]>
Commit | Line | Data |
---|---|---|
c913f08a | 1 | /* High-level loop manipulation functions. |
6ac01510 | 2 | Copyright (C) 2004, 2005, 2006, 2007 Free Software Foundation, Inc. |
c913f08a ZD |
3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it | |
7 | under the terms of the GNU General Public License as published by the | |
9dcd6f09 | 8 | Free Software Foundation; either version 3, or (at your option) any |
c913f08a ZD |
9 | later version. |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT | |
12 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
c913f08a ZD |
19 | |
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
23 | #include "tm.h" | |
24 | #include "tree.h" | |
25 | #include "rtl.h" | |
26 | #include "tm_p.h" | |
27 | #include "hard-reg-set.h" | |
28 | #include "basic-block.h" | |
29 | #include "output.h" | |
30 | #include "diagnostic.h" | |
31 | #include "tree-flow.h" | |
32 | #include "tree-dump.h" | |
33 | #include "timevar.h" | |
34 | #include "cfgloop.h" | |
35 | #include "tree-pass.h" | |
36 | #include "cfglayout.h" | |
37 | #include "tree-scalar-evolution.h" | |
17684618 | 38 | #include "params.h" |
7f9bc51b | 39 | #include "tree-inline.h" |
c913f08a | 40 | |
82b85a85 ZD |
41 | /* Creates an induction variable with value BASE + STEP * iteration in LOOP. |
42 | It is expected that neither BASE nor STEP are shared with other expressions | |
43 | (unless the sharing rules allow this). Use VAR as a base var_decl for it | |
44 | (if NULL, a new temporary will be created). The increment will occur at | |
92d2b330 SP |
45 | INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and |
46 | AFTER can be computed using standard_iv_increment_position. The ssa versions | |
82b85a85 ZD |
47 | of the variable before and after increment will be stored in VAR_BEFORE and |
48 | VAR_AFTER (unless they are NULL). */ | |
49 | ||
50 | void | |
51 | create_iv (tree base, tree step, tree var, struct loop *loop, | |
52 | block_stmt_iterator *incr_pos, bool after, | |
53 | tree *var_before, tree *var_after) | |
54 | { | |
8b11a64c | 55 | tree stmt, initial, step1, stmts; |
82b85a85 ZD |
56 | tree vb, va; |
57 | enum tree_code incr_op = PLUS_EXPR; | |
9be872b7 | 58 | edge pe = loop_preheader_edge (loop); |
82b85a85 ZD |
59 | |
60 | if (!var) | |
61 | { | |
62 | var = create_tmp_var (TREE_TYPE (base), "ivtmp"); | |
f004ab02 | 63 | add_referenced_var (var); |
82b85a85 ZD |
64 | } |
65 | ||
66 | vb = make_ssa_name (var, NULL_TREE); | |
67 | if (var_before) | |
68 | *var_before = vb; | |
69 | va = make_ssa_name (var, NULL_TREE); | |
70 | if (var_after) | |
71 | *var_after = va; | |
72 | ||
73 | /* For easier readability of the created code, produce MINUS_EXPRs | |
74 | when suitable. */ | |
75 | if (TREE_CODE (step) == INTEGER_CST) | |
76 | { | |
77 | if (TYPE_UNSIGNED (TREE_TYPE (step))) | |
78 | { | |
987b67bc | 79 | step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step); |
82b85a85 ZD |
80 | if (tree_int_cst_lt (step1, step)) |
81 | { | |
82 | incr_op = MINUS_EXPR; | |
83 | step = step1; | |
84 | } | |
85 | } | |
86 | else | |
87 | { | |
6ac01510 ILT |
88 | bool ovf; |
89 | ||
90 | if (!tree_expr_nonnegative_warnv_p (step, &ovf) | |
82b85a85 ZD |
91 | && may_negate_without_overflow_p (step)) |
92 | { | |
93 | incr_op = MINUS_EXPR; | |
987b67bc | 94 | step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step); |
82b85a85 ZD |
95 | } |
96 | } | |
97 | } | |
5be014d5 AP |
98 | if (POINTER_TYPE_P (TREE_TYPE (base))) |
99 | { | |
100 | step = fold_convert (sizetype, step); | |
101 | if (incr_op == MINUS_EXPR) | |
102 | step = fold_build1 (NEGATE_EXPR, sizetype, step); | |
103 | incr_op = POINTER_PLUS_EXPR; | |
104 | } | |
9be872b7 ZD |
105 | /* Gimplify the step if necessary. We put the computations in front of the |
106 | loop (i.e. the step should be loop invariant). */ | |
1ffe34d9 | 107 | step = force_gimple_operand (step, &stmts, true, NULL_TREE); |
9be872b7 | 108 | if (stmts) |
598ec7bd | 109 | bsi_insert_on_edge_immediate (pe, stmts); |
9be872b7 | 110 | |
939409af RS |
111 | stmt = build_gimple_modify_stmt (va, |
112 | build2 (incr_op, TREE_TYPE (base), | |
113 | vb, step)); | |
82b85a85 ZD |
114 | SSA_NAME_DEF_STMT (va) = stmt; |
115 | if (after) | |
116 | bsi_insert_after (incr_pos, stmt, BSI_NEW_STMT); | |
117 | else | |
118 | bsi_insert_before (incr_pos, stmt, BSI_NEW_STMT); | |
119 | ||
8b11a64c ZD |
120 | initial = force_gimple_operand (base, &stmts, true, var); |
121 | if (stmts) | |
598ec7bd | 122 | bsi_insert_on_edge_immediate (pe, stmts); |
82b85a85 ZD |
123 | |
124 | stmt = create_phi_node (vb, loop->header); | |
125 | SSA_NAME_DEF_STMT (vb) = stmt; | |
d2e398df KH |
126 | add_phi_arg (stmt, initial, loop_preheader_edge (loop)); |
127 | add_phi_arg (stmt, va, loop_latch_edge (loop)); | |
82b85a85 ZD |
128 | } |
129 | ||
c913f08a ZD |
130 | /* Add exit phis for the USE on EXIT. */ |
131 | ||
132 | static void | |
133 | add_exit_phis_edge (basic_block exit, tree use) | |
134 | { | |
135 | tree phi, def_stmt = SSA_NAME_DEF_STMT (use); | |
136 | basic_block def_bb = bb_for_stmt (def_stmt); | |
137 | struct loop *def_loop; | |
138 | edge e; | |
628f6a4e | 139 | edge_iterator ei; |
c913f08a ZD |
140 | |
141 | /* Check that some of the edges entering the EXIT block exits a loop in | |
142 | that USE is defined. */ | |
628f6a4e | 143 | FOR_EACH_EDGE (e, ei, exit->preds) |
c913f08a ZD |
144 | { |
145 | def_loop = find_common_loop (def_bb->loop_father, e->src->loop_father); | |
146 | if (!flow_bb_inside_loop_p (def_loop, e->dest)) | |
147 | break; | |
148 | } | |
149 | ||
150 | if (!e) | |
151 | return; | |
152 | ||
153 | phi = create_phi_node (use, exit); | |
84d65814 | 154 | create_new_def_for (PHI_RESULT (phi), phi, PHI_RESULT_PTR (phi)); |
628f6a4e | 155 | FOR_EACH_EDGE (e, ei, exit->preds) |
d2e398df | 156 | add_phi_arg (phi, use, e); |
c913f08a ZD |
157 | } |
158 | ||
159 | /* Add exit phis for VAR that is used in LIVEIN. | |
160 | Exits of the loops are stored in EXITS. */ | |
161 | ||
162 | static void | |
163 | add_exit_phis_var (tree var, bitmap livein, bitmap exits) | |
164 | { | |
165 | bitmap def; | |
3cd8c58a | 166 | unsigned index; |
c913f08a | 167 | basic_block def_bb = bb_for_stmt (SSA_NAME_DEF_STMT (var)); |
87c476a2 | 168 | bitmap_iterator bi; |
c913f08a | 169 | |
0bca51f0 DN |
170 | if (is_gimple_reg (var)) |
171 | bitmap_clear_bit (livein, def_bb->index); | |
172 | else | |
173 | bitmap_set_bit (livein, def_bb->index); | |
c913f08a | 174 | |
8bdbfff5 | 175 | def = BITMAP_ALLOC (NULL); |
c913f08a ZD |
176 | bitmap_set_bit (def, def_bb->index); |
177 | compute_global_livein (livein, def); | |
8bdbfff5 | 178 | BITMAP_FREE (def); |
c913f08a | 179 | |
87c476a2 ZD |
180 | EXECUTE_IF_AND_IN_BITMAP (exits, livein, 0, index, bi) |
181 | { | |
182 | add_exit_phis_edge (BASIC_BLOCK (index), var); | |
183 | } | |
c913f08a ZD |
184 | } |
185 | ||
186 | /* Add exit phis for the names marked in NAMES_TO_RENAME. | |
187 | Exits of the loops are stored in EXITS. Sets of blocks where the ssa | |
188 | names are used are stored in USE_BLOCKS. */ | |
189 | ||
190 | static void | |
191 | add_exit_phis (bitmap names_to_rename, bitmap *use_blocks, bitmap loop_exits) | |
192 | { | |
193 | unsigned i; | |
87c476a2 | 194 | bitmap_iterator bi; |
c913f08a | 195 | |
87c476a2 | 196 | EXECUTE_IF_SET_IN_BITMAP (names_to_rename, 0, i, bi) |
c913f08a ZD |
197 | { |
198 | add_exit_phis_var (ssa_name (i), use_blocks[i], loop_exits); | |
87c476a2 | 199 | } |
c913f08a ZD |
200 | } |
201 | ||
202 | /* Returns a bitmap of all loop exit edge targets. */ | |
203 | ||
204 | static bitmap | |
205 | get_loops_exits (void) | |
206 | { | |
8bdbfff5 | 207 | bitmap exits = BITMAP_ALLOC (NULL); |
c913f08a ZD |
208 | basic_block bb; |
209 | edge e; | |
628f6a4e | 210 | edge_iterator ei; |
c913f08a ZD |
211 | |
212 | FOR_EACH_BB (bb) | |
213 | { | |
628f6a4e | 214 | FOR_EACH_EDGE (e, ei, bb->preds) |
c913f08a ZD |
215 | if (e->src != ENTRY_BLOCK_PTR |
216 | && !flow_bb_inside_loop_p (e->src->loop_father, bb)) | |
217 | { | |
218 | bitmap_set_bit (exits, bb->index); | |
219 | break; | |
220 | } | |
221 | } | |
222 | ||
223 | return exits; | |
224 | } | |
225 | ||
226 | /* For USE in BB, if it is used outside of the loop it is defined in, | |
227 | mark it for rewrite. Record basic block BB where it is used | |
84d65814 | 228 | to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */ |
c913f08a ZD |
229 | |
230 | static void | |
84d65814 DN |
231 | find_uses_to_rename_use (basic_block bb, tree use, bitmap *use_blocks, |
232 | bitmap need_phis) | |
c913f08a ZD |
233 | { |
234 | unsigned ver; | |
235 | basic_block def_bb; | |
236 | struct loop *def_loop; | |
237 | ||
238 | if (TREE_CODE (use) != SSA_NAME) | |
239 | return; | |
240 | ||
84d65814 DN |
241 | /* We don't need to keep virtual operands in loop-closed form. */ |
242 | if (!is_gimple_reg (use)) | |
243 | return; | |
244 | ||
c913f08a ZD |
245 | ver = SSA_NAME_VERSION (use); |
246 | def_bb = bb_for_stmt (SSA_NAME_DEF_STMT (use)); | |
247 | if (!def_bb) | |
248 | return; | |
249 | def_loop = def_bb->loop_father; | |
250 | ||
d6e840ee | 251 | /* If the definition is not inside a loop, it is not interesting. */ |
9ba025a2 | 252 | if (!loop_outer (def_loop)) |
c913f08a ZD |
253 | return; |
254 | ||
d6e840ee RG |
255 | /* If the use is not outside of the loop it is defined in, it is not |
256 | interesting. */ | |
257 | if (flow_bb_inside_loop_p (def_loop, bb)) | |
258 | return; | |
259 | ||
c913f08a | 260 | if (!use_blocks[ver]) |
8bdbfff5 | 261 | use_blocks[ver] = BITMAP_ALLOC (NULL); |
c913f08a ZD |
262 | bitmap_set_bit (use_blocks[ver], bb->index); |
263 | ||
84d65814 | 264 | bitmap_set_bit (need_phis, ver); |
c913f08a ZD |
265 | } |
266 | ||
267 | /* For uses in STMT, mark names that are used outside of the loop they are | |
268 | defined to rewrite. Record the set of blocks in that the ssa | |
84d65814 DN |
269 | names are defined to USE_BLOCKS and the ssa names themselves to |
270 | NEED_PHIS. */ | |
c913f08a ZD |
271 | |
272 | static void | |
84d65814 | 273 | find_uses_to_rename_stmt (tree stmt, bitmap *use_blocks, bitmap need_phis) |
c913f08a | 274 | { |
4c124b4c AM |
275 | ssa_op_iter iter; |
276 | tree var; | |
c913f08a ZD |
277 | basic_block bb = bb_for_stmt (stmt); |
278 | ||
38635499 | 279 | FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_ALL_USES) |
84d65814 | 280 | find_uses_to_rename_use (bb, var, use_blocks, need_phis); |
c913f08a ZD |
281 | } |
282 | ||
2b271002 ZD |
283 | /* Marks names that are used in BB and outside of the loop they are |
284 | defined in for rewrite. Records the set of blocks in that the ssa | |
84d65814 DN |
285 | names are defined to USE_BLOCKS. Record the SSA names that will |
286 | need exit PHIs in NEED_PHIS. */ | |
c913f08a ZD |
287 | |
288 | static void | |
84d65814 | 289 | find_uses_to_rename_bb (basic_block bb, bitmap *use_blocks, bitmap need_phis) |
c913f08a | 290 | { |
c913f08a | 291 | block_stmt_iterator bsi; |
2b271002 ZD |
292 | edge e; |
293 | edge_iterator ei; | |
c913f08a | 294 | tree phi; |
c913f08a | 295 | |
2b271002 ZD |
296 | FOR_EACH_EDGE (e, ei, bb->succs) |
297 | for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi)) | |
298 | find_uses_to_rename_use (bb, PHI_ARG_DEF_FROM_EDGE (phi, e), | |
84d65814 | 299 | use_blocks, need_phis); |
2b271002 ZD |
300 | |
301 | for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) | |
84d65814 | 302 | find_uses_to_rename_stmt (bsi_stmt (bsi), use_blocks, need_phis); |
2b271002 ZD |
303 | } |
304 | ||
305 | /* Marks names that are used outside of the loop they are defined in | |
306 | for rewrite. Records the set of blocks in that the ssa | |
307 | names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL, | |
308 | scan only blocks in this set. */ | |
309 | ||
310 | static void | |
84d65814 | 311 | find_uses_to_rename (bitmap changed_bbs, bitmap *use_blocks, bitmap need_phis) |
2b271002 ZD |
312 | { |
313 | basic_block bb; | |
314 | unsigned index; | |
315 | bitmap_iterator bi; | |
c913f08a | 316 | |
84d65814 | 317 | if (changed_bbs && !bitmap_empty_p (changed_bbs)) |
2b271002 ZD |
318 | { |
319 | EXECUTE_IF_SET_IN_BITMAP (changed_bbs, 0, index, bi) | |
320 | { | |
84d65814 | 321 | find_uses_to_rename_bb (BASIC_BLOCK (index), use_blocks, need_phis); |
2b271002 ZD |
322 | } |
323 | } | |
324 | else | |
325 | { | |
326 | FOR_EACH_BB (bb) | |
327 | { | |
84d65814 | 328 | find_uses_to_rename_bb (bb, use_blocks, need_phis); |
2b271002 | 329 | } |
c913f08a ZD |
330 | } |
331 | } | |
332 | ||
333 | /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra | |
334 | phi nodes to ensure that no variable is used outside the loop it is | |
335 | defined in. | |
336 | ||
337 | This strengthening of the basic ssa form has several advantages: | |
338 | ||
339 | 1) Updating it during unrolling/peeling/versioning is trivial, since | |
340 | we do not need to care about the uses outside of the loop. | |
341 | 2) The behavior of all uses of an induction variable is the same. | |
342 | Without this, you need to distinguish the case when the variable | |
343 | is used outside of the loop it is defined in, for example | |
344 | ||
345 | for (i = 0; i < 100; i++) | |
346 | { | |
347 | for (j = 0; j < 100; j++) | |
348 | { | |
349 | k = i + j; | |
350 | use1 (k); | |
351 | } | |
352 | use2 (k); | |
353 | } | |
354 | ||
355 | Looking from the outer loop with the normal SSA form, the first use of k | |
356 | is not well-behaved, while the second one is an induction variable with | |
2b271002 ZD |
357 | base 99 and step 1. |
358 | ||
359 | If CHANGED_BBS is not NULL, we look for uses outside loops only in | |
84d65814 DN |
360 | the basic blocks in this set. |
361 | ||
362 | UPDATE_FLAG is used in the call to update_ssa. See | |
363 | TODO_update_ssa* for documentation. */ | |
c913f08a ZD |
364 | |
365 | void | |
84d65814 | 366 | rewrite_into_loop_closed_ssa (bitmap changed_bbs, unsigned update_flag) |
c913f08a | 367 | { |
c7b852c8 | 368 | bitmap loop_exits; |
c913f08a | 369 | bitmap *use_blocks; |
84d65814 | 370 | unsigned i, old_num_ssa_names; |
c7b852c8 ZD |
371 | bitmap names_to_rename; |
372 | ||
f87000d0 | 373 | loops_state_set (LOOP_CLOSED_SSA); |
d51157de | 374 | if (number_of_loops () <= 1) |
c7b852c8 ZD |
375 | return; |
376 | ||
377 | loop_exits = get_loops_exits (); | |
378 | names_to_rename = BITMAP_ALLOC (NULL); | |
c913f08a | 379 | |
84d65814 DN |
380 | /* If the pass has caused the SSA form to be out-of-date, update it |
381 | now. */ | |
382 | update_ssa (update_flag); | |
c913f08a | 383 | |
84d65814 | 384 | old_num_ssa_names = num_ssa_names; |
5ed6ace5 | 385 | use_blocks = XCNEWVEC (bitmap, old_num_ssa_names); |
c913f08a ZD |
386 | |
387 | /* Find the uses outside loops. */ | |
84d65814 | 388 | find_uses_to_rename (changed_bbs, use_blocks, names_to_rename); |
2b271002 | 389 | |
84d65814 | 390 | /* Add the PHI nodes on exits of the loops for the names we need to |
c913f08a | 391 | rewrite. */ |
c913f08a ZD |
392 | add_exit_phis (names_to_rename, use_blocks, loop_exits); |
393 | ||
84d65814 | 394 | for (i = 0; i < old_num_ssa_names; i++) |
8bdbfff5 | 395 | BITMAP_FREE (use_blocks[i]); |
c913f08a | 396 | free (use_blocks); |
8bdbfff5 NS |
397 | BITMAP_FREE (loop_exits); |
398 | BITMAP_FREE (names_to_rename); | |
c913f08a | 399 | |
84d65814 DN |
400 | /* Fix up all the names found to be used outside their original |
401 | loops. */ | |
402 | update_ssa (TODO_update_ssa); | |
c913f08a ZD |
403 | } |
404 | ||
405 | /* Check invariants of the loop closed ssa form for the USE in BB. */ | |
406 | ||
407 | static void | |
408 | check_loop_closed_ssa_use (basic_block bb, tree use) | |
409 | { | |
410 | tree def; | |
411 | basic_block def_bb; | |
412 | ||
84d65814 | 413 | if (TREE_CODE (use) != SSA_NAME || !is_gimple_reg (use)) |
c913f08a ZD |
414 | return; |
415 | ||
416 | def = SSA_NAME_DEF_STMT (use); | |
417 | def_bb = bb_for_stmt (def); | |
1e128c5f GB |
418 | gcc_assert (!def_bb |
419 | || flow_bb_inside_loop_p (def_bb->loop_father, bb)); | |
c913f08a ZD |
420 | } |
421 | ||
422 | /* Checks invariants of loop closed ssa form in statement STMT in BB. */ | |
423 | ||
424 | static void | |
425 | check_loop_closed_ssa_stmt (basic_block bb, tree stmt) | |
426 | { | |
4c124b4c AM |
427 | ssa_op_iter iter; |
428 | tree var; | |
c913f08a | 429 | |
38635499 | 430 | FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_ALL_USES) |
4c124b4c | 431 | check_loop_closed_ssa_use (bb, var); |
c913f08a ZD |
432 | } |
433 | ||
434 | /* Checks that invariants of the loop closed ssa form are preserved. */ | |
435 | ||
436 | void | |
437 | verify_loop_closed_ssa (void) | |
438 | { | |
439 | basic_block bb; | |
440 | block_stmt_iterator bsi; | |
441 | tree phi; | |
442 | unsigned i; | |
443 | ||
d51157de | 444 | if (number_of_loops () <= 1) |
84d65814 DN |
445 | return; |
446 | ||
f430bae8 | 447 | verify_ssa (false); |
c913f08a ZD |
448 | |
449 | FOR_EACH_BB (bb) | |
450 | { | |
bb29d951 | 451 | for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi)) |
c913f08a ZD |
452 | for (i = 0; i < (unsigned) PHI_NUM_ARGS (phi); i++) |
453 | check_loop_closed_ssa_use (PHI_ARG_EDGE (phi, i)->src, | |
454 | PHI_ARG_DEF (phi, i)); | |
455 | ||
456 | for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) | |
457 | check_loop_closed_ssa_stmt (bb, bsi_stmt (bsi)); | |
458 | } | |
459 | } | |
8b11a64c ZD |
460 | |
461 | /* Split loop exit edge EXIT. The things are a bit complicated by a need to | |
5f40b3cb | 462 | preserve the loop closed ssa form. The newly created block is returned. */ |
8b11a64c | 463 | |
5f40b3cb | 464 | basic_block |
8b11a64c ZD |
465 | split_loop_exit_edge (edge exit) |
466 | { | |
467 | basic_block dest = exit->dest; | |
598ec7bd | 468 | basic_block bb = split_edge (exit); |
7fac6722 | 469 | tree phi, new_phi, new_name, name; |
8b11a64c ZD |
470 | use_operand_p op_p; |
471 | ||
bb29d951 | 472 | for (phi = phi_nodes (dest); phi; phi = PHI_CHAIN (phi)) |
8b11a64c | 473 | { |
c5cbcccf | 474 | op_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (bb)); |
8b11a64c | 475 | |
7fac6722 ZD |
476 | name = USE_FROM_PTR (op_p); |
477 | ||
38635499 | 478 | /* If the argument of the PHI node is a constant, we do not need |
7fac6722 ZD |
479 | to keep it inside loop. */ |
480 | if (TREE_CODE (name) != SSA_NAME) | |
481 | continue; | |
482 | ||
483 | /* Otherwise create an auxiliary phi node that will copy the value | |
38635499 | 484 | of the SSA name out of the loop. */ |
7fac6722 | 485 | new_name = duplicate_ssa_name (name, NULL); |
8b11a64c ZD |
486 | new_phi = create_phi_node (new_name, bb); |
487 | SSA_NAME_DEF_STMT (new_name) = new_phi; | |
d2e398df | 488 | add_phi_arg (new_phi, name, exit); |
8b11a64c ZD |
489 | SET_USE (op_p, new_name); |
490 | } | |
5f40b3cb ZD |
491 | |
492 | return bb; | |
8b11a64c ZD |
493 | } |
494 | ||
8b11a64c ZD |
495 | /* Returns the basic block in that statements should be emitted for induction |
496 | variables incremented at the end of the LOOP. */ | |
497 | ||
498 | basic_block | |
499 | ip_end_pos (struct loop *loop) | |
500 | { | |
501 | return loop->latch; | |
502 | } | |
503 | ||
504 | /* Returns the basic block in that statements should be emitted for induction | |
505 | variables incremented just before exit condition of a LOOP. */ | |
506 | ||
507 | basic_block | |
508 | ip_normal_pos (struct loop *loop) | |
509 | { | |
510 | tree last; | |
511 | basic_block bb; | |
512 | edge exit; | |
513 | ||
c5cbcccf | 514 | if (!single_pred_p (loop->latch)) |
8b11a64c ZD |
515 | return NULL; |
516 | ||
c5cbcccf | 517 | bb = single_pred (loop->latch); |
8b11a64c | 518 | last = last_stmt (bb); |
ae2cf11b RG |
519 | if (!last |
520 | || TREE_CODE (last) != COND_EXPR) | |
8b11a64c ZD |
521 | return NULL; |
522 | ||
628f6a4e | 523 | exit = EDGE_SUCC (bb, 0); |
8b11a64c | 524 | if (exit->dest == loop->latch) |
628f6a4e | 525 | exit = EDGE_SUCC (bb, 1); |
8b11a64c ZD |
526 | |
527 | if (flow_bb_inside_loop_p (loop, exit->dest)) | |
528 | return NULL; | |
529 | ||
530 | return bb; | |
531 | } | |
532 | ||
533 | /* Stores the standard position for induction variable increment in LOOP | |
534 | (just before the exit condition if it is available and latch block is empty, | |
535 | end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if | |
536 | the increment should be inserted after *BSI. */ | |
537 | ||
538 | void | |
539 | standard_iv_increment_position (struct loop *loop, block_stmt_iterator *bsi, | |
540 | bool *insert_after) | |
541 | { | |
542 | basic_block bb = ip_normal_pos (loop), latch = ip_end_pos (loop); | |
543 | tree last = last_stmt (latch); | |
544 | ||
545 | if (!bb | |
546 | || (last && TREE_CODE (last) != LABEL_EXPR)) | |
547 | { | |
548 | *bsi = bsi_last (latch); | |
549 | *insert_after = true; | |
550 | } | |
551 | else | |
552 | { | |
553 | *bsi = bsi_last (bb); | |
554 | *insert_after = false; | |
555 | } | |
556 | } | |
92fc4a2f ZD |
557 | |
558 | /* Copies phi node arguments for duplicated blocks. The index of the first | |
559 | duplicated block is FIRST_NEW_BLOCK. */ | |
560 | ||
561 | static void | |
562 | copy_phi_node_args (unsigned first_new_block) | |
563 | { | |
564 | unsigned i; | |
565 | ||
566 | for (i = first_new_block; i < (unsigned) last_basic_block; i++) | |
6580ee77 | 567 | BASIC_BLOCK (i)->flags |= BB_DUPLICATED; |
92fc4a2f ZD |
568 | |
569 | for (i = first_new_block; i < (unsigned) last_basic_block; i++) | |
570 | add_phi_args_after_copy_bb (BASIC_BLOCK (i)); | |
571 | ||
572 | for (i = first_new_block; i < (unsigned) last_basic_block; i++) | |
6580ee77 | 573 | BASIC_BLOCK (i)->flags &= ~BB_DUPLICATED; |
92fc4a2f ZD |
574 | } |
575 | ||
92fc4a2f | 576 | |
84d65814 DN |
577 | /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also |
578 | updates the PHI nodes at start of the copied region. In order to | |
579 | achieve this, only loops whose exits all lead to the same location | |
580 | are handled. | |
92fc4a2f | 581 | |
84d65814 DN |
582 | Notice that we do not completely update the SSA web after |
583 | duplication. The caller is responsible for calling update_ssa | |
584 | after the loop has been duplicated. */ | |
92fc4a2f ZD |
585 | |
586 | bool | |
587 | tree_duplicate_loop_to_header_edge (struct loop *loop, edge e, | |
92fc4a2f | 588 | unsigned int ndupl, sbitmap wont_exit, |
ee8c1b05 ZD |
589 | edge orig, VEC (edge, heap) **to_remove, |
590 | int flags) | |
92fc4a2f ZD |
591 | { |
592 | unsigned first_new_block; | |
92fc4a2f | 593 | |
f87000d0 | 594 | if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES)) |
92fc4a2f | 595 | return false; |
f87000d0 | 596 | if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS)) |
92fc4a2f ZD |
597 | return false; |
598 | ||
599 | #ifdef ENABLE_CHECKING | |
d6e840ee RG |
600 | if (loops_state_satisfies_p (LOOP_CLOSED_SSA)) |
601 | verify_loop_closed_ssa (); | |
92fc4a2f ZD |
602 | #endif |
603 | ||
92fc4a2f | 604 | first_new_block = last_basic_block; |
d73be268 | 605 | if (!duplicate_loop_to_header_edge (loop, e, ndupl, wont_exit, |
ee8c1b05 | 606 | orig, to_remove, flags)) |
92fc4a2f ZD |
607 | return false; |
608 | ||
609 | /* Readd the removed phi args for e. */ | |
71882046 | 610 | flush_pending_stmts (e); |
92fc4a2f ZD |
611 | |
612 | /* Copy the phi node arguments. */ | |
613 | copy_phi_node_args (first_new_block); | |
614 | ||
92fc4a2f | 615 | scev_reset (); |
92fc4a2f ZD |
616 | |
617 | return true; | |
618 | } | |
17684618 | 619 | |
17684618 ZD |
620 | /* Returns true if we can unroll LOOP FACTOR times. Number |
621 | of iterations of the loop is returned in NITER. */ | |
622 | ||
623 | bool | |
624 | can_unroll_loop_p (struct loop *loop, unsigned factor, | |
625 | struct tree_niter_desc *niter) | |
626 | { | |
627 | edge exit; | |
628 | ||
629 | /* Check whether unrolling is possible. We only want to unroll loops | |
630 | for that we are able to determine number of iterations. We also | |
631 | want to split the extra iterations of the loop from its end, | |
632 | therefore we require that the loop has precisely one | |
633 | exit. */ | |
634 | ||
635 | exit = single_dom_exit (loop); | |
636 | if (!exit) | |
637 | return false; | |
638 | ||
639 | if (!number_of_iterations_exit (loop, exit, niter, false) | |
bf8dbe38 ZD |
640 | || niter->cmp == ERROR_MARK |
641 | /* Scalar evolutions analysis might have copy propagated | |
642 | the abnormal ssa names into these expressions, hence | |
2f8e468b | 643 | emitting the computations based on them during loop |
bf8dbe38 ZD |
644 | unrolling might create overlapping life ranges for |
645 | them, and failures in out-of-ssa. */ | |
646 | || contains_abnormal_ssa_name_p (niter->may_be_zero) | |
647 | || contains_abnormal_ssa_name_p (niter->control.base) | |
648 | || contains_abnormal_ssa_name_p (niter->control.step) | |
649 | || contains_abnormal_ssa_name_p (niter->bound)) | |
17684618 ZD |
650 | return false; |
651 | ||
652 | /* And of course, we must be able to duplicate the loop. */ | |
653 | if (!can_duplicate_loop_p (loop)) | |
654 | return false; | |
655 | ||
656 | /* The final loop should be small enough. */ | |
7f9bc51b | 657 | if (tree_num_loop_insns (loop, &eni_size_weights) * factor |
17684618 ZD |
658 | > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS)) |
659 | return false; | |
660 | ||
661 | return true; | |
662 | } | |
663 | ||
664 | /* Determines the conditions that control execution of LOOP unrolled FACTOR | |
665 | times. DESC is number of iterations of LOOP. ENTER_COND is set to | |
666 | condition that must be true if the main loop can be entered. | |
667 | EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing | |
668 | how the exit from the unrolled loop should be controlled. */ | |
669 | ||
670 | static void | |
671 | determine_exit_conditions (struct loop *loop, struct tree_niter_desc *desc, | |
672 | unsigned factor, tree *enter_cond, | |
673 | tree *exit_base, tree *exit_step, | |
674 | enum tree_code *exit_cmp, tree *exit_bound) | |
675 | { | |
676 | tree stmts; | |
677 | tree base = desc->control.base; | |
678 | tree step = desc->control.step; | |
679 | tree bound = desc->bound; | |
d24a32a1 | 680 | tree type = TREE_TYPE (step); |
17684618 ZD |
681 | tree bigstep, delta; |
682 | tree min = lower_bound_in_type (type, type); | |
683 | tree max = upper_bound_in_type (type, type); | |
684 | enum tree_code cmp = desc->cmp; | |
685 | tree cond = boolean_true_node, assum; | |
686 | ||
d24a32a1 ZD |
687 | /* For pointers, do the arithmetics in the type of step (sizetype). */ |
688 | base = fold_convert (type, base); | |
689 | bound = fold_convert (type, bound); | |
690 | ||
17684618 ZD |
691 | *enter_cond = boolean_false_node; |
692 | *exit_base = NULL_TREE; | |
693 | *exit_step = NULL_TREE; | |
694 | *exit_cmp = ERROR_MARK; | |
695 | *exit_bound = NULL_TREE; | |
696 | gcc_assert (cmp != ERROR_MARK); | |
697 | ||
698 | /* We only need to be correct when we answer question | |
699 | "Do at least FACTOR more iterations remain?" in the unrolled loop. | |
700 | Thus, transforming BASE + STEP * i <> BOUND to | |
701 | BASE + STEP * i < BOUND is ok. */ | |
702 | if (cmp == NE_EXPR) | |
703 | { | |
704 | if (tree_int_cst_sign_bit (step)) | |
705 | cmp = GT_EXPR; | |
706 | else | |
707 | cmp = LT_EXPR; | |
708 | } | |
709 | else if (cmp == LT_EXPR) | |
710 | { | |
711 | gcc_assert (!tree_int_cst_sign_bit (step)); | |
712 | } | |
713 | else if (cmp == GT_EXPR) | |
714 | { | |
715 | gcc_assert (tree_int_cst_sign_bit (step)); | |
716 | } | |
717 | else | |
718 | gcc_unreachable (); | |
719 | ||
720 | /* The main body of the loop may be entered iff: | |
721 | ||
722 | 1) desc->may_be_zero is false. | |
723 | 2) it is possible to check that there are at least FACTOR iterations | |
724 | of the loop, i.e., BOUND - step * FACTOR does not overflow. | |
725 | 3) # of iterations is at least FACTOR */ | |
726 | ||
6e682d7e | 727 | if (!integer_zerop (desc->may_be_zero)) |
17684618 ZD |
728 | cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
729 | invert_truthvalue (desc->may_be_zero), | |
730 | cond); | |
731 | ||
732 | bigstep = fold_build2 (MULT_EXPR, type, step, | |
733 | build_int_cst_type (type, factor)); | |
734 | delta = fold_build2 (MINUS_EXPR, type, bigstep, step); | |
735 | if (cmp == LT_EXPR) | |
736 | assum = fold_build2 (GE_EXPR, boolean_type_node, | |
737 | bound, | |
738 | fold_build2 (PLUS_EXPR, type, min, delta)); | |
739 | else | |
740 | assum = fold_build2 (LE_EXPR, boolean_type_node, | |
741 | bound, | |
742 | fold_build2 (PLUS_EXPR, type, max, delta)); | |
743 | cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond); | |
744 | ||
745 | bound = fold_build2 (MINUS_EXPR, type, bound, delta); | |
746 | assum = fold_build2 (cmp, boolean_type_node, base, bound); | |
747 | cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond); | |
748 | ||
749 | cond = force_gimple_operand (unshare_expr (cond), &stmts, false, NULL_TREE); | |
750 | if (stmts) | |
598ec7bd | 751 | bsi_insert_on_edge_immediate (loop_preheader_edge (loop), stmts); |
17684618 ZD |
752 | /* cond now may be a gimple comparison, which would be OK, but also any |
753 | other gimple rhs (say a && b). In this case we need to force it to | |
754 | operand. */ | |
755 | if (!is_gimple_condexpr (cond)) | |
756 | { | |
757 | cond = force_gimple_operand (cond, &stmts, true, NULL_TREE); | |
758 | if (stmts) | |
598ec7bd | 759 | bsi_insert_on_edge_immediate (loop_preheader_edge (loop), stmts); |
17684618 ZD |
760 | } |
761 | *enter_cond = cond; | |
762 | ||
763 | base = force_gimple_operand (unshare_expr (base), &stmts, true, NULL_TREE); | |
764 | if (stmts) | |
598ec7bd | 765 | bsi_insert_on_edge_immediate (loop_preheader_edge (loop), stmts); |
17684618 ZD |
766 | bound = force_gimple_operand (unshare_expr (bound), &stmts, true, NULL_TREE); |
767 | if (stmts) | |
598ec7bd | 768 | bsi_insert_on_edge_immediate (loop_preheader_edge (loop), stmts); |
17684618 ZD |
769 | |
770 | *exit_base = base; | |
771 | *exit_step = bigstep; | |
772 | *exit_cmp = cmp; | |
773 | *exit_bound = bound; | |
774 | } | |
775 | ||
14fa2cc0 ZD |
776 | /* Scales the frequencies of all basic blocks in LOOP that are strictly |
777 | dominated by BB by NUM/DEN. */ | |
778 | ||
779 | static void | |
780 | scale_dominated_blocks_in_loop (struct loop *loop, basic_block bb, | |
781 | int num, int den) | |
782 | { | |
783 | basic_block son; | |
784 | ||
785 | if (den == 0) | |
786 | return; | |
787 | ||
788 | for (son = first_dom_son (CDI_DOMINATORS, bb); | |
789 | son; | |
790 | son = next_dom_son (CDI_DOMINATORS, son)) | |
791 | { | |
792 | if (!flow_bb_inside_loop_p (loop, son)) | |
793 | continue; | |
794 | scale_bbs_frequencies_int (&son, 1, num, den); | |
795 | scale_dominated_blocks_in_loop (loop, son, num, den); | |
796 | } | |
797 | } | |
798 | ||
d73be268 ZD |
799 | /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP. |
800 | EXIT is the exit of the loop to that DESC corresponds. | |
801 | ||
17684618 ZD |
802 | If N is number of iterations of the loop and MAY_BE_ZERO is the condition |
803 | under that loop exits in the first iteration even if N != 0, | |
804 | ||
805 | while (1) | |
806 | { | |
807 | x = phi (init, next); | |
808 | ||
809 | pre; | |
810 | if (st) | |
811 | break; | |
812 | post; | |
813 | } | |
814 | ||
815 | becomes (with possibly the exit conditions formulated a bit differently, | |
816 | avoiding the need to create a new iv): | |
817 | ||
818 | if (MAY_BE_ZERO || N < FACTOR) | |
819 | goto rest; | |
820 | ||
821 | do | |
822 | { | |
823 | x = phi (init, next); | |
824 | ||
825 | pre; | |
826 | post; | |
827 | pre; | |
828 | post; | |
829 | ... | |
830 | pre; | |
831 | post; | |
832 | N -= FACTOR; | |
833 | ||
834 | } while (N >= FACTOR); | |
835 | ||
836 | rest: | |
837 | init' = phi (init, x); | |
838 | ||
839 | while (1) | |
840 | { | |
841 | x = phi (init', next); | |
842 | ||
843 | pre; | |
844 | if (st) | |
845 | break; | |
846 | post; | |
567b96ed ZD |
847 | } |
848 | ||
849 | Before the loop is unrolled, TRANSFORM is called for it (only for the | |
850 | unrolled loop, but not for its versioned copy). DATA is passed to | |
851 | TRANSFORM. */ | |
17684618 | 852 | |
03cb2019 ZD |
853 | /* Probability in % that the unrolled loop is entered. Just a guess. */ |
854 | #define PROB_UNROLLED_LOOP_ENTERED 90 | |
855 | ||
17684618 | 856 | void |
567b96ed ZD |
857 | tree_transform_and_unroll_loop (struct loop *loop, unsigned factor, |
858 | edge exit, struct tree_niter_desc *desc, | |
859 | transform_callback transform, | |
860 | void *data) | |
17684618 | 861 | { |
14fa2cc0 | 862 | tree exit_if, ctr_before, ctr_after; |
17684618 ZD |
863 | tree enter_main_cond, exit_base, exit_step, exit_bound; |
864 | enum tree_code exit_cmp; | |
865 | tree phi_old_loop, phi_new_loop, phi_rest, init, next, new_init, var; | |
866 | struct loop *new_loop; | |
867 | basic_block rest, exit_bb; | |
868 | edge old_entry, new_entry, old_latch, precond_edge, new_exit; | |
14fa2cc0 | 869 | edge new_nonexit, e; |
17684618 ZD |
870 | block_stmt_iterator bsi; |
871 | use_operand_p op; | |
872 | bool ok; | |
03cb2019 | 873 | unsigned est_niter, prob_entry, scale_unrolled, scale_rest, freq_e, freq_h; |
14fa2cc0 | 874 | unsigned new_est_niter, i, prob; |
8e08deeb | 875 | unsigned irr = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP; |
17684618 | 876 | sbitmap wont_exit; |
14fa2cc0 | 877 | VEC (edge, heap) *to_remove = NULL; |
17684618 ZD |
878 | |
879 | est_niter = expected_loop_iterations (loop); | |
880 | determine_exit_conditions (loop, desc, factor, | |
881 | &enter_main_cond, &exit_base, &exit_step, | |
882 | &exit_cmp, &exit_bound); | |
883 | ||
03cb2019 ZD |
884 | /* Let us assume that the unrolled loop is quite likely to be entered. */ |
885 | if (integer_nonzerop (enter_main_cond)) | |
886 | prob_entry = REG_BR_PROB_BASE; | |
887 | else | |
888 | prob_entry = PROB_UNROLLED_LOOP_ENTERED * REG_BR_PROB_BASE / 100; | |
889 | ||
890 | /* The values for scales should keep profile consistent, and somewhat close | |
891 | to correct. | |
892 | ||
893 | TODO: The current value of SCALE_REST makes it appear that the loop that | |
894 | is created by splitting the remaining iterations of the unrolled loop is | |
895 | executed the same number of times as the original loop, and with the same | |
896 | frequencies, which is obviously wrong. This does not appear to cause | |
897 | problems, so we do not bother with fixing it for now. To make the profile | |
898 | correct, we would need to change the probability of the exit edge of the | |
899 | loop, and recompute the distribution of frequencies in its body because | |
900 | of this change (scale the frequencies of blocks before and after the exit | |
901 | by appropriate factors). */ | |
902 | scale_unrolled = prob_entry; | |
903 | scale_rest = REG_BR_PROB_BASE; | |
904 | ||
905 | new_loop = loop_version (loop, enter_main_cond, NULL, | |
906 | prob_entry, scale_unrolled, scale_rest, true); | |
17684618 ZD |
907 | gcc_assert (new_loop != NULL); |
908 | update_ssa (TODO_update_ssa); | |
909 | ||
567b96ed | 910 | /* Determine the probability of the exit edge of the unrolled loop. */ |
03cb2019 ZD |
911 | new_est_niter = est_niter / factor; |
912 | ||
913 | /* Without profile feedback, loops for that we do not know a better estimate | |
914 | are assumed to roll 10 times. When we unroll such loop, it appears to | |
915 | roll too little, and it may even seem to be cold. To avoid this, we | |
916 | ensure that the created loop appears to roll at least 5 times (but at | |
917 | most as many times as before unrolling). */ | |
918 | if (new_est_niter < 5) | |
919 | { | |
920 | if (est_niter < 5) | |
921 | new_est_niter = est_niter; | |
922 | else | |
923 | new_est_niter = 5; | |
924 | } | |
925 | ||
14fa2cc0 ZD |
926 | /* Prepare the cfg and update the phi nodes. Move the loop exit to the |
927 | loop latch (and make its condition dummy, for the moment). */ | |
17684618 ZD |
928 | rest = loop_preheader_edge (new_loop)->src; |
929 | precond_edge = single_pred_edge (rest); | |
598ec7bd | 930 | split_edge (loop_latch_edge (loop)); |
17684618 ZD |
931 | exit_bb = single_pred (loop->latch); |
932 | ||
14fa2cc0 ZD |
933 | /* Since the exit edge will be removed, the frequency of all the blocks |
934 | in the loop that are dominated by it must be scaled by | |
935 | 1 / (1 - exit->probability). */ | |
936 | scale_dominated_blocks_in_loop (loop, exit->src, | |
937 | REG_BR_PROB_BASE, | |
938 | REG_BR_PROB_BASE - exit->probability); | |
939 | ||
567b96ed | 940 | bsi = bsi_last (exit_bb); |
a9b77cd1 ZD |
941 | exit_if = build3 (COND_EXPR, void_type_node, boolean_true_node, |
942 | NULL_TREE, NULL_TREE); | |
943 | ||
567b96ed | 944 | bsi_insert_after (&bsi, exit_if, BSI_NEW_STMT); |
8e08deeb | 945 | new_exit = make_edge (exit_bb, rest, EDGE_FALSE_VALUE | irr); |
6270df4c | 946 | rescan_loop_exit (new_exit, true, false); |
14fa2cc0 ZD |
947 | |
948 | /* Set the probability of new exit to the same of the old one. Fix | |
949 | the frequency of the latch block, by scaling it back by | |
950 | 1 - exit->probability. */ | |
951 | new_exit->count = exit->count; | |
952 | new_exit->probability = exit->probability; | |
17684618 | 953 | new_nonexit = single_pred_edge (loop->latch); |
14fa2cc0 | 954 | new_nonexit->probability = REG_BR_PROB_BASE - exit->probability; |
17684618 | 955 | new_nonexit->flags = EDGE_TRUE_VALUE; |
14fa2cc0 ZD |
956 | new_nonexit->count -= exit->count; |
957 | if (new_nonexit->count < 0) | |
958 | new_nonexit->count = 0; | |
959 | scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability, | |
960 | REG_BR_PROB_BASE); | |
17684618 ZD |
961 | |
962 | old_entry = loop_preheader_edge (loop); | |
963 | new_entry = loop_preheader_edge (new_loop); | |
964 | old_latch = loop_latch_edge (loop); | |
965 | for (phi_old_loop = phi_nodes (loop->header), | |
966 | phi_new_loop = phi_nodes (new_loop->header); | |
967 | phi_old_loop; | |
968 | phi_old_loop = PHI_CHAIN (phi_old_loop), | |
969 | phi_new_loop = PHI_CHAIN (phi_new_loop)) | |
970 | { | |
971 | init = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_entry); | |
972 | op = PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop, new_entry); | |
973 | gcc_assert (operand_equal_for_phi_arg_p (init, USE_FROM_PTR (op))); | |
974 | next = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_latch); | |
975 | ||
976 | /* Prefer using original variable as a base for the new ssa name. | |
977 | This is necessary for virtual ops, and useful in order to avoid | |
978 | losing debug info for real ops. */ | |
979 | if (TREE_CODE (next) == SSA_NAME) | |
980 | var = SSA_NAME_VAR (next); | |
981 | else if (TREE_CODE (init) == SSA_NAME) | |
982 | var = SSA_NAME_VAR (init); | |
983 | else | |
984 | { | |
985 | var = create_tmp_var (TREE_TYPE (init), "unrinittmp"); | |
f004ab02 | 986 | add_referenced_var (var); |
17684618 ZD |
987 | } |
988 | ||
989 | new_init = make_ssa_name (var, NULL_TREE); | |
990 | phi_rest = create_phi_node (new_init, rest); | |
991 | SSA_NAME_DEF_STMT (new_init) = phi_rest; | |
992 | ||
993 | add_phi_arg (phi_rest, init, precond_edge); | |
994 | add_phi_arg (phi_rest, next, new_exit); | |
995 | SET_USE (op, new_init); | |
996 | } | |
997 | ||
14fa2cc0 ZD |
998 | remove_path (exit); |
999 | ||
567b96ed ZD |
1000 | /* Transform the loop. */ |
1001 | if (transform) | |
1002 | (*transform) (loop, data); | |
1003 | ||
14fa2cc0 ZD |
1004 | /* Unroll the loop and remove the exits in all iterations except for the |
1005 | last one. */ | |
567b96ed ZD |
1006 | wont_exit = sbitmap_alloc (factor); |
1007 | sbitmap_ones (wont_exit); | |
14fa2cc0 ZD |
1008 | RESET_BIT (wont_exit, factor - 1); |
1009 | ||
567b96ed ZD |
1010 | ok = tree_duplicate_loop_to_header_edge |
1011 | (loop, loop_latch_edge (loop), factor - 1, | |
14fa2cc0 | 1012 | wont_exit, new_exit, &to_remove, DLTHE_FLAG_UPDATE_FREQ); |
567b96ed ZD |
1013 | free (wont_exit); |
1014 | gcc_assert (ok); | |
14fa2cc0 ZD |
1015 | |
1016 | for (i = 0; VEC_iterate (edge, to_remove, i, e); i++) | |
1017 | { | |
1018 | ok = remove_path (e); | |
1019 | gcc_assert (ok); | |
1020 | } | |
1021 | VEC_free (edge, heap, to_remove); | |
567b96ed ZD |
1022 | update_ssa (TODO_update_ssa); |
1023 | ||
1024 | /* Ensure that the frequencies in the loop match the new estimated | |
1025 | number of iterations, and change the probability of the new | |
1026 | exit edge. */ | |
1027 | freq_h = loop->header->frequency; | |
1028 | freq_e = EDGE_FREQUENCY (loop_preheader_edge (loop)); | |
1029 | if (freq_h != 0) | |
1030 | scale_loop_frequencies (loop, freq_e * (new_est_niter + 1), freq_h); | |
1031 | ||
1032 | exit_bb = single_pred (loop->latch); | |
1033 | new_exit = find_edge (exit_bb, rest); | |
1034 | new_exit->count = loop_preheader_edge (loop)->count; | |
1035 | new_exit->probability = REG_BR_PROB_BASE / (new_est_niter + 1); | |
1036 | ||
1037 | rest->count += new_exit->count; | |
1038 | rest->frequency += EDGE_FREQUENCY (new_exit); | |
1039 | ||
1040 | new_nonexit = single_pred_edge (loop->latch); | |
14fa2cc0 | 1041 | prob = new_nonexit->probability; |
567b96ed | 1042 | new_nonexit->probability = REG_BR_PROB_BASE - new_exit->probability; |
14fa2cc0 ZD |
1043 | new_nonexit->count = exit_bb->count - new_exit->count; |
1044 | if (new_nonexit->count < 0) | |
1045 | new_nonexit->count = 0; | |
87621e5f SE |
1046 | if (prob > 0) |
1047 | scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability, | |
1048 | prob); | |
567b96ed | 1049 | |
17684618 ZD |
1050 | /* Finally create the new counter for number of iterations and add the new |
1051 | exit instruction. */ | |
1052 | bsi = bsi_last (exit_bb); | |
567b96ed | 1053 | exit_if = bsi_stmt (bsi); |
17684618 | 1054 | create_iv (exit_base, exit_step, NULL_TREE, loop, |
567b96ed ZD |
1055 | &bsi, false, &ctr_before, &ctr_after); |
1056 | COND_EXPR_COND (exit_if) = build2 (exit_cmp, boolean_type_node, ctr_after, | |
1057 | exit_bound); | |
1058 | update_stmt (exit_if); | |
17684618 | 1059 | |
c0493b13 | 1060 | #ifdef ENABLE_CHECKING |
17684618 ZD |
1061 | verify_flow_info (); |
1062 | verify_dominators (CDI_DOMINATORS); | |
d73be268 | 1063 | verify_loop_structure (); |
17684618 | 1064 | verify_loop_closed_ssa (); |
c0493b13 | 1065 | #endif |
17684618 | 1066 | } |
567b96ed ZD |
1067 | |
1068 | /* Wrapper over tree_transform_and_unroll_loop for case we do not | |
1069 | want to transform the loop before unrolling. The meaning | |
1070 | of the arguments is the same as for tree_transform_and_unroll_loop. */ | |
1071 | ||
1072 | void | |
1073 | tree_unroll_loop (struct loop *loop, unsigned factor, | |
1074 | edge exit, struct tree_niter_desc *desc) | |
1075 | { | |
1076 | tree_transform_and_unroll_loop (loop, factor, exit, desc, | |
1077 | NULL, NULL); | |
1078 | } |