]>
Commit | Line | Data |
---|---|---|
6de9cd9a | 1 | /* SSA-PRE for trees. |
d1e082c2 | 2 | Copyright (C) 2001-2013 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" |
7a300452 | 30 | #include "tree-ssa.h" |
726a989a | 31 | #include "gimple.h" |
0823efed | 32 | #include "hash-table.h" |
6de9cd9a | 33 | #include "tree-iterator.h" |
6de9cd9a | 34 | #include "alloc-pool.h" |
5039610b | 35 | #include "obstack.h" |
6de9cd9a DN |
36 | #include "tree-pass.h" |
37 | #include "flags.h" | |
7e6eb623 DB |
38 | #include "bitmap.h" |
39 | #include "langhooks.h" | |
0fc6c492 | 40 | #include "cfgloop.h" |
89fb70a3 | 41 | #include "tree-ssa-sccvn.h" |
a8338640 | 42 | #include "tree-scalar-evolution.h" |
f0ed4cfb | 43 | #include "params.h" |
c9145754 | 44 | #include "dbgcnt.h" |
40b178f4 | 45 | #include "domwalk.h" |
e248d83f | 46 | #include "ipa-prop.h" |
33c94679 | 47 | |
7e6eb623 | 48 | /* TODO: |
b9c5e484 | 49 | |
bdee7684 | 50 | 1. Avail sets can be shared by making an avail_find_leader that |
7e6eb623 DB |
51 | walks up the dominator tree and looks in those avail sets. |
52 | This might affect code optimality, it's unclear right now. | |
c90186eb | 53 | 2. Strength reduction can be performed by anticipating expressions |
7e6eb623 | 54 | we can repair later on. |
c90186eb | 55 | 3. We can do back-substitution or smarter value numbering to catch |
0fc6c492 | 56 | commutative expressions split up over multiple statements. |
b9c5e484 | 57 | */ |
7e6eb623 DB |
58 | |
59 | /* For ease of terminology, "expression node" in the below refers to | |
726a989a | 60 | every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs |
07beea0d AH |
61 | represent the actual statement containing the expressions we care about, |
62 | and we cache the value number by putting it in the expression. */ | |
7e6eb623 DB |
63 | |
64 | /* Basic algorithm | |
b9c5e484 | 65 | |
56db793a DB |
66 | First we walk the statements to generate the AVAIL sets, the |
67 | EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the | |
68 | generation of values/expressions by a given block. We use them | |
69 | when computing the ANTIC sets. The AVAIL sets consist of | |
70 | SSA_NAME's that represent values, so we know what values are | |
71 | available in what blocks. AVAIL is a forward dataflow problem. In | |
72 | SSA, values are never killed, so we don't need a kill set, or a | |
73 | fixpoint iteration, in order to calculate the AVAIL sets. In | |
74 | traditional parlance, AVAIL sets tell us the downsafety of the | |
7e6eb623 | 75 | expressions/values. |
b9c5e484 | 76 | |
56db793a DB |
77 | Next, we generate the ANTIC sets. These sets represent the |
78 | anticipatable expressions. ANTIC is a backwards dataflow | |
d75dbccd | 79 | problem. An expression is anticipatable in a given block if it could |
56db793a DB |
80 | be generated in that block. This means that if we had to perform |
81 | an insertion in that block, of the value of that expression, we | |
82 | could. Calculating the ANTIC sets requires phi translation of | |
83 | expressions, because the flow goes backwards through phis. We must | |
84 | iterate to a fixpoint of the ANTIC sets, because we have a kill | |
85 | set. Even in SSA form, values are not live over the entire | |
86 | function, only from their definition point onwards. So we have to | |
87 | remove values from the ANTIC set once we go past the definition | |
88 | point of the leaders that make them up. | |
89 | compute_antic/compute_antic_aux performs this computation. | |
7e6eb623 DB |
90 | |
91 | Third, we perform insertions to make partially redundant | |
92 | expressions fully redundant. | |
93 | ||
94 | An expression is partially redundant (excluding partial | |
95 | anticipation) if: | |
96 | ||
97 | 1. It is AVAIL in some, but not all, of the predecessors of a | |
98 | given block. | |
99 | 2. It is ANTIC in all the predecessors. | |
100 | ||
101 | In order to make it fully redundant, we insert the expression into | |
102 | the predecessors where it is not available, but is ANTIC. | |
d75dbccd DB |
103 | |
104 | For the partial anticipation case, we only perform insertion if it | |
105 | is partially anticipated in some block, and fully available in all | |
106 | of the predecessors. | |
107 | ||
108 | insert/insert_aux/do_regular_insertion/do_partial_partial_insertion | |
109 | performs these steps. | |
7e6eb623 DB |
110 | |
111 | Fourth, we eliminate fully redundant expressions. | |
112 | This is a simple statement walk that replaces redundant | |
070b797d | 113 | calculations with the now available values. */ |
7e6eb623 DB |
114 | |
115 | /* Representations of value numbers: | |
116 | ||
c9145754 DB |
117 | Value numbers are represented by a representative SSA_NAME. We |
118 | will create fake SSA_NAME's in situations where we need a | |
119 | representative but do not have one (because it is a complex | |
120 | expression). In order to facilitate storing the value numbers in | |
121 | bitmaps, and keep the number of wasted SSA_NAME's down, we also | |
122 | associate a value_id with each value number, and create full blown | |
123 | ssa_name's only where we actually need them (IE in operands of | |
124 | existing expressions). | |
125 | ||
126 | Theoretically you could replace all the value_id's with | |
127 | SSA_NAME_VERSION, but this would allocate a large number of | |
128 | SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number. | |
129 | It would also require an additional indirection at each point we | |
130 | use the value id. */ | |
7e6eb623 | 131 | |
b9c5e484 | 132 | /* Representation of expressions on value numbers: |
7e6eb623 | 133 | |
249eb506 | 134 | Expressions consisting of value numbers are represented the same |
c9145754 DB |
135 | way as our VN internally represents them, with an additional |
136 | "pre_expr" wrapping around them in order to facilitate storing all | |
137 | of the expressions in the same sets. */ | |
7e6eb623 | 138 | |
c9145754 | 139 | /* Representation of sets: |
7e6eb623 | 140 | |
c9145754 DB |
141 | The dataflow sets do not need to be sorted in any particular order |
142 | for the majority of their lifetime, are simply represented as two | |
143 | bitmaps, one that keeps track of values present in the set, and one | |
144 | that keeps track of expressions present in the set. | |
7e6eb623 | 145 | |
c9145754 DB |
146 | When we need them in topological order, we produce it on demand by |
147 | transforming the bitmap into an array and sorting it into topo | |
148 | order. */ | |
7e6eb623 | 149 | |
c9145754 DB |
150 | /* Type of expression, used to know which member of the PRE_EXPR union |
151 | is valid. */ | |
7e6eb623 | 152 | |
c9145754 DB |
153 | enum pre_expr_kind |
154 | { | |
155 | NAME, | |
156 | NARY, | |
157 | REFERENCE, | |
158 | CONSTANT | |
159 | }; | |
160 | ||
161 | typedef union pre_expr_union_d | |
162 | { | |
163 | tree name; | |
164 | tree constant; | |
165 | vn_nary_op_t nary; | |
166 | vn_reference_t reference; | |
167 | } pre_expr_union; | |
b9c5e484 | 168 | |
5deac340 | 169 | typedef struct pre_expr_d : typed_noop_remove <pre_expr_d> |
c9145754 DB |
170 | { |
171 | enum pre_expr_kind kind; | |
172 | unsigned int id; | |
173 | pre_expr_union u; | |
5deac340 RG |
174 | |
175 | /* hash_table support. */ | |
5831a5f0 LC |
176 | typedef pre_expr_d value_type; |
177 | typedef pre_expr_d compare_type; | |
5deac340 RG |
178 | static inline hashval_t hash (const pre_expr_d *); |
179 | static inline int equal (const pre_expr_d *, const pre_expr_d *); | |
c9145754 | 180 | } *pre_expr; |
7e6eb623 | 181 | |
c9145754 DB |
182 | #define PRE_EXPR_NAME(e) (e)->u.name |
183 | #define PRE_EXPR_NARY(e) (e)->u.nary | |
184 | #define PRE_EXPR_REFERENCE(e) (e)->u.reference | |
185 | #define PRE_EXPR_CONSTANT(e) (e)->u.constant | |
7e6eb623 | 186 | |
0823efed | 187 | /* Compare E1 and E1 for equality. */ |
7e6eb623 | 188 | |
0823efed | 189 | inline int |
5831a5f0 | 190 | pre_expr_d::equal (const value_type *e1, const compare_type *e2) |
0823efed | 191 | { |
c9145754 DB |
192 | if (e1->kind != e2->kind) |
193 | return false; | |
194 | ||
195 | switch (e1->kind) | |
196 | { | |
197 | case CONSTANT: | |
726a989a RB |
198 | return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1), |
199 | PRE_EXPR_CONSTANT (e2)); | |
c9145754 DB |
200 | case NAME: |
201 | return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2); | |
202 | case NARY: | |
203 | return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2)); | |
204 | case REFERENCE: | |
205 | return vn_reference_eq (PRE_EXPR_REFERENCE (e1), | |
206 | PRE_EXPR_REFERENCE (e2)); | |
207 | default: | |
9708c51d | 208 | gcc_unreachable (); |
c9145754 DB |
209 | } |
210 | } | |
211 | ||
0823efed DN |
212 | /* Hash E. */ |
213 | ||
214 | inline hashval_t | |
5831a5f0 | 215 | pre_expr_d::hash (const value_type *e) |
c9145754 | 216 | { |
c9145754 DB |
217 | switch (e->kind) |
218 | { | |
219 | case CONSTANT: | |
726a989a | 220 | return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e)); |
c9145754 | 221 | case NAME: |
9708c51d | 222 | return SSA_NAME_VERSION (PRE_EXPR_NAME (e)); |
c9145754 | 223 | case NARY: |
85169114 | 224 | return PRE_EXPR_NARY (e)->hashcode; |
c9145754 | 225 | case REFERENCE: |
85169114 | 226 | return PRE_EXPR_REFERENCE (e)->hashcode; |
c9145754 | 227 | default: |
9708c51d | 228 | gcc_unreachable (); |
c9145754 DB |
229 | } |
230 | } | |
83737db2 | 231 | |
c9145754 DB |
232 | /* Next global expression id number. */ |
233 | static unsigned int next_expression_id; | |
070b797d | 234 | |
83737db2 | 235 | /* Mapping from expression to id number we can use in bitmap sets. */ |
9771b263 | 236 | static vec<pre_expr> expressions; |
5deac340 | 237 | static hash_table <pre_expr_d> expression_to_id; |
9771b263 | 238 | static vec<unsigned> name_to_id; |
81def1b7 | 239 | |
83737db2 DB |
240 | /* Allocate an expression id for EXPR. */ |
241 | ||
242 | static inline unsigned int | |
c9145754 | 243 | alloc_expression_id (pre_expr expr) |
6de9cd9a | 244 | { |
0823efed | 245 | struct pre_expr_d **slot; |
83737db2 DB |
246 | /* Make sure we won't overflow. */ |
247 | gcc_assert (next_expression_id + 1 > next_expression_id); | |
c9145754 | 248 | expr->id = next_expression_id++; |
9771b263 | 249 | expressions.safe_push (expr); |
13de9095 RG |
250 | if (expr->kind == NAME) |
251 | { | |
252 | unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr)); | |
9771b263 DN |
253 | /* vec::safe_grow_cleared allocates no headroom. Avoid frequent |
254 | re-allocations by using vec::reserve upfront. There is no | |
255 | vec::quick_grow_cleared unfortunately. */ | |
256 | unsigned old_len = name_to_id.length (); | |
257 | name_to_id.reserve (num_ssa_names - old_len); | |
258 | name_to_id.safe_grow_cleared (num_ssa_names); | |
259 | gcc_assert (name_to_id[version] == 0); | |
260 | name_to_id[version] = expr->id; | |
13de9095 RG |
261 | } |
262 | else | |
263 | { | |
0823efed | 264 | slot = expression_to_id.find_slot (expr, INSERT); |
13de9095 RG |
265 | gcc_assert (!*slot); |
266 | *slot = expr; | |
267 | } | |
83737db2 DB |
268 | return next_expression_id - 1; |
269 | } | |
270 | ||
271 | /* Return the expression id for tree EXPR. */ | |
272 | ||
273 | static inline unsigned int | |
c9145754 DB |
274 | get_expression_id (const pre_expr expr) |
275 | { | |
276 | return expr->id; | |
277 | } | |
278 | ||
279 | static inline unsigned int | |
280 | lookup_expression_id (const pre_expr expr) | |
7e6eb623 | 281 | { |
0823efed | 282 | struct pre_expr_d **slot; |
b71b4522 | 283 | |
13de9095 RG |
284 | if (expr->kind == NAME) |
285 | { | |
286 | unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr)); | |
9771b263 | 287 | if (name_to_id.length () <= version) |
13de9095 | 288 | return 0; |
9771b263 | 289 | return name_to_id[version]; |
13de9095 RG |
290 | } |
291 | else | |
292 | { | |
0823efed | 293 | slot = expression_to_id.find_slot (expr, NO_INSERT); |
13de9095 RG |
294 | if (!slot) |
295 | return 0; | |
296 | return ((pre_expr)*slot)->id; | |
297 | } | |
83737db2 | 298 | } |
b9c5e484 | 299 | |
83737db2 DB |
300 | /* Return the existing expression id for EXPR, or create one if one |
301 | does not exist yet. */ | |
b9c5e484 | 302 | |
83737db2 | 303 | static inline unsigned int |
c9145754 | 304 | get_or_alloc_expression_id (pre_expr expr) |
83737db2 | 305 | { |
c9145754 DB |
306 | unsigned int id = lookup_expression_id (expr); |
307 | if (id == 0) | |
83737db2 | 308 | return alloc_expression_id (expr); |
c9145754 | 309 | return expr->id = id; |
83737db2 DB |
310 | } |
311 | ||
312 | /* Return the expression that has expression id ID */ | |
313 | ||
c9145754 | 314 | static inline pre_expr |
83737db2 DB |
315 | expression_for_id (unsigned int id) |
316 | { | |
9771b263 | 317 | return expressions[id]; |
c5830edf DB |
318 | } |
319 | ||
b71b4522 DB |
320 | /* Free the expression id field in all of our expressions, |
321 | and then destroy the expressions array. */ | |
322 | ||
323 | static void | |
324 | clear_expression_ids (void) | |
325 | { | |
9771b263 | 326 | expressions.release (); |
c9145754 | 327 | } |
b71b4522 | 328 | |
c9145754 DB |
329 | static alloc_pool pre_expr_pool; |
330 | ||
331 | /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */ | |
332 | ||
333 | static pre_expr | |
334 | get_or_alloc_expr_for_name (tree name) | |
335 | { | |
5f5126d6 RG |
336 | struct pre_expr_d expr; |
337 | pre_expr result; | |
c9145754 DB |
338 | unsigned int result_id; |
339 | ||
5f5126d6 RG |
340 | expr.kind = NAME; |
341 | expr.id = 0; | |
342 | PRE_EXPR_NAME (&expr) = name; | |
343 | result_id = lookup_expression_id (&expr); | |
344 | if (result_id != 0) | |
345 | return expression_for_id (result_id); | |
346 | ||
347 | result = (pre_expr) pool_alloc (pre_expr_pool); | |
c9145754 | 348 | result->kind = NAME; |
c9145754 | 349 | PRE_EXPR_NAME (result) = name; |
5f5126d6 | 350 | alloc_expression_id (result); |
c9145754 | 351 | return result; |
b71b4522 DB |
352 | } |
353 | ||
bdee7684 | 354 | /* An unordered bitmap set. One bitmap tracks values, the other, |
8c27b7d4 | 355 | expressions. */ |
bdee7684 DB |
356 | typedef struct bitmap_set |
357 | { | |
5c72d561 JH |
358 | bitmap_head expressions; |
359 | bitmap_head values; | |
bdee7684 DB |
360 | } *bitmap_set_t; |
361 | ||
83737db2 | 362 | #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \ |
5c72d561 | 363 | EXECUTE_IF_SET_IN_BITMAP(&(set)->expressions, 0, (id), (bi)) |
c9145754 | 364 | |
85169114 | 365 | #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \ |
5c72d561 | 366 | EXECUTE_IF_SET_IN_BITMAP(&(set)->values, 0, (id), (bi)) |
85169114 | 367 | |
c9145754 | 368 | /* Mapping from value id to expressions with that value_id. */ |
9771b263 | 369 | static vec<bitmap> value_expressions; |
83737db2 | 370 | |
6b416da1 | 371 | /* Sets that we need to keep track of. */ |
83737db2 | 372 | typedef struct bb_bitmap_sets |
6de9cd9a | 373 | { |
ca072a31 DB |
374 | /* The EXP_GEN set, which represents expressions/values generated in |
375 | a basic block. */ | |
83737db2 | 376 | bitmap_set_t exp_gen; |
ca072a31 DB |
377 | |
378 | /* The PHI_GEN set, which represents PHI results generated in a | |
379 | basic block. */ | |
6b416da1 | 380 | bitmap_set_t phi_gen; |
ca072a31 | 381 | |
f6fe65dc | 382 | /* The TMP_GEN set, which represents results/temporaries generated |
ca072a31 | 383 | in a basic block. IE the LHS of an expression. */ |
6b416da1 | 384 | bitmap_set_t tmp_gen; |
ca072a31 DB |
385 | |
386 | /* The AVAIL_OUT set, which represents which values are available in | |
387 | a given basic block. */ | |
bdee7684 | 388 | bitmap_set_t avail_out; |
ca072a31 | 389 | |
c90186eb | 390 | /* The ANTIC_IN set, which represents which values are anticipatable |
ca072a31 | 391 | in a given basic block. */ |
83737db2 | 392 | bitmap_set_t antic_in; |
ca072a31 | 393 | |
d75dbccd DB |
394 | /* The PA_IN set, which represents which values are |
395 | partially anticipatable in a given basic block. */ | |
396 | bitmap_set_t pa_in; | |
397 | ||
ca072a31 DB |
398 | /* The NEW_SETS set, which is used during insertion to augment the |
399 | AVAIL_OUT set of blocks with the new insertions performed during | |
400 | the current iteration. */ | |
6b416da1 | 401 | bitmap_set_t new_sets; |
c90186eb | 402 | |
5006671f RG |
403 | /* A cache for value_dies_in_block_x. */ |
404 | bitmap expr_dies; | |
405 | ||
d75dbccd | 406 | /* True if we have visited this block during ANTIC calculation. */ |
a19eb9d2 | 407 | unsigned int visited : 1; |
d75dbccd DB |
408 | |
409 | /* True we have deferred processing this block during ANTIC | |
410 | calculation until its successor is processed. */ | |
411 | unsigned int deferred : 1; | |
a19eb9d2 RG |
412 | |
413 | /* True when the block contains a call that might not return. */ | |
414 | unsigned int contains_may_not_return_call : 1; | |
d75dbccd DB |
415 | } *bb_value_sets_t; |
416 | ||
417 | #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen | |
418 | #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen | |
419 | #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen | |
420 | #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out | |
421 | #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in | |
422 | #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in | |
d75dbccd | 423 | #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets |
5006671f RG |
424 | #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies |
425 | #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited | |
d75dbccd | 426 | #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred |
a19eb9d2 | 427 | #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call |
d75dbccd | 428 | |
c9145754 | 429 | |
83737db2 DB |
430 | /* Basic block list in postorder. */ |
431 | static int *postorder; | |
585d0dc4 | 432 | static int postorder_num; |
6de9cd9a | 433 | |
ca072a31 DB |
434 | /* This structure is used to keep track of statistics on what |
435 | optimization PRE was able to perform. */ | |
7e6eb623 | 436 | static struct |
6de9cd9a | 437 | { |
ca072a31 | 438 | /* The number of RHS computations eliminated by PRE. */ |
7e6eb623 | 439 | int eliminations; |
ca072a31 DB |
440 | |
441 | /* The number of new expressions/temporaries generated by PRE. */ | |
7e6eb623 | 442 | int insertions; |
ca072a31 | 443 | |
d75dbccd DB |
444 | /* The number of inserts found due to partial anticipation */ |
445 | int pa_insert; | |
446 | ||
ca072a31 | 447 | /* The number of new PHI nodes added by PRE. */ |
7e6eb623 DB |
448 | int phis; |
449 | } pre_stats; | |
6de9cd9a | 450 | |
d75dbccd | 451 | static bool do_partial_partial; |
e076271b | 452 | static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int); |
c9145754 DB |
453 | static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr); |
454 | static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr); | |
bdee7684 | 455 | static void bitmap_set_copy (bitmap_set_t, bitmap_set_t); |
c9145754 DB |
456 | static bool bitmap_set_contains_value (bitmap_set_t, unsigned int); |
457 | static void bitmap_insert_into_set (bitmap_set_t, pre_expr); | |
9708c51d RG |
458 | static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, |
459 | unsigned int, bool); | |
bdee7684 | 460 | static bitmap_set_t bitmap_set_new (void); |
726a989a | 461 | static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *, |
e076271b RG |
462 | tree); |
463 | static tree find_or_generate_expression (basic_block, tree, gimple_seq *); | |
de081cfd | 464 | static unsigned int get_expr_value_id (pre_expr); |
6de9cd9a | 465 | |
7e6eb623 DB |
466 | /* We can add and remove elements and entries to and from sets |
467 | and hash tables, so we use alloc pools for them. */ | |
6de9cd9a | 468 | |
bdee7684 | 469 | static alloc_pool bitmap_set_pool; |
7932a3db | 470 | static bitmap_obstack grand_bitmap_obstack; |
6de9cd9a | 471 | |
30fd5881 | 472 | /* Set of blocks with statements that have had their EH properties changed. */ |
53b4bf74 DN |
473 | static bitmap need_eh_cleanup; |
474 | ||
30fd5881 EB |
475 | /* Set of blocks with statements that have had their AB properties changed. */ |
476 | static bitmap need_ab_cleanup; | |
477 | ||
7e6eb623 DB |
478 | /* A three tuple {e, pred, v} used to cache phi translations in the |
479 | phi_translate_table. */ | |
6de9cd9a | 480 | |
5deac340 | 481 | typedef struct expr_pred_trans_d : typed_free_remove<expr_pred_trans_d> |
6de9cd9a | 482 | { |
8c27b7d4 | 483 | /* The expression. */ |
c9145754 | 484 | pre_expr e; |
ca072a31 DB |
485 | |
486 | /* The predecessor block along which we translated the expression. */ | |
7e6eb623 | 487 | basic_block pred; |
ca072a31 DB |
488 | |
489 | /* The value that resulted from the translation. */ | |
c9145754 | 490 | pre_expr v; |
ca072a31 DB |
491 | |
492 | /* The hashcode for the expression, pred pair. This is cached for | |
493 | speed reasons. */ | |
7e6eb623 | 494 | hashval_t hashcode; |
5deac340 RG |
495 | |
496 | /* hash_table support. */ | |
5831a5f0 LC |
497 | typedef expr_pred_trans_d value_type; |
498 | typedef expr_pred_trans_d compare_type; | |
499 | static inline hashval_t hash (const value_type *); | |
500 | static inline int equal (const value_type *, const compare_type *); | |
7e6eb623 | 501 | } *expr_pred_trans_t; |
741ac903 | 502 | typedef const struct expr_pred_trans_d *const_expr_pred_trans_t; |
6de9cd9a | 503 | |
0823efed | 504 | inline hashval_t |
5deac340 | 505 | expr_pred_trans_d::hash (const expr_pred_trans_d *e) |
7e6eb623 | 506 | { |
5deac340 | 507 | return e->hashcode; |
7e6eb623 | 508 | } |
6de9cd9a | 509 | |
0823efed | 510 | inline int |
5831a5f0 LC |
511 | expr_pred_trans_d::equal (const value_type *ve1, |
512 | const compare_type *ve2) | |
7e6eb623 | 513 | { |
7e6eb623 DB |
514 | basic_block b1 = ve1->pred; |
515 | basic_block b2 = ve2->pred; | |
b9c5e484 | 516 | |
ca072a31 DB |
517 | /* If they are not translations for the same basic block, they can't |
518 | be equal. */ | |
7e6eb623 | 519 | if (b1 != b2) |
6de9cd9a | 520 | return false; |
5deac340 | 521 | return pre_expr_d::equal (ve1->e, ve2->e); |
6de9cd9a DN |
522 | } |
523 | ||
0823efed DN |
524 | /* The phi_translate_table caches phi translations for a given |
525 | expression and predecessor. */ | |
5deac340 | 526 | static hash_table <expr_pred_trans_d> phi_translate_table; |
0823efed | 527 | |
ca072a31 | 528 | /* Search in the phi translation table for the translation of |
c9145754 | 529 | expression E in basic block PRED. |
c90186eb | 530 | Return the translated value, if found, NULL otherwise. */ |
6de9cd9a | 531 | |
c9145754 DB |
532 | static inline pre_expr |
533 | phi_trans_lookup (pre_expr e, basic_block pred) | |
6de9cd9a | 534 | { |
0823efed | 535 | expr_pred_trans_t *slot; |
ca072a31 | 536 | struct expr_pred_trans_d ept; |
c90186eb | 537 | |
ca072a31 DB |
538 | ept.e = e; |
539 | ept.pred = pred; | |
5deac340 | 540 | ept.hashcode = iterative_hash_hashval_t (pre_expr_d::hash (e), pred->index); |
0823efed | 541 | slot = phi_translate_table.find_slot_with_hash (&ept, ept.hashcode, |
7e6eb623 DB |
542 | NO_INSERT); |
543 | if (!slot) | |
544 | return NULL; | |
545 | else | |
0823efed | 546 | return (*slot)->v; |
7e6eb623 | 547 | } |
6de9cd9a | 548 | |
6de9cd9a | 549 | |
c9145754 | 550 | /* Add the tuple mapping from {expression E, basic block PRED} to |
ca072a31 | 551 | value V, to the phi translation table. */ |
7e6eb623 DB |
552 | |
553 | static inline void | |
c9145754 | 554 | phi_trans_add (pre_expr e, pre_expr v, basic_block pred) |
7e6eb623 | 555 | { |
0823efed | 556 | expr_pred_trans_t *slot; |
e1111e8e | 557 | expr_pred_trans_t new_pair = XNEW (struct expr_pred_trans_d); |
7e6eb623 DB |
558 | new_pair->e = e; |
559 | new_pair->pred = pred; | |
560 | new_pair->v = v; | |
5deac340 | 561 | new_pair->hashcode = iterative_hash_hashval_t (pre_expr_d::hash (e), |
c9145754 DB |
562 | pred->index); |
563 | ||
0823efed | 564 | slot = phi_translate_table.find_slot_with_hash (new_pair, |
7e6eb623 | 565 | new_pair->hashcode, INSERT); |
04695783 | 566 | free (*slot); |
0823efed | 567 | *slot = new_pair; |
6de9cd9a DN |
568 | } |
569 | ||
ff2ad0f7 | 570 | |
c9145754 | 571 | /* Add expression E to the expression set of value id V. */ |
33c94679 | 572 | |
6bcfb753 | 573 | static void |
c9145754 | 574 | add_to_value (unsigned int v, pre_expr e) |
7e6eb623 | 575 | { |
4d4b1b30 | 576 | bitmap set; |
c9145754 | 577 | |
4d4b1b30 | 578 | gcc_checking_assert (get_expr_value_id (e) == v); |
de081cfd | 579 | |
9771b263 | 580 | if (v >= value_expressions.length ()) |
c9145754 | 581 | { |
9771b263 | 582 | value_expressions.safe_grow_cleared (v + 1); |
c9145754 | 583 | } |
33c94679 | 584 | |
9771b263 | 585 | set = value_expressions[v]; |
c9145754 DB |
586 | if (!set) |
587 | { | |
4d4b1b30 | 588 | set = BITMAP_ALLOC (&grand_bitmap_obstack); |
9771b263 | 589 | value_expressions[v] = set; |
c9145754 | 590 | } |
33c94679 | 591 | |
4d4b1b30 | 592 | bitmap_set_bit (set, get_or_alloc_expression_id (e)); |
7e6eb623 | 593 | } |
6de9cd9a | 594 | |
bdee7684 DB |
595 | /* Create a new bitmap set and return it. */ |
596 | ||
b9c5e484 | 597 | static bitmap_set_t |
bdee7684 DB |
598 | bitmap_set_new (void) |
599 | { | |
e1111e8e | 600 | bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool); |
5c72d561 JH |
601 | bitmap_initialize (&ret->expressions, &grand_bitmap_obstack); |
602 | bitmap_initialize (&ret->values, &grand_bitmap_obstack); | |
bdee7684 DB |
603 | return ret; |
604 | } | |
605 | ||
c9145754 DB |
606 | /* Return the value id for a PRE expression EXPR. */ |
607 | ||
608 | static unsigned int | |
609 | get_expr_value_id (pre_expr expr) | |
610 | { | |
e3815735 | 611 | unsigned int id; |
c9145754 DB |
612 | switch (expr->kind) |
613 | { | |
614 | case CONSTANT: | |
5ba5e8ec | 615 | id = get_constant_value_id (PRE_EXPR_CONSTANT (expr)); |
e3815735 | 616 | break; |
c9145754 | 617 | case NAME: |
e3815735 RB |
618 | id = VN_INFO (PRE_EXPR_NAME (expr))->value_id; |
619 | break; | |
c9145754 | 620 | case NARY: |
e3815735 RB |
621 | id = PRE_EXPR_NARY (expr)->value_id; |
622 | break; | |
c9145754 | 623 | case REFERENCE: |
e3815735 RB |
624 | id = PRE_EXPR_REFERENCE (expr)->value_id; |
625 | break; | |
c9145754 DB |
626 | default: |
627 | gcc_unreachable (); | |
628 | } | |
e3815735 RB |
629 | /* ??? We cannot assert that expr has a value-id (it can be 0), because |
630 | we assign value-ids only to expressions that have a result | |
631 | in set_hashtable_value_ids. */ | |
632 | return id; | |
c9145754 DB |
633 | } |
634 | ||
40b178f4 RG |
635 | /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */ |
636 | ||
637 | static tree | |
638 | sccvn_valnum_from_value_id (unsigned int val) | |
639 | { | |
640 | bitmap_iterator bi; | |
641 | unsigned int i; | |
9771b263 | 642 | bitmap exprset = value_expressions[val]; |
40b178f4 RG |
643 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
644 | { | |
645 | pre_expr vexpr = expression_for_id (i); | |
646 | if (vexpr->kind == NAME) | |
647 | return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum; | |
648 | else if (vexpr->kind == CONSTANT) | |
649 | return PRE_EXPR_CONSTANT (vexpr); | |
650 | } | |
651 | return NULL_TREE; | |
652 | } | |
653 | ||
83737db2 | 654 | /* Remove an expression EXPR from a bitmapped set. */ |
bdee7684 DB |
655 | |
656 | static void | |
c9145754 | 657 | bitmap_remove_from_set (bitmap_set_t set, pre_expr expr) |
bdee7684 | 658 | { |
c9145754 DB |
659 | unsigned int val = get_expr_value_id (expr); |
660 | if (!value_id_constant_p (val)) | |
83737db2 | 661 | { |
5c72d561 JH |
662 | bitmap_clear_bit (&set->values, val); |
663 | bitmap_clear_bit (&set->expressions, get_expression_id (expr)); | |
83737db2 | 664 | } |
bdee7684 | 665 | } |
6de9cd9a | 666 | |
7e6eb623 | 667 | static void |
c9145754 | 668 | bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr, |
9708c51d | 669 | unsigned int val, bool allow_constants) |
7e6eb623 | 670 | { |
c9145754 | 671 | if (allow_constants || !value_id_constant_p (val)) |
7e6eb623 | 672 | { |
c9145754 DB |
673 | /* We specifically expect this and only this function to be able to |
674 | insert constants into a set. */ | |
5c72d561 JH |
675 | bitmap_set_bit (&set->values, val); |
676 | bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr)); | |
7e6eb623 DB |
677 | } |
678 | } | |
6de9cd9a | 679 | |
c9145754 DB |
680 | /* Insert an expression EXPR into a bitmapped set. */ |
681 | ||
682 | static void | |
683 | bitmap_insert_into_set (bitmap_set_t set, pre_expr expr) | |
684 | { | |
9708c51d | 685 | bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false); |
c9145754 DB |
686 | } |
687 | ||
bdee7684 DB |
688 | /* Copy a bitmapped set ORIG, into bitmapped set DEST. */ |
689 | ||
690 | static void | |
691 | bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig) | |
692 | { | |
5c72d561 JH |
693 | bitmap_copy (&dest->expressions, &orig->expressions); |
694 | bitmap_copy (&dest->values, &orig->values); | |
bdee7684 DB |
695 | } |
696 | ||
ff3fdad2 | 697 | |
83737db2 | 698 | /* Free memory used up by SET. */ |
ff3fdad2 | 699 | static void |
83737db2 | 700 | bitmap_set_free (bitmap_set_t set) |
ff3fdad2 | 701 | { |
5c72d561 JH |
702 | bitmap_clear (&set->expressions); |
703 | bitmap_clear (&set->values); | |
ff3fdad2 DB |
704 | } |
705 | ||
ff3fdad2 | 706 | |
83737db2 | 707 | /* Generate an topological-ordered array of bitmap set SET. */ |
ff3fdad2 | 708 | |
9771b263 | 709 | static vec<pre_expr> |
83737db2 | 710 | sorted_array_from_bitmap_set (bitmap_set_t set) |
ff3fdad2 | 711 | { |
85169114 PB |
712 | unsigned int i, j; |
713 | bitmap_iterator bi, bj; | |
9771b263 | 714 | vec<pre_expr> result; |
a7d04a53 RG |
715 | |
716 | /* Pre-allocate roughly enough space for the array. */ | |
9771b263 | 717 | result.create (bitmap_count_bits (&set->values)); |
ff3fdad2 | 718 | |
85169114 PB |
719 | FOR_EACH_VALUE_ID_IN_SET (set, i, bi) |
720 | { | |
721 | /* The number of expressions having a given value is usually | |
722 | relatively small. Thus, rather than making a vector of all | |
723 | the expressions and sorting it by value-id, we walk the values | |
724 | and check in the reverse mapping that tells us what expressions | |
725 | have a given value, to filter those in our set. As a result, | |
726 | the expressions are inserted in value-id order, which means | |
727 | topological order. | |
728 | ||
729 | If this is somehow a significant lose for some cases, we can | |
730 | choose which set to walk based on the set size. */ | |
9771b263 | 731 | bitmap exprset = value_expressions[i]; |
4d4b1b30 | 732 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj) |
85169114 | 733 | { |
5c72d561 | 734 | if (bitmap_bit_p (&set->expressions, j)) |
9771b263 | 735 | result.safe_push (expression_for_id (j)); |
85169114 PB |
736 | } |
737 | } | |
6de9cd9a | 738 | |
83737db2 | 739 | return result; |
7e6eb623 | 740 | } |
6de9cd9a | 741 | |
83737db2 | 742 | /* Perform bitmapped set operation DEST &= ORIG. */ |
6de9cd9a DN |
743 | |
744 | static void | |
83737db2 | 745 | bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig) |
6de9cd9a | 746 | { |
83737db2 DB |
747 | bitmap_iterator bi; |
748 | unsigned int i; | |
6de9cd9a | 749 | |
d75dbccd | 750 | if (dest != orig) |
83737db2 | 751 | { |
5c72d561 JH |
752 | bitmap_head temp; |
753 | bitmap_initialize (&temp, &grand_bitmap_obstack); | |
89fb70a3 | 754 | |
5c72d561 JH |
755 | bitmap_and_into (&dest->values, &orig->values); |
756 | bitmap_copy (&temp, &dest->expressions); | |
757 | EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi) | |
d75dbccd | 758 | { |
c9145754 DB |
759 | pre_expr expr = expression_for_id (i); |
760 | unsigned int value_id = get_expr_value_id (expr); | |
5c72d561 JH |
761 | if (!bitmap_bit_p (&dest->values, value_id)) |
762 | bitmap_clear_bit (&dest->expressions, i); | |
d75dbccd | 763 | } |
5c72d561 | 764 | bitmap_clear (&temp); |
6de9cd9a DN |
765 | } |
766 | } | |
767 | ||
83737db2 | 768 | /* Subtract all values and expressions contained in ORIG from DEST. */ |
7e6eb623 | 769 | |
83737db2 DB |
770 | static bitmap_set_t |
771 | bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig) | |
7e6eb623 | 772 | { |
83737db2 DB |
773 | bitmap_set_t result = bitmap_set_new (); |
774 | bitmap_iterator bi; | |
775 | unsigned int i; | |
b9c5e484 | 776 | |
5c72d561 JH |
777 | bitmap_and_compl (&result->expressions, &dest->expressions, |
778 | &orig->expressions); | |
6de9cd9a | 779 | |
83737db2 DB |
780 | FOR_EACH_EXPR_ID_IN_SET (result, i, bi) |
781 | { | |
c9145754 DB |
782 | pre_expr expr = expression_for_id (i); |
783 | unsigned int value_id = get_expr_value_id (expr); | |
5c72d561 | 784 | bitmap_set_bit (&result->values, value_id); |
83737db2 | 785 | } |
af75a7ea | 786 | |
83737db2 | 787 | return result; |
6b416da1 DB |
788 | } |
789 | ||
d75dbccd DB |
790 | /* Subtract all the values in bitmap set B from bitmap set A. */ |
791 | ||
792 | static void | |
793 | bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b) | |
794 | { | |
795 | unsigned int i; | |
796 | bitmap_iterator bi; | |
5c72d561 | 797 | bitmap_head temp; |
d75dbccd | 798 | |
5c72d561 JH |
799 | bitmap_initialize (&temp, &grand_bitmap_obstack); |
800 | ||
801 | bitmap_copy (&temp, &a->expressions); | |
802 | EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi) | |
d75dbccd | 803 | { |
c9145754 DB |
804 | pre_expr expr = expression_for_id (i); |
805 | if (bitmap_set_contains_value (b, get_expr_value_id (expr))) | |
d75dbccd DB |
806 | bitmap_remove_from_set (a, expr); |
807 | } | |
5c72d561 | 808 | bitmap_clear (&temp); |
d75dbccd DB |
809 | } |
810 | ||
811 | ||
c9145754 | 812 | /* Return true if bitmapped set SET contains the value VALUE_ID. */ |
6de9cd9a | 813 | |
bdee7684 | 814 | static bool |
c9145754 | 815 | bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id) |
6de9cd9a | 816 | { |
c9145754 | 817 | if (value_id_constant_p (value_id)) |
bdee7684 | 818 | return true; |
83737db2 | 819 | |
5c72d561 | 820 | if (!set || bitmap_empty_p (&set->expressions)) |
83737db2 DB |
821 | return false; |
822 | ||
5c72d561 | 823 | return bitmap_bit_p (&set->values, value_id); |
bdee7684 | 824 | } |
7e6eb623 | 825 | |
d75dbccd | 826 | static inline bool |
c9145754 | 827 | bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr) |
d75dbccd | 828 | { |
5c72d561 | 829 | return bitmap_bit_p (&set->expressions, get_expression_id (expr)); |
d75dbccd DB |
830 | } |
831 | ||
bdee7684 | 832 | /* Replace an instance of value LOOKFOR with expression EXPR in SET. */ |
7e6eb623 | 833 | |
bdee7684 | 834 | static void |
c9145754 DB |
835 | bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor, |
836 | const pre_expr expr) | |
bdee7684 | 837 | { |
4d4b1b30 | 838 | bitmap exprset; |
83737db2 DB |
839 | unsigned int i; |
840 | bitmap_iterator bi; | |
841 | ||
c9145754 | 842 | if (value_id_constant_p (lookfor)) |
bdee7684 | 843 | return; |
83737db2 | 844 | |
bdee7684 DB |
845 | if (!bitmap_set_contains_value (set, lookfor)) |
846 | return; | |
e9284566 | 847 | |
6b416da1 DB |
848 | /* The number of expressions having a given value is usually |
849 | significantly less than the total number of expressions in SET. | |
850 | Thus, rather than check, for each expression in SET, whether it | |
851 | has the value LOOKFOR, we walk the reverse mapping that tells us | |
852 | what expressions have a given value, and see if any of those | |
853 | expressions are in our set. For large testcases, this is about | |
854 | 5-10x faster than walking the bitmap. If this is somehow a | |
855 | significant lose for some cases, we can choose which set to walk | |
856 | based on the set size. */ | |
9771b263 | 857 | exprset = value_expressions[lookfor]; |
4d4b1b30 | 858 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
6de9cd9a | 859 | { |
fcaa4ca4 | 860 | if (bitmap_clear_bit (&set->expressions, i)) |
6de9cd9a | 861 | { |
5c72d561 | 862 | bitmap_set_bit (&set->expressions, get_expression_id (expr)); |
83737db2 | 863 | return; |
6de9cd9a | 864 | } |
6de9cd9a | 865 | } |
4bcc5786 RB |
866 | |
867 | gcc_unreachable (); | |
6de9cd9a DN |
868 | } |
869 | ||
83737db2 | 870 | /* Return true if two bitmap sets are equal. */ |
6de9cd9a | 871 | |
7e6eb623 | 872 | static bool |
83737db2 | 873 | bitmap_set_equal (bitmap_set_t a, bitmap_set_t b) |
6de9cd9a | 874 | { |
5c72d561 | 875 | return bitmap_equal_p (&a->values, &b->values); |
6de9cd9a DN |
876 | } |
877 | ||
e9284566 | 878 | /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists, |
6c6cfbfd | 879 | and add it otherwise. */ |
bdee7684 | 880 | |
7e6eb623 | 881 | static void |
c9145754 | 882 | bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr) |
7e6eb623 | 883 | { |
c9145754 | 884 | unsigned int val = get_expr_value_id (expr); |
83737db2 | 885 | |
e9284566 DB |
886 | if (bitmap_set_contains_value (set, val)) |
887 | bitmap_set_replace_value (set, val, expr); | |
888 | else | |
889 | bitmap_insert_into_set (set, expr); | |
bdee7684 | 890 | } |
7e6eb623 | 891 | |
bdee7684 DB |
892 | /* Insert EXPR into SET if EXPR's value is not already present in |
893 | SET. */ | |
894 | ||
895 | static void | |
c9145754 | 896 | bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr) |
bdee7684 | 897 | { |
c9145754 | 898 | unsigned int val = get_expr_value_id (expr); |
af75a7ea | 899 | |
77a74ed7 | 900 | gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr)); |
1befacc8 RG |
901 | |
902 | /* Constant values are always considered to be part of the set. */ | |
903 | if (value_id_constant_p (val)) | |
904 | return; | |
905 | ||
906 | /* If the value membership changed, add the expression. */ | |
5c72d561 JH |
907 | if (bitmap_set_bit (&set->values, val)) |
908 | bitmap_set_bit (&set->expressions, expr->id); | |
7e6eb623 | 909 | } |
6de9cd9a | 910 | |
c9145754 DB |
911 | /* Print out EXPR to outfile. */ |
912 | ||
913 | static void | |
914 | print_pre_expr (FILE *outfile, const pre_expr expr) | |
915 | { | |
916 | switch (expr->kind) | |
917 | { | |
918 | case CONSTANT: | |
919 | print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0); | |
920 | break; | |
921 | case NAME: | |
922 | print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0); | |
923 | break; | |
924 | case NARY: | |
925 | { | |
926 | unsigned int i; | |
927 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); | |
928 | fprintf (outfile, "{%s,", tree_code_name [nary->opcode]); | |
929 | for (i = 0; i < nary->length; i++) | |
930 | { | |
931 | print_generic_expr (outfile, nary->op[i], 0); | |
932 | if (i != (unsigned) nary->length - 1) | |
933 | fprintf (outfile, ","); | |
934 | } | |
935 | fprintf (outfile, "}"); | |
936 | } | |
937 | break; | |
938 | ||
939 | case REFERENCE: | |
940 | { | |
941 | vn_reference_op_t vro; | |
942 | unsigned int i; | |
943 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); | |
944 | fprintf (outfile, "{"); | |
945 | for (i = 0; | |
9771b263 | 946 | ref->operands.iterate (i, &vro); |
c9145754 DB |
947 | i++) |
948 | { | |
5006671f | 949 | bool closebrace = false; |
c9145754 DB |
950 | if (vro->opcode != SSA_NAME |
951 | && TREE_CODE_CLASS (vro->opcode) != tcc_declaration) | |
5006671f RG |
952 | { |
953 | fprintf (outfile, "%s", tree_code_name [vro->opcode]); | |
954 | if (vro->op0) | |
955 | { | |
956 | fprintf (outfile, "<"); | |
957 | closebrace = true; | |
958 | } | |
959 | } | |
c9145754 DB |
960 | if (vro->op0) |
961 | { | |
c9145754 DB |
962 | print_generic_expr (outfile, vro->op0, 0); |
963 | if (vro->op1) | |
964 | { | |
965 | fprintf (outfile, ","); | |
966 | print_generic_expr (outfile, vro->op1, 0); | |
967 | } | |
5006671f RG |
968 | if (vro->op2) |
969 | { | |
970 | fprintf (outfile, ","); | |
971 | print_generic_expr (outfile, vro->op2, 0); | |
972 | } | |
c9145754 | 973 | } |
5006671f RG |
974 | if (closebrace) |
975 | fprintf (outfile, ">"); | |
9771b263 | 976 | if (i != ref->operands.length () - 1) |
c9145754 DB |
977 | fprintf (outfile, ","); |
978 | } | |
979 | fprintf (outfile, "}"); | |
5006671f RG |
980 | if (ref->vuse) |
981 | { | |
982 | fprintf (outfile, "@"); | |
983 | print_generic_expr (outfile, ref->vuse, 0); | |
984 | } | |
c9145754 DB |
985 | } |
986 | break; | |
987 | } | |
988 | } | |
989 | void debug_pre_expr (pre_expr); | |
990 | ||
991 | /* Like print_pre_expr but always prints to stderr. */ | |
24e47c76 | 992 | DEBUG_FUNCTION void |
c9145754 DB |
993 | debug_pre_expr (pre_expr e) |
994 | { | |
995 | print_pre_expr (stderr, e); | |
996 | fprintf (stderr, "\n"); | |
997 | } | |
998 | ||
bdee7684 DB |
999 | /* Print out SET to OUTFILE. */ |
1000 | ||
1001 | static void | |
83737db2 DB |
1002 | print_bitmap_set (FILE *outfile, bitmap_set_t set, |
1003 | const char *setname, int blockindex) | |
bdee7684 DB |
1004 | { |
1005 | fprintf (outfile, "%s[%d] := { ", setname, blockindex); | |
1006 | if (set) | |
1007 | { | |
cf6b9ef1 | 1008 | bool first = true; |
3cd8c58a | 1009 | unsigned i; |
87c476a2 ZD |
1010 | bitmap_iterator bi; |
1011 | ||
83737db2 | 1012 | FOR_EACH_EXPR_ID_IN_SET (set, i, bi) |
87c476a2 | 1013 | { |
c9145754 | 1014 | const pre_expr expr = expression_for_id (i); |
83737db2 | 1015 | |
65a6f342 NS |
1016 | if (!first) |
1017 | fprintf (outfile, ", "); | |
1018 | first = false; | |
c9145754 | 1019 | print_pre_expr (outfile, expr); |
b9c5e484 | 1020 | |
c9145754 | 1021 | fprintf (outfile, " (%04d)", get_expr_value_id (expr)); |
87c476a2 | 1022 | } |
bdee7684 DB |
1023 | } |
1024 | fprintf (outfile, " }\n"); | |
1025 | } | |
6de9cd9a | 1026 | |
83737db2 | 1027 | void debug_bitmap_set (bitmap_set_t); |
6de9cd9a | 1028 | |
24e47c76 | 1029 | DEBUG_FUNCTION void |
83737db2 DB |
1030 | debug_bitmap_set (bitmap_set_t set) |
1031 | { | |
1032 | print_bitmap_set (stderr, set, "debug", 0); | |
7e6eb623 DB |
1033 | } |
1034 | ||
19372838 RG |
1035 | void debug_bitmap_sets_for (basic_block); |
1036 | ||
1037 | DEBUG_FUNCTION void | |
1038 | debug_bitmap_sets_for (basic_block bb) | |
1039 | { | |
1040 | print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index); | |
40b178f4 RG |
1041 | print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index); |
1042 | print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index); | |
1043 | print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index); | |
1044 | print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index); | |
1045 | if (do_partial_partial) | |
1046 | print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index); | |
1047 | print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index); | |
19372838 RG |
1048 | } |
1049 | ||
7e6eb623 | 1050 | /* Print out the expressions that have VAL to OUTFILE. */ |
33c94679 | 1051 | |
6bcfb753 | 1052 | static void |
c9145754 | 1053 | print_value_expressions (FILE *outfile, unsigned int val) |
7e6eb623 | 1054 | { |
9771b263 | 1055 | bitmap set = value_expressions[val]; |
c9145754 | 1056 | if (set) |
33c94679 | 1057 | { |
4d4b1b30 | 1058 | bitmap_set x; |
33c94679 | 1059 | char s[10]; |
c9145754 | 1060 | sprintf (s, "%04d", val); |
4d4b1b30 RG |
1061 | x.expressions = *set; |
1062 | print_bitmap_set (outfile, &x, s, 0); | |
33c94679 | 1063 | } |
6de9cd9a DN |
1064 | } |
1065 | ||
6de9cd9a | 1066 | |
24e47c76 | 1067 | DEBUG_FUNCTION void |
c9145754 | 1068 | debug_value_expressions (unsigned int val) |
6de9cd9a | 1069 | { |
7e6eb623 DB |
1070 | print_value_expressions (stderr, val); |
1071 | } | |
1072 | ||
c9145754 DB |
1073 | /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to |
1074 | represent it. */ | |
0995a441 | 1075 | |
c9145754 DB |
1076 | static pre_expr |
1077 | get_or_alloc_expr_for_constant (tree constant) | |
b9c5e484 | 1078 | { |
c9145754 DB |
1079 | unsigned int result_id; |
1080 | unsigned int value_id; | |
5f5126d6 RG |
1081 | struct pre_expr_d expr; |
1082 | pre_expr newexpr; | |
1083 | ||
1084 | expr.kind = CONSTANT; | |
1085 | PRE_EXPR_CONSTANT (&expr) = constant; | |
1086 | result_id = lookup_expression_id (&expr); | |
1087 | if (result_id != 0) | |
1088 | return expression_for_id (result_id); | |
1089 | ||
1090 | newexpr = (pre_expr) pool_alloc (pre_expr_pool); | |
c9145754 DB |
1091 | newexpr->kind = CONSTANT; |
1092 | PRE_EXPR_CONSTANT (newexpr) = constant; | |
5f5126d6 | 1093 | alloc_expression_id (newexpr); |
c9145754 | 1094 | value_id = get_or_alloc_constant_value_id (constant); |
c9145754 DB |
1095 | add_to_value (value_id, newexpr); |
1096 | return newexpr; | |
0995a441 SB |
1097 | } |
1098 | ||
c9145754 DB |
1099 | /* Given a value id V, find the actual tree representing the constant |
1100 | value if there is one, and return it. Return NULL if we can't find | |
1101 | a constant. */ | |
43da81be DB |
1102 | |
1103 | static tree | |
726a989a | 1104 | get_constant_for_value_id (unsigned int v) |
43da81be | 1105 | { |
c9145754 DB |
1106 | if (value_id_constant_p (v)) |
1107 | { | |
1108 | unsigned int i; | |
1109 | bitmap_iterator bi; | |
9771b263 | 1110 | bitmap exprset = value_expressions[v]; |
c9145754 | 1111 | |
4d4b1b30 | 1112 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
c9145754 DB |
1113 | { |
1114 | pre_expr expr = expression_for_id (i); | |
726a989a | 1115 | if (expr->kind == CONSTANT) |
c9145754 DB |
1116 | return PRE_EXPR_CONSTANT (expr); |
1117 | } | |
1118 | } | |
1119 | return NULL; | |
1120 | } | |
1121 | ||
1122 | /* Get or allocate a pre_expr for a piece of GIMPLE, and return it. | |
1123 | Currently only supports constants and SSA_NAMES. */ | |
1124 | static pre_expr | |
1125 | get_or_alloc_expr_for (tree t) | |
1126 | { | |
1127 | if (TREE_CODE (t) == SSA_NAME) | |
1128 | return get_or_alloc_expr_for_name (t); | |
1d65f45c | 1129 | else if (is_gimple_min_invariant (t)) |
c9145754 | 1130 | return get_or_alloc_expr_for_constant (t); |
726a989a RB |
1131 | else |
1132 | { | |
1133 | /* More complex expressions can result from SCCVN expression | |
1134 | simplification that inserts values for them. As they all | |
1135 | do not have VOPs the get handled by the nary ops struct. */ | |
1136 | vn_nary_op_t result; | |
1137 | unsigned int result_id; | |
1138 | vn_nary_op_lookup (t, &result); | |
1139 | if (result != NULL) | |
1140 | { | |
1141 | pre_expr e = (pre_expr) pool_alloc (pre_expr_pool); | |
1142 | e->kind = NARY; | |
1143 | PRE_EXPR_NARY (e) = result; | |
1144 | result_id = lookup_expression_id (e); | |
1145 | if (result_id != 0) | |
1146 | { | |
1147 | pool_free (pre_expr_pool, e); | |
1148 | e = expression_for_id (result_id); | |
1149 | return e; | |
1150 | } | |
1151 | alloc_expression_id (e); | |
1152 | return e; | |
1153 | } | |
1154 | } | |
c9145754 DB |
1155 | return NULL; |
1156 | } | |
1157 | ||
1158 | /* Return the folded version of T if T, when folded, is a gimple | |
1159 | min_invariant. Otherwise, return T. */ | |
1160 | ||
1161 | static pre_expr | |
1162 | fully_constant_expression (pre_expr e) | |
1163 | { | |
1164 | switch (e->kind) | |
1165 | { | |
1166 | case CONSTANT: | |
1167 | return e; | |
1168 | case NARY: | |
1169 | { | |
1170 | vn_nary_op_t nary = PRE_EXPR_NARY (e); | |
1171 | switch (TREE_CODE_CLASS (nary->opcode)) | |
1172 | { | |
1173 | case tcc_binary: | |
40f64141 | 1174 | case tcc_comparison: |
c9145754 DB |
1175 | { |
1176 | /* We have to go from trees to pre exprs to value ids to | |
1177 | constants. */ | |
1178 | tree naryop0 = nary->op[0]; | |
1179 | tree naryop1 = nary->op[1]; | |
40f64141 RG |
1180 | tree result; |
1181 | if (!is_gimple_min_invariant (naryop0)) | |
726a989a RB |
1182 | { |
1183 | pre_expr rep0 = get_or_alloc_expr_for (naryop0); | |
1184 | unsigned int vrep0 = get_expr_value_id (rep0); | |
40f64141 RG |
1185 | tree const0 = get_constant_for_value_id (vrep0); |
1186 | if (const0) | |
1187 | naryop0 = fold_convert (TREE_TYPE (naryop0), const0); | |
726a989a | 1188 | } |
40f64141 | 1189 | if (!is_gimple_min_invariant (naryop1)) |
726a989a RB |
1190 | { |
1191 | pre_expr rep1 = get_or_alloc_expr_for (naryop1); | |
1192 | unsigned int vrep1 = get_expr_value_id (rep1); | |
40f64141 RG |
1193 | tree const1 = get_constant_for_value_id (vrep1); |
1194 | if (const1) | |
1195 | naryop1 = fold_convert (TREE_TYPE (naryop1), const1); | |
b463e8de | 1196 | } |
40f64141 RG |
1197 | result = fold_binary (nary->opcode, nary->type, |
1198 | naryop0, naryop1); | |
c9145754 DB |
1199 | if (result && is_gimple_min_invariant (result)) |
1200 | return get_or_alloc_expr_for_constant (result); | |
40f64141 RG |
1201 | /* We might have simplified the expression to a |
1202 | SSA_NAME for example from x_1 * 1. But we cannot | |
1203 | insert a PHI for x_1 unconditionally as x_1 might | |
1204 | not be available readily. */ | |
c9145754 DB |
1205 | return e; |
1206 | } | |
40f64141 RG |
1207 | case tcc_reference: |
1208 | if (nary->opcode != REALPART_EXPR | |
b8698a0f | 1209 | && nary->opcode != IMAGPART_EXPR |
40f64141 RG |
1210 | && nary->opcode != VIEW_CONVERT_EXPR) |
1211 | return e; | |
1212 | /* Fallthrough. */ | |
c9145754 DB |
1213 | case tcc_unary: |
1214 | { | |
40f64141 RG |
1215 | /* We have to go from trees to pre exprs to value ids to |
1216 | constants. */ | |
c9145754 | 1217 | tree naryop0 = nary->op[0]; |
726a989a RB |
1218 | tree const0, result; |
1219 | if (is_gimple_min_invariant (naryop0)) | |
1220 | const0 = naryop0; | |
1221 | else | |
1222 | { | |
1223 | pre_expr rep0 = get_or_alloc_expr_for (naryop0); | |
1224 | unsigned int vrep0 = get_expr_value_id (rep0); | |
1225 | const0 = get_constant_for_value_id (vrep0); | |
1226 | } | |
1227 | result = NULL; | |
c9145754 | 1228 | if (const0) |
b463e8de DB |
1229 | { |
1230 | tree type1 = TREE_TYPE (nary->op[0]); | |
1231 | const0 = fold_convert (type1, const0); | |
1232 | result = fold_unary (nary->opcode, nary->type, const0); | |
1233 | } | |
c9145754 DB |
1234 | if (result && is_gimple_min_invariant (result)) |
1235 | return get_or_alloc_expr_for_constant (result); | |
1236 | return e; | |
1237 | } | |
1238 | default: | |
1239 | return e; | |
1240 | } | |
1241 | } | |
47af7a5c RG |
1242 | case REFERENCE: |
1243 | { | |
1244 | vn_reference_t ref = PRE_EXPR_REFERENCE (e); | |
12bd5a1e RG |
1245 | tree folded; |
1246 | if ((folded = fully_constant_vn_reference_p (ref))) | |
1247 | return get_or_alloc_expr_for_constant (folded); | |
1248 | return e; | |
1249 | } | |
c9145754 DB |
1250 | default: |
1251 | return e; | |
1252 | } | |
1253 | return e; | |
43da81be DB |
1254 | } |
1255 | ||
5006671f | 1256 | /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that |
84280917 MM |
1257 | it has the value it would have in BLOCK. Set *SAME_VALID to true |
1258 | in case the new vuse doesn't change the value id of the OPERANDS. */ | |
c90186eb | 1259 | |
5006671f | 1260 | static tree |
9771b263 | 1261 | translate_vuse_through_block (vec<vn_reference_op_s> operands, |
b45d2719 | 1262 | alias_set_type set, tree type, tree vuse, |
5006671f | 1263 | basic_block phiblock, |
84280917 | 1264 | basic_block block, bool *same_valid) |
c90186eb | 1265 | { |
5006671f | 1266 | gimple phi = SSA_NAME_DEF_STMT (vuse); |
b45d2719 | 1267 | ao_ref ref; |
84280917 MM |
1268 | edge e = NULL; |
1269 | bool use_oracle; | |
1270 | ||
1271 | *same_valid = true; | |
43da81be | 1272 | |
5006671f RG |
1273 | if (gimple_bb (phi) != phiblock) |
1274 | return vuse; | |
1275 | ||
84280917 | 1276 | use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands); |
5006671f RG |
1277 | |
1278 | /* Use the alias-oracle to find either the PHI node in this block, | |
1279 | the first VUSE used in this block that is equivalent to vuse or | |
1280 | the first VUSE which definition in this block kills the value. */ | |
84280917 MM |
1281 | if (gimple_code (phi) == GIMPLE_PHI) |
1282 | e = find_edge (block, phiblock); | |
1283 | else if (use_oracle) | |
1284 | while (!stmt_may_clobber_ref_p_1 (phi, &ref)) | |
1285 | { | |
1286 | vuse = gimple_vuse (phi); | |
1287 | phi = SSA_NAME_DEF_STMT (vuse); | |
1288 | if (gimple_bb (phi) != phiblock) | |
1289 | return vuse; | |
1290 | if (gimple_code (phi) == GIMPLE_PHI) | |
1291 | { | |
1292 | e = find_edge (block, phiblock); | |
1293 | break; | |
1294 | } | |
1295 | } | |
1296 | else | |
1297 | return NULL_TREE; | |
1298 | ||
1299 | if (e) | |
c90186eb | 1300 | { |
84280917 | 1301 | if (use_oracle) |
5006671f | 1302 | { |
84280917 | 1303 | bitmap visited = NULL; |
9bb06c2a | 1304 | unsigned int cnt; |
84280917 MM |
1305 | /* Try to find a vuse that dominates this phi node by skipping |
1306 | non-clobbering statements. */ | |
511c229c | 1307 | vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false); |
84280917 MM |
1308 | if (visited) |
1309 | BITMAP_FREE (visited); | |
5006671f | 1310 | } |
84280917 MM |
1311 | else |
1312 | vuse = NULL_TREE; | |
1313 | if (!vuse) | |
1314 | { | |
1315 | /* If we didn't find any, the value ID can't stay the same, | |
1316 | but return the translated vuse. */ | |
1317 | *same_valid = false; | |
1318 | vuse = PHI_ARG_DEF (phi, e->dest_idx); | |
1319 | } | |
1320 | /* ??? We would like to return vuse here as this is the canonical | |
1321 | upmost vdef that this reference is associated with. But during | |
1322 | insertion of the references into the hash tables we only ever | |
1323 | directly insert with their direct gimple_vuse, hence returning | |
1324 | something else would make us not find the other expression. */ | |
1325 | return PHI_ARG_DEF (phi, e->dest_idx); | |
c90186eb | 1326 | } |
c90186eb | 1327 | |
5006671f | 1328 | return NULL_TREE; |
c90186eb | 1329 | } |
83737db2 | 1330 | |
249eb506 | 1331 | /* Like bitmap_find_leader, but checks for the value existing in SET1 *or* |
83737db2 DB |
1332 | SET2. This is used to avoid making a set consisting of the union |
1333 | of PA_IN and ANTIC_IN during insert. */ | |
1334 | ||
c9145754 DB |
1335 | static inline pre_expr |
1336 | find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2) | |
83737db2 | 1337 | { |
c9145754 | 1338 | pre_expr result; |
83737db2 | 1339 | |
e076271b | 1340 | result = bitmap_find_leader (set1, val); |
83737db2 | 1341 | if (!result && set2) |
e076271b | 1342 | result = bitmap_find_leader (set2, val); |
83737db2 DB |
1343 | return result; |
1344 | } | |
1345 | ||
c9145754 DB |
1346 | /* Get the tree type for our PRE expression e. */ |
1347 | ||
1348 | static tree | |
1349 | get_expr_type (const pre_expr e) | |
1350 | { | |
1351 | switch (e->kind) | |
1352 | { | |
1353 | case NAME: | |
1354 | return TREE_TYPE (PRE_EXPR_NAME (e)); | |
1355 | case CONSTANT: | |
1356 | return TREE_TYPE (PRE_EXPR_CONSTANT (e)); | |
1357 | case REFERENCE: | |
b45d2719 | 1358 | return PRE_EXPR_REFERENCE (e)->type; |
c9145754 DB |
1359 | case NARY: |
1360 | return PRE_EXPR_NARY (e)->type; | |
1361 | } | |
1362 | gcc_unreachable(); | |
1363 | } | |
1364 | ||
1365 | /* Get a representative SSA_NAME for a given expression. | |
1366 | Since all of our sub-expressions are treated as values, we require | |
1367 | them to be SSA_NAME's for simplicity. | |
1368 | Prior versions of GVNPRE used to use "value handles" here, so that | |
1369 | an expression would be VH.11 + VH.10 instead of d_3 + e_6. In | |
1370 | either case, the operands are really values (IE we do not expect | |
1371 | them to be usable without finding leaders). */ | |
1372 | ||
1373 | static tree | |
1374 | get_representative_for (const pre_expr e) | |
1375 | { | |
c9145754 DB |
1376 | tree name; |
1377 | unsigned int value_id = get_expr_value_id (e); | |
1378 | ||
1379 | switch (e->kind) | |
1380 | { | |
1381 | case NAME: | |
1382 | return PRE_EXPR_NAME (e); | |
1383 | case CONSTANT: | |
726a989a | 1384 | return PRE_EXPR_CONSTANT (e); |
c9145754 DB |
1385 | case NARY: |
1386 | case REFERENCE: | |
1387 | { | |
1388 | /* Go through all of the expressions representing this value | |
1389 | and pick out an SSA_NAME. */ | |
1390 | unsigned int i; | |
1391 | bitmap_iterator bi; | |
9771b263 | 1392 | bitmap exprs = value_expressions[value_id]; |
4d4b1b30 | 1393 | EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi) |
c9145754 DB |
1394 | { |
1395 | pre_expr rep = expression_for_id (i); | |
1396 | if (rep->kind == NAME) | |
1397 | return PRE_EXPR_NAME (rep); | |
a044f0e7 RB |
1398 | else if (rep->kind == CONSTANT) |
1399 | return PRE_EXPR_CONSTANT (rep); | |
c9145754 DB |
1400 | } |
1401 | } | |
1402 | break; | |
1403 | } | |
a044f0e7 | 1404 | |
c9145754 DB |
1405 | /* If we reached here we couldn't find an SSA_NAME. This can |
1406 | happen when we've discovered a value that has never appeared in | |
a044f0e7 RB |
1407 | the program as set to an SSA_NAME, as the result of phi translation. |
1408 | Create one here. | |
1409 | ??? We should be able to re-use this when we insert the statement | |
1410 | to compute it. */ | |
83d5977e | 1411 | name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp"); |
c9145754 | 1412 | VN_INFO_GET (name)->value_id = value_id; |
a044f0e7 RB |
1413 | VN_INFO (name)->valnum = name; |
1414 | /* ??? For now mark this SSA name for release by SCCVN. */ | |
1415 | VN_INFO (name)->needs_insertion = true; | |
c9145754 | 1416 | add_to_value (value_id, get_or_alloc_expr_for_name (name)); |
a044f0e7 | 1417 | if (dump_file && (dump_flags & TDF_DETAILS)) |
c9145754 DB |
1418 | { |
1419 | fprintf (dump_file, "Created SSA_NAME representative "); | |
1420 | print_generic_expr (dump_file, name, 0); | |
1421 | fprintf (dump_file, " for expression:"); | |
1422 | print_pre_expr (dump_file, e); | |
1423 | fprintf (dump_file, "\n"); | |
1424 | } | |
1425 | ||
1426 | return name; | |
1427 | } | |
1428 | ||
1429 | ||
1430 | ||
3e999e7b RG |
1431 | static pre_expr |
1432 | phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, | |
1433 | basic_block pred, basic_block phiblock); | |
c9145754 | 1434 | |
7e6eb623 | 1435 | /* Translate EXPR using phis in PHIBLOCK, so that it has the values of |
788d04b2 RG |
1436 | the phis in PRED. Return NULL if we can't find a leader for each part |
1437 | of the translated expression. */ | |
6de9cd9a | 1438 | |
c9145754 | 1439 | static pre_expr |
3e999e7b RG |
1440 | phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, |
1441 | basic_block pred, basic_block phiblock) | |
6de9cd9a | 1442 | { |
c9145754 | 1443 | switch (expr->kind) |
6de9cd9a | 1444 | { |
c9145754 | 1445 | case NARY: |
43da81be | 1446 | { |
c9145754 DB |
1447 | unsigned int i; |
1448 | bool changed = false; | |
1449 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); | |
5a7d7f9c RG |
1450 | vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s, |
1451 | sizeof_vn_nary_op (nary->length)); | |
1452 | memcpy (newnary, nary, sizeof_vn_nary_op (nary->length)); | |
c9145754 | 1453 | |
5a7d7f9c | 1454 | for (i = 0; i < newnary->length; i++) |
43da81be | 1455 | { |
5a7d7f9c | 1456 | if (TREE_CODE (newnary->op[i]) != SSA_NAME) |
c9145754 DB |
1457 | continue; |
1458 | else | |
43da81be | 1459 | { |
af078bb0 | 1460 | pre_expr leader, result; |
5a7d7f9c | 1461 | unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id; |
af078bb0 | 1462 | leader = find_leader_in_sets (op_val_id, set1, set2); |
788d04b2 | 1463 | result = phi_translate (leader, set1, set2, pred, phiblock); |
c9145754 | 1464 | if (result && result != leader) |
43da81be | 1465 | { |
c9145754 DB |
1466 | tree name = get_representative_for (result); |
1467 | if (!name) | |
85300b46 | 1468 | return NULL; |
5a7d7f9c | 1469 | newnary->op[i] = name; |
43da81be | 1470 | } |
c9145754 DB |
1471 | else if (!result) |
1472 | return NULL; | |
0778d4e8 | 1473 | |
5a7d7f9c | 1474 | changed |= newnary->op[i] != nary->op[i]; |
0778d4e8 | 1475 | } |
c9145754 DB |
1476 | } |
1477 | if (changed) | |
1478 | { | |
1479 | pre_expr constant; | |
5f5126d6 | 1480 | unsigned int new_val_id; |
c9145754 | 1481 | |
5a7d7f9c RG |
1482 | tree result = vn_nary_op_lookup_pieces (newnary->length, |
1483 | newnary->opcode, | |
1484 | newnary->type, | |
1485 | &newnary->op[0], | |
c9145754 | 1486 | &nary); |
5f5126d6 RG |
1487 | if (result && is_gimple_min_invariant (result)) |
1488 | return get_or_alloc_expr_for_constant (result); | |
c9145754 DB |
1489 | |
1490 | expr = (pre_expr) pool_alloc (pre_expr_pool); | |
1491 | expr->kind = NARY; | |
1492 | expr->id = 0; | |
c9145754 DB |
1493 | if (nary) |
1494 | { | |
1495 | PRE_EXPR_NARY (expr) = nary; | |
1496 | constant = fully_constant_expression (expr); | |
1497 | if (constant != expr) | |
1498 | return constant; | |
0778d4e8 | 1499 | |
c9145754 DB |
1500 | new_val_id = nary->value_id; |
1501 | get_or_alloc_expression_id (expr); | |
1502 | } | |
1503 | else | |
43da81be | 1504 | { |
c9145754 | 1505 | new_val_id = get_next_value_id (); |
9771b263 | 1506 | value_expressions.safe_grow_cleared (get_max_value_id() + 1); |
5a7d7f9c RG |
1507 | nary = vn_nary_op_insert_pieces (newnary->length, |
1508 | newnary->opcode, | |
1509 | newnary->type, | |
1510 | &newnary->op[0], | |
c9145754 DB |
1511 | result, new_val_id); |
1512 | PRE_EXPR_NARY (expr) = nary; | |
1513 | constant = fully_constant_expression (expr); | |
1514 | if (constant != expr) | |
1515 | return constant; | |
1516 | get_or_alloc_expression_id (expr); | |
43da81be | 1517 | } |
b0a0ab2d | 1518 | add_to_value (new_val_id, expr); |
c5830edf | 1519 | } |
c9145754 | 1520 | return expr; |
c90186eb | 1521 | } |
c9145754 | 1522 | break; |
47af7a5c | 1523 | |
c9145754 | 1524 | case REFERENCE: |
c90186eb | 1525 | { |
c9145754 | 1526 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); |
9771b263 | 1527 | vec<vn_reference_op_s> operands = ref->operands; |
5006671f RG |
1528 | tree vuse = ref->vuse; |
1529 | tree newvuse = vuse; | |
6e1aa848 | 1530 | vec<vn_reference_op_s> newoperands = vNULL; |
84280917 | 1531 | bool changed = false, same_valid = true; |
4da80bfb | 1532 | unsigned int i, j, n; |
c9145754 DB |
1533 | vn_reference_op_t operand; |
1534 | vn_reference_t newref; | |
1535 | ||
aa7069aa | 1536 | for (i = 0, j = 0; |
9771b263 | 1537 | operands.iterate (i, &operand); i++, j++) |
e13f1c14 | 1538 | { |
c9145754 DB |
1539 | pre_expr opresult; |
1540 | pre_expr leader; | |
4da80bfb | 1541 | tree op[3]; |
c9145754 DB |
1542 | tree type = operand->type; |
1543 | vn_reference_op_s newop = *operand; | |
4da80bfb RG |
1544 | op[0] = operand->op0; |
1545 | op[1] = operand->op1; | |
1546 | op[2] = operand->op2; | |
1547 | for (n = 0; n < 3; ++n) | |
e13f1c14 | 1548 | { |
4da80bfb RG |
1549 | unsigned int op_val_id; |
1550 | if (!op[n]) | |
1551 | continue; | |
1552 | if (TREE_CODE (op[n]) != SSA_NAME) | |
c9145754 | 1553 | { |
4da80bfb RG |
1554 | /* We can't possibly insert these. */ |
1555 | if (n != 0 | |
1556 | && !is_gimple_min_invariant (op[n])) | |
b5d76df4 | 1557 | break; |
4da80bfb | 1558 | continue; |
c9145754 | 1559 | } |
4da80bfb | 1560 | op_val_id = VN_INFO (op[n])->value_id; |
c9145754 | 1561 | leader = find_leader_in_sets (op_val_id, set1, set2); |
4da80bfb RG |
1562 | if (!leader) |
1563 | break; | |
1564 | /* Make sure we do not recursively translate ourselves | |
1565 | like for translating a[n_1] with the leader for | |
1566 | n_1 being a[n_1]. */ | |
1567 | if (get_expression_id (leader) != get_expression_id (expr)) | |
c9145754 | 1568 | { |
4da80bfb RG |
1569 | opresult = phi_translate (leader, set1, set2, |
1570 | pred, phiblock); | |
1571 | if (!opresult) | |
b5d76df4 | 1572 | break; |
4da80bfb RG |
1573 | if (opresult != leader) |
1574 | { | |
1575 | tree name = get_representative_for (opresult); | |
1576 | if (!name) | |
1577 | break; | |
1578 | changed |= name != op[n]; | |
1579 | op[n] = name; | |
1580 | } | |
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 | { |
4da80bfb | 1605 | double_int off = tree_to_double_int (op[0]); |
27bcd47c LC |
1606 | off += -tree_to_double_int (op[1]); |
1607 | off *= tree_to_double_int (op[2]); | |
1608 | if (off.fits_shwi ()) | |
70f34814 RG |
1609 | newop.off = off.low; |
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 (); | |
9771b263 | 1698 | value_expressions.safe_grow_cleared(get_max_value_id() + 1); |
84280917 MM |
1699 | } |
1700 | else | |
1701 | new_val_id = ref->value_id; | |
b45d2719 RG |
1702 | newref = vn_reference_insert_pieces (newvuse, ref->set, |
1703 | ref->type, | |
c9145754 DB |
1704 | newoperands, |
1705 | result, new_val_id); | |
9771b263 | 1706 | newoperands.create (0); |
c9145754 | 1707 | PRE_EXPR_REFERENCE (expr) = newref; |
47af7a5c RG |
1708 | constant = fully_constant_expression (expr); |
1709 | if (constant != expr) | |
1710 | return constant; | |
c9145754 | 1711 | get_or_alloc_expression_id (expr); |
0995a441 | 1712 | } |
b0a0ab2d | 1713 | add_to_value (new_val_id, expr); |
7e6eb623 | 1714 | } |
9771b263 | 1715 | newoperands.release (); |
c9145754 | 1716 | return expr; |
7e6eb623 | 1717 | } |
c9145754 | 1718 | break; |
47af7a5c | 1719 | |
c9145754 | 1720 | case NAME: |
7e6eb623 | 1721 | { |
c9145754 | 1722 | tree name = PRE_EXPR_NAME (expr); |
e076271b RG |
1723 | gimple def_stmt = SSA_NAME_DEF_STMT (name); |
1724 | /* If the SSA name is defined by a PHI node in this block, | |
1725 | translate it. */ | |
726a989a RB |
1726 | if (gimple_code (def_stmt) == GIMPLE_PHI |
1727 | && gimple_bb (def_stmt) == phiblock) | |
9323afae | 1728 | { |
e076271b RG |
1729 | edge e = find_edge (pred, gimple_bb (def_stmt)); |
1730 | tree def = PHI_ARG_DEF (def_stmt, e->dest_idx); | |
73a63870 | 1731 | |
c9145754 | 1732 | /* Handle constant. */ |
89fb70a3 | 1733 | if (is_gimple_min_invariant (def)) |
c9145754 | 1734 | return get_or_alloc_expr_for_constant (def); |
070b797d | 1735 | |
e076271b | 1736 | return get_or_alloc_expr_for_name (def); |
9323afae | 1737 | } |
e076271b RG |
1738 | /* Otherwise return it unchanged - it will get cleaned if its |
1739 | value is not available in PREDs AVAIL_OUT set of expressions. */ | |
1740 | return expr; | |
7e6eb623 | 1741 | } |
6615c446 JO |
1742 | |
1743 | default: | |
1744 | gcc_unreachable (); | |
6de9cd9a DN |
1745 | } |
1746 | } | |
f8b04195 | 1747 | |
3e999e7b RG |
1748 | /* Wrapper around phi_translate_1 providing caching functionality. */ |
1749 | ||
1750 | static pre_expr | |
1751 | phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, | |
1752 | basic_block pred, basic_block phiblock) | |
1753 | { | |
1754 | pre_expr phitrans; | |
1755 | ||
1756 | if (!expr) | |
1757 | return NULL; | |
1758 | ||
1759 | /* Constants contain no values that need translation. */ | |
1760 | if (expr->kind == CONSTANT) | |
1761 | return expr; | |
1762 | ||
1763 | if (value_id_constant_p (get_expr_value_id (expr))) | |
1764 | return expr; | |
1765 | ||
1766 | if (expr->kind != NAME) | |
1767 | { | |
1768 | phitrans = phi_trans_lookup (expr, pred); | |
1769 | if (phitrans) | |
1770 | return phitrans; | |
1771 | } | |
1772 | ||
1773 | /* Translate. */ | |
1774 | phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock); | |
1775 | ||
1776 | /* Don't add empty translations to the cache. Neither add | |
1777 | translations of NAMEs as those are cheap to translate. */ | |
1778 | if (phitrans | |
1779 | && expr->kind != NAME) | |
1780 | phi_trans_add (expr, phitrans, pred); | |
1781 | ||
1782 | return phitrans; | |
1783 | } | |
1784 | ||
1785 | ||
c9145754 | 1786 | /* For each expression in SET, translate the values through phi nodes |
f5594471 DB |
1787 | in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting |
1788 | expressions in DEST. */ | |
1789 | ||
6de9cd9a | 1790 | static void |
83737db2 | 1791 | phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred, |
7e6eb623 | 1792 | basic_block phiblock) |
6de9cd9a | 1793 | { |
9771b263 | 1794 | vec<pre_expr> exprs; |
c9145754 | 1795 | pre_expr expr; |
83737db2 DB |
1796 | int i; |
1797 | ||
9adf0570 | 1798 | if (gimple_seq_empty_p (phi_nodes (phiblock))) |
6de9cd9a | 1799 | { |
83737db2 DB |
1800 | bitmap_set_copy (dest, set); |
1801 | return; | |
1802 | } | |
b9c5e484 | 1803 | |
83737db2 | 1804 | exprs = sorted_array_from_bitmap_set (set); |
9771b263 | 1805 | FOR_EACH_VEC_ELT (exprs, i, expr) |
83737db2 | 1806 | { |
c9145754 | 1807 | pre_expr translated; |
f8b04195 | 1808 | translated = phi_translate (expr, set, NULL, pred, phiblock); |
3ed7d068 RG |
1809 | if (!translated) |
1810 | continue; | |
c90186eb | 1811 | |
3ed7d068 RG |
1812 | /* We might end up with multiple expressions from SET being |
1813 | translated to the same value. In this case we do not want | |
1814 | to retain the NARY or REFERENCE expression but prefer a NAME | |
1815 | which would be the leader. */ | |
1816 | if (translated->kind == NAME) | |
1817 | bitmap_value_replace_in_set (dest, translated); | |
1818 | else | |
83737db2 | 1819 | bitmap_value_insert_into_set (dest, translated); |
b9c5e484 | 1820 | } |
9771b263 | 1821 | exprs.release (); |
6de9cd9a DN |
1822 | } |
1823 | ||
bdee7684 | 1824 | /* Find the leader for a value (i.e., the name representing that |
3d45dd59 RG |
1825 | value) in a given set, and return it. If STMT is non-NULL it |
1826 | makes sure the defining statement for the leader dominates it. | |
1827 | Return NULL if no leader is found. */ | |
bdee7684 | 1828 | |
c9145754 | 1829 | static pre_expr |
e076271b | 1830 | bitmap_find_leader (bitmap_set_t set, unsigned int val) |
bdee7684 | 1831 | { |
c9145754 DB |
1832 | if (value_id_constant_p (val)) |
1833 | { | |
1834 | unsigned int i; | |
1835 | bitmap_iterator bi; | |
9771b263 | 1836 | bitmap exprset = value_expressions[val]; |
83737db2 | 1837 | |
4d4b1b30 | 1838 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
c9145754 DB |
1839 | { |
1840 | pre_expr expr = expression_for_id (i); | |
1841 | if (expr->kind == CONSTANT) | |
1842 | return expr; | |
1843 | } | |
1844 | } | |
bdee7684 DB |
1845 | if (bitmap_set_contains_value (set, val)) |
1846 | { | |
6b416da1 DB |
1847 | /* Rather than walk the entire bitmap of expressions, and see |
1848 | whether any of them has the value we are looking for, we look | |
1849 | at the reverse mapping, which tells us the set of expressions | |
1850 | that have a given value (IE value->expressions with that | |
1851 | value) and see if any of those expressions are in our set. | |
1852 | The number of expressions per value is usually significantly | |
1853 | less than the number of expressions in the set. In fact, for | |
1854 | large testcases, doing it this way is roughly 5-10x faster | |
1855 | than walking the bitmap. | |
1856 | If this is somehow a significant lose for some cases, we can | |
b9c5e484 | 1857 | choose which set to walk based on which set is smaller. */ |
83737db2 DB |
1858 | unsigned int i; |
1859 | bitmap_iterator bi; | |
9771b263 | 1860 | bitmap exprset = value_expressions[val]; |
b9c5e484 | 1861 | |
4d4b1b30 | 1862 | EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi) |
e076271b | 1863 | return expression_for_id (i); |
6de9cd9a | 1864 | } |
7e6eb623 | 1865 | return NULL; |
6de9cd9a DN |
1866 | } |
1867 | ||
cea618ac | 1868 | /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of |
1e4816bc DB |
1869 | BLOCK by seeing if it is not killed in the block. Note that we are |
1870 | only determining whether there is a store that kills it. Because | |
1871 | of the order in which clean iterates over values, we are guaranteed | |
1872 | that altered operands will have caused us to be eliminated from the | |
1873 | ANTIC_IN set already. */ | |
c90186eb DB |
1874 | |
1875 | static bool | |
c9145754 | 1876 | value_dies_in_block_x (pre_expr expr, basic_block block) |
c90186eb | 1877 | { |
5006671f RG |
1878 | tree vuse = PRE_EXPR_REFERENCE (expr)->vuse; |
1879 | vn_reference_t refx = PRE_EXPR_REFERENCE (expr); | |
1880 | gimple def; | |
5006671f RG |
1881 | gimple_stmt_iterator gsi; |
1882 | unsigned id = get_expression_id (expr); | |
1883 | bool res = false; | |
b45d2719 | 1884 | ao_ref ref; |
89fb70a3 | 1885 | |
5006671f RG |
1886 | if (!vuse) |
1887 | return false; | |
1888 | ||
1889 | /* Lookup a previously calculated result. */ | |
1890 | if (EXPR_DIES (block) | |
1891 | && bitmap_bit_p (EXPR_DIES (block), id * 2)) | |
1892 | return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1); | |
1893 | ||
1894 | /* A memory expression {e, VUSE} dies in the block if there is a | |
1895 | statement that may clobber e. If, starting statement walk from the | |
1896 | top of the basic block, a statement uses VUSE there can be no kill | |
1897 | inbetween that use and the original statement that loaded {e, VUSE}, | |
1898 | so we can stop walking. */ | |
b45d2719 | 1899 | ref.base = NULL_TREE; |
5006671f | 1900 | for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi)) |
c90186eb | 1901 | { |
5006671f RG |
1902 | tree def_vuse, def_vdef; |
1903 | def = gsi_stmt (gsi); | |
1904 | def_vuse = gimple_vuse (def); | |
1905 | def_vdef = gimple_vdef (def); | |
89fb70a3 | 1906 | |
5006671f RG |
1907 | /* Not a memory statement. */ |
1908 | if (!def_vuse) | |
1e4816bc | 1909 | continue; |
5006671f RG |
1910 | |
1911 | /* Not a may-def. */ | |
1912 | if (!def_vdef) | |
1913 | { | |
1914 | /* A load with the same VUSE, we're done. */ | |
1915 | if (def_vuse == vuse) | |
1916 | break; | |
1917 | ||
1918 | continue; | |
1919 | } | |
1920 | ||
1921 | /* Init ref only if we really need it. */ | |
b45d2719 RG |
1922 | if (ref.base == NULL_TREE |
1923 | && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type, | |
1924 | refx->operands)) | |
5006671f | 1925 | { |
b45d2719 RG |
1926 | res = true; |
1927 | break; | |
5006671f RG |
1928 | } |
1929 | /* If the statement may clobber expr, it dies. */ | |
b45d2719 | 1930 | if (stmt_may_clobber_ref_p_1 (def, &ref)) |
5006671f RG |
1931 | { |
1932 | res = true; | |
1933 | break; | |
1934 | } | |
c90186eb | 1935 | } |
5006671f RG |
1936 | |
1937 | /* Remember the result. */ | |
1938 | if (!EXPR_DIES (block)) | |
1939 | EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack); | |
1940 | bitmap_set_bit (EXPR_DIES (block), id * 2); | |
1941 | if (res) | |
1942 | bitmap_set_bit (EXPR_DIES (block), id * 2 + 1); | |
1943 | ||
1944 | return res; | |
c90186eb DB |
1945 | } |
1946 | ||
6de9cd9a | 1947 | |
19372838 RG |
1948 | /* Determine if OP is valid in SET1 U SET2, which it is when the union |
1949 | contains its value-id. */ | |
83737db2 | 1950 | |
6de9cd9a | 1951 | static bool |
19372838 | 1952 | op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op) |
6de9cd9a | 1953 | { |
19372838 | 1954 | if (op && TREE_CODE (op) == SSA_NAME) |
c9145754 | 1955 | { |
19372838 | 1956 | unsigned int value_id = VN_INFO (op)->value_id; |
92290a18 RG |
1957 | if (!(bitmap_set_contains_value (set1, value_id) |
1958 | || (set2 && bitmap_set_contains_value (set2, value_id)))) | |
c9145754 DB |
1959 | return false; |
1960 | } | |
c9145754 DB |
1961 | return true; |
1962 | } | |
b9c5e484 | 1963 | |
c9145754 DB |
1964 | /* Determine if the expression EXPR is valid in SET1 U SET2. |
1965 | ONLY SET2 CAN BE NULL. | |
1966 | This means that we have a leader for each part of the expression | |
1967 | (if it consists of values), or the expression is an SSA_NAME. | |
7763473e | 1968 | For loads/calls, we also see if the vuse is killed in this block. */ |
5039610b | 1969 | |
c9145754 DB |
1970 | static bool |
1971 | valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr, | |
1972 | basic_block block) | |
1973 | { | |
1974 | switch (expr->kind) | |
1975 | { | |
1976 | case NAME: | |
e167c04d RB |
1977 | return bitmap_find_leader (AVAIL_OUT (block), |
1978 | get_expr_value_id (expr)) != NULL; | |
c9145754 | 1979 | case NARY: |
43da81be | 1980 | { |
c9145754 DB |
1981 | unsigned int i; |
1982 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); | |
1983 | for (i = 0; i < nary->length; i++) | |
19372838 RG |
1984 | if (!op_valid_in_sets (set1, set2, nary->op[i])) |
1985 | return false; | |
c9145754 | 1986 | return true; |
43da81be | 1987 | } |
c9145754 DB |
1988 | break; |
1989 | case REFERENCE: | |
c90186eb | 1990 | { |
c9145754 DB |
1991 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); |
1992 | vn_reference_op_t vro; | |
1993 | unsigned int i; | |
1994 | ||
9771b263 | 1995 | FOR_EACH_VEC_ELT (ref->operands, i, vro) |
c90186eb | 1996 | { |
19372838 RG |
1997 | if (!op_valid_in_sets (set1, set2, vro->op0) |
1998 | || !op_valid_in_sets (set1, set2, vro->op1) | |
1999 | || !op_valid_in_sets (set1, set2, vro->op2)) | |
e13f1c14 | 2000 | return false; |
c90186eb | 2001 | } |
cd32bb90 | 2002 | return true; |
c90186eb | 2003 | } |
6615c446 | 2004 | default: |
b9c5e484 | 2005 | gcc_unreachable (); |
c9145754 | 2006 | } |
6de9cd9a DN |
2007 | } |
2008 | ||
d75dbccd DB |
2009 | /* Clean the set of expressions that are no longer valid in SET1 or |
2010 | SET2. This means expressions that are made up of values we have no | |
2011 | leaders for in SET1 or SET2. This version is used for partial | |
2012 | anticipation, which means it is not valid in either ANTIC_IN or | |
2013 | PA_IN. */ | |
2014 | ||
2015 | static void | |
2016 | dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block) | |
2017 | { | |
9771b263 | 2018 | vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1); |
c9145754 | 2019 | pre_expr expr; |
d75dbccd DB |
2020 | int i; |
2021 | ||
9771b263 | 2022 | FOR_EACH_VEC_ELT (exprs, i, expr) |
d75dbccd DB |
2023 | { |
2024 | if (!valid_in_sets (set1, set2, expr, block)) | |
2025 | bitmap_remove_from_set (set1, expr); | |
2026 | } | |
9771b263 | 2027 | exprs.release (); |
d75dbccd DB |
2028 | } |
2029 | ||
ca072a31 DB |
2030 | /* Clean the set of expressions that are no longer valid in SET. This |
2031 | means expressions that are made up of values we have no leaders for | |
2032 | in SET. */ | |
6de9cd9a DN |
2033 | |
2034 | static void | |
83737db2 | 2035 | clean (bitmap_set_t set, basic_block block) |
6de9cd9a | 2036 | { |
9771b263 | 2037 | vec<pre_expr> exprs = sorted_array_from_bitmap_set (set); |
c9145754 | 2038 | pre_expr expr; |
83737db2 DB |
2039 | int i; |
2040 | ||
9771b263 | 2041 | FOR_EACH_VEC_ELT (exprs, i, expr) |
6de9cd9a | 2042 | { |
83737db2 DB |
2043 | if (!valid_in_sets (set, NULL, expr, block)) |
2044 | bitmap_remove_from_set (set, expr); | |
6de9cd9a | 2045 | } |
9771b263 | 2046 | exprs.release (); |
6de9cd9a DN |
2047 | } |
2048 | ||
cd32bb90 | 2049 | /* Clean the set of expressions that are no longer valid in SET because |
bea966c2 | 2050 | they are clobbered in BLOCK or because they trap and may not be executed. */ |
cd32bb90 RG |
2051 | |
2052 | static void | |
2053 | prune_clobbered_mems (bitmap_set_t set, basic_block block) | |
2054 | { | |
bea966c2 RG |
2055 | bitmap_iterator bi; |
2056 | unsigned i; | |
cd32bb90 | 2057 | |
bea966c2 | 2058 | FOR_EACH_EXPR_ID_IN_SET (set, i, bi) |
cd32bb90 | 2059 | { |
bea966c2 RG |
2060 | pre_expr expr = expression_for_id (i); |
2061 | if (expr->kind == REFERENCE) | |
2062 | { | |
2063 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); | |
2064 | if (ref->vuse) | |
2065 | { | |
2066 | gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse); | |
2067 | if (!gimple_nop_p (def_stmt) | |
2068 | && ((gimple_bb (def_stmt) != block | |
2069 | && !dominated_by_p (CDI_DOMINATORS, | |
2070 | block, gimple_bb (def_stmt))) | |
2071 | || (gimple_bb (def_stmt) == block | |
2072 | && value_dies_in_block_x (expr, block)))) | |
2073 | bitmap_remove_from_set (set, expr); | |
2074 | } | |
2075 | } | |
2076 | else if (expr->kind == NARY) | |
cd32bb90 | 2077 | { |
bea966c2 RG |
2078 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); |
2079 | /* If the NARY may trap make sure the block does not contain | |
2080 | a possible exit point. | |
2081 | ??? This is overly conservative if we translate AVAIL_OUT | |
2082 | as the available expression might be after the exit point. */ | |
2083 | if (BB_MAY_NOTRETURN (block) | |
2084 | && vn_nary_may_trap (nary)) | |
cd32bb90 RG |
2085 | bitmap_remove_from_set (set, expr); |
2086 | } | |
2087 | } | |
cd32bb90 RG |
2088 | } |
2089 | ||
d4222d43 | 2090 | static sbitmap has_abnormal_preds; |
89fb70a3 | 2091 | |
83737db2 DB |
2092 | /* List of blocks that may have changed during ANTIC computation and |
2093 | thus need to be iterated over. */ | |
2094 | ||
2095 | static sbitmap changed_blocks; | |
1e4816bc DB |
2096 | |
2097 | /* Decide whether to defer a block for a later iteration, or PHI | |
2098 | translate SOURCE to DEST using phis in PHIBLOCK. Return false if we | |
2099 | should defer the block, and true if we processed it. */ | |
2100 | ||
2101 | static bool | |
2102 | defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source, | |
2103 | basic_block block, basic_block phiblock) | |
2104 | { | |
2105 | if (!BB_VISITED (phiblock)) | |
2106 | { | |
d7c028c0 | 2107 | bitmap_set_bit (changed_blocks, block->index); |
1e4816bc DB |
2108 | BB_VISITED (block) = 0; |
2109 | BB_DEFERRED (block) = 1; | |
2110 | return false; | |
2111 | } | |
2112 | else | |
2113 | phi_translate_set (dest, source, block, phiblock); | |
2114 | return true; | |
2115 | } | |
2116 | ||
7e6eb623 | 2117 | /* Compute the ANTIC set for BLOCK. |
6de9cd9a | 2118 | |
665fcad8 SB |
2119 | If succs(BLOCK) > 1 then |
2120 | ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK) | |
2121 | else if succs(BLOCK) == 1 then | |
2122 | ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)]) | |
6de9cd9a | 2123 | |
665fcad8 | 2124 | ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK]) |
83737db2 | 2125 | */ |
6de9cd9a | 2126 | |
7e6eb623 | 2127 | static bool |
a28fee03 | 2128 | compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge) |
6de9cd9a | 2129 | { |
7e6eb623 | 2130 | bool changed = false; |
83737db2 DB |
2131 | bitmap_set_t S, old, ANTIC_OUT; |
2132 | bitmap_iterator bi; | |
2133 | unsigned int bii; | |
2134 | edge e; | |
2135 | edge_iterator ei; | |
a28fee03 | 2136 | |
83737db2 | 2137 | old = ANTIC_OUT = S = NULL; |
d75dbccd | 2138 | BB_VISITED (block) = 1; |
a28fee03 SB |
2139 | |
2140 | /* If any edges from predecessors are abnormal, antic_in is empty, | |
2141 | so do nothing. */ | |
2142 | if (block_has_abnormal_pred_edge) | |
2143 | goto maybe_dump_sets; | |
6de9cd9a | 2144 | |
83737db2 DB |
2145 | old = ANTIC_IN (block); |
2146 | ANTIC_OUT = bitmap_set_new (); | |
6de9cd9a | 2147 | |
a28fee03 SB |
2148 | /* If the block has no successors, ANTIC_OUT is empty. */ |
2149 | if (EDGE_COUNT (block->succs) == 0) | |
2150 | ; | |
7e6eb623 DB |
2151 | /* If we have one successor, we could have some phi nodes to |
2152 | translate through. */ | |
c5cbcccf | 2153 | else if (single_succ_p (block)) |
6de9cd9a | 2154 | { |
83737db2 | 2155 | basic_block succ_bb = single_succ (block); |
d75dbccd DB |
2156 | |
2157 | /* We trade iterations of the dataflow equations for having to | |
2158 | phi translate the maximal set, which is incredibly slow | |
2159 | (since the maximal set often has 300+ members, even when you | |
2160 | have a small number of blocks). | |
2161 | Basically, we defer the computation of ANTIC for this block | |
2f8e468b | 2162 | until we have processed it's successor, which will inevitably |
d75dbccd DB |
2163 | have a *much* smaller set of values to phi translate once |
2164 | clean has been run on it. | |
2165 | The cost of doing this is that we technically perform more | |
2166 | iterations, however, they are lower cost iterations. | |
2167 | ||
2168 | Timings for PRE on tramp3d-v4: | |
2169 | without maximal set fix: 11 seconds | |
2170 | with maximal set fix/without deferring: 26 seconds | |
2171 | with maximal set fix/with deferring: 11 seconds | |
2172 | */ | |
2173 | ||
1e4816bc DB |
2174 | if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb), |
2175 | block, succ_bb)) | |
d75dbccd DB |
2176 | { |
2177 | changed = true; | |
d75dbccd DB |
2178 | goto maybe_dump_sets; |
2179 | } | |
6de9cd9a | 2180 | } |
7e6eb623 | 2181 | /* If we have multiple successors, we take the intersection of all of |
1e4816bc DB |
2182 | them. Note that in the case of loop exit phi nodes, we may have |
2183 | phis to translate through. */ | |
7e6eb623 | 2184 | else |
6de9cd9a | 2185 | { |
9771b263 | 2186 | vec<basic_block> worklist; |
7e6eb623 | 2187 | size_t i; |
70a6b17e | 2188 | basic_block bprime, first = NULL; |
7e6eb623 | 2189 | |
9771b263 | 2190 | worklist.create (EDGE_COUNT (block->succs)); |
628f6a4e | 2191 | FOR_EACH_EDGE (e, ei, block->succs) |
1e4816bc | 2192 | { |
70a6b17e RG |
2193 | if (!first |
2194 | && BB_VISITED (e->dest)) | |
2195 | first = e->dest; | |
2196 | else if (BB_VISITED (e->dest)) | |
9771b263 | 2197 | worklist.quick_push (e->dest); |
1e4816bc | 2198 | } |
70a6b17e RG |
2199 | |
2200 | /* Of multiple successors we have to have visited one already. */ | |
2201 | if (!first) | |
1e4816bc | 2202 | { |
d7c028c0 | 2203 | bitmap_set_bit (changed_blocks, block->index); |
70a6b17e RG |
2204 | BB_VISITED (block) = 0; |
2205 | BB_DEFERRED (block) = 1; | |
2206 | changed = true; | |
9771b263 | 2207 | worklist.release (); |
70a6b17e | 2208 | goto maybe_dump_sets; |
1e4816bc | 2209 | } |
c90186eb | 2210 | |
9adf0570 | 2211 | if (!gimple_seq_empty_p (phi_nodes (first))) |
70a6b17e RG |
2212 | phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first); |
2213 | else | |
2214 | bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first)); | |
2215 | ||
9771b263 | 2216 | FOR_EACH_VEC_ELT (worklist, i, bprime) |
d75dbccd | 2217 | { |
9adf0570 | 2218 | if (!gimple_seq_empty_p (phi_nodes (bprime))) |
1e4816bc DB |
2219 | { |
2220 | bitmap_set_t tmp = bitmap_set_new (); | |
70a6b17e | 2221 | phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime); |
1e4816bc DB |
2222 | bitmap_set_and (ANTIC_OUT, tmp); |
2223 | bitmap_set_free (tmp); | |
2224 | } | |
89fb70a3 | 2225 | else |
70a6b17e | 2226 | bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime)); |
6de9cd9a | 2227 | } |
9771b263 | 2228 | worklist.release (); |
6de9cd9a | 2229 | } |
6de9cd9a | 2230 | |
cd32bb90 RG |
2231 | /* Prune expressions that are clobbered in block and thus become |
2232 | invalid if translated from ANTIC_OUT to ANTIC_IN. */ | |
2233 | prune_clobbered_mems (ANTIC_OUT, block); | |
2234 | ||
ea4b7848 | 2235 | /* Generate ANTIC_OUT - TMP_GEN. */ |
83737db2 | 2236 | S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block)); |
6de9cd9a | 2237 | |
d75dbccd | 2238 | /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */ |
83737db2 DB |
2239 | ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block), |
2240 | TMP_GEN (block)); | |
c33bae88 | 2241 | |
a28fee03 SB |
2242 | /* Then union in the ANTIC_OUT - TMP_GEN values, |
2243 | to get ANTIC_OUT U EXP_GEN - TMP_GEN */ | |
83737db2 DB |
2244 | FOR_EACH_EXPR_ID_IN_SET (S, bii, bi) |
2245 | bitmap_value_insert_into_set (ANTIC_IN (block), | |
2246 | expression_for_id (bii)); | |
6de9cd9a | 2247 | |
c90186eb | 2248 | clean (ANTIC_IN (block), block); |
d75dbccd | 2249 | |
5c72d561 | 2250 | if (!bitmap_set_equal (old, ANTIC_IN (block))) |
83737db2 DB |
2251 | { |
2252 | changed = true; | |
d7c028c0 | 2253 | bitmap_set_bit (changed_blocks, block->index); |
83737db2 | 2254 | FOR_EACH_EDGE (e, ei, block->preds) |
d7c028c0 | 2255 | bitmap_set_bit (changed_blocks, e->src->index); |
83737db2 DB |
2256 | } |
2257 | else | |
d7c028c0 | 2258 | bitmap_clear_bit (changed_blocks, block->index); |
6de9cd9a | 2259 | |
a28fee03 | 2260 | maybe_dump_sets: |
7e6eb623 DB |
2261 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2262 | { | |
d75dbccd DB |
2263 | if (!BB_DEFERRED (block) || BB_VISITED (block)) |
2264 | { | |
2265 | if (ANTIC_OUT) | |
2266 | print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index); | |
85300b46 | 2267 | |
d75dbccd DB |
2268 | print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN", |
2269 | block->index); | |
85300b46 | 2270 | |
d75dbccd DB |
2271 | if (S) |
2272 | print_bitmap_set (dump_file, S, "S", block->index); | |
2273 | } | |
2274 | else | |
2275 | { | |
2276 | fprintf (dump_file, | |
2277 | "Block %d was deferred for a future iteration.\n", | |
2278 | block->index); | |
2279 | } | |
83737db2 DB |
2280 | } |
2281 | if (old) | |
2282 | bitmap_set_free (old); | |
2283 | if (S) | |
2284 | bitmap_set_free (S); | |
2285 | if (ANTIC_OUT) | |
2286 | bitmap_set_free (ANTIC_OUT); | |
7e6eb623 | 2287 | return changed; |
6de9cd9a DN |
2288 | } |
2289 | ||
d75dbccd DB |
2290 | /* Compute PARTIAL_ANTIC for BLOCK. |
2291 | ||
2292 | If succs(BLOCK) > 1 then | |
2293 | PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not | |
2294 | in ANTIC_OUT for all succ(BLOCK) | |
2295 | else if succs(BLOCK) == 1 then | |
2296 | PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)]) | |
2297 | ||
2298 | PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK] | |
2299 | - ANTIC_IN[BLOCK]) | |
2300 | ||
2301 | */ | |
2302 | static bool | |
2303 | compute_partial_antic_aux (basic_block block, | |
2304 | bool block_has_abnormal_pred_edge) | |
2305 | { | |
2306 | bool changed = false; | |
2307 | bitmap_set_t old_PA_IN; | |
2308 | bitmap_set_t PA_OUT; | |
2309 | edge e; | |
2310 | edge_iterator ei; | |
f0ed4cfb | 2311 | unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH); |
d75dbccd DB |
2312 | |
2313 | old_PA_IN = PA_OUT = NULL; | |
2314 | ||
2315 | /* If any edges from predecessors are abnormal, antic_in is empty, | |
2316 | so do nothing. */ | |
2317 | if (block_has_abnormal_pred_edge) | |
2318 | goto maybe_dump_sets; | |
2319 | ||
f0ed4cfb NC |
2320 | /* If there are too many partially anticipatable values in the |
2321 | block, phi_translate_set can take an exponential time: stop | |
2322 | before the translation starts. */ | |
2323 | if (max_pa | |
2324 | && single_succ_p (block) | |
5c72d561 | 2325 | && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa) |
f0ed4cfb NC |
2326 | goto maybe_dump_sets; |
2327 | ||
d75dbccd DB |
2328 | old_PA_IN = PA_IN (block); |
2329 | PA_OUT = bitmap_set_new (); | |
2330 | ||
2331 | /* If the block has no successors, ANTIC_OUT is empty. */ | |
2332 | if (EDGE_COUNT (block->succs) == 0) | |
2333 | ; | |
2334 | /* If we have one successor, we could have some phi nodes to | |
2335 | translate through. Note that we can't phi translate across DFS | |
89fb70a3 DB |
2336 | back edges in partial antic, because it uses a union operation on |
2337 | the successors. For recurrences like IV's, we will end up | |
2338 | generating a new value in the set on each go around (i + 3 (VH.1) | |
2339 | VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */ | |
d75dbccd DB |
2340 | else if (single_succ_p (block)) |
2341 | { | |
2342 | basic_block succ = single_succ (block); | |
2343 | if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK)) | |
2344 | phi_translate_set (PA_OUT, PA_IN (succ), block, succ); | |
2345 | } | |
2346 | /* If we have multiple successors, we take the union of all of | |
2347 | them. */ | |
2348 | else | |
2349 | { | |
9771b263 | 2350 | vec<basic_block> worklist; |
d75dbccd DB |
2351 | size_t i; |
2352 | basic_block bprime; | |
2353 | ||
9771b263 | 2354 | worklist.create (EDGE_COUNT (block->succs)); |
d75dbccd DB |
2355 | FOR_EACH_EDGE (e, ei, block->succs) |
2356 | { | |
2357 | if (e->flags & EDGE_DFS_BACK) | |
2358 | continue; | |
9771b263 | 2359 | worklist.quick_push (e->dest); |
d75dbccd | 2360 | } |
9771b263 | 2361 | if (worklist.length () > 0) |
d75dbccd | 2362 | { |
9771b263 | 2363 | FOR_EACH_VEC_ELT (worklist, i, bprime) |
d75dbccd DB |
2364 | { |
2365 | unsigned int i; | |
2366 | bitmap_iterator bi; | |
2367 | ||
2368 | FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi) | |
2369 | bitmap_value_insert_into_set (PA_OUT, | |
2370 | expression_for_id (i)); | |
9adf0570 | 2371 | if (!gimple_seq_empty_p (phi_nodes (bprime))) |
1e4816bc DB |
2372 | { |
2373 | bitmap_set_t pa_in = bitmap_set_new (); | |
2374 | phi_translate_set (pa_in, PA_IN (bprime), block, bprime); | |
2375 | FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi) | |
2376 | bitmap_value_insert_into_set (PA_OUT, | |
2377 | expression_for_id (i)); | |
2378 | bitmap_set_free (pa_in); | |
2379 | } | |
2380 | else | |
2381 | FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi) | |
2382 | bitmap_value_insert_into_set (PA_OUT, | |
2383 | expression_for_id (i)); | |
d75dbccd DB |
2384 | } |
2385 | } | |
9771b263 | 2386 | worklist.release (); |
d75dbccd DB |
2387 | } |
2388 | ||
cd32bb90 RG |
2389 | /* Prune expressions that are clobbered in block and thus become |
2390 | invalid if translated from PA_OUT to PA_IN. */ | |
2391 | prune_clobbered_mems (PA_OUT, block); | |
2392 | ||
d75dbccd DB |
2393 | /* PA_IN starts with PA_OUT - TMP_GEN. |
2394 | Then we subtract things from ANTIC_IN. */ | |
2395 | PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block)); | |
2396 | ||
2397 | /* For partial antic, we want to put back in the phi results, since | |
2398 | we will properly avoid making them partially antic over backedges. */ | |
5c72d561 JH |
2399 | bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values); |
2400 | bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions); | |
d75dbccd DB |
2401 | |
2402 | /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */ | |
2403 | bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block)); | |
2404 | ||
2405 | dependent_clean (PA_IN (block), ANTIC_IN (block), block); | |
2406 | ||
2407 | if (!bitmap_set_equal (old_PA_IN, PA_IN (block))) | |
2408 | { | |
2409 | changed = true; | |
d7c028c0 | 2410 | bitmap_set_bit (changed_blocks, block->index); |
d75dbccd | 2411 | FOR_EACH_EDGE (e, ei, block->preds) |
d7c028c0 | 2412 | bitmap_set_bit (changed_blocks, e->src->index); |
d75dbccd DB |
2413 | } |
2414 | else | |
d7c028c0 | 2415 | bitmap_clear_bit (changed_blocks, block->index); |
d75dbccd DB |
2416 | |
2417 | maybe_dump_sets: | |
2418 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2419 | { | |
2420 | if (PA_OUT) | |
2421 | print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index); | |
2422 | ||
2423 | print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index); | |
2424 | } | |
2425 | if (old_PA_IN) | |
2426 | bitmap_set_free (old_PA_IN); | |
2427 | if (PA_OUT) | |
2428 | bitmap_set_free (PA_OUT); | |
2429 | return changed; | |
2430 | } | |
2431 | ||
83737db2 | 2432 | /* Compute ANTIC and partial ANTIC sets. */ |
6de9cd9a DN |
2433 | |
2434 | static void | |
7e6eb623 | 2435 | compute_antic (void) |
6de9cd9a | 2436 | { |
c33bae88 | 2437 | bool changed = true; |
7e6eb623 | 2438 | int num_iterations = 0; |
c33bae88 | 2439 | basic_block block; |
83737db2 | 2440 | int i; |
a28fee03 SB |
2441 | |
2442 | /* If any predecessor edges are abnormal, we punt, so antic_in is empty. | |
2443 | We pre-build the map of blocks with incoming abnormal edges here. */ | |
2444 | has_abnormal_preds = sbitmap_alloc (last_basic_block); | |
f61e445a | 2445 | bitmap_clear (has_abnormal_preds); |
83737db2 | 2446 | |
585d0dc4 | 2447 | FOR_ALL_BB (block) |
7e6eb623 | 2448 | { |
a28fee03 SB |
2449 | edge_iterator ei; |
2450 | edge e; | |
2451 | ||
2452 | FOR_EACH_EDGE (e, ei, block->preds) | |
83737db2 DB |
2453 | { |
2454 | e->flags &= ~EDGE_DFS_BACK; | |
2455 | if (e->flags & EDGE_ABNORMAL) | |
2456 | { | |
d7c028c0 | 2457 | bitmap_set_bit (has_abnormal_preds, block->index); |
83737db2 DB |
2458 | break; |
2459 | } | |
2460 | } | |
a28fee03 | 2461 | |
83737db2 | 2462 | BB_VISITED (block) = 0; |
d75dbccd | 2463 | BB_DEFERRED (block) = 0; |
a19eb9d2 | 2464 | |
a28fee03 | 2465 | /* While we are here, give empty ANTIC_IN sets to each block. */ |
83737db2 | 2466 | ANTIC_IN (block) = bitmap_set_new (); |
d75dbccd | 2467 | PA_IN (block) = bitmap_set_new (); |
7e6eb623 | 2468 | } |
83737db2 | 2469 | |
a28fee03 | 2470 | /* At the exit block we anticipate nothing. */ |
83737db2 | 2471 | BB_VISITED (EXIT_BLOCK_PTR) = 1; |
a28fee03 | 2472 | |
83737db2 | 2473 | changed_blocks = sbitmap_alloc (last_basic_block + 1); |
f61e445a | 2474 | bitmap_ones (changed_blocks); |
7e6eb623 DB |
2475 | while (changed) |
2476 | { | |
83737db2 DB |
2477 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2478 | fprintf (dump_file, "Starting iteration %d\n", num_iterations); | |
3bc27de7 RG |
2479 | /* ??? We need to clear our PHI translation cache here as the |
2480 | ANTIC sets shrink and we restrict valid translations to | |
2481 | those having operands with leaders in ANTIC. Same below | |
2482 | for PA ANTIC computation. */ | |
a28fee03 | 2483 | num_iterations++; |
c33bae88 | 2484 | changed = false; |
585d0dc4 | 2485 | for (i = postorder_num - 1; i >= 0; i--) |
83737db2 | 2486 | { |
d7c028c0 | 2487 | if (bitmap_bit_p (changed_blocks, postorder[i])) |
83737db2 DB |
2488 | { |
2489 | basic_block block = BASIC_BLOCK (postorder[i]); | |
2490 | changed |= compute_antic_aux (block, | |
d7c028c0 | 2491 | bitmap_bit_p (has_abnormal_preds, |
83737db2 DB |
2492 | block->index)); |
2493 | } | |
2494 | } | |
d75dbccd | 2495 | /* Theoretically possible, but *highly* unlikely. */ |
77a74ed7 | 2496 | gcc_checking_assert (num_iterations < 500); |
2e24fa83 | 2497 | } |
a28fee03 | 2498 | |
9fe0cb7d RG |
2499 | statistics_histogram_event (cfun, "compute_antic iterations", |
2500 | num_iterations); | |
83737db2 | 2501 | |
d75dbccd DB |
2502 | if (do_partial_partial) |
2503 | { | |
f61e445a | 2504 | bitmap_ones (changed_blocks); |
d75dbccd DB |
2505 | mark_dfs_back_edges (); |
2506 | num_iterations = 0; | |
2507 | changed = true; | |
2508 | while (changed) | |
2509 | { | |
2510 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2511 | fprintf (dump_file, "Starting iteration %d\n", num_iterations); | |
2512 | num_iterations++; | |
2513 | changed = false; | |
585d0dc4 | 2514 | for (i = postorder_num - 1 ; i >= 0; i--) |
d75dbccd | 2515 | { |
d7c028c0 | 2516 | if (bitmap_bit_p (changed_blocks, postorder[i])) |
d75dbccd DB |
2517 | { |
2518 | basic_block block = BASIC_BLOCK (postorder[i]); | |
2519 | changed | |
2520 | |= compute_partial_antic_aux (block, | |
d7c028c0 | 2521 | bitmap_bit_p (has_abnormal_preds, |
d75dbccd DB |
2522 | block->index)); |
2523 | } | |
2524 | } | |
2525 | /* Theoretically possible, but *highly* unlikely. */ | |
77a74ed7 | 2526 | gcc_checking_assert (num_iterations < 500); |
d75dbccd | 2527 | } |
9fe0cb7d RG |
2528 | statistics_histogram_event (cfun, "compute_partial_antic iterations", |
2529 | num_iterations); | |
d75dbccd | 2530 | } |
83737db2 DB |
2531 | sbitmap_free (has_abnormal_preds); |
2532 | sbitmap_free (changed_blocks); | |
6de9cd9a DN |
2533 | } |
2534 | ||
c90186eb DB |
2535 | |
2536 | /* Inserted expressions are placed onto this worklist, which is used | |
2537 | for performing quick dead code elimination of insertions we made | |
2538 | that didn't turn out to be necessary. */ | |
0ab555de | 2539 | static bitmap inserted_exprs; |
c90186eb | 2540 | |
ce94d354 | 2541 | /* The actual worker for create_component_ref_by_pieces. */ |
b9c5e484 | 2542 | |
85300b46 | 2543 | static tree |
ce94d354 | 2544 | create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref, |
e076271b | 2545 | unsigned int *operand, gimple_seq *stmts) |
85300b46 | 2546 | { |
9771b263 | 2547 | vn_reference_op_t currop = &ref->operands[*operand]; |
c9145754 | 2548 | tree genop; |
ce94d354 | 2549 | ++*operand; |
c9145754 | 2550 | switch (currop->opcode) |
85300b46 | 2551 | { |
c9145754 DB |
2552 | case CALL_EXPR: |
2553 | { | |
b3be2694 | 2554 | tree folded, sc = NULL_TREE; |
ce94d354 | 2555 | unsigned int nargs = 0; |
b3be2694 RG |
2556 | tree fn, *args; |
2557 | if (TREE_CODE (currop->op0) == FUNCTION_DECL) | |
2558 | fn = currop->op0; | |
2559 | else | |
e076271b | 2560 | fn = find_or_generate_expression (block, currop->op0, stmts); |
c3dd8dd7 RB |
2561 | if (!fn) |
2562 | return NULL_TREE; | |
b3be2694 | 2563 | if (currop->op1) |
c3dd8dd7 RB |
2564 | { |
2565 | sc = find_or_generate_expression (block, currop->op1, stmts); | |
2566 | if (!sc) | |
2567 | return NULL_TREE; | |
2568 | } | |
9771b263 DN |
2569 | args = XNEWVEC (tree, ref->operands.length () - 1); |
2570 | while (*operand < ref->operands.length ()) | |
c9145754 | 2571 | { |
ce94d354 | 2572 | args[nargs] = create_component_ref_by_pieces_1 (block, ref, |
e076271b | 2573 | operand, stmts); |
c3dd8dd7 RB |
2574 | if (!args[nargs]) |
2575 | return NULL_TREE; | |
ce94d354 | 2576 | nargs++; |
c9145754 | 2577 | } |
726a989a | 2578 | folded = build_call_array (currop->type, |
b3be2694 RG |
2579 | (TREE_CODE (fn) == FUNCTION_DECL |
2580 | ? build_fold_addr_expr (fn) : fn), | |
726a989a | 2581 | nargs, args); |
c9145754 | 2582 | free (args); |
7aec7a38 | 2583 | if (sc) |
b3be2694 | 2584 | CALL_EXPR_STATIC_CHAIN (folded) = sc; |
c9145754 DB |
2585 | return folded; |
2586 | } | |
ea814c66 | 2587 | |
70f34814 RG |
2588 | case MEM_REF: |
2589 | { | |
2590 | tree baseop = create_component_ref_by_pieces_1 (block, ref, operand, | |
e076271b | 2591 | stmts); |
c3dd8dd7 RB |
2592 | if (!baseop) |
2593 | return NULL_TREE; | |
70f34814 | 2594 | tree offset = currop->op0; |
70f34814 RG |
2595 | if (TREE_CODE (baseop) == ADDR_EXPR |
2596 | && handled_component_p (TREE_OPERAND (baseop, 0))) | |
2597 | { | |
2598 | HOST_WIDE_INT off; | |
2599 | tree base; | |
2600 | base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0), | |
2601 | &off); | |
2602 | gcc_assert (base); | |
2603 | offset = int_const_binop (PLUS_EXPR, offset, | |
2604 | build_int_cst (TREE_TYPE (offset), | |
d35936ab | 2605 | off)); |
70f34814 RG |
2606 | baseop = build_fold_addr_expr (base); |
2607 | } | |
2608 | return fold_build2 (MEM_REF, currop->type, baseop, offset); | |
2609 | } | |
ea814c66 | 2610 | |
150e3929 RG |
2611 | case TARGET_MEM_REF: |
2612 | { | |
4d948885 | 2613 | tree genop0 = NULL_TREE, genop1 = NULL_TREE; |
9771b263 | 2614 | vn_reference_op_t nextop = &ref->operands[++*operand]; |
150e3929 | 2615 | tree baseop = create_component_ref_by_pieces_1 (block, ref, operand, |
e076271b | 2616 | stmts); |
c3dd8dd7 RB |
2617 | if (!baseop) |
2618 | return NULL_TREE; | |
150e3929 | 2619 | if (currop->op0) |
c3dd8dd7 RB |
2620 | { |
2621 | genop0 = find_or_generate_expression (block, currop->op0, stmts); | |
2622 | if (!genop0) | |
2623 | return NULL_TREE; | |
2624 | } | |
4d948885 | 2625 | if (nextop->op0) |
c3dd8dd7 RB |
2626 | { |
2627 | genop1 = find_or_generate_expression (block, nextop->op0, stmts); | |
2628 | if (!genop1) | |
2629 | return NULL_TREE; | |
2630 | } | |
4d948885 RG |
2631 | return build5 (TARGET_MEM_REF, currop->type, |
2632 | baseop, currop->op2, genop0, currop->op1, genop1); | |
150e3929 | 2633 | } |
ea814c66 | 2634 | |
ce94d354 RG |
2635 | case ADDR_EXPR: |
2636 | if (currop->op0) | |
2637 | { | |
2638 | gcc_assert (is_gimple_min_invariant (currop->op0)); | |
2639 | return currop->op0; | |
2640 | } | |
2641 | /* Fallthrough. */ | |
c9145754 DB |
2642 | case REALPART_EXPR: |
2643 | case IMAGPART_EXPR: | |
2644 | case VIEW_CONVERT_EXPR: | |
2645 | { | |
c3dd8dd7 RB |
2646 | tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand, |
2647 | stmts); | |
2648 | if (!genop0) | |
2649 | return NULL_TREE; | |
ea814c66 | 2650 | return fold_build1 (currop->opcode, currop->type, genop0); |
c9145754 | 2651 | } |
ea814c66 | 2652 | |
842679dc TV |
2653 | case WITH_SIZE_EXPR: |
2654 | { | |
2655 | tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand, | |
e076271b | 2656 | stmts); |
c3dd8dd7 RB |
2657 | if (!genop0) |
2658 | return NULL_TREE; | |
e076271b | 2659 | tree genop1 = find_or_generate_expression (block, currop->op0, stmts); |
c3dd8dd7 RB |
2660 | if (!genop1) |
2661 | return NULL_TREE; | |
842679dc TV |
2662 | return fold_build2 (currop->opcode, currop->type, genop0, genop1); |
2663 | } | |
ea814c66 | 2664 | |
c9145754 | 2665 | case BIT_FIELD_REF: |
e13f1c14 | 2666 | { |
ce94d354 | 2667 | tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand, |
e076271b | 2668 | stmts); |
c3dd8dd7 RB |
2669 | if (!genop0) |
2670 | return NULL_TREE; | |
ea814c66 EB |
2671 | tree op1 = currop->op0; |
2672 | tree op2 = currop->op1; | |
ea814c66 | 2673 | return fold_build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2); |
e13f1c14 | 2674 | } |
c9145754 DB |
2675 | |
2676 | /* For array ref vn_reference_op's, operand 1 of the array ref | |
2677 | is op0 of the reference op and operand 3 of the array ref is | |
2678 | op1. */ | |
2679 | case ARRAY_RANGE_REF: | |
2680 | case ARRAY_REF: | |
2681 | { | |
c9145754 DB |
2682 | tree genop0; |
2683 | tree genop1 = currop->op0; | |
c9145754 | 2684 | tree genop2 = currop->op1; |
e52201b6 | 2685 | tree genop3 = currop->op2; |
c3dd8dd7 RB |
2686 | genop0 = create_component_ref_by_pieces_1 (block, ref, operand, |
2687 | stmts); | |
2688 | if (!genop0) | |
2689 | return NULL_TREE; | |
e076271b | 2690 | genop1 = find_or_generate_expression (block, genop1, stmts); |
c3dd8dd7 RB |
2691 | if (!genop1) |
2692 | return NULL_TREE; | |
c9145754 DB |
2693 | if (genop2) |
2694 | { | |
d4d73ce2 EB |
2695 | tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0)); |
2696 | /* Drop zero minimum index if redundant. */ | |
2697 | if (integer_zerop (genop2) | |
2698 | && (!domain_type | |
2699 | || integer_zerop (TYPE_MIN_VALUE (domain_type)))) | |
d53bed0b RG |
2700 | genop2 = NULL_TREE; |
2701 | else | |
c3dd8dd7 RB |
2702 | { |
2703 | genop2 = find_or_generate_expression (block, genop2, stmts); | |
2704 | if (!genop2) | |
2705 | return NULL_TREE; | |
2706 | } | |
c9145754 | 2707 | } |
e52201b6 RG |
2708 | if (genop3) |
2709 | { | |
2710 | tree elmt_type = TREE_TYPE (TREE_TYPE (genop0)); | |
d53bed0b RG |
2711 | /* We can't always put a size in units of the element alignment |
2712 | here as the element alignment may be not visible. See | |
2713 | PR43783. Simply drop the element size for constant | |
2714 | sizes. */ | |
2715 | if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type))) | |
2716 | genop3 = NULL_TREE; | |
2717 | else | |
2718 | { | |
2719 | genop3 = size_binop (EXACT_DIV_EXPR, genop3, | |
2720 | size_int (TYPE_ALIGN_UNIT (elmt_type))); | |
e076271b | 2721 | genop3 = find_or_generate_expression (block, genop3, stmts); |
c3dd8dd7 RB |
2722 | if (!genop3) |
2723 | return NULL_TREE; | |
d53bed0b | 2724 | } |
e52201b6 | 2725 | } |
c9145754 DB |
2726 | return build4 (currop->opcode, currop->type, genop0, genop1, |
2727 | genop2, genop3); | |
2728 | } | |
85300b46 DB |
2729 | case COMPONENT_REF: |
2730 | { | |
2731 | tree op0; | |
2732 | tree op1; | |
c9145754 | 2733 | tree genop2 = currop->op1; |
e076271b | 2734 | op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts); |
c3dd8dd7 RB |
2735 | if (!op0) |
2736 | return NULL_TREE; | |
e076271b | 2737 | /* op1 should be a FIELD_DECL, which are represented by themselves. */ |
c9145754 DB |
2738 | op1 = currop->op0; |
2739 | if (genop2) | |
c3dd8dd7 RB |
2740 | { |
2741 | genop2 = find_or_generate_expression (block, genop2, stmts); | |
2742 | if (!genop2) | |
2743 | return NULL_TREE; | |
2744 | } | |
ea814c66 | 2745 | return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2); |
85300b46 | 2746 | } |
ea814c66 | 2747 | |
c9145754 | 2748 | case SSA_NAME: |
85300b46 | 2749 | { |
e076271b | 2750 | genop = find_or_generate_expression (block, currop->op0, stmts); |
c9145754 | 2751 | return genop; |
85300b46 | 2752 | } |
c9145754 DB |
2753 | case STRING_CST: |
2754 | case INTEGER_CST: | |
2755 | case COMPLEX_CST: | |
2756 | case VECTOR_CST: | |
2757 | case REAL_CST: | |
2758 | case CONSTRUCTOR: | |
85300b46 DB |
2759 | case VAR_DECL: |
2760 | case PARM_DECL: | |
c9145754 | 2761 | case CONST_DECL: |
5230d884 | 2762 | case RESULT_DECL: |
c9145754 | 2763 | case FUNCTION_DECL: |
c9145754 DB |
2764 | return currop->op0; |
2765 | ||
85300b46 | 2766 | default: |
b9c5e484 | 2767 | gcc_unreachable (); |
85300b46 | 2768 | } |
85300b46 | 2769 | } |
c90186eb | 2770 | |
ce94d354 | 2771 | /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the |
70f34814 | 2772 | COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with |
ce94d354 RG |
2773 | trying to rename aggregates into ssa form directly, which is a no no. |
2774 | ||
2775 | Thus, this routine doesn't create temporaries, it just builds a | |
2776 | single access expression for the array, calling | |
2777 | find_or_generate_expression to build the innermost pieces. | |
2778 | ||
2779 | This function is a subroutine of create_expression_by_pieces, and | |
2780 | should not be called on it's own unless you really know what you | |
2781 | are doing. */ | |
2782 | ||
2783 | static tree | |
2784 | create_component_ref_by_pieces (basic_block block, vn_reference_t ref, | |
e076271b | 2785 | gimple_seq *stmts) |
ce94d354 RG |
2786 | { |
2787 | unsigned int op = 0; | |
e076271b | 2788 | return create_component_ref_by_pieces_1 (block, ref, &op, stmts); |
ce94d354 RG |
2789 | } |
2790 | ||
c3dd8dd7 RB |
2791 | /* Find a simple leader for an expression, or generate one using |
2792 | create_expression_by_pieces from a NARY expression for the value. | |
56db793a | 2793 | BLOCK is the basic_block we are looking for leaders in. |
e076271b | 2794 | OP is the tree expression to find a leader for or generate. |
c3dd8dd7 | 2795 | Returns the leader or NULL_TREE on failure. */ |
56db793a DB |
2796 | |
2797 | static tree | |
e076271b | 2798 | find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts) |
56db793a | 2799 | { |
e076271b RG |
2800 | pre_expr expr = get_or_alloc_expr_for (op); |
2801 | unsigned int lookfor = get_expr_value_id (expr); | |
2802 | pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor); | |
c9145754 DB |
2803 | if (leader) |
2804 | { | |
2805 | if (leader->kind == NAME) | |
e076271b | 2806 | return PRE_EXPR_NAME (leader); |
c9145754 | 2807 | else if (leader->kind == CONSTANT) |
e076271b | 2808 | return PRE_EXPR_CONSTANT (leader); |
c3dd8dd7 RB |
2809 | |
2810 | /* Defer. */ | |
2811 | return NULL_TREE; | |
c9145754 | 2812 | } |
e9284566 | 2813 | |
c3dd8dd7 RB |
2814 | /* It must be a complex expression, so generate it recursively. Note |
2815 | that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c | |
2816 | where the insert algorithm fails to insert a required expression. */ | |
9771b263 | 2817 | bitmap exprset = value_expressions[lookfor]; |
e076271b RG |
2818 | bitmap_iterator bi; |
2819 | unsigned int i; | |
2820 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) | |
56db793a | 2821 | { |
e076271b | 2822 | pre_expr temp = expression_for_id (i); |
c3dd8dd7 RB |
2823 | /* We cannot insert random REFERENCE expressions at arbitrary |
2824 | places. We can insert NARYs which eventually re-materializes | |
2825 | its operand values. */ | |
2826 | if (temp->kind == NARY) | |
e076271b RG |
2827 | return create_expression_by_pieces (block, temp, stmts, |
2828 | get_expr_type (expr)); | |
56db793a | 2829 | } |
e076271b | 2830 | |
c3dd8dd7 RB |
2831 | /* Defer. */ |
2832 | return NULL_TREE; | |
56db793a DB |
2833 | } |
2834 | ||
726a989a | 2835 | #define NECESSARY GF_PLF_1 |
c9145754 | 2836 | |
56db793a | 2837 | /* Create an expression in pieces, so that we can handle very complex |
b9c5e484 | 2838 | expressions that may be ANTIC, but not necessary GIMPLE. |
56db793a DB |
2839 | BLOCK is the basic block the expression will be inserted into, |
2840 | EXPR is the expression to insert (in value form) | |
2841 | STMTS is a statement list to append the necessary insertions into. | |
2842 | ||
0e61db61 | 2843 | This function will die if we hit some value that shouldn't be |
56db793a | 2844 | ANTIC but is (IE there is no leader for it, or its components). |
c3dd8dd7 RB |
2845 | The function returns NULL_TREE in case a different antic expression |
2846 | has to be inserted first. | |
56db793a DB |
2847 | This function may also generate expressions that are themselves |
2848 | partially or fully redundant. Those that are will be either made | |
2849 | fully redundant during the next iteration of insert (for partially | |
2850 | redundant ones), or eliminated by eliminate (for fully redundant | |
c3dd8dd7 | 2851 | ones). */ |
56db793a DB |
2852 | |
2853 | static tree | |
726a989a | 2854 | create_expression_by_pieces (basic_block block, pre_expr expr, |
e076271b | 2855 | gimple_seq *stmts, tree type) |
56db793a | 2856 | { |
83d5977e | 2857 | tree name; |
150e3929 RG |
2858 | tree folded; |
2859 | gimple_seq forced_stmts = NULL; | |
c9145754 | 2860 | unsigned int value_id; |
726a989a | 2861 | gimple_stmt_iterator gsi; |
c9145754 DB |
2862 | tree exprtype = type ? type : get_expr_type (expr); |
2863 | pre_expr nameexpr; | |
726a989a | 2864 | gimple newstmt; |
81c4f554 | 2865 | |
c9145754 | 2866 | switch (expr->kind) |
56db793a | 2867 | { |
c9145754 DB |
2868 | /* We may hit the NAME/CONSTANT case if we have to convert types |
2869 | that value numbering saw through. */ | |
2870 | case NAME: | |
2871 | folded = PRE_EXPR_NAME (expr); | |
2872 | break; | |
2873 | case CONSTANT: | |
2874 | folded = PRE_EXPR_CONSTANT (expr); | |
2875 | break; | |
2876 | case REFERENCE: | |
43da81be | 2877 | { |
c9145754 | 2878 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); |
e076271b | 2879 | folded = create_component_ref_by_pieces (block, ref, stmts); |
c3dd8dd7 RB |
2880 | if (!folded) |
2881 | return NULL_TREE; | |
43da81be DB |
2882 | } |
2883 | break; | |
c9145754 | 2884 | case NARY: |
c90186eb | 2885 | { |
c9145754 | 2886 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); |
f843144b | 2887 | tree *genop = XALLOCAVEC (tree, nary->length); |
5a7d7f9c RG |
2888 | unsigned i; |
2889 | for (i = 0; i < nary->length; ++i) | |
85300b46 | 2890 | { |
e076271b | 2891 | genop[i] = find_or_generate_expression (block, nary->op[i], stmts); |
c3dd8dd7 RB |
2892 | if (!genop[i]) |
2893 | return NULL_TREE; | |
48acf1b7 RG |
2894 | /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It |
2895 | may have conversions stripped. */ | |
2896 | if (nary->opcode == POINTER_PLUS_EXPR) | |
2897 | { | |
2898 | if (i == 0) | |
2899 | genop[i] = fold_convert (nary->type, genop[i]); | |
2900 | else if (i == 1) | |
2901 | genop[i] = convert_to_ptrofftype (genop[i]); | |
2902 | } | |
5a7d7f9c RG |
2903 | else |
2904 | genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]); | |
2905 | } | |
2906 | if (nary->opcode == CONSTRUCTOR) | |
2907 | { | |
9771b263 | 2908 | vec<constructor_elt, va_gc> *elts = NULL; |
5a7d7f9c RG |
2909 | for (i = 0; i < nary->length; ++i) |
2910 | CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]); | |
2911 | folded = build_constructor (nary->type, elts); | |
2912 | } | |
2913 | else | |
2914 | { | |
2915 | switch (nary->length) | |
2916 | { | |
2917 | case 1: | |
2918 | folded = fold_build1 (nary->opcode, nary->type, | |
2919 | genop[0]); | |
2920 | break; | |
2921 | case 2: | |
2922 | folded = fold_build2 (nary->opcode, nary->type, | |
2923 | genop[0], genop[1]); | |
2924 | break; | |
2925 | case 3: | |
2926 | folded = fold_build3 (nary->opcode, nary->type, | |
a475fd3d | 2927 | genop[0], genop[1], genop[2]); |
5a7d7f9c RG |
2928 | break; |
2929 | default: | |
2930 | gcc_unreachable (); | |
2931 | } | |
85300b46 | 2932 | } |
56db793a | 2933 | } |
c9145754 | 2934 | break; |
56db793a | 2935 | default: |
e076271b | 2936 | gcc_unreachable (); |
56db793a | 2937 | } |
150e3929 RG |
2938 | |
2939 | if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded))) | |
2940 | folded = fold_convert (exprtype, folded); | |
2941 | ||
81c4f554 SB |
2942 | /* Force the generated expression to be a sequence of GIMPLE |
2943 | statements. | |
2944 | We have to call unshare_expr because force_gimple_operand may | |
2945 | modify the tree we pass to it. */ | |
150e3929 RG |
2946 | folded = force_gimple_operand (unshare_expr (folded), &forced_stmts, |
2947 | false, NULL); | |
81c4f554 SB |
2948 | |
2949 | /* If we have any intermediate expressions to the value sets, add them | |
a7849637 | 2950 | to the value sets and chain them in the instruction stream. */ |
81c4f554 SB |
2951 | if (forced_stmts) |
2952 | { | |
726a989a RB |
2953 | gsi = gsi_start (forced_stmts); |
2954 | for (; !gsi_end_p (gsi); gsi_next (&gsi)) | |
81c4f554 | 2955 | { |
726a989a RB |
2956 | gimple stmt = gsi_stmt (gsi); |
2957 | tree forcedname = gimple_get_lhs (stmt); | |
c9145754 | 2958 | pre_expr nameexpr; |
b9c5e484 | 2959 | |
c9145754 DB |
2960 | if (TREE_CODE (forcedname) == SSA_NAME) |
2961 | { | |
0ab555de | 2962 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname)); |
c9145754 DB |
2963 | VN_INFO_GET (forcedname)->valnum = forcedname; |
2964 | VN_INFO (forcedname)->value_id = get_next_value_id (); | |
2965 | nameexpr = get_or_alloc_expr_for_name (forcedname); | |
2966 | add_to_value (VN_INFO (forcedname)->value_id, nameexpr); | |
40b178f4 | 2967 | bitmap_value_replace_in_set (NEW_SETS (block), nameexpr); |
c9145754 DB |
2968 | bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr); |
2969 | } | |
81c4f554 | 2970 | } |
726a989a | 2971 | gimple_seq_add_seq (stmts, forced_stmts); |
81c4f554 SB |
2972 | } |
2973 | ||
83d5977e RG |
2974 | name = make_temp_ssa_name (exprtype, NULL, "pretmp"); |
2975 | newstmt = gimple_build_assign (name, folded); | |
726a989a | 2976 | gimple_set_plf (newstmt, NECESSARY, false); |
c90186eb | 2977 | |
726a989a | 2978 | gimple_seq_add_stmt (stmts, newstmt); |
0ab555de | 2979 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name)); |
cfaab3a9 | 2980 | |
0aa16586 RG |
2981 | /* Fold the last statement. */ |
2982 | gsi = gsi_last (*stmts); | |
748c5114 RG |
2983 | if (fold_stmt_inplace (&gsi)) |
2984 | update_stmt (gsi_stmt (gsi)); | |
0aa16586 | 2985 | |
c9145754 | 2986 | /* Add a value number to the temporary. |
81c4f554 | 2987 | The value may already exist in either NEW_SETS, or AVAIL_OUT, because |
e9284566 DB |
2988 | we are creating the expression by pieces, and this particular piece of |
2989 | the expression may have been represented. There is no harm in replacing | |
2990 | here. */ | |
c9145754 | 2991 | value_id = get_expr_value_id (expr); |
40b178f4 RG |
2992 | VN_INFO_GET (name)->value_id = value_id; |
2993 | VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id); | |
2994 | if (VN_INFO (name)->valnum == NULL_TREE) | |
2995 | VN_INFO (name)->valnum = name; | |
2996 | gcc_assert (VN_INFO (name)->valnum != NULL_TREE); | |
c9145754 DB |
2997 | nameexpr = get_or_alloc_expr_for_name (name); |
2998 | add_to_value (value_id, nameexpr); | |
70f34814 | 2999 | if (NEW_SETS (block)) |
c9145754 DB |
3000 | bitmap_value_replace_in_set (NEW_SETS (block), nameexpr); |
3001 | bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr); | |
81c4f554 SB |
3002 | |
3003 | pre_stats.insertions++; | |
56db793a | 3004 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b9c5e484 | 3005 | { |
56db793a | 3006 | fprintf (dump_file, "Inserted "); |
726a989a | 3007 | print_gimple_stmt (dump_file, newstmt, 0, 0); |
56db793a DB |
3008 | fprintf (dump_file, " in predecessor %d\n", block->index); |
3009 | } | |
81c4f554 | 3010 | |
56db793a DB |
3011 | return name; |
3012 | } | |
e9284566 | 3013 | |
c9145754 | 3014 | |
a8338640 MM |
3015 | /* Returns true if we want to inhibit the insertions of PHI nodes |
3016 | for the given EXPR for basic block BB (a member of a loop). | |
3017 | We want to do this, when we fear that the induction variable we | |
3018 | create might inhibit vectorization. */ | |
3019 | ||
3020 | static bool | |
3021 | inhibit_phi_insertion (basic_block bb, pre_expr expr) | |
3022 | { | |
3023 | vn_reference_t vr = PRE_EXPR_REFERENCE (expr); | |
9771b263 | 3024 | vec<vn_reference_op_s> ops = vr->operands; |
a8338640 MM |
3025 | vn_reference_op_t op; |
3026 | unsigned i; | |
3027 | ||
3028 | /* If we aren't going to vectorize we don't inhibit anything. */ | |
3029 | if (!flag_tree_vectorize) | |
3030 | return false; | |
3031 | ||
3032 | /* Otherwise we inhibit the insertion when the address of the | |
3033 | memory reference is a simple induction variable. In other | |
3034 | cases the vectorizer won't do anything anyway (either it's | |
3035 | loop invariant or a complicated expression). */ | |
9771b263 | 3036 | FOR_EACH_VEC_ELT (ops, i, op) |
a8338640 MM |
3037 | { |
3038 | switch (op->opcode) | |
3039 | { | |
109a16c2 RG |
3040 | case CALL_EXPR: |
3041 | /* Calls are not a problem. */ | |
3042 | return false; | |
3043 | ||
a8338640 MM |
3044 | case ARRAY_REF: |
3045 | case ARRAY_RANGE_REF: | |
3046 | if (TREE_CODE (op->op0) != SSA_NAME) | |
3047 | break; | |
3048 | /* Fallthru. */ | |
3049 | case SSA_NAME: | |
3050 | { | |
3051 | basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0)); | |
3052 | affine_iv iv; | |
3053 | /* Default defs are loop invariant. */ | |
3054 | if (!defbb) | |
3055 | break; | |
3056 | /* Defined outside this loop, also loop invariant. */ | |
3057 | if (!flow_bb_inside_loop_p (bb->loop_father, defbb)) | |
3058 | break; | |
3059 | /* If it's a simple induction variable inhibit insertion, | |
3060 | the vectorizer might be interested in this one. */ | |
3061 | if (simple_iv (bb->loop_father, bb->loop_father, | |
3062 | op->op0, &iv, true)) | |
3063 | return true; | |
3064 | /* No simple IV, vectorizer can't do anything, hence no | |
3065 | reason to inhibit the transformation for this operand. */ | |
3066 | break; | |
3067 | } | |
3068 | default: | |
3069 | break; | |
3070 | } | |
3071 | } | |
3072 | return false; | |
3073 | } | |
3074 | ||
83737db2 | 3075 | /* Insert the to-be-made-available values of expression EXPRNUM for each |
c90186eb | 3076 | predecessor, stored in AVAIL, into the predecessors of BLOCK, and |
c9145754 | 3077 | merge the result with a phi node, given the same value number as |
c90186eb | 3078 | NODE. Return true if we have inserted new stuff. */ |
e9284566 DB |
3079 | |
3080 | static bool | |
83737db2 | 3081 | insert_into_preds_of_block (basic_block block, unsigned int exprnum, |
9771b263 | 3082 | vec<pre_expr> avail) |
e9284566 | 3083 | { |
c9145754 DB |
3084 | pre_expr expr = expression_for_id (exprnum); |
3085 | pre_expr newphi; | |
3086 | unsigned int val = get_expr_value_id (expr); | |
e9284566 | 3087 | edge pred; |
0fc6c492 DB |
3088 | bool insertions = false; |
3089 | bool nophi = false; | |
e9284566 | 3090 | basic_block bprime; |
c9145754 | 3091 | pre_expr eprime; |
e9284566 | 3092 | edge_iterator ei; |
c9145754 | 3093 | tree type = get_expr_type (expr); |
de081cfd | 3094 | tree temp; |
726a989a | 3095 | gimple phi; |
b9c5e484 | 3096 | |
0fc6c492 | 3097 | /* Make sure we aren't creating an induction variable. */ |
391886c8 | 3098 | if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2) |
0fc6c492 DB |
3099 | { |
3100 | bool firstinsideloop = false; | |
3101 | bool secondinsideloop = false; | |
b9c5e484 | 3102 | firstinsideloop = flow_bb_inside_loop_p (block->loop_father, |
0fc6c492 DB |
3103 | EDGE_PRED (block, 0)->src); |
3104 | secondinsideloop = flow_bb_inside_loop_p (block->loop_father, | |
3105 | EDGE_PRED (block, 1)->src); | |
3106 | /* Induction variables only have one edge inside the loop. */ | |
a8338640 MM |
3107 | if ((firstinsideloop ^ secondinsideloop) |
3108 | && (expr->kind != REFERENCE | |
3109 | || inhibit_phi_insertion (block, expr))) | |
0fc6c492 DB |
3110 | { |
3111 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3112 | fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n"); | |
3113 | nophi = true; | |
3114 | } | |
3115 | } | |
b9c5e484 | 3116 | |
c9145754 DB |
3117 | /* Make the necessary insertions. */ |
3118 | FOR_EACH_EDGE (pred, ei, block->preds) | |
3119 | { | |
726a989a | 3120 | gimple_seq stmts = NULL; |
c9145754 DB |
3121 | tree builtexpr; |
3122 | bprime = pred->src; | |
9771b263 | 3123 | eprime = avail[pred->dest_idx]; |
c9145754 DB |
3124 | |
3125 | if (eprime->kind != NAME && eprime->kind != CONSTANT) | |
3126 | { | |
e076271b RG |
3127 | builtexpr = create_expression_by_pieces (bprime, eprime, |
3128 | &stmts, type); | |
c9145754 | 3129 | gcc_assert (!(pred->flags & EDGE_ABNORMAL)); |
726a989a | 3130 | gsi_insert_seq_on_edge (pred, stmts); |
c3dd8dd7 RB |
3131 | if (!builtexpr) |
3132 | { | |
3133 | /* We cannot insert a PHI node if we failed to insert | |
3134 | on one edge. */ | |
3135 | nophi = true; | |
3136 | continue; | |
3137 | } | |
9771b263 | 3138 | avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr); |
c9145754 DB |
3139 | insertions = true; |
3140 | } | |
3141 | else if (eprime->kind == CONSTANT) | |
3142 | { | |
3143 | /* Constants may not have the right type, fold_convert | |
1c8f7377 | 3144 | should give us back a constant with the right type. */ |
c9145754 | 3145 | tree constant = PRE_EXPR_CONSTANT (eprime); |
15d5fe33 | 3146 | if (!useless_type_conversion_p (type, TREE_TYPE (constant))) |
c9145754 DB |
3147 | { |
3148 | tree builtexpr = fold_convert (type, constant); | |
b8698a0f | 3149 | if (!is_gimple_min_invariant (builtexpr)) |
c9145754 DB |
3150 | { |
3151 | tree forcedexpr = force_gimple_operand (builtexpr, | |
3152 | &stmts, true, | |
3153 | NULL); | |
15d5fe33 | 3154 | if (!is_gimple_min_invariant (forcedexpr)) |
c9145754 DB |
3155 | { |
3156 | if (forcedexpr != builtexpr) | |
3157 | { | |
3158 | VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime); | |
3159 | VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime); | |
3160 | } | |
3161 | if (stmts) | |
3162 | { | |
726a989a RB |
3163 | gimple_stmt_iterator gsi; |
3164 | gsi = gsi_start (stmts); | |
3165 | for (; !gsi_end_p (gsi); gsi_next (&gsi)) | |
c9145754 | 3166 | { |
726a989a | 3167 | gimple stmt = gsi_stmt (gsi); |
0ab555de MM |
3168 | tree lhs = gimple_get_lhs (stmt); |
3169 | if (TREE_CODE (lhs) == SSA_NAME) | |
3170 | bitmap_set_bit (inserted_exprs, | |
3171 | SSA_NAME_VERSION (lhs)); | |
726a989a | 3172 | gimple_set_plf (stmt, NECESSARY, false); |
c9145754 | 3173 | } |
726a989a | 3174 | gsi_insert_seq_on_edge (pred, stmts); |
c9145754 | 3175 | } |
9771b263 DN |
3176 | avail[pred->dest_idx] |
3177 | = get_or_alloc_expr_for_name (forcedexpr); | |
c9145754 DB |
3178 | } |
3179 | } | |
70f34814 | 3180 | else |
9771b263 DN |
3181 | avail[pred->dest_idx] |
3182 | = get_or_alloc_expr_for_constant (builtexpr); | |
c9145754 DB |
3183 | } |
3184 | } | |
3185 | else if (eprime->kind == NAME) | |
3186 | { | |
3187 | /* We may have to do a conversion because our value | |
3188 | numbering can look through types in certain cases, but | |
3189 | our IL requires all operands of a phi node have the same | |
3190 | type. */ | |
3191 | tree name = PRE_EXPR_NAME (eprime); | |
f709638a | 3192 | if (!useless_type_conversion_p (type, TREE_TYPE (name))) |
c9145754 DB |
3193 | { |
3194 | tree builtexpr; | |
3195 | tree forcedexpr; | |
f709638a | 3196 | builtexpr = fold_convert (type, name); |
c9145754 DB |
3197 | forcedexpr = force_gimple_operand (builtexpr, |
3198 | &stmts, true, | |
3199 | NULL); | |
3200 | ||
3201 | if (forcedexpr != name) | |
3202 | { | |
3203 | VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum; | |
3204 | VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id; | |
3205 | } | |
c90186eb | 3206 | |
c9145754 DB |
3207 | if (stmts) |
3208 | { | |
726a989a RB |
3209 | gimple_stmt_iterator gsi; |
3210 | gsi = gsi_start (stmts); | |
3211 | for (; !gsi_end_p (gsi); gsi_next (&gsi)) | |
c9145754 | 3212 | { |
726a989a | 3213 | gimple stmt = gsi_stmt (gsi); |
0ab555de MM |
3214 | tree lhs = gimple_get_lhs (stmt); |
3215 | if (TREE_CODE (lhs) == SSA_NAME) | |
3216 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs)); | |
726a989a | 3217 | gimple_set_plf (stmt, NECESSARY, false); |
c9145754 | 3218 | } |
726a989a | 3219 | gsi_insert_seq_on_edge (pred, stmts); |
c9145754 | 3220 | } |
9771b263 | 3221 | avail[pred->dest_idx] = get_or_alloc_expr_for_name (forcedexpr); |
c9145754 | 3222 | } |
b9c5e484 | 3223 | } |
e9284566 | 3224 | } |
0fc6c492 DB |
3225 | /* If we didn't want a phi node, and we made insertions, we still have |
3226 | inserted new stuff, and thus return true. If we didn't want a phi node, | |
3227 | and didn't make insertions, we haven't added anything new, so return | |
3228 | false. */ | |
3229 | if (nophi && insertions) | |
3230 | return true; | |
3231 | else if (nophi && !insertions) | |
3232 | return false; | |
3233 | ||
e9284566 | 3234 | /* Now build a phi for the new variable. */ |
83d5977e | 3235 | temp = make_temp_ssa_name (type, NULL, "prephitmp"); |
1e52075c | 3236 | phi = create_phi_node (temp, block); |
de081cfd RG |
3237 | |
3238 | gimple_set_plf (phi, NECESSARY, false); | |
40b178f4 RG |
3239 | VN_INFO_GET (temp)->value_id = val; |
3240 | VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val); | |
3241 | if (VN_INFO (temp)->valnum == NULL_TREE) | |
3242 | VN_INFO (temp)->valnum = temp; | |
3243 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp)); | |
e9284566 | 3244 | FOR_EACH_EDGE (pred, ei, block->preds) |
c9145754 | 3245 | { |
9771b263 | 3246 | pre_expr ae = avail[pred->dest_idx]; |
c9145754 DB |
3247 | gcc_assert (get_expr_type (ae) == type |
3248 | || useless_type_conversion_p (type, get_expr_type (ae))); | |
3249 | if (ae->kind == CONSTANT) | |
2724573f RB |
3250 | add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)), |
3251 | pred, UNKNOWN_LOCATION); | |
c9145754 | 3252 | else |
1c8f7377 | 3253 | add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION); |
c9145754 | 3254 | } |
b9c5e484 | 3255 | |
40b178f4 | 3256 | newphi = get_or_alloc_expr_for_name (temp); |
c9145754 | 3257 | add_to_value (val, newphi); |
b9c5e484 | 3258 | |
e9284566 DB |
3259 | /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing |
3260 | this insertion, since we test for the existence of this value in PHI_GEN | |
3261 | before proceeding with the partial redundancy checks in insert_aux. | |
b9c5e484 | 3262 | |
e9284566 DB |
3263 | The value may exist in AVAIL_OUT, in particular, it could be represented |
3264 | by the expression we are trying to eliminate, in which case we want the | |
3265 | replacement to occur. If it's not existing in AVAIL_OUT, we want it | |
3266 | inserted there. | |
b9c5e484 | 3267 | |
e9284566 DB |
3268 | Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of |
3269 | this block, because if it did, it would have existed in our dominator's | |
3270 | AVAIL_OUT, and would have been skipped due to the full redundancy check. | |
3271 | */ | |
3272 | ||
c9145754 | 3273 | bitmap_insert_into_set (PHI_GEN (block), newphi); |
b9c5e484 | 3274 | bitmap_value_replace_in_set (AVAIL_OUT (block), |
c9145754 | 3275 | newphi); |
e9284566 | 3276 | bitmap_insert_into_set (NEW_SETS (block), |
c9145754 | 3277 | newphi); |
b9c5e484 | 3278 | |
e9284566 DB |
3279 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3280 | { | |
3281 | fprintf (dump_file, "Created phi "); | |
726a989a | 3282 | print_gimple_stmt (dump_file, phi, 0, 0); |
e9284566 DB |
3283 | fprintf (dump_file, " in block %d\n", block->index); |
3284 | } | |
3285 | pre_stats.phis++; | |
3286 | return true; | |
3287 | } | |
3288 | ||
3289 | ||
b9c5e484 | 3290 | |
7e6eb623 DB |
3291 | /* Perform insertion of partially redundant values. |
3292 | For BLOCK, do the following: | |
3293 | 1. Propagate the NEW_SETS of the dominator into the current block. | |
b9c5e484 | 3294 | If the block has multiple predecessors, |
7e6eb623 | 3295 | 2a. Iterate over the ANTIC expressions for the block to see if |
83737db2 | 3296 | any of them are partially redundant. |
7e6eb623 | 3297 | 2b. If so, insert them into the necessary predecessors to make |
83737db2 | 3298 | the expression fully redundant. |
7e6eb623 DB |
3299 | 2c. Insert a new PHI merging the values of the predecessors. |
3300 | 2d. Insert the new PHI, and the new expressions, into the | |
83737db2 | 3301 | NEW_SETS set. |
7e6eb623 | 3302 | 3. Recursively call ourselves on the dominator children of BLOCK. |
6de9cd9a | 3303 | |
83737db2 | 3304 | Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by |
d75dbccd | 3305 | do_regular_insertion and do_partial_insertion. |
83737db2 | 3306 | |
7e6eb623 | 3307 | */ |
e9284566 | 3308 | |
83737db2 DB |
3309 | static bool |
3310 | do_regular_insertion (basic_block block, basic_block dom) | |
3311 | { | |
3312 | bool new_stuff = false; | |
9771b263 | 3313 | vec<pre_expr> exprs; |
c9145754 | 3314 | pre_expr expr; |
6e1aa848 | 3315 | vec<pre_expr> avail = vNULL; |
83737db2 DB |
3316 | int i; |
3317 | ||
1c8f7377 | 3318 | exprs = sorted_array_from_bitmap_set (ANTIC_IN (block)); |
9771b263 | 3319 | avail.safe_grow (EDGE_COUNT (block->preds)); |
1c8f7377 | 3320 | |
9771b263 | 3321 | FOR_EACH_VEC_ELT (exprs, i, expr) |
83737db2 | 3322 | { |
4bcc5786 RB |
3323 | if (expr->kind == NARY |
3324 | || expr->kind == REFERENCE) | |
83737db2 | 3325 | { |
c9145754 | 3326 | unsigned int val; |
83737db2 DB |
3327 | bool by_some = false; |
3328 | bool cant_insert = false; | |
3329 | bool all_same = true; | |
c9145754 | 3330 | pre_expr first_s = NULL; |
83737db2 DB |
3331 | edge pred; |
3332 | basic_block bprime; | |
c9145754 | 3333 | pre_expr eprime = NULL; |
83737db2 | 3334 | edge_iterator ei; |
f11d2f1e | 3335 | pre_expr edoubleprime = NULL; |
5813994e | 3336 | bool do_insertion = false; |
83737db2 | 3337 | |
c9145754 | 3338 | val = get_expr_value_id (expr); |
83737db2 DB |
3339 | if (bitmap_set_contains_value (PHI_GEN (block), val)) |
3340 | continue; | |
3341 | if (bitmap_set_contains_value (AVAIL_OUT (dom), val)) | |
3342 | { | |
3343 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
ead69dac JL |
3344 | { |
3345 | fprintf (dump_file, "Found fully redundant value: "); | |
3346 | print_pre_expr (dump_file, expr); | |
3347 | fprintf (dump_file, "\n"); | |
3348 | } | |
83737db2 DB |
3349 | continue; |
3350 | } | |
3351 | ||
83737db2 DB |
3352 | FOR_EACH_EDGE (pred, ei, block->preds) |
3353 | { | |
c9145754 | 3354 | unsigned int vprime; |
83737db2 | 3355 | |
c4ab2baa RG |
3356 | /* We should never run insertion for the exit block |
3357 | and so not come across fake pred edges. */ | |
3358 | gcc_assert (!(pred->flags & EDGE_FAKE)); | |
83737db2 DB |
3359 | bprime = pred->src; |
3360 | eprime = phi_translate (expr, ANTIC_IN (block), NULL, | |
3361 | bprime, block); | |
3362 | ||
3363 | /* eprime will generally only be NULL if the | |
3364 | value of the expression, translated | |
3365 | through the PHI for this predecessor, is | |
3366 | undefined. If that is the case, we can't | |
3367 | make the expression fully redundant, | |
3368 | because its value is undefined along a | |
3369 | predecessor path. We can thus break out | |
3370 | early because it doesn't matter what the | |
3371 | rest of the results are. */ | |
3372 | if (eprime == NULL) | |
3373 | { | |
9771b263 | 3374 | avail[pred->dest_idx] = NULL; |
83737db2 DB |
3375 | cant_insert = true; |
3376 | break; | |
3377 | } | |
3378 | ||
3379 | eprime = fully_constant_expression (eprime); | |
726a989a RB |
3380 | vprime = get_expr_value_id (eprime); |
3381 | edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), | |
e076271b | 3382 | vprime); |
83737db2 DB |
3383 | if (edoubleprime == NULL) |
3384 | { | |
9771b263 | 3385 | avail[pred->dest_idx] = eprime; |
83737db2 DB |
3386 | all_same = false; |
3387 | } | |
3388 | else | |
3389 | { | |
9771b263 | 3390 | avail[pred->dest_idx] = edoubleprime; |
83737db2 | 3391 | by_some = true; |
5813994e RG |
3392 | /* We want to perform insertions to remove a redundancy on |
3393 | a path in the CFG we want to optimize for speed. */ | |
3394 | if (optimize_edge_for_speed_p (pred)) | |
3395 | do_insertion = true; | |
83737db2 DB |
3396 | if (first_s == NULL) |
3397 | first_s = edoubleprime; | |
5deac340 | 3398 | else if (!pre_expr_d::equal (first_s, edoubleprime)) |
83737db2 DB |
3399 | all_same = false; |
3400 | } | |
3401 | } | |
3402 | /* If we can insert it, it's not the same value | |
3403 | already existing along every predecessor, and | |
3404 | it's defined by some predecessor, it is | |
3405 | partially redundant. */ | |
3bc27de7 | 3406 | if (!cant_insert && !all_same && by_some) |
83737db2 | 3407 | { |
3bc27de7 RG |
3408 | if (!do_insertion) |
3409 | { | |
3410 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3411 | { | |
3412 | fprintf (dump_file, "Skipping partial redundancy for " | |
3413 | "expression "); | |
3414 | print_pre_expr (dump_file, expr); | |
3415 | fprintf (dump_file, " (%04d), no redundancy on to be " | |
3416 | "optimized for speed edge\n", val); | |
3417 | } | |
3418 | } | |
19372838 RG |
3419 | else if (dbg_cnt (treepre_insert)) |
3420 | { | |
3421 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3422 | { | |
3423 | fprintf (dump_file, "Found partial redundancy for " | |
3424 | "expression "); | |
3425 | print_pre_expr (dump_file, expr); | |
3426 | fprintf (dump_file, " (%04d)\n", | |
3427 | get_expr_value_id (expr)); | |
3428 | } | |
3429 | if (insert_into_preds_of_block (block, | |
3430 | get_expression_id (expr), | |
3431 | avail)) | |
3432 | new_stuff = true; | |
3433 | } | |
83737db2 DB |
3434 | } |
3435 | /* If all edges produce the same value and that value is | |
3436 | an invariant, then the PHI has the same value on all | |
3437 | edges. Note this. */ | |
4bcc5786 | 3438 | else if (!cant_insert && all_same) |
83737db2 | 3439 | { |
4bcc5786 RB |
3440 | gcc_assert (edoubleprime->kind == CONSTANT |
3441 | || edoubleprime->kind == NAME); | |
3442 | ||
3443 | tree temp = make_temp_ssa_name (get_expr_type (expr), | |
3444 | NULL, "pretmp"); | |
3445 | gimple assign = gimple_build_assign (temp, | |
3446 | edoubleprime->kind == CONSTANT ? PRE_EXPR_CONSTANT (edoubleprime) : PRE_EXPR_NAME (edoubleprime)); | |
3447 | gimple_stmt_iterator gsi = gsi_after_labels (block); | |
3448 | gsi_insert_before (&gsi, assign, GSI_NEW_STMT); | |
3449 | ||
3450 | gimple_set_plf (assign, NECESSARY, false); | |
3451 | VN_INFO_GET (temp)->value_id = val; | |
3452 | VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val); | |
3453 | if (VN_INFO (temp)->valnum == NULL_TREE) | |
3454 | VN_INFO (temp)->valnum = temp; | |
3455 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp)); | |
3456 | pre_expr newe = get_or_alloc_expr_for_name (temp); | |
3457 | add_to_value (val, newe); | |
3458 | bitmap_value_replace_in_set (AVAIL_OUT (block), newe); | |
3459 | bitmap_insert_into_set (NEW_SETS (block), newe); | |
83737db2 | 3460 | } |
83737db2 DB |
3461 | } |
3462 | } | |
3463 | ||
9771b263 DN |
3464 | exprs.release (); |
3465 | avail.release (); | |
83737db2 DB |
3466 | return new_stuff; |
3467 | } | |
3468 | ||
3469 | ||
d75dbccd DB |
3470 | /* Perform insertion for partially anticipatable expressions. There |
3471 | is only one case we will perform insertion for these. This case is | |
3472 | if the expression is partially anticipatable, and fully available. | |
3473 | In this case, we know that putting it earlier will enable us to | |
3474 | remove the later computation. */ | |
3475 | ||
3476 | ||
3477 | static bool | |
3478 | do_partial_partial_insertion (basic_block block, basic_block dom) | |
3479 | { | |
3480 | bool new_stuff = false; | |
9771b263 | 3481 | vec<pre_expr> exprs; |
c9145754 | 3482 | pre_expr expr; |
6e1aa848 | 3483 | vec<pre_expr> avail = vNULL; |
d75dbccd DB |
3484 | int i; |
3485 | ||
1c8f7377 | 3486 | exprs = sorted_array_from_bitmap_set (PA_IN (block)); |
9771b263 | 3487 | avail.safe_grow (EDGE_COUNT (block->preds)); |
1c8f7377 | 3488 | |
9771b263 | 3489 | FOR_EACH_VEC_ELT (exprs, i, expr) |
d75dbccd | 3490 | { |
4bcc5786 RB |
3491 | if (expr->kind == NARY |
3492 | || expr->kind == REFERENCE) | |
d75dbccd | 3493 | { |
c9145754 | 3494 | unsigned int val; |
d75dbccd DB |
3495 | bool by_all = true; |
3496 | bool cant_insert = false; | |
3497 | edge pred; | |
3498 | basic_block bprime; | |
c9145754 | 3499 | pre_expr eprime = NULL; |
d75dbccd DB |
3500 | edge_iterator ei; |
3501 | ||
c9145754 | 3502 | val = get_expr_value_id (expr); |
d75dbccd DB |
3503 | if (bitmap_set_contains_value (PHI_GEN (block), val)) |
3504 | continue; | |
3505 | if (bitmap_set_contains_value (AVAIL_OUT (dom), val)) | |
3506 | continue; | |
3507 | ||
d75dbccd DB |
3508 | FOR_EACH_EDGE (pred, ei, block->preds) |
3509 | { | |
c9145754 DB |
3510 | unsigned int vprime; |
3511 | pre_expr edoubleprime; | |
d75dbccd | 3512 | |
c4ab2baa RG |
3513 | /* We should never run insertion for the exit block |
3514 | and so not come across fake pred edges. */ | |
3515 | gcc_assert (!(pred->flags & EDGE_FAKE)); | |
d75dbccd DB |
3516 | bprime = pred->src; |
3517 | eprime = phi_translate (expr, ANTIC_IN (block), | |
3518 | PA_IN (block), | |
3519 | bprime, block); | |
3520 | ||
3521 | /* eprime will generally only be NULL if the | |
3522 | value of the expression, translated | |
3523 | through the PHI for this predecessor, is | |
3524 | undefined. If that is the case, we can't | |
3525 | make the expression fully redundant, | |
3526 | because its value is undefined along a | |
3527 | predecessor path. We can thus break out | |
3528 | early because it doesn't matter what the | |
3529 | rest of the results are. */ | |
3530 | if (eprime == NULL) | |
3531 | { | |
9771b263 | 3532 | avail[pred->dest_idx] = NULL; |
d75dbccd DB |
3533 | cant_insert = true; |
3534 | break; | |
3535 | } | |
3536 | ||
3537 | eprime = fully_constant_expression (eprime); | |
726a989a | 3538 | vprime = get_expr_value_id (eprime); |
e076271b | 3539 | edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime); |
9771b263 | 3540 | avail[pred->dest_idx] = edoubleprime; |
d75dbccd DB |
3541 | if (edoubleprime == NULL) |
3542 | { | |
3543 | by_all = false; | |
3544 | break; | |
3545 | } | |
d75dbccd DB |
3546 | } |
3547 | ||
3548 | /* If we can insert it, it's not the same value | |
3549 | already existing along every predecessor, and | |
3550 | it's defined by some predecessor, it is | |
3551 | partially redundant. */ | |
fa06ad0d | 3552 | if (!cant_insert && by_all) |
d75dbccd | 3553 | { |
fa06ad0d JR |
3554 | edge succ; |
3555 | bool do_insertion = false; | |
3556 | ||
3557 | /* Insert only if we can remove a later expression on a path | |
3558 | that we want to optimize for speed. | |
3559 | The phi node that we will be inserting in BLOCK is not free, | |
3560 | and inserting it for the sake of !optimize_for_speed successor | |
3561 | may cause regressions on the speed path. */ | |
3562 | FOR_EACH_EDGE (succ, ei, block->succs) | |
3563 | { | |
0c46ead2 JH |
3564 | if (bitmap_set_contains_value (PA_IN (succ->dest), val) |
3565 | || bitmap_set_contains_value (ANTIC_IN (succ->dest), val)) | |
fa06ad0d JR |
3566 | { |
3567 | if (optimize_edge_for_speed_p (succ)) | |
3568 | do_insertion = true; | |
3569 | } | |
3570 | } | |
3571 | ||
3572 | if (!do_insertion) | |
3573 | { | |
3574 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3575 | { | |
3576 | fprintf (dump_file, "Skipping partial partial redundancy " | |
3577 | "for expression "); | |
3578 | print_pre_expr (dump_file, expr); | |
0c46ead2 | 3579 | fprintf (dump_file, " (%04d), not (partially) anticipated " |
fa06ad0d JR |
3580 | "on any to be optimized for speed edges\n", val); |
3581 | } | |
3582 | } | |
3583 | else if (dbg_cnt (treepre_insert)) | |
3584 | { | |
3585 | pre_stats.pa_insert++; | |
19372838 RG |
3586 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3587 | { | |
3588 | fprintf (dump_file, "Found partial partial redundancy " | |
3589 | "for expression "); | |
3590 | print_pre_expr (dump_file, expr); | |
3591 | fprintf (dump_file, " (%04d)\n", | |
3592 | get_expr_value_id (expr)); | |
3593 | } | |
fa06ad0d JR |
3594 | if (insert_into_preds_of_block (block, |
3595 | get_expression_id (expr), | |
3596 | avail)) | |
3597 | new_stuff = true; | |
3598 | } | |
3599 | } | |
d75dbccd DB |
3600 | } |
3601 | } | |
3602 | ||
9771b263 DN |
3603 | exprs.release (); |
3604 | avail.release (); | |
d75dbccd DB |
3605 | return new_stuff; |
3606 | } | |
83737db2 | 3607 | |
7e6eb623 DB |
3608 | static bool |
3609 | insert_aux (basic_block block) | |
6de9cd9a | 3610 | { |
7e6eb623 DB |
3611 | basic_block son; |
3612 | bool new_stuff = false; | |
6de9cd9a | 3613 | |
7e6eb623 | 3614 | if (block) |
6de9cd9a | 3615 | { |
7e6eb623 DB |
3616 | basic_block dom; |
3617 | dom = get_immediate_dominator (CDI_DOMINATORS, block); | |
3618 | if (dom) | |
a32b97a2 | 3619 | { |
3cd8c58a | 3620 | unsigned i; |
87c476a2 | 3621 | bitmap_iterator bi; |
6b416da1 | 3622 | bitmap_set_t newset = NEW_SETS (dom); |
e9284566 | 3623 | if (newset) |
87c476a2 | 3624 | { |
e9284566 DB |
3625 | /* Note that we need to value_replace both NEW_SETS, and |
3626 | AVAIL_OUT. For both the case of NEW_SETS, the value may be | |
3627 | represented by some non-simple expression here that we want | |
3628 | to replace it with. */ | |
83737db2 | 3629 | FOR_EACH_EXPR_ID_IN_SET (newset, i, bi) |
e9284566 | 3630 | { |
c9145754 | 3631 | pre_expr expr = expression_for_id (i); |
83737db2 DB |
3632 | bitmap_value_replace_in_set (NEW_SETS (block), expr); |
3633 | bitmap_value_replace_in_set (AVAIL_OUT (block), expr); | |
e9284566 | 3634 | } |
87c476a2 | 3635 | } |
c5cbcccf | 3636 | if (!single_pred_p (block)) |
6de9cd9a | 3637 | { |
83737db2 | 3638 | new_stuff |= do_regular_insertion (block, dom); |
d75dbccd DB |
3639 | if (do_partial_partial) |
3640 | new_stuff |= do_partial_partial_insertion (block, dom); | |
6de9cd9a DN |
3641 | } |
3642 | } | |
3643 | } | |
7e6eb623 DB |
3644 | for (son = first_dom_son (CDI_DOMINATORS, block); |
3645 | son; | |
3646 | son = next_dom_son (CDI_DOMINATORS, son)) | |
3647 | { | |
3648 | new_stuff |= insert_aux (son); | |
3649 | } | |
3650 | ||
3651 | return new_stuff; | |
6de9cd9a DN |
3652 | } |
3653 | ||
7e6eb623 | 3654 | /* Perform insertion of partially redundant values. */ |
6de9cd9a | 3655 | |
7e6eb623 DB |
3656 | static void |
3657 | insert (void) | |
6de9cd9a | 3658 | { |
7e6eb623 | 3659 | bool new_stuff = true; |
6de9cd9a | 3660 | basic_block bb; |
7e6eb623 | 3661 | int num_iterations = 0; |
b9c5e484 | 3662 | |
7e6eb623 | 3663 | FOR_ALL_BB (bb) |
6b416da1 | 3664 | NEW_SETS (bb) = bitmap_set_new (); |
b9c5e484 | 3665 | |
7e6eb623 | 3666 | while (new_stuff) |
6de9cd9a | 3667 | { |
7e6eb623 | 3668 | num_iterations++; |
19372838 RG |
3669 | if (dump_file && dump_flags & TDF_DETAILS) |
3670 | fprintf (dump_file, "Starting insert iteration %d\n", num_iterations); | |
7e6eb623 | 3671 | new_stuff = insert_aux (ENTRY_BLOCK_PTR); |
3dbc97a9 RB |
3672 | |
3673 | /* Clear the NEW sets before the next iteration. We have already | |
3674 | fully propagated its contents. */ | |
3675 | if (new_stuff) | |
3676 | FOR_ALL_BB (bb) | |
3677 | bitmap_set_free (NEW_SETS (bb)); | |
6de9cd9a | 3678 | } |
9fe0cb7d | 3679 | statistics_histogram_event (cfun, "insert iterations", num_iterations); |
7e6eb623 | 3680 | } |
6de9cd9a | 3681 | |
33c94679 | 3682 | |
665fcad8 SB |
3683 | /* Compute the AVAIL set for all basic blocks. |
3684 | ||
3685 | This function performs value numbering of the statements in each basic | |
3686 | block. The AVAIL sets are built from information we glean while doing | |
3687 | this value numbering, since the AVAIL sets contain only one entry per | |
7e6eb623 | 3688 | value. |
b9c5e484 | 3689 | |
7e6eb623 | 3690 | AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)]. |
ff2ad0f7 | 3691 | AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */ |
6de9cd9a | 3692 | |
7e6eb623 | 3693 | static void |
665fcad8 | 3694 | compute_avail (void) |
6de9cd9a | 3695 | { |
c9145754 | 3696 | |
665fcad8 SB |
3697 | basic_block block, son; |
3698 | basic_block *worklist; | |
3699 | size_t sp = 0; | |
b005da11 RG |
3700 | unsigned i; |
3701 | ||
3702 | /* We pretend that default definitions are defined in the entry block. | |
3703 | This includes function arguments and the static chain decl. */ | |
3704 | for (i = 1; i < num_ssa_names; ++i) | |
3705 | { | |
3706 | tree name = ssa_name (i); | |
3707 | pre_expr e; | |
3708 | if (!name | |
3709 | || !SSA_NAME_IS_DEFAULT_DEF (name) | |
3710 | || has_zero_uses (name) | |
ea057359 | 3711 | || virtual_operand_p (name)) |
b005da11 | 3712 | continue; |
665fcad8 | 3713 | |
b005da11 RG |
3714 | e = get_or_alloc_expr_for_name (name); |
3715 | add_to_value (get_expr_value_id (e), e); | |
40b178f4 | 3716 | bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e); |
b005da11 | 3717 | bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e); |
2984956b AP |
3718 | } |
3719 | ||
e3815735 RB |
3720 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3721 | { | |
3722 | print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR), | |
3723 | "tmp_gen", ENTRY_BLOCK); | |
3724 | print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR), | |
3725 | "avail_out", ENTRY_BLOCK); | |
3726 | } | |
3727 | ||
665fcad8 | 3728 | /* Allocate the worklist. */ |
e1111e8e | 3729 | worklist = XNEWVEC (basic_block, n_basic_blocks); |
665fcad8 SB |
3730 | |
3731 | /* Seed the algorithm by putting the dominator children of the entry | |
3732 | block on the worklist. */ | |
3733 | for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR); | |
3734 | son; | |
3735 | son = next_dom_son (CDI_DOMINATORS, son)) | |
3736 | worklist[sp++] = son; | |
3737 | ||
3738 | /* Loop until the worklist is empty. */ | |
3739 | while (sp) | |
6de9cd9a | 3740 | { |
726a989a RB |
3741 | gimple_stmt_iterator gsi; |
3742 | gimple stmt; | |
7e6eb623 DB |
3743 | basic_block dom; |
3744 | ||
665fcad8 SB |
3745 | /* Pick a block from the worklist. */ |
3746 | block = worklist[--sp]; | |
3747 | ||
ff2ad0f7 DN |
3748 | /* Initially, the set of available values in BLOCK is that of |
3749 | its immediate dominator. */ | |
7e6eb623 DB |
3750 | dom = get_immediate_dominator (CDI_DOMINATORS, block); |
3751 | if (dom) | |
bdee7684 | 3752 | bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom)); |
33c94679 | 3753 | |
ff2ad0f7 | 3754 | /* Generate values for PHI nodes. */ |
726a989a | 3755 | for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi)) |
e3815735 RB |
3756 | { |
3757 | tree result = gimple_phi_result (gsi_stmt (gsi)); | |
3758 | ||
3759 | /* We have no need for virtual phis, as they don't represent | |
3760 | actual computations. */ | |
3761 | if (virtual_operand_p (result)) | |
3762 | continue; | |
3763 | ||
3764 | pre_expr e = get_or_alloc_expr_for_name (result); | |
3765 | add_to_value (get_expr_value_id (e), e); | |
3766 | bitmap_value_insert_into_set (AVAIL_OUT (block), e); | |
3767 | bitmap_insert_into_set (PHI_GEN (block), e); | |
3768 | } | |
7e6eb623 | 3769 | |
a19eb9d2 RG |
3770 | BB_MAY_NOTRETURN (block) = 0; |
3771 | ||
ff2ad0f7 DN |
3772 | /* Now compute value numbers and populate value sets with all |
3773 | the expressions computed in BLOCK. */ | |
726a989a | 3774 | for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi)) |
6de9cd9a | 3775 | { |
f47c96aa AM |
3776 | ssa_op_iter iter; |
3777 | tree op; | |
ff2ad0f7 | 3778 | |
726a989a | 3779 | stmt = gsi_stmt (gsi); |
ff2ad0f7 | 3780 | |
a19eb9d2 RG |
3781 | /* Cache whether the basic-block has any non-visible side-effect |
3782 | or control flow. | |
3783 | If this isn't a call or it is the last stmt in the | |
3784 | basic-block then the CFG represents things correctly. */ | |
9cbbba28 | 3785 | if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt)) |
a19eb9d2 RG |
3786 | { |
3787 | /* Non-looping const functions always return normally. | |
3788 | Otherwise the call might not return or have side-effects | |
3789 | that forbids hoisting possibly trapping expressions | |
3790 | before it. */ | |
3791 | int flags = gimple_call_flags (stmt); | |
3792 | if (!(flags & ECF_CONST) | |
3793 | || (flags & ECF_LOOPING_CONST_OR_PURE)) | |
3794 | BB_MAY_NOTRETURN (block) = 1; | |
3795 | } | |
3796 | ||
c9145754 | 3797 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF) |
83737db2 | 3798 | { |
c9145754 | 3799 | pre_expr e = get_or_alloc_expr_for_name (op); |
83737db2 | 3800 | |
c9145754 | 3801 | add_to_value (get_expr_value_id (e), e); |
40b178f4 | 3802 | bitmap_insert_into_set (TMP_GEN (block), e); |
c9145754 | 3803 | bitmap_value_insert_into_set (AVAIL_OUT (block), e); |
83737db2 DB |
3804 | } |
3805 | ||
17742d62 RG |
3806 | if (gimple_has_side_effects (stmt) |
3807 | || stmt_could_throw_p (stmt) | |
3808 | || is_gimple_debug (stmt)) | |
726a989a RB |
3809 | continue; |
3810 | ||
17742d62 | 3811 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) |
e3815735 RB |
3812 | { |
3813 | if (ssa_undefined_value_p (op)) | |
3814 | continue; | |
3815 | pre_expr e = get_or_alloc_expr_for_name (op); | |
3816 | bitmap_value_insert_into_set (EXP_GEN (block), e); | |
3817 | } | |
17742d62 | 3818 | |
726a989a | 3819 | switch (gimple_code (stmt)) |
7e6eb623 | 3820 | { |
726a989a | 3821 | case GIMPLE_RETURN: |
c9145754 | 3822 | continue; |
726a989a RB |
3823 | |
3824 | case GIMPLE_CALL: | |
c9145754 | 3825 | { |
726a989a | 3826 | vn_reference_t ref; |
726a989a | 3827 | pre_expr result = NULL; |
6e1aa848 | 3828 | vec<vn_reference_op_s> ops = vNULL; |
c9145754 | 3829 | |
9cbbba28 EB |
3830 | /* We can value number only calls to real functions. */ |
3831 | if (gimple_call_internal_p (stmt)) | |
726a989a | 3832 | continue; |
c9145754 | 3833 | |
726a989a | 3834 | copy_reference_ops_from_call (stmt, &ops); |
b45d2719 RG |
3835 | vn_reference_lookup_pieces (gimple_vuse (stmt), 0, |
3836 | gimple_expr_type (stmt), | |
3bc27de7 | 3837 | ops, &ref, VN_NOWALK); |
9771b263 | 3838 | ops.release (); |
726a989a RB |
3839 | if (!ref) |
3840 | continue; | |
c9145754 | 3841 | |
cd32bb90 RG |
3842 | /* If the value of the call is not invalidated in |
3843 | this block until it is computed, add the expression | |
3844 | to EXP_GEN. */ | |
3845 | if (!gimple_vuse (stmt) | |
3846 | || gimple_code | |
3847 | (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI | |
3848 | || gimple_bb (SSA_NAME_DEF_STMT | |
3849 | (gimple_vuse (stmt))) != block) | |
3850 | { | |
3851 | result = (pre_expr) pool_alloc (pre_expr_pool); | |
3852 | result->kind = REFERENCE; | |
3853 | result->id = 0; | |
3854 | PRE_EXPR_REFERENCE (result) = ref; | |
3855 | ||
3856 | get_or_alloc_expression_id (result); | |
3857 | add_to_value (get_expr_value_id (result), result); | |
3c1d57d0 | 3858 | bitmap_value_insert_into_set (EXP_GEN (block), result); |
cd32bb90 | 3859 | } |
726a989a RB |
3860 | continue; |
3861 | } | |
c9145754 | 3862 | |
726a989a RB |
3863 | case GIMPLE_ASSIGN: |
3864 | { | |
3865 | pre_expr result = NULL; | |
17742d62 | 3866 | switch (vn_get_stmt_kind (stmt)) |
726a989a | 3867 | { |
17742d62 | 3868 | case VN_NARY: |
726a989a | 3869 | { |
61514fe4 | 3870 | enum tree_code code = gimple_assign_rhs_code (stmt); |
726a989a | 3871 | vn_nary_op_t nary; |
61514fe4 RG |
3872 | |
3873 | /* COND_EXPR and VEC_COND_EXPR are awkward in | |
3874 | that they contain an embedded complex expression. | |
3875 | Don't even try to shove those through PRE. */ | |
3876 | if (code == COND_EXPR | |
3877 | || code == VEC_COND_EXPR) | |
3878 | continue; | |
3879 | ||
91af9dc9 | 3880 | vn_nary_op_lookup_stmt (stmt, &nary); |
726a989a RB |
3881 | if (!nary) |
3882 | continue; | |
3883 | ||
073a8998 | 3884 | /* If the NARY traps and there was a preceding |
bea966c2 RG |
3885 | point in the block that might not return avoid |
3886 | adding the nary to EXP_GEN. */ | |
3887 | if (BB_MAY_NOTRETURN (block) | |
3888 | && vn_nary_may_trap (nary)) | |
3889 | continue; | |
3890 | ||
726a989a RB |
3891 | result = (pre_expr) pool_alloc (pre_expr_pool); |
3892 | result->kind = NARY; | |
3893 | result->id = 0; | |
3894 | PRE_EXPR_NARY (result) = nary; | |
3895 | break; | |
3896 | } | |
c9145754 | 3897 | |
17742d62 | 3898 | case VN_REFERENCE: |
726a989a RB |
3899 | { |
3900 | vn_reference_t ref; | |
726a989a | 3901 | vn_reference_lookup (gimple_assign_rhs1 (stmt), |
5006671f | 3902 | gimple_vuse (stmt), |
3bc27de7 | 3903 | VN_WALK, &ref); |
726a989a RB |
3904 | if (!ref) |
3905 | continue; | |
3906 | ||
cd32bb90 RG |
3907 | /* If the value of the reference is not invalidated in |
3908 | this block until it is computed, add the expression | |
3909 | to EXP_GEN. */ | |
3910 | if (gimple_vuse (stmt)) | |
3911 | { | |
3912 | gimple def_stmt; | |
3913 | bool ok = true; | |
3914 | def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt)); | |
3915 | while (!gimple_nop_p (def_stmt) | |
3916 | && gimple_code (def_stmt) != GIMPLE_PHI | |
3917 | && gimple_bb (def_stmt) == block) | |
3918 | { | |
3919 | if (stmt_may_clobber_ref_p | |
3920 | (def_stmt, gimple_assign_rhs1 (stmt))) | |
3921 | { | |
3922 | ok = false; | |
3923 | break; | |
3924 | } | |
3925 | def_stmt | |
3926 | = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt)); | |
3927 | } | |
3928 | if (!ok) | |
3929 | continue; | |
3930 | } | |
3931 | ||
726a989a RB |
3932 | result = (pre_expr) pool_alloc (pre_expr_pool); |
3933 | result->kind = REFERENCE; | |
3934 | result->id = 0; | |
3935 | PRE_EXPR_REFERENCE (result) = ref; | |
3936 | break; | |
3937 | } | |
c9145754 | 3938 | |
726a989a | 3939 | default: |
726a989a | 3940 | continue; |
c9145754 | 3941 | } |
726a989a RB |
3942 | |
3943 | get_or_alloc_expression_id (result); | |
3944 | add_to_value (get_expr_value_id (result), result); | |
3c1d57d0 | 3945 | bitmap_value_insert_into_set (EXP_GEN (block), result); |
dda243de | 3946 | continue; |
c9145754 DB |
3947 | } |
3948 | default: | |
3949 | break; | |
7e6eb623 | 3950 | } |
6de9cd9a | 3951 | } |
726a989a | 3952 | |
e3815735 RB |
3953 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3954 | { | |
3955 | print_bitmap_set (dump_file, EXP_GEN (block), | |
3956 | "exp_gen", block->index); | |
3957 | print_bitmap_set (dump_file, PHI_GEN (block), | |
3958 | "phi_gen", block->index); | |
3959 | print_bitmap_set (dump_file, TMP_GEN (block), | |
3960 | "tmp_gen", block->index); | |
3961 | print_bitmap_set (dump_file, AVAIL_OUT (block), | |
3962 | "avail_out", block->index); | |
3963 | } | |
3964 | ||
665fcad8 SB |
3965 | /* Put the dominator children of BLOCK on the worklist of blocks |
3966 | to compute available sets for. */ | |
3967 | for (son = first_dom_son (CDI_DOMINATORS, block); | |
3968 | son; | |
3969 | son = next_dom_son (CDI_DOMINATORS, son)) | |
3970 | worklist[sp++] = son; | |
6de9cd9a | 3971 | } |
33c94679 | 3972 | |
665fcad8 | 3973 | free (worklist); |
7e6eb623 DB |
3974 | } |
3975 | ||
40b178f4 RG |
3976 | |
3977 | /* Local state for the eliminate domwalk. */ | |
9771b263 DN |
3978 | static vec<gimple> el_to_remove; |
3979 | static vec<gimple> el_to_update; | |
40b178f4 | 3980 | static unsigned int el_todo; |
9771b263 DN |
3981 | static vec<tree> el_avail; |
3982 | static vec<tree> el_avail_stack; | |
40b178f4 RG |
3983 | |
3984 | /* Return a leader for OP that is available at the current point of the | |
3985 | eliminate domwalk. */ | |
3d45dd59 RG |
3986 | |
3987 | static tree | |
40b178f4 | 3988 | eliminate_avail (tree op) |
3d45dd59 | 3989 | { |
40b178f4 RG |
3990 | tree valnum = VN_INFO (op)->valnum; |
3991 | if (TREE_CODE (valnum) == SSA_NAME) | |
3992 | { | |
3993 | if (SSA_NAME_IS_DEFAULT_DEF (valnum)) | |
3994 | return valnum; | |
9771b263 DN |
3995 | if (el_avail.length () > SSA_NAME_VERSION (valnum)) |
3996 | return el_avail[SSA_NAME_VERSION (valnum)]; | |
40b178f4 RG |
3997 | } |
3998 | else if (is_gimple_min_invariant (valnum)) | |
3999 | return valnum; | |
4000 | return NULL_TREE; | |
4001 | } | |
4002 | ||
4003 | /* At the current point of the eliminate domwalk make OP available. */ | |
3d45dd59 | 4004 | |
40b178f4 RG |
4005 | static void |
4006 | eliminate_push_avail (tree op) | |
4007 | { | |
4008 | tree valnum = VN_INFO (op)->valnum; | |
4009 | if (TREE_CODE (valnum) == SSA_NAME) | |
4010 | { | |
9771b263 DN |
4011 | if (el_avail.length () <= SSA_NAME_VERSION (valnum)) |
4012 | el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1); | |
4013 | el_avail[SSA_NAME_VERSION (valnum)] = op; | |
4014 | el_avail_stack.safe_push (op); | |
40b178f4 RG |
4015 | } |
4016 | } | |
4017 | ||
4018 | /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns | |
4019 | the leader for the expression if insertion was successful. */ | |
4020 | ||
4021 | static tree | |
4022 | eliminate_insert (gimple_stmt_iterator *gsi, tree val) | |
4023 | { | |
4024 | tree expr = vn_get_expr_for (val); | |
4025 | if (!CONVERT_EXPR_P (expr) | |
4026 | && TREE_CODE (expr) != VIEW_CONVERT_EXPR) | |
3d45dd59 | 4027 | return NULL_TREE; |
3d45dd59 | 4028 | |
40b178f4 RG |
4029 | tree op = TREE_OPERAND (expr, 0); |
4030 | tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op; | |
4031 | if (!leader) | |
3d45dd59 | 4032 | return NULL_TREE; |
3d45dd59 | 4033 | |
40b178f4 RG |
4034 | tree res = make_temp_ssa_name (TREE_TYPE (val), NULL, "pretmp"); |
4035 | gimple tem = gimple_build_assign (res, | |
c96cab6e RB |
4036 | fold_build1 (TREE_CODE (expr), |
4037 | TREE_TYPE (expr), leader)); | |
40b178f4 RG |
4038 | gsi_insert_before (gsi, tem, GSI_SAME_STMT); |
4039 | VN_INFO_GET (res)->valnum = val; | |
4040 | ||
4041 | if (TREE_CODE (leader) == SSA_NAME) | |
4042 | gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true); | |
4043 | ||
4044 | pre_stats.insertions++; | |
4045 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4046 | { | |
4047 | fprintf (dump_file, "Inserted "); | |
4048 | print_gimple_stmt (dump_file, tem, 0, 0); | |
4049 | } | |
4050 | ||
4051 | return res; | |
3d45dd59 | 4052 | } |
33c94679 | 4053 | |
40b178f4 | 4054 | /* Perform elimination for the basic-block B during the domwalk. */ |
7e6eb623 | 4055 | |
40b178f4 RG |
4056 | static void |
4057 | eliminate_bb (dom_walk_data *, basic_block b) | |
7e6eb623 | 4058 | { |
5006671f RG |
4059 | gimple_stmt_iterator gsi; |
4060 | gimple stmt; | |
7e6eb623 | 4061 | |
40b178f4 | 4062 | /* Mark new bb. */ |
9771b263 | 4063 | el_avail_stack.safe_push (NULL_TREE); |
40b178f4 RG |
4064 | |
4065 | for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);) | |
7e6eb623 | 4066 | { |
40b178f4 RG |
4067 | gimple stmt, phi = gsi_stmt (gsi); |
4068 | tree sprime = NULL_TREE, res = PHI_RESULT (phi); | |
4069 | gimple_stmt_iterator gsi2; | |
4070 | ||
4071 | /* We want to perform redundant PHI elimination. Do so by | |
4072 | replacing the PHI with a single copy if possible. | |
4073 | Do not touch inserted, single-argument or virtual PHIs. */ | |
4074 | if (gimple_phi_num_args (phi) == 1 | |
4075 | || virtual_operand_p (res)) | |
83737db2 | 4076 | { |
40b178f4 RG |
4077 | gsi_next (&gsi); |
4078 | continue; | |
4079 | } | |
c9145754 | 4080 | |
40b178f4 RG |
4081 | sprime = eliminate_avail (res); |
4082 | if (!sprime | |
4083 | || sprime == res) | |
4084 | { | |
4085 | eliminate_push_avail (res); | |
4086 | gsi_next (&gsi); | |
4087 | continue; | |
4088 | } | |
4089 | else if (is_gimple_min_invariant (sprime)) | |
4090 | { | |
4091 | if (!useless_type_conversion_p (TREE_TYPE (res), | |
4092 | TREE_TYPE (sprime))) | |
4093 | sprime = fold_convert (TREE_TYPE (res), sprime); | |
4094 | } | |
c9145754 | 4095 | |
40b178f4 RG |
4096 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4097 | { | |
4098 | fprintf (dump_file, "Replaced redundant PHI node defining "); | |
4099 | print_generic_expr (dump_file, res, 0); | |
4100 | fprintf (dump_file, " with "); | |
4101 | print_generic_expr (dump_file, sprime, 0); | |
4102 | fprintf (dump_file, "\n"); | |
4103 | } | |
6999afe1 | 4104 | |
40b178f4 RG |
4105 | remove_phi_node (&gsi, false); |
4106 | ||
4107 | if (inserted_exprs | |
4108 | && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)) | |
4109 | && TREE_CODE (sprime) == SSA_NAME) | |
4110 | gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true); | |
4111 | ||
4112 | if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime))) | |
4113 | sprime = fold_convert (TREE_TYPE (res), sprime); | |
4114 | stmt = gimple_build_assign (res, sprime); | |
4115 | SSA_NAME_DEF_STMT (res) = stmt; | |
4116 | gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY)); | |
4117 | ||
4118 | gsi2 = gsi_after_labels (b); | |
4119 | gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT); | |
4120 | /* Queue the copy for eventual removal. */ | |
9771b263 | 4121 | el_to_remove.safe_push (stmt); |
40b178f4 RG |
4122 | /* If we inserted this PHI node ourself, it's not an elimination. */ |
4123 | if (inserted_exprs | |
4124 | && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))) | |
4125 | pre_stats.phis--; | |
4126 | else | |
4127 | pre_stats.eliminations++; | |
4128 | } | |
ff2ad0f7 | 4129 | |
40b178f4 RG |
4130 | for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi)) |
4131 | { | |
4132 | tree lhs = NULL_TREE; | |
4133 | tree rhs = NULL_TREE; | |
4134 | ||
4135 | stmt = gsi_stmt (gsi); | |
4136 | ||
4137 | if (gimple_has_lhs (stmt)) | |
4138 | lhs = gimple_get_lhs (stmt); | |
4139 | ||
4140 | if (gimple_assign_single_p (stmt)) | |
4141 | rhs = gimple_assign_rhs1 (stmt); | |
4142 | ||
4143 | /* Lookup the RHS of the expression, see if we have an | |
4144 | available computation for it. If so, replace the RHS with | |
4bcc5786 | 4145 | the available computation. */ |
40b178f4 RG |
4146 | if (gimple_has_lhs (stmt) |
4147 | && TREE_CODE (lhs) == SSA_NAME | |
40b178f4 RG |
4148 | && !gimple_has_volatile_ops (stmt)) |
4149 | { | |
4150 | tree sprime; | |
4151 | gimple orig_stmt = stmt; | |
3d45dd59 | 4152 | |
40b178f4 | 4153 | sprime = eliminate_avail (lhs); |
4bcc5786 RB |
4154 | /* If there is no usable leader mark lhs as leader for its value. */ |
4155 | if (!sprime) | |
4156 | eliminate_push_avail (lhs); | |
4157 | ||
4158 | /* See PR43491. Do not replace a global register variable when | |
4159 | it is a the RHS of an assignment. Do replace local register | |
4160 | variables since gcc does not guarantee a local variable will | |
4161 | be allocated in register. | |
4162 | Do not perform copy propagation or undo constant propagation. */ | |
4163 | if (gimple_assign_single_p (stmt) | |
4164 | && (TREE_CODE (rhs) == SSA_NAME | |
4165 | || is_gimple_min_invariant (rhs) | |
4166 | || (TREE_CODE (rhs) == VAR_DECL | |
4167 | && is_global_var (rhs) | |
4168 | && DECL_HARD_REGISTER (rhs)))) | |
4169 | continue; | |
4170 | ||
40b178f4 RG |
4171 | if (!sprime) |
4172 | { | |
3d45dd59 RG |
4173 | /* If there is no existing usable leader but SCCVN thinks |
4174 | it has an expression it wants to use as replacement, | |
4175 | insert that. */ | |
40b178f4 RG |
4176 | tree val = VN_INFO (lhs)->valnum; |
4177 | if (val != VN_TOP | |
4178 | && TREE_CODE (val) == SSA_NAME | |
4179 | && VN_INFO (val)->needs_insertion | |
a044f0e7 | 4180 | && VN_INFO (val)->expr != NULL_TREE |
40b178f4 RG |
4181 | && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE) |
4182 | eliminate_push_avail (sprime); | |
4183 | } | |
4184 | else if (is_gimple_min_invariant (sprime)) | |
4185 | { | |
4186 | /* If there is no existing leader but SCCVN knows this | |
4187 | value is constant, use that constant. */ | |
4188 | if (!useless_type_conversion_p (TREE_TYPE (lhs), | |
4189 | TREE_TYPE (sprime))) | |
4190 | sprime = fold_convert (TREE_TYPE (lhs), sprime); | |
4191 | ||
4192 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3d45dd59 | 4193 | { |
40b178f4 RG |
4194 | fprintf (dump_file, "Replaced "); |
4195 | print_gimple_expr (dump_file, stmt, 0, 0); | |
4196 | fprintf (dump_file, " with "); | |
4197 | print_generic_expr (dump_file, sprime, 0); | |
4198 | fprintf (dump_file, " in "); | |
4199 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
3d45dd59 | 4200 | } |
40b178f4 RG |
4201 | pre_stats.eliminations++; |
4202 | propagate_tree_value_into_stmt (&gsi, sprime); | |
4203 | stmt = gsi_stmt (gsi); | |
4204 | update_stmt (stmt); | |
4205 | ||
4206 | /* If we removed EH side-effects from the statement, clean | |
4207 | its EH information. */ | |
4208 | if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt)) | |
ff2ad0f7 | 4209 | { |
40b178f4 RG |
4210 | bitmap_set_bit (need_eh_cleanup, |
4211 | gimple_bb (stmt)->index); | |
4212 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4213 | fprintf (dump_file, " Removed EH side-effects.\n"); | |
4214 | } | |
4215 | continue; | |
4216 | } | |
30fd5881 | 4217 | |
40b178f4 RG |
4218 | if (sprime |
4219 | && sprime != lhs | |
4220 | && (rhs == NULL_TREE | |
4221 | || TREE_CODE (rhs) != SSA_NAME | |
4222 | || may_propagate_copy (rhs, sprime))) | |
4223 | { | |
4224 | bool can_make_abnormal_goto | |
4225 | = is_gimple_call (stmt) | |
4226 | && stmt_can_make_abnormal_goto (stmt); | |
b9c5e484 | 4227 | |
40b178f4 | 4228 | gcc_assert (sprime != rhs); |
30fd5881 | 4229 | |
40b178f4 RG |
4230 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4231 | { | |
4232 | fprintf (dump_file, "Replaced "); | |
4233 | print_gimple_expr (dump_file, stmt, 0, 0); | |
4234 | fprintf (dump_file, " with "); | |
4235 | print_generic_expr (dump_file, sprime, 0); | |
4236 | fprintf (dump_file, " in "); | |
4237 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
ff2ad0f7 | 4238 | } |
40b178f4 RG |
4239 | |
4240 | if (TREE_CODE (sprime) == SSA_NAME) | |
4241 | gimple_set_plf (SSA_NAME_DEF_STMT (sprime), | |
4242 | NECESSARY, true); | |
4243 | /* We need to make sure the new and old types actually match, | |
4244 | which may require adding a simple cast, which fold_convert | |
4245 | will do for us. */ | |
4246 | if ((!rhs || TREE_CODE (rhs) != SSA_NAME) | |
4247 | && !useless_type_conversion_p (gimple_expr_type (stmt), | |
4248 | TREE_TYPE (sprime))) | |
4249 | sprime = fold_convert (gimple_expr_type (stmt), sprime); | |
4250 | ||
4251 | pre_stats.eliminations++; | |
4252 | propagate_tree_value_into_stmt (&gsi, sprime); | |
4253 | stmt = gsi_stmt (gsi); | |
4254 | update_stmt (stmt); | |
4255 | ||
4256 | /* If we removed EH side-effects from the statement, clean | |
4257 | its EH information. */ | |
4258 | if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt)) | |
6cdb0ee3 | 4259 | { |
40b178f4 RG |
4260 | bitmap_set_bit (need_eh_cleanup, |
4261 | gimple_bb (stmt)->index); | |
6cdb0ee3 | 4262 | if (dump_file && (dump_flags & TDF_DETAILS)) |
40b178f4 RG |
4263 | fprintf (dump_file, " Removed EH side-effects.\n"); |
4264 | } | |
6cdb0ee3 | 4265 | |
40b178f4 RG |
4266 | /* Likewise for AB side-effects. */ |
4267 | if (can_make_abnormal_goto | |
4268 | && !stmt_can_make_abnormal_goto (stmt)) | |
4269 | { | |
4270 | bitmap_set_bit (need_ab_cleanup, | |
4271 | gimple_bb (stmt)->index); | |
4272 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4273 | fprintf (dump_file, " Removed AB side-effects.\n"); | |
6cdb0ee3 RG |
4274 | } |
4275 | } | |
40b178f4 RG |
4276 | } |
4277 | /* If the statement is a scalar store, see if the expression | |
4278 | has the same value number as its rhs. If so, the store is | |
4279 | dead. */ | |
4280 | else if (gimple_assign_single_p (stmt) | |
4281 | && !gimple_has_volatile_ops (stmt) | |
4282 | && !is_gimple_reg (gimple_assign_lhs (stmt)) | |
4283 | && (TREE_CODE (rhs) == SSA_NAME | |
4284 | || is_gimple_min_invariant (rhs))) | |
4285 | { | |
4286 | tree val; | |
4287 | val = vn_reference_lookup (gimple_assign_lhs (stmt), | |
4288 | gimple_vuse (stmt), VN_WALK, NULL); | |
4289 | if (TREE_CODE (rhs) == SSA_NAME) | |
4290 | rhs = VN_INFO (rhs)->valnum; | |
4291 | if (val | |
4292 | && operand_equal_p (val, rhs, 0)) | |
b80280f2 | 4293 | { |
40b178f4 | 4294 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b80280f2 | 4295 | { |
40b178f4 RG |
4296 | fprintf (dump_file, "Deleted redundant store "); |
4297 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
b80280f2 | 4298 | } |
40b178f4 RG |
4299 | |
4300 | /* Queue stmt for removal. */ | |
9771b263 | 4301 | el_to_remove.safe_push (stmt); |
b80280f2 | 4302 | } |
40b178f4 RG |
4303 | } |
4304 | /* Visit COND_EXPRs and fold the comparison with the | |
4305 | available value-numbers. */ | |
4306 | else if (gimple_code (stmt) == GIMPLE_COND) | |
4307 | { | |
4308 | tree op0 = gimple_cond_lhs (stmt); | |
4309 | tree op1 = gimple_cond_rhs (stmt); | |
4310 | tree result; | |
4311 | ||
4312 | if (TREE_CODE (op0) == SSA_NAME) | |
4313 | op0 = VN_INFO (op0)->valnum; | |
4314 | if (TREE_CODE (op1) == SSA_NAME) | |
4315 | op1 = VN_INFO (op1)->valnum; | |
4316 | result = fold_binary (gimple_cond_code (stmt), boolean_type_node, | |
4317 | op0, op1); | |
4318 | if (result && TREE_CODE (result) == INTEGER_CST) | |
aa7069aa | 4319 | { |
40b178f4 RG |
4320 | if (integer_zerop (result)) |
4321 | gimple_cond_make_false (stmt); | |
3b45a007 | 4322 | else |
40b178f4 RG |
4323 | gimple_cond_make_true (stmt); |
4324 | update_stmt (stmt); | |
4325 | el_todo = TODO_cleanup_cfg; | |
4326 | } | |
4327 | } | |
4328 | /* Visit indirect calls and turn them into direct calls if | |
4329 | possible. */ | |
4330 | if (is_gimple_call (stmt)) | |
4331 | { | |
4332 | tree orig_fn = gimple_call_fn (stmt); | |
4333 | tree fn; | |
4334 | if (!orig_fn) | |
4335 | continue; | |
4336 | if (TREE_CODE (orig_fn) == SSA_NAME) | |
4337 | fn = VN_INFO (orig_fn)->valnum; | |
4338 | else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF | |
4339 | && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME) | |
e248d83f MJ |
4340 | { |
4341 | fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum; | |
4342 | if (!gimple_call_addr_fndecl (fn)) | |
4343 | { | |
4344 | fn = ipa_intraprocedural_devirtualization (stmt); | |
4345 | if (fn) | |
4346 | fn = build_fold_addr_expr (fn); | |
4347 | } | |
4348 | } | |
40b178f4 RG |
4349 | else |
4350 | continue; | |
4351 | if (gimple_call_addr_fndecl (fn) != NULL_TREE | |
4352 | && useless_type_conversion_p (TREE_TYPE (orig_fn), | |
4353 | TREE_TYPE (fn))) | |
4354 | { | |
4355 | bool can_make_abnormal_goto | |
4356 | = stmt_can_make_abnormal_goto (stmt); | |
4357 | bool was_noreturn = gimple_call_noreturn_p (stmt); | |
aa7069aa | 4358 | |
40b178f4 RG |
4359 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4360 | { | |
4361 | fprintf (dump_file, "Replacing call target with "); | |
4362 | print_generic_expr (dump_file, fn, 0); | |
4363 | fprintf (dump_file, " in "); | |
4364 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
4365 | } | |
30fd5881 | 4366 | |
40b178f4 | 4367 | gimple_call_set_fn (stmt, fn); |
9771b263 | 4368 | el_to_update.safe_push (stmt); |
87c20fe7 | 4369 | |
40b178f4 RG |
4370 | /* When changing a call into a noreturn call, cfg cleanup |
4371 | is needed to fix up the noreturn call. */ | |
4372 | if (!was_noreturn && gimple_call_noreturn_p (stmt)) | |
4373 | el_todo |= TODO_cleanup_cfg; | |
30fd5881 | 4374 | |
40b178f4 RG |
4375 | /* If we removed EH side-effects from the statement, clean |
4376 | its EH information. */ | |
4377 | if (maybe_clean_or_replace_eh_stmt (stmt, stmt)) | |
4378 | { | |
4379 | bitmap_set_bit (need_eh_cleanup, | |
4380 | gimple_bb (stmt)->index); | |
4381 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4382 | fprintf (dump_file, " Removed EH side-effects.\n"); | |
4383 | } | |
de0b4ad5 | 4384 | |
40b178f4 RG |
4385 | /* Likewise for AB side-effects. */ |
4386 | if (can_make_abnormal_goto | |
4387 | && !stmt_can_make_abnormal_goto (stmt)) | |
4388 | { | |
4389 | bitmap_set_bit (need_ab_cleanup, | |
4390 | gimple_bb (stmt)->index); | |
4391 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4392 | fprintf (dump_file, " Removed AB side-effects.\n"); | |
aa7069aa | 4393 | } |
40b178f4 RG |
4394 | |
4395 | /* Changing an indirect call to a direct call may | |
4396 | have exposed different semantics. This may | |
4397 | require an SSA update. */ | |
4398 | el_todo |= TODO_update_ssa_only_virtuals; | |
aa7069aa | 4399 | } |
83737db2 | 4400 | } |
40b178f4 RG |
4401 | } |
4402 | } | |
439ef907 | 4403 | |
40b178f4 | 4404 | /* Make no longer available leaders no longer available. */ |
439ef907 | 4405 | |
40b178f4 RG |
4406 | static void |
4407 | eliminate_leave_block (dom_walk_data *, basic_block) | |
4408 | { | |
4409 | tree entry; | |
9771b263 DN |
4410 | while ((entry = el_avail_stack.pop ()) != NULL_TREE) |
4411 | el_avail[SSA_NAME_VERSION (VN_INFO (entry)->valnum)] = NULL_TREE; | |
40b178f4 | 4412 | } |
439ef907 | 4413 | |
40b178f4 | 4414 | /* Eliminate fully redundant computations. */ |
439ef907 | 4415 | |
40b178f4 RG |
4416 | static unsigned int |
4417 | eliminate (void) | |
4418 | { | |
4419 | struct dom_walk_data walk_data; | |
4420 | gimple_stmt_iterator gsi; | |
4421 | gimple stmt; | |
4422 | unsigned i; | |
439ef907 | 4423 | |
40b178f4 RG |
4424 | need_eh_cleanup = BITMAP_ALLOC (NULL); |
4425 | need_ab_cleanup = BITMAP_ALLOC (NULL); | |
4d2ab9e3 | 4426 | |
9771b263 DN |
4427 | el_to_remove.create (0); |
4428 | el_to_update.create (0); | |
40b178f4 | 4429 | el_todo = 0; |
9771b263 DN |
4430 | el_avail.create (0); |
4431 | el_avail_stack.create (0); | |
40b178f4 RG |
4432 | |
4433 | walk_data.dom_direction = CDI_DOMINATORS; | |
4434 | walk_data.initialize_block_local_data = NULL; | |
4435 | walk_data.before_dom_children = eliminate_bb; | |
4436 | walk_data.after_dom_children = eliminate_leave_block; | |
4437 | walk_data.global_data = NULL; | |
4438 | walk_data.block_local_data_size = 0; | |
4439 | init_walk_dominator_tree (&walk_data); | |
4440 | walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR); | |
4441 | fini_walk_dominator_tree (&walk_data); | |
4442 | ||
9771b263 DN |
4443 | el_avail.release (); |
4444 | el_avail_stack.release (); | |
b80280f2 | 4445 | |
5006671f | 4446 | /* We cannot remove stmts during BB walk, especially not release SSA |
439ef907 | 4447 | names there as this confuses the VN machinery. The stmts ending |
40b178f4 | 4448 | up in el_to_remove are either stores or simple copies. */ |
9771b263 | 4449 | FOR_EACH_VEC_ELT (el_to_remove, i, stmt) |
5006671f | 4450 | { |
439ef907 | 4451 | tree lhs = gimple_assign_lhs (stmt); |
84ae6d7b | 4452 | tree rhs = gimple_assign_rhs1 (stmt); |
439ef907 RG |
4453 | use_operand_p use_p; |
4454 | gimple use_stmt; | |
4455 | ||
4456 | /* If there is a single use only, propagate the equivalency | |
4457 | instead of keeping the copy. */ | |
4458 | if (TREE_CODE (lhs) == SSA_NAME | |
84ae6d7b | 4459 | && TREE_CODE (rhs) == SSA_NAME |
85e59f3a | 4460 | && single_imm_use (lhs, &use_p, &use_stmt) |
84ae6d7b | 4461 | && may_propagate_copy (USE_FROM_PTR (use_p), rhs)) |
439ef907 | 4462 | { |
43896afb | 4463 | SET_USE (use_p, rhs); |
b4104018 | 4464 | update_stmt (use_stmt); |
40b178f4 RG |
4465 | if (inserted_exprs |
4466 | && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs)) | |
43896afb RG |
4467 | && TREE_CODE (rhs) == SSA_NAME) |
4468 | gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true); | |
439ef907 RG |
4469 | } |
4470 | ||
4471 | /* If this is a store or a now unused copy, remove it. */ | |
4472 | if (TREE_CODE (lhs) != SSA_NAME | |
4473 | || has_zero_uses (lhs)) | |
4474 | { | |
a2c0ed2e | 4475 | basic_block bb = gimple_bb (stmt); |
439ef907 RG |
4476 | gsi = gsi_for_stmt (stmt); |
4477 | unlink_stmt_vdef (stmt); | |
b5b3ec3e RG |
4478 | if (gsi_remove (&gsi, true)) |
4479 | bitmap_set_bit (need_eh_cleanup, bb->index); | |
40b178f4 RG |
4480 | if (inserted_exprs |
4481 | && TREE_CODE (lhs) == SSA_NAME) | |
0ab555de | 4482 | bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs)); |
439ef907 RG |
4483 | release_defs (stmt); |
4484 | } | |
5006671f | 4485 | } |
9771b263 | 4486 | el_to_remove.release (); |
5006671f | 4487 | |
9aef53ee RG |
4488 | /* We cannot update call statements with virtual operands during |
4489 | SSA walk. This might remove them which in turn makes our | |
4490 | VN lattice invalid. */ | |
9771b263 | 4491 | FOR_EACH_VEC_ELT (el_to_update, i, stmt) |
9aef53ee | 4492 | update_stmt (stmt); |
9771b263 | 4493 | el_to_update.release (); |
9aef53ee | 4494 | |
40b178f4 RG |
4495 | return el_todo; |
4496 | } | |
4497 | ||
4498 | /* Perform CFG cleanups made necessary by elimination. */ | |
4499 | ||
9c370032 | 4500 | static unsigned |
40b178f4 RG |
4501 | fini_eliminate (void) |
4502 | { | |
4503 | bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup); | |
4504 | bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup); | |
4505 | ||
4506 | if (do_eh_cleanup) | |
4507 | gimple_purge_all_dead_eh_edges (need_eh_cleanup); | |
4508 | ||
4509 | if (do_ab_cleanup) | |
4510 | gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup); | |
4511 | ||
4512 | BITMAP_FREE (need_eh_cleanup); | |
4513 | BITMAP_FREE (need_ab_cleanup); | |
4514 | ||
4515 | if (do_eh_cleanup || do_ab_cleanup) | |
9c370032 RB |
4516 | return TODO_cleanup_cfg; |
4517 | return 0; | |
6de9cd9a DN |
4518 | } |
4519 | ||
0fc6c492 DB |
4520 | /* Borrow a bit of tree-ssa-dce.c for the moment. |
4521 | XXX: In 4.1, we should be able to just run a DCE pass after PRE, though | |
4522 | this may be a bit faster, and we may want critical edges kept split. */ | |
4523 | ||
4524 | /* If OP's defining statement has not already been determined to be necessary, | |
d4e6fecb | 4525 | mark that statement necessary. Return the stmt, if it is newly |
b9c5e484 | 4526 | necessary. */ |
0fc6c492 | 4527 | |
726a989a | 4528 | static inline gimple |
d4e6fecb | 4529 | mark_operand_necessary (tree op) |
0fc6c492 | 4530 | { |
726a989a | 4531 | gimple stmt; |
0fc6c492 DB |
4532 | |
4533 | gcc_assert (op); | |
4534 | ||
c90186eb DB |
4535 | if (TREE_CODE (op) != SSA_NAME) |
4536 | return NULL; | |
4537 | ||
0fc6c492 DB |
4538 | stmt = SSA_NAME_DEF_STMT (op); |
4539 | gcc_assert (stmt); | |
4540 | ||
726a989a RB |
4541 | if (gimple_plf (stmt, NECESSARY) |
4542 | || gimple_nop_p (stmt)) | |
d4e6fecb | 4543 | return NULL; |
0fc6c492 | 4544 | |
726a989a | 4545 | gimple_set_plf (stmt, NECESSARY, true); |
d4e6fecb | 4546 | return stmt; |
0fc6c492 DB |
4547 | } |
4548 | ||
4549 | /* Because we don't follow exactly the standard PRE algorithm, and decide not | |
4550 | to insert PHI nodes sometimes, and because value numbering of casts isn't | |
4551 | perfect, we sometimes end up inserting dead code. This simple DCE-like | |
4552 | pass removes any insertions we made that weren't actually used. */ | |
4553 | ||
4554 | static void | |
4555 | remove_dead_inserted_code (void) | |
4556 | { | |
0ab555de MM |
4557 | bitmap worklist; |
4558 | unsigned i; | |
4559 | bitmap_iterator bi; | |
726a989a | 4560 | gimple t; |
0fc6c492 | 4561 | |
0ab555de MM |
4562 | worklist = BITMAP_ALLOC (NULL); |
4563 | EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi) | |
0fc6c492 | 4564 | { |
0ab555de | 4565 | t = SSA_NAME_DEF_STMT (ssa_name (i)); |
726a989a | 4566 | if (gimple_plf (t, NECESSARY)) |
0ab555de | 4567 | bitmap_set_bit (worklist, i); |
0fc6c492 | 4568 | } |
0ab555de | 4569 | while (!bitmap_empty_p (worklist)) |
0fc6c492 | 4570 | { |
0ab555de MM |
4571 | i = bitmap_first_set_bit (worklist); |
4572 | bitmap_clear_bit (worklist, i); | |
4573 | t = SSA_NAME_DEF_STMT (ssa_name (i)); | |
c90186eb DB |
4574 | |
4575 | /* PHI nodes are somewhat special in that each PHI alternative has | |
4576 | data and control dependencies. All the statements feeding the | |
4577 | PHI node's arguments are always necessary. */ | |
726a989a | 4578 | if (gimple_code (t) == GIMPLE_PHI) |
0fc6c492 | 4579 | { |
726a989a | 4580 | unsigned k; |
d4e6fecb | 4581 | |
726a989a | 4582 | for (k = 0; k < gimple_phi_num_args (t); k++) |
83737db2 | 4583 | { |
0fc6c492 DB |
4584 | tree arg = PHI_ARG_DEF (t, k); |
4585 | if (TREE_CODE (arg) == SSA_NAME) | |
d4e6fecb | 4586 | { |
726a989a RB |
4587 | gimple n = mark_operand_necessary (arg); |
4588 | if (n) | |
0ab555de | 4589 | bitmap_set_bit (worklist, SSA_NAME_VERSION (arg)); |
d4e6fecb | 4590 | } |
0fc6c492 DB |
4591 | } |
4592 | } | |
4593 | else | |
4594 | { | |
4595 | /* Propagate through the operands. Examine all the USE, VUSE and | |
89fb70a3 | 4596 | VDEF operands in this statement. Mark all the statements |
0fc6c492 DB |
4597 | which feed this statement's uses as necessary. */ |
4598 | ssa_op_iter iter; | |
4599 | tree use; | |
4600 | ||
38635499 | 4601 | /* The operands of VDEF expressions are also needed as they |
0fc6c492 | 4602 | represent potential definitions that may reach this |
89fb70a3 | 4603 | statement (VDEF operands allow us to follow def-def |
0fc6c492 | 4604 | links). */ |
33c94679 | 4605 | |
0fc6c492 | 4606 | FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES) |
d4e6fecb | 4607 | { |
726a989a | 4608 | gimple n = mark_operand_necessary (use); |
d4e6fecb | 4609 | if (n) |
0ab555de | 4610 | bitmap_set_bit (worklist, SSA_NAME_VERSION (use)); |
d4e6fecb | 4611 | } |
0fc6c492 DB |
4612 | } |
4613 | } | |
c90186eb | 4614 | |
0ab555de | 4615 | EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi) |
0fc6c492 | 4616 | { |
0ab555de | 4617 | t = SSA_NAME_DEF_STMT (ssa_name (i)); |
726a989a | 4618 | if (!gimple_plf (t, NECESSARY)) |
0fc6c492 | 4619 | { |
726a989a | 4620 | gimple_stmt_iterator gsi; |
c90186eb | 4621 | |
0fc6c492 DB |
4622 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4623 | { | |
4624 | fprintf (dump_file, "Removing unnecessary insertion:"); | |
726a989a | 4625 | print_gimple_stmt (dump_file, t, 0, 0); |
0fc6c492 | 4626 | } |
c90186eb | 4627 | |
726a989a RB |
4628 | gsi = gsi_for_stmt (t); |
4629 | if (gimple_code (t) == GIMPLE_PHI) | |
4630 | remove_phi_node (&gsi, true); | |
0fc6c492 | 4631 | else |
b70fdfe4 AO |
4632 | { |
4633 | gsi_remove (&gsi, true); | |
4634 | release_defs (t); | |
4635 | } | |
0fc6c492 DB |
4636 | } |
4637 | } | |
0ab555de | 4638 | BITMAP_FREE (worklist); |
0fc6c492 | 4639 | } |
c90186eb | 4640 | |
9ca87236 | 4641 | |
ff2ad0f7 | 4642 | /* Initialize data structures used by PRE. */ |
6de9cd9a DN |
4643 | |
4644 | static void | |
40b178f4 | 4645 | init_pre (void) |
6de9cd9a | 4646 | { |
c9145754 DB |
4647 | basic_block bb; |
4648 | ||
4649 | next_expression_id = 1; | |
9771b263 DN |
4650 | expressions.create (0); |
4651 | expressions.safe_push (NULL); | |
4652 | value_expressions.create (get_max_value_id () + 1); | |
4653 | value_expressions.safe_grow_cleared (get_max_value_id() + 1); | |
4654 | name_to_id.create (0); | |
c9145754 | 4655 | |
0ab555de | 4656 | inserted_exprs = BITMAP_ALLOC (NULL); |
c90186eb | 4657 | |
b71b4522 DB |
4658 | connect_infinite_loops_to_exit (); |
4659 | memset (&pre_stats, 0, sizeof (pre_stats)); | |
4660 | ||
585d0dc4 RG |
4661 | postorder = XNEWVEC (int, n_basic_blocks); |
4662 | postorder_num = inverted_post_order_compute (postorder); | |
c9145754 | 4663 | |
a72ae88a | 4664 | alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets)); |
c9145754 | 4665 | |
b71b4522 | 4666 | calculate_dominance_info (CDI_POST_DOMINATORS); |
d75dbccd DB |
4667 | calculate_dominance_info (CDI_DOMINATORS); |
4668 | ||
c9145754 | 4669 | bitmap_obstack_initialize (&grand_bitmap_obstack); |
0823efed DN |
4670 | phi_translate_table.create (5110); |
4671 | expression_to_id.create (num_ssa_names * 3); | |
c9145754 DB |
4672 | bitmap_set_pool = create_alloc_pool ("Bitmap sets", |
4673 | sizeof (struct bitmap_set), 30); | |
4674 | pre_expr_pool = create_alloc_pool ("pre_expr nodes", | |
4675 | sizeof (struct pre_expr_d), 30); | |
4676 | FOR_ALL_BB (bb) | |
4677 | { | |
40b178f4 RG |
4678 | EXP_GEN (bb) = bitmap_set_new (); |
4679 | PHI_GEN (bb) = bitmap_set_new (); | |
4680 | TMP_GEN (bb) = bitmap_set_new (); | |
c9145754 DB |
4681 | AVAIL_OUT (bb) = bitmap_set_new (); |
4682 | } | |
ff2ad0f7 DN |
4683 | } |
4684 | ||
4685 | ||
4686 | /* Deallocate data structures used by PRE. */ | |
6de9cd9a | 4687 | |
ff2ad0f7 | 4688 | static void |
40b178f4 | 4689 | fini_pre () |
ff2ad0f7 | 4690 | { |
c9145754 | 4691 | free (postorder); |
9771b263 | 4692 | value_expressions.release (); |
0ab555de | 4693 | BITMAP_FREE (inserted_exprs); |
c9145754 DB |
4694 | bitmap_obstack_release (&grand_bitmap_obstack); |
4695 | free_alloc_pool (bitmap_set_pool); | |
4696 | free_alloc_pool (pre_expr_pool); | |
0823efed DN |
4697 | phi_translate_table.dispose (); |
4698 | expression_to_id.dispose (); | |
9771b263 | 4699 | name_to_id.release (); |
50265400 | 4700 | |
a72ae88a | 4701 | free_aux_for_blocks (); |
6809cbf9 | 4702 | |
b71b4522 | 4703 | free_dominance_info (CDI_POST_DOMINATORS); |
ff2ad0f7 DN |
4704 | } |
4705 | ||
40b178f4 | 4706 | /* Gate and execute functions for PRE. */ |
ff2ad0f7 | 4707 | |
b80280f2 | 4708 | static unsigned int |
40b178f4 | 4709 | do_pre (void) |
ff2ad0f7 | 4710 | { |
b80280f2 | 4711 | unsigned int todo = 0; |
83737db2 | 4712 | |
fa06ad0d JR |
4713 | do_partial_partial = |
4714 | flag_tree_partial_pre && optimize_function_for_speed_p (cfun); | |
ff2ad0f7 | 4715 | |
c9145754 DB |
4716 | /* This has to happen before SCCVN runs because |
4717 | loop_optimizer_init may create new phis, etc. */ | |
40b178f4 | 4718 | loop_optimizer_init (LOOPS_NORMAL); |
c90186eb | 4719 | |
40b178f4 | 4720 | if (!run_scc_vn (VN_WALK)) |
863d2a57 | 4721 | { |
40b178f4 | 4722 | loop_optimizer_finalize (); |
b80280f2 | 4723 | return 0; |
863d2a57 | 4724 | } |
a7d4562a | 4725 | |
40b178f4 | 4726 | init_pre (); |
a8338640 | 4727 | scev_initialize (); |
c9145754 | 4728 | |
c9145754 | 4729 | /* Collect and value number expressions computed in each basic block. */ |
665fcad8 | 4730 | compute_avail (); |
6de9cd9a | 4731 | |
7e6eb623 DB |
4732 | /* Insert can get quite slow on an incredibly large number of basic |
4733 | blocks due to some quadratic behavior. Until this behavior is | |
4734 | fixed, don't run it when he have an incredibly large number of | |
4735 | bb's. If we aren't going to run insert, there is no point in | |
4736 | computing ANTIC, either, even though it's plenty fast. */ | |
40b178f4 | 4737 | if (n_basic_blocks < 4000) |
6de9cd9a | 4738 | { |
7e6eb623 | 4739 | compute_antic (); |
7e6eb623 DB |
4740 | insert (); |
4741 | } | |
ff2ad0f7 | 4742 | |
0ab555de MM |
4743 | /* Make sure to remove fake edges before committing our inserts. |
4744 | This makes sure we don't end up with extra critical edges that | |
4745 | we would need to split. */ | |
4746 | remove_fake_exit_edges (); | |
4747 | gsi_commit_edge_inserts (); | |
4748 | ||
ff2ad0f7 | 4749 | /* Remove all the redundant expressions. */ |
b80280f2 | 4750 | todo |= eliminate (); |
0fc6c492 | 4751 | |
9fe0cb7d RG |
4752 | statistics_counter_event (cfun, "Insertions", pre_stats.insertions); |
4753 | statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert); | |
4754 | statistics_counter_event (cfun, "New PHIs", pre_stats.phis); | |
4755 | statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations); | |
c4ab2baa | 4756 | |
b71b4522 | 4757 | clear_expression_ids (); |
40b178f4 RG |
4758 | remove_dead_inserted_code (); |
4759 | todo |= TODO_verify_flow; | |
c90186eb | 4760 | |
a8338640 | 4761 | scev_finalize (); |
40b178f4 | 4762 | fini_pre (); |
9c370032 | 4763 | todo |= fini_eliminate (); |
40b178f4 RG |
4764 | loop_optimizer_finalize (); |
4765 | ||
4766 | /* TODO: tail_merge_optimize may merge all predecessors of a block, in which | |
4767 | case we can merge the block with the remaining predecessor of the block. | |
4768 | It should either: | |
4769 | - call merge_blocks after each tail merge iteration | |
4770 | - call merge_blocks after all tail merge iterations | |
4771 | - mark TODO_cleanup_cfg when necessary | |
4772 | - share the cfg cleanup with fini_pre. */ | |
4773 | todo |= tail_merge_optimize (todo); | |
4774 | ||
c9e93168 TV |
4775 | free_scc_vn (); |
4776 | ||
678771ad RG |
4777 | /* Tail merging invalidates the virtual SSA web, together with |
4778 | cfg-cleanup opportunities exposed by PRE this will wreck the | |
4779 | SSA updating machinery. So make sure to run update-ssa | |
4780 | manually, before eventually scheduling cfg-cleanup as part of | |
4781 | the todo. */ | |
4782 | update_ssa (TODO_update_ssa_only_virtuals); | |
4783 | ||
b80280f2 | 4784 | return todo; |
ff2ad0f7 DN |
4785 | } |
4786 | ||
6de9cd9a DN |
4787 | static bool |
4788 | gate_pre (void) | |
4789 | { | |
5813994e | 4790 | return flag_tree_pre != 0; |
6de9cd9a DN |
4791 | } |
4792 | ||
27a4cd48 DM |
4793 | namespace { |
4794 | ||
4795 | const pass_data pass_data_pre = | |
6de9cd9a | 4796 | { |
27a4cd48 DM |
4797 | GIMPLE_PASS, /* type */ |
4798 | "pre", /* name */ | |
4799 | OPTGROUP_NONE, /* optinfo_flags */ | |
4800 | true, /* has_gate */ | |
4801 | true, /* has_execute */ | |
4802 | TV_TREE_PRE, /* tv_id */ | |
4803 | ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */ | |
4804 | 0, /* properties_provided */ | |
4805 | 0, /* properties_destroyed */ | |
4806 | TODO_rebuild_alias, /* todo_flags_start */ | |
4807 | TODO_verify_ssa, /* todo_flags_finish */ | |
6de9cd9a | 4808 | }; |
ff2ad0f7 | 4809 | |
27a4cd48 DM |
4810 | class pass_pre : public gimple_opt_pass |
4811 | { | |
4812 | public: | |
4813 | pass_pre(gcc::context *ctxt) | |
4814 | : gimple_opt_pass(pass_data_pre, ctxt) | |
4815 | {} | |
4816 | ||
4817 | /* opt_pass methods: */ | |
4818 | bool gate () { return gate_pre (); } | |
4819 | unsigned int execute () { return do_pre (); } | |
4820 | ||
4821 | }; // class pass_pre | |
4822 | ||
4823 | } // anon namespace | |
4824 | ||
4825 | gimple_opt_pass * | |
4826 | make_pass_pre (gcc::context *ctxt) | |
4827 | { | |
4828 | return new pass_pre (ctxt); | |
4829 | } | |
4830 | ||
ff2ad0f7 DN |
4831 | |
4832 | /* Gate and execute functions for FRE. */ | |
4833 | ||
c2924966 | 4834 | static unsigned int |
b89361c6 | 4835 | execute_fre (void) |
ff2ad0f7 | 4836 | { |
40b178f4 RG |
4837 | unsigned int todo = 0; |
4838 | ||
4839 | if (!run_scc_vn (VN_WALKREWRITE)) | |
4840 | return 0; | |
4841 | ||
4842 | memset (&pre_stats, 0, sizeof (pre_stats)); | |
4843 | ||
4844 | /* Remove all the redundant expressions. */ | |
4845 | todo |= eliminate (); | |
4846 | ||
9c370032 | 4847 | todo |= fini_eliminate (); |
40b178f4 RG |
4848 | |
4849 | free_scc_vn (); | |
4850 | ||
4851 | statistics_counter_event (cfun, "Insertions", pre_stats.insertions); | |
4852 | statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations); | |
40b178f4 RG |
4853 | |
4854 | return todo; | |
ff2ad0f7 DN |
4855 | } |
4856 | ||
4857 | static bool | |
4858 | gate_fre (void) | |
4859 | { | |
4860 | return flag_tree_fre != 0; | |
4861 | } | |
4862 | ||
27a4cd48 DM |
4863 | namespace { |
4864 | ||
4865 | const pass_data pass_data_fre = | |
ff2ad0f7 | 4866 | { |
27a4cd48 DM |
4867 | GIMPLE_PASS, /* type */ |
4868 | "fre", /* name */ | |
4869 | OPTGROUP_NONE, /* optinfo_flags */ | |
4870 | true, /* has_gate */ | |
4871 | true, /* has_execute */ | |
4872 | TV_TREE_FRE, /* tv_id */ | |
4873 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
4874 | 0, /* properties_provided */ | |
4875 | 0, /* properties_destroyed */ | |
4876 | 0, /* todo_flags_start */ | |
4877 | TODO_verify_ssa, /* todo_flags_finish */ | |
ff2ad0f7 | 4878 | }; |
27a4cd48 DM |
4879 | |
4880 | class pass_fre : public gimple_opt_pass | |
4881 | { | |
4882 | public: | |
4883 | pass_fre(gcc::context *ctxt) | |
4884 | : gimple_opt_pass(pass_data_fre, ctxt) | |
4885 | {} | |
4886 | ||
4887 | /* opt_pass methods: */ | |
4888 | opt_pass * clone () { return new pass_fre (ctxt_); } | |
4889 | bool gate () { return gate_fre (); } | |
4890 | unsigned int execute () { return execute_fre (); } | |
4891 | ||
4892 | }; // class pass_fre | |
4893 | ||
4894 | } // anon namespace | |
4895 | ||
4896 | gimple_opt_pass * | |
4897 | make_pass_fre (gcc::context *ctxt) | |
4898 | { | |
4899 | return new pass_fre (ctxt); | |
4900 | } |