]>
Commit | Line | Data |
---|---|---|
6de9cd9a | 1 | /* SSA-PRE for trees. |
23a5b65a | 2 | Copyright (C) 2001-2014 Free Software Foundation, Inc. |
7e6eb623 | 3 | Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher |
b9c5e484 | 4 | <stevenb@suse.de> |
6de9cd9a DN |
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 | |
9dcd6f09 | 10 | the Free Software Foundation; either version 3, or (at your option) |
6de9cd9a DN |
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 | |
9dcd6f09 NC |
19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ | |
33c94679 | 21 | |
6de9cd9a DN |
22 | #include "config.h" |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
25 | #include "tm.h" | |
6de9cd9a | 26 | #include "tree.h" |
6de9cd9a | 27 | #include "basic-block.h" |
cf835838 | 28 | #include "gimple-pretty-print.h" |
6de9cd9a | 29 | #include "tree-inline.h" |
6d8eb96b | 30 | #include "inchash.h" |
2fb9a547 AM |
31 | #include "hash-table.h" |
32 | #include "tree-ssa-alias.h" | |
33 | #include "internal-fn.h" | |
34 | #include "gimple-fold.h" | |
35 | #include "tree-eh.h" | |
36 | #include "gimple-expr.h" | |
37 | #include "is-a.h" | |
18f429e2 | 38 | #include "gimple.h" |
45b0be94 | 39 | #include "gimplify.h" |
5be5c238 | 40 | #include "gimple-iterator.h" |
18f429e2 | 41 | #include "gimplify-me.h" |
442b4905 AM |
42 | #include "gimple-ssa.h" |
43 | #include "tree-cfg.h" | |
44 | #include "tree-phinodes.h" | |
45 | #include "ssa-iterators.h" | |
d8a2d370 | 46 | #include "stringpool.h" |
442b4905 AM |
47 | #include "tree-ssanames.h" |
48 | #include "tree-ssa-loop.h" | |
49 | #include "tree-into-ssa.h" | |
d8a2d370 | 50 | #include "expr.h" |
442b4905 AM |
51 | #include "tree-dfa.h" |
52 | #include "tree-ssa.h" | |
6de9cd9a | 53 | #include "tree-iterator.h" |
6de9cd9a | 54 | #include "alloc-pool.h" |
5039610b | 55 | #include "obstack.h" |
6de9cd9a DN |
56 | #include "tree-pass.h" |
57 | #include "flags.h" | |
7e6eb623 | 58 | #include "langhooks.h" |
0fc6c492 | 59 | #include "cfgloop.h" |
89fb70a3 | 60 | #include "tree-ssa-sccvn.h" |
a8338640 | 61 | #include "tree-scalar-evolution.h" |
f0ed4cfb | 62 | #include "params.h" |
c9145754 | 63 | #include "dbgcnt.h" |
40b178f4 | 64 | #include "domwalk.h" |
e248d83f | 65 | #include "ipa-prop.h" |
744730a4 | 66 | #include "tree-ssa-propagate.h" |
7d0aa05b | 67 | #include "ipa-utils.h" |
33c94679 | 68 | |
7e6eb623 | 69 | /* TODO: |
b9c5e484 | 70 | |
bdee7684 | 71 | 1. Avail sets can be shared by making an avail_find_leader that |
7e6eb623 DB |
72 | walks up the dominator tree and looks in those avail sets. |
73 | This might affect code optimality, it's unclear right now. | |
c90186eb | 74 | 2. Strength reduction can be performed by anticipating expressions |
7e6eb623 | 75 | we can repair later on. |
c90186eb | 76 | 3. We can do back-substitution or smarter value numbering to catch |
0fc6c492 | 77 | commutative expressions split up over multiple statements. |
b9c5e484 | 78 | */ |
7e6eb623 DB |
79 | |
80 | /* For ease of terminology, "expression node" in the below refers to | |
726a989a | 81 | every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs |
07beea0d AH |
82 | represent the actual statement containing the expressions we care about, |
83 | and we cache the value number by putting it in the expression. */ | |
7e6eb623 DB |
84 | |
85 | /* Basic algorithm | |
b9c5e484 | 86 | |
56db793a DB |
87 | First we walk the statements to generate the AVAIL sets, the |
88 | EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the | |
89 | generation of values/expressions by a given block. We use them | |
90 | when computing the ANTIC sets. The AVAIL sets consist of | |
91 | SSA_NAME's that represent values, so we know what values are | |
92 | available in what blocks. AVAIL is a forward dataflow problem. In | |
93 | SSA, values are never killed, so we don't need a kill set, or a | |
94 | fixpoint iteration, in order to calculate the AVAIL sets. In | |
95 | traditional parlance, AVAIL sets tell us the downsafety of the | |
7e6eb623 | 96 | expressions/values. |
b9c5e484 | 97 | |
56db793a DB |
98 | Next, we generate the ANTIC sets. These sets represent the |
99 | anticipatable expressions. ANTIC is a backwards dataflow | |
d75dbccd | 100 | problem. An expression is anticipatable in a given block if it could |
56db793a DB |
101 | be generated in that block. This means that if we had to perform |
102 | an insertion in that block, of the value of that expression, we | |
103 | could. Calculating the ANTIC sets requires phi translation of | |
104 | expressions, because the flow goes backwards through phis. We must | |
105 | iterate to a fixpoint of the ANTIC sets, because we have a kill | |
106 | set. Even in SSA form, values are not live over the entire | |
107 | function, only from their definition point onwards. So we have to | |
108 | remove values from the ANTIC set once we go past the definition | |
109 | point of the leaders that make them up. | |
110 | compute_antic/compute_antic_aux performs this computation. | |
7e6eb623 DB |
111 | |
112 | Third, we perform insertions to make partially redundant | |
113 | expressions fully redundant. | |
114 | ||
115 | An expression is partially redundant (excluding partial | |
116 | anticipation) if: | |
117 | ||
118 | 1. It is AVAIL in some, but not all, of the predecessors of a | |
119 | given block. | |
120 | 2. It is ANTIC in all the predecessors. | |
121 | ||
122 | In order to make it fully redundant, we insert the expression into | |
123 | the predecessors where it is not available, but is ANTIC. | |
d75dbccd DB |
124 | |
125 | For the partial anticipation case, we only perform insertion if it | |
126 | is partially anticipated in some block, and fully available in all | |
127 | of the predecessors. | |
128 | ||
129 | insert/insert_aux/do_regular_insertion/do_partial_partial_insertion | |
130 | performs these steps. | |
7e6eb623 DB |
131 | |
132 | Fourth, we eliminate fully redundant expressions. | |
133 | This is a simple statement walk that replaces redundant | |
070b797d | 134 | calculations with the now available values. */ |
7e6eb623 DB |
135 | |
136 | /* Representations of value numbers: | |
137 | ||
c9145754 DB |
138 | Value numbers are represented by a representative SSA_NAME. We |
139 | will create fake SSA_NAME's in situations where we need a | |
140 | representative but do not have one (because it is a complex | |
141 | expression). In order to facilitate storing the value numbers in | |
142 | bitmaps, and keep the number of wasted SSA_NAME's down, we also | |
143 | associate a value_id with each value number, and create full blown | |
144 | ssa_name's only where we actually need them (IE in operands of | |
145 | existing expressions). | |
146 | ||
147 | Theoretically you could replace all the value_id's with | |
148 | SSA_NAME_VERSION, but this would allocate a large number of | |
149 | SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number. | |
150 | It would also require an additional indirection at each point we | |
151 | use the value id. */ | |
7e6eb623 | 152 | |
b9c5e484 | 153 | /* Representation of expressions on value numbers: |
7e6eb623 | 154 | |
249eb506 | 155 | Expressions consisting of value numbers are represented the same |
c9145754 DB |
156 | way as our VN internally represents them, with an additional |
157 | "pre_expr" wrapping around them in order to facilitate storing all | |
158 | of the expressions in the same sets. */ | |
7e6eb623 | 159 | |
c9145754 | 160 | /* Representation of sets: |
7e6eb623 | 161 | |
c9145754 DB |
162 | The dataflow sets do not need to be sorted in any particular order |
163 | for the majority of their lifetime, are simply represented as two | |
164 | bitmaps, one that keeps track of values present in the set, and one | |
165 | that keeps track of expressions present in the set. | |
7e6eb623 | 166 | |
c9145754 DB |
167 | When we need them in topological order, we produce it on demand by |
168 | transforming the bitmap into an array and sorting it into topo | |
169 | order. */ | |
7e6eb623 | 170 | |
c9145754 DB |
171 | /* Type of expression, used to know which member of the PRE_EXPR union |
172 | is valid. */ | |
7e6eb623 | 173 | |
c9145754 DB |
174 | enum pre_expr_kind |
175 | { | |
176 | NAME, | |
177 | NARY, | |
178 | REFERENCE, | |
179 | CONSTANT | |
180 | }; | |
181 | ||
182 | typedef union pre_expr_union_d | |
183 | { | |
184 | tree name; | |
185 | tree constant; | |
186 | vn_nary_op_t nary; | |
187 | vn_reference_t reference; | |
188 | } pre_expr_union; | |
b9c5e484 | 189 | |
5deac340 | 190 | typedef struct pre_expr_d : typed_noop_remove <pre_expr_d> |
c9145754 DB |
191 | { |
192 | enum pre_expr_kind kind; | |
193 | unsigned int id; | |
194 | pre_expr_union u; | |
5deac340 RG |
195 | |
196 | /* hash_table support. */ | |
5831a5f0 LC |
197 | typedef pre_expr_d value_type; |
198 | typedef pre_expr_d compare_type; | |
5deac340 RG |
199 | static inline hashval_t hash (const pre_expr_d *); |
200 | static inline int equal (const pre_expr_d *, const pre_expr_d *); | |
c9145754 | 201 | } *pre_expr; |
7e6eb623 | 202 | |
c9145754 DB |
203 | #define PRE_EXPR_NAME(e) (e)->u.name |
204 | #define PRE_EXPR_NARY(e) (e)->u.nary | |
205 | #define PRE_EXPR_REFERENCE(e) (e)->u.reference | |
206 | #define PRE_EXPR_CONSTANT(e) (e)->u.constant | |
7e6eb623 | 207 | |
0823efed | 208 | /* Compare E1 and E1 for equality. */ |
7e6eb623 | 209 | |
0823efed | 210 | inline int |
5831a5f0 | 211 | pre_expr_d::equal (const value_type *e1, const compare_type *e2) |
0823efed | 212 | { |
c9145754 DB |
213 | if (e1->kind != e2->kind) |
214 | return false; | |
215 | ||
216 | switch (e1->kind) | |
217 | { | |
218 | case CONSTANT: | |
726a989a RB |
219 | return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1), |
220 | PRE_EXPR_CONSTANT (e2)); | |
c9145754 DB |
221 | case NAME: |
222 | return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2); | |
223 | case NARY: | |
224 | return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2)); | |
225 | case REFERENCE: | |
226 | return vn_reference_eq (PRE_EXPR_REFERENCE (e1), | |
227 | PRE_EXPR_REFERENCE (e2)); | |
228 | default: | |
9708c51d | 229 | gcc_unreachable (); |
c9145754 DB |
230 | } |
231 | } | |
232 | ||
0823efed DN |
233 | /* Hash E. */ |
234 | ||
235 | inline hashval_t | |
5831a5f0 | 236 | pre_expr_d::hash (const value_type *e) |
c9145754 | 237 | { |
c9145754 DB |
238 | switch (e->kind) |
239 | { | |
240 | case CONSTANT: | |
726a989a | 241 | return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e)); |
c9145754 | 242 | case NAME: |
9708c51d | 243 | return SSA_NAME_VERSION (PRE_EXPR_NAME (e)); |
c9145754 | 244 | case NARY: |
85169114 | 245 | return PRE_EXPR_NARY (e)->hashcode; |
c9145754 | 246 | case REFERENCE: |
85169114 | 247 | return PRE_EXPR_REFERENCE (e)->hashcode; |
c9145754 | 248 | default: |
9708c51d | 249 | gcc_unreachable (); |
c9145754 DB |
250 | } |
251 | } | |
83737db2 | 252 | |
c9145754 DB |
253 | /* Next global expression id number. */ |
254 | static unsigned int next_expression_id; | |
070b797d | 255 | |
83737db2 | 256 | /* Mapping from expression to id number we can use in bitmap sets. */ |
9771b263 | 257 | static vec<pre_expr> expressions; |
c203e8a7 | 258 | static hash_table<pre_expr_d> *expression_to_id; |
9771b263 | 259 | static vec<unsigned> name_to_id; |
81def1b7 | 260 | |
83737db2 DB |
261 | /* Allocate an expression id for EXPR. */ |
262 | ||
263 | static inline unsigned int | |
c9145754 | 264 | alloc_expression_id (pre_expr expr) |
6de9cd9a | 265 | { |
0823efed | 266 | struct pre_expr_d **slot; |
83737db2 DB |
267 | /* Make sure we won't overflow. */ |
268 | gcc_assert (next_expression_id + 1 > next_expression_id); | |
c9145754 | 269 | expr->id = next_expression_id++; |
9771b263 | 270 | expressions.safe_push (expr); |
13de9095 RG |
271 | if (expr->kind == NAME) |
272 | { | |
273 | unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr)); | |
9771b263 DN |
274 | /* vec::safe_grow_cleared allocates no headroom. Avoid frequent |
275 | re-allocations by using vec::reserve upfront. There is no | |
276 | vec::quick_grow_cleared unfortunately. */ | |
277 | unsigned old_len = name_to_id.length (); | |
278 | name_to_id.reserve (num_ssa_names - old_len); | |
279 | name_to_id.safe_grow_cleared (num_ssa_names); | |
280 | gcc_assert (name_to_id[version] == 0); | |
281 | name_to_id[version] = expr->id; | |
13de9095 RG |
282 | } |
283 | else | |
284 | { | |
c203e8a7 | 285 | slot = expression_to_id->find_slot (expr, INSERT); |
13de9095 RG |
286 | gcc_assert (!*slot); |
287 | *slot = expr; | |
288 | } | |
83737db2 DB |
289 | return next_expression_id - 1; |
290 | } | |
291 | ||
292 | /* Return the expression id for tree EXPR. */ | |
293 | ||
294 | static inline unsigned int | |
c9145754 DB |
295 | get_expression_id (const pre_expr expr) |
296 | { | |
297 | return expr->id; | |
298 | } | |
299 | ||
300 | static inline unsigned int | |
301 | lookup_expression_id (const pre_expr expr) | |
7e6eb623 | 302 | { |
0823efed | 303 | struct pre_expr_d **slot; |
b71b4522 | 304 | |
13de9095 RG |
305 | if (expr->kind == NAME) |
306 | { | |
307 | unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr)); | |
9771b263 | 308 | if (name_to_id.length () <= version) |
13de9095 | 309 | return 0; |
9771b263 | 310 | return name_to_id[version]; |
13de9095 RG |
311 | } |
312 | else | |
313 | { | |
c203e8a7 | 314 | slot = expression_to_id->find_slot (expr, NO_INSERT); |
13de9095 RG |
315 | if (!slot) |
316 | return 0; | |
317 | return ((pre_expr)*slot)->id; | |
318 | } | |
83737db2 | 319 | } |
b9c5e484 | 320 | |
83737db2 DB |
321 | /* Return the existing expression id for EXPR, or create one if one |
322 | does not exist yet. */ | |
b9c5e484 | 323 | |
83737db2 | 324 | static inline unsigned int |
c9145754 | 325 | get_or_alloc_expression_id (pre_expr expr) |
83737db2 | 326 | { |
c9145754 DB |
327 | unsigned int id = lookup_expression_id (expr); |
328 | if (id == 0) | |
83737db2 | 329 | return alloc_expression_id (expr); |
c9145754 | 330 | return expr->id = id; |
83737db2 DB |
331 | } |
332 | ||
333 | /* Return the expression that has expression id ID */ | |
334 | ||
c9145754 | 335 | static inline pre_expr |
83737db2 DB |
336 | expression_for_id (unsigned int id) |
337 | { | |
9771b263 | 338 | return expressions[id]; |
c5830edf DB |
339 | } |
340 | ||
b71b4522 DB |
341 | /* Free the expression id field in all of our expressions, |
342 | and then destroy the expressions array. */ | |
343 | ||
344 | static void | |
345 | clear_expression_ids (void) | |
346 | { | |
9771b263 | 347 | expressions.release (); |
c9145754 | 348 | } |
b71b4522 | 349 | |
c9145754 DB |
350 | static alloc_pool pre_expr_pool; |
351 | ||
352 | /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */ | |
353 | ||
354 | static pre_expr | |
355 | get_or_alloc_expr_for_name (tree name) | |
356 | { | |
5f5126d6 RG |
357 | struct pre_expr_d expr; |
358 | pre_expr result; | |
c9145754 DB |
359 | unsigned int result_id; |
360 | ||
5f5126d6 RG |
361 | expr.kind = NAME; |
362 | expr.id = 0; | |
363 | PRE_EXPR_NAME (&expr) = name; | |
364 | result_id = lookup_expression_id (&expr); | |
365 | if (result_id != 0) | |
366 | return expression_for_id (result_id); | |
367 | ||
368 | result = (pre_expr) pool_alloc (pre_expr_pool); | |
c9145754 | 369 | result->kind = NAME; |
c9145754 | 370 | PRE_EXPR_NAME (result) = name; |
5f5126d6 | 371 | alloc_expression_id (result); |
c9145754 | 372 | return result; |
b71b4522 DB |
373 | } |
374 | ||
bdee7684 | 375 | /* An unordered bitmap set. One bitmap tracks values, the other, |
8c27b7d4 | 376 | expressions. */ |
bdee7684 DB |
377 | typedef struct bitmap_set |
378 | { | |
5c72d561 JH |
379 | bitmap_head expressions; |
380 | bitmap_head values; | |
bdee7684 DB |
381 | } *bitmap_set_t; |
382 | ||
83737db2 | 383 | #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \ |
c3284718 | 384 | EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi)) |
c9145754 | 385 | |
85169114 | 386 | #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \ |
c3284718 | 387 | EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi)) |
85169114 | 388 | |
c9145754 | 389 | /* Mapping from value id to expressions with that value_id. */ |
9771b263 | 390 | static vec<bitmap> value_expressions; |
83737db2 | 391 | |
6b416da1 | 392 | /* Sets that we need to keep track of. */ |
83737db2 | 393 | typedef struct bb_bitmap_sets |
6de9cd9a | 394 | { |
ca072a31 DB |
395 | /* The EXP_GEN set, which represents expressions/values generated in |
396 | a basic block. */ | |
83737db2 | 397 | bitmap_set_t exp_gen; |
ca072a31 DB |
398 | |
399 | /* The PHI_GEN set, which represents PHI results generated in a | |
400 | basic block. */ | |
6b416da1 | 401 | bitmap_set_t phi_gen; |
ca072a31 | 402 | |
f6fe65dc | 403 | /* The TMP_GEN set, which represents results/temporaries generated |
ca072a31 | 404 | in a basic block. IE the LHS of an expression. */ |
6b416da1 | 405 | bitmap_set_t tmp_gen; |
ca072a31 DB |
406 | |
407 | /* The AVAIL_OUT set, which represents which values are available in | |
408 | a given basic block. */ | |
bdee7684 | 409 | bitmap_set_t avail_out; |
ca072a31 | 410 | |
c90186eb | 411 | /* The ANTIC_IN set, which represents which values are anticipatable |
ca072a31 | 412 | in a given basic block. */ |
83737db2 | 413 | bitmap_set_t antic_in; |
ca072a31 | 414 | |
d75dbccd DB |
415 | /* The PA_IN set, which represents which values are |
416 | partially anticipatable in a given basic block. */ | |
417 | bitmap_set_t pa_in; | |
418 | ||
ca072a31 DB |
419 | /* The NEW_SETS set, which is used during insertion to augment the |
420 | AVAIL_OUT set of blocks with the new insertions performed during | |
421 | the current iteration. */ | |
6b416da1 | 422 | bitmap_set_t new_sets; |
c90186eb | 423 | |
5006671f RG |
424 | /* A cache for value_dies_in_block_x. */ |
425 | bitmap expr_dies; | |
426 | ||
d75dbccd | 427 | /* True if we have visited this block during ANTIC calculation. */ |
a19eb9d2 | 428 | unsigned int visited : 1; |
d75dbccd DB |
429 | |
430 | /* True we have deferred processing this block during ANTIC | |
431 | calculation until its successor is processed. */ | |
432 | unsigned int deferred : 1; | |
a19eb9d2 RG |
433 | |
434 | /* True when the block contains a call that might not return. */ | |
435 | unsigned int contains_may_not_return_call : 1; | |
d75dbccd DB |
436 | } *bb_value_sets_t; |
437 | ||
438 | #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen | |
439 | #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen | |
440 | #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen | |
441 | #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out | |
442 | #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in | |
443 | #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in | |
d75dbccd | 444 | #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets |
5006671f RG |
445 | #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies |
446 | #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited | |
d75dbccd | 447 | #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred |
a19eb9d2 | 448 | #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call |
d75dbccd | 449 | |
c9145754 | 450 | |
83737db2 DB |
451 | /* Basic block list in postorder. */ |
452 | static int *postorder; | |
585d0dc4 | 453 | static int postorder_num; |
6de9cd9a | 454 | |
ca072a31 DB |
455 | /* This structure is used to keep track of statistics on what |
456 | optimization PRE was able to perform. */ | |
7e6eb623 | 457 | static struct |
6de9cd9a | 458 | { |
ca072a31 | 459 | /* The number of RHS computations eliminated by PRE. */ |
7e6eb623 | 460 | int eliminations; |
ca072a31 DB |
461 | |
462 | /* The number of new expressions/temporaries generated by PRE. */ | |
7e6eb623 | 463 | int insertions; |
ca072a31 | 464 | |
d75dbccd DB |
465 | /* The number of inserts found due to partial anticipation */ |
466 | int pa_insert; | |
467 | ||
ca072a31 | 468 | /* The number of new PHI nodes added by PRE. */ |
7e6eb623 DB |
469 | int phis; |
470 | } pre_stats; | |
6de9cd9a | 471 | |
d75dbccd | 472 | static bool do_partial_partial; |
e076271b | 473 | static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int); |
c9145754 DB |
474 | static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr); |
475 | static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr); | |
bdee7684 | 476 | static void bitmap_set_copy (bitmap_set_t, bitmap_set_t); |
c9145754 DB |
477 | static bool bitmap_set_contains_value (bitmap_set_t, unsigned int); |
478 | static void bitmap_insert_into_set (bitmap_set_t, pre_expr); | |
9708c51d RG |
479 | static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, |
480 | unsigned int, bool); | |
bdee7684 | 481 | static bitmap_set_t bitmap_set_new (void); |
726a989a | 482 | static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *, |
e076271b RG |
483 | tree); |
484 | static tree find_or_generate_expression (basic_block, tree, gimple_seq *); | |
de081cfd | 485 | static unsigned int get_expr_value_id (pre_expr); |
6de9cd9a | 486 | |
7e6eb623 DB |
487 | /* We can add and remove elements and entries to and from sets |
488 | and hash tables, so we use alloc pools for them. */ | |
6de9cd9a | 489 | |
bdee7684 | 490 | static alloc_pool bitmap_set_pool; |
7932a3db | 491 | static bitmap_obstack grand_bitmap_obstack; |
6de9cd9a | 492 | |
30fd5881 | 493 | /* Set of blocks with statements that have had their EH properties changed. */ |
53b4bf74 DN |
494 | static bitmap need_eh_cleanup; |
495 | ||
30fd5881 EB |
496 | /* Set of blocks with statements that have had their AB properties changed. */ |
497 | static bitmap need_ab_cleanup; | |
498 | ||
7e6eb623 DB |
499 | /* A three tuple {e, pred, v} used to cache phi translations in the |
500 | phi_translate_table. */ | |
6de9cd9a | 501 | |
5deac340 | 502 | typedef struct expr_pred_trans_d : typed_free_remove<expr_pred_trans_d> |
6de9cd9a | 503 | { |
8c27b7d4 | 504 | /* The expression. */ |
c9145754 | 505 | pre_expr e; |
ca072a31 DB |
506 | |
507 | /* The predecessor block along which we translated the expression. */ | |
7e6eb623 | 508 | basic_block pred; |
ca072a31 DB |
509 | |
510 | /* The value that resulted from the translation. */ | |
c9145754 | 511 | pre_expr v; |
ca072a31 DB |
512 | |
513 | /* The hashcode for the expression, pred pair. This is cached for | |
514 | speed reasons. */ | |
7e6eb623 | 515 | hashval_t hashcode; |
5deac340 RG |
516 | |
517 | /* hash_table support. */ | |
5831a5f0 LC |
518 | typedef expr_pred_trans_d value_type; |
519 | typedef expr_pred_trans_d compare_type; | |
520 | static inline hashval_t hash (const value_type *); | |
521 | static inline int equal (const value_type *, const compare_type *); | |
7e6eb623 | 522 | } *expr_pred_trans_t; |
741ac903 | 523 | typedef const struct expr_pred_trans_d *const_expr_pred_trans_t; |
6de9cd9a | 524 | |
0823efed | 525 | inline hashval_t |
5deac340 | 526 | expr_pred_trans_d::hash (const expr_pred_trans_d *e) |
7e6eb623 | 527 | { |
5deac340 | 528 | return e->hashcode; |
7e6eb623 | 529 | } |
6de9cd9a | 530 | |
0823efed | 531 | inline int |
5831a5f0 LC |
532 | expr_pred_trans_d::equal (const value_type *ve1, |
533 | const compare_type *ve2) | |
7e6eb623 | 534 | { |
7e6eb623 DB |
535 | basic_block b1 = ve1->pred; |
536 | basic_block b2 = ve2->pred; | |
b9c5e484 | 537 | |
ca072a31 DB |
538 | /* If they are not translations for the same basic block, they can't |
539 | be equal. */ | |
7e6eb623 | 540 | if (b1 != b2) |
6de9cd9a | 541 | return false; |
5deac340 | 542 | return pre_expr_d::equal (ve1->e, ve2->e); |
6de9cd9a DN |
543 | } |
544 | ||
0823efed DN |
545 | /* The phi_translate_table caches phi translations for a given |
546 | expression and predecessor. */ | |
c203e8a7 | 547 | static hash_table<expr_pred_trans_d> *phi_translate_table; |
0823efed | 548 | |
c9145754 | 549 | /* Add the tuple mapping from {expression E, basic block PRED} to |
984af6ac | 550 | the phi translation table and return whether it pre-existed. */ |
7e6eb623 | 551 | |
984af6ac RB |
552 | static inline bool |
553 | phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred) | |
7e6eb623 | 554 | { |
0823efed | 555 | expr_pred_trans_t *slot; |
984af6ac RB |
556 | expr_pred_trans_d tem; |
557 | hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e), | |
558 | pred->index); | |
559 | tem.e = e; | |
560 | tem.pred = pred; | |
561 | tem.hashcode = hash; | |
c203e8a7 | 562 | slot = phi_translate_table->find_slot_with_hash (&tem, hash, INSERT); |
984af6ac RB |
563 | if (*slot) |
564 | { | |
565 | *entry = *slot; | |
566 | return true; | |
567 | } | |
c9145754 | 568 | |
984af6ac RB |
569 | *entry = *slot = XNEW (struct expr_pred_trans_d); |
570 | (*entry)->e = e; | |
571 | (*entry)->pred = pred; | |
572 | (*entry)->hashcode = hash; | |
573 | return false; | |
6de9cd9a DN |
574 | } |
575 | ||
ff2ad0f7 | 576 | |
c9145754 | 577 | /* Add expression E to the expression set of value id V. */ |
33c94679 | 578 | |
6bcfb753 | 579 | static void |
c9145754 | 580 | add_to_value (unsigned int v, pre_expr e) |
7e6eb623 | 581 | { |
4d4b1b30 | 582 | bitmap set; |
c9145754 | 583 | |
4d4b1b30 | 584 | gcc_checking_assert (get_expr_value_id (e) == v); |
de081cfd | 585 | |
9771b263 | 586 | if (v >= value_expressions.length ()) |
c9145754 | 587 | { |
9771b263 | 588 | value_expressions.safe_grow_cleared (v + 1); |
c9145754 | 589 | } |
33c94679 | 590 | |
9771b263 | 591 | set = value_expressions[v]; |
c9145754 DB |
592 | if (!set) |
593 | { | |
4d4b1b30 | 594 | set = BITMAP_ALLOC (&grand_bitmap_obstack); |
9771b263 | 595 | value_expressions[v] = set; |
c9145754 | 596 | } |
33c94679 | 597 | |
4d4b1b30 | 598 | bitmap_set_bit (set, get_or_alloc_expression_id (e)); |
7e6eb623 | 599 | } |
6de9cd9a | 600 | |
bdee7684 DB |
601 | /* Create a new bitmap set and return it. */ |
602 | ||
b9c5e484 | 603 | static bitmap_set_t |
bdee7684 DB |
604 | bitmap_set_new (void) |
605 | { | |
e1111e8e | 606 | bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool); |
5c72d561 JH |
607 | bitmap_initialize (&ret->expressions, &grand_bitmap_obstack); |
608 | bitmap_initialize (&ret->values, &grand_bitmap_obstack); | |
bdee7684 DB |
609 | return ret; |
610 | } | |
611 | ||
c9145754 DB |
612 | /* Return the value id for a PRE expression EXPR. */ |
613 | ||
614 | static unsigned int | |
615 | get_expr_value_id (pre_expr expr) | |
616 | { | |
e3815735 | 617 | unsigned int id; |
c9145754 DB |
618 | switch (expr->kind) |
619 | { | |
620 | case CONSTANT: | |
5ba5e8ec | 621 | id = get_constant_value_id (PRE_EXPR_CONSTANT (expr)); |
e3815735 | 622 | break; |
c9145754 | 623 | case NAME: |
e3815735 RB |
624 | id = VN_INFO (PRE_EXPR_NAME (expr))->value_id; |
625 | break; | |
c9145754 | 626 | case NARY: |
e3815735 RB |
627 | id = PRE_EXPR_NARY (expr)->value_id; |
628 | break; | |
c9145754 | 629 | case REFERENCE: |
e3815735 RB |
630 | id = PRE_EXPR_REFERENCE (expr)->value_id; |
631 | break; | |
c9145754 DB |
632 | default: |
633 | gcc_unreachable (); | |
634 | } | |
e3815735 RB |
635 | /* ??? We cannot assert that expr has a value-id (it can be 0), because |
636 | we assign value-ids only to expressions that have a result | |
637 | in set_hashtable_value_ids. */ | |
638 | return id; | |
c9145754 DB |
639 | } |
640 | ||
40b178f4 RG |
641 | /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */ |
642 | ||
643 | static tree | |
644 | sccvn_valnum_from_value_id (unsigned int val) | |
645 | { | |
646 | bitmap_iterator bi; | |
647 | unsigned int i; | |
9771b263 | 648 | bitmap exprset = value_expressions[val]; |
40b178f4 RG |
649 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
650 | { | |
651 | pre_expr vexpr = expression_for_id (i); | |
652 | if (vexpr->kind == NAME) | |
653 | return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum; | |
654 | else if (vexpr->kind == CONSTANT) | |
655 | return PRE_EXPR_CONSTANT (vexpr); | |
656 | } | |
657 | return NULL_TREE; | |
658 | } | |
659 | ||
83737db2 | 660 | /* Remove an expression EXPR from a bitmapped set. */ |
bdee7684 DB |
661 | |
662 | static void | |
c9145754 | 663 | bitmap_remove_from_set (bitmap_set_t set, pre_expr expr) |
bdee7684 | 664 | { |
c9145754 DB |
665 | unsigned int val = get_expr_value_id (expr); |
666 | if (!value_id_constant_p (val)) | |
83737db2 | 667 | { |
5c72d561 JH |
668 | bitmap_clear_bit (&set->values, val); |
669 | bitmap_clear_bit (&set->expressions, get_expression_id (expr)); | |
83737db2 | 670 | } |
bdee7684 | 671 | } |
6de9cd9a | 672 | |
7e6eb623 | 673 | static void |
c9145754 | 674 | bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr, |
9708c51d | 675 | unsigned int val, bool allow_constants) |
7e6eb623 | 676 | { |
c9145754 | 677 | if (allow_constants || !value_id_constant_p (val)) |
7e6eb623 | 678 | { |
c9145754 DB |
679 | /* We specifically expect this and only this function to be able to |
680 | insert constants into a set. */ | |
5c72d561 JH |
681 | bitmap_set_bit (&set->values, val); |
682 | bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr)); | |
7e6eb623 DB |
683 | } |
684 | } | |
6de9cd9a | 685 | |
c9145754 DB |
686 | /* Insert an expression EXPR into a bitmapped set. */ |
687 | ||
688 | static void | |
689 | bitmap_insert_into_set (bitmap_set_t set, pre_expr expr) | |
690 | { | |
9708c51d | 691 | bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false); |
c9145754 DB |
692 | } |
693 | ||
bdee7684 DB |
694 | /* Copy a bitmapped set ORIG, into bitmapped set DEST. */ |
695 | ||
696 | static void | |
697 | bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig) | |
698 | { | |
5c72d561 JH |
699 | bitmap_copy (&dest->expressions, &orig->expressions); |
700 | bitmap_copy (&dest->values, &orig->values); | |
bdee7684 DB |
701 | } |
702 | ||
ff3fdad2 | 703 | |
83737db2 | 704 | /* Free memory used up by SET. */ |
ff3fdad2 | 705 | static void |
83737db2 | 706 | bitmap_set_free (bitmap_set_t set) |
ff3fdad2 | 707 | { |
5c72d561 JH |
708 | bitmap_clear (&set->expressions); |
709 | bitmap_clear (&set->values); | |
ff3fdad2 DB |
710 | } |
711 | ||
ff3fdad2 | 712 | |
83737db2 | 713 | /* Generate an topological-ordered array of bitmap set SET. */ |
ff3fdad2 | 714 | |
9771b263 | 715 | static vec<pre_expr> |
83737db2 | 716 | sorted_array_from_bitmap_set (bitmap_set_t set) |
ff3fdad2 | 717 | { |
85169114 PB |
718 | unsigned int i, j; |
719 | bitmap_iterator bi, bj; | |
9771b263 | 720 | vec<pre_expr> result; |
a7d04a53 RG |
721 | |
722 | /* Pre-allocate roughly enough space for the array. */ | |
9771b263 | 723 | result.create (bitmap_count_bits (&set->values)); |
ff3fdad2 | 724 | |
85169114 PB |
725 | FOR_EACH_VALUE_ID_IN_SET (set, i, bi) |
726 | { | |
727 | /* The number of expressions having a given value is usually | |
728 | relatively small. Thus, rather than making a vector of all | |
729 | the expressions and sorting it by value-id, we walk the values | |
730 | and check in the reverse mapping that tells us what expressions | |
731 | have a given value, to filter those in our set. As a result, | |
732 | the expressions are inserted in value-id order, which means | |
733 | topological order. | |
734 | ||
735 | If this is somehow a significant lose for some cases, we can | |
736 | choose which set to walk based on the set size. */ | |
9771b263 | 737 | bitmap exprset = value_expressions[i]; |
4d4b1b30 | 738 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj) |
85169114 | 739 | { |
5c72d561 | 740 | if (bitmap_bit_p (&set->expressions, j)) |
9771b263 | 741 | result.safe_push (expression_for_id (j)); |
85169114 PB |
742 | } |
743 | } | |
6de9cd9a | 744 | |
83737db2 | 745 | return result; |
7e6eb623 | 746 | } |
6de9cd9a | 747 | |
83737db2 | 748 | /* Perform bitmapped set operation DEST &= ORIG. */ |
6de9cd9a DN |
749 | |
750 | static void | |
83737db2 | 751 | bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig) |
6de9cd9a | 752 | { |
83737db2 DB |
753 | bitmap_iterator bi; |
754 | unsigned int i; | |
6de9cd9a | 755 | |
d75dbccd | 756 | if (dest != orig) |
83737db2 | 757 | { |
5c72d561 JH |
758 | bitmap_head temp; |
759 | bitmap_initialize (&temp, &grand_bitmap_obstack); | |
89fb70a3 | 760 | |
5c72d561 JH |
761 | bitmap_and_into (&dest->values, &orig->values); |
762 | bitmap_copy (&temp, &dest->expressions); | |
763 | EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi) | |
d75dbccd | 764 | { |
c9145754 DB |
765 | pre_expr expr = expression_for_id (i); |
766 | unsigned int value_id = get_expr_value_id (expr); | |
5c72d561 JH |
767 | if (!bitmap_bit_p (&dest->values, value_id)) |
768 | bitmap_clear_bit (&dest->expressions, i); | |
d75dbccd | 769 | } |
5c72d561 | 770 | bitmap_clear (&temp); |
6de9cd9a DN |
771 | } |
772 | } | |
773 | ||
83737db2 | 774 | /* Subtract all values and expressions contained in ORIG from DEST. */ |
7e6eb623 | 775 | |
83737db2 DB |
776 | static bitmap_set_t |
777 | bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig) | |
7e6eb623 | 778 | { |
83737db2 DB |
779 | bitmap_set_t result = bitmap_set_new (); |
780 | bitmap_iterator bi; | |
781 | unsigned int i; | |
b9c5e484 | 782 | |
5c72d561 JH |
783 | bitmap_and_compl (&result->expressions, &dest->expressions, |
784 | &orig->expressions); | |
6de9cd9a | 785 | |
83737db2 DB |
786 | FOR_EACH_EXPR_ID_IN_SET (result, i, bi) |
787 | { | |
c9145754 DB |
788 | pre_expr expr = expression_for_id (i); |
789 | unsigned int value_id = get_expr_value_id (expr); | |
5c72d561 | 790 | bitmap_set_bit (&result->values, value_id); |
83737db2 | 791 | } |
af75a7ea | 792 | |
83737db2 | 793 | return result; |
6b416da1 DB |
794 | } |
795 | ||
d75dbccd DB |
796 | /* Subtract all the values in bitmap set B from bitmap set A. */ |
797 | ||
798 | static void | |
799 | bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b) | |
800 | { | |
801 | unsigned int i; | |
802 | bitmap_iterator bi; | |
5c72d561 | 803 | bitmap_head temp; |
d75dbccd | 804 | |
5c72d561 JH |
805 | bitmap_initialize (&temp, &grand_bitmap_obstack); |
806 | ||
807 | bitmap_copy (&temp, &a->expressions); | |
808 | EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi) | |
d75dbccd | 809 | { |
c9145754 DB |
810 | pre_expr expr = expression_for_id (i); |
811 | if (bitmap_set_contains_value (b, get_expr_value_id (expr))) | |
d75dbccd DB |
812 | bitmap_remove_from_set (a, expr); |
813 | } | |
5c72d561 | 814 | bitmap_clear (&temp); |
d75dbccd DB |
815 | } |
816 | ||
817 | ||
c9145754 | 818 | /* Return true if bitmapped set SET contains the value VALUE_ID. */ |
6de9cd9a | 819 | |
bdee7684 | 820 | static bool |
c9145754 | 821 | bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id) |
6de9cd9a | 822 | { |
c9145754 | 823 | if (value_id_constant_p (value_id)) |
bdee7684 | 824 | return true; |
83737db2 | 825 | |
5c72d561 | 826 | if (!set || bitmap_empty_p (&set->expressions)) |
83737db2 DB |
827 | return false; |
828 | ||
5c72d561 | 829 | return bitmap_bit_p (&set->values, value_id); |
bdee7684 | 830 | } |
7e6eb623 | 831 | |
d75dbccd | 832 | static inline bool |
c9145754 | 833 | bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr) |
d75dbccd | 834 | { |
5c72d561 | 835 | return bitmap_bit_p (&set->expressions, get_expression_id (expr)); |
d75dbccd DB |
836 | } |
837 | ||
bdee7684 | 838 | /* Replace an instance of value LOOKFOR with expression EXPR in SET. */ |
7e6eb623 | 839 | |
bdee7684 | 840 | static void |
c9145754 DB |
841 | bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor, |
842 | const pre_expr expr) | |
bdee7684 | 843 | { |
4d4b1b30 | 844 | bitmap exprset; |
83737db2 DB |
845 | unsigned int i; |
846 | bitmap_iterator bi; | |
847 | ||
c9145754 | 848 | if (value_id_constant_p (lookfor)) |
bdee7684 | 849 | return; |
83737db2 | 850 | |
bdee7684 DB |
851 | if (!bitmap_set_contains_value (set, lookfor)) |
852 | return; | |
e9284566 | 853 | |
6b416da1 DB |
854 | /* The number of expressions having a given value is usually |
855 | significantly less than the total number of expressions in SET. | |
856 | Thus, rather than check, for each expression in SET, whether it | |
857 | has the value LOOKFOR, we walk the reverse mapping that tells us | |
858 | what expressions have a given value, and see if any of those | |
859 | expressions are in our set. For large testcases, this is about | |
860 | 5-10x faster than walking the bitmap. If this is somehow a | |
861 | significant lose for some cases, we can choose which set to walk | |
862 | based on the set size. */ | |
9771b263 | 863 | exprset = value_expressions[lookfor]; |
4d4b1b30 | 864 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
6de9cd9a | 865 | { |
fcaa4ca4 | 866 | if (bitmap_clear_bit (&set->expressions, i)) |
6de9cd9a | 867 | { |
5c72d561 | 868 | bitmap_set_bit (&set->expressions, get_expression_id (expr)); |
83737db2 | 869 | return; |
6de9cd9a | 870 | } |
6de9cd9a | 871 | } |
4bcc5786 RB |
872 | |
873 | gcc_unreachable (); | |
6de9cd9a DN |
874 | } |
875 | ||
83737db2 | 876 | /* Return true if two bitmap sets are equal. */ |
6de9cd9a | 877 | |
7e6eb623 | 878 | static bool |
83737db2 | 879 | bitmap_set_equal (bitmap_set_t a, bitmap_set_t b) |
6de9cd9a | 880 | { |
5c72d561 | 881 | return bitmap_equal_p (&a->values, &b->values); |
6de9cd9a DN |
882 | } |
883 | ||
e9284566 | 884 | /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists, |
6c6cfbfd | 885 | and add it otherwise. */ |
bdee7684 | 886 | |
7e6eb623 | 887 | static void |
c9145754 | 888 | bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr) |
7e6eb623 | 889 | { |
c9145754 | 890 | unsigned int val = get_expr_value_id (expr); |
83737db2 | 891 | |
e9284566 DB |
892 | if (bitmap_set_contains_value (set, val)) |
893 | bitmap_set_replace_value (set, val, expr); | |
894 | else | |
895 | bitmap_insert_into_set (set, expr); | |
bdee7684 | 896 | } |
7e6eb623 | 897 | |
bdee7684 DB |
898 | /* Insert EXPR into SET if EXPR's value is not already present in |
899 | SET. */ | |
900 | ||
901 | static void | |
c9145754 | 902 | bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr) |
bdee7684 | 903 | { |
c9145754 | 904 | unsigned int val = get_expr_value_id (expr); |
af75a7ea | 905 | |
77a74ed7 | 906 | gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr)); |
1befacc8 RG |
907 | |
908 | /* Constant values are always considered to be part of the set. */ | |
909 | if (value_id_constant_p (val)) | |
910 | return; | |
911 | ||
912 | /* If the value membership changed, add the expression. */ | |
5c72d561 JH |
913 | if (bitmap_set_bit (&set->values, val)) |
914 | bitmap_set_bit (&set->expressions, expr->id); | |
7e6eb623 | 915 | } |
6de9cd9a | 916 | |
c9145754 DB |
917 | /* Print out EXPR to outfile. */ |
918 | ||
919 | static void | |
920 | print_pre_expr (FILE *outfile, const pre_expr expr) | |
921 | { | |
922 | switch (expr->kind) | |
923 | { | |
924 | case CONSTANT: | |
925 | print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0); | |
926 | break; | |
927 | case NAME: | |
928 | print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0); | |
929 | break; | |
930 | case NARY: | |
931 | { | |
932 | unsigned int i; | |
933 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); | |
5806f481 | 934 | fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode)); |
c9145754 DB |
935 | for (i = 0; i < nary->length; i++) |
936 | { | |
937 | print_generic_expr (outfile, nary->op[i], 0); | |
938 | if (i != (unsigned) nary->length - 1) | |
939 | fprintf (outfile, ","); | |
940 | } | |
941 | fprintf (outfile, "}"); | |
942 | } | |
943 | break; | |
944 | ||
945 | case REFERENCE: | |
946 | { | |
947 | vn_reference_op_t vro; | |
948 | unsigned int i; | |
949 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); | |
950 | fprintf (outfile, "{"); | |
951 | for (i = 0; | |
9771b263 | 952 | ref->operands.iterate (i, &vro); |
c9145754 DB |
953 | i++) |
954 | { | |
5006671f | 955 | bool closebrace = false; |
c9145754 DB |
956 | if (vro->opcode != SSA_NAME |
957 | && TREE_CODE_CLASS (vro->opcode) != tcc_declaration) | |
5006671f | 958 | { |
5806f481 | 959 | fprintf (outfile, "%s", get_tree_code_name (vro->opcode)); |
5006671f RG |
960 | if (vro->op0) |
961 | { | |
962 | fprintf (outfile, "<"); | |
963 | closebrace = true; | |
964 | } | |
965 | } | |
c9145754 DB |
966 | if (vro->op0) |
967 | { | |
c9145754 DB |
968 | print_generic_expr (outfile, vro->op0, 0); |
969 | if (vro->op1) | |
970 | { | |
971 | fprintf (outfile, ","); | |
972 | print_generic_expr (outfile, vro->op1, 0); | |
973 | } | |
5006671f RG |
974 | if (vro->op2) |
975 | { | |
976 | fprintf (outfile, ","); | |
977 | print_generic_expr (outfile, vro->op2, 0); | |
978 | } | |
c9145754 | 979 | } |
5006671f RG |
980 | if (closebrace) |
981 | fprintf (outfile, ">"); | |
9771b263 | 982 | if (i != ref->operands.length () - 1) |
c9145754 DB |
983 | fprintf (outfile, ","); |
984 | } | |
985 | fprintf (outfile, "}"); | |
5006671f RG |
986 | if (ref->vuse) |
987 | { | |
988 | fprintf (outfile, "@"); | |
989 | print_generic_expr (outfile, ref->vuse, 0); | |
990 | } | |
c9145754 DB |
991 | } |
992 | break; | |
993 | } | |
994 | } | |
995 | void debug_pre_expr (pre_expr); | |
996 | ||
997 | /* Like print_pre_expr but always prints to stderr. */ | |
24e47c76 | 998 | DEBUG_FUNCTION void |
c9145754 DB |
999 | debug_pre_expr (pre_expr e) |
1000 | { | |
1001 | print_pre_expr (stderr, e); | |
1002 | fprintf (stderr, "\n"); | |
1003 | } | |
1004 | ||
bdee7684 DB |
1005 | /* Print out SET to OUTFILE. */ |
1006 | ||
1007 | static void | |
83737db2 DB |
1008 | print_bitmap_set (FILE *outfile, bitmap_set_t set, |
1009 | const char *setname, int blockindex) | |
bdee7684 DB |
1010 | { |
1011 | fprintf (outfile, "%s[%d] := { ", setname, blockindex); | |
1012 | if (set) | |
1013 | { | |
cf6b9ef1 | 1014 | bool first = true; |
3cd8c58a | 1015 | unsigned i; |
87c476a2 ZD |
1016 | bitmap_iterator bi; |
1017 | ||
83737db2 | 1018 | FOR_EACH_EXPR_ID_IN_SET (set, i, bi) |
87c476a2 | 1019 | { |
c9145754 | 1020 | const pre_expr expr = expression_for_id (i); |
83737db2 | 1021 | |
65a6f342 NS |
1022 | if (!first) |
1023 | fprintf (outfile, ", "); | |
1024 | first = false; | |
c9145754 | 1025 | print_pre_expr (outfile, expr); |
b9c5e484 | 1026 | |
c9145754 | 1027 | fprintf (outfile, " (%04d)", get_expr_value_id (expr)); |
87c476a2 | 1028 | } |
bdee7684 DB |
1029 | } |
1030 | fprintf (outfile, " }\n"); | |
1031 | } | |
6de9cd9a | 1032 | |
83737db2 | 1033 | void debug_bitmap_set (bitmap_set_t); |
6de9cd9a | 1034 | |
24e47c76 | 1035 | DEBUG_FUNCTION void |
83737db2 DB |
1036 | debug_bitmap_set (bitmap_set_t set) |
1037 | { | |
1038 | print_bitmap_set (stderr, set, "debug", 0); | |
7e6eb623 DB |
1039 | } |
1040 | ||
19372838 RG |
1041 | void debug_bitmap_sets_for (basic_block); |
1042 | ||
1043 | DEBUG_FUNCTION void | |
1044 | debug_bitmap_sets_for (basic_block bb) | |
1045 | { | |
1046 | print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index); | |
40b178f4 RG |
1047 | print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index); |
1048 | print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index); | |
1049 | print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index); | |
1050 | print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index); | |
1051 | if (do_partial_partial) | |
1052 | print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index); | |
1053 | print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index); | |
19372838 RG |
1054 | } |
1055 | ||
7e6eb623 | 1056 | /* Print out the expressions that have VAL to OUTFILE. */ |
33c94679 | 1057 | |
6bcfb753 | 1058 | static void |
c9145754 | 1059 | print_value_expressions (FILE *outfile, unsigned int val) |
7e6eb623 | 1060 | { |
9771b263 | 1061 | bitmap set = value_expressions[val]; |
c9145754 | 1062 | if (set) |
33c94679 | 1063 | { |
4d4b1b30 | 1064 | bitmap_set x; |
33c94679 | 1065 | char s[10]; |
c9145754 | 1066 | sprintf (s, "%04d", val); |
4d4b1b30 RG |
1067 | x.expressions = *set; |
1068 | print_bitmap_set (outfile, &x, s, 0); | |
33c94679 | 1069 | } |
6de9cd9a DN |
1070 | } |
1071 | ||
6de9cd9a | 1072 | |
24e47c76 | 1073 | DEBUG_FUNCTION void |
c9145754 | 1074 | debug_value_expressions (unsigned int val) |
6de9cd9a | 1075 | { |
7e6eb623 DB |
1076 | print_value_expressions (stderr, val); |
1077 | } | |
1078 | ||
c9145754 DB |
1079 | /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to |
1080 | represent it. */ | |
0995a441 | 1081 | |
c9145754 DB |
1082 | static pre_expr |
1083 | get_or_alloc_expr_for_constant (tree constant) | |
b9c5e484 | 1084 | { |
c9145754 DB |
1085 | unsigned int result_id; |
1086 | unsigned int value_id; | |
5f5126d6 RG |
1087 | struct pre_expr_d expr; |
1088 | pre_expr newexpr; | |
1089 | ||
1090 | expr.kind = CONSTANT; | |
1091 | PRE_EXPR_CONSTANT (&expr) = constant; | |
1092 | result_id = lookup_expression_id (&expr); | |
1093 | if (result_id != 0) | |
1094 | return expression_for_id (result_id); | |
1095 | ||
1096 | newexpr = (pre_expr) pool_alloc (pre_expr_pool); | |
c9145754 DB |
1097 | newexpr->kind = CONSTANT; |
1098 | PRE_EXPR_CONSTANT (newexpr) = constant; | |
5f5126d6 | 1099 | alloc_expression_id (newexpr); |
c9145754 | 1100 | value_id = get_or_alloc_constant_value_id (constant); |
c9145754 DB |
1101 | add_to_value (value_id, newexpr); |
1102 | return newexpr; | |
0995a441 SB |
1103 | } |
1104 | ||
c9145754 DB |
1105 | /* Given a value id V, find the actual tree representing the constant |
1106 | value if there is one, and return it. Return NULL if we can't find | |
1107 | a constant. */ | |
43da81be DB |
1108 | |
1109 | static tree | |
726a989a | 1110 | get_constant_for_value_id (unsigned int v) |
43da81be | 1111 | { |
c9145754 DB |
1112 | if (value_id_constant_p (v)) |
1113 | { | |
1114 | unsigned int i; | |
1115 | bitmap_iterator bi; | |
9771b263 | 1116 | bitmap exprset = value_expressions[v]; |
c9145754 | 1117 | |
4d4b1b30 | 1118 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
c9145754 DB |
1119 | { |
1120 | pre_expr expr = expression_for_id (i); | |
726a989a | 1121 | if (expr->kind == CONSTANT) |
c9145754 DB |
1122 | return PRE_EXPR_CONSTANT (expr); |
1123 | } | |
1124 | } | |
1125 | return NULL; | |
1126 | } | |
1127 | ||
1128 | /* Get or allocate a pre_expr for a piece of GIMPLE, and return it. | |
1129 | Currently only supports constants and SSA_NAMES. */ | |
1130 | static pre_expr | |
1131 | get_or_alloc_expr_for (tree t) | |
1132 | { | |
1133 | if (TREE_CODE (t) == SSA_NAME) | |
1134 | return get_or_alloc_expr_for_name (t); | |
1d65f45c | 1135 | else if (is_gimple_min_invariant (t)) |
c9145754 | 1136 | return get_or_alloc_expr_for_constant (t); |
726a989a RB |
1137 | else |
1138 | { | |
1139 | /* More complex expressions can result from SCCVN expression | |
1140 | simplification that inserts values for them. As they all | |
1141 | do not have VOPs the get handled by the nary ops struct. */ | |
1142 | vn_nary_op_t result; | |
1143 | unsigned int result_id; | |
1144 | vn_nary_op_lookup (t, &result); | |
1145 | if (result != NULL) | |
1146 | { | |
1147 | pre_expr e = (pre_expr) pool_alloc (pre_expr_pool); | |
1148 | e->kind = NARY; | |
1149 | PRE_EXPR_NARY (e) = result; | |
1150 | result_id = lookup_expression_id (e); | |
1151 | if (result_id != 0) | |
1152 | { | |
1153 | pool_free (pre_expr_pool, e); | |
1154 | e = expression_for_id (result_id); | |
1155 | return e; | |
1156 | } | |
1157 | alloc_expression_id (e); | |
1158 | return e; | |
1159 | } | |
1160 | } | |
c9145754 DB |
1161 | return NULL; |
1162 | } | |
1163 | ||
1164 | /* Return the folded version of T if T, when folded, is a gimple | |
1165 | min_invariant. Otherwise, return T. */ | |
1166 | ||
1167 | static pre_expr | |
1168 | fully_constant_expression (pre_expr e) | |
1169 | { | |
1170 | switch (e->kind) | |
1171 | { | |
1172 | case CONSTANT: | |
1173 | return e; | |
1174 | case NARY: | |
1175 | { | |
1176 | vn_nary_op_t nary = PRE_EXPR_NARY (e); | |
1177 | switch (TREE_CODE_CLASS (nary->opcode)) | |
1178 | { | |
1179 | case tcc_binary: | |
40f64141 | 1180 | case tcc_comparison: |
c9145754 DB |
1181 | { |
1182 | /* We have to go from trees to pre exprs to value ids to | |
1183 | constants. */ | |
1184 | tree naryop0 = nary->op[0]; | |
1185 | tree naryop1 = nary->op[1]; | |
40f64141 RG |
1186 | tree result; |
1187 | if (!is_gimple_min_invariant (naryop0)) | |
726a989a RB |
1188 | { |
1189 | pre_expr rep0 = get_or_alloc_expr_for (naryop0); | |
1190 | unsigned int vrep0 = get_expr_value_id (rep0); | |
40f64141 RG |
1191 | tree const0 = get_constant_for_value_id (vrep0); |
1192 | if (const0) | |
1193 | naryop0 = fold_convert (TREE_TYPE (naryop0), const0); | |
726a989a | 1194 | } |
40f64141 | 1195 | if (!is_gimple_min_invariant (naryop1)) |
726a989a RB |
1196 | { |
1197 | pre_expr rep1 = get_or_alloc_expr_for (naryop1); | |
1198 | unsigned int vrep1 = get_expr_value_id (rep1); | |
40f64141 RG |
1199 | tree const1 = get_constant_for_value_id (vrep1); |
1200 | if (const1) | |
1201 | naryop1 = fold_convert (TREE_TYPE (naryop1), const1); | |
b463e8de | 1202 | } |
40f64141 RG |
1203 | result = fold_binary (nary->opcode, nary->type, |
1204 | naryop0, naryop1); | |
c9145754 DB |
1205 | if (result && is_gimple_min_invariant (result)) |
1206 | return get_or_alloc_expr_for_constant (result); | |
40f64141 RG |
1207 | /* We might have simplified the expression to a |
1208 | SSA_NAME for example from x_1 * 1. But we cannot | |
1209 | insert a PHI for x_1 unconditionally as x_1 might | |
1210 | not be available readily. */ | |
c9145754 DB |
1211 | return e; |
1212 | } | |
40f64141 RG |
1213 | case tcc_reference: |
1214 | if (nary->opcode != REALPART_EXPR | |
b8698a0f | 1215 | && nary->opcode != IMAGPART_EXPR |
40f64141 RG |
1216 | && nary->opcode != VIEW_CONVERT_EXPR) |
1217 | return e; | |
1218 | /* Fallthrough. */ | |
c9145754 DB |
1219 | case tcc_unary: |
1220 | { | |
40f64141 RG |
1221 | /* We have to go from trees to pre exprs to value ids to |
1222 | constants. */ | |
c9145754 | 1223 | tree naryop0 = nary->op[0]; |
726a989a RB |
1224 | tree const0, result; |
1225 | if (is_gimple_min_invariant (naryop0)) | |
1226 | const0 = naryop0; | |
1227 | else | |
1228 | { | |
1229 | pre_expr rep0 = get_or_alloc_expr_for (naryop0); | |
1230 | unsigned int vrep0 = get_expr_value_id (rep0); | |
1231 | const0 = get_constant_for_value_id (vrep0); | |
1232 | } | |
1233 | result = NULL; | |
c9145754 | 1234 | if (const0) |
b463e8de DB |
1235 | { |
1236 | tree type1 = TREE_TYPE (nary->op[0]); | |
1237 | const0 = fold_convert (type1, const0); | |
1238 | result = fold_unary (nary->opcode, nary->type, const0); | |
1239 | } | |
c9145754 DB |
1240 | if (result && is_gimple_min_invariant (result)) |
1241 | return get_or_alloc_expr_for_constant (result); | |
1242 | return e; | |
1243 | } | |
1244 | default: | |
1245 | return e; | |
1246 | } | |
1247 | } | |
47af7a5c RG |
1248 | case REFERENCE: |
1249 | { | |
1250 | vn_reference_t ref = PRE_EXPR_REFERENCE (e); | |
12bd5a1e RG |
1251 | tree folded; |
1252 | if ((folded = fully_constant_vn_reference_p (ref))) | |
1253 | return get_or_alloc_expr_for_constant (folded); | |
1254 | return e; | |
1255 | } | |
c9145754 DB |
1256 | default: |
1257 | return e; | |
1258 | } | |
1259 | return e; | |
43da81be DB |
1260 | } |
1261 | ||
5006671f | 1262 | /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that |
84280917 MM |
1263 | it has the value it would have in BLOCK. Set *SAME_VALID to true |
1264 | in case the new vuse doesn't change the value id of the OPERANDS. */ | |
c90186eb | 1265 | |
5006671f | 1266 | static tree |
9771b263 | 1267 | translate_vuse_through_block (vec<vn_reference_op_s> operands, |
b45d2719 | 1268 | alias_set_type set, tree type, tree vuse, |
5006671f | 1269 | basic_block phiblock, |
84280917 | 1270 | basic_block block, bool *same_valid) |
c90186eb | 1271 | { |
5006671f | 1272 | gimple phi = SSA_NAME_DEF_STMT (vuse); |
b45d2719 | 1273 | ao_ref ref; |
84280917 MM |
1274 | edge e = NULL; |
1275 | bool use_oracle; | |
1276 | ||
1277 | *same_valid = true; | |
43da81be | 1278 | |
5006671f RG |
1279 | if (gimple_bb (phi) != phiblock) |
1280 | return vuse; | |
1281 | ||
84280917 | 1282 | use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands); |
5006671f RG |
1283 | |
1284 | /* Use the alias-oracle to find either the PHI node in this block, | |
1285 | the first VUSE used in this block that is equivalent to vuse or | |
1286 | the first VUSE which definition in this block kills the value. */ | |
84280917 MM |
1287 | if (gimple_code (phi) == GIMPLE_PHI) |
1288 | e = find_edge (block, phiblock); | |
1289 | else if (use_oracle) | |
1290 | while (!stmt_may_clobber_ref_p_1 (phi, &ref)) | |
1291 | { | |
1292 | vuse = gimple_vuse (phi); | |
1293 | phi = SSA_NAME_DEF_STMT (vuse); | |
1294 | if (gimple_bb (phi) != phiblock) | |
1295 | return vuse; | |
1296 | if (gimple_code (phi) == GIMPLE_PHI) | |
1297 | { | |
1298 | e = find_edge (block, phiblock); | |
1299 | break; | |
1300 | } | |
1301 | } | |
1302 | else | |
1303 | return NULL_TREE; | |
1304 | ||
1305 | if (e) | |
c90186eb | 1306 | { |
84280917 | 1307 | if (use_oracle) |
5006671f | 1308 | { |
84280917 | 1309 | bitmap visited = NULL; |
9bb06c2a | 1310 | unsigned int cnt; |
84280917 MM |
1311 | /* Try to find a vuse that dominates this phi node by skipping |
1312 | non-clobbering statements. */ | |
50f0aa20 RB |
1313 | vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false, |
1314 | NULL, NULL); | |
84280917 MM |
1315 | if (visited) |
1316 | BITMAP_FREE (visited); | |
5006671f | 1317 | } |
84280917 MM |
1318 | else |
1319 | vuse = NULL_TREE; | |
1320 | if (!vuse) | |
1321 | { | |
1322 | /* If we didn't find any, the value ID can't stay the same, | |
1323 | but return the translated vuse. */ | |
1324 | *same_valid = false; | |
1325 | vuse = PHI_ARG_DEF (phi, e->dest_idx); | |
1326 | } | |
1327 | /* ??? We would like to return vuse here as this is the canonical | |
1328 | upmost vdef that this reference is associated with. But during | |
1329 | insertion of the references into the hash tables we only ever | |
1330 | directly insert with their direct gimple_vuse, hence returning | |
1331 | something else would make us not find the other expression. */ | |
1332 | return PHI_ARG_DEF (phi, e->dest_idx); | |
c90186eb | 1333 | } |
c90186eb | 1334 | |
5006671f | 1335 | return NULL_TREE; |
c90186eb | 1336 | } |
83737db2 | 1337 | |
249eb506 | 1338 | /* Like bitmap_find_leader, but checks for the value existing in SET1 *or* |
83737db2 DB |
1339 | SET2. This is used to avoid making a set consisting of the union |
1340 | of PA_IN and ANTIC_IN during insert. */ | |
1341 | ||
c9145754 DB |
1342 | static inline pre_expr |
1343 | find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2) | |
83737db2 | 1344 | { |
c9145754 | 1345 | pre_expr result; |
83737db2 | 1346 | |
e076271b | 1347 | result = bitmap_find_leader (set1, val); |
83737db2 | 1348 | if (!result && set2) |
e076271b | 1349 | result = bitmap_find_leader (set2, val); |
83737db2 DB |
1350 | return result; |
1351 | } | |
1352 | ||
c9145754 DB |
1353 | /* Get the tree type for our PRE expression e. */ |
1354 | ||
1355 | static tree | |
1356 | get_expr_type (const pre_expr e) | |
1357 | { | |
1358 | switch (e->kind) | |
1359 | { | |
1360 | case NAME: | |
1361 | return TREE_TYPE (PRE_EXPR_NAME (e)); | |
1362 | case CONSTANT: | |
1363 | return TREE_TYPE (PRE_EXPR_CONSTANT (e)); | |
1364 | case REFERENCE: | |
b45d2719 | 1365 | return PRE_EXPR_REFERENCE (e)->type; |
c9145754 DB |
1366 | case NARY: |
1367 | return PRE_EXPR_NARY (e)->type; | |
1368 | } | |
c3284718 | 1369 | gcc_unreachable (); |
c9145754 DB |
1370 | } |
1371 | ||
1372 | /* Get a representative SSA_NAME for a given expression. | |
1373 | Since all of our sub-expressions are treated as values, we require | |
1374 | them to be SSA_NAME's for simplicity. | |
1375 | Prior versions of GVNPRE used to use "value handles" here, so that | |
1376 | an expression would be VH.11 + VH.10 instead of d_3 + e_6. In | |
1377 | either case, the operands are really values (IE we do not expect | |
1378 | them to be usable without finding leaders). */ | |
1379 | ||
1380 | static tree | |
1381 | get_representative_for (const pre_expr e) | |
1382 | { | |
c9145754 DB |
1383 | tree name; |
1384 | unsigned int value_id = get_expr_value_id (e); | |
1385 | ||
1386 | switch (e->kind) | |
1387 | { | |
1388 | case NAME: | |
1389 | return PRE_EXPR_NAME (e); | |
1390 | case CONSTANT: | |
726a989a | 1391 | return PRE_EXPR_CONSTANT (e); |
c9145754 DB |
1392 | case NARY: |
1393 | case REFERENCE: | |
1394 | { | |
1395 | /* Go through all of the expressions representing this value | |
1396 | and pick out an SSA_NAME. */ | |
1397 | unsigned int i; | |
1398 | bitmap_iterator bi; | |
9771b263 | 1399 | bitmap exprs = value_expressions[value_id]; |
4d4b1b30 | 1400 | EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi) |
c9145754 DB |
1401 | { |
1402 | pre_expr rep = expression_for_id (i); | |
1403 | if (rep->kind == NAME) | |
1404 | return PRE_EXPR_NAME (rep); | |
a044f0e7 RB |
1405 | else if (rep->kind == CONSTANT) |
1406 | return PRE_EXPR_CONSTANT (rep); | |
c9145754 DB |
1407 | } |
1408 | } | |
1409 | break; | |
1410 | } | |
a044f0e7 | 1411 | |
c9145754 DB |
1412 | /* If we reached here we couldn't find an SSA_NAME. This can |
1413 | happen when we've discovered a value that has never appeared in | |
a044f0e7 RB |
1414 | the program as set to an SSA_NAME, as the result of phi translation. |
1415 | Create one here. | |
1416 | ??? We should be able to re-use this when we insert the statement | |
1417 | to compute it. */ | |
83d5977e | 1418 | name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp"); |
c9145754 | 1419 | VN_INFO_GET (name)->value_id = value_id; |
a044f0e7 RB |
1420 | VN_INFO (name)->valnum = name; |
1421 | /* ??? For now mark this SSA name for release by SCCVN. */ | |
1422 | VN_INFO (name)->needs_insertion = true; | |
c9145754 | 1423 | add_to_value (value_id, get_or_alloc_expr_for_name (name)); |
a044f0e7 | 1424 | if (dump_file && (dump_flags & TDF_DETAILS)) |
c9145754 DB |
1425 | { |
1426 | fprintf (dump_file, "Created SSA_NAME representative "); | |
1427 | print_generic_expr (dump_file, name, 0); | |
1428 | fprintf (dump_file, " for expression:"); | |
1429 | print_pre_expr (dump_file, e); | |
984af6ac | 1430 | fprintf (dump_file, " (%04d)\n", value_id); |
c9145754 DB |
1431 | } |
1432 | ||
1433 | return name; | |
1434 | } | |
1435 | ||
1436 | ||
1437 | ||
3e999e7b RG |
1438 | static pre_expr |
1439 | phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, | |
1440 | basic_block pred, basic_block phiblock); | |
c9145754 | 1441 | |
7e6eb623 | 1442 | /* Translate EXPR using phis in PHIBLOCK, so that it has the values of |
788d04b2 RG |
1443 | the phis in PRED. Return NULL if we can't find a leader for each part |
1444 | of the translated expression. */ | |
6de9cd9a | 1445 | |
c9145754 | 1446 | static pre_expr |
3e999e7b RG |
1447 | phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, |
1448 | basic_block pred, basic_block phiblock) | |
6de9cd9a | 1449 | { |
c9145754 | 1450 | switch (expr->kind) |
6de9cd9a | 1451 | { |
c9145754 | 1452 | case NARY: |
43da81be | 1453 | { |
c9145754 DB |
1454 | unsigned int i; |
1455 | bool changed = false; | |
1456 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); | |
5a7d7f9c RG |
1457 | vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s, |
1458 | sizeof_vn_nary_op (nary->length)); | |
1459 | memcpy (newnary, nary, sizeof_vn_nary_op (nary->length)); | |
c9145754 | 1460 | |
5a7d7f9c | 1461 | for (i = 0; i < newnary->length; i++) |
43da81be | 1462 | { |
5a7d7f9c | 1463 | if (TREE_CODE (newnary->op[i]) != SSA_NAME) |
c9145754 DB |
1464 | continue; |
1465 | else | |
43da81be | 1466 | { |
af078bb0 | 1467 | pre_expr leader, result; |
5a7d7f9c | 1468 | unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id; |
af078bb0 | 1469 | leader = find_leader_in_sets (op_val_id, set1, set2); |
788d04b2 | 1470 | result = phi_translate (leader, set1, set2, pred, phiblock); |
c9145754 | 1471 | if (result && result != leader) |
43da81be | 1472 | { |
c9145754 DB |
1473 | tree name = get_representative_for (result); |
1474 | if (!name) | |
85300b46 | 1475 | return NULL; |
5a7d7f9c | 1476 | newnary->op[i] = name; |
43da81be | 1477 | } |
c9145754 DB |
1478 | else if (!result) |
1479 | return NULL; | |
0778d4e8 | 1480 | |
5a7d7f9c | 1481 | changed |= newnary->op[i] != nary->op[i]; |
0778d4e8 | 1482 | } |
c9145754 DB |
1483 | } |
1484 | if (changed) | |
1485 | { | |
1486 | pre_expr constant; | |
5f5126d6 | 1487 | unsigned int new_val_id; |
c9145754 | 1488 | |
5a7d7f9c RG |
1489 | tree result = vn_nary_op_lookup_pieces (newnary->length, |
1490 | newnary->opcode, | |
1491 | newnary->type, | |
1492 | &newnary->op[0], | |
c9145754 | 1493 | &nary); |
5f5126d6 RG |
1494 | if (result && is_gimple_min_invariant (result)) |
1495 | return get_or_alloc_expr_for_constant (result); | |
c9145754 DB |
1496 | |
1497 | expr = (pre_expr) pool_alloc (pre_expr_pool); | |
1498 | expr->kind = NARY; | |
1499 | expr->id = 0; | |
c9145754 DB |
1500 | if (nary) |
1501 | { | |
1502 | PRE_EXPR_NARY (expr) = nary; | |
1503 | constant = fully_constant_expression (expr); | |
1504 | if (constant != expr) | |
1505 | return constant; | |
0778d4e8 | 1506 | |
c9145754 DB |
1507 | new_val_id = nary->value_id; |
1508 | get_or_alloc_expression_id (expr); | |
1509 | } | |
1510 | else | |
43da81be | 1511 | { |
c9145754 | 1512 | new_val_id = get_next_value_id (); |
c3284718 | 1513 | value_expressions.safe_grow_cleared (get_max_value_id () + 1); |
5a7d7f9c RG |
1514 | nary = vn_nary_op_insert_pieces (newnary->length, |
1515 | newnary->opcode, | |
1516 | newnary->type, | |
1517 | &newnary->op[0], | |
c9145754 DB |
1518 | result, new_val_id); |
1519 | PRE_EXPR_NARY (expr) = nary; | |
1520 | constant = fully_constant_expression (expr); | |
1521 | if (constant != expr) | |
1522 | return constant; | |
1523 | get_or_alloc_expression_id (expr); | |
43da81be | 1524 | } |
b0a0ab2d | 1525 | add_to_value (new_val_id, expr); |
c5830edf | 1526 | } |
c9145754 | 1527 | return expr; |
c90186eb | 1528 | } |
c9145754 | 1529 | break; |
47af7a5c | 1530 | |
c9145754 | 1531 | case REFERENCE: |
c90186eb | 1532 | { |
c9145754 | 1533 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); |
9771b263 | 1534 | vec<vn_reference_op_s> operands = ref->operands; |
5006671f RG |
1535 | tree vuse = ref->vuse; |
1536 | tree newvuse = vuse; | |
6e1aa848 | 1537 | vec<vn_reference_op_s> newoperands = vNULL; |
84280917 | 1538 | bool changed = false, same_valid = true; |
4da80bfb | 1539 | unsigned int i, j, n; |
c9145754 DB |
1540 | vn_reference_op_t operand; |
1541 | vn_reference_t newref; | |
1542 | ||
aa7069aa | 1543 | for (i = 0, j = 0; |
9771b263 | 1544 | operands.iterate (i, &operand); i++, j++) |
e13f1c14 | 1545 | { |
c9145754 DB |
1546 | pre_expr opresult; |
1547 | pre_expr leader; | |
4da80bfb | 1548 | tree op[3]; |
c9145754 DB |
1549 | tree type = operand->type; |
1550 | vn_reference_op_s newop = *operand; | |
4da80bfb RG |
1551 | op[0] = operand->op0; |
1552 | op[1] = operand->op1; | |
1553 | op[2] = operand->op2; | |
1554 | for (n = 0; n < 3; ++n) | |
e13f1c14 | 1555 | { |
4da80bfb RG |
1556 | unsigned int op_val_id; |
1557 | if (!op[n]) | |
1558 | continue; | |
1559 | if (TREE_CODE (op[n]) != SSA_NAME) | |
c9145754 | 1560 | { |
4da80bfb RG |
1561 | /* We can't possibly insert these. */ |
1562 | if (n != 0 | |
1563 | && !is_gimple_min_invariant (op[n])) | |
b5d76df4 | 1564 | break; |
4da80bfb | 1565 | continue; |
c9145754 | 1566 | } |
4da80bfb | 1567 | op_val_id = VN_INFO (op[n])->value_id; |
c9145754 | 1568 | leader = find_leader_in_sets (op_val_id, set1, set2); |
4da80bfb RG |
1569 | if (!leader) |
1570 | break; | |
984af6ac RB |
1571 | opresult = phi_translate (leader, set1, set2, pred, phiblock); |
1572 | if (!opresult) | |
1573 | break; | |
1574 | if (opresult != leader) | |
c9145754 | 1575 | { |
984af6ac RB |
1576 | tree name = get_representative_for (opresult); |
1577 | if (!name) | |
b5d76df4 | 1578 | break; |
984af6ac RB |
1579 | changed |= name != op[n]; |
1580 | op[n] = name; | |
c9145754 | 1581 | } |
e13f1c14 | 1582 | } |
4da80bfb | 1583 | if (n != 3) |
e13f1c14 | 1584 | { |
9771b263 | 1585 | newoperands.release (); |
4da80bfb | 1586 | return NULL; |
e13f1c14 | 1587 | } |
9771b263 DN |
1588 | if (!newoperands.exists ()) |
1589 | newoperands = operands.copy (); | |
c9145754 | 1590 | /* We may have changed from an SSA_NAME to a constant */ |
4da80bfb RG |
1591 | if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME) |
1592 | newop.opcode = TREE_CODE (op[0]); | |
c9145754 | 1593 | newop.type = type; |
4da80bfb RG |
1594 | newop.op0 = op[0]; |
1595 | newop.op1 = op[1]; | |
1596 | newop.op2 = op[2]; | |
70f34814 RG |
1597 | /* If it transforms a non-constant ARRAY_REF into a constant |
1598 | one, adjust the constant offset. */ | |
1599 | if (newop.opcode == ARRAY_REF | |
1600 | && newop.off == -1 | |
4da80bfb RG |
1601 | && TREE_CODE (op[0]) == INTEGER_CST |
1602 | && TREE_CODE (op[1]) == INTEGER_CST | |
1603 | && TREE_CODE (op[2]) == INTEGER_CST) | |
70f34814 | 1604 | { |
807e902e KZ |
1605 | offset_int off = ((wi::to_offset (op[0]) |
1606 | - wi::to_offset (op[1])) | |
1607 | * wi::to_offset (op[2])); | |
1608 | if (wi::fits_shwi_p (off)) | |
1609 | newop.off = off.to_shwi (); | |
70f34814 | 1610 | } |
9771b263 | 1611 | newoperands[j] = newop; |
aa7069aa RG |
1612 | /* If it transforms from an SSA_NAME to an address, fold with |
1613 | a preceding indirect reference. */ | |
4da80bfb | 1614 | if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR |
9771b263 | 1615 | && newoperands[j - 1].opcode == MEM_REF) |
aa7069aa | 1616 | vn_reference_fold_indirect (&newoperands, &j); |
e13f1c14 | 1617 | } |
9771b263 | 1618 | if (i != operands.length ()) |
b5d76df4 | 1619 | { |
9771b263 | 1620 | newoperands.release (); |
b5d76df4 RG |
1621 | return NULL; |
1622 | } | |
c90186eb | 1623 | |
5006671f RG |
1624 | if (vuse) |
1625 | { | |
1626 | newvuse = translate_vuse_through_block (newoperands, | |
b45d2719 | 1627 | ref->set, ref->type, |
84280917 MM |
1628 | vuse, phiblock, pred, |
1629 | &same_valid); | |
5006671f RG |
1630 | if (newvuse == NULL_TREE) |
1631 | { | |
9771b263 | 1632 | newoperands.release (); |
5006671f RG |
1633 | return NULL; |
1634 | } | |
1635 | } | |
b9c5e484 | 1636 | |
84280917 | 1637 | if (changed || newvuse != vuse) |
c90186eb | 1638 | { |
47af7a5c RG |
1639 | unsigned int new_val_id; |
1640 | pre_expr constant; | |
1641 | ||
b45d2719 RG |
1642 | tree result = vn_reference_lookup_pieces (newvuse, ref->set, |
1643 | ref->type, | |
c9145754 | 1644 | newoperands, |
3bc27de7 | 1645 | &newref, VN_WALK); |
12bd5a1e | 1646 | if (result) |
9771b263 | 1647 | newoperands.release (); |
c90186eb | 1648 | |
efe7068b RG |
1649 | /* We can always insert constants, so if we have a partial |
1650 | redundant constant load of another type try to translate it | |
1651 | to a constant of appropriate type. */ | |
1652 | if (result && is_gimple_min_invariant (result)) | |
70f34814 | 1653 | { |
efe7068b RG |
1654 | tree tem = result; |
1655 | if (!useless_type_conversion_p (ref->type, TREE_TYPE (result))) | |
1656 | { | |
1657 | tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result); | |
1658 | if (tem && !is_gimple_min_invariant (tem)) | |
1659 | tem = NULL_TREE; | |
1660 | } | |
1661 | if (tem) | |
1662 | return get_or_alloc_expr_for_constant (tem); | |
70f34814 | 1663 | } |
efe7068b RG |
1664 | |
1665 | /* If we'd have to convert things we would need to validate | |
1666 | if we can insert the translated expression. So fail | |
1667 | here for now - we cannot insert an alias with a different | |
1668 | type in the VN tables either, as that would assert. */ | |
1669 | if (result | |
1670 | && !useless_type_conversion_p (ref->type, TREE_TYPE (result))) | |
1671 | return NULL; | |
48feba28 RG |
1672 | else if (!result && newref |
1673 | && !useless_type_conversion_p (ref->type, newref->type)) | |
1674 | { | |
9771b263 | 1675 | newoperands.release (); |
48feba28 RG |
1676 | return NULL; |
1677 | } | |
70f34814 | 1678 | |
c9145754 DB |
1679 | expr = (pre_expr) pool_alloc (pre_expr_pool); |
1680 | expr->kind = REFERENCE; | |
1681 | expr->id = 0; | |
6615c446 | 1682 | |
efe7068b | 1683 | if (newref) |
0995a441 | 1684 | { |
c9145754 | 1685 | PRE_EXPR_REFERENCE (expr) = newref; |
47af7a5c RG |
1686 | constant = fully_constant_expression (expr); |
1687 | if (constant != expr) | |
1688 | return constant; | |
1689 | ||
c9145754 DB |
1690 | new_val_id = newref->value_id; |
1691 | get_or_alloc_expression_id (expr); | |
0995a441 SB |
1692 | } |
1693 | else | |
1694 | { | |
84280917 MM |
1695 | if (changed || !same_valid) |
1696 | { | |
1697 | new_val_id = get_next_value_id (); | |
c3284718 RS |
1698 | value_expressions.safe_grow_cleared |
1699 | (get_max_value_id () + 1); | |
84280917 MM |
1700 | } |
1701 | else | |
1702 | new_val_id = ref->value_id; | |
b45d2719 RG |
1703 | newref = vn_reference_insert_pieces (newvuse, ref->set, |
1704 | ref->type, | |
c9145754 DB |
1705 | newoperands, |
1706 | result, new_val_id); | |
9771b263 | 1707 | newoperands.create (0); |
c9145754 | 1708 | PRE_EXPR_REFERENCE (expr) = newref; |
47af7a5c RG |
1709 | constant = fully_constant_expression (expr); |
1710 | if (constant != expr) | |
1711 | return constant; | |
c9145754 | 1712 | get_or_alloc_expression_id (expr); |
0995a441 | 1713 | } |
b0a0ab2d | 1714 | add_to_value (new_val_id, expr); |
7e6eb623 | 1715 | } |
9771b263 | 1716 | newoperands.release (); |
c9145754 | 1717 | return expr; |
7e6eb623 | 1718 | } |
c9145754 | 1719 | break; |
47af7a5c | 1720 | |
c9145754 | 1721 | case NAME: |
7e6eb623 | 1722 | { |
c9145754 | 1723 | tree name = PRE_EXPR_NAME (expr); |
e076271b RG |
1724 | gimple def_stmt = SSA_NAME_DEF_STMT (name); |
1725 | /* If the SSA name is defined by a PHI node in this block, | |
1726 | translate it. */ | |
726a989a RB |
1727 | if (gimple_code (def_stmt) == GIMPLE_PHI |
1728 | && gimple_bb (def_stmt) == phiblock) | |
9323afae | 1729 | { |
e076271b RG |
1730 | edge e = find_edge (pred, gimple_bb (def_stmt)); |
1731 | tree def = PHI_ARG_DEF (def_stmt, e->dest_idx); | |
73a63870 | 1732 | |
c9145754 | 1733 | /* Handle constant. */ |
89fb70a3 | 1734 | if (is_gimple_min_invariant (def)) |
c9145754 | 1735 | return get_or_alloc_expr_for_constant (def); |
070b797d | 1736 | |
e076271b | 1737 | return get_or_alloc_expr_for_name (def); |
9323afae | 1738 | } |
e076271b RG |
1739 | /* Otherwise return it unchanged - it will get cleaned if its |
1740 | value is not available in PREDs AVAIL_OUT set of expressions. */ | |
1741 | return expr; | |
7e6eb623 | 1742 | } |
6615c446 JO |
1743 | |
1744 | default: | |
1745 | gcc_unreachable (); | |
6de9cd9a DN |
1746 | } |
1747 | } | |
f8b04195 | 1748 | |
3e999e7b RG |
1749 | /* Wrapper around phi_translate_1 providing caching functionality. */ |
1750 | ||
1751 | static pre_expr | |
1752 | phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, | |
1753 | basic_block pred, basic_block phiblock) | |
1754 | { | |
984af6ac | 1755 | expr_pred_trans_t slot = NULL; |
3e999e7b RG |
1756 | pre_expr phitrans; |
1757 | ||
1758 | if (!expr) | |
1759 | return NULL; | |
1760 | ||
1761 | /* Constants contain no values that need translation. */ | |
1762 | if (expr->kind == CONSTANT) | |
1763 | return expr; | |
1764 | ||
1765 | if (value_id_constant_p (get_expr_value_id (expr))) | |
1766 | return expr; | |
1767 | ||
984af6ac | 1768 | /* Don't add translations of NAMEs as those are cheap to translate. */ |
3e999e7b RG |
1769 | if (expr->kind != NAME) |
1770 | { | |
984af6ac RB |
1771 | if (phi_trans_add (&slot, expr, pred)) |
1772 | return slot->v; | |
1773 | /* Store NULL for the value we want to return in the case of | |
1774 | recursing. */ | |
1775 | slot->v = NULL; | |
3e999e7b RG |
1776 | } |
1777 | ||
1778 | /* Translate. */ | |
1779 | phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock); | |
1780 | ||
984af6ac | 1781 | if (slot) |
e2c2fde2 RB |
1782 | { |
1783 | if (phitrans) | |
1784 | slot->v = phitrans; | |
1785 | else | |
1786 | /* Remove failed translations again, they cause insert | |
1787 | iteration to not pick up new opportunities reliably. */ | |
c203e8a7 | 1788 | phi_translate_table->remove_elt_with_hash (slot, slot->hashcode); |
e2c2fde2 | 1789 | } |
3e999e7b RG |
1790 | |
1791 | return phitrans; | |
1792 | } | |
1793 | ||
1794 | ||
c9145754 | 1795 | /* For each expression in SET, translate the values through phi nodes |
f5594471 DB |
1796 | in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting |
1797 | expressions in DEST. */ | |
1798 | ||
6de9cd9a | 1799 | static void |
83737db2 | 1800 | phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred, |
7e6eb623 | 1801 | basic_block phiblock) |
6de9cd9a | 1802 | { |
9771b263 | 1803 | vec<pre_expr> exprs; |
c9145754 | 1804 | pre_expr expr; |
83737db2 DB |
1805 | int i; |
1806 | ||
9adf0570 | 1807 | if (gimple_seq_empty_p (phi_nodes (phiblock))) |
6de9cd9a | 1808 | { |
83737db2 DB |
1809 | bitmap_set_copy (dest, set); |
1810 | return; | |
1811 | } | |
b9c5e484 | 1812 | |
83737db2 | 1813 | exprs = sorted_array_from_bitmap_set (set); |
9771b263 | 1814 | FOR_EACH_VEC_ELT (exprs, i, expr) |
83737db2 | 1815 | { |
c9145754 | 1816 | pre_expr translated; |
f8b04195 | 1817 | translated = phi_translate (expr, set, NULL, pred, phiblock); |
3ed7d068 RG |
1818 | if (!translated) |
1819 | continue; | |
c90186eb | 1820 | |
3ed7d068 RG |
1821 | /* We might end up with multiple expressions from SET being |
1822 | translated to the same value. In this case we do not want | |
1823 | to retain the NARY or REFERENCE expression but prefer a NAME | |
1824 | which would be the leader. */ | |
1825 | if (translated->kind == NAME) | |
1826 | bitmap_value_replace_in_set (dest, translated); | |
1827 | else | |
83737db2 | 1828 | bitmap_value_insert_into_set (dest, translated); |
b9c5e484 | 1829 | } |
9771b263 | 1830 | exprs.release (); |
6de9cd9a DN |
1831 | } |
1832 | ||
bdee7684 | 1833 | /* Find the leader for a value (i.e., the name representing that |
984af6ac RB |
1834 | value) in a given set, and return it. Return NULL if no leader |
1835 | is found. */ | |
bdee7684 | 1836 | |
c9145754 | 1837 | static pre_expr |
e076271b | 1838 | bitmap_find_leader (bitmap_set_t set, unsigned int val) |
bdee7684 | 1839 | { |
c9145754 DB |
1840 | if (value_id_constant_p (val)) |
1841 | { | |
1842 | unsigned int i; | |
1843 | bitmap_iterator bi; | |
9771b263 | 1844 | bitmap exprset = value_expressions[val]; |
83737db2 | 1845 | |
4d4b1b30 | 1846 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
c9145754 DB |
1847 | { |
1848 | pre_expr expr = expression_for_id (i); | |
1849 | if (expr->kind == CONSTANT) | |
1850 | return expr; | |
1851 | } | |
1852 | } | |
bdee7684 DB |
1853 | if (bitmap_set_contains_value (set, val)) |
1854 | { | |
6b416da1 DB |
1855 | /* Rather than walk the entire bitmap of expressions, and see |
1856 | whether any of them has the value we are looking for, we look | |
1857 | at the reverse mapping, which tells us the set of expressions | |
1858 | that have a given value (IE value->expressions with that | |
1859 | value) and see if any of those expressions are in our set. | |
1860 | The number of expressions per value is usually significantly | |
1861 | less than the number of expressions in the set. In fact, for | |
1862 | large testcases, doing it this way is roughly 5-10x faster | |
1863 | than walking the bitmap. | |
1864 | If this is somehow a significant lose for some cases, we can | |
b9c5e484 | 1865 | choose which set to walk based on which set is smaller. */ |
83737db2 DB |
1866 | unsigned int i; |
1867 | bitmap_iterator bi; | |
9771b263 | 1868 | bitmap exprset = value_expressions[val]; |
b9c5e484 | 1869 | |
4d4b1b30 | 1870 | EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi) |
e076271b | 1871 | return expression_for_id (i); |
6de9cd9a | 1872 | } |
7e6eb623 | 1873 | return NULL; |
6de9cd9a DN |
1874 | } |
1875 | ||
cea618ac | 1876 | /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of |
1e4816bc DB |
1877 | BLOCK by seeing if it is not killed in the block. Note that we are |
1878 | only determining whether there is a store that kills it. Because | |
1879 | of the order in which clean iterates over values, we are guaranteed | |
1880 | that altered operands will have caused us to be eliminated from the | |
1881 | ANTIC_IN set already. */ | |
c90186eb DB |
1882 | |
1883 | static bool | |
c9145754 | 1884 | value_dies_in_block_x (pre_expr expr, basic_block block) |
c90186eb | 1885 | { |
5006671f RG |
1886 | tree vuse = PRE_EXPR_REFERENCE (expr)->vuse; |
1887 | vn_reference_t refx = PRE_EXPR_REFERENCE (expr); | |
1888 | gimple def; | |
5006671f RG |
1889 | gimple_stmt_iterator gsi; |
1890 | unsigned id = get_expression_id (expr); | |
1891 | bool res = false; | |
b45d2719 | 1892 | ao_ref ref; |
89fb70a3 | 1893 | |
5006671f RG |
1894 | if (!vuse) |
1895 | return false; | |
1896 | ||
1897 | /* Lookup a previously calculated result. */ | |
1898 | if (EXPR_DIES (block) | |
1899 | && bitmap_bit_p (EXPR_DIES (block), id * 2)) | |
1900 | return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1); | |
1901 | ||
1902 | /* A memory expression {e, VUSE} dies in the block if there is a | |
1903 | statement that may clobber e. If, starting statement walk from the | |
1904 | top of the basic block, a statement uses VUSE there can be no kill | |
1905 | inbetween that use and the original statement that loaded {e, VUSE}, | |
1906 | so we can stop walking. */ | |
b45d2719 | 1907 | ref.base = NULL_TREE; |
5006671f | 1908 | for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi)) |
c90186eb | 1909 | { |
5006671f RG |
1910 | tree def_vuse, def_vdef; |
1911 | def = gsi_stmt (gsi); | |
1912 | def_vuse = gimple_vuse (def); | |
1913 | def_vdef = gimple_vdef (def); | |
89fb70a3 | 1914 | |
5006671f RG |
1915 | /* Not a memory statement. */ |
1916 | if (!def_vuse) | |
1e4816bc | 1917 | continue; |
5006671f RG |
1918 | |
1919 | /* Not a may-def. */ | |
1920 | if (!def_vdef) | |
1921 | { | |
1922 | /* A load with the same VUSE, we're done. */ | |
1923 | if (def_vuse == vuse) | |
1924 | break; | |
1925 | ||
1926 | continue; | |
1927 | } | |
1928 | ||
1929 | /* Init ref only if we really need it. */ | |
b45d2719 RG |
1930 | if (ref.base == NULL_TREE |
1931 | && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type, | |
1932 | refx->operands)) | |
5006671f | 1933 | { |
b45d2719 RG |
1934 | res = true; |
1935 | break; | |
5006671f RG |
1936 | } |
1937 | /* If the statement may clobber expr, it dies. */ | |
b45d2719 | 1938 | if (stmt_may_clobber_ref_p_1 (def, &ref)) |
5006671f RG |
1939 | { |
1940 | res = true; | |
1941 | break; | |
1942 | } | |
c90186eb | 1943 | } |
5006671f RG |
1944 | |
1945 | /* Remember the result. */ | |
1946 | if (!EXPR_DIES (block)) | |
1947 | EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack); | |
1948 | bitmap_set_bit (EXPR_DIES (block), id * 2); | |
1949 | if (res) | |
1950 | bitmap_set_bit (EXPR_DIES (block), id * 2 + 1); | |
1951 | ||
1952 | return res; | |
c90186eb DB |
1953 | } |
1954 | ||
6de9cd9a | 1955 | |
19372838 RG |
1956 | /* Determine if OP is valid in SET1 U SET2, which it is when the union |
1957 | contains its value-id. */ | |
83737db2 | 1958 | |
6de9cd9a | 1959 | static bool |
19372838 | 1960 | op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op) |
6de9cd9a | 1961 | { |
19372838 | 1962 | if (op && TREE_CODE (op) == SSA_NAME) |
c9145754 | 1963 | { |
19372838 | 1964 | unsigned int value_id = VN_INFO (op)->value_id; |
92290a18 RG |
1965 | if (!(bitmap_set_contains_value (set1, value_id) |
1966 | || (set2 && bitmap_set_contains_value (set2, value_id)))) | |
c9145754 DB |
1967 | return false; |
1968 | } | |
c9145754 DB |
1969 | return true; |
1970 | } | |
b9c5e484 | 1971 | |
c9145754 DB |
1972 | /* Determine if the expression EXPR is valid in SET1 U SET2. |
1973 | ONLY SET2 CAN BE NULL. | |
1974 | This means that we have a leader for each part of the expression | |
1975 | (if it consists of values), or the expression is an SSA_NAME. | |
7763473e | 1976 | For loads/calls, we also see if the vuse is killed in this block. */ |
5039610b | 1977 | |
c9145754 DB |
1978 | static bool |
1979 | valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr, | |
1980 | basic_block block) | |
1981 | { | |
1982 | switch (expr->kind) | |
1983 | { | |
1984 | case NAME: | |
e167c04d RB |
1985 | return bitmap_find_leader (AVAIL_OUT (block), |
1986 | get_expr_value_id (expr)) != NULL; | |
c9145754 | 1987 | case NARY: |
43da81be | 1988 | { |
c9145754 DB |
1989 | unsigned int i; |
1990 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); | |
1991 | for (i = 0; i < nary->length; i++) | |
19372838 RG |
1992 | if (!op_valid_in_sets (set1, set2, nary->op[i])) |
1993 | return false; | |
c9145754 | 1994 | return true; |
43da81be | 1995 | } |
c9145754 DB |
1996 | break; |
1997 | case REFERENCE: | |
c90186eb | 1998 | { |
c9145754 DB |
1999 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); |
2000 | vn_reference_op_t vro; | |
2001 | unsigned int i; | |
2002 | ||
9771b263 | 2003 | FOR_EACH_VEC_ELT (ref->operands, i, vro) |
c90186eb | 2004 | { |
19372838 RG |
2005 | if (!op_valid_in_sets (set1, set2, vro->op0) |
2006 | || !op_valid_in_sets (set1, set2, vro->op1) | |
2007 | || !op_valid_in_sets (set1, set2, vro->op2)) | |
e13f1c14 | 2008 | return false; |
c90186eb | 2009 | } |
cd32bb90 | 2010 | return true; |
c90186eb | 2011 | } |
6615c446 | 2012 | default: |
b9c5e484 | 2013 | gcc_unreachable (); |
c9145754 | 2014 | } |
6de9cd9a DN |
2015 | } |
2016 | ||
d75dbccd DB |
2017 | /* Clean the set of expressions that are no longer valid in SET1 or |
2018 | SET2. This means expressions that are made up of values we have no | |
2019 | leaders for in SET1 or SET2. This version is used for partial | |
2020 | anticipation, which means it is not valid in either ANTIC_IN or | |
2021 | PA_IN. */ | |
2022 | ||
2023 | static void | |
2024 | dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block) | |
2025 | { | |
9771b263 | 2026 | vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1); |
c9145754 | 2027 | pre_expr expr; |
d75dbccd DB |
2028 | int i; |
2029 | ||
9771b263 | 2030 | FOR_EACH_VEC_ELT (exprs, i, expr) |
d75dbccd DB |
2031 | { |
2032 | if (!valid_in_sets (set1, set2, expr, block)) | |
2033 | bitmap_remove_from_set (set1, expr); | |
2034 | } | |
9771b263 | 2035 | exprs.release (); |
d75dbccd DB |
2036 | } |
2037 | ||
ca072a31 DB |
2038 | /* Clean the set of expressions that are no longer valid in SET. This |
2039 | means expressions that are made up of values we have no leaders for | |
2040 | in SET. */ | |
6de9cd9a DN |
2041 | |
2042 | static void | |
83737db2 | 2043 | clean (bitmap_set_t set, basic_block block) |
6de9cd9a | 2044 | { |
9771b263 | 2045 | vec<pre_expr> exprs = sorted_array_from_bitmap_set (set); |
c9145754 | 2046 | pre_expr expr; |
83737db2 DB |
2047 | int i; |
2048 | ||
9771b263 | 2049 | FOR_EACH_VEC_ELT (exprs, i, expr) |
6de9cd9a | 2050 | { |
83737db2 DB |
2051 | if (!valid_in_sets (set, NULL, expr, block)) |
2052 | bitmap_remove_from_set (set, expr); | |
6de9cd9a | 2053 | } |
9771b263 | 2054 | exprs.release (); |
6de9cd9a DN |
2055 | } |
2056 | ||
cd32bb90 | 2057 | /* Clean the set of expressions that are no longer valid in SET because |
bea966c2 | 2058 | they are clobbered in BLOCK or because they trap and may not be executed. */ |
cd32bb90 RG |
2059 | |
2060 | static void | |
2061 | prune_clobbered_mems (bitmap_set_t set, basic_block block) | |
2062 | { | |
bea966c2 RG |
2063 | bitmap_iterator bi; |
2064 | unsigned i; | |
cd32bb90 | 2065 | |
bea966c2 | 2066 | FOR_EACH_EXPR_ID_IN_SET (set, i, bi) |
cd32bb90 | 2067 | { |
bea966c2 RG |
2068 | pre_expr expr = expression_for_id (i); |
2069 | if (expr->kind == REFERENCE) | |
2070 | { | |
2071 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); | |
2072 | if (ref->vuse) | |
2073 | { | |
2074 | gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse); | |
2075 | if (!gimple_nop_p (def_stmt) | |
2076 | && ((gimple_bb (def_stmt) != block | |
2077 | && !dominated_by_p (CDI_DOMINATORS, | |
2078 | block, gimple_bb (def_stmt))) | |
2079 | || (gimple_bb (def_stmt) == block | |
2080 | && value_dies_in_block_x (expr, block)))) | |
2081 | bitmap_remove_from_set (set, expr); | |
2082 | } | |
2083 | } | |
2084 | else if (expr->kind == NARY) | |
cd32bb90 | 2085 | { |
bea966c2 RG |
2086 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); |
2087 | /* If the NARY may trap make sure the block does not contain | |
2088 | a possible exit point. | |
2089 | ??? This is overly conservative if we translate AVAIL_OUT | |
2090 | as the available expression might be after the exit point. */ | |
2091 | if (BB_MAY_NOTRETURN (block) | |
2092 | && vn_nary_may_trap (nary)) | |
cd32bb90 RG |
2093 | bitmap_remove_from_set (set, expr); |
2094 | } | |
2095 | } | |
cd32bb90 RG |
2096 | } |
2097 | ||
d4222d43 | 2098 | static sbitmap has_abnormal_preds; |
89fb70a3 | 2099 | |
83737db2 DB |
2100 | /* List of blocks that may have changed during ANTIC computation and |
2101 | thus need to be iterated over. */ | |
2102 | ||
2103 | static sbitmap changed_blocks; | |
1e4816bc DB |
2104 | |
2105 | /* Decide whether to defer a block for a later iteration, or PHI | |
2106 | translate SOURCE to DEST using phis in PHIBLOCK. Return false if we | |
2107 | should defer the block, and true if we processed it. */ | |
2108 | ||
2109 | static bool | |
2110 | defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source, | |
2111 | basic_block block, basic_block phiblock) | |
2112 | { | |
2113 | if (!BB_VISITED (phiblock)) | |
2114 | { | |
d7c028c0 | 2115 | bitmap_set_bit (changed_blocks, block->index); |
1e4816bc DB |
2116 | BB_VISITED (block) = 0; |
2117 | BB_DEFERRED (block) = 1; | |
2118 | return false; | |
2119 | } | |
2120 | else | |
2121 | phi_translate_set (dest, source, block, phiblock); | |
2122 | return true; | |
2123 | } | |
2124 | ||
7e6eb623 | 2125 | /* Compute the ANTIC set for BLOCK. |
6de9cd9a | 2126 | |
665fcad8 SB |
2127 | If succs(BLOCK) > 1 then |
2128 | ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK) | |
2129 | else if succs(BLOCK) == 1 then | |
2130 | ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)]) | |
6de9cd9a | 2131 | |
665fcad8 | 2132 | ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK]) |
83737db2 | 2133 | */ |
6de9cd9a | 2134 | |
7e6eb623 | 2135 | static bool |
a28fee03 | 2136 | compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge) |
6de9cd9a | 2137 | { |
7e6eb623 | 2138 | bool changed = false; |
83737db2 DB |
2139 | bitmap_set_t S, old, ANTIC_OUT; |
2140 | bitmap_iterator bi; | |
2141 | unsigned int bii; | |
2142 | edge e; | |
2143 | edge_iterator ei; | |
a28fee03 | 2144 | |
83737db2 | 2145 | old = ANTIC_OUT = S = NULL; |
d75dbccd | 2146 | BB_VISITED (block) = 1; |
a28fee03 SB |
2147 | |
2148 | /* If any edges from predecessors are abnormal, antic_in is empty, | |
2149 | so do nothing. */ | |
2150 | if (block_has_abnormal_pred_edge) | |
2151 | goto maybe_dump_sets; | |
6de9cd9a | 2152 | |
83737db2 DB |
2153 | old = ANTIC_IN (block); |
2154 | ANTIC_OUT = bitmap_set_new (); | |
6de9cd9a | 2155 | |
a28fee03 SB |
2156 | /* If the block has no successors, ANTIC_OUT is empty. */ |
2157 | if (EDGE_COUNT (block->succs) == 0) | |
2158 | ; | |
7e6eb623 DB |
2159 | /* If we have one successor, we could have some phi nodes to |
2160 | translate through. */ | |
c5cbcccf | 2161 | else if (single_succ_p (block)) |
6de9cd9a | 2162 | { |
83737db2 | 2163 | basic_block succ_bb = single_succ (block); |
d75dbccd DB |
2164 | |
2165 | /* We trade iterations of the dataflow equations for having to | |
2166 | phi translate the maximal set, which is incredibly slow | |
2167 | (since the maximal set often has 300+ members, even when you | |
2168 | have a small number of blocks). | |
2169 | Basically, we defer the computation of ANTIC for this block | |
2f8e468b | 2170 | until we have processed it's successor, which will inevitably |
d75dbccd DB |
2171 | have a *much* smaller set of values to phi translate once |
2172 | clean has been run on it. | |
2173 | The cost of doing this is that we technically perform more | |
2174 | iterations, however, they are lower cost iterations. | |
2175 | ||
2176 | Timings for PRE on tramp3d-v4: | |
2177 | without maximal set fix: 11 seconds | |
2178 | with maximal set fix/without deferring: 26 seconds | |
2179 | with maximal set fix/with deferring: 11 seconds | |
2180 | */ | |
2181 | ||
1e4816bc DB |
2182 | if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb), |
2183 | block, succ_bb)) | |
d75dbccd DB |
2184 | { |
2185 | changed = true; | |
d75dbccd DB |
2186 | goto maybe_dump_sets; |
2187 | } | |
6de9cd9a | 2188 | } |
7e6eb623 | 2189 | /* If we have multiple successors, we take the intersection of all of |
1e4816bc DB |
2190 | them. Note that in the case of loop exit phi nodes, we may have |
2191 | phis to translate through. */ | |
7e6eb623 | 2192 | else |
6de9cd9a | 2193 | { |
7e6eb623 | 2194 | size_t i; |
70a6b17e | 2195 | basic_block bprime, first = NULL; |
7e6eb623 | 2196 | |
ef062b13 | 2197 | auto_vec<basic_block> worklist (EDGE_COUNT (block->succs)); |
628f6a4e | 2198 | FOR_EACH_EDGE (e, ei, block->succs) |
1e4816bc | 2199 | { |
70a6b17e RG |
2200 | if (!first |
2201 | && BB_VISITED (e->dest)) | |
2202 | first = e->dest; | |
2203 | else if (BB_VISITED (e->dest)) | |
9771b263 | 2204 | worklist.quick_push (e->dest); |
1e4816bc | 2205 | } |
70a6b17e RG |
2206 | |
2207 | /* Of multiple successors we have to have visited one already. */ | |
2208 | if (!first) | |
1e4816bc | 2209 | { |
d7c028c0 | 2210 | bitmap_set_bit (changed_blocks, block->index); |
70a6b17e RG |
2211 | BB_VISITED (block) = 0; |
2212 | BB_DEFERRED (block) = 1; | |
2213 | changed = true; | |
70a6b17e | 2214 | goto maybe_dump_sets; |
1e4816bc | 2215 | } |
c90186eb | 2216 | |
9adf0570 | 2217 | if (!gimple_seq_empty_p (phi_nodes (first))) |
70a6b17e RG |
2218 | phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first); |
2219 | else | |
2220 | bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first)); | |
2221 | ||
9771b263 | 2222 | FOR_EACH_VEC_ELT (worklist, i, bprime) |
d75dbccd | 2223 | { |
9adf0570 | 2224 | if (!gimple_seq_empty_p (phi_nodes (bprime))) |
1e4816bc DB |
2225 | { |
2226 | bitmap_set_t tmp = bitmap_set_new (); | |
70a6b17e | 2227 | phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime); |
1e4816bc DB |
2228 | bitmap_set_and (ANTIC_OUT, tmp); |
2229 | bitmap_set_free (tmp); | |
2230 | } | |
89fb70a3 | 2231 | else |
70a6b17e | 2232 | bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime)); |
6de9cd9a DN |
2233 | } |
2234 | } | |
6de9cd9a | 2235 | |
cd32bb90 RG |
2236 | /* Prune expressions that are clobbered in block and thus become |
2237 | invalid if translated from ANTIC_OUT to ANTIC_IN. */ | |
2238 | prune_clobbered_mems (ANTIC_OUT, block); | |
2239 | ||
ea4b7848 | 2240 | /* Generate ANTIC_OUT - TMP_GEN. */ |
83737db2 | 2241 | S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block)); |
6de9cd9a | 2242 | |
d75dbccd | 2243 | /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */ |
83737db2 DB |
2244 | ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block), |
2245 | TMP_GEN (block)); | |
c33bae88 | 2246 | |
a28fee03 SB |
2247 | /* Then union in the ANTIC_OUT - TMP_GEN values, |
2248 | to get ANTIC_OUT U EXP_GEN - TMP_GEN */ | |
83737db2 DB |
2249 | FOR_EACH_EXPR_ID_IN_SET (S, bii, bi) |
2250 | bitmap_value_insert_into_set (ANTIC_IN (block), | |
2251 | expression_for_id (bii)); | |
6de9cd9a | 2252 | |
c90186eb | 2253 | clean (ANTIC_IN (block), block); |
d75dbccd | 2254 | |
5c72d561 | 2255 | if (!bitmap_set_equal (old, ANTIC_IN (block))) |
83737db2 DB |
2256 | { |
2257 | changed = true; | |
d7c028c0 | 2258 | bitmap_set_bit (changed_blocks, block->index); |
83737db2 | 2259 | FOR_EACH_EDGE (e, ei, block->preds) |
d7c028c0 | 2260 | bitmap_set_bit (changed_blocks, e->src->index); |
83737db2 DB |
2261 | } |
2262 | else | |
d7c028c0 | 2263 | bitmap_clear_bit (changed_blocks, block->index); |
6de9cd9a | 2264 | |
a28fee03 | 2265 | maybe_dump_sets: |
7e6eb623 DB |
2266 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2267 | { | |
d75dbccd DB |
2268 | if (!BB_DEFERRED (block) || BB_VISITED (block)) |
2269 | { | |
2270 | if (ANTIC_OUT) | |
2271 | print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index); | |
85300b46 | 2272 | |
d75dbccd DB |
2273 | print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN", |
2274 | block->index); | |
85300b46 | 2275 | |
d75dbccd DB |
2276 | if (S) |
2277 | print_bitmap_set (dump_file, S, "S", block->index); | |
2278 | } | |
2279 | else | |
2280 | { | |
2281 | fprintf (dump_file, | |
2282 | "Block %d was deferred for a future iteration.\n", | |
2283 | block->index); | |
2284 | } | |
83737db2 DB |
2285 | } |
2286 | if (old) | |
2287 | bitmap_set_free (old); | |
2288 | if (S) | |
2289 | bitmap_set_free (S); | |
2290 | if (ANTIC_OUT) | |
2291 | bitmap_set_free (ANTIC_OUT); | |
7e6eb623 | 2292 | return changed; |
6de9cd9a DN |
2293 | } |
2294 | ||
d75dbccd DB |
2295 | /* Compute PARTIAL_ANTIC for BLOCK. |
2296 | ||
2297 | If succs(BLOCK) > 1 then | |
2298 | PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not | |
2299 | in ANTIC_OUT for all succ(BLOCK) | |
2300 | else if succs(BLOCK) == 1 then | |
2301 | PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)]) | |
2302 | ||
2303 | PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK] | |
2304 | - ANTIC_IN[BLOCK]) | |
2305 | ||
2306 | */ | |
2307 | static bool | |
2308 | compute_partial_antic_aux (basic_block block, | |
2309 | bool block_has_abnormal_pred_edge) | |
2310 | { | |
2311 | bool changed = false; | |
2312 | bitmap_set_t old_PA_IN; | |
2313 | bitmap_set_t PA_OUT; | |
2314 | edge e; | |
2315 | edge_iterator ei; | |
f0ed4cfb | 2316 | unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH); |
d75dbccd DB |
2317 | |
2318 | old_PA_IN = PA_OUT = NULL; | |
2319 | ||
2320 | /* If any edges from predecessors are abnormal, antic_in is empty, | |
2321 | so do nothing. */ | |
2322 | if (block_has_abnormal_pred_edge) | |
2323 | goto maybe_dump_sets; | |
2324 | ||
f0ed4cfb NC |
2325 | /* If there are too many partially anticipatable values in the |
2326 | block, phi_translate_set can take an exponential time: stop | |
2327 | before the translation starts. */ | |
2328 | if (max_pa | |
2329 | && single_succ_p (block) | |
5c72d561 | 2330 | && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa) |
f0ed4cfb NC |
2331 | goto maybe_dump_sets; |
2332 | ||
d75dbccd DB |
2333 | old_PA_IN = PA_IN (block); |
2334 | PA_OUT = bitmap_set_new (); | |
2335 | ||
2336 | /* If the block has no successors, ANTIC_OUT is empty. */ | |
2337 | if (EDGE_COUNT (block->succs) == 0) | |
2338 | ; | |
2339 | /* If we have one successor, we could have some phi nodes to | |
2340 | translate through. Note that we can't phi translate across DFS | |
89fb70a3 DB |
2341 | back edges in partial antic, because it uses a union operation on |
2342 | the successors. For recurrences like IV's, we will end up | |
2343 | generating a new value in the set on each go around (i + 3 (VH.1) | |
2344 | VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */ | |
d75dbccd DB |
2345 | else if (single_succ_p (block)) |
2346 | { | |
2347 | basic_block succ = single_succ (block); | |
2348 | if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK)) | |
2349 | phi_translate_set (PA_OUT, PA_IN (succ), block, succ); | |
2350 | } | |
2351 | /* If we have multiple successors, we take the union of all of | |
2352 | them. */ | |
2353 | else | |
2354 | { | |
d75dbccd DB |
2355 | size_t i; |
2356 | basic_block bprime; | |
2357 | ||
ef062b13 | 2358 | auto_vec<basic_block> worklist (EDGE_COUNT (block->succs)); |
d75dbccd DB |
2359 | FOR_EACH_EDGE (e, ei, block->succs) |
2360 | { | |
2361 | if (e->flags & EDGE_DFS_BACK) | |
2362 | continue; | |
9771b263 | 2363 | worklist.quick_push (e->dest); |
d75dbccd | 2364 | } |
9771b263 | 2365 | if (worklist.length () > 0) |
d75dbccd | 2366 | { |
9771b263 | 2367 | FOR_EACH_VEC_ELT (worklist, i, bprime) |
d75dbccd DB |
2368 | { |
2369 | unsigned int i; | |
2370 | bitmap_iterator bi; | |
2371 | ||
2372 | FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi) | |
2373 | bitmap_value_insert_into_set (PA_OUT, | |
2374 | expression_for_id (i)); | |
9adf0570 | 2375 | if (!gimple_seq_empty_p (phi_nodes (bprime))) |
1e4816bc DB |
2376 | { |
2377 | bitmap_set_t pa_in = bitmap_set_new (); | |
2378 | phi_translate_set (pa_in, PA_IN (bprime), block, bprime); | |
2379 | FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi) | |
2380 | bitmap_value_insert_into_set (PA_OUT, | |
2381 | expression_for_id (i)); | |
2382 | bitmap_set_free (pa_in); | |
2383 | } | |
2384 | else | |
2385 | FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi) | |
2386 | bitmap_value_insert_into_set (PA_OUT, | |
2387 | expression_for_id (i)); | |
d75dbccd DB |
2388 | } |
2389 | } | |
d75dbccd DB |
2390 | } |
2391 | ||
cd32bb90 RG |
2392 | /* Prune expressions that are clobbered in block and thus become |
2393 | invalid if translated from PA_OUT to PA_IN. */ | |
2394 | prune_clobbered_mems (PA_OUT, block); | |
2395 | ||
d75dbccd DB |
2396 | /* PA_IN starts with PA_OUT - TMP_GEN. |
2397 | Then we subtract things from ANTIC_IN. */ | |
2398 | PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block)); | |
2399 | ||
2400 | /* For partial antic, we want to put back in the phi results, since | |
2401 | we will properly avoid making them partially antic over backedges. */ | |
5c72d561 JH |
2402 | bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values); |
2403 | bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions); | |
d75dbccd DB |
2404 | |
2405 | /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */ | |
2406 | bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block)); | |
2407 | ||
2408 | dependent_clean (PA_IN (block), ANTIC_IN (block), block); | |
2409 | ||
2410 | if (!bitmap_set_equal (old_PA_IN, PA_IN (block))) | |
2411 | { | |
2412 | changed = true; | |
d7c028c0 | 2413 | bitmap_set_bit (changed_blocks, block->index); |
d75dbccd | 2414 | FOR_EACH_EDGE (e, ei, block->preds) |
d7c028c0 | 2415 | bitmap_set_bit (changed_blocks, e->src->index); |
d75dbccd DB |
2416 | } |
2417 | else | |
d7c028c0 | 2418 | bitmap_clear_bit (changed_blocks, block->index); |
d75dbccd DB |
2419 | |
2420 | maybe_dump_sets: | |
2421 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2422 | { | |
2423 | if (PA_OUT) | |
2424 | print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index); | |
2425 | ||
2426 | print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index); | |
2427 | } | |
2428 | if (old_PA_IN) | |
2429 | bitmap_set_free (old_PA_IN); | |
2430 | if (PA_OUT) | |
2431 | bitmap_set_free (PA_OUT); | |
2432 | return changed; | |
2433 | } | |
2434 | ||
83737db2 | 2435 | /* Compute ANTIC and partial ANTIC sets. */ |
6de9cd9a DN |
2436 | |
2437 | static void | |
7e6eb623 | 2438 | compute_antic (void) |
6de9cd9a | 2439 | { |
c33bae88 | 2440 | bool changed = true; |
7e6eb623 | 2441 | int num_iterations = 0; |
c33bae88 | 2442 | basic_block block; |
83737db2 | 2443 | int i; |
a28fee03 SB |
2444 | |
2445 | /* If any predecessor edges are abnormal, we punt, so antic_in is empty. | |
2446 | We pre-build the map of blocks with incoming abnormal edges here. */ | |
8b1c6fd7 | 2447 | has_abnormal_preds = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
f61e445a | 2448 | bitmap_clear (has_abnormal_preds); |
83737db2 | 2449 | |
04a90bec | 2450 | FOR_ALL_BB_FN (block, cfun) |
7e6eb623 | 2451 | { |
a28fee03 SB |
2452 | edge_iterator ei; |
2453 | edge e; | |
2454 | ||
2455 | FOR_EACH_EDGE (e, ei, block->preds) | |
83737db2 DB |
2456 | { |
2457 | e->flags &= ~EDGE_DFS_BACK; | |
2458 | if (e->flags & EDGE_ABNORMAL) | |
2459 | { | |
d7c028c0 | 2460 | bitmap_set_bit (has_abnormal_preds, block->index); |
83737db2 DB |
2461 | break; |
2462 | } | |
2463 | } | |
a28fee03 | 2464 | |
83737db2 | 2465 | BB_VISITED (block) = 0; |
d75dbccd | 2466 | BB_DEFERRED (block) = 0; |
a19eb9d2 | 2467 | |
a28fee03 | 2468 | /* While we are here, give empty ANTIC_IN sets to each block. */ |
83737db2 | 2469 | ANTIC_IN (block) = bitmap_set_new (); |
d75dbccd | 2470 | PA_IN (block) = bitmap_set_new (); |
7e6eb623 | 2471 | } |
83737db2 | 2472 | |
a28fee03 | 2473 | /* At the exit block we anticipate nothing. */ |
fefa31b5 | 2474 | BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1; |
a28fee03 | 2475 | |
8b1c6fd7 | 2476 | changed_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun) + 1); |
f61e445a | 2477 | bitmap_ones (changed_blocks); |
7e6eb623 DB |
2478 | while (changed) |
2479 | { | |
83737db2 DB |
2480 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2481 | fprintf (dump_file, "Starting iteration %d\n", num_iterations); | |
3bc27de7 RG |
2482 | /* ??? We need to clear our PHI translation cache here as the |
2483 | ANTIC sets shrink and we restrict valid translations to | |
2484 | those having operands with leaders in ANTIC. Same below | |
2485 | for PA ANTIC computation. */ | |
a28fee03 | 2486 | num_iterations++; |
c33bae88 | 2487 | changed = false; |
585d0dc4 | 2488 | for (i = postorder_num - 1; i >= 0; i--) |
83737db2 | 2489 | { |
d7c028c0 | 2490 | if (bitmap_bit_p (changed_blocks, postorder[i])) |
83737db2 | 2491 | { |
06e28de2 | 2492 | basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]); |
83737db2 | 2493 | changed |= compute_antic_aux (block, |
d7c028c0 | 2494 | bitmap_bit_p (has_abnormal_preds, |
83737db2 DB |
2495 | block->index)); |
2496 | } | |
2497 | } | |
d75dbccd | 2498 | /* Theoretically possible, but *highly* unlikely. */ |
77a74ed7 | 2499 | gcc_checking_assert (num_iterations < 500); |
2e24fa83 | 2500 | } |
a28fee03 | 2501 | |
9fe0cb7d RG |
2502 | statistics_histogram_event (cfun, "compute_antic iterations", |
2503 | num_iterations); | |
83737db2 | 2504 | |
d75dbccd DB |
2505 | if (do_partial_partial) |
2506 | { | |
f61e445a | 2507 | bitmap_ones (changed_blocks); |
d75dbccd DB |
2508 | mark_dfs_back_edges (); |
2509 | num_iterations = 0; | |
2510 | changed = true; | |
2511 | while (changed) | |
2512 | { | |
2513 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2514 | fprintf (dump_file, "Starting iteration %d\n", num_iterations); | |
2515 | num_iterations++; | |
2516 | changed = false; | |
585d0dc4 | 2517 | for (i = postorder_num - 1 ; i >= 0; i--) |
d75dbccd | 2518 | { |
d7c028c0 | 2519 | if (bitmap_bit_p (changed_blocks, postorder[i])) |
d75dbccd | 2520 | { |
06e28de2 | 2521 | basic_block block = BASIC_BLOCK_FOR_FN (cfun, postorder[i]); |
d75dbccd DB |
2522 | changed |
2523 | |= compute_partial_antic_aux (block, | |
d7c028c0 | 2524 | bitmap_bit_p (has_abnormal_preds, |
d75dbccd DB |
2525 | block->index)); |
2526 | } | |
2527 | } | |
2528 | /* Theoretically possible, but *highly* unlikely. */ | |
77a74ed7 | 2529 | gcc_checking_assert (num_iterations < 500); |
d75dbccd | 2530 | } |
9fe0cb7d RG |
2531 | statistics_histogram_event (cfun, "compute_partial_antic iterations", |
2532 | num_iterations); | |
d75dbccd | 2533 | } |
83737db2 DB |
2534 | sbitmap_free (has_abnormal_preds); |
2535 | sbitmap_free (changed_blocks); | |
6de9cd9a DN |
2536 | } |
2537 | ||
c90186eb DB |
2538 | |
2539 | /* Inserted expressions are placed onto this worklist, which is used | |
2540 | for performing quick dead code elimination of insertions we made | |
2541 | that didn't turn out to be necessary. */ | |
0ab555de | 2542 | static bitmap inserted_exprs; |
c90186eb | 2543 | |
ce94d354 | 2544 | /* The actual worker for create_component_ref_by_pieces. */ |
b9c5e484 | 2545 | |
85300b46 | 2546 | static tree |
ce94d354 | 2547 | create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref, |
e076271b | 2548 | unsigned int *operand, gimple_seq *stmts) |
85300b46 | 2549 | { |
9771b263 | 2550 | vn_reference_op_t currop = &ref->operands[*operand]; |
c9145754 | 2551 | tree genop; |
ce94d354 | 2552 | ++*operand; |
c9145754 | 2553 | switch (currop->opcode) |
85300b46 | 2554 | { |
c9145754 DB |
2555 | case CALL_EXPR: |
2556 | { | |
b3be2694 | 2557 | tree folded, sc = NULL_TREE; |
ce94d354 | 2558 | unsigned int nargs = 0; |
b3be2694 RG |
2559 | tree fn, *args; |
2560 | if (TREE_CODE (currop->op0) == FUNCTION_DECL) | |
2561 | fn = currop->op0; | |
2562 | else | |
e076271b | 2563 | fn = find_or_generate_expression (block, currop->op0, stmts); |
c3dd8dd7 RB |
2564 | if (!fn) |
2565 | return NULL_TREE; | |
b3be2694 | 2566 | if (currop->op1) |
c3dd8dd7 RB |
2567 | { |
2568 | sc = find_or_generate_expression (block, currop->op1, stmts); | |
2569 | if (!sc) | |
2570 | return NULL_TREE; | |
2571 | } | |
9771b263 DN |
2572 | args = XNEWVEC (tree, ref->operands.length () - 1); |
2573 | while (*operand < ref->operands.length ()) | |
c9145754 | 2574 | { |
ce94d354 | 2575 | args[nargs] = create_component_ref_by_pieces_1 (block, ref, |
e076271b | 2576 | operand, stmts); |
c3dd8dd7 RB |
2577 | if (!args[nargs]) |
2578 | return NULL_TREE; | |
ce94d354 | 2579 | nargs++; |
c9145754 | 2580 | } |
726a989a | 2581 | folded = build_call_array (currop->type, |
b3be2694 RG |
2582 | (TREE_CODE (fn) == FUNCTION_DECL |
2583 | ? build_fold_addr_expr (fn) : fn), | |
726a989a | 2584 | nargs, args); |
c9145754 | 2585 | free (args); |
7aec7a38 | 2586 | if (sc) |
b3be2694 | 2587 | CALL_EXPR_STATIC_CHAIN (folded) = sc; |
c9145754 DB |
2588 | return folded; |
2589 | } | |
ea814c66 | 2590 | |
70f34814 RG |
2591 | case MEM_REF: |
2592 | { | |
2593 | tree baseop = create_component_ref_by_pieces_1 (block, ref, operand, | |
e076271b | 2594 | stmts); |
c3dd8dd7 RB |
2595 | if (!baseop) |
2596 | return NULL_TREE; | |
70f34814 | 2597 | tree offset = currop->op0; |
70f34814 RG |
2598 | if (TREE_CODE (baseop) == ADDR_EXPR |
2599 | && handled_component_p (TREE_OPERAND (baseop, 0))) | |
2600 | { | |
2601 | HOST_WIDE_INT off; | |
2602 | tree base; | |
2603 | base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0), | |
2604 | &off); | |
2605 | gcc_assert (base); | |
2606 | offset = int_const_binop (PLUS_EXPR, offset, | |
2607 | build_int_cst (TREE_TYPE (offset), | |
d35936ab | 2608 | off)); |
70f34814 RG |
2609 | baseop = build_fold_addr_expr (base); |
2610 | } | |
2611 | return fold_build2 (MEM_REF, currop->type, baseop, offset); | |
2612 | } | |
ea814c66 | 2613 | |
150e3929 RG |
2614 | case TARGET_MEM_REF: |
2615 | { | |
4d948885 | 2616 | tree genop0 = NULL_TREE, genop1 = NULL_TREE; |
9771b263 | 2617 | vn_reference_op_t nextop = &ref->operands[++*operand]; |
150e3929 | 2618 | tree baseop = create_component_ref_by_pieces_1 (block, ref, operand, |
e076271b | 2619 | stmts); |
c3dd8dd7 RB |
2620 | if (!baseop) |
2621 | return NULL_TREE; | |
150e3929 | 2622 | if (currop->op0) |
c3dd8dd7 RB |
2623 | { |
2624 | genop0 = find_or_generate_expression (block, currop->op0, stmts); | |
2625 | if (!genop0) | |
2626 | return NULL_TREE; | |
2627 | } | |
4d948885 | 2628 | if (nextop->op0) |
c3dd8dd7 RB |
2629 | { |
2630 | genop1 = find_or_generate_expression (block, nextop->op0, stmts); | |
2631 | if (!genop1) | |
2632 | return NULL_TREE; | |
2633 | } | |
4d948885 RG |
2634 | return build5 (TARGET_MEM_REF, currop->type, |
2635 | baseop, currop->op2, genop0, currop->op1, genop1); | |
150e3929 | 2636 | } |
ea814c66 | 2637 | |
ce94d354 RG |
2638 | case ADDR_EXPR: |
2639 | if (currop->op0) | |
2640 | { | |
2641 | gcc_assert (is_gimple_min_invariant (currop->op0)); | |
2642 | return currop->op0; | |
2643 | } | |
2644 | /* Fallthrough. */ | |
c9145754 DB |
2645 | case REALPART_EXPR: |
2646 | case IMAGPART_EXPR: | |
2647 | case VIEW_CONVERT_EXPR: | |
2648 | { | |
c3dd8dd7 RB |
2649 | tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand, |
2650 | stmts); | |
2651 | if (!genop0) | |
2652 | return NULL_TREE; | |
ea814c66 | 2653 | return fold_build1 (currop->opcode, currop->type, genop0); |
c9145754 | 2654 | } |
ea814c66 | 2655 | |
842679dc TV |
2656 | case WITH_SIZE_EXPR: |
2657 | { | |
2658 | tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand, | |
e076271b | 2659 | stmts); |
c3dd8dd7 RB |
2660 | if (!genop0) |
2661 | return NULL_TREE; | |
e076271b | 2662 | tree genop1 = find_or_generate_expression (block, currop->op0, stmts); |
c3dd8dd7 RB |
2663 | if (!genop1) |
2664 | return NULL_TREE; | |
842679dc TV |
2665 | return fold_build2 (currop->opcode, currop->type, genop0, genop1); |
2666 | } | |
ea814c66 | 2667 | |
c9145754 | 2668 | case BIT_FIELD_REF: |
e13f1c14 | 2669 | { |
ce94d354 | 2670 | tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand, |
e076271b | 2671 | stmts); |
c3dd8dd7 RB |
2672 | if (!genop0) |
2673 | return NULL_TREE; | |
ea814c66 EB |
2674 | tree op1 = currop->op0; |
2675 | tree op2 = currop->op1; | |
ea814c66 | 2676 | return fold_build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2); |
e13f1c14 | 2677 | } |
c9145754 DB |
2678 | |
2679 | /* For array ref vn_reference_op's, operand 1 of the array ref | |
2680 | is op0 of the reference op and operand 3 of the array ref is | |
2681 | op1. */ | |
2682 | case ARRAY_RANGE_REF: | |
2683 | case ARRAY_REF: | |
2684 | { | |
c9145754 DB |
2685 | tree genop0; |
2686 | tree genop1 = currop->op0; | |
c9145754 | 2687 | tree genop2 = currop->op1; |
e52201b6 | 2688 | tree genop3 = currop->op2; |
c3dd8dd7 RB |
2689 | genop0 = create_component_ref_by_pieces_1 (block, ref, operand, |
2690 | stmts); | |
2691 | if (!genop0) | |
2692 | return NULL_TREE; | |
e076271b | 2693 | genop1 = find_or_generate_expression (block, genop1, stmts); |
c3dd8dd7 RB |
2694 | if (!genop1) |
2695 | return NULL_TREE; | |
c9145754 DB |
2696 | if (genop2) |
2697 | { | |
d4d73ce2 EB |
2698 | tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0)); |
2699 | /* Drop zero minimum index if redundant. */ | |
2700 | if (integer_zerop (genop2) | |
2701 | && (!domain_type | |
2702 | || integer_zerop (TYPE_MIN_VALUE (domain_type)))) | |
d53bed0b RG |
2703 | genop2 = NULL_TREE; |
2704 | else | |
c3dd8dd7 RB |
2705 | { |
2706 | genop2 = find_or_generate_expression (block, genop2, stmts); | |
2707 | if (!genop2) | |
2708 | return NULL_TREE; | |
2709 | } | |
c9145754 | 2710 | } |
e52201b6 RG |
2711 | if (genop3) |
2712 | { | |
2713 | tree elmt_type = TREE_TYPE (TREE_TYPE (genop0)); | |
d53bed0b RG |
2714 | /* We can't always put a size in units of the element alignment |
2715 | here as the element alignment may be not visible. See | |
2716 | PR43783. Simply drop the element size for constant | |
2717 | sizes. */ | |
2718 | if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type))) | |
2719 | genop3 = NULL_TREE; | |
2720 | else | |
2721 | { | |
2722 | genop3 = size_binop (EXACT_DIV_EXPR, genop3, | |
2723 | size_int (TYPE_ALIGN_UNIT (elmt_type))); | |
e076271b | 2724 | genop3 = find_or_generate_expression (block, genop3, stmts); |
c3dd8dd7 RB |
2725 | if (!genop3) |
2726 | return NULL_TREE; | |
d53bed0b | 2727 | } |
e52201b6 | 2728 | } |
c9145754 DB |
2729 | return build4 (currop->opcode, currop->type, genop0, genop1, |
2730 | genop2, genop3); | |
2731 | } | |
85300b46 DB |
2732 | case COMPONENT_REF: |
2733 | { | |
2734 | tree op0; | |
2735 | tree op1; | |
c9145754 | 2736 | tree genop2 = currop->op1; |
e076271b | 2737 | op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts); |
c3dd8dd7 RB |
2738 | if (!op0) |
2739 | return NULL_TREE; | |
e076271b | 2740 | /* op1 should be a FIELD_DECL, which are represented by themselves. */ |
c9145754 DB |
2741 | op1 = currop->op0; |
2742 | if (genop2) | |
c3dd8dd7 RB |
2743 | { |
2744 | genop2 = find_or_generate_expression (block, genop2, stmts); | |
2745 | if (!genop2) | |
2746 | return NULL_TREE; | |
2747 | } | |
ea814c66 | 2748 | return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2); |
85300b46 | 2749 | } |
ea814c66 | 2750 | |
c9145754 | 2751 | case SSA_NAME: |
85300b46 | 2752 | { |
e076271b | 2753 | genop = find_or_generate_expression (block, currop->op0, stmts); |
c9145754 | 2754 | return genop; |
85300b46 | 2755 | } |
c9145754 DB |
2756 | case STRING_CST: |
2757 | case INTEGER_CST: | |
2758 | case COMPLEX_CST: | |
2759 | case VECTOR_CST: | |
2760 | case REAL_CST: | |
2761 | case CONSTRUCTOR: | |
85300b46 DB |
2762 | case VAR_DECL: |
2763 | case PARM_DECL: | |
c9145754 | 2764 | case CONST_DECL: |
5230d884 | 2765 | case RESULT_DECL: |
c9145754 | 2766 | case FUNCTION_DECL: |
c9145754 DB |
2767 | return currop->op0; |
2768 | ||
85300b46 | 2769 | default: |
b9c5e484 | 2770 | gcc_unreachable (); |
85300b46 | 2771 | } |
85300b46 | 2772 | } |
c90186eb | 2773 | |
ce94d354 | 2774 | /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the |
70f34814 | 2775 | COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with |
ce94d354 RG |
2776 | trying to rename aggregates into ssa form directly, which is a no no. |
2777 | ||
2778 | Thus, this routine doesn't create temporaries, it just builds a | |
2779 | single access expression for the array, calling | |
2780 | find_or_generate_expression to build the innermost pieces. | |
2781 | ||
2782 | This function is a subroutine of create_expression_by_pieces, and | |
2783 | should not be called on it's own unless you really know what you | |
2784 | are doing. */ | |
2785 | ||
2786 | static tree | |
2787 | create_component_ref_by_pieces (basic_block block, vn_reference_t ref, | |
e076271b | 2788 | gimple_seq *stmts) |
ce94d354 RG |
2789 | { |
2790 | unsigned int op = 0; | |
e076271b | 2791 | return create_component_ref_by_pieces_1 (block, ref, &op, stmts); |
ce94d354 RG |
2792 | } |
2793 | ||
c3dd8dd7 RB |
2794 | /* Find a simple leader for an expression, or generate one using |
2795 | create_expression_by_pieces from a NARY expression for the value. | |
56db793a | 2796 | BLOCK is the basic_block we are looking for leaders in. |
e076271b | 2797 | OP is the tree expression to find a leader for or generate. |
c3dd8dd7 | 2798 | Returns the leader or NULL_TREE on failure. */ |
56db793a DB |
2799 | |
2800 | static tree | |
e076271b | 2801 | find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts) |
56db793a | 2802 | { |
e076271b RG |
2803 | pre_expr expr = get_or_alloc_expr_for (op); |
2804 | unsigned int lookfor = get_expr_value_id (expr); | |
2805 | pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor); | |
c9145754 DB |
2806 | if (leader) |
2807 | { | |
2808 | if (leader->kind == NAME) | |
e076271b | 2809 | return PRE_EXPR_NAME (leader); |
c9145754 | 2810 | else if (leader->kind == CONSTANT) |
e076271b | 2811 | return PRE_EXPR_CONSTANT (leader); |
c3dd8dd7 RB |
2812 | |
2813 | /* Defer. */ | |
2814 | return NULL_TREE; | |
c9145754 | 2815 | } |
e9284566 | 2816 | |
c3dd8dd7 RB |
2817 | /* It must be a complex expression, so generate it recursively. Note |
2818 | that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c | |
2819 | where the insert algorithm fails to insert a required expression. */ | |
9771b263 | 2820 | bitmap exprset = value_expressions[lookfor]; |
e076271b RG |
2821 | bitmap_iterator bi; |
2822 | unsigned int i; | |
2823 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) | |
56db793a | 2824 | { |
e076271b | 2825 | pre_expr temp = expression_for_id (i); |
c3dd8dd7 RB |
2826 | /* We cannot insert random REFERENCE expressions at arbitrary |
2827 | places. We can insert NARYs which eventually re-materializes | |
2828 | its operand values. */ | |
2829 | if (temp->kind == NARY) | |
e076271b RG |
2830 | return create_expression_by_pieces (block, temp, stmts, |
2831 | get_expr_type (expr)); | |
56db793a | 2832 | } |
e076271b | 2833 | |
c3dd8dd7 RB |
2834 | /* Defer. */ |
2835 | return NULL_TREE; | |
56db793a DB |
2836 | } |
2837 | ||
726a989a | 2838 | #define NECESSARY GF_PLF_1 |
c9145754 | 2839 | |
56db793a | 2840 | /* Create an expression in pieces, so that we can handle very complex |
b9c5e484 | 2841 | expressions that may be ANTIC, but not necessary GIMPLE. |
56db793a DB |
2842 | BLOCK is the basic block the expression will be inserted into, |
2843 | EXPR is the expression to insert (in value form) | |
2844 | STMTS is a statement list to append the necessary insertions into. | |
2845 | ||
0e61db61 | 2846 | This function will die if we hit some value that shouldn't be |
56db793a | 2847 | ANTIC but is (IE there is no leader for it, or its components). |
c3dd8dd7 RB |
2848 | The function returns NULL_TREE in case a different antic expression |
2849 | has to be inserted first. | |
56db793a DB |
2850 | This function may also generate expressions that are themselves |
2851 | partially or fully redundant. Those that are will be either made | |
2852 | fully redundant during the next iteration of insert (for partially | |
2853 | redundant ones), or eliminated by eliminate (for fully redundant | |
c3dd8dd7 | 2854 | ones). */ |
56db793a DB |
2855 | |
2856 | static tree | |
726a989a | 2857 | create_expression_by_pieces (basic_block block, pre_expr expr, |
e076271b | 2858 | gimple_seq *stmts, tree type) |
56db793a | 2859 | { |
83d5977e | 2860 | tree name; |
150e3929 RG |
2861 | tree folded; |
2862 | gimple_seq forced_stmts = NULL; | |
c9145754 | 2863 | unsigned int value_id; |
726a989a | 2864 | gimple_stmt_iterator gsi; |
c9145754 DB |
2865 | tree exprtype = type ? type : get_expr_type (expr); |
2866 | pre_expr nameexpr; | |
726a989a | 2867 | gimple newstmt; |
81c4f554 | 2868 | |
c9145754 | 2869 | switch (expr->kind) |
56db793a | 2870 | { |
c9145754 DB |
2871 | /* We may hit the NAME/CONSTANT case if we have to convert types |
2872 | that value numbering saw through. */ | |
2873 | case NAME: | |
2874 | folded = PRE_EXPR_NAME (expr); | |
2875 | break; | |
2876 | case CONSTANT: | |
2877 | folded = PRE_EXPR_CONSTANT (expr); | |
2878 | break; | |
2879 | case REFERENCE: | |
43da81be | 2880 | { |
c9145754 | 2881 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); |
e076271b | 2882 | folded = create_component_ref_by_pieces (block, ref, stmts); |
c3dd8dd7 RB |
2883 | if (!folded) |
2884 | return NULL_TREE; | |
43da81be DB |
2885 | } |
2886 | break; | |
c9145754 | 2887 | case NARY: |
c90186eb | 2888 | { |
c9145754 | 2889 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); |
f843144b | 2890 | tree *genop = XALLOCAVEC (tree, nary->length); |
5a7d7f9c RG |
2891 | unsigned i; |
2892 | for (i = 0; i < nary->length; ++i) | |
85300b46 | 2893 | { |
e076271b | 2894 | genop[i] = find_or_generate_expression (block, nary->op[i], stmts); |
c3dd8dd7 RB |
2895 | if (!genop[i]) |
2896 | return NULL_TREE; | |
48acf1b7 RG |
2897 | /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It |
2898 | may have conversions stripped. */ | |
2899 | if (nary->opcode == POINTER_PLUS_EXPR) | |
2900 | { | |
2901 | if (i == 0) | |
2902 | genop[i] = fold_convert (nary->type, genop[i]); | |
2903 | else if (i == 1) | |
2904 | genop[i] = convert_to_ptrofftype (genop[i]); | |
2905 | } | |
5a7d7f9c RG |
2906 | else |
2907 | genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]); | |
2908 | } | |
2909 | if (nary->opcode == CONSTRUCTOR) | |
2910 | { | |
9771b263 | 2911 | vec<constructor_elt, va_gc> *elts = NULL; |
5a7d7f9c RG |
2912 | for (i = 0; i < nary->length; ++i) |
2913 | CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]); | |
2914 | folded = build_constructor (nary->type, elts); | |
2915 | } | |
2916 | else | |
2917 | { | |
2918 | switch (nary->length) | |
2919 | { | |
2920 | case 1: | |
2921 | folded = fold_build1 (nary->opcode, nary->type, | |
2922 | genop[0]); | |
2923 | break; | |
2924 | case 2: | |
2925 | folded = fold_build2 (nary->opcode, nary->type, | |
2926 | genop[0], genop[1]); | |
2927 | break; | |
2928 | case 3: | |
2929 | folded = fold_build3 (nary->opcode, nary->type, | |
a475fd3d | 2930 | genop[0], genop[1], genop[2]); |
5a7d7f9c RG |
2931 | break; |
2932 | default: | |
2933 | gcc_unreachable (); | |
2934 | } | |
85300b46 | 2935 | } |
56db793a | 2936 | } |
c9145754 | 2937 | break; |
56db793a | 2938 | default: |
e076271b | 2939 | gcc_unreachable (); |
56db793a | 2940 | } |
150e3929 RG |
2941 | |
2942 | if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded))) | |
2943 | folded = fold_convert (exprtype, folded); | |
2944 | ||
81c4f554 SB |
2945 | /* Force the generated expression to be a sequence of GIMPLE |
2946 | statements. | |
2947 | We have to call unshare_expr because force_gimple_operand may | |
2948 | modify the tree we pass to it. */ | |
150e3929 RG |
2949 | folded = force_gimple_operand (unshare_expr (folded), &forced_stmts, |
2950 | false, NULL); | |
81c4f554 SB |
2951 | |
2952 | /* If we have any intermediate expressions to the value sets, add them | |
a7849637 | 2953 | to the value sets and chain them in the instruction stream. */ |
81c4f554 SB |
2954 | if (forced_stmts) |
2955 | { | |
726a989a RB |
2956 | gsi = gsi_start (forced_stmts); |
2957 | for (; !gsi_end_p (gsi); gsi_next (&gsi)) | |
81c4f554 | 2958 | { |
726a989a RB |
2959 | gimple stmt = gsi_stmt (gsi); |
2960 | tree forcedname = gimple_get_lhs (stmt); | |
c9145754 | 2961 | pre_expr nameexpr; |
b9c5e484 | 2962 | |
c9145754 DB |
2963 | if (TREE_CODE (forcedname) == SSA_NAME) |
2964 | { | |
0ab555de | 2965 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname)); |
c9145754 DB |
2966 | VN_INFO_GET (forcedname)->valnum = forcedname; |
2967 | VN_INFO (forcedname)->value_id = get_next_value_id (); | |
2968 | nameexpr = get_or_alloc_expr_for_name (forcedname); | |
2969 | add_to_value (VN_INFO (forcedname)->value_id, nameexpr); | |
40b178f4 | 2970 | bitmap_value_replace_in_set (NEW_SETS (block), nameexpr); |
c9145754 DB |
2971 | bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr); |
2972 | } | |
81c4f554 | 2973 | } |
726a989a | 2974 | gimple_seq_add_seq (stmts, forced_stmts); |
81c4f554 SB |
2975 | } |
2976 | ||
83d5977e RG |
2977 | name = make_temp_ssa_name (exprtype, NULL, "pretmp"); |
2978 | newstmt = gimple_build_assign (name, folded); | |
726a989a | 2979 | gimple_set_plf (newstmt, NECESSARY, false); |
c90186eb | 2980 | |
726a989a | 2981 | gimple_seq_add_stmt (stmts, newstmt); |
0ab555de | 2982 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name)); |
cfaab3a9 | 2983 | |
0aa16586 RG |
2984 | /* Fold the last statement. */ |
2985 | gsi = gsi_last (*stmts); | |
748c5114 RG |
2986 | if (fold_stmt_inplace (&gsi)) |
2987 | update_stmt (gsi_stmt (gsi)); | |
0aa16586 | 2988 | |
c9145754 | 2989 | /* Add a value number to the temporary. |
81c4f554 | 2990 | The value may already exist in either NEW_SETS, or AVAIL_OUT, because |
e9284566 DB |
2991 | we are creating the expression by pieces, and this particular piece of |
2992 | the expression may have been represented. There is no harm in replacing | |
2993 | here. */ | |
c9145754 | 2994 | value_id = get_expr_value_id (expr); |
40b178f4 RG |
2995 | VN_INFO_GET (name)->value_id = value_id; |
2996 | VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id); | |
2997 | if (VN_INFO (name)->valnum == NULL_TREE) | |
2998 | VN_INFO (name)->valnum = name; | |
2999 | gcc_assert (VN_INFO (name)->valnum != NULL_TREE); | |
c9145754 DB |
3000 | nameexpr = get_or_alloc_expr_for_name (name); |
3001 | add_to_value (value_id, nameexpr); | |
70f34814 | 3002 | if (NEW_SETS (block)) |
c9145754 DB |
3003 | bitmap_value_replace_in_set (NEW_SETS (block), nameexpr); |
3004 | bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr); | |
81c4f554 SB |
3005 | |
3006 | pre_stats.insertions++; | |
56db793a | 3007 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b9c5e484 | 3008 | { |
56db793a | 3009 | fprintf (dump_file, "Inserted "); |
726a989a | 3010 | print_gimple_stmt (dump_file, newstmt, 0, 0); |
984af6ac RB |
3011 | fprintf (dump_file, " in predecessor %d (%04d)\n", |
3012 | block->index, value_id); | |
56db793a | 3013 | } |
81c4f554 | 3014 | |
56db793a DB |
3015 | return name; |
3016 | } | |
e9284566 | 3017 | |
c9145754 | 3018 | |
83737db2 | 3019 | /* Insert the to-be-made-available values of expression EXPRNUM for each |
c90186eb | 3020 | predecessor, stored in AVAIL, into the predecessors of BLOCK, and |
c9145754 | 3021 | merge the result with a phi node, given the same value number as |
c90186eb | 3022 | NODE. Return true if we have inserted new stuff. */ |
e9284566 DB |
3023 | |
3024 | static bool | |
83737db2 | 3025 | insert_into_preds_of_block (basic_block block, unsigned int exprnum, |
9771b263 | 3026 | vec<pre_expr> avail) |
e9284566 | 3027 | { |
c9145754 DB |
3028 | pre_expr expr = expression_for_id (exprnum); |
3029 | pre_expr newphi; | |
3030 | unsigned int val = get_expr_value_id (expr); | |
e9284566 | 3031 | edge pred; |
0fc6c492 DB |
3032 | bool insertions = false; |
3033 | bool nophi = false; | |
e9284566 | 3034 | basic_block bprime; |
c9145754 | 3035 | pre_expr eprime; |
e9284566 | 3036 | edge_iterator ei; |
c9145754 | 3037 | tree type = get_expr_type (expr); |
de081cfd | 3038 | tree temp; |
726a989a | 3039 | gimple phi; |
b9c5e484 | 3040 | |
0fc6c492 | 3041 | /* Make sure we aren't creating an induction variable. */ |
391886c8 | 3042 | if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2) |
0fc6c492 DB |
3043 | { |
3044 | bool firstinsideloop = false; | |
3045 | bool secondinsideloop = false; | |
b9c5e484 | 3046 | firstinsideloop = flow_bb_inside_loop_p (block->loop_father, |
0fc6c492 DB |
3047 | EDGE_PRED (block, 0)->src); |
3048 | secondinsideloop = flow_bb_inside_loop_p (block->loop_father, | |
3049 | EDGE_PRED (block, 1)->src); | |
3050 | /* Induction variables only have one edge inside the loop. */ | |
a8338640 | 3051 | if ((firstinsideloop ^ secondinsideloop) |
cddaefa3 | 3052 | && expr->kind != REFERENCE) |
0fc6c492 DB |
3053 | { |
3054 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3055 | fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n"); | |
3056 | nophi = true; | |
3057 | } | |
3058 | } | |
b9c5e484 | 3059 | |
c9145754 DB |
3060 | /* Make the necessary insertions. */ |
3061 | FOR_EACH_EDGE (pred, ei, block->preds) | |
3062 | { | |
726a989a | 3063 | gimple_seq stmts = NULL; |
c9145754 DB |
3064 | tree builtexpr; |
3065 | bprime = pred->src; | |
9771b263 | 3066 | eprime = avail[pred->dest_idx]; |
c9145754 DB |
3067 | |
3068 | if (eprime->kind != NAME && eprime->kind != CONSTANT) | |
3069 | { | |
e076271b RG |
3070 | builtexpr = create_expression_by_pieces (bprime, eprime, |
3071 | &stmts, type); | |
c9145754 | 3072 | gcc_assert (!(pred->flags & EDGE_ABNORMAL)); |
726a989a | 3073 | gsi_insert_seq_on_edge (pred, stmts); |
c3dd8dd7 RB |
3074 | if (!builtexpr) |
3075 | { | |
3076 | /* We cannot insert a PHI node if we failed to insert | |
3077 | on one edge. */ | |
3078 | nophi = true; | |
3079 | continue; | |
3080 | } | |
9771b263 | 3081 | avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr); |
c9145754 DB |
3082 | insertions = true; |
3083 | } | |
3084 | else if (eprime->kind == CONSTANT) | |
3085 | { | |
3086 | /* Constants may not have the right type, fold_convert | |
1c8f7377 | 3087 | should give us back a constant with the right type. */ |
c9145754 | 3088 | tree constant = PRE_EXPR_CONSTANT (eprime); |
15d5fe33 | 3089 | if (!useless_type_conversion_p (type, TREE_TYPE (constant))) |
c9145754 DB |
3090 | { |
3091 | tree builtexpr = fold_convert (type, constant); | |
b8698a0f | 3092 | if (!is_gimple_min_invariant (builtexpr)) |
c9145754 DB |
3093 | { |
3094 | tree forcedexpr = force_gimple_operand (builtexpr, | |
3095 | &stmts, true, | |
3096 | NULL); | |
15d5fe33 | 3097 | if (!is_gimple_min_invariant (forcedexpr)) |
c9145754 DB |
3098 | { |
3099 | if (forcedexpr != builtexpr) | |
3100 | { | |
3101 | VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime); | |
3102 | VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime); | |
3103 | } | |
3104 | if (stmts) | |
3105 | { | |
726a989a RB |
3106 | gimple_stmt_iterator gsi; |
3107 | gsi = gsi_start (stmts); | |
3108 | for (; !gsi_end_p (gsi); gsi_next (&gsi)) | |
c9145754 | 3109 | { |
726a989a | 3110 | gimple stmt = gsi_stmt (gsi); |
0ab555de MM |
3111 | tree lhs = gimple_get_lhs (stmt); |
3112 | if (TREE_CODE (lhs) == SSA_NAME) | |
3113 | bitmap_set_bit (inserted_exprs, | |
3114 | SSA_NAME_VERSION (lhs)); | |
726a989a | 3115 | gimple_set_plf (stmt, NECESSARY, false); |
c9145754 | 3116 | } |
726a989a | 3117 | gsi_insert_seq_on_edge (pred, stmts); |
c9145754 | 3118 | } |
9771b263 DN |
3119 | avail[pred->dest_idx] |
3120 | = get_or_alloc_expr_for_name (forcedexpr); | |
c9145754 DB |
3121 | } |
3122 | } | |
70f34814 | 3123 | else |
9771b263 DN |
3124 | avail[pred->dest_idx] |
3125 | = get_or_alloc_expr_for_constant (builtexpr); | |
c9145754 DB |
3126 | } |
3127 | } | |
3128 | else if (eprime->kind == NAME) | |
3129 | { | |
3130 | /* We may have to do a conversion because our value | |
3131 | numbering can look through types in certain cases, but | |
3132 | our IL requires all operands of a phi node have the same | |
3133 | type. */ | |
3134 | tree name = PRE_EXPR_NAME (eprime); | |
f709638a | 3135 | if (!useless_type_conversion_p (type, TREE_TYPE (name))) |
c9145754 DB |
3136 | { |
3137 | tree builtexpr; | |
3138 | tree forcedexpr; | |
f709638a | 3139 | builtexpr = fold_convert (type, name); |
c9145754 DB |
3140 | forcedexpr = force_gimple_operand (builtexpr, |
3141 | &stmts, true, | |
3142 | NULL); | |
3143 | ||
3144 | if (forcedexpr != name) | |
3145 | { | |
3146 | VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum; | |
3147 | VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id; | |
3148 | } | |
c90186eb | 3149 | |
c9145754 DB |
3150 | if (stmts) |
3151 | { | |
726a989a RB |
3152 | gimple_stmt_iterator gsi; |
3153 | gsi = gsi_start (stmts); | |
3154 | for (; !gsi_end_p (gsi); gsi_next (&gsi)) | |
c9145754 | 3155 | { |
726a989a | 3156 | gimple stmt = gsi_stmt (gsi); |
0ab555de MM |
3157 | tree lhs = gimple_get_lhs (stmt); |
3158 | if (TREE_CODE (lhs) == SSA_NAME) | |
3159 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs)); | |
726a989a | 3160 | gimple_set_plf (stmt, NECESSARY, false); |
c9145754 | 3161 | } |
726a989a | 3162 | gsi_insert_seq_on_edge (pred, stmts); |
c9145754 | 3163 | } |
9771b263 | 3164 | avail[pred->dest_idx] = get_or_alloc_expr_for_name (forcedexpr); |
c9145754 | 3165 | } |
b9c5e484 | 3166 | } |
e9284566 | 3167 | } |
0fc6c492 DB |
3168 | /* If we didn't want a phi node, and we made insertions, we still have |
3169 | inserted new stuff, and thus return true. If we didn't want a phi node, | |
3170 | and didn't make insertions, we haven't added anything new, so return | |
3171 | false. */ | |
3172 | if (nophi && insertions) | |
3173 | return true; | |
3174 | else if (nophi && !insertions) | |
3175 | return false; | |
3176 | ||
e9284566 | 3177 | /* Now build a phi for the new variable. */ |
83d5977e | 3178 | temp = make_temp_ssa_name (type, NULL, "prephitmp"); |
1e52075c | 3179 | phi = create_phi_node (temp, block); |
de081cfd RG |
3180 | |
3181 | gimple_set_plf (phi, NECESSARY, false); | |
40b178f4 RG |
3182 | VN_INFO_GET (temp)->value_id = val; |
3183 | VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val); | |
3184 | if (VN_INFO (temp)->valnum == NULL_TREE) | |
3185 | VN_INFO (temp)->valnum = temp; | |
3186 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp)); | |
e9284566 | 3187 | FOR_EACH_EDGE (pred, ei, block->preds) |
c9145754 | 3188 | { |
9771b263 | 3189 | pre_expr ae = avail[pred->dest_idx]; |
c9145754 DB |
3190 | gcc_assert (get_expr_type (ae) == type |
3191 | || useless_type_conversion_p (type, get_expr_type (ae))); | |
3192 | if (ae->kind == CONSTANT) | |
2724573f RB |
3193 | add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)), |
3194 | pred, UNKNOWN_LOCATION); | |
c9145754 | 3195 | else |
1c8f7377 | 3196 | add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION); |
c9145754 | 3197 | } |
b9c5e484 | 3198 | |
40b178f4 | 3199 | newphi = get_or_alloc_expr_for_name (temp); |
c9145754 | 3200 | add_to_value (val, newphi); |
b9c5e484 | 3201 | |
e9284566 DB |
3202 | /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing |
3203 | this insertion, since we test for the existence of this value in PHI_GEN | |
3204 | before proceeding with the partial redundancy checks in insert_aux. | |
b9c5e484 | 3205 | |
e9284566 DB |
3206 | The value may exist in AVAIL_OUT, in particular, it could be represented |
3207 | by the expression we are trying to eliminate, in which case we want the | |
3208 | replacement to occur. If it's not existing in AVAIL_OUT, we want it | |
3209 | inserted there. | |
b9c5e484 | 3210 | |
e9284566 DB |
3211 | Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of |
3212 | this block, because if it did, it would have existed in our dominator's | |
3213 | AVAIL_OUT, and would have been skipped due to the full redundancy check. | |
3214 | */ | |
3215 | ||
c9145754 | 3216 | bitmap_insert_into_set (PHI_GEN (block), newphi); |
b9c5e484 | 3217 | bitmap_value_replace_in_set (AVAIL_OUT (block), |
c9145754 | 3218 | newphi); |
e9284566 | 3219 | bitmap_insert_into_set (NEW_SETS (block), |
c9145754 | 3220 | newphi); |
b9c5e484 | 3221 | |
e9284566 DB |
3222 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3223 | { | |
3224 | fprintf (dump_file, "Created phi "); | |
726a989a | 3225 | print_gimple_stmt (dump_file, phi, 0, 0); |
984af6ac | 3226 | fprintf (dump_file, " in block %d (%04d)\n", block->index, val); |
e9284566 DB |
3227 | } |
3228 | pre_stats.phis++; | |
3229 | return true; | |
3230 | } | |
3231 | ||
3232 | ||
b9c5e484 | 3233 | |
7e6eb623 DB |
3234 | /* Perform insertion of partially redundant values. |
3235 | For BLOCK, do the following: | |
3236 | 1. Propagate the NEW_SETS of the dominator into the current block. | |
b9c5e484 | 3237 | If the block has multiple predecessors, |
7e6eb623 | 3238 | 2a. Iterate over the ANTIC expressions for the block to see if |
83737db2 | 3239 | any of them are partially redundant. |
7e6eb623 | 3240 | 2b. If so, insert them into the necessary predecessors to make |
83737db2 | 3241 | the expression fully redundant. |
7e6eb623 DB |
3242 | 2c. Insert a new PHI merging the values of the predecessors. |
3243 | 2d. Insert the new PHI, and the new expressions, into the | |
83737db2 | 3244 | NEW_SETS set. |
7e6eb623 | 3245 | 3. Recursively call ourselves on the dominator children of BLOCK. |
6de9cd9a | 3246 | |
83737db2 | 3247 | Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by |
d75dbccd | 3248 | do_regular_insertion and do_partial_insertion. |
83737db2 | 3249 | |
7e6eb623 | 3250 | */ |
e9284566 | 3251 | |
83737db2 DB |
3252 | static bool |
3253 | do_regular_insertion (basic_block block, basic_block dom) | |
3254 | { | |
3255 | bool new_stuff = false; | |
9771b263 | 3256 | vec<pre_expr> exprs; |
c9145754 | 3257 | pre_expr expr; |
6e1aa848 | 3258 | vec<pre_expr> avail = vNULL; |
83737db2 DB |
3259 | int i; |
3260 | ||
1c8f7377 | 3261 | exprs = sorted_array_from_bitmap_set (ANTIC_IN (block)); |
9771b263 | 3262 | avail.safe_grow (EDGE_COUNT (block->preds)); |
1c8f7377 | 3263 | |
9771b263 | 3264 | FOR_EACH_VEC_ELT (exprs, i, expr) |
83737db2 | 3265 | { |
4bcc5786 RB |
3266 | if (expr->kind == NARY |
3267 | || expr->kind == REFERENCE) | |
83737db2 | 3268 | { |
c9145754 | 3269 | unsigned int val; |
83737db2 DB |
3270 | bool by_some = false; |
3271 | bool cant_insert = false; | |
3272 | bool all_same = true; | |
c9145754 | 3273 | pre_expr first_s = NULL; |
83737db2 DB |
3274 | edge pred; |
3275 | basic_block bprime; | |
c9145754 | 3276 | pre_expr eprime = NULL; |
83737db2 | 3277 | edge_iterator ei; |
f11d2f1e | 3278 | pre_expr edoubleprime = NULL; |
5813994e | 3279 | bool do_insertion = false; |
83737db2 | 3280 | |
c9145754 | 3281 | val = get_expr_value_id (expr); |
83737db2 DB |
3282 | if (bitmap_set_contains_value (PHI_GEN (block), val)) |
3283 | continue; | |
3284 | if (bitmap_set_contains_value (AVAIL_OUT (dom), val)) | |
3285 | { | |
3286 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
ead69dac JL |
3287 | { |
3288 | fprintf (dump_file, "Found fully redundant value: "); | |
3289 | print_pre_expr (dump_file, expr); | |
3290 | fprintf (dump_file, "\n"); | |
3291 | } | |
83737db2 DB |
3292 | continue; |
3293 | } | |
3294 | ||
83737db2 DB |
3295 | FOR_EACH_EDGE (pred, ei, block->preds) |
3296 | { | |
c9145754 | 3297 | unsigned int vprime; |
83737db2 | 3298 | |
c4ab2baa RG |
3299 | /* We should never run insertion for the exit block |
3300 | and so not come across fake pred edges. */ | |
3301 | gcc_assert (!(pred->flags & EDGE_FAKE)); | |
83737db2 DB |
3302 | bprime = pred->src; |
3303 | eprime = phi_translate (expr, ANTIC_IN (block), NULL, | |
3304 | bprime, block); | |
3305 | ||
3306 | /* eprime will generally only be NULL if the | |
3307 | value of the expression, translated | |
3308 | through the PHI for this predecessor, is | |
3309 | undefined. If that is the case, we can't | |
3310 | make the expression fully redundant, | |
3311 | because its value is undefined along a | |
3312 | predecessor path. We can thus break out | |
3313 | early because it doesn't matter what the | |
3314 | rest of the results are. */ | |
3315 | if (eprime == NULL) | |
3316 | { | |
9771b263 | 3317 | avail[pred->dest_idx] = NULL; |
83737db2 DB |
3318 | cant_insert = true; |
3319 | break; | |
3320 | } | |
3321 | ||
3322 | eprime = fully_constant_expression (eprime); | |
726a989a RB |
3323 | vprime = get_expr_value_id (eprime); |
3324 | edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), | |
e076271b | 3325 | vprime); |
83737db2 DB |
3326 | if (edoubleprime == NULL) |
3327 | { | |
9771b263 | 3328 | avail[pred->dest_idx] = eprime; |
83737db2 DB |
3329 | all_same = false; |
3330 | } | |
3331 | else | |
3332 | { | |
9771b263 | 3333 | avail[pred->dest_idx] = edoubleprime; |
83737db2 | 3334 | by_some = true; |
5813994e RG |
3335 | /* We want to perform insertions to remove a redundancy on |
3336 | a path in the CFG we want to optimize for speed. */ | |
3337 | if (optimize_edge_for_speed_p (pred)) | |
3338 | do_insertion = true; | |
83737db2 DB |
3339 | if (first_s == NULL) |
3340 | first_s = edoubleprime; | |
5deac340 | 3341 | else if (!pre_expr_d::equal (first_s, edoubleprime)) |
83737db2 DB |
3342 | all_same = false; |
3343 | } | |
3344 | } | |
3345 | /* If we can insert it, it's not the same value | |
3346 | already existing along every predecessor, and | |
3347 | it's defined by some predecessor, it is | |
3348 | partially redundant. */ | |
3bc27de7 | 3349 | if (!cant_insert && !all_same && by_some) |
83737db2 | 3350 | { |
3bc27de7 RG |
3351 | if (!do_insertion) |
3352 | { | |
3353 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3354 | { | |
3355 | fprintf (dump_file, "Skipping partial redundancy for " | |
3356 | "expression "); | |
3357 | print_pre_expr (dump_file, expr); | |
3358 | fprintf (dump_file, " (%04d), no redundancy on to be " | |
3359 | "optimized for speed edge\n", val); | |
3360 | } | |
3361 | } | |
19372838 RG |
3362 | else if (dbg_cnt (treepre_insert)) |
3363 | { | |
3364 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3365 | { | |
3366 | fprintf (dump_file, "Found partial redundancy for " | |
3367 | "expression "); | |
3368 | print_pre_expr (dump_file, expr); | |
3369 | fprintf (dump_file, " (%04d)\n", | |
3370 | get_expr_value_id (expr)); | |
3371 | } | |
3372 | if (insert_into_preds_of_block (block, | |
3373 | get_expression_id (expr), | |
3374 | avail)) | |
3375 | new_stuff = true; | |
3376 | } | |
83737db2 DB |
3377 | } |
3378 | /* If all edges produce the same value and that value is | |
3379 | an invariant, then the PHI has the same value on all | |
3380 | edges. Note this. */ | |
4bcc5786 | 3381 | else if (!cant_insert && all_same) |
83737db2 | 3382 | { |
4bcc5786 RB |
3383 | gcc_assert (edoubleprime->kind == CONSTANT |
3384 | || edoubleprime->kind == NAME); | |
3385 | ||
3386 | tree temp = make_temp_ssa_name (get_expr_type (expr), | |
3387 | NULL, "pretmp"); | |
3388 | gimple assign = gimple_build_assign (temp, | |
3389 | edoubleprime->kind == CONSTANT ? PRE_EXPR_CONSTANT (edoubleprime) : PRE_EXPR_NAME (edoubleprime)); | |
3390 | gimple_stmt_iterator gsi = gsi_after_labels (block); | |
3391 | gsi_insert_before (&gsi, assign, GSI_NEW_STMT); | |
3392 | ||
3393 | gimple_set_plf (assign, NECESSARY, false); | |
3394 | VN_INFO_GET (temp)->value_id = val; | |
3395 | VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val); | |
3396 | if (VN_INFO (temp)->valnum == NULL_TREE) | |
3397 | VN_INFO (temp)->valnum = temp; | |
3398 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp)); | |
3399 | pre_expr newe = get_or_alloc_expr_for_name (temp); | |
3400 | add_to_value (val, newe); | |
3401 | bitmap_value_replace_in_set (AVAIL_OUT (block), newe); | |
3402 | bitmap_insert_into_set (NEW_SETS (block), newe); | |
83737db2 | 3403 | } |
83737db2 DB |
3404 | } |
3405 | } | |
3406 | ||
9771b263 | 3407 | exprs.release (); |
83737db2 DB |
3408 | return new_stuff; |
3409 | } | |
3410 | ||
3411 | ||
d75dbccd DB |
3412 | /* Perform insertion for partially anticipatable expressions. There |
3413 | is only one case we will perform insertion for these. This case is | |
3414 | if the expression is partially anticipatable, and fully available. | |
3415 | In this case, we know that putting it earlier will enable us to | |
3416 | remove the later computation. */ | |
3417 | ||
3418 | ||
3419 | static bool | |
3420 | do_partial_partial_insertion (basic_block block, basic_block dom) | |
3421 | { | |
3422 | bool new_stuff = false; | |
9771b263 | 3423 | vec<pre_expr> exprs; |
c9145754 | 3424 | pre_expr expr; |
ef062b13 | 3425 | auto_vec<pre_expr> avail; |
d75dbccd DB |
3426 | int i; |
3427 | ||
1c8f7377 | 3428 | exprs = sorted_array_from_bitmap_set (PA_IN (block)); |
9771b263 | 3429 | avail.safe_grow (EDGE_COUNT (block->preds)); |
1c8f7377 | 3430 | |
9771b263 | 3431 | FOR_EACH_VEC_ELT (exprs, i, expr) |
d75dbccd | 3432 | { |
4bcc5786 RB |
3433 | if (expr->kind == NARY |
3434 | || expr->kind == REFERENCE) | |
d75dbccd | 3435 | { |
c9145754 | 3436 | unsigned int val; |
d75dbccd DB |
3437 | bool by_all = true; |
3438 | bool cant_insert = false; | |
3439 | edge pred; | |
3440 | basic_block bprime; | |
c9145754 | 3441 | pre_expr eprime = NULL; |
d75dbccd DB |
3442 | edge_iterator ei; |
3443 | ||
c9145754 | 3444 | val = get_expr_value_id (expr); |
d75dbccd DB |
3445 | if (bitmap_set_contains_value (PHI_GEN (block), val)) |
3446 | continue; | |
3447 | if (bitmap_set_contains_value (AVAIL_OUT (dom), val)) | |
3448 | continue; | |
3449 | ||
d75dbccd DB |
3450 | FOR_EACH_EDGE (pred, ei, block->preds) |
3451 | { | |
c9145754 DB |
3452 | unsigned int vprime; |
3453 | pre_expr edoubleprime; | |
d75dbccd | 3454 | |
c4ab2baa RG |
3455 | /* We should never run insertion for the exit block |
3456 | and so not come across fake pred edges. */ | |
3457 | gcc_assert (!(pred->flags & EDGE_FAKE)); | |
d75dbccd DB |
3458 | bprime = pred->src; |
3459 | eprime = phi_translate (expr, ANTIC_IN (block), | |
3460 | PA_IN (block), | |
3461 | bprime, block); | |
3462 | ||
3463 | /* eprime will generally only be NULL if the | |
3464 | value of the expression, translated | |
3465 | through the PHI for this predecessor, is | |
3466 | undefined. If that is the case, we can't | |
3467 | make the expression fully redundant, | |
3468 | because its value is undefined along a | |
3469 | predecessor path. We can thus break out | |
3470 | early because it doesn't matter what the | |
3471 | rest of the results are. */ | |
3472 | if (eprime == NULL) | |
3473 | { | |
9771b263 | 3474 | avail[pred->dest_idx] = NULL; |
d75dbccd DB |
3475 | cant_insert = true; |
3476 | break; | |
3477 | } | |
3478 | ||
3479 | eprime = fully_constant_expression (eprime); | |
726a989a | 3480 | vprime = get_expr_value_id (eprime); |
e076271b | 3481 | edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime); |
9771b263 | 3482 | avail[pred->dest_idx] = edoubleprime; |
d75dbccd DB |
3483 | if (edoubleprime == NULL) |
3484 | { | |
3485 | by_all = false; | |
3486 | break; | |
3487 | } | |
d75dbccd DB |
3488 | } |
3489 | ||
3490 | /* If we can insert it, it's not the same value | |
3491 | already existing along every predecessor, and | |
3492 | it's defined by some predecessor, it is | |
3493 | partially redundant. */ | |
fa06ad0d | 3494 | if (!cant_insert && by_all) |
d75dbccd | 3495 | { |
fa06ad0d JR |
3496 | edge succ; |
3497 | bool do_insertion = false; | |
3498 | ||
3499 | /* Insert only if we can remove a later expression on a path | |
3500 | that we want to optimize for speed. | |
3501 | The phi node that we will be inserting in BLOCK is not free, | |
3502 | and inserting it for the sake of !optimize_for_speed successor | |
3503 | may cause regressions on the speed path. */ | |
3504 | FOR_EACH_EDGE (succ, ei, block->succs) | |
3505 | { | |
0c46ead2 JH |
3506 | if (bitmap_set_contains_value (PA_IN (succ->dest), val) |
3507 | || bitmap_set_contains_value (ANTIC_IN (succ->dest), val)) | |
fa06ad0d JR |
3508 | { |
3509 | if (optimize_edge_for_speed_p (succ)) | |
3510 | do_insertion = true; | |
3511 | } | |
3512 | } | |
3513 | ||
3514 | if (!do_insertion) | |
3515 | { | |
3516 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3517 | { | |
3518 | fprintf (dump_file, "Skipping partial partial redundancy " | |
3519 | "for expression "); | |
3520 | print_pre_expr (dump_file, expr); | |
0c46ead2 | 3521 | fprintf (dump_file, " (%04d), not (partially) anticipated " |
fa06ad0d JR |
3522 | "on any to be optimized for speed edges\n", val); |
3523 | } | |
3524 | } | |
3525 | else if (dbg_cnt (treepre_insert)) | |
3526 | { | |
3527 | pre_stats.pa_insert++; | |
19372838 RG |
3528 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3529 | { | |
3530 | fprintf (dump_file, "Found partial partial redundancy " | |
3531 | "for expression "); | |
3532 | print_pre_expr (dump_file, expr); | |
3533 | fprintf (dump_file, " (%04d)\n", | |
3534 | get_expr_value_id (expr)); | |
3535 | } | |
fa06ad0d JR |
3536 | if (insert_into_preds_of_block (block, |
3537 | get_expression_id (expr), | |
3538 | avail)) | |
3539 | new_stuff = true; | |
3540 | } | |
3541 | } | |
d75dbccd DB |
3542 | } |
3543 | } | |
3544 | ||
9771b263 | 3545 | exprs.release (); |
d75dbccd DB |
3546 | return new_stuff; |
3547 | } | |
83737db2 | 3548 | |
7e6eb623 DB |
3549 | static bool |
3550 | insert_aux (basic_block block) | |
6de9cd9a | 3551 | { |
7e6eb623 DB |
3552 | basic_block son; |
3553 | bool new_stuff = false; | |
6de9cd9a | 3554 | |
7e6eb623 | 3555 | if (block) |
6de9cd9a | 3556 | { |
7e6eb623 DB |
3557 | basic_block dom; |
3558 | dom = get_immediate_dominator (CDI_DOMINATORS, block); | |
3559 | if (dom) | |
a32b97a2 | 3560 | { |
3cd8c58a | 3561 | unsigned i; |
87c476a2 | 3562 | bitmap_iterator bi; |
6b416da1 | 3563 | bitmap_set_t newset = NEW_SETS (dom); |
e9284566 | 3564 | if (newset) |
87c476a2 | 3565 | { |
e9284566 DB |
3566 | /* Note that we need to value_replace both NEW_SETS, and |
3567 | AVAIL_OUT. For both the case of NEW_SETS, the value may be | |
3568 | represented by some non-simple expression here that we want | |
3569 | to replace it with. */ | |
83737db2 | 3570 | FOR_EACH_EXPR_ID_IN_SET (newset, i, bi) |
e9284566 | 3571 | { |
c9145754 | 3572 | pre_expr expr = expression_for_id (i); |
83737db2 DB |
3573 | bitmap_value_replace_in_set (NEW_SETS (block), expr); |
3574 | bitmap_value_replace_in_set (AVAIL_OUT (block), expr); | |
e9284566 | 3575 | } |
87c476a2 | 3576 | } |
c5cbcccf | 3577 | if (!single_pred_p (block)) |
6de9cd9a | 3578 | { |
83737db2 | 3579 | new_stuff |= do_regular_insertion (block, dom); |
d75dbccd DB |
3580 | if (do_partial_partial) |
3581 | new_stuff |= do_partial_partial_insertion (block, dom); | |
6de9cd9a DN |
3582 | } |
3583 | } | |
3584 | } | |
7e6eb623 DB |
3585 | for (son = first_dom_son (CDI_DOMINATORS, block); |
3586 | son; | |
3587 | son = next_dom_son (CDI_DOMINATORS, son)) | |
3588 | { | |
3589 | new_stuff |= insert_aux (son); | |
3590 | } | |
3591 | ||
3592 | return new_stuff; | |
6de9cd9a DN |
3593 | } |
3594 | ||
7e6eb623 | 3595 | /* Perform insertion of partially redundant values. */ |
6de9cd9a | 3596 | |
7e6eb623 DB |
3597 | static void |
3598 | insert (void) | |
6de9cd9a | 3599 | { |
7e6eb623 | 3600 | bool new_stuff = true; |
6de9cd9a | 3601 | basic_block bb; |
7e6eb623 | 3602 | int num_iterations = 0; |
b9c5e484 | 3603 | |
04a90bec | 3604 | FOR_ALL_BB_FN (bb, cfun) |
6b416da1 | 3605 | NEW_SETS (bb) = bitmap_set_new (); |
b9c5e484 | 3606 | |
7e6eb623 | 3607 | while (new_stuff) |
6de9cd9a | 3608 | { |
7e6eb623 | 3609 | num_iterations++; |
19372838 RG |
3610 | if (dump_file && dump_flags & TDF_DETAILS) |
3611 | fprintf (dump_file, "Starting insert iteration %d\n", num_iterations); | |
fefa31b5 | 3612 | new_stuff = insert_aux (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
3dbc97a9 RB |
3613 | |
3614 | /* Clear the NEW sets before the next iteration. We have already | |
3615 | fully propagated its contents. */ | |
3616 | if (new_stuff) | |
04a90bec | 3617 | FOR_ALL_BB_FN (bb, cfun) |
3dbc97a9 | 3618 | bitmap_set_free (NEW_SETS (bb)); |
6de9cd9a | 3619 | } |
9fe0cb7d | 3620 | statistics_histogram_event (cfun, "insert iterations", num_iterations); |
7e6eb623 | 3621 | } |
6de9cd9a | 3622 | |
33c94679 | 3623 | |
665fcad8 SB |
3624 | /* Compute the AVAIL set for all basic blocks. |
3625 | ||
3626 | This function performs value numbering of the statements in each basic | |
3627 | block. The AVAIL sets are built from information we glean while doing | |
3628 | this value numbering, since the AVAIL sets contain only one entry per | |
7e6eb623 | 3629 | value. |
b9c5e484 | 3630 | |
7e6eb623 | 3631 | AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)]. |
ff2ad0f7 | 3632 | AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */ |
6de9cd9a | 3633 | |
7e6eb623 | 3634 | static void |
665fcad8 | 3635 | compute_avail (void) |
6de9cd9a | 3636 | { |
c9145754 | 3637 | |
665fcad8 SB |
3638 | basic_block block, son; |
3639 | basic_block *worklist; | |
3640 | size_t sp = 0; | |
b005da11 RG |
3641 | unsigned i; |
3642 | ||
3643 | /* We pretend that default definitions are defined in the entry block. | |
3644 | This includes function arguments and the static chain decl. */ | |
3645 | for (i = 1; i < num_ssa_names; ++i) | |
3646 | { | |
3647 | tree name = ssa_name (i); | |
3648 | pre_expr e; | |
3649 | if (!name | |
3650 | || !SSA_NAME_IS_DEFAULT_DEF (name) | |
3651 | || has_zero_uses (name) | |
ea057359 | 3652 | || virtual_operand_p (name)) |
b005da11 | 3653 | continue; |
665fcad8 | 3654 | |
b005da11 RG |
3655 | e = get_or_alloc_expr_for_name (name); |
3656 | add_to_value (get_expr_value_id (e), e); | |
fefa31b5 DM |
3657 | bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)), e); |
3658 | bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)), | |
3659 | e); | |
2984956b AP |
3660 | } |
3661 | ||
e3815735 RB |
3662 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3663 | { | |
fefa31b5 | 3664 | print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (cfun)), |
e3815735 | 3665 | "tmp_gen", ENTRY_BLOCK); |
fefa31b5 | 3666 | print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (cfun)), |
e3815735 RB |
3667 | "avail_out", ENTRY_BLOCK); |
3668 | } | |
3669 | ||
665fcad8 | 3670 | /* Allocate the worklist. */ |
0cae8d31 | 3671 | worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun)); |
665fcad8 SB |
3672 | |
3673 | /* Seed the algorithm by putting the dominator children of the entry | |
3674 | block on the worklist. */ | |
fefa31b5 | 3675 | for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
665fcad8 SB |
3676 | son; |
3677 | son = next_dom_son (CDI_DOMINATORS, son)) | |
3678 | worklist[sp++] = son; | |
3679 | ||
3680 | /* Loop until the worklist is empty. */ | |
3681 | while (sp) | |
6de9cd9a | 3682 | { |
726a989a RB |
3683 | gimple_stmt_iterator gsi; |
3684 | gimple stmt; | |
7e6eb623 DB |
3685 | basic_block dom; |
3686 | ||
665fcad8 SB |
3687 | /* Pick a block from the worklist. */ |
3688 | block = worklist[--sp]; | |
3689 | ||
ff2ad0f7 DN |
3690 | /* Initially, the set of available values in BLOCK is that of |
3691 | its immediate dominator. */ | |
7e6eb623 DB |
3692 | dom = get_immediate_dominator (CDI_DOMINATORS, block); |
3693 | if (dom) | |
bdee7684 | 3694 | bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom)); |
33c94679 | 3695 | |
ff2ad0f7 | 3696 | /* Generate values for PHI nodes. */ |
726a989a | 3697 | for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi)) |
e3815735 RB |
3698 | { |
3699 | tree result = gimple_phi_result (gsi_stmt (gsi)); | |
3700 | ||
3701 | /* We have no need for virtual phis, as they don't represent | |
3702 | actual computations. */ | |
3703 | if (virtual_operand_p (result)) | |
3704 | continue; | |
3705 | ||
3706 | pre_expr e = get_or_alloc_expr_for_name (result); | |
3707 | add_to_value (get_expr_value_id (e), e); | |
3708 | bitmap_value_insert_into_set (AVAIL_OUT (block), e); | |
3709 | bitmap_insert_into_set (PHI_GEN (block), e); | |
3710 | } | |
7e6eb623 | 3711 | |
a19eb9d2 RG |
3712 | BB_MAY_NOTRETURN (block) = 0; |
3713 | ||
ff2ad0f7 DN |
3714 | /* Now compute value numbers and populate value sets with all |
3715 | the expressions computed in BLOCK. */ | |
726a989a | 3716 | for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi)) |
6de9cd9a | 3717 | { |
f47c96aa AM |
3718 | ssa_op_iter iter; |
3719 | tree op; | |
ff2ad0f7 | 3720 | |
726a989a | 3721 | stmt = gsi_stmt (gsi); |
ff2ad0f7 | 3722 | |
a19eb9d2 RG |
3723 | /* Cache whether the basic-block has any non-visible side-effect |
3724 | or control flow. | |
3725 | If this isn't a call or it is the last stmt in the | |
3726 | basic-block then the CFG represents things correctly. */ | |
9cbbba28 | 3727 | if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt)) |
a19eb9d2 RG |
3728 | { |
3729 | /* Non-looping const functions always return normally. | |
3730 | Otherwise the call might not return or have side-effects | |
3731 | that forbids hoisting possibly trapping expressions | |
3732 | before it. */ | |
3733 | int flags = gimple_call_flags (stmt); | |
3734 | if (!(flags & ECF_CONST) | |
3735 | || (flags & ECF_LOOPING_CONST_OR_PURE)) | |
3736 | BB_MAY_NOTRETURN (block) = 1; | |
3737 | } | |
3738 | ||
c9145754 | 3739 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF) |
83737db2 | 3740 | { |
c9145754 | 3741 | pre_expr e = get_or_alloc_expr_for_name (op); |
83737db2 | 3742 | |
c9145754 | 3743 | add_to_value (get_expr_value_id (e), e); |
40b178f4 | 3744 | bitmap_insert_into_set (TMP_GEN (block), e); |
c9145754 | 3745 | bitmap_value_insert_into_set (AVAIL_OUT (block), e); |
83737db2 DB |
3746 | } |
3747 | ||
17742d62 RG |
3748 | if (gimple_has_side_effects (stmt) |
3749 | || stmt_could_throw_p (stmt) | |
3750 | || is_gimple_debug (stmt)) | |
726a989a RB |
3751 | continue; |
3752 | ||
17742d62 | 3753 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) |
e3815735 RB |
3754 | { |
3755 | if (ssa_undefined_value_p (op)) | |
3756 | continue; | |
3757 | pre_expr e = get_or_alloc_expr_for_name (op); | |
3758 | bitmap_value_insert_into_set (EXP_GEN (block), e); | |
3759 | } | |
17742d62 | 3760 | |
726a989a | 3761 | switch (gimple_code (stmt)) |
7e6eb623 | 3762 | { |
726a989a | 3763 | case GIMPLE_RETURN: |
c9145754 | 3764 | continue; |
726a989a RB |
3765 | |
3766 | case GIMPLE_CALL: | |
c9145754 | 3767 | { |
726a989a | 3768 | vn_reference_t ref; |
726a989a | 3769 | pre_expr result = NULL; |
ef062b13 | 3770 | auto_vec<vn_reference_op_s> ops; |
c9145754 | 3771 | |
9cbbba28 EB |
3772 | /* We can value number only calls to real functions. */ |
3773 | if (gimple_call_internal_p (stmt)) | |
726a989a | 3774 | continue; |
c9145754 | 3775 | |
726a989a | 3776 | copy_reference_ops_from_call (stmt, &ops); |
b45d2719 RG |
3777 | vn_reference_lookup_pieces (gimple_vuse (stmt), 0, |
3778 | gimple_expr_type (stmt), | |
3bc27de7 | 3779 | ops, &ref, VN_NOWALK); |
726a989a RB |
3780 | if (!ref) |
3781 | continue; | |
c9145754 | 3782 | |
cd32bb90 RG |
3783 | /* If the value of the call is not invalidated in |
3784 | this block until it is computed, add the expression | |
3785 | to EXP_GEN. */ | |
3786 | if (!gimple_vuse (stmt) | |
3787 | || gimple_code | |
3788 | (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI | |
3789 | || gimple_bb (SSA_NAME_DEF_STMT | |
3790 | (gimple_vuse (stmt))) != block) | |
3791 | { | |
3792 | result = (pre_expr) pool_alloc (pre_expr_pool); | |
3793 | result->kind = REFERENCE; | |
3794 | result->id = 0; | |
3795 | PRE_EXPR_REFERENCE (result) = ref; | |
3796 | ||
3797 | get_or_alloc_expression_id (result); | |
3798 | add_to_value (get_expr_value_id (result), result); | |
3c1d57d0 | 3799 | bitmap_value_insert_into_set (EXP_GEN (block), result); |
cd32bb90 | 3800 | } |
726a989a RB |
3801 | continue; |
3802 | } | |
c9145754 | 3803 | |
726a989a RB |
3804 | case GIMPLE_ASSIGN: |
3805 | { | |
3806 | pre_expr result = NULL; | |
17742d62 | 3807 | switch (vn_get_stmt_kind (stmt)) |
726a989a | 3808 | { |
17742d62 | 3809 | case VN_NARY: |
726a989a | 3810 | { |
61514fe4 | 3811 | enum tree_code code = gimple_assign_rhs_code (stmt); |
726a989a | 3812 | vn_nary_op_t nary; |
61514fe4 RG |
3813 | |
3814 | /* COND_EXPR and VEC_COND_EXPR are awkward in | |
3815 | that they contain an embedded complex expression. | |
3816 | Don't even try to shove those through PRE. */ | |
3817 | if (code == COND_EXPR | |
3818 | || code == VEC_COND_EXPR) | |
3819 | continue; | |
3820 | ||
91af9dc9 | 3821 | vn_nary_op_lookup_stmt (stmt, &nary); |
726a989a RB |
3822 | if (!nary) |
3823 | continue; | |
3824 | ||
073a8998 | 3825 | /* If the NARY traps and there was a preceding |
bea966c2 RG |
3826 | point in the block that might not return avoid |
3827 | adding the nary to EXP_GEN. */ | |
3828 | if (BB_MAY_NOTRETURN (block) | |
3829 | && vn_nary_may_trap (nary)) | |
3830 | continue; | |
3831 | ||
726a989a RB |
3832 | result = (pre_expr) pool_alloc (pre_expr_pool); |
3833 | result->kind = NARY; | |
3834 | result->id = 0; | |
3835 | PRE_EXPR_NARY (result) = nary; | |
3836 | break; | |
3837 | } | |
c9145754 | 3838 | |
17742d62 | 3839 | case VN_REFERENCE: |
726a989a RB |
3840 | { |
3841 | vn_reference_t ref; | |
726a989a | 3842 | vn_reference_lookup (gimple_assign_rhs1 (stmt), |
5006671f | 3843 | gimple_vuse (stmt), |
3bc27de7 | 3844 | VN_WALK, &ref); |
726a989a RB |
3845 | if (!ref) |
3846 | continue; | |
3847 | ||
cd32bb90 RG |
3848 | /* If the value of the reference is not invalidated in |
3849 | this block until it is computed, add the expression | |
3850 | to EXP_GEN. */ | |
3851 | if (gimple_vuse (stmt)) | |
3852 | { | |
3853 | gimple def_stmt; | |
3854 | bool ok = true; | |
3855 | def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt)); | |
3856 | while (!gimple_nop_p (def_stmt) | |
3857 | && gimple_code (def_stmt) != GIMPLE_PHI | |
3858 | && gimple_bb (def_stmt) == block) | |
3859 | { | |
3860 | if (stmt_may_clobber_ref_p | |
3861 | (def_stmt, gimple_assign_rhs1 (stmt))) | |
3862 | { | |
3863 | ok = false; | |
3864 | break; | |
3865 | } | |
3866 | def_stmt | |
3867 | = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt)); | |
3868 | } | |
3869 | if (!ok) | |
3870 | continue; | |
3871 | } | |
3872 | ||
726a989a RB |
3873 | result = (pre_expr) pool_alloc (pre_expr_pool); |
3874 | result->kind = REFERENCE; | |
3875 | result->id = 0; | |
3876 | PRE_EXPR_REFERENCE (result) = ref; | |
3877 | break; | |
3878 | } | |
c9145754 | 3879 | |
726a989a | 3880 | default: |
726a989a | 3881 | continue; |
c9145754 | 3882 | } |
726a989a RB |
3883 | |
3884 | get_or_alloc_expression_id (result); | |
3885 | add_to_value (get_expr_value_id (result), result); | |
3c1d57d0 | 3886 | bitmap_value_insert_into_set (EXP_GEN (block), result); |
dda243de | 3887 | continue; |
c9145754 DB |
3888 | } |
3889 | default: | |
3890 | break; | |
7e6eb623 | 3891 | } |
6de9cd9a | 3892 | } |
726a989a | 3893 | |
e3815735 RB |
3894 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3895 | { | |
3896 | print_bitmap_set (dump_file, EXP_GEN (block), | |
3897 | "exp_gen", block->index); | |
3898 | print_bitmap_set (dump_file, PHI_GEN (block), | |
3899 | "phi_gen", block->index); | |
3900 | print_bitmap_set (dump_file, TMP_GEN (block), | |
3901 | "tmp_gen", block->index); | |
3902 | print_bitmap_set (dump_file, AVAIL_OUT (block), | |
3903 | "avail_out", block->index); | |
3904 | } | |
3905 | ||
665fcad8 SB |
3906 | /* Put the dominator children of BLOCK on the worklist of blocks |
3907 | to compute available sets for. */ | |
3908 | for (son = first_dom_son (CDI_DOMINATORS, block); | |
3909 | son; | |
3910 | son = next_dom_son (CDI_DOMINATORS, son)) | |
3911 | worklist[sp++] = son; | |
6de9cd9a | 3912 | } |
33c94679 | 3913 | |
665fcad8 | 3914 | free (worklist); |
7e6eb623 DB |
3915 | } |
3916 | ||
40b178f4 RG |
3917 | |
3918 | /* Local state for the eliminate domwalk. */ | |
9771b263 | 3919 | static vec<gimple> el_to_remove; |
40b178f4 | 3920 | static unsigned int el_todo; |
9771b263 DN |
3921 | static vec<tree> el_avail; |
3922 | static vec<tree> el_avail_stack; | |
40b178f4 RG |
3923 | |
3924 | /* Return a leader for OP that is available at the current point of the | |
3925 | eliminate domwalk. */ | |
3d45dd59 RG |
3926 | |
3927 | static tree | |
40b178f4 | 3928 | eliminate_avail (tree op) |
3d45dd59 | 3929 | { |
40b178f4 RG |
3930 | tree valnum = VN_INFO (op)->valnum; |
3931 | if (TREE_CODE (valnum) == SSA_NAME) | |
3932 | { | |
3933 | if (SSA_NAME_IS_DEFAULT_DEF (valnum)) | |
3934 | return valnum; | |
9771b263 DN |
3935 | if (el_avail.length () > SSA_NAME_VERSION (valnum)) |
3936 | return el_avail[SSA_NAME_VERSION (valnum)]; | |
40b178f4 RG |
3937 | } |
3938 | else if (is_gimple_min_invariant (valnum)) | |
3939 | return valnum; | |
3940 | return NULL_TREE; | |
3941 | } | |
3942 | ||
3943 | /* At the current point of the eliminate domwalk make OP available. */ | |
3d45dd59 | 3944 | |
40b178f4 RG |
3945 | static void |
3946 | eliminate_push_avail (tree op) | |
3947 | { | |
3948 | tree valnum = VN_INFO (op)->valnum; | |
3949 | if (TREE_CODE (valnum) == SSA_NAME) | |
3950 | { | |
9771b263 DN |
3951 | if (el_avail.length () <= SSA_NAME_VERSION (valnum)) |
3952 | el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1); | |
3953 | el_avail[SSA_NAME_VERSION (valnum)] = op; | |
3954 | el_avail_stack.safe_push (op); | |
40b178f4 RG |
3955 | } |
3956 | } | |
3957 | ||
3958 | /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns | |
3959 | the leader for the expression if insertion was successful. */ | |
3960 | ||
3961 | static tree | |
3962 | eliminate_insert (gimple_stmt_iterator *gsi, tree val) | |
3963 | { | |
3964 | tree expr = vn_get_expr_for (val); | |
3965 | if (!CONVERT_EXPR_P (expr) | |
3966 | && TREE_CODE (expr) != VIEW_CONVERT_EXPR) | |
3d45dd59 | 3967 | return NULL_TREE; |
3d45dd59 | 3968 | |
40b178f4 RG |
3969 | tree op = TREE_OPERAND (expr, 0); |
3970 | tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op; | |
3971 | if (!leader) | |
3d45dd59 | 3972 | return NULL_TREE; |
3d45dd59 | 3973 | |
40b178f4 RG |
3974 | tree res = make_temp_ssa_name (TREE_TYPE (val), NULL, "pretmp"); |
3975 | gimple tem = gimple_build_assign (res, | |
c96cab6e RB |
3976 | fold_build1 (TREE_CODE (expr), |
3977 | TREE_TYPE (expr), leader)); | |
40b178f4 RG |
3978 | gsi_insert_before (gsi, tem, GSI_SAME_STMT); |
3979 | VN_INFO_GET (res)->valnum = val; | |
3980 | ||
3981 | if (TREE_CODE (leader) == SSA_NAME) | |
3982 | gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true); | |
3983 | ||
3984 | pre_stats.insertions++; | |
3985 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3986 | { | |
3987 | fprintf (dump_file, "Inserted "); | |
3988 | print_gimple_stmt (dump_file, tem, 0, 0); | |
3989 | } | |
3990 | ||
3991 | return res; | |
3d45dd59 | 3992 | } |
33c94679 | 3993 | |
4d9192b5 TS |
3994 | class eliminate_dom_walker : public dom_walker |
3995 | { | |
3996 | public: | |
b82ef848 RB |
3997 | eliminate_dom_walker (cdi_direction direction, bool do_pre_) |
3998 | : dom_walker (direction), do_pre (do_pre_) {} | |
4d9192b5 TS |
3999 | |
4000 | virtual void before_dom_children (basic_block); | |
4001 | virtual void after_dom_children (basic_block); | |
b82ef848 RB |
4002 | |
4003 | bool do_pre; | |
4d9192b5 TS |
4004 | }; |
4005 | ||
40b178f4 | 4006 | /* Perform elimination for the basic-block B during the domwalk. */ |
7e6eb623 | 4007 | |
4d9192b5 TS |
4008 | void |
4009 | eliminate_dom_walker::before_dom_children (basic_block b) | |
7e6eb623 | 4010 | { |
5006671f RG |
4011 | gimple_stmt_iterator gsi; |
4012 | gimple stmt; | |
7e6eb623 | 4013 | |
40b178f4 | 4014 | /* Mark new bb. */ |
9771b263 | 4015 | el_avail_stack.safe_push (NULL_TREE); |
40b178f4 | 4016 | |
6aa4c5b6 RB |
4017 | /* ??? If we do nothing for unreachable blocks then this will confuse |
4018 | tailmerging. Eventually we can reduce its reliance on SCCVN now | |
4019 | that we fully copy/constant-propagate (most) things. */ | |
1d44def2 | 4020 | |
40b178f4 | 4021 | for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);) |
7e6eb623 | 4022 | { |
6aa4c5b6 RB |
4023 | gimple phi = gsi_stmt (gsi); |
4024 | tree res = PHI_RESULT (phi); | |
c9145754 | 4025 | |
6aa4c5b6 | 4026 | if (virtual_operand_p (res)) |
40b178f4 | 4027 | { |
40b178f4 RG |
4028 | gsi_next (&gsi); |
4029 | continue; | |
4030 | } | |
c9145754 | 4031 | |
6aa4c5b6 RB |
4032 | tree sprime = eliminate_avail (res); |
4033 | if (sprime | |
4034 | && sprime != res) | |
40b178f4 | 4035 | { |
6aa4c5b6 RB |
4036 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4037 | { | |
4038 | fprintf (dump_file, "Replaced redundant PHI node defining "); | |
4039 | print_generic_expr (dump_file, res, 0); | |
4040 | fprintf (dump_file, " with "); | |
4041 | print_generic_expr (dump_file, sprime, 0); | |
4042 | fprintf (dump_file, "\n"); | |
4043 | } | |
40b178f4 | 4044 | |
6aa4c5b6 RB |
4045 | /* If we inserted this PHI node ourself, it's not an elimination. */ |
4046 | if (inserted_exprs | |
4047 | && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))) | |
4048 | pre_stats.phis--; | |
4049 | else | |
4050 | pre_stats.eliminations++; | |
ff2ad0f7 | 4051 | |
6aa4c5b6 RB |
4052 | /* If we will propagate into all uses don't bother to do |
4053 | anything. */ | |
4054 | if (may_propagate_copy (res, sprime)) | |
4055 | { | |
4056 | /* Mark the PHI for removal. */ | |
4057 | el_to_remove.safe_push (phi); | |
4058 | gsi_next (&gsi); | |
4059 | continue; | |
4060 | } | |
40b178f4 | 4061 | |
6aa4c5b6 | 4062 | remove_phi_node (&gsi, false); |
40b178f4 | 4063 | |
6aa4c5b6 RB |
4064 | if (inserted_exprs |
4065 | && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)) | |
4066 | && TREE_CODE (sprime) == SSA_NAME) | |
4067 | gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true); | |
40b178f4 | 4068 | |
6aa4c5b6 RB |
4069 | if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime))) |
4070 | sprime = fold_convert (TREE_TYPE (res), sprime); | |
4071 | gimple stmt = gimple_build_assign (res, sprime); | |
4072 | /* ??? It cannot yet be necessary (DOM walk). */ | |
4073 | gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY)); | |
40b178f4 | 4074 | |
6aa4c5b6 RB |
4075 | gimple_stmt_iterator gsi2 = gsi_after_labels (b); |
4076 | gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT); | |
4077 | continue; | |
4078 | } | |
3d45dd59 | 4079 | |
6aa4c5b6 RB |
4080 | eliminate_push_avail (res); |
4081 | gsi_next (&gsi); | |
4082 | } | |
4bcc5786 | 4083 | |
6aa4c5b6 RB |
4084 | for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi)) |
4085 | { | |
4086 | tree sprime = NULL_TREE; | |
4087 | stmt = gsi_stmt (gsi); | |
4088 | tree lhs = gimple_get_lhs (stmt); | |
4089 | if (lhs && TREE_CODE (lhs) == SSA_NAME | |
4090 | && !gimple_has_volatile_ops (stmt) | |
4bcc5786 RB |
4091 | /* See PR43491. Do not replace a global register variable when |
4092 | it is a the RHS of an assignment. Do replace local register | |
4093 | variables since gcc does not guarantee a local variable will | |
4094 | be allocated in register. | |
6aa4c5b6 RB |
4095 | ??? The fix isn't effective here. This should instead |
4096 | be ensured by not value-numbering them the same but treating | |
4097 | them like volatiles? */ | |
4098 | && !(gimple_assign_single_p (stmt) | |
4099 | && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL | |
4100 | && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt)) | |
4101 | && is_global_var (gimple_assign_rhs1 (stmt))))) | |
4102 | { | |
4103 | sprime = eliminate_avail (lhs); | |
40b178f4 RG |
4104 | if (!sprime) |
4105 | { | |
3d45dd59 RG |
4106 | /* If there is no existing usable leader but SCCVN thinks |
4107 | it has an expression it wants to use as replacement, | |
4108 | insert that. */ | |
40b178f4 RG |
4109 | tree val = VN_INFO (lhs)->valnum; |
4110 | if (val != VN_TOP | |
4111 | && TREE_CODE (val) == SSA_NAME | |
4112 | && VN_INFO (val)->needs_insertion | |
a044f0e7 | 4113 | && VN_INFO (val)->expr != NULL_TREE |
40b178f4 RG |
4114 | && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE) |
4115 | eliminate_push_avail (sprime); | |
4116 | } | |
40b178f4 | 4117 | |
6aa4c5b6 RB |
4118 | /* If this now constitutes a copy duplicate points-to |
4119 | and range info appropriately. This is especially | |
4120 | important for inserted code. See tree-ssa-copy.c | |
4121 | for similar code. */ | |
4122 | if (sprime | |
4123 | && TREE_CODE (sprime) == SSA_NAME) | |
4124 | { | |
4125 | basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime)); | |
4126 | if (POINTER_TYPE_P (TREE_TYPE (lhs)) | |
4127 | && SSA_NAME_PTR_INFO (lhs) | |
4128 | && !SSA_NAME_PTR_INFO (sprime)) | |
ff2ad0f7 | 4129 | { |
6aa4c5b6 RB |
4130 | duplicate_ssa_name_ptr_info (sprime, |
4131 | SSA_NAME_PTR_INFO (lhs)); | |
4132 | if (b != sprime_b) | |
4133 | mark_ptr_info_alignment_unknown | |
4134 | (SSA_NAME_PTR_INFO (sprime)); | |
40b178f4 | 4135 | } |
6aa4c5b6 RB |
4136 | else if (!POINTER_TYPE_P (TREE_TYPE (lhs)) |
4137 | && SSA_NAME_RANGE_INFO (lhs) | |
4138 | && !SSA_NAME_RANGE_INFO (sprime) | |
4139 | && b == sprime_b) | |
4140 | duplicate_ssa_name_range_info (sprime, | |
4141 | SSA_NAME_RANGE_TYPE (lhs), | |
4142 | SSA_NAME_RANGE_INFO (lhs)); | |
40b178f4 | 4143 | } |
30fd5881 | 4144 | |
6aa4c5b6 RB |
4145 | /* Inhibit the use of an inserted PHI on a loop header when |
4146 | the address of the memory reference is a simple induction | |
4147 | variable. In other cases the vectorizer won't do anything | |
4148 | anyway (either it's loop invariant or a complicated | |
4149 | expression). */ | |
40b178f4 | 4150 | if (sprime |
6aa4c5b6 RB |
4151 | && TREE_CODE (sprime) == SSA_NAME |
4152 | && do_pre | |
4153 | && flag_tree_loop_vectorize | |
4154 | && loop_outer (b->loop_father) | |
4155 | && has_zero_uses (sprime) | |
4156 | && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime)) | |
4157 | && gimple_assign_load_p (stmt)) | |
40b178f4 | 4158 | { |
6aa4c5b6 RB |
4159 | gimple def_stmt = SSA_NAME_DEF_STMT (sprime); |
4160 | basic_block def_bb = gimple_bb (def_stmt); | |
4161 | if (gimple_code (def_stmt) == GIMPLE_PHI | |
4162 | && b->loop_father->header == def_bb) | |
cddaefa3 | 4163 | { |
6aa4c5b6 RB |
4164 | ssa_op_iter iter; |
4165 | tree op; | |
4166 | bool found = false; | |
4167 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) | |
cddaefa3 | 4168 | { |
6aa4c5b6 RB |
4169 | affine_iv iv; |
4170 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (op)); | |
4171 | if (def_bb | |
4172 | && flow_bb_inside_loop_p (b->loop_father, def_bb) | |
4173 | && simple_iv (b->loop_father, | |
4174 | b->loop_father, op, &iv, true)) | |
cddaefa3 | 4175 | { |
6aa4c5b6 RB |
4176 | found = true; |
4177 | break; | |
cddaefa3 | 4178 | } |
6aa4c5b6 RB |
4179 | } |
4180 | if (found) | |
4181 | { | |
4182 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
cddaefa3 | 4183 | { |
6aa4c5b6 RB |
4184 | fprintf (dump_file, "Not replacing "); |
4185 | print_gimple_expr (dump_file, stmt, 0, 0); | |
4186 | fprintf (dump_file, " with "); | |
4187 | print_generic_expr (dump_file, sprime, 0); | |
4188 | fprintf (dump_file, " which would add a loop" | |
4189 | " carried dependence to loop %d\n", | |
4190 | b->loop_father->num); | |
cddaefa3 | 4191 | } |
6aa4c5b6 | 4192 | /* Don't keep sprime available. */ |
6aa4c5b6 | 4193 | sprime = NULL_TREE; |
cddaefa3 RB |
4194 | } |
4195 | } | |
6aa4c5b6 RB |
4196 | } |
4197 | ||
4198 | if (sprime) | |
4199 | { | |
4200 | /* If we can propagate the value computed for LHS into | |
4201 | all uses don't bother doing anything with this stmt. */ | |
4202 | if (may_propagate_copy (lhs, sprime)) | |
4203 | { | |
4204 | /* Mark it for removal. */ | |
4205 | el_to_remove.safe_push (stmt); | |
4206 | ||
4207 | /* ??? Don't count copy/constant propagations. */ | |
4208 | if (gimple_assign_single_p (stmt) | |
4209 | && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME | |
4210 | || gimple_assign_rhs1 (stmt) == sprime)) | |
4211 | continue; | |
4212 | ||
4213 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4214 | { | |
4215 | fprintf (dump_file, "Replaced "); | |
4216 | print_gimple_expr (dump_file, stmt, 0, 0); | |
4217 | fprintf (dump_file, " with "); | |
4218 | print_generic_expr (dump_file, sprime, 0); | |
4219 | fprintf (dump_file, " in all uses of "); | |
4220 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
4221 | } | |
4222 | ||
4223 | pre_stats.eliminations++; | |
4224 | continue; | |
4225 | } | |
4226 | ||
4227 | /* If this is an assignment from our leader (which | |
4228 | happens in the case the value-number is a constant) | |
4229 | then there is nothing to do. */ | |
4230 | if (gimple_assign_single_p (stmt) | |
4231 | && sprime == gimple_assign_rhs1 (stmt)) | |
4232 | continue; | |
4233 | ||
4234 | /* Else replace its RHS. */ | |
4235 | bool can_make_abnormal_goto | |
4236 | = is_gimple_call (stmt) | |
4237 | && stmt_can_make_abnormal_goto (stmt); | |
cddaefa3 | 4238 | |
40b178f4 RG |
4239 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4240 | { | |
4241 | fprintf (dump_file, "Replaced "); | |
4242 | print_gimple_expr (dump_file, stmt, 0, 0); | |
4243 | fprintf (dump_file, " with "); | |
4244 | print_generic_expr (dump_file, sprime, 0); | |
4245 | fprintf (dump_file, " in "); | |
4246 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
ff2ad0f7 | 4247 | } |
40b178f4 RG |
4248 | |
4249 | if (TREE_CODE (sprime) == SSA_NAME) | |
4250 | gimple_set_plf (SSA_NAME_DEF_STMT (sprime), | |
4251 | NECESSARY, true); | |
40b178f4 RG |
4252 | |
4253 | pre_stats.eliminations++; | |
6aa4c5b6 RB |
4254 | gimple orig_stmt = stmt; |
4255 | if (!useless_type_conversion_p (TREE_TYPE (lhs), | |
4256 | TREE_TYPE (sprime))) | |
4257 | sprime = fold_convert (TREE_TYPE (lhs), sprime); | |
d1c0308e RB |
4258 | tree vdef = gimple_vdef (stmt); |
4259 | tree vuse = gimple_vuse (stmt); | |
40b178f4 RG |
4260 | propagate_tree_value_into_stmt (&gsi, sprime); |
4261 | stmt = gsi_stmt (gsi); | |
4262 | update_stmt (stmt); | |
d1c0308e RB |
4263 | if (vdef != gimple_vdef (stmt)) |
4264 | VN_INFO (vdef)->valnum = vuse; | |
40b178f4 RG |
4265 | |
4266 | /* If we removed EH side-effects from the statement, clean | |
4267 | its EH information. */ | |
4268 | if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt)) | |
6cdb0ee3 | 4269 | { |
40b178f4 RG |
4270 | bitmap_set_bit (need_eh_cleanup, |
4271 | gimple_bb (stmt)->index); | |
6cdb0ee3 | 4272 | if (dump_file && (dump_flags & TDF_DETAILS)) |
40b178f4 RG |
4273 | fprintf (dump_file, " Removed EH side-effects.\n"); |
4274 | } | |
6cdb0ee3 | 4275 | |
40b178f4 RG |
4276 | /* Likewise for AB side-effects. */ |
4277 | if (can_make_abnormal_goto | |
4278 | && !stmt_can_make_abnormal_goto (stmt)) | |
4279 | { | |
4280 | bitmap_set_bit (need_ab_cleanup, | |
4281 | gimple_bb (stmt)->index); | |
4282 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4283 | fprintf (dump_file, " Removed AB side-effects.\n"); | |
6cdb0ee3 | 4284 | } |
6aa4c5b6 RB |
4285 | |
4286 | continue; | |
6cdb0ee3 | 4287 | } |
40b178f4 | 4288 | } |
6aa4c5b6 | 4289 | |
40b178f4 | 4290 | /* If the statement is a scalar store, see if the expression |
6aa4c5b6 RB |
4291 | has the same value number as its rhs. If so, the store is |
4292 | dead. */ | |
4293 | if (gimple_assign_single_p (stmt) | |
4294 | && !gimple_has_volatile_ops (stmt) | |
4295 | && !is_gimple_reg (gimple_assign_lhs (stmt)) | |
4296 | && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME | |
4297 | || is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))) | |
4298 | { | |
4299 | tree val; | |
4300 | tree rhs = gimple_assign_rhs1 (stmt); | |
4301 | val = vn_reference_lookup (gimple_assign_lhs (stmt), | |
4302 | gimple_vuse (stmt), VN_WALK, NULL); | |
4303 | if (TREE_CODE (rhs) == SSA_NAME) | |
4304 | rhs = VN_INFO (rhs)->valnum; | |
4305 | if (val | |
4306 | && operand_equal_p (val, rhs, 0)) | |
4307 | { | |
4308 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4309 | { | |
4310 | fprintf (dump_file, "Deleted redundant store "); | |
4311 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
4312 | } | |
4313 | ||
4314 | /* Queue stmt for removal. */ | |
4315 | el_to_remove.safe_push (stmt); | |
4316 | continue; | |
4317 | } | |
4318 | } | |
40b178f4 | 4319 | |
6aa4c5b6 RB |
4320 | bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt); |
4321 | bool was_noreturn = (is_gimple_call (stmt) | |
4322 | && gimple_call_noreturn_p (stmt)); | |
4323 | tree vdef = gimple_vdef (stmt); | |
4324 | tree vuse = gimple_vuse (stmt); | |
4325 | ||
4326 | /* If we didn't replace the whole stmt (or propagate the result | |
4327 | into all uses), replace all uses on this stmt with their | |
4328 | leaders. */ | |
4329 | use_operand_p use_p; | |
4330 | ssa_op_iter iter; | |
4331 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) | |
40b178f4 | 4332 | { |
6aa4c5b6 RB |
4333 | tree use = USE_FROM_PTR (use_p); |
4334 | /* ??? The call code above leaves stmt operands un-updated. */ | |
4335 | if (TREE_CODE (use) != SSA_NAME) | |
4336 | continue; | |
4337 | tree sprime = eliminate_avail (use); | |
4338 | if (sprime && sprime != use | |
4339 | && may_propagate_copy (use, sprime) | |
4340 | /* We substitute into debug stmts to avoid excessive | |
4341 | debug temporaries created by removed stmts, but we need | |
4342 | to avoid doing so for inserted sprimes as we never want | |
4343 | to create debug temporaries for them. */ | |
4344 | && (!inserted_exprs | |
4345 | || TREE_CODE (sprime) != SSA_NAME | |
4346 | || !is_gimple_debug (stmt) | |
4347 | || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime)))) | |
aa7069aa | 4348 | { |
6aa4c5b6 RB |
4349 | propagate_value (use_p, sprime); |
4350 | gimple_set_modified (stmt, true); | |
4351 | if (TREE_CODE (sprime) == SSA_NAME | |
4352 | && !is_gimple_debug (stmt)) | |
4353 | gimple_set_plf (SSA_NAME_DEF_STMT (sprime), | |
4354 | NECESSARY, true); | |
40b178f4 RG |
4355 | } |
4356 | } | |
6aa4c5b6 | 4357 | |
40b178f4 | 4358 | /* Visit indirect calls and turn them into direct calls if |
6aa4c5b6 | 4359 | possible using the devirtualization machinery. */ |
40b178f4 RG |
4360 | if (is_gimple_call (stmt)) |
4361 | { | |
6aa4c5b6 RB |
4362 | tree fn = gimple_call_fn (stmt); |
4363 | if (fn | |
7d0aa05b JH |
4364 | && flag_devirtualize |
4365 | && virtual_method_call_p (fn)) | |
e248d83f | 4366 | { |
7d0aa05b JH |
4367 | tree otr_type; |
4368 | HOST_WIDE_INT otr_token; | |
4369 | ipa_polymorphic_call_context context; | |
4370 | tree instance; | |
4371 | bool final; | |
4372 | ||
4373 | instance = get_polymorphic_call_info (current_function_decl, | |
4374 | fn, | |
4375 | &otr_type, &otr_token, &context, stmt); | |
4376 | ||
4d7cf10d | 4377 | context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn), otr_type, stmt); |
7d0aa05b JH |
4378 | |
4379 | vec <cgraph_node *>targets | |
4380 | = possible_polymorphic_call_targets (obj_type_ref_class (fn), | |
4381 | tree_to_uhwi | |
4382 | (OBJ_TYPE_REF_TOKEN (fn)), | |
4383 | context, | |
4384 | &final); | |
4385 | if (dump_enabled_p ()) | |
4386 | dump_possible_polymorphic_call_targets (dump_file, | |
4387 | obj_type_ref_class (fn), | |
4388 | tree_to_uhwi | |
4389 | (OBJ_TYPE_REF_TOKEN (fn)), | |
4390 | context); | |
4391 | if (final && targets.length () <= 1 && dbg_cnt (devirt)) | |
e248d83f | 4392 | { |
7d0aa05b JH |
4393 | tree fn; |
4394 | if (targets.length () == 1) | |
4395 | fn = targets[0]->decl; | |
4396 | else | |
4397 | fn = builtin_decl_implicit (BUILT_IN_UNREACHABLE); | |
6aa4c5b6 RB |
4398 | if (dump_enabled_p ()) |
4399 | { | |
807b7d62 | 4400 | location_t loc = gimple_location_safe (stmt); |
6aa4c5b6 RB |
4401 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc, |
4402 | "converting indirect call to " | |
4403 | "function %s\n", | |
d52f5295 | 4404 | cgraph_node::get (fn)->name ()); |
6aa4c5b6 RB |
4405 | } |
4406 | gimple_call_set_fndecl (stmt, fn); | |
4407 | gimple_set_modified (stmt, true); | |
e248d83f | 4408 | } |
7d0aa05b JH |
4409 | else |
4410 | gcc_assert (!ipa_intraprocedural_devirtualization (stmt)); | |
e248d83f | 4411 | } |
6aa4c5b6 | 4412 | } |
87c20fe7 | 4413 | |
6aa4c5b6 RB |
4414 | if (gimple_modified_p (stmt)) |
4415 | { | |
4416 | /* If a formerly non-invariant ADDR_EXPR is turned into an | |
4417 | invariant one it was on a separate stmt. */ | |
4418 | if (gimple_assign_single_p (stmt) | |
4419 | && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR) | |
4420 | recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (stmt)); | |
4421 | gimple old_stmt = stmt; | |
4422 | if (is_gimple_call (stmt)) | |
4423 | { | |
4424 | /* ??? Only fold calls inplace for now, this may create new | |
4425 | SSA names which in turn will confuse free_scc_vn SSA name | |
4426 | release code. */ | |
4427 | fold_stmt_inplace (&gsi); | |
40b178f4 RG |
4428 | /* When changing a call into a noreturn call, cfg cleanup |
4429 | is needed to fix up the noreturn call. */ | |
4430 | if (!was_noreturn && gimple_call_noreturn_p (stmt)) | |
4431 | el_todo |= TODO_cleanup_cfg; | |
6aa4c5b6 RB |
4432 | } |
4433 | else | |
4434 | { | |
4435 | fold_stmt (&gsi); | |
4436 | stmt = gsi_stmt (gsi); | |
4437 | if ((gimple_code (stmt) == GIMPLE_COND | |
4438 | && (gimple_cond_true_p (stmt) | |
4439 | || gimple_cond_false_p (stmt))) | |
4440 | || (gimple_code (stmt) == GIMPLE_SWITCH | |
4441 | && TREE_CODE (gimple_switch_index (stmt)) == INTEGER_CST)) | |
4442 | el_todo |= TODO_cleanup_cfg; | |
4443 | } | |
4444 | /* If we removed EH side-effects from the statement, clean | |
4445 | its EH information. */ | |
4446 | if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt)) | |
4447 | { | |
4448 | bitmap_set_bit (need_eh_cleanup, | |
4449 | gimple_bb (stmt)->index); | |
4450 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4451 | fprintf (dump_file, " Removed EH side-effects.\n"); | |
4452 | } | |
4453 | /* Likewise for AB side-effects. */ | |
4454 | if (can_make_abnormal_goto | |
4455 | && !stmt_can_make_abnormal_goto (stmt)) | |
4456 | { | |
4457 | bitmap_set_bit (need_ab_cleanup, | |
4458 | gimple_bb (stmt)->index); | |
4459 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4460 | fprintf (dump_file, " Removed AB side-effects.\n"); | |
4461 | } | |
4462 | update_stmt (stmt); | |
4463 | if (vdef != gimple_vdef (stmt)) | |
4464 | VN_INFO (vdef)->valnum = vuse; | |
4465 | } | |
30fd5881 | 4466 | |
5d5cb4d4 RB |
4467 | /* Make new values available - for fully redundant LHS we |
4468 | continue with the next stmt above and skip this. */ | |
4469 | def_operand_p defp; | |
4470 | FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_DEF) | |
4471 | eliminate_push_avail (DEF_FROM_PTR (defp)); | |
6aa4c5b6 | 4472 | } |
40b178f4 | 4473 | |
6aa4c5b6 RB |
4474 | /* Replace destination PHI arguments. */ |
4475 | edge_iterator ei; | |
4476 | edge e; | |
4477 | FOR_EACH_EDGE (e, ei, b->succs) | |
4478 | { | |
4479 | for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
4480 | { | |
4481 | gimple phi = gsi_stmt (gsi); | |
4482 | use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e); | |
4483 | tree arg = USE_FROM_PTR (use_p); | |
4484 | if (TREE_CODE (arg) != SSA_NAME | |
4485 | || virtual_operand_p (arg)) | |
4486 | continue; | |
4487 | tree sprime = eliminate_avail (arg); | |
4488 | if (sprime && may_propagate_copy (arg, sprime)) | |
4489 | { | |
4490 | propagate_value (use_p, sprime); | |
4491 | if (TREE_CODE (sprime) == SSA_NAME) | |
4492 | gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true); | |
aa7069aa | 4493 | } |
83737db2 | 4494 | } |
40b178f4 RG |
4495 | } |
4496 | } | |
439ef907 | 4497 | |
40b178f4 | 4498 | /* Make no longer available leaders no longer available. */ |
439ef907 | 4499 | |
4d9192b5 TS |
4500 | void |
4501 | eliminate_dom_walker::after_dom_children (basic_block) | |
40b178f4 RG |
4502 | { |
4503 | tree entry; | |
9771b263 DN |
4504 | while ((entry = el_avail_stack.pop ()) != NULL_TREE) |
4505 | el_avail[SSA_NAME_VERSION (VN_INFO (entry)->valnum)] = NULL_TREE; | |
40b178f4 | 4506 | } |
439ef907 | 4507 | |
40b178f4 | 4508 | /* Eliminate fully redundant computations. */ |
439ef907 | 4509 | |
40b178f4 | 4510 | static unsigned int |
b82ef848 | 4511 | eliminate (bool do_pre) |
40b178f4 | 4512 | { |
40b178f4 RG |
4513 | gimple_stmt_iterator gsi; |
4514 | gimple stmt; | |
439ef907 | 4515 | |
40b178f4 RG |
4516 | need_eh_cleanup = BITMAP_ALLOC (NULL); |
4517 | need_ab_cleanup = BITMAP_ALLOC (NULL); | |
4d2ab9e3 | 4518 | |
9771b263 | 4519 | el_to_remove.create (0); |
40b178f4 | 4520 | el_todo = 0; |
9771b263 DN |
4521 | el_avail.create (0); |
4522 | el_avail_stack.create (0); | |
40b178f4 | 4523 | |
b82ef848 RB |
4524 | eliminate_dom_walker (CDI_DOMINATORS, |
4525 | do_pre).walk (cfun->cfg->x_entry_block_ptr); | |
40b178f4 | 4526 | |
9771b263 DN |
4527 | el_avail.release (); |
4528 | el_avail_stack.release (); | |
b80280f2 | 4529 | |
5006671f | 4530 | /* We cannot remove stmts during BB walk, especially not release SSA |
439ef907 | 4531 | names there as this confuses the VN machinery. The stmts ending |
6aa4c5b6 RB |
4532 | up in el_to_remove are either stores or simple copies. |
4533 | Remove stmts in reverse order to make debug stmt creation possible. */ | |
4534 | while (!el_to_remove.is_empty ()) | |
5006671f | 4535 | { |
6aa4c5b6 RB |
4536 | stmt = el_to_remove.pop (); |
4537 | ||
4538 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
439ef907 | 4539 | { |
6aa4c5b6 RB |
4540 | fprintf (dump_file, "Removing dead stmt "); |
4541 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
439ef907 RG |
4542 | } |
4543 | ||
6aa4c5b6 RB |
4544 | tree lhs; |
4545 | if (gimple_code (stmt) == GIMPLE_PHI) | |
4546 | lhs = gimple_phi_result (stmt); | |
4547 | else | |
4548 | lhs = gimple_get_lhs (stmt); | |
4549 | ||
4550 | if (inserted_exprs | |
4551 | && TREE_CODE (lhs) == SSA_NAME) | |
4552 | bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs)); | |
4553 | ||
4554 | gsi = gsi_for_stmt (stmt); | |
4555 | if (gimple_code (stmt) == GIMPLE_PHI) | |
25b7069a | 4556 | remove_phi_node (&gsi, true); |
6aa4c5b6 | 4557 | else |
439ef907 | 4558 | { |
a2c0ed2e | 4559 | basic_block bb = gimple_bb (stmt); |
439ef907 | 4560 | unlink_stmt_vdef (stmt); |
b5b3ec3e RG |
4561 | if (gsi_remove (&gsi, true)) |
4562 | bitmap_set_bit (need_eh_cleanup, bb->index); | |
439ef907 RG |
4563 | release_defs (stmt); |
4564 | } | |
25b7069a RB |
4565 | |
4566 | /* Removing a stmt may expose a forwarder block. */ | |
4567 | el_todo |= TODO_cleanup_cfg; | |
5006671f | 4568 | } |
9771b263 | 4569 | el_to_remove.release (); |
5006671f | 4570 | |
40b178f4 RG |
4571 | return el_todo; |
4572 | } | |
4573 | ||
4574 | /* Perform CFG cleanups made necessary by elimination. */ | |
4575 | ||
9c370032 | 4576 | static unsigned |
40b178f4 RG |
4577 | fini_eliminate (void) |
4578 | { | |
4579 | bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup); | |
4580 | bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup); | |
4581 | ||
4582 | if (do_eh_cleanup) | |
4583 | gimple_purge_all_dead_eh_edges (need_eh_cleanup); | |
4584 | ||
4585 | if (do_ab_cleanup) | |
4586 | gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup); | |
4587 | ||
4588 | BITMAP_FREE (need_eh_cleanup); | |
4589 | BITMAP_FREE (need_ab_cleanup); | |
4590 | ||
4591 | if (do_eh_cleanup || do_ab_cleanup) | |
9c370032 RB |
4592 | return TODO_cleanup_cfg; |
4593 | return 0; | |
6de9cd9a DN |
4594 | } |
4595 | ||
0fc6c492 DB |
4596 | /* Borrow a bit of tree-ssa-dce.c for the moment. |
4597 | XXX: In 4.1, we should be able to just run a DCE pass after PRE, though | |
4598 | this may be a bit faster, and we may want critical edges kept split. */ | |
4599 | ||
4600 | /* If OP's defining statement has not already been determined to be necessary, | |
d4e6fecb | 4601 | mark that statement necessary. Return the stmt, if it is newly |
b9c5e484 | 4602 | necessary. */ |
0fc6c492 | 4603 | |
726a989a | 4604 | static inline gimple |
d4e6fecb | 4605 | mark_operand_necessary (tree op) |
0fc6c492 | 4606 | { |
726a989a | 4607 | gimple stmt; |
0fc6c492 DB |
4608 | |
4609 | gcc_assert (op); | |
4610 | ||
c90186eb DB |
4611 | if (TREE_CODE (op) != SSA_NAME) |
4612 | return NULL; | |
4613 | ||
0fc6c492 DB |
4614 | stmt = SSA_NAME_DEF_STMT (op); |
4615 | gcc_assert (stmt); | |
4616 | ||
726a989a RB |
4617 | if (gimple_plf (stmt, NECESSARY) |
4618 | || gimple_nop_p (stmt)) | |
d4e6fecb | 4619 | return NULL; |
0fc6c492 | 4620 | |
726a989a | 4621 | gimple_set_plf (stmt, NECESSARY, true); |
d4e6fecb | 4622 | return stmt; |
0fc6c492 DB |
4623 | } |
4624 | ||
4625 | /* Because we don't follow exactly the standard PRE algorithm, and decide not | |
4626 | to insert PHI nodes sometimes, and because value numbering of casts isn't | |
4627 | perfect, we sometimes end up inserting dead code. This simple DCE-like | |
4628 | pass removes any insertions we made that weren't actually used. */ | |
4629 | ||
4630 | static void | |
4631 | remove_dead_inserted_code (void) | |
4632 | { | |
0ab555de MM |
4633 | bitmap worklist; |
4634 | unsigned i; | |
4635 | bitmap_iterator bi; | |
726a989a | 4636 | gimple t; |
0fc6c492 | 4637 | |
0ab555de MM |
4638 | worklist = BITMAP_ALLOC (NULL); |
4639 | EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi) | |
0fc6c492 | 4640 | { |
0ab555de | 4641 | t = SSA_NAME_DEF_STMT (ssa_name (i)); |
726a989a | 4642 | if (gimple_plf (t, NECESSARY)) |
0ab555de | 4643 | bitmap_set_bit (worklist, i); |
0fc6c492 | 4644 | } |
0ab555de | 4645 | while (!bitmap_empty_p (worklist)) |
0fc6c492 | 4646 | { |
0ab555de MM |
4647 | i = bitmap_first_set_bit (worklist); |
4648 | bitmap_clear_bit (worklist, i); | |
4649 | t = SSA_NAME_DEF_STMT (ssa_name (i)); | |
c90186eb DB |
4650 | |
4651 | /* PHI nodes are somewhat special in that each PHI alternative has | |
4652 | data and control dependencies. All the statements feeding the | |
4653 | PHI node's arguments are always necessary. */ | |
726a989a | 4654 | if (gimple_code (t) == GIMPLE_PHI) |
0fc6c492 | 4655 | { |
726a989a | 4656 | unsigned k; |
d4e6fecb | 4657 | |
726a989a | 4658 | for (k = 0; k < gimple_phi_num_args (t); k++) |
83737db2 | 4659 | { |
0fc6c492 DB |
4660 | tree arg = PHI_ARG_DEF (t, k); |
4661 | if (TREE_CODE (arg) == SSA_NAME) | |
d4e6fecb | 4662 | { |
726a989a RB |
4663 | gimple n = mark_operand_necessary (arg); |
4664 | if (n) | |
0ab555de | 4665 | bitmap_set_bit (worklist, SSA_NAME_VERSION (arg)); |
d4e6fecb | 4666 | } |
0fc6c492 DB |
4667 | } |
4668 | } | |
4669 | else | |
4670 | { | |
4671 | /* Propagate through the operands. Examine all the USE, VUSE and | |
89fb70a3 | 4672 | VDEF operands in this statement. Mark all the statements |
0fc6c492 DB |
4673 | which feed this statement's uses as necessary. */ |
4674 | ssa_op_iter iter; | |
4675 | tree use; | |
4676 | ||
38635499 | 4677 | /* The operands of VDEF expressions are also needed as they |
0fc6c492 | 4678 | represent potential definitions that may reach this |
89fb70a3 | 4679 | statement (VDEF operands allow us to follow def-def |
0fc6c492 | 4680 | links). */ |
33c94679 | 4681 | |
0fc6c492 | 4682 | FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES) |
d4e6fecb | 4683 | { |
726a989a | 4684 | gimple n = mark_operand_necessary (use); |
d4e6fecb | 4685 | if (n) |
0ab555de | 4686 | bitmap_set_bit (worklist, SSA_NAME_VERSION (use)); |
d4e6fecb | 4687 | } |
0fc6c492 DB |
4688 | } |
4689 | } | |
c90186eb | 4690 | |
0ab555de | 4691 | EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi) |
0fc6c492 | 4692 | { |
0ab555de | 4693 | t = SSA_NAME_DEF_STMT (ssa_name (i)); |
726a989a | 4694 | if (!gimple_plf (t, NECESSARY)) |
0fc6c492 | 4695 | { |
726a989a | 4696 | gimple_stmt_iterator gsi; |
c90186eb | 4697 | |
0fc6c492 DB |
4698 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4699 | { | |
4700 | fprintf (dump_file, "Removing unnecessary insertion:"); | |
726a989a | 4701 | print_gimple_stmt (dump_file, t, 0, 0); |
0fc6c492 | 4702 | } |
c90186eb | 4703 | |
726a989a RB |
4704 | gsi = gsi_for_stmt (t); |
4705 | if (gimple_code (t) == GIMPLE_PHI) | |
4706 | remove_phi_node (&gsi, true); | |
0fc6c492 | 4707 | else |
b70fdfe4 AO |
4708 | { |
4709 | gsi_remove (&gsi, true); | |
4710 | release_defs (t); | |
4711 | } | |
0fc6c492 DB |
4712 | } |
4713 | } | |
0ab555de | 4714 | BITMAP_FREE (worklist); |
0fc6c492 | 4715 | } |
c90186eb | 4716 | |
9ca87236 | 4717 | |
ff2ad0f7 | 4718 | /* Initialize data structures used by PRE. */ |
6de9cd9a DN |
4719 | |
4720 | static void | |
40b178f4 | 4721 | init_pre (void) |
6de9cd9a | 4722 | { |
c9145754 DB |
4723 | basic_block bb; |
4724 | ||
4725 | next_expression_id = 1; | |
9771b263 DN |
4726 | expressions.create (0); |
4727 | expressions.safe_push (NULL); | |
4728 | value_expressions.create (get_max_value_id () + 1); | |
c3284718 | 4729 | value_expressions.safe_grow_cleared (get_max_value_id () + 1); |
9771b263 | 4730 | name_to_id.create (0); |
c9145754 | 4731 | |
0ab555de | 4732 | inserted_exprs = BITMAP_ALLOC (NULL); |
c90186eb | 4733 | |
b71b4522 DB |
4734 | connect_infinite_loops_to_exit (); |
4735 | memset (&pre_stats, 0, sizeof (pre_stats)); | |
4736 | ||
0cae8d31 | 4737 | postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); |
585d0dc4 | 4738 | postorder_num = inverted_post_order_compute (postorder); |
c9145754 | 4739 | |
a72ae88a | 4740 | alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets)); |
c9145754 | 4741 | |
b71b4522 | 4742 | calculate_dominance_info (CDI_POST_DOMINATORS); |
d75dbccd DB |
4743 | calculate_dominance_info (CDI_DOMINATORS); |
4744 | ||
c9145754 | 4745 | bitmap_obstack_initialize (&grand_bitmap_obstack); |
c203e8a7 TS |
4746 | phi_translate_table = new hash_table<expr_pred_trans_d> (5110); |
4747 | expression_to_id = new hash_table<pre_expr_d> (num_ssa_names * 3); | |
c9145754 DB |
4748 | bitmap_set_pool = create_alloc_pool ("Bitmap sets", |
4749 | sizeof (struct bitmap_set), 30); | |
4750 | pre_expr_pool = create_alloc_pool ("pre_expr nodes", | |
4751 | sizeof (struct pre_expr_d), 30); | |
04a90bec | 4752 | FOR_ALL_BB_FN (bb, cfun) |
c9145754 | 4753 | { |
40b178f4 RG |
4754 | EXP_GEN (bb) = bitmap_set_new (); |
4755 | PHI_GEN (bb) = bitmap_set_new (); | |
4756 | TMP_GEN (bb) = bitmap_set_new (); | |
c9145754 DB |
4757 | AVAIL_OUT (bb) = bitmap_set_new (); |
4758 | } | |
ff2ad0f7 DN |
4759 | } |
4760 | ||
4761 | ||
4762 | /* Deallocate data structures used by PRE. */ | |
6de9cd9a | 4763 | |
ff2ad0f7 | 4764 | static void |
40b178f4 | 4765 | fini_pre () |
ff2ad0f7 | 4766 | { |
c9145754 | 4767 | free (postorder); |
9771b263 | 4768 | value_expressions.release (); |
0ab555de | 4769 | BITMAP_FREE (inserted_exprs); |
c9145754 DB |
4770 | bitmap_obstack_release (&grand_bitmap_obstack); |
4771 | free_alloc_pool (bitmap_set_pool); | |
4772 | free_alloc_pool (pre_expr_pool); | |
c203e8a7 TS |
4773 | delete phi_translate_table; |
4774 | phi_translate_table = NULL; | |
4775 | delete expression_to_id; | |
4776 | expression_to_id = NULL; | |
9771b263 | 4777 | name_to_id.release (); |
50265400 | 4778 | |
a72ae88a | 4779 | free_aux_for_blocks (); |
6809cbf9 | 4780 | |
b71b4522 | 4781 | free_dominance_info (CDI_POST_DOMINATORS); |
ff2ad0f7 DN |
4782 | } |
4783 | ||
be55bfe6 | 4784 | namespace { |
ff2ad0f7 | 4785 | |
be55bfe6 TS |
4786 | const pass_data pass_data_pre = |
4787 | { | |
4788 | GIMPLE_PASS, /* type */ | |
4789 | "pre", /* name */ | |
4790 | OPTGROUP_NONE, /* optinfo_flags */ | |
be55bfe6 TS |
4791 | TV_TREE_PRE, /* tv_id */ |
4792 | /* PROP_no_crit_edges is ensured by placing pass_split_crit_edges before | |
4793 | pass_pre. */ | |
4794 | ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */ | |
4795 | 0, /* properties_provided */ | |
4796 | PROP_no_crit_edges, /* properties_destroyed */ | |
4797 | TODO_rebuild_alias, /* todo_flags_start */ | |
3bea341f | 4798 | 0, /* todo_flags_finish */ |
be55bfe6 TS |
4799 | }; |
4800 | ||
4801 | class pass_pre : public gimple_opt_pass | |
4802 | { | |
4803 | public: | |
4804 | pass_pre (gcc::context *ctxt) | |
4805 | : gimple_opt_pass (pass_data_pre, ctxt) | |
4806 | {} | |
4807 | ||
4808 | /* opt_pass methods: */ | |
4809 | virtual bool gate (function *) { return flag_tree_pre != 0; } | |
4810 | virtual unsigned int execute (function *); | |
4811 | ||
4812 | }; // class pass_pre | |
4813 | ||
4814 | unsigned int | |
4815 | pass_pre::execute (function *fun) | |
ff2ad0f7 | 4816 | { |
b80280f2 | 4817 | unsigned int todo = 0; |
83737db2 | 4818 | |
fa06ad0d | 4819 | do_partial_partial = |
be55bfe6 | 4820 | flag_tree_partial_pre && optimize_function_for_speed_p (fun); |
ff2ad0f7 | 4821 | |
c9145754 DB |
4822 | /* This has to happen before SCCVN runs because |
4823 | loop_optimizer_init may create new phis, etc. */ | |
40b178f4 | 4824 | loop_optimizer_init (LOOPS_NORMAL); |
c90186eb | 4825 | |
40b178f4 | 4826 | if (!run_scc_vn (VN_WALK)) |
863d2a57 | 4827 | { |
40b178f4 | 4828 | loop_optimizer_finalize (); |
b80280f2 | 4829 | return 0; |
863d2a57 | 4830 | } |
a7d4562a | 4831 | |
40b178f4 | 4832 | init_pre (); |
a8338640 | 4833 | scev_initialize (); |
c9145754 | 4834 | |
c9145754 | 4835 | /* Collect and value number expressions computed in each basic block. */ |
665fcad8 | 4836 | compute_avail (); |
6de9cd9a | 4837 | |
7e6eb623 DB |
4838 | /* Insert can get quite slow on an incredibly large number of basic |
4839 | blocks due to some quadratic behavior. Until this behavior is | |
4840 | fixed, don't run it when he have an incredibly large number of | |
4841 | bb's. If we aren't going to run insert, there is no point in | |
4842 | computing ANTIC, either, even though it's plenty fast. */ | |
be55bfe6 | 4843 | if (n_basic_blocks_for_fn (fun) < 4000) |
6de9cd9a | 4844 | { |
7e6eb623 | 4845 | compute_antic (); |
7e6eb623 DB |
4846 | insert (); |
4847 | } | |
ff2ad0f7 | 4848 | |
0ab555de MM |
4849 | /* Make sure to remove fake edges before committing our inserts. |
4850 | This makes sure we don't end up with extra critical edges that | |
4851 | we would need to split. */ | |
4852 | remove_fake_exit_edges (); | |
4853 | gsi_commit_edge_inserts (); | |
4854 | ||
ff2ad0f7 | 4855 | /* Remove all the redundant expressions. */ |
b82ef848 | 4856 | todo |= eliminate (true); |
0fc6c492 | 4857 | |
be55bfe6 TS |
4858 | statistics_counter_event (fun, "Insertions", pre_stats.insertions); |
4859 | statistics_counter_event (fun, "PA inserted", pre_stats.pa_insert); | |
4860 | statistics_counter_event (fun, "New PHIs", pre_stats.phis); | |
4861 | statistics_counter_event (fun, "Eliminated", pre_stats.eliminations); | |
c4ab2baa | 4862 | |
b71b4522 | 4863 | clear_expression_ids (); |
40b178f4 | 4864 | remove_dead_inserted_code (); |
c90186eb | 4865 | |
a8338640 | 4866 | scev_finalize (); |
40b178f4 | 4867 | fini_pre (); |
9c370032 | 4868 | todo |= fini_eliminate (); |
40b178f4 RG |
4869 | loop_optimizer_finalize (); |
4870 | ||
4871 | /* TODO: tail_merge_optimize may merge all predecessors of a block, in which | |
4872 | case we can merge the block with the remaining predecessor of the block. | |
4873 | It should either: | |
4874 | - call merge_blocks after each tail merge iteration | |
4875 | - call merge_blocks after all tail merge iterations | |
4876 | - mark TODO_cleanup_cfg when necessary | |
4877 | - share the cfg cleanup with fini_pre. */ | |
4878 | todo |= tail_merge_optimize (todo); | |
4879 | ||
c9e93168 TV |
4880 | free_scc_vn (); |
4881 | ||
678771ad RG |
4882 | /* Tail merging invalidates the virtual SSA web, together with |
4883 | cfg-cleanup opportunities exposed by PRE this will wreck the | |
4884 | SSA updating machinery. So make sure to run update-ssa | |
4885 | manually, before eventually scheduling cfg-cleanup as part of | |
4886 | the todo. */ | |
4887 | update_ssa (TODO_update_ssa_only_virtuals); | |
4888 | ||
b80280f2 | 4889 | return todo; |
ff2ad0f7 DN |
4890 | } |
4891 | ||
27a4cd48 DM |
4892 | } // anon namespace |
4893 | ||
4894 | gimple_opt_pass * | |
4895 | make_pass_pre (gcc::context *ctxt) | |
4896 | { | |
4897 | return new pass_pre (ctxt); | |
4898 | } | |
4899 | ||
27a4cd48 DM |
4900 | namespace { |
4901 | ||
4902 | const pass_data pass_data_fre = | |
ff2ad0f7 | 4903 | { |
27a4cd48 DM |
4904 | GIMPLE_PASS, /* type */ |
4905 | "fre", /* name */ | |
4906 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
4907 | TV_TREE_FRE, /* tv_id */ |
4908 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
4909 | 0, /* properties_provided */ | |
4910 | 0, /* properties_destroyed */ | |
4911 | 0, /* todo_flags_start */ | |
3bea341f | 4912 | 0, /* todo_flags_finish */ |
ff2ad0f7 | 4913 | }; |
27a4cd48 DM |
4914 | |
4915 | class pass_fre : public gimple_opt_pass | |
4916 | { | |
4917 | public: | |
c3284718 RS |
4918 | pass_fre (gcc::context *ctxt) |
4919 | : gimple_opt_pass (pass_data_fre, ctxt) | |
27a4cd48 DM |
4920 | {} |
4921 | ||
4922 | /* opt_pass methods: */ | |
65d3284b | 4923 | opt_pass * clone () { return new pass_fre (m_ctxt); } |
1a3d085c | 4924 | virtual bool gate (function *) { return flag_tree_fre != 0; } |
be55bfe6 | 4925 | virtual unsigned int execute (function *); |
27a4cd48 DM |
4926 | |
4927 | }; // class pass_fre | |
4928 | ||
be55bfe6 TS |
4929 | unsigned int |
4930 | pass_fre::execute (function *fun) | |
4931 | { | |
4932 | unsigned int todo = 0; | |
4933 | ||
4934 | if (!run_scc_vn (VN_WALKREWRITE)) | |
4935 | return 0; | |
4936 | ||
4937 | memset (&pre_stats, 0, sizeof (pre_stats)); | |
4938 | ||
4939 | /* Remove all the redundant expressions. */ | |
b82ef848 | 4940 | todo |= eliminate (false); |
be55bfe6 TS |
4941 | |
4942 | todo |= fini_eliminate (); | |
4943 | ||
4944 | free_scc_vn (); | |
4945 | ||
4946 | statistics_counter_event (fun, "Insertions", pre_stats.insertions); | |
4947 | statistics_counter_event (fun, "Eliminated", pre_stats.eliminations); | |
4948 | ||
4949 | return todo; | |
4950 | } | |
4951 | ||
27a4cd48 DM |
4952 | } // anon namespace |
4953 | ||
4954 | gimple_opt_pass * | |
4955 | make_pass_fre (gcc::context *ctxt) | |
4956 | { | |
4957 | return new pass_fre (ctxt); | |
4958 | } |