]>
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, | |
726a989a | 52 | gimple_stmt_iterator *incr_pos, bool after, |
82b85a85 ZD |
53 | tree *var_before, tree *var_after) |
54 | { | |
726a989a RB |
55 | gimple stmt; |
56 | tree initial, step1; | |
57 | gimple_seq stmts; | |
82b85a85 ZD |
58 | tree vb, va; |
59 | enum tree_code incr_op = PLUS_EXPR; | |
9be872b7 | 60 | edge pe = loop_preheader_edge (loop); |
82b85a85 ZD |
61 | |
62 | if (!var) | |
63 | { | |
64 | var = create_tmp_var (TREE_TYPE (base), "ivtmp"); | |
f004ab02 | 65 | add_referenced_var (var); |
82b85a85 ZD |
66 | } |
67 | ||
726a989a | 68 | vb = make_ssa_name (var, NULL); |
82b85a85 ZD |
69 | if (var_before) |
70 | *var_before = vb; | |
726a989a | 71 | va = make_ssa_name (var, NULL); |
82b85a85 ZD |
72 | if (var_after) |
73 | *var_after = va; | |
74 | ||
75 | /* For easier readability of the created code, produce MINUS_EXPRs | |
76 | when suitable. */ | |
77 | if (TREE_CODE (step) == INTEGER_CST) | |
78 | { | |
79 | if (TYPE_UNSIGNED (TREE_TYPE (step))) | |
80 | { | |
987b67bc | 81 | step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step); |
82b85a85 ZD |
82 | if (tree_int_cst_lt (step1, step)) |
83 | { | |
84 | incr_op = MINUS_EXPR; | |
85 | step = step1; | |
86 | } | |
87 | } | |
88 | else | |
89 | { | |
6ac01510 ILT |
90 | bool ovf; |
91 | ||
92 | if (!tree_expr_nonnegative_warnv_p (step, &ovf) | |
82b85a85 ZD |
93 | && may_negate_without_overflow_p (step)) |
94 | { | |
95 | incr_op = MINUS_EXPR; | |
987b67bc | 96 | step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step); |
82b85a85 ZD |
97 | } |
98 | } | |
99 | } | |
5be014d5 AP |
100 | if (POINTER_TYPE_P (TREE_TYPE (base))) |
101 | { | |
102 | step = fold_convert (sizetype, step); | |
103 | if (incr_op == MINUS_EXPR) | |
104 | step = fold_build1 (NEGATE_EXPR, sizetype, step); | |
105 | incr_op = POINTER_PLUS_EXPR; | |
106 | } | |
9be872b7 ZD |
107 | /* Gimplify the step if necessary. We put the computations in front of the |
108 | loop (i.e. the step should be loop invariant). */ | |
1ffe34d9 | 109 | step = force_gimple_operand (step, &stmts, true, NULL_TREE); |
9be872b7 | 110 | if (stmts) |
726a989a | 111 | gsi_insert_seq_on_edge_immediate (pe, stmts); |
9be872b7 | 112 | |
726a989a | 113 | stmt = gimple_build_assign_with_ops (incr_op, va, vb, step); |
82b85a85 | 114 | if (after) |
726a989a | 115 | gsi_insert_after (incr_pos, stmt, GSI_NEW_STMT); |
82b85a85 | 116 | else |
726a989a | 117 | gsi_insert_before (incr_pos, stmt, GSI_NEW_STMT); |
82b85a85 | 118 | |
8b11a64c ZD |
119 | initial = force_gimple_operand (base, &stmts, true, var); |
120 | if (stmts) | |
726a989a | 121 | gsi_insert_seq_on_edge_immediate (pe, stmts); |
82b85a85 ZD |
122 | |
123 | stmt = create_phi_node (vb, loop->header); | |
124 | SSA_NAME_DEF_STMT (vb) = stmt; | |
d2e398df KH |
125 | add_phi_arg (stmt, initial, loop_preheader_edge (loop)); |
126 | add_phi_arg (stmt, va, loop_latch_edge (loop)); | |
82b85a85 ZD |
127 | } |
128 | ||
c913f08a ZD |
129 | /* Add exit phis for the USE on EXIT. */ |
130 | ||
131 | static void | |
132 | add_exit_phis_edge (basic_block exit, tree use) | |
133 | { | |
726a989a RB |
134 | gimple phi, def_stmt = SSA_NAME_DEF_STMT (use); |
135 | basic_block def_bb = gimple_bb (def_stmt); | |
c913f08a ZD |
136 | struct loop *def_loop; |
137 | edge e; | |
628f6a4e | 138 | edge_iterator ei; |
c913f08a ZD |
139 | |
140 | /* Check that some of the edges entering the EXIT block exits a loop in | |
141 | that USE is defined. */ | |
628f6a4e | 142 | FOR_EACH_EDGE (e, ei, exit->preds) |
c913f08a ZD |
143 | { |
144 | def_loop = find_common_loop (def_bb->loop_father, e->src->loop_father); | |
145 | if (!flow_bb_inside_loop_p (def_loop, e->dest)) | |
146 | break; | |
147 | } | |
148 | ||
149 | if (!e) | |
150 | return; | |
151 | ||
152 | phi = create_phi_node (use, exit); | |
726a989a RB |
153 | create_new_def_for (gimple_phi_result (phi), phi, |
154 | gimple_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; |
726a989a | 167 | basic_block def_bb = gimple_bb (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 | 245 | ver = SSA_NAME_VERSION (use); |
726a989a | 246 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (use)); |
c913f08a ZD |
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 | |
726a989a | 273 | find_uses_to_rename_stmt (gimple stmt, bitmap *use_blocks, bitmap need_phis) |
c913f08a | 274 | { |
4c124b4c AM |
275 | ssa_op_iter iter; |
276 | tree var; | |
726a989a | 277 | basic_block bb = gimple_bb (stmt); |
c913f08a | 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 | { |
726a989a | 291 | gimple_stmt_iterator bsi; |
2b271002 ZD |
292 | edge e; |
293 | edge_iterator ei; | |
c913f08a | 294 | |
2b271002 | 295 | FOR_EACH_EDGE (e, ei, bb->succs) |
726a989a RB |
296 | for (bsi = gsi_start_phis (e->dest); !gsi_end_p (bsi); gsi_next (&bsi)) |
297 | find_uses_to_rename_use (bb, PHI_ARG_DEF_FROM_EDGE (gsi_stmt (bsi), e), | |
84d65814 | 298 | use_blocks, need_phis); |
2b271002 | 299 | |
726a989a RB |
300 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
301 | find_uses_to_rename_stmt (gsi_stmt (bsi), use_blocks, need_phis); | |
2b271002 ZD |
302 | } |
303 | ||
304 | /* Marks names that are used outside of the loop they are defined in | |
305 | for rewrite. Records the set of blocks in that the ssa | |
306 | names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL, | |
307 | scan only blocks in this set. */ | |
308 | ||
309 | static void | |
84d65814 | 310 | find_uses_to_rename (bitmap changed_bbs, bitmap *use_blocks, bitmap need_phis) |
2b271002 ZD |
311 | { |
312 | basic_block bb; | |
313 | unsigned index; | |
314 | bitmap_iterator bi; | |
c913f08a | 315 | |
84d65814 | 316 | if (changed_bbs && !bitmap_empty_p (changed_bbs)) |
2b271002 ZD |
317 | { |
318 | EXECUTE_IF_SET_IN_BITMAP (changed_bbs, 0, index, bi) | |
319 | { | |
84d65814 | 320 | find_uses_to_rename_bb (BASIC_BLOCK (index), use_blocks, need_phis); |
2b271002 ZD |
321 | } |
322 | } | |
323 | else | |
324 | { | |
325 | FOR_EACH_BB (bb) | |
326 | { | |
84d65814 | 327 | find_uses_to_rename_bb (bb, use_blocks, need_phis); |
2b271002 | 328 | } |
c913f08a ZD |
329 | } |
330 | } | |
331 | ||
332 | /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra | |
333 | phi nodes to ensure that no variable is used outside the loop it is | |
334 | defined in. | |
335 | ||
336 | This strengthening of the basic ssa form has several advantages: | |
337 | ||
338 | 1) Updating it during unrolling/peeling/versioning is trivial, since | |
339 | we do not need to care about the uses outside of the loop. | |
340 | 2) The behavior of all uses of an induction variable is the same. | |
341 | Without this, you need to distinguish the case when the variable | |
342 | is used outside of the loop it is defined in, for example | |
343 | ||
344 | for (i = 0; i < 100; i++) | |
345 | { | |
346 | for (j = 0; j < 100; j++) | |
347 | { | |
348 | k = i + j; | |
349 | use1 (k); | |
350 | } | |
351 | use2 (k); | |
352 | } | |
353 | ||
354 | Looking from the outer loop with the normal SSA form, the first use of k | |
355 | is not well-behaved, while the second one is an induction variable with | |
2b271002 ZD |
356 | base 99 and step 1. |
357 | ||
358 | If CHANGED_BBS is not NULL, we look for uses outside loops only in | |
84d65814 DN |
359 | the basic blocks in this set. |
360 | ||
361 | UPDATE_FLAG is used in the call to update_ssa. See | |
362 | TODO_update_ssa* for documentation. */ | |
c913f08a ZD |
363 | |
364 | void | |
84d65814 | 365 | rewrite_into_loop_closed_ssa (bitmap changed_bbs, unsigned update_flag) |
c913f08a | 366 | { |
c7b852c8 | 367 | bitmap loop_exits; |
c913f08a | 368 | bitmap *use_blocks; |
84d65814 | 369 | unsigned i, old_num_ssa_names; |
c7b852c8 ZD |
370 | bitmap names_to_rename; |
371 | ||
f87000d0 | 372 | loops_state_set (LOOP_CLOSED_SSA); |
d51157de | 373 | if (number_of_loops () <= 1) |
c7b852c8 ZD |
374 | return; |
375 | ||
376 | loop_exits = get_loops_exits (); | |
377 | names_to_rename = BITMAP_ALLOC (NULL); | |
c913f08a | 378 | |
84d65814 DN |
379 | /* If the pass has caused the SSA form to be out-of-date, update it |
380 | now. */ | |
381 | update_ssa (update_flag); | |
c913f08a | 382 | |
84d65814 | 383 | old_num_ssa_names = num_ssa_names; |
5ed6ace5 | 384 | use_blocks = XCNEWVEC (bitmap, old_num_ssa_names); |
c913f08a ZD |
385 | |
386 | /* Find the uses outside loops. */ | |
84d65814 | 387 | find_uses_to_rename (changed_bbs, use_blocks, names_to_rename); |
2b271002 | 388 | |
84d65814 | 389 | /* Add the PHI nodes on exits of the loops for the names we need to |
c913f08a | 390 | rewrite. */ |
c913f08a ZD |
391 | add_exit_phis (names_to_rename, use_blocks, loop_exits); |
392 | ||
84d65814 | 393 | for (i = 0; i < old_num_ssa_names; i++) |
8bdbfff5 | 394 | BITMAP_FREE (use_blocks[i]); |
c913f08a | 395 | free (use_blocks); |
8bdbfff5 NS |
396 | BITMAP_FREE (loop_exits); |
397 | BITMAP_FREE (names_to_rename); | |
c913f08a | 398 | |
84d65814 DN |
399 | /* Fix up all the names found to be used outside their original |
400 | loops. */ | |
401 | update_ssa (TODO_update_ssa); | |
c913f08a ZD |
402 | } |
403 | ||
404 | /* Check invariants of the loop closed ssa form for the USE in BB. */ | |
405 | ||
406 | static void | |
407 | check_loop_closed_ssa_use (basic_block bb, tree use) | |
408 | { | |
726a989a | 409 | gimple def; |
c913f08a ZD |
410 | basic_block def_bb; |
411 | ||
84d65814 | 412 | if (TREE_CODE (use) != SSA_NAME || !is_gimple_reg (use)) |
c913f08a ZD |
413 | return; |
414 | ||
415 | def = SSA_NAME_DEF_STMT (use); | |
726a989a | 416 | def_bb = gimple_bb (def); |
1e128c5f GB |
417 | gcc_assert (!def_bb |
418 | || flow_bb_inside_loop_p (def_bb->loop_father, bb)); | |
c913f08a ZD |
419 | } |
420 | ||
421 | /* Checks invariants of loop closed ssa form in statement STMT in BB. */ | |
422 | ||
423 | static void | |
726a989a | 424 | check_loop_closed_ssa_stmt (basic_block bb, gimple stmt) |
c913f08a | 425 | { |
4c124b4c AM |
426 | ssa_op_iter iter; |
427 | tree var; | |
c913f08a | 428 | |
38635499 | 429 | FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_ALL_USES) |
4c124b4c | 430 | check_loop_closed_ssa_use (bb, var); |
c913f08a ZD |
431 | } |
432 | ||
433 | /* Checks that invariants of the loop closed ssa form are preserved. */ | |
434 | ||
435 | void | |
436 | verify_loop_closed_ssa (void) | |
437 | { | |
438 | basic_block bb; | |
726a989a RB |
439 | gimple_stmt_iterator bsi; |
440 | gimple phi; | |
441 | edge e; | |
442 | edge_iterator ei; | |
c913f08a | 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 | { | |
726a989a RB |
451 | for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
452 | { | |
453 | phi = gsi_stmt (bsi); | |
454 | FOR_EACH_EDGE (e, ei, bb->preds) | |
455 | check_loop_closed_ssa_use (e->src, | |
456 | PHI_ARG_DEF_FROM_EDGE (phi, e)); | |
457 | } | |
c913f08a | 458 | |
726a989a RB |
459 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
460 | check_loop_closed_ssa_stmt (bb, gsi_stmt (bsi)); | |
c913f08a ZD |
461 | } |
462 | } | |
8b11a64c ZD |
463 | |
464 | /* Split loop exit edge EXIT. The things are a bit complicated by a need to | |
5f40b3cb | 465 | preserve the loop closed ssa form. The newly created block is returned. */ |
8b11a64c | 466 | |
5f40b3cb | 467 | basic_block |
8b11a64c ZD |
468 | split_loop_exit_edge (edge exit) |
469 | { | |
470 | basic_block dest = exit->dest; | |
598ec7bd | 471 | basic_block bb = split_edge (exit); |
726a989a RB |
472 | gimple phi, new_phi; |
473 | tree new_name, name; | |
8b11a64c | 474 | use_operand_p op_p; |
726a989a | 475 | gimple_stmt_iterator psi; |
8b11a64c | 476 | |
726a989a | 477 | for (psi = gsi_start_phis (dest); !gsi_end_p (psi); gsi_next (&psi)) |
8b11a64c | 478 | { |
726a989a | 479 | phi = gsi_stmt (psi); |
c5cbcccf | 480 | op_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (bb)); |
8b11a64c | 481 | |
7fac6722 ZD |
482 | name = USE_FROM_PTR (op_p); |
483 | ||
38635499 | 484 | /* If the argument of the PHI node is a constant, we do not need |
7fac6722 ZD |
485 | to keep it inside loop. */ |
486 | if (TREE_CODE (name) != SSA_NAME) | |
487 | continue; | |
488 | ||
489 | /* Otherwise create an auxiliary phi node that will copy the value | |
38635499 | 490 | of the SSA name out of the loop. */ |
7fac6722 | 491 | new_name = duplicate_ssa_name (name, NULL); |
8b11a64c ZD |
492 | new_phi = create_phi_node (new_name, bb); |
493 | SSA_NAME_DEF_STMT (new_name) = new_phi; | |
d2e398df | 494 | add_phi_arg (new_phi, name, exit); |
8b11a64c ZD |
495 | SET_USE (op_p, new_name); |
496 | } | |
5f40b3cb ZD |
497 | |
498 | return bb; | |
8b11a64c ZD |
499 | } |
500 | ||
8b11a64c ZD |
501 | /* Returns the basic block in that statements should be emitted for induction |
502 | variables incremented at the end of the LOOP. */ | |
503 | ||
504 | basic_block | |
505 | ip_end_pos (struct loop *loop) | |
506 | { | |
507 | return loop->latch; | |
508 | } | |
509 | ||
510 | /* Returns the basic block in that statements should be emitted for induction | |
511 | variables incremented just before exit condition of a LOOP. */ | |
512 | ||
513 | basic_block | |
514 | ip_normal_pos (struct loop *loop) | |
515 | { | |
726a989a | 516 | gimple last; |
8b11a64c ZD |
517 | basic_block bb; |
518 | edge exit; | |
519 | ||
c5cbcccf | 520 | if (!single_pred_p (loop->latch)) |
8b11a64c ZD |
521 | return NULL; |
522 | ||
c5cbcccf | 523 | bb = single_pred (loop->latch); |
8b11a64c | 524 | last = last_stmt (bb); |
ae2cf11b | 525 | if (!last |
726a989a | 526 | || gimple_code (last) != GIMPLE_COND) |
8b11a64c ZD |
527 | return NULL; |
528 | ||
628f6a4e | 529 | exit = EDGE_SUCC (bb, 0); |
8b11a64c | 530 | if (exit->dest == loop->latch) |
628f6a4e | 531 | exit = EDGE_SUCC (bb, 1); |
8b11a64c ZD |
532 | |
533 | if (flow_bb_inside_loop_p (loop, exit->dest)) | |
534 | return NULL; | |
535 | ||
536 | return bb; | |
537 | } | |
538 | ||
539 | /* Stores the standard position for induction variable increment in LOOP | |
540 | (just before the exit condition if it is available and latch block is empty, | |
541 | end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if | |
542 | the increment should be inserted after *BSI. */ | |
543 | ||
544 | void | |
726a989a | 545 | standard_iv_increment_position (struct loop *loop, gimple_stmt_iterator *bsi, |
8b11a64c ZD |
546 | bool *insert_after) |
547 | { | |
548 | basic_block bb = ip_normal_pos (loop), latch = ip_end_pos (loop); | |
726a989a | 549 | gimple last = last_stmt (latch); |
8b11a64c ZD |
550 | |
551 | if (!bb | |
726a989a | 552 | || (last && gimple_code (last) != GIMPLE_LABEL)) |
8b11a64c | 553 | { |
726a989a | 554 | *bsi = gsi_last_bb (latch); |
8b11a64c ZD |
555 | *insert_after = true; |
556 | } | |
557 | else | |
558 | { | |
726a989a | 559 | *bsi = gsi_last_bb (bb); |
8b11a64c ZD |
560 | *insert_after = false; |
561 | } | |
562 | } | |
92fc4a2f ZD |
563 | |
564 | /* Copies phi node arguments for duplicated blocks. The index of the first | |
565 | duplicated block is FIRST_NEW_BLOCK. */ | |
566 | ||
567 | static void | |
568 | copy_phi_node_args (unsigned first_new_block) | |
569 | { | |
570 | unsigned 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 | for (i = first_new_block; i < (unsigned) last_basic_block; i++) | |
576 | add_phi_args_after_copy_bb (BASIC_BLOCK (i)); | |
577 | ||
578 | for (i = first_new_block; i < (unsigned) last_basic_block; i++) | |
6580ee77 | 579 | BASIC_BLOCK (i)->flags &= ~BB_DUPLICATED; |
92fc4a2f ZD |
580 | } |
581 | ||
92fc4a2f | 582 | |
84d65814 DN |
583 | /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also |
584 | updates the PHI nodes at start of the copied region. In order to | |
585 | achieve this, only loops whose exits all lead to the same location | |
586 | are handled. | |
92fc4a2f | 587 | |
84d65814 DN |
588 | Notice that we do not completely update the SSA web after |
589 | duplication. The caller is responsible for calling update_ssa | |
590 | after the loop has been duplicated. */ | |
92fc4a2f ZD |
591 | |
592 | bool | |
726a989a | 593 | gimple_duplicate_loop_to_header_edge (struct loop *loop, edge e, |
92fc4a2f | 594 | unsigned int ndupl, sbitmap wont_exit, |
ee8c1b05 ZD |
595 | edge orig, VEC (edge, heap) **to_remove, |
596 | int flags) | |
92fc4a2f ZD |
597 | { |
598 | unsigned first_new_block; | |
92fc4a2f | 599 | |
f87000d0 | 600 | if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES)) |
92fc4a2f | 601 | return false; |
f87000d0 | 602 | if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS)) |
92fc4a2f ZD |
603 | return false; |
604 | ||
605 | #ifdef ENABLE_CHECKING | |
d6e840ee RG |
606 | if (loops_state_satisfies_p (LOOP_CLOSED_SSA)) |
607 | verify_loop_closed_ssa (); | |
92fc4a2f ZD |
608 | #endif |
609 | ||
92fc4a2f | 610 | first_new_block = last_basic_block; |
d73be268 | 611 | if (!duplicate_loop_to_header_edge (loop, e, ndupl, wont_exit, |
ee8c1b05 | 612 | orig, to_remove, flags)) |
92fc4a2f ZD |
613 | return false; |
614 | ||
615 | /* Readd the removed phi args for e. */ | |
71882046 | 616 | flush_pending_stmts (e); |
92fc4a2f ZD |
617 | |
618 | /* Copy the phi node arguments. */ | |
619 | copy_phi_node_args (first_new_block); | |
620 | ||
92fc4a2f | 621 | scev_reset (); |
92fc4a2f ZD |
622 | |
623 | return true; | |
624 | } | |
17684618 | 625 | |
17684618 ZD |
626 | /* Returns true if we can unroll LOOP FACTOR times. Number |
627 | of iterations of the loop is returned in NITER. */ | |
628 | ||
629 | bool | |
630 | can_unroll_loop_p (struct loop *loop, unsigned factor, | |
631 | struct tree_niter_desc *niter) | |
632 | { | |
633 | edge exit; | |
634 | ||
635 | /* Check whether unrolling is possible. We only want to unroll loops | |
636 | for that we are able to determine number of iterations. We also | |
637 | want to split the extra iterations of the loop from its end, | |
638 | therefore we require that the loop has precisely one | |
639 | exit. */ | |
640 | ||
641 | exit = single_dom_exit (loop); | |
642 | if (!exit) | |
643 | return false; | |
644 | ||
645 | if (!number_of_iterations_exit (loop, exit, niter, false) | |
bf8dbe38 ZD |
646 | || niter->cmp == ERROR_MARK |
647 | /* Scalar evolutions analysis might have copy propagated | |
648 | the abnormal ssa names into these expressions, hence | |
2f8e468b | 649 | emitting the computations based on them during loop |
bf8dbe38 ZD |
650 | unrolling might create overlapping life ranges for |
651 | them, and failures in out-of-ssa. */ | |
652 | || contains_abnormal_ssa_name_p (niter->may_be_zero) | |
653 | || contains_abnormal_ssa_name_p (niter->control.base) | |
654 | || contains_abnormal_ssa_name_p (niter->control.step) | |
655 | || contains_abnormal_ssa_name_p (niter->bound)) | |
17684618 ZD |
656 | return false; |
657 | ||
658 | /* And of course, we must be able to duplicate the loop. */ | |
659 | if (!can_duplicate_loop_p (loop)) | |
660 | return false; | |
661 | ||
662 | /* The final loop should be small enough. */ | |
7f9bc51b | 663 | if (tree_num_loop_insns (loop, &eni_size_weights) * factor |
17684618 ZD |
664 | > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS)) |
665 | return false; | |
666 | ||
667 | return true; | |
668 | } | |
669 | ||
670 | /* Determines the conditions that control execution of LOOP unrolled FACTOR | |
671 | times. DESC is number of iterations of LOOP. ENTER_COND is set to | |
672 | condition that must be true if the main loop can be entered. | |
673 | EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing | |
674 | how the exit from the unrolled loop should be controlled. */ | |
675 | ||
676 | static void | |
677 | determine_exit_conditions (struct loop *loop, struct tree_niter_desc *desc, | |
678 | unsigned factor, tree *enter_cond, | |
679 | tree *exit_base, tree *exit_step, | |
680 | enum tree_code *exit_cmp, tree *exit_bound) | |
681 | { | |
726a989a | 682 | gimple_seq stmts; |
17684618 ZD |
683 | tree base = desc->control.base; |
684 | tree step = desc->control.step; | |
685 | tree bound = desc->bound; | |
d24a32a1 | 686 | tree type = TREE_TYPE (step); |
17684618 ZD |
687 | tree bigstep, delta; |
688 | tree min = lower_bound_in_type (type, type); | |
689 | tree max = upper_bound_in_type (type, type); | |
690 | enum tree_code cmp = desc->cmp; | |
691 | tree cond = boolean_true_node, assum; | |
692 | ||
d24a32a1 ZD |
693 | /* For pointers, do the arithmetics in the type of step (sizetype). */ |
694 | base = fold_convert (type, base); | |
695 | bound = fold_convert (type, bound); | |
696 | ||
17684618 ZD |
697 | *enter_cond = boolean_false_node; |
698 | *exit_base = NULL_TREE; | |
699 | *exit_step = NULL_TREE; | |
700 | *exit_cmp = ERROR_MARK; | |
701 | *exit_bound = NULL_TREE; | |
702 | gcc_assert (cmp != ERROR_MARK); | |
703 | ||
704 | /* We only need to be correct when we answer question | |
705 | "Do at least FACTOR more iterations remain?" in the unrolled loop. | |
706 | Thus, transforming BASE + STEP * i <> BOUND to | |
707 | BASE + STEP * i < BOUND is ok. */ | |
708 | if (cmp == NE_EXPR) | |
709 | { | |
710 | if (tree_int_cst_sign_bit (step)) | |
711 | cmp = GT_EXPR; | |
712 | else | |
713 | cmp = LT_EXPR; | |
714 | } | |
715 | else if (cmp == LT_EXPR) | |
716 | { | |
717 | gcc_assert (!tree_int_cst_sign_bit (step)); | |
718 | } | |
719 | else if (cmp == GT_EXPR) | |
720 | { | |
721 | gcc_assert (tree_int_cst_sign_bit (step)); | |
722 | } | |
723 | else | |
724 | gcc_unreachable (); | |
725 | ||
726 | /* The main body of the loop may be entered iff: | |
727 | ||
728 | 1) desc->may_be_zero is false. | |
729 | 2) it is possible to check that there are at least FACTOR iterations | |
730 | of the loop, i.e., BOUND - step * FACTOR does not overflow. | |
731 | 3) # of iterations is at least FACTOR */ | |
732 | ||
6e682d7e | 733 | if (!integer_zerop (desc->may_be_zero)) |
17684618 ZD |
734 | cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
735 | invert_truthvalue (desc->may_be_zero), | |
736 | cond); | |
737 | ||
738 | bigstep = fold_build2 (MULT_EXPR, type, step, | |
739 | build_int_cst_type (type, factor)); | |
740 | delta = fold_build2 (MINUS_EXPR, type, bigstep, step); | |
741 | if (cmp == LT_EXPR) | |
742 | assum = fold_build2 (GE_EXPR, boolean_type_node, | |
743 | bound, | |
744 | fold_build2 (PLUS_EXPR, type, min, delta)); | |
745 | else | |
746 | assum = fold_build2 (LE_EXPR, boolean_type_node, | |
747 | bound, | |
748 | fold_build2 (PLUS_EXPR, type, max, delta)); | |
749 | cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond); | |
750 | ||
751 | bound = fold_build2 (MINUS_EXPR, type, bound, delta); | |
752 | assum = fold_build2 (cmp, boolean_type_node, base, bound); | |
753 | cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond); | |
754 | ||
755 | cond = force_gimple_operand (unshare_expr (cond), &stmts, false, NULL_TREE); | |
756 | if (stmts) | |
726a989a | 757 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
17684618 ZD |
758 | /* cond now may be a gimple comparison, which would be OK, but also any |
759 | other gimple rhs (say a && b). In this case we need to force it to | |
760 | operand. */ | |
761 | if (!is_gimple_condexpr (cond)) | |
762 | { | |
763 | cond = force_gimple_operand (cond, &stmts, true, NULL_TREE); | |
764 | if (stmts) | |
726a989a | 765 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
17684618 ZD |
766 | } |
767 | *enter_cond = cond; | |
768 | ||
769 | base = force_gimple_operand (unshare_expr (base), &stmts, true, NULL_TREE); | |
770 | if (stmts) | |
726a989a | 771 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
17684618 ZD |
772 | bound = force_gimple_operand (unshare_expr (bound), &stmts, true, NULL_TREE); |
773 | if (stmts) | |
726a989a | 774 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
17684618 ZD |
775 | |
776 | *exit_base = base; | |
777 | *exit_step = bigstep; | |
778 | *exit_cmp = cmp; | |
779 | *exit_bound = bound; | |
780 | } | |
781 | ||
14fa2cc0 ZD |
782 | /* Scales the frequencies of all basic blocks in LOOP that are strictly |
783 | dominated by BB by NUM/DEN. */ | |
784 | ||
785 | static void | |
786 | scale_dominated_blocks_in_loop (struct loop *loop, basic_block bb, | |
787 | int num, int den) | |
788 | { | |
789 | basic_block son; | |
790 | ||
791 | if (den == 0) | |
792 | return; | |
793 | ||
794 | for (son = first_dom_son (CDI_DOMINATORS, bb); | |
795 | son; | |
796 | son = next_dom_son (CDI_DOMINATORS, son)) | |
797 | { | |
798 | if (!flow_bb_inside_loop_p (loop, son)) | |
799 | continue; | |
800 | scale_bbs_frequencies_int (&son, 1, num, den); | |
801 | scale_dominated_blocks_in_loop (loop, son, num, den); | |
802 | } | |
803 | } | |
804 | ||
d73be268 ZD |
805 | /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP. |
806 | EXIT is the exit of the loop to that DESC corresponds. | |
807 | ||
17684618 ZD |
808 | If N is number of iterations of the loop and MAY_BE_ZERO is the condition |
809 | under that loop exits in the first iteration even if N != 0, | |
810 | ||
811 | while (1) | |
812 | { | |
813 | x = phi (init, next); | |
814 | ||
815 | pre; | |
816 | if (st) | |
817 | break; | |
818 | post; | |
819 | } | |
820 | ||
821 | becomes (with possibly the exit conditions formulated a bit differently, | |
822 | avoiding the need to create a new iv): | |
823 | ||
824 | if (MAY_BE_ZERO || N < FACTOR) | |
825 | goto rest; | |
826 | ||
827 | do | |
828 | { | |
829 | x = phi (init, next); | |
830 | ||
831 | pre; | |
832 | post; | |
833 | pre; | |
834 | post; | |
835 | ... | |
836 | pre; | |
837 | post; | |
838 | N -= FACTOR; | |
839 | ||
840 | } while (N >= FACTOR); | |
841 | ||
842 | rest: | |
843 | init' = phi (init, x); | |
844 | ||
845 | while (1) | |
846 | { | |
847 | x = phi (init', next); | |
848 | ||
849 | pre; | |
850 | if (st) | |
851 | break; | |
852 | post; | |
567b96ed ZD |
853 | } |
854 | ||
855 | Before the loop is unrolled, TRANSFORM is called for it (only for the | |
856 | unrolled loop, but not for its versioned copy). DATA is passed to | |
857 | TRANSFORM. */ | |
17684618 | 858 | |
03cb2019 ZD |
859 | /* Probability in % that the unrolled loop is entered. Just a guess. */ |
860 | #define PROB_UNROLLED_LOOP_ENTERED 90 | |
861 | ||
17684618 | 862 | void |
567b96ed ZD |
863 | tree_transform_and_unroll_loop (struct loop *loop, unsigned factor, |
864 | edge exit, struct tree_niter_desc *desc, | |
865 | transform_callback transform, | |
866 | void *data) | |
17684618 | 867 | { |
726a989a RB |
868 | gimple exit_if; |
869 | tree ctr_before, ctr_after; | |
17684618 ZD |
870 | tree enter_main_cond, exit_base, exit_step, exit_bound; |
871 | enum tree_code exit_cmp; | |
726a989a RB |
872 | gimple phi_old_loop, phi_new_loop, phi_rest; |
873 | gimple_stmt_iterator psi_old_loop, psi_new_loop; | |
874 | tree init, next, new_init, var; | |
17684618 ZD |
875 | struct loop *new_loop; |
876 | basic_block rest, exit_bb; | |
877 | edge old_entry, new_entry, old_latch, precond_edge, new_exit; | |
14fa2cc0 | 878 | edge new_nonexit, e; |
726a989a | 879 | gimple_stmt_iterator bsi; |
17684618 ZD |
880 | use_operand_p op; |
881 | bool ok; | |
03cb2019 | 882 | unsigned est_niter, prob_entry, scale_unrolled, scale_rest, freq_e, freq_h; |
14fa2cc0 | 883 | unsigned new_est_niter, i, prob; |
8e08deeb | 884 | unsigned irr = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP; |
17684618 | 885 | sbitmap wont_exit; |
14fa2cc0 | 886 | VEC (edge, heap) *to_remove = NULL; |
17684618 ZD |
887 | |
888 | est_niter = expected_loop_iterations (loop); | |
889 | determine_exit_conditions (loop, desc, factor, | |
890 | &enter_main_cond, &exit_base, &exit_step, | |
891 | &exit_cmp, &exit_bound); | |
892 | ||
03cb2019 ZD |
893 | /* Let us assume that the unrolled loop is quite likely to be entered. */ |
894 | if (integer_nonzerop (enter_main_cond)) | |
895 | prob_entry = REG_BR_PROB_BASE; | |
896 | else | |
897 | prob_entry = PROB_UNROLLED_LOOP_ENTERED * REG_BR_PROB_BASE / 100; | |
898 | ||
899 | /* The values for scales should keep profile consistent, and somewhat close | |
900 | to correct. | |
901 | ||
902 | TODO: The current value of SCALE_REST makes it appear that the loop that | |
903 | is created by splitting the remaining iterations of the unrolled loop is | |
904 | executed the same number of times as the original loop, and with the same | |
905 | frequencies, which is obviously wrong. This does not appear to cause | |
906 | problems, so we do not bother with fixing it for now. To make the profile | |
907 | correct, we would need to change the probability of the exit edge of the | |
908 | loop, and recompute the distribution of frequencies in its body because | |
909 | of this change (scale the frequencies of blocks before and after the exit | |
910 | by appropriate factors). */ | |
911 | scale_unrolled = prob_entry; | |
912 | scale_rest = REG_BR_PROB_BASE; | |
913 | ||
914 | new_loop = loop_version (loop, enter_main_cond, NULL, | |
915 | prob_entry, scale_unrolled, scale_rest, true); | |
17684618 ZD |
916 | gcc_assert (new_loop != NULL); |
917 | update_ssa (TODO_update_ssa); | |
918 | ||
567b96ed | 919 | /* Determine the probability of the exit edge of the unrolled loop. */ |
03cb2019 ZD |
920 | new_est_niter = est_niter / factor; |
921 | ||
922 | /* Without profile feedback, loops for that we do not know a better estimate | |
923 | are assumed to roll 10 times. When we unroll such loop, it appears to | |
924 | roll too little, and it may even seem to be cold. To avoid this, we | |
925 | ensure that the created loop appears to roll at least 5 times (but at | |
926 | most as many times as before unrolling). */ | |
927 | if (new_est_niter < 5) | |
928 | { | |
929 | if (est_niter < 5) | |
930 | new_est_niter = est_niter; | |
931 | else | |
932 | new_est_niter = 5; | |
933 | } | |
934 | ||
14fa2cc0 ZD |
935 | /* Prepare the cfg and update the phi nodes. Move the loop exit to the |
936 | loop latch (and make its condition dummy, for the moment). */ | |
17684618 ZD |
937 | rest = loop_preheader_edge (new_loop)->src; |
938 | precond_edge = single_pred_edge (rest); | |
598ec7bd | 939 | split_edge (loop_latch_edge (loop)); |
17684618 ZD |
940 | exit_bb = single_pred (loop->latch); |
941 | ||
14fa2cc0 ZD |
942 | /* Since the exit edge will be removed, the frequency of all the blocks |
943 | in the loop that are dominated by it must be scaled by | |
944 | 1 / (1 - exit->probability). */ | |
945 | scale_dominated_blocks_in_loop (loop, exit->src, | |
946 | REG_BR_PROB_BASE, | |
947 | REG_BR_PROB_BASE - exit->probability); | |
948 | ||
726a989a RB |
949 | bsi = gsi_last_bb (exit_bb); |
950 | exit_if = gimple_build_cond (EQ_EXPR, integer_zero_node, | |
951 | integer_zero_node, | |
952 | NULL_TREE, NULL_TREE); | |
a9b77cd1 | 953 | |
726a989a | 954 | gsi_insert_after (&bsi, exit_if, GSI_NEW_STMT); |
8e08deeb | 955 | new_exit = make_edge (exit_bb, rest, EDGE_FALSE_VALUE | irr); |
6270df4c | 956 | rescan_loop_exit (new_exit, true, false); |
14fa2cc0 ZD |
957 | |
958 | /* Set the probability of new exit to the same of the old one. Fix | |
959 | the frequency of the latch block, by scaling it back by | |
960 | 1 - exit->probability. */ | |
961 | new_exit->count = exit->count; | |
962 | new_exit->probability = exit->probability; | |
17684618 | 963 | new_nonexit = single_pred_edge (loop->latch); |
14fa2cc0 | 964 | new_nonexit->probability = REG_BR_PROB_BASE - exit->probability; |
17684618 | 965 | new_nonexit->flags = EDGE_TRUE_VALUE; |
14fa2cc0 ZD |
966 | new_nonexit->count -= exit->count; |
967 | if (new_nonexit->count < 0) | |
968 | new_nonexit->count = 0; | |
969 | scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability, | |
970 | REG_BR_PROB_BASE); | |
17684618 ZD |
971 | |
972 | old_entry = loop_preheader_edge (loop); | |
973 | new_entry = loop_preheader_edge (new_loop); | |
974 | old_latch = loop_latch_edge (loop); | |
726a989a RB |
975 | for (psi_old_loop = gsi_start_phis (loop->header), |
976 | psi_new_loop = gsi_start_phis (new_loop->header); | |
977 | !gsi_end_p (psi_old_loop); | |
978 | gsi_next (&psi_old_loop), gsi_next (&psi_new_loop)) | |
17684618 | 979 | { |
726a989a RB |
980 | phi_old_loop = gsi_stmt (psi_old_loop); |
981 | phi_new_loop = gsi_stmt (psi_new_loop); | |
982 | ||
17684618 ZD |
983 | init = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_entry); |
984 | op = PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop, new_entry); | |
985 | gcc_assert (operand_equal_for_phi_arg_p (init, USE_FROM_PTR (op))); | |
986 | next = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_latch); | |
987 | ||
988 | /* Prefer using original variable as a base for the new ssa name. | |
989 | This is necessary for virtual ops, and useful in order to avoid | |
990 | losing debug info for real ops. */ | |
991 | if (TREE_CODE (next) == SSA_NAME) | |
992 | var = SSA_NAME_VAR (next); | |
993 | else if (TREE_CODE (init) == SSA_NAME) | |
994 | var = SSA_NAME_VAR (init); | |
995 | else | |
996 | { | |
997 | var = create_tmp_var (TREE_TYPE (init), "unrinittmp"); | |
f004ab02 | 998 | add_referenced_var (var); |
17684618 ZD |
999 | } |
1000 | ||
726a989a | 1001 | new_init = make_ssa_name (var, NULL); |
17684618 ZD |
1002 | phi_rest = create_phi_node (new_init, rest); |
1003 | SSA_NAME_DEF_STMT (new_init) = phi_rest; | |
1004 | ||
1005 | add_phi_arg (phi_rest, init, precond_edge); | |
1006 | add_phi_arg (phi_rest, next, new_exit); | |
1007 | SET_USE (op, new_init); | |
1008 | } | |
1009 | ||
14fa2cc0 ZD |
1010 | remove_path (exit); |
1011 | ||
567b96ed ZD |
1012 | /* Transform the loop. */ |
1013 | if (transform) | |
1014 | (*transform) (loop, data); | |
1015 | ||
14fa2cc0 ZD |
1016 | /* Unroll the loop and remove the exits in all iterations except for the |
1017 | last one. */ | |
567b96ed ZD |
1018 | wont_exit = sbitmap_alloc (factor); |
1019 | sbitmap_ones (wont_exit); | |
14fa2cc0 ZD |
1020 | RESET_BIT (wont_exit, factor - 1); |
1021 | ||
726a989a | 1022 | ok = gimple_duplicate_loop_to_header_edge |
567b96ed | 1023 | (loop, loop_latch_edge (loop), factor - 1, |
14fa2cc0 | 1024 | wont_exit, new_exit, &to_remove, DLTHE_FLAG_UPDATE_FREQ); |
567b96ed ZD |
1025 | free (wont_exit); |
1026 | gcc_assert (ok); | |
14fa2cc0 ZD |
1027 | |
1028 | for (i = 0; VEC_iterate (edge, to_remove, i, e); i++) | |
1029 | { | |
1030 | ok = remove_path (e); | |
1031 | gcc_assert (ok); | |
1032 | } | |
1033 | VEC_free (edge, heap, to_remove); | |
567b96ed ZD |
1034 | update_ssa (TODO_update_ssa); |
1035 | ||
1036 | /* Ensure that the frequencies in the loop match the new estimated | |
1037 | number of iterations, and change the probability of the new | |
1038 | exit edge. */ | |
1039 | freq_h = loop->header->frequency; | |
1040 | freq_e = EDGE_FREQUENCY (loop_preheader_edge (loop)); | |
1041 | if (freq_h != 0) | |
1042 | scale_loop_frequencies (loop, freq_e * (new_est_niter + 1), freq_h); | |
1043 | ||
1044 | exit_bb = single_pred (loop->latch); | |
1045 | new_exit = find_edge (exit_bb, rest); | |
1046 | new_exit->count = loop_preheader_edge (loop)->count; | |
1047 | new_exit->probability = REG_BR_PROB_BASE / (new_est_niter + 1); | |
1048 | ||
1049 | rest->count += new_exit->count; | |
1050 | rest->frequency += EDGE_FREQUENCY (new_exit); | |
1051 | ||
1052 | new_nonexit = single_pred_edge (loop->latch); | |
14fa2cc0 | 1053 | prob = new_nonexit->probability; |
567b96ed | 1054 | new_nonexit->probability = REG_BR_PROB_BASE - new_exit->probability; |
14fa2cc0 ZD |
1055 | new_nonexit->count = exit_bb->count - new_exit->count; |
1056 | if (new_nonexit->count < 0) | |
1057 | new_nonexit->count = 0; | |
87621e5f SE |
1058 | if (prob > 0) |
1059 | scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability, | |
1060 | prob); | |
567b96ed | 1061 | |
17684618 ZD |
1062 | /* Finally create the new counter for number of iterations and add the new |
1063 | exit instruction. */ | |
726a989a RB |
1064 | bsi = gsi_last_bb (exit_bb); |
1065 | exit_if = gsi_stmt (bsi); | |
17684618 | 1066 | create_iv (exit_base, exit_step, NULL_TREE, loop, |
567b96ed | 1067 | &bsi, false, &ctr_before, &ctr_after); |
726a989a RB |
1068 | gimple_cond_set_code (exit_if, exit_cmp); |
1069 | gimple_cond_set_lhs (exit_if, ctr_after); | |
1070 | gimple_cond_set_rhs (exit_if, exit_bound); | |
567b96ed | 1071 | update_stmt (exit_if); |
17684618 | 1072 | |
c0493b13 | 1073 | #ifdef ENABLE_CHECKING |
17684618 ZD |
1074 | verify_flow_info (); |
1075 | verify_dominators (CDI_DOMINATORS); | |
d73be268 | 1076 | verify_loop_structure (); |
17684618 | 1077 | verify_loop_closed_ssa (); |
c0493b13 | 1078 | #endif |
17684618 | 1079 | } |
567b96ed ZD |
1080 | |
1081 | /* Wrapper over tree_transform_and_unroll_loop for case we do not | |
1082 | want to transform the loop before unrolling. The meaning | |
1083 | of the arguments is the same as for tree_transform_and_unroll_loop. */ | |
1084 | ||
1085 | void | |
1086 | tree_unroll_loop (struct loop *loop, unsigned factor, | |
1087 | edge exit, struct tree_niter_desc *desc) | |
1088 | { | |
1089 | tree_transform_and_unroll_loop (loop, factor, exit, desc, | |
1090 | NULL, NULL); | |
1091 | } |