022 |
|
|
stmt = gimple_build_assign (tmp_then, val);
|
3023 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
3024 |
|
|
|
3025 |
|
|
gsi = gsi_start_bb (else_bb);
|
3026 |
|
|
stmt = gimple_build_assign
|
3027 |
|
|
(tmp_else, build_int_cst (unsigned_type_node, 1));
|
3028 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
3029 |
|
|
|
3030 |
|
|
make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE);
|
3031 |
|
|
make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE);
|
3032 |
|
|
e_then = make_edge (then_bb, bb, EDGE_FALLTHRU);
|
3033 |
|
|
e_else = make_edge (else_bb, bb, EDGE_FALLTHRU);
|
3034 |
|
|
|
3035 |
|
|
if (gimple_in_ssa_p (cfun))
|
3036 |
|
|
{
|
3037 |
|
|
gimple phi = create_phi_node (tmp_join, bb);
|
3038 |
|
|
SSA_NAME_DEF_STMT (tmp_join) = phi;
|
3039 |
|
|
add_phi_arg (phi, tmp_then, e_then, UNKNOWN_LOCATION);
|
3040 |
|
|
add_phi_arg (phi, tmp_else, e_else, UNKNOWN_LOCATION);
|
3041 |
|
|
}
|
3042 |
|
|
|
3043 |
|
|
val = tmp_join;
|
3044 |
|
|
}
|
3045 |
|
|
|
3046 |
|
|
gsi = gsi_start_bb (bb);
|
3047 |
|
|
val = force_gimple_operand_gsi (&gsi, val, true, NULL_TREE,
|
3048 |
|
|
false, GSI_CONTINUE_LINKING);
|
3049 |
|
|
}
|
3050 |
|
|
|
3051 |
|
|
gsi = gsi_last_bb (bb);
|
3052 |
|
|
t = gimple_omp_parallel_data_arg (entry_stmt);
|
3053 |
|
|
if (t == NULL)
|
3054 |
|
|
t1 = null_pointer_node;
|
3055 |
|
|
else
|
3056 |
|
|
t1 = build_fold_addr_expr (t);
|
3057 |
|
|
t2 = build_fold_addr_expr (gimple_omp_parallel_child_fn (entry_stmt));
|
3058 |
|
|
|
3059 |
|
|
args = VEC_alloc (tree, gc, 3 + VEC_length (tree, ws_args));
|
3060 |
|
|
VEC_quick_push (tree, args, t2);
|
3061 |
|
|
VEC_quick_push (tree, args, t1);
|
3062 |
|
|
VEC_quick_push (tree, args, val);
|
3063 |
|
|
VEC_splice (tree, args, ws_args);
|
3064 |
|
|
|
3065 |
|
|
t = build_call_expr_loc_vec (UNKNOWN_LOCATION,
|
3066 |
|
|
builtin_decl_explicit (start_ix), args);
|
3067 |
|
|
|
3068 |
|
|
force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
|
3069 |
|
|
false, GSI_CONTINUE_LINKING);
|
3070 |
|
|
|
3071 |
|
|
t = gimple_omp_parallel_data_arg (entry_stmt);
|
3072 |
|
|
if (t == NULL)
|
3073 |
|
|
t = null_pointer_node;
|
3074 |
|
|
else
|
3075 |
|
|
t = build_fold_addr_expr (t);
|
3076 |
|
|
t = build_call_expr_loc (gimple_location (entry_stmt),
|
3077 |
|
|
gimple_omp_parallel_child_fn (entry_stmt), 1, t);
|
3078 |
|
|
force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
|
3079 |
|
|
false, GSI_CONTINUE_LINKING);
|
3080 |
|
|
|
3081 |
|
|
t = build_call_expr_loc (gimple_location (entry_stmt),
|
3082 |
|
|
builtin_decl_explicit (BUILT_IN_GOMP_PARALLEL_END),
|
3083 |
|
|
0);
|
3084 |
|
|
force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
|
3085 |
|
|
false, GSI_CONTINUE_LINKING);
|
3086 |
|
|
}
|
3087 |
|
|
|
3088 |
|
|
|
3089 |
|
|
/* Build the function call to GOMP_task to actually
|
3090 |
|
|
generate the task operation. BB is the block where to insert the code. */
|
3091 |
|
|
|
3092 |
|
|
static void
|
3093 |
|
|
expand_task_call (basic_block bb, gimple entry_stmt)
|
3094 |
|
|
{
|
3095 |
|
|
tree t, t1, t2, t3, flags, cond, c, c2, clauses;
|
3096 |
|
|
gimple_stmt_iterator gsi;
|
3097 |
|
|
location_t loc = gimple_location (entry_stmt);
|
3098 |
|
|
|
3099 |
|
|
clauses = gimple_omp_task_clauses (entry_stmt);
|
3100 |
|
|
|
3101 |
|
|
c = find_omp_clause (clauses, OMP_CLAUSE_IF);
|
3102 |
|
|
if (c)
|
3103 |
|
|
cond = gimple_boolify (OMP_CLAUSE_IF_EXPR (c));
|
3104 |
|
|
else
|
3105 |
|
|
cond = boolean_true_node;
|
3106 |
|
|
|
3107 |
|
|
c = find_omp_clause (clauses, OMP_CLAUSE_UNTIED);
|
3108 |
|
|
c2 = find_omp_clause (clauses, OMP_CLAUSE_MERGEABLE);
|
3109 |
|
|
flags = build_int_cst (unsigned_type_node,
|
3110 |
|
|
(c ? 1 : 0) + (c2 ? 4 : 0));
|
3111 |
|
|
|
3112 |
|
|
c = find_omp_clause (clauses, OMP_CLAUSE_FINAL);
|
3113 |
|
|
if (c)
|
3114 |
|
|
{
|
3115 |
|
|
c = gimple_boolify (OMP_CLAUSE_FINAL_EXPR (c));
|
3116 |
|
|
c = fold_build3_loc (loc, COND_EXPR, unsigned_type_node, c,
|
3117 |
|
|
build_int_cst (unsigned_type_node, 2),
|
3118 |
|
|
build_int_cst (unsigned_type_node, 0));
|
3119 |
|
|
flags = fold_build2_loc (loc, PLUS_EXPR, unsigned_type_node, flags, c);
|
3120 |
|
|
}
|
3121 |
|
|
|
3122 |
|
|
gsi = gsi_last_bb (bb);
|
3123 |
|
|
t = gimple_omp_task_data_arg (entry_stmt);
|
3124 |
|
|
if (t == NULL)
|
3125 |
|
|
t2 = null_pointer_node;
|
3126 |
|
|
else
|
3127 |
|
|
t2 = build_fold_addr_expr_loc (loc, t);
|
3128 |
|
|
t1 = build_fold_addr_expr_loc (loc, gimple_omp_task_child_fn (entry_stmt));
|
3129 |
|
|
t = gimple_omp_task_copy_fn (entry_stmt);
|
3130 |
|
|
if (t == NULL)
|
3131 |
|
|
t3 = null_pointer_node;
|
3132 |
|
|
else
|
3133 |
|
|
t3 = build_fold_addr_expr_loc (loc, t);
|
3134 |
|
|
|
3135 |
|
|
t = build_call_expr (builtin_decl_explicit (BUILT_IN_GOMP_TASK),
|
3136 |
|
|
7, t1, t2, t3,
|
3137 |
|
|
gimple_omp_task_arg_size (entry_stmt),
|
3138 |
|
|
gimple_omp_task_arg_align (entry_stmt), cond, flags);
|
3139 |
|
|
|
3140 |
|
|
force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
|
3141 |
|
|
false, GSI_CONTINUE_LINKING);
|
3142 |
|
|
}
|
3143 |
|
|
|
3144 |
|
|
|
3145 |
|
|
/* If exceptions are enabled, wrap the statements in BODY in a MUST_NOT_THROW
|
3146 |
|
|
catch handler and return it. This prevents programs from violating the
|
3147 |
|
|
structured block semantics with throws. */
|
3148 |
|
|
|
3149 |
|
|
static gimple_seq
|
3150 |
|
|
maybe_catch_exception (gimple_seq body)
|
3151 |
|
|
{
|
3152 |
|
|
gimple g;
|
3153 |
|
|
tree decl;
|
3154 |
|
|
|
3155 |
|
|
if (!flag_exceptions)
|
3156 |
|
|
return body;
|
3157 |
|
|
|
3158 |
|
|
if (lang_hooks.eh_protect_cleanup_actions != NULL)
|
3159 |
|
|
decl = lang_hooks.eh_protect_cleanup_actions ();
|
3160 |
|
|
else
|
3161 |
|
|
decl = builtin_decl_explicit (BUILT_IN_TRAP);
|
3162 |
|
|
|
3163 |
|
|
g = gimple_build_eh_must_not_throw (decl);
|
3164 |
|
|
g = gimple_build_try (body, gimple_seq_alloc_with_stmt (g),
|
3165 |
|
|
GIMPLE_TRY_CATCH);
|
3166 |
|
|
|
3167 |
|
|
return gimple_seq_alloc_with_stmt (g);
|
3168 |
|
|
}
|
3169 |
|
|
|
3170 |
|
|
/* Chain all the DECLs in LIST by their TREE_CHAIN fields. */
|
3171 |
|
|
|
3172 |
|
|
static tree
|
3173 |
|
|
vec2chain (VEC(tree,gc) *v)
|
3174 |
|
|
{
|
3175 |
|
|
tree chain = NULL_TREE, t;
|
3176 |
|
|
unsigned ix;
|
3177 |
|
|
|
3178 |
|
|
FOR_EACH_VEC_ELT_REVERSE (tree, v, ix, t)
|
3179 |
|
|
{
|
3180 |
|
|
DECL_CHAIN (t) = chain;
|
3181 |
|
|
chain = t;
|
3182 |
|
|
}
|
3183 |
|
|
|
3184 |
|
|
return chain;
|
3185 |
|
|
}
|
3186 |
|
|
|
3187 |
|
|
|
3188 |
|
|
/* Remove barriers in REGION->EXIT's block. Note that this is only
|
3189 |
|
|
valid for GIMPLE_OMP_PARALLEL regions. Since the end of a parallel region
|
3190 |
|
|
is an implicit barrier, any workshare inside the GIMPLE_OMP_PARALLEL that
|
3191 |
|
|
left a barrier at the end of the GIMPLE_OMP_PARALLEL region can now be
|
3192 |
|
|
removed. */
|
3193 |
|
|
|
3194 |
|
|
static void
|
3195 |
|
|
remove_exit_barrier (struct omp_region *region)
|
3196 |
|
|
{
|
3197 |
|
|
gimple_stmt_iterator gsi;
|
3198 |
|
|
basic_block exit_bb;
|
3199 |
|
|
edge_iterator ei;
|
3200 |
|
|
edge e;
|
3201 |
|
|
gimple stmt;
|
3202 |
|
|
int any_addressable_vars = -1;
|
3203 |
|
|
|
3204 |
|
|
exit_bb = region->exit;
|
3205 |
|
|
|
3206 |
|
|
/* If the parallel region doesn't return, we don't have REGION->EXIT
|
3207 |
|
|
block at all. */
|
3208 |
|
|
if (! exit_bb)
|
3209 |
|
|
return;
|
3210 |
|
|
|
3211 |
|
|
/* The last insn in the block will be the parallel's GIMPLE_OMP_RETURN. The
|
3212 |
|
|
workshare's GIMPLE_OMP_RETURN will be in a preceding block. The kinds of
|
3213 |
|
|
statements that can appear in between are extremely limited -- no
|
3214 |
|
|
memory operations at all. Here, we allow nothing at all, so the
|
3215 |
|
|
only thing we allow to precede this GIMPLE_OMP_RETURN is a label. */
|
3216 |
|
|
gsi = gsi_last_bb (exit_bb);
|
3217 |
|
|
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN);
|
3218 |
|
|
gsi_prev (&gsi);
|
3219 |
|
|
if (!gsi_end_p (gsi) && gimple_code (gsi_stmt (gsi)) != GIMPLE_LABEL)
|
3220 |
|
|
return;
|
3221 |
|
|
|
3222 |
|
|
FOR_EACH_EDGE (e, ei, exit_bb->preds)
|
3223 |
|
|
{
|
3224 |
|
|
gsi = gsi_last_bb (e->src);
|
3225 |
|
|
if (gsi_end_p (gsi))
|
3226 |
|
|
continue;
|
3227 |
|
|
stmt = gsi_stmt (gsi);
|
3228 |
|
|
if (gimple_code (stmt) == GIMPLE_OMP_RETURN
|
3229 |
|
|
&& !gimple_omp_return_nowait_p (stmt))
|
3230 |
|
|
{
|
3231 |
|
|
/* OpenMP 3.0 tasks unfortunately prevent this optimization
|
3232 |
|
|
in many cases. If there could be tasks queued, the barrier
|
3233 |
|
|
might be needed to let the tasks run before some local
|
3234 |
|
|
variable of the parallel that the task uses as shared
|
3235 |
|
|
runs out of scope. The task can be spawned either
|
3236 |
|
|
from within current function (this would be easy to check)
|
3237 |
|
|
or from some function it calls and gets passed an address
|
3238 |
|
|
of such a variable. */
|
3239 |
|
|
if (any_addressable_vars < 0)
|
3240 |
|
|
{
|
3241 |
|
|
gimple parallel_stmt = last_stmt (region->entry);
|
3242 |
|
|
tree child_fun = gimple_omp_parallel_child_fn (parallel_stmt);
|
3243 |
|
|
tree local_decls, block, decl;
|
3244 |
|
|
unsigned ix;
|
3245 |
|
|
|
3246 |
|
|
any_addressable_vars = 0;
|
3247 |
|
|
FOR_EACH_LOCAL_DECL (DECL_STRUCT_FUNCTION (child_fun), ix, decl)
|
3248 |
|
|
if (TREE_ADDRESSABLE (decl))
|
3249 |
|
|
{
|
3250 |
|
|
any_addressable_vars = 1;
|
3251 |
|
|
break;
|
3252 |
|
|
}
|
3253 |
|
|
for (block = gimple_block (stmt);
|
3254 |
|
|
!any_addressable_vars
|
3255 |
|
|
&& block
|
3256 |
|
|
&& TREE_CODE (block) == BLOCK;
|
3257 |
|
|
block = BLOCK_SUPERCONTEXT (block))
|
3258 |
|
|
{
|
3259 |
|
|
for (local_decls = BLOCK_VARS (block);
|
3260 |
|
|
local_decls;
|
3261 |
|
|
local_decls = DECL_CHAIN (local_decls))
|
3262 |
|
|
if (TREE_ADDRESSABLE (local_decls))
|
3263 |
|
|
{
|
3264 |
|
|
any_addressable_vars = 1;
|
3265 |
|
|
break;
|
3266 |
|
|
}
|
3267 |
|
|
if (block == gimple_block (parallel_stmt))
|
3268 |
|
|
break;
|
3269 |
|
|
}
|
3270 |
|
|
}
|
3271 |
|
|
if (!any_addressable_vars)
|
3272 |
|
|
gimple_omp_return_set_nowait (stmt);
|
3273 |
|
|
}
|
3274 |
|
|
}
|
3275 |
|
|
}
|
3276 |
|
|
|
3277 |
|
|
static void
|
3278 |
|
|
remove_exit_barriers (struct omp_region *region)
|
3279 |
|
|
{
|
3280 |
|
|
if (region->type == GIMPLE_OMP_PARALLEL)
|
3281 |
|
|
remove_exit_barrier (region);
|
3282 |
|
|
|
3283 |
|
|
if (region->inner)
|
3284 |
|
|
{
|
3285 |
|
|
region = region->inner;
|
3286 |
|
|
remove_exit_barriers (region);
|
3287 |
|
|
while (region->next)
|
3288 |
|
|
{
|
3289 |
|
|
region = region->next;
|
3290 |
|
|
remove_exit_barriers (region);
|
3291 |
|
|
}
|
3292 |
|
|
}
|
3293 |
|
|
}
|
3294 |
|
|
|
3295 |
|
|
/* Optimize omp_get_thread_num () and omp_get_num_threads ()
|
3296 |
|
|
calls. These can't be declared as const functions, but
|
3297 |
|
|
within one parallel body they are constant, so they can be
|
3298 |
|
|
transformed there into __builtin_omp_get_{thread_num,num_threads} ()
|
3299 |
|
|
which are declared const. Similarly for task body, except
|
3300 |
|
|
that in untied task omp_get_thread_num () can change at any task
|
3301 |
|
|
scheduling point. */
|
3302 |
|
|
|
3303 |
|
|
static void
|
3304 |
|
|
optimize_omp_library_calls (gimple entry_stmt)
|
3305 |
|
|
{
|
3306 |
|
|
basic_block bb;
|
3307 |
|
|
gimple_stmt_iterator gsi;
|
3308 |
|
|
tree thr_num_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM);
|
3309 |
|
|
tree thr_num_id = DECL_ASSEMBLER_NAME (thr_num_tree);
|
3310 |
|
|
tree num_thr_tree = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS);
|
3311 |
|
|
tree num_thr_id = DECL_ASSEMBLER_NAME (num_thr_tree);
|
3312 |
|
|
bool untied_task = (gimple_code (entry_stmt) == GIMPLE_OMP_TASK
|
3313 |
|
|
&& find_omp_clause (gimple_omp_task_clauses (entry_stmt),
|
3314 |
|
|
OMP_CLAUSE_UNTIED) != NULL);
|
3315 |
|
|
|
3316 |
|
|
FOR_EACH_BB (bb)
|
3317 |
|
|
for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
|
3318 |
|
|
{
|
3319 |
|
|
gimple call = gsi_stmt (gsi);
|
3320 |
|
|
tree decl;
|
3321 |
|
|
|
3322 |
|
|
if (is_gimple_call (call)
|
3323 |
|
|
&& (decl = gimple_call_fndecl (call))
|
3324 |
|
|
&& DECL_EXTERNAL (decl)
|
3325 |
|
|
&& TREE_PUBLIC (decl)
|
3326 |
|
|
&& DECL_INITIAL (decl) == NULL)
|
3327 |
|
|
{
|
3328 |
|
|
tree built_in;
|
3329 |
|
|
|
3330 |
|
|
if (DECL_NAME (decl) == thr_num_id)
|
3331 |
|
|
{
|
3332 |
|
|
/* In #pragma omp task untied omp_get_thread_num () can change
|
3333 |
|
|
during the execution of the task region. */
|
3334 |
|
|
if (untied_task)
|
3335 |
|
|
continue;
|
3336 |
|
|
built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM);
|
3337 |
|
|
}
|
3338 |
|
|
else if (DECL_NAME (decl) == num_thr_id)
|
3339 |
|
|
built_in = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS);
|
3340 |
|
|
else
|
3341 |
|
|
continue;
|
3342 |
|
|
|
3343 |
|
|
if (DECL_ASSEMBLER_NAME (decl) != DECL_ASSEMBLER_NAME (built_in)
|
3344 |
|
|
|| gimple_call_num_args (call) != 0)
|
3345 |
|
|
continue;
|
3346 |
|
|
|
3347 |
|
|
if (flag_exceptions && !TREE_NOTHROW (decl))
|
3348 |
|
|
continue;
|
3349 |
|
|
|
3350 |
|
|
if (TREE_CODE (TREE_TYPE (decl)) != FUNCTION_TYPE
|
3351 |
|
|
|| !types_compatible_p (TREE_TYPE (TREE_TYPE (decl)),
|
3352 |
|
|
TREE_TYPE (TREE_TYPE (built_in))))
|
3353 |
|
|
continue;
|
3354 |
|
|
|
3355 |
|
|
gimple_call_set_fndecl (call, built_in);
|
3356 |
|
|
}
|
3357 |
|
|
}
|
3358 |
|
|
}
|
3359 |
|
|
|
3360 |
|
|
/* Expand the OpenMP parallel or task directive starting at REGION. */
|
3361 |
|
|
|
3362 |
|
|
static void
|
3363 |
|
|
expand_omp_taskreg (struct omp_region *region)
|
3364 |
|
|
{
|
3365 |
|
|
basic_block entry_bb, exit_bb, new_bb;
|
3366 |
|
|
struct function *child_cfun;
|
3367 |
|
|
tree child_fn, block, t;
|
3368 |
|
|
tree save_current;
|
3369 |
|
|
gimple_stmt_iterator gsi;
|
3370 |
|
|
gimple entry_stmt, stmt;
|
3371 |
|
|
edge e;
|
3372 |
|
|
VEC(tree,gc) *ws_args;
|
3373 |
|
|
|
3374 |
|
|
entry_stmt = last_stmt (region->entry);
|
3375 |
|
|
child_fn = gimple_omp_taskreg_child_fn (entry_stmt);
|
3376 |
|
|
child_cfun = DECL_STRUCT_FUNCTION (child_fn);
|
3377 |
|
|
/* If this function has been already instrumented, make sure
|
3378 |
|
|
the child function isn't instrumented again. */
|
3379 |
|
|
child_cfun->after_tree_profile = cfun->after_tree_profile;
|
3380 |
|
|
|
3381 |
|
|
entry_bb = region->entry;
|
3382 |
|
|
exit_bb = region->exit;
|
3383 |
|
|
|
3384 |
|
|
if (is_combined_parallel (region))
|
3385 |
|
|
ws_args = region->ws_args;
|
3386 |
|
|
else
|
3387 |
|
|
ws_args = NULL;
|
3388 |
|
|
|
3389 |
|
|
if (child_cfun->cfg)
|
3390 |
|
|
{
|
3391 |
|
|
/* Due to inlining, it may happen that we have already outlined
|
3392 |
|
|
the region, in which case all we need to do is make the
|
3393 |
|
|
sub-graph unreachable and emit the parallel call. */
|
3394 |
|
|
edge entry_succ_e, exit_succ_e;
|
3395 |
|
|
gimple_stmt_iterator gsi;
|
3396 |
|
|
|
3397 |
|
|
entry_succ_e = single_succ_edge (entry_bb);
|
3398 |
|
|
|
3399 |
|
|
gsi = gsi_last_bb (entry_bb);
|
3400 |
|
|
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_PARALLEL
|
3401 |
|
|
|| gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_TASK);
|
3402 |
|
|
gsi_remove (&gsi, true);
|
3403 |
|
|
|
3404 |
|
|
new_bb = entry_bb;
|
3405 |
|
|
if (exit_bb)
|
3406 |
|
|
{
|
3407 |
|
|
exit_succ_e = single_succ_edge (exit_bb);
|
3408 |
|
|
make_edge (new_bb, exit_succ_e->dest, EDGE_FALLTHRU);
|
3409 |
|
|
}
|
3410 |
|
|
remove_edge_and_dominated_blocks (entry_succ_e);
|
3411 |
|
|
}
|
3412 |
|
|
else
|
3413 |
|
|
{
|
3414 |
|
|
unsigned srcidx, dstidx, num;
|
3415 |
|
|
|
3416 |
|
|
/* If the parallel region needs data sent from the parent
|
3417 |
|
|
function, then the very first statement (except possible
|
3418 |
|
|
tree profile counter updates) of the parallel body
|
3419 |
|
|
is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since
|
3420 |
|
|
&.OMP_DATA_O is passed as an argument to the child function,
|
3421 |
|
|
we need to replace it with the argument as seen by the child
|
3422 |
|
|
function.
|
3423 |
|
|
|
3424 |
|
|
In most cases, this will end up being the identity assignment
|
3425 |
|
|
.OMP_DATA_I = .OMP_DATA_I. However, if the parallel body had
|
3426 |
|
|
a function call that has been inlined, the original PARM_DECL
|
3427 |
|
|
.OMP_DATA_I may have been converted into a different local
|
3428 |
|
|
variable. In which case, we need to keep the assignment. */
|
3429 |
|
|
if (gimple_omp_taskreg_data_arg (entry_stmt))
|
3430 |
|
|
{
|
3431 |
|
|
basic_block entry_succ_bb = single_succ (entry_bb);
|
3432 |
|
|
gimple_stmt_iterator gsi;
|
3433 |
|
|
tree arg, narg;
|
3434 |
|
|
gimple parcopy_stmt = NULL;
|
3435 |
|
|
|
3436 |
|
|
for (gsi = gsi_start_bb (entry_succ_bb); ; gsi_next (&gsi))
|
3437 |
|
|
{
|
3438 |
|
|
gimple stmt;
|
3439 |
|
|
|
3440 |
|
|
gcc_assert (!gsi_end_p (gsi));
|
3441 |
|
|
stmt = gsi_stmt (gsi);
|
3442 |
|
|
if (gimple_code (stmt) != GIMPLE_ASSIGN)
|
3443 |
|
|
continue;
|
3444 |
|
|
|
3445 |
|
|
if (gimple_num_ops (stmt) == 2)
|
3446 |
|
|
{
|
3447 |
|
|
tree arg = gimple_assign_rhs1 (stmt);
|
3448 |
|
|
|
3449 |
|
|
/* We're ignore the subcode because we're
|
3450 |
|
|
effectively doing a STRIP_NOPS. */
|
3451 |
|
|
|
3452 |
|
|
if (TREE_CODE (arg) == ADDR_EXPR
|
3453 |
|
|
&& TREE_OPERAND (arg, 0)
|
3454 |
|
|
== gimple_omp_taskreg_data_arg (entry_stmt))
|
3455 |
|
|
{
|
3456 |
|
|
parcopy_stmt = stmt;
|
3457 |
|
|
break;
|
3458 |
|
|
}
|
3459 |
|
|
}
|
3460 |
|
|
}
|
3461 |
|
|
|
3462 |
|
|
gcc_assert (parcopy_stmt != NULL);
|
3463 |
|
|
arg = DECL_ARGUMENTS (child_fn);
|
3464 |
|
|
|
3465 |
|
|
if (!gimple_in_ssa_p (cfun))
|
3466 |
|
|
{
|
3467 |
|
|
if (gimple_assign_lhs (parcopy_stmt) == arg)
|
3468 |
|
|
gsi_remove (&gsi, true);
|
3469 |
|
|
else
|
3470 |
|
|
{
|
3471 |
|
|
/* ?? Is setting the subcode really necessary ?? */
|
3472 |
|
|
gimple_omp_set_subcode (parcopy_stmt, TREE_CODE (arg));
|
3473 |
|
|
gimple_assign_set_rhs1 (parcopy_stmt, arg);
|
3474 |
|
|
}
|
3475 |
|
|
}
|
3476 |
|
|
else
|
3477 |
|
|
{
|
3478 |
|
|
/* If we are in ssa form, we must load the value from the default
|
3479 |
|
|
definition of the argument. That should not be defined now,
|
3480 |
|
|
since the argument is not used uninitialized. */
|
3481 |
|
|
gcc_assert (gimple_default_def (cfun, arg) == NULL);
|
3482 |
|
|
narg = make_ssa_name (arg, gimple_build_nop ());
|
3483 |
|
|
set_default_def (arg, narg);
|
3484 |
|
|
/* ?? Is setting the subcode really necessary ?? */
|
3485 |
|
|
gimple_omp_set_subcode (parcopy_stmt, TREE_CODE (narg));
|
3486 |
|
|
gimple_assign_set_rhs1 (parcopy_stmt, narg);
|
3487 |
|
|
update_stmt (parcopy_stmt);
|
3488 |
|
|
}
|
3489 |
|
|
}
|
3490 |
|
|
|
3491 |
|
|
/* Declare local variables needed in CHILD_CFUN. */
|
3492 |
|
|
block = DECL_INITIAL (child_fn);
|
3493 |
|
|
BLOCK_VARS (block) = vec2chain (child_cfun->local_decls);
|
3494 |
|
|
/* The gimplifier could record temporaries in parallel/task block
|
3495 |
|
|
rather than in containing function's local_decls chain,
|
3496 |
|
|
which would mean cgraph missed finalizing them. Do it now. */
|
3497 |
|
|
for (t = BLOCK_VARS (block); t; t = DECL_CHAIN (t))
|
3498 |
|
|
if (TREE_CODE (t) == VAR_DECL
|
3499 |
|
|
&& TREE_STATIC (t)
|
3500 |
|
|
&& !DECL_EXTERNAL (t))
|
3501 |
|
|
varpool_finalize_decl (t);
|
3502 |
|
|
DECL_SAVED_TREE (child_fn) = NULL;
|
3503 |
|
|
gimple_set_body (child_fn, bb_seq (single_succ (entry_bb)));
|
3504 |
|
|
TREE_USED (block) = 1;
|
3505 |
|
|
|
3506 |
|
|
/* Reset DECL_CONTEXT on function arguments. */
|
3507 |
|
|
for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t))
|
3508 |
|
|
DECL_CONTEXT (t) = child_fn;
|
3509 |
|
|
|
3510 |
|
|
/* Split ENTRY_BB at GIMPLE_OMP_PARALLEL or GIMPLE_OMP_TASK,
|
3511 |
|
|
so that it can be moved to the child function. */
|
3512 |
|
|
gsi = gsi_last_bb (entry_bb);
|
3513 |
|
|
stmt = gsi_stmt (gsi);
|
3514 |
|
|
gcc_assert (stmt && (gimple_code (stmt) == GIMPLE_OMP_PARALLEL
|
3515 |
|
|
|| gimple_code (stmt) == GIMPLE_OMP_TASK));
|
3516 |
|
|
gsi_remove (&gsi, true);
|
3517 |
|
|
e = split_block (entry_bb, stmt);
|
3518 |
|
|
entry_bb = e->dest;
|
3519 |
|
|
single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU;
|
3520 |
|
|
|
3521 |
|
|
/* Convert GIMPLE_OMP_RETURN into a RETURN_EXPR. */
|
3522 |
|
|
if (exit_bb)
|
3523 |
|
|
{
|
3524 |
|
|
gsi = gsi_last_bb (exit_bb);
|
3525 |
|
|
gcc_assert (!gsi_end_p (gsi)
|
3526 |
|
|
&& gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_RETURN);
|
3527 |
|
|
stmt = gimple_build_return (NULL);
|
3528 |
|
|
gsi_insert_after (&gsi, stmt, GSI_SAME_STMT);
|
3529 |
|
|
gsi_remove (&gsi, true);
|
3530 |
|
|
}
|
3531 |
|
|
|
3532 |
|
|
/* Move the parallel region into CHILD_CFUN. */
|
3533 |
|
|
|
3534 |
|
|
if (gimple_in_ssa_p (cfun))
|
3535 |
|
|
{
|
3536 |
|
|
push_cfun (child_cfun);
|
3537 |
|
|
init_tree_ssa (child_cfun);
|
3538 |
|
|
init_ssa_operands ();
|
3539 |
|
|
cfun->gimple_df->in_ssa_p = true;
|
3540 |
|
|
pop_cfun ();
|
3541 |
|
|
block = NULL_TREE;
|
3542 |
|
|
}
|
3543 |
|
|
else
|
3544 |
|
|
block = gimple_block (entry_stmt);
|
3545 |
|
|
|
3546 |
|
|
new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb, block);
|
3547 |
|
|
if (exit_bb)
|
3548 |
|
|
single_succ_edge (new_bb)->flags = EDGE_FALLTHRU;
|
3549 |
|
|
|
3550 |
|
|
/* Remove non-local VAR_DECLs from child_cfun->local_decls list. */
|
3551 |
|
|
num = VEC_length (tree, child_cfun->local_decls);
|
3552 |
|
|
for (srcidx = 0, dstidx = 0; srcidx < num; srcidx++)
|
3553 |
|
|
{
|
3554 |
|
|
t = VEC_index (tree, child_cfun->local_decls, srcidx);
|
3555 |
|
|
if (DECL_CONTEXT (t) == cfun->decl)
|
3556 |
|
|
continue;
|
3557 |
|
|
if (srcidx != dstidx)
|
3558 |
|
|
VEC_replace (tree, child_cfun->local_decls, dstidx, t);
|
3559 |
|
|
dstidx++;
|
3560 |
|
|
}
|
3561 |
|
|
if (dstidx != num)
|
3562 |
|
|
VEC_truncate (tree, child_cfun->local_decls, dstidx);
|
3563 |
|
|
|
3564 |
|
|
/* Inform the callgraph about the new function. */
|
3565 |
|
|
DECL_STRUCT_FUNCTION (child_fn)->curr_properties
|
3566 |
|
|
= cfun->curr_properties;
|
3567 |
|
|
cgraph_add_new_function (child_fn, true);
|
3568 |
|
|
|
3569 |
|
|
/* Fix the callgraph edges for child_cfun. Those for cfun will be
|
3570 |
|
|
fixed in a following pass. */
|
3571 |
|
|
push_cfun (child_cfun);
|
3572 |
|
|
save_current = current_function_decl;
|
3573 |
|
|
current_function_decl = child_fn;
|
3574 |
|
|
if (optimize)
|
3575 |
|
|
optimize_omp_library_calls (entry_stmt);
|
3576 |
|
|
rebuild_cgraph_edges ();
|
3577 |
|
|
|
3578 |
|
|
/* Some EH regions might become dead, see PR34608. If
|
3579 |
|
|
pass_cleanup_cfg isn't the first pass to happen with the
|
3580 |
|
|
new child, these dead EH edges might cause problems.
|
3581 |
|
|
Clean them up now. */
|
3582 |
|
|
if (flag_exceptions)
|
3583 |
|
|
{
|
3584 |
|
|
basic_block bb;
|
3585 |
|
|
bool changed = false;
|
3586 |
|
|
|
3587 |
|
|
FOR_EACH_BB (bb)
|
3588 |
|
|
changed |= gimple_purge_dead_eh_edges (bb);
|
3589 |
|
|
if (changed)
|
3590 |
|
|
cleanup_tree_cfg ();
|
3591 |
|
|
}
|
3592 |
|
|
if (gimple_in_ssa_p (cfun))
|
3593 |
|
|
update_ssa (TODO_update_ssa);
|
3594 |
|
|
current_function_decl = save_current;
|
3595 |
|
|
pop_cfun ();
|
3596 |
|
|
}
|
3597 |
|
|
|
3598 |
|
|
/* Emit a library call to launch the children threads. */
|
3599 |
|
|
if (gimple_code (entry_stmt) == GIMPLE_OMP_PARALLEL)
|
3600 |
|
|
expand_parallel_call (region, new_bb, entry_stmt, ws_args);
|
3601 |
|
|
else
|
3602 |
|
|
expand_task_call (new_bb, entry_stmt);
|
3603 |
|
|
update_ssa (TODO_update_ssa_only_virtuals);
|
3604 |
|
|
}
|
3605 |
|
|
|
3606 |
|
|
|
3607 |
|
|
/* A subroutine of expand_omp_for. Generate code for a parallel
|
3608 |
|
|
loop with any schedule. Given parameters:
|
3609 |
|
|
|
3610 |
|
|
for (V = N1; V cond N2; V += STEP) BODY;
|
3611 |
|
|
|
3612 |
|
|
where COND is "<" or ">", we generate pseudocode
|
3613 |
|
|
|
3614 |
|
|
more = GOMP_loop_foo_start (N1, N2, STEP, CHUNK, &istart0, &iend0);
|
3615 |
|
|
if (more) goto L0; else goto L3;
|
3616 |
|
|
L0:
|
3617 |
|
|
V = istart0;
|
3618 |
|
|
iend = iend0;
|
3619 |
|
|
L1:
|
3620 |
|
|
BODY;
|
3621 |
|
|
V += STEP;
|
3622 |
|
|
if (V cond iend) goto L1; else goto L2;
|
3623 |
|
|
L2:
|
3624 |
|
|
if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3;
|
3625 |
|
|
L3:
|
3626 |
|
|
|
3627 |
|
|
If this is a combined omp parallel loop, instead of the call to
|
3628 |
|
|
GOMP_loop_foo_start, we call GOMP_loop_foo_next.
|
3629 |
|
|
|
3630 |
|
|
For collapsed loops, given parameters:
|
3631 |
|
|
collapse(3)
|
3632 |
|
|
for (V1 = N11; V1 cond1 N12; V1 += STEP1)
|
3633 |
|
|
for (V2 = N21; V2 cond2 N22; V2 += STEP2)
|
3634 |
|
|
for (V3 = N31; V3 cond3 N32; V3 += STEP3)
|
3635 |
|
|
BODY;
|
3636 |
|
|
|
3637 |
|
|
we generate pseudocode
|
3638 |
|
|
|
3639 |
|
|
if (cond3 is <)
|
3640 |
|
|
adj = STEP3 - 1;
|
3641 |
|
|
else
|
3642 |
|
|
adj = STEP3 + 1;
|
3643 |
|
|
count3 = (adj + N32 - N31) / STEP3;
|
3644 |
|
|
if (cond2 is <)
|
3645 |
|
|
adj = STEP2 - 1;
|
3646 |
|
|
else
|
3647 |
|
|
adj = STEP2 + 1;
|
3648 |
|
|
count2 = (adj + N22 - N21) / STEP2;
|
3649 |
|
|
if (cond1 is <)
|
3650 |
|
|
adj = STEP1 - 1;
|
3651 |
|
|
else
|
3652 |
|
|
adj = STEP1 + 1;
|
3653 |
|
|
count1 = (adj + N12 - N11) / STEP1;
|
3654 |
|
|
count = count1 * count2 * count3;
|
3655 |
|
|
more = GOMP_loop_foo_start (0, count, 1, CHUNK, &istart0, &iend0);
|
3656 |
|
|
if (more) goto L0; else goto L3;
|
3657 |
|
|
L0:
|
3658 |
|
|
V = istart0;
|
3659 |
|
|
T = V;
|
3660 |
|
|
V3 = N31 + (T % count3) * STEP3;
|
3661 |
|
|
T = T / count3;
|
3662 |
|
|
V2 = N21 + (T % count2) * STEP2;
|
3663 |
|
|
T = T / count2;
|
3664 |
|
|
V1 = N11 + T * STEP1;
|
3665 |
|
|
iend = iend0;
|
3666 |
|
|
L1:
|
3667 |
|
|
BODY;
|
3668 |
|
|
V += 1;
|
3669 |
|
|
if (V < iend) goto L10; else goto L2;
|
3670 |
|
|
L10:
|
3671 |
|
|
V3 += STEP3;
|
3672 |
|
|
if (V3 cond3 N32) goto L1; else goto L11;
|
3673 |
|
|
L11:
|
3674 |
|
|
V3 = N31;
|
3675 |
|
|
V2 += STEP2;
|
3676 |
|
|
if (V2 cond2 N22) goto L1; else goto L12;
|
3677 |
|
|
L12:
|
3678 |
|
|
V2 = N21;
|
3679 |
|
|
V1 += STEP1;
|
3680 |
|
|
goto L1;
|
3681 |
|
|
L2:
|
3682 |
|
|
if (GOMP_loop_foo_next (&istart0, &iend0)) goto L0; else goto L3;
|
3683 |
|
|
L3:
|
3684 |
|
|
|
3685 |
|
|
*/
|
3686 |
|
|
|
3687 |
|
|
static void
|
3688 |
|
|
expand_omp_for_generic (struct omp_region *region,
|
3689 |
|
|
struct omp_for_data *fd,
|
3690 |
|
|
enum built_in_function start_fn,
|
3691 |
|
|
enum built_in_function next_fn)
|
3692 |
|
|
{
|
3693 |
|
|
tree type, istart0, iend0, iend;
|
3694 |
|
|
tree t, vmain, vback, bias = NULL_TREE;
|
3695 |
|
|
basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb, collapse_bb;
|
3696 |
|
|
basic_block l2_bb = NULL, l3_bb = NULL;
|
3697 |
|
|
gimple_stmt_iterator gsi;
|
3698 |
|
|
gimple stmt;
|
3699 |
|
|
bool in_combined_parallel = is_combined_parallel (region);
|
3700 |
|
|
bool broken_loop = region->cont == NULL;
|
3701 |
|
|
edge e, ne;
|
3702 |
|
|
tree *counts = NULL;
|
3703 |
|
|
int i;
|
3704 |
|
|
|
3705 |
|
|
gcc_assert (!broken_loop || !in_combined_parallel);
|
3706 |
|
|
gcc_assert (fd->iter_type == long_integer_type_node
|
3707 |
|
|
|| !in_combined_parallel);
|
3708 |
|
|
|
3709 |
|
|
type = TREE_TYPE (fd->loop.v);
|
3710 |
|
|
istart0 = create_tmp_var (fd->iter_type, ".istart0");
|
3711 |
|
|
iend0 = create_tmp_var (fd->iter_type, ".iend0");
|
3712 |
|
|
TREE_ADDRESSABLE (istart0) = 1;
|
3713 |
|
|
TREE_ADDRESSABLE (iend0) = 1;
|
3714 |
|
|
if (gimple_in_ssa_p (cfun))
|
3715 |
|
|
{
|
3716 |
|
|
add_referenced_var (istart0);
|
3717 |
|
|
add_referenced_var (iend0);
|
3718 |
|
|
}
|
3719 |
|
|
|
3720 |
|
|
/* See if we need to bias by LLONG_MIN. */
|
3721 |
|
|
if (fd->iter_type == long_long_unsigned_type_node
|
3722 |
|
|
&& TREE_CODE (type) == INTEGER_TYPE
|
3723 |
|
|
&& !TYPE_UNSIGNED (type))
|
3724 |
|
|
{
|
3725 |
|
|
tree n1, n2;
|
3726 |
|
|
|
3727 |
|
|
if (fd->loop.cond_code == LT_EXPR)
|
3728 |
|
|
{
|
3729 |
|
|
n1 = fd->loop.n1;
|
3730 |
|
|
n2 = fold_build2 (PLUS_EXPR, type, fd->loop.n2, fd->loop.step);
|
3731 |
|
|
}
|
3732 |
|
|
else
|
3733 |
|
|
{
|
3734 |
|
|
n1 = fold_build2 (MINUS_EXPR, type, fd->loop.n2, fd->loop.step);
|
3735 |
|
|
n2 = fd->loop.n1;
|
3736 |
|
|
}
|
3737 |
|
|
if (TREE_CODE (n1) != INTEGER_CST
|
3738 |
|
|
|| TREE_CODE (n2) != INTEGER_CST
|
3739 |
|
|
|| ((tree_int_cst_sgn (n1) < 0) ^ (tree_int_cst_sgn (n2) < 0)))
|
3740 |
|
|
bias = fold_convert (fd->iter_type, TYPE_MIN_VALUE (type));
|
3741 |
|
|
}
|
3742 |
|
|
|
3743 |
|
|
entry_bb = region->entry;
|
3744 |
|
|
cont_bb = region->cont;
|
3745 |
|
|
collapse_bb = NULL;
|
3746 |
|
|
gcc_assert (EDGE_COUNT (entry_bb->succs) == 2);
|
3747 |
|
|
gcc_assert (broken_loop
|
3748 |
|
|
|| BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest);
|
3749 |
|
|
l0_bb = split_edge (FALLTHRU_EDGE (entry_bb));
|
3750 |
|
|
l1_bb = single_succ (l0_bb);
|
3751 |
|
|
if (!broken_loop)
|
3752 |
|
|
{
|
3753 |
|
|
l2_bb = create_empty_bb (cont_bb);
|
3754 |
|
|
gcc_assert (BRANCH_EDGE (cont_bb)->dest == l1_bb);
|
3755 |
|
|
gcc_assert (EDGE_COUNT (cont_bb->succs) == 2);
|
3756 |
|
|
}
|
3757 |
|
|
else
|
3758 |
|
|
l2_bb = NULL;
|
3759 |
|
|
l3_bb = BRANCH_EDGE (entry_bb)->dest;
|
3760 |
|
|
exit_bb = region->exit;
|
3761 |
|
|
|
3762 |
|
|
gsi = gsi_last_bb (entry_bb);
|
3763 |
|
|
|
3764 |
|
|
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR);
|
3765 |
|
|
if (fd->collapse > 1)
|
3766 |
|
|
{
|
3767 |
|
|
/* collapsed loops need work for expansion in SSA form. */
|
3768 |
|
|
gcc_assert (!gimple_in_ssa_p (cfun));
|
3769 |
|
|
counts = (tree *) alloca (fd->collapse * sizeof (tree));
|
3770 |
|
|
for (i = 0; i < fd->collapse; i++)
|
3771 |
|
|
{
|
3772 |
|
|
tree itype = TREE_TYPE (fd->loops[i].v);
|
3773 |
|
|
|
3774 |
|
|
if (POINTER_TYPE_P (itype))
|
3775 |
|
|
itype = lang_hooks.types.type_for_size (TYPE_PRECISION (itype), 0);
|
3776 |
|
|
t = build_int_cst (itype, (fd->loops[i].cond_code == LT_EXPR
|
3777 |
|
|
? -1 : 1));
|
3778 |
|
|
t = fold_build2 (PLUS_EXPR, itype,
|
3779 |
|
|
fold_convert (itype, fd->loops[i].step), t);
|
3780 |
|
|
t = fold_build2 (PLUS_EXPR, itype, t,
|
3781 |
|
|
fold_convert (itype, fd->loops[i].n2));
|
3782 |
|
|
t = fold_build2 (MINUS_EXPR, itype, t,
|
3783 |
|
|
fold_convert (itype, fd->loops[i].n1));
|
3784 |
|
|
if (TYPE_UNSIGNED (itype) && fd->loops[i].cond_code == GT_EXPR)
|
3785 |
|
|
t = fold_build2 (TRUNC_DIV_EXPR, itype,
|
3786 |
|
|
fold_build1 (NEGATE_EXPR, itype, t),
|
3787 |
|
|
fold_build1 (NEGATE_EXPR, itype,
|
3788 |
|
|
fold_convert (itype,
|
3789 |
|
|
fd->loops[i].step)));
|
3790 |
|
|
else
|
3791 |
|
|
t = fold_build2 (TRUNC_DIV_EXPR, itype, t,
|
3792 |
|
|
fold_convert (itype, fd->loops[i].step));
|
3793 |
|
|
t = fold_convert (type, t);
|
3794 |
|
|
if (TREE_CODE (t) == INTEGER_CST)
|
3795 |
|
|
counts[i] = t;
|
3796 |
|
|
else
|
3797 |
|
|
{
|
3798 |
|
|
counts[i] = create_tmp_var (type, ".count");
|
3799 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE,
|
3800 |
|
|
true, GSI_SAME_STMT);
|
3801 |
|
|
stmt = gimple_build_assign (counts[i], t);
|
3802 |
|
|
gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
|
3803 |
|
|
}
|
3804 |
|
|
if (SSA_VAR_P (fd->loop.n2))
|
3805 |
|
|
{
|
3806 |
|
|
if (i == 0)
|
3807 |
|
|
t = counts[0];
|
3808 |
|
|
else
|
3809 |
|
|
{
|
3810 |
|
|
t = fold_build2 (MULT_EXPR, type, fd->loop.n2, counts[i]);
|
3811 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE,
|
3812 |
|
|
true, GSI_SAME_STMT);
|
3813 |
|
|
}
|
3814 |
|
|
stmt = gimple_build_assign (fd->loop.n2, t);
|
3815 |
|
|
gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
|
3816 |
|
|
}
|
3817 |
|
|
}
|
3818 |
|
|
}
|
3819 |
|
|
if (in_combined_parallel)
|
3820 |
|
|
{
|
3821 |
|
|
/* In a combined parallel loop, emit a call to
|
3822 |
|
|
GOMP_loop_foo_next. */
|
3823 |
|
|
t = build_call_expr (builtin_decl_explicit (next_fn), 2,
|
3824 |
|
|
build_fold_addr_expr (istart0),
|
3825 |
|
|
build_fold_addr_expr (iend0));
|
3826 |
|
|
}
|
3827 |
|
|
else
|
3828 |
|
|
{
|
3829 |
|
|
tree t0, t1, t2, t3, t4;
|
3830 |
|
|
/* If this is not a combined parallel loop, emit a call to
|
3831 |
|
|
GOMP_loop_foo_start in ENTRY_BB. */
|
3832 |
|
|
t4 = build_fold_addr_expr (iend0);
|
3833 |
|
|
t3 = build_fold_addr_expr (istart0);
|
3834 |
|
|
t2 = fold_convert (fd->iter_type, fd->loop.step);
|
3835 |
|
|
if (POINTER_TYPE_P (type)
|
3836 |
|
|
&& TYPE_PRECISION (type) != TYPE_PRECISION (fd->iter_type))
|
3837 |
|
|
{
|
3838 |
|
|
/* Avoid casting pointers to integer of a different size. */
|
3839 |
|
|
tree itype
|
3840 |
|
|
= lang_hooks.types.type_for_size (TYPE_PRECISION (type), 0);
|
3841 |
|
|
t1 = fold_convert (fd->iter_type, fold_convert (itype, fd->loop.n2));
|
3842 |
|
|
t0 = fold_convert (fd->iter_type, fold_convert (itype, fd->loop.n1));
|
3843 |
|
|
}
|
3844 |
|
|
else
|
3845 |
|
|
{
|
3846 |
|
|
t1 = fold_convert (fd->iter_type, fd->loop.n2);
|
3847 |
|
|
t0 = fold_convert (fd->iter_type, fd->loop.n1);
|
3848 |
|
|
}
|
3849 |
|
|
if (bias)
|
3850 |
|
|
{
|
3851 |
|
|
t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias);
|
3852 |
|
|
t0 = fold_build2 (PLUS_EXPR, fd->iter_type, t0, bias);
|
3853 |
|
|
}
|
3854 |
|
|
if (fd->iter_type == long_integer_type_node)
|
3855 |
|
|
{
|
3856 |
|
|
if (fd->chunk_size)
|
3857 |
|
|
{
|
3858 |
|
|
t = fold_convert (fd->iter_type, fd->chunk_size);
|
3859 |
|
|
t = build_call_expr (builtin_decl_explicit (start_fn),
|
3860 |
|
|
6, t0, t1, t2, t, t3, t4);
|
3861 |
|
|
}
|
3862 |
|
|
else
|
3863 |
|
|
t = build_call_expr (builtin_decl_explicit (start_fn),
|
3864 |
|
|
5, t0, t1, t2, t3, t4);
|
3865 |
|
|
}
|
3866 |
|
|
else
|
3867 |
|
|
{
|
3868 |
|
|
tree t5;
|
3869 |
|
|
tree c_bool_type;
|
3870 |
|
|
tree bfn_decl;
|
3871 |
|
|
|
3872 |
|
|
/* The GOMP_loop_ull_*start functions have additional boolean
|
3873 |
|
|
argument, true for < loops and false for > loops.
|
3874 |
|
|
In Fortran, the C bool type can be different from
|
3875 |
|
|
boolean_type_node. */
|
3876 |
|
|
bfn_decl = builtin_decl_explicit (start_fn);
|
3877 |
|
|
c_bool_type = TREE_TYPE (TREE_TYPE (bfn_decl));
|
3878 |
|
|
t5 = build_int_cst (c_bool_type,
|
3879 |
|
|
fd->loop.cond_code == LT_EXPR ? 1 : 0);
|
3880 |
|
|
if (fd->chunk_size)
|
3881 |
|
|
{
|
3882 |
|
|
tree bfn_decl = builtin_decl_explicit (start_fn);
|
3883 |
|
|
t = fold_convert (fd->iter_type, fd->chunk_size);
|
3884 |
|
|
t = build_call_expr (bfn_decl, 7, t5, t0, t1, t2, t, t3, t4);
|
3885 |
|
|
}
|
3886 |
|
|
else
|
3887 |
|
|
t = build_call_expr (builtin_decl_explicit (start_fn),
|
3888 |
|
|
6, t5, t0, t1, t2, t3, t4);
|
3889 |
|
|
}
|
3890 |
|
|
}
|
3891 |
|
|
if (TREE_TYPE (t) != boolean_type_node)
|
3892 |
|
|
t = fold_build2 (NE_EXPR, boolean_type_node,
|
3893 |
|
|
t, build_int_cst (TREE_TYPE (t), 0));
|
3894 |
|
|
t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
|
3895 |
|
|
true, GSI_SAME_STMT);
|
3896 |
|
|
gsi_insert_after (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT);
|
3897 |
|
|
|
3898 |
|
|
/* Remove the GIMPLE_OMP_FOR statement. */
|
3899 |
|
|
gsi_remove (&gsi, true);
|
3900 |
|
|
|
3901 |
|
|
/* Iteration setup for sequential loop goes in L0_BB. */
|
3902 |
|
|
gsi = gsi_start_bb (l0_bb);
|
3903 |
|
|
t = istart0;
|
3904 |
|
|
if (bias)
|
3905 |
|
|
t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias);
|
3906 |
|
|
if (POINTER_TYPE_P (type))
|
3907 |
|
|
t = fold_convert (lang_hooks.types.type_for_size (TYPE_PRECISION (type),
|
3908 |
|
|
0), t);
|
3909 |
|
|
t = fold_convert (type, t);
|
3910 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE,
|
3911 |
|
|
false, GSI_CONTINUE_LINKING);
|
3912 |
|
|
stmt = gimple_build_assign (fd->loop.v, t);
|
3913 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
3914 |
|
|
|
3915 |
|
|
t = iend0;
|
3916 |
|
|
if (bias)
|
3917 |
|
|
t = fold_build2 (MINUS_EXPR, fd->iter_type, t, bias);
|
3918 |
|
|
if (POINTER_TYPE_P (type))
|
3919 |
|
|
t = fold_convert (lang_hooks.types.type_for_size (TYPE_PRECISION (type),
|
3920 |
|
|
0), t);
|
3921 |
|
|
t = fold_convert (type, t);
|
3922 |
|
|
iend = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
|
3923 |
|
|
false, GSI_CONTINUE_LINKING);
|
3924 |
|
|
if (fd->collapse > 1)
|
3925 |
|
|
{
|
3926 |
|
|
tree tem = create_tmp_var (type, ".tem");
|
3927 |
|
|
|
3928 |
|
|
stmt = gimple_build_assign (tem, fd->loop.v);
|
3929 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
3930 |
|
|
for (i = fd->collapse - 1; i >= 0; i--)
|
3931 |
|
|
{
|
3932 |
|
|
tree vtype = TREE_TYPE (fd->loops[i].v), itype;
|
3933 |
|
|
itype = vtype;
|
3934 |
|
|
if (POINTER_TYPE_P (vtype))
|
3935 |
|
|
itype = lang_hooks.types.type_for_size (TYPE_PRECISION (vtype), 0);
|
3936 |
|
|
t = fold_build2 (TRUNC_MOD_EXPR, type, tem, counts[i]);
|
3937 |
|
|
t = fold_convert (itype, t);
|
3938 |
|
|
t = fold_build2 (MULT_EXPR, itype, t,
|
3939 |
|
|
fold_convert (itype, fd->loops[i].step));
|
3940 |
|
|
if (POINTER_TYPE_P (vtype))
|
3941 |
|
|
t = fold_build_pointer_plus (fd->loops[i].n1, t);
|
3942 |
|
|
else
|
3943 |
|
|
t = fold_build2 (PLUS_EXPR, itype, fd->loops[i].n1, t);
|
3944 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE,
|
3945 |
|
|
false, GSI_CONTINUE_LINKING);
|
3946 |
|
|
stmt = gimple_build_assign (fd->loops[i].v, t);
|
3947 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
3948 |
|
|
if (i != 0)
|
3949 |
|
|
{
|
3950 |
|
|
t = fold_build2 (TRUNC_DIV_EXPR, type, tem, counts[i]);
|
3951 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE,
|
3952 |
|
|
false, GSI_CONTINUE_LINKING);
|
3953 |
|
|
stmt = gimple_build_assign (tem, t);
|
3954 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
3955 |
|
|
}
|
3956 |
|
|
}
|
3957 |
|
|
}
|
3958 |
|
|
|
3959 |
|
|
if (!broken_loop)
|
3960 |
|
|
{
|
3961 |
|
|
/* Code to control the increment and predicate for the sequential
|
3962 |
|
|
loop goes in the CONT_BB. */
|
3963 |
|
|
gsi = gsi_last_bb (cont_bb);
|
3964 |
|
|
stmt = gsi_stmt (gsi);
|
3965 |
|
|
gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE);
|
3966 |
|
|
vmain = gimple_omp_continue_control_use (stmt);
|
3967 |
|
|
vback = gimple_omp_continue_control_def (stmt);
|
3968 |
|
|
|
3969 |
|
|
if (POINTER_TYPE_P (type))
|
3970 |
|
|
t = fold_build_pointer_plus (vmain, fd->loop.step);
|
3971 |
|
|
else
|
3972 |
|
|
t = fold_build2 (PLUS_EXPR, type, vmain, fd->loop.step);
|
3973 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE,
|
3974 |
|
|
true, GSI_SAME_STMT);
|
3975 |
|
|
stmt = gimple_build_assign (vback, t);
|
3976 |
|
|
gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
|
3977 |
|
|
|
3978 |
|
|
t = build2 (fd->loop.cond_code, boolean_type_node, vback, iend);
|
3979 |
|
|
stmt = gimple_build_cond_empty (t);
|
3980 |
|
|
gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
|
3981 |
|
|
|
3982 |
|
|
/* Remove GIMPLE_OMP_CONTINUE. */
|
3983 |
|
|
gsi_remove (&gsi, true);
|
3984 |
|
|
|
3985 |
|
|
if (fd->collapse > 1)
|
3986 |
|
|
{
|
3987 |
|
|
basic_block last_bb, bb;
|
3988 |
|
|
|
3989 |
|
|
last_bb = cont_bb;
|
3990 |
|
|
for (i = fd->collapse - 1; i >= 0; i--)
|
3991 |
|
|
{
|
3992 |
|
|
tree vtype = TREE_TYPE (fd->loops[i].v);
|
3993 |
|
|
|
3994 |
|
|
bb = create_empty_bb (last_bb);
|
3995 |
|
|
gsi = gsi_start_bb (bb);
|
3996 |
|
|
|
3997 |
|
|
if (i < fd->collapse - 1)
|
3998 |
|
|
{
|
3999 |
|
|
e = make_edge (last_bb, bb, EDGE_FALSE_VALUE);
|
4000 |
|
|
e->probability = REG_BR_PROB_BASE / 8;
|
4001 |
|
|
|
4002 |
|
|
t = fd->loops[i + 1].n1;
|
4003 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE,
|
4004 |
|
|
false, GSI_CONTINUE_LINKING);
|
4005 |
|
|
stmt = gimple_build_assign (fd->loops[i + 1].v, t);
|
4006 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
4007 |
|
|
}
|
4008 |
|
|
else
|
4009 |
|
|
collapse_bb = bb;
|
4010 |
|
|
|
4011 |
|
|
set_immediate_dominator (CDI_DOMINATORS, bb, last_bb);
|
4012 |
|
|
|
4013 |
|
|
if (POINTER_TYPE_P (vtype))
|
4014 |
|
|
t = fold_build_pointer_plus (fd->loops[i].v, fd->loops[i].step);
|
4015 |
|
|
else
|
4016 |
|
|
t = fold_build2 (PLUS_EXPR, vtype, fd->loops[i].v,
|
4017 |
|
|
fd->loops[i].step);
|
4018 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE,
|
4019 |
|
|
false, GSI_CONTINUE_LINKING);
|
4020 |
|
|
stmt = gimple_build_assign (fd->loops[i].v, t);
|
4021 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
4022 |
|
|
|
4023 |
|
|
if (i > 0)
|
4024 |
|
|
{
|
4025 |
|
|
t = fd->loops[i].n2;
|
4026 |
|
|
t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
|
4027 |
|
|
false, GSI_CONTINUE_LINKING);
|
4028 |
|
|
t = fold_build2 (fd->loops[i].cond_code, boolean_type_node,
|
4029 |
|
|
fd->loops[i].v, t);
|
4030 |
|
|
stmt = gimple_build_cond_empty (t);
|
4031 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
4032 |
|
|
e = make_edge (bb, l1_bb, EDGE_TRUE_VALUE);
|
4033 |
|
|
e->probability = REG_BR_PROB_BASE * 7 / 8;
|
4034 |
|
|
}
|
4035 |
|
|
else
|
4036 |
|
|
make_edge (bb, l1_bb, EDGE_FALLTHRU);
|
4037 |
|
|
last_bb = bb;
|
4038 |
|
|
}
|
4039 |
|
|
}
|
4040 |
|
|
|
4041 |
|
|
/* Emit code to get the next parallel iteration in L2_BB. */
|
4042 |
|
|
gsi = gsi_start_bb (l2_bb);
|
4043 |
|
|
|
4044 |
|
|
t = build_call_expr (builtin_decl_explicit (next_fn), 2,
|
4045 |
|
|
build_fold_addr_expr (istart0),
|
4046 |
|
|
build_fold_addr_expr (iend0));
|
4047 |
|
|
t = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
|
4048 |
|
|
false, GSI_CONTINUE_LINKING);
|
4049 |
|
|
if (TREE_TYPE (t) != boolean_type_node)
|
4050 |
|
|
t = fold_build2 (NE_EXPR, boolean_type_node,
|
4051 |
|
|
t, build_int_cst (TREE_TYPE (t), 0));
|
4052 |
|
|
stmt = gimple_build_cond_empty (t);
|
4053 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
4054 |
|
|
}
|
4055 |
|
|
|
4056 |
|
|
/* Add the loop cleanup function. */
|
4057 |
|
|
gsi = gsi_last_bb (exit_bb);
|
4058 |
|
|
if (gimple_omp_return_nowait_p (gsi_stmt (gsi)))
|
4059 |
|
|
t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END_NOWAIT);
|
4060 |
|
|
else
|
4061 |
|
|
t = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_END);
|
4062 |
|
|
stmt = gimple_build_call (t, 0);
|
4063 |
|
|
gsi_insert_after (&gsi, stmt, GSI_SAME_STMT);
|
4064 |
|
|
gsi_remove (&gsi, true);
|
4065 |
|
|
|
4066 |
|
|
/* Connect the new blocks. */
|
4067 |
|
|
find_edge (entry_bb, l0_bb)->flags = EDGE_TRUE_VALUE;
|
4068 |
|
|
find_edge (entry_bb, l3_bb)->flags = EDGE_FALSE_VALUE;
|
4069 |
|
|
|
4070 |
|
|
if (!broken_loop)
|
4071 |
|
|
{
|
4072 |
|
|
gimple_seq phis;
|
4073 |
|
|
|
4074 |
|
|
e = find_edge (cont_bb, l3_bb);
|
4075 |
|
|
ne = make_edge (l2_bb, l3_bb, EDGE_FALSE_VALUE);
|
4076 |
|
|
|
4077 |
|
|
phis = phi_nodes (l3_bb);
|
4078 |
|
|
for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi))
|
4079 |
|
|
{
|
4080 |
|
|
gimple phi = gsi_stmt (gsi);
|
4081 |
|
|
SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, ne),
|
4082 |
|
|
PHI_ARG_DEF_FROM_EDGE (phi, e));
|
4083 |
|
|
}
|
4084 |
|
|
remove_edge (e);
|
4085 |
|
|
|
4086 |
|
|
make_edge (cont_bb, l2_bb, EDGE_FALSE_VALUE);
|
4087 |
|
|
if (fd->collapse > 1)
|
4088 |
|
|
{
|
4089 |
|
|
e = find_edge (cont_bb, l1_bb);
|
4090 |
|
|
remove_edge (e);
|
4091 |
|
|
e = make_edge (cont_bb, collapse_bb, EDGE_TRUE_VALUE);
|
4092 |
|
|
}
|
4093 |
|
|
else
|
4094 |
|
|
{
|
4095 |
|
|
e = find_edge (cont_bb, l1_bb);
|
4096 |
|
|
e->flags = EDGE_TRUE_VALUE;
|
4097 |
|
|
}
|
4098 |
|
|
e->probability = REG_BR_PROB_BASE * 7 / 8;
|
4099 |
|
|
find_edge (cont_bb, l2_bb)->probability = REG_BR_PROB_BASE / 8;
|
4100 |
|
|
make_edge (l2_bb, l0_bb, EDGE_TRUE_VALUE);
|
4101 |
|
|
|
4102 |
|
|
set_immediate_dominator (CDI_DOMINATORS, l2_bb,
|
4103 |
|
|
recompute_dominator (CDI_DOMINATORS, l2_bb));
|
4104 |
|
|
set_immediate_dominator (CDI_DOMINATORS, l3_bb,
|
4105 |
|
|
recompute_dominator (CDI_DOMINATORS, l3_bb));
|
4106 |
|
|
set_immediate_dominator (CDI_DOMINATORS, l0_bb,
|
4107 |
|
|
recompute_dominator (CDI_DOMINATORS, l0_bb));
|
4108 |
|
|
set_immediate_dominator (CDI_DOMINATORS, l1_bb,
|
4109 |
|
|
recompute_dominator (CDI_DOMINATORS, l1_bb));
|
4110 |
|
|
}
|
4111 |
|
|
}
|
4112 |
|
|
|
4113 |
|
|
|
4114 |
|
|
/* A subroutine of expand_omp_for. Generate code for a parallel
|
4115 |
|
|
loop with static schedule and no specified chunk size. Given
|
4116 |
|
|
parameters:
|
4117 |
|
|
|
4118 |
|
|
for (V = N1; V cond N2; V += STEP) BODY;
|
4119 |
|
|
|
4120 |
|
|
where COND is "<" or ">", we generate pseudocode
|
4121 |
|
|
|
4122 |
|
|
if (cond is <)
|
4123 |
|
|
adj = STEP - 1;
|
4124 |
|
|
else
|
4125 |
|
|
adj = STEP + 1;
|
4126 |
|
|
if ((__typeof (V)) -1 > 0 && cond is >)
|
4127 |
|
|
n = -(adj + N2 - N1) / -STEP;
|
4128 |
|
|
else
|
4129 |
|
|
n = (adj + N2 - N1) / STEP;
|
4130 |
|
|
q = n / nthreads;
|
4131 |
|
|
tt = n % nthreads;
|
4132 |
|
|
if (threadid < tt) goto L3; else goto L4;
|
4133 |
|
|
L3:
|
4134 |
|
|
tt = 0;
|
4135 |
|
|
q = q + 1;
|
4136 |
|
|
L4:
|
4137 |
|
|
s0 = q * threadid + tt;
|
4138 |
|
|
e0 = s0 + q;
|
4139 |
|
|
V = s0 * STEP + N1;
|
4140 |
|
|
if (s0 >= e0) goto L2; else goto L0;
|
4141 |
|
|
L0:
|
4142 |
|
|
e = e0 * STEP + N1;
|
4143 |
|
|
L1:
|
4144 |
|
|
BODY;
|
4145 |
|
|
V += STEP;
|
4146 |
|
|
if (V cond e) goto L1;
|
4147 |
|
|
L2:
|
4148 |
|
|
*/
|
4149 |
|
|
|
4150 |
|
|
static void
|
4151 |
|
|
expand_omp_for_static_nochunk (struct omp_region *region,
|
4152 |
|
|
struct omp_for_data *fd)
|
4153 |
|
|
{
|
4154 |
|
|
tree n, q, s0, e0, e, t, tt, nthreads, threadid;
|
4155 |
|
|
tree type, itype, vmain, vback;
|
4156 |
|
|
basic_block entry_bb, second_bb, third_bb, exit_bb, seq_start_bb;
|
4157 |
|
|
basic_block body_bb, cont_bb;
|
4158 |
|
|
basic_block fin_bb;
|
4159 |
|
|
gimple_stmt_iterator gsi;
|
4160 |
|
|
gimple stmt;
|
4161 |
|
|
edge ep;
|
4162 |
|
|
|
4163 |
|
|
itype = type = TREE_TYPE (fd->loop.v);
|
4164 |
|
|
if (POINTER_TYPE_P (type))
|
4165 |
|
|
itype = lang_hooks.types.type_for_size (TYPE_PRECISION (type), 0);
|
4166 |
|
|
|
4167 |
|
|
entry_bb = region->entry;
|
4168 |
|
|
cont_bb = region->cont;
|
4169 |
|
|
gcc_assert (EDGE_COUNT (entry_bb->succs) == 2);
|
4170 |
|
|
gcc_assert (BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest);
|
4171 |
|
|
seq_start_bb = split_edge (FALLTHRU_EDGE (entry_bb));
|
4172 |
|
|
body_bb = single_succ (seq_start_bb);
|
4173 |
|
|
gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb);
|
4174 |
|
|
gcc_assert (EDGE_COUNT (cont_bb->succs) == 2);
|
4175 |
|
|
fin_bb = FALLTHRU_EDGE (cont_bb)->dest;
|
4176 |
|
|
exit_bb = region->exit;
|
4177 |
|
|
|
4178 |
|
|
/* Iteration space partitioning goes in ENTRY_BB. */
|
4179 |
|
|
gsi = gsi_last_bb (entry_bb);
|
4180 |
|
|
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR);
|
4181 |
|
|
|
4182 |
|
|
t = build_call_expr (builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS), 0);
|
4183 |
|
|
t = fold_convert (itype, t);
|
4184 |
|
|
nthreads = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
|
4185 |
|
|
true, GSI_SAME_STMT);
|
4186 |
|
|
|
4187 |
|
|
t = build_call_expr (builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM), 0);
|
4188 |
|
|
t = fold_convert (itype, t);
|
4189 |
|
|
threadid = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
|
4190 |
|
|
true, GSI_SAME_STMT);
|
4191 |
|
|
|
4192 |
|
|
fd->loop.n1
|
4193 |
|
|
= force_gimple_operand_gsi (&gsi, fold_convert (type, fd->loop.n1),
|
4194 |
|
|
true, NULL_TREE, true, GSI_SAME_STMT);
|
4195 |
|
|
fd->loop.n2
|
4196 |
|
|
= force_gimple_operand_gsi (&gsi, fold_convert (itype, fd->loop.n2),
|
4197 |
|
|
true, NULL_TREE, true, GSI_SAME_STMT);
|
4198 |
|
|
fd->loop.step
|
4199 |
|
|
= force_gimple_operand_gsi (&gsi, fold_convert (itype, fd->loop.step),
|
4200 |
|
|
true, NULL_TREE, true, GSI_SAME_STMT);
|
4201 |
|
|
|
4202 |
|
|
t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1));
|
4203 |
|
|
t = fold_build2 (PLUS_EXPR, itype, fd->loop.step, t);
|
4204 |
|
|
t = fold_build2 (PLUS_EXPR, itype, t, fd->loop.n2);
|
4205 |
|
|
t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, fd->loop.n1));
|
4206 |
|
|
if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR)
|
4207 |
|
|
t = fold_build2 (TRUNC_DIV_EXPR, itype,
|
4208 |
|
|
fold_build1 (NEGATE_EXPR, itype, t),
|
4209 |
|
|
fold_build1 (NEGATE_EXPR, itype, fd->loop.step));
|
4210 |
|
|
else
|
4211 |
|
|
t = fold_build2 (TRUNC_DIV_EXPR, itype, t, fd->loop.step);
|
4212 |
|
|
t = fold_convert (itype, t);
|
4213 |
|
|
n = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT);
|
4214 |
|
|
|
4215 |
|
|
q = create_tmp_var (itype, "q");
|
4216 |
|
|
t = fold_build2 (TRUNC_DIV_EXPR, itype, n, nthreads);
|
4217 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT);
|
4218 |
|
|
gsi_insert_before (&gsi, gimple_build_assign (q, t), GSI_SAME_STMT);
|
4219 |
|
|
|
4220 |
|
|
tt = create_tmp_var (itype, "tt");
|
4221 |
|
|
t = fold_build2 (TRUNC_MOD_EXPR, itype, n, nthreads);
|
4222 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE, true, GSI_SAME_STMT);
|
4223 |
|
|
gsi_insert_before (&gsi, gimple_build_assign (tt, t), GSI_SAME_STMT);
|
4224 |
|
|
|
4225 |
|
|
t = build2 (LT_EXPR, boolean_type_node, threadid, tt);
|
4226 |
|
|
stmt = gimple_build_cond_empty (t);
|
4227 |
|
|
gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
|
4228 |
|
|
|
4229 |
|
|
second_bb = split_block (entry_bb, stmt)->dest;
|
4230 |
|
|
gsi = gsi_last_bb (second_bb);
|
4231 |
|
|
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR);
|
4232 |
|
|
|
4233 |
|
|
gsi_insert_before (&gsi, gimple_build_assign (tt, build_int_cst (itype, 0)),
|
4234 |
|
|
GSI_SAME_STMT);
|
4235 |
|
|
stmt = gimple_build_assign_with_ops (PLUS_EXPR, q, q,
|
4236 |
|
|
build_int_cst (itype, 1));
|
4237 |
|
|
gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
|
4238 |
|
|
|
4239 |
|
|
third_bb = split_block (second_bb, stmt)->dest;
|
4240 |
|
|
gsi = gsi_last_bb (third_bb);
|
4241 |
|
|
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_FOR);
|
4242 |
|
|
|
4243 |
|
|
t = build2 (MULT_EXPR, itype, q, threadid);
|
4244 |
|
|
t = build2 (PLUS_EXPR, itype, t, tt);
|
4245 |
|
|
s0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT);
|
4246 |
|
|
|
4247 |
|
|
t = fold_build2 (PLUS_EXPR, itype, s0, q);
|
4248 |
|
|
e0 = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE, true, GSI_SAME_STMT);
|
4249 |
|
|
|
4250 |
|
|
t = build2 (GE_EXPR, boolean_type_node, s0, e0);
|
4251 |
|
|
gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT);
|
4252 |
|
|
|
4253 |
|
|
/* Remove the GIMPLE_OMP_FOR statement. */
|
4254 |
|
|
gsi_remove (&gsi, true);
|
4255 |
|
|
|
4256 |
|
|
/* Setup code for sequential iteration goes in SEQ_START_BB. */
|
4257 |
|
|
gsi = gsi_start_bb (seq_start_bb);
|
4258 |
|
|
|
4259 |
|
|
t = fold_convert (itype, s0);
|
4260 |
|
|
t = fold_build2 (MULT_EXPR, itype, t, fd->loop.step);
|
4261 |
|
|
if (POINTER_TYPE_P (type))
|
4262 |
|
|
t = fold_build_pointer_plus (fd->loop.n1, t);
|
4263 |
|
|
else
|
4264 |
|
|
t = fold_build2 (PLUS_EXPR, type, t, fd->loop.n1);
|
4265 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE,
|
4266 |
|
|
false, GSI_CONTINUE_LINKING);
|
4267 |
|
|
stmt = gimple_build_assign (fd->loop.v, t);
|
4268 |
|
|
gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
|
4269 |
|
|
|
4270 |
|
|
t = fold_convert (itype, e0);
|
4271 |
|
|
t = fold_build2 (MULT_EXPR, itype, t, fd->loop.step);
|
4272 |
|
|
if (POINTER_TYPE_P (type))
|
4273 |
|
|
t = fold_build_pointer_plus (fd->loop.n1, t);
|
4274 |
|
|
else
|
4275 |
|
|
t = fold_build2 (PLUS_EXPR, type, t, fd->loop.n1);
|
4276 |
|
|
e = force_gimple_operand_gsi (&gsi, t, true, NULL_TREE,
|
4277 |
|
|
false, GSI_CONTINUE_LINKING);
|
4278 |
|
|
|
4279 |
|
|
/* The code controlling the sequential loop replaces the
|
4280 |
|
|
GIMPLE_OMP_CONTINUE. */
|
4281 |
|
|
gsi = gsi_last_bb (cont_bb);
|
4282 |
|
|
stmt = gsi_stmt (gsi);
|
4283 |
|
|
gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE);
|
4284 |
|
|
vmain = gimple_omp_continue_control_use (stmt);
|
4285 |
|
|
vback = gimple_omp_continue_control_def (stmt);
|
4286 |
|
|
|
4287 |
|
|
if (POINTER_TYPE_P (type))
|
4288 |
|
|
t = fold_build_pointer_plus (vmain, fd->loop.step);
|
4289 |
|
|
else
|
4290 |
|
|
t = fold_build2 (PLUS_EXPR, type, vmain, fd->loop.step);
|
4291 |
|
|
t = force_gimple_operand_gsi (&gsi, t, false, NULL_TREE,
|
4292 |
|
|
true, GSI_SAME_STMT);
|
4293 |
|
|
stmt = gimple_build_assign (vback, t);
|
4294 |
|
|
gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
|
4295 |
|
|
|
4296 |
|
|
t = build2 (fd->loop.cond_code, boolean_type_node, vback, e);
|
4297 |
|
|
gsi_insert_before (&gsi, gimple_build_cond_empty (t), GSI_SAME_STMT);
|
4298 |
|
|
|
4299 |
|
|
/* Remove the GIMPLE_OMP_CONTINUE statement. */
|
4300 |
|
|
gsi_remove (&gsi, true);
|
4301 |
|
|
|
4302 |
|
|
/* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */
|
4303 |
|
|
gsi = gsi_last_bb (exit_bb);
|
4304 |
|
|
if (!gimple_omp_return_nowait_p (gsi_stmt (gsi)))
|
4305 |
|
|
force_gimple_operand_gsi (&gsi, build_omp_barrier (), false, NULL_TREE,
|
4306 |
|
|
false, GSI_SAME_STMT);
|
4307 |
|
|
gsi_remove (&gsi, true);
|
4308 |
|
|
|
4309 |
|
|
/* Connect all the blocks. */
|
4310 |
|
|
ep = make_edge (entry_bb, third_bb, EDGE_FALSE_VALUE);
|
4311 |
|
|
ep->probability = REG_BR_PROB_BASE / 4 * 3;
|
4312 |
|
|
ep = find_edge (entry_bb, second_bb);
|
4313 |
|
|
ep->flags = EDGE_TRUE_VALUE;
|
4314 |
|
|
ep->probability = REG_BR_PROB_BASE / 4;
|
4315 |
|
|
find_edge (third_bb, seq_start_bb)->flags = EDGE_FALSE_VALUE;
|
4316 |
|
|
find_edge (third_bb, fin_bb)->flags = EDGE_TRUE_VALUE;
|
4317 |
|
|
|
4318 |
|
|
find_edge (cont_bb, body_bb)->flags = EDGE_TRUE_VALUE;
|
4319 |
|
|
find_edge (cont_bb, fin_bb)->flags = EDGE_FALSE_VALUE;
|
4320 |
|
|
|
4321 |
|
|
set_immediate_dominator (CDI_DOMINATORS, second_bb, entry_bb);
|
4322 |
|
|
set_immediate_dominator (CDI_DOMINATORS, third_bb, entry_bb);
|
4323 |
|
|
set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, third_bb);
|
4324 |
|
|
set_immediate_dominator (CDI_DOMINATORS, body_bb,
|
4325 |
|
|
recompute_dominator (CDI_DOMINATORS, body_bb));
|
4326 |
|
|
set_immediate_dominator (CDI_DOMINATORS, fin_bb,
|
4327 |
|
|
recompute_dominator (CDI_DOMINATORS, fin_bb));
|
4328 |
|
|
}
|
4329 |
|
|
|
4330 |
|
|
|
4331 |
|
|
/* A subroutine of expand_omp_for. Generate code for a parallel
|
4332 |
|
|
loop with static schedule and a specified chunk size. Given
|
4333 |
|
|
parameters:
|
4334 |
|
|
|
4335 |
|
|
for (V = N1; V cond N2; V += STEP) BODY;
|
4336 |
|
|
|
4337 |
|
|
where COND is "<" or ">", we generate pseudocode
|
4338 |
|
|
|
4339 |
|
|
if (cond is <)
|
4340 |
|
|
adj = STEP - 1;
|
4341 |
|
|
else
|
4342 |
|
|
adj = STEP + 1;
|
4343 |
|
|
if ((__typeof (V)) -1 > 0 && cond is >)
|
4344 |
|
|
n = -(adj + N2 - N1) / -STEP;
|
4345 |
|
|
else
|
4346 |
|
|
n = (adj + N2 - N1) / STEP;
|
4347 |
|
|
trip = 0;
|
4348 |
|
|
V = threadid * CHUNK * STEP + N1; -- this extra definition of V is
|
4349 |
|
|
here so that V is defined
|
4350 |
|
|
if the loop is not entered
|
4351 |
|
|
L0:
|
4352 |
|
|
s0 = (trip * nthreads + threadid) * CHUNK;
|
4353 |
|
|
e0 = min(s0 + CHUNK, n);
|
4354 |
|
|
if (s0 < n) goto L1; else goto L4;
|
4355 |
|
|
L1:
|
4356 |
|
|
V = s0 * STEP + N1;
|
4357 |
|
|
e = e0 * STEP + N1;
|
4358 |
|
|
L2:
|
4359 |
|
|
BODY;
|
4360 |
|
|
V += STEP;
|
4361 |
|
|
if (V cond e) goto L2; else goto L3;
|
4362 |
|
|
L3:
|
4363 |
|
|
trip += 1;
|
4364 |
|
|
goto L0;
|
4365 |
|
|
L4:
|
4366 |
|
|
*/
|
4367 |
|
|
|
4368 |
|
|
static void
|
4369 |
|
|
expand_omp_for_static_chunk (struct omp_region *region, struct omp_for_data *fd)
|
4370 |
|
|
{
|
4371 |
|
|
tree n, s0, e0, e, t;
|
4372 |
|
|
tree trip_var, trip_init, trip_main, trip_back, nthreads, threadid;
|
4373 |
|
|
tree type, itype, v_main, v_back, v_extra;
|
4374 |
|
|
basic_block entry_bb, exit_bb, body_bb, seq_start_bb, iter_part_bb;
|
4375 |
|
|
basic_block trip_update_bb, cont_bb, fin_bb;
|
4376 |
|
|
gimple_stmt_iterator si;
|
4377 |
|
|
gimple stmt;
|
4378 |
|
|
edge se;
|
4379 |
|
|
|
4380 |
|
|
itype = type = TREE_TYPE (fd->loop.v);
|
4381 |
|
|
if (POINTER_TYPE_P (type))
|
4382 |
|
|
itype = lang_hooks.types.type_for_size (TYPE_PRECISION (type), 0);
|
4383 |
|
|
|
4384 |
|
|
entry_bb = region->entry;
|
4385 |
|
|
se = split_block (entry_bb, last_stmt (entry_bb));
|
4386 |
|
|
entry_bb = se->src;
|
4387 |
|
|
iter_part_bb = se->dest;
|
4388 |
|
|
cont_bb = region->cont;
|
4389 |
|
|
gcc_assert (EDGE_COUNT (iter_part_bb->succs) == 2);
|
4390 |
|
|
gcc_assert (BRANCH_EDGE (iter_part_bb)->dest
|
4391 |
|
|
== FALLTHRU_EDGE (cont_bb)->dest);
|
4392 |
|
|
seq_start_bb = split_edge (FALLTHRU_EDGE (iter_part_bb));
|
4393 |
|
|
body_bb = single_succ (seq_start_bb);
|
4394 |
|
|
gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb);
|
4395 |
|
|
gcc_assert (EDGE_COUNT (cont_bb->succs) == 2);
|
4396 |
|
|
fin_bb = FALLTHRU_EDGE (cont_bb)->dest;
|
4397 |
|
|
trip_update_bb = split_edge (FALLTHRU_EDGE (cont_bb));
|
4398 |
|
|
exit_bb = region->exit;
|
4399 |
|
|
|
4400 |
|
|
/* Trip and adjustment setup goes in ENTRY_BB. */
|
4401 |
|
|
si = gsi_last_bb (entry_bb);
|
4402 |
|
|
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_FOR);
|
4403 |
|
|
|
4404 |
|
|
t = build_call_expr (builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS), 0);
|
4405 |
|
|
t = fold_convert (itype, t);
|
4406 |
|
|
nthreads = force_gimple_operand_gsi (&si, t, true, NULL_TREE,
|
4407 |
|
|
true, GSI_SAME_STMT);
|
4408 |
|
|
|
4409 |
|
|
t = build_call_expr (builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM), 0);
|
4410 |
|
|
t = fold_convert (itype, t);
|
4411 |
|
|
threadid = force_gimple_operand_gsi (&si, t, true, NULL_TREE,
|
4412 |
|
|
true, GSI_SAME_STMT);
|
4413 |
|
|
|
4414 |
|
|
fd->loop.n1
|
4415 |
|
|
= force_gimple_operand_gsi (&si, fold_convert (type, fd->loop.n1),
|
4416 |
|
|
true, NULL_TREE, true, GSI_SAME_STMT);
|
4417 |
|
|
fd->loop.n2
|
4418 |
|
|
= force_gimple_operand_gsi (&si, fold_convert (itype, fd->loop.n2),
|
4419 |
|
|
true, NULL_TREE, true, GSI_SAME_STMT);
|
4420 |
|
|
fd->loop.step
|
4421 |
|
|
= force_gimple_operand_gsi (&si, fold_convert (itype, fd->loop.step),
|
4422 |
|
|
true, NULL_TREE, true, GSI_SAME_STMT);
|
4423 |
|
|
fd->chunk_size
|
4424 |
|
|
= force_gimple_operand_gsi (&si, fold_convert (itype, fd->chunk_size),
|
4425 |
|
|
true, NULL_TREE, true, GSI_SAME_STMT);
|
4426 |
|
|
|
4427 |
|
|
t = build_int_cst (itype, (fd->loop.cond_code == LT_EXPR ? -1 : 1));
|
4428 |
|
|
t = fold_build2 (PLUS_EXPR, itype, fd->loop.step, t);
|
4429 |
|
|
t = fold_build2 (PLUS_EXPR, itype, t, fd->loop.n2);
|
4430 |
|
|
t = fold_build2 (MINUS_EXPR, itype, t, fold_convert (itype, fd->loop.n1));
|
4431 |
|
|
if (TYPE_UNSIGNED (itype) && fd->loop.cond_code == GT_EXPR)
|
4432 |
|
|
t = fold_build2 (TRUNC_DIV_EXPR, itype,
|
4433 |
|
|
fold_build1 (NEGATE_EXPR, itype, t),
|
4434 |
|
|
fold_build1 (NEGATE_EXPR, itype, fd->loop.step));
|
4435 |
|
|
else
|
4436 |
|
|
t = fold_build2 (TRUNC_DIV_EXPR, itype, t, fd->loop.step);
|
4437 |
|
|
t = fold_convert (itype, t);
|
4438 |
|
|
n = force_gimple_operand_gsi (&si, t, true, NULL_TREE,
|
4439 |
|
|
true, GSI_SAME_STMT);
|
4440 |
|
|
|
4441 |
|
|
trip_var = create_tmp_var (itype, ".trip");
|
4442 |
|
|
if (gimple_in_ssa_p (cfun))
|
4443 |
|
|
{
|
4444 |
|
|
add_referenced_var (trip_var);
|
4445 |
|
|
trip_init = make_ssa_name (trip_var, NULL);
|
4446 |
|
|
trip_main = make_ssa_name (trip_var, NULL);
|
4447 |
|
|
trip_back = make_ssa_name (trip_var, NULL);
|
4448 |
|
|
}
|
4449 |
|
|
else
|
4450 |
|
|
{
|
4451 |
|
|
trip_init = trip_var;
|
4452 |
|
|
trip_main = trip_var;
|
4453 |
|
|
trip_back = trip_var;
|
4454 |
|
|
}
|
4455 |
|
|
|
4456 |
|
|
stmt = gimple_build_assign (trip_init, build_int_cst (itype, 0));
|
4457 |
|
|
gsi_insert_before (&si, stmt, GSI_SAME_STMT);
|
4458 |
|
|
|
4459 |
|
|
t = fold_build2 (MULT_EXPR, itype, threadid, fd->chunk_size);
|
4460 |
|
|
t = fold_build2 (MULT_EXPR, itype, t, fd->loop.step);
|
4461 |
|
|
if (POINTER_TYPE_P (type))
|
4462 |
|
|
t = fold_build_pointer_plus (fd->loop.n1, t);
|
4463 |
|
|
else
|
4464 |
|
|
t = fold_build2 (PLUS_EXPR, type, t, fd->loop.n1);
|
4465 |
|
|
v_extra = force_gimple_operand_gsi (&si, t, true, NULL_TREE,
|
4466 |
|
|
true, GSI_SAME_STMT);
|
4467 |
|
|
|
4468 |
|
|
/* Remove the GIMPLE_OMP_FOR. */
|
4469 |
|
|
gsi_remove (&si, true);
|
4470 |
|
|
|
4471 |
|
|
/* Iteration space partitioning goes in ITER_PART_BB. */
|
4472 |
|
|
si = gsi_last_bb (iter_part_bb);
|
4473 |
|
|
|
4474 |
|
|
t = fold_build2 (MULT_EXPR, itype, trip_main, nthreads);
|
4475 |
|
|
t = fold_build2 (PLUS_EXPR, itype, t, threadid);
|
4476 |
|
|
t = fold_build2 (MULT_EXPR, itype, t, fd->chunk_size);
|
4477 |
|
|
s0 = force_gimple_operand_gsi (&si, t, true, NULL_TREE,
|
4478 |
|
|
false, GSI_CONTINUE_LINKING);
|
4479 |
|
|
|
4480 |
|
|
t = fold_build2 (PLUS_EXPR, itype, s0, fd->chunk_size);
|
4481 |
|
|
t = fold_build2 (MIN_EXPR, itype, t, n);
|
4482 |
|
|
e0 = force_gimple_operand_gsi (&si, t, true, NULL_TREE,
|
4483 |
|
|
false, GSI_CONTINUE_LINKING);
|
4484 |
|
|
|
4485 |
|
|
t = build2 (LT_EXPR, boolean_type_node, s0, n);
|
4486 |
|
|
gsi_insert_after (&si, gimple_build_cond_empty (t), GSI_CONTINUE_LINKING);
|
4487 |
|
|
|
4488 |
|
|
/* Setup code for sequential iteration goes in SEQ_START_BB. */
|
4489 |
|
|
si = gsi_start_bb (seq_start_bb);
|
4490 |
|
|
|
4491 |
|
|
t = fold_convert (itype, s0);
|
4492 |
|
|
t = fold_build2 (MULT_EXPR, itype, t, fd->loop.step);
|
4493 |
|
|
if (POINTER_TYPE_P (type))
|
4494 |
|
|
t = fold_build_pointer_plus (fd->loop.n1, t);
|
4495 |
|
|
else
|
4496 |
|
|
t = fold_build2 (PLUS_EXPR, type, t, fd->loop.n1);
|
4497 |
|
|
t = force_gimple_operand_gsi (&si, t, false, NULL_TREE,
|
4498 |
|
|
false, GSI_CONTINUE_LINKING);
|
4499 |
|
|
stmt = gimple_build_assign (fd->loop.v, t);
|
4500 |
|
|
gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING);
|
4501 |
|
|
|
4502 |
|
|
t = fold_convert (itype, e0);
|
4503 |
|
|
t = fold_build2 (MULT_EXPR, itype, t, fd->loop.step);
|
4504 |
|
|
if (POINTER_TYPE_P (type))
|
4505 |
|
|
t = fold_build_pointer_plus (fd->loop.n1, t);
|
4506 |
|
|
else
|
4507 |
|
|
t = fold_build2 (PLUS_EXPR, type, t, fd->loop.n1);
|
4508 |
|
|
e = force_gimple_operand_gsi (&si, t, true, NULL_TREE,
|
4509 |
|
|
false, GSI_CONTINUE_LINKING);
|
4510 |
|
|
|
4511 |
|
|
/* The code controlling the sequential loop goes in CONT_BB,
|
4512 |
|
|
replacing the GIMPLE_OMP_CONTINUE. */
|
4513 |
|
|
si = gsi_last_bb (cont_bb);
|
4514 |
|
|
stmt = gsi_stmt (si);
|
4515 |
|
|
gcc_assert (gimple_code (stmt) == GIMPLE_OMP_CONTINUE);
|
4516 |
|
|
v_main = gimple_omp_continue_control_use (stmt);
|
4517 |
|
|
v_back = gimple_omp_continue_control_def (stmt);
|
4518 |
|
|
|
4519 |
|
|
if (POINTER_TYPE_P (type))
|
4520 |
|
|
t = fold_build_pointer_plus (v_main, fd->loop.step);
|
4521 |
|
|
else
|
4522 |
|
|
t = fold_build2 (PLUS_EXPR, type, v_main, fd->loop.step);
|
4523 |
|
|
stmt = gimple_build_assign (v_back, t);
|
4524 |
|
|
gsi_insert_before (&si, stmt, GSI_SAME_STMT);
|
4525 |
|
|
|
4526 |
|
|
t = build2 (fd->loop.cond_code, boolean_type_node, v_back, e);
|
4527 |
|
|
gsi_insert_before (&si, gimple_build_cond_empty (t), GSI_SAME_STMT);
|
4528 |
|
|
|
4529 |
|
|
/* Remove GIMPLE_OMP_CONTINUE. */
|
4530 |
|
|
gsi_remove (&si, true);
|
4531 |
|
|
|
4532 |
|
|
/* Trip update code goes into TRIP_UPDATE_BB. */
|
4533 |
|
|
si = gsi_start_bb (trip_update_bb);
|
4534 |
|
|
|
4535 |
|
|
t = build_int_cst (itype, 1);
|
4536 |
|
|
t = build2 (PLUS_EXPR, itype, trip_main, t);
|
4537 |
|
|
stmt = gimple_build_assign (trip_back, t);
|
4538 |
|
|
gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING);
|
4539 |
|
|
|
4540 |
|
|
/* Replace the GIMPLE_OMP_RETURN with a barrier, or nothing. */
|
4541 |
|
|
si = gsi_last_bb (exit_bb);
|
4542 |
|
|
if (!gimple_omp_return_nowait_p (gsi_stmt (si)))
|
4543 |
|
|
force_gimple_operand_gsi (&si, build_omp_barrier (), false, NULL_TREE,
|
4544 |
|
|
false, GSI_SAME_STMT);
|
4545 |
|
|
gsi_remove (&si, true);
|
4546 |
|
|
|
4547 |
|
|
/* Connect the new blocks. */
|
4548 |
|
|
find_edge (iter_part_bb, seq_start_bb)->flags = EDGE_TRUE_VALUE;
|
4549 |
|
|
find_edge (iter_part_bb, fin_bb)->flags = EDGE_FALSE_VALUE;
|
4550 |
|
|
|
4551 |
|
|
find_edge (cont_bb, body_bb)->flags = EDGE_TRUE_VALUE;
|
4552 |
|
|
find_edge (cont_bb, trip_update_bb)->flags = EDGE_FALSE_VALUE;
|
4553 |
|
|
|
4554 |
|
|
redirect_edge_and_branch (single_succ_edge (trip_update_bb), iter_part_bb);
|
4555 |
|
|
|
4556 |
|
|
if (gimple_in_ssa_p (cfun))
|
4557 |
|
|
{
|
4558 |
|
|
gimple_stmt_iterator psi;
|
4559 |
|
|
gimple phi;
|
4560 |
|
|
edge re, ene;
|
4561 |
|
|
edge_var_map_vector head;
|
4562 |
|
|
edge_var_map *vm;
|
4563 |
|
|
size_t i;
|
4564 |
|
|
|
4565 |
|
|
/* When we redirect the edge from trip_update_bb to iter_part_bb, we
|
4566 |
|
|
remove arguments of the phi nodes in fin_bb. We need to create
|
4567 |
|
|
appropriate phi nodes in iter_part_bb instead. */
|
4568 |
|
|
se = single_pred_edge (fin_bb);
|
4569 |
|
|
re = single_succ_edge (trip_update_bb);
|
4570 |
|
|
head = redirect_edge_var_map_vector (re);
|
4571 |
|
|
ene = single_succ_edge (entry_bb);
|
4572 |
|
|
|
4573 |
|
|
psi = gsi_start_phis (fin_bb);
|
4574 |
|
|
for (i = 0; !gsi_end_p (psi) && VEC_iterate (edge_var_map, head, i, vm);
|
4575 |
|
|
gsi_next (&psi), ++i)
|
4576 |
|
|
{
|
4577 |
|
|
gimple nphi;
|
4578 |
|
|
source_location locus;
|
4579 |
|
|
|
4580 |
|
|
phi = gsi_stmt (psi);
|
4581 |
|
|
t = gimple_phi_result (phi);
|
4582 |
|
|
gcc_assert (t == redirect_edge_var_map_result (vm));
|
4583 |
|
|
nphi = create_phi_node (t, iter_part_bb);
|
4584 |
|
|
SSA_NAME_DEF_STMT (t) = nphi;
|
4585 |
|
|
|
4586 |
|
|
t = PHI_ARG_DEF_FROM_EDGE (phi, se);
|
4587 |
|
|
locus = gimple_phi_arg_location_from_edge (phi, se);
|
4588 |
|
|
|
4589 |
|
|
/* A special case -- fd->loop.v is not yet computed in
|
4590 |
|
|
iter_part_bb, we need to use v_extra instead. */
|
4591 |
|
|
if (t == fd->loop.v)
|
4592 |
|
|
t = v_extra;
|
4593 |
|
|
add_phi_arg (nphi, t, ene, locus);
|
4594 |
|
|
locus = redirect_edge_var_map_location (vm);
|
4595 |
|
|
add_phi_arg (nphi, redirect_edge_var_map_def (vm), re, locus);
|
4596 |
|
|
}
|
4597 |
|
|
gcc_assert (!gsi_end_p (psi) && i == VEC_length (edge_var_map, head));
|
4598 |
|
|
redirect_edge_var_map_clear (re);
|
4599 |
|
|
while (1)
|
4600 |
|
|
{
|
4601 |
|
|
psi = gsi_start_phis (fin_bb);
|
4602 |
|
|
if (gsi_end_p (psi))
|
4603 |
|
|
break;
|
4604 |
|
|
remove_phi_node (&psi, false);
|
4605 |
|
|
}
|
4606 |
|
|
|
4607 |
|
|
/* Make phi node for trip. */
|
4608 |
|
|
phi = create_phi_node (trip_main, iter_part_bb);
|
4609 |
|
|
SSA_NAME_DEF_STMT (trip_main) = phi;
|
4610 |
|
|
add_phi_arg (phi, trip_back, single_succ_edge (trip_update_bb),
|
4611 |
|
|
UNKNOWN_LOCATION);
|
4612 |
|
|
add_phi_arg (phi, trip_init, single_succ_edge (entry_bb),
|
4613 |
|
|
UNKNOWN_LOCATION);
|
4614 |
|
|
}
|
4615 |
|
|
|
4616 |
|
|
set_immediate_dominator (CDI_DOMINATORS, trip_update_bb, cont_bb);
|
4617 |
|
|
set_immediate_dominator (CDI_DOMINATORS, iter_part_bb,
|
4618 |
|
|
recompute_dominator (CDI_DOMINATORS, iter_part_bb));
|
4619 |
|
|
set_immediate_dominator (CDI_DOMINATORS, fin_bb,
|
4620 |
|
|
recompute_dominator (CDI_DOMINATORS, fin_bb));
|
4621 |
|
|
set_immediate_dominator (CDI_DOMINATORS, seq_start_bb,
|
4622 |
|
|
recompute_dominator (CDI_DOMINATORS, seq_start_bb));
|
4623 |
|
|
set_immediate_dominator (CDI_DOMINATORS, body_bb,
|
4624 |
|
|
recompute_dominator (CDI_DOMINATORS, body_bb));
|
4625 |
|
|
}
|
4626 |
|
|
|
4627 |
|
|
|
4628 |
|
|
/* Expand the OpenMP loop defined by REGION. */
|
4629 |
|
|
|
4630 |
|
|
static void
|
4631 |
|
|
expand_omp_for (struct omp_region *region)
|
4632 |
|
|
{
|
4633 |
|
|
struct omp_for_data fd;
|
4634 |
|
|
struct omp_for_data_loop *loops;
|
4635 |
|
|
|
4636 |
|
|
loops
|
4637 |
|
|
= (struct omp_for_data_loop *)
|
4638 |
|
|
alloca (gimple_omp_for_collapse (last_stmt (region->entry))
|
4639 |
|
|
* sizeof (struct omp_for_data_loop));
|
4640 |
|
|
extract_omp_for_data (last_stmt (region->entry), &fd, loops);
|
4641 |
|
|
region->sched_kind = fd.sched_kind;
|
4642 |
|
|
|
4643 |
|
|
gcc_assert (EDGE_COUNT (region->entry->succs) == 2);
|
4644 |
|
|
BRANCH_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL;
|
4645 |
|
|
FALLTHRU_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL;
|
4646 |
|
|
if (region->cont)
|
4647 |
|
|
{
|
4648 |
|
|
gcc_assert (EDGE_COUNT (region->cont->succs) == 2);
|
4649 |
|
|
BRANCH_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL;
|
4650 |
|
|
FALLTHRU_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL;
|
4651 |
|
|
}
|
4652 |
|
|
|
4653 |
|
|
if (fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC
|
4654 |
|
|
&& !fd.have_ordered
|
4655 |
|
|
&& fd.collapse == 1
|
4656 |
|
|
&& region->cont != NULL)
|
4657 |
|
|
{
|
4658 |
|
|
if (fd.chunk_size == NULL)
|
4659 |
|
|
expand_omp_for_static_nochunk (region, &fd);
|
4660 |
|
|
else
|
4661 |
|
|
expand_omp_for_static_chunk (region, &fd);
|
4662 |
|
|
}
|
4663 |
|
|
else
|
4664 |
|
|
{
|
4665 |
|
|
int fn_index, start_ix, next_ix;
|
4666 |
|
|
|
4667 |
|
|
if (fd.chunk_size == NULL
|
4668 |
|
|
&& fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC)
|
4669 |
|
|
fd.chunk_size = integer_zero_node;
|
4670 |
|
|
gcc_assert (fd.sched_kind != OMP_CLAUSE_SCHEDULE_AUTO);
|
4671 |
|
|
fn_index = (fd.sched_kind == OMP_CLAUSE_SCHEDULE_RUNTIME)
|
4672 |
|
|
? 3 : fd.sched_kind;
|
4673 |
|
|
fn_index += fd.have_ordered * 4;
|
4674 |
|
|
start_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_START) + fn_index;
|
4675 |
|
|
next_ix = ((int)BUILT_IN_GOMP_LOOP_STATIC_NEXT) + fn_index;
|
4676 |
|
|
if (fd.iter_type == long_long_unsigned_type_node)
|
4677 |
|
|
{
|
4678 |
|
|
start_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_START
|
4679 |
|
|
- (int)BUILT_IN_GOMP_LOOP_STATIC_START);
|
4680 |
|
|
next_ix += ((int)BUILT_IN_GOMP_LOOP_ULL_STATIC_NEXT
|
4681 |
|
|
- (int)BUILT_IN_GOMP_LOOP_STATIC_NEXT);
|
4682 |
|
|
}
|
4683 |
|
|
expand_omp_for_generic (region, &fd, (enum built_in_function) start_ix,
|
4684 |
|
|
(enum built_in_function) next_ix);
|
4685 |
|
|
}
|
4686 |
|
|
|
4687 |
|
|
update_ssa (TODO_update_ssa_only_virtuals);
|
4688 |
|
|
}
|
4689 |
|
|
|
4690 |
|
|
|
4691 |
|
|
/* Expand code for an OpenMP sections directive. In pseudo code, we generate
|
4692 |
|
|
|
4693 |
|
|
v = GOMP_sections_start (n);
|
4694 |
|
|
L0:
|
4695 |
|
|
switch (v)
|
4696 |
|
|
{
|
4697 |
|
|
case 0:
|
4698 |
|
|
goto L2;
|
4699 |
|
|
case 1:
|
4700 |
|
|
section 1;
|
4701 |
|
|
goto L1;
|
4702 |
|
|
case 2:
|
4703 |
|
|
...
|
4704 |
|
|
case n:
|
4705 |
|
|
...
|
4706 |
|
|
default:
|
4707 |
|
|
abort ();
|
4708 |
|
|
}
|
4709 |
|
|
L1:
|
4710 |
|
|
v = GOMP_sections_next ();
|
4711 |
|
|
goto L0;
|
4712 |
|
|
L2:
|
4713 |
|
|
reduction;
|
4714 |
|
|
|
4715 |
|
|
If this is a combined parallel sections, replace the call to
|
4716 |
|
|
GOMP_sections_start with call to GOMP_sections_next. */
|
4717 |
|
|
|
4718 |
|
|
static void
|
4719 |
|
|
expand_omp_sections (struct omp_region *region)
|
4720 |
|
|
{
|
4721 |
|
|
tree t, u, vin = NULL, vmain, vnext, l2;
|
4722 |
|
|
VEC (tree,heap) *label_vec;
|
4723 |
|
|
unsigned len;
|
4724 |
|
|
basic_block entry_bb, l0_bb, l1_bb, l2_bb, default_bb;
|
4725 |
|
|
gimple_stmt_iterator si, switch_si;
|
4726 |
|
|
gimple sections_stmt, stmt, cont;
|
4727 |
|
|
edge_iterator ei;
|
4728 |
|
|
edge e;
|
4729 |
|
|
struct omp_region *inner;
|
4730 |
|
|
unsigned i, casei;
|
4731 |
|
|
bool exit_reachable = region->cont != NULL;
|
4732 |
|
|
|
4733 |
|
|
gcc_assert (exit_reachable == (region->exit != NULL));
|
4734 |
|
|
entry_bb = region->entry;
|
4735 |
|
|
l0_bb = single_succ (entry_bb);
|
4736 |
|
|
l1_bb = region->cont;
|
4737 |
|
|
l2_bb = region->exit;
|
4738 |
|
|
if (exit_reachable)
|
4739 |
|
|
{
|
4740 |
|
|
if (single_pred_p (l2_bb) && single_pred (l2_bb) == l0_bb)
|
4741 |
|
|
l2 = gimple_block_label (l2_bb);
|
4742 |
|
|
else
|
4743 |
|
|
{
|
4744 |
|
|
/* This can happen if there are reductions. */
|
4745 |
|
|
len = EDGE_COUNT (l0_bb->succs);
|
4746 |
|
|
gcc_assert (len > 0);
|
4747 |
|
|
e = EDGE_SUCC (l0_bb, len - 1);
|
4748 |
|
|
si = gsi_last_bb (e->dest);
|
4749 |
|
|
l2 = NULL_TREE;
|
4750 |
|
|
if (gsi_end_p (si)
|
4751 |
|
|
|| gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION)
|
4752 |
|
|
l2 = gimple_block_label (e->dest);
|
4753 |
|
|
else
|
4754 |
|
|
FOR_EACH_EDGE (e, ei, l0_bb->succs)
|
4755 |
|
|
{
|
4756 |
|
|
si = gsi_last_bb (e->dest);
|
4757 |
|
|
if (gsi_end_p (si)
|
4758 |
|
|
|| gimple_code (gsi_stmt (si)) != GIMPLE_OMP_SECTION)
|
4759 |
|
|
{
|
4760 |
|
|
l2 = gimple_block_label (e->dest);
|
4761 |
|
|
break;
|
4762 |
|
|
}
|
4763 |
|
|
}
|
4764 |
|
|
}
|
4765 |
|
|
default_bb = create_empty_bb (l1_bb->prev_bb);
|
4766 |
|
|
}
|
4767 |
|
|
else
|
4768 |
|
|
{
|
4769 |
|
|
default_bb = create_empty_bb (l0_bb);
|
4770 |
|
|
l2 = gimple_block_label (default_bb);
|
4771 |
|
|
}
|
4772 |
|
|
|
4773 |
|
|
/* We will build a switch() with enough cases for all the
|
4774 |
|
|
GIMPLE_OMP_SECTION regions, a '0' case to handle the end of more work
|
4775 |
|
|
and a default case to abort if something goes wrong. */
|
4776 |
|
|
len = EDGE_COUNT (l0_bb->succs);
|
4777 |
|
|
|
4778 |
|
|
/* Use VEC_quick_push on label_vec throughout, since we know the size
|
4779 |
|
|
in advance. */
|
4780 |
|
|
label_vec = VEC_alloc (tree, heap, len);
|
4781 |
|
|
|
4782 |
|
|
/* The call to GOMP_sections_start goes in ENTRY_BB, replacing the
|
4783 |
|
|
GIMPLE_OMP_SECTIONS statement. */
|
4784 |
|
|
si = gsi_last_bb (entry_bb);
|
4785 |
|
|
sections_stmt = gsi_stmt (si);
|
4786 |
|
|
gcc_assert (gimple_code (sections_stmt) == GIMPLE_OMP_SECTIONS);
|
4787 |
|
|
vin = gimple_omp_sections_control (sections_stmt);
|
4788 |
|
|
if (!is_combined_parallel (region))
|
4789 |
|
|
{
|
4790 |
|
|
/* If we are not inside a combined parallel+sections region,
|
4791 |
|
|
call GOMP_sections_start. */
|
4792 |
|
|
t = build_int_cst (unsigned_type_node,
|
4793 |
|
|
exit_reachable ? len - 1 : len);
|
4794 |
|
|
u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_START);
|
4795 |
|
|
stmt = gimple_build_call (u, 1, t);
|
4796 |
|
|
}
|
4797 |
|
|
else
|
4798 |
|
|
{
|
4799 |
|
|
/* Otherwise, call GOMP_sections_next. */
|
4800 |
|
|
u = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT);
|
4801 |
|
|
stmt = gimple_build_call (u, 0);
|
4802 |
|
|
}
|
4803 |
|
|
gimple_call_set_lhs (stmt, vin);
|
4804 |
|
|
gsi_insert_after (&si, stmt, GSI_SAME_STMT);
|
4805 |
|
|
gsi_remove (&si, true);
|
4806 |
|
|
|
4807 |
|
|
/* The switch() statement replacing GIMPLE_OMP_SECTIONS_SWITCH goes in
|
4808 |
|
|
L0_BB. */
|
4809 |
|
|
switch_si = gsi_last_bb (l0_bb);
|
4810 |
|
|
gcc_assert (gimple_code (gsi_stmt (switch_si)) == GIMPLE_OMP_SECTIONS_SWITCH);
|
4811 |
|
|
if (exit_reachable)
|
4812 |
|
|
{
|
4813 |
|
|
cont = last_stmt (l1_bb);
|
4814 |
|
|
gcc_assert (gimple_code (cont) == GIMPLE_OMP_CONTINUE);
|
4815 |
|
|
vmain = gimple_omp_continue_control_use (cont);
|
4816 |
|
|
vnext = gimple_omp_continue_control_def (cont);
|
4817 |
|
|
}
|
4818 |
|
|
else
|
4819 |
|
|
{
|
4820 |
|
|
vmain = vin;
|
4821 |
|
|
vnext = NULL_TREE;
|
4822 |
|
|
}
|
4823 |
|
|
|
4824 |
|
|
i = 0;
|
4825 |
|
|
if (exit_reachable)
|
4826 |
|
|
{
|
4827 |
|
|
t = build_case_label (build_int_cst (unsigned_type_node, 0), NULL, l2);
|
4828 |
|
|
VEC_quick_push (tree, label_vec, t);
|
4829 |
|
|
i++;
|
4830 |
|
|
}
|
4831 |
|
|
|
4832 |
|
|
/* Convert each GIMPLE_OMP_SECTION into a CASE_LABEL_EXPR. */
|
4833 |
|
|
for (inner = region->inner, casei = 1;
|
4834 |
|
|
inner;
|
4835 |
|
|
inner = inner->next, i++, casei++)
|
4836 |
|
|
{
|
4837 |
|
|
basic_block s_entry_bb, s_exit_bb;
|
4838 |
|
|
|
4839 |
|
|
/* Skip optional reduction region. */
|
4840 |
|
|
if (inner->type == GIMPLE_OMP_ATOMIC_LOAD)
|
4841 |
|
|
{
|
4842 |
|
|
--i;
|
4843 |
|
|
--casei;
|
4844 |
|
|
continue;
|
4845 |
|
|
}
|
4846 |
|
|
|
4847 |
|
|
s_entry_bb = inner->entry;
|
4848 |
|
|
s_exit_bb = inner->exit;
|
4849 |
|
|
|
4850 |
|
|
t = gimple_block_label (s_entry_bb);
|
4851 |
|
|
u = build_int_cst (unsigned_type_node, casei);
|
4852 |
|
|
u = build_case_label (u, NULL, t);
|
4853 |
|
|
VEC_quick_push (tree, label_vec, u);
|
4854 |
|
|
|
4855 |
|
|
si = gsi_last_bb (s_entry_bb);
|
4856 |
|
|
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SECTION);
|
4857 |
|
|
gcc_assert (i < len || gimple_omp_section_last_p (gsi_stmt (si)));
|
4858 |
|
|
gsi_remove (&si, true);
|
4859 |
|
|
single_succ_edge (s_entry_bb)->flags = EDGE_FALLTHRU;
|
4860 |
|
|
|
4861 |
|
|
if (s_exit_bb == NULL)
|
4862 |
|
|
continue;
|
4863 |
|
|
|
4864 |
|
|
si = gsi_last_bb (s_exit_bb);
|
4865 |
|
|
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN);
|
4866 |
|
|
gsi_remove (&si, true);
|
4867 |
|
|
|
4868 |
|
|
single_succ_edge (s_exit_bb)->flags = EDGE_FALLTHRU;
|
4869 |
|
|
}
|
4870 |
|
|
|
4871 |
|
|
/* Error handling code goes in DEFAULT_BB. */
|
4872 |
|
|
t = gimple_block_label (default_bb);
|
4873 |
|
|
u = build_case_label (NULL, NULL, t);
|
4874 |
|
|
make_edge (l0_bb, default_bb, 0);
|
4875 |
|
|
|
4876 |
|
|
stmt = gimple_build_switch_vec (vmain, u, label_vec);
|
4877 |
|
|
gsi_insert_after (&switch_si, stmt, GSI_SAME_STMT);
|
4878 |
|
|
gsi_remove (&switch_si, true);
|
4879 |
|
|
VEC_free (tree, heap, label_vec);
|
4880 |
|
|
|
4881 |
|
|
si = gsi_start_bb (default_bb);
|
4882 |
|
|
stmt = gimple_build_call (builtin_decl_explicit (BUILT_IN_TRAP), 0);
|
4883 |
|
|
gsi_insert_after (&si, stmt, GSI_CONTINUE_LINKING);
|
4884 |
|
|
|
4885 |
|
|
if (exit_reachable)
|
4886 |
|
|
{
|
4887 |
|
|
tree bfn_decl;
|
4888 |
|
|
|
4889 |
|
|
/* Code to get the next section goes in L1_BB. */
|
4890 |
|
|
si = gsi_last_bb (l1_bb);
|
4891 |
|
|
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CONTINUE);
|
4892 |
|
|
|
4893 |
|
|
bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_NEXT);
|
4894 |
|
|
stmt = gimple_build_call (bfn_decl, 0);
|
4895 |
|
|
gimple_call_set_lhs (stmt, vnext);
|
4896 |
|
|
gsi_insert_after (&si, stmt, GSI_SAME_STMT);
|
4897 |
|
|
gsi_remove (&si, true);
|
4898 |
|
|
|
4899 |
|
|
single_succ_edge (l1_bb)->flags = EDGE_FALLTHRU;
|
4900 |
|
|
|
4901 |
|
|
/* Cleanup function replaces GIMPLE_OMP_RETURN in EXIT_BB. */
|
4902 |
|
|
si = gsi_last_bb (l2_bb);
|
4903 |
|
|
if (gimple_omp_return_nowait_p (gsi_stmt (si)))
|
4904 |
|
|
t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END_NOWAIT);
|
4905 |
|
|
else
|
4906 |
|
|
t = builtin_decl_explicit (BUILT_IN_GOMP_SECTIONS_END);
|
4907 |
|
|
stmt = gimple_build_call (t, 0);
|
4908 |
|
|
gsi_insert_after (&si, stmt, GSI_SAME_STMT);
|
4909 |
|
|
gsi_remove (&si, true);
|
4910 |
|
|
}
|
4911 |
|
|
|
4912 |
|
|
set_immediate_dominator (CDI_DOMINATORS, default_bb, l0_bb);
|
4913 |
|
|
}
|
4914 |
|
|
|
4915 |
|
|
|
4916 |
|
|
/* Expand code for an OpenMP single directive. We've already expanded
|
4917 |
|
|
much of the code, here we simply place the GOMP_barrier call. */
|
4918 |
|
|
|
4919 |
|
|
static void
|
4920 |
|
|
expand_omp_single (struct omp_region *region)
|
4921 |
|
|
{
|
4922 |
|
|
basic_block entry_bb, exit_bb;
|
4923 |
|
|
gimple_stmt_iterator si;
|
4924 |
|
|
bool need_barrier = false;
|
4925 |
|
|
|
4926 |
|
|
entry_bb = region->entry;
|
4927 |
|
|
exit_bb = region->exit;
|
4928 |
|
|
|
4929 |
|
|
si = gsi_last_bb (entry_bb);
|
4930 |
|
|
/* The terminal barrier at the end of a GOMP_single_copy sequence cannot
|
4931 |
|
|
be removed. We need to ensure that the thread that entered the single
|
4932 |
|
|
does not exit before the data is copied out by the other threads. */
|
4933 |
|
|
if (find_omp_clause (gimple_omp_single_clauses (gsi_stmt (si)),
|
4934 |
|
|
OMP_CLAUSE_COPYPRIVATE))
|
4935 |
|
|
need_barrier = true;
|
4936 |
|
|
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE);
|
4937 |
|
|
gsi_remove (&si, true);
|
4938 |
|
|
single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU;
|
4939 |
|
|
|
4940 |
|
|
si = gsi_last_bb (exit_bb);
|
4941 |
|
|
if (!gimple_omp_return_nowait_p (gsi_stmt (si)) || need_barrier)
|
4942 |
|
|
force_gimple_operand_gsi (&si, build_omp_barrier (), false, NULL_TREE,
|
4943 |
|
|
false, GSI_SAME_STMT);
|
4944 |
|
|
gsi_remove (&si, true);
|
4945 |
|
|
single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU;
|
4946 |
|
|
}
|
4947 |
|
|
|
4948 |
|
|
|
4949 |
|
|
/* Generic expansion for OpenMP synchronization directives: master,
|
4950 |
|
|
ordered and critical. All we need to do here is remove the entry
|
4951 |
|
|
and exit markers for REGION. */
|
4952 |
|
|
|
4953 |
|
|
static void
|
4954 |
|
|
expand_omp_synch (struct omp_region *region)
|
4955 |
|
|
{
|
4956 |
|
|
basic_block entry_bb, exit_bb;
|
4957 |
|
|
gimple_stmt_iterator si;
|
4958 |
|
|
|
4959 |
|
|
entry_bb = region->entry;
|
4960 |
|
|
exit_bb = region->exit;
|
4961 |
|
|
|
4962 |
|
|
si = gsi_last_bb (entry_bb);
|
4963 |
|
|
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_SINGLE
|
4964 |
|
|
|| gimple_code (gsi_stmt (si)) == GIMPLE_OMP_MASTER
|
4965 |
|
|
|| gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ORDERED
|
4966 |
|
|
|| gimple_code (gsi_stmt (si)) == GIMPLE_OMP_CRITICAL);
|
4967 |
|
|
gsi_remove (&si, true);
|
4968 |
|
|
single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU;
|
4969 |
|
|
|
4970 |
|
|
if (exit_bb)
|
4971 |
|
|
{
|
4972 |
|
|
si = gsi_last_bb (exit_bb);
|
4973 |
|
|
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_RETURN);
|
4974 |
|
|
gsi_remove (&si, true);
|
4975 |
|
|
single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU;
|
4976 |
|
|
}
|
4977 |
|
|
}
|
4978 |
|
|
|
4979 |
|
|
/* A subroutine of expand_omp_atomic. Attempt to implement the atomic
|
4980 |
|
|
operation as a normal volatile load. */
|
4981 |
|
|
|
4982 |
|
|
static bool
|
4983 |
|
|
expand_omp_atomic_load (basic_block load_bb, tree addr,
|
4984 |
|
|
tree loaded_val, int index)
|
4985 |
|
|
{
|
4986 |
|
|
enum built_in_function tmpbase;
|
4987 |
|
|
gimple_stmt_iterator gsi;
|
4988 |
|
|
basic_block store_bb;
|
4989 |
|
|
location_t loc;
|
4990 |
|
|
gimple stmt;
|
4991 |
|
|
tree decl, call, type, itype;
|
4992 |
|
|
|
4993 |
|
|
gsi = gsi_last_bb (load_bb);
|
4994 |
|
|
stmt = gsi_stmt (gsi);
|
4995 |
|
|
gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD);
|
4996 |
|
|
loc = gimple_location (stmt);
|
4997 |
|
|
|
4998 |
|
|
/* ??? If the target does not implement atomic_load_optab[mode], and mode
|
4999 |
|
|
is smaller than word size, then expand_atomic_load assumes that the load
|
5000 |
|
|
is atomic. We could avoid the builtin entirely in this case. */
|
5001 |
|
|
|
5002 |
|
|
tmpbase = (enum built_in_function) (BUILT_IN_ATOMIC_LOAD_N + index + 1);
|
5003 |
|
|
decl = builtin_decl_explicit (tmpbase);
|
5004 |
|
|
if (decl == NULL_TREE)
|
5005 |
|
|
return false;
|
5006 |
|
|
|
5007 |
|
|
type = TREE_TYPE (loaded_val);
|
5008 |
|
|
itype = TREE_TYPE (TREE_TYPE (decl));
|
5009 |
|
|
|
5010 |
|
|
call = build_call_expr_loc (loc, decl, 2, addr,
|
5011 |
|
|
build_int_cst (NULL, MEMMODEL_RELAXED));
|
5012 |
|
|
if (!useless_type_conversion_p (type, itype))
|
5013 |
|
|
call = fold_build1_loc (loc, VIEW_CONVERT_EXPR, type, call);
|
5014 |
|
|
call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call);
|
5015 |
|
|
|
5016 |
|
|
force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT);
|
5017 |
|
|
gsi_remove (&gsi, true);
|
5018 |
|
|
|
5019 |
|
|
store_bb = single_succ (load_bb);
|
5020 |
|
|
gsi = gsi_last_bb (store_bb);
|
5021 |
|
|
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE);
|
5022 |
|
|
gsi_remove (&gsi, true);
|
5023 |
|
|
|
5024 |
|
|
if (gimple_in_ssa_p (cfun))
|
5025 |
|
|
update_ssa (TODO_update_ssa_no_phi);
|
5026 |
|
|
|
5027 |
|
|
return true;
|
5028 |
|
|
}
|
5029 |
|
|
|
5030 |
|
|
/* A subroutine of expand_omp_atomic. Attempt to implement the atomic
|
5031 |
|
|
operation as a normal volatile store. */
|
5032 |
|
|
|
5033 |
|
|
static bool
|
5034 |
|
|
expand_omp_atomic_store (basic_block load_bb, tree addr,
|
5035 |
|
|
tree loaded_val, tree stored_val, int index)
|
5036 |
|
|
{
|
5037 |
|
|
enum built_in_function tmpbase;
|
5038 |
|
|
gimple_stmt_iterator gsi;
|
5039 |
|
|
basic_block store_bb = single_succ (load_bb);
|
5040 |
|
|
location_t loc;
|
5041 |
|
|
gimple stmt;
|
5042 |
|
|
tree decl, call, type, itype;
|
5043 |
|
|
enum machine_mode imode;
|
5044 |
|
|
bool exchange;
|
5045 |
|
|
|
5046 |
|
|
gsi = gsi_last_bb (load_bb);
|
5047 |
|
|
stmt = gsi_stmt (gsi);
|
5048 |
|
|
gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_LOAD);
|
5049 |
|
|
|
5050 |
|
|
/* If the load value is needed, then this isn't a store but an exchange. */
|
5051 |
|
|
exchange = gimple_omp_atomic_need_value_p (stmt);
|
5052 |
|
|
|
5053 |
|
|
gsi = gsi_last_bb (store_bb);
|
5054 |
|
|
stmt = gsi_stmt (gsi);
|
5055 |
|
|
gcc_assert (gimple_code (stmt) == GIMPLE_OMP_ATOMIC_STORE);
|
5056 |
|
|
loc = gimple_location (stmt);
|
5057 |
|
|
|
5058 |
|
|
/* ??? If the target does not implement atomic_store_optab[mode], and mode
|
5059 |
|
|
is smaller than word size, then expand_atomic_store assumes that the store
|
5060 |
|
|
is atomic. We could avoid the builtin entirely in this case. */
|
5061 |
|
|
|
5062 |
|
|
tmpbase = (exchange ? BUILT_IN_ATOMIC_EXCHANGE_N : BUILT_IN_ATOMIC_STORE_N);
|
5063 |
|
|
tmpbase = (enum built_in_function) ((int) tmpbase + index + 1);
|
5064 |
|
|
decl = builtin_decl_explicit (tmpbase);
|
5065 |
|
|
if (decl == NULL_TREE)
|
5066 |
|
|
return false;
|
5067 |
|
|
|
5068 |
|
|
type = TREE_TYPE (stored_val);
|
5069 |
|
|
|
5070 |
|
|
/* Dig out the type of the function's second argument. */
|
5071 |
|
|
itype = TREE_TYPE (decl);
|
5072 |
|
|
itype = TYPE_ARG_TYPES (itype);
|
5073 |
|
|
itype = TREE_CHAIN (itype);
|
5074 |
|
|
itype = TREE_VALUE (itype);
|
5075 |
|
|
imode = TYPE_MODE (itype);
|
5076 |
|
|
|
5077 |
|
|
if (exchange && !can_atomic_exchange_p (imode, true))
|
5078 |
|
|
return false;
|
5079 |
|
|
|
5080 |
|
|
if (!useless_type_conversion_p (itype, type))
|
5081 |
|
|
stored_val = fold_build1_loc (loc, VIEW_CONVERT_EXPR, itype, stored_val);
|
5082 |
|
|
call = build_call_expr_loc (loc, decl, 3, addr, stored_val,
|
5083 |
|
|
build_int_cst (NULL, MEMMODEL_RELAXED));
|
5084 |
|
|
if (exchange)
|
5085 |
|
|
{
|
5086 |
|
|
if (!useless_type_conversion_p (type, itype))
|
5087 |
|
|
call = build1_loc (loc, VIEW_CONVERT_EXPR, type, call);
|
5088 |
|
|
call = build2_loc (loc, MODIFY_EXPR, void_type_node, loaded_val, call);
|
5089 |
|
|
}
|
5090 |
|
|
|
5091 |
|
|
force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT);
|
5092 |
|
|
gsi_remove (&gsi, true);
|
5093 |
|
|
|
5094 |
|
|
/* Remove the GIMPLE_OMP_ATOMIC_LOAD that we verified above. */
|
5095 |
|
|
gsi = gsi_last_bb (load_bb);
|
5096 |
|
|
gsi_remove (&gsi, true);
|
5097 |
|
|
|
5098 |
|
|
if (gimple_in_ssa_p (cfun))
|
5099 |
|
|
update_ssa (TODO_update_ssa_no_phi);
|
5100 |
|
|
|
5101 |
|
|
return true;
|
5102 |
|
|
}
|
5103 |
|
|
|
5104 |
|
|
/* A subroutine of expand_omp_atomic. Attempt to implement the atomic
|
5105 |
|
|
operation as a __atomic_fetch_op builtin. INDEX is log2 of the
|
5106 |
|
|
size of the data type, and thus usable to find the index of the builtin
|
5107 |
|
|
decl. Returns false if the expression is not of the proper form. */
|
5108 |
|
|
|
5109 |
|
|
static bool
|
5110 |
|
|
expand_omp_atomic_fetch_op (basic_block load_bb,
|
5111 |
|
|
tree addr, tree loaded_val,
|
5112 |
|
|
tree stored_val, int index)
|
5113 |
|
|
{
|
5114 |
|
|
enum built_in_function oldbase, newbase, tmpbase;
|
5115 |
|
|
tree decl, itype, call;
|
5116 |
|
|
tree lhs, rhs;
|
5117 |
|
|
basic_block store_bb = single_succ (load_bb);
|
5118 |
|
|
gimple_stmt_iterator gsi;
|
5119 |
|
|
gimple stmt;
|
5120 |
|
|
location_t loc;
|
5121 |
|
|
enum tree_code code;
|
5122 |
|
|
bool need_old, need_new;
|
5123 |
|
|
enum machine_mode imode;
|
5124 |
|
|
|
5125 |
|
|
/* We expect to find the following sequences:
|
5126 |
|
|
|
5127 |
|
|
load_bb:
|
5128 |
|
|
GIMPLE_OMP_ATOMIC_LOAD (tmp, mem)
|
5129 |
|
|
|
5130 |
|
|
store_bb:
|
5131 |
|
|
val = tmp OP something; (or: something OP tmp)
|
5132 |
|
|
GIMPLE_OMP_STORE (val)
|
5133 |
|
|
|
5134 |
|
|
???FIXME: Allow a more flexible sequence.
|
5135 |
|
|
Perhaps use data flow to pick the statements.
|
5136 |
|
|
|
5137 |
|
|
*/
|
5138 |
|
|
|
5139 |
|
|
gsi = gsi_after_labels (store_bb);
|
5140 |
|
|
stmt = gsi_stmt (gsi);
|
5141 |
|
|
loc = gimple_location (stmt);
|
5142 |
|
|
if (!is_gimple_assign (stmt))
|
5143 |
|
|
return false;
|
5144 |
|
|
gsi_next (&gsi);
|
5145 |
|
|
if (gimple_code (gsi_stmt (gsi)) != GIMPLE_OMP_ATOMIC_STORE)
|
5146 |
|
|
return false;
|
5147 |
|
|
need_new = gimple_omp_atomic_need_value_p (gsi_stmt (gsi));
|
5148 |
|
|
need_old = gimple_omp_atomic_need_value_p (last_stmt (load_bb));
|
5149 |
|
|
gcc_checking_assert (!need_old || !need_new);
|
5150 |
|
|
|
5151 |
|
|
if (!operand_equal_p (gimple_assign_lhs (stmt), stored_val, 0))
|
5152 |
|
|
return false;
|
5153 |
|
|
|
5154 |
|
|
/* Check for one of the supported fetch-op operations. */
|
5155 |
|
|
code = gimple_assign_rhs_code (stmt);
|
5156 |
|
|
switch (code)
|
5157 |
|
|
{
|
5158 |
|
|
case PLUS_EXPR:
|
5159 |
|
|
case POINTER_PLUS_EXPR:
|
5160 |
|
|
oldbase = BUILT_IN_ATOMIC_FETCH_ADD_N;
|
5161 |
|
|
newbase = BUILT_IN_ATOMIC_ADD_FETCH_N;
|
5162 |
|
|
break;
|
5163 |
|
|
case MINUS_EXPR:
|
5164 |
|
|
oldbase = BUILT_IN_ATOMIC_FETCH_SUB_N;
|
5165 |
|
|
newbase = BUILT_IN_ATOMIC_SUB_FETCH_N;
|
5166 |
|
|
break;
|
5167 |
|
|
case BIT_AND_EXPR:
|
5168 |
|
|
oldbase = BUILT_IN_ATOMIC_FETCH_AND_N;
|
5169 |
|
|
newbase = BUILT_IN_ATOMIC_AND_FETCH_N;
|
5170 |
|
|
break;
|
5171 |
|
|
case BIT_IOR_EXPR:
|
5172 |
|
|
oldbase = BUILT_IN_ATOMIC_FETCH_OR_N;
|
5173 |
|
|
newbase = BUILT_IN_ATOMIC_OR_FETCH_N;
|
5174 |
|
|
break;
|
5175 |
|
|
case BIT_XOR_EXPR:
|
5176 |
|
|
oldbase = BUILT_IN_ATOMIC_FETCH_XOR_N;
|
5177 |
|
|
newbase = BUILT_IN_ATOMIC_XOR_FETCH_N;
|
5178 |
|
|
break;
|
5179 |
|
|
default:
|
5180 |
|
|
return false;
|
5181 |
|
|
}
|
5182 |
|
|
|
5183 |
|
|
/* Make sure the expression is of the proper form. */
|
5184 |
|
|
if (operand_equal_p (gimple_assign_rhs1 (stmt), loaded_val, 0))
|
5185 |
|
|
rhs = gimple_assign_rhs2 (stmt);
|
5186 |
|
|
else if (commutative_tree_code (gimple_assign_rhs_code (stmt))
|
5187 |
|
|
&& operand_equal_p (gimple_assign_rhs2 (stmt), loaded_val, 0))
|
5188 |
|
|
rhs = gimple_assign_rhs1 (stmt);
|
5189 |
|
|
else
|
5190 |
|
|
return false;
|
5191 |
|
|
|
5192 |
|
|
tmpbase = ((enum built_in_function)
|
5193 |
|
|
((need_new ? newbase : oldbase) + index + 1));
|
5194 |
|
|
decl = builtin_decl_explicit (tmpbase);
|
5195 |
|
|
if (decl == NULL_TREE)
|
5196 |
|
|
return false;
|
5197 |
|
|
itype = TREE_TYPE (TREE_TYPE (decl));
|
5198 |
|
|
imode = TYPE_MODE (itype);
|
5199 |
|
|
|
5200 |
|
|
/* We could test all of the various optabs involved, but the fact of the
|
5201 |
|
|
matter is that (with the exception of i486 vs i586 and xadd) all targets
|
5202 |
|
|
that support any atomic operaton optab also implements compare-and-swap.
|
5203 |
|
|
Let optabs.c take care of expanding any compare-and-swap loop. */
|
5204 |
|
|
if (!can_compare_and_swap_p (imode, true))
|
5205 |
|
|
return false;
|
5206 |
|
|
|
5207 |
|
|
gsi = gsi_last_bb (load_bb);
|
5208 |
|
|
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_LOAD);
|
5209 |
|
|
|
5210 |
|
|
/* OpenMP does not imply any barrier-like semantics on its atomic ops.
|
5211 |
|
|
It only requires that the operation happen atomically. Thus we can
|
5212 |
|
|
use the RELAXED memory model. */
|
5213 |
|
|
call = build_call_expr_loc (loc, decl, 3, addr,
|
5214 |
|
|
fold_convert_loc (loc, itype, rhs),
|
5215 |
|
|
build_int_cst (NULL, MEMMODEL_RELAXED));
|
5216 |
|
|
|
5217 |
|
|
if (need_old || need_new)
|
5218 |
|
|
{
|
5219 |
|
|
lhs = need_old ? loaded_val : stored_val;
|
5220 |
|
|
call = fold_convert_loc (loc, TREE_TYPE (lhs), call);
|
5221 |
|
|
call = build2_loc (loc, MODIFY_EXPR, void_type_node, lhs, call);
|
5222 |
|
|
}
|
5223 |
|
|
else
|
5224 |
|
|
call = fold_convert_loc (loc, void_type_node, call);
|
5225 |
|
|
force_gimple_operand_gsi (&gsi, call, true, NULL_TREE, true, GSI_SAME_STMT);
|
5226 |
|
|
gsi_remove (&gsi, true);
|
5227 |
|
|
|
5228 |
|
|
gsi = gsi_last_bb (store_bb);
|
5229 |
|
|
gcc_assert (gimple_code (gsi_stmt (gsi)) == GIMPLE_OMP_ATOMIC_STORE);
|
5230 |
|
|
gsi_remove (&gsi, true);
|
5231 |
|
|
gsi = gsi_last_bb (store_bb);
|
5232 |
|
|
gsi_remove (&gsi, true);
|
5233 |
|
|
|
5234 |
|
|
if (gimple_in_ssa_p (cfun))
|
5235 |
|
|
update_ssa (TODO_update_ssa_no_phi);
|
5236 |
|
|
|
5237 |
|
|
return true;
|
5238 |
|
|
}
|
5239 |
|
|
|
5240 |
|
|
/* A subroutine of expand_omp_atomic. Implement the atomic operation as:
|
5241 |
|
|
|
5242 |
|
|
oldval = *addr;
|
5243 |
|
|
repeat:
|
5244 |
|
|
newval = rhs; // with oldval replacing *addr in rhs
|
5245 |
|
|
oldval = __sync_val_compare_and_swap (addr, oldval, newval);
|
5246 |
|
|
if (oldval != newval)
|
5247 |
|
|
goto repeat;
|
5248 |
|
|
|
5249 |
|
|
INDEX is log2 of the size of the data type, and thus usable to find the
|
5250 |
|
|
index of the builtin decl. */
|
5251 |
|
|
|
5252 |
|
|
static bool
|
5253 |
|
|
expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb,
|
5254 |
|
|
tree addr, tree loaded_val, tree stored_val,
|
5255 |
|
|
int index)
|
5256 |
|
|
{
|
5257 |
|
|
tree loadedi, storedi, initial, new_storedi, old_vali;
|
5258 |
|
|
tree type, itype, cmpxchg, iaddr;
|
5259 |
|
|
gimple_stmt_iterator si;
|
5260 |
|
|
basic_block loop_header = single_succ (load_bb);
|
5261 |
|
|
gimple phi, stmt;
|
5262 |
|
|
edge e;
|
5263 |
|
|
enum built_in_function fncode;
|
5264 |
|
|
|
5265 |
|
|
/* ??? We need a non-pointer interface to __atomic_compare_exchange in
|
5266 |
|
|
order to use the RELAXED memory model effectively. */
|
5267 |
|
|
fncode = (enum built_in_function)((int)BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_N
|
5268 |
|
|
+ index + 1);
|
5269 |
|
|
cmpxchg = builtin_decl_explicit (fncode);
|
5270 |
|
|
if (cmpxchg == NULL_TREE)
|
5271 |
|
|
return false;
|
5272 |
|
|
type = TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (addr)));
|
5273 |
|
|
itype = TREE_TYPE (TREE_TYPE (cmpxchg));
|
5274 |
|
|
|
5275 |
|
|
if (!can_compare_and_swap_p (TYPE_MODE (itype), true))
|
5276 |
|
|
return false;
|
5277 |
|
|
|
5278 |
|
|
/* Load the initial value, replacing the GIMPLE_OMP_ATOMIC_LOAD. */
|
5279 |
|
|
si = gsi_last_bb (load_bb);
|
5280 |
|
|
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD);
|
5281 |
|
|
|
5282 |
|
|
/* For floating-point values, we'll need to view-convert them to integers
|
5283 |
|
|
so that we can perform the atomic compare and swap. Simplify the
|
5284 |
|
|
following code by always setting up the "i"ntegral variables. */
|
5285 |
|
|
if (!INTEGRAL_TYPE_P (type) && !POINTER_TYPE_P (type))
|
5286 |
|
|
{
|
5287 |
|
|
tree iaddr_val;
|
5288 |
|
|
|
5289 |
|
|
iaddr = create_tmp_var (build_pointer_type_for_mode (itype, ptr_mode,
|
5290 |
|
|
true), NULL);
|
5291 |
|
|
iaddr_val
|
5292 |
|
|
= force_gimple_operand_gsi (&si,
|
5293 |
|
|
fold_convert (TREE_TYPE (iaddr), addr),
|
5294 |
|
|
false, NULL_TREE, true, GSI_SAME_STMT);
|
5295 |
|
|
stmt = gimple_build_assign (iaddr, iaddr_val);
|
5296 |
|
|
gsi_insert_before (&si, stmt, GSI_SAME_STMT);
|
5297 |
|
|
loadedi = create_tmp_var (itype, NULL);
|
5298 |
|
|
if (gimple_in_ssa_p (cfun))
|
5299 |
|
|
{
|
5300 |
|
|
add_referenced_var (iaddr);
|
5301 |
|
|
add_referenced_var (loadedi);
|
5302 |
|
|
loadedi = make_ssa_name (loadedi, NULL);
|
5303 |
|
|
}
|
5304 |
|
|
}
|
5305 |
|
|
else
|
5306 |
|
|
{
|
5307 |
|
|
iaddr = addr;
|
5308 |
|
|
loadedi = loaded_val;
|
5309 |
|
|
}
|
5310 |
|
|
|
5311 |
|
|
initial
|
5312 |
|
|
= force_gimple_operand_gsi (&si,
|
5313 |
|
|
build2 (MEM_REF, TREE_TYPE (TREE_TYPE (iaddr)),
|
5314 |
|
|
iaddr,
|
5315 |
|
|
build_int_cst (TREE_TYPE (iaddr), 0)),
|
5316 |
|
|
true, NULL_TREE, true, GSI_SAME_STMT);
|
5317 |
|
|
|
5318 |
|
|
/* Move the value to the LOADEDI temporary. */
|
5319 |
|
|
if (gimple_in_ssa_p (cfun))
|
5320 |
|
|
{
|
5321 |
|
|
gcc_assert (gimple_seq_empty_p (phi_nodes (loop_header)));
|
5322 |
|
|
phi = create_phi_node (loadedi, loop_header);
|
5323 |
|
|
SSA_NAME_DEF_STMT (loadedi) = phi;
|
5324 |
|
|
SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (load_bb)),
|
5325 |
|
|
initial);
|
5326 |
|
|
}
|
5327 |
|
|
else
|
5328 |
|
|
gsi_insert_before (&si,
|
5329 |
|
|
gimple_build_assign (loadedi, initial),
|
5330 |
|
|
GSI_SAME_STMT);
|
5331 |
|
|
if (loadedi != loaded_val)
|
5332 |
|
|
{
|
5333 |
|
|
gimple_stmt_iterator gsi2;
|
5334 |
|
|
tree x;
|
5335 |
|
|
|
5336 |
|
|
x = build1 (VIEW_CONVERT_EXPR, type, loadedi);
|
5337 |
|
|
gsi2 = gsi_start_bb (loop_header);
|
5338 |
|
|
if (gimple_in_ssa_p (cfun))
|
5339 |
|
|
{
|
5340 |
|
|
gimple stmt;
|
5341 |
|
|
x = force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE,
|
5342 |
|
|
true, GSI_SAME_STMT);
|
5343 |
|
|
stmt = gimple_build_assign (loaded_val, x);
|
5344 |
|
|
gsi_insert_before (&gsi2, stmt, GSI_SAME_STMT);
|
5345 |
|
|
}
|
5346 |
|
|
else
|
5347 |
|
|
{
|
5348 |
|
|
x = build2 (MODIFY_EXPR, TREE_TYPE (loaded_val), loaded_val, x);
|
5349 |
|
|
force_gimple_operand_gsi (&gsi2, x, true, NULL_TREE,
|
5350 |
|
|
true, GSI_SAME_STMT);
|
5351 |
|
|
}
|
5352 |
|
|
}
|
5353 |
|
|
gsi_remove (&si, true);
|
5354 |
|
|
|
5355 |
|
|
si = gsi_last_bb (store_bb);
|
5356 |
|
|
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE);
|
5357 |
|
|
|
5358 |
|
|
if (iaddr == addr)
|
5359 |
|
|
storedi = stored_val;
|
5360 |
|
|
else
|
5361 |
|
|
storedi =
|
5362 |
|
|
force_gimple_operand_gsi (&si,
|
5363 |
|
|
build1 (VIEW_CONVERT_EXPR, itype,
|
5364 |
|
|
stored_val), true, NULL_TREE, true,
|
5365 |
|
|
GSI_SAME_STMT);
|
5366 |
|
|
|
5367 |
|
|
/* Build the compare&swap statement. */
|
5368 |
|
|
new_storedi = build_call_expr (cmpxchg, 3, iaddr, loadedi, storedi);
|
5369 |
|
|
new_storedi = force_gimple_operand_gsi (&si,
|
5370 |
|
|
fold_convert (TREE_TYPE (loadedi),
|
5371 |
|
|
new_storedi),
|
5372 |
|
|
true, NULL_TREE,
|
5373 |
|
|
true, GSI_SAME_STMT);
|
5374 |
|
|
|
5375 |
|
|
if (gimple_in_ssa_p (cfun))
|
5376 |
|
|
old_vali = loadedi;
|
5377 |
|
|
else
|
5378 |
|
|
{
|
5379 |
|
|
old_vali = create_tmp_var (TREE_TYPE (loadedi), NULL);
|
5380 |
|
|
if (gimple_in_ssa_p (cfun))
|
5381 |
|
|
add_referenced_var (old_vali);
|
5382 |
|
|
stmt = gimple_build_assign (old_vali, loadedi);
|
5383 |
|
|
gsi_insert_before (&si, stmt, GSI_SAME_STMT);
|
5384 |
|
|
|
5385 |
|
|
stmt = gimple_build_assign (loadedi, new_storedi);
|
5386 |
|
|
gsi_insert_before (&si, stmt, GSI_SAME_STMT);
|
5387 |
|
|
}
|
5388 |
|
|
|
5389 |
|
|
/* Note that we always perform the comparison as an integer, even for
|
5390 |
|
|
floating point. This allows the atomic operation to properly
|
5391 |
|
|
succeed even with NaNs and -0.0. */
|
5392 |
|
|
stmt = gimple_build_cond_empty
|
5393 |
|
|
(build2 (NE_EXPR, boolean_type_node,
|
5394 |
|
|
new_storedi, old_vali));
|
5395 |
|
|
gsi_insert_before (&si, stmt, GSI_SAME_STMT);
|
5396 |
|
|
|
5397 |
|
|
/* Update cfg. */
|
5398 |
|
|
e = single_succ_edge (store_bb);
|
5399 |
|
|
e->flags &= ~EDGE_FALLTHRU;
|
5400 |
|
|
e->flags |= EDGE_FALSE_VALUE;
|
5401 |
|
|
|
5402 |
|
|
e = make_edge (store_bb, loop_header, EDGE_TRUE_VALUE);
|
5403 |
|
|
|
5404 |
|
|
/* Copy the new value to loadedi (we already did that before the condition
|
5405 |
|
|
if we are not in SSA). */
|
5406 |
|
|
if (gimple_in_ssa_p (cfun))
|
5407 |
|
|
{
|
5408 |
|
|
phi = gimple_seq_first_stmt (phi_nodes (loop_header));
|
5409 |
|
|
SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e), new_storedi);
|
5410 |
|
|
}
|
5411 |
|
|
|
5412 |
|
|
/* Remove GIMPLE_OMP_ATOMIC_STORE. */
|
5413 |
|
|
gsi_remove (&si, true);
|
5414 |
|
|
|
5415 |
|
|
if (gimple_in_ssa_p (cfun))
|
5416 |
|
|
update_ssa (TODO_update_ssa_no_phi);
|
5417 |
|
|
|
5418 |
|
|
return true;
|
5419 |
|
|
}
|
5420 |
|
|
|
5421 |
|
|
/* A subroutine of expand_omp_atomic. Implement the atomic operation as:
|
5422 |
|
|
|
5423 |
|
|
GOMP_atomic_start ();
|
5424 |
|
|
*addr = rhs;
|
5425 |
|
|
GOMP_atomic_end ();
|
5426 |
|
|
|
5427 |
|
|
The result is not globally atomic, but works so long as all parallel
|
5428 |
|
|
references are within #pragma omp atomic directives. According to
|
5429 |
|
|
responses received from omp@openmp.org, appears to be within spec.
|
5430 |
|
|
Which makes sense, since that's how several other compilers handle
|
5431 |
|
|
this situation as well.
|
5432 |
|
|
LOADED_VAL and ADDR are the operands of GIMPLE_OMP_ATOMIC_LOAD we're
|
5433 |
|
|
expanding. STORED_VAL is the operand of the matching
|
5434 |
|
|
GIMPLE_OMP_ATOMIC_STORE.
|
5435 |
|
|
|
5436 |
|
|
We replace
|
5437 |
|
|
GIMPLE_OMP_ATOMIC_LOAD (loaded_val, addr) with
|
5438 |
|
|
loaded_val = *addr;
|
5439 |
|
|
|
5440 |
|
|
and replace
|
5441 |
|
|
GIMPLE_OMP_ATOMIC_STORE (stored_val) with
|
5442 |
|
|
*addr = stored_val;
|
5443 |
|
|
*/
|
5444 |
|
|
|
5445 |
|
|
static bool
|
5446 |
|
|
expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb,
|
5447 |
|
|
tree addr, tree loaded_val, tree stored_val)
|
5448 |
|
|
{
|
5449 |
|
|
gimple_stmt_iterator si;
|
5450 |
|
|
gimple stmt;
|
5451 |
|
|
tree t;
|
5452 |
|
|
|
5453 |
|
|
si = gsi_last_bb (load_bb);
|
5454 |
|
|
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_LOAD);
|
5455 |
|
|
|
5456 |
|
|
t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_START);
|
5457 |
|
|
t = build_call_expr (t, 0);
|
5458 |
|
|
force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT);
|
5459 |
|
|
|
5460 |
|
|
stmt = gimple_build_assign (loaded_val, build_simple_mem_ref (addr));
|
5461 |
|
|
gsi_insert_before (&si, stmt, GSI_SAME_STMT);
|
5462 |
|
|
gsi_remove (&si, true);
|
5463 |
|
|
|
5464 |
|
|
si = gsi_last_bb (store_bb);
|
5465 |
|
|
gcc_assert (gimple_code (gsi_stmt (si)) == GIMPLE_OMP_ATOMIC_STORE);
|
5466 |
|
|
|
5467 |
|
|
stmt = gimple_build_assign (build_simple_mem_ref (unshare_expr (addr)),
|
5468 |
|
|
stored_val);
|
5469 |
|
|
gsi_insert_before (&si, stmt, GSI_SAME_STMT);
|
5470 |
|
|
|
5471 |
|
|
t = builtin_decl_explicit (BUILT_IN_GOMP_ATOMIC_END);
|
5472 |
|
|
t = build_call_expr (t, 0);
|
5473 |
|
|
force_gimple_operand_gsi (&si, t, true, NULL_TREE, true, GSI_SAME_STMT);
|
5474 |
|
|
gsi_remove (&si, true);
|
5475 |
|
|
|
5476 |
|
|
if (gimple_in_ssa_p (cfun))
|
5477 |
|
|
update_ssa (TODO_update_ssa_no_phi);
|
5478 |
|
|
return true;
|
5479 |
|
|
}
|
5480 |
|
|
|
5481 |
|
|
/* Expand an GIMPLE_OMP_ATOMIC statement. We try to expand
|
5482 |
|
|
using expand_omp_atomic_fetch_op. If it failed, we try to
|
5483 |
|
|
call expand_omp_atomic_pipeline, and if it fails too, the
|
5484 |
|
|
ultimate fallback is wrapping the operation in a mutex
|
5485 |
|
|
(expand_omp_atomic_mutex). REGION is the atomic region built
|
5486 |
|
|
by build_omp_regions_1(). */
|
5487 |
|
|
|
5488 |
|
|
static void
|
5489 |
|
|
expand_omp_atomic (struct omp_region *region)
|
5490 |
|
|
{
|
5491 |
|
|
basic_block load_bb = region->entry, store_bb = region->exit;
|
5492 |
|
|
gimple load = last_stmt (load_bb), store = last_stmt (store_bb);
|
5493 |
|
|
tree loaded_val = gimple_omp_atomic_load_lhs (load);
|
5494 |
|
|
tree addr = gimple_omp_atomic_load_rhs (load);
|
5495 |
|
|
tree stored_val = gimple_omp_atomic_store_val (store);
|
5496 |
|
|
tree type = TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (addr)));
|
5497 |
|
|
HOST_WIDE_INT index;
|
5498 |
|
|
|
5499 |
|
|
/* Make sure the type is one of the supported sizes. */
|
5500 |
|
|
index = tree_low_cst (TYPE_SIZE_UNIT (type), 1);
|
5501 |
|
|
index = exact_log2 (index);
|
5502 |
|
|
if (index >= 0 && index <= 4)
|
5503 |
|
|
{
|
5504 |
|
|
unsigned int align = TYPE_ALIGN_UNIT (type);
|
5505 |
|
|
|
5506 |
|
|
/* __sync builtins require strict data alignment. */
|
5507 |
|
|
/* ??? Assume BIGGEST_ALIGNMENT *is* aligned. */
|
5508 |
|
|
if (exact_log2 (align) >= index
|
5509 |
|
|
|| align * BITS_PER_UNIT >= BIGGEST_ALIGNMENT)
|
5510 |
|
|
{
|
5511 |
|
|
/* Atomic load. */
|
5512 |
|
|
if (loaded_val == stored_val
|
5513 |
|
|
&& (GET_MODE_CLASS (TYPE_MODE (type)) == MODE_INT
|
5514 |
|
|
|| GET_MODE_CLASS (TYPE_MODE (type)) == MODE_FLOAT)
|
5515 |
|
|
&& GET_MODE_BITSIZE (TYPE_MODE (type)) <= BITS_PER_WORD
|
5516 |
|
|
&& expand_omp_atomic_load (load_bb, addr, loaded_val, index))
|
5517 |
|
|
return;
|
5518 |
|
|
|
5519 |
|
|
/* Atomic store. */
|
5520 |
|
|
if ((GET_MODE_CLASS (TYPE_MODE (type)) == MODE_INT
|
5521 |
|
|
|| GET_MODE_CLASS (TYPE_MODE (type)) == MODE_FLOAT)
|
5522 |
|
|
&& GET_MODE_BITSIZE (TYPE_MODE (type)) <= BITS_PER_WORD
|
5523 |
|
|
&& store_bb == single_succ (load_bb)
|
5524 |
|
|
&& first_stmt (store_bb) == store
|
5525 |
|
|
&& expand_omp_atomic_store (load_bb, addr, loaded_val,
|
5526 |
|
|
stored_val, index))
|
5527 |
|
|
return;
|
5528 |
|
|
|
5529 |
|
|
/* When possible, use specialized atomic update functions. */
|
5530 |
|
|
if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
|
5531 |
|
|
&& store_bb == single_succ (load_bb)
|
5532 |
|
|
&& expand_omp_atomic_fetch_op (load_bb, addr,
|
5533 |
|
|
loaded_val, stored_val, index))
|
5534 |
|
|
return;
|
5535 |
|
|
|
5536 |
|
|
/* If we don't have specialized __sync builtins, try and implement
|
5537 |
|
|
as a compare and swap loop. */
|
5538 |
|
|
if (expand_omp_atomic_pipeline (load_bb, store_bb, addr,
|
5539 |
|
|
loaded_val, stored_val, index))
|
5540 |
|
|
return;
|
5541 |
|
|
}
|
5542 |
|
|
}
|
5543 |
|
|
|
5544 |
|
|
/* The ultimate fallback is wrapping the operation in a mutex. */
|
5545 |
|
|
expand_omp_atomic_mutex (load_bb, store_bb, addr, loaded_val, stored_val);
|
5546 |
|
|
}
|
5547 |
|
|
|
5548 |
|
|
|
5549 |
|
|
/* Expand the parallel region tree rooted at REGION. Expansion
|
5550 |
|
|
proceeds in depth-first order. Innermost regions are expanded
|
5551 |
|
|
first. This way, parallel regions that require a new function to
|
5552 |
|
|
be created (e.g., GIMPLE_OMP_PARALLEL) can be expanded without having any
|
5553 |
|
|
internal dependencies in their body. */
|
5554 |
|
|
|
5555 |
|
|
static void
|
5556 |
|
|
expand_omp (struct omp_region *region)
|
5557 |
|
|
{
|
5558 |
|
|
while (region)
|
5559 |
|
|
{
|
5560 |
|
|
location_t saved_location;
|
5561 |
|
|
|
5562 |
|
|
/* First, determine whether this is a combined parallel+workshare
|
5563 |
|
|
region. */
|
5564 |
|
|
if (region->type == GIMPLE_OMP_PARALLEL)
|
5565 |
|
|
determine_parallel_type (region);
|
5566 |
|
|
|
5567 |
|
|
if (region->inner)
|
5568 |
|
|
expand_omp (region->inner);
|
5569 |
|
|
|
5570 |
|
|
saved_location = input_location;
|
5571 |
|
|
if (gimple_has_location (last_stmt (region->entry)))
|
5572 |
|
|
input_location = gimple_location (last_stmt (region->entry));
|
5573 |
|
|
|
5574 |
|
|
switch (region->type)
|
5575 |
|
|
{
|
5576 |
|
|
case GIMPLE_OMP_PARALLEL:
|
5577 |
|
|
case GIMPLE_OMP_TASK:
|
5578 |
|
|
expand_omp_taskreg (region);
|
5579 |
|
|
break;
|
5580 |
|
|
|
5581 |
|
|
case GIMPLE_OMP_FOR:
|
5582 |
|
|
expand_omp_for (region);
|
5583 |
|
|
break;
|
5584 |
|
|
|
5585 |
|
|
case GIMPLE_OMP_SECTIONS:
|
5586 |
|
|
expand_omp_sections (region);
|
5587 |
|
|
break;
|
5588 |
|
|
|
5589 |
|
|
case GIMPLE_OMP_SECTION:
|
5590 |
|
|
/* Individual omp sections are handled together with their
|
5591 |
|
|
parent GIMPLE_OMP_SECTIONS region. */
|
5592 |
|
|
break;
|
5593 |
|
|
|
5594 |
|
|
case GIMPLE_OMP_SINGLE:
|
5595 |
|
|
expand_omp_single (region);
|
5596 |
|
|
break;
|
5597 |
|
|
|
5598 |
|
|
case GIMPLE_OMP_MASTER:
|
5599 |
|
|
case GIMPLE_OMP_ORDERED:
|
5600 |
|
|
case GIMPLE_OMP_CRITICAL:
|
5601 |
|
|
expand_omp_synch (region);
|
5602 |
|
|
break;
|
5603 |
|
|
|
5604 |
|
|
case GIMPLE_OMP_ATOMIC_LOAD:
|
5605 |
|
|
expand_omp_atomic (region);
|
5606 |
|
|
break;
|
5607 |
|
|
|
5608 |
|
|
default:
|
5609 |
|
|
gcc_unreachable ();
|
5610 |
|
|
}
|
5611 |
|
|
|
5612 |
|
|
input_location = saved_location;
|
5613 |
|
|
region = region->next;
|
5614 |
|
|
}
|
5615 |
|
|
}
|
5616 |
|
|
|
5617 |
|
|
|
5618 |
|
|
/* Helper for build_omp_regions. Scan the dominator tree starting at
|
5619 |
|
|
block BB. PARENT is the region that contains BB. If SINGLE_TREE is
|
5620 |
|
|
true, the function ends once a single tree is built (otherwise, whole
|
5621 |
|
|
forest of OMP constructs may be built). */
|
5622 |
|
|
|
5623 |
|
|
static void
|
5624 |
|
|
build_omp_regions_1 (basic_block bb, struct omp_region *parent,
|
5625 |
|
|
bool single_tree)
|
5626 |
|
|
{
|
5627 |
|
|
gimple_stmt_iterator gsi;
|
5628 |
|
|
gimple stmt;
|
5629 |
|
|
basic_block son;
|
5630 |
|
|
|
5631 |
|
|
gsi = gsi_last_bb (bb);
|
5632 |
|
|
if (!gsi_end_p (gsi) && is_gimple_omp (gsi_stmt (gsi)))
|
5633 |
|
|
{
|
5634 |
|
|
struct omp_region *region;
|
5635 |
|
|
enum gimple_code code;
|
5636 |
|
|
|
5637 |
|
|
stmt = gsi_stmt (gsi);
|
5638 |
|
|
code = gimple_code (stmt);
|
5639 |
|
|
if (code == GIMPLE_OMP_RETURN)
|
5640 |
|
|
{
|
5641 |
|
|
/* STMT is the return point out of region PARENT. Mark it
|
5642 |
|
|
as the exit point and make PARENT the immediately
|
5643 |
|
|
enclosing region. */
|
5644 |
|
|
gcc_assert (parent);
|
5645 |
|
|
region = parent;
|
5646 |
|
|
region->exit = bb;
|
5647 |
|
|
parent = parent->outer;
|
5648 |
|
|
}
|
5649 |
|
|
else if (code == GIMPLE_OMP_ATOMIC_STORE)
|
5650 |
|
|
{
|
5651 |
|
|
/* GIMPLE_OMP_ATOMIC_STORE is analoguous to
|
5652 |
|
|
GIMPLE_OMP_RETURN, but matches with
|
5653 |
|
|
GIMPLE_OMP_ATOMIC_LOAD. */
|
5654 |
|
|
gcc_assert (parent);
|
5655 |
|
|
gcc_assert (parent->type == GIMPLE_OMP_ATOMIC_LOAD);
|
5656 |
|
|
region = parent;
|
5657 |
|
|
region->exit = bb;
|
5658 |
|
|
parent = parent->outer;
|
5659 |
|
|
}
|
5660 |
|
|
|
5661 |
|
|
else if (code == GIMPLE_OMP_CONTINUE)
|
5662 |
|
|
{
|
5663 |
|
|
gcc_assert (parent);
|
5664 |
|
|
parent->cont = bb;
|
5665 |
|
|
}
|
5666 |
|
|
else if (code == GIMPLE_OMP_SECTIONS_SWITCH)
|
5667 |
|
|
{
|
5668 |
|
|
/* GIMPLE_OMP_SECTIONS_SWITCH is part of
|
5669 |
|
|
GIMPLE_OMP_SECTIONS, and we do nothing for it. */
|
5670 |
|
|
;
|
5671 |
|
|
}
|
5672 |
|
|
else
|
5673 |
|
|
{
|
5674 |
|
|
/* Otherwise, this directive becomes the parent for a new
|
5675 |
|
|
region. */
|
5676 |
|
|
region = new_omp_region (bb, code, parent);
|
5677 |
|
|
parent = region;
|
5678 |
|
|
}
|
5679 |
|
|
}
|
5680 |
|
|
|
5681 |
|
|
if (single_tree && !parent)
|
5682 |
|
|
return;
|
5683 |
|
|
|
5684 |
|
|
for (son = first_dom_son (CDI_DOMINATORS, bb);
|
5685 |
|
|
son;
|
5686 |
|
|
son = next_dom_son (CDI_DOMINATORS, son))
|
5687 |
|
|
build_omp_regions_1 (son, parent, single_tree);
|
5688 |
|
|
}
|
5689 |
|
|
|
5690 |
|
|
/* Builds the tree of OMP regions rooted at ROOT, storing it to
|
5691 |
|
|
root_omp_region. */
|
5692 |
|
|
|
5693 |
|
|
static void
|
5694 |
|
|
build_omp_regions_root (basic_block root)
|
5695 |
|
|
{
|
5696 |
|
|
gcc_assert (root_omp_region == NULL);
|
5697 |
|
|
build_omp_regions_1 (root, NULL, true);
|
5698 |
|
|
gcc_assert (root_omp_region != NULL);
|
5699 |
|
|
}
|
5700 |
|
|
|
5701 |
|
|
/* Expands omp construct (and its subconstructs) starting in HEAD. */
|
5702 |
|
|
|
5703 |
|
|
void
|
5704 |
|
|
omp_expand_local (basic_block head)
|
5705 |
|
|
{
|
5706 |
|
|
build_omp_regions_root (head);
|
5707 |
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
5708 |
|
|
{
|
5709 |
|
|
fprintf (dump_file, "\nOMP region tree\n\n");
|
5710 |
|
|
dump_omp_region (dump_file, root_omp_region, 0);
|
5711 |
|
|
fprintf (dump_file, "\n");
|
5712 |
|
|
}
|
5713 |
|
|
|
5714 |
|
|
remove_exit_barriers (root_omp_region);
|
5715 |
|
|
expand_omp (root_omp_region);
|
5716 |
|
|
|
5717 |
|
|
free_omp_regions ();
|
5718 |
|
|
}
|
5719 |
|
|
|
5720 |
|
|
/* Scan the CFG and build a tree of OMP regions. Return the root of
|
5721 |
|
|
the OMP region tree. */
|
5722 |
|
|
|
5723 |
|
|
static void
|
5724 |
|
|
build_omp_regions (void)
|
5725 |
|
|
{
|
5726 |
|
|
gcc_assert (root_omp_region == NULL);
|
5727 |
|
|
calculate_dominance_info (CDI_DOMINATORS);
|
5728 |
|
|
build_omp_regions_1 (ENTRY_BLOCK_PTR, NULL, false);
|
5729 |
|
|
}
|
5730 |
|
|
|
5731 |
|
|
/* Main entry point for expanding OMP-GIMPLE into runtime calls. */
|
5732 |
|
|
|
5733 |
|
|
static unsigned int
|
5734 |
|
|
execute_expand_omp (void)
|
5735 |
|
|
{
|
5736 |
|
|
build_omp_regions ();
|
5737 |
|
|
|
5738 |
|
|
if (!root_omp_region)
|
5739 |
|
|
return 0;
|
5740 |
|
|
|
5741 |
|
|
if (dump_file)
|
5742 |
|
|
{
|
5743 |
|
|
fprintf (dump_file, "\nOMP region tree\n\n");
|
5744 |
|
|
dump_omp_region (dump_file, root_omp_region, 0);
|
5745 |
|
|
fprintf (dump_file, "\n");
|
5746 |
|
|
}
|
5747 |
|
|
|
5748 |
|
|
remove_exit_barriers (root_omp_region);
|
5749 |
|
|
|
5750 |
|
|
expand_omp (root_omp_region);
|
5751 |
|
|
|
5752 |
|
|
cleanup_tree_cfg ();
|
5753 |
|
|
|
5754 |
|
|
free_omp_regions ();
|
5755 |
|
|
|
5756 |
|
|
return 0;
|
5757 |
|
|
}
|
5758 |
|
|
|
5759 |
|
|
/* OMP expansion -- the default pass, run before creation of SSA form. */
|
5760 |
|
|
|
5761 |
|
|
static bool
|
5762 |
|
|
gate_expand_omp (void)
|
5763 |
|
|
{
|
5764 |
|
|
return (flag_openmp != 0 && !seen_error ());
|
5765 |
|
|
}
|
5766 |
|
|
|
5767 |
|
|
struct gimple_opt_pass pass_expand_omp =
|
5768 |
|
|
{
|
5769 |
|
|
{
|
5770 |
|
|
GIMPLE_PASS,
|
5771 |
|
|
"ompexp", /* name */
|
5772 |
|
|
gate_expand_omp, /* gate */
|
5773 |
|
|
execute_expand_omp, /* execute */
|
5774 |
|
|
NULL, /* sub */
|
5775 |
|
|
NULL, /* next */
|
5776 |
|
|
0, /* static_pass_number */
|
5777 |
|
|
TV_NONE, /* tv_id */
|
5778 |
|
|
PROP_gimple_any, /* properties_required */
|
5779 |
|
|
0, /* properties_provided */
|
5780 |
|
|
0, /* properties_destroyed */
|
5781 |
|
|
0, /* todo_flags_start */
|
5782 |
|
|
|
5783 |
|
|
}
|
5784 |
|
|
};
|
5785 |
|
|
|
5786 |
|
|
/* Routines to lower OpenMP directives into OMP-GIMPLE. */
|
5787 |
|
|
|
5788 |
|
|
/* Lower the OpenMP sections directive in the current statement in GSI_P.
|
5789 |
|
|
CTX is the enclosing OMP context for the current statement. */
|
5790 |
|
|
|
5791 |
|
|
static void
|
5792 |
|
|
lower_omp_sections (gimple_stmt_iterator *gsi_p, omp_context *ctx)
|
5793 |
|
|
{
|
5794 |
|
|
tree block, control;
|
5795 |
|
|
gimple_stmt_iterator tgsi;
|
5796 |
|
|
unsigned i, len;
|
5797 |
|
|
gimple stmt, new_stmt, bind, t;
|
5798 |
|
|
gimple_seq ilist, dlist, olist, new_body, body;
|
5799 |
|
|
struct gimplify_ctx gctx;
|
5800 |
|
|
|
5801 |
|
|
stmt = gsi_stmt (*gsi_p);
|
5802 |
|
|
|
5803 |
|
|
push_gimplify_context (&gctx);
|
5804 |
|
|
|
5805 |
|
|
dlist = NULL;
|
5806 |
|
|
ilist = NULL;
|
5807 |
|
|
lower_rec_input_clauses (gimple_omp_sections_clauses (stmt),
|
5808 |
|
|
&ilist, &dlist, ctx);
|
5809 |
|
|
|
5810 |
|
|
tgsi = gsi_start (gimple_omp_body (stmt));
|
5811 |
|
|
for (len = 0; !gsi_end_p (tgsi); len++, gsi_next (&tgsi))
|
5812 |
|
|
continue;
|
5813 |
|
|
|
5814 |
|
|
tgsi = gsi_start (gimple_omp_body (stmt));
|
5815 |
|
|
body = NULL;
|
5816 |
|
|
for (i = 0; i < len; i++, gsi_next (&tgsi))
|
5817 |
|
|
{
|
5818 |
|
|
omp_context *sctx;
|
5819 |
|
|
gimple sec_start;
|
5820 |
|
|
|
5821 |
|
|
sec_start = gsi_stmt (tgsi);
|
5822 |
|
|
sctx = maybe_lookup_ctx (sec_start);
|
5823 |
|
|
gcc_assert (sctx);
|
5824 |
|
|
|
5825 |
|
|
gimple_seq_add_stmt (&body, sec_start);
|
5826 |
|
|
|
5827 |
|
|
lower_omp (gimple_omp_body (sec_start), sctx);
|
5828 |
|
|
gimple_seq_add_seq (&body, gimple_omp_body (sec_start));
|
5829 |
|
|
gimple_omp_set_body (sec_start, NULL);
|
5830 |
|
|
|
5831 |
|
|
if (i == len - 1)
|
5832 |
|
|
{
|
5833 |
|
|
gimple_seq l = NULL;
|
5834 |
|
|
lower_lastprivate_clauses (gimple_omp_sections_clauses (stmt), NULL,
|
5835 |
|
|
&l, ctx);
|
5836 |
|
|
gimple_seq_add_seq (&body, l);
|
5837 |
|
|
gimple_omp_section_set_last (sec_start);
|
5838 |
|
|
}
|
5839 |
|
|
|
5840 |
|
|
gimple_seq_add_stmt (&body, gimple_build_omp_return (false));
|
5841 |
|
|
}
|
5842 |
|
|
|
5843 |
|
|
block = make_node (BLOCK);
|
5844 |
|
|
bind = gimple_build_bind (NULL, body, block);
|
5845 |
|
|
|
5846 |
|
|
olist = NULL;
|
5847 |
|
|
lower_reduction_clauses (gimple_omp_sections_clauses (stmt), &olist, ctx);
|
5848 |
|
|
|
5849 |
|
|
block = make_node (BLOCK);
|
5850 |
|
|
new_stmt = gimple_build_bind (NULL, NULL, block);
|
5851 |
|
|
|
5852 |
|
|
pop_gimplify_context (new_stmt);
|
5853 |
|
|
gimple_bind_append_vars (new_stmt, ctx->block_vars);
|
5854 |
|
|
BLOCK_VARS (block) = gimple_bind_vars (bind);
|
5855 |
|
|
if (BLOCK_VARS (block))
|
5856 |
|
|
TREE_USED (block) = 1;
|
5857 |
|
|
|
5858 |
|
|
new_body = NULL;
|
5859 |
|
|
gimple_seq_add_seq (&new_body, ilist);
|
5860 |
|
|
gimple_seq_add_stmt (&new_body, stmt);
|
5861 |
|
|
gimple_seq_add_stmt (&new_body, gimple_build_omp_sections_switch ());
|
5862 |
|
|
gimple_seq_add_stmt (&new_body, bind);
|
5863 |
|
|
|
5864 |
|
|
control = create_tmp_var (unsigned_type_node, ".section");
|
5865 |
|
|
t = gimple_build_omp_continue (control, control);
|
5866 |
|
|
gimple_omp_sections_set_control (stmt, control);
|
5867 |
|
|
gimple_seq_add_stmt (&new_body, t);
|
5868 |
|
|
|
5869 |
|
|
gimple_seq_add_seq (&new_body, olist);
|
5870 |
|
|
gimple_seq_add_seq (&new_body, dlist);
|
5871 |
|
|
|
5872 |
|
|
new_body = maybe_catch_exception (new_body);
|
5873 |
|
|
|
5874 |
|
|
t = gimple_build_omp_return
|
5875 |
|
|
(!!find_omp_clause (gimple_omp_sections_clauses (stmt),
|
5876 |
|
|
OMP_CLAUSE_NOWAIT));
|
5877 |
|
|
gimple_seq_add_stmt (&new_body, t);
|
5878 |
|
|
|
5879 |
|
|
gimple_bind_set_body (new_stmt, new_body);
|
5880 |
|
|
gimple_omp_set_body (stmt, NULL);
|
5881 |
|
|
|
5882 |
|
|
gsi_replace (gsi_p, new_stmt, true);
|
5883 |
|
|
}
|
5884 |
|
|
|
5885 |
|
|
|
5886 |
|
|
/* A subroutine of lower_omp_single. Expand the simple form of
|
5887 |
|
|
a GIMPLE_OMP_SINGLE, without a copyprivate clause:
|
5888 |
|
|
|
5889 |
|
|
if (GOMP_single_start ())
|
5890 |
|
|
BODY;
|
5891 |
|
|
[ GOMP_barrier (); ] -> unless 'nowait' is present.
|
5892 |
|
|
|
5893 |
|
|
FIXME. It may be better to delay expanding the logic of this until
|
5894 |
|
|
pass_expand_omp. The expanded logic may make the job more difficult
|
5895 |
|
|
to a synchronization analysis pass. */
|
5896 |
|
|
|
5897 |
|
|
static void
|
5898 |
|
|
lower_omp_single_simple (gimple single_stmt, gimple_seq *pre_p)
|
5899 |
|
|
{
|
5900 |
|
|
location_t loc = gimple_location (single_stmt);
|
5901 |
|
|
tree tlabel = create_artificial_label (loc);
|
5902 |
|
|
tree flabel = create_artificial_label (loc);
|
5903 |
|
|
gimple call, cond;
|
5904 |
|
|
tree lhs, decl;
|
5905 |
|
|
|
5906 |
|
|
decl = builtin_decl_explicit (BUILT_IN_GOMP_SINGLE_START);
|
5907 |
|
|
lhs = create_tmp_var (TREE_TYPE (TREE_TYPE (decl)), NULL);
|
5908 |
|
|
call = gimple_build_call (decl, 0);
|
5909 |
|
|
gimple_call_set_lhs (call, lhs);
|
5910 |
|
|
gimple_seq_add_stmt (pre_p, call);
|
5911 |
|
|
|
5912 |
|
|
cond = gimple_build_cond (EQ_EXPR, lhs,
|
5913 |
|
|
fold_convert_loc (loc, TREE_TYPE (lhs),
|
5914 |
|
|
boolean_true_node),
|
5915 |
|
|
tlabel, flabel);
|
5916 |
|
|
gimple_seq_add_stmt (pre_p, cond);
|
5917 |
|
|
gimple_seq_add_stmt (pre_p, gimple_build_label (tlabel));
|
5918 |
|
|
gimple_seq_add_seq (pre_p, gimple_omp_body (single_stmt));
|
5919 |
|
|
gimple_seq_add_stmt (pre_p, gimple_build_label (flabel));
|
5920 |
|
|
}
|
5921 |
|
|
|
5922 |
|
|
|
5923 |
|
|
/* A subroutine of lower_omp_single. Expand the simple form of
|
5924 |
|
|
a GIMPLE_OMP_SINGLE, with a copyprivate clause:
|
5925 |
|
|
|
5926 |
|
|
#pragma omp single copyprivate (a, b, c)
|
5927 |
|
|
|
5928 |
|
|
Create a new structure to hold copies of 'a', 'b' and 'c' and emit:
|
5929 |
|
|
|
5930 |
|
|
{
|
5931 |
|
|
if ((copyout_p = GOMP_single_copy_start ()) == NULL)
|
5932 |
|
|
{
|
5933 |
|
|
BODY;
|
5934 |
|
|
copyout.a = a;
|
5935 |
|
|
copyout.b = b;
|
5936 |
|
|
copyout.c = c;
|
5937 |
|
|
GOMP_single_copy_end (©out);
|
5938 |
|
|
}
|
5939 |
|
|
else
|
5940 |
|
|
{
|
5941 |
|
|
a = copyout_p->a;
|
5942 |
|
|
b = copyout_p->b;
|
5943 |
|
|
c = copyout_p->c;
|
5944 |
|
|
}
|
5945 |
|
|
GOMP_barrier ();
|
5946 |
|
|
}
|
5947 |
|
|
|
5948 |
|
|
FIXME. It may be better to delay expanding the logic of this until
|
5949 |
|
|
pass_expand_omp. The expanded logic may make the job more difficult
|
5950 |
|
|
to a synchronization analysis pass. */
|
5951 |
|
|
|
5952 |
|
|
static void
|
5953 |
|
|
lower_omp_single_copy (gimple single_stmt, gimple_seq *pre_p, omp_context *ctx)
|
5954 |
|
|
{
|
5955 |
|
|
tree ptr_type, t, l0, l1, l2, bfn_decl;
|
5956 |
|
|
gimple_seq copyin_seq;
|
5957 |
|
|
location_t loc = gimple_location (single_stmt);
|
5958 |
|
|
|
5959 |
|
|
ctx->sender_decl = create_tmp_var (ctx->record_type, ".omp_copy_o");
|
5960 |
|
|
|
5961 |
|
|
ptr_type = build_pointer_type (ctx->record_type);
|
5962 |
|
|
ctx->receiver_decl = create_tmp_var (ptr_type, ".omp_copy_i");
|
5963 |
|
|
|
5964 |
|
|
l0 = create_artificial_label (loc);
|
5965 |
|
|
l1 = create_artificial_label (loc);
|
5966 |
|
|
l2 = create_artificial_label (loc);
|
5967 |
|
|
|
5968 |
|
|
bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SINGLE_COPY_START);
|
5969 |
|
|
t = build_call_expr_loc (loc, bfn_decl, 0);
|
5970 |
|
|
t = fold_convert_loc (loc, ptr_type, t);
|
5971 |
|
|
gimplify_assign (ctx->receiver_decl, t, pre_p);
|
5972 |
|
|
|
5973 |
|
|
t = build2 (EQ_EXPR, boolean_type_node, ctx->receiver_decl,
|
5974 |
|
|
build_int_cst (ptr_type, 0));
|
5975 |
|
|
t = build3 (COND_EXPR, void_type_node, t,
|
5976 |
|
|
build_and_jump (&l0), build_and_jump (&l1));
|
5977 |
|
|
gimplify_and_add (t, pre_p);
|
5978 |
|
|
|
5979 |
|
|
gimple_seq_add_stmt (pre_p, gimple_build_label (l0));
|
5980 |
|
|
|
5981 |
|
|
gimple_seq_add_seq (pre_p, gimple_omp_body (single_stmt));
|
5982 |
|
|
|
5983 |
|
|
copyin_seq = NULL;
|
5984 |
|
|
lower_copyprivate_clauses (gimple_omp_single_clauses (single_stmt), pre_p,
|
5985 |
|
|
©in_seq, ctx);
|
5986 |
|
|
|
5987 |
|
|
t = build_fold_addr_expr_loc (loc, ctx->sender_decl);
|
5988 |
|
|
bfn_decl = builtin_decl_explicit (BUILT_IN_GOMP_SINGLE_COPY_END);
|
5989 |
|
|
t = build_call_expr_loc (loc, bfn_decl, 1, t);
|
5990 |
|
|
gimplify_and_add (t, pre_p);
|
5991 |
|
|
|
5992 |
|
|
t = build_and_jump (&l2);
|
5993 |
|
|
gimplify_and_add (t, pre_p);
|
5994 |
|
|
|
5995 |
|
|
gimple_seq_add_stmt (pre_p, gimple_build_label (l1));
|
5996 |
|
|
|
5997 |
|
|
gimple_seq_add_seq (pre_p, copyin_seq);
|
5998 |
|
|
|
5999 |
|
|
gimple_seq_add_stmt (pre_p, gimple_build_label (l2));
|
6000 |
|
|
}
|
6001 |
|
|
|
6002 |
|
|
|
6003 |
|
|
/* Expand code for an OpenMP single directive. */
|
6004 |
|
|
|
6005 |
|
|
static void
|
6006 |
|
|
lower_omp_single (gimple_stmt_iterator *gsi_p, omp_context *ctx)
|
6007 |
|
|
{
|
6008 |
|
|
tree block;
|
6009 |
|
|
gimple t, bind, single_stmt = gsi_stmt (*gsi_p);
|
6010 |
|
|
gimple_seq bind_body, dlist;
|
6011 |
|
|
struct gimplify_ctx gctx;
|
6012 |
|
|
|
6013 |
|
|
push_gimplify_context (&gctx);
|
6014 |
|
|
|
6015 |
|
|
bind_body = NULL;
|
6016 |
|
|
lower_rec_input_clauses (gimple_omp_single_clauses (single_stmt),
|
6017 |
|
|
&bind_body, &dlist, ctx);
|
6018 |
|
|
lower_omp (gimple_omp_body (single_stmt), ctx);
|
6019 |
|
|
|
6020 |
|
|
gimple_seq_add_stmt (&bind_body, single_stmt);
|
6021 |
|
|
|
6022 |
|
|
if (ctx->record_type)
|
6023 |
|
|
lower_omp_single_copy (single_stmt, &bind_body, ctx);
|
6024 |
|
|
else
|
6025 |
|
|
lower_omp_single_simple (single_stmt, &bind_body);
|
6026 |
|
|
|
6027 |
|
|
gimple_omp_set_body (single_stmt, NULL);
|
6028 |
|
|
|
6029 |
|
|
gimple_seq_add_seq (&bind_body, dlist);
|
6030 |
|
|
|
6031 |
|
|
bind_body = maybe_catch_exception (bind_body);
|
6032 |
|
|
|
6033 |
|
|
t = gimple_build_omp_return
|
6034 |
|
|
(!!find_omp_clause (gimple_omp_single_clauses (single_stmt),
|
6035 |
|
|
OMP_CLAUSE_NOWAIT));
|
6036 |
|
|
gimple_seq_add_stmt (&bind_body, t);
|
6037 |
|
|
|
6038 |
|
|
block = make_node (BLOCK);
|
6039 |
|
|
bind = gimple_build_bind (NULL, bind_body, block);
|
6040 |
|
|
|
6041 |
|
|
pop_gimplify_context (bind);
|
6042 |
|
|
|
6043 |
|
|
gimple_bind_append_vars (bind, ctx->block_vars);
|
6044 |
|
|
BLOCK_VARS (block) = ctx->block_vars;
|
6045 |
|
|
gsi_replace (gsi_p, bind, true);
|
6046 |
|
|
if (BLOCK_VARS (block))
|
6047 |
|
|
TREE_USED (block) = 1;
|
6048 |
|
|
}
|
6049 |
|
|
|
6050 |
|
|
|
6051 |
|
|
/* Expand code for an OpenMP master directive. */
|
6052 |
|
|
|
6053 |
|
|
static void
|
6054 |
|
|
lower_omp_master (gimple_stmt_iterator *gsi_p, omp_context *ctx)
|
6055 |
|
|
{
|
6056 |
|
|
tree block, lab = NULL, x, bfn_decl;
|
6057 |
|
|
gimple stmt = gsi_stmt (*gsi_p), bind;
|
6058 |
|
|
location_t loc = gimple_location (stmt);
|
6059 |
|
|
gimple_seq tseq;
|
6060 |
|
|
struct gimplify_ctx gctx;
|
6061 |
|
|
|
6062 |
|
|
push_gimplify_context (&gctx);
|
6063 |
|
|
|
6064 |
|
|
block = make_node (BLOCK);
|
6065 |
|
|
bind = gimple_build_bind (NULL, gimple_seq_alloc_with_stmt (stmt),
|
6066 |
|
|
block);
|
6067 |
|
|
|
6068 |
|
|
bfn_decl = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM);
|
6069 |
|
|
x = build_call_expr_loc (loc, bfn_decl, 0);
|
6070 |
|
|
x = build2 (EQ_EXPR, boolean_type_node, x, integer_zero_node);
|
6071 |
|
|
x = build3 (COND_EXPR, void_type_node, x, NULL, build_and_jump (&lab));
|
6072 |
|
|
tseq = NULL;
|
6073 |
|
|
gimplify_and_add (x, &tseq);
|
6074 |
|
|
gimple_bind_add_seq (bind, tseq);
|
6075 |
|
|
|
6076 |
|
|
lower_omp (gimple_omp_body (stmt), ctx);
|
6077 |
|
|
gimple_omp_set_body (stmt, maybe_catch_exception (gimple_omp_body (stmt)));
|
6078 |
|
|
gimple_bind_add_seq (bind, gimple_omp_body (stmt));
|
6079 |
|
|
gimple_omp_set_body (stmt, NULL);
|
6080 |
|
|
|
6081 |
|
|
gimple_bind_add_stmt (bind, gimple_build_label (lab));
|
6082 |
|
|
|
6083 |
|
|
gimple_bind_add_stmt (bind, gimple_build_omp_return (true));
|
6084 |
|
|
|
6085 |
|
|
pop_gimplify_context (bind);
|
6086 |
|
|
|
6087 |
|
|
gimple_bind_append_vars (bind, ctx->block_vars);
|
6088 |
|
|
BLOCK_VARS (block) = ctx->block_vars;
|
6089 |
|
|
gsi_replace (gsi_p, bind, true);
|
6090 |
|
|
}
|
6091 |
|
|
|
6092 |
|
|
|
6093 |
|
|
/* Expand code for an OpenMP ordered directive. */
|
6094 |
|
|
|
6095 |
|
|
static void
|
6096 |
|
|
lower_omp_ordered (gimple_stmt_iterator *gsi_p, omp_context *ctx)
|
6097 |
|
|
{
|
6098 |
|
|
tree block;
|
6099 |
|
|
gimple stmt = gsi_stmt (*gsi_p), bind, x;
|
6100 |
|
|
struct gimplify_ctx gctx;
|
6101 |
|
|
|
6102 |
|
|
push_gimplify_context (&gctx);
|
6103 |
|
|
|
6104 |
|
|
block = make_node (BLOCK);
|
6105 |
|
|
bind = gimple_build_bind (NULL, gimple_seq_alloc_with_stmt (stmt),
|
6106 |
|
|
block);
|
6107 |
|
|
|
6108 |
|
|
x = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ORDERED_START),
|
6109 |
|
|
0);
|
6110 |
|
|
gimple_bind_add_stmt (bind, x);
|
6111 |
|
|
|
6112 |
|
|
lower_omp (gimple_omp_body (stmt), ctx);
|
6113 |
|
|
gimple_omp_set_body (stmt, maybe_catch_exception (gimple_omp_body (stmt)));
|
6114 |
|
|
gimple_bind_add_seq (bind, gimple_omp_body (stmt));
|
6115 |
|
|
gimple_omp_set_body (stmt, NULL);
|
6116 |
|
|
|
6117 |
|
|
x = gimple_build_call (builtin_decl_explicit (BUILT_IN_GOMP_ORDERED_END), 0);
|
6118 |
|
|
gimple_bind_add_stmt (bind, x);
|
6119 |
|
|
|
6120 |
|
|
gimple_bind_add_stmt (bind, gimple_build_omp_return (true));
|
6121 |
|
|
|
6122 |
|
|
pop_gimplify_context (bind);
|
6123 |
|
|
|
6124 |
|
|
gimple_bind_append_vars (bind, ctx->block_vars);
|
6125 |
|
|
BLOCK_VARS (block) = gimple_bind_vars (bind);
|
6126 |
|
|
gsi_replace (gsi_p, bind, true);
|
6127 |
|
|
}
|
6128 |
|
|
|
6129 |
|
|
|
6130 |
|
|
/* Gimplify a GIMPLE_OMP_CRITICAL statement. This is a relatively simple
|
6131 |
|
|
substitution of a couple of function calls. But in the NAMED case,
|
6132 |
|
|
requires that languages coordinate a symbol name. It is therefore
|
6133 |
|
|
best put here in common code. */
|
6134 |
|
|
|
6135 |
|
|
static GTY((param1_is (tree), param2_is (tree)))
|
6136 |
|
|
splay_tree critical_name_mutexes;
|
6137 |
|
|
|
6138 |
|
|
static void
|
6139 |
|
|
lower_omp_critical (gimple_stmt_iterator *gsi_p, omp_context *ctx)
|
6140 |
|
|
{
|
6141 |
|
|
tree block;
|
6142 |
|
|
tree name, lock, unlock;
|
6143 |
|
|
gimple stmt = gsi_stmt (*gsi_p), bind;
|
6144 |
|
|
location_t loc = gimple_location (stmt);
|
6145 |
|
|
gimple_seq tbody;
|
6146 |
|
|
struct gimplify_ctx gctx;
|
6147 |
|
|
|
6148 |
|
|
name = gimple_omp_critical_name (stmt);
|
6149 |
|
|
if (name)
|
6150 |
|
|
{
|
6151 |
|
|
tree decl;
|
6152 |
|
|
splay_tree_node n;
|
6153 |
|
|
|
6154 |
|
|
if (!critical_name_mutexes)
|
6155 |
|
|
critical_name_mutexes
|
6156 |
|
|
= splay_tree_new_ggc (splay_tree_compare_pointers,
|
6157 |
|
|
ggc_alloc_splay_tree_tree_node_tree_node_splay_tree_s,
|
6158 |
|
|
ggc_alloc_splay_tree_tree_node_tree_node_splay_tree_node_s);
|
6159 |
|
|
|
6160 |
|
|
n = splay_tree_lookup (critical_name_mutexes, (splay_tree_key) name);
|
6161 |
|
|
if (n == NULL)
|
6162 |
|
|
{
|
6163 |
|
|
char *new_str;
|
6164 |
|
|
|
6165 |
|
|
decl = create_tmp_var_raw (ptr_type_node, NULL);
|
6166 |
|
|
|
6167 |
|
|
new_str = ACONCAT ((".gomp_critical_user_",
|
6168 |
|
|
IDENTIFIER_POINTER (name), NULL));
|
6169 |
|
|
DECL_NAME (decl) = get_identifier (new_str);
|
6170 |
|
|
TREE_PUBLIC (decl) = 1;
|
6171 |
|
|
TREE_STATIC (decl) = 1;
|
6172 |
|
|
DECL_COMMON (decl) = 1;
|
6173 |
|
|
DECL_ARTIFICIAL (decl) = 1;
|
6174 |
|
|
DECL_IGNORED_P (decl) = 1;
|
6175 |
|
|
varpool_finalize_decl (decl);
|
6176 |
|
|
|
6177 |
|
|
splay_tree_insert (critical_name_mutexes, (splay_tree_key) name,
|
6178 |
|
|
(splay_tree_value) decl);
|
6179 |
|
|
}
|
6180 |
|
|
else
|
6181 |
|
|
decl = (tree) n->value;
|
6182 |
|
|
|
6183 |
|
|
lock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_NAME_START);
|
6184 |
|
|
lock = build_call_expr_loc (loc, lock, 1, build_fold_addr_expr_loc (loc, decl));
|
6185 |
|
|
|
6186 |
|
|
unlock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_NAME_END);
|
6187 |
|
|
unlock = build_call_expr_loc (loc, unlock, 1,
|
6188 |
|
|
build_fold_addr_expr_loc (loc, decl));
|
6189 |
|
|
}
|
6190 |
|
|
else
|
6191 |
|
|
{
|
6192 |
|
|
lock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_START);
|
6193 |
|
|
lock = build_call_expr_loc (loc, lock, 0);
|
6194 |
|
|
|
6195 |
|
|
unlock = builtin_decl_explicit (BUILT_IN_GOMP_CRITICAL_END);
|
6196 |
|
|
unlock = build_call_expr_loc (loc, unlock, 0);
|
6197 |
|
|
}
|
6198 |
|
|
|
6199 |
|
|
push_gimplify_context (&gctx);
|
6200 |
|
|
|
6201 |
|
|
block = make_node (BLOCK);
|
6202 |
|
|
bind = gimple_build_bind (NULL, gimple_seq_alloc_with_stmt (stmt), block);
|
6203 |
|
|
|
6204 |
|
|
tbody = gimple_bind_body (bind);
|
6205 |
|
|
gimplify_and_add (lock, &tbody);
|
6206 |
|
|
gimple_bind_set_body (bind, tbody);
|
6207 |
|
|
|
6208 |
|
|
lower_omp (gimple_omp_body (stmt), ctx);
|
6209 |
|
|
gimple_omp_set_body (stmt, maybe_catch_exception (gimple_omp_body (stmt)));
|
6210 |
|
|
gimple_bind_add_seq (bind, gimple_omp_body (stmt));
|
6211 |
|
|
gimple_omp_set_body (stmt, NULL);
|
6212 |
|
|
|
6213 |
|
|
tbody = gimple_bind_body (bind);
|
6214 |
|
|
gimplify_and_add (unlock, &tbody);
|
6215 |
|
|
gimple_bind_set_body (bind, tbody);
|
6216 |
|
|
|
6217 |
|
|
gimple_bind_add_stmt (bind, gimple_build_omp_return (true));
|
6218 |
|
|
|
6219 |
|
|
pop_gimplify_context (bind);
|
6220 |
|
|
gimple_bind_append_vars (bind, ctx->block_vars);
|
6221 |
|
|
BLOCK_VARS (block) = gimple_bind_vars (bind);
|
6222 |
|
|
gsi_replace (gsi_p, bind, true);
|
6223 |
|
|
}
|
6224 |
|
|
|
6225 |
|
|
|
6226 |
|
|
/* A subroutine of lower_omp_for. Generate code to emit the predicate
|
6227 |
|
|
for a lastprivate clause. Given a loop control predicate of (V
|
6228 |
|
|
cond N2), we gate the clause on (!(V cond N2)). The lowered form
|
6229 |
|
|
is appended to *DLIST, iterator initialization is appended to
|
6230 |
|
|
*BODY_P. */
|
6231 |
|
|
|
6232 |
|
|
static void
|
6233 |
|
|
lower_omp_for_lastprivate (struct omp_for_data *fd, gimple_seq *body_p,
|
6234 |
|
|
gimple_seq *dlist, struct omp_context *ctx)
|
6235 |
|
|
{
|
6236 |
|
|
tree clauses, cond, vinit;
|
6237 |
|
|
enum tree_code cond_code;
|
6238 |
|
|
gimple_seq stmts;
|
6239 |
|
|
|
6240 |
|
|
cond_code = fd->loop.cond_code;
|
6241 |
|
|
cond_code = cond_code == LT_EXPR ? GE_EXPR : LE_EXPR;
|
6242 |
|
|
|
6243 |
|
|
/* When possible, use a strict equality expression. This can let VRP
|
6244 |
|
|
type optimizations deduce the value and remove a copy. */
|
6245 |
|
|
if (host_integerp (fd->loop.step, 0))
|
6246 |
|
|
{
|
6247 |
|
|
HOST_WIDE_INT step = TREE_INT_CST_LOW (fd->loop.step);
|
6248 |
|
|
if (step == 1 || step == -1)
|
6249 |
|
|
cond_code = EQ_EXPR;
|
6250 |
|
|
}
|
6251 |
|
|
|
6252 |
|
|
cond = build2 (cond_code, boolean_type_node, fd->loop.v, fd->loop.n2);
|
6253 |
|
|
|
6254 |
|
|
clauses = gimple_omp_for_clauses (fd->for_stmt);
|
6255 |
|
|
stmts = NULL;
|
6256 |
|
|
lower_lastprivate_clauses (clauses, cond, &stmts, ctx);
|
6257 |
|
|
if (!gimple_seq_empty_p (stmts))
|
6258 |
|
|
{
|
6259 |
|
|
gimple_seq_add_seq (&stmts, *dlist);
|
6260 |
|
|
*dlist = stmts;
|
6261 |
|
|
|
6262 |
|
|
/* Optimize: v = 0; is usually cheaper than v = some_other_constant. */
|
6263 |
|
|
vinit = fd->loop.n1;
|
6264 |
|
|
if (cond_code == EQ_EXPR
|
6265 |
|
|
&& host_integerp (fd->loop.n2, 0)
|
6266 |
|
|
&& ! integer_zerop (fd->loop.n2))
|
6267 |
|
|
vinit = build_int_cst (TREE_TYPE (fd->loop.v), 0);
|
6268 |
|
|
|
6269 |
|
|
/* Initialize the iterator variable, so that threads that don't execute
|
6270 |
|
|
any iterations don't execute the lastprivate clauses by accident. */
|
6271 |
|
|
gimplify_assign (fd->loop.v, vinit, body_p);
|
6272 |
|
|
}
|
6273 |
|
|
}
|
6274 |
|
|
|
6275 |
|
|
|
6276 |
|
|
/* Lower code for an OpenMP loop directive. */
|
6277 |
|
|
|
6278 |
|
|
static void
|
6279 |
|
|
lower_omp_for (gimple_stmt_iterator *gsi_p, omp_context *ctx)
|
6280 |
|
|
{
|
6281 |
|
|
tree *rhs_p, block;
|
6282 |
|
|
struct omp_for_data fd;
|
6283 |
|
|
gimple stmt = gsi_stmt (*gsi_p), new_stmt;
|
6284 |
|
|
gimple_seq omp_for_body, body, dlist;
|
6285 |
|
|
size_t i;
|
6286 |
|
|
struct gimplify_ctx gctx;
|
6287 |
|
|
|
6288 |
|
|
push_gimplify_context (&gctx);
|
6289 |
|
|
|
6290 |
|
|
lower_omp (gimple_omp_for_pre_body (stmt), ctx);
|
6291 |
|
|
lower_omp (gimple_omp_body (stmt), ctx);
|
6292 |
|
|
|
6293 |
|
|
block = make_node (BLOCK);
|
6294 |
|
|
new_stmt = gimple_build_bind (NULL, NULL, block);
|
6295 |
|
|
|
6296 |
|
|
/* Move declaration of temporaries in the loop body before we make
|
6297 |
|
|
it go away. */
|
6298 |
|
|
omp_for_body = gimple_omp_body (stmt);
|
6299 |
|
|
if (!gimple_seq_empty_p (omp_for_body)
|
6300 |
|
|
&& gimple_code (gimple_seq_first_stmt (omp_for_body)) == GIMPLE_BIND)
|
6301 |
|
|
{
|
6302 |
|
|
tree vars = gimple_bind_vars (gimple_seq_first_stmt (omp_for_body));
|
6303 |
|
|
gimple_bind_append_vars (new_stmt, vars);
|
6304 |
|
|
}
|
6305 |
|
|
|
6306 |
|
|
/* The pre-body and input clauses go before the lowered GIMPLE_OMP_FOR. */
|
6307 |
|
|
dlist = NULL;
|
6308 |
|
|
body = NULL;
|
6309 |
|
|
lower_rec_input_clauses (gimple_omp_for_clauses (stmt), &body, &dlist, ctx);
|
6310 |
|
|
gimple_seq_add_seq (&body, gimple_omp_for_pre_body (stmt));
|
6311 |
|
|
|
6312 |
|
|
/* Lower the header expressions. At this point, we can assume that
|
6313 |
|
|
the header is of the form:
|
6314 |
|
|
|
6315 |
|
|
#pragma omp for (V = VAL1; V {<|>|<=|>=} VAL2; V = V [+-] VAL3)
|
6316 |
|
|
|
6317 |
|
|
We just need to make sure that VAL1, VAL2 and VAL3 are lowered
|
6318 |
|
|
using the .omp_data_s mapping, if needed. */
|
6319 |
|
|
for (i = 0; i < gimple_omp_for_collapse (stmt); i++)
|
6320 |
|
|
{
|
6321 |
|
|
rhs_p = gimple_omp_for_initial_ptr (stmt, i);
|
6322 |
|
|
if (!is_gimple_min_invariant (*rhs_p))
|
6323 |
|
|
*rhs_p = get_formal_tmp_var (*rhs_p, &body);
|
6324 |
|
|
|
6325 |
|
|
rhs_p = gimple_omp_for_final_ptr (stmt, i);
|
6326 |
|
|
if (!is_gimple_min_invariant (*rhs_p))
|
6327 |
|
|
*rhs_p = get_formal_tmp_var (*rhs_p, &body);
|
6328 |
|
|
|
6329 |
|
|
rhs_p = &TREE_OPERAND (gimple_omp_for_incr (stmt, i), 1);
|
6330 |
|
|
if (!is_gimple_min_invariant (*rhs_p))
|
6331 |
|
|
*rhs_p = get_formal_tmp_var (*rhs_p, &body);
|
6332 |
|
|
}
|
6333 |
|
|
|
6334 |
|
|
/* Once lowered, extract the bounds and clauses. */
|
6335 |
|
|
extract_omp_for_data (stmt, &fd, NULL);
|
6336 |
|
|
|
6337 |
|
|
lower_omp_for_lastprivate (&fd, &body, &dlist, ctx);
|
6338 |
|
|
|
6339 |
|
|
gimple_seq_add_stmt (&body, stmt);
|
6340 |
|
|
gimple_seq_add_seq (&body, gimple_omp_body (stmt));
|
6341 |
|
|
|
6342 |
|
|
gimple_seq_add_stmt (&body, gimple_build_omp_continue (fd.loop.v,
|
6343 |
|
|
fd.loop.v));
|
6344 |
|
|
|
6345 |
|
|
/* After the loop, add exit clauses. */
|
6346 |
|
|
lower_reduction_clauses (gimple_omp_for_clauses (stmt), &body, ctx);
|
6347 |
|
|
gimple_seq_add_seq (&body, dlist);
|
6348 |
|
|
|
6349 |
|
|
body = maybe_catch_exception (body);
|
6350 |
|
|
|
6351 |
|
|
/* Region exit marker goes at the end of the loop body. */
|
6352 |
|
|
gimple_seq_add_stmt (&body, gimple_build_omp_return (fd.have_nowait));
|
6353 |
|
|
|
6354 |
|
|
pop_gimplify_context (new_stmt);
|
6355 |
|
|
|
6356 |
|
|
gimple_bind_append_vars (new_stmt, ctx->block_vars);
|
6357 |
|
|
BLOCK_VARS (block) = gimple_bind_vars (new_stmt);
|
6358 |
|
|
if (BLOCK_VARS (block))
|
6359 |
|
|
TREE_USED (block) = 1;
|
6360 |
|
|
|
6361 |
|
|
gimple_bind_set_body (new_stmt, body);
|
6362 |
|
|
gimple_omp_set_body (stmt, NULL);
|
6363 |
|
|
gimple_omp_for_set_pre_body (stmt, NULL);
|
6364 |
|
|
gsi_replace (gsi_p, new_stmt, true);
|
6365 |
|
|
}
|
6366 |
|
|
|
6367 |
|
|
/* Callback for walk_stmts. Check if the current statement only contains
|
6368 |
|
|
GIMPLE_OMP_FOR or GIMPLE_OMP_PARALLEL. */
|
6369 |
|
|
|
6370 |
|
|
static tree
|
6371 |
|
|
check_combined_parallel (gimple_stmt_iterator *gsi_p,
|
6372 |
|
|
bool *handled_ops_p,
|
6373 |
|
|
struct walk_stmt_info *wi)
|
6374 |
|
|
{
|
6375 |
|
|
int *info = (int *) wi->info;
|
6376 |
|
|
gimple stmt = gsi_stmt (*gsi_p);
|
6377 |
|
|
|
6378 |
|
|
*handled_ops_p = true;
|
6379 |
|
|
switch (gimple_code (stmt))
|
6380 |
|
|
{
|
6381 |
|
|
WALK_SUBSTMTS;
|
6382 |
|
|
|
6383 |
|
|
case GIMPLE_OMP_FOR:
|
6384 |
|
|
case GIMPLE_OMP_SECTIONS:
|
6385 |
|
|
*info = *info == 0 ? 1 : -1;
|
6386 |
|
|
break;
|
6387 |
|
|
default:
|
6388 |
|
|
*info = -1;
|
6389 |
|
|
break;
|
6390 |
|
|
}
|
6391 |
|
|
return NULL;
|
6392 |
|
|
}
|
6393 |
|
|
|
6394 |
|
|
struct omp_taskcopy_context
|
6395 |
|
|
{
|
6396 |
|
|
/* This field must be at the beginning, as we do "inheritance": Some
|
6397 |
|
|
callback functions for tree-inline.c (e.g., omp_copy_decl)
|
6398 |
|
|
receive a copy_body_data pointer that is up-casted to an
|
6399 |
|
|
omp_context pointer. */
|
6400 |
|
|
copy_body_data cb;
|
6401 |
|
|
omp_context *ctx;
|
6402 |
|
|
};
|
6403 |
|
|
|
6404 |
|
|
static tree
|
6405 |
|
|
task_copyfn_copy_decl (tree var, copy_body_data *cb)
|
6406 |
|
|
{
|
6407 |
|
|
struct omp_taskcopy_context *tcctx = (struct omp_taskcopy_context *) cb;
|
6408 |
|
|
|
6409 |
|
|
if (splay_tree_lookup (tcctx->ctx->sfield_map, (splay_tree_key) var))
|
6410 |
|
|
return create_tmp_var (TREE_TYPE (var), NULL);
|
6411 |
|
|
|
6412 |
|
|
return var;
|
6413 |
|
|
}
|
6414 |
|
|
|
6415 |
|
|
static tree
|
6416 |
|
|
task_copyfn_remap_type (struct omp_taskcopy_context *tcctx, tree orig_type)
|
6417 |
|
|
{
|
6418 |
|
|
tree name, new_fields = NULL, type, f;
|
6419 |
|
|
|
6420 |
|
|
type = lang_hooks.types.make_type (RECORD_TYPE);
|
6421 |
|
|
name = DECL_NAME (TYPE_NAME (orig_type));
|
6422 |
|
|
name = build_decl (gimple_location (tcctx->ctx->stmt),
|
6423 |
|
|
TYPE_DECL, name, type);
|
6424 |
|
|
TYPE_NAME (type) = name;
|
6425 |
|
|
|
6426 |
|
|
for (f = TYPE_FIELDS (orig_type); f ; f = TREE_CHAIN (f))
|
6427 |
|
|
{
|
6428 |
|
|
tree new_f = copy_node (f);
|
6429 |
|
|
DECL_CONTEXT (new_f) = type;
|
6430 |
|
|
TREE_TYPE (new_f) = remap_type (TREE_TYPE (f), &tcctx->cb);
|
6431 |
|
|
TREE_CHAIN (new_f) = new_fields;
|
6432 |
|
|
walk_tree (&DECL_SIZE (new_f), copy_tree_body_r, &tcctx->cb, NULL);
|
6433 |
|
|
walk_tree (&DECL_SIZE_UNIT (new_f), copy_tree_body_r, &tcctx->cb, NULL);
|
6434 |
|
|
walk_tree (&DECL_FIELD_OFFSET (new_f), copy_tree_body_r,
|
6435 |
|
|
&tcctx->cb, NULL);
|
6436 |
|
|
new_fields = new_f;
|
6437 |
|
|
*pointer_map_insert (tcctx->cb.decl_map, f) = new_f;
|
6438 |
|
|
}
|
6439 |
|
|
TYPE_FIELDS (type) = nreverse (new_fields);
|
6440 |
|
|
layout_type (type);
|
6441 |
|
|
return type;
|
6442 |
|
|
}
|
6443 |
|
|
|
6444 |
|
|
/* Create task copyfn. */
|
6445 |
|
|
|
6446 |
|
|
static void
|
6447 |
|
|
create_task_copyfn (gimple task_stmt, omp_context *ctx)
|
6448 |
|
|
{
|
6449 |
|
|
struct function *child_cfun;
|
6450 |
|
|
tree child_fn, t, c, src, dst, f, sf, arg, sarg, decl;
|
6451 |
|
|
tree record_type, srecord_type, bind, list;
|
6452 |
|
|
bool record_needs_remap = false, srecord_needs_remap = false;
|
6453 |
|
|
splay_tree_node n;
|
6454 |
|
|
struct omp_taskcopy_context tcctx;
|
6455 |
|
|
struct gimplify_ctx gctx;
|
6456 |
|
|
location_t loc = gimple_location (task_stmt);
|
6457 |
|
|
|
6458 |
|
|
child_fn = gimple_omp_task_copy_fn (task_stmt);
|
6459 |
|
|
child_cfun = DECL_STRUCT_FUNCTION (child_fn);
|
6460 |
|
|
gcc_assert (child_cfun->cfg == NULL);
|
6461 |
|
|
DECL_SAVED_TREE (child_fn) = alloc_stmt_list ();
|
6462 |
|
|
|
6463 |
|
|
/* Reset DECL_CONTEXT on function arguments. */
|
6464 |
|
|
for (t = DECL_ARGUMENTS (child_fn); t; t = DECL_CHAIN (t))
|
6465 |
|
|
DECL_CONTEXT (t) = child_fn;
|
6466 |
|
|
|
6467 |
|
|
/* Populate the function. */
|
6468 |
|
|
push_gimplify_context (&gctx);
|
6469 |
|
|
current_function_decl = child_fn;
|
6470 |
|
|
|
6471 |
|
|
bind = build3 (BIND_EXPR, void_type_node, NULL, NULL, NULL);
|
6472 |
|
|
TREE_SIDE_EFFECTS (bind) = 1;
|
6473 |
|
|
list = NULL;
|
6474 |
|
|
DECL_SAVED_TREE (child_fn) = bind;
|
6475 |
|
|
DECL_SOURCE_LOCATION (child_fn) = gimple_location (task_stmt);
|
6476 |
|
|
|
6477 |
|
|
/* Remap src and dst argument types if needed. */
|
6478 |
|
|
record_type = ctx->record_type;
|
6479 |
|
|
srecord_type = ctx->srecord_type;
|
6480 |
|
|
for (f = TYPE_FIELDS (record_type); f ; f = DECL_CHAIN (f))
|
6481 |
|
|
if (variably_modified_type_p (TREE_TYPE (f), ctx->cb.src_fn))
|
6482 |
|
|
{
|
6483 |
|
|
record_needs_remap = true;
|
6484 |
|
|
break;
|
6485 |
|
|
}
|
6486 |
|
|
for (f = TYPE_FIELDS (srecord_type); f ; f = DECL_CHAIN (f))
|
6487 |
|
|
if (variably_modified_type_p (TREE_TYPE (f), ctx->cb.src_fn))
|
6488 |
|
|
{
|
6489 |
|
|
srecord_needs_remap = true;
|
6490 |
|
|
break;
|
6491 |
|
|
}
|
6492 |
|
|
|
6493 |
|
|
if (record_needs_remap || srecord_needs_remap)
|
6494 |
|
|
{
|
6495 |
|
|
memset (&tcctx, '\0', sizeof (tcctx));
|
6496 |
|
|
tcctx.cb.src_fn = ctx->cb.src_fn;
|
6497 |
|
|
tcctx.cb.dst_fn = child_fn;
|
6498 |
|
|
tcctx.cb.src_node = cgraph_get_node (tcctx.cb.src_fn);
|
6499 |
|
|
gcc_checking_assert (tcctx.cb.src_node);
|
6500 |
|
|
tcctx.cb.dst_node = tcctx.cb.src_node;
|
6501 |
|
|
tcctx.cb.src_cfun = ctx->cb.src_cfun;
|
6502 |
|
|
tcctx.cb.copy_decl = task_copyfn_copy_decl;
|
6503 |
|
|
tcctx.cb.eh_lp_nr = 0;
|
6504 |
|
|
tcctx.cb.transform_call_graph_edges = CB_CGE_MOVE;
|
6505 |
|
|
tcctx.cb.decl_map = pointer_map_create ();
|
6506 |
|
|
tcctx.ctx = ctx;
|
6507 |
|
|
|
6508 |
|
|
if (record_needs_remap)
|
6509 |
|
|
record_type = task_copyfn_remap_type (&tcctx, record_type);
|
6510 |
|
|
if (srecord_needs_remap)
|
6511 |
|
|
srecord_type = task_copyfn_remap_type (&tcctx, srecord_type);
|
6512 |
|
|
}
|
6513 |
|
|
else
|
6514 |
|
|
tcctx.cb.decl_map = NULL;
|
6515 |
|
|
|
6516 |
|
|
push_cfun (child_cfun);
|
6517 |
|
|
|
6518 |
|
|
arg = DECL_ARGUMENTS (child_fn);
|
6519 |
|
|
TREE_TYPE (arg) = build_pointer_type (record_type);
|
6520 |
|
|
sarg = DECL_CHAIN (arg);
|
6521 |
|
|
TREE_TYPE (sarg) = build_pointer_type (srecord_type);
|
6522 |
|
|
|
6523 |
|
|
/* First pass: initialize temporaries used in record_type and srecord_type
|
6524 |
|
|
sizes and field offsets. */
|
6525 |
|
|
if (tcctx.cb.decl_map)
|
6526 |
|
|
for (c = gimple_omp_task_clauses (task_stmt); c; c = OMP_CLAUSE_CHAIN (c))
|
6527 |
|
|
if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_FIRSTPRIVATE)
|
6528 |
|
|
{
|
6529 |
|
|
tree *p;
|
6530 |
|
|
|
6531 |
|
|
decl = OMP_CLAUSE_DECL (c);
|
6532 |
|
|
p = (tree *) pointer_map_contains (tcctx.cb.decl_map, decl);
|
6533 |
|
|
if (p == NULL)
|
6534 |
|
|
continue;
|
6535 |
|
|
n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl);
|
6536 |
|
|
sf = (tree) n->value;
|
6537 |
|
|
sf = *(tree *) pointer_map_contains (tcctx.cb.decl_map, sf);
|
6538 |
|
|
src = build_simple_mem_ref_loc (loc, sarg);
|
6539 |
|
|
src = build3 (COMPONENT_REF, TREE_TYPE (sf), src, sf, NULL);
|
6540 |
|
|
t = build2 (MODIFY_EXPR, TREE_TYPE (*p), *p, src);
|
6541 |
|
|
append_to_statement_list (t, &list);
|
6542 |
|
|
}
|
6543 |
|
|
|
6544 |
|
|
/* Second pass: copy shared var pointers and copy construct non-VLA
|
6545 |
|
|
firstprivate vars. */
|
6546 |
|
|
for (c = gimple_omp_task_clauses (task_stmt); c; c = OMP_CLAUSE_CHAIN (c))
|
6547 |
|
|
switch (OMP_CLAUSE_CODE (c))
|
6548 |
|
|
{
|
6549 |
|
|
case OMP_CLAUSE_SHARED:
|
6550 |
|
|
decl = OMP_CLAUSE_DECL (c);
|
6551 |
|
|
n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl);
|
6552 |
|
|
if (n == NULL)
|
6553 |
|
|
break;
|
6554 |
|
|
f = (tree) n->value;
|
6555 |
|
|
if (tcctx.cb.decl_map)
|
6556 |
|
|
f = *(tree *) pointer_map_contains (tcctx.cb.decl_map, f);
|
6557 |
|
|
n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl);
|
6558 |
|
|
sf = (tree) n->value;
|
6559 |
|
|
if (tcctx.cb.decl_map)
|
6560 |
|
|
sf = *(tree *) pointer_map_contains (tcctx.cb.decl_map, sf);
|
6561 |
|
|
src = build_simple_mem_ref_loc (loc, sarg);
|
6562 |
|
|
src = build3 (COMPONENT_REF, TREE_TYPE (sf), src, sf, NULL);
|
6563 |
|
|
dst = build_simple_mem_ref_loc (loc, arg);
|
6564 |
|
|
dst = build3 (COMPONENT_REF, TREE_TYPE (f), dst, f, NULL);
|
6565 |
|
|
t = build2 (MODIFY_EXPR, TREE_TYPE (dst), dst, src);
|
6566 |
|
|
append_to_statement_list (t, &list);
|
6567 |
|
|
break;
|
6568 |
|
|
case OMP_CLAUSE_FIRSTPRIVATE:
|
6569 |
|
|
decl = OMP_CLAUSE_DECL (c);
|
6570 |
|
|
if (is_variable_sized (decl))
|
6571 |
|
|
break;
|
6572 |
|
|
n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl);
|
6573 |
|
|
if (n == NULL)
|
6574 |
|
|
break;
|
6575 |
|
|
f = (tree) n->value;
|
6576 |
|
|
if (tcctx.cb.decl_map)
|
6577 |
|
|
f = *(tree *) pointer_map_contains (tcctx.cb.decl_map, f);
|
6578 |
|
|
n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl);
|
6579 |
|
|
if (n != NULL)
|
6580 |
|
|
{
|
6581 |
|
|
sf = (tree) n->value;
|
6582 |
|
|
if (tcctx.cb.decl_map)
|
6583 |
|
|
sf = *(tree *) pointer_map_contains (tcctx.cb.decl_map, sf);
|
6584 |
|
|
src = build_simple_mem_ref_loc (loc, sarg);
|
6585 |
|
|
src = build3 (COMPONENT_REF, TREE_TYPE (sf), src, sf, NULL);
|
6586 |
|
|
if (use_pointer_for_field (decl, NULL) || is_reference (decl))
|
6587 |
|
|
src = build_simple_mem_ref_loc (loc, src);
|
6588 |
|
|
}
|
6589 |
|
|
else
|
6590 |
|
|
src = decl;
|
6591 |
|
|
dst = build_simple_mem_ref_loc (loc, arg);
|
6592 |
|
|
dst = build3 (COMPONENT_REF, TREE_TYPE (f), dst, f, NULL);
|
6593 |
|
|
t = lang_hooks.decls.omp_clause_copy_ctor (c, dst, src);
|
6594 |
|
|
append_to_statement_list (t, &list);
|
6595 |
|
|
break;
|
6596 |
|
|
case OMP_CLAUSE_PRIVATE:
|
6597 |
|
|
if (! OMP_CLAUSE_PRIVATE_OUTER_REF (c))
|
6598 |
|
|
break;
|
6599 |
|
|
decl = OMP_CLAUSE_DECL (c);
|
6600 |
|
|
n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl);
|
6601 |
|
|
f = (tree) n->value;
|
6602 |
|
|
if (tcctx.cb.decl_map)
|
6603 |
|
|
f = *(tree *) pointer_map_contains (tcctx.cb.decl_map, f);
|
6604 |
|
|
n = splay_tree_lookup (ctx->sfield_map, (splay_tree_key) decl);
|
6605 |
|
|
if (n != NULL)
|
6606 |
|
|
{
|
6607 |
|
|
sf = (tree) n->value;
|
6608 |
|
|
if (tcctx.cb.decl_map)
|
6609 |
|
|
sf = *(tree *) pointer_map_contains (tcctx.cb.decl_map, sf);
|
6610 |
|
|
src = build_simple_mem_ref_loc (loc, sarg);
|
6611 |
|
|
src = build3 (COMPONENT_REF, TREE_TYPE (sf), src, sf, NULL);
|
6612 |
|
|
if (use_pointer_for_field (decl, NULL))
|
6613 |
|
|
src = build_simple_mem_ref_loc (loc, src);
|
6614 |
|
|
}
|
6615 |
|
|
else
|
6616 |
|
|
src = decl;
|
6617 |
|
|
dst = build_simple_mem_ref_loc (loc, arg);
|
6618 |
|
|
dst = build3 (COMPONENT_REF, TREE_TYPE (f), dst, f, NULL);
|
6619 |
|
|
t = build2 (MODIFY_EXPR, TREE_TYPE (dst), dst, src);
|
6620 |
|
|
append_to_statement_list (t, &list);
|
6621 |
|
|
break;
|
6622 |
|
|
default:
|
6623 |
|
|
break;
|
6624 |
|
|
}
|
6625 |
|
|
|
6626 |
|
|
/* Last pass: handle VLA firstprivates. */
|
6627 |
|
|
if (tcctx.cb.decl_map)
|
6628 |
|
|
for (c = gimple_omp_task_clauses (task_stmt); c; c = OMP_CLAUSE_CHAIN (c))
|
6629 |
|
|
if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_FIRSTPRIVATE)
|
6630 |
|
|
{
|
6631 |
|
|
tree ind, ptr, df;
|
6632 |
|
|
|
6633 |
|
|
decl = OMP_CLAUSE_DECL (c);
|
6634 |
|
|
if (!is_variable_sized (decl))
|
6635 |
|
|
continue;
|
6636 |
|
|
n = splay_tree_lookup (ctx->field_map, (splay_tree_key) decl);
|
6637 |
|
|
if (n == NULL)
|
6638 |
|
|
continue;
|
6639 |
|
|
f = (tree) n->value;
|
6640 |
|
|
f = *(tree *) pointer_map_contains (tcctx.cb.decl_map, f);
|
6641 |
|
|
gcc_assert (DECL_HAS_VALUE_EXPR_P (decl));
|
6642 |
|
|
ind = DECL_VALUE_EXPR (decl);
|
6643 |
|
|
gcc_assert (TREE_CODE (ind) == INDIRECT_REF);
|
6644 |
|
|
gcc_assert (DECL_P (TREE_OPERAND (ind, 0)));
|
6645 |
|
|
n = splay_tree_lookup (ctx->sfield_map,
|
6646 |
|
|
(splay_tree_key) TREE_OPERAND (ind, 0));
|
6647 |
|
|
sf = (tree) n->value;
|
6648 |
|
|
sf = *(tree *) pointer_map_contains (tcctx.cb.decl_map, sf);
|
6649 |
|
|
src = build_simple_mem_ref_loc (loc, sarg);
|
6650 |
|
|
src = build3 (COMPONENT_REF, TREE_TYPE (sf), src, sf, NULL);
|
6651 |
|
|
src = build_simple_mem_ref_loc (loc, src);
|
6652 |
|
|
dst = build_simple_mem_ref_loc (loc, arg);
|
6653 |
|
|
dst = build3 (COMPONENT_REF, TREE_TYPE (f), dst, f, NULL);
|
6654 |
|
|
t = lang_hooks.decls.omp_clause_copy_ctor (c, dst, src);
|
6655 |
|
|
append_to_statement_list (t, &list);
|
6656 |
|
|
n = splay_tree_lookup (ctx->field_map,
|
6657 |
|
|
(splay_tree_key) TREE_OPERAND (ind, 0));
|
6658 |
|
|
df = (tree) n->value;
|
6659 |
|
|
df = *(tree *) pointer_map_contains (tcctx.cb.decl_map, df);
|
6660 |
|
|
ptr = build_simple_mem_ref_loc (loc, arg);
|
6661 |
|
|
ptr = build3 (COMPONENT_REF, TREE_TYPE (df), ptr, df, NULL);
|
6662 |
|
|
t = build2 (MODIFY_EXPR, TREE_TYPE (ptr), ptr,
|
6663 |
|
|
build_fold_addr_expr_loc (loc, dst));
|
6664 |
|
|
append_to_statement_list (t, &list);
|
6665 |
|
|
}
|
6666 |
|
|
|
6667 |
|
|
t = build1 (RETURN_EXPR, void_type_node, NULL);
|
6668 |
|
|
append_to_statement_list (t, &list);
|
6669 |
|
|
|
6670 |
|
|
if (tcctx.cb.decl_map)
|
6671 |
|
|
pointer_map_destroy (tcctx.cb.decl_map);
|
6672 |
|
|
pop_gimplify_context (NULL);
|
6673 |
|
|
BIND_EXPR_BODY (bind) = list;
|
6674 |
|
|
pop_cfun ();
|
6675 |
|
|
current_function_decl = ctx->cb.src_fn;
|
6676 |
|
|
}
|
6677 |
|
|
|
6678 |
|
|
/* Lower the OpenMP parallel or task directive in the current statement
|
6679 |
|
|
in GSI_P. CTX holds context information for the directive. */
|
6680 |
|
|
|
6681 |
|
|
static void
|
6682 |
|
|
lower_omp_taskreg (gimple_stmt_iterator *gsi_p, omp_context *ctx)
|
6683 |
|
|
{
|
6684 |
|
|
tree clauses;
|
6685 |
|
|
tree child_fn, t;
|
6686 |
|
|
gimple stmt = gsi_stmt (*gsi_p);
|
6687 |
|
|
gimple par_bind, bind;
|
6688 |
|
|
gimple_seq par_body, olist, ilist, par_olist, par_ilist, new_body;
|
6689 |
|
|
struct gimplify_ctx gctx;
|
6690 |
|
|
location_t loc = gimple_location (stmt);
|
6691 |
|
|
|
6692 |
|
|
clauses = gimple_omp_taskreg_clauses (stmt);
|
6693 |
|
|
par_bind = gimple_seq_first_stmt (gimple_omp_body (stmt));
|
6694 |
|
|
par_body = gimple_bind_body (par_bind);
|
6695 |
|
|
child_fn = ctx->cb.dst_fn;
|
6696 |
|
|
if (gimple_code (stmt) == GIMPLE_OMP_PARALLEL
|
6697 |
|
|
&& !gimple_omp_parallel_combined_p (stmt))
|
6698 |
|
|
{
|
6699 |
|
|
struct walk_stmt_info wi;
|
6700 |
|
|
int ws_num = 0;
|
6701 |
|
|
|
6702 |
|
|
memset (&wi, 0, sizeof (wi));
|
6703 |
|
|
wi.info = &ws_num;
|
6704 |
|
|
wi.val_only = true;
|
6705 |
|
|
walk_gimple_seq (par_body, check_combined_parallel, NULL, &wi);
|
6706 |
|
|
if (ws_num == 1)
|
6707 |
|
|
gimple_omp_parallel_set_combined_p (stmt, true);
|
6708 |
|
|
}
|
6709 |
|
|
if (ctx->srecord_type)
|
6710 |
|
|
create_task_copyfn (stmt, ctx);
|
6711 |
|
|
|
6712 |
|
|
push_gimplify_context (&gctx);
|
6713 |
|
|
|
6714 |
|
|
par_olist = NULL;
|
6715 |
|
|
par_ilist = NULL;
|
6716 |
|
|
lower_rec_input_clauses (clauses, &par_ilist, &par_olist, ctx);
|
6717 |
|
|
lower_omp (par_body, ctx);
|
6718 |
|
|
if (gimple_code (stmt) == GIMPLE_OMP_PARALLEL)
|
6719 |
|
|
lower_reduction_clauses (clauses, &par_olist, ctx);
|
6720 |
|
|
|
6721 |
|
|
/* Declare all the variables created by mapping and the variables
|
6722 |
|
|
declared in the scope of the parallel body. */
|
6723 |
|
|
record_vars_into (ctx->block_vars, child_fn);
|
6724 |
|
|
record_vars_into (gimple_bind_vars (par_bind), child_fn);
|
6725 |
|
|
|
6726 |
|
|
if (ctx->record_type)
|
6727 |
|
|
{
|
6728 |
|
|
ctx->sender_decl
|
6729 |
|
|
= create_tmp_var (ctx->srecord_type ? ctx->srecord_type
|
6730 |
|
|
: ctx->record_type, ".omp_data_o");
|
6731 |
|
|
DECL_NAMELESS (ctx->sender_decl) = 1;
|
6732 |
|
|
TREE_ADDRESSABLE (ctx->sender_decl) = 1;
|
6733 |
|
|
gimple_omp_taskreg_set_data_arg (stmt, ctx->sender_decl);
|
6734 |
|
|
}
|
6735 |
|
|
|
6736 |
|
|
olist = NULL;
|
6737 |
|
|
ilist = NULL;
|
6738 |
|
|
lower_send_clauses (clauses, &ilist, &olist, ctx);
|
6739 |
|
|
lower_send_shared_vars (&ilist, &olist, ctx);
|
6740 |
|
|
|
6741 |
|
|
/* Once all the expansions are done, sequence all the different
|
6742 |
|
|
fragments inside gimple_omp_body. */
|
6743 |
|
|
|
6744 |
|
|
new_body = NULL;
|
6745 |
|
|
|
6746 |
|
|
if (ctx->record_type)
|
6747 |
|
|
{
|
6748 |
|
|
t = build_fold_addr_expr_loc (loc, ctx->sender_decl);
|
6749 |
|
|
/* fixup_child_record_type might have changed receiver_decl's type. */
|
6750 |
|
|
t = fold_convert_loc (loc, TREE_TYPE (ctx->receiver_decl), t);
|
6751 |
|
|
gimple_seq_add_stmt (&new_body,
|
6752 |
|
|
gimple_build_assign (ctx->receiver_decl, t));
|
6753 |
|
|
}
|
6754 |
|
|
|
6755 |
|
|
gimple_seq_add_seq (&new_body, par_ilist);
|
6756 |
|
|
gimple_seq_add_seq (&new_body, par_body);
|
6757 |
|
|
gimple_seq_add_seq (&new_body, par_olist);
|
6758 |
|
|
new_body = maybe_catch_exception (new_body);
|
6759 |
|
|
gimple_seq_add_stmt (&new_body, gimple_build_omp_return (false));
|
6760 |
|
|
gimple_omp_set_body (stmt, new_body);
|
6761 |
|
|
|
6762 |
|
|
bind = gimple_build_bind (NULL, NULL, gimple_bind_block (par_bind));
|
6763 |
|
|
gimple_bind_add_stmt (bind, stmt);
|
6764 |
|
|
if (ilist || olist)
|
6765 |
|
|
{
|
6766 |
|
|
gimple_seq_add_stmt (&ilist, bind);
|
6767 |
|
|
gimple_seq_add_seq (&ilist, olist);
|
6768 |
|
|
bind = gimple_build_bind (NULL, ilist, NULL);
|
6769 |
|
|
}
|
6770 |
|
|
|
6771 |
|
|
gsi_replace (gsi_p, bind, true);
|
6772 |
|
|
|
6773 |
|
|
pop_gimplify_context (NULL);
|
6774 |
|
|
}
|
6775 |
|
|
|
6776 |
|
|
/* Callback for lower_omp_1. Return non-NULL if *tp needs to be
|
6777 |
|
|
regimplified. If DATA is non-NULL, lower_omp_1 is outside
|
6778 |
|
|
of OpenMP context, but with task_shared_vars set. */
|
6779 |
|
|
|
6780 |
|
|
static tree
|
6781 |
|
|
lower_omp_regimplify_p (tree *tp, int *walk_subtrees,
|
6782 |
|
|
void *data)
|
6783 |
|
|
{
|
6784 |
|
|
tree t = *tp;
|
6785 |
|
|
|
6786 |
|
|
/* Any variable with DECL_VALUE_EXPR needs to be regimplified. */
|
6787 |
|
|
if (TREE_CODE (t) == VAR_DECL && data == NULL && DECL_HAS_VALUE_EXPR_P (t))
|
6788 |
|
|
return t;
|
6789 |
|
|
|
6790 |
|
|
if (task_shared_vars
|
6791 |
|
|
&& DECL_P (t)
|
6792 |
|
|
&& bitmap_bit_p (task_shared_vars, DECL_UID (t)))
|
6793 |
|
|
return t;
|
6794 |
|
|
|
6795 |
|
|
/* If a global variable has been privatized, TREE_CONSTANT on
|
6796 |
|
|
ADDR_EXPR might be wrong. */
|
6797 |
|
|
if (data == NULL && TREE_CODE (t) == ADDR_EXPR)
|
6798 |
|
|
recompute_tree_invariant_for_addr_expr (t);
|
6799 |
|
|
|
6800 |
|
|
*walk_subtrees = !TYPE_P (t) && !DECL_P (t);
|
6801 |
|
|
return NULL_TREE;
|
6802 |
|
|
}
|
6803 |
|
|
|
6804 |
|
|
static void
|
6805 |
|
|
lower_omp_1 (gimple_stmt_iterator *gsi_p, omp_context *ctx)
|
6806 |
|
|
{
|
6807 |
|
|
gimple stmt = gsi_stmt (*gsi_p);
|
6808 |
|
|
struct walk_stmt_info wi;
|
6809 |
|
|
|
6810 |
|
|
if (gimple_has_location (stmt))
|
6811 |
|
|
input_location = gimple_location (stmt);
|
6812 |
|
|
|
6813 |
|
|
if (task_shared_vars)
|
6814 |
|
|
memset (&wi, '\0', sizeof (wi));
|
6815 |
|
|
|
6816 |
|
|
/* If we have issued syntax errors, avoid doing any heavy lifting.
|
6817 |
|
|
Just replace the OpenMP directives with a NOP to avoid
|
6818 |
|
|
confusing RTL expansion. */
|
6819 |
|
|
if (seen_error () && is_gimple_omp (stmt))
|
6820 |
|
|
{
|
6821 |
|
|
gsi_replace (gsi_p, gimple_build_nop (), true);
|
6822 |
|
|
return;
|
6823 |
|
|
}
|
6824 |
|
|
|
6825 |
|
|
switch (gimple_code (stmt))
|
6826 |
|
|
{
|
6827 |
|
|
case GIMPLE_COND:
|
6828 |
|
|
if ((ctx || task_shared_vars)
|
6829 |
|
|
&& (walk_tree (gimple_cond_lhs_ptr (stmt), lower_omp_regimplify_p,
|
6830 |
|
|
ctx ? NULL : &wi, NULL)
|
6831 |
|
|
|| walk_tree (gimple_cond_rhs_ptr (stmt), lower_omp_regimplify_p,
|
6832 |
|
|
ctx ? NULL : &wi, NULL)))
|
6833 |
|
|
gimple_regimplify_operands (stmt, gsi_p);
|
6834 |
|
|
break;
|
6835 |
|
|
case GIMPLE_CATCH:
|
6836 |
|
|
lower_omp (gimple_catch_handler (stmt), ctx);
|
6837 |
|
|
break;
|
6838 |
|
|
case GIMPLE_EH_FILTER:
|
6839 |
|
|
lower_omp (gimple_eh_filter_failure (stmt), ctx);
|
6840 |
|
|
break;
|
6841 |
|
|
case GIMPLE_TRY:
|
6842 |
|
|
lower_omp (gimple_try_eval (stmt), ctx);
|
6843 |
|
|
lower_omp (gimple_try_cleanup (stmt), ctx);
|
6844 |
|
|
break;
|
6845 |
|
|
case GIMPLE_BIND:
|
6846 |
|
|
lower_omp (gimple_bind_body (stmt), ctx);
|
6847 |
|
|
break;
|
6848 |
|
|
case GIMPLE_OMP_PARALLEL:
|
6849 |
|
|
case GIMPLE_OMP_TASK:
|
6850 |
|
|
ctx = maybe_lookup_ctx (stmt);
|
6851 |
|
|
lower_omp_taskreg (gsi_p, ctx);
|
6852 |
|
|
break;
|
6853 |
|
|
case GIMPLE_OMP_FOR:
|
6854 |
|
|
ctx = maybe_lookup_ctx (stmt);
|
6855 |
|
|
gcc_assert (ctx);
|
6856 |
|
|
lower_omp_for (gsi_p, ctx);
|
6857 |
|
|
break;
|
6858 |
|
|
case GIMPLE_OMP_SECTIONS:
|
6859 |
|
|
ctx = maybe_lookup_ctx (stmt);
|
6860 |
|
|
gcc_assert (ctx);
|
6861 |
|
|
lower_omp_sections (gsi_p, ctx);
|
6862 |
|
|
break;
|
6863 |
|
|
case GIMPLE_OMP_SINGLE:
|
6864 |
|
|
ctx = maybe_lookup_ctx (stmt);
|
6865 |
|
|
gcc_assert (ctx);
|
6866 |
|
|
lower_omp_single (gsi_p, ctx);
|
6867 |
|
|
break;
|
6868 |
|
|
case GIMPLE_OMP_MASTER:
|
6869 |
|
|
ctx = maybe_lookup_ctx (stmt);
|
6870 |
|
|
gcc_assert (ctx);
|
6871 |
|
|
lower_omp_master (gsi_p, ctx);
|
6872 |
|
|
break;
|
6873 |
|
|
case GIMPLE_OMP_ORDERED:
|
6874 |
|
|
ctx = maybe_lookup_ctx (stmt);
|
6875 |
|
|
gcc_assert (ctx);
|
6876 |
|
|
lower_omp_ordered (gsi_p, ctx);
|
6877 |
|
|
break;
|
6878 |
|
|
case GIMPLE_OMP_CRITICAL:
|
6879 |
|
|
ctx = maybe_lookup_ctx (stmt);
|
6880 |
|
|
gcc_assert (ctx);
|
6881 |
|
|
lower_omp_critical (gsi_p, ctx);
|
6882 |
|
|
break;
|
6883 |
|
|
case GIMPLE_OMP_ATOMIC_LOAD:
|
6884 |
|
|
if ((ctx || task_shared_vars)
|
6885 |
|
|
&& walk_tree (gimple_omp_atomic_load_rhs_ptr (stmt),
|
6886 |
|
|
lower_omp_regimplify_p, ctx ? NULL : &wi, NULL))
|
6887 |
|
|
gimple_regimplify_operands (stmt, gsi_p);
|
6888 |
|
|
break;
|
6889 |
|
|
default:
|
6890 |
|
|
if ((ctx || task_shared_vars)
|
6891 |
|
|
&& walk_gimple_op (stmt, lower_omp_regimplify_p,
|
6892 |
|
|
ctx ? NULL : &wi))
|
6893 |
|
|
gimple_regimplify_operands (stmt, gsi_p);
|
6894 |
|
|
break;
|
6895 |
|
|
}
|
6896 |
|
|
}
|
6897 |
|
|
|
6898 |
|
|
static void
|
6899 |
|
|
lower_omp (gimple_seq body, omp_context *ctx)
|
6900 |
|
|
{
|
6901 |
|
|
location_t saved_location = input_location;
|
6902 |
|
|
gimple_stmt_iterator gsi = gsi_start (body);
|
6903 |
|
|
for (gsi = gsi_start (body); !gsi_end_p (gsi); gsi_next (&gsi))
|
6904 |
|
|
lower_omp_1 (&gsi, ctx);
|
6905 |
|
|
input_location = saved_location;
|
6906 |
|
|
}
|
6907 |
|
|
|
6908 |
|
|
/* Main entry point. */
|
6909 |
|
|
|
6910 |
|
|
static unsigned int
|
6911 |
|
|
execute_lower_omp (void)
|
6912 |
|
|
{
|
6913 |
|
|
gimple_seq body;
|
6914 |
|
|
|
6915 |
|
|
/* This pass always runs, to provide PROP_gimple_lomp.
|
6916 |
|
|
But there is nothing to do unless -fopenmp is given. */
|
6917 |
|
|
if (flag_openmp == 0)
|
6918 |
|
|
return 0;
|
6919 |
|
|
|
6920 |
|
|
all_contexts = splay_tree_new (splay_tree_compare_pointers, 0,
|
6921 |
|
|
delete_omp_context);
|
6922 |
|
|
|
6923 |
|
|
body = gimple_body (current_function_decl);
|
6924 |
|
|
scan_omp (body, NULL);
|
6925 |
|
|
gcc_assert (taskreg_nesting_level == 0);
|
6926 |
|
|
|
6927 |
|
|
if (all_contexts->root)
|
6928 |
|
|
{
|
6929 |
|
|
struct gimplify_ctx gctx;
|
6930 |
|
|
|
6931 |
|
|
if (task_shared_vars)
|
6932 |
|
|
push_gimplify_context (&gctx);
|
6933 |
|
|
lower_omp (body, NULL);
|
6934 |
|
|
if (task_shared_vars)
|
6935 |
|
|
pop_gimplify_context (NULL);
|
6936 |
|
|
}
|
6937 |
|
|
|
6938 |
|
|
if (all_contexts)
|
6939 |
|
|
{
|
6940 |
|
|
splay_tree_delete (all_contexts);
|
6941 |
|
|
all_contexts = NULL;
|
6942 |
|
|
}
|
6943 |
|
|
BITMAP_FREE (task_shared_vars);
|
6944 |
|
|
return 0;
|
6945 |
|
|
}
|
6946 |
|
|
|
6947 |
|
|
struct gimple_opt_pass pass_lower_omp =
|
6948 |
|
|
{
|
6949 |
|
|
{
|
6950 |
|
|
GIMPLE_PASS,
|
6951 |
|
|
"omplower", /* name */
|
6952 |
|
|
NULL, /* gate */
|
6953 |
|
|
execute_lower_omp, /* execute */
|
6954 |
|
|
NULL, /* sub */
|
6955 |
|
|
NULL, /* next */
|
6956 |
|
|
0, /* static_pass_number */
|
6957 |
|
|
TV_NONE, /* tv_id */
|
6958 |
|
|
PROP_gimple_any, /* properties_required */
|
6959 |
|
|
PROP_gimple_lomp, /* properties_provided */
|
6960 |
|
|
0, /* properties_destroyed */
|
6961 |
|
|
0, /* todo_flags_start */
|
6962 |
|
|
|
6963 |
|
|
}
|
6964 |
|
|
};
|
6965 |
|
|
|
6966 |
|
|
/* The following is a utility to diagnose OpenMP structured block violations.
|
6967 |
|
|
It is not part of the "omplower" pass, as that's invoked too late. It
|
6968 |
|
|
should be invoked by the respective front ends after gimplification. */
|
6969 |
|
|
|
6970 |
|
|
static splay_tree all_labels;
|
6971 |
|
|
|
6972 |
|
|
/* Check for mismatched contexts and generate an error if needed. Return
|
6973 |
|
|
true if an error is detected. */
|
6974 |
|
|
|
6975 |
|
|
static bool
|
6976 |
|
|
diagnose_sb_0 (gimple_stmt_iterator *gsi_p,
|
6977 |
|
|
gimple branch_ctx, gimple label_ctx)
|
6978 |
|
|
{
|
6979 |
|
|
if (label_ctx == branch_ctx)
|
6980 |
|
|
return false;
|
6981 |
|
|
|
6982 |
|
|
|
6983 |
|
|
/*
|
6984 |
|
|
Previously we kept track of the label's entire context in diagnose_sb_[12]
|
6985 |
|
|
so we could traverse it and issue a correct "exit" or "enter" error
|
6986 |
|
|
message upon a structured block violation.
|
6987 |
|
|
|
6988 |
|
|
We built the context by building a list with tree_cons'ing, but there is
|
6989 |
|
|
no easy counterpart in gimple tuples. It seems like far too much work
|
6990 |
|
|
for issuing exit/enter error messages. If someone really misses the
|
6991 |
|
|
distinct error message... patches welcome.
|
6992 |
|
|
*/
|
6993 |
|
|
|
6994 |
|
|
#if 0
|
6995 |
|
|
/* Try to avoid confusing the user by producing and error message
|
6996 |
|
|
with correct "exit" or "enter" verbiage. We prefer "exit"
|
6997 |
|
|
unless we can show that LABEL_CTX is nested within BRANCH_CTX. */
|
6998 |
|
|
if (branch_ctx == NULL)
|
6999 |
|
|
exit_p = false;
|
7000 |
|
|
else
|
7001 |
|
|
{
|
7002 |
|
|
while (label_ctx)
|
7003 |
|
|
{
|
7004 |
|
|
if (TREE_VALUE (label_ctx) == branch_ctx)
|
7005 |
|
|
{
|
7006 |
|
|
exit_p = false;
|
7007 |
|
|
break;
|
7008 |
|
|
}
|
7009 |
|
|
label_ctx = TREE_CHAIN (label_ctx);
|
7010 |
|
|
}
|
7011 |
|
|
}
|
7012 |
|
|
|
7013 |
|
|
if (exit_p)
|
7014 |
|
|
error ("invalid exit from OpenMP structured block");
|
7015 |
|
|
else
|
7016 |
|
|
error ("invalid entry to OpenMP structured block");
|
7017 |
|
|
#endif
|
7018 |
|
|
|
7019 |
|
|
/* If it's obvious we have an invalid entry, be specific about the error. */
|
7020 |
|
|
if (branch_ctx == NULL)
|
7021 |
|
|
error ("invalid entry to OpenMP structured block");
|
7022 |
|
|
else
|
7023 |
|
|
/* Otherwise, be vague and lazy, but efficient. */
|
7024 |
|
|
error ("invalid branch to/from an OpenMP structured block");
|
7025 |
|
|
|
7026 |
|
|
gsi_replace (gsi_p, gimple_build_nop (), false);
|
7027 |
|
|
return true;
|
7028 |
|
|
}
|
7029 |
|
|
|
7030 |
|
|
/* Pass 1: Create a minimal tree of OpenMP structured blocks, and record
|
7031 |
|
|
where each label is found. */
|
7032 |
|
|
|
7033 |
|
|
static tree
|
7034 |
|
|
diagnose_sb_1 (gimple_stmt_iterator *gsi_p, bool *handled_ops_p,
|
7035 |
|
|
struct walk_stmt_info *wi)
|
7036 |
|
|
{
|
7037 |
|
|
gimple context = (gimple) wi->info;
|
7038 |
|
|
gimple inner_context;
|
7039 |
|
|
gimple stmt = gsi_stmt (*gsi_p);
|
7040 |
|
|
|
7041 |
|
|
*handled_ops_p = true;
|
7042 |
|
|
|
7043 |
|
|
switch (gimple_code (stmt))
|
7044 |
|
|
{
|
7045 |
|
|
WALK_SUBSTMTS;
|
7046 |
|
|
|
7047 |
|
|
case GIMPLE_OMP_PARALLEL:
|
7048 |
|
|
case GIMPLE_OMP_TASK:
|
7049 |
|
|
case GIMPLE_OMP_SECTIONS:
|
7050 |
|
|
case GIMPLE_OMP_SINGLE:
|
7051 |
|
|
case GIMPLE_OMP_SECTION:
|
7052 |
|
|
case GIMPLE_OMP_MASTER:
|
7053 |
|
|
case GIMPLE_OMP_ORDERED:
|
7054 |
|
|
case GIMPLE_OMP_CRITICAL:
|
7055 |
|
|
/* The minimal context here is just the current OMP construct. */
|
7056 |
|
|
inner_context = stmt;
|
7057 |
|
|
wi->info = inner_context;
|
7058 |
|
|
walk_gimple_seq (gimple_omp_body (stmt), diagnose_sb_1, NULL, wi);
|
7059 |
|
|
wi->info = context;
|
7060 |
|
|
break;
|
7061 |
|
|
|
7062 |
|
|
case GIMPLE_OMP_FOR:
|
7063 |
|
|
inner_context = stmt;
|
7064 |
|
|
wi->info = inner_context;
|
7065 |
|
|
/* gimple_omp_for_{index,initial,final} are all DECLs; no need to
|
7066 |
|
|
walk them. */
|
7067 |
|
|
walk_gimple_seq (gimple_omp_for_pre_body (stmt),
|
7068 |
|
|
diagnose_sb_1, NULL, wi);
|
7069 |
|
|
walk_gimple_seq (gimple_omp_body (stmt), diagnose_sb_1, NULL, wi);
|
7070 |
|
|
wi->info = context;
|
7071 |
|
|
break;
|
7072 |
|
|
|
7073 |
|
|
case GIMPLE_LABEL:
|
7074 |
|
|
splay_tree_insert (all_labels, (splay_tree_key) gimple_label_label (stmt),
|
7075 |
|
|
(splay_tree_value) context);
|
7076 |
|
|
break;
|
7077 |
|
|
|
7078 |
|
|
default:
|
7079 |
|
|
break;
|
7080 |
|
|
}
|
7081 |
|
|
|
7082 |
|
|
return NULL_TREE;
|
7083 |
|
|
}
|
7084 |
|
|
|
7085 |
|
|
/* Pass 2: Check each branch and see if its context differs from that of
|
7086 |
|
|
the destination label's context. */
|
7087 |
|
|
|
7088 |
|
|
static tree
|
7089 |
|
|
diagnose_sb_2 (gimple_stmt_iterator *gsi_p, bool *handled_ops_p,
|
7090 |
|
|
struct walk_stmt_info *wi)
|
7091 |
|
|
{
|
7092 |
|
|
gimple context = (gimple) wi->info;
|
7093 |
|
|
splay_tree_node n;
|
7094 |
|
|
gimple stmt = gsi_stmt (*gsi_p);
|
7095 |
|
|
|
7096 |
|
|
*handled_ops_p = true;
|
7097 |
|
|
|
7098 |
|
|
switch (gimple_code (stmt))
|
7099 |
|
|
{
|
7100 |
|
|
WALK_SUBSTMTS;
|
7101 |
|
|
|
7102 |
|
|
case GIMPLE_OMP_PARALLEL:
|
7103 |
|
|
case GIMPLE_OMP_TASK:
|
7104 |
|
|
case GIMPLE_OMP_SECTIONS:
|
7105 |
|
|
case GIMPLE_OMP_SINGLE:
|
7106 |
|
|
case GIMPLE_OMP_SECTION:
|
7107 |
|
|
case GIMPLE_OMP_MASTER:
|
7108 |
|
|
case GIMPLE_OMP_ORDERED:
|
7109 |
|
|
case GIMPLE_OMP_CRITICAL:
|
7110 |
|
|
wi->info = stmt;
|
7111 |
|
|
walk_gimple_seq (gimple_omp_body (stmt), diagnose_sb_2, NULL, wi);
|
7112 |
|
|
wi->info = context;
|
7113 |
|
|
break;
|
7114 |
|
|
|
7115 |
|
|
case GIMPLE_OMP_FOR:
|
7116 |
|
|
wi->info = stmt;
|
7117 |
|
|
/* gimple_omp_for_{index,initial,final} are all DECLs; no need to
|
7118 |
|
|
walk them. */
|
7119 |
|
|
walk_gimple_seq (gimple_omp_for_pre_body (stmt),
|
7120 |
|
|
diagnose_sb_2, NULL, wi);
|
7121 |
|
|
walk_gimple_seq (gimple_omp_body (stmt), diagnose_sb_2, NULL, wi);
|
7122 |
|
|
wi->info = context;
|
7123 |
|
|
break;
|
7124 |
|
|
|
7125 |
|
|
case GIMPLE_COND:
|
7126 |
|
|
{
|
7127 |
|
|
tree lab = gimple_cond_true_label (stmt);
|
7128 |
|
|
if (lab)
|
7129 |
|
|
{
|
7130 |
|
|
n = splay_tree_lookup (all_labels,
|
7131 |
|
|
(splay_tree_key) lab);
|
7132 |
|
|
diagnose_sb_0 (gsi_p, context,
|
7133 |
|
|
n ? (gimple) n->value : NULL);
|
7134 |
|
|
}
|
7135 |
|
|
lab = gimple_cond_false_label (stmt);
|
7136 |
|
|
if (lab)
|
7137 |
|
|
{
|
7138 |
|
|
n = splay_tree_lookup (all_labels,
|
7139 |
|
|
(splay_tree_key) lab);
|
7140 |
|
|
diagnose_sb_0 (gsi_p, context,
|
7141 |
|
|
n ? (gimple) n->value : NULL);
|
7142 |
|
|
}
|
7143 |
|
|
}
|
7144 |
|
|
break;
|
7145 |
|
|
|
7146 |
|
|
case GIMPLE_GOTO:
|
7147 |
|
|
{
|
7148 |
|
|
tree lab = gimple_goto_dest (stmt);
|
7149 |
|
|
if (TREE_CODE (lab) != LABEL_DECL)
|
7150 |
|
|
break;
|
7151 |
|
|
|
7152 |
|
|
n = splay_tree_lookup (all_labels, (splay_tree_key) lab);
|
7153 |
|
|
diagnose_sb_0 (gsi_p, context, n ? (gimple) n->value : NULL);
|
7154 |
|
|
}
|
7155 |
|
|
break;
|
7156 |
|
|
|
7157 |
|
|
case GIMPLE_SWITCH:
|
7158 |
|
|
{
|
7159 |
|
|
unsigned int i;
|
7160 |
|
|
for (i = 0; i < gimple_switch_num_labels (stmt); ++i)
|
7161 |
|
|
{
|
7162 |
|
|
tree lab = CASE_LABEL (gimple_switch_label (stmt, i));
|
7163 |
|
|
n = splay_tree_lookup (all_labels, (splay_tree_key) lab);
|
7164 |
|
|
if (n && diagnose_sb_0 (gsi_p, context, (gimple) n->value))
|
7165 |
|
|
break;
|
7166 |
|
|
}
|
7167 |
|
|
}
|
7168 |
|
|
break;
|
7169 |
|
|
|
7170 |
|
|
case GIMPLE_RETURN:
|
7171 |
|
|
diagnose_sb_0 (gsi_p, context, NULL);
|
7172 |
|
|
break;
|
7173 |
|
|
|
7174 |
|
|
default:
|
7175 |
|
|
break;
|
7176 |
|
|
}
|
7177 |
|
|
|
7178 |
|
|
return NULL_TREE;
|
7179 |
|
|
}
|
7180 |
|
|
|
7181 |
|
|
static unsigned int
|
7182 |
|
|
diagnose_omp_structured_block_errors (void)
|
7183 |
|
|
{
|
7184 |
|
|
struct walk_stmt_info wi;
|
7185 |
|
|
gimple_seq body = gimple_body (current_function_decl);
|
7186 |
|
|
|
7187 |
|
|
all_labels = splay_tree_new (splay_tree_compare_pointers, 0, 0);
|
7188 |
|
|
|
7189 |
|
|
memset (&wi, 0, sizeof (wi));
|
7190 |
|
|
walk_gimple_seq (body, diagnose_sb_1, NULL, &wi);
|
7191 |
|
|
|
7192 |
|
|
memset (&wi, 0, sizeof (wi));
|
7193 |
|
|
wi.want_locations = true;
|
7194 |
|
|
walk_gimple_seq (body, diagnose_sb_2, NULL, &wi);
|
7195 |
|
|
|
7196 |
|
|
splay_tree_delete (all_labels);
|
7197 |
|
|
all_labels = NULL;
|
7198 |
|
|
|
7199 |
|
|
return 0;
|
7200 |
|
|
}
|
7201 |
|
|
|
7202 |
|
|
static bool
|
7203 |
|
|
gate_diagnose_omp_blocks (void)
|
7204 |
|
|
{
|
7205 |
|
|
return flag_openmp != 0;
|
7206 |
|
|
}
|
7207 |
|
|
|
7208 |
|
|
struct gimple_opt_pass pass_diagnose_omp_blocks =
|
7209 |
|
|
{
|
7210 |
|
|
{
|
7211 |
|
|
GIMPLE_PASS,
|
7212 |
|
|
"*diagnose_omp_blocks", /* name */
|
7213 |
|
|
gate_diagnose_omp_blocks, /* gate */
|
7214 |
|
|
diagnose_omp_structured_block_errors, /* execute */
|
7215 |
|
|
NULL, /* sub */
|
7216 |
|
|
NULL, /* next */
|
7217 |
|
|
0, /* static_pass_number */
|
7218 |
|
|
TV_NONE, /* tv_id */
|
7219 |
|
|
PROP_gimple_any, /* properties_required */
|
7220 |
|
|
0, /* properties_provided */
|
7221 |
|
|
0, /* properties_destroyed */
|
7222 |
|
|
0, /* todo_flags_start */
|
7223 |
|
|
0, /* todo_flags_finish */
|
7224 |
|
|
}
|
7225 |
|
|
};
|
7226 |
|
|
|
7227 |
|
|
#include "gt-omp-low.h"
|