]>
Commit | Line | Data |
---|---|---|
2abae5f1 | 1 | /* Detection of Static Control Parts (SCoP) for Graphite. |
d1e082c2 | 2 | Copyright (C) 2009-2013 Free Software Foundation, Inc. |
2abae5f1 SP |
3 | Contributed by Sebastian Pop <sebastian.pop@amd.com> and |
4 | Tobias Grosser <grosser@fim.uni-passau.de>. | |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify | |
9 | it under the terms of the GNU General Public License as published by | |
10 | the Free Software Foundation; either version 3, or (at your option) | |
11 | any later version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, | |
14 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
16 | GNU General Public License for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
33ad93b9 RG |
23 | |
24 | #ifdef HAVE_cloog | |
25 | #include <isl/set.h> | |
26 | #include <isl/map.h> | |
27 | #include <isl/union_map.h> | |
28 | #include <cloog/cloog.h> | |
29 | #include <cloog/isl/domain.h> | |
30 | #endif | |
31 | ||
2abae5f1 SP |
32 | #include "system.h" |
33 | #include "coretypes.h" | |
4d648807 | 34 | #include "tree.h" |
2fb9a547 AM |
35 | #include "basic-block.h" |
36 | #include "tree-ssa-alias.h" | |
37 | #include "internal-fn.h" | |
38 | #include "gimple-expr.h" | |
39 | #include "is-a.h" | |
442b4905 | 40 | #include "gimple.h" |
5be5c238 | 41 | #include "gimple-iterator.h" |
442b4905 AM |
42 | #include "gimple-ssa.h" |
43 | #include "tree-phinodes.h" | |
44 | #include "ssa-iterators.h" | |
e28030cf AM |
45 | #include "tree-ssa-loop-manip.h" |
46 | #include "tree-ssa-loop-niter.h" | |
442b4905 AM |
47 | #include "tree-ssa-loop.h" |
48 | #include "tree-into-ssa.h" | |
7a300452 | 49 | #include "tree-ssa.h" |
2abae5f1 SP |
50 | #include "cfgloop.h" |
51 | #include "tree-chrec.h" | |
52 | #include "tree-data-ref.h" | |
53 | #include "tree-scalar-evolution.h" | |
54 | #include "tree-pass.h" | |
2abae5f1 | 55 | #include "sese.h" |
440917de | 56 | #include "tree-ssa-propagate.h" |
2abae5f1 SP |
57 | |
58 | #ifdef HAVE_cloog | |
2abae5f1 SP |
59 | #include "graphite-poly.h" |
60 | #include "graphite-scop-detection.h" | |
61 | ||
99e2796b BS |
62 | /* Forward declarations. */ |
63 | static void make_close_phi_nodes_unique (basic_block); | |
64 | ||
2abae5f1 SP |
65 | /* The type of the analyzed basic block. */ |
66 | ||
67 | typedef enum gbb_type { | |
68 | GBB_UNKNOWN, | |
69 | GBB_LOOP_SING_EXIT_HEADER, | |
70 | GBB_LOOP_MULT_EXIT_HEADER, | |
71 | GBB_LOOP_EXIT, | |
72 | GBB_COND_HEADER, | |
73 | GBB_SIMPLE, | |
74 | GBB_LAST | |
75 | } gbb_type; | |
76 | ||
77 | /* Detect the type of BB. Loop headers are only marked, if they are | |
78 | new. This means their loop_father is different to LAST_LOOP. | |
79 | Otherwise they are treated like any other bb and their type can be | |
80 | any other type. */ | |
81 | ||
82 | static gbb_type | |
83 | get_bb_type (basic_block bb, struct loop *last_loop) | |
84 | { | |
9771b263 | 85 | vec<basic_block> dom; |
d630245f | 86 | int nb_dom; |
2abae5f1 SP |
87 | struct loop *loop = bb->loop_father; |
88 | ||
89 | /* Check, if we entry into a new loop. */ | |
90 | if (loop != last_loop) | |
91 | { | |
92 | if (single_exit (loop) != NULL) | |
93 | return GBB_LOOP_SING_EXIT_HEADER; | |
94 | else if (loop->num != 0) | |
95 | return GBB_LOOP_MULT_EXIT_HEADER; | |
96 | else | |
97 | return GBB_COND_HEADER; | |
98 | } | |
99 | ||
100 | dom = get_dominated_by (CDI_DOMINATORS, bb); | |
9771b263 DN |
101 | nb_dom = dom.length (); |
102 | dom.release (); | |
2abae5f1 SP |
103 | |
104 | if (nb_dom == 0) | |
105 | return GBB_LAST; | |
106 | ||
d630245f | 107 | if (nb_dom == 1 && single_succ_p (bb)) |
2abae5f1 SP |
108 | return GBB_SIMPLE; |
109 | ||
110 | return GBB_COND_HEADER; | |
111 | } | |
112 | ||
113 | /* A SCoP detection region, defined using bbs as borders. | |
114 | ||
115 | All control flow touching this region, comes in passing basic_block | |
116 | ENTRY and leaves passing basic_block EXIT. By using bbs instead of | |
117 | edges for the borders we are able to represent also regions that do | |
118 | not have a single entry or exit edge. | |
119 | ||
120 | But as they have a single entry basic_block and a single exit | |
121 | basic_block, we are able to generate for every sd_region a single | |
122 | entry and exit edge. | |
123 | ||
124 | 1 2 | |
125 | \ / | |
126 | 3 <- entry | |
127 | | | |
128 | 4 | |
129 | / \ This region contains: {3, 4, 5, 6, 7, 8} | |
130 | 5 6 | |
131 | | | | |
132 | 7 8 | |
133 | \ / | |
134 | 9 <- exit */ | |
135 | ||
136 | ||
137 | typedef struct sd_region_p | |
138 | { | |
139 | /* The entry bb dominates all bbs in the sd_region. It is part of | |
140 | the region. */ | |
141 | basic_block entry; | |
142 | ||
143 | /* The exit bb postdominates all bbs in the sd_region, but is not | |
144 | part of the region. */ | |
145 | basic_block exit; | |
146 | } sd_region; | |
147 | ||
2abae5f1 SP |
148 | |
149 | ||
150 | /* Moves the scops from SOURCE to TARGET and clean up SOURCE. */ | |
151 | ||
152 | static void | |
9771b263 | 153 | move_sd_regions (vec<sd_region> *source, vec<sd_region> *target) |
2abae5f1 SP |
154 | { |
155 | sd_region *s; | |
156 | int i; | |
157 | ||
9771b263 DN |
158 | FOR_EACH_VEC_ELT (*source, i, s) |
159 | target->safe_push (*s); | |
2abae5f1 | 160 | |
9771b263 | 161 | source->release (); |
2abae5f1 SP |
162 | } |
163 | ||
164 | /* Something like "n * m" is not allowed. */ | |
165 | ||
166 | static bool | |
167 | graphite_can_represent_init (tree e) | |
168 | { | |
169 | switch (TREE_CODE (e)) | |
170 | { | |
171 | case POLYNOMIAL_CHREC: | |
172 | return graphite_can_represent_init (CHREC_LEFT (e)) | |
173 | && graphite_can_represent_init (CHREC_RIGHT (e)); | |
174 | ||
175 | case MULT_EXPR: | |
176 | if (chrec_contains_symbols (TREE_OPERAND (e, 0))) | |
d505015a | 177 | return graphite_can_represent_init (TREE_OPERAND (e, 0)) |
9541ffee | 178 | && tree_fits_shwi_p (TREE_OPERAND (e, 1)); |
2abae5f1 | 179 | else |
d505015a | 180 | return graphite_can_represent_init (TREE_OPERAND (e, 1)) |
9541ffee | 181 | && tree_fits_shwi_p (TREE_OPERAND (e, 0)); |
2abae5f1 SP |
182 | |
183 | case PLUS_EXPR: | |
184 | case POINTER_PLUS_EXPR: | |
185 | case MINUS_EXPR: | |
186 | return graphite_can_represent_init (TREE_OPERAND (e, 0)) | |
187 | && graphite_can_represent_init (TREE_OPERAND (e, 1)); | |
188 | ||
189 | case NEGATE_EXPR: | |
190 | case BIT_NOT_EXPR: | |
191 | CASE_CONVERT: | |
192 | case NON_LVALUE_EXPR: | |
193 | return graphite_can_represent_init (TREE_OPERAND (e, 0)); | |
194 | ||
195 | default: | |
196 | break; | |
197 | } | |
198 | ||
199 | return true; | |
200 | } | |
201 | ||
202 | /* Return true when SCEV can be represented in the polyhedral model. | |
203 | ||
204 | An expression can be represented, if it can be expressed as an | |
205 | affine expression. For loops (i, j) and parameters (m, n) all | |
206 | affine expressions are of the form: | |
207 | ||
208 | x1 * i + x2 * j + x3 * m + x4 * n + x5 * 1 where x1..x5 element of Z | |
209 | ||
210 | 1 i + 20 j + (-2) m + 25 | |
211 | ||
56f30f65 | 212 | Something like "i * n" or "n * m" is not allowed. */ |
2abae5f1 SP |
213 | |
214 | static bool | |
56f30f65 | 215 | graphite_can_represent_scev (tree scev) |
2abae5f1 SP |
216 | { |
217 | if (chrec_contains_undetermined (scev)) | |
218 | return false; | |
219 | ||
4b216ab0 SP |
220 | switch (TREE_CODE (scev)) |
221 | { | |
222 | case PLUS_EXPR: | |
223 | case MINUS_EXPR: | |
56f30f65 VK |
224 | return graphite_can_represent_scev (TREE_OPERAND (scev, 0)) |
225 | && graphite_can_represent_scev (TREE_OPERAND (scev, 1)); | |
2abae5f1 | 226 | |
4b216ab0 SP |
227 | case MULT_EXPR: |
228 | return !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev, 0))) | |
229 | && !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev, 1))) | |
230 | && !(chrec_contains_symbols (TREE_OPERAND (scev, 0)) | |
231 | && chrec_contains_symbols (TREE_OPERAND (scev, 1))) | |
c4c4983e | 232 | && graphite_can_represent_init (scev) |
56f30f65 VK |
233 | && graphite_can_represent_scev (TREE_OPERAND (scev, 0)) |
234 | && graphite_can_represent_scev (TREE_OPERAND (scev, 1)); | |
2abae5f1 | 235 | |
4b216ab0 SP |
236 | case POLYNOMIAL_CHREC: |
237 | /* Check for constant strides. With a non constant stride of | |
238 | 'n' we would have a value of 'iv * n'. Also check that the | |
239 | initial value can represented: for example 'n * m' cannot be | |
240 | represented. */ | |
241 | if (!evolution_function_right_is_integer_cst (scev) | |
242 | || !graphite_can_represent_init (scev)) | |
243 | return false; | |
244 | ||
245 | default: | |
246 | break; | |
247 | } | |
2abae5f1 SP |
248 | |
249 | /* Only affine functions can be represented. */ | |
250 | if (!scev_is_linear_expression (scev)) | |
251 | return false; | |
252 | ||
d9ae7906 | 253 | return true; |
2abae5f1 SP |
254 | } |
255 | ||
256 | ||
257 | /* Return true when EXPR can be represented in the polyhedral model. | |
258 | ||
259 | This means an expression can be represented, if it is linear with | |
260 | respect to the loops and the strides are non parametric. | |
56f30f65 | 261 | LOOP is the place where the expr will be evaluated. SCOP_ENTRY defines the |
2abae5f1 SP |
262 | entry of the region we analyse. */ |
263 | ||
264 | static bool | |
265 | graphite_can_represent_expr (basic_block scop_entry, loop_p loop, | |
56f30f65 | 266 | tree expr) |
2abae5f1 SP |
267 | { |
268 | tree scev = analyze_scalar_evolution (loop, expr); | |
269 | ||
270 | scev = instantiate_scev (scop_entry, loop, scev); | |
271 | ||
56f30f65 | 272 | return graphite_can_represent_scev (scev); |
2abae5f1 SP |
273 | } |
274 | ||
2abae5f1 SP |
275 | /* Return true if the data references of STMT can be represented by |
276 | Graphite. */ | |
277 | ||
278 | static bool | |
390b24dc RG |
279 | stmt_has_simple_data_refs_p (loop_p outermost_loop ATTRIBUTE_UNUSED, |
280 | gimple stmt) | |
2abae5f1 SP |
281 | { |
282 | data_reference_p dr; | |
283 | unsigned i; | |
284 | int j; | |
285 | bool res = true; | |
6e1aa848 | 286 | vec<data_reference_p> drs = vNULL; |
390b24dc RG |
287 | loop_p outer; |
288 | ||
289 | for (outer = loop_containing_stmt (stmt); outer; outer = loop_outer (outer)) | |
290 | { | |
291 | graphite_find_data_references_in_stmt (outer, | |
292 | loop_containing_stmt (stmt), | |
293 | stmt, &drs); | |
294 | ||
9771b263 | 295 | FOR_EACH_VEC_ELT (drs, j, dr) |
390b24dc RG |
296 | for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++) |
297 | if (!graphite_can_represent_scev (DR_ACCESS_FN (dr, i))) | |
298 | { | |
299 | res = false; | |
300 | goto done; | |
301 | } | |
302 | ||
303 | free_data_refs (drs); | |
9771b263 | 304 | drs.create (0); |
390b24dc | 305 | } |
2abae5f1 SP |
306 | |
307 | done: | |
308 | free_data_refs (drs); | |
309 | return res; | |
310 | } | |
311 | ||
2abae5f1 SP |
312 | /* Return true only when STMT is simple enough for being handled by |
313 | Graphite. This depends on SCOP_ENTRY, as the parameters are | |
314 | initialized relatively to this basic block, the linear functions | |
315 | are initialized to OUTERMOST_LOOP and BB is the place where we try | |
316 | to evaluate the STMT. */ | |
317 | ||
318 | static bool | |
319 | stmt_simple_for_scop_p (basic_block scop_entry, loop_p outermost_loop, | |
320 | gimple stmt, basic_block bb) | |
321 | { | |
322 | loop_p loop = bb->loop_father; | |
323 | ||
324 | gcc_assert (scop_entry); | |
325 | ||
326 | /* GIMPLE_ASM and GIMPLE_CALL may embed arbitrary side effects. | |
327 | Calls have side-effects, except those to const or pure | |
328 | functions. */ | |
329 | if (gimple_has_volatile_ops (stmt) | |
330 | || (gimple_code (stmt) == GIMPLE_CALL | |
331 | && !(gimple_call_flags (stmt) & (ECF_CONST | ECF_PURE))) | |
332 | || (gimple_code (stmt) == GIMPLE_ASM)) | |
333 | return false; | |
334 | ||
a3201927 AO |
335 | if (is_gimple_debug (stmt)) |
336 | return true; | |
337 | ||
2abae5f1 SP |
338 | if (!stmt_has_simple_data_refs_p (outermost_loop, stmt)) |
339 | return false; | |
340 | ||
341 | switch (gimple_code (stmt)) | |
342 | { | |
343 | case GIMPLE_RETURN: | |
344 | case GIMPLE_LABEL: | |
345 | return true; | |
346 | ||
347 | case GIMPLE_COND: | |
348 | { | |
349 | tree op; | |
350 | ssa_op_iter op_iter; | |
351 | enum tree_code code = gimple_cond_code (stmt); | |
352 | ||
353 | /* We can handle all binary comparisons. Inequalities are | |
354 | also supported as they can be represented with union of | |
355 | polyhedra. */ | |
356 | if (!(code == LT_EXPR | |
357 | || code == GT_EXPR | |
358 | || code == LE_EXPR | |
359 | || code == GE_EXPR | |
360 | || code == EQ_EXPR | |
361 | || code == NE_EXPR)) | |
362 | return false; | |
363 | ||
364 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, op_iter, SSA_OP_ALL_USES) | |
56f30f65 | 365 | if (!graphite_can_represent_expr (scop_entry, loop, op) |
2abae5f1 SP |
366 | /* We can not handle REAL_TYPE. Failed for pr39260. */ |
367 | || TREE_CODE (TREE_TYPE (op)) == REAL_TYPE) | |
368 | return false; | |
369 | ||
370 | return true; | |
371 | } | |
372 | ||
373 | case GIMPLE_ASSIGN: | |
2abae5f1 | 374 | case GIMPLE_CALL: |
c8ae0613 | 375 | return true; |
2abae5f1 SP |
376 | |
377 | default: | |
378 | /* These nodes cut a new scope. */ | |
379 | return false; | |
380 | } | |
381 | ||
382 | return false; | |
383 | } | |
384 | ||
385 | /* Returns the statement of BB that contains a harmful operation: that | |
386 | can be a function call with side effects, the induction variables | |
387 | are not linear with respect to SCOP_ENTRY, etc. The current open | |
388 | scop should end before this statement. The evaluation is limited using | |
389 | OUTERMOST_LOOP as outermost loop that may change. */ | |
390 | ||
391 | static gimple | |
392 | harmful_stmt_in_bb (basic_block scop_entry, loop_p outer_loop, basic_block bb) | |
393 | { | |
394 | gimple_stmt_iterator gsi; | |
395 | ||
396 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
397 | if (!stmt_simple_for_scop_p (scop_entry, outer_loop, gsi_stmt (gsi), bb)) | |
398 | return gsi_stmt (gsi); | |
399 | ||
400 | return NULL; | |
401 | } | |
402 | ||
45fc26fc VK |
403 | /* Return true if LOOP can be represented in the polyhedral |
404 | representation. This is evaluated taking SCOP_ENTRY and | |
405 | OUTERMOST_LOOP in mind. */ | |
2abae5f1 SP |
406 | |
407 | static bool | |
56f30f65 | 408 | graphite_can_represent_loop (basic_block scop_entry, loop_p loop) |
2abae5f1 | 409 | { |
cbc1994b SP |
410 | tree niter; |
411 | struct tree_niter_desc niter_desc; | |
2abae5f1 | 412 | |
cbc1994b SP |
413 | /* FIXME: For the moment, graphite cannot be used on loops that |
414 | iterate using induction variables that wrap. */ | |
2abae5f1 | 415 | |
cbc1994b SP |
416 | return number_of_iterations_exit (loop, single_exit (loop), &niter_desc, false) |
417 | && niter_desc.control.no_overflow | |
418 | && (niter = number_of_latch_executions (loop)) | |
419 | && !chrec_contains_undetermined (niter) | |
420 | && graphite_can_represent_expr (scop_entry, loop, niter); | |
2abae5f1 SP |
421 | } |
422 | ||
423 | /* Store information needed by scopdet_* functions. */ | |
424 | ||
425 | struct scopdet_info | |
426 | { | |
427 | /* Exit of the open scop would stop if the current BB is harmful. */ | |
428 | basic_block exit; | |
429 | ||
430 | /* Where the next scop would start if the current BB is harmful. */ | |
431 | basic_block next; | |
432 | ||
433 | /* The bb or one of its children contains open loop exits. That means | |
434 | loop exit nodes that are not surrounded by a loop dominated by bb. */ | |
435 | bool exits; | |
436 | ||
437 | /* The bb or one of its children contains only structures we can handle. */ | |
438 | bool difficult; | |
439 | }; | |
440 | ||
441 | static struct scopdet_info build_scops_1 (basic_block, loop_p, | |
9771b263 | 442 | vec<sd_region> *, loop_p); |
2abae5f1 SP |
443 | |
444 | /* Calculates BB infos. If bb is difficult we add valid SCoPs dominated by BB | |
445 | to SCOPS. TYPE is the gbb_type of BB. */ | |
446 | ||
447 | static struct scopdet_info | |
448 | scopdet_basic_block_info (basic_block bb, loop_p outermost_loop, | |
9771b263 | 449 | vec<sd_region> *scops, gbb_type type) |
2abae5f1 SP |
450 | { |
451 | loop_p loop = bb->loop_father; | |
452 | struct scopdet_info result; | |
453 | gimple stmt; | |
454 | ||
455 | /* XXX: ENTRY_BLOCK_PTR could be optimized in later steps. */ | |
fefa31b5 | 456 | basic_block entry_block = ENTRY_BLOCK_PTR_FOR_FN (cfun); |
2abae5f1 SP |
457 | stmt = harmful_stmt_in_bb (entry_block, outermost_loop, bb); |
458 | result.difficult = (stmt != NULL); | |
459 | result.exit = NULL; | |
460 | ||
461 | switch (type) | |
462 | { | |
463 | case GBB_LAST: | |
464 | result.next = NULL; | |
465 | result.exits = false; | |
466 | ||
467 | /* Mark bbs terminating a SESE region difficult, if they start | |
468 | a condition. */ | |
469 | if (!single_succ_p (bb)) | |
470 | result.difficult = true; | |
471 | else | |
472 | result.exit = single_succ (bb); | |
473 | ||
474 | break; | |
475 | ||
476 | case GBB_SIMPLE: | |
477 | result.next = single_succ (bb); | |
478 | result.exits = false; | |
479 | result.exit = single_succ (bb); | |
480 | break; | |
481 | ||
482 | case GBB_LOOP_SING_EXIT_HEADER: | |
483 | { | |
07687835 | 484 | stack_vec<sd_region, 3> regions; |
2abae5f1 SP |
485 | struct scopdet_info sinfo; |
486 | edge exit_e = single_exit (loop); | |
487 | ||
488 | sinfo = build_scops_1 (bb, outermost_loop, ®ions, loop); | |
489 | ||
56f30f65 | 490 | if (!graphite_can_represent_loop (entry_block, loop)) |
2abae5f1 SP |
491 | result.difficult = true; |
492 | ||
493 | result.difficult |= sinfo.difficult; | |
494 | ||
495 | /* Try again with another loop level. */ | |
496 | if (result.difficult | |
497 | && loop_depth (outermost_loop) + 1 == loop_depth (loop)) | |
498 | { | |
499 | outermost_loop = loop; | |
500 | ||
9771b263 DN |
501 | regions.release (); |
502 | regions.create (3); | |
2abae5f1 SP |
503 | |
504 | sinfo = scopdet_basic_block_info (bb, outermost_loop, scops, type); | |
505 | ||
506 | result = sinfo; | |
507 | result.difficult = true; | |
508 | ||
509 | if (sinfo.difficult) | |
510 | move_sd_regions (®ions, scops); | |
511 | else | |
512 | { | |
513 | sd_region open_scop; | |
514 | open_scop.entry = bb; | |
515 | open_scop.exit = exit_e->dest; | |
9771b263 DN |
516 | scops->safe_push (open_scop); |
517 | regions.release (); | |
2abae5f1 SP |
518 | } |
519 | } | |
520 | else | |
521 | { | |
522 | result.exit = exit_e->dest; | |
523 | result.next = exit_e->dest; | |
524 | ||
525 | /* If we do not dominate result.next, remove it. It's either | |
6626665f | 526 | the exit block, or another bb dominates it and will |
2abae5f1 SP |
527 | call the scop detection for this bb. */ |
528 | if (!dominated_by_p (CDI_DOMINATORS, result.next, bb)) | |
529 | result.next = NULL; | |
530 | ||
531 | if (exit_e->src->loop_father != loop) | |
532 | result.next = NULL; | |
533 | ||
534 | result.exits = false; | |
535 | ||
536 | if (result.difficult) | |
537 | move_sd_regions (®ions, scops); | |
538 | else | |
9771b263 | 539 | regions.release (); |
2abae5f1 SP |
540 | } |
541 | ||
542 | break; | |
543 | } | |
544 | ||
545 | case GBB_LOOP_MULT_EXIT_HEADER: | |
546 | { | |
547 | /* XXX: For now we just do not join loops with multiple exits. If the | |
548 | exits lead to the same bb it may be possible to join the loop. */ | |
07687835 | 549 | stack_vec<sd_region, 3> regions; |
9771b263 | 550 | vec<edge> exits = get_loop_exit_edges (loop); |
2abae5f1 SP |
551 | edge e; |
552 | int i; | |
553 | build_scops_1 (bb, loop, ®ions, loop); | |
554 | ||
555 | /* Scan the code dominated by this loop. This means all bbs, that are | |
556 | are dominated by a bb in this loop, but are not part of this loop. | |
557 | ||
558 | The easiest case: | |
559 | - The loop exit destination is dominated by the exit sources. | |
560 | ||
561 | TODO: We miss here the more complex cases: | |
562 | - The exit destinations are dominated by another bb inside | |
563 | the loop. | |
564 | - The loop dominates bbs, that are not exit destinations. */ | |
9771b263 | 565 | FOR_EACH_VEC_ELT (exits, i, e) |
2abae5f1 SP |
566 | if (e->src->loop_father == loop |
567 | && dominated_by_p (CDI_DOMINATORS, e->dest, e->src)) | |
568 | { | |
569 | if (loop_outer (outermost_loop)) | |
570 | outermost_loop = loop_outer (outermost_loop); | |
571 | ||
572 | /* Pass loop_outer to recognize e->dest as loop header in | |
573 | build_scops_1. */ | |
574 | if (e->dest->loop_father->header == e->dest) | |
575 | build_scops_1 (e->dest, outermost_loop, ®ions, | |
576 | loop_outer (e->dest->loop_father)); | |
577 | else | |
578 | build_scops_1 (e->dest, outermost_loop, ®ions, | |
579 | e->dest->loop_father); | |
580 | } | |
581 | ||
582 | result.next = NULL; | |
583 | result.exit = NULL; | |
584 | result.difficult = true; | |
585 | result.exits = false; | |
586 | move_sd_regions (®ions, scops); | |
9771b263 | 587 | exits.release (); |
2abae5f1 SP |
588 | break; |
589 | } | |
590 | case GBB_COND_HEADER: | |
591 | { | |
07687835 | 592 | stack_vec<sd_region, 3> regions; |
2abae5f1 | 593 | struct scopdet_info sinfo; |
9771b263 | 594 | vec<basic_block> dominated; |
2abae5f1 SP |
595 | int i; |
596 | basic_block dom_bb; | |
597 | basic_block last_exit = NULL; | |
598 | edge e; | |
599 | result.exits = false; | |
600 | ||
601 | /* First check the successors of BB, and check if it is | |
602 | possible to join the different branches. */ | |
9771b263 | 603 | FOR_EACH_VEC_SAFE_ELT (bb->succs, i, e) |
2abae5f1 SP |
604 | { |
605 | /* Ignore loop exits. They will be handled after the loop | |
606 | body. */ | |
f4ce375d | 607 | if (loop_exits_to_bb_p (loop, e->dest)) |
2abae5f1 SP |
608 | { |
609 | result.exits = true; | |
610 | continue; | |
611 | } | |
612 | ||
613 | /* Do not follow edges that lead to the end of the | |
614 | conditions block. For example, in | |
615 | ||
616 | | 0 | |
617 | | /|\ | |
618 | | 1 2 | | |
619 | | | | | | |
620 | | 3 4 | | |
621 | | \|/ | |
622 | | 6 | |
623 | ||
624 | the edge from 0 => 6. Only check if all paths lead to | |
625 | the same node 6. */ | |
626 | ||
627 | if (!single_pred_p (e->dest)) | |
628 | { | |
629 | /* Check, if edge leads directly to the end of this | |
630 | condition. */ | |
631 | if (!last_exit) | |
632 | last_exit = e->dest; | |
633 | ||
634 | if (e->dest != last_exit) | |
635 | result.difficult = true; | |
636 | ||
637 | continue; | |
638 | } | |
639 | ||
640 | if (!dominated_by_p (CDI_DOMINATORS, e->dest, bb)) | |
641 | { | |
642 | result.difficult = true; | |
643 | continue; | |
644 | } | |
645 | ||
646 | sinfo = build_scops_1 (e->dest, outermost_loop, ®ions, loop); | |
647 | ||
648 | result.exits |= sinfo.exits; | |
649 | result.difficult |= sinfo.difficult; | |
650 | ||
651 | /* Checks, if all branches end at the same point. | |
652 | If that is true, the condition stays joinable. | |
653 | Have a look at the example above. */ | |
654 | if (sinfo.exit) | |
655 | { | |
656 | if (!last_exit) | |
657 | last_exit = sinfo.exit; | |
658 | ||
659 | if (sinfo.exit != last_exit) | |
660 | result.difficult = true; | |
661 | } | |
662 | else | |
663 | result.difficult = true; | |
664 | } | |
665 | ||
666 | if (!last_exit) | |
667 | result.difficult = true; | |
668 | ||
669 | /* Join the branches of the condition if possible. */ | |
670 | if (!result.exits && !result.difficult) | |
671 | { | |
672 | /* Only return a next pointer if we dominate this pointer. | |
673 | Otherwise it will be handled by the bb dominating it. */ | |
674 | if (dominated_by_p (CDI_DOMINATORS, last_exit, bb) | |
675 | && last_exit != bb) | |
676 | result.next = last_exit; | |
677 | else | |
678 | result.next = NULL; | |
679 | ||
680 | result.exit = last_exit; | |
681 | ||
9771b263 | 682 | regions.release (); |
2abae5f1 SP |
683 | break; |
684 | } | |
685 | ||
686 | /* Scan remaining bbs dominated by BB. */ | |
687 | dominated = get_dominated_by (CDI_DOMINATORS, bb); | |
688 | ||
9771b263 | 689 | FOR_EACH_VEC_ELT (dominated, i, dom_bb) |
2abae5f1 SP |
690 | { |
691 | /* Ignore loop exits: they will be handled after the loop body. */ | |
692 | if (loop_depth (find_common_loop (loop, dom_bb->loop_father)) | |
693 | < loop_depth (loop)) | |
694 | { | |
695 | result.exits = true; | |
696 | continue; | |
697 | } | |
698 | ||
699 | /* Ignore the bbs processed above. */ | |
700 | if (single_pred_p (dom_bb) && single_pred (dom_bb) == bb) | |
701 | continue; | |
702 | ||
703 | if (loop_depth (loop) > loop_depth (dom_bb->loop_father)) | |
704 | sinfo = build_scops_1 (dom_bb, outermost_loop, ®ions, | |
705 | loop_outer (loop)); | |
706 | else | |
707 | sinfo = build_scops_1 (dom_bb, outermost_loop, ®ions, loop); | |
708 | ||
709 | result.exits |= sinfo.exits; | |
710 | result.difficult = true; | |
711 | result.exit = NULL; | |
712 | } | |
713 | ||
9771b263 | 714 | dominated.release (); |
2abae5f1 SP |
715 | |
716 | result.next = NULL; | |
717 | move_sd_regions (®ions, scops); | |
718 | ||
719 | break; | |
720 | } | |
721 | ||
722 | default: | |
723 | gcc_unreachable (); | |
724 | } | |
725 | ||
726 | return result; | |
727 | } | |
728 | ||
729 | /* Starting from CURRENT we walk the dominance tree and add new sd_regions to | |
730 | SCOPS. The analyse if a sd_region can be handled is based on the value | |
731 | of OUTERMOST_LOOP. Only loops inside OUTERMOST loops may change. LOOP | |
732 | is the loop in which CURRENT is handled. | |
733 | ||
734 | TODO: These functions got a little bit big. They definitely should be cleaned | |
735 | up. */ | |
736 | ||
737 | static struct scopdet_info | |
738 | build_scops_1 (basic_block current, loop_p outermost_loop, | |
9771b263 | 739 | vec<sd_region> *scops, loop_p loop) |
2abae5f1 SP |
740 | { |
741 | bool in_scop = false; | |
742 | sd_region open_scop; | |
743 | struct scopdet_info sinfo; | |
744 | ||
745 | /* Initialize result. */ | |
746 | struct scopdet_info result; | |
747 | result.exits = false; | |
748 | result.difficult = false; | |
749 | result.next = NULL; | |
750 | result.exit = NULL; | |
751 | open_scop.entry = NULL; | |
752 | open_scop.exit = NULL; | |
753 | sinfo.exit = NULL; | |
754 | ||
755 | /* Loop over the dominance tree. If we meet a difficult bb, close | |
756 | the current SCoP. Loop and condition header start a new layer, | |
757 | and can only be added if all bbs in deeper layers are simple. */ | |
758 | while (current != NULL) | |
759 | { | |
760 | sinfo = scopdet_basic_block_info (current, outermost_loop, scops, | |
761 | get_bb_type (current, loop)); | |
762 | ||
763 | if (!in_scop && !(sinfo.exits || sinfo.difficult)) | |
764 | { | |
765 | open_scop.entry = current; | |
766 | open_scop.exit = NULL; | |
767 | in_scop = true; | |
768 | } | |
769 | else if (in_scop && (sinfo.exits || sinfo.difficult)) | |
770 | { | |
771 | open_scop.exit = current; | |
9771b263 | 772 | scops->safe_push (open_scop); |
2abae5f1 SP |
773 | in_scop = false; |
774 | } | |
775 | ||
776 | result.difficult |= sinfo.difficult; | |
777 | result.exits |= sinfo.exits; | |
778 | ||
779 | current = sinfo.next; | |
780 | } | |
781 | ||
782 | /* Try to close open_scop, if we are still in an open SCoP. */ | |
783 | if (in_scop) | |
784 | { | |
785 | open_scop.exit = sinfo.exit; | |
786 | gcc_assert (open_scop.exit); | |
9771b263 | 787 | scops->safe_push (open_scop); |
2abae5f1 SP |
788 | } |
789 | ||
790 | result.exit = sinfo.exit; | |
791 | return result; | |
792 | } | |
793 | ||
794 | /* Checks if a bb is contained in REGION. */ | |
795 | ||
796 | static bool | |
797 | bb_in_sd_region (basic_block bb, sd_region *region) | |
798 | { | |
799 | return bb_in_region (bb, region->entry, region->exit); | |
800 | } | |
801 | ||
802 | /* Returns the single entry edge of REGION, if it does not exits NULL. */ | |
803 | ||
804 | static edge | |
805 | find_single_entry_edge (sd_region *region) | |
806 | { | |
807 | edge e; | |
808 | edge_iterator ei; | |
809 | edge entry = NULL; | |
810 | ||
811 | FOR_EACH_EDGE (e, ei, region->entry->preds) | |
812 | if (!bb_in_sd_region (e->src, region)) | |
813 | { | |
814 | if (entry) | |
815 | { | |
816 | entry = NULL; | |
817 | break; | |
818 | } | |
819 | ||
820 | else | |
821 | entry = e; | |
822 | } | |
823 | ||
824 | return entry; | |
825 | } | |
826 | ||
827 | /* Returns the single exit edge of REGION, if it does not exits NULL. */ | |
828 | ||
829 | static edge | |
830 | find_single_exit_edge (sd_region *region) | |
831 | { | |
832 | edge e; | |
833 | edge_iterator ei; | |
834 | edge exit = NULL; | |
835 | ||
836 | FOR_EACH_EDGE (e, ei, region->exit->preds) | |
837 | if (bb_in_sd_region (e->src, region)) | |
838 | { | |
839 | if (exit) | |
840 | { | |
841 | exit = NULL; | |
842 | break; | |
843 | } | |
844 | ||
845 | else | |
846 | exit = e; | |
847 | } | |
848 | ||
849 | return exit; | |
850 | } | |
851 | ||
852 | /* Create a single entry edge for REGION. */ | |
853 | ||
854 | static void | |
855 | create_single_entry_edge (sd_region *region) | |
856 | { | |
857 | if (find_single_entry_edge (region)) | |
858 | return; | |
859 | ||
860 | /* There are multiple predecessors for bb_3 | |
861 | ||
862 | | 1 2 | |
863 | | | / | |
864 | | |/ | |
865 | | 3 <- entry | |
866 | | |\ | |
867 | | | | | |
868 | | 4 ^ | |
869 | | | | | |
870 | | |/ | |
871 | | 5 | |
872 | ||
873 | There are two edges (1->3, 2->3), that point from outside into the region, | |
874 | and another one (5->3), a loop latch, lead to bb_3. | |
875 | ||
876 | We split bb_3. | |
877 | ||
878 | | 1 2 | |
879 | | | / | |
880 | | |/ | |
881 | |3.0 | |
882 | | |\ (3.0 -> 3.1) = single entry edge | |
883 | |3.1 | <- entry | |
884 | | | | | |
885 | | | | | |
886 | | 4 ^ | |
887 | | | | | |
888 | | |/ | |
889 | | 5 | |
890 | ||
891 | If the loop is part of the SCoP, we have to redirect the loop latches. | |
892 | ||
893 | | 1 2 | |
894 | | | / | |
895 | | |/ | |
896 | |3.0 | |
897 | | | (3.0 -> 3.1) = entry edge | |
898 | |3.1 <- entry | |
899 | | |\ | |
900 | | | | | |
901 | | 4 ^ | |
902 | | | | | |
903 | | |/ | |
904 | | 5 */ | |
905 | ||
906 | if (region->entry->loop_father->header != region->entry | |
907 | || dominated_by_p (CDI_DOMINATORS, | |
908 | loop_latch_edge (region->entry->loop_father)->src, | |
909 | region->exit)) | |
910 | { | |
911 | edge forwarder = split_block_after_labels (region->entry); | |
912 | region->entry = forwarder->dest; | |
913 | } | |
914 | else | |
915 | /* This case is never executed, as the loop headers seem always to have a | |
916 | single edge pointing from outside into the loop. */ | |
917 | gcc_unreachable (); | |
918 | ||
77a74ed7 | 919 | gcc_checking_assert (find_single_entry_edge (region)); |
2abae5f1 SP |
920 | } |
921 | ||
922 | /* Check if the sd_region, mentioned in EDGE, has no exit bb. */ | |
923 | ||
924 | static bool | |
925 | sd_region_without_exit (edge e) | |
926 | { | |
927 | sd_region *r = (sd_region *) e->aux; | |
928 | ||
929 | if (r) | |
930 | return r->exit == NULL; | |
931 | else | |
932 | return false; | |
933 | } | |
934 | ||
935 | /* Create a single exit edge for REGION. */ | |
936 | ||
937 | static void | |
938 | create_single_exit_edge (sd_region *region) | |
939 | { | |
940 | edge e; | |
941 | edge_iterator ei; | |
942 | edge forwarder = NULL; | |
943 | basic_block exit; | |
944 | ||
2abae5f1 SP |
945 | /* We create a forwarder bb (5) for all edges leaving this region |
946 | (3->5, 4->5). All other edges leading to the same bb, are moved | |
947 | to a new bb (6). If these edges where part of another region (2->5) | |
948 | we update the region->exit pointer, of this region. | |
949 | ||
950 | To identify which edge belongs to which region we depend on the e->aux | |
951 | pointer in every edge. It points to the region of the edge or to NULL, | |
952 | if the edge is not part of any region. | |
953 | ||
954 | 1 2 3 4 1->5 no region, 2->5 region->exit = 5, | |
955 | \| |/ 3->5 region->exit = NULL, 4->5 region->exit = NULL | |
956 | 5 <- exit | |
957 | ||
958 | changes to | |
959 | ||
960 | 1 2 3 4 1->6 no region, 2->6 region->exit = 6, | |
961 | | | \/ 3->5 no region, 4->5 no region, | |
962 | | | 5 | |
963 | \| / 5->6 region->exit = 6 | |
964 | 6 | |
965 | ||
966 | Now there is only a single exit edge (5->6). */ | |
967 | exit = region->exit; | |
968 | region->exit = NULL; | |
969 | forwarder = make_forwarder_block (exit, &sd_region_without_exit, NULL); | |
970 | ||
971 | /* Unmark the edges, that are no longer exit edges. */ | |
972 | FOR_EACH_EDGE (e, ei, forwarder->src->preds) | |
973 | if (e->aux) | |
974 | e->aux = NULL; | |
975 | ||
976 | /* Mark the new exit edge. */ | |
977 | single_succ_edge (forwarder->src)->aux = region; | |
978 | ||
979 | /* Update the exit bb of all regions, where exit edges lead to | |
980 | forwarder->dest. */ | |
981 | FOR_EACH_EDGE (e, ei, forwarder->dest->preds) | |
982 | if (e->aux) | |
983 | ((sd_region *) e->aux)->exit = forwarder->dest; | |
984 | ||
77a74ed7 | 985 | gcc_checking_assert (find_single_exit_edge (region)); |
2abae5f1 SP |
986 | } |
987 | ||
988 | /* Unmark the exit edges of all REGIONS. | |
989 | See comment in "create_single_exit_edge". */ | |
990 | ||
991 | static void | |
9771b263 | 992 | unmark_exit_edges (vec<sd_region> regions) |
2abae5f1 SP |
993 | { |
994 | int i; | |
995 | sd_region *s; | |
996 | edge e; | |
997 | edge_iterator ei; | |
998 | ||
9771b263 | 999 | FOR_EACH_VEC_ELT (regions, i, s) |
2abae5f1 SP |
1000 | FOR_EACH_EDGE (e, ei, s->exit->preds) |
1001 | e->aux = NULL; | |
1002 | } | |
1003 | ||
1004 | ||
1005 | /* Mark the exit edges of all REGIONS. | |
1006 | See comment in "create_single_exit_edge". */ | |
1007 | ||
1008 | static void | |
9771b263 | 1009 | mark_exit_edges (vec<sd_region> regions) |
2abae5f1 SP |
1010 | { |
1011 | int i; | |
1012 | sd_region *s; | |
1013 | edge e; | |
1014 | edge_iterator ei; | |
1015 | ||
9771b263 | 1016 | FOR_EACH_VEC_ELT (regions, i, s) |
2abae5f1 SP |
1017 | FOR_EACH_EDGE (e, ei, s->exit->preds) |
1018 | if (bb_in_sd_region (e->src, s)) | |
1019 | e->aux = s; | |
1020 | } | |
1021 | ||
1022 | /* Create for all scop regions a single entry and a single exit edge. */ | |
1023 | ||
1024 | static void | |
9771b263 | 1025 | create_sese_edges (vec<sd_region> regions) |
2abae5f1 SP |
1026 | { |
1027 | int i; | |
1028 | sd_region *s; | |
1029 | ||
9771b263 | 1030 | FOR_EACH_VEC_ELT (regions, i, s) |
2abae5f1 SP |
1031 | create_single_entry_edge (s); |
1032 | ||
1033 | mark_exit_edges (regions); | |
1034 | ||
9771b263 | 1035 | FOR_EACH_VEC_ELT (regions, i, s) |
4caa8e21 SP |
1036 | /* Don't handle multiple edges exiting the function. */ |
1037 | if (!find_single_exit_edge (s) | |
fefa31b5 | 1038 | && s->exit != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
4caa8e21 | 1039 | create_single_exit_edge (s); |
2abae5f1 SP |
1040 | |
1041 | unmark_exit_edges (regions); | |
1042 | ||
cc360b36 | 1043 | calculate_dominance_info (CDI_DOMINATORS); |
2abae5f1 SP |
1044 | fix_loop_structure (NULL); |
1045 | ||
1046 | #ifdef ENABLE_CHECKING | |
1047 | verify_loop_structure (); | |
2abae5f1 SP |
1048 | verify_ssa (false); |
1049 | #endif | |
1050 | } | |
1051 | ||
1052 | /* Create graphite SCoPs from an array of scop detection REGIONS. */ | |
1053 | ||
1054 | static void | |
9771b263 DN |
1055 | build_graphite_scops (vec<sd_region> regions, |
1056 | vec<scop_p> *scops) | |
2abae5f1 SP |
1057 | { |
1058 | int i; | |
1059 | sd_region *s; | |
1060 | ||
9771b263 | 1061 | FOR_EACH_VEC_ELT (regions, i, s) |
2abae5f1 SP |
1062 | { |
1063 | edge entry = find_single_entry_edge (s); | |
1064 | edge exit = find_single_exit_edge (s); | |
4caa8e21 SP |
1065 | scop_p scop; |
1066 | ||
1067 | if (!exit) | |
1068 | continue; | |
1069 | ||
1070 | scop = new_scop (new_sese (entry, exit)); | |
9771b263 | 1071 | scops->safe_push (scop); |
2abae5f1 SP |
1072 | |
1073 | /* Are there overlapping SCoPs? */ | |
1074 | #ifdef ENABLE_CHECKING | |
1075 | { | |
1076 | int j; | |
1077 | sd_region *s2; | |
1078 | ||
9771b263 | 1079 | FOR_EACH_VEC_ELT (regions, j, s2) |
2abae5f1 SP |
1080 | if (s != s2) |
1081 | gcc_assert (!bb_in_sd_region (s->entry, s2)); | |
1082 | } | |
1083 | #endif | |
1084 | } | |
1085 | } | |
1086 | ||
1087 | /* Returns true when BB contains only close phi nodes. */ | |
1088 | ||
1089 | static bool | |
1090 | contains_only_close_phi_nodes (basic_block bb) | |
1091 | { | |
1092 | gimple_stmt_iterator gsi; | |
1093 | ||
1094 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1095 | if (gimple_code (gsi_stmt (gsi)) != GIMPLE_LABEL) | |
1096 | return false; | |
1097 | ||
1098 | return true; | |
1099 | } | |
1100 | ||
1101 | /* Print statistics for SCOP to FILE. */ | |
1102 | ||
1103 | static void | |
1104 | print_graphite_scop_statistics (FILE* file, scop_p scop) | |
1105 | { | |
1106 | long n_bbs = 0; | |
1107 | long n_loops = 0; | |
1108 | long n_stmts = 0; | |
1109 | long n_conditions = 0; | |
1110 | long n_p_bbs = 0; | |
1111 | long n_p_loops = 0; | |
1112 | long n_p_stmts = 0; | |
1113 | long n_p_conditions = 0; | |
1114 | ||
1115 | basic_block bb; | |
1116 | ||
1117 | FOR_ALL_BB (bb) | |
1118 | { | |
1119 | gimple_stmt_iterator psi; | |
1120 | loop_p loop = bb->loop_father; | |
1121 | ||
1122 | if (!bb_in_sese_p (bb, SCOP_REGION (scop))) | |
1123 | continue; | |
1124 | ||
1125 | n_bbs++; | |
1126 | n_p_bbs += bb->count; | |
1127 | ||
d630245f | 1128 | if (EDGE_COUNT (bb->succs) > 1) |
2abae5f1 SP |
1129 | { |
1130 | n_conditions++; | |
1131 | n_p_conditions += bb->count; | |
1132 | } | |
1133 | ||
1134 | for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi)) | |
1135 | { | |
1136 | n_stmts++; | |
1137 | n_p_stmts += bb->count; | |
1138 | } | |
1139 | ||
1140 | if (loop->header == bb && loop_in_sese_p (loop, SCOP_REGION (scop))) | |
1141 | { | |
1142 | n_loops++; | |
1143 | n_p_loops += bb->count; | |
1144 | } | |
1145 | ||
1146 | } | |
1147 | ||
1148 | fprintf (file, "\nBefore limit_scops SCoP statistics ("); | |
1149 | fprintf (file, "BBS:%ld, ", n_bbs); | |
1150 | fprintf (file, "LOOPS:%ld, ", n_loops); | |
1151 | fprintf (file, "CONDITIONS:%ld, ", n_conditions); | |
1152 | fprintf (file, "STMTS:%ld)\n", n_stmts); | |
1153 | fprintf (file, "\nBefore limit_scops SCoP profiling statistics ("); | |
1154 | fprintf (file, "BBS:%ld, ", n_p_bbs); | |
1155 | fprintf (file, "LOOPS:%ld, ", n_p_loops); | |
1156 | fprintf (file, "CONDITIONS:%ld, ", n_p_conditions); | |
1157 | fprintf (file, "STMTS:%ld)\n", n_p_stmts); | |
1158 | } | |
1159 | ||
1160 | /* Print statistics for SCOPS to FILE. */ | |
1161 | ||
1162 | static void | |
9771b263 | 1163 | print_graphite_statistics (FILE* file, vec<scop_p> scops) |
2abae5f1 SP |
1164 | { |
1165 | int i; | |
1166 | scop_p scop; | |
1167 | ||
9771b263 | 1168 | FOR_EACH_VEC_ELT (scops, i, scop) |
2abae5f1 SP |
1169 | print_graphite_scop_statistics (file, scop); |
1170 | } | |
1171 | ||
2abae5f1 SP |
1172 | /* We limit all SCoPs to SCoPs, that are completely surrounded by a loop. |
1173 | ||
1174 | Example: | |
1175 | ||
1176 | for (i | | |
1177 | { | | |
1178 | for (j | SCoP 1 | |
1179 | for (k | | |
1180 | } | | |
1181 | ||
1182 | * SCoP frontier, as this line is not surrounded by any loop. * | |
1183 | ||
1184 | for (l | SCoP 2 | |
1185 | ||
1186 | This is necessary as scalar evolution and parameter detection need a | |
1187 | outermost loop to initialize parameters correctly. | |
1188 | ||
1189 | TODO: FIX scalar evolution and parameter detection to allow more flexible | |
1190 | SCoP frontiers. */ | |
1191 | ||
1192 | static void | |
9771b263 | 1193 | limit_scops (vec<scop_p> *scops) |
2abae5f1 | 1194 | { |
07687835 | 1195 | stack_vec<sd_region, 3> regions; |
2abae5f1 SP |
1196 | |
1197 | int i; | |
1198 | scop_p scop; | |
1199 | ||
9771b263 | 1200 | FOR_EACH_VEC_ELT (*scops, i, scop) |
2abae5f1 SP |
1201 | { |
1202 | int j; | |
1203 | loop_p loop; | |
1204 | sese region = SCOP_REGION (scop); | |
2abae5f1 SP |
1205 | build_sese_loop_nests (region); |
1206 | ||
9771b263 | 1207 | FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), j, loop) |
2abae5f1 SP |
1208 | if (!loop_in_sese_p (loop_outer (loop), region) |
1209 | && single_exit (loop)) | |
1210 | { | |
1211 | sd_region open_scop; | |
1212 | open_scop.entry = loop->header; | |
1213 | open_scop.exit = single_exit (loop)->dest; | |
1214 | ||
1215 | /* This is a hack on top of the limit_scops hack. The | |
1216 | limit_scops hack should disappear all together. */ | |
1217 | if (single_succ_p (open_scop.exit) | |
1218 | && contains_only_close_phi_nodes (open_scop.exit)) | |
1219 | open_scop.exit = single_succ_edge (open_scop.exit)->dest; | |
1220 | ||
9771b263 | 1221 | regions.safe_push (open_scop); |
2abae5f1 SP |
1222 | } |
1223 | } | |
1224 | ||
7a521ff2 | 1225 | free_scops (*scops); |
9771b263 | 1226 | scops->create (3); |
2abae5f1 SP |
1227 | |
1228 | create_sese_edges (regions); | |
1229 | build_graphite_scops (regions, scops); | |
2abae5f1 SP |
1230 | } |
1231 | ||
3a292d59 SP |
1232 | /* Returns true when P1 and P2 are close phis with the same |
1233 | argument. */ | |
1234 | ||
1235 | static inline bool | |
1236 | same_close_phi_node (gimple p1, gimple p2) | |
1237 | { | |
1238 | return operand_equal_p (gimple_phi_arg_def (p1, 0), | |
1239 | gimple_phi_arg_def (p2, 0), 0); | |
1240 | } | |
1241 | ||
1242 | /* Remove the close phi node at GSI and replace its rhs with the rhs | |
1243 | of PHI. */ | |
1244 | ||
1245 | static void | |
1246 | remove_duplicate_close_phi (gimple phi, gimple_stmt_iterator *gsi) | |
1247 | { | |
1248 | gimple use_stmt; | |
1249 | use_operand_p use_p; | |
1250 | imm_use_iterator imm_iter; | |
1251 | tree res = gimple_phi_result (phi); | |
1252 | tree def = gimple_phi_result (gsi_stmt (*gsi)); | |
1253 | ||
1254 | gcc_assert (same_close_phi_node (phi, gsi_stmt (*gsi))); | |
1255 | ||
1256 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def) | |
1257 | { | |
1258 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
1259 | SET_USE (use_p, res); | |
1260 | ||
1261 | update_stmt (use_stmt); | |
99e2796b BS |
1262 | |
1263 | /* It is possible that we just created a duplicate close-phi | |
1264 | for an already-processed containing loop. Check for this | |
1265 | case and clean it up. */ | |
1266 | if (gimple_code (use_stmt) == GIMPLE_PHI | |
1267 | && gimple_phi_num_args (use_stmt) == 1) | |
1268 | make_close_phi_nodes_unique (gimple_bb (use_stmt)); | |
3a292d59 SP |
1269 | } |
1270 | ||
1271 | remove_phi_node (gsi, true); | |
1272 | } | |
1273 | ||
1274 | /* Removes all the close phi duplicates from BB. */ | |
1275 | ||
1276 | static void | |
1277 | make_close_phi_nodes_unique (basic_block bb) | |
1278 | { | |
1279 | gimple_stmt_iterator psi; | |
1280 | ||
1281 | for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi)) | |
1282 | { | |
1283 | gimple_stmt_iterator gsi = psi; | |
1284 | gimple phi = gsi_stmt (psi); | |
1285 | ||
1286 | /* At this point, PHI should be a close phi in normal form. */ | |
1287 | gcc_assert (gimple_phi_num_args (phi) == 1); | |
1288 | ||
1289 | /* Iterate over the next phis and remove duplicates. */ | |
1290 | gsi_next (&gsi); | |
1291 | while (!gsi_end_p (gsi)) | |
1292 | if (same_close_phi_node (phi, gsi_stmt (gsi))) | |
1293 | remove_duplicate_close_phi (phi, &gsi); | |
1294 | else | |
1295 | gsi_next (&gsi); | |
1296 | } | |
1297 | } | |
1298 | ||
2abae5f1 SP |
1299 | /* Transforms LOOP to the canonical loop closed SSA form. */ |
1300 | ||
1301 | static void | |
1302 | canonicalize_loop_closed_ssa (loop_p loop) | |
1303 | { | |
1304 | edge e = single_exit (loop); | |
1305 | basic_block bb; | |
1306 | ||
1307 | if (!e || e->flags & EDGE_ABNORMAL) | |
1308 | return; | |
1309 | ||
1310 | bb = e->dest; | |
1311 | ||
d630245f | 1312 | if (single_pred_p (bb)) |
3a292d59 SP |
1313 | { |
1314 | e = split_block_after_labels (bb); | |
1315 | make_close_phi_nodes_unique (e->src); | |
1316 | } | |
2abae5f1 SP |
1317 | else |
1318 | { | |
1319 | gimple_stmt_iterator psi; | |
1320 | basic_block close = split_edge (e); | |
1321 | ||
1322 | e = single_succ_edge (close); | |
1323 | ||
1324 | for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi)) | |
1325 | { | |
1326 | gimple phi = gsi_stmt (psi); | |
1327 | unsigned i; | |
1328 | ||
1329 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
1330 | if (gimple_phi_arg_edge (phi, i) == e) | |
1331 | { | |
1332 | tree res, arg = gimple_phi_arg_def (phi, i); | |
1333 | use_operand_p use_p; | |
1334 | gimple close_phi; | |
1335 | ||
1336 | if (TREE_CODE (arg) != SSA_NAME) | |
1337 | continue; | |
1338 | ||
dcc748dd RG |
1339 | close_phi = create_phi_node (NULL_TREE, close); |
1340 | res = create_new_def_for (arg, close_phi, | |
2abae5f1 SP |
1341 | gimple_phi_result_ptr (close_phi)); |
1342 | add_phi_arg (close_phi, arg, | |
1343 | gimple_phi_arg_edge (close_phi, 0), | |
9e227d60 | 1344 | UNKNOWN_LOCATION); |
2abae5f1 SP |
1345 | use_p = gimple_phi_arg_imm_use_ptr (phi, i); |
1346 | replace_exp (use_p, res); | |
1347 | update_stmt (phi); | |
1348 | } | |
1349 | } | |
3a292d59 SP |
1350 | |
1351 | make_close_phi_nodes_unique (close); | |
2abae5f1 | 1352 | } |
1c3ba85b SP |
1353 | |
1354 | /* The code above does not properly handle changes in the post dominance | |
1355 | information (yet). */ | |
1356 | free_dominance_info (CDI_POST_DOMINATORS); | |
2abae5f1 SP |
1357 | } |
1358 | ||
1359 | /* Converts the current loop closed SSA form to a canonical form | |
1360 | expected by the Graphite code generation. | |
1361 | ||
1362 | The loop closed SSA form has the following invariant: a variable | |
1363 | defined in a loop that is used outside the loop appears only in the | |
1364 | phi nodes in the destination of the loop exit. These phi nodes are | |
1365 | called close phi nodes. | |
1366 | ||
1367 | The canonical loop closed SSA form contains the extra invariants: | |
1368 | ||
1369 | - when the loop contains only one exit, the close phi nodes contain | |
1370 | only one argument. That implies that the basic block that contains | |
1371 | the close phi nodes has only one predecessor, that is a basic block | |
1372 | in the loop. | |
1373 | ||
1374 | - the basic block containing the close phi nodes does not contain | |
1375 | other statements. | |
3a292d59 SP |
1376 | |
1377 | - there exist only one phi node per definition in the loop. | |
2abae5f1 SP |
1378 | */ |
1379 | ||
1380 | static void | |
1381 | canonicalize_loop_closed_ssa_form (void) | |
1382 | { | |
2abae5f1 SP |
1383 | loop_p loop; |
1384 | ||
1385 | #ifdef ENABLE_CHECKING | |
a3b9e73c | 1386 | verify_loop_closed_ssa (true); |
2abae5f1 SP |
1387 | #endif |
1388 | ||
f0bd40b1 | 1389 | FOR_EACH_LOOP (loop, 0) |
2abae5f1 SP |
1390 | canonicalize_loop_closed_ssa (loop); |
1391 | ||
1392 | rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); | |
1393 | update_ssa (TODO_update_ssa); | |
1394 | ||
1395 | #ifdef ENABLE_CHECKING | |
a3b9e73c | 1396 | verify_loop_closed_ssa (true); |
2abae5f1 SP |
1397 | #endif |
1398 | } | |
1399 | ||
1400 | /* Find Static Control Parts (SCoP) in the current function and pushes | |
1401 | them to SCOPS. */ | |
1402 | ||
1403 | void | |
9771b263 | 1404 | build_scops (vec<scop_p> *scops) |
2abae5f1 SP |
1405 | { |
1406 | struct loop *loop = current_loops->tree_root; | |
07687835 | 1407 | stack_vec<sd_region, 3> regions; |
2abae5f1 SP |
1408 | |
1409 | canonicalize_loop_closed_ssa_form (); | |
fefa31b5 DM |
1410 | build_scops_1 (single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)), |
1411 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->loop_father, | |
4caa8e21 | 1412 | ®ions, loop); |
2abae5f1 SP |
1413 | create_sese_edges (regions); |
1414 | build_graphite_scops (regions, scops); | |
1415 | ||
1416 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1417 | print_graphite_statistics (dump_file, *scops); | |
1418 | ||
1419 | limit_scops (scops); | |
9771b263 | 1420 | regions.release (); |
2abae5f1 SP |
1421 | |
1422 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1423 | fprintf (dump_file, "\nnumber of SCoPs: %d\n", | |
9771b263 | 1424 | scops ? scops->length () : 0); |
2abae5f1 SP |
1425 | } |
1426 | ||
afae0207 SP |
1427 | /* Pretty print to FILE all the SCoPs in DOT format and mark them with |
1428 | different colors. If there are not enough colors, paint the | |
1429 | remaining SCoPs in gray. | |
1430 | ||
2abae5f1 | 1431 | Special nodes: |
afae0207 SP |
1432 | - "*" after the node number denotes the entry of a SCoP, |
1433 | - "#" after the node number denotes the exit of a SCoP, | |
1434 | - "()" around the node number denotes the entry or the | |
1435 | exit nodes of the SCOP. These are not part of SCoP. */ | |
2abae5f1 SP |
1436 | |
1437 | static void | |
9771b263 | 1438 | dot_all_scops_1 (FILE *file, vec<scop_p> scops) |
2abae5f1 SP |
1439 | { |
1440 | basic_block bb; | |
1441 | edge e; | |
1442 | edge_iterator ei; | |
1443 | scop_p scop; | |
1444 | const char* color; | |
1445 | int i; | |
1446 | ||
1447 | /* Disable debugging while printing graph. */ | |
1448 | int tmp_dump_flags = dump_flags; | |
1449 | dump_flags = 0; | |
1450 | ||
1451 | fprintf (file, "digraph all {\n"); | |
1452 | ||
1453 | FOR_ALL_BB (bb) | |
1454 | { | |
1455 | int part_of_scop = false; | |
1456 | ||
1457 | /* Use HTML for every bb label. So we are able to print bbs | |
1458 | which are part of two different SCoPs, with two different | |
1459 | background colors. */ | |
1460 | fprintf (file, "%d [label=<\n <TABLE BORDER=\"0\" CELLBORDER=\"1\" ", | |
1461 | bb->index); | |
1462 | fprintf (file, "CELLSPACING=\"0\">\n"); | |
1463 | ||
1464 | /* Select color for SCoP. */ | |
9771b263 | 1465 | FOR_EACH_VEC_ELT (scops, i, scop) |
2abae5f1 SP |
1466 | { |
1467 | sese region = SCOP_REGION (scop); | |
1468 | if (bb_in_sese_p (bb, region) | |
1469 | || (SESE_EXIT_BB (region) == bb) | |
1470 | || (SESE_ENTRY_BB (region) == bb)) | |
1471 | { | |
1472 | switch (i % 17) | |
1473 | { | |
1474 | case 0: /* red */ | |
1475 | color = "#e41a1c"; | |
1476 | break; | |
1477 | case 1: /* blue */ | |
1478 | color = "#377eb8"; | |
1479 | break; | |
1480 | case 2: /* green */ | |
1481 | color = "#4daf4a"; | |
1482 | break; | |
1483 | case 3: /* purple */ | |
1484 | color = "#984ea3"; | |
1485 | break; | |
1486 | case 4: /* orange */ | |
1487 | color = "#ff7f00"; | |
1488 | break; | |
1489 | case 5: /* yellow */ | |
1490 | color = "#ffff33"; | |
1491 | break; | |
1492 | case 6: /* brown */ | |
1493 | color = "#a65628"; | |
1494 | break; | |
1495 | case 7: /* rose */ | |
1496 | color = "#f781bf"; | |
1497 | break; | |
1498 | case 8: | |
1499 | color = "#8dd3c7"; | |
1500 | break; | |
1501 | case 9: | |
1502 | color = "#ffffb3"; | |
1503 | break; | |
1504 | case 10: | |
1505 | color = "#bebada"; | |
1506 | break; | |
1507 | case 11: | |
1508 | color = "#fb8072"; | |
1509 | break; | |
1510 | case 12: | |
1511 | color = "#80b1d3"; | |
1512 | break; | |
1513 | case 13: | |
1514 | color = "#fdb462"; | |
1515 | break; | |
1516 | case 14: | |
1517 | color = "#b3de69"; | |
1518 | break; | |
1519 | case 15: | |
1520 | color = "#fccde5"; | |
1521 | break; | |
1522 | case 16: | |
1523 | color = "#bc80bd"; | |
1524 | break; | |
1525 | default: /* gray */ | |
1526 | color = "#999999"; | |
1527 | } | |
1528 | ||
1529 | fprintf (file, " <TR><TD WIDTH=\"50\" BGCOLOR=\"%s\">", color); | |
1530 | ||
1531 | if (!bb_in_sese_p (bb, region)) | |
1532 | fprintf (file, " ("); | |
1533 | ||
1534 | if (bb == SESE_ENTRY_BB (region) | |
1535 | && bb == SESE_EXIT_BB (region)) | |
1536 | fprintf (file, " %d*# ", bb->index); | |
1537 | else if (bb == SESE_ENTRY_BB (region)) | |
1538 | fprintf (file, " %d* ", bb->index); | |
1539 | else if (bb == SESE_EXIT_BB (region)) | |
1540 | fprintf (file, " %d# ", bb->index); | |
1541 | else | |
1542 | fprintf (file, " %d ", bb->index); | |
1543 | ||
1544 | if (!bb_in_sese_p (bb,region)) | |
1545 | fprintf (file, ")"); | |
1546 | ||
1547 | fprintf (file, "</TD></TR>\n"); | |
1548 | part_of_scop = true; | |
1549 | } | |
1550 | } | |
1551 | ||
1552 | if (!part_of_scop) | |
1553 | { | |
1554 | fprintf (file, " <TR><TD WIDTH=\"50\" BGCOLOR=\"#ffffff\">"); | |
1555 | fprintf (file, " %d </TD></TR>\n", bb->index); | |
1556 | } | |
1557 | fprintf (file, " </TABLE>>, shape=box, style=\"setlinewidth(0)\"]\n"); | |
1558 | } | |
1559 | ||
1560 | FOR_ALL_BB (bb) | |
1561 | { | |
1562 | FOR_EACH_EDGE (e, ei, bb->succs) | |
1563 | fprintf (file, "%d -> %d;\n", bb->index, e->dest->index); | |
1564 | } | |
1565 | ||
1566 | fputs ("}\n\n", file); | |
1567 | ||
1568 | /* Enable debugging again. */ | |
1569 | dump_flags = tmp_dump_flags; | |
1570 | } | |
1571 | ||
1572 | /* Display all SCoPs using dotty. */ | |
1573 | ||
1f424006 | 1574 | DEBUG_FUNCTION void |
9771b263 | 1575 | dot_all_scops (vec<scop_p> scops) |
2abae5f1 SP |
1576 | { |
1577 | /* When debugging, enable the following code. This cannot be used | |
1578 | in production compilers because it calls "system". */ | |
1579 | #if 0 | |
1580 | int x; | |
1581 | FILE *stream = fopen ("/tmp/allscops.dot", "w"); | |
1582 | gcc_assert (stream); | |
1583 | ||
1584 | dot_all_scops_1 (stream, scops); | |
1585 | fclose (stream); | |
1586 | ||
4c8f3c48 | 1587 | x = system ("dotty /tmp/allscops.dot &"); |
2abae5f1 SP |
1588 | #else |
1589 | dot_all_scops_1 (stderr, scops); | |
1590 | #endif | |
1591 | } | |
1592 | ||
1593 | /* Display all SCoPs using dotty. */ | |
1594 | ||
1f424006 | 1595 | DEBUG_FUNCTION void |
2abae5f1 SP |
1596 | dot_scop (scop_p scop) |
1597 | { | |
07687835 | 1598 | stack_vec<scop_p, 1> scops; |
2abae5f1 SP |
1599 | |
1600 | if (scop) | |
9771b263 | 1601 | scops.safe_push (scop); |
2abae5f1 SP |
1602 | |
1603 | /* When debugging, enable the following code. This cannot be used | |
1604 | in production compilers because it calls "system". */ | |
1605 | #if 0 | |
1606 | { | |
1607 | int x; | |
1608 | FILE *stream = fopen ("/tmp/allscops.dot", "w"); | |
1609 | gcc_assert (stream); | |
1610 | ||
1611 | dot_all_scops_1 (stream, scops); | |
1612 | fclose (stream); | |
4c8f3c48 | 1613 | x = system ("dotty /tmp/allscops.dot &"); |
2abae5f1 SP |
1614 | } |
1615 | #else | |
1616 | dot_all_scops_1 (stderr, scops); | |
1617 | #endif | |
2abae5f1 SP |
1618 | } |
1619 | ||
1620 | #endif |