]>
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" |
726a989a | 30 | #include "gimple.h" |
442b4905 AM |
31 | #include "gimple-ssa.h" |
32 | #include "tree-cfg.h" | |
33 | #include "tree-phinodes.h" | |
34 | #include "ssa-iterators.h" | |
35 | #include "tree-ssanames.h" | |
36 | #include "tree-ssa-loop.h" | |
37 | #include "tree-into-ssa.h" | |
38 | #include "tree-dfa.h" | |
39 | #include "tree-ssa.h" | |
0823efed | 40 | #include "hash-table.h" |
6de9cd9a | 41 | #include "tree-iterator.h" |
6de9cd9a | 42 | #include "alloc-pool.h" |
5039610b | 43 | #include "obstack.h" |
6de9cd9a DN |
44 | #include "tree-pass.h" |
45 | #include "flags.h" | |
7e6eb623 | 46 | #include "langhooks.h" |
0fc6c492 | 47 | #include "cfgloop.h" |
89fb70a3 | 48 | #include "tree-ssa-sccvn.h" |
a8338640 | 49 | #include "tree-scalar-evolution.h" |
f0ed4cfb | 50 | #include "params.h" |
c9145754 | 51 | #include "dbgcnt.h" |
40b178f4 | 52 | #include "domwalk.h" |
e248d83f | 53 | #include "ipa-prop.h" |
744730a4 | 54 | #include "tree-ssa-propagate.h" |
33c94679 | 55 | |
7e6eb623 | 56 | /* TODO: |
b9c5e484 | 57 | |
bdee7684 | 58 | 1. Avail sets can be shared by making an avail_find_leader that |
7e6eb623 DB |
59 | walks up the dominator tree and looks in those avail sets. |
60 | This might affect code optimality, it's unclear right now. | |
c90186eb | 61 | 2. Strength reduction can be performed by anticipating expressions |
7e6eb623 | 62 | we can repair later on. |
c90186eb | 63 | 3. We can do back-substitution or smarter value numbering to catch |
0fc6c492 | 64 | commutative expressions split up over multiple statements. |
b9c5e484 | 65 | */ |
7e6eb623 DB |
66 | |
67 | /* For ease of terminology, "expression node" in the below refers to | |
726a989a | 68 | every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs |
07beea0d AH |
69 | represent the actual statement containing the expressions we care about, |
70 | and we cache the value number by putting it in the expression. */ | |
7e6eb623 DB |
71 | |
72 | /* Basic algorithm | |
b9c5e484 | 73 | |
56db793a DB |
74 | First we walk the statements to generate the AVAIL sets, the |
75 | EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the | |
76 | generation of values/expressions by a given block. We use them | |
77 | when computing the ANTIC sets. The AVAIL sets consist of | |
78 | SSA_NAME's that represent values, so we know what values are | |
79 | available in what blocks. AVAIL is a forward dataflow problem. In | |
80 | SSA, values are never killed, so we don't need a kill set, or a | |
81 | fixpoint iteration, in order to calculate the AVAIL sets. In | |
82 | traditional parlance, AVAIL sets tell us the downsafety of the | |
7e6eb623 | 83 | expressions/values. |
b9c5e484 | 84 | |
56db793a DB |
85 | Next, we generate the ANTIC sets. These sets represent the |
86 | anticipatable expressions. ANTIC is a backwards dataflow | |
d75dbccd | 87 | problem. An expression is anticipatable in a given block if it could |
56db793a DB |
88 | be generated in that block. This means that if we had to perform |
89 | an insertion in that block, of the value of that expression, we | |
90 | could. Calculating the ANTIC sets requires phi translation of | |
91 | expressions, because the flow goes backwards through phis. We must | |
92 | iterate to a fixpoint of the ANTIC sets, because we have a kill | |
93 | set. Even in SSA form, values are not live over the entire | |
94 | function, only from their definition point onwards. So we have to | |
95 | remove values from the ANTIC set once we go past the definition | |
96 | point of the leaders that make them up. | |
97 | compute_antic/compute_antic_aux performs this computation. | |
7e6eb623 DB |
98 | |
99 | Third, we perform insertions to make partially redundant | |
100 | expressions fully redundant. | |
101 | ||
102 | An expression is partially redundant (excluding partial | |
103 | anticipation) if: | |
104 | ||
105 | 1. It is AVAIL in some, but not all, of the predecessors of a | |
106 | given block. | |
107 | 2. It is ANTIC in all the predecessors. | |
108 | ||
109 | In order to make it fully redundant, we insert the expression into | |
110 | the predecessors where it is not available, but is ANTIC. | |
d75dbccd DB |
111 | |
112 | For the partial anticipation case, we only perform insertion if it | |
113 | is partially anticipated in some block, and fully available in all | |
114 | of the predecessors. | |
115 | ||
116 | insert/insert_aux/do_regular_insertion/do_partial_partial_insertion | |
117 | performs these steps. | |
7e6eb623 DB |
118 | |
119 | Fourth, we eliminate fully redundant expressions. | |
120 | This is a simple statement walk that replaces redundant | |
070b797d | 121 | calculations with the now available values. */ |
7e6eb623 DB |
122 | |
123 | /* Representations of value numbers: | |
124 | ||
c9145754 DB |
125 | Value numbers are represented by a representative SSA_NAME. We |
126 | will create fake SSA_NAME's in situations where we need a | |
127 | representative but do not have one (because it is a complex | |
128 | expression). In order to facilitate storing the value numbers in | |
129 | bitmaps, and keep the number of wasted SSA_NAME's down, we also | |
130 | associate a value_id with each value number, and create full blown | |
131 | ssa_name's only where we actually need them (IE in operands of | |
132 | existing expressions). | |
133 | ||
134 | Theoretically you could replace all the value_id's with | |
135 | SSA_NAME_VERSION, but this would allocate a large number of | |
136 | SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number. | |
137 | It would also require an additional indirection at each point we | |
138 | use the value id. */ | |
7e6eb623 | 139 | |
b9c5e484 | 140 | /* Representation of expressions on value numbers: |
7e6eb623 | 141 | |
249eb506 | 142 | Expressions consisting of value numbers are represented the same |
c9145754 DB |
143 | way as our VN internally represents them, with an additional |
144 | "pre_expr" wrapping around them in order to facilitate storing all | |
145 | of the expressions in the same sets. */ | |
7e6eb623 | 146 | |
c9145754 | 147 | /* Representation of sets: |
7e6eb623 | 148 | |
c9145754 DB |
149 | The dataflow sets do not need to be sorted in any particular order |
150 | for the majority of their lifetime, are simply represented as two | |
151 | bitmaps, one that keeps track of values present in the set, and one | |
152 | that keeps track of expressions present in the set. | |
7e6eb623 | 153 | |
c9145754 DB |
154 | When we need them in topological order, we produce it on demand by |
155 | transforming the bitmap into an array and sorting it into topo | |
156 | order. */ | |
7e6eb623 | 157 | |
c9145754 DB |
158 | /* Type of expression, used to know which member of the PRE_EXPR union |
159 | is valid. */ | |
7e6eb623 | 160 | |
c9145754 DB |
161 | enum pre_expr_kind |
162 | { | |
163 | NAME, | |
164 | NARY, | |
165 | REFERENCE, | |
166 | CONSTANT | |
167 | }; | |
168 | ||
169 | typedef union pre_expr_union_d | |
170 | { | |
171 | tree name; | |
172 | tree constant; | |
173 | vn_nary_op_t nary; | |
174 | vn_reference_t reference; | |
175 | } pre_expr_union; | |
b9c5e484 | 176 | |
5deac340 | 177 | typedef struct pre_expr_d : typed_noop_remove <pre_expr_d> |
c9145754 DB |
178 | { |
179 | enum pre_expr_kind kind; | |
180 | unsigned int id; | |
181 | pre_expr_union u; | |
5deac340 RG |
182 | |
183 | /* hash_table support. */ | |
5831a5f0 LC |
184 | typedef pre_expr_d value_type; |
185 | typedef pre_expr_d compare_type; | |
5deac340 RG |
186 | static inline hashval_t hash (const pre_expr_d *); |
187 | static inline int equal (const pre_expr_d *, const pre_expr_d *); | |
c9145754 | 188 | } *pre_expr; |
7e6eb623 | 189 | |
c9145754 DB |
190 | #define PRE_EXPR_NAME(e) (e)->u.name |
191 | #define PRE_EXPR_NARY(e) (e)->u.nary | |
192 | #define PRE_EXPR_REFERENCE(e) (e)->u.reference | |
193 | #define PRE_EXPR_CONSTANT(e) (e)->u.constant | |
7e6eb623 | 194 | |
0823efed | 195 | /* Compare E1 and E1 for equality. */ |
7e6eb623 | 196 | |
0823efed | 197 | inline int |
5831a5f0 | 198 | pre_expr_d::equal (const value_type *e1, const compare_type *e2) |
0823efed | 199 | { |
c9145754 DB |
200 | if (e1->kind != e2->kind) |
201 | return false; | |
202 | ||
203 | switch (e1->kind) | |
204 | { | |
205 | case CONSTANT: | |
726a989a RB |
206 | return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1), |
207 | PRE_EXPR_CONSTANT (e2)); | |
c9145754 DB |
208 | case NAME: |
209 | return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2); | |
210 | case NARY: | |
211 | return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2)); | |
212 | case REFERENCE: | |
213 | return vn_reference_eq (PRE_EXPR_REFERENCE (e1), | |
214 | PRE_EXPR_REFERENCE (e2)); | |
215 | default: | |
9708c51d | 216 | gcc_unreachable (); |
c9145754 DB |
217 | } |
218 | } | |
219 | ||
0823efed DN |
220 | /* Hash E. */ |
221 | ||
222 | inline hashval_t | |
5831a5f0 | 223 | pre_expr_d::hash (const value_type *e) |
c9145754 | 224 | { |
c9145754 DB |
225 | switch (e->kind) |
226 | { | |
227 | case CONSTANT: | |
726a989a | 228 | return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e)); |
c9145754 | 229 | case NAME: |
9708c51d | 230 | return SSA_NAME_VERSION (PRE_EXPR_NAME (e)); |
c9145754 | 231 | case NARY: |
85169114 | 232 | return PRE_EXPR_NARY (e)->hashcode; |
c9145754 | 233 | case REFERENCE: |
85169114 | 234 | return PRE_EXPR_REFERENCE (e)->hashcode; |
c9145754 | 235 | default: |
9708c51d | 236 | gcc_unreachable (); |
c9145754 DB |
237 | } |
238 | } | |
83737db2 | 239 | |
c9145754 DB |
240 | /* Next global expression id number. */ |
241 | static unsigned int next_expression_id; | |
070b797d | 242 | |
83737db2 | 243 | /* Mapping from expression to id number we can use in bitmap sets. */ |
9771b263 | 244 | static vec<pre_expr> expressions; |
5deac340 | 245 | static hash_table <pre_expr_d> expression_to_id; |
9771b263 | 246 | static vec<unsigned> name_to_id; |
81def1b7 | 247 | |
83737db2 DB |
248 | /* Allocate an expression id for EXPR. */ |
249 | ||
250 | static inline unsigned int | |
c9145754 | 251 | alloc_expression_id (pre_expr expr) |
6de9cd9a | 252 | { |
0823efed | 253 | struct pre_expr_d **slot; |
83737db2 DB |
254 | /* Make sure we won't overflow. */ |
255 | gcc_assert (next_expression_id + 1 > next_expression_id); | |
c9145754 | 256 | expr->id = next_expression_id++; |
9771b263 | 257 | expressions.safe_push (expr); |
13de9095 RG |
258 | if (expr->kind == NAME) |
259 | { | |
260 | unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr)); | |
9771b263 DN |
261 | /* vec::safe_grow_cleared allocates no headroom. Avoid frequent |
262 | re-allocations by using vec::reserve upfront. There is no | |
263 | vec::quick_grow_cleared unfortunately. */ | |
264 | unsigned old_len = name_to_id.length (); | |
265 | name_to_id.reserve (num_ssa_names - old_len); | |
266 | name_to_id.safe_grow_cleared (num_ssa_names); | |
267 | gcc_assert (name_to_id[version] == 0); | |
268 | name_to_id[version] = expr->id; | |
13de9095 RG |
269 | } |
270 | else | |
271 | { | |
0823efed | 272 | slot = expression_to_id.find_slot (expr, INSERT); |
13de9095 RG |
273 | gcc_assert (!*slot); |
274 | *slot = expr; | |
275 | } | |
83737db2 DB |
276 | return next_expression_id - 1; |
277 | } | |
278 | ||
279 | /* Return the expression id for tree EXPR. */ | |
280 | ||
281 | static inline unsigned int | |
c9145754 DB |
282 | get_expression_id (const pre_expr expr) |
283 | { | |
284 | return expr->id; | |
285 | } | |
286 | ||
287 | static inline unsigned int | |
288 | lookup_expression_id (const pre_expr expr) | |
7e6eb623 | 289 | { |
0823efed | 290 | struct pre_expr_d **slot; |
b71b4522 | 291 | |
13de9095 RG |
292 | if (expr->kind == NAME) |
293 | { | |
294 | unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr)); | |
9771b263 | 295 | if (name_to_id.length () <= version) |
13de9095 | 296 | return 0; |
9771b263 | 297 | return name_to_id[version]; |
13de9095 RG |
298 | } |
299 | else | |
300 | { | |
0823efed | 301 | slot = expression_to_id.find_slot (expr, NO_INSERT); |
13de9095 RG |
302 | if (!slot) |
303 | return 0; | |
304 | return ((pre_expr)*slot)->id; | |
305 | } | |
83737db2 | 306 | } |
b9c5e484 | 307 | |
83737db2 DB |
308 | /* Return the existing expression id for EXPR, or create one if one |
309 | does not exist yet. */ | |
b9c5e484 | 310 | |
83737db2 | 311 | static inline unsigned int |
c9145754 | 312 | get_or_alloc_expression_id (pre_expr expr) |
83737db2 | 313 | { |
c9145754 DB |
314 | unsigned int id = lookup_expression_id (expr); |
315 | if (id == 0) | |
83737db2 | 316 | return alloc_expression_id (expr); |
c9145754 | 317 | return expr->id = id; |
83737db2 DB |
318 | } |
319 | ||
320 | /* Return the expression that has expression id ID */ | |
321 | ||
c9145754 | 322 | static inline pre_expr |
83737db2 DB |
323 | expression_for_id (unsigned int id) |
324 | { | |
9771b263 | 325 | return expressions[id]; |
c5830edf DB |
326 | } |
327 | ||
b71b4522 DB |
328 | /* Free the expression id field in all of our expressions, |
329 | and then destroy the expressions array. */ | |
330 | ||
331 | static void | |
332 | clear_expression_ids (void) | |
333 | { | |
9771b263 | 334 | expressions.release (); |
c9145754 | 335 | } |
b71b4522 | 336 | |
c9145754 DB |
337 | static alloc_pool pre_expr_pool; |
338 | ||
339 | /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */ | |
340 | ||
341 | static pre_expr | |
342 | get_or_alloc_expr_for_name (tree name) | |
343 | { | |
5f5126d6 RG |
344 | struct pre_expr_d expr; |
345 | pre_expr result; | |
c9145754 DB |
346 | unsigned int result_id; |
347 | ||
5f5126d6 RG |
348 | expr.kind = NAME; |
349 | expr.id = 0; | |
350 | PRE_EXPR_NAME (&expr) = name; | |
351 | result_id = lookup_expression_id (&expr); | |
352 | if (result_id != 0) | |
353 | return expression_for_id (result_id); | |
354 | ||
355 | result = (pre_expr) pool_alloc (pre_expr_pool); | |
c9145754 | 356 | result->kind = NAME; |
c9145754 | 357 | PRE_EXPR_NAME (result) = name; |
5f5126d6 | 358 | alloc_expression_id (result); |
c9145754 | 359 | return result; |
b71b4522 DB |
360 | } |
361 | ||
bdee7684 | 362 | /* An unordered bitmap set. One bitmap tracks values, the other, |
8c27b7d4 | 363 | expressions. */ |
bdee7684 DB |
364 | typedef struct bitmap_set |
365 | { | |
5c72d561 JH |
366 | bitmap_head expressions; |
367 | bitmap_head values; | |
bdee7684 DB |
368 | } *bitmap_set_t; |
369 | ||
83737db2 | 370 | #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \ |
c3284718 | 371 | EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi)) |
c9145754 | 372 | |
85169114 | 373 | #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \ |
c3284718 | 374 | EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi)) |
85169114 | 375 | |
c9145754 | 376 | /* Mapping from value id to expressions with that value_id. */ |
9771b263 | 377 | static vec<bitmap> value_expressions; |
83737db2 | 378 | |
6b416da1 | 379 | /* Sets that we need to keep track of. */ |
83737db2 | 380 | typedef struct bb_bitmap_sets |
6de9cd9a | 381 | { |
ca072a31 DB |
382 | /* The EXP_GEN set, which represents expressions/values generated in |
383 | a basic block. */ | |
83737db2 | 384 | bitmap_set_t exp_gen; |
ca072a31 DB |
385 | |
386 | /* The PHI_GEN set, which represents PHI results generated in a | |
387 | basic block. */ | |
6b416da1 | 388 | bitmap_set_t phi_gen; |
ca072a31 | 389 | |
f6fe65dc | 390 | /* The TMP_GEN set, which represents results/temporaries generated |
ca072a31 | 391 | in a basic block. IE the LHS of an expression. */ |
6b416da1 | 392 | bitmap_set_t tmp_gen; |
ca072a31 DB |
393 | |
394 | /* The AVAIL_OUT set, which represents which values are available in | |
395 | a given basic block. */ | |
bdee7684 | 396 | bitmap_set_t avail_out; |
ca072a31 | 397 | |
c90186eb | 398 | /* The ANTIC_IN set, which represents which values are anticipatable |
ca072a31 | 399 | in a given basic block. */ |
83737db2 | 400 | bitmap_set_t antic_in; |
ca072a31 | 401 | |
d75dbccd DB |
402 | /* The PA_IN set, which represents which values are |
403 | partially anticipatable in a given basic block. */ | |
404 | bitmap_set_t pa_in; | |
405 | ||
ca072a31 DB |
406 | /* The NEW_SETS set, which is used during insertion to augment the |
407 | AVAIL_OUT set of blocks with the new insertions performed during | |
408 | the current iteration. */ | |
6b416da1 | 409 | bitmap_set_t new_sets; |
c90186eb | 410 | |
5006671f RG |
411 | /* A cache for value_dies_in_block_x. */ |
412 | bitmap expr_dies; | |
413 | ||
d75dbccd | 414 | /* True if we have visited this block during ANTIC calculation. */ |
a19eb9d2 | 415 | unsigned int visited : 1; |
d75dbccd DB |
416 | |
417 | /* True we have deferred processing this block during ANTIC | |
418 | calculation until its successor is processed. */ | |
419 | unsigned int deferred : 1; | |
a19eb9d2 RG |
420 | |
421 | /* True when the block contains a call that might not return. */ | |
422 | unsigned int contains_may_not_return_call : 1; | |
d75dbccd DB |
423 | } *bb_value_sets_t; |
424 | ||
425 | #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen | |
426 | #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen | |
427 | #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen | |
428 | #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out | |
429 | #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in | |
430 | #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in | |
d75dbccd | 431 | #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets |
5006671f RG |
432 | #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies |
433 | #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited | |
d75dbccd | 434 | #define BB_DEFERRED(BB) ((bb_value_sets_t) ((BB)->aux))->deferred |
a19eb9d2 | 435 | #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call |
d75dbccd | 436 | |
c9145754 | 437 | |
83737db2 DB |
438 | /* Basic block list in postorder. */ |
439 | static int *postorder; | |
585d0dc4 | 440 | static int postorder_num; |
6de9cd9a | 441 | |
ca072a31 DB |
442 | /* This structure is used to keep track of statistics on what |
443 | optimization PRE was able to perform. */ | |
7e6eb623 | 444 | static struct |
6de9cd9a | 445 | { |
ca072a31 | 446 | /* The number of RHS computations eliminated by PRE. */ |
7e6eb623 | 447 | int eliminations; |
ca072a31 DB |
448 | |
449 | /* The number of new expressions/temporaries generated by PRE. */ | |
7e6eb623 | 450 | int insertions; |
ca072a31 | 451 | |
d75dbccd DB |
452 | /* The number of inserts found due to partial anticipation */ |
453 | int pa_insert; | |
454 | ||
ca072a31 | 455 | /* The number of new PHI nodes added by PRE. */ |
7e6eb623 DB |
456 | int phis; |
457 | } pre_stats; | |
6de9cd9a | 458 | |
d75dbccd | 459 | static bool do_partial_partial; |
e076271b | 460 | static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int); |
c9145754 DB |
461 | static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr); |
462 | static void bitmap_value_replace_in_set (bitmap_set_t, pre_expr); | |
bdee7684 | 463 | static void bitmap_set_copy (bitmap_set_t, bitmap_set_t); |
c9145754 DB |
464 | static bool bitmap_set_contains_value (bitmap_set_t, unsigned int); |
465 | static void bitmap_insert_into_set (bitmap_set_t, pre_expr); | |
9708c51d RG |
466 | static void bitmap_insert_into_set_1 (bitmap_set_t, pre_expr, |
467 | unsigned int, bool); | |
bdee7684 | 468 | static bitmap_set_t bitmap_set_new (void); |
726a989a | 469 | static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *, |
e076271b RG |
470 | tree); |
471 | static tree find_or_generate_expression (basic_block, tree, gimple_seq *); | |
de081cfd | 472 | static unsigned int get_expr_value_id (pre_expr); |
6de9cd9a | 473 | |
7e6eb623 DB |
474 | /* We can add and remove elements and entries to and from sets |
475 | and hash tables, so we use alloc pools for them. */ | |
6de9cd9a | 476 | |
bdee7684 | 477 | static alloc_pool bitmap_set_pool; |
7932a3db | 478 | static bitmap_obstack grand_bitmap_obstack; |
6de9cd9a | 479 | |
30fd5881 | 480 | /* Set of blocks with statements that have had their EH properties changed. */ |
53b4bf74 DN |
481 | static bitmap need_eh_cleanup; |
482 | ||
30fd5881 EB |
483 | /* Set of blocks with statements that have had their AB properties changed. */ |
484 | static bitmap need_ab_cleanup; | |
485 | ||
7e6eb623 DB |
486 | /* A three tuple {e, pred, v} used to cache phi translations in the |
487 | phi_translate_table. */ | |
6de9cd9a | 488 | |
5deac340 | 489 | typedef struct expr_pred_trans_d : typed_free_remove<expr_pred_trans_d> |
6de9cd9a | 490 | { |
8c27b7d4 | 491 | /* The expression. */ |
c9145754 | 492 | pre_expr e; |
ca072a31 DB |
493 | |
494 | /* The predecessor block along which we translated the expression. */ | |
7e6eb623 | 495 | basic_block pred; |
ca072a31 DB |
496 | |
497 | /* The value that resulted from the translation. */ | |
c9145754 | 498 | pre_expr v; |
ca072a31 DB |
499 | |
500 | /* The hashcode for the expression, pred pair. This is cached for | |
501 | speed reasons. */ | |
7e6eb623 | 502 | hashval_t hashcode; |
5deac340 RG |
503 | |
504 | /* hash_table support. */ | |
5831a5f0 LC |
505 | typedef expr_pred_trans_d value_type; |
506 | typedef expr_pred_trans_d compare_type; | |
507 | static inline hashval_t hash (const value_type *); | |
508 | static inline int equal (const value_type *, const compare_type *); | |
7e6eb623 | 509 | } *expr_pred_trans_t; |
741ac903 | 510 | typedef const struct expr_pred_trans_d *const_expr_pred_trans_t; |
6de9cd9a | 511 | |
0823efed | 512 | inline hashval_t |
5deac340 | 513 | expr_pred_trans_d::hash (const expr_pred_trans_d *e) |
7e6eb623 | 514 | { |
5deac340 | 515 | return e->hashcode; |
7e6eb623 | 516 | } |
6de9cd9a | 517 | |
0823efed | 518 | inline int |
5831a5f0 LC |
519 | expr_pred_trans_d::equal (const value_type *ve1, |
520 | const compare_type *ve2) | |
7e6eb623 | 521 | { |
7e6eb623 DB |
522 | basic_block b1 = ve1->pred; |
523 | basic_block b2 = ve2->pred; | |
b9c5e484 | 524 | |
ca072a31 DB |
525 | /* If they are not translations for the same basic block, they can't |
526 | be equal. */ | |
7e6eb623 | 527 | if (b1 != b2) |
6de9cd9a | 528 | return false; |
5deac340 | 529 | return pre_expr_d::equal (ve1->e, ve2->e); |
6de9cd9a DN |
530 | } |
531 | ||
0823efed DN |
532 | /* The phi_translate_table caches phi translations for a given |
533 | expression and predecessor. */ | |
5deac340 | 534 | static hash_table <expr_pred_trans_d> phi_translate_table; |
0823efed | 535 | |
c9145754 | 536 | /* Add the tuple mapping from {expression E, basic block PRED} to |
984af6ac | 537 | the phi translation table and return whether it pre-existed. */ |
7e6eb623 | 538 | |
984af6ac RB |
539 | static inline bool |
540 | phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred) | |
7e6eb623 | 541 | { |
0823efed | 542 | expr_pred_trans_t *slot; |
984af6ac RB |
543 | expr_pred_trans_d tem; |
544 | hashval_t hash = iterative_hash_hashval_t (pre_expr_d::hash (e), | |
545 | pred->index); | |
546 | tem.e = e; | |
547 | tem.pred = pred; | |
548 | tem.hashcode = hash; | |
549 | slot = phi_translate_table.find_slot_with_hash (&tem, hash, INSERT); | |
550 | if (*slot) | |
551 | { | |
552 | *entry = *slot; | |
553 | return true; | |
554 | } | |
c9145754 | 555 | |
984af6ac RB |
556 | *entry = *slot = XNEW (struct expr_pred_trans_d); |
557 | (*entry)->e = e; | |
558 | (*entry)->pred = pred; | |
559 | (*entry)->hashcode = hash; | |
560 | return false; | |
6de9cd9a DN |
561 | } |
562 | ||
ff2ad0f7 | 563 | |
c9145754 | 564 | /* Add expression E to the expression set of value id V. */ |
33c94679 | 565 | |
6bcfb753 | 566 | static void |
c9145754 | 567 | add_to_value (unsigned int v, pre_expr e) |
7e6eb623 | 568 | { |
4d4b1b30 | 569 | bitmap set; |
c9145754 | 570 | |
4d4b1b30 | 571 | gcc_checking_assert (get_expr_value_id (e) == v); |
de081cfd | 572 | |
9771b263 | 573 | if (v >= value_expressions.length ()) |
c9145754 | 574 | { |
9771b263 | 575 | value_expressions.safe_grow_cleared (v + 1); |
c9145754 | 576 | } |
33c94679 | 577 | |
9771b263 | 578 | set = value_expressions[v]; |
c9145754 DB |
579 | if (!set) |
580 | { | |
4d4b1b30 | 581 | set = BITMAP_ALLOC (&grand_bitmap_obstack); |
9771b263 | 582 | value_expressions[v] = set; |
c9145754 | 583 | } |
33c94679 | 584 | |
4d4b1b30 | 585 | bitmap_set_bit (set, get_or_alloc_expression_id (e)); |
7e6eb623 | 586 | } |
6de9cd9a | 587 | |
bdee7684 DB |
588 | /* Create a new bitmap set and return it. */ |
589 | ||
b9c5e484 | 590 | static bitmap_set_t |
bdee7684 DB |
591 | bitmap_set_new (void) |
592 | { | |
e1111e8e | 593 | bitmap_set_t ret = (bitmap_set_t) pool_alloc (bitmap_set_pool); |
5c72d561 JH |
594 | bitmap_initialize (&ret->expressions, &grand_bitmap_obstack); |
595 | bitmap_initialize (&ret->values, &grand_bitmap_obstack); | |
bdee7684 DB |
596 | return ret; |
597 | } | |
598 | ||
c9145754 DB |
599 | /* Return the value id for a PRE expression EXPR. */ |
600 | ||
601 | static unsigned int | |
602 | get_expr_value_id (pre_expr expr) | |
603 | { | |
e3815735 | 604 | unsigned int id; |
c9145754 DB |
605 | switch (expr->kind) |
606 | { | |
607 | case CONSTANT: | |
5ba5e8ec | 608 | id = get_constant_value_id (PRE_EXPR_CONSTANT (expr)); |
e3815735 | 609 | break; |
c9145754 | 610 | case NAME: |
e3815735 RB |
611 | id = VN_INFO (PRE_EXPR_NAME (expr))->value_id; |
612 | break; | |
c9145754 | 613 | case NARY: |
e3815735 RB |
614 | id = PRE_EXPR_NARY (expr)->value_id; |
615 | break; | |
c9145754 | 616 | case REFERENCE: |
e3815735 RB |
617 | id = PRE_EXPR_REFERENCE (expr)->value_id; |
618 | break; | |
c9145754 DB |
619 | default: |
620 | gcc_unreachable (); | |
621 | } | |
e3815735 RB |
622 | /* ??? We cannot assert that expr has a value-id (it can be 0), because |
623 | we assign value-ids only to expressions that have a result | |
624 | in set_hashtable_value_ids. */ | |
625 | return id; | |
c9145754 DB |
626 | } |
627 | ||
40b178f4 RG |
628 | /* Return a SCCVN valnum (SSA name or constant) for the PRE value-id VAL. */ |
629 | ||
630 | static tree | |
631 | sccvn_valnum_from_value_id (unsigned int val) | |
632 | { | |
633 | bitmap_iterator bi; | |
634 | unsigned int i; | |
9771b263 | 635 | bitmap exprset = value_expressions[val]; |
40b178f4 RG |
636 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
637 | { | |
638 | pre_expr vexpr = expression_for_id (i); | |
639 | if (vexpr->kind == NAME) | |
640 | return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum; | |
641 | else if (vexpr->kind == CONSTANT) | |
642 | return PRE_EXPR_CONSTANT (vexpr); | |
643 | } | |
644 | return NULL_TREE; | |
645 | } | |
646 | ||
83737db2 | 647 | /* Remove an expression EXPR from a bitmapped set. */ |
bdee7684 DB |
648 | |
649 | static void | |
c9145754 | 650 | bitmap_remove_from_set (bitmap_set_t set, pre_expr expr) |
bdee7684 | 651 | { |
c9145754 DB |
652 | unsigned int val = get_expr_value_id (expr); |
653 | if (!value_id_constant_p (val)) | |
83737db2 | 654 | { |
5c72d561 JH |
655 | bitmap_clear_bit (&set->values, val); |
656 | bitmap_clear_bit (&set->expressions, get_expression_id (expr)); | |
83737db2 | 657 | } |
bdee7684 | 658 | } |
6de9cd9a | 659 | |
7e6eb623 | 660 | static void |
c9145754 | 661 | bitmap_insert_into_set_1 (bitmap_set_t set, pre_expr expr, |
9708c51d | 662 | unsigned int val, bool allow_constants) |
7e6eb623 | 663 | { |
c9145754 | 664 | if (allow_constants || !value_id_constant_p (val)) |
7e6eb623 | 665 | { |
c9145754 DB |
666 | /* We specifically expect this and only this function to be able to |
667 | insert constants into a set. */ | |
5c72d561 JH |
668 | bitmap_set_bit (&set->values, val); |
669 | bitmap_set_bit (&set->expressions, get_or_alloc_expression_id (expr)); | |
7e6eb623 DB |
670 | } |
671 | } | |
6de9cd9a | 672 | |
c9145754 DB |
673 | /* Insert an expression EXPR into a bitmapped set. */ |
674 | ||
675 | static void | |
676 | bitmap_insert_into_set (bitmap_set_t set, pre_expr expr) | |
677 | { | |
9708c51d | 678 | bitmap_insert_into_set_1 (set, expr, get_expr_value_id (expr), false); |
c9145754 DB |
679 | } |
680 | ||
bdee7684 DB |
681 | /* Copy a bitmapped set ORIG, into bitmapped set DEST. */ |
682 | ||
683 | static void | |
684 | bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig) | |
685 | { | |
5c72d561 JH |
686 | bitmap_copy (&dest->expressions, &orig->expressions); |
687 | bitmap_copy (&dest->values, &orig->values); | |
bdee7684 DB |
688 | } |
689 | ||
ff3fdad2 | 690 | |
83737db2 | 691 | /* Free memory used up by SET. */ |
ff3fdad2 | 692 | static void |
83737db2 | 693 | bitmap_set_free (bitmap_set_t set) |
ff3fdad2 | 694 | { |
5c72d561 JH |
695 | bitmap_clear (&set->expressions); |
696 | bitmap_clear (&set->values); | |
ff3fdad2 DB |
697 | } |
698 | ||
ff3fdad2 | 699 | |
83737db2 | 700 | /* Generate an topological-ordered array of bitmap set SET. */ |
ff3fdad2 | 701 | |
9771b263 | 702 | static vec<pre_expr> |
83737db2 | 703 | sorted_array_from_bitmap_set (bitmap_set_t set) |
ff3fdad2 | 704 | { |
85169114 PB |
705 | unsigned int i, j; |
706 | bitmap_iterator bi, bj; | |
9771b263 | 707 | vec<pre_expr> result; |
a7d04a53 RG |
708 | |
709 | /* Pre-allocate roughly enough space for the array. */ | |
9771b263 | 710 | result.create (bitmap_count_bits (&set->values)); |
ff3fdad2 | 711 | |
85169114 PB |
712 | FOR_EACH_VALUE_ID_IN_SET (set, i, bi) |
713 | { | |
714 | /* The number of expressions having a given value is usually | |
715 | relatively small. Thus, rather than making a vector of all | |
716 | the expressions and sorting it by value-id, we walk the values | |
717 | and check in the reverse mapping that tells us what expressions | |
718 | have a given value, to filter those in our set. As a result, | |
719 | the expressions are inserted in value-id order, which means | |
720 | topological order. | |
721 | ||
722 | If this is somehow a significant lose for some cases, we can | |
723 | choose which set to walk based on the set size. */ | |
9771b263 | 724 | bitmap exprset = value_expressions[i]; |
4d4b1b30 | 725 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, j, bj) |
85169114 | 726 | { |
5c72d561 | 727 | if (bitmap_bit_p (&set->expressions, j)) |
9771b263 | 728 | result.safe_push (expression_for_id (j)); |
85169114 PB |
729 | } |
730 | } | |
6de9cd9a | 731 | |
83737db2 | 732 | return result; |
7e6eb623 | 733 | } |
6de9cd9a | 734 | |
83737db2 | 735 | /* Perform bitmapped set operation DEST &= ORIG. */ |
6de9cd9a DN |
736 | |
737 | static void | |
83737db2 | 738 | bitmap_set_and (bitmap_set_t dest, bitmap_set_t orig) |
6de9cd9a | 739 | { |
83737db2 DB |
740 | bitmap_iterator bi; |
741 | unsigned int i; | |
6de9cd9a | 742 | |
d75dbccd | 743 | if (dest != orig) |
83737db2 | 744 | { |
5c72d561 JH |
745 | bitmap_head temp; |
746 | bitmap_initialize (&temp, &grand_bitmap_obstack); | |
89fb70a3 | 747 | |
5c72d561 JH |
748 | bitmap_and_into (&dest->values, &orig->values); |
749 | bitmap_copy (&temp, &dest->expressions); | |
750 | EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi) | |
d75dbccd | 751 | { |
c9145754 DB |
752 | pre_expr expr = expression_for_id (i); |
753 | unsigned int value_id = get_expr_value_id (expr); | |
5c72d561 JH |
754 | if (!bitmap_bit_p (&dest->values, value_id)) |
755 | bitmap_clear_bit (&dest->expressions, i); | |
d75dbccd | 756 | } |
5c72d561 | 757 | bitmap_clear (&temp); |
6de9cd9a DN |
758 | } |
759 | } | |
760 | ||
83737db2 | 761 | /* Subtract all values and expressions contained in ORIG from DEST. */ |
7e6eb623 | 762 | |
83737db2 DB |
763 | static bitmap_set_t |
764 | bitmap_set_subtract (bitmap_set_t dest, bitmap_set_t orig) | |
7e6eb623 | 765 | { |
83737db2 DB |
766 | bitmap_set_t result = bitmap_set_new (); |
767 | bitmap_iterator bi; | |
768 | unsigned int i; | |
b9c5e484 | 769 | |
5c72d561 JH |
770 | bitmap_and_compl (&result->expressions, &dest->expressions, |
771 | &orig->expressions); | |
6de9cd9a | 772 | |
83737db2 DB |
773 | FOR_EACH_EXPR_ID_IN_SET (result, i, bi) |
774 | { | |
c9145754 DB |
775 | pre_expr expr = expression_for_id (i); |
776 | unsigned int value_id = get_expr_value_id (expr); | |
5c72d561 | 777 | bitmap_set_bit (&result->values, value_id); |
83737db2 | 778 | } |
af75a7ea | 779 | |
83737db2 | 780 | return result; |
6b416da1 DB |
781 | } |
782 | ||
d75dbccd DB |
783 | /* Subtract all the values in bitmap set B from bitmap set A. */ |
784 | ||
785 | static void | |
786 | bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b) | |
787 | { | |
788 | unsigned int i; | |
789 | bitmap_iterator bi; | |
5c72d561 | 790 | bitmap_head temp; |
d75dbccd | 791 | |
5c72d561 JH |
792 | bitmap_initialize (&temp, &grand_bitmap_obstack); |
793 | ||
794 | bitmap_copy (&temp, &a->expressions); | |
795 | EXECUTE_IF_SET_IN_BITMAP (&temp, 0, i, bi) | |
d75dbccd | 796 | { |
c9145754 DB |
797 | pre_expr expr = expression_for_id (i); |
798 | if (bitmap_set_contains_value (b, get_expr_value_id (expr))) | |
d75dbccd DB |
799 | bitmap_remove_from_set (a, expr); |
800 | } | |
5c72d561 | 801 | bitmap_clear (&temp); |
d75dbccd DB |
802 | } |
803 | ||
804 | ||
c9145754 | 805 | /* Return true if bitmapped set SET contains the value VALUE_ID. */ |
6de9cd9a | 806 | |
bdee7684 | 807 | static bool |
c9145754 | 808 | bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id) |
6de9cd9a | 809 | { |
c9145754 | 810 | if (value_id_constant_p (value_id)) |
bdee7684 | 811 | return true; |
83737db2 | 812 | |
5c72d561 | 813 | if (!set || bitmap_empty_p (&set->expressions)) |
83737db2 DB |
814 | return false; |
815 | ||
5c72d561 | 816 | return bitmap_bit_p (&set->values, value_id); |
bdee7684 | 817 | } |
7e6eb623 | 818 | |
d75dbccd | 819 | static inline bool |
c9145754 | 820 | bitmap_set_contains_expr (bitmap_set_t set, const pre_expr expr) |
d75dbccd | 821 | { |
5c72d561 | 822 | return bitmap_bit_p (&set->expressions, get_expression_id (expr)); |
d75dbccd DB |
823 | } |
824 | ||
bdee7684 | 825 | /* Replace an instance of value LOOKFOR with expression EXPR in SET. */ |
7e6eb623 | 826 | |
bdee7684 | 827 | static void |
c9145754 DB |
828 | bitmap_set_replace_value (bitmap_set_t set, unsigned int lookfor, |
829 | const pre_expr expr) | |
bdee7684 | 830 | { |
4d4b1b30 | 831 | bitmap exprset; |
83737db2 DB |
832 | unsigned int i; |
833 | bitmap_iterator bi; | |
834 | ||
c9145754 | 835 | if (value_id_constant_p (lookfor)) |
bdee7684 | 836 | return; |
83737db2 | 837 | |
bdee7684 DB |
838 | if (!bitmap_set_contains_value (set, lookfor)) |
839 | return; | |
e9284566 | 840 | |
6b416da1 DB |
841 | /* The number of expressions having a given value is usually |
842 | significantly less than the total number of expressions in SET. | |
843 | Thus, rather than check, for each expression in SET, whether it | |
844 | has the value LOOKFOR, we walk the reverse mapping that tells us | |
845 | what expressions have a given value, and see if any of those | |
846 | expressions are in our set. For large testcases, this is about | |
847 | 5-10x faster than walking the bitmap. If this is somehow a | |
848 | significant lose for some cases, we can choose which set to walk | |
849 | based on the set size. */ | |
9771b263 | 850 | exprset = value_expressions[lookfor]; |
4d4b1b30 | 851 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
6de9cd9a | 852 | { |
fcaa4ca4 | 853 | if (bitmap_clear_bit (&set->expressions, i)) |
6de9cd9a | 854 | { |
5c72d561 | 855 | bitmap_set_bit (&set->expressions, get_expression_id (expr)); |
83737db2 | 856 | return; |
6de9cd9a | 857 | } |
6de9cd9a | 858 | } |
4bcc5786 RB |
859 | |
860 | gcc_unreachable (); | |
6de9cd9a DN |
861 | } |
862 | ||
83737db2 | 863 | /* Return true if two bitmap sets are equal. */ |
6de9cd9a | 864 | |
7e6eb623 | 865 | static bool |
83737db2 | 866 | bitmap_set_equal (bitmap_set_t a, bitmap_set_t b) |
6de9cd9a | 867 | { |
5c72d561 | 868 | return bitmap_equal_p (&a->values, &b->values); |
6de9cd9a DN |
869 | } |
870 | ||
e9284566 | 871 | /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists, |
6c6cfbfd | 872 | and add it otherwise. */ |
bdee7684 | 873 | |
7e6eb623 | 874 | static void |
c9145754 | 875 | bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr) |
7e6eb623 | 876 | { |
c9145754 | 877 | unsigned int val = get_expr_value_id (expr); |
83737db2 | 878 | |
e9284566 DB |
879 | if (bitmap_set_contains_value (set, val)) |
880 | bitmap_set_replace_value (set, val, expr); | |
881 | else | |
882 | bitmap_insert_into_set (set, expr); | |
bdee7684 | 883 | } |
7e6eb623 | 884 | |
bdee7684 DB |
885 | /* Insert EXPR into SET if EXPR's value is not already present in |
886 | SET. */ | |
887 | ||
888 | static void | |
c9145754 | 889 | bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr) |
bdee7684 | 890 | { |
c9145754 | 891 | unsigned int val = get_expr_value_id (expr); |
af75a7ea | 892 | |
77a74ed7 | 893 | gcc_checking_assert (expr->id == get_or_alloc_expression_id (expr)); |
1befacc8 RG |
894 | |
895 | /* Constant values are always considered to be part of the set. */ | |
896 | if (value_id_constant_p (val)) | |
897 | return; | |
898 | ||
899 | /* If the value membership changed, add the expression. */ | |
5c72d561 JH |
900 | if (bitmap_set_bit (&set->values, val)) |
901 | bitmap_set_bit (&set->expressions, expr->id); | |
7e6eb623 | 902 | } |
6de9cd9a | 903 | |
c9145754 DB |
904 | /* Print out EXPR to outfile. */ |
905 | ||
906 | static void | |
907 | print_pre_expr (FILE *outfile, const pre_expr expr) | |
908 | { | |
909 | switch (expr->kind) | |
910 | { | |
911 | case CONSTANT: | |
912 | print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr), 0); | |
913 | break; | |
914 | case NAME: | |
915 | print_generic_expr (outfile, PRE_EXPR_NAME (expr), 0); | |
916 | break; | |
917 | case NARY: | |
918 | { | |
919 | unsigned int i; | |
920 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); | |
5806f481 | 921 | fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode)); |
c9145754 DB |
922 | for (i = 0; i < nary->length; i++) |
923 | { | |
924 | print_generic_expr (outfile, nary->op[i], 0); | |
925 | if (i != (unsigned) nary->length - 1) | |
926 | fprintf (outfile, ","); | |
927 | } | |
928 | fprintf (outfile, "}"); | |
929 | } | |
930 | break; | |
931 | ||
932 | case REFERENCE: | |
933 | { | |
934 | vn_reference_op_t vro; | |
935 | unsigned int i; | |
936 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); | |
937 | fprintf (outfile, "{"); | |
938 | for (i = 0; | |
9771b263 | 939 | ref->operands.iterate (i, &vro); |
c9145754 DB |
940 | i++) |
941 | { | |
5006671f | 942 | bool closebrace = false; |
c9145754 DB |
943 | if (vro->opcode != SSA_NAME |
944 | && TREE_CODE_CLASS (vro->opcode) != tcc_declaration) | |
5006671f | 945 | { |
5806f481 | 946 | fprintf (outfile, "%s", get_tree_code_name (vro->opcode)); |
5006671f RG |
947 | if (vro->op0) |
948 | { | |
949 | fprintf (outfile, "<"); | |
950 | closebrace = true; | |
951 | } | |
952 | } | |
c9145754 DB |
953 | if (vro->op0) |
954 | { | |
c9145754 DB |
955 | print_generic_expr (outfile, vro->op0, 0); |
956 | if (vro->op1) | |
957 | { | |
958 | fprintf (outfile, ","); | |
959 | print_generic_expr (outfile, vro->op1, 0); | |
960 | } | |
5006671f RG |
961 | if (vro->op2) |
962 | { | |
963 | fprintf (outfile, ","); | |
964 | print_generic_expr (outfile, vro->op2, 0); | |
965 | } | |
c9145754 | 966 | } |
5006671f RG |
967 | if (closebrace) |
968 | fprintf (outfile, ">"); | |
9771b263 | 969 | if (i != ref->operands.length () - 1) |
c9145754 DB |
970 | fprintf (outfile, ","); |
971 | } | |
972 | fprintf (outfile, "}"); | |
5006671f RG |
973 | if (ref->vuse) |
974 | { | |
975 | fprintf (outfile, "@"); | |
976 | print_generic_expr (outfile, ref->vuse, 0); | |
977 | } | |
c9145754 DB |
978 | } |
979 | break; | |
980 | } | |
981 | } | |
982 | void debug_pre_expr (pre_expr); | |
983 | ||
984 | /* Like print_pre_expr but always prints to stderr. */ | |
24e47c76 | 985 | DEBUG_FUNCTION void |
c9145754 DB |
986 | debug_pre_expr (pre_expr e) |
987 | { | |
988 | print_pre_expr (stderr, e); | |
989 | fprintf (stderr, "\n"); | |
990 | } | |
991 | ||
bdee7684 DB |
992 | /* Print out SET to OUTFILE. */ |
993 | ||
994 | static void | |
83737db2 DB |
995 | print_bitmap_set (FILE *outfile, bitmap_set_t set, |
996 | const char *setname, int blockindex) | |
bdee7684 DB |
997 | { |
998 | fprintf (outfile, "%s[%d] := { ", setname, blockindex); | |
999 | if (set) | |
1000 | { | |
cf6b9ef1 | 1001 | bool first = true; |
3cd8c58a | 1002 | unsigned i; |
87c476a2 ZD |
1003 | bitmap_iterator bi; |
1004 | ||
83737db2 | 1005 | FOR_EACH_EXPR_ID_IN_SET (set, i, bi) |
87c476a2 | 1006 | { |
c9145754 | 1007 | const pre_expr expr = expression_for_id (i); |
83737db2 | 1008 | |
65a6f342 NS |
1009 | if (!first) |
1010 | fprintf (outfile, ", "); | |
1011 | first = false; | |
c9145754 | 1012 | print_pre_expr (outfile, expr); |
b9c5e484 | 1013 | |
c9145754 | 1014 | fprintf (outfile, " (%04d)", get_expr_value_id (expr)); |
87c476a2 | 1015 | } |
bdee7684 DB |
1016 | } |
1017 | fprintf (outfile, " }\n"); | |
1018 | } | |
6de9cd9a | 1019 | |
83737db2 | 1020 | void debug_bitmap_set (bitmap_set_t); |
6de9cd9a | 1021 | |
24e47c76 | 1022 | DEBUG_FUNCTION void |
83737db2 DB |
1023 | debug_bitmap_set (bitmap_set_t set) |
1024 | { | |
1025 | print_bitmap_set (stderr, set, "debug", 0); | |
7e6eb623 DB |
1026 | } |
1027 | ||
19372838 RG |
1028 | void debug_bitmap_sets_for (basic_block); |
1029 | ||
1030 | DEBUG_FUNCTION void | |
1031 | debug_bitmap_sets_for (basic_block bb) | |
1032 | { | |
1033 | print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index); | |
40b178f4 RG |
1034 | print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index); |
1035 | print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index); | |
1036 | print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index); | |
1037 | print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index); | |
1038 | if (do_partial_partial) | |
1039 | print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index); | |
1040 | print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index); | |
19372838 RG |
1041 | } |
1042 | ||
7e6eb623 | 1043 | /* Print out the expressions that have VAL to OUTFILE. */ |
33c94679 | 1044 | |
6bcfb753 | 1045 | static void |
c9145754 | 1046 | print_value_expressions (FILE *outfile, unsigned int val) |
7e6eb623 | 1047 | { |
9771b263 | 1048 | bitmap set = value_expressions[val]; |
c9145754 | 1049 | if (set) |
33c94679 | 1050 | { |
4d4b1b30 | 1051 | bitmap_set x; |
33c94679 | 1052 | char s[10]; |
c9145754 | 1053 | sprintf (s, "%04d", val); |
4d4b1b30 RG |
1054 | x.expressions = *set; |
1055 | print_bitmap_set (outfile, &x, s, 0); | |
33c94679 | 1056 | } |
6de9cd9a DN |
1057 | } |
1058 | ||
6de9cd9a | 1059 | |
24e47c76 | 1060 | DEBUG_FUNCTION void |
c9145754 | 1061 | debug_value_expressions (unsigned int val) |
6de9cd9a | 1062 | { |
7e6eb623 DB |
1063 | print_value_expressions (stderr, val); |
1064 | } | |
1065 | ||
c9145754 DB |
1066 | /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to |
1067 | represent it. */ | |
0995a441 | 1068 | |
c9145754 DB |
1069 | static pre_expr |
1070 | get_or_alloc_expr_for_constant (tree constant) | |
b9c5e484 | 1071 | { |
c9145754 DB |
1072 | unsigned int result_id; |
1073 | unsigned int value_id; | |
5f5126d6 RG |
1074 | struct pre_expr_d expr; |
1075 | pre_expr newexpr; | |
1076 | ||
1077 | expr.kind = CONSTANT; | |
1078 | PRE_EXPR_CONSTANT (&expr) = constant; | |
1079 | result_id = lookup_expression_id (&expr); | |
1080 | if (result_id != 0) | |
1081 | return expression_for_id (result_id); | |
1082 | ||
1083 | newexpr = (pre_expr) pool_alloc (pre_expr_pool); | |
c9145754 DB |
1084 | newexpr->kind = CONSTANT; |
1085 | PRE_EXPR_CONSTANT (newexpr) = constant; | |
5f5126d6 | 1086 | alloc_expression_id (newexpr); |
c9145754 | 1087 | value_id = get_or_alloc_constant_value_id (constant); |
c9145754 DB |
1088 | add_to_value (value_id, newexpr); |
1089 | return newexpr; | |
0995a441 SB |
1090 | } |
1091 | ||
c9145754 DB |
1092 | /* Given a value id V, find the actual tree representing the constant |
1093 | value if there is one, and return it. Return NULL if we can't find | |
1094 | a constant. */ | |
43da81be DB |
1095 | |
1096 | static tree | |
726a989a | 1097 | get_constant_for_value_id (unsigned int v) |
43da81be | 1098 | { |
c9145754 DB |
1099 | if (value_id_constant_p (v)) |
1100 | { | |
1101 | unsigned int i; | |
1102 | bitmap_iterator bi; | |
9771b263 | 1103 | bitmap exprset = value_expressions[v]; |
c9145754 | 1104 | |
4d4b1b30 | 1105 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
c9145754 DB |
1106 | { |
1107 | pre_expr expr = expression_for_id (i); | |
726a989a | 1108 | if (expr->kind == CONSTANT) |
c9145754 DB |
1109 | return PRE_EXPR_CONSTANT (expr); |
1110 | } | |
1111 | } | |
1112 | return NULL; | |
1113 | } | |
1114 | ||
1115 | /* Get or allocate a pre_expr for a piece of GIMPLE, and return it. | |
1116 | Currently only supports constants and SSA_NAMES. */ | |
1117 | static pre_expr | |
1118 | get_or_alloc_expr_for (tree t) | |
1119 | { | |
1120 | if (TREE_CODE (t) == SSA_NAME) | |
1121 | return get_or_alloc_expr_for_name (t); | |
1d65f45c | 1122 | else if (is_gimple_min_invariant (t)) |
c9145754 | 1123 | return get_or_alloc_expr_for_constant (t); |
726a989a RB |
1124 | else |
1125 | { | |
1126 | /* More complex expressions can result from SCCVN expression | |
1127 | simplification that inserts values for them. As they all | |
1128 | do not have VOPs the get handled by the nary ops struct. */ | |
1129 | vn_nary_op_t result; | |
1130 | unsigned int result_id; | |
1131 | vn_nary_op_lookup (t, &result); | |
1132 | if (result != NULL) | |
1133 | { | |
1134 | pre_expr e = (pre_expr) pool_alloc (pre_expr_pool); | |
1135 | e->kind = NARY; | |
1136 | PRE_EXPR_NARY (e) = result; | |
1137 | result_id = lookup_expression_id (e); | |
1138 | if (result_id != 0) | |
1139 | { | |
1140 | pool_free (pre_expr_pool, e); | |
1141 | e = expression_for_id (result_id); | |
1142 | return e; | |
1143 | } | |
1144 | alloc_expression_id (e); | |
1145 | return e; | |
1146 | } | |
1147 | } | |
c9145754 DB |
1148 | return NULL; |
1149 | } | |
1150 | ||
1151 | /* Return the folded version of T if T, when folded, is a gimple | |
1152 | min_invariant. Otherwise, return T. */ | |
1153 | ||
1154 | static pre_expr | |
1155 | fully_constant_expression (pre_expr e) | |
1156 | { | |
1157 | switch (e->kind) | |
1158 | { | |
1159 | case CONSTANT: | |
1160 | return e; | |
1161 | case NARY: | |
1162 | { | |
1163 | vn_nary_op_t nary = PRE_EXPR_NARY (e); | |
1164 | switch (TREE_CODE_CLASS (nary->opcode)) | |
1165 | { | |
1166 | case tcc_binary: | |
40f64141 | 1167 | case tcc_comparison: |
c9145754 DB |
1168 | { |
1169 | /* We have to go from trees to pre exprs to value ids to | |
1170 | constants. */ | |
1171 | tree naryop0 = nary->op[0]; | |
1172 | tree naryop1 = nary->op[1]; | |
40f64141 RG |
1173 | tree result; |
1174 | if (!is_gimple_min_invariant (naryop0)) | |
726a989a RB |
1175 | { |
1176 | pre_expr rep0 = get_or_alloc_expr_for (naryop0); | |
1177 | unsigned int vrep0 = get_expr_value_id (rep0); | |
40f64141 RG |
1178 | tree const0 = get_constant_for_value_id (vrep0); |
1179 | if (const0) | |
1180 | naryop0 = fold_convert (TREE_TYPE (naryop0), const0); | |
726a989a | 1181 | } |
40f64141 | 1182 | if (!is_gimple_min_invariant (naryop1)) |
726a989a RB |
1183 | { |
1184 | pre_expr rep1 = get_or_alloc_expr_for (naryop1); | |
1185 | unsigned int vrep1 = get_expr_value_id (rep1); | |
40f64141 RG |
1186 | tree const1 = get_constant_for_value_id (vrep1); |
1187 | if (const1) | |
1188 | naryop1 = fold_convert (TREE_TYPE (naryop1), const1); | |
b463e8de | 1189 | } |
40f64141 RG |
1190 | result = fold_binary (nary->opcode, nary->type, |
1191 | naryop0, naryop1); | |
c9145754 DB |
1192 | if (result && is_gimple_min_invariant (result)) |
1193 | return get_or_alloc_expr_for_constant (result); | |
40f64141 RG |
1194 | /* We might have simplified the expression to a |
1195 | SSA_NAME for example from x_1 * 1. But we cannot | |
1196 | insert a PHI for x_1 unconditionally as x_1 might | |
1197 | not be available readily. */ | |
c9145754 DB |
1198 | return e; |
1199 | } | |
40f64141 RG |
1200 | case tcc_reference: |
1201 | if (nary->opcode != REALPART_EXPR | |
b8698a0f | 1202 | && nary->opcode != IMAGPART_EXPR |
40f64141 RG |
1203 | && nary->opcode != VIEW_CONVERT_EXPR) |
1204 | return e; | |
1205 | /* Fallthrough. */ | |
c9145754 DB |
1206 | case tcc_unary: |
1207 | { | |
40f64141 RG |
1208 | /* We have to go from trees to pre exprs to value ids to |
1209 | constants. */ | |
c9145754 | 1210 | tree naryop0 = nary->op[0]; |
726a989a RB |
1211 | tree const0, result; |
1212 | if (is_gimple_min_invariant (naryop0)) | |
1213 | const0 = naryop0; | |
1214 | else | |
1215 | { | |
1216 | pre_expr rep0 = get_or_alloc_expr_for (naryop0); | |
1217 | unsigned int vrep0 = get_expr_value_id (rep0); | |
1218 | const0 = get_constant_for_value_id (vrep0); | |
1219 | } | |
1220 | result = NULL; | |
c9145754 | 1221 | if (const0) |
b463e8de DB |
1222 | { |
1223 | tree type1 = TREE_TYPE (nary->op[0]); | |
1224 | const0 = fold_convert (type1, const0); | |
1225 | result = fold_unary (nary->opcode, nary->type, const0); | |
1226 | } | |
c9145754 DB |
1227 | if (result && is_gimple_min_invariant (result)) |
1228 | return get_or_alloc_expr_for_constant (result); | |
1229 | return e; | |
1230 | } | |
1231 | default: | |
1232 | return e; | |
1233 | } | |
1234 | } | |
47af7a5c RG |
1235 | case REFERENCE: |
1236 | { | |
1237 | vn_reference_t ref = PRE_EXPR_REFERENCE (e); | |
12bd5a1e RG |
1238 | tree folded; |
1239 | if ((folded = fully_constant_vn_reference_p (ref))) | |
1240 | return get_or_alloc_expr_for_constant (folded); | |
1241 | return e; | |
1242 | } | |
c9145754 DB |
1243 | default: |
1244 | return e; | |
1245 | } | |
1246 | return e; | |
43da81be DB |
1247 | } |
1248 | ||
5006671f | 1249 | /* Translate the VUSE backwards through phi nodes in PHIBLOCK, so that |
84280917 MM |
1250 | it has the value it would have in BLOCK. Set *SAME_VALID to true |
1251 | in case the new vuse doesn't change the value id of the OPERANDS. */ | |
c90186eb | 1252 | |
5006671f | 1253 | static tree |
9771b263 | 1254 | translate_vuse_through_block (vec<vn_reference_op_s> operands, |
b45d2719 | 1255 | alias_set_type set, tree type, tree vuse, |
5006671f | 1256 | basic_block phiblock, |
84280917 | 1257 | basic_block block, bool *same_valid) |
c90186eb | 1258 | { |
5006671f | 1259 | gimple phi = SSA_NAME_DEF_STMT (vuse); |
b45d2719 | 1260 | ao_ref ref; |
84280917 MM |
1261 | edge e = NULL; |
1262 | bool use_oracle; | |
1263 | ||
1264 | *same_valid = true; | |
43da81be | 1265 | |
5006671f RG |
1266 | if (gimple_bb (phi) != phiblock) |
1267 | return vuse; | |
1268 | ||
84280917 | 1269 | use_oracle = ao_ref_init_from_vn_reference (&ref, set, type, operands); |
5006671f RG |
1270 | |
1271 | /* Use the alias-oracle to find either the PHI node in this block, | |
1272 | the first VUSE used in this block that is equivalent to vuse or | |
1273 | the first VUSE which definition in this block kills the value. */ | |
84280917 MM |
1274 | if (gimple_code (phi) == GIMPLE_PHI) |
1275 | e = find_edge (block, phiblock); | |
1276 | else if (use_oracle) | |
1277 | while (!stmt_may_clobber_ref_p_1 (phi, &ref)) | |
1278 | { | |
1279 | vuse = gimple_vuse (phi); | |
1280 | phi = SSA_NAME_DEF_STMT (vuse); | |
1281 | if (gimple_bb (phi) != phiblock) | |
1282 | return vuse; | |
1283 | if (gimple_code (phi) == GIMPLE_PHI) | |
1284 | { | |
1285 | e = find_edge (block, phiblock); | |
1286 | break; | |
1287 | } | |
1288 | } | |
1289 | else | |
1290 | return NULL_TREE; | |
1291 | ||
1292 | if (e) | |
c90186eb | 1293 | { |
84280917 | 1294 | if (use_oracle) |
5006671f | 1295 | { |
84280917 | 1296 | bitmap visited = NULL; |
9bb06c2a | 1297 | unsigned int cnt; |
84280917 MM |
1298 | /* Try to find a vuse that dominates this phi node by skipping |
1299 | non-clobbering statements. */ | |
511c229c | 1300 | vuse = get_continuation_for_phi (phi, &ref, &cnt, &visited, false); |
84280917 MM |
1301 | if (visited) |
1302 | BITMAP_FREE (visited); | |
5006671f | 1303 | } |
84280917 MM |
1304 | else |
1305 | vuse = NULL_TREE; | |
1306 | if (!vuse) | |
1307 | { | |
1308 | /* If we didn't find any, the value ID can't stay the same, | |
1309 | but return the translated vuse. */ | |
1310 | *same_valid = false; | |
1311 | vuse = PHI_ARG_DEF (phi, e->dest_idx); | |
1312 | } | |
1313 | /* ??? We would like to return vuse here as this is the canonical | |
1314 | upmost vdef that this reference is associated with. But during | |
1315 | insertion of the references into the hash tables we only ever | |
1316 | directly insert with their direct gimple_vuse, hence returning | |
1317 | something else would make us not find the other expression. */ | |
1318 | return PHI_ARG_DEF (phi, e->dest_idx); | |
c90186eb | 1319 | } |
c90186eb | 1320 | |
5006671f | 1321 | return NULL_TREE; |
c90186eb | 1322 | } |
83737db2 | 1323 | |
249eb506 | 1324 | /* Like bitmap_find_leader, but checks for the value existing in SET1 *or* |
83737db2 DB |
1325 | SET2. This is used to avoid making a set consisting of the union |
1326 | of PA_IN and ANTIC_IN during insert. */ | |
1327 | ||
c9145754 DB |
1328 | static inline pre_expr |
1329 | find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2) | |
83737db2 | 1330 | { |
c9145754 | 1331 | pre_expr result; |
83737db2 | 1332 | |
e076271b | 1333 | result = bitmap_find_leader (set1, val); |
83737db2 | 1334 | if (!result && set2) |
e076271b | 1335 | result = bitmap_find_leader (set2, val); |
83737db2 DB |
1336 | return result; |
1337 | } | |
1338 | ||
c9145754 DB |
1339 | /* Get the tree type for our PRE expression e. */ |
1340 | ||
1341 | static tree | |
1342 | get_expr_type (const pre_expr e) | |
1343 | { | |
1344 | switch (e->kind) | |
1345 | { | |
1346 | case NAME: | |
1347 | return TREE_TYPE (PRE_EXPR_NAME (e)); | |
1348 | case CONSTANT: | |
1349 | return TREE_TYPE (PRE_EXPR_CONSTANT (e)); | |
1350 | case REFERENCE: | |
b45d2719 | 1351 | return PRE_EXPR_REFERENCE (e)->type; |
c9145754 DB |
1352 | case NARY: |
1353 | return PRE_EXPR_NARY (e)->type; | |
1354 | } | |
c3284718 | 1355 | gcc_unreachable (); |
c9145754 DB |
1356 | } |
1357 | ||
1358 | /* Get a representative SSA_NAME for a given expression. | |
1359 | Since all of our sub-expressions are treated as values, we require | |
1360 | them to be SSA_NAME's for simplicity. | |
1361 | Prior versions of GVNPRE used to use "value handles" here, so that | |
1362 | an expression would be VH.11 + VH.10 instead of d_3 + e_6. In | |
1363 | either case, the operands are really values (IE we do not expect | |
1364 | them to be usable without finding leaders). */ | |
1365 | ||
1366 | static tree | |
1367 | get_representative_for (const pre_expr e) | |
1368 | { | |
c9145754 DB |
1369 | tree name; |
1370 | unsigned int value_id = get_expr_value_id (e); | |
1371 | ||
1372 | switch (e->kind) | |
1373 | { | |
1374 | case NAME: | |
1375 | return PRE_EXPR_NAME (e); | |
1376 | case CONSTANT: | |
726a989a | 1377 | return PRE_EXPR_CONSTANT (e); |
c9145754 DB |
1378 | case NARY: |
1379 | case REFERENCE: | |
1380 | { | |
1381 | /* Go through all of the expressions representing this value | |
1382 | and pick out an SSA_NAME. */ | |
1383 | unsigned int i; | |
1384 | bitmap_iterator bi; | |
9771b263 | 1385 | bitmap exprs = value_expressions[value_id]; |
4d4b1b30 | 1386 | EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi) |
c9145754 DB |
1387 | { |
1388 | pre_expr rep = expression_for_id (i); | |
1389 | if (rep->kind == NAME) | |
1390 | return PRE_EXPR_NAME (rep); | |
a044f0e7 RB |
1391 | else if (rep->kind == CONSTANT) |
1392 | return PRE_EXPR_CONSTANT (rep); | |
c9145754 DB |
1393 | } |
1394 | } | |
1395 | break; | |
1396 | } | |
a044f0e7 | 1397 | |
c9145754 DB |
1398 | /* If we reached here we couldn't find an SSA_NAME. This can |
1399 | happen when we've discovered a value that has never appeared in | |
a044f0e7 RB |
1400 | the program as set to an SSA_NAME, as the result of phi translation. |
1401 | Create one here. | |
1402 | ??? We should be able to re-use this when we insert the statement | |
1403 | to compute it. */ | |
83d5977e | 1404 | name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp"); |
c9145754 | 1405 | VN_INFO_GET (name)->value_id = value_id; |
a044f0e7 RB |
1406 | VN_INFO (name)->valnum = name; |
1407 | /* ??? For now mark this SSA name for release by SCCVN. */ | |
1408 | VN_INFO (name)->needs_insertion = true; | |
c9145754 | 1409 | add_to_value (value_id, get_or_alloc_expr_for_name (name)); |
a044f0e7 | 1410 | if (dump_file && (dump_flags & TDF_DETAILS)) |
c9145754 DB |
1411 | { |
1412 | fprintf (dump_file, "Created SSA_NAME representative "); | |
1413 | print_generic_expr (dump_file, name, 0); | |
1414 | fprintf (dump_file, " for expression:"); | |
1415 | print_pre_expr (dump_file, e); | |
984af6ac | 1416 | fprintf (dump_file, " (%04d)\n", value_id); |
c9145754 DB |
1417 | } |
1418 | ||
1419 | return name; | |
1420 | } | |
1421 | ||
1422 | ||
1423 | ||
3e999e7b RG |
1424 | static pre_expr |
1425 | phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, | |
1426 | basic_block pred, basic_block phiblock); | |
c9145754 | 1427 | |
7e6eb623 | 1428 | /* Translate EXPR using phis in PHIBLOCK, so that it has the values of |
788d04b2 RG |
1429 | the phis in PRED. Return NULL if we can't find a leader for each part |
1430 | of the translated expression. */ | |
6de9cd9a | 1431 | |
c9145754 | 1432 | static pre_expr |
3e999e7b RG |
1433 | phi_translate_1 (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, |
1434 | basic_block pred, basic_block phiblock) | |
6de9cd9a | 1435 | { |
c9145754 | 1436 | switch (expr->kind) |
6de9cd9a | 1437 | { |
c9145754 | 1438 | case NARY: |
43da81be | 1439 | { |
c9145754 DB |
1440 | unsigned int i; |
1441 | bool changed = false; | |
1442 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); | |
5a7d7f9c RG |
1443 | vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s, |
1444 | sizeof_vn_nary_op (nary->length)); | |
1445 | memcpy (newnary, nary, sizeof_vn_nary_op (nary->length)); | |
c9145754 | 1446 | |
5a7d7f9c | 1447 | for (i = 0; i < newnary->length; i++) |
43da81be | 1448 | { |
5a7d7f9c | 1449 | if (TREE_CODE (newnary->op[i]) != SSA_NAME) |
c9145754 DB |
1450 | continue; |
1451 | else | |
43da81be | 1452 | { |
af078bb0 | 1453 | pre_expr leader, result; |
5a7d7f9c | 1454 | unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id; |
af078bb0 | 1455 | leader = find_leader_in_sets (op_val_id, set1, set2); |
788d04b2 | 1456 | result = phi_translate (leader, set1, set2, pred, phiblock); |
c9145754 | 1457 | if (result && result != leader) |
43da81be | 1458 | { |
c9145754 DB |
1459 | tree name = get_representative_for (result); |
1460 | if (!name) | |
85300b46 | 1461 | return NULL; |
5a7d7f9c | 1462 | newnary->op[i] = name; |
43da81be | 1463 | } |
c9145754 DB |
1464 | else if (!result) |
1465 | return NULL; | |
0778d4e8 | 1466 | |
5a7d7f9c | 1467 | changed |= newnary->op[i] != nary->op[i]; |
0778d4e8 | 1468 | } |
c9145754 DB |
1469 | } |
1470 | if (changed) | |
1471 | { | |
1472 | pre_expr constant; | |
5f5126d6 | 1473 | unsigned int new_val_id; |
c9145754 | 1474 | |
5a7d7f9c RG |
1475 | tree result = vn_nary_op_lookup_pieces (newnary->length, |
1476 | newnary->opcode, | |
1477 | newnary->type, | |
1478 | &newnary->op[0], | |
c9145754 | 1479 | &nary); |
5f5126d6 RG |
1480 | if (result && is_gimple_min_invariant (result)) |
1481 | return get_or_alloc_expr_for_constant (result); | |
c9145754 DB |
1482 | |
1483 | expr = (pre_expr) pool_alloc (pre_expr_pool); | |
1484 | expr->kind = NARY; | |
1485 | expr->id = 0; | |
c9145754 DB |
1486 | if (nary) |
1487 | { | |
1488 | PRE_EXPR_NARY (expr) = nary; | |
1489 | constant = fully_constant_expression (expr); | |
1490 | if (constant != expr) | |
1491 | return constant; | |
0778d4e8 | 1492 | |
c9145754 DB |
1493 | new_val_id = nary->value_id; |
1494 | get_or_alloc_expression_id (expr); | |
1495 | } | |
1496 | else | |
43da81be | 1497 | { |
c9145754 | 1498 | new_val_id = get_next_value_id (); |
c3284718 | 1499 | value_expressions.safe_grow_cleared (get_max_value_id () + 1); |
5a7d7f9c RG |
1500 | nary = vn_nary_op_insert_pieces (newnary->length, |
1501 | newnary->opcode, | |
1502 | newnary->type, | |
1503 | &newnary->op[0], | |
c9145754 DB |
1504 | result, new_val_id); |
1505 | PRE_EXPR_NARY (expr) = nary; | |
1506 | constant = fully_constant_expression (expr); | |
1507 | if (constant != expr) | |
1508 | return constant; | |
1509 | get_or_alloc_expression_id (expr); | |
43da81be | 1510 | } |
b0a0ab2d | 1511 | add_to_value (new_val_id, expr); |
c5830edf | 1512 | } |
c9145754 | 1513 | return expr; |
c90186eb | 1514 | } |
c9145754 | 1515 | break; |
47af7a5c | 1516 | |
c9145754 | 1517 | case REFERENCE: |
c90186eb | 1518 | { |
c9145754 | 1519 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); |
9771b263 | 1520 | vec<vn_reference_op_s> operands = ref->operands; |
5006671f RG |
1521 | tree vuse = ref->vuse; |
1522 | tree newvuse = vuse; | |
6e1aa848 | 1523 | vec<vn_reference_op_s> newoperands = vNULL; |
84280917 | 1524 | bool changed = false, same_valid = true; |
4da80bfb | 1525 | unsigned int i, j, n; |
c9145754 DB |
1526 | vn_reference_op_t operand; |
1527 | vn_reference_t newref; | |
1528 | ||
aa7069aa | 1529 | for (i = 0, j = 0; |
9771b263 | 1530 | operands.iterate (i, &operand); i++, j++) |
e13f1c14 | 1531 | { |
c9145754 DB |
1532 | pre_expr opresult; |
1533 | pre_expr leader; | |
4da80bfb | 1534 | tree op[3]; |
c9145754 DB |
1535 | tree type = operand->type; |
1536 | vn_reference_op_s newop = *operand; | |
4da80bfb RG |
1537 | op[0] = operand->op0; |
1538 | op[1] = operand->op1; | |
1539 | op[2] = operand->op2; | |
1540 | for (n = 0; n < 3; ++n) | |
e13f1c14 | 1541 | { |
4da80bfb RG |
1542 | unsigned int op_val_id; |
1543 | if (!op[n]) | |
1544 | continue; | |
1545 | if (TREE_CODE (op[n]) != SSA_NAME) | |
c9145754 | 1546 | { |
4da80bfb RG |
1547 | /* We can't possibly insert these. */ |
1548 | if (n != 0 | |
1549 | && !is_gimple_min_invariant (op[n])) | |
b5d76df4 | 1550 | break; |
4da80bfb | 1551 | continue; |
c9145754 | 1552 | } |
4da80bfb | 1553 | op_val_id = VN_INFO (op[n])->value_id; |
c9145754 | 1554 | leader = find_leader_in_sets (op_val_id, set1, set2); |
4da80bfb RG |
1555 | if (!leader) |
1556 | break; | |
984af6ac RB |
1557 | opresult = phi_translate (leader, set1, set2, pred, phiblock); |
1558 | if (!opresult) | |
1559 | break; | |
1560 | if (opresult != leader) | |
c9145754 | 1561 | { |
984af6ac RB |
1562 | tree name = get_representative_for (opresult); |
1563 | if (!name) | |
b5d76df4 | 1564 | break; |
984af6ac RB |
1565 | changed |= name != op[n]; |
1566 | op[n] = name; | |
c9145754 | 1567 | } |
e13f1c14 | 1568 | } |
4da80bfb | 1569 | if (n != 3) |
e13f1c14 | 1570 | { |
9771b263 | 1571 | newoperands.release (); |
4da80bfb | 1572 | return NULL; |
e13f1c14 | 1573 | } |
9771b263 DN |
1574 | if (!newoperands.exists ()) |
1575 | newoperands = operands.copy (); | |
c9145754 | 1576 | /* We may have changed from an SSA_NAME to a constant */ |
4da80bfb RG |
1577 | if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME) |
1578 | newop.opcode = TREE_CODE (op[0]); | |
c9145754 | 1579 | newop.type = type; |
4da80bfb RG |
1580 | newop.op0 = op[0]; |
1581 | newop.op1 = op[1]; | |
1582 | newop.op2 = op[2]; | |
70f34814 RG |
1583 | /* If it transforms a non-constant ARRAY_REF into a constant |
1584 | one, adjust the constant offset. */ | |
1585 | if (newop.opcode == ARRAY_REF | |
1586 | && newop.off == -1 | |
4da80bfb RG |
1587 | && TREE_CODE (op[0]) == INTEGER_CST |
1588 | && TREE_CODE (op[1]) == INTEGER_CST | |
1589 | && TREE_CODE (op[2]) == INTEGER_CST) | |
70f34814 | 1590 | { |
4da80bfb | 1591 | double_int off = tree_to_double_int (op[0]); |
27bcd47c LC |
1592 | off += -tree_to_double_int (op[1]); |
1593 | off *= tree_to_double_int (op[2]); | |
1594 | if (off.fits_shwi ()) | |
70f34814 RG |
1595 | newop.off = off.low; |
1596 | } | |
9771b263 | 1597 | newoperands[j] = newop; |
aa7069aa RG |
1598 | /* If it transforms from an SSA_NAME to an address, fold with |
1599 | a preceding indirect reference. */ | |
4da80bfb | 1600 | if (j > 0 && op[0] && TREE_CODE (op[0]) == ADDR_EXPR |
9771b263 | 1601 | && newoperands[j - 1].opcode == MEM_REF) |
aa7069aa | 1602 | vn_reference_fold_indirect (&newoperands, &j); |
e13f1c14 | 1603 | } |
9771b263 | 1604 | if (i != operands.length ()) |
b5d76df4 | 1605 | { |
9771b263 | 1606 | newoperands.release (); |
b5d76df4 RG |
1607 | return NULL; |
1608 | } | |
c90186eb | 1609 | |
5006671f RG |
1610 | if (vuse) |
1611 | { | |
1612 | newvuse = translate_vuse_through_block (newoperands, | |
b45d2719 | 1613 | ref->set, ref->type, |
84280917 MM |
1614 | vuse, phiblock, pred, |
1615 | &same_valid); | |
5006671f RG |
1616 | if (newvuse == NULL_TREE) |
1617 | { | |
9771b263 | 1618 | newoperands.release (); |
5006671f RG |
1619 | return NULL; |
1620 | } | |
1621 | } | |
b9c5e484 | 1622 | |
84280917 | 1623 | if (changed || newvuse != vuse) |
c90186eb | 1624 | { |
47af7a5c RG |
1625 | unsigned int new_val_id; |
1626 | pre_expr constant; | |
1627 | ||
b45d2719 RG |
1628 | tree result = vn_reference_lookup_pieces (newvuse, ref->set, |
1629 | ref->type, | |
c9145754 | 1630 | newoperands, |
3bc27de7 | 1631 | &newref, VN_WALK); |
12bd5a1e | 1632 | if (result) |
9771b263 | 1633 | newoperands.release (); |
c90186eb | 1634 | |
efe7068b RG |
1635 | /* We can always insert constants, so if we have a partial |
1636 | redundant constant load of another type try to translate it | |
1637 | to a constant of appropriate type. */ | |
1638 | if (result && is_gimple_min_invariant (result)) | |
70f34814 | 1639 | { |
efe7068b RG |
1640 | tree tem = result; |
1641 | if (!useless_type_conversion_p (ref->type, TREE_TYPE (result))) | |
1642 | { | |
1643 | tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result); | |
1644 | if (tem && !is_gimple_min_invariant (tem)) | |
1645 | tem = NULL_TREE; | |
1646 | } | |
1647 | if (tem) | |
1648 | return get_or_alloc_expr_for_constant (tem); | |
70f34814 | 1649 | } |
efe7068b RG |
1650 | |
1651 | /* If we'd have to convert things we would need to validate | |
1652 | if we can insert the translated expression. So fail | |
1653 | here for now - we cannot insert an alias with a different | |
1654 | type in the VN tables either, as that would assert. */ | |
1655 | if (result | |
1656 | && !useless_type_conversion_p (ref->type, TREE_TYPE (result))) | |
1657 | return NULL; | |
48feba28 RG |
1658 | else if (!result && newref |
1659 | && !useless_type_conversion_p (ref->type, newref->type)) | |
1660 | { | |
9771b263 | 1661 | newoperands.release (); |
48feba28 RG |
1662 | return NULL; |
1663 | } | |
70f34814 | 1664 | |
c9145754 DB |
1665 | expr = (pre_expr) pool_alloc (pre_expr_pool); |
1666 | expr->kind = REFERENCE; | |
1667 | expr->id = 0; | |
6615c446 | 1668 | |
efe7068b | 1669 | if (newref) |
0995a441 | 1670 | { |
c9145754 | 1671 | PRE_EXPR_REFERENCE (expr) = newref; |
47af7a5c RG |
1672 | constant = fully_constant_expression (expr); |
1673 | if (constant != expr) | |
1674 | return constant; | |
1675 | ||
c9145754 DB |
1676 | new_val_id = newref->value_id; |
1677 | get_or_alloc_expression_id (expr); | |
0995a441 SB |
1678 | } |
1679 | else | |
1680 | { | |
84280917 MM |
1681 | if (changed || !same_valid) |
1682 | { | |
1683 | new_val_id = get_next_value_id (); | |
c3284718 RS |
1684 | value_expressions.safe_grow_cleared |
1685 | (get_max_value_id () + 1); | |
84280917 MM |
1686 | } |
1687 | else | |
1688 | new_val_id = ref->value_id; | |
b45d2719 RG |
1689 | newref = vn_reference_insert_pieces (newvuse, ref->set, |
1690 | ref->type, | |
c9145754 DB |
1691 | newoperands, |
1692 | result, new_val_id); | |
9771b263 | 1693 | newoperands.create (0); |
c9145754 | 1694 | PRE_EXPR_REFERENCE (expr) = newref; |
47af7a5c RG |
1695 | constant = fully_constant_expression (expr); |
1696 | if (constant != expr) | |
1697 | return constant; | |
c9145754 | 1698 | get_or_alloc_expression_id (expr); |
0995a441 | 1699 | } |
b0a0ab2d | 1700 | add_to_value (new_val_id, expr); |
7e6eb623 | 1701 | } |
9771b263 | 1702 | newoperands.release (); |
c9145754 | 1703 | return expr; |
7e6eb623 | 1704 | } |
c9145754 | 1705 | break; |
47af7a5c | 1706 | |
c9145754 | 1707 | case NAME: |
7e6eb623 | 1708 | { |
c9145754 | 1709 | tree name = PRE_EXPR_NAME (expr); |
e076271b RG |
1710 | gimple def_stmt = SSA_NAME_DEF_STMT (name); |
1711 | /* If the SSA name is defined by a PHI node in this block, | |
1712 | translate it. */ | |
726a989a RB |
1713 | if (gimple_code (def_stmt) == GIMPLE_PHI |
1714 | && gimple_bb (def_stmt) == phiblock) | |
9323afae | 1715 | { |
e076271b RG |
1716 | edge e = find_edge (pred, gimple_bb (def_stmt)); |
1717 | tree def = PHI_ARG_DEF (def_stmt, e->dest_idx); | |
73a63870 | 1718 | |
c9145754 | 1719 | /* Handle constant. */ |
89fb70a3 | 1720 | if (is_gimple_min_invariant (def)) |
c9145754 | 1721 | return get_or_alloc_expr_for_constant (def); |
070b797d | 1722 | |
e076271b | 1723 | return get_or_alloc_expr_for_name (def); |
9323afae | 1724 | } |
e076271b RG |
1725 | /* Otherwise return it unchanged - it will get cleaned if its |
1726 | value is not available in PREDs AVAIL_OUT set of expressions. */ | |
1727 | return expr; | |
7e6eb623 | 1728 | } |
6615c446 JO |
1729 | |
1730 | default: | |
1731 | gcc_unreachable (); | |
6de9cd9a DN |
1732 | } |
1733 | } | |
f8b04195 | 1734 | |
3e999e7b RG |
1735 | /* Wrapper around phi_translate_1 providing caching functionality. */ |
1736 | ||
1737 | static pre_expr | |
1738 | phi_translate (pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, | |
1739 | basic_block pred, basic_block phiblock) | |
1740 | { | |
984af6ac | 1741 | expr_pred_trans_t slot = NULL; |
3e999e7b RG |
1742 | pre_expr phitrans; |
1743 | ||
1744 | if (!expr) | |
1745 | return NULL; | |
1746 | ||
1747 | /* Constants contain no values that need translation. */ | |
1748 | if (expr->kind == CONSTANT) | |
1749 | return expr; | |
1750 | ||
1751 | if (value_id_constant_p (get_expr_value_id (expr))) | |
1752 | return expr; | |
1753 | ||
984af6ac | 1754 | /* Don't add translations of NAMEs as those are cheap to translate. */ |
3e999e7b RG |
1755 | if (expr->kind != NAME) |
1756 | { | |
984af6ac RB |
1757 | if (phi_trans_add (&slot, expr, pred)) |
1758 | return slot->v; | |
1759 | /* Store NULL for the value we want to return in the case of | |
1760 | recursing. */ | |
1761 | slot->v = NULL; | |
3e999e7b RG |
1762 | } |
1763 | ||
1764 | /* Translate. */ | |
1765 | phitrans = phi_translate_1 (expr, set1, set2, pred, phiblock); | |
1766 | ||
984af6ac | 1767 | if (slot) |
e2c2fde2 RB |
1768 | { |
1769 | if (phitrans) | |
1770 | slot->v = phitrans; | |
1771 | else | |
1772 | /* Remove failed translations again, they cause insert | |
1773 | iteration to not pick up new opportunities reliably. */ | |
1774 | phi_translate_table.remove_elt_with_hash (slot, slot->hashcode); | |
1775 | } | |
3e999e7b RG |
1776 | |
1777 | return phitrans; | |
1778 | } | |
1779 | ||
1780 | ||
c9145754 | 1781 | /* For each expression in SET, translate the values through phi nodes |
f5594471 DB |
1782 | in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting |
1783 | expressions in DEST. */ | |
1784 | ||
6de9cd9a | 1785 | static void |
83737db2 | 1786 | phi_translate_set (bitmap_set_t dest, bitmap_set_t set, basic_block pred, |
7e6eb623 | 1787 | basic_block phiblock) |
6de9cd9a | 1788 | { |
9771b263 | 1789 | vec<pre_expr> exprs; |
c9145754 | 1790 | pre_expr expr; |
83737db2 DB |
1791 | int i; |
1792 | ||
9adf0570 | 1793 | if (gimple_seq_empty_p (phi_nodes (phiblock))) |
6de9cd9a | 1794 | { |
83737db2 DB |
1795 | bitmap_set_copy (dest, set); |
1796 | return; | |
1797 | } | |
b9c5e484 | 1798 | |
83737db2 | 1799 | exprs = sorted_array_from_bitmap_set (set); |
9771b263 | 1800 | FOR_EACH_VEC_ELT (exprs, i, expr) |
83737db2 | 1801 | { |
c9145754 | 1802 | pre_expr translated; |
f8b04195 | 1803 | translated = phi_translate (expr, set, NULL, pred, phiblock); |
3ed7d068 RG |
1804 | if (!translated) |
1805 | continue; | |
c90186eb | 1806 | |
3ed7d068 RG |
1807 | /* We might end up with multiple expressions from SET being |
1808 | translated to the same value. In this case we do not want | |
1809 | to retain the NARY or REFERENCE expression but prefer a NAME | |
1810 | which would be the leader. */ | |
1811 | if (translated->kind == NAME) | |
1812 | bitmap_value_replace_in_set (dest, translated); | |
1813 | else | |
83737db2 | 1814 | bitmap_value_insert_into_set (dest, translated); |
b9c5e484 | 1815 | } |
9771b263 | 1816 | exprs.release (); |
6de9cd9a DN |
1817 | } |
1818 | ||
bdee7684 | 1819 | /* Find the leader for a value (i.e., the name representing that |
984af6ac RB |
1820 | value) in a given set, and return it. Return NULL if no leader |
1821 | is found. */ | |
bdee7684 | 1822 | |
c9145754 | 1823 | static pre_expr |
e076271b | 1824 | bitmap_find_leader (bitmap_set_t set, unsigned int val) |
bdee7684 | 1825 | { |
c9145754 DB |
1826 | if (value_id_constant_p (val)) |
1827 | { | |
1828 | unsigned int i; | |
1829 | bitmap_iterator bi; | |
9771b263 | 1830 | bitmap exprset = value_expressions[val]; |
83737db2 | 1831 | |
4d4b1b30 | 1832 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) |
c9145754 DB |
1833 | { |
1834 | pre_expr expr = expression_for_id (i); | |
1835 | if (expr->kind == CONSTANT) | |
1836 | return expr; | |
1837 | } | |
1838 | } | |
bdee7684 DB |
1839 | if (bitmap_set_contains_value (set, val)) |
1840 | { | |
6b416da1 DB |
1841 | /* Rather than walk the entire bitmap of expressions, and see |
1842 | whether any of them has the value we are looking for, we look | |
1843 | at the reverse mapping, which tells us the set of expressions | |
1844 | that have a given value (IE value->expressions with that | |
1845 | value) and see if any of those expressions are in our set. | |
1846 | The number of expressions per value is usually significantly | |
1847 | less than the number of expressions in the set. In fact, for | |
1848 | large testcases, doing it this way is roughly 5-10x faster | |
1849 | than walking the bitmap. | |
1850 | If this is somehow a significant lose for some cases, we can | |
b9c5e484 | 1851 | choose which set to walk based on which set is smaller. */ |
83737db2 DB |
1852 | unsigned int i; |
1853 | bitmap_iterator bi; | |
9771b263 | 1854 | bitmap exprset = value_expressions[val]; |
b9c5e484 | 1855 | |
4d4b1b30 | 1856 | EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi) |
e076271b | 1857 | return expression_for_id (i); |
6de9cd9a | 1858 | } |
7e6eb623 | 1859 | return NULL; |
6de9cd9a DN |
1860 | } |
1861 | ||
cea618ac | 1862 | /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of |
1e4816bc DB |
1863 | BLOCK by seeing if it is not killed in the block. Note that we are |
1864 | only determining whether there is a store that kills it. Because | |
1865 | of the order in which clean iterates over values, we are guaranteed | |
1866 | that altered operands will have caused us to be eliminated from the | |
1867 | ANTIC_IN set already. */ | |
c90186eb DB |
1868 | |
1869 | static bool | |
c9145754 | 1870 | value_dies_in_block_x (pre_expr expr, basic_block block) |
c90186eb | 1871 | { |
5006671f RG |
1872 | tree vuse = PRE_EXPR_REFERENCE (expr)->vuse; |
1873 | vn_reference_t refx = PRE_EXPR_REFERENCE (expr); | |
1874 | gimple def; | |
5006671f RG |
1875 | gimple_stmt_iterator gsi; |
1876 | unsigned id = get_expression_id (expr); | |
1877 | bool res = false; | |
b45d2719 | 1878 | ao_ref ref; |
89fb70a3 | 1879 | |
5006671f RG |
1880 | if (!vuse) |
1881 | return false; | |
1882 | ||
1883 | /* Lookup a previously calculated result. */ | |
1884 | if (EXPR_DIES (block) | |
1885 | && bitmap_bit_p (EXPR_DIES (block), id * 2)) | |
1886 | return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1); | |
1887 | ||
1888 | /* A memory expression {e, VUSE} dies in the block if there is a | |
1889 | statement that may clobber e. If, starting statement walk from the | |
1890 | top of the basic block, a statement uses VUSE there can be no kill | |
1891 | inbetween that use and the original statement that loaded {e, VUSE}, | |
1892 | so we can stop walking. */ | |
b45d2719 | 1893 | ref.base = NULL_TREE; |
5006671f | 1894 | for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi)) |
c90186eb | 1895 | { |
5006671f RG |
1896 | tree def_vuse, def_vdef; |
1897 | def = gsi_stmt (gsi); | |
1898 | def_vuse = gimple_vuse (def); | |
1899 | def_vdef = gimple_vdef (def); | |
89fb70a3 | 1900 | |
5006671f RG |
1901 | /* Not a memory statement. */ |
1902 | if (!def_vuse) | |
1e4816bc | 1903 | continue; |
5006671f RG |
1904 | |
1905 | /* Not a may-def. */ | |
1906 | if (!def_vdef) | |
1907 | { | |
1908 | /* A load with the same VUSE, we're done. */ | |
1909 | if (def_vuse == vuse) | |
1910 | break; | |
1911 | ||
1912 | continue; | |
1913 | } | |
1914 | ||
1915 | /* Init ref only if we really need it. */ | |
b45d2719 RG |
1916 | if (ref.base == NULL_TREE |
1917 | && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->type, | |
1918 | refx->operands)) | |
5006671f | 1919 | { |
b45d2719 RG |
1920 | res = true; |
1921 | break; | |
5006671f RG |
1922 | } |
1923 | /* If the statement may clobber expr, it dies. */ | |
b45d2719 | 1924 | if (stmt_may_clobber_ref_p_1 (def, &ref)) |
5006671f RG |
1925 | { |
1926 | res = true; | |
1927 | break; | |
1928 | } | |
c90186eb | 1929 | } |
5006671f RG |
1930 | |
1931 | /* Remember the result. */ | |
1932 | if (!EXPR_DIES (block)) | |
1933 | EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack); | |
1934 | bitmap_set_bit (EXPR_DIES (block), id * 2); | |
1935 | if (res) | |
1936 | bitmap_set_bit (EXPR_DIES (block), id * 2 + 1); | |
1937 | ||
1938 | return res; | |
c90186eb DB |
1939 | } |
1940 | ||
6de9cd9a | 1941 | |
19372838 RG |
1942 | /* Determine if OP is valid in SET1 U SET2, which it is when the union |
1943 | contains its value-id. */ | |
83737db2 | 1944 | |
6de9cd9a | 1945 | static bool |
19372838 | 1946 | op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op) |
6de9cd9a | 1947 | { |
19372838 | 1948 | if (op && TREE_CODE (op) == SSA_NAME) |
c9145754 | 1949 | { |
19372838 | 1950 | unsigned int value_id = VN_INFO (op)->value_id; |
92290a18 RG |
1951 | if (!(bitmap_set_contains_value (set1, value_id) |
1952 | || (set2 && bitmap_set_contains_value (set2, value_id)))) | |
c9145754 DB |
1953 | return false; |
1954 | } | |
c9145754 DB |
1955 | return true; |
1956 | } | |
b9c5e484 | 1957 | |
c9145754 DB |
1958 | /* Determine if the expression EXPR is valid in SET1 U SET2. |
1959 | ONLY SET2 CAN BE NULL. | |
1960 | This means that we have a leader for each part of the expression | |
1961 | (if it consists of values), or the expression is an SSA_NAME. | |
7763473e | 1962 | For loads/calls, we also see if the vuse is killed in this block. */ |
5039610b | 1963 | |
c9145754 DB |
1964 | static bool |
1965 | valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr, | |
1966 | basic_block block) | |
1967 | { | |
1968 | switch (expr->kind) | |
1969 | { | |
1970 | case NAME: | |
e167c04d RB |
1971 | return bitmap_find_leader (AVAIL_OUT (block), |
1972 | get_expr_value_id (expr)) != NULL; | |
c9145754 | 1973 | case NARY: |
43da81be | 1974 | { |
c9145754 DB |
1975 | unsigned int i; |
1976 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); | |
1977 | for (i = 0; i < nary->length; i++) | |
19372838 RG |
1978 | if (!op_valid_in_sets (set1, set2, nary->op[i])) |
1979 | return false; | |
c9145754 | 1980 | return true; |
43da81be | 1981 | } |
c9145754 DB |
1982 | break; |
1983 | case REFERENCE: | |
c90186eb | 1984 | { |
c9145754 DB |
1985 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); |
1986 | vn_reference_op_t vro; | |
1987 | unsigned int i; | |
1988 | ||
9771b263 | 1989 | FOR_EACH_VEC_ELT (ref->operands, i, vro) |
c90186eb | 1990 | { |
19372838 RG |
1991 | if (!op_valid_in_sets (set1, set2, vro->op0) |
1992 | || !op_valid_in_sets (set1, set2, vro->op1) | |
1993 | || !op_valid_in_sets (set1, set2, vro->op2)) | |
e13f1c14 | 1994 | return false; |
c90186eb | 1995 | } |
cd32bb90 | 1996 | return true; |
c90186eb | 1997 | } |
6615c446 | 1998 | default: |
b9c5e484 | 1999 | gcc_unreachable (); |
c9145754 | 2000 | } |
6de9cd9a DN |
2001 | } |
2002 | ||
d75dbccd DB |
2003 | /* Clean the set of expressions that are no longer valid in SET1 or |
2004 | SET2. This means expressions that are made up of values we have no | |
2005 | leaders for in SET1 or SET2. This version is used for partial | |
2006 | anticipation, which means it is not valid in either ANTIC_IN or | |
2007 | PA_IN. */ | |
2008 | ||
2009 | static void | |
2010 | dependent_clean (bitmap_set_t set1, bitmap_set_t set2, basic_block block) | |
2011 | { | |
9771b263 | 2012 | vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1); |
c9145754 | 2013 | pre_expr expr; |
d75dbccd DB |
2014 | int i; |
2015 | ||
9771b263 | 2016 | FOR_EACH_VEC_ELT (exprs, i, expr) |
d75dbccd DB |
2017 | { |
2018 | if (!valid_in_sets (set1, set2, expr, block)) | |
2019 | bitmap_remove_from_set (set1, expr); | |
2020 | } | |
9771b263 | 2021 | exprs.release (); |
d75dbccd DB |
2022 | } |
2023 | ||
ca072a31 DB |
2024 | /* Clean the set of expressions that are no longer valid in SET. This |
2025 | means expressions that are made up of values we have no leaders for | |
2026 | in SET. */ | |
6de9cd9a DN |
2027 | |
2028 | static void | |
83737db2 | 2029 | clean (bitmap_set_t set, basic_block block) |
6de9cd9a | 2030 | { |
9771b263 | 2031 | vec<pre_expr> exprs = sorted_array_from_bitmap_set (set); |
c9145754 | 2032 | pre_expr expr; |
83737db2 DB |
2033 | int i; |
2034 | ||
9771b263 | 2035 | FOR_EACH_VEC_ELT (exprs, i, expr) |
6de9cd9a | 2036 | { |
83737db2 DB |
2037 | if (!valid_in_sets (set, NULL, expr, block)) |
2038 | bitmap_remove_from_set (set, expr); | |
6de9cd9a | 2039 | } |
9771b263 | 2040 | exprs.release (); |
6de9cd9a DN |
2041 | } |
2042 | ||
cd32bb90 | 2043 | /* Clean the set of expressions that are no longer valid in SET because |
bea966c2 | 2044 | they are clobbered in BLOCK or because they trap and may not be executed. */ |
cd32bb90 RG |
2045 | |
2046 | static void | |
2047 | prune_clobbered_mems (bitmap_set_t set, basic_block block) | |
2048 | { | |
bea966c2 RG |
2049 | bitmap_iterator bi; |
2050 | unsigned i; | |
cd32bb90 | 2051 | |
bea966c2 | 2052 | FOR_EACH_EXPR_ID_IN_SET (set, i, bi) |
cd32bb90 | 2053 | { |
bea966c2 RG |
2054 | pre_expr expr = expression_for_id (i); |
2055 | if (expr->kind == REFERENCE) | |
2056 | { | |
2057 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); | |
2058 | if (ref->vuse) | |
2059 | { | |
2060 | gimple def_stmt = SSA_NAME_DEF_STMT (ref->vuse); | |
2061 | if (!gimple_nop_p (def_stmt) | |
2062 | && ((gimple_bb (def_stmt) != block | |
2063 | && !dominated_by_p (CDI_DOMINATORS, | |
2064 | block, gimple_bb (def_stmt))) | |
2065 | || (gimple_bb (def_stmt) == block | |
2066 | && value_dies_in_block_x (expr, block)))) | |
2067 | bitmap_remove_from_set (set, expr); | |
2068 | } | |
2069 | } | |
2070 | else if (expr->kind == NARY) | |
cd32bb90 | 2071 | { |
bea966c2 RG |
2072 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); |
2073 | /* If the NARY may trap make sure the block does not contain | |
2074 | a possible exit point. | |
2075 | ??? This is overly conservative if we translate AVAIL_OUT | |
2076 | as the available expression might be after the exit point. */ | |
2077 | if (BB_MAY_NOTRETURN (block) | |
2078 | && vn_nary_may_trap (nary)) | |
cd32bb90 RG |
2079 | bitmap_remove_from_set (set, expr); |
2080 | } | |
2081 | } | |
cd32bb90 RG |
2082 | } |
2083 | ||
d4222d43 | 2084 | static sbitmap has_abnormal_preds; |
89fb70a3 | 2085 | |
83737db2 DB |
2086 | /* List of blocks that may have changed during ANTIC computation and |
2087 | thus need to be iterated over. */ | |
2088 | ||
2089 | static sbitmap changed_blocks; | |
1e4816bc DB |
2090 | |
2091 | /* Decide whether to defer a block for a later iteration, or PHI | |
2092 | translate SOURCE to DEST using phis in PHIBLOCK. Return false if we | |
2093 | should defer the block, and true if we processed it. */ | |
2094 | ||
2095 | static bool | |
2096 | defer_or_phi_translate_block (bitmap_set_t dest, bitmap_set_t source, | |
2097 | basic_block block, basic_block phiblock) | |
2098 | { | |
2099 | if (!BB_VISITED (phiblock)) | |
2100 | { | |
d7c028c0 | 2101 | bitmap_set_bit (changed_blocks, block->index); |
1e4816bc DB |
2102 | BB_VISITED (block) = 0; |
2103 | BB_DEFERRED (block) = 1; | |
2104 | return false; | |
2105 | } | |
2106 | else | |
2107 | phi_translate_set (dest, source, block, phiblock); | |
2108 | return true; | |
2109 | } | |
2110 | ||
7e6eb623 | 2111 | /* Compute the ANTIC set for BLOCK. |
6de9cd9a | 2112 | |
665fcad8 SB |
2113 | If succs(BLOCK) > 1 then |
2114 | ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK) | |
2115 | else if succs(BLOCK) == 1 then | |
2116 | ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)]) | |
6de9cd9a | 2117 | |
665fcad8 | 2118 | ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK]) |
83737db2 | 2119 | */ |
6de9cd9a | 2120 | |
7e6eb623 | 2121 | static bool |
a28fee03 | 2122 | compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge) |
6de9cd9a | 2123 | { |
7e6eb623 | 2124 | bool changed = false; |
83737db2 DB |
2125 | bitmap_set_t S, old, ANTIC_OUT; |
2126 | bitmap_iterator bi; | |
2127 | unsigned int bii; | |
2128 | edge e; | |
2129 | edge_iterator ei; | |
a28fee03 | 2130 | |
83737db2 | 2131 | old = ANTIC_OUT = S = NULL; |
d75dbccd | 2132 | BB_VISITED (block) = 1; |
a28fee03 SB |
2133 | |
2134 | /* If any edges from predecessors are abnormal, antic_in is empty, | |
2135 | so do nothing. */ | |
2136 | if (block_has_abnormal_pred_edge) | |
2137 | goto maybe_dump_sets; | |
6de9cd9a | 2138 | |
83737db2 DB |
2139 | old = ANTIC_IN (block); |
2140 | ANTIC_OUT = bitmap_set_new (); | |
6de9cd9a | 2141 | |
a28fee03 SB |
2142 | /* If the block has no successors, ANTIC_OUT is empty. */ |
2143 | if (EDGE_COUNT (block->succs) == 0) | |
2144 | ; | |
7e6eb623 DB |
2145 | /* If we have one successor, we could have some phi nodes to |
2146 | translate through. */ | |
c5cbcccf | 2147 | else if (single_succ_p (block)) |
6de9cd9a | 2148 | { |
83737db2 | 2149 | basic_block succ_bb = single_succ (block); |
d75dbccd DB |
2150 | |
2151 | /* We trade iterations of the dataflow equations for having to | |
2152 | phi translate the maximal set, which is incredibly slow | |
2153 | (since the maximal set often has 300+ members, even when you | |
2154 | have a small number of blocks). | |
2155 | Basically, we defer the computation of ANTIC for this block | |
2f8e468b | 2156 | until we have processed it's successor, which will inevitably |
d75dbccd DB |
2157 | have a *much* smaller set of values to phi translate once |
2158 | clean has been run on it. | |
2159 | The cost of doing this is that we technically perform more | |
2160 | iterations, however, they are lower cost iterations. | |
2161 | ||
2162 | Timings for PRE on tramp3d-v4: | |
2163 | without maximal set fix: 11 seconds | |
2164 | with maximal set fix/without deferring: 26 seconds | |
2165 | with maximal set fix/with deferring: 11 seconds | |
2166 | */ | |
2167 | ||
1e4816bc DB |
2168 | if (!defer_or_phi_translate_block (ANTIC_OUT, ANTIC_IN (succ_bb), |
2169 | block, succ_bb)) | |
d75dbccd DB |
2170 | { |
2171 | changed = true; | |
d75dbccd DB |
2172 | goto maybe_dump_sets; |
2173 | } | |
6de9cd9a | 2174 | } |
7e6eb623 | 2175 | /* If we have multiple successors, we take the intersection of all of |
1e4816bc DB |
2176 | them. Note that in the case of loop exit phi nodes, we may have |
2177 | phis to translate through. */ | |
7e6eb623 | 2178 | else |
6de9cd9a | 2179 | { |
9771b263 | 2180 | vec<basic_block> worklist; |
7e6eb623 | 2181 | size_t i; |
70a6b17e | 2182 | basic_block bprime, first = NULL; |
7e6eb623 | 2183 | |
9771b263 | 2184 | worklist.create (EDGE_COUNT (block->succs)); |
628f6a4e | 2185 | FOR_EACH_EDGE (e, ei, block->succs) |
1e4816bc | 2186 | { |
70a6b17e RG |
2187 | if (!first |
2188 | && BB_VISITED (e->dest)) | |
2189 | first = e->dest; | |
2190 | else if (BB_VISITED (e->dest)) | |
9771b263 | 2191 | worklist.quick_push (e->dest); |
1e4816bc | 2192 | } |
70a6b17e RG |
2193 | |
2194 | /* Of multiple successors we have to have visited one already. */ | |
2195 | if (!first) | |
1e4816bc | 2196 | { |
d7c028c0 | 2197 | bitmap_set_bit (changed_blocks, block->index); |
70a6b17e RG |
2198 | BB_VISITED (block) = 0; |
2199 | BB_DEFERRED (block) = 1; | |
2200 | changed = true; | |
9771b263 | 2201 | worklist.release (); |
70a6b17e | 2202 | goto maybe_dump_sets; |
1e4816bc | 2203 | } |
c90186eb | 2204 | |
9adf0570 | 2205 | if (!gimple_seq_empty_p (phi_nodes (first))) |
70a6b17e RG |
2206 | phi_translate_set (ANTIC_OUT, ANTIC_IN (first), block, first); |
2207 | else | |
2208 | bitmap_set_copy (ANTIC_OUT, ANTIC_IN (first)); | |
2209 | ||
9771b263 | 2210 | FOR_EACH_VEC_ELT (worklist, i, bprime) |
d75dbccd | 2211 | { |
9adf0570 | 2212 | if (!gimple_seq_empty_p (phi_nodes (bprime))) |
1e4816bc DB |
2213 | { |
2214 | bitmap_set_t tmp = bitmap_set_new (); | |
70a6b17e | 2215 | phi_translate_set (tmp, ANTIC_IN (bprime), block, bprime); |
1e4816bc DB |
2216 | bitmap_set_and (ANTIC_OUT, tmp); |
2217 | bitmap_set_free (tmp); | |
2218 | } | |
89fb70a3 | 2219 | else |
70a6b17e | 2220 | bitmap_set_and (ANTIC_OUT, ANTIC_IN (bprime)); |
6de9cd9a | 2221 | } |
9771b263 | 2222 | worklist.release (); |
6de9cd9a | 2223 | } |
6de9cd9a | 2224 | |
cd32bb90 RG |
2225 | /* Prune expressions that are clobbered in block and thus become |
2226 | invalid if translated from ANTIC_OUT to ANTIC_IN. */ | |
2227 | prune_clobbered_mems (ANTIC_OUT, block); | |
2228 | ||
ea4b7848 | 2229 | /* Generate ANTIC_OUT - TMP_GEN. */ |
83737db2 | 2230 | S = bitmap_set_subtract (ANTIC_OUT, TMP_GEN (block)); |
6de9cd9a | 2231 | |
d75dbccd | 2232 | /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */ |
83737db2 DB |
2233 | ANTIC_IN (block) = bitmap_set_subtract (EXP_GEN (block), |
2234 | TMP_GEN (block)); | |
c33bae88 | 2235 | |
a28fee03 SB |
2236 | /* Then union in the ANTIC_OUT - TMP_GEN values, |
2237 | to get ANTIC_OUT U EXP_GEN - TMP_GEN */ | |
83737db2 DB |
2238 | FOR_EACH_EXPR_ID_IN_SET (S, bii, bi) |
2239 | bitmap_value_insert_into_set (ANTIC_IN (block), | |
2240 | expression_for_id (bii)); | |
6de9cd9a | 2241 | |
c90186eb | 2242 | clean (ANTIC_IN (block), block); |
d75dbccd | 2243 | |
5c72d561 | 2244 | if (!bitmap_set_equal (old, ANTIC_IN (block))) |
83737db2 DB |
2245 | { |
2246 | changed = true; | |
d7c028c0 | 2247 | bitmap_set_bit (changed_blocks, block->index); |
83737db2 | 2248 | FOR_EACH_EDGE (e, ei, block->preds) |
d7c028c0 | 2249 | bitmap_set_bit (changed_blocks, e->src->index); |
83737db2 DB |
2250 | } |
2251 | else | |
d7c028c0 | 2252 | bitmap_clear_bit (changed_blocks, block->index); |
6de9cd9a | 2253 | |
a28fee03 | 2254 | maybe_dump_sets: |
7e6eb623 DB |
2255 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2256 | { | |
d75dbccd DB |
2257 | if (!BB_DEFERRED (block) || BB_VISITED (block)) |
2258 | { | |
2259 | if (ANTIC_OUT) | |
2260 | print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index); | |
85300b46 | 2261 | |
d75dbccd DB |
2262 | print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN", |
2263 | block->index); | |
85300b46 | 2264 | |
d75dbccd DB |
2265 | if (S) |
2266 | print_bitmap_set (dump_file, S, "S", block->index); | |
2267 | } | |
2268 | else | |
2269 | { | |
2270 | fprintf (dump_file, | |
2271 | "Block %d was deferred for a future iteration.\n", | |
2272 | block->index); | |
2273 | } | |
83737db2 DB |
2274 | } |
2275 | if (old) | |
2276 | bitmap_set_free (old); | |
2277 | if (S) | |
2278 | bitmap_set_free (S); | |
2279 | if (ANTIC_OUT) | |
2280 | bitmap_set_free (ANTIC_OUT); | |
7e6eb623 | 2281 | return changed; |
6de9cd9a DN |
2282 | } |
2283 | ||
d75dbccd DB |
2284 | /* Compute PARTIAL_ANTIC for BLOCK. |
2285 | ||
2286 | If succs(BLOCK) > 1 then | |
2287 | PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not | |
2288 | in ANTIC_OUT for all succ(BLOCK) | |
2289 | else if succs(BLOCK) == 1 then | |
2290 | PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)]) | |
2291 | ||
2292 | PA_IN[BLOCK] = dependent_clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK] | |
2293 | - ANTIC_IN[BLOCK]) | |
2294 | ||
2295 | */ | |
2296 | static bool | |
2297 | compute_partial_antic_aux (basic_block block, | |
2298 | bool block_has_abnormal_pred_edge) | |
2299 | { | |
2300 | bool changed = false; | |
2301 | bitmap_set_t old_PA_IN; | |
2302 | bitmap_set_t PA_OUT; | |
2303 | edge e; | |
2304 | edge_iterator ei; | |
f0ed4cfb | 2305 | unsigned long max_pa = PARAM_VALUE (PARAM_MAX_PARTIAL_ANTIC_LENGTH); |
d75dbccd DB |
2306 | |
2307 | old_PA_IN = PA_OUT = NULL; | |
2308 | ||
2309 | /* If any edges from predecessors are abnormal, antic_in is empty, | |
2310 | so do nothing. */ | |
2311 | if (block_has_abnormal_pred_edge) | |
2312 | goto maybe_dump_sets; | |
2313 | ||
f0ed4cfb NC |
2314 | /* If there are too many partially anticipatable values in the |
2315 | block, phi_translate_set can take an exponential time: stop | |
2316 | before the translation starts. */ | |
2317 | if (max_pa | |
2318 | && single_succ_p (block) | |
5c72d561 | 2319 | && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa) |
f0ed4cfb NC |
2320 | goto maybe_dump_sets; |
2321 | ||
d75dbccd DB |
2322 | old_PA_IN = PA_IN (block); |
2323 | PA_OUT = bitmap_set_new (); | |
2324 | ||
2325 | /* If the block has no successors, ANTIC_OUT is empty. */ | |
2326 | if (EDGE_COUNT (block->succs) == 0) | |
2327 | ; | |
2328 | /* If we have one successor, we could have some phi nodes to | |
2329 | translate through. Note that we can't phi translate across DFS | |
89fb70a3 DB |
2330 | back edges in partial antic, because it uses a union operation on |
2331 | the successors. For recurrences like IV's, we will end up | |
2332 | generating a new value in the set on each go around (i + 3 (VH.1) | |
2333 | VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */ | |
d75dbccd DB |
2334 | else if (single_succ_p (block)) |
2335 | { | |
2336 | basic_block succ = single_succ (block); | |
2337 | if (!(single_succ_edge (block)->flags & EDGE_DFS_BACK)) | |
2338 | phi_translate_set (PA_OUT, PA_IN (succ), block, succ); | |
2339 | } | |
2340 | /* If we have multiple successors, we take the union of all of | |
2341 | them. */ | |
2342 | else | |
2343 | { | |
9771b263 | 2344 | vec<basic_block> worklist; |
d75dbccd DB |
2345 | size_t i; |
2346 | basic_block bprime; | |
2347 | ||
9771b263 | 2348 | worklist.create (EDGE_COUNT (block->succs)); |
d75dbccd DB |
2349 | FOR_EACH_EDGE (e, ei, block->succs) |
2350 | { | |
2351 | if (e->flags & EDGE_DFS_BACK) | |
2352 | continue; | |
9771b263 | 2353 | worklist.quick_push (e->dest); |
d75dbccd | 2354 | } |
9771b263 | 2355 | if (worklist.length () > 0) |
d75dbccd | 2356 | { |
9771b263 | 2357 | FOR_EACH_VEC_ELT (worklist, i, bprime) |
d75dbccd DB |
2358 | { |
2359 | unsigned int i; | |
2360 | bitmap_iterator bi; | |
2361 | ||
2362 | FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (bprime), i, bi) | |
2363 | bitmap_value_insert_into_set (PA_OUT, | |
2364 | expression_for_id (i)); | |
9adf0570 | 2365 | if (!gimple_seq_empty_p (phi_nodes (bprime))) |
1e4816bc DB |
2366 | { |
2367 | bitmap_set_t pa_in = bitmap_set_new (); | |
2368 | phi_translate_set (pa_in, PA_IN (bprime), block, bprime); | |
2369 | FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi) | |
2370 | bitmap_value_insert_into_set (PA_OUT, | |
2371 | expression_for_id (i)); | |
2372 | bitmap_set_free (pa_in); | |
2373 | } | |
2374 | else | |
2375 | FOR_EACH_EXPR_ID_IN_SET (PA_IN (bprime), i, bi) | |
2376 | bitmap_value_insert_into_set (PA_OUT, | |
2377 | expression_for_id (i)); | |
d75dbccd DB |
2378 | } |
2379 | } | |
9771b263 | 2380 | worklist.release (); |
d75dbccd DB |
2381 | } |
2382 | ||
cd32bb90 RG |
2383 | /* Prune expressions that are clobbered in block and thus become |
2384 | invalid if translated from PA_OUT to PA_IN. */ | |
2385 | prune_clobbered_mems (PA_OUT, block); | |
2386 | ||
d75dbccd DB |
2387 | /* PA_IN starts with PA_OUT - TMP_GEN. |
2388 | Then we subtract things from ANTIC_IN. */ | |
2389 | PA_IN (block) = bitmap_set_subtract (PA_OUT, TMP_GEN (block)); | |
2390 | ||
2391 | /* For partial antic, we want to put back in the phi results, since | |
2392 | we will properly avoid making them partially antic over backedges. */ | |
5c72d561 JH |
2393 | bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values); |
2394 | bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions); | |
d75dbccd DB |
2395 | |
2396 | /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */ | |
2397 | bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block)); | |
2398 | ||
2399 | dependent_clean (PA_IN (block), ANTIC_IN (block), block); | |
2400 | ||
2401 | if (!bitmap_set_equal (old_PA_IN, PA_IN (block))) | |
2402 | { | |
2403 | changed = true; | |
d7c028c0 | 2404 | bitmap_set_bit (changed_blocks, block->index); |
d75dbccd | 2405 | FOR_EACH_EDGE (e, ei, block->preds) |
d7c028c0 | 2406 | bitmap_set_bit (changed_blocks, e->src->index); |
d75dbccd DB |
2407 | } |
2408 | else | |
d7c028c0 | 2409 | bitmap_clear_bit (changed_blocks, block->index); |
d75dbccd DB |
2410 | |
2411 | maybe_dump_sets: | |
2412 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2413 | { | |
2414 | if (PA_OUT) | |
2415 | print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index); | |
2416 | ||
2417 | print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index); | |
2418 | } | |
2419 | if (old_PA_IN) | |
2420 | bitmap_set_free (old_PA_IN); | |
2421 | if (PA_OUT) | |
2422 | bitmap_set_free (PA_OUT); | |
2423 | return changed; | |
2424 | } | |
2425 | ||
83737db2 | 2426 | /* Compute ANTIC and partial ANTIC sets. */ |
6de9cd9a DN |
2427 | |
2428 | static void | |
7e6eb623 | 2429 | compute_antic (void) |
6de9cd9a | 2430 | { |
c33bae88 | 2431 | bool changed = true; |
7e6eb623 | 2432 | int num_iterations = 0; |
c33bae88 | 2433 | basic_block block; |
83737db2 | 2434 | int i; |
a28fee03 SB |
2435 | |
2436 | /* If any predecessor edges are abnormal, we punt, so antic_in is empty. | |
2437 | We pre-build the map of blocks with incoming abnormal edges here. */ | |
2438 | has_abnormal_preds = sbitmap_alloc (last_basic_block); | |
f61e445a | 2439 | bitmap_clear (has_abnormal_preds); |
83737db2 | 2440 | |
585d0dc4 | 2441 | FOR_ALL_BB (block) |
7e6eb623 | 2442 | { |
a28fee03 SB |
2443 | edge_iterator ei; |
2444 | edge e; | |
2445 | ||
2446 | FOR_EACH_EDGE (e, ei, block->preds) | |
83737db2 DB |
2447 | { |
2448 | e->flags &= ~EDGE_DFS_BACK; | |
2449 | if (e->flags & EDGE_ABNORMAL) | |
2450 | { | |
d7c028c0 | 2451 | bitmap_set_bit (has_abnormal_preds, block->index); |
83737db2 DB |
2452 | break; |
2453 | } | |
2454 | } | |
a28fee03 | 2455 | |
83737db2 | 2456 | BB_VISITED (block) = 0; |
d75dbccd | 2457 | BB_DEFERRED (block) = 0; |
a19eb9d2 | 2458 | |
a28fee03 | 2459 | /* While we are here, give empty ANTIC_IN sets to each block. */ |
83737db2 | 2460 | ANTIC_IN (block) = bitmap_set_new (); |
d75dbccd | 2461 | PA_IN (block) = bitmap_set_new (); |
7e6eb623 | 2462 | } |
83737db2 | 2463 | |
a28fee03 | 2464 | /* At the exit block we anticipate nothing. */ |
83737db2 | 2465 | BB_VISITED (EXIT_BLOCK_PTR) = 1; |
a28fee03 | 2466 | |
83737db2 | 2467 | changed_blocks = sbitmap_alloc (last_basic_block + 1); |
f61e445a | 2468 | bitmap_ones (changed_blocks); |
7e6eb623 DB |
2469 | while (changed) |
2470 | { | |
83737db2 DB |
2471 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2472 | fprintf (dump_file, "Starting iteration %d\n", num_iterations); | |
3bc27de7 RG |
2473 | /* ??? We need to clear our PHI translation cache here as the |
2474 | ANTIC sets shrink and we restrict valid translations to | |
2475 | those having operands with leaders in ANTIC. Same below | |
2476 | for PA ANTIC computation. */ | |
a28fee03 | 2477 | num_iterations++; |
c33bae88 | 2478 | changed = false; |
585d0dc4 | 2479 | for (i = postorder_num - 1; i >= 0; i--) |
83737db2 | 2480 | { |
d7c028c0 | 2481 | if (bitmap_bit_p (changed_blocks, postorder[i])) |
83737db2 DB |
2482 | { |
2483 | basic_block block = BASIC_BLOCK (postorder[i]); | |
2484 | changed |= compute_antic_aux (block, | |
d7c028c0 | 2485 | bitmap_bit_p (has_abnormal_preds, |
83737db2 DB |
2486 | block->index)); |
2487 | } | |
2488 | } | |
d75dbccd | 2489 | /* Theoretically possible, but *highly* unlikely. */ |
77a74ed7 | 2490 | gcc_checking_assert (num_iterations < 500); |
2e24fa83 | 2491 | } |
a28fee03 | 2492 | |
9fe0cb7d RG |
2493 | statistics_histogram_event (cfun, "compute_antic iterations", |
2494 | num_iterations); | |
83737db2 | 2495 | |
d75dbccd DB |
2496 | if (do_partial_partial) |
2497 | { | |
f61e445a | 2498 | bitmap_ones (changed_blocks); |
d75dbccd DB |
2499 | mark_dfs_back_edges (); |
2500 | num_iterations = 0; | |
2501 | changed = true; | |
2502 | while (changed) | |
2503 | { | |
2504 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2505 | fprintf (dump_file, "Starting iteration %d\n", num_iterations); | |
2506 | num_iterations++; | |
2507 | changed = false; | |
585d0dc4 | 2508 | for (i = postorder_num - 1 ; i >= 0; i--) |
d75dbccd | 2509 | { |
d7c028c0 | 2510 | if (bitmap_bit_p (changed_blocks, postorder[i])) |
d75dbccd DB |
2511 | { |
2512 | basic_block block = BASIC_BLOCK (postorder[i]); | |
2513 | changed | |
2514 | |= compute_partial_antic_aux (block, | |
d7c028c0 | 2515 | bitmap_bit_p (has_abnormal_preds, |
d75dbccd DB |
2516 | block->index)); |
2517 | } | |
2518 | } | |
2519 | /* Theoretically possible, but *highly* unlikely. */ | |
77a74ed7 | 2520 | gcc_checking_assert (num_iterations < 500); |
d75dbccd | 2521 | } |
9fe0cb7d RG |
2522 | statistics_histogram_event (cfun, "compute_partial_antic iterations", |
2523 | num_iterations); | |
d75dbccd | 2524 | } |
83737db2 DB |
2525 | sbitmap_free (has_abnormal_preds); |
2526 | sbitmap_free (changed_blocks); | |
6de9cd9a DN |
2527 | } |
2528 | ||
c90186eb DB |
2529 | |
2530 | /* Inserted expressions are placed onto this worklist, which is used | |
2531 | for performing quick dead code elimination of insertions we made | |
2532 | that didn't turn out to be necessary. */ | |
0ab555de | 2533 | static bitmap inserted_exprs; |
c90186eb | 2534 | |
ce94d354 | 2535 | /* The actual worker for create_component_ref_by_pieces. */ |
b9c5e484 | 2536 | |
85300b46 | 2537 | static tree |
ce94d354 | 2538 | create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref, |
e076271b | 2539 | unsigned int *operand, gimple_seq *stmts) |
85300b46 | 2540 | { |
9771b263 | 2541 | vn_reference_op_t currop = &ref->operands[*operand]; |
c9145754 | 2542 | tree genop; |
ce94d354 | 2543 | ++*operand; |
c9145754 | 2544 | switch (currop->opcode) |
85300b46 | 2545 | { |
c9145754 DB |
2546 | case CALL_EXPR: |
2547 | { | |
b3be2694 | 2548 | tree folded, sc = NULL_TREE; |
ce94d354 | 2549 | unsigned int nargs = 0; |
b3be2694 RG |
2550 | tree fn, *args; |
2551 | if (TREE_CODE (currop->op0) == FUNCTION_DECL) | |
2552 | fn = currop->op0; | |
2553 | else | |
e076271b | 2554 | fn = find_or_generate_expression (block, currop->op0, stmts); |
c3dd8dd7 RB |
2555 | if (!fn) |
2556 | return NULL_TREE; | |
b3be2694 | 2557 | if (currop->op1) |
c3dd8dd7 RB |
2558 | { |
2559 | sc = find_or_generate_expression (block, currop->op1, stmts); | |
2560 | if (!sc) | |
2561 | return NULL_TREE; | |
2562 | } | |
9771b263 DN |
2563 | args = XNEWVEC (tree, ref->operands.length () - 1); |
2564 | while (*operand < ref->operands.length ()) | |
c9145754 | 2565 | { |
ce94d354 | 2566 | args[nargs] = create_component_ref_by_pieces_1 (block, ref, |
e076271b | 2567 | operand, stmts); |
c3dd8dd7 RB |
2568 | if (!args[nargs]) |
2569 | return NULL_TREE; | |
ce94d354 | 2570 | nargs++; |
c9145754 | 2571 | } |
726a989a | 2572 | folded = build_call_array (currop->type, |
b3be2694 RG |
2573 | (TREE_CODE (fn) == FUNCTION_DECL |
2574 | ? build_fold_addr_expr (fn) : fn), | |
726a989a | 2575 | nargs, args); |
c9145754 | 2576 | free (args); |
7aec7a38 | 2577 | if (sc) |
b3be2694 | 2578 | CALL_EXPR_STATIC_CHAIN (folded) = sc; |
c9145754 DB |
2579 | return folded; |
2580 | } | |
ea814c66 | 2581 | |
70f34814 RG |
2582 | case MEM_REF: |
2583 | { | |
2584 | tree baseop = create_component_ref_by_pieces_1 (block, ref, operand, | |
e076271b | 2585 | stmts); |
c3dd8dd7 RB |
2586 | if (!baseop) |
2587 | return NULL_TREE; | |
70f34814 | 2588 | tree offset = currop->op0; |
70f34814 RG |
2589 | if (TREE_CODE (baseop) == ADDR_EXPR |
2590 | && handled_component_p (TREE_OPERAND (baseop, 0))) | |
2591 | { | |
2592 | HOST_WIDE_INT off; | |
2593 | tree base; | |
2594 | base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0), | |
2595 | &off); | |
2596 | gcc_assert (base); | |
2597 | offset = int_const_binop (PLUS_EXPR, offset, | |
2598 | build_int_cst (TREE_TYPE (offset), | |
d35936ab | 2599 | off)); |
70f34814 RG |
2600 | baseop = build_fold_addr_expr (base); |
2601 | } | |
2602 | return fold_build2 (MEM_REF, currop->type, baseop, offset); | |
2603 | } | |
ea814c66 | 2604 | |
150e3929 RG |
2605 | case TARGET_MEM_REF: |
2606 | { | |
4d948885 | 2607 | tree genop0 = NULL_TREE, genop1 = NULL_TREE; |
9771b263 | 2608 | vn_reference_op_t nextop = &ref->operands[++*operand]; |
150e3929 | 2609 | tree baseop = create_component_ref_by_pieces_1 (block, ref, operand, |
e076271b | 2610 | stmts); |
c3dd8dd7 RB |
2611 | if (!baseop) |
2612 | return NULL_TREE; | |
150e3929 | 2613 | if (currop->op0) |
c3dd8dd7 RB |
2614 | { |
2615 | genop0 = find_or_generate_expression (block, currop->op0, stmts); | |
2616 | if (!genop0) | |
2617 | return NULL_TREE; | |
2618 | } | |
4d948885 | 2619 | if (nextop->op0) |
c3dd8dd7 RB |
2620 | { |
2621 | genop1 = find_or_generate_expression (block, nextop->op0, stmts); | |
2622 | if (!genop1) | |
2623 | return NULL_TREE; | |
2624 | } | |
4d948885 RG |
2625 | return build5 (TARGET_MEM_REF, currop->type, |
2626 | baseop, currop->op2, genop0, currop->op1, genop1); | |
150e3929 | 2627 | } |
ea814c66 | 2628 | |
ce94d354 RG |
2629 | case ADDR_EXPR: |
2630 | if (currop->op0) | |
2631 | { | |
2632 | gcc_assert (is_gimple_min_invariant (currop->op0)); | |
2633 | return currop->op0; | |
2634 | } | |
2635 | /* Fallthrough. */ | |
c9145754 DB |
2636 | case REALPART_EXPR: |
2637 | case IMAGPART_EXPR: | |
2638 | case VIEW_CONVERT_EXPR: | |
2639 | { | |
c3dd8dd7 RB |
2640 | tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand, |
2641 | stmts); | |
2642 | if (!genop0) | |
2643 | return NULL_TREE; | |
ea814c66 | 2644 | return fold_build1 (currop->opcode, currop->type, genop0); |
c9145754 | 2645 | } |
ea814c66 | 2646 | |
842679dc TV |
2647 | case WITH_SIZE_EXPR: |
2648 | { | |
2649 | tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand, | |
e076271b | 2650 | stmts); |
c3dd8dd7 RB |
2651 | if (!genop0) |
2652 | return NULL_TREE; | |
e076271b | 2653 | tree genop1 = find_or_generate_expression (block, currop->op0, stmts); |
c3dd8dd7 RB |
2654 | if (!genop1) |
2655 | return NULL_TREE; | |
842679dc TV |
2656 | return fold_build2 (currop->opcode, currop->type, genop0, genop1); |
2657 | } | |
ea814c66 | 2658 | |
c9145754 | 2659 | case BIT_FIELD_REF: |
e13f1c14 | 2660 | { |
ce94d354 | 2661 | tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand, |
e076271b | 2662 | stmts); |
c3dd8dd7 RB |
2663 | if (!genop0) |
2664 | return NULL_TREE; | |
ea814c66 EB |
2665 | tree op1 = currop->op0; |
2666 | tree op2 = currop->op1; | |
ea814c66 | 2667 | return fold_build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2); |
e13f1c14 | 2668 | } |
c9145754 DB |
2669 | |
2670 | /* For array ref vn_reference_op's, operand 1 of the array ref | |
2671 | is op0 of the reference op and operand 3 of the array ref is | |
2672 | op1. */ | |
2673 | case ARRAY_RANGE_REF: | |
2674 | case ARRAY_REF: | |
2675 | { | |
c9145754 DB |
2676 | tree genop0; |
2677 | tree genop1 = currop->op0; | |
c9145754 | 2678 | tree genop2 = currop->op1; |
e52201b6 | 2679 | tree genop3 = currop->op2; |
c3dd8dd7 RB |
2680 | genop0 = create_component_ref_by_pieces_1 (block, ref, operand, |
2681 | stmts); | |
2682 | if (!genop0) | |
2683 | return NULL_TREE; | |
e076271b | 2684 | genop1 = find_or_generate_expression (block, genop1, stmts); |
c3dd8dd7 RB |
2685 | if (!genop1) |
2686 | return NULL_TREE; | |
c9145754 DB |
2687 | if (genop2) |
2688 | { | |
d4d73ce2 EB |
2689 | tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0)); |
2690 | /* Drop zero minimum index if redundant. */ | |
2691 | if (integer_zerop (genop2) | |
2692 | && (!domain_type | |
2693 | || integer_zerop (TYPE_MIN_VALUE (domain_type)))) | |
d53bed0b RG |
2694 | genop2 = NULL_TREE; |
2695 | else | |
c3dd8dd7 RB |
2696 | { |
2697 | genop2 = find_or_generate_expression (block, genop2, stmts); | |
2698 | if (!genop2) | |
2699 | return NULL_TREE; | |
2700 | } | |
c9145754 | 2701 | } |
e52201b6 RG |
2702 | if (genop3) |
2703 | { | |
2704 | tree elmt_type = TREE_TYPE (TREE_TYPE (genop0)); | |
d53bed0b RG |
2705 | /* We can't always put a size in units of the element alignment |
2706 | here as the element alignment may be not visible. See | |
2707 | PR43783. Simply drop the element size for constant | |
2708 | sizes. */ | |
2709 | if (tree_int_cst_equal (genop3, TYPE_SIZE_UNIT (elmt_type))) | |
2710 | genop3 = NULL_TREE; | |
2711 | else | |
2712 | { | |
2713 | genop3 = size_binop (EXACT_DIV_EXPR, genop3, | |
2714 | size_int (TYPE_ALIGN_UNIT (elmt_type))); | |
e076271b | 2715 | genop3 = find_or_generate_expression (block, genop3, stmts); |
c3dd8dd7 RB |
2716 | if (!genop3) |
2717 | return NULL_TREE; | |
d53bed0b | 2718 | } |
e52201b6 | 2719 | } |
c9145754 DB |
2720 | return build4 (currop->opcode, currop->type, genop0, genop1, |
2721 | genop2, genop3); | |
2722 | } | |
85300b46 DB |
2723 | case COMPONENT_REF: |
2724 | { | |
2725 | tree op0; | |
2726 | tree op1; | |
c9145754 | 2727 | tree genop2 = currop->op1; |
e076271b | 2728 | op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts); |
c3dd8dd7 RB |
2729 | if (!op0) |
2730 | return NULL_TREE; | |
e076271b | 2731 | /* op1 should be a FIELD_DECL, which are represented by themselves. */ |
c9145754 DB |
2732 | op1 = currop->op0; |
2733 | if (genop2) | |
c3dd8dd7 RB |
2734 | { |
2735 | genop2 = find_or_generate_expression (block, genop2, stmts); | |
2736 | if (!genop2) | |
2737 | return NULL_TREE; | |
2738 | } | |
ea814c66 | 2739 | return fold_build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2); |
85300b46 | 2740 | } |
ea814c66 | 2741 | |
c9145754 | 2742 | case SSA_NAME: |
85300b46 | 2743 | { |
e076271b | 2744 | genop = find_or_generate_expression (block, currop->op0, stmts); |
c9145754 | 2745 | return genop; |
85300b46 | 2746 | } |
c9145754 DB |
2747 | case STRING_CST: |
2748 | case INTEGER_CST: | |
2749 | case COMPLEX_CST: | |
2750 | case VECTOR_CST: | |
2751 | case REAL_CST: | |
2752 | case CONSTRUCTOR: | |
85300b46 DB |
2753 | case VAR_DECL: |
2754 | case PARM_DECL: | |
c9145754 | 2755 | case CONST_DECL: |
5230d884 | 2756 | case RESULT_DECL: |
c9145754 | 2757 | case FUNCTION_DECL: |
c9145754 DB |
2758 | return currop->op0; |
2759 | ||
85300b46 | 2760 | default: |
b9c5e484 | 2761 | gcc_unreachable (); |
85300b46 | 2762 | } |
85300b46 | 2763 | } |
c90186eb | 2764 | |
ce94d354 | 2765 | /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the |
70f34814 | 2766 | COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with |
ce94d354 RG |
2767 | trying to rename aggregates into ssa form directly, which is a no no. |
2768 | ||
2769 | Thus, this routine doesn't create temporaries, it just builds a | |
2770 | single access expression for the array, calling | |
2771 | find_or_generate_expression to build the innermost pieces. | |
2772 | ||
2773 | This function is a subroutine of create_expression_by_pieces, and | |
2774 | should not be called on it's own unless you really know what you | |
2775 | are doing. */ | |
2776 | ||
2777 | static tree | |
2778 | create_component_ref_by_pieces (basic_block block, vn_reference_t ref, | |
e076271b | 2779 | gimple_seq *stmts) |
ce94d354 RG |
2780 | { |
2781 | unsigned int op = 0; | |
e076271b | 2782 | return create_component_ref_by_pieces_1 (block, ref, &op, stmts); |
ce94d354 RG |
2783 | } |
2784 | ||
c3dd8dd7 RB |
2785 | /* Find a simple leader for an expression, or generate one using |
2786 | create_expression_by_pieces from a NARY expression for the value. | |
56db793a | 2787 | BLOCK is the basic_block we are looking for leaders in. |
e076271b | 2788 | OP is the tree expression to find a leader for or generate. |
c3dd8dd7 | 2789 | Returns the leader or NULL_TREE on failure. */ |
56db793a DB |
2790 | |
2791 | static tree | |
e076271b | 2792 | find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts) |
56db793a | 2793 | { |
e076271b RG |
2794 | pre_expr expr = get_or_alloc_expr_for (op); |
2795 | unsigned int lookfor = get_expr_value_id (expr); | |
2796 | pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor); | |
c9145754 DB |
2797 | if (leader) |
2798 | { | |
2799 | if (leader->kind == NAME) | |
e076271b | 2800 | return PRE_EXPR_NAME (leader); |
c9145754 | 2801 | else if (leader->kind == CONSTANT) |
e076271b | 2802 | return PRE_EXPR_CONSTANT (leader); |
c3dd8dd7 RB |
2803 | |
2804 | /* Defer. */ | |
2805 | return NULL_TREE; | |
c9145754 | 2806 | } |
e9284566 | 2807 | |
c3dd8dd7 RB |
2808 | /* It must be a complex expression, so generate it recursively. Note |
2809 | that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c | |
2810 | where the insert algorithm fails to insert a required expression. */ | |
9771b263 | 2811 | bitmap exprset = value_expressions[lookfor]; |
e076271b RG |
2812 | bitmap_iterator bi; |
2813 | unsigned int i; | |
2814 | EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi) | |
56db793a | 2815 | { |
e076271b | 2816 | pre_expr temp = expression_for_id (i); |
c3dd8dd7 RB |
2817 | /* We cannot insert random REFERENCE expressions at arbitrary |
2818 | places. We can insert NARYs which eventually re-materializes | |
2819 | its operand values. */ | |
2820 | if (temp->kind == NARY) | |
e076271b RG |
2821 | return create_expression_by_pieces (block, temp, stmts, |
2822 | get_expr_type (expr)); | |
56db793a | 2823 | } |
e076271b | 2824 | |
c3dd8dd7 RB |
2825 | /* Defer. */ |
2826 | return NULL_TREE; | |
56db793a DB |
2827 | } |
2828 | ||
726a989a | 2829 | #define NECESSARY GF_PLF_1 |
c9145754 | 2830 | |
56db793a | 2831 | /* Create an expression in pieces, so that we can handle very complex |
b9c5e484 | 2832 | expressions that may be ANTIC, but not necessary GIMPLE. |
56db793a DB |
2833 | BLOCK is the basic block the expression will be inserted into, |
2834 | EXPR is the expression to insert (in value form) | |
2835 | STMTS is a statement list to append the necessary insertions into. | |
2836 | ||
0e61db61 | 2837 | This function will die if we hit some value that shouldn't be |
56db793a | 2838 | ANTIC but is (IE there is no leader for it, or its components). |
c3dd8dd7 RB |
2839 | The function returns NULL_TREE in case a different antic expression |
2840 | has to be inserted first. | |
56db793a DB |
2841 | This function may also generate expressions that are themselves |
2842 | partially or fully redundant. Those that are will be either made | |
2843 | fully redundant during the next iteration of insert (for partially | |
2844 | redundant ones), or eliminated by eliminate (for fully redundant | |
c3dd8dd7 | 2845 | ones). */ |
56db793a DB |
2846 | |
2847 | static tree | |
726a989a | 2848 | create_expression_by_pieces (basic_block block, pre_expr expr, |
e076271b | 2849 | gimple_seq *stmts, tree type) |
56db793a | 2850 | { |
83d5977e | 2851 | tree name; |
150e3929 RG |
2852 | tree folded; |
2853 | gimple_seq forced_stmts = NULL; | |
c9145754 | 2854 | unsigned int value_id; |
726a989a | 2855 | gimple_stmt_iterator gsi; |
c9145754 DB |
2856 | tree exprtype = type ? type : get_expr_type (expr); |
2857 | pre_expr nameexpr; | |
726a989a | 2858 | gimple newstmt; |
81c4f554 | 2859 | |
c9145754 | 2860 | switch (expr->kind) |
56db793a | 2861 | { |
c9145754 DB |
2862 | /* We may hit the NAME/CONSTANT case if we have to convert types |
2863 | that value numbering saw through. */ | |
2864 | case NAME: | |
2865 | folded = PRE_EXPR_NAME (expr); | |
2866 | break; | |
2867 | case CONSTANT: | |
2868 | folded = PRE_EXPR_CONSTANT (expr); | |
2869 | break; | |
2870 | case REFERENCE: | |
43da81be | 2871 | { |
c9145754 | 2872 | vn_reference_t ref = PRE_EXPR_REFERENCE (expr); |
e076271b | 2873 | folded = create_component_ref_by_pieces (block, ref, stmts); |
c3dd8dd7 RB |
2874 | if (!folded) |
2875 | return NULL_TREE; | |
43da81be DB |
2876 | } |
2877 | break; | |
c9145754 | 2878 | case NARY: |
c90186eb | 2879 | { |
c9145754 | 2880 | vn_nary_op_t nary = PRE_EXPR_NARY (expr); |
f843144b | 2881 | tree *genop = XALLOCAVEC (tree, nary->length); |
5a7d7f9c RG |
2882 | unsigned i; |
2883 | for (i = 0; i < nary->length; ++i) | |
85300b46 | 2884 | { |
e076271b | 2885 | genop[i] = find_or_generate_expression (block, nary->op[i], stmts); |
c3dd8dd7 RB |
2886 | if (!genop[i]) |
2887 | return NULL_TREE; | |
48acf1b7 RG |
2888 | /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It |
2889 | may have conversions stripped. */ | |
2890 | if (nary->opcode == POINTER_PLUS_EXPR) | |
2891 | { | |
2892 | if (i == 0) | |
2893 | genop[i] = fold_convert (nary->type, genop[i]); | |
2894 | else if (i == 1) | |
2895 | genop[i] = convert_to_ptrofftype (genop[i]); | |
2896 | } | |
5a7d7f9c RG |
2897 | else |
2898 | genop[i] = fold_convert (TREE_TYPE (nary->op[i]), genop[i]); | |
2899 | } | |
2900 | if (nary->opcode == CONSTRUCTOR) | |
2901 | { | |
9771b263 | 2902 | vec<constructor_elt, va_gc> *elts = NULL; |
5a7d7f9c RG |
2903 | for (i = 0; i < nary->length; ++i) |
2904 | CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]); | |
2905 | folded = build_constructor (nary->type, elts); | |
2906 | } | |
2907 | else | |
2908 | { | |
2909 | switch (nary->length) | |
2910 | { | |
2911 | case 1: | |
2912 | folded = fold_build1 (nary->opcode, nary->type, | |
2913 | genop[0]); | |
2914 | break; | |
2915 | case 2: | |
2916 | folded = fold_build2 (nary->opcode, nary->type, | |
2917 | genop[0], genop[1]); | |
2918 | break; | |
2919 | case 3: | |
2920 | folded = fold_build3 (nary->opcode, nary->type, | |
a475fd3d | 2921 | genop[0], genop[1], genop[2]); |
5a7d7f9c RG |
2922 | break; |
2923 | default: | |
2924 | gcc_unreachable (); | |
2925 | } | |
85300b46 | 2926 | } |
56db793a | 2927 | } |
c9145754 | 2928 | break; |
56db793a | 2929 | default: |
e076271b | 2930 | gcc_unreachable (); |
56db793a | 2931 | } |
150e3929 RG |
2932 | |
2933 | if (!useless_type_conversion_p (exprtype, TREE_TYPE (folded))) | |
2934 | folded = fold_convert (exprtype, folded); | |
2935 | ||
81c4f554 SB |
2936 | /* Force the generated expression to be a sequence of GIMPLE |
2937 | statements. | |
2938 | We have to call unshare_expr because force_gimple_operand may | |
2939 | modify the tree we pass to it. */ | |
150e3929 RG |
2940 | folded = force_gimple_operand (unshare_expr (folded), &forced_stmts, |
2941 | false, NULL); | |
81c4f554 SB |
2942 | |
2943 | /* If we have any intermediate expressions to the value sets, add them | |
a7849637 | 2944 | to the value sets and chain them in the instruction stream. */ |
81c4f554 SB |
2945 | if (forced_stmts) |
2946 | { | |
726a989a RB |
2947 | gsi = gsi_start (forced_stmts); |
2948 | for (; !gsi_end_p (gsi); gsi_next (&gsi)) | |
81c4f554 | 2949 | { |
726a989a RB |
2950 | gimple stmt = gsi_stmt (gsi); |
2951 | tree forcedname = gimple_get_lhs (stmt); | |
c9145754 | 2952 | pre_expr nameexpr; |
b9c5e484 | 2953 | |
c9145754 DB |
2954 | if (TREE_CODE (forcedname) == SSA_NAME) |
2955 | { | |
0ab555de | 2956 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname)); |
c9145754 DB |
2957 | VN_INFO_GET (forcedname)->valnum = forcedname; |
2958 | VN_INFO (forcedname)->value_id = get_next_value_id (); | |
2959 | nameexpr = get_or_alloc_expr_for_name (forcedname); | |
2960 | add_to_value (VN_INFO (forcedname)->value_id, nameexpr); | |
40b178f4 | 2961 | bitmap_value_replace_in_set (NEW_SETS (block), nameexpr); |
c9145754 DB |
2962 | bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr); |
2963 | } | |
81c4f554 | 2964 | } |
726a989a | 2965 | gimple_seq_add_seq (stmts, forced_stmts); |
81c4f554 SB |
2966 | } |
2967 | ||
83d5977e RG |
2968 | name = make_temp_ssa_name (exprtype, NULL, "pretmp"); |
2969 | newstmt = gimple_build_assign (name, folded); | |
726a989a | 2970 | gimple_set_plf (newstmt, NECESSARY, false); |
c90186eb | 2971 | |
726a989a | 2972 | gimple_seq_add_stmt (stmts, newstmt); |
0ab555de | 2973 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name)); |
cfaab3a9 | 2974 | |
0aa16586 RG |
2975 | /* Fold the last statement. */ |
2976 | gsi = gsi_last (*stmts); | |
748c5114 RG |
2977 | if (fold_stmt_inplace (&gsi)) |
2978 | update_stmt (gsi_stmt (gsi)); | |
0aa16586 | 2979 | |
c9145754 | 2980 | /* Add a value number to the temporary. |
81c4f554 | 2981 | The value may already exist in either NEW_SETS, or AVAIL_OUT, because |
e9284566 DB |
2982 | we are creating the expression by pieces, and this particular piece of |
2983 | the expression may have been represented. There is no harm in replacing | |
2984 | here. */ | |
c9145754 | 2985 | value_id = get_expr_value_id (expr); |
40b178f4 RG |
2986 | VN_INFO_GET (name)->value_id = value_id; |
2987 | VN_INFO (name)->valnum = sccvn_valnum_from_value_id (value_id); | |
2988 | if (VN_INFO (name)->valnum == NULL_TREE) | |
2989 | VN_INFO (name)->valnum = name; | |
2990 | gcc_assert (VN_INFO (name)->valnum != NULL_TREE); | |
c9145754 DB |
2991 | nameexpr = get_or_alloc_expr_for_name (name); |
2992 | add_to_value (value_id, nameexpr); | |
70f34814 | 2993 | if (NEW_SETS (block)) |
c9145754 DB |
2994 | bitmap_value_replace_in_set (NEW_SETS (block), nameexpr); |
2995 | bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr); | |
81c4f554 SB |
2996 | |
2997 | pre_stats.insertions++; | |
56db793a | 2998 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b9c5e484 | 2999 | { |
56db793a | 3000 | fprintf (dump_file, "Inserted "); |
726a989a | 3001 | print_gimple_stmt (dump_file, newstmt, 0, 0); |
984af6ac RB |
3002 | fprintf (dump_file, " in predecessor %d (%04d)\n", |
3003 | block->index, value_id); | |
56db793a | 3004 | } |
81c4f554 | 3005 | |
56db793a DB |
3006 | return name; |
3007 | } | |
e9284566 | 3008 | |
c9145754 | 3009 | |
a8338640 MM |
3010 | /* Returns true if we want to inhibit the insertions of PHI nodes |
3011 | for the given EXPR for basic block BB (a member of a loop). | |
3012 | We want to do this, when we fear that the induction variable we | |
3013 | create might inhibit vectorization. */ | |
3014 | ||
3015 | static bool | |
3016 | inhibit_phi_insertion (basic_block bb, pre_expr expr) | |
3017 | { | |
3018 | vn_reference_t vr = PRE_EXPR_REFERENCE (expr); | |
9771b263 | 3019 | vec<vn_reference_op_s> ops = vr->operands; |
a8338640 MM |
3020 | vn_reference_op_t op; |
3021 | unsigned i; | |
3022 | ||
3023 | /* If we aren't going to vectorize we don't inhibit anything. */ | |
ea0f3e87 | 3024 | if (!flag_tree_loop_vectorize) |
a8338640 MM |
3025 | return false; |
3026 | ||
3027 | /* Otherwise we inhibit the insertion when the address of the | |
3028 | memory reference is a simple induction variable. In other | |
3029 | cases the vectorizer won't do anything anyway (either it's | |
3030 | loop invariant or a complicated expression). */ | |
9771b263 | 3031 | FOR_EACH_VEC_ELT (ops, i, op) |
a8338640 MM |
3032 | { |
3033 | switch (op->opcode) | |
3034 | { | |
109a16c2 RG |
3035 | case CALL_EXPR: |
3036 | /* Calls are not a problem. */ | |
3037 | return false; | |
3038 | ||
a8338640 MM |
3039 | case ARRAY_REF: |
3040 | case ARRAY_RANGE_REF: | |
3041 | if (TREE_CODE (op->op0) != SSA_NAME) | |
3042 | break; | |
3043 | /* Fallthru. */ | |
3044 | case SSA_NAME: | |
3045 | { | |
3046 | basic_block defbb = gimple_bb (SSA_NAME_DEF_STMT (op->op0)); | |
3047 | affine_iv iv; | |
3048 | /* Default defs are loop invariant. */ | |
3049 | if (!defbb) | |
3050 | break; | |
3051 | /* Defined outside this loop, also loop invariant. */ | |
3052 | if (!flow_bb_inside_loop_p (bb->loop_father, defbb)) | |
3053 | break; | |
3054 | /* If it's a simple induction variable inhibit insertion, | |
3055 | the vectorizer might be interested in this one. */ | |
3056 | if (simple_iv (bb->loop_father, bb->loop_father, | |
3057 | op->op0, &iv, true)) | |
3058 | return true; | |
3059 | /* No simple IV, vectorizer can't do anything, hence no | |
3060 | reason to inhibit the transformation for this operand. */ | |
3061 | break; | |
3062 | } | |
3063 | default: | |
3064 | break; | |
3065 | } | |
3066 | } | |
3067 | return false; | |
3068 | } | |
3069 | ||
83737db2 | 3070 | /* Insert the to-be-made-available values of expression EXPRNUM for each |
c90186eb | 3071 | predecessor, stored in AVAIL, into the predecessors of BLOCK, and |
c9145754 | 3072 | merge the result with a phi node, given the same value number as |
c90186eb | 3073 | NODE. Return true if we have inserted new stuff. */ |
e9284566 DB |
3074 | |
3075 | static bool | |
83737db2 | 3076 | insert_into_preds_of_block (basic_block block, unsigned int exprnum, |
9771b263 | 3077 | vec<pre_expr> avail) |
e9284566 | 3078 | { |
c9145754 DB |
3079 | pre_expr expr = expression_for_id (exprnum); |
3080 | pre_expr newphi; | |
3081 | unsigned int val = get_expr_value_id (expr); | |
e9284566 | 3082 | edge pred; |
0fc6c492 DB |
3083 | bool insertions = false; |
3084 | bool nophi = false; | |
e9284566 | 3085 | basic_block bprime; |
c9145754 | 3086 | pre_expr eprime; |
e9284566 | 3087 | edge_iterator ei; |
c9145754 | 3088 | tree type = get_expr_type (expr); |
de081cfd | 3089 | tree temp; |
726a989a | 3090 | gimple phi; |
b9c5e484 | 3091 | |
0fc6c492 | 3092 | /* Make sure we aren't creating an induction variable. */ |
391886c8 | 3093 | if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2) |
0fc6c492 DB |
3094 | { |
3095 | bool firstinsideloop = false; | |
3096 | bool secondinsideloop = false; | |
b9c5e484 | 3097 | firstinsideloop = flow_bb_inside_loop_p (block->loop_father, |
0fc6c492 DB |
3098 | EDGE_PRED (block, 0)->src); |
3099 | secondinsideloop = flow_bb_inside_loop_p (block->loop_father, | |
3100 | EDGE_PRED (block, 1)->src); | |
3101 | /* Induction variables only have one edge inside the loop. */ | |
a8338640 MM |
3102 | if ((firstinsideloop ^ secondinsideloop) |
3103 | && (expr->kind != REFERENCE | |
3104 | || inhibit_phi_insertion (block, expr))) | |
0fc6c492 DB |
3105 | { |
3106 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3107 | fprintf (dump_file, "Skipping insertion of phi for partial redundancy: Looks like an induction variable\n"); | |
3108 | nophi = true; | |
3109 | } | |
3110 | } | |
b9c5e484 | 3111 | |
c9145754 DB |
3112 | /* Make the necessary insertions. */ |
3113 | FOR_EACH_EDGE (pred, ei, block->preds) | |
3114 | { | |
726a989a | 3115 | gimple_seq stmts = NULL; |
c9145754 DB |
3116 | tree builtexpr; |
3117 | bprime = pred->src; | |
9771b263 | 3118 | eprime = avail[pred->dest_idx]; |
c9145754 DB |
3119 | |
3120 | if (eprime->kind != NAME && eprime->kind != CONSTANT) | |
3121 | { | |
e076271b RG |
3122 | builtexpr = create_expression_by_pieces (bprime, eprime, |
3123 | &stmts, type); | |
c9145754 | 3124 | gcc_assert (!(pred->flags & EDGE_ABNORMAL)); |
726a989a | 3125 | gsi_insert_seq_on_edge (pred, stmts); |
c3dd8dd7 RB |
3126 | if (!builtexpr) |
3127 | { | |
3128 | /* We cannot insert a PHI node if we failed to insert | |
3129 | on one edge. */ | |
3130 | nophi = true; | |
3131 | continue; | |
3132 | } | |
9771b263 | 3133 | avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr); |
c9145754 DB |
3134 | insertions = true; |
3135 | } | |
3136 | else if (eprime->kind == CONSTANT) | |
3137 | { | |
3138 | /* Constants may not have the right type, fold_convert | |
1c8f7377 | 3139 | should give us back a constant with the right type. */ |
c9145754 | 3140 | tree constant = PRE_EXPR_CONSTANT (eprime); |
15d5fe33 | 3141 | if (!useless_type_conversion_p (type, TREE_TYPE (constant))) |
c9145754 DB |
3142 | { |
3143 | tree builtexpr = fold_convert (type, constant); | |
b8698a0f | 3144 | if (!is_gimple_min_invariant (builtexpr)) |
c9145754 DB |
3145 | { |
3146 | tree forcedexpr = force_gimple_operand (builtexpr, | |
3147 | &stmts, true, | |
3148 | NULL); | |
15d5fe33 | 3149 | if (!is_gimple_min_invariant (forcedexpr)) |
c9145754 DB |
3150 | { |
3151 | if (forcedexpr != builtexpr) | |
3152 | { | |
3153 | VN_INFO_GET (forcedexpr)->valnum = PRE_EXPR_CONSTANT (eprime); | |
3154 | VN_INFO (forcedexpr)->value_id = get_expr_value_id (eprime); | |
3155 | } | |
3156 | if (stmts) | |
3157 | { | |
726a989a RB |
3158 | gimple_stmt_iterator gsi; |
3159 | gsi = gsi_start (stmts); | |
3160 | for (; !gsi_end_p (gsi); gsi_next (&gsi)) | |
c9145754 | 3161 | { |
726a989a | 3162 | gimple stmt = gsi_stmt (gsi); |
0ab555de MM |
3163 | tree lhs = gimple_get_lhs (stmt); |
3164 | if (TREE_CODE (lhs) == SSA_NAME) | |
3165 | bitmap_set_bit (inserted_exprs, | |
3166 | SSA_NAME_VERSION (lhs)); | |
726a989a | 3167 | gimple_set_plf (stmt, NECESSARY, false); |
c9145754 | 3168 | } |
726a989a | 3169 | gsi_insert_seq_on_edge (pred, stmts); |
c9145754 | 3170 | } |
9771b263 DN |
3171 | avail[pred->dest_idx] |
3172 | = get_or_alloc_expr_for_name (forcedexpr); | |
c9145754 DB |
3173 | } |
3174 | } | |
70f34814 | 3175 | else |
9771b263 DN |
3176 | avail[pred->dest_idx] |
3177 | = get_or_alloc_expr_for_constant (builtexpr); | |
c9145754 DB |
3178 | } |
3179 | } | |
3180 | else if (eprime->kind == NAME) | |
3181 | { | |
3182 | /* We may have to do a conversion because our value | |
3183 | numbering can look through types in certain cases, but | |
3184 | our IL requires all operands of a phi node have the same | |
3185 | type. */ | |
3186 | tree name = PRE_EXPR_NAME (eprime); | |
f709638a | 3187 | if (!useless_type_conversion_p (type, TREE_TYPE (name))) |
c9145754 DB |
3188 | { |
3189 | tree builtexpr; | |
3190 | tree forcedexpr; | |
f709638a | 3191 | builtexpr = fold_convert (type, name); |
c9145754 DB |
3192 | forcedexpr = force_gimple_operand (builtexpr, |
3193 | &stmts, true, | |
3194 | NULL); | |
3195 | ||
3196 | if (forcedexpr != name) | |
3197 | { | |
3198 | VN_INFO_GET (forcedexpr)->valnum = VN_INFO (name)->valnum; | |
3199 | VN_INFO (forcedexpr)->value_id = VN_INFO (name)->value_id; | |
3200 | } | |
c90186eb | 3201 | |
c9145754 DB |
3202 | if (stmts) |
3203 | { | |
726a989a RB |
3204 | gimple_stmt_iterator gsi; |
3205 | gsi = gsi_start (stmts); | |
3206 | for (; !gsi_end_p (gsi); gsi_next (&gsi)) | |
c9145754 | 3207 | { |
726a989a | 3208 | gimple stmt = gsi_stmt (gsi); |
0ab555de MM |
3209 | tree lhs = gimple_get_lhs (stmt); |
3210 | if (TREE_CODE (lhs) == SSA_NAME) | |
3211 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (lhs)); | |
726a989a | 3212 | gimple_set_plf (stmt, NECESSARY, false); |
c9145754 | 3213 | } |
726a989a | 3214 | gsi_insert_seq_on_edge (pred, stmts); |
c9145754 | 3215 | } |
9771b263 | 3216 | avail[pred->dest_idx] = get_or_alloc_expr_for_name (forcedexpr); |
c9145754 | 3217 | } |
b9c5e484 | 3218 | } |
e9284566 | 3219 | } |
0fc6c492 DB |
3220 | /* If we didn't want a phi node, and we made insertions, we still have |
3221 | inserted new stuff, and thus return true. If we didn't want a phi node, | |
3222 | and didn't make insertions, we haven't added anything new, so return | |
3223 | false. */ | |
3224 | if (nophi && insertions) | |
3225 | return true; | |
3226 | else if (nophi && !insertions) | |
3227 | return false; | |
3228 | ||
e9284566 | 3229 | /* Now build a phi for the new variable. */ |
83d5977e | 3230 | temp = make_temp_ssa_name (type, NULL, "prephitmp"); |
1e52075c | 3231 | phi = create_phi_node (temp, block); |
de081cfd RG |
3232 | |
3233 | gimple_set_plf (phi, NECESSARY, false); | |
40b178f4 RG |
3234 | VN_INFO_GET (temp)->value_id = val; |
3235 | VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val); | |
3236 | if (VN_INFO (temp)->valnum == NULL_TREE) | |
3237 | VN_INFO (temp)->valnum = temp; | |
3238 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp)); | |
e9284566 | 3239 | FOR_EACH_EDGE (pred, ei, block->preds) |
c9145754 | 3240 | { |
9771b263 | 3241 | pre_expr ae = avail[pred->dest_idx]; |
c9145754 DB |
3242 | gcc_assert (get_expr_type (ae) == type |
3243 | || useless_type_conversion_p (type, get_expr_type (ae))); | |
3244 | if (ae->kind == CONSTANT) | |
2724573f RB |
3245 | add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)), |
3246 | pred, UNKNOWN_LOCATION); | |
c9145754 | 3247 | else |
1c8f7377 | 3248 | add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION); |
c9145754 | 3249 | } |
b9c5e484 | 3250 | |
40b178f4 | 3251 | newphi = get_or_alloc_expr_for_name (temp); |
c9145754 | 3252 | add_to_value (val, newphi); |
b9c5e484 | 3253 | |
e9284566 DB |
3254 | /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing |
3255 | this insertion, since we test for the existence of this value in PHI_GEN | |
3256 | before proceeding with the partial redundancy checks in insert_aux. | |
b9c5e484 | 3257 | |
e9284566 DB |
3258 | The value may exist in AVAIL_OUT, in particular, it could be represented |
3259 | by the expression we are trying to eliminate, in which case we want the | |
3260 | replacement to occur. If it's not existing in AVAIL_OUT, we want it | |
3261 | inserted there. | |
b9c5e484 | 3262 | |
e9284566 DB |
3263 | Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of |
3264 | this block, because if it did, it would have existed in our dominator's | |
3265 | AVAIL_OUT, and would have been skipped due to the full redundancy check. | |
3266 | */ | |
3267 | ||
c9145754 | 3268 | bitmap_insert_into_set (PHI_GEN (block), newphi); |
b9c5e484 | 3269 | bitmap_value_replace_in_set (AVAIL_OUT (block), |
c9145754 | 3270 | newphi); |
e9284566 | 3271 | bitmap_insert_into_set (NEW_SETS (block), |
c9145754 | 3272 | newphi); |
b9c5e484 | 3273 | |
e9284566 DB |
3274 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3275 | { | |
3276 | fprintf (dump_file, "Created phi "); | |
726a989a | 3277 | print_gimple_stmt (dump_file, phi, 0, 0); |
984af6ac | 3278 | fprintf (dump_file, " in block %d (%04d)\n", block->index, val); |
e9284566 DB |
3279 | } |
3280 | pre_stats.phis++; | |
3281 | return true; | |
3282 | } | |
3283 | ||
3284 | ||
b9c5e484 | 3285 | |
7e6eb623 DB |
3286 | /* Perform insertion of partially redundant values. |
3287 | For BLOCK, do the following: | |
3288 | 1. Propagate the NEW_SETS of the dominator into the current block. | |
b9c5e484 | 3289 | If the block has multiple predecessors, |
7e6eb623 | 3290 | 2a. Iterate over the ANTIC expressions for the block to see if |
83737db2 | 3291 | any of them are partially redundant. |
7e6eb623 | 3292 | 2b. If so, insert them into the necessary predecessors to make |
83737db2 | 3293 | the expression fully redundant. |
7e6eb623 DB |
3294 | 2c. Insert a new PHI merging the values of the predecessors. |
3295 | 2d. Insert the new PHI, and the new expressions, into the | |
83737db2 | 3296 | NEW_SETS set. |
7e6eb623 | 3297 | 3. Recursively call ourselves on the dominator children of BLOCK. |
6de9cd9a | 3298 | |
83737db2 | 3299 | Steps 1, 2a, and 3 are done by insert_aux. 2b, 2c and 2d are done by |
d75dbccd | 3300 | do_regular_insertion and do_partial_insertion. |
83737db2 | 3301 | |
7e6eb623 | 3302 | */ |
e9284566 | 3303 | |
83737db2 DB |
3304 | static bool |
3305 | do_regular_insertion (basic_block block, basic_block dom) | |
3306 | { | |
3307 | bool new_stuff = false; | |
9771b263 | 3308 | vec<pre_expr> exprs; |
c9145754 | 3309 | pre_expr expr; |
6e1aa848 | 3310 | vec<pre_expr> avail = vNULL; |
83737db2 DB |
3311 | int i; |
3312 | ||
1c8f7377 | 3313 | exprs = sorted_array_from_bitmap_set (ANTIC_IN (block)); |
9771b263 | 3314 | avail.safe_grow (EDGE_COUNT (block->preds)); |
1c8f7377 | 3315 | |
9771b263 | 3316 | FOR_EACH_VEC_ELT (exprs, i, expr) |
83737db2 | 3317 | { |
4bcc5786 RB |
3318 | if (expr->kind == NARY |
3319 | || expr->kind == REFERENCE) | |
83737db2 | 3320 | { |
c9145754 | 3321 | unsigned int val; |
83737db2 DB |
3322 | bool by_some = false; |
3323 | bool cant_insert = false; | |
3324 | bool all_same = true; | |
c9145754 | 3325 | pre_expr first_s = NULL; |
83737db2 DB |
3326 | edge pred; |
3327 | basic_block bprime; | |
c9145754 | 3328 | pre_expr eprime = NULL; |
83737db2 | 3329 | edge_iterator ei; |
f11d2f1e | 3330 | pre_expr edoubleprime = NULL; |
5813994e | 3331 | bool do_insertion = false; |
83737db2 | 3332 | |
c9145754 | 3333 | val = get_expr_value_id (expr); |
83737db2 DB |
3334 | if (bitmap_set_contains_value (PHI_GEN (block), val)) |
3335 | continue; | |
3336 | if (bitmap_set_contains_value (AVAIL_OUT (dom), val)) | |
3337 | { | |
3338 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
ead69dac JL |
3339 | { |
3340 | fprintf (dump_file, "Found fully redundant value: "); | |
3341 | print_pre_expr (dump_file, expr); | |
3342 | fprintf (dump_file, "\n"); | |
3343 | } | |
83737db2 DB |
3344 | continue; |
3345 | } | |
3346 | ||
83737db2 DB |
3347 | FOR_EACH_EDGE (pred, ei, block->preds) |
3348 | { | |
c9145754 | 3349 | unsigned int vprime; |
83737db2 | 3350 | |
c4ab2baa RG |
3351 | /* We should never run insertion for the exit block |
3352 | and so not come across fake pred edges. */ | |
3353 | gcc_assert (!(pred->flags & EDGE_FAKE)); | |
83737db2 DB |
3354 | bprime = pred->src; |
3355 | eprime = phi_translate (expr, ANTIC_IN (block), NULL, | |
3356 | bprime, block); | |
3357 | ||
3358 | /* eprime will generally only be NULL if the | |
3359 | value of the expression, translated | |
3360 | through the PHI for this predecessor, is | |
3361 | undefined. If that is the case, we can't | |
3362 | make the expression fully redundant, | |
3363 | because its value is undefined along a | |
3364 | predecessor path. We can thus break out | |
3365 | early because it doesn't matter what the | |
3366 | rest of the results are. */ | |
3367 | if (eprime == NULL) | |
3368 | { | |
9771b263 | 3369 | avail[pred->dest_idx] = NULL; |
83737db2 DB |
3370 | cant_insert = true; |
3371 | break; | |
3372 | } | |
3373 | ||
3374 | eprime = fully_constant_expression (eprime); | |
726a989a RB |
3375 | vprime = get_expr_value_id (eprime); |
3376 | edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), | |
e076271b | 3377 | vprime); |
83737db2 DB |
3378 | if (edoubleprime == NULL) |
3379 | { | |
9771b263 | 3380 | avail[pred->dest_idx] = eprime; |
83737db2 DB |
3381 | all_same = false; |
3382 | } | |
3383 | else | |
3384 | { | |
9771b263 | 3385 | avail[pred->dest_idx] = edoubleprime; |
83737db2 | 3386 | by_some = true; |
5813994e RG |
3387 | /* We want to perform insertions to remove a redundancy on |
3388 | a path in the CFG we want to optimize for speed. */ | |
3389 | if (optimize_edge_for_speed_p (pred)) | |
3390 | do_insertion = true; | |
83737db2 DB |
3391 | if (first_s == NULL) |
3392 | first_s = edoubleprime; | |
5deac340 | 3393 | else if (!pre_expr_d::equal (first_s, edoubleprime)) |
83737db2 DB |
3394 | all_same = false; |
3395 | } | |
3396 | } | |
3397 | /* If we can insert it, it's not the same value | |
3398 | already existing along every predecessor, and | |
3399 | it's defined by some predecessor, it is | |
3400 | partially redundant. */ | |
3bc27de7 | 3401 | if (!cant_insert && !all_same && by_some) |
83737db2 | 3402 | { |
3bc27de7 RG |
3403 | if (!do_insertion) |
3404 | { | |
3405 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3406 | { | |
3407 | fprintf (dump_file, "Skipping partial redundancy for " | |
3408 | "expression "); | |
3409 | print_pre_expr (dump_file, expr); | |
3410 | fprintf (dump_file, " (%04d), no redundancy on to be " | |
3411 | "optimized for speed edge\n", val); | |
3412 | } | |
3413 | } | |
19372838 RG |
3414 | else if (dbg_cnt (treepre_insert)) |
3415 | { | |
3416 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3417 | { | |
3418 | fprintf (dump_file, "Found partial redundancy for " | |
3419 | "expression "); | |
3420 | print_pre_expr (dump_file, expr); | |
3421 | fprintf (dump_file, " (%04d)\n", | |
3422 | get_expr_value_id (expr)); | |
3423 | } | |
3424 | if (insert_into_preds_of_block (block, | |
3425 | get_expression_id (expr), | |
3426 | avail)) | |
3427 | new_stuff = true; | |
3428 | } | |
83737db2 DB |
3429 | } |
3430 | /* If all edges produce the same value and that value is | |
3431 | an invariant, then the PHI has the same value on all | |
3432 | edges. Note this. */ | |
4bcc5786 | 3433 | else if (!cant_insert && all_same) |
83737db2 | 3434 | { |
4bcc5786 RB |
3435 | gcc_assert (edoubleprime->kind == CONSTANT |
3436 | || edoubleprime->kind == NAME); | |
3437 | ||
3438 | tree temp = make_temp_ssa_name (get_expr_type (expr), | |
3439 | NULL, "pretmp"); | |
3440 | gimple assign = gimple_build_assign (temp, | |
3441 | edoubleprime->kind == CONSTANT ? PRE_EXPR_CONSTANT (edoubleprime) : PRE_EXPR_NAME (edoubleprime)); | |
3442 | gimple_stmt_iterator gsi = gsi_after_labels (block); | |
3443 | gsi_insert_before (&gsi, assign, GSI_NEW_STMT); | |
3444 | ||
3445 | gimple_set_plf (assign, NECESSARY, false); | |
3446 | VN_INFO_GET (temp)->value_id = val; | |
3447 | VN_INFO (temp)->valnum = sccvn_valnum_from_value_id (val); | |
3448 | if (VN_INFO (temp)->valnum == NULL_TREE) | |
3449 | VN_INFO (temp)->valnum = temp; | |
3450 | bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp)); | |
3451 | pre_expr newe = get_or_alloc_expr_for_name (temp); | |
3452 | add_to_value (val, newe); | |
3453 | bitmap_value_replace_in_set (AVAIL_OUT (block), newe); | |
3454 | bitmap_insert_into_set (NEW_SETS (block), newe); | |
83737db2 | 3455 | } |
83737db2 DB |
3456 | } |
3457 | } | |
3458 | ||
9771b263 DN |
3459 | exprs.release (); |
3460 | avail.release (); | |
83737db2 DB |
3461 | return new_stuff; |
3462 | } | |
3463 | ||
3464 | ||
d75dbccd DB |
3465 | /* Perform insertion for partially anticipatable expressions. There |
3466 | is only one case we will perform insertion for these. This case is | |
3467 | if the expression is partially anticipatable, and fully available. | |
3468 | In this case, we know that putting it earlier will enable us to | |
3469 | remove the later computation. */ | |
3470 | ||
3471 | ||
3472 | static bool | |
3473 | do_partial_partial_insertion (basic_block block, basic_block dom) | |
3474 | { | |
3475 | bool new_stuff = false; | |
9771b263 | 3476 | vec<pre_expr> exprs; |
c9145754 | 3477 | pre_expr expr; |
6e1aa848 | 3478 | vec<pre_expr> avail = vNULL; |
d75dbccd DB |
3479 | int i; |
3480 | ||
1c8f7377 | 3481 | exprs = sorted_array_from_bitmap_set (PA_IN (block)); |
9771b263 | 3482 | avail.safe_grow (EDGE_COUNT (block->preds)); |
1c8f7377 | 3483 | |
9771b263 | 3484 | FOR_EACH_VEC_ELT (exprs, i, expr) |
d75dbccd | 3485 | { |
4bcc5786 RB |
3486 | if (expr->kind == NARY |
3487 | || expr->kind == REFERENCE) | |
d75dbccd | 3488 | { |
c9145754 | 3489 | unsigned int val; |
d75dbccd DB |
3490 | bool by_all = true; |
3491 | bool cant_insert = false; | |
3492 | edge pred; | |
3493 | basic_block bprime; | |
c9145754 | 3494 | pre_expr eprime = NULL; |
d75dbccd DB |
3495 | edge_iterator ei; |
3496 | ||
c9145754 | 3497 | val = get_expr_value_id (expr); |
d75dbccd DB |
3498 | if (bitmap_set_contains_value (PHI_GEN (block), val)) |
3499 | continue; | |
3500 | if (bitmap_set_contains_value (AVAIL_OUT (dom), val)) | |
3501 | continue; | |
3502 | ||
d75dbccd DB |
3503 | FOR_EACH_EDGE (pred, ei, block->preds) |
3504 | { | |
c9145754 DB |
3505 | unsigned int vprime; |
3506 | pre_expr edoubleprime; | |
d75dbccd | 3507 | |
c4ab2baa RG |
3508 | /* We should never run insertion for the exit block |
3509 | and so not come across fake pred edges. */ | |
3510 | gcc_assert (!(pred->flags & EDGE_FAKE)); | |
d75dbccd DB |
3511 | bprime = pred->src; |
3512 | eprime = phi_translate (expr, ANTIC_IN (block), | |
3513 | PA_IN (block), | |
3514 | bprime, block); | |
3515 | ||
3516 | /* eprime will generally only be NULL if the | |
3517 | value of the expression, translated | |
3518 | through the PHI for this predecessor, is | |
3519 | undefined. If that is the case, we can't | |
3520 | make the expression fully redundant, | |
3521 | because its value is undefined along a | |
3522 | predecessor path. We can thus break out | |
3523 | early because it doesn't matter what the | |
3524 | rest of the results are. */ | |
3525 | if (eprime == NULL) | |
3526 | { | |
9771b263 | 3527 | avail[pred->dest_idx] = NULL; |
d75dbccd DB |
3528 | cant_insert = true; |
3529 | break; | |
3530 | } | |
3531 | ||
3532 | eprime = fully_constant_expression (eprime); | |
726a989a | 3533 | vprime = get_expr_value_id (eprime); |
e076271b | 3534 | edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime); |
9771b263 | 3535 | avail[pred->dest_idx] = edoubleprime; |
d75dbccd DB |
3536 | if (edoubleprime == NULL) |
3537 | { | |
3538 | by_all = false; | |
3539 | break; | |
3540 | } | |
d75dbccd DB |
3541 | } |
3542 | ||
3543 | /* If we can insert it, it's not the same value | |
3544 | already existing along every predecessor, and | |
3545 | it's defined by some predecessor, it is | |
3546 | partially redundant. */ | |
fa06ad0d | 3547 | if (!cant_insert && by_all) |
d75dbccd | 3548 | { |
fa06ad0d JR |
3549 | edge succ; |
3550 | bool do_insertion = false; | |
3551 | ||
3552 | /* Insert only if we can remove a later expression on a path | |
3553 | that we want to optimize for speed. | |
3554 | The phi node that we will be inserting in BLOCK is not free, | |
3555 | and inserting it for the sake of !optimize_for_speed successor | |
3556 | may cause regressions on the speed path. */ | |
3557 | FOR_EACH_EDGE (succ, ei, block->succs) | |
3558 | { | |
0c46ead2 JH |
3559 | if (bitmap_set_contains_value (PA_IN (succ->dest), val) |
3560 | || bitmap_set_contains_value (ANTIC_IN (succ->dest), val)) | |
fa06ad0d JR |
3561 | { |
3562 | if (optimize_edge_for_speed_p (succ)) | |
3563 | do_insertion = true; | |
3564 | } | |
3565 | } | |
3566 | ||
3567 | if (!do_insertion) | |
3568 | { | |
3569 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3570 | { | |
3571 | fprintf (dump_file, "Skipping partial partial redundancy " | |
3572 | "for expression "); | |
3573 | print_pre_expr (dump_file, expr); | |
0c46ead2 | 3574 | fprintf (dump_file, " (%04d), not (partially) anticipated " |
fa06ad0d JR |
3575 | "on any to be optimized for speed edges\n", val); |
3576 | } | |
3577 | } | |
3578 | else if (dbg_cnt (treepre_insert)) | |
3579 | { | |
3580 | pre_stats.pa_insert++; | |
19372838 RG |
3581 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3582 | { | |
3583 | fprintf (dump_file, "Found partial partial redundancy " | |
3584 | "for expression "); | |
3585 | print_pre_expr (dump_file, expr); | |
3586 | fprintf (dump_file, " (%04d)\n", | |
3587 | get_expr_value_id (expr)); | |
3588 | } | |
fa06ad0d JR |
3589 | if (insert_into_preds_of_block (block, |
3590 | get_expression_id (expr), | |
3591 | avail)) | |
3592 | new_stuff = true; | |
3593 | } | |
3594 | } | |
d75dbccd DB |
3595 | } |
3596 | } | |
3597 | ||
9771b263 DN |
3598 | exprs.release (); |
3599 | avail.release (); | |
d75dbccd DB |
3600 | return new_stuff; |
3601 | } | |
83737db2 | 3602 | |
7e6eb623 DB |
3603 | static bool |
3604 | insert_aux (basic_block block) | |
6de9cd9a | 3605 | { |
7e6eb623 DB |
3606 | basic_block son; |
3607 | bool new_stuff = false; | |
6de9cd9a | 3608 | |
7e6eb623 | 3609 | if (block) |
6de9cd9a | 3610 | { |
7e6eb623 DB |
3611 | basic_block dom; |
3612 | dom = get_immediate_dominator (CDI_DOMINATORS, block); | |
3613 | if (dom) | |
a32b97a2 | 3614 | { |
3cd8c58a | 3615 | unsigned i; |
87c476a2 | 3616 | bitmap_iterator bi; |
6b416da1 | 3617 | bitmap_set_t newset = NEW_SETS (dom); |
e9284566 | 3618 | if (newset) |
87c476a2 | 3619 | { |
e9284566 DB |
3620 | /* Note that we need to value_replace both NEW_SETS, and |
3621 | AVAIL_OUT. For both the case of NEW_SETS, the value may be | |
3622 | represented by some non-simple expression here that we want | |
3623 | to replace it with. */ | |
83737db2 | 3624 | FOR_EACH_EXPR_ID_IN_SET (newset, i, bi) |
e9284566 | 3625 | { |
c9145754 | 3626 | pre_expr expr = expression_for_id (i); |
83737db2 DB |
3627 | bitmap_value_replace_in_set (NEW_SETS (block), expr); |
3628 | bitmap_value_replace_in_set (AVAIL_OUT (block), expr); | |
e9284566 | 3629 | } |
87c476a2 | 3630 | } |
c5cbcccf | 3631 | if (!single_pred_p (block)) |
6de9cd9a | 3632 | { |
83737db2 | 3633 | new_stuff |= do_regular_insertion (block, dom); |
d75dbccd DB |
3634 | if (do_partial_partial) |
3635 | new_stuff |= do_partial_partial_insertion (block, dom); | |
6de9cd9a DN |
3636 | } |
3637 | } | |
3638 | } | |
7e6eb623 DB |
3639 | for (son = first_dom_son (CDI_DOMINATORS, block); |
3640 | son; | |
3641 | son = next_dom_son (CDI_DOMINATORS, son)) | |
3642 | { | |
3643 | new_stuff |= insert_aux (son); | |
3644 | } | |
3645 | ||
3646 | return new_stuff; | |
6de9cd9a DN |
3647 | } |
3648 | ||
7e6eb623 | 3649 | /* Perform insertion of partially redundant values. */ |
6de9cd9a | 3650 | |
7e6eb623 DB |
3651 | static void |
3652 | insert (void) | |
6de9cd9a | 3653 | { |
7e6eb623 | 3654 | bool new_stuff = true; |
6de9cd9a | 3655 | basic_block bb; |
7e6eb623 | 3656 | int num_iterations = 0; |
b9c5e484 | 3657 | |
7e6eb623 | 3658 | FOR_ALL_BB (bb) |
6b416da1 | 3659 | NEW_SETS (bb) = bitmap_set_new (); |
b9c5e484 | 3660 | |
7e6eb623 | 3661 | while (new_stuff) |
6de9cd9a | 3662 | { |
7e6eb623 | 3663 | num_iterations++; |
19372838 RG |
3664 | if (dump_file && dump_flags & TDF_DETAILS) |
3665 | fprintf (dump_file, "Starting insert iteration %d\n", num_iterations); | |
7e6eb623 | 3666 | new_stuff = insert_aux (ENTRY_BLOCK_PTR); |
3dbc97a9 RB |
3667 | |
3668 | /* Clear the NEW sets before the next iteration. We have already | |
3669 | fully propagated its contents. */ | |
3670 | if (new_stuff) | |
3671 | FOR_ALL_BB (bb) | |
3672 | bitmap_set_free (NEW_SETS (bb)); | |
6de9cd9a | 3673 | } |
9fe0cb7d | 3674 | statistics_histogram_event (cfun, "insert iterations", num_iterations); |
7e6eb623 | 3675 | } |
6de9cd9a | 3676 | |
33c94679 | 3677 | |
665fcad8 SB |
3678 | /* Compute the AVAIL set for all basic blocks. |
3679 | ||
3680 | This function performs value numbering of the statements in each basic | |
3681 | block. The AVAIL sets are built from information we glean while doing | |
3682 | this value numbering, since the AVAIL sets contain only one entry per | |
7e6eb623 | 3683 | value. |
b9c5e484 | 3684 | |
7e6eb623 | 3685 | AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)]. |
ff2ad0f7 | 3686 | AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */ |
6de9cd9a | 3687 | |
7e6eb623 | 3688 | static void |
665fcad8 | 3689 | compute_avail (void) |
6de9cd9a | 3690 | { |
c9145754 | 3691 | |
665fcad8 SB |
3692 | basic_block block, son; |
3693 | basic_block *worklist; | |
3694 | size_t sp = 0; | |
b005da11 RG |
3695 | unsigned i; |
3696 | ||
3697 | /* We pretend that default definitions are defined in the entry block. | |
3698 | This includes function arguments and the static chain decl. */ | |
3699 | for (i = 1; i < num_ssa_names; ++i) | |
3700 | { | |
3701 | tree name = ssa_name (i); | |
3702 | pre_expr e; | |
3703 | if (!name | |
3704 | || !SSA_NAME_IS_DEFAULT_DEF (name) | |
3705 | || has_zero_uses (name) | |
ea057359 | 3706 | || virtual_operand_p (name)) |
b005da11 | 3707 | continue; |
665fcad8 | 3708 | |
b005da11 RG |
3709 | e = get_or_alloc_expr_for_name (name); |
3710 | add_to_value (get_expr_value_id (e), e); | |
40b178f4 | 3711 | bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR), e); |
b005da11 | 3712 | bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR), e); |
2984956b AP |
3713 | } |
3714 | ||
e3815735 RB |
3715 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3716 | { | |
3717 | print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR), | |
3718 | "tmp_gen", ENTRY_BLOCK); | |
3719 | print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR), | |
3720 | "avail_out", ENTRY_BLOCK); | |
3721 | } | |
3722 | ||
665fcad8 | 3723 | /* Allocate the worklist. */ |
e1111e8e | 3724 | worklist = XNEWVEC (basic_block, n_basic_blocks); |
665fcad8 SB |
3725 | |
3726 | /* Seed the algorithm by putting the dominator children of the entry | |
3727 | block on the worklist. */ | |
3728 | for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR); | |
3729 | son; | |
3730 | son = next_dom_son (CDI_DOMINATORS, son)) | |
3731 | worklist[sp++] = son; | |
3732 | ||
3733 | /* Loop until the worklist is empty. */ | |
3734 | while (sp) | |
6de9cd9a | 3735 | { |
726a989a RB |
3736 | gimple_stmt_iterator gsi; |
3737 | gimple stmt; | |
7e6eb623 DB |
3738 | basic_block dom; |
3739 | ||
665fcad8 SB |
3740 | /* Pick a block from the worklist. */ |
3741 | block = worklist[--sp]; | |
3742 | ||
ff2ad0f7 DN |
3743 | /* Initially, the set of available values in BLOCK is that of |
3744 | its immediate dominator. */ | |
7e6eb623 DB |
3745 | dom = get_immediate_dominator (CDI_DOMINATORS, block); |
3746 | if (dom) | |
bdee7684 | 3747 | bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom)); |
33c94679 | 3748 | |
ff2ad0f7 | 3749 | /* Generate values for PHI nodes. */ |
726a989a | 3750 | for (gsi = gsi_start_phis (block); !gsi_end_p (gsi); gsi_next (&gsi)) |
e3815735 RB |
3751 | { |
3752 | tree result = gimple_phi_result (gsi_stmt (gsi)); | |
3753 | ||
3754 | /* We have no need for virtual phis, as they don't represent | |
3755 | actual computations. */ | |
3756 | if (virtual_operand_p (result)) | |
3757 | continue; | |
3758 | ||
3759 | pre_expr e = get_or_alloc_expr_for_name (result); | |
3760 | add_to_value (get_expr_value_id (e), e); | |
3761 | bitmap_value_insert_into_set (AVAIL_OUT (block), e); | |
3762 | bitmap_insert_into_set (PHI_GEN (block), e); | |
3763 | } | |
7e6eb623 | 3764 | |
a19eb9d2 RG |
3765 | BB_MAY_NOTRETURN (block) = 0; |
3766 | ||
ff2ad0f7 DN |
3767 | /* Now compute value numbers and populate value sets with all |
3768 | the expressions computed in BLOCK. */ | |
726a989a | 3769 | for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi)) |
6de9cd9a | 3770 | { |
f47c96aa AM |
3771 | ssa_op_iter iter; |
3772 | tree op; | |
ff2ad0f7 | 3773 | |
726a989a | 3774 | stmt = gsi_stmt (gsi); |
ff2ad0f7 | 3775 | |
a19eb9d2 RG |
3776 | /* Cache whether the basic-block has any non-visible side-effect |
3777 | or control flow. | |
3778 | If this isn't a call or it is the last stmt in the | |
3779 | basic-block then the CFG represents things correctly. */ | |
9cbbba28 | 3780 | if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt)) |
a19eb9d2 RG |
3781 | { |
3782 | /* Non-looping const functions always return normally. | |
3783 | Otherwise the call might not return or have side-effects | |
3784 | that forbids hoisting possibly trapping expressions | |
3785 | before it. */ | |
3786 | int flags = gimple_call_flags (stmt); | |
3787 | if (!(flags & ECF_CONST) | |
3788 | || (flags & ECF_LOOPING_CONST_OR_PURE)) | |
3789 | BB_MAY_NOTRETURN (block) = 1; | |
3790 | } | |
3791 | ||
c9145754 | 3792 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF) |
83737db2 | 3793 | { |
c9145754 | 3794 | pre_expr e = get_or_alloc_expr_for_name (op); |
83737db2 | 3795 | |
c9145754 | 3796 | add_to_value (get_expr_value_id (e), e); |
40b178f4 | 3797 | bitmap_insert_into_set (TMP_GEN (block), e); |
c9145754 | 3798 | bitmap_value_insert_into_set (AVAIL_OUT (block), e); |
83737db2 DB |
3799 | } |
3800 | ||
17742d62 RG |
3801 | if (gimple_has_side_effects (stmt) |
3802 | || stmt_could_throw_p (stmt) | |
3803 | || is_gimple_debug (stmt)) | |
726a989a RB |
3804 | continue; |
3805 | ||
17742d62 | 3806 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) |
e3815735 RB |
3807 | { |
3808 | if (ssa_undefined_value_p (op)) | |
3809 | continue; | |
3810 | pre_expr e = get_or_alloc_expr_for_name (op); | |
3811 | bitmap_value_insert_into_set (EXP_GEN (block), e); | |
3812 | } | |
17742d62 | 3813 | |
726a989a | 3814 | switch (gimple_code (stmt)) |
7e6eb623 | 3815 | { |
726a989a | 3816 | case GIMPLE_RETURN: |
c9145754 | 3817 | continue; |
726a989a RB |
3818 | |
3819 | case GIMPLE_CALL: | |
c9145754 | 3820 | { |
726a989a | 3821 | vn_reference_t ref; |
726a989a | 3822 | pre_expr result = NULL; |
6e1aa848 | 3823 | vec<vn_reference_op_s> ops = vNULL; |
c9145754 | 3824 | |
9cbbba28 EB |
3825 | /* We can value number only calls to real functions. */ |
3826 | if (gimple_call_internal_p (stmt)) | |
726a989a | 3827 | continue; |
c9145754 | 3828 | |
726a989a | 3829 | copy_reference_ops_from_call (stmt, &ops); |
b45d2719 RG |
3830 | vn_reference_lookup_pieces (gimple_vuse (stmt), 0, |
3831 | gimple_expr_type (stmt), | |
3bc27de7 | 3832 | ops, &ref, VN_NOWALK); |
9771b263 | 3833 | ops.release (); |
726a989a RB |
3834 | if (!ref) |
3835 | continue; | |
c9145754 | 3836 | |
cd32bb90 RG |
3837 | /* If the value of the call is not invalidated in |
3838 | this block until it is computed, add the expression | |
3839 | to EXP_GEN. */ | |
3840 | if (!gimple_vuse (stmt) | |
3841 | || gimple_code | |
3842 | (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI | |
3843 | || gimple_bb (SSA_NAME_DEF_STMT | |
3844 | (gimple_vuse (stmt))) != block) | |
3845 | { | |
3846 | result = (pre_expr) pool_alloc (pre_expr_pool); | |
3847 | result->kind = REFERENCE; | |
3848 | result->id = 0; | |
3849 | PRE_EXPR_REFERENCE (result) = ref; | |
3850 | ||
3851 | get_or_alloc_expression_id (result); | |
3852 | add_to_value (get_expr_value_id (result), result); | |
3c1d57d0 | 3853 | bitmap_value_insert_into_set (EXP_GEN (block), result); |
cd32bb90 | 3854 | } |
726a989a RB |
3855 | continue; |
3856 | } | |
c9145754 | 3857 | |
726a989a RB |
3858 | case GIMPLE_ASSIGN: |
3859 | { | |
3860 | pre_expr result = NULL; | |
17742d62 | 3861 | switch (vn_get_stmt_kind (stmt)) |
726a989a | 3862 | { |
17742d62 | 3863 | case VN_NARY: |
726a989a | 3864 | { |
61514fe4 | 3865 | enum tree_code code = gimple_assign_rhs_code (stmt); |
726a989a | 3866 | vn_nary_op_t nary; |
61514fe4 RG |
3867 | |
3868 | /* COND_EXPR and VEC_COND_EXPR are awkward in | |
3869 | that they contain an embedded complex expression. | |
3870 | Don't even try to shove those through PRE. */ | |
3871 | if (code == COND_EXPR | |
3872 | || code == VEC_COND_EXPR) | |
3873 | continue; | |
3874 | ||
91af9dc9 | 3875 | vn_nary_op_lookup_stmt (stmt, &nary); |
726a989a RB |
3876 | if (!nary) |
3877 | continue; | |
3878 | ||
073a8998 | 3879 | /* If the NARY traps and there was a preceding |
bea966c2 RG |
3880 | point in the block that might not return avoid |
3881 | adding the nary to EXP_GEN. */ | |
3882 | if (BB_MAY_NOTRETURN (block) | |
3883 | && vn_nary_may_trap (nary)) | |
3884 | continue; | |
3885 | ||
726a989a RB |
3886 | result = (pre_expr) pool_alloc (pre_expr_pool); |
3887 | result->kind = NARY; | |
3888 | result->id = 0; | |
3889 | PRE_EXPR_NARY (result) = nary; | |
3890 | break; | |
3891 | } | |
c9145754 | 3892 | |
17742d62 | 3893 | case VN_REFERENCE: |
726a989a RB |
3894 | { |
3895 | vn_reference_t ref; | |
726a989a | 3896 | vn_reference_lookup (gimple_assign_rhs1 (stmt), |
5006671f | 3897 | gimple_vuse (stmt), |
3bc27de7 | 3898 | VN_WALK, &ref); |
726a989a RB |
3899 | if (!ref) |
3900 | continue; | |
3901 | ||
cd32bb90 RG |
3902 | /* If the value of the reference is not invalidated in |
3903 | this block until it is computed, add the expression | |
3904 | to EXP_GEN. */ | |
3905 | if (gimple_vuse (stmt)) | |
3906 | { | |
3907 | gimple def_stmt; | |
3908 | bool ok = true; | |
3909 | def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt)); | |
3910 | while (!gimple_nop_p (def_stmt) | |
3911 | && gimple_code (def_stmt) != GIMPLE_PHI | |
3912 | && gimple_bb (def_stmt) == block) | |
3913 | { | |
3914 | if (stmt_may_clobber_ref_p | |
3915 | (def_stmt, gimple_assign_rhs1 (stmt))) | |
3916 | { | |
3917 | ok = false; | |
3918 | break; | |
3919 | } | |
3920 | def_stmt | |
3921 | = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt)); | |
3922 | } | |
3923 | if (!ok) | |
3924 | continue; | |
3925 | } | |
3926 | ||
726a989a RB |
3927 | result = (pre_expr) pool_alloc (pre_expr_pool); |
3928 | result->kind = REFERENCE; | |
3929 | result->id = 0; | |
3930 | PRE_EXPR_REFERENCE (result) = ref; | |
3931 | break; | |
3932 | } | |
c9145754 | 3933 | |
726a989a | 3934 | default: |
726a989a | 3935 | continue; |
c9145754 | 3936 | } |
726a989a RB |
3937 | |
3938 | get_or_alloc_expression_id (result); | |
3939 | add_to_value (get_expr_value_id (result), result); | |
3c1d57d0 | 3940 | bitmap_value_insert_into_set (EXP_GEN (block), result); |
dda243de | 3941 | continue; |
c9145754 DB |
3942 | } |
3943 | default: | |
3944 | break; | |
7e6eb623 | 3945 | } |
6de9cd9a | 3946 | } |
726a989a | 3947 | |
e3815735 RB |
3948 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3949 | { | |
3950 | print_bitmap_set (dump_file, EXP_GEN (block), | |
3951 | "exp_gen", block->index); | |
3952 | print_bitmap_set (dump_file, PHI_GEN (block), | |
3953 | "phi_gen", block->index); | |
3954 | print_bitmap_set (dump_file, TMP_GEN (block), | |
3955 | "tmp_gen", block->index); | |
3956 | print_bitmap_set (dump_file, AVAIL_OUT (block), | |
3957 | "avail_out", block->index); | |
3958 | } | |
3959 | ||
665fcad8 SB |
3960 | /* Put the dominator children of BLOCK on the worklist of blocks |
3961 | to compute available sets for. */ | |
3962 | for (son = first_dom_son (CDI_DOMINATORS, block); | |
3963 | son; | |
3964 | son = next_dom_son (CDI_DOMINATORS, son)) | |
3965 | worklist[sp++] = son; | |
6de9cd9a | 3966 | } |
33c94679 | 3967 | |
665fcad8 | 3968 | free (worklist); |
7e6eb623 DB |
3969 | } |
3970 | ||
40b178f4 RG |
3971 | |
3972 | /* Local state for the eliminate domwalk. */ | |
9771b263 DN |
3973 | static vec<gimple> el_to_remove; |
3974 | static vec<gimple> el_to_update; | |
40b178f4 | 3975 | static unsigned int el_todo; |
9771b263 DN |
3976 | static vec<tree> el_avail; |
3977 | static vec<tree> el_avail_stack; | |
40b178f4 RG |
3978 | |
3979 | /* Return a leader for OP that is available at the current point of the | |
3980 | eliminate domwalk. */ | |
3d45dd59 RG |
3981 | |
3982 | static tree | |
40b178f4 | 3983 | eliminate_avail (tree op) |
3d45dd59 | 3984 | { |
40b178f4 RG |
3985 | tree valnum = VN_INFO (op)->valnum; |
3986 | if (TREE_CODE (valnum) == SSA_NAME) | |
3987 | { | |
3988 | if (SSA_NAME_IS_DEFAULT_DEF (valnum)) | |
3989 | return valnum; | |
9771b263 DN |
3990 | if (el_avail.length () > SSA_NAME_VERSION (valnum)) |
3991 | return el_avail[SSA_NAME_VERSION (valnum)]; | |
40b178f4 RG |
3992 | } |
3993 | else if (is_gimple_min_invariant (valnum)) | |
3994 | return valnum; | |
3995 | return NULL_TREE; | |
3996 | } | |
3997 | ||
3998 | /* At the current point of the eliminate domwalk make OP available. */ | |
3d45dd59 | 3999 | |
40b178f4 RG |
4000 | static void |
4001 | eliminate_push_avail (tree op) | |
4002 | { | |
4003 | tree valnum = VN_INFO (op)->valnum; | |
4004 | if (TREE_CODE (valnum) == SSA_NAME) | |
4005 | { | |
9771b263 DN |
4006 | if (el_avail.length () <= SSA_NAME_VERSION (valnum)) |
4007 | el_avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1); | |
4008 | el_avail[SSA_NAME_VERSION (valnum)] = op; | |
4009 | el_avail_stack.safe_push (op); | |
40b178f4 RG |
4010 | } |
4011 | } | |
4012 | ||
4013 | /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns | |
4014 | the leader for the expression if insertion was successful. */ | |
4015 | ||
4016 | static tree | |
4017 | eliminate_insert (gimple_stmt_iterator *gsi, tree val) | |
4018 | { | |
4019 | tree expr = vn_get_expr_for (val); | |
4020 | if (!CONVERT_EXPR_P (expr) | |
4021 | && TREE_CODE (expr) != VIEW_CONVERT_EXPR) | |
3d45dd59 | 4022 | return NULL_TREE; |
3d45dd59 | 4023 | |
40b178f4 RG |
4024 | tree op = TREE_OPERAND (expr, 0); |
4025 | tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (op) : op; | |
4026 | if (!leader) | |
3d45dd59 | 4027 | return NULL_TREE; |
3d45dd59 | 4028 | |
40b178f4 RG |
4029 | tree res = make_temp_ssa_name (TREE_TYPE (val), NULL, "pretmp"); |
4030 | gimple tem = gimple_build_assign (res, | |
c96cab6e RB |
4031 | fold_build1 (TREE_CODE (expr), |
4032 | TREE_TYPE (expr), leader)); | |
40b178f4 RG |
4033 | gsi_insert_before (gsi, tem, GSI_SAME_STMT); |
4034 | VN_INFO_GET (res)->valnum = val; | |
4035 | ||
4036 | if (TREE_CODE (leader) == SSA_NAME) | |
4037 | gimple_set_plf (SSA_NAME_DEF_STMT (leader), NECESSARY, true); | |
4038 | ||
4039 | pre_stats.insertions++; | |
4040 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4041 | { | |
4042 | fprintf (dump_file, "Inserted "); | |
4043 | print_gimple_stmt (dump_file, tem, 0, 0); | |
4044 | } | |
4045 | ||
4046 | return res; | |
3d45dd59 | 4047 | } |
33c94679 | 4048 | |
4d9192b5 TS |
4049 | class eliminate_dom_walker : public dom_walker |
4050 | { | |
4051 | public: | |
4052 | eliminate_dom_walker (cdi_direction direction) : dom_walker (direction) {} | |
4053 | ||
4054 | virtual void before_dom_children (basic_block); | |
4055 | virtual void after_dom_children (basic_block); | |
4056 | }; | |
4057 | ||
40b178f4 | 4058 | /* Perform elimination for the basic-block B during the domwalk. */ |
7e6eb623 | 4059 | |
4d9192b5 TS |
4060 | void |
4061 | eliminate_dom_walker::before_dom_children (basic_block b) | |
7e6eb623 | 4062 | { |
5006671f RG |
4063 | gimple_stmt_iterator gsi; |
4064 | gimple stmt; | |
7e6eb623 | 4065 | |
40b178f4 | 4066 | /* Mark new bb. */ |
9771b263 | 4067 | el_avail_stack.safe_push (NULL_TREE); |
40b178f4 RG |
4068 | |
4069 | for (gsi = gsi_start_phis (b); !gsi_end_p (gsi);) | |
7e6eb623 | 4070 | { |
40b178f4 RG |
4071 | gimple stmt, phi = gsi_stmt (gsi); |
4072 | tree sprime = NULL_TREE, res = PHI_RESULT (phi); | |
4073 | gimple_stmt_iterator gsi2; | |
4074 | ||
4075 | /* We want to perform redundant PHI elimination. Do so by | |
4076 | replacing the PHI with a single copy if possible. | |
4077 | Do not touch inserted, single-argument or virtual PHIs. */ | |
4078 | if (gimple_phi_num_args (phi) == 1 | |
4079 | || virtual_operand_p (res)) | |
83737db2 | 4080 | { |
40b178f4 RG |
4081 | gsi_next (&gsi); |
4082 | continue; | |
4083 | } | |
c9145754 | 4084 | |
40b178f4 RG |
4085 | sprime = eliminate_avail (res); |
4086 | if (!sprime | |
4087 | || sprime == res) | |
4088 | { | |
4089 | eliminate_push_avail (res); | |
4090 | gsi_next (&gsi); | |
4091 | continue; | |
4092 | } | |
4093 | else if (is_gimple_min_invariant (sprime)) | |
4094 | { | |
4095 | if (!useless_type_conversion_p (TREE_TYPE (res), | |
4096 | TREE_TYPE (sprime))) | |
4097 | sprime = fold_convert (TREE_TYPE (res), sprime); | |
4098 | } | |
c9145754 | 4099 | |
40b178f4 RG |
4100 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4101 | { | |
4102 | fprintf (dump_file, "Replaced redundant PHI node defining "); | |
4103 | print_generic_expr (dump_file, res, 0); | |
4104 | fprintf (dump_file, " with "); | |
4105 | print_generic_expr (dump_file, sprime, 0); | |
4106 | fprintf (dump_file, "\n"); | |
4107 | } | |
6999afe1 | 4108 | |
40b178f4 RG |
4109 | remove_phi_node (&gsi, false); |
4110 | ||
4111 | if (inserted_exprs | |
4112 | && !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res)) | |
4113 | && TREE_CODE (sprime) == SSA_NAME) | |
4114 | gimple_set_plf (SSA_NAME_DEF_STMT (sprime), NECESSARY, true); | |
4115 | ||
4116 | if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime))) | |
4117 | sprime = fold_convert (TREE_TYPE (res), sprime); | |
4118 | stmt = gimple_build_assign (res, sprime); | |
4119 | SSA_NAME_DEF_STMT (res) = stmt; | |
4120 | gimple_set_plf (stmt, NECESSARY, gimple_plf (phi, NECESSARY)); | |
4121 | ||
4122 | gsi2 = gsi_after_labels (b); | |
4123 | gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT); | |
4124 | /* Queue the copy for eventual removal. */ | |
9771b263 | 4125 | el_to_remove.safe_push (stmt); |
40b178f4 RG |
4126 | /* If we inserted this PHI node ourself, it's not an elimination. */ |
4127 | if (inserted_exprs | |
4128 | && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))) | |
4129 | pre_stats.phis--; | |
4130 | else | |
4131 | pre_stats.eliminations++; | |
4132 | } | |
ff2ad0f7 | 4133 | |
40b178f4 RG |
4134 | for (gsi = gsi_start_bb (b); !gsi_end_p (gsi); gsi_next (&gsi)) |
4135 | { | |
4136 | tree lhs = NULL_TREE; | |
4137 | tree rhs = NULL_TREE; | |
4138 | ||
4139 | stmt = gsi_stmt (gsi); | |
4140 | ||
4141 | if (gimple_has_lhs (stmt)) | |
4142 | lhs = gimple_get_lhs (stmt); | |
4143 | ||
4144 | if (gimple_assign_single_p (stmt)) | |
4145 | rhs = gimple_assign_rhs1 (stmt); | |
4146 | ||
4147 | /* Lookup the RHS of the expression, see if we have an | |
4148 | available computation for it. If so, replace the RHS with | |
4bcc5786 | 4149 | the available computation. */ |
40b178f4 RG |
4150 | if (gimple_has_lhs (stmt) |
4151 | && TREE_CODE (lhs) == SSA_NAME | |
40b178f4 RG |
4152 | && !gimple_has_volatile_ops (stmt)) |
4153 | { | |
4154 | tree sprime; | |
4155 | gimple orig_stmt = stmt; | |
3d45dd59 | 4156 | |
40b178f4 | 4157 | sprime = eliminate_avail (lhs); |
4bcc5786 RB |
4158 | /* If there is no usable leader mark lhs as leader for its value. */ |
4159 | if (!sprime) | |
4160 | eliminate_push_avail (lhs); | |
4161 | ||
4162 | /* See PR43491. Do not replace a global register variable when | |
4163 | it is a the RHS of an assignment. Do replace local register | |
4164 | variables since gcc does not guarantee a local variable will | |
4165 | be allocated in register. | |
4166 | Do not perform copy propagation or undo constant propagation. */ | |
4167 | if (gimple_assign_single_p (stmt) | |
4168 | && (TREE_CODE (rhs) == SSA_NAME | |
4169 | || is_gimple_min_invariant (rhs) | |
4170 | || (TREE_CODE (rhs) == VAR_DECL | |
4171 | && is_global_var (rhs) | |
4172 | && DECL_HARD_REGISTER (rhs)))) | |
4173 | continue; | |
4174 | ||
40b178f4 RG |
4175 | if (!sprime) |
4176 | { | |
3d45dd59 RG |
4177 | /* If there is no existing usable leader but SCCVN thinks |
4178 | it has an expression it wants to use as replacement, | |
4179 | insert that. */ | |
40b178f4 RG |
4180 | tree val = VN_INFO (lhs)->valnum; |
4181 | if (val != VN_TOP | |
4182 | && TREE_CODE (val) == SSA_NAME | |
4183 | && VN_INFO (val)->needs_insertion | |
a044f0e7 | 4184 | && VN_INFO (val)->expr != NULL_TREE |
40b178f4 RG |
4185 | && (sprime = eliminate_insert (&gsi, val)) != NULL_TREE) |
4186 | eliminate_push_avail (sprime); | |
4187 | } | |
4188 | else if (is_gimple_min_invariant (sprime)) | |
4189 | { | |
4190 | /* If there is no existing leader but SCCVN knows this | |
4191 | value is constant, use that constant. */ | |
4192 | if (!useless_type_conversion_p (TREE_TYPE (lhs), | |
4193 | TREE_TYPE (sprime))) | |
4194 | sprime = fold_convert (TREE_TYPE (lhs), sprime); | |
4195 | ||
4196 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3d45dd59 | 4197 | { |
40b178f4 RG |
4198 | fprintf (dump_file, "Replaced "); |
4199 | print_gimple_expr (dump_file, stmt, 0, 0); | |
4200 | fprintf (dump_file, " with "); | |
4201 | print_generic_expr (dump_file, sprime, 0); | |
4202 | fprintf (dump_file, " in "); | |
4203 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
3d45dd59 | 4204 | } |
40b178f4 RG |
4205 | pre_stats.eliminations++; |
4206 | propagate_tree_value_into_stmt (&gsi, sprime); | |
4207 | stmt = gsi_stmt (gsi); | |
4208 | update_stmt (stmt); | |
4209 | ||
4210 | /* If we removed EH side-effects from the statement, clean | |
4211 | its EH information. */ | |
4212 | if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt)) | |
ff2ad0f7 | 4213 | { |
40b178f4 RG |
4214 | bitmap_set_bit (need_eh_cleanup, |
4215 | gimple_bb (stmt)->index); | |
4216 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4217 | fprintf (dump_file, " Removed EH side-effects.\n"); | |
4218 | } | |
4219 | continue; | |
4220 | } | |
30fd5881 | 4221 | |
40b178f4 RG |
4222 | if (sprime |
4223 | && sprime != lhs | |
4224 | && (rhs == NULL_TREE | |
4225 | || TREE_CODE (rhs) != SSA_NAME | |
4226 | || may_propagate_copy (rhs, sprime))) | |
4227 | { | |
4228 | bool can_make_abnormal_goto | |
4229 | = is_gimple_call (stmt) | |
4230 | && stmt_can_make_abnormal_goto (stmt); | |
b9c5e484 | 4231 | |
40b178f4 | 4232 | gcc_assert (sprime != rhs); |
30fd5881 | 4233 | |
40b178f4 RG |
4234 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4235 | { | |
4236 | fprintf (dump_file, "Replaced "); | |
4237 | print_gimple_expr (dump_file, stmt, 0, 0); | |
4238 | fprintf (dump_file, " with "); | |
4239 | print_generic_expr (dump_file, sprime, 0); | |
4240 | fprintf (dump_file, " in "); | |
4241 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
ff2ad0f7 | 4242 | } |
40b178f4 RG |
4243 | |
4244 | if (TREE_CODE (sprime) == SSA_NAME) | |
4245 | gimple_set_plf (SSA_NAME_DEF_STMT (sprime), | |
4246 | NECESSARY, true); | |
4247 | /* We need to make sure the new and old types actually match, | |
4248 | which may require adding a simple cast, which fold_convert | |
4249 | will do for us. */ | |
4250 | if ((!rhs || TREE_CODE (rhs) != SSA_NAME) | |
4251 | && !useless_type_conversion_p (gimple_expr_type (stmt), | |
4252 | TREE_TYPE (sprime))) | |
4253 | sprime = fold_convert (gimple_expr_type (stmt), sprime); | |
4254 | ||
4255 | pre_stats.eliminations++; | |
4256 | propagate_tree_value_into_stmt (&gsi, sprime); | |
4257 | stmt = gsi_stmt (gsi); | |
4258 | update_stmt (stmt); | |
4259 | ||
4260 | /* If we removed EH side-effects from the statement, clean | |
4261 | its EH information. */ | |
4262 | if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt)) | |
6cdb0ee3 | 4263 | { |
40b178f4 RG |
4264 | bitmap_set_bit (need_eh_cleanup, |
4265 | gimple_bb (stmt)->index); | |
6cdb0ee3 | 4266 | if (dump_file && (dump_flags & TDF_DETAILS)) |
40b178f4 RG |
4267 | fprintf (dump_file, " Removed EH side-effects.\n"); |
4268 | } | |
6cdb0ee3 | 4269 | |
40b178f4 RG |
4270 | /* Likewise for AB side-effects. */ |
4271 | if (can_make_abnormal_goto | |
4272 | && !stmt_can_make_abnormal_goto (stmt)) | |
4273 | { | |
4274 | bitmap_set_bit (need_ab_cleanup, | |
4275 | gimple_bb (stmt)->index); | |
4276 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4277 | fprintf (dump_file, " Removed AB side-effects.\n"); | |
6cdb0ee3 RG |
4278 | } |
4279 | } | |
40b178f4 RG |
4280 | } |
4281 | /* If the statement is a scalar store, see if the expression | |
4282 | has the same value number as its rhs. If so, the store is | |
4283 | dead. */ | |
4284 | else if (gimple_assign_single_p (stmt) | |
4285 | && !gimple_has_volatile_ops (stmt) | |
4286 | && !is_gimple_reg (gimple_assign_lhs (stmt)) | |
4287 | && (TREE_CODE (rhs) == SSA_NAME | |
4288 | || is_gimple_min_invariant (rhs))) | |
4289 | { | |
4290 | tree val; | |
4291 | val = vn_reference_lookup (gimple_assign_lhs (stmt), | |
4292 | gimple_vuse (stmt), VN_WALK, NULL); | |
4293 | if (TREE_CODE (rhs) == SSA_NAME) | |
4294 | rhs = VN_INFO (rhs)->valnum; | |
4295 | if (val | |
4296 | && operand_equal_p (val, rhs, 0)) | |
b80280f2 | 4297 | { |
40b178f4 | 4298 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b80280f2 | 4299 | { |
40b178f4 RG |
4300 | fprintf (dump_file, "Deleted redundant store "); |
4301 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
b80280f2 | 4302 | } |
40b178f4 RG |
4303 | |
4304 | /* Queue stmt for removal. */ | |
9771b263 | 4305 | el_to_remove.safe_push (stmt); |
b80280f2 | 4306 | } |
40b178f4 RG |
4307 | } |
4308 | /* Visit COND_EXPRs and fold the comparison with the | |
4309 | available value-numbers. */ | |
4310 | else if (gimple_code (stmt) == GIMPLE_COND) | |
4311 | { | |
4312 | tree op0 = gimple_cond_lhs (stmt); | |
4313 | tree op1 = gimple_cond_rhs (stmt); | |
4314 | tree result; | |
4315 | ||
4316 | if (TREE_CODE (op0) == SSA_NAME) | |
4317 | op0 = VN_INFO (op0)->valnum; | |
4318 | if (TREE_CODE (op1) == SSA_NAME) | |
4319 | op1 = VN_INFO (op1)->valnum; | |
4320 | result = fold_binary (gimple_cond_code (stmt), boolean_type_node, | |
4321 | op0, op1); | |
4322 | if (result && TREE_CODE (result) == INTEGER_CST) | |
aa7069aa | 4323 | { |
40b178f4 RG |
4324 | if (integer_zerop (result)) |
4325 | gimple_cond_make_false (stmt); | |
3b45a007 | 4326 | else |
40b178f4 RG |
4327 | gimple_cond_make_true (stmt); |
4328 | update_stmt (stmt); | |
4329 | el_todo = TODO_cleanup_cfg; | |
4330 | } | |
4331 | } | |
4332 | /* Visit indirect calls and turn them into direct calls if | |
4333 | possible. */ | |
4334 | if (is_gimple_call (stmt)) | |
4335 | { | |
4336 | tree orig_fn = gimple_call_fn (stmt); | |
4337 | tree fn; | |
4338 | if (!orig_fn) | |
4339 | continue; | |
4340 | if (TREE_CODE (orig_fn) == SSA_NAME) | |
4341 | fn = VN_INFO (orig_fn)->valnum; | |
4342 | else if (TREE_CODE (orig_fn) == OBJ_TYPE_REF | |
4343 | && TREE_CODE (OBJ_TYPE_REF_EXPR (orig_fn)) == SSA_NAME) | |
e248d83f MJ |
4344 | { |
4345 | fn = VN_INFO (OBJ_TYPE_REF_EXPR (orig_fn))->valnum; | |
4346 | if (!gimple_call_addr_fndecl (fn)) | |
4347 | { | |
4348 | fn = ipa_intraprocedural_devirtualization (stmt); | |
4349 | if (fn) | |
4350 | fn = build_fold_addr_expr (fn); | |
4351 | } | |
4352 | } | |
40b178f4 RG |
4353 | else |
4354 | continue; | |
4355 | if (gimple_call_addr_fndecl (fn) != NULL_TREE | |
4356 | && useless_type_conversion_p (TREE_TYPE (orig_fn), | |
4357 | TREE_TYPE (fn))) | |
4358 | { | |
4359 | bool can_make_abnormal_goto | |
4360 | = stmt_can_make_abnormal_goto (stmt); | |
4361 | bool was_noreturn = gimple_call_noreturn_p (stmt); | |
aa7069aa | 4362 | |
40b178f4 RG |
4363 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4364 | { | |
4365 | fprintf (dump_file, "Replacing call target with "); | |
4366 | print_generic_expr (dump_file, fn, 0); | |
4367 | fprintf (dump_file, " in "); | |
4368 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
4369 | } | |
30fd5881 | 4370 | |
40b178f4 | 4371 | gimple_call_set_fn (stmt, fn); |
9771b263 | 4372 | el_to_update.safe_push (stmt); |
87c20fe7 | 4373 | |
40b178f4 RG |
4374 | /* When changing a call into a noreturn call, cfg cleanup |
4375 | is needed to fix up the noreturn call. */ | |
4376 | if (!was_noreturn && gimple_call_noreturn_p (stmt)) | |
4377 | el_todo |= TODO_cleanup_cfg; | |
30fd5881 | 4378 | |
40b178f4 RG |
4379 | /* If we removed EH side-effects from the statement, clean |
4380 | its EH information. */ | |
4381 | if (maybe_clean_or_replace_eh_stmt (stmt, stmt)) | |
4382 | { | |
4383 | bitmap_set_bit (need_eh_cleanup, | |
4384 | gimple_bb (stmt)->index); | |
4385 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4386 | fprintf (dump_file, " Removed EH side-effects.\n"); | |
4387 | } | |
de0b4ad5 | 4388 | |
40b178f4 RG |
4389 | /* Likewise for AB side-effects. */ |
4390 | if (can_make_abnormal_goto | |
4391 | && !stmt_can_make_abnormal_goto (stmt)) | |
4392 | { | |
4393 | bitmap_set_bit (need_ab_cleanup, | |
4394 | gimple_bb (stmt)->index); | |
4395 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4396 | fprintf (dump_file, " Removed AB side-effects.\n"); | |
aa7069aa | 4397 | } |
40b178f4 RG |
4398 | |
4399 | /* Changing an indirect call to a direct call may | |
4400 | have exposed different semantics. This may | |
4401 | require an SSA update. */ | |
4402 | el_todo |= TODO_update_ssa_only_virtuals; | |
aa7069aa | 4403 | } |
83737db2 | 4404 | } |
40b178f4 RG |
4405 | } |
4406 | } | |
439ef907 | 4407 | |
40b178f4 | 4408 | /* Make no longer available leaders no longer available. */ |
439ef907 | 4409 | |
4d9192b5 TS |
4410 | void |
4411 | eliminate_dom_walker::after_dom_children (basic_block) | |
40b178f4 RG |
4412 | { |
4413 | tree entry; | |
9771b263 DN |
4414 | while ((entry = el_avail_stack.pop ()) != NULL_TREE) |
4415 | el_avail[SSA_NAME_VERSION (VN_INFO (entry)->valnum)] = NULL_TREE; | |
40b178f4 | 4416 | } |
439ef907 | 4417 | |
40b178f4 | 4418 | /* Eliminate fully redundant computations. */ |
439ef907 | 4419 | |
40b178f4 RG |
4420 | static unsigned int |
4421 | eliminate (void) | |
4422 | { | |
40b178f4 RG |
4423 | gimple_stmt_iterator gsi; |
4424 | gimple stmt; | |
4425 | unsigned i; | |
439ef907 | 4426 | |
40b178f4 RG |
4427 | need_eh_cleanup = BITMAP_ALLOC (NULL); |
4428 | need_ab_cleanup = BITMAP_ALLOC (NULL); | |
4d2ab9e3 | 4429 | |
9771b263 DN |
4430 | el_to_remove.create (0); |
4431 | el_to_update.create (0); | |
40b178f4 | 4432 | el_todo = 0; |
9771b263 DN |
4433 | el_avail.create (0); |
4434 | el_avail_stack.create (0); | |
40b178f4 | 4435 | |
4d9192b5 | 4436 | eliminate_dom_walker (CDI_DOMINATORS).walk (cfun->cfg->x_entry_block_ptr); |
40b178f4 | 4437 | |
9771b263 DN |
4438 | el_avail.release (); |
4439 | el_avail_stack.release (); | |
b80280f2 | 4440 | |
5006671f | 4441 | /* We cannot remove stmts during BB walk, especially not release SSA |
439ef907 | 4442 | names there as this confuses the VN machinery. The stmts ending |
40b178f4 | 4443 | up in el_to_remove are either stores or simple copies. */ |
9771b263 | 4444 | FOR_EACH_VEC_ELT (el_to_remove, i, stmt) |
5006671f | 4445 | { |
439ef907 | 4446 | tree lhs = gimple_assign_lhs (stmt); |
84ae6d7b | 4447 | tree rhs = gimple_assign_rhs1 (stmt); |
439ef907 RG |
4448 | use_operand_p use_p; |
4449 | gimple use_stmt; | |
4450 | ||
4451 | /* If there is a single use only, propagate the equivalency | |
4452 | instead of keeping the copy. */ | |
4453 | if (TREE_CODE (lhs) == SSA_NAME | |
84ae6d7b | 4454 | && TREE_CODE (rhs) == SSA_NAME |
85e59f3a | 4455 | && single_imm_use (lhs, &use_p, &use_stmt) |
84ae6d7b | 4456 | && may_propagate_copy (USE_FROM_PTR (use_p), rhs)) |
439ef907 | 4457 | { |
43896afb | 4458 | SET_USE (use_p, rhs); |
b4104018 | 4459 | update_stmt (use_stmt); |
40b178f4 RG |
4460 | if (inserted_exprs |
4461 | && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (lhs)) | |
43896afb RG |
4462 | && TREE_CODE (rhs) == SSA_NAME) |
4463 | gimple_set_plf (SSA_NAME_DEF_STMT (rhs), NECESSARY, true); | |
439ef907 RG |
4464 | } |
4465 | ||
4466 | /* If this is a store or a now unused copy, remove it. */ | |
4467 | if (TREE_CODE (lhs) != SSA_NAME | |
4468 | || has_zero_uses (lhs)) | |
4469 | { | |
a2c0ed2e | 4470 | basic_block bb = gimple_bb (stmt); |
439ef907 RG |
4471 | gsi = gsi_for_stmt (stmt); |
4472 | unlink_stmt_vdef (stmt); | |
b5b3ec3e RG |
4473 | if (gsi_remove (&gsi, true)) |
4474 | bitmap_set_bit (need_eh_cleanup, bb->index); | |
40b178f4 RG |
4475 | if (inserted_exprs |
4476 | && TREE_CODE (lhs) == SSA_NAME) | |
0ab555de | 4477 | bitmap_clear_bit (inserted_exprs, SSA_NAME_VERSION (lhs)); |
439ef907 RG |
4478 | release_defs (stmt); |
4479 | } | |
5006671f | 4480 | } |
9771b263 | 4481 | el_to_remove.release (); |
5006671f | 4482 | |
9aef53ee RG |
4483 | /* We cannot update call statements with virtual operands during |
4484 | SSA walk. This might remove them which in turn makes our | |
4485 | VN lattice invalid. */ | |
9771b263 | 4486 | FOR_EACH_VEC_ELT (el_to_update, i, stmt) |
9aef53ee | 4487 | update_stmt (stmt); |
9771b263 | 4488 | el_to_update.release (); |
9aef53ee | 4489 | |
40b178f4 RG |
4490 | return el_todo; |
4491 | } | |
4492 | ||
4493 | /* Perform CFG cleanups made necessary by elimination. */ | |
4494 | ||
9c370032 | 4495 | static unsigned |
40b178f4 RG |
4496 | fini_eliminate (void) |
4497 | { | |
4498 | bool do_eh_cleanup = !bitmap_empty_p (need_eh_cleanup); | |
4499 | bool do_ab_cleanup = !bitmap_empty_p (need_ab_cleanup); | |
4500 | ||
4501 | if (do_eh_cleanup) | |
4502 | gimple_purge_all_dead_eh_edges (need_eh_cleanup); | |
4503 | ||
4504 | if (do_ab_cleanup) | |
4505 | gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup); | |
4506 | ||
4507 | BITMAP_FREE (need_eh_cleanup); | |
4508 | BITMAP_FREE (need_ab_cleanup); | |
4509 | ||
4510 | if (do_eh_cleanup || do_ab_cleanup) | |
9c370032 RB |
4511 | return TODO_cleanup_cfg; |
4512 | return 0; | |
6de9cd9a DN |
4513 | } |
4514 | ||
0fc6c492 DB |
4515 | /* Borrow a bit of tree-ssa-dce.c for the moment. |
4516 | XXX: In 4.1, we should be able to just run a DCE pass after PRE, though | |
4517 | this may be a bit faster, and we may want critical edges kept split. */ | |
4518 | ||
4519 | /* If OP's defining statement has not already been determined to be necessary, | |
d4e6fecb | 4520 | mark that statement necessary. Return the stmt, if it is newly |
b9c5e484 | 4521 | necessary. */ |
0fc6c492 | 4522 | |
726a989a | 4523 | static inline gimple |
d4e6fecb | 4524 | mark_operand_necessary (tree op) |
0fc6c492 | 4525 | { |
726a989a | 4526 | gimple stmt; |
0fc6c492 DB |
4527 | |
4528 | gcc_assert (op); | |
4529 | ||
c90186eb DB |
4530 | if (TREE_CODE (op) != SSA_NAME) |
4531 | return NULL; | |
4532 | ||
0fc6c492 DB |
4533 | stmt = SSA_NAME_DEF_STMT (op); |
4534 | gcc_assert (stmt); | |
4535 | ||
726a989a RB |
4536 | if (gimple_plf (stmt, NECESSARY) |
4537 | || gimple_nop_p (stmt)) | |
d4e6fecb | 4538 | return NULL; |
0fc6c492 | 4539 | |
726a989a | 4540 | gimple_set_plf (stmt, NECESSARY, true); |
d4e6fecb | 4541 | return stmt; |
0fc6c492 DB |
4542 | } |
4543 | ||
4544 | /* Because we don't follow exactly the standard PRE algorithm, and decide not | |
4545 | to insert PHI nodes sometimes, and because value numbering of casts isn't | |
4546 | perfect, we sometimes end up inserting dead code. This simple DCE-like | |
4547 | pass removes any insertions we made that weren't actually used. */ | |
4548 | ||
4549 | static void | |
4550 | remove_dead_inserted_code (void) | |
4551 | { | |
0ab555de MM |
4552 | bitmap worklist; |
4553 | unsigned i; | |
4554 | bitmap_iterator bi; | |
726a989a | 4555 | gimple t; |
0fc6c492 | 4556 | |
0ab555de MM |
4557 | worklist = BITMAP_ALLOC (NULL); |
4558 | EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi) | |
0fc6c492 | 4559 | { |
0ab555de | 4560 | t = SSA_NAME_DEF_STMT (ssa_name (i)); |
726a989a | 4561 | if (gimple_plf (t, NECESSARY)) |
0ab555de | 4562 | bitmap_set_bit (worklist, i); |
0fc6c492 | 4563 | } |
0ab555de | 4564 | while (!bitmap_empty_p (worklist)) |
0fc6c492 | 4565 | { |
0ab555de MM |
4566 | i = bitmap_first_set_bit (worklist); |
4567 | bitmap_clear_bit (worklist, i); | |
4568 | t = SSA_NAME_DEF_STMT (ssa_name (i)); | |
c90186eb DB |
4569 | |
4570 | /* PHI nodes are somewhat special in that each PHI alternative has | |
4571 | data and control dependencies. All the statements feeding the | |
4572 | PHI node's arguments are always necessary. */ | |
726a989a | 4573 | if (gimple_code (t) == GIMPLE_PHI) |
0fc6c492 | 4574 | { |
726a989a | 4575 | unsigned k; |
d4e6fecb | 4576 | |
726a989a | 4577 | for (k = 0; k < gimple_phi_num_args (t); k++) |
83737db2 | 4578 | { |
0fc6c492 DB |
4579 | tree arg = PHI_ARG_DEF (t, k); |
4580 | if (TREE_CODE (arg) == SSA_NAME) | |
d4e6fecb | 4581 | { |
726a989a RB |
4582 | gimple n = mark_operand_necessary (arg); |
4583 | if (n) | |
0ab555de | 4584 | bitmap_set_bit (worklist, SSA_NAME_VERSION (arg)); |
d4e6fecb | 4585 | } |
0fc6c492 DB |
4586 | } |
4587 | } | |
4588 | else | |
4589 | { | |
4590 | /* Propagate through the operands. Examine all the USE, VUSE and | |
89fb70a3 | 4591 | VDEF operands in this statement. Mark all the statements |
0fc6c492 DB |
4592 | which feed this statement's uses as necessary. */ |
4593 | ssa_op_iter iter; | |
4594 | tree use; | |
4595 | ||
38635499 | 4596 | /* The operands of VDEF expressions are also needed as they |
0fc6c492 | 4597 | represent potential definitions that may reach this |
89fb70a3 | 4598 | statement (VDEF operands allow us to follow def-def |
0fc6c492 | 4599 | links). */ |
33c94679 | 4600 | |
0fc6c492 | 4601 | FOR_EACH_SSA_TREE_OPERAND (use, t, iter, SSA_OP_ALL_USES) |
d4e6fecb | 4602 | { |
726a989a | 4603 | gimple n = mark_operand_necessary (use); |
d4e6fecb | 4604 | if (n) |
0ab555de | 4605 | bitmap_set_bit (worklist, SSA_NAME_VERSION (use)); |
d4e6fecb | 4606 | } |
0fc6c492 DB |
4607 | } |
4608 | } | |
c90186eb | 4609 | |
0ab555de | 4610 | EXECUTE_IF_SET_IN_BITMAP (inserted_exprs, 0, i, bi) |
0fc6c492 | 4611 | { |
0ab555de | 4612 | t = SSA_NAME_DEF_STMT (ssa_name (i)); |
726a989a | 4613 | if (!gimple_plf (t, NECESSARY)) |
0fc6c492 | 4614 | { |
726a989a | 4615 | gimple_stmt_iterator gsi; |
c90186eb | 4616 | |
0fc6c492 DB |
4617 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4618 | { | |
4619 | fprintf (dump_file, "Removing unnecessary insertion:"); | |
726a989a | 4620 | print_gimple_stmt (dump_file, t, 0, 0); |
0fc6c492 | 4621 | } |
c90186eb | 4622 | |
726a989a RB |
4623 | gsi = gsi_for_stmt (t); |
4624 | if (gimple_code (t) == GIMPLE_PHI) | |
4625 | remove_phi_node (&gsi, true); | |
0fc6c492 | 4626 | else |
b70fdfe4 AO |
4627 | { |
4628 | gsi_remove (&gsi, true); | |
4629 | release_defs (t); | |
4630 | } | |
0fc6c492 DB |
4631 | } |
4632 | } | |
0ab555de | 4633 | BITMAP_FREE (worklist); |
0fc6c492 | 4634 | } |
c90186eb | 4635 | |
9ca87236 | 4636 | |
ff2ad0f7 | 4637 | /* Initialize data structures used by PRE. */ |
6de9cd9a DN |
4638 | |
4639 | static void | |
40b178f4 | 4640 | init_pre (void) |
6de9cd9a | 4641 | { |
c9145754 DB |
4642 | basic_block bb; |
4643 | ||
4644 | next_expression_id = 1; | |
9771b263 DN |
4645 | expressions.create (0); |
4646 | expressions.safe_push (NULL); | |
4647 | value_expressions.create (get_max_value_id () + 1); | |
c3284718 | 4648 | value_expressions.safe_grow_cleared (get_max_value_id () + 1); |
9771b263 | 4649 | name_to_id.create (0); |
c9145754 | 4650 | |
0ab555de | 4651 | inserted_exprs = BITMAP_ALLOC (NULL); |
c90186eb | 4652 | |
b71b4522 DB |
4653 | connect_infinite_loops_to_exit (); |
4654 | memset (&pre_stats, 0, sizeof (pre_stats)); | |
4655 | ||
585d0dc4 RG |
4656 | postorder = XNEWVEC (int, n_basic_blocks); |
4657 | postorder_num = inverted_post_order_compute (postorder); | |
c9145754 | 4658 | |
a72ae88a | 4659 | alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets)); |
c9145754 | 4660 | |
b71b4522 | 4661 | calculate_dominance_info (CDI_POST_DOMINATORS); |
d75dbccd DB |
4662 | calculate_dominance_info (CDI_DOMINATORS); |
4663 | ||
c9145754 | 4664 | bitmap_obstack_initialize (&grand_bitmap_obstack); |
0823efed DN |
4665 | phi_translate_table.create (5110); |
4666 | expression_to_id.create (num_ssa_names * 3); | |
c9145754 DB |
4667 | bitmap_set_pool = create_alloc_pool ("Bitmap sets", |
4668 | sizeof (struct bitmap_set), 30); | |
4669 | pre_expr_pool = create_alloc_pool ("pre_expr nodes", | |
4670 | sizeof (struct pre_expr_d), 30); | |
4671 | FOR_ALL_BB (bb) | |
4672 | { | |
40b178f4 RG |
4673 | EXP_GEN (bb) = bitmap_set_new (); |
4674 | PHI_GEN (bb) = bitmap_set_new (); | |
4675 | TMP_GEN (bb) = bitmap_set_new (); | |
c9145754 DB |
4676 | AVAIL_OUT (bb) = bitmap_set_new (); |
4677 | } | |
ff2ad0f7 DN |
4678 | } |
4679 | ||
4680 | ||
4681 | /* Deallocate data structures used by PRE. */ | |
6de9cd9a | 4682 | |
ff2ad0f7 | 4683 | static void |
40b178f4 | 4684 | fini_pre () |
ff2ad0f7 | 4685 | { |
c9145754 | 4686 | free (postorder); |
9771b263 | 4687 | value_expressions.release (); |
0ab555de | 4688 | BITMAP_FREE (inserted_exprs); |
c9145754 DB |
4689 | bitmap_obstack_release (&grand_bitmap_obstack); |
4690 | free_alloc_pool (bitmap_set_pool); | |
4691 | free_alloc_pool (pre_expr_pool); | |
0823efed DN |
4692 | phi_translate_table.dispose (); |
4693 | expression_to_id.dispose (); | |
9771b263 | 4694 | name_to_id.release (); |
50265400 | 4695 | |
a72ae88a | 4696 | free_aux_for_blocks (); |
6809cbf9 | 4697 | |
b71b4522 | 4698 | free_dominance_info (CDI_POST_DOMINATORS); |
ff2ad0f7 DN |
4699 | } |
4700 | ||
40b178f4 | 4701 | /* Gate and execute functions for PRE. */ |
ff2ad0f7 | 4702 | |
b80280f2 | 4703 | static unsigned int |
40b178f4 | 4704 | do_pre (void) |
ff2ad0f7 | 4705 | { |
b80280f2 | 4706 | unsigned int todo = 0; |
83737db2 | 4707 | |
fa06ad0d JR |
4708 | do_partial_partial = |
4709 | flag_tree_partial_pre && optimize_function_for_speed_p (cfun); | |
ff2ad0f7 | 4710 | |
c9145754 DB |
4711 | /* This has to happen before SCCVN runs because |
4712 | loop_optimizer_init may create new phis, etc. */ | |
40b178f4 | 4713 | loop_optimizer_init (LOOPS_NORMAL); |
c90186eb | 4714 | |
40b178f4 | 4715 | if (!run_scc_vn (VN_WALK)) |
863d2a57 | 4716 | { |
40b178f4 | 4717 | loop_optimizer_finalize (); |
b80280f2 | 4718 | return 0; |
863d2a57 | 4719 | } |
a7d4562a | 4720 | |
40b178f4 | 4721 | init_pre (); |
a8338640 | 4722 | scev_initialize (); |
c9145754 | 4723 | |
c9145754 | 4724 | /* Collect and value number expressions computed in each basic block. */ |
665fcad8 | 4725 | compute_avail (); |
6de9cd9a | 4726 | |
7e6eb623 DB |
4727 | /* Insert can get quite slow on an incredibly large number of basic |
4728 | blocks due to some quadratic behavior. Until this behavior is | |
4729 | fixed, don't run it when he have an incredibly large number of | |
4730 | bb's. If we aren't going to run insert, there is no point in | |
4731 | computing ANTIC, either, even though it's plenty fast. */ | |
40b178f4 | 4732 | if (n_basic_blocks < 4000) |
6de9cd9a | 4733 | { |
7e6eb623 | 4734 | compute_antic (); |
7e6eb623 DB |
4735 | insert (); |
4736 | } | |
ff2ad0f7 | 4737 | |
0ab555de MM |
4738 | /* Make sure to remove fake edges before committing our inserts. |
4739 | This makes sure we don't end up with extra critical edges that | |
4740 | we would need to split. */ | |
4741 | remove_fake_exit_edges (); | |
4742 | gsi_commit_edge_inserts (); | |
4743 | ||
ff2ad0f7 | 4744 | /* Remove all the redundant expressions. */ |
b80280f2 | 4745 | todo |= eliminate (); |
0fc6c492 | 4746 | |
9fe0cb7d RG |
4747 | statistics_counter_event (cfun, "Insertions", pre_stats.insertions); |
4748 | statistics_counter_event (cfun, "PA inserted", pre_stats.pa_insert); | |
4749 | statistics_counter_event (cfun, "New PHIs", pre_stats.phis); | |
4750 | statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations); | |
c4ab2baa | 4751 | |
b71b4522 | 4752 | clear_expression_ids (); |
40b178f4 RG |
4753 | remove_dead_inserted_code (); |
4754 | todo |= TODO_verify_flow; | |
c90186eb | 4755 | |
a8338640 | 4756 | scev_finalize (); |
40b178f4 | 4757 | fini_pre (); |
9c370032 | 4758 | todo |= fini_eliminate (); |
40b178f4 RG |
4759 | loop_optimizer_finalize (); |
4760 | ||
4761 | /* TODO: tail_merge_optimize may merge all predecessors of a block, in which | |
4762 | case we can merge the block with the remaining predecessor of the block. | |
4763 | It should either: | |
4764 | - call merge_blocks after each tail merge iteration | |
4765 | - call merge_blocks after all tail merge iterations | |
4766 | - mark TODO_cleanup_cfg when necessary | |
4767 | - share the cfg cleanup with fini_pre. */ | |
4768 | todo |= tail_merge_optimize (todo); | |
4769 | ||
c9e93168 TV |
4770 | free_scc_vn (); |
4771 | ||
678771ad RG |
4772 | /* Tail merging invalidates the virtual SSA web, together with |
4773 | cfg-cleanup opportunities exposed by PRE this will wreck the | |
4774 | SSA updating machinery. So make sure to run update-ssa | |
4775 | manually, before eventually scheduling cfg-cleanup as part of | |
4776 | the todo. */ | |
4777 | update_ssa (TODO_update_ssa_only_virtuals); | |
4778 | ||
b80280f2 | 4779 | return todo; |
ff2ad0f7 DN |
4780 | } |
4781 | ||
6de9cd9a DN |
4782 | static bool |
4783 | gate_pre (void) | |
4784 | { | |
5813994e | 4785 | return flag_tree_pre != 0; |
6de9cd9a DN |
4786 | } |
4787 | ||
27a4cd48 DM |
4788 | namespace { |
4789 | ||
4790 | const pass_data pass_data_pre = | |
6de9cd9a | 4791 | { |
27a4cd48 DM |
4792 | GIMPLE_PASS, /* type */ |
4793 | "pre", /* name */ | |
4794 | OPTGROUP_NONE, /* optinfo_flags */ | |
4795 | true, /* has_gate */ | |
4796 | true, /* has_execute */ | |
4797 | TV_TREE_PRE, /* tv_id */ | |
4798 | ( PROP_no_crit_edges | PROP_cfg | PROP_ssa ), /* properties_required */ | |
4799 | 0, /* properties_provided */ | |
4800 | 0, /* properties_destroyed */ | |
4801 | TODO_rebuild_alias, /* todo_flags_start */ | |
4802 | TODO_verify_ssa, /* todo_flags_finish */ | |
6de9cd9a | 4803 | }; |
ff2ad0f7 | 4804 | |
27a4cd48 DM |
4805 | class pass_pre : public gimple_opt_pass |
4806 | { | |
4807 | public: | |
c3284718 RS |
4808 | pass_pre (gcc::context *ctxt) |
4809 | : gimple_opt_pass (pass_data_pre, ctxt) | |
27a4cd48 DM |
4810 | {} |
4811 | ||
4812 | /* opt_pass methods: */ | |
4813 | bool gate () { return gate_pre (); } | |
4814 | unsigned int execute () { return do_pre (); } | |
4815 | ||
4816 | }; // class pass_pre | |
4817 | ||
4818 | } // anon namespace | |
4819 | ||
4820 | gimple_opt_pass * | |
4821 | make_pass_pre (gcc::context *ctxt) | |
4822 | { | |
4823 | return new pass_pre (ctxt); | |
4824 | } | |
4825 | ||
ff2ad0f7 DN |
4826 | |
4827 | /* Gate and execute functions for FRE. */ | |
4828 | ||
c2924966 | 4829 | static unsigned int |
b89361c6 | 4830 | execute_fre (void) |
ff2ad0f7 | 4831 | { |
40b178f4 RG |
4832 | unsigned int todo = 0; |
4833 | ||
4834 | if (!run_scc_vn (VN_WALKREWRITE)) | |
4835 | return 0; | |
4836 | ||
4837 | memset (&pre_stats, 0, sizeof (pre_stats)); | |
4838 | ||
4839 | /* Remove all the redundant expressions. */ | |
4840 | todo |= eliminate (); | |
4841 | ||
9c370032 | 4842 | todo |= fini_eliminate (); |
40b178f4 RG |
4843 | |
4844 | free_scc_vn (); | |
4845 | ||
4846 | statistics_counter_event (cfun, "Insertions", pre_stats.insertions); | |
4847 | statistics_counter_event (cfun, "Eliminated", pre_stats.eliminations); | |
40b178f4 RG |
4848 | |
4849 | return todo; | |
ff2ad0f7 DN |
4850 | } |
4851 | ||
4852 | static bool | |
4853 | gate_fre (void) | |
4854 | { | |
4855 | return flag_tree_fre != 0; | |
4856 | } | |
4857 | ||
27a4cd48 DM |
4858 | namespace { |
4859 | ||
4860 | const pass_data pass_data_fre = | |
ff2ad0f7 | 4861 | { |
27a4cd48 DM |
4862 | GIMPLE_PASS, /* type */ |
4863 | "fre", /* name */ | |
4864 | OPTGROUP_NONE, /* optinfo_flags */ | |
4865 | true, /* has_gate */ | |
4866 | true, /* has_execute */ | |
4867 | TV_TREE_FRE, /* tv_id */ | |
4868 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
4869 | 0, /* properties_provided */ | |
4870 | 0, /* properties_destroyed */ | |
4871 | 0, /* todo_flags_start */ | |
4872 | TODO_verify_ssa, /* todo_flags_finish */ | |
ff2ad0f7 | 4873 | }; |
27a4cd48 DM |
4874 | |
4875 | class pass_fre : public gimple_opt_pass | |
4876 | { | |
4877 | public: | |
c3284718 RS |
4878 | pass_fre (gcc::context *ctxt) |
4879 | : gimple_opt_pass (pass_data_fre, ctxt) | |
27a4cd48 DM |
4880 | {} |
4881 | ||
4882 | /* opt_pass methods: */ | |
65d3284b | 4883 | opt_pass * clone () { return new pass_fre (m_ctxt); } |
27a4cd48 DM |
4884 | bool gate () { return gate_fre (); } |
4885 | unsigned int execute () { return execute_fre (); } | |
4886 | ||
4887 | }; // class pass_fre | |
4888 | ||
4889 | } // anon namespace | |
4890 | ||
4891 | gimple_opt_pass * | |
4892 | make_pass_fre (gcc::context *ctxt) | |
4893 | { | |
4894 | return new pass_fre (ctxt); | |
4895 | } |