1 | /* Interprocedural constant propagation |
2 | Copyright (C) 2005-2024 Free Software Foundation, Inc. |
3 | |
4 | Contributed by Razya Ladelsky <RAZYA@il.ibm.com> and Martin Jambor |
5 | <mjambor@suse.cz> |
6 | |
7 | This file is part of GCC. |
8 | |
9 | GCC is free software; you can redistribute it and/or modify it under |
10 | the terms of the GNU General Public License as published by the Free |
11 | Software Foundation; either version 3, or (at your option) any later |
12 | version. |
13 | |
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
17 | for more details. |
18 | |
19 | You should have received a copy of the GNU General Public License |
20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ |
22 | |
23 | /* Interprocedural constant propagation (IPA-CP). |
24 | |
25 | The goal of this transformation is to |
26 | |
27 | 1) discover functions which are always invoked with some arguments with the |
28 | same known constant values and modify the functions so that the |
29 | subsequent optimizations can take advantage of the knowledge, and |
30 | |
31 | 2) partial specialization - create specialized versions of functions |
32 | transformed in this way if some parameters are known constants only in |
33 | certain contexts but the estimated tradeoff between speedup and cost size |
34 | is deemed good. |
35 | |
36 | The algorithm also propagates types and attempts to perform type based |
37 | devirtualization. Types are propagated much like constants. |
38 | |
39 | The algorithm basically consists of three stages. In the first, functions |
40 | are analyzed one at a time and jump functions are constructed for all known |
41 | call-sites. In the second phase, the pass propagates information from the |
42 | jump functions across the call to reveal what values are available at what |
43 | call sites, performs estimations of effects of known values on functions and |
44 | their callees, and finally decides what specialized extra versions should be |
45 | created. In the third, the special versions materialize and appropriate |
46 | calls are redirected. |
47 | |
48 | The algorithm used is to a certain extent based on "Interprocedural Constant |
49 | Propagation", by David Callahan, Keith D Cooper, Ken Kennedy, Linda Torczon, |
50 | Comp86, pg 152-161 and "A Methodology for Procedure Cloning" by Keith D |
51 | Cooper, Mary W. Hall, and Ken Kennedy. |
52 | |
53 | |
54 | First stage - intraprocedural analysis |
55 | ======================================= |
56 | |
57 | This phase computes jump_function and modification flags. |
58 | |
59 | A jump function for a call-site represents the values passed as an actual |
60 | arguments of a given call-site. In principle, there are three types of |
61 | values: |
62 | |
63 | Pass through - the caller's formal parameter is passed as an actual |
64 | argument, plus an operation on it can be performed. |
65 | Constant - a constant is passed as an actual argument. |
66 | Unknown - neither of the above. |
67 | |
68 | All jump function types are described in detail in ipa-prop.h, together with |
69 | the data structures that represent them and methods of accessing them. |
70 | |
71 | ipcp_generate_summary() is the main function of the first stage. |
72 | |
73 | Second stage - interprocedural analysis |
74 | ======================================== |
75 | |
76 | This stage is itself divided into two phases. In the first, we propagate |
77 | known values over the call graph, in the second, we make cloning decisions. |
78 | It uses a different algorithm than the original Callahan's paper. |
79 | |
80 | First, we traverse the functions topologically from callers to callees and, |
81 | for each strongly connected component (SCC), we propagate constants |
82 | according to previously computed jump functions. We also record what known |
83 | values depend on other known values and estimate local effects. Finally, we |
84 | propagate cumulative information about these effects from dependent values |
85 | to those on which they depend. |
86 | |
87 | Second, we again traverse the call graph in the same topological order and |
88 | make clones for functions which we know are called with the same values in |
89 | all contexts and decide about extra specialized clones of functions just for |
90 | some contexts - these decisions are based on both local estimates and |
91 | cumulative estimates propagated from callees. |
92 | |
93 | ipcp_propagate_stage() and ipcp_decision_stage() together constitute the |
94 | third stage. |
95 | |
96 | Third phase - materialization of clones, call statement updates. |
97 | ============================================ |
98 | |
99 | This stage is currently performed by call graph code (mainly in cgraphunit.cc |
100 | and tree-inline.cc) according to instructions inserted to the call graph by |
101 | the second stage. */ |
102 | |
103 | #define INCLUDE_ALGORITHM |
104 | #include "config.h" |
105 | #include "system.h" |
106 | #include "coretypes.h" |
107 | #include "backend.h" |
108 | #include "tree.h" |
109 | #include "gimple-expr.h" |
110 | #include "gimple.h" |
111 | #include "predict.h" |
112 | #include "sreal.h" |
113 | #include "alloc-pool.h" |
114 | #include "tree-pass.h" |
115 | #include "cgraph.h" |
116 | #include "diagnostic.h" |
117 | #include "fold-const.h" |
118 | #include "gimple-iterator.h" |
119 | #include "gimple-fold.h" |
120 | #include "symbol-summary.h" |
121 | #include "tree-vrp.h" |
122 | #include "ipa-cp.h" |
123 | #include "ipa-prop.h" |
124 | #include "tree-pretty-print.h" |
125 | #include "tree-inline.h" |
126 | #include "ipa-fnsummary.h" |
127 | #include "ipa-utils.h" |
128 | #include "tree-ssa-ccp.h" |
129 | #include "stringpool.h" |
130 | #include "attribs.h" |
131 | #include "dbgcnt.h" |
132 | #include "symtab-clones.h" |
133 | #include "gimple-range.h" |
134 | |
135 | |
136 | /* Allocation pools for values and their sources in ipa-cp. */ |
137 | |
138 | object_allocator<ipcp_value<tree> > ipcp_cst_values_pool |
139 | ("IPA-CP constant values" ); |
140 | |
141 | object_allocator<ipcp_value<ipa_polymorphic_call_context> > |
142 | ipcp_poly_ctx_values_pool ("IPA-CP polymorphic contexts" ); |
143 | |
144 | object_allocator<ipcp_value_source<tree> > ipcp_sources_pool |
145 | ("IPA-CP value sources" ); |
146 | |
147 | object_allocator<ipcp_agg_lattice> ipcp_agg_lattice_pool |
148 | ("IPA_CP aggregate lattices" ); |
149 | |
150 | /* Base count to use in heuristics when using profile feedback. */ |
151 | |
152 | static profile_count base_count; |
153 | |
154 | /* Original overall size of the program. */ |
155 | |
156 | static long overall_size, orig_overall_size; |
157 | |
158 | /* Node name to unique clone suffix number map. */ |
159 | static hash_map<const char *, unsigned> *clone_num_suffixes; |
160 | |
161 | /* Return the param lattices structure corresponding to the Ith formal |
162 | parameter of the function described by INFO. */ |
163 | static inline class ipcp_param_lattices * |
164 | ipa_get_parm_lattices (class ipa_node_params *info, int i) |
165 | { |
166 | gcc_assert (i >= 0 && i < ipa_get_param_count (info)); |
167 | gcc_checking_assert (!info->ipcp_orig_node); |
168 | return &(info->lattices[i]); |
169 | } |
170 | |
171 | /* Return the lattice corresponding to the scalar value of the Ith formal |
172 | parameter of the function described by INFO. */ |
173 | static inline ipcp_lattice<tree> * |
174 | ipa_get_scalar_lat (class ipa_node_params *info, int i) |
175 | { |
176 | class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i); |
177 | return &plats->itself; |
178 | } |
179 | |
180 | /* Return the lattice corresponding to the scalar value of the Ith formal |
181 | parameter of the function described by INFO. */ |
182 | static inline ipcp_lattice<ipa_polymorphic_call_context> * |
183 | ipa_get_poly_ctx_lat (class ipa_node_params *info, int i) |
184 | { |
185 | class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i); |
186 | return &plats->ctxlat; |
187 | } |
188 | |
189 | /* Return whether LAT is a lattice with a single constant and without an |
190 | undefined value. */ |
191 | |
192 | template <typename valtype> |
193 | inline bool |
194 | ipcp_lattice<valtype>::is_single_const () |
195 | { |
196 | if (bottom || contains_variable || values_count != 1) |
197 | return false; |
198 | else |
199 | return true; |
200 | } |
201 | |
202 | /* Return true iff X and Y should be considered equal values by IPA-CP. */ |
203 | |
204 | bool |
205 | values_equal_for_ipcp_p (tree x, tree y) |
206 | { |
207 | gcc_checking_assert (x != NULL_TREE && y != NULL_TREE); |
208 | |
209 | if (x == y) |
210 | return true; |
211 | |
212 | if (TREE_CODE (x) == ADDR_EXPR |
213 | && TREE_CODE (y) == ADDR_EXPR |
214 | && (TREE_CODE (TREE_OPERAND (x, 0)) == CONST_DECL |
215 | || (TREE_CODE (TREE_OPERAND (x, 0)) == VAR_DECL |
216 | && DECL_IN_CONSTANT_POOL (TREE_OPERAND (x, 0)))) |
217 | && (TREE_CODE (TREE_OPERAND (y, 0)) == CONST_DECL |
218 | || (TREE_CODE (TREE_OPERAND (y, 0)) == VAR_DECL |
219 | && DECL_IN_CONSTANT_POOL (TREE_OPERAND (y, 0))))) |
220 | return TREE_OPERAND (x, 0) == TREE_OPERAND (y, 0) |
221 | || operand_equal_p (DECL_INITIAL (TREE_OPERAND (x, 0)), |
222 | DECL_INITIAL (TREE_OPERAND (y, 0)), flags: 0); |
223 | else |
224 | return operand_equal_p (x, y, flags: 0); |
225 | } |
226 | |
227 | /* Print V which is extracted from a value in a lattice to F. */ |
228 | |
229 | static void |
230 | print_ipcp_constant_value (FILE * f, ipa_polymorphic_call_context v) |
231 | { |
232 | v.dump(f, newline: false); |
233 | } |
234 | |
235 | /* Print a lattice LAT to F. */ |
236 | |
237 | template <typename valtype> |
238 | void |
239 | ipcp_lattice<valtype>::print (FILE * f, bool dump_sources, bool dump_benefits) |
240 | { |
241 | ipcp_value<valtype> *val; |
242 | bool prev = false; |
243 | |
244 | if (bottom) |
245 | { |
246 | fprintf (stream: f, format: "BOTTOM\n" ); |
247 | return; |
248 | } |
249 | |
250 | if (!values_count && !contains_variable) |
251 | { |
252 | fprintf (stream: f, format: "TOP\n" ); |
253 | return; |
254 | } |
255 | |
256 | if (contains_variable) |
257 | { |
258 | fprintf (stream: f, format: "VARIABLE" ); |
259 | prev = true; |
260 | if (dump_benefits) |
261 | fprintf (stream: f, format: "\n" ); |
262 | } |
263 | |
264 | for (val = values; val; val = val->next) |
265 | { |
266 | if (dump_benefits && prev) |
267 | fprintf (stream: f, format: " " ); |
268 | else if (!dump_benefits && prev) |
269 | fprintf (stream: f, format: ", " ); |
270 | else |
271 | prev = true; |
272 | |
273 | print_ipcp_constant_value (f, val->value); |
274 | |
275 | if (dump_sources) |
276 | { |
277 | ipcp_value_source<valtype> *s; |
278 | |
279 | if (val->self_recursion_generated_p ()) |
280 | fprintf (f, " [self_gen(%i), from:" , |
281 | val->self_recursion_generated_level); |
282 | else |
283 | fprintf (f, " [scc: %i, from:" , val->scc_no); |
284 | for (s = val->sources; s; s = s->next) |
285 | fprintf (f, " %i(%f)" , s->cs->caller->order, |
286 | s->cs->sreal_frequency ().to_double ()); |
287 | fprintf (stream: f, format: "]" ); |
288 | } |
289 | |
290 | if (dump_benefits) |
291 | fprintf (f, " [loc_time: %g, loc_size: %i, " |
292 | "prop_time: %g, prop_size: %i]\n" , |
293 | val->local_time_benefit.to_double (), val->local_size_cost, |
294 | val->prop_time_benefit.to_double (), val->prop_size_cost); |
295 | } |
296 | if (!dump_benefits) |
297 | fprintf (stream: f, format: "\n" ); |
298 | } |
299 | |
300 | void |
301 | ipcp_bits_lattice::print (FILE *f) |
302 | { |
303 | if (top_p ()) |
304 | fprintf (stream: f, format: " Bits unknown (TOP)\n" ); |
305 | else if (bottom_p ()) |
306 | fprintf (stream: f, format: " Bits unusable (BOTTOM)\n" ); |
307 | else |
308 | { |
309 | fprintf (stream: f, format: " Bits: value = " ); print_hex (wi: get_value (), file: f); |
310 | fprintf (stream: f, format: ", mask = " ); print_hex (wi: get_mask (), file: f); |
311 | fprintf (stream: f, format: "\n" ); |
312 | } |
313 | } |
314 | |
315 | /* Print value range lattice to F. */ |
316 | |
317 | void |
318 | ipcp_vr_lattice::print (FILE * f) |
319 | { |
320 | m_vr.dump (f); |
321 | } |
322 | |
323 | /* Print all ipcp_lattices of all functions to F. */ |
324 | |
325 | static void |
326 | print_all_lattices (FILE * f, bool dump_sources, bool dump_benefits) |
327 | { |
328 | struct cgraph_node *node; |
329 | int i, count; |
330 | |
331 | fprintf (stream: f, format: "\nLattices:\n" ); |
332 | FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node) |
333 | { |
334 | class ipa_node_params *info; |
335 | |
336 | info = ipa_node_params_sum->get (node); |
337 | /* Skip unoptimized functions and constprop clones since we don't make |
338 | lattices for them. */ |
339 | if (!info || info->ipcp_orig_node) |
340 | continue; |
341 | fprintf (stream: f, format: " Node: %s:\n" , node->dump_name ()); |
342 | count = ipa_get_param_count (info); |
343 | for (i = 0; i < count; i++) |
344 | { |
345 | struct ipcp_agg_lattice *aglat; |
346 | class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i); |
347 | fprintf (stream: f, format: " param [%d]: " , i); |
348 | plats->itself.print (f, dump_sources, dump_benefits); |
349 | fprintf (stream: f, format: " ctxs: " ); |
350 | plats->ctxlat.print (f, dump_sources, dump_benefits); |
351 | plats->bits_lattice.print (f); |
352 | fprintf (stream: f, format: " " ); |
353 | plats->m_value_range.print (f); |
354 | fprintf (stream: f, format: "\n" ); |
355 | if (plats->virt_call) |
356 | fprintf (stream: f, format: " virt_call flag set\n" ); |
357 | |
358 | if (plats->aggs_bottom) |
359 | { |
360 | fprintf (stream: f, format: " AGGS BOTTOM\n" ); |
361 | continue; |
362 | } |
363 | if (plats->aggs_contain_variable) |
364 | fprintf (stream: f, format: " AGGS VARIABLE\n" ); |
365 | for (aglat = plats->aggs; aglat; aglat = aglat->next) |
366 | { |
367 | fprintf (stream: f, format: " %soffset " HOST_WIDE_INT_PRINT_DEC ": " , |
368 | plats->aggs_by_ref ? "ref " : "" , aglat->offset); |
369 | aglat->print (f, dump_sources, dump_benefits); |
370 | } |
371 | } |
372 | } |
373 | } |
374 | |
375 | /* Determine whether it is at all technically possible to create clones of NODE |
376 | and store this information in the ipa_node_params structure associated |
377 | with NODE. */ |
378 | |
379 | static void |
380 | determine_versionability (struct cgraph_node *node, |
381 | class ipa_node_params *info) |
382 | { |
383 | const char *reason = NULL; |
384 | |
385 | /* There are a number of generic reasons functions cannot be versioned. We |
386 | also cannot remove parameters if there are type attributes such as fnspec |
387 | present. */ |
388 | if (node->alias || node->thunk) |
389 | reason = "alias or thunk" ; |
390 | else if (!node->versionable) |
391 | reason = "not a tree_versionable_function" ; |
392 | else if (node->get_availability () <= AVAIL_INTERPOSABLE) |
393 | reason = "insufficient body availability" ; |
394 | else if (!opt_for_fn (node->decl, optimize) |
395 | || !opt_for_fn (node->decl, flag_ipa_cp)) |
396 | reason = "non-optimized function" ; |
397 | else if (lookup_attribute (attr_name: "omp declare simd" , DECL_ATTRIBUTES (node->decl))) |
398 | { |
399 | /* Ideally we should clone the SIMD clones themselves and create |
400 | vector copies of them, so IPA-cp and SIMD clones can happily |
401 | coexist, but that may not be worth the effort. */ |
402 | reason = "function has SIMD clones" ; |
403 | } |
404 | else if (lookup_attribute (attr_name: "target_clones" , DECL_ATTRIBUTES (node->decl))) |
405 | { |
406 | /* Ideally we should clone the target clones themselves and create |
407 | copies of them, so IPA-cp and target clones can happily |
408 | coexist, but that may not be worth the effort. */ |
409 | reason = "function target_clones attribute" ; |
410 | } |
411 | /* Don't clone decls local to a comdat group; it breaks and for C++ |
412 | decloned constructors, inlining is always better anyway. */ |
413 | else if (node->comdat_local_p ()) |
414 | reason = "comdat-local function" ; |
415 | else if (node->calls_comdat_local) |
416 | { |
417 | /* TODO: call is versionable if we make sure that all |
418 | callers are inside of a comdat group. */ |
419 | reason = "calls comdat-local function" ; |
420 | } |
421 | |
422 | /* Functions calling BUILT_IN_VA_ARG_PACK and BUILT_IN_VA_ARG_PACK_LEN |
423 | work only when inlined. Cloning them may still lead to better code |
424 | because ipa-cp will not give up on cloning further. If the function is |
425 | external this however leads to wrong code because we may end up producing |
426 | offline copy of the function. */ |
427 | if (DECL_EXTERNAL (node->decl)) |
428 | for (cgraph_edge *edge = node->callees; !reason && edge; |
429 | edge = edge->next_callee) |
430 | if (fndecl_built_in_p (node: edge->callee->decl, klass: BUILT_IN_NORMAL)) |
431 | { |
432 | if (DECL_FUNCTION_CODE (decl: edge->callee->decl) == BUILT_IN_VA_ARG_PACK) |
433 | reason = "external function which calls va_arg_pack" ; |
434 | if (DECL_FUNCTION_CODE (decl: edge->callee->decl) |
435 | == BUILT_IN_VA_ARG_PACK_LEN) |
436 | reason = "external function which calls va_arg_pack_len" ; |
437 | } |
438 | |
439 | if (reason && dump_file && !node->alias && !node->thunk) |
440 | fprintf (stream: dump_file, format: "Function %s is not versionable, reason: %s.\n" , |
441 | node->dump_name (), reason); |
442 | |
443 | info->versionable = (reason == NULL); |
444 | } |
445 | |
446 | /* Return true if it is at all technically possible to create clones of a |
447 | NODE. */ |
448 | |
449 | static bool |
450 | ipcp_versionable_function_p (struct cgraph_node *node) |
451 | { |
452 | ipa_node_params *info = ipa_node_params_sum->get (node); |
453 | return info && info->versionable; |
454 | } |
455 | |
456 | /* Structure holding accumulated information about callers of a node. */ |
457 | |
458 | struct caller_statistics |
459 | { |
460 | /* If requested (see below), self-recursive call counts are summed into this |
461 | field. */ |
462 | profile_count rec_count_sum; |
463 | /* The sum of all ipa counts of all the other (non-recursive) calls. */ |
464 | profile_count count_sum; |
465 | /* Sum of all frequencies for all calls. */ |
466 | sreal freq_sum; |
467 | /* Number of calls and hot calls respectively. */ |
468 | int n_calls, n_hot_calls; |
469 | /* If itself is set up, also count the number of non-self-recursive |
470 | calls. */ |
471 | int n_nonrec_calls; |
472 | /* If non-NULL, this is the node itself and calls from it should have their |
473 | counts included in rec_count_sum and not count_sum. */ |
474 | cgraph_node *itself; |
475 | }; |
476 | |
477 | /* Initialize fields of STAT to zeroes and optionally set it up so that edges |
478 | from IGNORED_CALLER are not counted. */ |
479 | |
480 | static inline void |
481 | init_caller_stats (caller_statistics *stats, cgraph_node *itself = NULL) |
482 | { |
483 | stats->rec_count_sum = profile_count::zero (); |
484 | stats->count_sum = profile_count::zero (); |
485 | stats->n_calls = 0; |
486 | stats->n_hot_calls = 0; |
487 | stats->n_nonrec_calls = 0; |
488 | stats->freq_sum = 0; |
489 | stats->itself = itself; |
490 | } |
491 | |
492 | /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of |
493 | non-thunk incoming edges to NODE. */ |
494 | |
495 | static bool |
496 | gather_caller_stats (struct cgraph_node *node, void *data) |
497 | { |
498 | struct caller_statistics *stats = (struct caller_statistics *) data; |
499 | struct cgraph_edge *cs; |
500 | |
501 | for (cs = node->callers; cs; cs = cs->next_caller) |
502 | if (!cs->caller->thunk) |
503 | { |
504 | ipa_node_params *info = ipa_node_params_sum->get (node: cs->caller); |
505 | if (info && info->node_dead) |
506 | continue; |
507 | |
508 | if (cs->count.ipa ().initialized_p ()) |
509 | { |
510 | if (stats->itself && stats->itself == cs->caller) |
511 | stats->rec_count_sum += cs->count.ipa (); |
512 | else |
513 | stats->count_sum += cs->count.ipa (); |
514 | } |
515 | stats->freq_sum += cs->sreal_frequency (); |
516 | stats->n_calls++; |
517 | if (stats->itself && stats->itself != cs->caller) |
518 | stats->n_nonrec_calls++; |
519 | |
520 | if (cs->maybe_hot_p ()) |
521 | stats->n_hot_calls ++; |
522 | } |
523 | return false; |
524 | |
525 | } |
526 | |
527 | /* Return true if this NODE is viable candidate for cloning. */ |
528 | |
529 | static bool |
530 | ipcp_cloning_candidate_p (struct cgraph_node *node) |
531 | { |
532 | struct caller_statistics stats; |
533 | |
534 | gcc_checking_assert (node->has_gimple_body_p ()); |
535 | |
536 | if (!opt_for_fn (node->decl, flag_ipa_cp_clone)) |
537 | { |
538 | if (dump_file) |
539 | fprintf (stream: dump_file, format: "Not considering %s for cloning; " |
540 | "-fipa-cp-clone disabled.\n" , |
541 | node->dump_name ()); |
542 | return false; |
543 | } |
544 | |
545 | if (node->optimize_for_size_p ()) |
546 | { |
547 | if (dump_file) |
548 | fprintf (stream: dump_file, format: "Not considering %s for cloning; " |
549 | "optimizing it for size.\n" , |
550 | node->dump_name ()); |
551 | return false; |
552 | } |
553 | |
554 | init_caller_stats (stats: &stats); |
555 | node->call_for_symbol_thunks_and_aliases (callback: gather_caller_stats, data: &stats, include_overwritable: false); |
556 | |
557 | if (ipa_size_summaries->get (node)->self_size < stats.n_calls) |
558 | { |
559 | if (dump_file) |
560 | fprintf (stream: dump_file, format: "Considering %s for cloning; code might shrink.\n" , |
561 | node->dump_name ()); |
562 | return true; |
563 | } |
564 | |
565 | /* When profile is available and function is hot, propagate into it even if |
566 | calls seems cold; constant propagation can improve function's speed |
567 | significantly. */ |
568 | if (stats.count_sum > profile_count::zero () |
569 | && node->count.ipa ().initialized_p ()) |
570 | { |
571 | if (stats.count_sum > node->count.ipa ().apply_scale (num: 90, den: 100)) |
572 | { |
573 | if (dump_file) |
574 | fprintf (stream: dump_file, format: "Considering %s for cloning; " |
575 | "usually called directly.\n" , |
576 | node->dump_name ()); |
577 | return true; |
578 | } |
579 | } |
580 | if (!stats.n_hot_calls) |
581 | { |
582 | if (dump_file) |
583 | fprintf (stream: dump_file, format: "Not considering %s for cloning; no hot calls.\n" , |
584 | node->dump_name ()); |
585 | return false; |
586 | } |
587 | if (dump_file) |
588 | fprintf (stream: dump_file, format: "Considering %s for cloning.\n" , |
589 | node->dump_name ()); |
590 | return true; |
591 | } |
592 | |
593 | template <typename valtype> |
594 | class value_topo_info |
595 | { |
596 | public: |
597 | /* Head of the linked list of topologically sorted values. */ |
598 | ipcp_value<valtype> *values_topo; |
599 | /* Stack for creating SCCs, represented by a linked list too. */ |
600 | ipcp_value<valtype> *stack; |
601 | /* Counter driving the algorithm in add_val_to_toposort. */ |
602 | int dfs_counter; |
603 | |
604 | value_topo_info () : values_topo (NULL), stack (NULL), dfs_counter (0) |
605 | {} |
606 | void add_val (ipcp_value<valtype> *cur_val); |
607 | void propagate_effects (); |
608 | }; |
609 | |
610 | /* Arrays representing a topological ordering of call graph nodes and a stack |
611 | of nodes used during constant propagation and also data required to perform |
612 | topological sort of values and propagation of benefits in the determined |
613 | order. */ |
614 | |
615 | class ipa_topo_info |
616 | { |
617 | public: |
618 | /* Array with obtained topological order of cgraph nodes. */ |
619 | struct cgraph_node **order; |
620 | /* Stack of cgraph nodes used during propagation within SCC until all values |
621 | in the SCC stabilize. */ |
622 | struct cgraph_node **stack; |
623 | int nnodes, stack_top; |
624 | |
625 | value_topo_info<tree> constants; |
626 | value_topo_info<ipa_polymorphic_call_context> contexts; |
627 | |
628 | ipa_topo_info () : order(NULL), stack(NULL), nnodes(0), stack_top(0), |
629 | constants () |
630 | {} |
631 | }; |
632 | |
633 | /* Skip edges from and to nodes without ipa_cp enabled. |
634 | Ignore not available symbols. */ |
635 | |
636 | static bool |
637 | ignore_edge_p (cgraph_edge *e) |
638 | { |
639 | enum availability avail; |
640 | cgraph_node *ultimate_target |
641 | = e->callee->function_or_virtual_thunk_symbol (avail: &avail, ref: e->caller); |
642 | |
643 | return (avail <= AVAIL_INTERPOSABLE |
644 | || !opt_for_fn (ultimate_target->decl, optimize) |
645 | || !opt_for_fn (ultimate_target->decl, flag_ipa_cp)); |
646 | } |
647 | |
648 | /* Allocate the arrays in TOPO and topologically sort the nodes into order. */ |
649 | |
650 | static void |
651 | build_toporder_info (class ipa_topo_info *topo) |
652 | { |
653 | topo->order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count); |
654 | topo->stack = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count); |
655 | |
656 | gcc_checking_assert (topo->stack_top == 0); |
657 | topo->nnodes = ipa_reduced_postorder (topo->order, true, |
658 | ignore_edge: ignore_edge_p); |
659 | } |
660 | |
661 | /* Free information about strongly connected components and the arrays in |
662 | TOPO. */ |
663 | |
664 | static void |
665 | free_toporder_info (class ipa_topo_info *topo) |
666 | { |
667 | ipa_free_postorder_info (); |
668 | free (ptr: topo->order); |
669 | free (ptr: topo->stack); |
670 | } |
671 | |
672 | /* Add NODE to the stack in TOPO, unless it is already there. */ |
673 | |
674 | static inline void |
675 | push_node_to_stack (class ipa_topo_info *topo, struct cgraph_node *node) |
676 | { |
677 | ipa_node_params *info = ipa_node_params_sum->get (node); |
678 | if (info->node_enqueued) |
679 | return; |
680 | info->node_enqueued = 1; |
681 | topo->stack[topo->stack_top++] = node; |
682 | } |
683 | |
684 | /* Pop a node from the stack in TOPO and return it or return NULL if the stack |
685 | is empty. */ |
686 | |
687 | static struct cgraph_node * |
688 | pop_node_from_stack (class ipa_topo_info *topo) |
689 | { |
690 | if (topo->stack_top) |
691 | { |
692 | struct cgraph_node *node; |
693 | topo->stack_top--; |
694 | node = topo->stack[topo->stack_top]; |
695 | ipa_node_params_sum->get (node)->node_enqueued = 0; |
696 | return node; |
697 | } |
698 | else |
699 | return NULL; |
700 | } |
701 | |
702 | /* Set lattice LAT to bottom and return true if it previously was not set as |
703 | such. */ |
704 | |
705 | template <typename valtype> |
706 | inline bool |
707 | ipcp_lattice<valtype>::set_to_bottom () |
708 | { |
709 | bool ret = !bottom; |
710 | bottom = true; |
711 | return ret; |
712 | } |
713 | |
714 | /* Mark lattice as containing an unknown value and return true if it previously |
715 | was not marked as such. */ |
716 | |
717 | template <typename valtype> |
718 | inline bool |
719 | ipcp_lattice<valtype>::set_contains_variable () |
720 | { |
721 | bool ret = !contains_variable; |
722 | contains_variable = true; |
723 | return ret; |
724 | } |
725 | |
726 | /* Set all aggregate lattices in PLATS to bottom and return true if they were |
727 | not previously set as such. */ |
728 | |
729 | static inline bool |
730 | set_agg_lats_to_bottom (class ipcp_param_lattices *plats) |
731 | { |
732 | bool ret = !plats->aggs_bottom; |
733 | plats->aggs_bottom = true; |
734 | return ret; |
735 | } |
736 | |
737 | /* Mark all aggregate lattices in PLATS as containing an unknown value and |
738 | return true if they were not previously marked as such. */ |
739 | |
740 | static inline bool |
741 | set_agg_lats_contain_variable (class ipcp_param_lattices *plats) |
742 | { |
743 | bool ret = !plats->aggs_contain_variable; |
744 | plats->aggs_contain_variable = true; |
745 | return ret; |
746 | } |
747 | |
748 | bool |
749 | ipcp_vr_lattice::meet_with (const ipcp_vr_lattice &other) |
750 | { |
751 | return meet_with_1 (other_vr: other.m_vr); |
752 | } |
753 | |
754 | /* Meet the current value of the lattice with the range described by |
755 | P_VR. */ |
756 | |
757 | bool |
758 | ipcp_vr_lattice::meet_with (const vrange &p_vr) |
759 | { |
760 | return meet_with_1 (other_vr: p_vr); |
761 | } |
762 | |
763 | /* Meet the current value of the lattice with the range described by |
764 | OTHER_VR. Return TRUE if anything changed. */ |
765 | |
766 | bool |
767 | ipcp_vr_lattice::meet_with_1 (const vrange &other_vr) |
768 | { |
769 | if (bottom_p ()) |
770 | return false; |
771 | |
772 | if (other_vr.varying_p ()) |
773 | return set_to_bottom (); |
774 | |
775 | bool res; |
776 | if (flag_checking) |
777 | { |
778 | Value_Range save (m_vr); |
779 | res = m_vr.union_ (r: other_vr); |
780 | gcc_assert (res == (m_vr != save)); |
781 | } |
782 | else |
783 | res = m_vr.union_ (r: other_vr); |
784 | return res; |
785 | } |
786 | |
787 | /* Return true if value range information in the lattice is yet unknown. */ |
788 | |
789 | bool |
790 | ipcp_vr_lattice::top_p () const |
791 | { |
792 | return m_vr.undefined_p (); |
793 | } |
794 | |
795 | /* Return true if value range information in the lattice is known to be |
796 | unusable. */ |
797 | |
798 | bool |
799 | ipcp_vr_lattice::bottom_p () const |
800 | { |
801 | return m_vr.varying_p (); |
802 | } |
803 | |
804 | /* Set value range information in the lattice to bottom. Return true if it |
805 | previously was in a different state. */ |
806 | |
807 | bool |
808 | ipcp_vr_lattice::set_to_bottom () |
809 | { |
810 | if (m_vr.varying_p ()) |
811 | return false; |
812 | |
813 | /* Setting an unsupported type here forces the temporary to default |
814 | to unsupported_range, which can handle VARYING/DEFINED ranges, |
815 | but nothing else (union, intersect, etc). This allows us to set |
816 | bottoms on any ranges, and is safe as all users of the lattice |
817 | check for bottom first. */ |
818 | m_vr.set_type (void_type_node); |
819 | m_vr.set_varying (void_type_node); |
820 | |
821 | return true; |
822 | } |
823 | |
824 | /* Set lattice value to bottom, if it already isn't the case. */ |
825 | |
826 | bool |
827 | ipcp_bits_lattice::set_to_bottom () |
828 | { |
829 | if (bottom_p ()) |
830 | return false; |
831 | m_lattice_val = IPA_BITS_VARYING; |
832 | m_value = 0; |
833 | m_mask = -1; |
834 | return true; |
835 | } |
836 | |
837 | /* Set to constant if it isn't already. Only meant to be called |
838 | when switching state from TOP. */ |
839 | |
840 | bool |
841 | ipcp_bits_lattice::set_to_constant (widest_int value, widest_int mask) |
842 | { |
843 | gcc_assert (top_p ()); |
844 | m_lattice_val = IPA_BITS_CONSTANT; |
845 | m_value = wi::bit_and (x: wi::bit_not (x: mask), y: value); |
846 | m_mask = mask; |
847 | return true; |
848 | } |
849 | |
850 | /* Return true if any of the known bits are non-zero. */ |
851 | |
852 | bool |
853 | ipcp_bits_lattice::known_nonzero_p () const |
854 | { |
855 | if (!constant_p ()) |
856 | return false; |
857 | return wi::ne_p (x: wi::bit_and (x: wi::bit_not (x: m_mask), y: m_value), y: 0); |
858 | } |
859 | |
860 | /* Convert operand to value, mask form. */ |
861 | |
862 | void |
863 | ipcp_bits_lattice::get_value_and_mask (tree operand, widest_int *valuep, widest_int *maskp) |
864 | { |
865 | wide_int get_nonzero_bits (const_tree); |
866 | |
867 | if (TREE_CODE (operand) == INTEGER_CST) |
868 | { |
869 | *valuep = wi::to_widest (t: operand); |
870 | *maskp = 0; |
871 | } |
872 | else |
873 | { |
874 | *valuep = 0; |
875 | *maskp = -1; |
876 | } |
877 | } |
878 | |
879 | /* Meet operation, similar to ccp_lattice_meet, we xor values |
880 | if this->value, value have different values at same bit positions, we want |
881 | to drop that bit to varying. Return true if mask is changed. |
882 | This function assumes that the lattice value is in CONSTANT state. If |
883 | DROP_ALL_ONES, mask out any known bits with value one afterwards. */ |
884 | |
885 | bool |
886 | ipcp_bits_lattice::meet_with_1 (widest_int value, widest_int mask, |
887 | unsigned precision, bool drop_all_ones) |
888 | { |
889 | gcc_assert (constant_p ()); |
890 | |
891 | widest_int old_mask = m_mask; |
892 | m_mask = (m_mask | mask) | (m_value ^ value); |
893 | if (drop_all_ones) |
894 | m_mask |= m_value; |
895 | m_value &= ~m_mask; |
896 | |
897 | if (wi::sext (x: m_mask, offset: precision) == -1) |
898 | return set_to_bottom (); |
899 | |
900 | return m_mask != old_mask; |
901 | } |
902 | |
903 | /* Meet the bits lattice with operand |
904 | described by <value, mask, sgn, precision. */ |
905 | |
906 | bool |
907 | ipcp_bits_lattice::meet_with (widest_int value, widest_int mask, |
908 | unsigned precision) |
909 | { |
910 | if (bottom_p ()) |
911 | return false; |
912 | |
913 | if (top_p ()) |
914 | { |
915 | if (wi::sext (x: mask, offset: precision) == -1) |
916 | return set_to_bottom (); |
917 | return set_to_constant (value, mask); |
918 | } |
919 | |
920 | return meet_with_1 (value, mask, precision, drop_all_ones: false); |
921 | } |
922 | |
923 | /* Meet bits lattice with the result of bit_value_binop (other, operand) |
924 | if code is binary operation or bit_value_unop (other) if code is unary op. |
925 | In the case when code is nop_expr, no adjustment is required. If |
926 | DROP_ALL_ONES, mask out any known bits with value one afterwards. */ |
927 | |
928 | bool |
929 | ipcp_bits_lattice::meet_with (ipcp_bits_lattice& other, unsigned precision, |
930 | signop sgn, enum tree_code code, tree operand, |
931 | bool drop_all_ones) |
932 | { |
933 | if (other.bottom_p ()) |
934 | return set_to_bottom (); |
935 | |
936 | if (bottom_p () || other.top_p ()) |
937 | return false; |
938 | |
939 | widest_int adjusted_value, adjusted_mask; |
940 | |
941 | if (TREE_CODE_CLASS (code) == tcc_binary) |
942 | { |
943 | tree type = TREE_TYPE (operand); |
944 | widest_int o_value, o_mask; |
945 | get_value_and_mask (operand, valuep: &o_value, maskp: &o_mask); |
946 | |
947 | bit_value_binop (code, sgn, precision, &adjusted_value, &adjusted_mask, |
948 | sgn, precision, other.get_value (), other.get_mask (), |
949 | TYPE_SIGN (type), TYPE_PRECISION (type), o_value, o_mask); |
950 | |
951 | if (wi::sext (x: adjusted_mask, offset: precision) == -1) |
952 | return set_to_bottom (); |
953 | } |
954 | |
955 | else if (TREE_CODE_CLASS (code) == tcc_unary) |
956 | { |
957 | bit_value_unop (code, sgn, precision, &adjusted_value, |
958 | &adjusted_mask, sgn, precision, other.get_value (), |
959 | other.get_mask ()); |
960 | |
961 | if (wi::sext (x: adjusted_mask, offset: precision) == -1) |
962 | return set_to_bottom (); |
963 | } |
964 | |
965 | else |
966 | return set_to_bottom (); |
967 | |
968 | if (top_p ()) |
969 | { |
970 | if (drop_all_ones) |
971 | { |
972 | adjusted_mask |= adjusted_value; |
973 | adjusted_value &= ~adjusted_mask; |
974 | } |
975 | if (wi::sext (x: adjusted_mask, offset: precision) == -1) |
976 | return set_to_bottom (); |
977 | return set_to_constant (value: adjusted_value, mask: adjusted_mask); |
978 | } |
979 | else |
980 | return meet_with_1 (value: adjusted_value, mask: adjusted_mask, precision, |
981 | drop_all_ones); |
982 | } |
983 | |
984 | /* Dump the contents of the list to FILE. */ |
985 | |
986 | void |
987 | ipa_argagg_value_list::dump (FILE *f) |
988 | { |
989 | bool comma = false; |
990 | for (const ipa_argagg_value &av : m_elts) |
991 | { |
992 | fprintf (stream: f, format: "%s %i[%u]=" , comma ? "," : "" , |
993 | av.index, av.unit_offset); |
994 | print_generic_expr (f, av.value); |
995 | if (av.by_ref) |
996 | fprintf (stream: f, format: "(by_ref)" ); |
997 | if (av.killed) |
998 | fprintf (stream: f, format: "(killed)" ); |
999 | comma = true; |
1000 | } |
1001 | fprintf (stream: f, format: "\n" ); |
1002 | } |
1003 | |
1004 | /* Dump the contents of the list to stderr. */ |
1005 | |
1006 | void |
1007 | ipa_argagg_value_list::debug () |
1008 | { |
1009 | dump (stderr); |
1010 | } |
1011 | |
1012 | /* Return the item describing a constant stored for INDEX at UNIT_OFFSET or |
1013 | NULL if there is no such constant. */ |
1014 | |
1015 | const ipa_argagg_value * |
1016 | ipa_argagg_value_list::get_elt (int index, unsigned unit_offset) const |
1017 | { |
1018 | ipa_argagg_value key; |
1019 | key.index = index; |
1020 | key.unit_offset = unit_offset; |
1021 | const ipa_argagg_value *res |
1022 | = std::lower_bound (first: m_elts.begin (), last: m_elts.end (), val: key, |
1023 | comp: [] (const ipa_argagg_value &elt, |
1024 | const ipa_argagg_value &val) |
1025 | { |
1026 | if (elt.index < val.index) |
1027 | return true; |
1028 | if (elt.index > val.index) |
1029 | return false; |
1030 | if (elt.unit_offset < val.unit_offset) |
1031 | return true; |
1032 | return false; |
1033 | }); |
1034 | |
1035 | if (res == m_elts.end () |
1036 | || res->index != index |
1037 | || res->unit_offset != unit_offset) |
1038 | res = nullptr; |
1039 | |
1040 | /* TODO: perhaps remove the check (that the underlying array is indeed |
1041 | sorted) if it turns out it can be too slow? */ |
1042 | if (!flag_checking) |
1043 | return res; |
1044 | |
1045 | const ipa_argagg_value *slow_res = NULL; |
1046 | int prev_index = -1; |
1047 | unsigned prev_unit_offset = 0; |
1048 | for (const ipa_argagg_value &av : m_elts) |
1049 | { |
1050 | gcc_assert (prev_index < 0 |
1051 | || prev_index < av.index |
1052 | || prev_unit_offset < av.unit_offset); |
1053 | prev_index = av.index; |
1054 | prev_unit_offset = av.unit_offset; |
1055 | if (av.index == index |
1056 | && av.unit_offset == unit_offset) |
1057 | slow_res = &av; |
1058 | } |
1059 | gcc_assert (res == slow_res); |
1060 | |
1061 | return res; |
1062 | } |
1063 | |
1064 | /* Return the first item describing a constant stored for parameter with INDEX, |
1065 | regardless of offset or reference, or NULL if there is no such constant. */ |
1066 | |
1067 | const ipa_argagg_value * |
1068 | ipa_argagg_value_list::get_elt_for_index (int index) const |
1069 | { |
1070 | const ipa_argagg_value *res |
1071 | = std::lower_bound (first: m_elts.begin (), last: m_elts.end (), val: index, |
1072 | comp: [] (const ipa_argagg_value &elt, unsigned idx) |
1073 | { |
1074 | return elt.index < idx; |
1075 | }); |
1076 | if (res == m_elts.end () |
1077 | || res->index != index) |
1078 | res = nullptr; |
1079 | return res; |
1080 | } |
1081 | |
1082 | /* Return the aggregate constant stored for INDEX at UNIT_OFFSET, not |
1083 | performing any check of whether value is passed by reference, or NULL_TREE |
1084 | if there is no such constant. */ |
1085 | |
1086 | tree |
1087 | ipa_argagg_value_list::get_value (int index, unsigned unit_offset) const |
1088 | { |
1089 | const ipa_argagg_value *av = get_elt (index, unit_offset); |
1090 | return av ? av->value : NULL_TREE; |
1091 | } |
1092 | |
1093 | /* Return the aggregate constant stored for INDEX at UNIT_OFFSET, if it is |
1094 | passed by reference or not according to BY_REF, or NULL_TREE if there is |
1095 | no such constant. */ |
1096 | |
1097 | tree |
1098 | ipa_argagg_value_list::get_value (int index, unsigned unit_offset, |
1099 | bool by_ref) const |
1100 | { |
1101 | const ipa_argagg_value *av = get_elt (index, unit_offset); |
1102 | if (av && av->by_ref == by_ref) |
1103 | return av->value; |
1104 | return NULL_TREE; |
1105 | } |
1106 | |
1107 | /* Return true if all elements present in OTHER are also present in this |
1108 | list. */ |
1109 | |
1110 | bool |
1111 | ipa_argagg_value_list::superset_of_p (const ipa_argagg_value_list &other) const |
1112 | { |
1113 | unsigned j = 0; |
1114 | for (unsigned i = 0; i < other.m_elts.size (); i++) |
1115 | { |
1116 | unsigned other_index = other.m_elts[i].index; |
1117 | unsigned other_offset = other.m_elts[i].unit_offset; |
1118 | |
1119 | while (j < m_elts.size () |
1120 | && (m_elts[j].index < other_index |
1121 | || (m_elts[j].index == other_index |
1122 | && m_elts[j].unit_offset < other_offset))) |
1123 | j++; |
1124 | |
1125 | if (j >= m_elts.size () |
1126 | || m_elts[j].index != other_index |
1127 | || m_elts[j].unit_offset != other_offset |
1128 | || m_elts[j].by_ref != other.m_elts[i].by_ref |
1129 | || !m_elts[j].value |
1130 | || !values_equal_for_ipcp_p (x: m_elts[j].value, y: other.m_elts[i].value)) |
1131 | return false; |
1132 | } |
1133 | return true; |
1134 | } |
1135 | |
1136 | /* Push all items in this list that describe parameter SRC_INDEX into RES as |
1137 | ones describing DST_INDEX while subtracting UNIT_DELTA from their unit |
1138 | offsets but skip those which would end up with a negative offset. */ |
1139 | |
1140 | void |
1141 | ipa_argagg_value_list::push_adjusted_values (unsigned src_index, |
1142 | unsigned dest_index, |
1143 | unsigned unit_delta, |
1144 | vec<ipa_argagg_value> *res) const |
1145 | { |
1146 | const ipa_argagg_value *av = get_elt_for_index (index: src_index); |
1147 | if (!av) |
1148 | return; |
1149 | unsigned prev_unit_offset = 0; |
1150 | bool first = true; |
1151 | for (; av < m_elts.end (); ++av) |
1152 | { |
1153 | if (av->index > src_index) |
1154 | return; |
1155 | if (av->index == src_index |
1156 | && (av->unit_offset >= unit_delta) |
1157 | && av->value) |
1158 | { |
1159 | ipa_argagg_value new_av; |
1160 | gcc_checking_assert (av->value); |
1161 | new_av.value = av->value; |
1162 | new_av.unit_offset = av->unit_offset - unit_delta; |
1163 | new_av.index = dest_index; |
1164 | new_av.by_ref = av->by_ref; |
1165 | gcc_assert (!av->killed); |
1166 | new_av.killed = false; |
1167 | |
1168 | /* Quick check that the offsets we push are indeed increasing. */ |
1169 | gcc_assert (first |
1170 | || new_av.unit_offset > prev_unit_offset); |
1171 | prev_unit_offset = new_av.unit_offset; |
1172 | first = false; |
1173 | |
1174 | res->safe_push (obj: new_av); |
1175 | } |
1176 | } |
1177 | } |
1178 | |
1179 | /* Push to RES information about single lattices describing aggregate values in |
1180 | PLATS as those describing parameter DEST_INDEX and the original offset minus |
1181 | UNIT_DELTA. Return true if any item has been pushed to RES. */ |
1182 | |
1183 | static bool |
1184 | push_agg_values_from_plats (ipcp_param_lattices *plats, int dest_index, |
1185 | unsigned unit_delta, |
1186 | vec<ipa_argagg_value> *res) |
1187 | { |
1188 | if (plats->aggs_contain_variable) |
1189 | return false; |
1190 | |
1191 | bool pushed_sth = false; |
1192 | bool first = true; |
1193 | unsigned prev_unit_offset = 0; |
1194 | for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next) |
1195 | if (aglat->is_single_const () |
1196 | && (aglat->offset / BITS_PER_UNIT - unit_delta) >= 0) |
1197 | { |
1198 | ipa_argagg_value iav; |
1199 | iav.value = aglat->values->value; |
1200 | iav.unit_offset = aglat->offset / BITS_PER_UNIT - unit_delta; |
1201 | iav.index = dest_index; |
1202 | iav.by_ref = plats->aggs_by_ref; |
1203 | iav.killed = false; |
1204 | |
1205 | gcc_assert (first |
1206 | || iav.unit_offset > prev_unit_offset); |
1207 | prev_unit_offset = iav.unit_offset; |
1208 | first = false; |
1209 | |
1210 | pushed_sth = true; |
1211 | res->safe_push (obj: iav); |
1212 | } |
1213 | return pushed_sth; |
1214 | } |
1215 | |
1216 | /* Turn all values in LIST that are not present in OTHER into NULL_TREEs. |
1217 | Return the number of remaining valid entries. */ |
1218 | |
1219 | static unsigned |
1220 | intersect_argaggs_with (vec<ipa_argagg_value> &elts, |
1221 | const vec<ipa_argagg_value> &other) |
1222 | { |
1223 | unsigned valid_entries = 0; |
1224 | unsigned j = 0; |
1225 | for (unsigned i = 0; i < elts.length (); i++) |
1226 | { |
1227 | if (!elts[i].value) |
1228 | continue; |
1229 | |
1230 | unsigned this_index = elts[i].index; |
1231 | unsigned this_offset = elts[i].unit_offset; |
1232 | |
1233 | while (j < other.length () |
1234 | && (other[j].index < this_index |
1235 | || (other[j].index == this_index |
1236 | && other[j].unit_offset < this_offset))) |
1237 | j++; |
1238 | |
1239 | if (j >= other.length ()) |
1240 | { |
1241 | elts[i].value = NULL_TREE; |
1242 | continue; |
1243 | } |
1244 | |
1245 | if (other[j].index == this_index |
1246 | && other[j].unit_offset == this_offset |
1247 | && other[j].by_ref == elts[i].by_ref |
1248 | && other[j].value |
1249 | && values_equal_for_ipcp_p (x: other[j].value, y: elts[i].value)) |
1250 | valid_entries++; |
1251 | else |
1252 | elts[i].value = NULL_TREE; |
1253 | } |
1254 | return valid_entries; |
1255 | } |
1256 | |
1257 | /* Mark bot aggregate and scalar lattices as containing an unknown variable, |
1258 | return true is any of them has not been marked as such so far. */ |
1259 | |
1260 | static inline bool |
1261 | set_all_contains_variable (class ipcp_param_lattices *plats) |
1262 | { |
1263 | bool ret; |
1264 | ret = plats->itself.set_contains_variable (); |
1265 | ret |= plats->ctxlat.set_contains_variable (); |
1266 | ret |= set_agg_lats_contain_variable (plats); |
1267 | ret |= plats->bits_lattice.set_to_bottom (); |
1268 | ret |= plats->m_value_range.set_to_bottom (); |
1269 | return ret; |
1270 | } |
1271 | |
1272 | /* Worker of call_for_symbol_thunks_and_aliases, increment the integer DATA |
1273 | points to by the number of callers to NODE. */ |
1274 | |
1275 | static bool |
1276 | count_callers (cgraph_node *node, void *data) |
1277 | { |
1278 | int *caller_count = (int *) data; |
1279 | |
1280 | for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller) |
1281 | /* Local thunks can be handled transparently, but if the thunk cannot |
1282 | be optimized out, count it as a real use. */ |
1283 | if (!cs->caller->thunk || !cs->caller->local) |
1284 | ++*caller_count; |
1285 | return false; |
1286 | } |
1287 | |
1288 | /* Worker of call_for_symbol_thunks_and_aliases, it is supposed to be called on |
1289 | the one caller of some other node. Set the caller's corresponding flag. */ |
1290 | |
1291 | static bool |
1292 | set_single_call_flag (cgraph_node *node, void *) |
1293 | { |
1294 | cgraph_edge *cs = node->callers; |
1295 | /* Local thunks can be handled transparently, skip them. */ |
1296 | while (cs && cs->caller->thunk && cs->caller->local) |
1297 | cs = cs->next_caller; |
1298 | if (cs) |
1299 | if (ipa_node_params* info = ipa_node_params_sum->get (node: cs->caller)) |
1300 | { |
1301 | info->node_calling_single_call = true; |
1302 | return true; |
1303 | } |
1304 | return false; |
1305 | } |
1306 | |
1307 | /* Initialize ipcp_lattices. */ |
1308 | |
1309 | static void |
1310 | initialize_node_lattices (struct cgraph_node *node) |
1311 | { |
1312 | ipa_node_params *info = ipa_node_params_sum->get (node); |
1313 | struct cgraph_edge *ie; |
1314 | bool disable = false, variable = false; |
1315 | int i; |
1316 | |
1317 | gcc_checking_assert (node->has_gimple_body_p ()); |
1318 | |
1319 | if (!ipa_get_param_count (info)) |
1320 | disable = true; |
1321 | else if (node->local) |
1322 | { |
1323 | int caller_count = 0; |
1324 | node->call_for_symbol_thunks_and_aliases (callback: count_callers, data: &caller_count, |
1325 | include_overwritable: true); |
1326 | gcc_checking_assert (caller_count > 0); |
1327 | if (caller_count == 1) |
1328 | node->call_for_symbol_thunks_and_aliases (callback: set_single_call_flag, |
1329 | NULL, include_overwritable: true); |
1330 | } |
1331 | else |
1332 | { |
1333 | /* When cloning is allowed, we can assume that externally visible |
1334 | functions are not called. We will compensate this by cloning |
1335 | later. */ |
1336 | if (ipcp_versionable_function_p (node) |
1337 | && ipcp_cloning_candidate_p (node)) |
1338 | variable = true; |
1339 | else |
1340 | disable = true; |
1341 | } |
1342 | |
1343 | if (dump_file && (dump_flags & TDF_DETAILS) |
1344 | && !node->alias && !node->thunk) |
1345 | { |
1346 | fprintf (stream: dump_file, format: "Initializing lattices of %s\n" , |
1347 | node->dump_name ()); |
1348 | if (disable || variable) |
1349 | fprintf (stream: dump_file, format: " Marking all lattices as %s\n" , |
1350 | disable ? "BOTTOM" : "VARIABLE" ); |
1351 | } |
1352 | |
1353 | auto_vec<bool, 16> surviving_params; |
1354 | bool pre_modified = false; |
1355 | |
1356 | clone_info *cinfo = clone_info::get (node); |
1357 | |
1358 | if (!disable && cinfo && cinfo->param_adjustments) |
1359 | { |
1360 | /* At the moment all IPA optimizations should use the number of |
1361 | parameters of the prevailing decl as the m_always_copy_start. |
1362 | Handling any other value would complicate the code below, so for the |
1363 | time bing let's only assert it is so. */ |
1364 | gcc_assert ((cinfo->param_adjustments->m_always_copy_start |
1365 | == ipa_get_param_count (info)) |
1366 | || cinfo->param_adjustments->m_always_copy_start < 0); |
1367 | |
1368 | pre_modified = true; |
1369 | cinfo->param_adjustments->get_surviving_params (surviving_params: &surviving_params); |
1370 | |
1371 | if (dump_file && (dump_flags & TDF_DETAILS) |
1372 | && !node->alias && !node->thunk) |
1373 | { |
1374 | bool first = true; |
1375 | for (int j = 0; j < ipa_get_param_count (info); j++) |
1376 | { |
1377 | if (j < (int) surviving_params.length () |
1378 | && surviving_params[j]) |
1379 | continue; |
1380 | if (first) |
1381 | { |
1382 | fprintf (stream: dump_file, |
1383 | format: " The following parameters are dead on arrival:" ); |
1384 | first = false; |
1385 | } |
1386 | fprintf (stream: dump_file, format: " %u" , j); |
1387 | } |
1388 | if (!first) |
1389 | fprintf (stream: dump_file, format: "\n" ); |
1390 | } |
1391 | } |
1392 | |
1393 | for (i = 0; i < ipa_get_param_count (info); i++) |
1394 | { |
1395 | ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i); |
1396 | tree type = ipa_get_type (info, i); |
1397 | if (disable |
1398 | || !ipa_get_type (info, i) |
1399 | || (pre_modified && (surviving_params.length () <= (unsigned) i |
1400 | || !surviving_params[i]))) |
1401 | { |
1402 | plats->itself.set_to_bottom (); |
1403 | plats->ctxlat.set_to_bottom (); |
1404 | set_agg_lats_to_bottom (plats); |
1405 | plats->bits_lattice.set_to_bottom (); |
1406 | plats->m_value_range.init (type); |
1407 | plats->m_value_range.set_to_bottom (); |
1408 | } |
1409 | else |
1410 | { |
1411 | plats->m_value_range.init (type); |
1412 | if (variable) |
1413 | set_all_contains_variable (plats); |
1414 | } |
1415 | } |
1416 | |
1417 | for (ie = node->indirect_calls; ie; ie = ie->next_callee) |
1418 | if (ie->indirect_info->polymorphic |
1419 | && ie->indirect_info->param_index >= 0) |
1420 | { |
1421 | gcc_checking_assert (ie->indirect_info->param_index >= 0); |
1422 | ipa_get_parm_lattices (info, |
1423 | i: ie->indirect_info->param_index)->virt_call = 1; |
1424 | } |
1425 | } |
1426 | |
1427 | /* Return true if VALUE can be safely IPA-CP propagated to a parameter of type |
1428 | PARAM_TYPE. */ |
1429 | |
1430 | static bool |
1431 | ipacp_value_safe_for_type (tree param_type, tree value) |
1432 | { |
1433 | tree val_type = TREE_TYPE (value); |
1434 | if (param_type == val_type |
1435 | || useless_type_conversion_p (param_type, val_type) |
1436 | || fold_convertible_p (param_type, value)) |
1437 | return true; |
1438 | else |
1439 | return false; |
1440 | } |
1441 | |
1442 | /* Return the result of a (possibly arithmetic) operation on the constant |
1443 | value INPUT. OPERAND is 2nd operand for binary operation. RES_TYPE is |
1444 | the type of the parameter to which the result is passed. Return |
1445 | NULL_TREE if that cannot be determined or be considered an |
1446 | interprocedural invariant. */ |
1447 | |
1448 | static tree |
1449 | ipa_get_jf_arith_result (enum tree_code opcode, tree input, tree operand, |
1450 | tree res_type) |
1451 | { |
1452 | tree res; |
1453 | |
1454 | if (opcode == NOP_EXPR) |
1455 | return input; |
1456 | if (!is_gimple_ip_invariant (input)) |
1457 | return NULL_TREE; |
1458 | |
1459 | if (opcode == ASSERT_EXPR) |
1460 | { |
1461 | if (values_equal_for_ipcp_p (x: input, y: operand)) |
1462 | return input; |
1463 | else |
1464 | return NULL_TREE; |
1465 | } |
1466 | |
1467 | if (!res_type) |
1468 | { |
1469 | if (TREE_CODE_CLASS (opcode) == tcc_comparison) |
1470 | res_type = boolean_type_node; |
1471 | else if (expr_type_first_operand_type_p (opcode)) |
1472 | res_type = TREE_TYPE (input); |
1473 | else |
1474 | return NULL_TREE; |
1475 | } |
1476 | |
1477 | if (TREE_CODE_CLASS (opcode) == tcc_unary) |
1478 | res = fold_unary (opcode, res_type, input); |
1479 | else |
1480 | res = fold_binary (opcode, res_type, input, operand); |
1481 | |
1482 | if (res && !is_gimple_ip_invariant (res)) |
1483 | return NULL_TREE; |
1484 | |
1485 | return res; |
1486 | } |
1487 | |
1488 | /* Return the result of a (possibly arithmetic) pass through jump function |
1489 | JFUNC on the constant value INPUT. RES_TYPE is the type of the parameter |
1490 | to which the result is passed. Return NULL_TREE if that cannot be |
1491 | determined or be considered an interprocedural invariant. */ |
1492 | |
1493 | static tree |
1494 | ipa_get_jf_pass_through_result (struct ipa_jump_func *jfunc, tree input, |
1495 | tree res_type) |
1496 | { |
1497 | return ipa_get_jf_arith_result (opcode: ipa_get_jf_pass_through_operation (jfunc), |
1498 | input, |
1499 | operand: ipa_get_jf_pass_through_operand (jfunc), |
1500 | res_type); |
1501 | } |
1502 | |
1503 | /* Return the result of an ancestor jump function JFUNC on the constant value |
1504 | INPUT. Return NULL_TREE if that cannot be determined. */ |
1505 | |
1506 | static tree |
1507 | ipa_get_jf_ancestor_result (struct ipa_jump_func *jfunc, tree input) |
1508 | { |
1509 | gcc_checking_assert (TREE_CODE (input) != TREE_BINFO); |
1510 | if (TREE_CODE (input) == ADDR_EXPR) |
1511 | { |
1512 | gcc_checking_assert (is_gimple_ip_invariant_address (input)); |
1513 | poly_int64 off = ipa_get_jf_ancestor_offset (jfunc); |
1514 | if (known_eq (off, 0)) |
1515 | return input; |
1516 | poly_int64 byte_offset = exact_div (a: off, BITS_PER_UNIT); |
1517 | return build1 (ADDR_EXPR, TREE_TYPE (input), |
1518 | fold_build2 (MEM_REF, TREE_TYPE (TREE_TYPE (input)), input, |
1519 | build_int_cst (ptr_type_node, byte_offset))); |
1520 | } |
1521 | else if (ipa_get_jf_ancestor_keep_null (jfunc) |
1522 | && zerop (input)) |
1523 | return input; |
1524 | else |
1525 | return NULL_TREE; |
1526 | } |
1527 | |
1528 | /* Determine whether JFUNC evaluates to a single known constant value and if |
1529 | so, return it. Otherwise return NULL. INFO describes the caller node or |
1530 | the one it is inlined to, so that pass-through jump functions can be |
1531 | evaluated. PARM_TYPE is the type of the parameter to which the result is |
1532 | passed. */ |
1533 | |
1534 | tree |
1535 | ipa_value_from_jfunc (class ipa_node_params *info, struct ipa_jump_func *jfunc, |
1536 | tree parm_type) |
1537 | { |
1538 | if (jfunc->type == IPA_JF_CONST) |
1539 | return ipa_get_jf_constant (jfunc); |
1540 | else if (jfunc->type == IPA_JF_PASS_THROUGH |
1541 | || jfunc->type == IPA_JF_ANCESTOR) |
1542 | { |
1543 | tree input; |
1544 | int idx; |
1545 | |
1546 | if (jfunc->type == IPA_JF_PASS_THROUGH) |
1547 | idx = ipa_get_jf_pass_through_formal_id (jfunc); |
1548 | else |
1549 | idx = ipa_get_jf_ancestor_formal_id (jfunc); |
1550 | |
1551 | if (info->ipcp_orig_node) |
1552 | input = info->known_csts[idx]; |
1553 | else |
1554 | { |
1555 | ipcp_lattice<tree> *lat; |
1556 | |
1557 | if (info->lattices.is_empty () |
1558 | || idx >= ipa_get_param_count (info)) |
1559 | return NULL_TREE; |
1560 | lat = ipa_get_scalar_lat (info, i: idx); |
1561 | if (!lat->is_single_const ()) |
1562 | return NULL_TREE; |
1563 | input = lat->values->value; |
1564 | } |
1565 | |
1566 | if (!input) |
1567 | return NULL_TREE; |
1568 | |
1569 | if (jfunc->type == IPA_JF_PASS_THROUGH) |
1570 | return ipa_get_jf_pass_through_result (jfunc, input, res_type: parm_type); |
1571 | else |
1572 | return ipa_get_jf_ancestor_result (jfunc, input); |
1573 | } |
1574 | else |
1575 | return NULL_TREE; |
1576 | } |
1577 | |
1578 | /* Determine whether JFUNC evaluates to single known polymorphic context, given |
1579 | that INFO describes the caller node or the one it is inlined to, CS is the |
1580 | call graph edge corresponding to JFUNC and CSIDX index of the described |
1581 | parameter. */ |
1582 | |
1583 | ipa_polymorphic_call_context |
1584 | ipa_context_from_jfunc (ipa_node_params *info, cgraph_edge *cs, int csidx, |
1585 | ipa_jump_func *jfunc) |
1586 | { |
1587 | ipa_edge_args *args = ipa_edge_args_sum->get (edge: cs); |
1588 | ipa_polymorphic_call_context ctx; |
1589 | ipa_polymorphic_call_context *edge_ctx |
1590 | = cs ? ipa_get_ith_polymorhic_call_context (args, i: csidx) : NULL; |
1591 | |
1592 | if (edge_ctx && !edge_ctx->useless_p ()) |
1593 | ctx = *edge_ctx; |
1594 | |
1595 | if (jfunc->type == IPA_JF_PASS_THROUGH |
1596 | || jfunc->type == IPA_JF_ANCESTOR) |
1597 | { |
1598 | ipa_polymorphic_call_context srcctx; |
1599 | int srcidx; |
1600 | bool type_preserved = true; |
1601 | if (jfunc->type == IPA_JF_PASS_THROUGH) |
1602 | { |
1603 | if (ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR) |
1604 | return ctx; |
1605 | type_preserved = ipa_get_jf_pass_through_type_preserved (jfunc); |
1606 | srcidx = ipa_get_jf_pass_through_formal_id (jfunc); |
1607 | } |
1608 | else |
1609 | { |
1610 | type_preserved = ipa_get_jf_ancestor_type_preserved (jfunc); |
1611 | srcidx = ipa_get_jf_ancestor_formal_id (jfunc); |
1612 | } |
1613 | if (info->ipcp_orig_node) |
1614 | { |
1615 | if (info->known_contexts.exists ()) |
1616 | srcctx = info->known_contexts[srcidx]; |
1617 | } |
1618 | else |
1619 | { |
1620 | if (info->lattices.is_empty () |
1621 | || srcidx >= ipa_get_param_count (info)) |
1622 | return ctx; |
1623 | ipcp_lattice<ipa_polymorphic_call_context> *lat; |
1624 | lat = ipa_get_poly_ctx_lat (info, i: srcidx); |
1625 | if (!lat->is_single_const ()) |
1626 | return ctx; |
1627 | srcctx = lat->values->value; |
1628 | } |
1629 | if (srcctx.useless_p ()) |
1630 | return ctx; |
1631 | if (jfunc->type == IPA_JF_ANCESTOR) |
1632 | srcctx.offset_by (off: ipa_get_jf_ancestor_offset (jfunc)); |
1633 | if (!type_preserved) |
1634 | srcctx.possible_dynamic_type_change (cs->in_polymorphic_cdtor); |
1635 | srcctx.combine_with (ctx); |
1636 | return srcctx; |
1637 | } |
1638 | |
1639 | return ctx; |
1640 | } |
1641 | |
1642 | /* Emulate effects of unary OPERATION and/or conversion from SRC_TYPE to |
1643 | DST_TYPE on value range in SRC_VR and store it to DST_VR. Return true if |
1644 | the result is a range that is not VARYING nor UNDEFINED. */ |
1645 | |
1646 | static bool |
1647 | ipa_vr_operation_and_type_effects (vrange &dst_vr, |
1648 | const vrange &src_vr, |
1649 | enum tree_code operation, |
1650 | tree dst_type, tree src_type) |
1651 | { |
1652 | if (!irange::supports_p (type: dst_type) || !irange::supports_p (type: src_type)) |
1653 | return false; |
1654 | |
1655 | range_op_handler handler (operation); |
1656 | if (!handler) |
1657 | return false; |
1658 | |
1659 | Value_Range varying (dst_type); |
1660 | varying.set_varying (dst_type); |
1661 | |
1662 | return (handler.operand_check_p (dst_type, src_type, dst_type) |
1663 | && handler.fold_range (r&: dst_vr, type: dst_type, lh: src_vr, rh: varying) |
1664 | && !dst_vr.varying_p () |
1665 | && !dst_vr.undefined_p ()); |
1666 | } |
1667 | |
1668 | /* Same as above, but the SRC_VR argument is an IPA_VR which must |
1669 | first be extracted onto a vrange. */ |
1670 | |
1671 | static bool |
1672 | ipa_vr_operation_and_type_effects (vrange &dst_vr, |
1673 | const ipa_vr &src_vr, |
1674 | enum tree_code operation, |
1675 | tree dst_type, tree src_type) |
1676 | { |
1677 | Value_Range tmp; |
1678 | src_vr.get_vrange (tmp); |
1679 | return ipa_vr_operation_and_type_effects (dst_vr, src_vr: tmp, operation, |
1680 | dst_type, src_type); |
1681 | } |
1682 | |
1683 | /* Determine range of JFUNC given that INFO describes the caller node or |
1684 | the one it is inlined to, CS is the call graph edge corresponding to JFUNC |
1685 | and PARM_TYPE of the parameter. */ |
1686 | |
1687 | void |
1688 | ipa_value_range_from_jfunc (vrange &vr, |
1689 | ipa_node_params *info, cgraph_edge *cs, |
1690 | ipa_jump_func *jfunc, tree parm_type) |
1691 | { |
1692 | vr.set_undefined (); |
1693 | |
1694 | if (jfunc->m_vr) |
1695 | ipa_vr_operation_and_type_effects (dst_vr&: vr, |
1696 | src_vr: *jfunc->m_vr, |
1697 | operation: NOP_EXPR, dst_type: parm_type, |
1698 | src_type: jfunc->m_vr->type ()); |
1699 | if (vr.singleton_p ()) |
1700 | return; |
1701 | if (jfunc->type == IPA_JF_PASS_THROUGH) |
1702 | { |
1703 | int idx; |
1704 | ipcp_transformation *sum |
1705 | = ipcp_get_transformation_summary (node: cs->caller->inlined_to |
1706 | ? cs->caller->inlined_to |
1707 | : cs->caller); |
1708 | if (!sum || !sum->m_vr) |
1709 | return; |
1710 | |
1711 | idx = ipa_get_jf_pass_through_formal_id (jfunc); |
1712 | |
1713 | if (!(*sum->m_vr)[idx].known_p ()) |
1714 | return; |
1715 | tree vr_type = ipa_get_type (info, i: idx); |
1716 | Value_Range srcvr; |
1717 | (*sum->m_vr)[idx].get_vrange (srcvr); |
1718 | |
1719 | enum tree_code operation = ipa_get_jf_pass_through_operation (jfunc); |
1720 | |
1721 | if (TREE_CODE_CLASS (operation) == tcc_unary) |
1722 | { |
1723 | Value_Range res (vr_type); |
1724 | |
1725 | if (ipa_vr_operation_and_type_effects (dst_vr&: res, |
1726 | src_vr: srcvr, |
1727 | operation, dst_type: parm_type, |
1728 | src_type: vr_type)) |
1729 | vr.intersect (res); |
1730 | } |
1731 | else |
1732 | { |
1733 | Value_Range op_res (vr_type); |
1734 | Value_Range res (vr_type); |
1735 | tree op = ipa_get_jf_pass_through_operand (jfunc); |
1736 | Value_Range op_vr (vr_type); |
1737 | range_op_handler handler (operation); |
1738 | |
1739 | ipa_range_set_and_normalize (r&: op_vr, val: op); |
1740 | |
1741 | if (!handler |
1742 | || !op_res.supports_type_p (type: vr_type) |
1743 | || !handler.fold_range (r&: op_res, type: vr_type, lh: srcvr, rh: op_vr)) |
1744 | op_res.set_varying (vr_type); |
1745 | |
1746 | if (ipa_vr_operation_and_type_effects (dst_vr&: res, |
1747 | src_vr: op_res, |
1748 | operation: NOP_EXPR, dst_type: parm_type, |
1749 | src_type: vr_type)) |
1750 | vr.intersect (res); |
1751 | } |
1752 | } |
1753 | } |
1754 | |
1755 | /* Determine whether ITEM, jump function for an aggregate part, evaluates to a |
1756 | single known constant value and if so, return it. Otherwise return NULL. |
1757 | NODE and INFO describes the caller node or the one it is inlined to, and |
1758 | its related info. */ |
1759 | |
1760 | tree |
1761 | ipa_agg_value_from_jfunc (ipa_node_params *info, cgraph_node *node, |
1762 | const ipa_agg_jf_item *item) |
1763 | { |
1764 | tree value = NULL_TREE; |
1765 | int src_idx; |
1766 | |
1767 | if (item->offset < 0 |
1768 | || item->jftype == IPA_JF_UNKNOWN |
1769 | || item->offset >= (HOST_WIDE_INT) UINT_MAX * BITS_PER_UNIT) |
1770 | return NULL_TREE; |
1771 | |
1772 | if (item->jftype == IPA_JF_CONST) |
1773 | return item->value.constant; |
1774 | |
1775 | gcc_checking_assert (item->jftype == IPA_JF_PASS_THROUGH |
1776 | || item->jftype == IPA_JF_LOAD_AGG); |
1777 | |
1778 | src_idx = item->value.pass_through.formal_id; |
1779 | |
1780 | if (info->ipcp_orig_node) |
1781 | { |
1782 | if (item->jftype == IPA_JF_PASS_THROUGH) |
1783 | value = info->known_csts[src_idx]; |
1784 | else if (ipcp_transformation *ts = ipcp_get_transformation_summary (node)) |
1785 | { |
1786 | ipa_argagg_value_list avl (ts); |
1787 | value = avl.get_value (index: src_idx, |
1788 | unit_offset: item->value.load_agg.offset / BITS_PER_UNIT, |
1789 | by_ref: item->value.load_agg.by_ref); |
1790 | } |
1791 | } |
1792 | else if (!info->lattices.is_empty ()) |
1793 | { |
1794 | class ipcp_param_lattices *src_plats |
1795 | = ipa_get_parm_lattices (info, i: src_idx); |
1796 | |
1797 | if (item->jftype == IPA_JF_PASS_THROUGH) |
1798 | { |
1799 | struct ipcp_lattice<tree> *lat = &src_plats->itself; |
1800 | |
1801 | if (!lat->is_single_const ()) |
1802 | return NULL_TREE; |
1803 | |
1804 | value = lat->values->value; |
1805 | } |
1806 | else if (src_plats->aggs |
1807 | && !src_plats->aggs_bottom |
1808 | && !src_plats->aggs_contain_variable |
1809 | && src_plats->aggs_by_ref == item->value.load_agg.by_ref) |
1810 | { |
1811 | struct ipcp_agg_lattice *aglat; |
1812 | |
1813 | for (aglat = src_plats->aggs; aglat; aglat = aglat->next) |
1814 | { |
1815 | if (aglat->offset > item->value.load_agg.offset) |
1816 | break; |
1817 | |
1818 | if (aglat->offset == item->value.load_agg.offset) |
1819 | { |
1820 | if (aglat->is_single_const ()) |
1821 | value = aglat->values->value; |
1822 | break; |
1823 | } |
1824 | } |
1825 | } |
1826 | } |
1827 | |
1828 | if (!value) |
1829 | return NULL_TREE; |
1830 | |
1831 | if (item->jftype == IPA_JF_LOAD_AGG) |
1832 | { |
1833 | tree load_type = item->value.load_agg.type; |
1834 | tree value_type = TREE_TYPE (value); |
1835 | |
1836 | /* Ensure value type is compatible with load type. */ |
1837 | if (!useless_type_conversion_p (load_type, value_type)) |
1838 | return NULL_TREE; |
1839 | } |
1840 | |
1841 | return ipa_get_jf_arith_result (opcode: item->value.pass_through.operation, |
1842 | input: value, |
1843 | operand: item->value.pass_through.operand, |
1844 | res_type: item->type); |
1845 | } |
1846 | |
1847 | /* Process all items in AGG_JFUNC relative to caller (or the node the original |
1848 | caller is inlined to) NODE which described by INFO and push the results to |
1849 | RES as describing values passed in parameter DST_INDEX. */ |
1850 | |
1851 | void |
1852 | ipa_push_agg_values_from_jfunc (ipa_node_params *info, cgraph_node *node, |
1853 | ipa_agg_jump_function *agg_jfunc, |
1854 | unsigned dst_index, |
1855 | vec<ipa_argagg_value> *res) |
1856 | { |
1857 | unsigned prev_unit_offset = 0; |
1858 | bool first = true; |
1859 | |
1860 | for (const ipa_agg_jf_item &item : agg_jfunc->items) |
1861 | { |
1862 | tree value = ipa_agg_value_from_jfunc (info, node, item: &item); |
1863 | if (!value) |
1864 | continue; |
1865 | |
1866 | ipa_argagg_value iav; |
1867 | iav.value = value; |
1868 | iav.unit_offset = item.offset / BITS_PER_UNIT; |
1869 | iav.index = dst_index; |
1870 | iav.by_ref = agg_jfunc->by_ref; |
1871 | iav.killed = 0; |
1872 | |
1873 | gcc_assert (first |
1874 | || iav.unit_offset > prev_unit_offset); |
1875 | prev_unit_offset = iav.unit_offset; |
1876 | first = false; |
1877 | |
1878 | res->safe_push (obj: iav); |
1879 | } |
1880 | } |
1881 | |
1882 | /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not |
1883 | bottom, not containing a variable component and without any known value at |
1884 | the same time. */ |
1885 | |
1886 | DEBUG_FUNCTION void |
1887 | ipcp_verify_propagated_values (void) |
1888 | { |
1889 | struct cgraph_node *node; |
1890 | |
1891 | FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node) |
1892 | { |
1893 | ipa_node_params *info = ipa_node_params_sum->get (node); |
1894 | if (!opt_for_fn (node->decl, flag_ipa_cp) |
1895 | || !opt_for_fn (node->decl, optimize)) |
1896 | continue; |
1897 | int i, count = ipa_get_param_count (info); |
1898 | |
1899 | for (i = 0; i < count; i++) |
1900 | { |
1901 | ipcp_lattice<tree> *lat = ipa_get_scalar_lat (info, i); |
1902 | |
1903 | if (!lat->bottom |
1904 | && !lat->contains_variable |
1905 | && lat->values_count == 0) |
1906 | { |
1907 | if (dump_file) |
1908 | { |
1909 | symtab->dump (f: dump_file); |
1910 | fprintf (stream: dump_file, format: "\nIPA lattices after constant " |
1911 | "propagation, before gcc_unreachable:\n" ); |
1912 | print_all_lattices (f: dump_file, dump_sources: true, dump_benefits: false); |
1913 | } |
1914 | |
1915 | gcc_unreachable (); |
1916 | } |
1917 | } |
1918 | } |
1919 | } |
1920 | |
1921 | /* Return true iff X and Y should be considered equal contexts by IPA-CP. */ |
1922 | |
1923 | static bool |
1924 | values_equal_for_ipcp_p (ipa_polymorphic_call_context x, |
1925 | ipa_polymorphic_call_context y) |
1926 | { |
1927 | return x.equal_to (x: y); |
1928 | } |
1929 | |
1930 | |
1931 | /* Add a new value source to the value represented by THIS, marking that a |
1932 | value comes from edge CS and (if the underlying jump function is a |
1933 | pass-through or an ancestor one) from a caller value SRC_VAL of a caller |
1934 | parameter described by SRC_INDEX. OFFSET is negative if the source was the |
1935 | scalar value of the parameter itself or the offset within an aggregate. */ |
1936 | |
1937 | template <typename valtype> |
1938 | void |
1939 | ipcp_value<valtype>::add_source (cgraph_edge *cs, ipcp_value *src_val, |
1940 | int src_idx, HOST_WIDE_INT offset) |
1941 | { |
1942 | ipcp_value_source<valtype> *src; |
1943 | |
1944 | src = new (ipcp_sources_pool.allocate ()) ipcp_value_source<valtype>; |
1945 | src->offset = offset; |
1946 | src->cs = cs; |
1947 | src->val = src_val; |
1948 | src->index = src_idx; |
1949 | |
1950 | src->next = sources; |
1951 | sources = src; |
1952 | } |
1953 | |
1954 | /* Allocate a new ipcp_value holding a tree constant, initialize its value to |
1955 | SOURCE and clear all other fields. */ |
1956 | |
1957 | static ipcp_value<tree> * |
1958 | allocate_and_init_ipcp_value (tree cst, unsigned same_lat_gen_level) |
1959 | { |
1960 | ipcp_value<tree> *val; |
1961 | |
1962 | val = new (ipcp_cst_values_pool.allocate ()) ipcp_value<tree>(); |
1963 | val->value = cst; |
1964 | val->self_recursion_generated_level = same_lat_gen_level; |
1965 | return val; |
1966 | } |
1967 | |
1968 | /* Allocate a new ipcp_value holding a polymorphic context, initialize its |
1969 | value to SOURCE and clear all other fields. */ |
1970 | |
1971 | static ipcp_value<ipa_polymorphic_call_context> * |
1972 | allocate_and_init_ipcp_value (ipa_polymorphic_call_context ctx, |
1973 | unsigned same_lat_gen_level) |
1974 | { |
1975 | ipcp_value<ipa_polymorphic_call_context> *val; |
1976 | |
1977 | val = new (ipcp_poly_ctx_values_pool.allocate ()) |
1978 | ipcp_value<ipa_polymorphic_call_context>(); |
1979 | val->value = ctx; |
1980 | val->self_recursion_generated_level = same_lat_gen_level; |
1981 | return val; |
1982 | } |
1983 | |
1984 | /* Try to add NEWVAL to LAT, potentially creating a new ipcp_value for it. CS, |
1985 | SRC_VAL SRC_INDEX and OFFSET are meant for add_source and have the same |
1986 | meaning. OFFSET -1 means the source is scalar and not a part of an |
1987 | aggregate. If non-NULL, VAL_P records address of existing or newly added |
1988 | ipcp_value. |
1989 | |
1990 | If the value is generated for a self-recursive call as a result of an |
1991 | arithmetic pass-through jump-function acting on a value in the same lattice, |
1992 | SAME_LAT_GEN_LEVEL must be the length of such chain, otherwise it must be |
1993 | zero. If it is non-zero, PARAM_IPA_CP_VALUE_LIST_SIZE limit is ignored. */ |
1994 | |
1995 | template <typename valtype> |
1996 | bool |
1997 | ipcp_lattice<valtype>::add_value (valtype newval, cgraph_edge *cs, |
1998 | ipcp_value<valtype> *src_val, |
1999 | int src_idx, HOST_WIDE_INT offset, |
2000 | ipcp_value<valtype> **val_p, |
2001 | unsigned same_lat_gen_level) |
2002 | { |
2003 | ipcp_value<valtype> *val, *last_val = NULL; |
2004 | |
2005 | if (val_p) |
2006 | *val_p = NULL; |
2007 | |
2008 | if (bottom) |
2009 | return false; |
2010 | |
2011 | for (val = values; val; last_val = val, val = val->next) |
2012 | if (values_equal_for_ipcp_p (val->value, newval)) |
2013 | { |
2014 | if (val_p) |
2015 | *val_p = val; |
2016 | |
2017 | if (val->self_recursion_generated_level < same_lat_gen_level) |
2018 | val->self_recursion_generated_level = same_lat_gen_level; |
2019 | |
2020 | if (ipa_edge_within_scc (cs)) |
2021 | { |
2022 | ipcp_value_source<valtype> *s; |
2023 | for (s = val->sources; s; s = s->next) |
2024 | if (s->cs == cs && s->val == src_val) |
2025 | break; |
2026 | if (s) |
2027 | return false; |
2028 | } |
2029 | |
2030 | val->add_source (cs, src_val, src_idx, offset); |
2031 | return false; |
2032 | } |
2033 | |
2034 | if (!same_lat_gen_level && values_count >= opt_for_fn (cs->callee->decl, |
2035 | param_ipa_cp_value_list_size)) |
2036 | { |
2037 | /* We can only free sources, not the values themselves, because sources |
2038 | of other values in this SCC might point to them. */ |
2039 | for (val = values; val; val = val->next) |
2040 | { |
2041 | while (val->sources) |
2042 | { |
2043 | ipcp_value_source<valtype> *src = val->sources; |
2044 | val->sources = src->next; |
2045 | ipcp_sources_pool.remove (object: (ipcp_value_source<tree>*)src); |
2046 | } |
2047 | } |
2048 | values = NULL; |
2049 | return set_to_bottom (); |
2050 | } |
2051 | |
2052 | values_count++; |
2053 | val = allocate_and_init_ipcp_value (newval, same_lat_gen_level); |
2054 | val->add_source (cs, src_val, src_idx, offset); |
2055 | val->next = NULL; |
2056 | |
2057 | /* Add the new value to end of value list, which can reduce iterations |
2058 | of propagation stage for recursive function. */ |
2059 | if (last_val) |
2060 | last_val->next = val; |
2061 | else |
2062 | values = val; |
2063 | |
2064 | if (val_p) |
2065 | *val_p = val; |
2066 | |
2067 | return true; |
2068 | } |
2069 | |
2070 | /* A helper function that returns result of operation specified by OPCODE on |
2071 | the value of SRC_VAL. If non-NULL, OPND1_TYPE is expected type for the |
2072 | value of SRC_VAL. If the operation is binary, OPND2 is a constant value |
2073 | acting as its second operand. If non-NULL, RES_TYPE is expected type of |
2074 | the result. */ |
2075 | |
2076 | static tree |
2077 | get_val_across_arith_op (enum tree_code opcode, |
2078 | tree opnd1_type, |
2079 | tree opnd2, |
2080 | ipcp_value<tree> *src_val, |
2081 | tree res_type) |
2082 | { |
2083 | tree opnd1 = src_val->value; |
2084 | |
2085 | /* Skip source values that is incompatible with specified type. */ |
2086 | if (opnd1_type |
2087 | && !useless_type_conversion_p (opnd1_type, TREE_TYPE (opnd1))) |
2088 | return NULL_TREE; |
2089 | |
2090 | return ipa_get_jf_arith_result (opcode, input: opnd1, operand: opnd2, res_type); |
2091 | } |
2092 | |
2093 | /* Propagate values through an arithmetic transformation described by a jump |
2094 | function associated with edge CS, taking values from SRC_LAT and putting |
2095 | them into DEST_LAT. OPND1_TYPE is expected type for the values in SRC_LAT. |
2096 | OPND2 is a constant value if transformation is a binary operation. |
2097 | SRC_OFFSET specifies offset in an aggregate if SRC_LAT describes lattice of |
2098 | a part of the aggregate. SRC_IDX is the index of the source parameter. |
2099 | RES_TYPE is the value type of result being propagated into. Return true if |
2100 | DEST_LAT changed. */ |
2101 | |
2102 | static bool |
2103 | propagate_vals_across_arith_jfunc (cgraph_edge *cs, |
2104 | enum tree_code opcode, |
2105 | tree opnd1_type, |
2106 | tree opnd2, |
2107 | ipcp_lattice<tree> *src_lat, |
2108 | ipcp_lattice<tree> *dest_lat, |
2109 | HOST_WIDE_INT src_offset, |
2110 | int src_idx, |
2111 | tree res_type) |
2112 | { |
2113 | ipcp_value<tree> *src_val; |
2114 | bool ret = false; |
2115 | |
2116 | /* Due to circular dependencies, propagating within an SCC through arithmetic |
2117 | transformation would create infinite number of values. But for |
2118 | self-feeding recursive function, we could allow propagation in a limited |
2119 | count, and this can enable a simple kind of recursive function versioning. |
2120 | For other scenario, we would just make lattices bottom. */ |
2121 | if (opcode != NOP_EXPR && ipa_edge_within_scc (cs)) |
2122 | { |
2123 | int i; |
2124 | |
2125 | int max_recursive_depth = opt_for_fn(cs->caller->decl, |
2126 | param_ipa_cp_max_recursive_depth); |
2127 | if (src_lat != dest_lat || max_recursive_depth < 1) |
2128 | return dest_lat->set_contains_variable (); |
2129 | |
2130 | /* No benefit if recursive execution is in low probability. */ |
2131 | if (cs->sreal_frequency () * 100 |
2132 | <= ((sreal) 1) * opt_for_fn (cs->caller->decl, |
2133 | param_ipa_cp_min_recursive_probability)) |
2134 | return dest_lat->set_contains_variable (); |
2135 | |
2136 | auto_vec<ipcp_value<tree> *, 8> val_seeds; |
2137 | |
2138 | for (src_val = src_lat->values; src_val; src_val = src_val->next) |
2139 | { |
2140 | /* Now we do not use self-recursively generated value as propagation |
2141 | source, this is absolutely conservative, but could avoid explosion |
2142 | of lattice's value space, especially when one recursive function |
2143 | calls another recursive. */ |
2144 | if (src_val->self_recursion_generated_p ()) |
2145 | { |
2146 | ipcp_value_source<tree> *s; |
2147 | |
2148 | /* If the lattice has already been propagated for the call site, |
2149 | no need to do that again. */ |
2150 | for (s = src_val->sources; s; s = s->next) |
2151 | if (s->cs == cs) |
2152 | return dest_lat->set_contains_variable (); |
2153 | } |
2154 | else |
2155 | val_seeds.safe_push (obj: src_val); |
2156 | } |
2157 | |
2158 | gcc_assert ((int) val_seeds.length () <= param_ipa_cp_value_list_size); |
2159 | |
2160 | /* Recursively generate lattice values with a limited count. */ |
2161 | FOR_EACH_VEC_ELT (val_seeds, i, src_val) |
2162 | { |
2163 | for (int j = 1; j < max_recursive_depth; j++) |
2164 | { |
2165 | tree cstval = get_val_across_arith_op (opcode, opnd1_type, opnd2, |
2166 | src_val, res_type); |
2167 | if (!cstval |
2168 | || !ipacp_value_safe_for_type (param_type: res_type, value: cstval)) |
2169 | break; |
2170 | |
2171 | ret |= dest_lat->add_value (newval: cstval, cs, src_val, src_idx, |
2172 | offset: src_offset, val_p: &src_val, same_lat_gen_level: j); |
2173 | gcc_checking_assert (src_val); |
2174 | } |
2175 | } |
2176 | ret |= dest_lat->set_contains_variable (); |
2177 | } |
2178 | else |
2179 | for (src_val = src_lat->values; src_val; src_val = src_val->next) |
2180 | { |
2181 | /* Now we do not use self-recursively generated value as propagation |
2182 | source, otherwise it is easy to make value space of normal lattice |
2183 | overflow. */ |
2184 | if (src_val->self_recursion_generated_p ()) |
2185 | { |
2186 | ret |= dest_lat->set_contains_variable (); |
2187 | continue; |
2188 | } |
2189 | |
2190 | tree cstval = get_val_across_arith_op (opcode, opnd1_type, opnd2, |
2191 | src_val, res_type); |
2192 | if (cstval |
2193 | && ipacp_value_safe_for_type (param_type: res_type, value: cstval)) |
2194 | ret |= dest_lat->add_value (newval: cstval, cs, src_val, src_idx, |
2195 | offset: src_offset); |
2196 | else |
2197 | ret |= dest_lat->set_contains_variable (); |
2198 | } |
2199 | |
2200 | return ret; |
2201 | } |
2202 | |
2203 | /* Propagate values through a pass-through jump function JFUNC associated with |
2204 | edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX |
2205 | is the index of the source parameter. PARM_TYPE is the type of the |
2206 | parameter to which the result is passed. */ |
2207 | |
2208 | static bool |
2209 | propagate_vals_across_pass_through (cgraph_edge *cs, ipa_jump_func *jfunc, |
2210 | ipcp_lattice<tree> *src_lat, |
2211 | ipcp_lattice<tree> *dest_lat, int src_idx, |
2212 | tree parm_type) |
2213 | { |
2214 | return propagate_vals_across_arith_jfunc (cs, |
2215 | opcode: ipa_get_jf_pass_through_operation (jfunc), |
2216 | NULL_TREE, |
2217 | opnd2: ipa_get_jf_pass_through_operand (jfunc), |
2218 | src_lat, dest_lat, src_offset: -1, src_idx, res_type: parm_type); |
2219 | } |
2220 | |
2221 | /* Propagate values through an ancestor jump function JFUNC associated with |
2222 | edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX |
2223 | is the index of the source parameter. */ |
2224 | |
2225 | static bool |
2226 | propagate_vals_across_ancestor (struct cgraph_edge *cs, |
2227 | struct ipa_jump_func *jfunc, |
2228 | ipcp_lattice<tree> *src_lat, |
2229 | ipcp_lattice<tree> *dest_lat, int src_idx, |
2230 | tree param_type) |
2231 | { |
2232 | ipcp_value<tree> *src_val; |
2233 | bool ret = false; |
2234 | |
2235 | if (ipa_edge_within_scc (cs)) |
2236 | return dest_lat->set_contains_variable (); |
2237 | |
2238 | for (src_val = src_lat->values; src_val; src_val = src_val->next) |
2239 | { |
2240 | tree t = ipa_get_jf_ancestor_result (jfunc, input: src_val->value); |
2241 | |
2242 | if (t && ipacp_value_safe_for_type (param_type, value: t)) |
2243 | ret |= dest_lat->add_value (newval: t, cs, src_val, src_idx); |
2244 | else |
2245 | ret |= dest_lat->set_contains_variable (); |
2246 | } |
2247 | |
2248 | return ret; |
2249 | } |
2250 | |
2251 | /* Propagate scalar values across jump function JFUNC that is associated with |
2252 | edge CS and put the values into DEST_LAT. PARM_TYPE is the type of the |
2253 | parameter to which the result is passed. */ |
2254 | |
2255 | static bool |
2256 | propagate_scalar_across_jump_function (struct cgraph_edge *cs, |
2257 | struct ipa_jump_func *jfunc, |
2258 | ipcp_lattice<tree> *dest_lat, |
2259 | tree param_type) |
2260 | { |
2261 | if (dest_lat->bottom) |
2262 | return false; |
2263 | |
2264 | if (jfunc->type == IPA_JF_CONST) |
2265 | { |
2266 | tree val = ipa_get_jf_constant (jfunc); |
2267 | if (ipacp_value_safe_for_type (param_type, value: val)) |
2268 | return dest_lat->add_value (newval: val, cs, NULL, src_idx: 0); |
2269 | else |
2270 | return dest_lat->set_contains_variable (); |
2271 | } |
2272 | else if (jfunc->type == IPA_JF_PASS_THROUGH |
2273 | || jfunc->type == IPA_JF_ANCESTOR) |
2274 | { |
2275 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: cs->caller); |
2276 | ipcp_lattice<tree> *src_lat; |
2277 | int src_idx; |
2278 | bool ret; |
2279 | |
2280 | if (jfunc->type == IPA_JF_PASS_THROUGH) |
2281 | src_idx = ipa_get_jf_pass_through_formal_id (jfunc); |
2282 | else |
2283 | src_idx = ipa_get_jf_ancestor_formal_id (jfunc); |
2284 | |
2285 | src_lat = ipa_get_scalar_lat (info: caller_info, i: src_idx); |
2286 | if (src_lat->bottom) |
2287 | return dest_lat->set_contains_variable (); |
2288 | |
2289 | /* If we would need to clone the caller and cannot, do not propagate. */ |
2290 | if (!ipcp_versionable_function_p (node: cs->caller) |
2291 | && (src_lat->contains_variable |
2292 | || (src_lat->values_count > 1))) |
2293 | return dest_lat->set_contains_variable (); |
2294 | |
2295 | if (jfunc->type == IPA_JF_PASS_THROUGH) |
2296 | ret = propagate_vals_across_pass_through (cs, jfunc, src_lat, |
2297 | dest_lat, src_idx, |
2298 | parm_type: param_type); |
2299 | else |
2300 | ret = propagate_vals_across_ancestor (cs, jfunc, src_lat, dest_lat, |
2301 | src_idx, param_type); |
2302 | |
2303 | if (src_lat->contains_variable) |
2304 | ret |= dest_lat->set_contains_variable (); |
2305 | |
2306 | return ret; |
2307 | } |
2308 | |
2309 | /* TODO: We currently do not handle member method pointers in IPA-CP (we only |
2310 | use it for indirect inlining), we should propagate them too. */ |
2311 | return dest_lat->set_contains_variable (); |
2312 | } |
2313 | |
2314 | /* Propagate scalar values across jump function JFUNC that is associated with |
2315 | edge CS and describes argument IDX and put the values into DEST_LAT. */ |
2316 | |
2317 | static bool |
2318 | propagate_context_across_jump_function (cgraph_edge *cs, |
2319 | ipa_jump_func *jfunc, int idx, |
2320 | ipcp_lattice<ipa_polymorphic_call_context> *dest_lat) |
2321 | { |
2322 | if (dest_lat->bottom) |
2323 | return false; |
2324 | ipa_edge_args *args = ipa_edge_args_sum->get (edge: cs); |
2325 | bool ret = false; |
2326 | bool added_sth = false; |
2327 | bool type_preserved = true; |
2328 | |
2329 | ipa_polymorphic_call_context edge_ctx, *edge_ctx_ptr |
2330 | = ipa_get_ith_polymorhic_call_context (args, i: idx); |
2331 | |
2332 | if (edge_ctx_ptr) |
2333 | edge_ctx = *edge_ctx_ptr; |
2334 | |
2335 | if (jfunc->type == IPA_JF_PASS_THROUGH |
2336 | || jfunc->type == IPA_JF_ANCESTOR) |
2337 | { |
2338 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: cs->caller); |
2339 | int src_idx; |
2340 | ipcp_lattice<ipa_polymorphic_call_context> *src_lat; |
2341 | |
2342 | /* TODO: Once we figure out how to propagate speculations, it will |
2343 | probably be a good idea to switch to speculation if type_preserved is |
2344 | not set instead of punting. */ |
2345 | if (jfunc->type == IPA_JF_PASS_THROUGH) |
2346 | { |
2347 | if (ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR) |
2348 | goto prop_fail; |
2349 | type_preserved = ipa_get_jf_pass_through_type_preserved (jfunc); |
2350 | src_idx = ipa_get_jf_pass_through_formal_id (jfunc); |
2351 | } |
2352 | else |
2353 | { |
2354 | type_preserved = ipa_get_jf_ancestor_type_preserved (jfunc); |
2355 | src_idx = ipa_get_jf_ancestor_formal_id (jfunc); |
2356 | } |
2357 | |
2358 | src_lat = ipa_get_poly_ctx_lat (info: caller_info, i: src_idx); |
2359 | /* If we would need to clone the caller and cannot, do not propagate. */ |
2360 | if (!ipcp_versionable_function_p (node: cs->caller) |
2361 | && (src_lat->contains_variable |
2362 | || (src_lat->values_count > 1))) |
2363 | goto prop_fail; |
2364 | |
2365 | ipcp_value<ipa_polymorphic_call_context> *src_val; |
2366 | for (src_val = src_lat->values; src_val; src_val = src_val->next) |
2367 | { |
2368 | ipa_polymorphic_call_context cur = src_val->value; |
2369 | |
2370 | if (!type_preserved) |
2371 | cur.possible_dynamic_type_change (cs->in_polymorphic_cdtor); |
2372 | if (jfunc->type == IPA_JF_ANCESTOR) |
2373 | cur.offset_by (off: ipa_get_jf_ancestor_offset (jfunc)); |
2374 | /* TODO: In cases we know how the context is going to be used, |
2375 | we can improve the result by passing proper OTR_TYPE. */ |
2376 | cur.combine_with (edge_ctx); |
2377 | if (!cur.useless_p ()) |
2378 | { |
2379 | if (src_lat->contains_variable |
2380 | && !edge_ctx.equal_to (x: cur)) |
2381 | ret |= dest_lat->set_contains_variable (); |
2382 | ret |= dest_lat->add_value (newval: cur, cs, src_val, src_idx); |
2383 | added_sth = true; |
2384 | } |
2385 | } |
2386 | } |
2387 | |
2388 | prop_fail: |
2389 | if (!added_sth) |
2390 | { |
2391 | if (!edge_ctx.useless_p ()) |
2392 | ret |= dest_lat->add_value (newval: edge_ctx, cs); |
2393 | else |
2394 | ret |= dest_lat->set_contains_variable (); |
2395 | } |
2396 | |
2397 | return ret; |
2398 | } |
2399 | |
2400 | /* Propagate bits across jfunc that is associated with |
2401 | edge cs and update dest_lattice accordingly. */ |
2402 | |
2403 | bool |
2404 | propagate_bits_across_jump_function (cgraph_edge *cs, int idx, |
2405 | ipa_jump_func *jfunc, |
2406 | ipcp_bits_lattice *dest_lattice) |
2407 | { |
2408 | if (dest_lattice->bottom_p ()) |
2409 | return false; |
2410 | |
2411 | enum availability availability; |
2412 | cgraph_node *callee = cs->callee->function_symbol (avail: &availability); |
2413 | ipa_node_params *callee_info = ipa_node_params_sum->get (node: callee); |
2414 | tree parm_type = ipa_get_type (info: callee_info, i: idx); |
2415 | |
2416 | /* For K&R C programs, ipa_get_type() could return NULL_TREE. Avoid the |
2417 | transform for these cases. Similarly, we can have bad type mismatches |
2418 | with LTO, avoid doing anything with those too. */ |
2419 | if (!parm_type |
2420 | || (!INTEGRAL_TYPE_P (parm_type) && !POINTER_TYPE_P (parm_type))) |
2421 | { |
2422 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2423 | fprintf (stream: dump_file, format: "Setting dest_lattice to bottom, because type of " |
2424 | "param %i of %s is NULL or unsuitable for bits propagation\n" , |
2425 | idx, cs->callee->dump_name ()); |
2426 | |
2427 | return dest_lattice->set_to_bottom (); |
2428 | } |
2429 | |
2430 | unsigned precision = TYPE_PRECISION (parm_type); |
2431 | signop sgn = TYPE_SIGN (parm_type); |
2432 | |
2433 | if (jfunc->type == IPA_JF_PASS_THROUGH |
2434 | || jfunc->type == IPA_JF_ANCESTOR) |
2435 | { |
2436 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: cs->caller); |
2437 | tree operand = NULL_TREE; |
2438 | enum tree_code code; |
2439 | unsigned src_idx; |
2440 | bool keep_null = false; |
2441 | |
2442 | if (jfunc->type == IPA_JF_PASS_THROUGH) |
2443 | { |
2444 | code = ipa_get_jf_pass_through_operation (jfunc); |
2445 | src_idx = ipa_get_jf_pass_through_formal_id (jfunc); |
2446 | if (code != NOP_EXPR) |
2447 | operand = ipa_get_jf_pass_through_operand (jfunc); |
2448 | } |
2449 | else |
2450 | { |
2451 | code = POINTER_PLUS_EXPR; |
2452 | src_idx = ipa_get_jf_ancestor_formal_id (jfunc); |
2453 | unsigned HOST_WIDE_INT offset |
2454 | = ipa_get_jf_ancestor_offset (jfunc) / BITS_PER_UNIT; |
2455 | keep_null = (ipa_get_jf_ancestor_keep_null (jfunc) || !offset); |
2456 | operand = build_int_cstu (size_type_node, offset); |
2457 | } |
2458 | |
2459 | class ipcp_param_lattices *src_lats |
2460 | = ipa_get_parm_lattices (info: caller_info, i: src_idx); |
2461 | |
2462 | /* Try to propagate bits if src_lattice is bottom, but jfunc is known. |
2463 | for eg consider: |
2464 | int f(int x) |
2465 | { |
2466 | g (x & 0xff); |
2467 | } |
2468 | Assume lattice for x is bottom, however we can still propagate |
2469 | result of x & 0xff == 0xff, which gets computed during ccp1 pass |
2470 | and we store it in jump function during analysis stage. */ |
2471 | |
2472 | if (!src_lats->bits_lattice.bottom_p ()) |
2473 | { |
2474 | bool drop_all_ones |
2475 | = keep_null && !src_lats->bits_lattice.known_nonzero_p (); |
2476 | |
2477 | return dest_lattice->meet_with (other&: src_lats->bits_lattice, precision, |
2478 | sgn, code, operand, drop_all_ones); |
2479 | } |
2480 | } |
2481 | |
2482 | Value_Range vr (parm_type); |
2483 | if (jfunc->m_vr) |
2484 | { |
2485 | jfunc->m_vr->get_vrange (vr); |
2486 | if (!vr.undefined_p () && !vr.varying_p ()) |
2487 | { |
2488 | irange &r = as_a <irange> (v&: vr); |
2489 | irange_bitmask bm = r.get_bitmask (); |
2490 | widest_int mask |
2491 | = widest_int::from (x: bm.mask (), TYPE_SIGN (parm_type)); |
2492 | widest_int value |
2493 | = widest_int::from (x: bm.value (), TYPE_SIGN (parm_type)); |
2494 | return dest_lattice->meet_with (value, mask, precision); |
2495 | } |
2496 | } |
2497 | return dest_lattice->set_to_bottom (); |
2498 | } |
2499 | |
2500 | /* Propagate value range across jump function JFUNC that is associated with |
2501 | edge CS with param of callee of PARAM_TYPE and update DEST_PLATS |
2502 | accordingly. */ |
2503 | |
2504 | static bool |
2505 | propagate_vr_across_jump_function (cgraph_edge *cs, ipa_jump_func *jfunc, |
2506 | class ipcp_param_lattices *dest_plats, |
2507 | tree param_type) |
2508 | { |
2509 | ipcp_vr_lattice *dest_lat = &dest_plats->m_value_range; |
2510 | |
2511 | if (dest_lat->bottom_p ()) |
2512 | return false; |
2513 | |
2514 | if (!param_type |
2515 | || (!INTEGRAL_TYPE_P (param_type) |
2516 | && !POINTER_TYPE_P (param_type))) |
2517 | return dest_lat->set_to_bottom (); |
2518 | |
2519 | if (jfunc->type == IPA_JF_PASS_THROUGH) |
2520 | { |
2521 | enum tree_code operation = ipa_get_jf_pass_through_operation (jfunc); |
2522 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: cs->caller); |
2523 | int src_idx = ipa_get_jf_pass_through_formal_id (jfunc); |
2524 | class ipcp_param_lattices *src_lats |
2525 | = ipa_get_parm_lattices (info: caller_info, i: src_idx); |
2526 | tree operand_type = ipa_get_type (info: caller_info, i: src_idx); |
2527 | |
2528 | if (src_lats->m_value_range.bottom_p ()) |
2529 | return dest_lat->set_to_bottom (); |
2530 | |
2531 | Value_Range vr (operand_type); |
2532 | if (TREE_CODE_CLASS (operation) == tcc_unary) |
2533 | ipa_vr_operation_and_type_effects (dst_vr&: vr, |
2534 | src_vr: src_lats->m_value_range.m_vr, |
2535 | operation, dst_type: param_type, |
2536 | src_type: operand_type); |
2537 | /* A crude way to prevent unbounded number of value range updates |
2538 | in SCC components. We should allow limited number of updates within |
2539 | SCC, too. */ |
2540 | else if (!ipa_edge_within_scc (cs)) |
2541 | { |
2542 | tree op = ipa_get_jf_pass_through_operand (jfunc); |
2543 | Value_Range op_vr (TREE_TYPE (op)); |
2544 | Value_Range op_res (operand_type); |
2545 | range_op_handler handler (operation); |
2546 | |
2547 | ipa_range_set_and_normalize (r&: op_vr, val: op); |
2548 | |
2549 | if (!handler |
2550 | || !op_res.supports_type_p (type: operand_type) |
2551 | || !handler.fold_range (r&: op_res, type: operand_type, |
2552 | lh: src_lats->m_value_range.m_vr, rh: op_vr)) |
2553 | op_res.set_varying (operand_type); |
2554 | |
2555 | ipa_vr_operation_and_type_effects (dst_vr&: vr, |
2556 | src_vr: op_res, |
2557 | operation: NOP_EXPR, dst_type: param_type, |
2558 | src_type: operand_type); |
2559 | } |
2560 | if (!vr.undefined_p () && !vr.varying_p ()) |
2561 | { |
2562 | if (jfunc->m_vr) |
2563 | { |
2564 | Value_Range jvr (param_type); |
2565 | if (ipa_vr_operation_and_type_effects (dst_vr&: jvr, src_vr: *jfunc->m_vr, |
2566 | operation: NOP_EXPR, |
2567 | dst_type: param_type, |
2568 | src_type: jfunc->m_vr->type ())) |
2569 | vr.intersect (r: jvr); |
2570 | } |
2571 | return dest_lat->meet_with (p_vr: vr); |
2572 | } |
2573 | } |
2574 | else if (jfunc->type == IPA_JF_CONST) |
2575 | { |
2576 | tree val = ipa_get_jf_constant (jfunc); |
2577 | if (TREE_CODE (val) == INTEGER_CST) |
2578 | { |
2579 | val = fold_convert (param_type, val); |
2580 | if (TREE_OVERFLOW_P (val)) |
2581 | val = drop_tree_overflow (val); |
2582 | |
2583 | Value_Range tmpvr (val, val); |
2584 | return dest_lat->meet_with (p_vr: tmpvr); |
2585 | } |
2586 | } |
2587 | |
2588 | Value_Range vr (param_type); |
2589 | if (jfunc->m_vr |
2590 | && ipa_vr_operation_and_type_effects (dst_vr&: vr, src_vr: *jfunc->m_vr, operation: NOP_EXPR, |
2591 | dst_type: param_type, |
2592 | src_type: jfunc->m_vr->type ())) |
2593 | return dest_lat->meet_with (p_vr: vr); |
2594 | else |
2595 | return dest_lat->set_to_bottom (); |
2596 | } |
2597 | |
2598 | /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches |
2599 | NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all |
2600 | other cases, return false). If there are no aggregate items, set |
2601 | aggs_by_ref to NEW_AGGS_BY_REF. */ |
2602 | |
2603 | static bool |
2604 | set_check_aggs_by_ref (class ipcp_param_lattices *dest_plats, |
2605 | bool new_aggs_by_ref) |
2606 | { |
2607 | if (dest_plats->aggs) |
2608 | { |
2609 | if (dest_plats->aggs_by_ref != new_aggs_by_ref) |
2610 | { |
2611 | set_agg_lats_to_bottom (dest_plats); |
2612 | return true; |
2613 | } |
2614 | } |
2615 | else |
2616 | dest_plats->aggs_by_ref = new_aggs_by_ref; |
2617 | return false; |
2618 | } |
2619 | |
2620 | /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an |
2621 | already existing lattice for the given OFFSET and SIZE, marking all skipped |
2622 | lattices as containing variable and checking for overlaps. If there is no |
2623 | already existing lattice for the OFFSET and VAL_SIZE, create one, initialize |
2624 | it with offset, size and contains_variable to PRE_EXISTING, and return true, |
2625 | unless there are too many already. If there are two many, return false. If |
2626 | there are overlaps turn whole DEST_PLATS to bottom and return false. If any |
2627 | skipped lattices were newly marked as containing variable, set *CHANGE to |
2628 | true. MAX_AGG_ITEMS is the maximum number of lattices. */ |
2629 | |
2630 | static bool |
2631 | merge_agg_lats_step (class ipcp_param_lattices *dest_plats, |
2632 | HOST_WIDE_INT offset, HOST_WIDE_INT val_size, |
2633 | struct ipcp_agg_lattice ***aglat, |
2634 | bool pre_existing, bool *change, int max_agg_items) |
2635 | { |
2636 | gcc_checking_assert (offset >= 0); |
2637 | |
2638 | while (**aglat && (**aglat)->offset < offset) |
2639 | { |
2640 | if ((**aglat)->offset + (**aglat)->size > offset) |
2641 | { |
2642 | set_agg_lats_to_bottom (dest_plats); |
2643 | return false; |
2644 | } |
2645 | *change |= (**aglat)->set_contains_variable (); |
2646 | *aglat = &(**aglat)->next; |
2647 | } |
2648 | |
2649 | if (**aglat && (**aglat)->offset == offset) |
2650 | { |
2651 | if ((**aglat)->size != val_size) |
2652 | { |
2653 | set_agg_lats_to_bottom (dest_plats); |
2654 | return false; |
2655 | } |
2656 | gcc_assert (!(**aglat)->next |
2657 | || (**aglat)->next->offset >= offset + val_size); |
2658 | return true; |
2659 | } |
2660 | else |
2661 | { |
2662 | struct ipcp_agg_lattice *new_al; |
2663 | |
2664 | if (**aglat && (**aglat)->offset < offset + val_size) |
2665 | { |
2666 | set_agg_lats_to_bottom (dest_plats); |
2667 | return false; |
2668 | } |
2669 | if (dest_plats->aggs_count == max_agg_items) |
2670 | return false; |
2671 | dest_plats->aggs_count++; |
2672 | new_al = ipcp_agg_lattice_pool.allocate (); |
2673 | |
2674 | new_al->offset = offset; |
2675 | new_al->size = val_size; |
2676 | new_al->contains_variable = pre_existing; |
2677 | |
2678 | new_al->next = **aglat; |
2679 | **aglat = new_al; |
2680 | return true; |
2681 | } |
2682 | } |
2683 | |
2684 | /* Set all AGLAT and all other aggregate lattices reachable by next pointers as |
2685 | containing an unknown value. */ |
2686 | |
2687 | static bool |
2688 | set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice *aglat) |
2689 | { |
2690 | bool ret = false; |
2691 | while (aglat) |
2692 | { |
2693 | ret |= aglat->set_contains_variable (); |
2694 | aglat = aglat->next; |
2695 | } |
2696 | return ret; |
2697 | } |
2698 | |
2699 | /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting |
2700 | DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source |
2701 | parameter used for lattice value sources. Return true if DEST_PLATS changed |
2702 | in any way. */ |
2703 | |
2704 | static bool |
2705 | merge_aggregate_lattices (struct cgraph_edge *cs, |
2706 | class ipcp_param_lattices *dest_plats, |
2707 | class ipcp_param_lattices *src_plats, |
2708 | int src_idx, HOST_WIDE_INT offset_delta) |
2709 | { |
2710 | bool pre_existing = dest_plats->aggs != NULL; |
2711 | struct ipcp_agg_lattice **dst_aglat; |
2712 | bool ret = false; |
2713 | |
2714 | if (set_check_aggs_by_ref (dest_plats, new_aggs_by_ref: src_plats->aggs_by_ref)) |
2715 | return true; |
2716 | if (src_plats->aggs_bottom) |
2717 | return set_agg_lats_contain_variable (dest_plats); |
2718 | if (src_plats->aggs_contain_variable) |
2719 | ret |= set_agg_lats_contain_variable (dest_plats); |
2720 | dst_aglat = &dest_plats->aggs; |
2721 | |
2722 | int max_agg_items = opt_for_fn (cs->callee->function_symbol ()->decl, |
2723 | param_ipa_max_agg_items); |
2724 | for (struct ipcp_agg_lattice *src_aglat = src_plats->aggs; |
2725 | src_aglat; |
2726 | src_aglat = src_aglat->next) |
2727 | { |
2728 | HOST_WIDE_INT new_offset = src_aglat->offset - offset_delta; |
2729 | |
2730 | if (new_offset < 0) |
2731 | continue; |
2732 | if (merge_agg_lats_step (dest_plats, offset: new_offset, val_size: src_aglat->size, |
2733 | aglat: &dst_aglat, pre_existing, change: &ret, max_agg_items)) |
2734 | { |
2735 | struct ipcp_agg_lattice *new_al = *dst_aglat; |
2736 | |
2737 | dst_aglat = &(*dst_aglat)->next; |
2738 | if (src_aglat->bottom) |
2739 | { |
2740 | ret |= new_al->set_contains_variable (); |
2741 | continue; |
2742 | } |
2743 | if (src_aglat->contains_variable) |
2744 | ret |= new_al->set_contains_variable (); |
2745 | for (ipcp_value<tree> *val = src_aglat->values; |
2746 | val; |
2747 | val = val->next) |
2748 | ret |= new_al->add_value (newval: val->value, cs, src_val: val, src_idx, |
2749 | offset: src_aglat->offset); |
2750 | } |
2751 | else if (dest_plats->aggs_bottom) |
2752 | return true; |
2753 | } |
2754 | ret |= set_chain_of_aglats_contains_variable (*dst_aglat); |
2755 | return ret; |
2756 | } |
2757 | |
2758 | /* Determine whether there is anything to propagate FROM SRC_PLATS through a |
2759 | pass-through JFUNC and if so, whether it has conform and conforms to the |
2760 | rules about propagating values passed by reference. */ |
2761 | |
2762 | static bool |
2763 | agg_pass_through_permissible_p (class ipcp_param_lattices *src_plats, |
2764 | struct ipa_jump_func *jfunc) |
2765 | { |
2766 | return src_plats->aggs |
2767 | && (!src_plats->aggs_by_ref |
2768 | || ipa_get_jf_pass_through_agg_preserved (jfunc)); |
2769 | } |
2770 | |
2771 | /* Propagate values through ITEM, jump function for a part of an aggregate, |
2772 | into corresponding aggregate lattice AGLAT. CS is the call graph edge |
2773 | associated with the jump function. Return true if AGLAT changed in any |
2774 | way. */ |
2775 | |
2776 | static bool |
2777 | propagate_aggregate_lattice (struct cgraph_edge *cs, |
2778 | struct ipa_agg_jf_item *item, |
2779 | struct ipcp_agg_lattice *aglat) |
2780 | { |
2781 | class ipa_node_params *caller_info; |
2782 | class ipcp_param_lattices *src_plats; |
2783 | struct ipcp_lattice<tree> *src_lat; |
2784 | HOST_WIDE_INT src_offset; |
2785 | int src_idx; |
2786 | tree load_type; |
2787 | bool ret; |
2788 | |
2789 | if (item->jftype == IPA_JF_CONST) |
2790 | { |
2791 | tree value = item->value.constant; |
2792 | |
2793 | gcc_checking_assert (is_gimple_ip_invariant (value)); |
2794 | return aglat->add_value (newval: value, cs, NULL, src_idx: 0); |
2795 | } |
2796 | |
2797 | gcc_checking_assert (item->jftype == IPA_JF_PASS_THROUGH |
2798 | || item->jftype == IPA_JF_LOAD_AGG); |
2799 | |
2800 | caller_info = ipa_node_params_sum->get (node: cs->caller); |
2801 | src_idx = item->value.pass_through.formal_id; |
2802 | src_plats = ipa_get_parm_lattices (info: caller_info, i: src_idx); |
2803 | |
2804 | if (item->jftype == IPA_JF_PASS_THROUGH) |
2805 | { |
2806 | load_type = NULL_TREE; |
2807 | src_lat = &src_plats->itself; |
2808 | src_offset = -1; |
2809 | } |
2810 | else |
2811 | { |
2812 | HOST_WIDE_INT load_offset = item->value.load_agg.offset; |
2813 | struct ipcp_agg_lattice *src_aglat; |
2814 | |
2815 | for (src_aglat = src_plats->aggs; src_aglat; src_aglat = src_aglat->next) |
2816 | if (src_aglat->offset >= load_offset) |
2817 | break; |
2818 | |
2819 | load_type = item->value.load_agg.type; |
2820 | if (!src_aglat |
2821 | || src_aglat->offset > load_offset |
2822 | || src_aglat->size != tree_to_shwi (TYPE_SIZE (load_type)) |
2823 | || src_plats->aggs_by_ref != item->value.load_agg.by_ref) |
2824 | return aglat->set_contains_variable (); |
2825 | |
2826 | src_lat = src_aglat; |
2827 | src_offset = load_offset; |
2828 | } |
2829 | |
2830 | if (src_lat->bottom |
2831 | || (!ipcp_versionable_function_p (node: cs->caller) |
2832 | && !src_lat->is_single_const ())) |
2833 | return aglat->set_contains_variable (); |
2834 | |
2835 | ret = propagate_vals_across_arith_jfunc (cs, |
2836 | opcode: item->value.pass_through.operation, |
2837 | opnd1_type: load_type, |
2838 | opnd2: item->value.pass_through.operand, |
2839 | src_lat, dest_lat: aglat, |
2840 | src_offset, |
2841 | src_idx, |
2842 | res_type: item->type); |
2843 | |
2844 | if (src_lat->contains_variable) |
2845 | ret |= aglat->set_contains_variable (); |
2846 | |
2847 | return ret; |
2848 | } |
2849 | |
2850 | /* Propagate scalar values across jump function JFUNC that is associated with |
2851 | edge CS and put the values into DEST_LAT. */ |
2852 | |
2853 | static bool |
2854 | propagate_aggs_across_jump_function (struct cgraph_edge *cs, |
2855 | struct ipa_jump_func *jfunc, |
2856 | class ipcp_param_lattices *dest_plats) |
2857 | { |
2858 | bool ret = false; |
2859 | |
2860 | if (dest_plats->aggs_bottom) |
2861 | return false; |
2862 | |
2863 | if (jfunc->type == IPA_JF_PASS_THROUGH |
2864 | && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR) |
2865 | { |
2866 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: cs->caller); |
2867 | int src_idx = ipa_get_jf_pass_through_formal_id (jfunc); |
2868 | class ipcp_param_lattices *src_plats; |
2869 | |
2870 | src_plats = ipa_get_parm_lattices (info: caller_info, i: src_idx); |
2871 | if (agg_pass_through_permissible_p (src_plats, jfunc)) |
2872 | { |
2873 | /* Currently we do not produce clobber aggregate jump |
2874 | functions, replace with merging when we do. */ |
2875 | gcc_assert (!jfunc->agg.items); |
2876 | ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, |
2877 | src_idx, offset_delta: 0); |
2878 | return ret; |
2879 | } |
2880 | } |
2881 | else if (jfunc->type == IPA_JF_ANCESTOR |
2882 | && ipa_get_jf_ancestor_agg_preserved (jfunc)) |
2883 | { |
2884 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: cs->caller); |
2885 | int src_idx = ipa_get_jf_ancestor_formal_id (jfunc); |
2886 | class ipcp_param_lattices *src_plats; |
2887 | |
2888 | src_plats = ipa_get_parm_lattices (info: caller_info, i: src_idx); |
2889 | if (src_plats->aggs && src_plats->aggs_by_ref) |
2890 | { |
2891 | /* Currently we do not produce clobber aggregate jump |
2892 | functions, replace with merging when we do. */ |
2893 | gcc_assert (!jfunc->agg.items); |
2894 | ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, src_idx, |
2895 | offset_delta: ipa_get_jf_ancestor_offset (jfunc)); |
2896 | } |
2897 | else if (!src_plats->aggs_by_ref) |
2898 | ret |= set_agg_lats_to_bottom (dest_plats); |
2899 | else |
2900 | ret |= set_agg_lats_contain_variable (dest_plats); |
2901 | return ret; |
2902 | } |
2903 | |
2904 | if (jfunc->agg.items) |
2905 | { |
2906 | bool pre_existing = dest_plats->aggs != NULL; |
2907 | struct ipcp_agg_lattice **aglat = &dest_plats->aggs; |
2908 | struct ipa_agg_jf_item *item; |
2909 | int i; |
2910 | |
2911 | if (set_check_aggs_by_ref (dest_plats, new_aggs_by_ref: jfunc->agg.by_ref)) |
2912 | return true; |
2913 | |
2914 | int max_agg_items = opt_for_fn (cs->callee->function_symbol ()->decl, |
2915 | param_ipa_max_agg_items); |
2916 | FOR_EACH_VEC_ELT (*jfunc->agg.items, i, item) |
2917 | { |
2918 | HOST_WIDE_INT val_size; |
2919 | |
2920 | if (item->offset < 0 || item->jftype == IPA_JF_UNKNOWN) |
2921 | continue; |
2922 | val_size = tree_to_shwi (TYPE_SIZE (item->type)); |
2923 | |
2924 | if (merge_agg_lats_step (dest_plats, offset: item->offset, val_size, |
2925 | aglat: &aglat, pre_existing, change: &ret, max_agg_items)) |
2926 | { |
2927 | ret |= propagate_aggregate_lattice (cs, item, aglat: *aglat); |
2928 | aglat = &(*aglat)->next; |
2929 | } |
2930 | else if (dest_plats->aggs_bottom) |
2931 | return true; |
2932 | } |
2933 | |
2934 | ret |= set_chain_of_aglats_contains_variable (*aglat); |
2935 | } |
2936 | else |
2937 | ret |= set_agg_lats_contain_variable (dest_plats); |
2938 | |
2939 | return ret; |
2940 | } |
2941 | |
2942 | /* Return true if on the way cfrom CS->caller to the final (non-alias and |
2943 | non-thunk) destination, the call passes through a thunk. */ |
2944 | |
2945 | static bool |
2946 | call_passes_through_thunk (cgraph_edge *cs) |
2947 | { |
2948 | cgraph_node *alias_or_thunk = cs->callee; |
2949 | while (alias_or_thunk->alias) |
2950 | alias_or_thunk = alias_or_thunk->get_alias_target (); |
2951 | return alias_or_thunk->thunk; |
2952 | } |
2953 | |
2954 | /* Propagate constants from the caller to the callee of CS. INFO describes the |
2955 | caller. */ |
2956 | |
2957 | static bool |
2958 | propagate_constants_across_call (struct cgraph_edge *cs) |
2959 | { |
2960 | class ipa_node_params *callee_info; |
2961 | enum availability availability; |
2962 | cgraph_node *callee; |
2963 | class ipa_edge_args *args; |
2964 | bool ret = false; |
2965 | int i, args_count, parms_count; |
2966 | |
2967 | callee = cs->callee->function_symbol (avail: &availability); |
2968 | if (!callee->definition) |
2969 | return false; |
2970 | gcc_checking_assert (callee->has_gimple_body_p ()); |
2971 | callee_info = ipa_node_params_sum->get (node: callee); |
2972 | if (!callee_info) |
2973 | return false; |
2974 | |
2975 | args = ipa_edge_args_sum->get (edge: cs); |
2976 | parms_count = ipa_get_param_count (info: callee_info); |
2977 | if (parms_count == 0) |
2978 | return false; |
2979 | if (!args |
2980 | || !opt_for_fn (cs->caller->decl, flag_ipa_cp) |
2981 | || !opt_for_fn (cs->caller->decl, optimize)) |
2982 | { |
2983 | for (i = 0; i < parms_count; i++) |
2984 | ret |= set_all_contains_variable (ipa_get_parm_lattices (info: callee_info, |
2985 | i)); |
2986 | return ret; |
2987 | } |
2988 | args_count = ipa_get_cs_argument_count (args); |
2989 | |
2990 | /* If this call goes through a thunk we must not propagate to the first (0th) |
2991 | parameter. However, we might need to uncover a thunk from below a series |
2992 | of aliases first. */ |
2993 | if (call_passes_through_thunk (cs)) |
2994 | { |
2995 | ret |= set_all_contains_variable (ipa_get_parm_lattices (info: callee_info, |
2996 | i: 0)); |
2997 | i = 1; |
2998 | } |
2999 | else |
3000 | i = 0; |
3001 | |
3002 | for (; (i < args_count) && (i < parms_count); i++) |
3003 | { |
3004 | struct ipa_jump_func *jump_func = ipa_get_ith_jump_func (args, i); |
3005 | class ipcp_param_lattices *dest_plats; |
3006 | tree param_type = ipa_get_type (info: callee_info, i); |
3007 | |
3008 | dest_plats = ipa_get_parm_lattices (info: callee_info, i); |
3009 | if (availability == AVAIL_INTERPOSABLE) |
3010 | ret |= set_all_contains_variable (dest_plats); |
3011 | else |
3012 | { |
3013 | ret |= propagate_scalar_across_jump_function (cs, jfunc: jump_func, |
3014 | dest_lat: &dest_plats->itself, |
3015 | param_type); |
3016 | ret |= propagate_context_across_jump_function (cs, jfunc: jump_func, idx: i, |
3017 | dest_lat: &dest_plats->ctxlat); |
3018 | ret |
3019 | |= propagate_bits_across_jump_function (cs, idx: i, jfunc: jump_func, |
3020 | dest_lattice: &dest_plats->bits_lattice); |
3021 | ret |= propagate_aggs_across_jump_function (cs, jfunc: jump_func, |
3022 | dest_plats); |
3023 | if (opt_for_fn (callee->decl, flag_ipa_vrp)) |
3024 | ret |= propagate_vr_across_jump_function (cs, jfunc: jump_func, |
3025 | dest_plats, param_type); |
3026 | else |
3027 | ret |= dest_plats->m_value_range.set_to_bottom (); |
3028 | } |
3029 | } |
3030 | for (; i < parms_count; i++) |
3031 | ret |= set_all_contains_variable (ipa_get_parm_lattices (info: callee_info, i)); |
3032 | |
3033 | return ret; |
3034 | } |
3035 | |
3036 | /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS |
3037 | KNOWN_CONTEXTS, and known aggregates either in AVS or KNOWN_AGGS return |
3038 | the destination. The latter three can be NULL. If AGG_REPS is not NULL, |
3039 | KNOWN_AGGS is ignored. */ |
3040 | |
3041 | static tree |
3042 | ipa_get_indirect_edge_target_1 (struct cgraph_edge *ie, |
3043 | const vec<tree> &known_csts, |
3044 | const vec<ipa_polymorphic_call_context> &known_contexts, |
3045 | const ipa_argagg_value_list &avs, |
3046 | bool *speculative) |
3047 | { |
3048 | int param_index = ie->indirect_info->param_index; |
3049 | HOST_WIDE_INT anc_offset; |
3050 | tree t = NULL; |
3051 | tree target = NULL; |
3052 | |
3053 | *speculative = false; |
3054 | |
3055 | if (param_index == -1) |
3056 | return NULL_TREE; |
3057 | |
3058 | if (!ie->indirect_info->polymorphic) |
3059 | { |
3060 | tree t = NULL; |
3061 | |
3062 | if (ie->indirect_info->agg_contents) |
3063 | { |
3064 | t = NULL; |
3065 | if ((unsigned) param_index < known_csts.length () |
3066 | && known_csts[param_index]) |
3067 | t = ipa_find_agg_cst_from_init (scalar: known_csts[param_index], |
3068 | offset: ie->indirect_info->offset, |
3069 | by_ref: ie->indirect_info->by_ref); |
3070 | |
3071 | if (!t && ie->indirect_info->guaranteed_unmodified) |
3072 | t = avs.get_value (index: param_index, |
3073 | unit_offset: ie->indirect_info->offset / BITS_PER_UNIT, |
3074 | by_ref: ie->indirect_info->by_ref); |
3075 | } |
3076 | else if ((unsigned) param_index < known_csts.length ()) |
3077 | t = known_csts[param_index]; |
3078 | |
3079 | if (t |
3080 | && TREE_CODE (t) == ADDR_EXPR |
3081 | && TREE_CODE (TREE_OPERAND (t, 0)) == FUNCTION_DECL) |
3082 | return TREE_OPERAND (t, 0); |
3083 | else |
3084 | return NULL_TREE; |
3085 | } |
3086 | |
3087 | if (!opt_for_fn (ie->caller->decl, flag_devirtualize)) |
3088 | return NULL_TREE; |
3089 | |
3090 | gcc_assert (!ie->indirect_info->agg_contents); |
3091 | gcc_assert (!ie->indirect_info->by_ref); |
3092 | anc_offset = ie->indirect_info->offset; |
3093 | |
3094 | t = NULL; |
3095 | |
3096 | if ((unsigned) param_index < known_csts.length () |
3097 | && known_csts[param_index]) |
3098 | t = ipa_find_agg_cst_from_init (scalar: known_csts[param_index], |
3099 | offset: ie->indirect_info->offset, by_ref: true); |
3100 | |
3101 | /* Try to work out value of virtual table pointer value in replacements. */ |
3102 | /* or known aggregate values. */ |
3103 | if (!t) |
3104 | t = avs.get_value (index: param_index, |
3105 | unit_offset: ie->indirect_info->offset / BITS_PER_UNIT, |
3106 | by_ref: true); |
3107 | |
3108 | /* If we found the virtual table pointer, lookup the target. */ |
3109 | if (t) |
3110 | { |
3111 | tree vtable; |
3112 | unsigned HOST_WIDE_INT offset; |
3113 | if (vtable_pointer_value_to_vtable (t, &vtable, &offset)) |
3114 | { |
3115 | bool can_refer; |
3116 | target = gimple_get_virt_method_for_vtable (ie->indirect_info->otr_token, |
3117 | vtable, offset, can_refer: &can_refer); |
3118 | if (can_refer) |
3119 | { |
3120 | if (!target |
3121 | || fndecl_built_in_p (node: target, name1: BUILT_IN_UNREACHABLE) |
3122 | || !possible_polymorphic_call_target_p |
3123 | (e: ie, n: cgraph_node::get (decl: target))) |
3124 | { |
3125 | /* Do not speculate builtin_unreachable, it is stupid! */ |
3126 | if (ie->indirect_info->vptr_changed) |
3127 | return NULL; |
3128 | target = ipa_impossible_devirt_target (ie, target); |
3129 | } |
3130 | *speculative = ie->indirect_info->vptr_changed; |
3131 | if (!*speculative) |
3132 | return target; |
3133 | } |
3134 | } |
3135 | } |
3136 | |
3137 | /* Do we know the constant value of pointer? */ |
3138 | if (!t && (unsigned) param_index < known_csts.length ()) |
3139 | t = known_csts[param_index]; |
3140 | |
3141 | gcc_checking_assert (!t || TREE_CODE (t) != TREE_BINFO); |
3142 | |
3143 | ipa_polymorphic_call_context context; |
3144 | if (known_contexts.length () > (unsigned int) param_index) |
3145 | { |
3146 | context = known_contexts[param_index]; |
3147 | context.offset_by (off: anc_offset); |
3148 | if (ie->indirect_info->vptr_changed) |
3149 | context.possible_dynamic_type_change (ie->in_polymorphic_cdtor, |
3150 | otr_type: ie->indirect_info->otr_type); |
3151 | if (t) |
3152 | { |
3153 | ipa_polymorphic_call_context ctx2 = ipa_polymorphic_call_context |
3154 | (t, ie->indirect_info->otr_type, anc_offset); |
3155 | if (!ctx2.useless_p ()) |
3156 | context.combine_with (ctx2, otr_type: ie->indirect_info->otr_type); |
3157 | } |
3158 | } |
3159 | else if (t) |
3160 | { |
3161 | context = ipa_polymorphic_call_context (t, ie->indirect_info->otr_type, |
3162 | anc_offset); |
3163 | if (ie->indirect_info->vptr_changed) |
3164 | context.possible_dynamic_type_change (ie->in_polymorphic_cdtor, |
3165 | otr_type: ie->indirect_info->otr_type); |
3166 | } |
3167 | else |
3168 | return NULL_TREE; |
3169 | |
3170 | vec <cgraph_node *>targets; |
3171 | bool final; |
3172 | |
3173 | targets = possible_polymorphic_call_targets |
3174 | (ie->indirect_info->otr_type, |
3175 | ie->indirect_info->otr_token, |
3176 | context, copletep: &final); |
3177 | if (!final || targets.length () > 1) |
3178 | { |
3179 | struct cgraph_node *node; |
3180 | if (*speculative) |
3181 | return target; |
3182 | if (!opt_for_fn (ie->caller->decl, flag_devirtualize_speculatively) |
3183 | || ie->speculative || !ie->maybe_hot_p ()) |
3184 | return NULL; |
3185 | node = try_speculative_devirtualization (ie->indirect_info->otr_type, |
3186 | ie->indirect_info->otr_token, |
3187 | context); |
3188 | if (node) |
3189 | { |
3190 | *speculative = true; |
3191 | target = node->decl; |
3192 | } |
3193 | else |
3194 | return NULL; |
3195 | } |
3196 | else |
3197 | { |
3198 | *speculative = false; |
3199 | if (targets.length () == 1) |
3200 | target = targets[0]->decl; |
3201 | else |
3202 | target = ipa_impossible_devirt_target (ie, NULL_TREE); |
3203 | } |
3204 | |
3205 | if (target && !possible_polymorphic_call_target_p (e: ie, |
3206 | n: cgraph_node::get (decl: target))) |
3207 | { |
3208 | if (*speculative) |
3209 | return NULL; |
3210 | target = ipa_impossible_devirt_target (ie, target); |
3211 | } |
3212 | |
3213 | return target; |
3214 | } |
3215 | |
3216 | /* If an indirect edge IE can be turned into a direct one based on data in |
3217 | AVALS, return the destination. Store into *SPECULATIVE a boolean determinig |
3218 | whether the discovered target is only speculative guess. */ |
3219 | |
3220 | tree |
3221 | ipa_get_indirect_edge_target (struct cgraph_edge *ie, |
3222 | ipa_call_arg_values *avals, |
3223 | bool *speculative) |
3224 | { |
3225 | ipa_argagg_value_list avl (avals); |
3226 | return ipa_get_indirect_edge_target_1 (ie, known_csts: avals->m_known_vals, |
3227 | known_contexts: avals->m_known_contexts, |
3228 | avs: avl, speculative); |
3229 | } |
3230 | |
3231 | /* Calculate devirtualization time bonus for NODE, assuming we know information |
3232 | about arguments stored in AVALS. */ |
3233 | |
3234 | static int |
3235 | devirtualization_time_bonus (struct cgraph_node *node, |
3236 | ipa_auto_call_arg_values *avals) |
3237 | { |
3238 | struct cgraph_edge *ie; |
3239 | int res = 0; |
3240 | |
3241 | for (ie = node->indirect_calls; ie; ie = ie->next_callee) |
3242 | { |
3243 | struct cgraph_node *callee; |
3244 | class ipa_fn_summary *isummary; |
3245 | enum availability avail; |
3246 | tree target; |
3247 | bool speculative; |
3248 | |
3249 | ipa_argagg_value_list avl (avals); |
3250 | target = ipa_get_indirect_edge_target_1 (ie, known_csts: avals->m_known_vals, |
3251 | known_contexts: avals->m_known_contexts, |
3252 | avs: avl, speculative: &speculative); |
3253 | if (!target) |
3254 | continue; |
3255 | |
3256 | /* Only bare minimum benefit for clearly un-inlineable targets. */ |
3257 | res += 1; |
3258 | callee = cgraph_node::get (decl: target); |
3259 | if (!callee || !callee->definition) |
3260 | continue; |
3261 | callee = callee->function_symbol (avail: &avail); |
3262 | if (avail < AVAIL_AVAILABLE) |
3263 | continue; |
3264 | isummary = ipa_fn_summaries->get (node: callee); |
3265 | if (!isummary || !isummary->inlinable) |
3266 | continue; |
3267 | |
3268 | int size = ipa_size_summaries->get (node: callee)->size; |
3269 | /* FIXME: The values below need re-considering and perhaps also |
3270 | integrating into the cost metrics, at lest in some very basic way. */ |
3271 | int max_inline_insns_auto |
3272 | = opt_for_fn (callee->decl, param_max_inline_insns_auto); |
3273 | if (size <= max_inline_insns_auto / 4) |
3274 | res += 31 / ((int)speculative + 1); |
3275 | else if (size <= max_inline_insns_auto / 2) |
3276 | res += 15 / ((int)speculative + 1); |
3277 | else if (size <= max_inline_insns_auto |
3278 | || DECL_DECLARED_INLINE_P (callee->decl)) |
3279 | res += 7 / ((int)speculative + 1); |
3280 | } |
3281 | |
3282 | return res; |
3283 | } |
3284 | |
3285 | /* Return time bonus incurred because of hints stored in ESTIMATES. */ |
3286 | |
3287 | static int |
3288 | hint_time_bonus (cgraph_node *node, const ipa_call_estimates &estimates) |
3289 | { |
3290 | int result = 0; |
3291 | ipa_hints hints = estimates.hints; |
3292 | if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride)) |
3293 | result += opt_for_fn (node->decl, param_ipa_cp_loop_hint_bonus); |
3294 | |
3295 | sreal bonus_for_one = opt_for_fn (node->decl, param_ipa_cp_loop_hint_bonus); |
3296 | |
3297 | if (hints & INLINE_HINT_loop_iterations) |
3298 | result += (estimates.loops_with_known_iterations * bonus_for_one).to_int (); |
3299 | |
3300 | if (hints & INLINE_HINT_loop_stride) |
3301 | result += (estimates.loops_with_known_strides * bonus_for_one).to_int (); |
3302 | |
3303 | return result; |
3304 | } |
3305 | |
3306 | /* If there is a reason to penalize the function described by INFO in the |
3307 | cloning goodness evaluation, do so. */ |
3308 | |
3309 | static inline sreal |
3310 | incorporate_penalties (cgraph_node *node, ipa_node_params *info, |
3311 | sreal evaluation) |
3312 | { |
3313 | if (info->node_within_scc && !info->node_is_self_scc) |
3314 | evaluation = (evaluation |
3315 | * (100 - opt_for_fn (node->decl, |
3316 | param_ipa_cp_recursion_penalty))) / 100; |
3317 | |
3318 | if (info->node_calling_single_call) |
3319 | evaluation = (evaluation |
3320 | * (100 - opt_for_fn (node->decl, |
3321 | param_ipa_cp_single_call_penalty))) |
3322 | / 100; |
3323 | |
3324 | return evaluation; |
3325 | } |
3326 | |
3327 | /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT |
3328 | and SIZE_COST and with the sum of frequencies of incoming edges to the |
3329 | potential new clone in FREQUENCIES. */ |
3330 | |
3331 | static bool |
3332 | good_cloning_opportunity_p (struct cgraph_node *node, sreal time_benefit, |
3333 | sreal freq_sum, profile_count count_sum, |
3334 | int size_cost) |
3335 | { |
3336 | if (time_benefit == 0 |
3337 | || !opt_for_fn (node->decl, flag_ipa_cp_clone) |
3338 | || node->optimize_for_size_p ()) |
3339 | return false; |
3340 | |
3341 | gcc_assert (size_cost > 0); |
3342 | |
3343 | ipa_node_params *info = ipa_node_params_sum->get (node); |
3344 | int eval_threshold = opt_for_fn (node->decl, param_ipa_cp_eval_threshold); |
3345 | if (count_sum.nonzero_p ()) |
3346 | { |
3347 | gcc_assert (base_count.nonzero_p ()); |
3348 | sreal factor = count_sum.probability_in (overall: base_count).to_sreal (); |
3349 | sreal evaluation = (time_benefit * factor) / size_cost; |
3350 | evaluation = incorporate_penalties (node, info, evaluation); |
3351 | evaluation *= 1000; |
3352 | |
3353 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3354 | { |
3355 | fprintf (stream: dump_file, format: " good_cloning_opportunity_p (time: %g, " |
3356 | "size: %i, count_sum: " , time_benefit.to_double (), |
3357 | size_cost); |
3358 | count_sum.dump (f: dump_file); |
3359 | fprintf (stream: dump_file, format: "%s%s) -> evaluation: %.2f, threshold: %i\n" , |
3360 | info->node_within_scc |
3361 | ? (info->node_is_self_scc ? ", self_scc" : ", scc" ) : "" , |
3362 | info->node_calling_single_call ? ", single_call" : "" , |
3363 | evaluation.to_double (), eval_threshold); |
3364 | } |
3365 | |
3366 | return evaluation.to_int () >= eval_threshold; |
3367 | } |
3368 | else |
3369 | { |
3370 | sreal evaluation = (time_benefit * freq_sum) / size_cost; |
3371 | evaluation = incorporate_penalties (node, info, evaluation); |
3372 | evaluation *= 1000; |
3373 | |
3374 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3375 | fprintf (stream: dump_file, format: " good_cloning_opportunity_p (time: %g, " |
3376 | "size: %i, freq_sum: %g%s%s) -> evaluation: %.2f, " |
3377 | "threshold: %i\n" , |
3378 | time_benefit.to_double (), size_cost, freq_sum.to_double (), |
3379 | info->node_within_scc |
3380 | ? (info->node_is_self_scc ? ", self_scc" : ", scc" ) : "" , |
3381 | info->node_calling_single_call ? ", single_call" : "" , |
3382 | evaluation.to_double (), eval_threshold); |
3383 | |
3384 | return evaluation.to_int () >= eval_threshold; |
3385 | } |
3386 | } |
3387 | |
3388 | /* Grow vectors in AVALS and fill them with information about values of |
3389 | parameters that are known to be independent of the context. Only calculate |
3390 | m_known_aggs if CALCULATE_AGGS is true. INFO describes the function. If |
3391 | REMOVABLE_PARAMS_COST is non-NULL, the movement cost of all removable |
3392 | parameters will be stored in it. |
3393 | |
3394 | TODO: Also grow context independent value range vectors. */ |
3395 | |
3396 | static bool |
3397 | gather_context_independent_values (class ipa_node_params *info, |
3398 | ipa_auto_call_arg_values *avals, |
3399 | bool calculate_aggs, |
3400 | int *removable_params_cost) |
3401 | { |
3402 | int i, count = ipa_get_param_count (info); |
3403 | bool ret = false; |
3404 | |
3405 | avals->m_known_vals.safe_grow_cleared (len: count, exact: true); |
3406 | avals->m_known_contexts.safe_grow_cleared (len: count, exact: true); |
3407 | |
3408 | if (removable_params_cost) |
3409 | *removable_params_cost = 0; |
3410 | |
3411 | for (i = 0; i < count; i++) |
3412 | { |
3413 | class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i); |
3414 | ipcp_lattice<tree> *lat = &plats->itself; |
3415 | |
3416 | if (lat->is_single_const ()) |
3417 | { |
3418 | ipcp_value<tree> *val = lat->values; |
3419 | gcc_checking_assert (TREE_CODE (val->value) != TREE_BINFO); |
3420 | avals->m_known_vals[i] = val->value; |
3421 | if (removable_params_cost) |
3422 | *removable_params_cost |
3423 | += estimate_move_cost (TREE_TYPE (val->value), false); |
3424 | ret = true; |
3425 | } |
3426 | else if (removable_params_cost |
3427 | && !ipa_is_param_used (info, i)) |
3428 | *removable_params_cost |
3429 | += ipa_get_param_move_cost (info, i); |
3430 | |
3431 | if (!ipa_is_param_used (info, i)) |
3432 | continue; |
3433 | |
3434 | ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat; |
3435 | /* Do not account known context as reason for cloning. We can see |
3436 | if it permits devirtualization. */ |
3437 | if (ctxlat->is_single_const ()) |
3438 | avals->m_known_contexts[i] = ctxlat->values->value; |
3439 | |
3440 | if (calculate_aggs) |
3441 | ret |= push_agg_values_from_plats (plats, dest_index: i, unit_delta: 0, res: &avals->m_known_aggs); |
3442 | } |
3443 | |
3444 | return ret; |
3445 | } |
3446 | |
3447 | /* Perform time and size measurement of NODE with the context given in AVALS, |
3448 | calculate the benefit compared to the node without specialization and store |
3449 | it into VAL. Take into account REMOVABLE_PARAMS_COST of all |
3450 | context-independent or unused removable parameters and EST_MOVE_COST, the |
3451 | estimated movement of the considered parameter. */ |
3452 | |
3453 | static void |
3454 | perform_estimation_of_a_value (cgraph_node *node, |
3455 | ipa_auto_call_arg_values *avals, |
3456 | int removable_params_cost, int est_move_cost, |
3457 | ipcp_value_base *val) |
3458 | { |
3459 | sreal time_benefit; |
3460 | ipa_call_estimates estimates; |
3461 | |
3462 | estimate_ipcp_clone_size_and_time (node, avals, estimates: &estimates); |
3463 | |
3464 | /* Extern inline functions have no cloning local time benefits because they |
3465 | will be inlined anyway. The only reason to clone them is if it enables |
3466 | optimization in any of the functions they call. */ |
3467 | if (DECL_EXTERNAL (node->decl) && DECL_DECLARED_INLINE_P (node->decl)) |
3468 | time_benefit = 0; |
3469 | else |
3470 | time_benefit = (estimates.nonspecialized_time - estimates.time) |
3471 | + (devirtualization_time_bonus (node, avals) |
3472 | + hint_time_bonus (node, estimates) |
3473 | + removable_params_cost + est_move_cost); |
3474 | |
3475 | int size = estimates.size; |
3476 | gcc_checking_assert (size >=0); |
3477 | /* The inliner-heuristics based estimates may think that in certain |
3478 | contexts some functions do not have any size at all but we want |
3479 | all specializations to have at least a tiny cost, not least not to |
3480 | divide by zero. */ |
3481 | if (size == 0) |
3482 | size = 1; |
3483 | |
3484 | val->local_time_benefit = time_benefit; |
3485 | val->local_size_cost = size; |
3486 | } |
3487 | |
3488 | /* Get the overall limit oof growth based on parameters extracted from growth. |
3489 | it does not really make sense to mix functions with different overall growth |
3490 | limits but it is possible and if it happens, we do not want to select one |
3491 | limit at random. */ |
3492 | |
3493 | static long |
3494 | get_max_overall_size (cgraph_node *node) |
3495 | { |
3496 | long max_new_size = orig_overall_size; |
3497 | long large_unit = opt_for_fn (node->decl, param_ipa_cp_large_unit_insns); |
3498 | if (max_new_size < large_unit) |
3499 | max_new_size = large_unit; |
3500 | int unit_growth = opt_for_fn (node->decl, param_ipa_cp_unit_growth); |
3501 | max_new_size += max_new_size * unit_growth / 100 + 1; |
3502 | return max_new_size; |
3503 | } |
3504 | |
3505 | /* Return true if NODE should be cloned just for a parameter removal, possibly |
3506 | dumping a reason if not. */ |
3507 | |
3508 | static bool |
3509 | clone_for_param_removal_p (cgraph_node *node) |
3510 | { |
3511 | if (!node->can_change_signature) |
3512 | { |
3513 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3514 | fprintf (stream: dump_file, format: " Not considering cloning to remove parameters, " |
3515 | "function cannot change signature.\n" ); |
3516 | return false; |
3517 | } |
3518 | if (node->can_be_local_p ()) |
3519 | { |
3520 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3521 | fprintf (stream: dump_file, format: " Not considering cloning to remove parameters, " |
3522 | "IPA-SRA can do it potentially better.\n" ); |
3523 | return false; |
3524 | } |
3525 | return true; |
3526 | } |
3527 | |
3528 | /* Iterate over known values of parameters of NODE and estimate the local |
3529 | effects in terms of time and size they have. */ |
3530 | |
3531 | static void |
3532 | estimate_local_effects (struct cgraph_node *node) |
3533 | { |
3534 | ipa_node_params *info = ipa_node_params_sum->get (node); |
3535 | int count = ipa_get_param_count (info); |
3536 | bool always_const; |
3537 | int removable_params_cost; |
3538 | |
3539 | if (!count || !ipcp_versionable_function_p (node)) |
3540 | return; |
3541 | |
3542 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3543 | fprintf (stream: dump_file, format: "\nEstimating effects for %s.\n" , node->dump_name ()); |
3544 | |
3545 | ipa_auto_call_arg_values avals; |
3546 | always_const = gather_context_independent_values (info, avals: &avals, calculate_aggs: true, |
3547 | removable_params_cost: &removable_params_cost); |
3548 | int devirt_bonus = devirtualization_time_bonus (node, avals: &avals); |
3549 | if (always_const || devirt_bonus |
3550 | || (removable_params_cost && clone_for_param_removal_p (node))) |
3551 | { |
3552 | struct caller_statistics stats; |
3553 | ipa_call_estimates estimates; |
3554 | |
3555 | init_caller_stats (stats: &stats); |
3556 | node->call_for_symbol_thunks_and_aliases (callback: gather_caller_stats, data: &stats, |
3557 | include_overwritable: false); |
3558 | estimate_ipcp_clone_size_and_time (node, avals: &avals, estimates: &estimates); |
3559 | sreal time = estimates.nonspecialized_time - estimates.time; |
3560 | time += devirt_bonus; |
3561 | time += hint_time_bonus (node, estimates); |
3562 | time += removable_params_cost; |
3563 | int size = estimates.size - stats.n_calls * removable_params_cost; |
3564 | |
3565 | if (dump_file) |
3566 | fprintf (stream: dump_file, format: " - context independent values, size: %i, " |
3567 | "time_benefit: %f\n" , size, (time).to_double ()); |
3568 | |
3569 | if (size <= 0 || node->local) |
3570 | { |
3571 | info->do_clone_for_all_contexts = true; |
3572 | |
3573 | if (dump_file) |
3574 | fprintf (stream: dump_file, format: " Decided to specialize for all " |
3575 | "known contexts, code not going to grow.\n" ); |
3576 | } |
3577 | else if (good_cloning_opportunity_p (node, time_benefit: time, freq_sum: stats.freq_sum, |
3578 | count_sum: stats.count_sum, size_cost: size)) |
3579 | { |
3580 | if (size + overall_size <= get_max_overall_size (node)) |
3581 | { |
3582 | info->do_clone_for_all_contexts = true; |
3583 | overall_size += size; |
3584 | |
3585 | if (dump_file) |
3586 | fprintf (stream: dump_file, format: " Decided to specialize for all " |
3587 | "known contexts, growth (to %li) deemed " |
3588 | "beneficial.\n" , overall_size); |
3589 | } |
3590 | else if (dump_file && (dump_flags & TDF_DETAILS)) |
3591 | fprintf (stream: dump_file, format: " Not cloning for all contexts because " |
3592 | "maximum unit size would be reached with %li.\n" , |
3593 | size + overall_size); |
3594 | } |
3595 | else if (dump_file && (dump_flags & TDF_DETAILS)) |
3596 | fprintf (stream: dump_file, format: " Not cloning for all contexts because " |
3597 | "!good_cloning_opportunity_p.\n" ); |
3598 | |
3599 | } |
3600 | |
3601 | for (int i = 0; i < count; i++) |
3602 | { |
3603 | class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i); |
3604 | ipcp_lattice<tree> *lat = &plats->itself; |
3605 | ipcp_value<tree> *val; |
3606 | |
3607 | if (lat->bottom |
3608 | || !lat->values |
3609 | || avals.m_known_vals[i]) |
3610 | continue; |
3611 | |
3612 | for (val = lat->values; val; val = val->next) |
3613 | { |
3614 | gcc_checking_assert (TREE_CODE (val->value) != TREE_BINFO); |
3615 | avals.m_known_vals[i] = val->value; |
3616 | |
3617 | int emc = estimate_move_cost (TREE_TYPE (val->value), true); |
3618 | perform_estimation_of_a_value (node, avals: &avals, removable_params_cost, |
3619 | est_move_cost: emc, val); |
3620 | |
3621 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3622 | { |
3623 | fprintf (stream: dump_file, format: " - estimates for value " ); |
3624 | print_ipcp_constant_value (f: dump_file, v: val->value); |
3625 | fprintf (stream: dump_file, format: " for " ); |
3626 | ipa_dump_param (dump_file, info, i); |
3627 | fprintf (stream: dump_file, format: ": time_benefit: %g, size: %i\n" , |
3628 | val->local_time_benefit.to_double (), |
3629 | val->local_size_cost); |
3630 | } |
3631 | } |
3632 | avals.m_known_vals[i] = NULL_TREE; |
3633 | } |
3634 | |
3635 | for (int i = 0; i < count; i++) |
3636 | { |
3637 | class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i); |
3638 | |
3639 | if (!plats->virt_call) |
3640 | continue; |
3641 | |
3642 | ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat; |
3643 | ipcp_value<ipa_polymorphic_call_context> *val; |
3644 | |
3645 | if (ctxlat->bottom |
3646 | || !ctxlat->values |
3647 | || !avals.m_known_contexts[i].useless_p ()) |
3648 | continue; |
3649 | |
3650 | for (val = ctxlat->values; val; val = val->next) |
3651 | { |
3652 | avals.m_known_contexts[i] = val->value; |
3653 | perform_estimation_of_a_value (node, avals: &avals, removable_params_cost, |
3654 | est_move_cost: 0, val); |
3655 | |
3656 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3657 | { |
3658 | fprintf (stream: dump_file, format: " - estimates for polymorphic context " ); |
3659 | print_ipcp_constant_value (f: dump_file, v: val->value); |
3660 | fprintf (stream: dump_file, format: " for " ); |
3661 | ipa_dump_param (dump_file, info, i); |
3662 | fprintf (stream: dump_file, format: ": time_benefit: %g, size: %i\n" , |
3663 | val->local_time_benefit.to_double (), |
3664 | val->local_size_cost); |
3665 | } |
3666 | } |
3667 | avals.m_known_contexts[i] = ipa_polymorphic_call_context (); |
3668 | } |
3669 | |
3670 | unsigned all_ctx_len = avals.m_known_aggs.length (); |
3671 | auto_vec<ipa_argagg_value, 32> all_ctx; |
3672 | all_ctx.reserve_exact (nelems: all_ctx_len); |
3673 | all_ctx.splice (src: avals.m_known_aggs); |
3674 | avals.m_known_aggs.safe_grow_cleared (len: all_ctx_len + 1); |
3675 | |
3676 | unsigned j = 0; |
3677 | for (int index = 0; index < count; index++) |
3678 | { |
3679 | class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i: index); |
3680 | |
3681 | if (plats->aggs_bottom || !plats->aggs) |
3682 | continue; |
3683 | |
3684 | for (ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next) |
3685 | { |
3686 | ipcp_value<tree> *val; |
3687 | if (aglat->bottom || !aglat->values |
3688 | /* If the following is true, the one value is already part of all |
3689 | context estimations. */ |
3690 | || (!plats->aggs_contain_variable |
3691 | && aglat->is_single_const ())) |
3692 | continue; |
3693 | |
3694 | unsigned unit_offset = aglat->offset / BITS_PER_UNIT; |
3695 | while (j < all_ctx_len |
3696 | && (all_ctx[j].index < index |
3697 | || (all_ctx[j].index == index |
3698 | && all_ctx[j].unit_offset < unit_offset))) |
3699 | { |
3700 | avals.m_known_aggs[j] = all_ctx[j]; |
3701 | j++; |
3702 | } |
3703 | |
3704 | for (unsigned k = j; k < all_ctx_len; k++) |
3705 | avals.m_known_aggs[k+1] = all_ctx[k]; |
3706 | |
3707 | for (val = aglat->values; val; val = val->next) |
3708 | { |
3709 | avals.m_known_aggs[j].value = val->value; |
3710 | avals.m_known_aggs[j].unit_offset = unit_offset; |
3711 | avals.m_known_aggs[j].index = index; |
3712 | avals.m_known_aggs[j].by_ref = plats->aggs_by_ref; |
3713 | avals.m_known_aggs[j].killed = false; |
3714 | |
3715 | perform_estimation_of_a_value (node, avals: &avals, |
3716 | removable_params_cost, est_move_cost: 0, val); |
3717 | |
3718 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3719 | { |
3720 | fprintf (stream: dump_file, format: " - estimates for value " ); |
3721 | print_ipcp_constant_value (f: dump_file, v: val->value); |
3722 | fprintf (stream: dump_file, format: " for " ); |
3723 | ipa_dump_param (dump_file, info, i: index); |
3724 | fprintf (stream: dump_file, format: "[%soffset: " HOST_WIDE_INT_PRINT_DEC |
3725 | "]: time_benefit: %g, size: %i\n" , |
3726 | plats->aggs_by_ref ? "ref " : "" , |
3727 | aglat->offset, |
3728 | val->local_time_benefit.to_double (), |
3729 | val->local_size_cost); |
3730 | } |
3731 | } |
3732 | } |
3733 | } |
3734 | } |
3735 | |
3736 | |
3737 | /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the |
3738 | topological sort of values. */ |
3739 | |
3740 | template <typename valtype> |
3741 | void |
3742 | value_topo_info<valtype>::add_val (ipcp_value<valtype> *cur_val) |
3743 | { |
3744 | ipcp_value_source<valtype> *src; |
3745 | |
3746 | if (cur_val->dfs) |
3747 | return; |
3748 | |
3749 | dfs_counter++; |
3750 | cur_val->dfs = dfs_counter; |
3751 | cur_val->low_link = dfs_counter; |
3752 | |
3753 | cur_val->topo_next = stack; |
3754 | stack = cur_val; |
3755 | cur_val->on_stack = true; |
3756 | |
3757 | for (src = cur_val->sources; src; src = src->next) |
3758 | if (src->val) |
3759 | { |
3760 | if (src->val->dfs == 0) |
3761 | { |
3762 | add_val (cur_val: src->val); |
3763 | if (src->val->low_link < cur_val->low_link) |
3764 | cur_val->low_link = src->val->low_link; |
3765 | } |
3766 | else if (src->val->on_stack |
3767 | && src->val->dfs < cur_val->low_link) |
3768 | cur_val->low_link = src->val->dfs; |
3769 | } |
3770 | |
3771 | if (cur_val->dfs == cur_val->low_link) |
3772 | { |
3773 | ipcp_value<valtype> *v, *scc_list = NULL; |
3774 | |
3775 | do |
3776 | { |
3777 | v = stack; |
3778 | stack = v->topo_next; |
3779 | v->on_stack = false; |
3780 | v->scc_no = cur_val->dfs; |
3781 | |
3782 | v->scc_next = scc_list; |
3783 | scc_list = v; |
3784 | } |
3785 | while (v != cur_val); |
3786 | |
3787 | cur_val->topo_next = values_topo; |
3788 | values_topo = cur_val; |
3789 | } |
3790 | } |
3791 | |
3792 | /* Add all values in lattices associated with NODE to the topological sort if |
3793 | they are not there yet. */ |
3794 | |
3795 | static void |
3796 | add_all_node_vals_to_toposort (cgraph_node *node, ipa_topo_info *topo) |
3797 | { |
3798 | ipa_node_params *info = ipa_node_params_sum->get (node); |
3799 | int i, count = ipa_get_param_count (info); |
3800 | |
3801 | for (i = 0; i < count; i++) |
3802 | { |
3803 | class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i); |
3804 | ipcp_lattice<tree> *lat = &plats->itself; |
3805 | struct ipcp_agg_lattice *aglat; |
3806 | |
3807 | if (!lat->bottom) |
3808 | { |
3809 | ipcp_value<tree> *val; |
3810 | for (val = lat->values; val; val = val->next) |
3811 | topo->constants.add_val (cur_val: val); |
3812 | } |
3813 | |
3814 | if (!plats->aggs_bottom) |
3815 | for (aglat = plats->aggs; aglat; aglat = aglat->next) |
3816 | if (!aglat->bottom) |
3817 | { |
3818 | ipcp_value<tree> *val; |
3819 | for (val = aglat->values; val; val = val->next) |
3820 | topo->constants.add_val (cur_val: val); |
3821 | } |
3822 | |
3823 | ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat; |
3824 | if (!ctxlat->bottom) |
3825 | { |
3826 | ipcp_value<ipa_polymorphic_call_context> *ctxval; |
3827 | for (ctxval = ctxlat->values; ctxval; ctxval = ctxval->next) |
3828 | topo->contexts.add_val (cur_val: ctxval); |
3829 | } |
3830 | } |
3831 | } |
3832 | |
3833 | /* One pass of constants propagation along the call graph edges, from callers |
3834 | to callees (requires topological ordering in TOPO), iterate over strongly |
3835 | connected components. */ |
3836 | |
3837 | static void |
3838 | propagate_constants_topo (class ipa_topo_info *topo) |
3839 | { |
3840 | int i; |
3841 | |
3842 | for (i = topo->nnodes - 1; i >= 0; i--) |
3843 | { |
3844 | unsigned j; |
3845 | struct cgraph_node *v, *node = topo->order[i]; |
3846 | vec<cgraph_node *> cycle_nodes = ipa_get_nodes_in_cycle (node); |
3847 | |
3848 | /* First, iteratively propagate within the strongly connected component |
3849 | until all lattices stabilize. */ |
3850 | FOR_EACH_VEC_ELT (cycle_nodes, j, v) |
3851 | if (v->has_gimple_body_p ()) |
3852 | { |
3853 | if (opt_for_fn (v->decl, flag_ipa_cp) |
3854 | && opt_for_fn (v->decl, optimize)) |
3855 | push_node_to_stack (topo, node: v); |
3856 | /* When V is not optimized, we can not push it to stack, but |
3857 | still we need to set all its callees lattices to bottom. */ |
3858 | else |
3859 | { |
3860 | for (cgraph_edge *cs = v->callees; cs; cs = cs->next_callee) |
3861 | propagate_constants_across_call (cs); |
3862 | } |
3863 | } |
3864 | |
3865 | v = pop_node_from_stack (topo); |
3866 | while (v) |
3867 | { |
3868 | struct cgraph_edge *cs; |
3869 | class ipa_node_params *info = NULL; |
3870 | bool self_scc = true; |
3871 | |
3872 | for (cs = v->callees; cs; cs = cs->next_callee) |
3873 | if (ipa_edge_within_scc (cs)) |
3874 | { |
3875 | cgraph_node *callee = cs->callee->function_symbol (); |
3876 | |
3877 | if (v != callee) |
3878 | self_scc = false; |
3879 | |
3880 | if (!info) |
3881 | { |
3882 | info = ipa_node_params_sum->get (node: v); |
3883 | info->node_within_scc = true; |
3884 | } |
3885 | |
3886 | if (propagate_constants_across_call (cs)) |
3887 | push_node_to_stack (topo, node: callee); |
3888 | } |
3889 | |
3890 | if (info) |
3891 | info->node_is_self_scc = self_scc; |
3892 | |
3893 | v = pop_node_from_stack (topo); |
3894 | } |
3895 | |
3896 | /* Afterwards, propagate along edges leading out of the SCC, calculates |
3897 | the local effects of the discovered constants and all valid values to |
3898 | their topological sort. */ |
3899 | FOR_EACH_VEC_ELT (cycle_nodes, j, v) |
3900 | if (v->has_gimple_body_p () |
3901 | && opt_for_fn (v->decl, flag_ipa_cp) |
3902 | && opt_for_fn (v->decl, optimize)) |
3903 | { |
3904 | struct cgraph_edge *cs; |
3905 | |
3906 | estimate_local_effects (node: v); |
3907 | add_all_node_vals_to_toposort (node: v, topo); |
3908 | for (cs = v->callees; cs; cs = cs->next_callee) |
3909 | if (!ipa_edge_within_scc (cs)) |
3910 | propagate_constants_across_call (cs); |
3911 | } |
3912 | cycle_nodes.release (); |
3913 | } |
3914 | } |
3915 | |
3916 | /* Propagate the estimated effects of individual values along the topological |
3917 | from the dependent values to those they depend on. */ |
3918 | |
3919 | template <typename valtype> |
3920 | void |
3921 | value_topo_info<valtype>::propagate_effects () |
3922 | { |
3923 | ipcp_value<valtype> *base; |
3924 | hash_set<ipcp_value<valtype> *> processed_srcvals; |
3925 | |
3926 | for (base = values_topo; base; base = base->topo_next) |
3927 | { |
3928 | ipcp_value_source<valtype> *src; |
3929 | ipcp_value<valtype> *val; |
3930 | sreal time = 0; |
3931 | HOST_WIDE_INT size = 0; |
3932 | |
3933 | for (val = base; val; val = val->scc_next) |
3934 | { |
3935 | time = time + val->local_time_benefit + val->prop_time_benefit; |
3936 | size = size + val->local_size_cost + val->prop_size_cost; |
3937 | } |
3938 | |
3939 | for (val = base; val; val = val->scc_next) |
3940 | { |
3941 | processed_srcvals.empty (); |
3942 | for (src = val->sources; src; src = src->next) |
3943 | if (src->val |
3944 | && src->cs->maybe_hot_p ()) |
3945 | { |
3946 | if (!processed_srcvals.add (src->val)) |
3947 | { |
3948 | HOST_WIDE_INT prop_size = size + src->val->prop_size_cost; |
3949 | if (prop_size < INT_MAX) |
3950 | src->val->prop_size_cost = prop_size; |
3951 | else |
3952 | continue; |
3953 | } |
3954 | |
3955 | int special_factor = 1; |
3956 | if (val->same_scc (src->val)) |
3957 | special_factor |
3958 | = opt_for_fn(src->cs->caller->decl, |
3959 | param_ipa_cp_recursive_freq_factor); |
3960 | else if (val->self_recursion_generated_p () |
3961 | && (src->cs->callee->function_symbol () |
3962 | == src->cs->caller)) |
3963 | { |
3964 | int max_recur_gen_depth |
3965 | = opt_for_fn(src->cs->caller->decl, |
3966 | param_ipa_cp_max_recursive_depth); |
3967 | special_factor = max_recur_gen_depth |
3968 | - val->self_recursion_generated_level + 1; |
3969 | } |
3970 | |
3971 | src->val->prop_time_benefit |
3972 | += time * special_factor * src->cs->sreal_frequency (); |
3973 | } |
3974 | |
3975 | if (size < INT_MAX) |
3976 | { |
3977 | val->prop_time_benefit = time; |
3978 | val->prop_size_cost = size; |
3979 | } |
3980 | else |
3981 | { |
3982 | val->prop_time_benefit = 0; |
3983 | val->prop_size_cost = 0; |
3984 | } |
3985 | } |
3986 | } |
3987 | } |
3988 | |
3989 | /* Callback for qsort to sort counts of all edges. */ |
3990 | |
3991 | static int |
3992 | compare_edge_profile_counts (const void *a, const void *b) |
3993 | { |
3994 | const profile_count *cnt1 = (const profile_count *) a; |
3995 | const profile_count *cnt2 = (const profile_count *) b; |
3996 | |
3997 | if (*cnt1 < *cnt2) |
3998 | return 1; |
3999 | if (*cnt1 > *cnt2) |
4000 | return -1; |
4001 | return 0; |
4002 | } |
4003 | |
4004 | |
4005 | /* Propagate constants, polymorphic contexts and their effects from the |
4006 | summaries interprocedurally. */ |
4007 | |
4008 | static void |
4009 | ipcp_propagate_stage (class ipa_topo_info *topo) |
4010 | { |
4011 | struct cgraph_node *node; |
4012 | |
4013 | if (dump_file) |
4014 | fprintf (stream: dump_file, format: "\n Propagating constants:\n\n" ); |
4015 | |
4016 | base_count = profile_count::uninitialized (); |
4017 | |
4018 | bool compute_count_base = false; |
4019 | unsigned base_count_pos_percent = 0; |
4020 | FOR_EACH_DEFINED_FUNCTION (node) |
4021 | { |
4022 | if (node->has_gimple_body_p () |
4023 | && opt_for_fn (node->decl, flag_ipa_cp) |
4024 | && opt_for_fn (node->decl, optimize)) |
4025 | { |
4026 | ipa_node_params *info = ipa_node_params_sum->get (node); |
4027 | determine_versionability (node, info); |
4028 | |
4029 | unsigned nlattices = ipa_get_param_count (info); |
4030 | info->lattices.safe_grow_cleared (len: nlattices, exact: true); |
4031 | initialize_node_lattices (node); |
4032 | } |
4033 | ipa_size_summary *s = ipa_size_summaries->get (node); |
4034 | if (node->definition && !node->alias && s != NULL) |
4035 | overall_size += s->self_size; |
4036 | if (node->count.ipa ().initialized_p ()) |
4037 | { |
4038 | compute_count_base = true; |
4039 | unsigned pos_percent = opt_for_fn (node->decl, |
4040 | param_ipa_cp_profile_count_base); |
4041 | base_count_pos_percent = MAX (base_count_pos_percent, pos_percent); |
4042 | } |
4043 | } |
4044 | |
4045 | if (compute_count_base) |
4046 | { |
4047 | auto_vec<profile_count> all_edge_counts; |
4048 | all_edge_counts.reserve_exact (nelems: symtab->edges_count); |
4049 | FOR_EACH_DEFINED_FUNCTION (node) |
4050 | for (cgraph_edge *cs = node->callees; cs; cs = cs->next_callee) |
4051 | { |
4052 | profile_count count = cs->count.ipa (); |
4053 | if (!count.nonzero_p ()) |
4054 | continue; |
4055 | |
4056 | enum availability avail; |
4057 | cgraph_node *tgt |
4058 | = cs->callee->function_or_virtual_thunk_symbol (avail: &avail); |
4059 | ipa_node_params *info = ipa_node_params_sum->get (node: tgt); |
4060 | if (info && info->versionable) |
4061 | all_edge_counts.quick_push (obj: count); |
4062 | } |
4063 | |
4064 | if (!all_edge_counts.is_empty ()) |
4065 | { |
4066 | gcc_assert (base_count_pos_percent <= 100); |
4067 | all_edge_counts.qsort (compare_edge_profile_counts); |
4068 | |
4069 | unsigned base_count_pos |
4070 | = ((all_edge_counts.length () * (base_count_pos_percent)) / 100); |
4071 | base_count = all_edge_counts[base_count_pos]; |
4072 | |
4073 | if (dump_file) |
4074 | { |
4075 | fprintf (stream: dump_file, format: "\nSelected base_count from %u edges at " |
4076 | "position %u, arriving at: " , all_edge_counts.length (), |
4077 | base_count_pos); |
4078 | base_count.dump (f: dump_file); |
4079 | fprintf (stream: dump_file, format: "\n" ); |
4080 | } |
4081 | } |
4082 | else if (dump_file) |
4083 | fprintf (stream: dump_file, format: "\nNo candidates with non-zero call count found, " |
4084 | "continuing as if without profile feedback.\n" ); |
4085 | } |
4086 | |
4087 | orig_overall_size = overall_size; |
4088 | |
4089 | if (dump_file) |
4090 | fprintf (stream: dump_file, format: "\noverall_size: %li\n" , overall_size); |
4091 | |
4092 | propagate_constants_topo (topo); |
4093 | if (flag_checking) |
4094 | ipcp_verify_propagated_values (); |
4095 | topo->constants.propagate_effects (); |
4096 | topo->contexts.propagate_effects (); |
4097 | |
4098 | if (dump_file) |
4099 | { |
4100 | fprintf (stream: dump_file, format: "\nIPA lattices after all propagation:\n" ); |
4101 | print_all_lattices (f: dump_file, dump_sources: (dump_flags & TDF_DETAILS), dump_benefits: true); |
4102 | } |
4103 | } |
4104 | |
4105 | /* Discover newly direct outgoing edges from NODE which is a new clone with |
4106 | known KNOWN_CSTS and make them direct. */ |
4107 | |
4108 | static void |
4109 | ipcp_discover_new_direct_edges (struct cgraph_node *node, |
4110 | vec<tree> known_csts, |
4111 | vec<ipa_polymorphic_call_context> |
4112 | known_contexts, |
4113 | vec<ipa_argagg_value, va_gc> *aggvals) |
4114 | { |
4115 | struct cgraph_edge *ie, *next_ie; |
4116 | bool found = false; |
4117 | |
4118 | for (ie = node->indirect_calls; ie; ie = next_ie) |
4119 | { |
4120 | tree target; |
4121 | bool speculative; |
4122 | |
4123 | next_ie = ie->next_callee; |
4124 | ipa_argagg_value_list avs (aggvals); |
4125 | target = ipa_get_indirect_edge_target_1 (ie, known_csts, known_contexts, |
4126 | avs, speculative: &speculative); |
4127 | if (target) |
4128 | { |
4129 | bool agg_contents = ie->indirect_info->agg_contents; |
4130 | bool polymorphic = ie->indirect_info->polymorphic; |
4131 | int param_index = ie->indirect_info->param_index; |
4132 | struct cgraph_edge *cs = ipa_make_edge_direct_to_target (ie, target, |
4133 | speculative); |
4134 | found = true; |
4135 | |
4136 | if (cs && !agg_contents && !polymorphic) |
4137 | { |
4138 | ipa_node_params *info = ipa_node_params_sum->get (node); |
4139 | int c = ipa_get_controlled_uses (info, i: param_index); |
4140 | if (c != IPA_UNDESCRIBED_USE |
4141 | && !ipa_get_param_load_dereferenced (info, i: param_index)) |
4142 | { |
4143 | struct ipa_ref *to_del; |
4144 | |
4145 | c--; |
4146 | ipa_set_controlled_uses (info, i: param_index, val: c); |
4147 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4148 | fprintf (stream: dump_file, format: " controlled uses count of param " |
4149 | "%i bumped down to %i\n" , param_index, c); |
4150 | if (c == 0 |
4151 | && (to_del = node->find_reference (referred_node: cs->callee, NULL, lto_stmt_uid: 0, |
4152 | use_type: IPA_REF_ADDR))) |
4153 | { |
4154 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4155 | fprintf (stream: dump_file, format: " and even removing its " |
4156 | "cloning-created reference\n" ); |
4157 | to_del->remove_reference (); |
4158 | } |
4159 | } |
4160 | } |
4161 | } |
4162 | } |
4163 | /* Turning calls to direct calls will improve overall summary. */ |
4164 | if (found) |
4165 | ipa_update_overall_fn_summary (node); |
4166 | } |
4167 | |
4168 | class edge_clone_summary; |
4169 | static call_summary <edge_clone_summary *> *edge_clone_summaries = NULL; |
4170 | |
4171 | /* Edge clone summary. */ |
4172 | |
4173 | class edge_clone_summary |
4174 | { |
4175 | public: |
4176 | /* Default constructor. */ |
4177 | edge_clone_summary (): prev_clone (NULL), next_clone (NULL) {} |
4178 | |
4179 | /* Default destructor. */ |
4180 | ~edge_clone_summary () |
4181 | { |
4182 | if (prev_clone) |
4183 | edge_clone_summaries->get (edge: prev_clone)->next_clone = next_clone; |
4184 | if (next_clone) |
4185 | edge_clone_summaries->get (edge: next_clone)->prev_clone = prev_clone; |
4186 | } |
4187 | |
4188 | cgraph_edge *prev_clone; |
4189 | cgraph_edge *next_clone; |
4190 | }; |
4191 | |
4192 | class edge_clone_summary_t: |
4193 | public call_summary <edge_clone_summary *> |
4194 | { |
4195 | public: |
4196 | edge_clone_summary_t (symbol_table *symtab): |
4197 | call_summary <edge_clone_summary *> (symtab) |
4198 | { |
4199 | m_initialize_when_cloning = true; |
4200 | } |
4201 | |
4202 | void duplicate (cgraph_edge *src_edge, cgraph_edge *dst_edge, |
4203 | edge_clone_summary *src_data, |
4204 | edge_clone_summary *dst_data) final override; |
4205 | }; |
4206 | |
4207 | /* Edge duplication hook. */ |
4208 | |
4209 | void |
4210 | edge_clone_summary_t::duplicate (cgraph_edge *src_edge, cgraph_edge *dst_edge, |
4211 | edge_clone_summary *src_data, |
4212 | edge_clone_summary *dst_data) |
4213 | { |
4214 | if (src_data->next_clone) |
4215 | edge_clone_summaries->get (edge: src_data->next_clone)->prev_clone = dst_edge; |
4216 | dst_data->prev_clone = src_edge; |
4217 | dst_data->next_clone = src_data->next_clone; |
4218 | src_data->next_clone = dst_edge; |
4219 | } |
4220 | |
4221 | /* Return true is CS calls DEST or its clone for all contexts. When |
4222 | ALLOW_RECURSION_TO_CLONE is false, also return false for self-recursive |
4223 | edges from/to an all-context clone. */ |
4224 | |
4225 | static bool |
4226 | calls_same_node_or_its_all_contexts_clone_p (cgraph_edge *cs, cgraph_node *dest, |
4227 | bool allow_recursion_to_clone) |
4228 | { |
4229 | enum availability availability; |
4230 | cgraph_node *callee = cs->callee->function_symbol (avail: &availability); |
4231 | |
4232 | if (availability <= AVAIL_INTERPOSABLE) |
4233 | return false; |
4234 | if (callee == dest) |
4235 | return true; |
4236 | if (!allow_recursion_to_clone && cs->caller == callee) |
4237 | return false; |
4238 | |
4239 | ipa_node_params *info = ipa_node_params_sum->get (node: callee); |
4240 | return info->is_all_contexts_clone && info->ipcp_orig_node == dest; |
4241 | } |
4242 | |
4243 | /* Return true if edge CS does bring about the value described by SRC to |
4244 | DEST_VAL of node DEST or its clone for all contexts. */ |
4245 | |
4246 | static bool |
4247 | cgraph_edge_brings_value_p (cgraph_edge *cs, ipcp_value_source<tree> *src, |
4248 | cgraph_node *dest, ipcp_value<tree> *dest_val) |
4249 | { |
4250 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: cs->caller); |
4251 | |
4252 | if (!calls_same_node_or_its_all_contexts_clone_p (cs, dest, allow_recursion_to_clone: !src->val) |
4253 | || caller_info->node_dead) |
4254 | return false; |
4255 | |
4256 | if (!src->val) |
4257 | return true; |
4258 | |
4259 | if (caller_info->ipcp_orig_node) |
4260 | { |
4261 | tree t = NULL_TREE; |
4262 | if (src->offset == -1) |
4263 | t = caller_info->known_csts[src->index]; |
4264 | else if (ipcp_transformation *ts |
4265 | = ipcp_get_transformation_summary (node: cs->caller)) |
4266 | { |
4267 | ipa_argagg_value_list avl (ts); |
4268 | t = avl.get_value (index: src->index, unit_offset: src->offset / BITS_PER_UNIT); |
4269 | } |
4270 | return (t != NULL_TREE |
4271 | && values_equal_for_ipcp_p (x: src->val->value, y: t)); |
4272 | } |
4273 | else |
4274 | { |
4275 | if (src->val == dest_val) |
4276 | return true; |
4277 | |
4278 | struct ipcp_agg_lattice *aglat; |
4279 | class ipcp_param_lattices *plats = ipa_get_parm_lattices (info: caller_info, |
4280 | i: src->index); |
4281 | if (src->offset == -1) |
4282 | return (plats->itself.is_single_const () |
4283 | && values_equal_for_ipcp_p (x: src->val->value, |
4284 | y: plats->itself.values->value)); |
4285 | else |
4286 | { |
4287 | if (plats->aggs_bottom || plats->aggs_contain_variable) |
4288 | return false; |
4289 | for (aglat = plats->aggs; aglat; aglat = aglat->next) |
4290 | if (aglat->offset == src->offset) |
4291 | return (aglat->is_single_const () |
4292 | && values_equal_for_ipcp_p (x: src->val->value, |
4293 | y: aglat->values->value)); |
4294 | } |
4295 | return false; |
4296 | } |
4297 | } |
4298 | |
4299 | /* Return true if edge CS does bring about the value described by SRC to |
4300 | DST_VAL of node DEST or its clone for all contexts. */ |
4301 | |
4302 | static bool |
4303 | cgraph_edge_brings_value_p (cgraph_edge *cs, |
4304 | ipcp_value_source<ipa_polymorphic_call_context> *src, |
4305 | cgraph_node *dest, |
4306 | ipcp_value<ipa_polymorphic_call_context> *) |
4307 | { |
4308 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: cs->caller); |
4309 | |
4310 | if (!calls_same_node_or_its_all_contexts_clone_p (cs, dest, allow_recursion_to_clone: true) |
4311 | || caller_info->node_dead) |
4312 | return false; |
4313 | if (!src->val) |
4314 | return true; |
4315 | |
4316 | if (caller_info->ipcp_orig_node) |
4317 | return (caller_info->known_contexts.length () > (unsigned) src->index) |
4318 | && values_equal_for_ipcp_p (x: src->val->value, |
4319 | y: caller_info->known_contexts[src->index]); |
4320 | |
4321 | class ipcp_param_lattices *plats = ipa_get_parm_lattices (info: caller_info, |
4322 | i: src->index); |
4323 | return plats->ctxlat.is_single_const () |
4324 | && values_equal_for_ipcp_p (x: src->val->value, |
4325 | y: plats->ctxlat.values->value); |
4326 | } |
4327 | |
4328 | /* Get the next clone in the linked list of clones of an edge. */ |
4329 | |
4330 | static inline struct cgraph_edge * |
4331 | get_next_cgraph_edge_clone (struct cgraph_edge *cs) |
4332 | { |
4333 | edge_clone_summary *s = edge_clone_summaries->get (edge: cs); |
4334 | return s != NULL ? s->next_clone : NULL; |
4335 | } |
4336 | |
4337 | /* Given VAL that is intended for DEST, iterate over all its sources and if any |
4338 | of them is viable and hot, return true. In that case, for those that still |
4339 | hold, add their edge frequency and their number and cumulative profile |
4340 | counts of self-ecursive and other edges into *FREQUENCY, *CALLER_COUNT, |
4341 | REC_COUNT_SUM and NONREC_COUNT_SUM respectively. */ |
4342 | |
4343 | template <typename valtype> |
4344 | static bool |
4345 | get_info_about_necessary_edges (ipcp_value<valtype> *val, cgraph_node *dest, |
4346 | sreal *freq_sum, int *caller_count, |
4347 | profile_count *rec_count_sum, |
4348 | profile_count *nonrec_count_sum) |
4349 | { |
4350 | ipcp_value_source<valtype> *src; |
4351 | sreal freq = 0; |
4352 | int count = 0; |
4353 | profile_count rec_cnt = profile_count::zero (); |
4354 | profile_count nonrec_cnt = profile_count::zero (); |
4355 | bool hot = false; |
4356 | bool non_self_recursive = false; |
4357 | |
4358 | for (src = val->sources; src; src = src->next) |
4359 | { |
4360 | struct cgraph_edge *cs = src->cs; |
4361 | while (cs) |
4362 | { |
4363 | if (cgraph_edge_brings_value_p (cs, src, dest, val)) |
4364 | { |
4365 | count++; |
4366 | freq += cs->sreal_frequency (); |
4367 | hot |= cs->maybe_hot_p (); |
4368 | if (cs->caller != dest) |
4369 | { |
4370 | non_self_recursive = true; |
4371 | if (cs->count.ipa ().initialized_p ()) |
4372 | rec_cnt += cs->count.ipa (); |
4373 | } |
4374 | else if (cs->count.ipa ().initialized_p ()) |
4375 | nonrec_cnt += cs->count.ipa (); |
4376 | } |
4377 | cs = get_next_cgraph_edge_clone (cs); |
4378 | } |
4379 | } |
4380 | |
4381 | /* If the only edges bringing a value are self-recursive ones, do not bother |
4382 | evaluating it. */ |
4383 | if (!non_self_recursive) |
4384 | return false; |
4385 | |
4386 | *freq_sum = freq; |
4387 | *caller_count = count; |
4388 | *rec_count_sum = rec_cnt; |
4389 | *nonrec_count_sum = nonrec_cnt; |
4390 | |
4391 | if (!hot && ipa_node_params_sum->get (node: dest)->node_within_scc) |
4392 | { |
4393 | struct cgraph_edge *cs; |
4394 | |
4395 | /* Cold non-SCC source edge could trigger hot recursive execution of |
4396 | function. Consider the case as hot and rely on following cost model |
4397 | computation to further select right one. */ |
4398 | for (cs = dest->callers; cs; cs = cs->next_caller) |
4399 | if (cs->caller == dest && cs->maybe_hot_p ()) |
4400 | return true; |
4401 | } |
4402 | |
4403 | return hot; |
4404 | } |
4405 | |
4406 | /* Given a NODE, and a set of its CALLERS, try to adjust order of the callers |
4407 | to let a non-self-recursive caller be the first element. Thus, we can |
4408 | simplify intersecting operations on values that arrive from all of these |
4409 | callers, especially when there exists self-recursive call. Return true if |
4410 | this kind of adjustment is possible. */ |
4411 | |
4412 | static bool |
4413 | adjust_callers_for_value_intersection (vec<cgraph_edge *> &callers, |
4414 | cgraph_node *node) |
4415 | { |
4416 | for (unsigned i = 0; i < callers.length (); i++) |
4417 | { |
4418 | cgraph_edge *cs = callers[i]; |
4419 | |
4420 | if (cs->caller != node) |
4421 | { |
4422 | if (i > 0) |
4423 | { |
4424 | callers[i] = callers[0]; |
4425 | callers[0] = cs; |
4426 | } |
4427 | return true; |
4428 | } |
4429 | } |
4430 | return false; |
4431 | } |
4432 | |
4433 | /* Return a vector of incoming edges that do bring value VAL to node DEST. It |
4434 | is assumed their number is known and equal to CALLER_COUNT. */ |
4435 | |
4436 | template <typename valtype> |
4437 | static vec<cgraph_edge *> |
4438 | gather_edges_for_value (ipcp_value<valtype> *val, cgraph_node *dest, |
4439 | int caller_count) |
4440 | { |
4441 | ipcp_value_source<valtype> *src; |
4442 | vec<cgraph_edge *> ret; |
4443 | |
4444 | ret.create (nelems: caller_count); |
4445 | for (src = val->sources; src; src = src->next) |
4446 | { |
4447 | struct cgraph_edge *cs = src->cs; |
4448 | while (cs) |
4449 | { |
4450 | if (cgraph_edge_brings_value_p (cs, src, dest, val)) |
4451 | ret.quick_push (obj: cs); |
4452 | cs = get_next_cgraph_edge_clone (cs); |
4453 | } |
4454 | } |
4455 | |
4456 | if (caller_count > 1) |
4457 | adjust_callers_for_value_intersection (callers&: ret, node: dest); |
4458 | |
4459 | return ret; |
4460 | } |
4461 | |
4462 | /* Construct a replacement map for a know VALUE for a formal parameter PARAM. |
4463 | Return it or NULL if for some reason it cannot be created. FORCE_LOAD_REF |
4464 | should be set to true when the reference created for the constant should be |
4465 | a load one and not an address one because the corresponding parameter p is |
4466 | only used as *p. */ |
4467 | |
4468 | static struct ipa_replace_map * |
4469 | get_replacement_map (class ipa_node_params *info, tree value, int parm_num, |
4470 | bool force_load_ref) |
4471 | { |
4472 | struct ipa_replace_map *replace_map; |
4473 | |
4474 | replace_map = ggc_alloc<ipa_replace_map> (); |
4475 | if (dump_file) |
4476 | { |
4477 | fprintf (stream: dump_file, format: " replacing " ); |
4478 | ipa_dump_param (dump_file, info, i: parm_num); |
4479 | |
4480 | fprintf (stream: dump_file, format: " with const " ); |
4481 | print_generic_expr (dump_file, value); |
4482 | |
4483 | if (force_load_ref) |
4484 | fprintf (stream: dump_file, format: " - forcing load reference\n" ); |
4485 | else |
4486 | fprintf (stream: dump_file, format: "\n" ); |
4487 | } |
4488 | replace_map->parm_num = parm_num; |
4489 | replace_map->new_tree = value; |
4490 | replace_map->force_load_ref = force_load_ref; |
4491 | return replace_map; |
4492 | } |
4493 | |
4494 | /* Dump new profiling counts of NODE. SPEC is true when NODE is a specialzied |
4495 | one, otherwise it will be referred to as the original node. */ |
4496 | |
4497 | static void |
4498 | dump_profile_updates (cgraph_node *node, bool spec) |
4499 | { |
4500 | if (spec) |
4501 | fprintf (stream: dump_file, format: " setting count of the specialized node %s to " , |
4502 | node->dump_name ()); |
4503 | else |
4504 | fprintf (stream: dump_file, format: " setting count of the original node %s to " , |
4505 | node->dump_name ()); |
4506 | |
4507 | node->count.dump (f: dump_file); |
4508 | fprintf (stream: dump_file, format: "\n" ); |
4509 | for (cgraph_edge *cs = node->callees; cs; cs = cs->next_callee) |
4510 | { |
4511 | fprintf (stream: dump_file, format: " edge to %s has count " , |
4512 | cs->callee->dump_name ()); |
4513 | cs->count.dump (f: dump_file); |
4514 | fprintf (stream: dump_file, format: "\n" ); |
4515 | } |
4516 | } |
4517 | |
4518 | /* With partial train run we do not want to assume that original's count is |
4519 | zero whenever we redurect all executed edges to clone. Simply drop profile |
4520 | to local one in this case. In eany case, return the new value. ORIG_NODE |
4521 | is the original node and its count has not been updaed yet. */ |
4522 | |
4523 | profile_count |
4524 | lenient_count_portion_handling (profile_count remainder, cgraph_node *orig_node) |
4525 | { |
4526 | if (remainder.ipa_p () && !remainder.ipa ().nonzero_p () |
4527 | && orig_node->count.ipa_p () && orig_node->count.ipa ().nonzero_p () |
4528 | && opt_for_fn (orig_node->decl, flag_profile_partial_training)) |
4529 | remainder = remainder.guessed_local (); |
4530 | |
4531 | return remainder; |
4532 | } |
4533 | |
4534 | /* Structure to sum counts coming from nodes other than the original node and |
4535 | its clones. */ |
4536 | |
4537 | struct gather_other_count_struct |
4538 | { |
4539 | cgraph_node *orig; |
4540 | profile_count other_count; |
4541 | }; |
4542 | |
4543 | /* Worker callback of call_for_symbol_thunks_and_aliases summing the number of |
4544 | counts that come from non-self-recursive calls.. */ |
4545 | |
4546 | static bool |
4547 | gather_count_of_non_rec_edges (cgraph_node *node, void *data) |
4548 | { |
4549 | gather_other_count_struct *desc = (gather_other_count_struct *) data; |
4550 | for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller) |
4551 | if (cs->caller != desc->orig && cs->caller->clone_of != desc->orig) |
4552 | desc->other_count += cs->count.ipa (); |
4553 | return false; |
4554 | } |
4555 | |
4556 | /* Structure to help analyze if we need to boost counts of some clones of some |
4557 | non-recursive edges to match the new callee count. */ |
4558 | |
4559 | struct desc_incoming_count_struct |
4560 | { |
4561 | cgraph_node *orig; |
4562 | hash_set <cgraph_edge *> *processed_edges; |
4563 | profile_count count; |
4564 | unsigned unproc_orig_rec_edges; |
4565 | }; |
4566 | |
4567 | /* Go over edges calling NODE and its thunks and gather information about |
4568 | incoming counts so that we know if we need to make any adjustments. */ |
4569 | |
4570 | static void |
4571 | analyze_clone_icoming_counts (cgraph_node *node, |
4572 | desc_incoming_count_struct *desc) |
4573 | { |
4574 | for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller) |
4575 | if (cs->caller->thunk) |
4576 | { |
4577 | analyze_clone_icoming_counts (node: cs->caller, desc); |
4578 | continue; |
4579 | } |
4580 | else |
4581 | { |
4582 | if (cs->count.initialized_p ()) |
4583 | desc->count += cs->count.ipa (); |
4584 | if (!desc->processed_edges->contains (k: cs) |
4585 | && cs->caller->clone_of == desc->orig) |
4586 | desc->unproc_orig_rec_edges++; |
4587 | } |
4588 | } |
4589 | |
4590 | /* If caller edge counts of a clone created for a self-recursive arithmetic |
4591 | jump function must be adjusted because it is coming from a the "seed" clone |
4592 | for the first value and so has been excessively scaled back as if it was not |
4593 | a recursive call, adjust it so that the incoming counts of NODE match its |
4594 | count. NODE is the node or its thunk. */ |
4595 | |
4596 | static void |
4597 | adjust_clone_incoming_counts (cgraph_node *node, |
4598 | desc_incoming_count_struct *desc) |
4599 | { |
4600 | for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller) |
4601 | if (cs->caller->thunk) |
4602 | { |
4603 | adjust_clone_incoming_counts (node: cs->caller, desc); |
4604 | profile_count sum = profile_count::zero (); |
4605 | for (cgraph_edge *e = cs->caller->callers; e; e = e->next_caller) |
4606 | if (e->count.initialized_p ()) |
4607 | sum += e->count.ipa (); |
4608 | cs->count = cs->count.combine_with_ipa_count (ipa: sum); |
4609 | } |
4610 | else if (!desc->processed_edges->contains (k: cs) |
4611 | && cs->caller->clone_of == desc->orig) |
4612 | { |
4613 | cs->count += desc->count; |
4614 | if (dump_file) |
4615 | { |
4616 | fprintf (stream: dump_file, format: " Adjusted count of an incoming edge of " |
4617 | "a clone %s -> %s to " , cs->caller->dump_name (), |
4618 | cs->callee->dump_name ()); |
4619 | cs->count.dump (f: dump_file); |
4620 | fprintf (stream: dump_file, format: "\n" ); |
4621 | } |
4622 | } |
4623 | } |
4624 | |
4625 | /* When ORIG_NODE has been cloned for values which have been generated fora |
4626 | self-recursive call as a result of an arithmetic pass-through |
4627 | jump-functions, adjust its count together with counts of all such clones in |
4628 | SELF_GEN_CLONES which also at this point contains ORIG_NODE itself. |
4629 | |
4630 | The function sums the counts of the original node and all its clones that |
4631 | cannot be attributed to a specific clone because it comes from a |
4632 | non-recursive edge. This sum is then evenly divided between the clones and |
4633 | on top of that each one gets all the counts which can be attributed directly |
4634 | to it. */ |
4635 | |
4636 | static void |
4637 | update_counts_for_self_gen_clones (cgraph_node *orig_node, |
4638 | const vec<cgraph_node *> &self_gen_clones) |
4639 | { |
4640 | profile_count redist_sum = orig_node->count.ipa (); |
4641 | if (!(redist_sum > profile_count::zero ())) |
4642 | return; |
4643 | |
4644 | if (dump_file) |
4645 | fprintf (stream: dump_file, format: " Updating profile of self recursive clone " |
4646 | "series\n" ); |
4647 | |
4648 | gather_other_count_struct gocs; |
4649 | gocs.orig = orig_node; |
4650 | gocs.other_count = profile_count::zero (); |
4651 | |
4652 | auto_vec <profile_count, 8> other_edges_count; |
4653 | for (cgraph_node *n : self_gen_clones) |
4654 | { |
4655 | gocs.other_count = profile_count::zero (); |
4656 | n->call_for_symbol_thunks_and_aliases (callback: gather_count_of_non_rec_edges, |
4657 | data: &gocs, include_overwritable: false); |
4658 | other_edges_count.safe_push (obj: gocs.other_count); |
4659 | redist_sum -= gocs.other_count; |
4660 | } |
4661 | |
4662 | hash_set<cgraph_edge *> processed_edges; |
4663 | unsigned i = 0; |
4664 | for (cgraph_node *n : self_gen_clones) |
4665 | { |
4666 | profile_count orig_count = n->count; |
4667 | profile_count new_count |
4668 | = (redist_sum / self_gen_clones.length () + other_edges_count[i]); |
4669 | new_count = lenient_count_portion_handling (remainder: new_count, orig_node); |
4670 | n->count = new_count; |
4671 | profile_count::adjust_for_ipa_scaling (num: &new_count, den: &orig_count); |
4672 | for (cgraph_edge *cs = n->callees; cs; cs = cs->next_callee) |
4673 | { |
4674 | cs->count = cs->count.apply_scale (num: new_count, den: orig_count); |
4675 | processed_edges.add (k: cs); |
4676 | } |
4677 | for (cgraph_edge *cs = n->indirect_calls; cs; cs = cs->next_callee) |
4678 | cs->count = cs->count.apply_scale (num: new_count, den: orig_count); |
4679 | |
4680 | i++; |
4681 | } |
4682 | |
4683 | /* There are still going to be edges to ORIG_NODE that have one or more |
4684 | clones coming from another node clone in SELF_GEN_CLONES and which we |
4685 | scaled by the same amount, which means that the total incoming sum of |
4686 | counts to ORIG_NODE will be too high, scale such edges back. */ |
4687 | for (cgraph_edge *cs = orig_node->callees; cs; cs = cs->next_callee) |
4688 | { |
4689 | if (cs->callee->ultimate_alias_target () == orig_node) |
4690 | { |
4691 | unsigned den = 0; |
4692 | for (cgraph_edge *e = cs; e; e = get_next_cgraph_edge_clone (cs: e)) |
4693 | if (e->callee->ultimate_alias_target () == orig_node |
4694 | && processed_edges.contains (k: e)) |
4695 | den++; |
4696 | if (den > 0) |
4697 | for (cgraph_edge *e = cs; e; e = get_next_cgraph_edge_clone (cs: e)) |
4698 | if (e->callee->ultimate_alias_target () == orig_node |
4699 | && processed_edges.contains (k: e)) |
4700 | e->count /= den; |
4701 | } |
4702 | } |
4703 | |
4704 | /* Edges from the seeds of the valus generated for arithmetic jump-functions |
4705 | along self-recursive edges are likely to have fairly low count and so |
4706 | edges from them to nodes in the self_gen_clones do not correspond to the |
4707 | artificially distributed count of the nodes, the total sum of incoming |
4708 | edges to some clones might be too low. Detect this situation and correct |
4709 | it. */ |
4710 | for (cgraph_node *n : self_gen_clones) |
4711 | { |
4712 | if (!(n->count.ipa () > profile_count::zero ())) |
4713 | continue; |
4714 | |
4715 | desc_incoming_count_struct desc; |
4716 | desc.orig = orig_node; |
4717 | desc.processed_edges = &processed_edges; |
4718 | desc.count = profile_count::zero (); |
4719 | desc.unproc_orig_rec_edges = 0; |
4720 | analyze_clone_icoming_counts (node: n, desc: &desc); |
4721 | |
4722 | if (n->count.differs_from_p (other: desc.count)) |
4723 | { |
4724 | if (n->count > desc.count |
4725 | && desc.unproc_orig_rec_edges > 0) |
4726 | { |
4727 | desc.count = n->count - desc.count; |
4728 | desc.count = desc.count /= desc.unproc_orig_rec_edges; |
4729 | adjust_clone_incoming_counts (node: n, desc: &desc); |
4730 | } |
4731 | else if (dump_file) |
4732 | fprintf (stream: dump_file, |
4733 | format: " Unable to fix up incoming counts for %s.\n" , |
4734 | n->dump_name ()); |
4735 | } |
4736 | } |
4737 | |
4738 | if (dump_file) |
4739 | for (cgraph_node *n : self_gen_clones) |
4740 | dump_profile_updates (node: n, spec: n != orig_node); |
4741 | return; |
4742 | } |
4743 | |
4744 | /* After a specialized NEW_NODE version of ORIG_NODE has been created, update |
4745 | their profile information to reflect this. This function should not be used |
4746 | for clones generated for arithmetic pass-through jump functions on a |
4747 | self-recursive call graph edge, that situation is handled by |
4748 | update_counts_for_self_gen_clones. */ |
4749 | |
4750 | static void |
4751 | update_profiling_info (struct cgraph_node *orig_node, |
4752 | struct cgraph_node *new_node) |
4753 | { |
4754 | struct caller_statistics stats; |
4755 | profile_count new_sum; |
4756 | profile_count remainder, orig_node_count = orig_node->count.ipa (); |
4757 | |
4758 | if (!(orig_node_count > profile_count::zero ())) |
4759 | return; |
4760 | |
4761 | if (dump_file) |
4762 | { |
4763 | fprintf (stream: dump_file, format: " Updating profile from original count: " ); |
4764 | orig_node_count.dump (f: dump_file); |
4765 | fprintf (stream: dump_file, format: "\n" ); |
4766 | } |
4767 | |
4768 | init_caller_stats (stats: &stats, itself: new_node); |
4769 | new_node->call_for_symbol_thunks_and_aliases (callback: gather_caller_stats, data: &stats, |
4770 | include_overwritable: false); |
4771 | new_sum = stats.count_sum; |
4772 | |
4773 | bool orig_edges_processed = false; |
4774 | if (new_sum > orig_node_count) |
4775 | { |
4776 | /* TODO: Profile has alreay gone astray, keep what we have but lower it |
4777 | to global0 category. */ |
4778 | remainder = orig_node->count.global0 (); |
4779 | |
4780 | for (cgraph_edge *cs = orig_node->callees; cs; cs = cs->next_callee) |
4781 | cs->count = cs->count.global0 (); |
4782 | for (cgraph_edge *cs = orig_node->indirect_calls; |
4783 | cs; |
4784 | cs = cs->next_callee) |
4785 | cs->count = cs->count.global0 (); |
4786 | orig_edges_processed = true; |
4787 | } |
4788 | else if (stats.rec_count_sum.nonzero_p ()) |
4789 | { |
4790 | int new_nonrec_calls = stats.n_nonrec_calls; |
4791 | /* There are self-recursive edges which are likely to bring in the |
4792 | majority of calls but which we must divide in between the original and |
4793 | new node. */ |
4794 | init_caller_stats (stats: &stats, itself: orig_node); |
4795 | orig_node->call_for_symbol_thunks_and_aliases (callback: gather_caller_stats, |
4796 | data: &stats, include_overwritable: false); |
4797 | int orig_nonrec_calls = stats.n_nonrec_calls; |
4798 | profile_count orig_nonrec_call_count = stats.count_sum; |
4799 | |
4800 | if (orig_node->local) |
4801 | { |
4802 | if (!orig_nonrec_call_count.nonzero_p ()) |
4803 | { |
4804 | if (dump_file) |
4805 | fprintf (stream: dump_file, format: " The original is local and the only " |
4806 | "incoming edges from non-dead callers with nonzero " |
4807 | "counts are self-recursive, assuming it is cold.\n" ); |
4808 | /* The NEW_NODE count and counts of all its outgoing edges |
4809 | are still unmodified copies of ORIG_NODE's. Just clear |
4810 | the latter and bail out. */ |
4811 | profile_count zero; |
4812 | if (opt_for_fn (orig_node->decl, flag_profile_partial_training)) |
4813 | zero = profile_count::zero ().guessed_local (); |
4814 | else |
4815 | zero = profile_count::adjusted_zero (); |
4816 | orig_node->count = zero; |
4817 | for (cgraph_edge *cs = orig_node->callees; |
4818 | cs; |
4819 | cs = cs->next_callee) |
4820 | cs->count = zero; |
4821 | for (cgraph_edge *cs = orig_node->indirect_calls; |
4822 | cs; |
4823 | cs = cs->next_callee) |
4824 | cs->count = zero; |
4825 | return; |
4826 | } |
4827 | } |
4828 | else |
4829 | { |
4830 | /* Let's behave as if there was another caller that accounts for all |
4831 | the calls that were either indirect or from other compilation |
4832 | units. */ |
4833 | orig_nonrec_calls++; |
4834 | profile_count pretend_caller_count |
4835 | = (orig_node_count - new_sum - orig_nonrec_call_count |
4836 | - stats.rec_count_sum); |
4837 | orig_nonrec_call_count += pretend_caller_count; |
4838 | } |
4839 | |
4840 | /* Divide all "unexplained" counts roughly proportionally to sums of |
4841 | counts of non-recursive calls. |
4842 | |
4843 | We put rather arbitrary limits on how many counts we claim because the |
4844 | number of non-self-recursive incoming count is only a rough guideline |
4845 | and there are cases (such as mcf) where using it blindly just takes |
4846 | too many. And if lattices are considered in the opposite order we |
4847 | could also take too few. */ |
4848 | profile_count unexp = orig_node_count - new_sum - orig_nonrec_call_count; |
4849 | |
4850 | int limit_den = 2 * (orig_nonrec_calls + new_nonrec_calls); |
4851 | profile_count new_part |
4852 | = MAX(MIN (unexp.apply_scale (new_sum, |
4853 | new_sum + orig_nonrec_call_count), |
4854 | unexp.apply_scale (limit_den - 1, limit_den)), |
4855 | unexp.apply_scale (new_nonrec_calls, limit_den)); |
4856 | if (dump_file) |
4857 | { |
4858 | fprintf (stream: dump_file, format: " Claiming " ); |
4859 | new_part.dump (f: dump_file); |
4860 | fprintf (stream: dump_file, format: " of unexplained " ); |
4861 | unexp.dump (f: dump_file); |
4862 | fprintf (stream: dump_file, format: " counts because of self-recursive " |
4863 | "calls\n" ); |
4864 | } |
4865 | new_sum += new_part; |
4866 | remainder = lenient_count_portion_handling (remainder: orig_node_count - new_sum, |
4867 | orig_node); |
4868 | } |
4869 | else |
4870 | remainder = lenient_count_portion_handling (remainder: orig_node_count - new_sum, |
4871 | orig_node); |
4872 | |
4873 | new_sum = orig_node_count.combine_with_ipa_count (ipa: new_sum); |
4874 | new_node->count = new_sum; |
4875 | orig_node->count = remainder; |
4876 | |
4877 | profile_count orig_new_node_count = orig_node_count; |
4878 | profile_count::adjust_for_ipa_scaling (num: &new_sum, den: &orig_new_node_count); |
4879 | for (cgraph_edge *cs = new_node->callees; cs; cs = cs->next_callee) |
4880 | cs->count = cs->count.apply_scale (num: new_sum, den: orig_new_node_count); |
4881 | for (cgraph_edge *cs = new_node->indirect_calls; cs; cs = cs->next_callee) |
4882 | cs->count = cs->count.apply_scale (num: new_sum, den: orig_new_node_count); |
4883 | |
4884 | if (!orig_edges_processed) |
4885 | { |
4886 | profile_count::adjust_for_ipa_scaling (num: &remainder, den: &orig_node_count); |
4887 | for (cgraph_edge *cs = orig_node->callees; cs; cs = cs->next_callee) |
4888 | cs->count = cs->count.apply_scale (num: remainder, den: orig_node_count); |
4889 | for (cgraph_edge *cs = orig_node->indirect_calls; |
4890 | cs; |
4891 | cs = cs->next_callee) |
4892 | cs->count = cs->count.apply_scale (num: remainder, den: orig_node_count); |
4893 | } |
4894 | |
4895 | if (dump_file) |
4896 | { |
4897 | dump_profile_updates (node: new_node, spec: true); |
4898 | dump_profile_updates (node: orig_node, spec: false); |
4899 | } |
4900 | } |
4901 | |
4902 | /* Update the respective profile of specialized NEW_NODE and the original |
4903 | ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM |
4904 | have been redirected to the specialized version. */ |
4905 | |
4906 | static void |
4907 | update_specialized_profile (struct cgraph_node *new_node, |
4908 | struct cgraph_node *orig_node, |
4909 | profile_count redirected_sum) |
4910 | { |
4911 | struct cgraph_edge *cs; |
4912 | profile_count new_node_count, orig_node_count = orig_node->count.ipa (); |
4913 | |
4914 | if (dump_file) |
4915 | { |
4916 | fprintf (stream: dump_file, format: " the sum of counts of redirected edges is " ); |
4917 | redirected_sum.dump (f: dump_file); |
4918 | fprintf (stream: dump_file, format: "\n old ipa count of the original node is " ); |
4919 | orig_node_count.dump (f: dump_file); |
4920 | fprintf (stream: dump_file, format: "\n" ); |
4921 | } |
4922 | if (!(orig_node_count > profile_count::zero ())) |
4923 | return; |
4924 | |
4925 | new_node_count = new_node->count; |
4926 | new_node->count += redirected_sum; |
4927 | orig_node->count |
4928 | = lenient_count_portion_handling (remainder: orig_node->count - redirected_sum, |
4929 | orig_node); |
4930 | |
4931 | for (cs = new_node->callees; cs; cs = cs->next_callee) |
4932 | cs->count += cs->count.apply_scale (num: redirected_sum, den: new_node_count); |
4933 | |
4934 | for (cs = orig_node->callees; cs; cs = cs->next_callee) |
4935 | { |
4936 | profile_count dec = cs->count.apply_scale (num: redirected_sum, |
4937 | den: orig_node_count); |
4938 | cs->count -= dec; |
4939 | } |
4940 | |
4941 | if (dump_file) |
4942 | { |
4943 | dump_profile_updates (node: new_node, spec: true); |
4944 | dump_profile_updates (node: orig_node, spec: false); |
4945 | } |
4946 | } |
4947 | |
4948 | static void adjust_references_in_caller (cgraph_edge *cs, |
4949 | symtab_node *symbol, int index); |
4950 | |
4951 | /* Simple structure to pass a symbol and index (with same meaning as parameters |
4952 | of adjust_references_in_caller) through a void* parameter of a |
4953 | call_for_symbol_thunks_and_aliases callback. */ |
4954 | struct symbol_and_index_together |
4955 | { |
4956 | symtab_node *symbol; |
4957 | int index; |
4958 | }; |
4959 | |
4960 | /* Worker callback of call_for_symbol_thunks_and_aliases to recursively call |
4961 | adjust_references_in_caller on edges up in the call-graph, if necessary. */ |
4962 | static bool |
4963 | adjust_refs_in_act_callers (struct cgraph_node *node, void *data) |
4964 | { |
4965 | symbol_and_index_together *pack = (symbol_and_index_together *) data; |
4966 | for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller) |
4967 | if (!cs->caller->thunk) |
4968 | adjust_references_in_caller (cs, symbol: pack->symbol, index: pack->index); |
4969 | return false; |
4970 | } |
4971 | |
4972 | /* At INDEX of a function being called by CS there is an ADDR_EXPR of a |
4973 | variable which is only dereferenced and which is represented by SYMBOL. See |
4974 | if we can remove ADDR reference in callers assosiated witht the call. */ |
4975 | |
4976 | static void |
4977 | adjust_references_in_caller (cgraph_edge *cs, symtab_node *symbol, int index) |
4978 | { |
4979 | ipa_edge_args *args = ipa_edge_args_sum->get (edge: cs); |
4980 | ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i: index); |
4981 | if (jfunc->type == IPA_JF_CONST) |
4982 | { |
4983 | ipa_ref *to_del = cs->caller->find_reference (referred_node: symbol, stmt: cs->call_stmt, |
4984 | lto_stmt_uid: cs->lto_stmt_uid, |
4985 | use_type: IPA_REF_ADDR); |
4986 | if (!to_del) |
4987 | return; |
4988 | to_del->remove_reference (); |
4989 | ipa_zap_jf_refdesc (jfunc); |
4990 | if (dump_file) |
4991 | fprintf (stream: dump_file, format: " Removed a reference from %s to %s.\n" , |
4992 | cs->caller->dump_name (), symbol->dump_name ()); |
4993 | return; |
4994 | } |
4995 | |
4996 | if (jfunc->type != IPA_JF_PASS_THROUGH |
4997 | || ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR |
4998 | || ipa_get_jf_pass_through_refdesc_decremented (jfunc)) |
4999 | return; |
5000 | |
5001 | int fidx = ipa_get_jf_pass_through_formal_id (jfunc); |
5002 | cgraph_node *caller = cs->caller; |
5003 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: caller); |
5004 | /* TODO: This consistency check may be too big and not really |
5005 | that useful. Consider removing it. */ |
5006 | tree cst; |
5007 | if (caller_info->ipcp_orig_node) |
5008 | cst = caller_info->known_csts[fidx]; |
5009 | else |
5010 | { |
5011 | ipcp_lattice<tree> *lat = ipa_get_scalar_lat (info: caller_info, i: fidx); |
5012 | gcc_assert (lat->is_single_const ()); |
5013 | cst = lat->values->value; |
5014 | } |
5015 | gcc_assert (TREE_CODE (cst) == ADDR_EXPR |
5016 | && (symtab_node::get (get_base_address (TREE_OPERAND (cst, 0))) |
5017 | == symbol)); |
5018 | |
5019 | int cuses = ipa_get_controlled_uses (info: caller_info, i: fidx); |
5020 | if (cuses == IPA_UNDESCRIBED_USE) |
5021 | return; |
5022 | gcc_assert (cuses > 0); |
5023 | cuses--; |
5024 | ipa_set_controlled_uses (info: caller_info, i: fidx, val: cuses); |
5025 | ipa_set_jf_pass_through_refdesc_decremented (jfunc, value: true); |
5026 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5027 | fprintf (stream: dump_file, format: " Controlled uses of parameter %i of %s dropped " |
5028 | "to %i.\n" , fidx, caller->dump_name (), cuses); |
5029 | if (cuses) |
5030 | return; |
5031 | |
5032 | if (caller_info->ipcp_orig_node) |
5033 | { |
5034 | /* Cloning machinery has created a reference here, we need to either |
5035 | remove it or change it to a read one. */ |
5036 | ipa_ref *to_del = caller->find_reference (referred_node: symbol, NULL, lto_stmt_uid: 0, use_type: IPA_REF_ADDR); |
5037 | if (to_del) |
5038 | { |
5039 | to_del->remove_reference (); |
5040 | if (dump_file) |
5041 | fprintf (stream: dump_file, format: " Removed a reference from %s to %s.\n" , |
5042 | cs->caller->dump_name (), symbol->dump_name ()); |
5043 | if (ipa_get_param_load_dereferenced (info: caller_info, i: fidx)) |
5044 | { |
5045 | caller->create_reference (referred_node: symbol, use_type: IPA_REF_LOAD, NULL); |
5046 | if (dump_file) |
5047 | fprintf (stream: dump_file, |
5048 | format: " ...and replaced it with LOAD one.\n" ); |
5049 | } |
5050 | } |
5051 | } |
5052 | |
5053 | symbol_and_index_together pack; |
5054 | pack.symbol = symbol; |
5055 | pack.index = fidx; |
5056 | if (caller->can_change_signature) |
5057 | caller->call_for_symbol_thunks_and_aliases (callback: adjust_refs_in_act_callers, |
5058 | data: &pack, include_overwritable: true); |
5059 | } |
5060 | |
5061 | |
5062 | /* Return true if we would like to remove a parameter from NODE when cloning it |
5063 | with KNOWN_CSTS scalar constants. */ |
5064 | |
5065 | static bool |
5066 | want_remove_some_param_p (cgraph_node *node, vec<tree> known_csts) |
5067 | { |
5068 | auto_vec<bool, 16> surviving; |
5069 | bool filled_vec = false; |
5070 | ipa_node_params *info = ipa_node_params_sum->get (node); |
5071 | int i, count = ipa_get_param_count (info); |
5072 | |
5073 | for (i = 0; i < count; i++) |
5074 | { |
5075 | if (!known_csts[i] && ipa_is_param_used (info, i)) |
5076 | continue; |
5077 | |
5078 | if (!filled_vec) |
5079 | { |
5080 | clone_info *info = clone_info::get (node); |
5081 | if (!info || !info->param_adjustments) |
5082 | return true; |
5083 | info->param_adjustments->get_surviving_params (surviving_params: &surviving); |
5084 | filled_vec = true; |
5085 | } |
5086 | if (surviving.length() < (unsigned) i && surviving[i]) |
5087 | return true; |
5088 | } |
5089 | return false; |
5090 | } |
5091 | |
5092 | /* Create a specialized version of NODE with known constants in KNOWN_CSTS, |
5093 | known contexts in KNOWN_CONTEXTS and known aggregate values in AGGVALS and |
5094 | redirect all edges in CALLERS to it. */ |
5095 | |
5096 | static struct cgraph_node * |
5097 | create_specialized_node (struct cgraph_node *node, |
5098 | vec<tree> known_csts, |
5099 | vec<ipa_polymorphic_call_context> known_contexts, |
5100 | vec<ipa_argagg_value, va_gc> *aggvals, |
5101 | vec<cgraph_edge *> &callers) |
5102 | { |
5103 | ipa_node_params *new_info, *info = ipa_node_params_sum->get (node); |
5104 | vec<ipa_replace_map *, va_gc> *replace_trees = NULL; |
5105 | vec<ipa_adjusted_param, va_gc> *new_params = NULL; |
5106 | struct cgraph_node *new_node; |
5107 | int i, count = ipa_get_param_count (info); |
5108 | clone_info *cinfo = clone_info::get (node); |
5109 | ipa_param_adjustments *old_adjustments = cinfo |
5110 | ? cinfo->param_adjustments : NULL; |
5111 | ipa_param_adjustments *new_adjustments; |
5112 | gcc_assert (!info->ipcp_orig_node); |
5113 | gcc_assert (node->can_change_signature |
5114 | || !old_adjustments); |
5115 | |
5116 | if (old_adjustments) |
5117 | { |
5118 | /* At the moment all IPA optimizations should use the number of |
5119 | parameters of the prevailing decl as the m_always_copy_start. |
5120 | Handling any other value would complicate the code below, so for the |
5121 | time bing let's only assert it is so. */ |
5122 | gcc_assert (old_adjustments->m_always_copy_start == count |
5123 | || old_adjustments->m_always_copy_start < 0); |
5124 | int old_adj_count = vec_safe_length (v: old_adjustments->m_adj_params); |
5125 | for (i = 0; i < old_adj_count; i++) |
5126 | { |
5127 | ipa_adjusted_param *old_adj = &(*old_adjustments->m_adj_params)[i]; |
5128 | if (!node->can_change_signature |
5129 | || old_adj->op != IPA_PARAM_OP_COPY |
5130 | || (!known_csts[old_adj->base_index] |
5131 | && ipa_is_param_used (info, i: old_adj->base_index))) |
5132 | { |
5133 | ipa_adjusted_param new_adj = *old_adj; |
5134 | |
5135 | new_adj.prev_clone_adjustment = true; |
5136 | new_adj.prev_clone_index = i; |
5137 | vec_safe_push (v&: new_params, obj: new_adj); |
5138 | } |
5139 | } |
5140 | bool skip_return = old_adjustments->m_skip_return; |
5141 | new_adjustments = (new (ggc_alloc <ipa_param_adjustments> ()) |
5142 | ipa_param_adjustments (new_params, count, |
5143 | skip_return)); |
5144 | } |
5145 | else if (node->can_change_signature |
5146 | && want_remove_some_param_p (node, known_csts)) |
5147 | { |
5148 | ipa_adjusted_param adj; |
5149 | memset (s: &adj, c: 0, n: sizeof (adj)); |
5150 | adj.op = IPA_PARAM_OP_COPY; |
5151 | for (i = 0; i < count; i++) |
5152 | if (!known_csts[i] && ipa_is_param_used (info, i)) |
5153 | { |
5154 | adj.base_index = i; |
5155 | adj.prev_clone_index = i; |
5156 | vec_safe_push (v&: new_params, obj: adj); |
5157 | } |
5158 | new_adjustments = (new (ggc_alloc <ipa_param_adjustments> ()) |
5159 | ipa_param_adjustments (new_params, count, false)); |
5160 | } |
5161 | else |
5162 | new_adjustments = NULL; |
5163 | |
5164 | auto_vec<cgraph_edge *, 2> self_recursive_calls; |
5165 | for (i = callers.length () - 1; i >= 0; i--) |
5166 | { |
5167 | cgraph_edge *cs = callers[i]; |
5168 | if (cs->caller == node) |
5169 | { |
5170 | self_recursive_calls.safe_push (obj: cs); |
5171 | callers.unordered_remove (ix: i); |
5172 | } |
5173 | } |
5174 | replace_trees = cinfo ? vec_safe_copy (src: cinfo->tree_map) : NULL; |
5175 | for (i = 0; i < count; i++) |
5176 | { |
5177 | tree t = known_csts[i]; |
5178 | if (!t) |
5179 | continue; |
5180 | |
5181 | gcc_checking_assert (TREE_CODE (t) != TREE_BINFO); |
5182 | |
5183 | bool load_ref = false; |
5184 | symtab_node *ref_symbol; |
5185 | if (TREE_CODE (t) == ADDR_EXPR) |
5186 | { |
5187 | tree base = get_base_address (TREE_OPERAND (t, 0)); |
5188 | if (TREE_CODE (base) == VAR_DECL |
5189 | && ipa_get_controlled_uses (info, i) == 0 |
5190 | && ipa_get_param_load_dereferenced (info, i) |
5191 | && (ref_symbol = symtab_node::get (decl: base))) |
5192 | { |
5193 | load_ref = true; |
5194 | if (node->can_change_signature) |
5195 | for (cgraph_edge *caller : callers) |
5196 | adjust_references_in_caller (cs: caller, symbol: ref_symbol, index: i); |
5197 | } |
5198 | } |
5199 | |
5200 | ipa_replace_map *replace_map = get_replacement_map (info, value: t, parm_num: i, force_load_ref: load_ref); |
5201 | if (replace_map) |
5202 | vec_safe_push (v&: replace_trees, obj: replace_map); |
5203 | } |
5204 | |
5205 | unsigned &suffix_counter = clone_num_suffixes->get_or_insert ( |
5206 | IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME ( |
5207 | node->decl))); |
5208 | new_node = node->create_virtual_clone (redirect_callers: callers, tree_map: replace_trees, |
5209 | param_adjustments: new_adjustments, suffix: "constprop" , |
5210 | num_suffix: suffix_counter); |
5211 | suffix_counter++; |
5212 | |
5213 | bool have_self_recursive_calls = !self_recursive_calls.is_empty (); |
5214 | for (unsigned j = 0; j < self_recursive_calls.length (); j++) |
5215 | { |
5216 | cgraph_edge *cs = get_next_cgraph_edge_clone (cs: self_recursive_calls[j]); |
5217 | /* Cloned edges can disappear during cloning as speculation can be |
5218 | resolved, check that we have one and that it comes from the last |
5219 | cloning. */ |
5220 | if (cs && cs->caller == new_node) |
5221 | cs->redirect_callee_duplicating_thunks (n: new_node); |
5222 | /* Any future code that would make more than one clone of an outgoing |
5223 | edge would confuse this mechanism, so let's check that does not |
5224 | happen. */ |
5225 | gcc_checking_assert (!cs |
5226 | || !get_next_cgraph_edge_clone (cs) |
5227 | || get_next_cgraph_edge_clone (cs)->caller != new_node); |
5228 | } |
5229 | if (have_self_recursive_calls) |
5230 | new_node->expand_all_artificial_thunks (); |
5231 | |
5232 | ipa_set_node_agg_value_chain (node: new_node, aggs: aggvals); |
5233 | for (const ipa_argagg_value &av : aggvals) |
5234 | new_node->maybe_create_reference (val: av.value, NULL); |
5235 | |
5236 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5237 | { |
5238 | fprintf (stream: dump_file, format: " the new node is %s.\n" , new_node->dump_name ()); |
5239 | if (known_contexts.exists ()) |
5240 | { |
5241 | for (i = 0; i < count; i++) |
5242 | if (!known_contexts[i].useless_p ()) |
5243 | { |
5244 | fprintf (stream: dump_file, format: " known ctx %i is " , i); |
5245 | known_contexts[i].dump (f: dump_file); |
5246 | } |
5247 | } |
5248 | if (aggvals) |
5249 | { |
5250 | fprintf (stream: dump_file, format: " Aggregate replacements:" ); |
5251 | ipa_argagg_value_list avs (aggvals); |
5252 | avs.dump (f: dump_file); |
5253 | } |
5254 | } |
5255 | |
5256 | new_info = ipa_node_params_sum->get (node: new_node); |
5257 | new_info->ipcp_orig_node = node; |
5258 | new_node->ipcp_clone = true; |
5259 | new_info->known_csts = known_csts; |
5260 | new_info->known_contexts = known_contexts; |
5261 | |
5262 | ipcp_discover_new_direct_edges (node: new_node, known_csts, known_contexts, |
5263 | aggvals); |
5264 | |
5265 | return new_node; |
5266 | } |
5267 | |
5268 | /* Return true if JFUNC, which describes a i-th parameter of call CS, is a |
5269 | pass-through function to itself when the cgraph_node involved is not an |
5270 | IPA-CP clone. When SIMPLE is true, further check if JFUNC is a simple |
5271 | no-operation pass-through. */ |
5272 | |
5273 | static bool |
5274 | self_recursive_pass_through_p (cgraph_edge *cs, ipa_jump_func *jfunc, int i, |
5275 | bool simple = true) |
5276 | { |
5277 | enum availability availability; |
5278 | if (cs->caller == cs->callee->function_symbol (avail: &availability) |
5279 | && availability > AVAIL_INTERPOSABLE |
5280 | && jfunc->type == IPA_JF_PASS_THROUGH |
5281 | && (!simple || ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR) |
5282 | && ipa_get_jf_pass_through_formal_id (jfunc) == i |
5283 | && ipa_node_params_sum->get (node: cs->caller) |
5284 | && !ipa_node_params_sum->get (node: cs->caller)->ipcp_orig_node) |
5285 | return true; |
5286 | return false; |
5287 | } |
5288 | |
5289 | /* Return true if JFUNC, which describes a part of an aggregate represented or |
5290 | pointed to by the i-th parameter of call CS, is a pass-through function to |
5291 | itself when the cgraph_node involved is not an IPA-CP clone.. When |
5292 | SIMPLE is true, further check if JFUNC is a simple no-operation |
5293 | pass-through. */ |
5294 | |
5295 | static bool |
5296 | self_recursive_agg_pass_through_p (const cgraph_edge *cs, |
5297 | const ipa_agg_jf_item *jfunc, |
5298 | int i, bool simple = true) |
5299 | { |
5300 | enum availability availability; |
5301 | if (cs->caller == cs->callee->function_symbol (avail: &availability) |
5302 | && availability > AVAIL_INTERPOSABLE |
5303 | && jfunc->jftype == IPA_JF_LOAD_AGG |
5304 | && jfunc->offset == jfunc->value.load_agg.offset |
5305 | && (!simple || jfunc->value.pass_through.operation == NOP_EXPR) |
5306 | && jfunc->value.pass_through.formal_id == i |
5307 | && useless_type_conversion_p (jfunc->value.load_agg.type, jfunc->type) |
5308 | && ipa_node_params_sum->get (node: cs->caller) |
5309 | && !ipa_node_params_sum->get (node: cs->caller)->ipcp_orig_node) |
5310 | return true; |
5311 | return false; |
5312 | } |
5313 | |
5314 | /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in |
5315 | KNOWN_CSTS with constants that are also known for all of the CALLERS. */ |
5316 | |
5317 | static void |
5318 | find_more_scalar_values_for_callers_subset (struct cgraph_node *node, |
5319 | vec<tree> &known_csts, |
5320 | const vec<cgraph_edge *> &callers) |
5321 | { |
5322 | ipa_node_params *info = ipa_node_params_sum->get (node); |
5323 | int i, count = ipa_get_param_count (info); |
5324 | |
5325 | for (i = 0; i < count; i++) |
5326 | { |
5327 | struct cgraph_edge *cs; |
5328 | tree newval = NULL_TREE; |
5329 | int j; |
5330 | bool first = true; |
5331 | tree type = ipa_get_type (info, i); |
5332 | |
5333 | if (ipa_get_scalar_lat (info, i)->bottom || known_csts[i]) |
5334 | continue; |
5335 | |
5336 | FOR_EACH_VEC_ELT (callers, j, cs) |
5337 | { |
5338 | struct ipa_jump_func *jump_func; |
5339 | tree t; |
5340 | |
5341 | ipa_edge_args *args = ipa_edge_args_sum->get (edge: cs); |
5342 | if (!args |
5343 | || i >= ipa_get_cs_argument_count (args) |
5344 | || (i == 0 |
5345 | && call_passes_through_thunk (cs))) |
5346 | { |
5347 | newval = NULL_TREE; |
5348 | break; |
5349 | } |
5350 | jump_func = ipa_get_ith_jump_func (args, i); |
5351 | |
5352 | /* Besides simple pass-through jump function, arithmetic jump |
5353 | function could also introduce argument-direct-pass-through for |
5354 | self-feeding recursive call. For example, |
5355 | |
5356 | fn (int i) |
5357 | { |
5358 | fn (i & 1); |
5359 | } |
5360 | |
5361 | Given that i is 0, recursive propagation via (i & 1) also gets |
5362 | 0. */ |
5363 | if (self_recursive_pass_through_p (cs, jfunc: jump_func, i, simple: false)) |
5364 | { |
5365 | gcc_assert (newval); |
5366 | t = ipa_get_jf_arith_result ( |
5367 | opcode: ipa_get_jf_pass_through_operation (jfunc: jump_func), |
5368 | input: newval, |
5369 | operand: ipa_get_jf_pass_through_operand (jfunc: jump_func), |
5370 | res_type: type); |
5371 | } |
5372 | else |
5373 | t = ipa_value_from_jfunc (info: ipa_node_params_sum->get (node: cs->caller), |
5374 | jfunc: jump_func, parm_type: type); |
5375 | if (!t |
5376 | || (newval |
5377 | && !values_equal_for_ipcp_p (x: t, y: newval)) |
5378 | || (!first && !newval)) |
5379 | { |
5380 | newval = NULL_TREE; |
5381 | break; |
5382 | } |
5383 | else |
5384 | newval = t; |
5385 | first = false; |
5386 | } |
5387 | |
5388 | if (newval) |
5389 | { |
5390 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5391 | { |
5392 | fprintf (stream: dump_file, format: " adding an extra known scalar value " ); |
5393 | print_ipcp_constant_value (f: dump_file, v: newval); |
5394 | fprintf (stream: dump_file, format: " for " ); |
5395 | ipa_dump_param (dump_file, info, i); |
5396 | fprintf (stream: dump_file, format: "\n" ); |
5397 | } |
5398 | |
5399 | known_csts[i] = newval; |
5400 | } |
5401 | } |
5402 | } |
5403 | |
5404 | /* Given a NODE and a subset of its CALLERS, try to populate plank slots in |
5405 | KNOWN_CONTEXTS with polymorphic contexts that are also known for all of the |
5406 | CALLERS. */ |
5407 | |
5408 | static void |
5409 | find_more_contexts_for_caller_subset (cgraph_node *node, |
5410 | vec<ipa_polymorphic_call_context> |
5411 | *known_contexts, |
5412 | const vec<cgraph_edge *> &callers) |
5413 | { |
5414 | ipa_node_params *info = ipa_node_params_sum->get (node); |
5415 | int i, count = ipa_get_param_count (info); |
5416 | |
5417 | for (i = 0; i < count; i++) |
5418 | { |
5419 | cgraph_edge *cs; |
5420 | |
5421 | if (ipa_get_poly_ctx_lat (info, i)->bottom |
5422 | || (known_contexts->exists () |
5423 | && !(*known_contexts)[i].useless_p ())) |
5424 | continue; |
5425 | |
5426 | ipa_polymorphic_call_context newval; |
5427 | bool first = true; |
5428 | int j; |
5429 | |
5430 | FOR_EACH_VEC_ELT (callers, j, cs) |
5431 | { |
5432 | ipa_edge_args *args = ipa_edge_args_sum->get (edge: cs); |
5433 | if (!args |
5434 | || i >= ipa_get_cs_argument_count (args)) |
5435 | return; |
5436 | ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i); |
5437 | ipa_polymorphic_call_context ctx; |
5438 | ctx = ipa_context_from_jfunc (info: ipa_node_params_sum->get (node: cs->caller), |
5439 | cs, csidx: i, jfunc); |
5440 | if (first) |
5441 | { |
5442 | newval = ctx; |
5443 | first = false; |
5444 | } |
5445 | else |
5446 | newval.meet_with (ctx); |
5447 | if (newval.useless_p ()) |
5448 | break; |
5449 | } |
5450 | |
5451 | if (!newval.useless_p ()) |
5452 | { |
5453 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5454 | { |
5455 | fprintf (stream: dump_file, format: " adding an extra known polymorphic " |
5456 | "context " ); |
5457 | print_ipcp_constant_value (f: dump_file, v: newval); |
5458 | fprintf (stream: dump_file, format: " for " ); |
5459 | ipa_dump_param (dump_file, info, i); |
5460 | fprintf (stream: dump_file, format: "\n" ); |
5461 | } |
5462 | |
5463 | if (!known_contexts->exists ()) |
5464 | known_contexts->safe_grow_cleared (len: ipa_get_param_count (info), |
5465 | exact: true); |
5466 | (*known_contexts)[i] = newval; |
5467 | } |
5468 | |
5469 | } |
5470 | } |
5471 | |
5472 | /* Push all aggregate values coming along edge CS for parameter number INDEX to |
5473 | RES. If INTERIM is non-NULL, it contains the current interim state of |
5474 | collected aggregate values which can be used to compute values passed over |
5475 | self-recursive edges. |
5476 | |
5477 | This basically one iteration of push_agg_values_from_edge over one |
5478 | parameter, which allows for simpler early returns. */ |
5479 | |
5480 | static void |
5481 | push_agg_values_for_index_from_edge (struct cgraph_edge *cs, int index, |
5482 | vec<ipa_argagg_value> *res, |
5483 | const ipa_argagg_value_list *interim) |
5484 | { |
5485 | bool agg_values_from_caller = false; |
5486 | bool agg_jf_preserved = false; |
5487 | unsigned unit_delta = UINT_MAX; |
5488 | int src_idx = -1; |
5489 | ipa_jump_func *jfunc = ipa_get_ith_jump_func (args: ipa_edge_args_sum->get (edge: cs), |
5490 | i: index); |
5491 | |
5492 | if (jfunc->type == IPA_JF_PASS_THROUGH |
5493 | && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR) |
5494 | { |
5495 | agg_values_from_caller = true; |
5496 | agg_jf_preserved = ipa_get_jf_pass_through_agg_preserved (jfunc); |
5497 | src_idx = ipa_get_jf_pass_through_formal_id (jfunc); |
5498 | unit_delta = 0; |
5499 | } |
5500 | else if (jfunc->type == IPA_JF_ANCESTOR |
5501 | && ipa_get_jf_ancestor_agg_preserved (jfunc)) |
5502 | { |
5503 | agg_values_from_caller = true; |
5504 | agg_jf_preserved = true; |
5505 | src_idx = ipa_get_jf_ancestor_formal_id (jfunc); |
5506 | unit_delta = ipa_get_jf_ancestor_offset (jfunc) / BITS_PER_UNIT; |
5507 | } |
5508 | |
5509 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: cs->caller); |
5510 | if (agg_values_from_caller) |
5511 | { |
5512 | if (caller_info->ipcp_orig_node) |
5513 | { |
5514 | struct cgraph_node *orig_node = caller_info->ipcp_orig_node; |
5515 | ipcp_transformation *ts |
5516 | = ipcp_get_transformation_summary (node: cs->caller); |
5517 | ipa_node_params *orig_info = ipa_node_params_sum->get (node: orig_node); |
5518 | ipcp_param_lattices *orig_plats |
5519 | = ipa_get_parm_lattices (info: orig_info, i: src_idx); |
5520 | if (ts |
5521 | && orig_plats->aggs |
5522 | && (agg_jf_preserved || !orig_plats->aggs_by_ref)) |
5523 | { |
5524 | ipa_argagg_value_list src (ts); |
5525 | src.push_adjusted_values (src_index: src_idx, dest_index: index, unit_delta, res); |
5526 | return; |
5527 | } |
5528 | } |
5529 | else |
5530 | { |
5531 | ipcp_param_lattices *src_plats |
5532 | = ipa_get_parm_lattices (info: caller_info, i: src_idx); |
5533 | if (src_plats->aggs |
5534 | && !src_plats->aggs_bottom |
5535 | && (agg_jf_preserved || !src_plats->aggs_by_ref)) |
5536 | { |
5537 | if (interim && self_recursive_pass_through_p (cs, jfunc, i: index)) |
5538 | { |
5539 | interim->push_adjusted_values (src_index: src_idx, dest_index: index, unit_delta, |
5540 | res); |
5541 | return; |
5542 | } |
5543 | if (!src_plats->aggs_contain_variable) |
5544 | { |
5545 | push_agg_values_from_plats (plats: src_plats, dest_index: index, unit_delta, |
5546 | res); |
5547 | return; |
5548 | } |
5549 | } |
5550 | } |
5551 | } |
5552 | |
5553 | if (!jfunc->agg.items) |
5554 | return; |
5555 | bool first = true; |
5556 | unsigned prev_unit_offset = 0; |
5557 | for (const ipa_agg_jf_item &agg_jf : *jfunc->agg.items) |
5558 | { |
5559 | tree value, srcvalue; |
5560 | /* Besides simple pass-through aggregate jump function, arithmetic |
5561 | aggregate jump function could also bring same aggregate value as |
5562 | parameter passed-in for self-feeding recursive call. For example, |
5563 | |
5564 | fn (int *i) |
5565 | { |
5566 | int j = *i & 1; |
5567 | fn (&j); |
5568 | } |
5569 | |
5570 | Given that *i is 0, recursive propagation via (*i & 1) also gets 0. */ |
5571 | if (interim |
5572 | && self_recursive_agg_pass_through_p (cs, jfunc: &agg_jf, i: index, simple: false) |
5573 | && (srcvalue = interim->get_value(index, |
5574 | unit_offset: agg_jf.offset / BITS_PER_UNIT))) |
5575 | value = ipa_get_jf_arith_result (opcode: agg_jf.value.pass_through.operation, |
5576 | input: srcvalue, |
5577 | operand: agg_jf.value.pass_through.operand, |
5578 | res_type: agg_jf.type); |
5579 | else |
5580 | value = ipa_agg_value_from_jfunc (info: caller_info, node: cs->caller, |
5581 | item: &agg_jf); |
5582 | if (value) |
5583 | { |
5584 | struct ipa_argagg_value iav; |
5585 | iav.value = value; |
5586 | iav.unit_offset = agg_jf.offset / BITS_PER_UNIT; |
5587 | iav.index = index; |
5588 | iav.by_ref = jfunc->agg.by_ref; |
5589 | iav.killed = false; |
5590 | |
5591 | gcc_assert (first |
5592 | || iav.unit_offset > prev_unit_offset); |
5593 | prev_unit_offset = iav.unit_offset; |
5594 | first = false; |
5595 | |
5596 | res->safe_push (obj: iav); |
5597 | } |
5598 | } |
5599 | return; |
5600 | } |
5601 | |
5602 | /* Push all aggregate values coming along edge CS to RES. DEST_INFO is the |
5603 | description of ultimate callee of CS or the one it was cloned from (the |
5604 | summary where lattices are). If INTERIM is non-NULL, it contains the |
5605 | current interim state of collected aggregate values which can be used to |
5606 | compute values passed over self-recursive edges (if OPTIMIZE_SELF_RECURSION |
5607 | is true) and to skip values which clearly will not be part of intersection |
5608 | with INTERIM. */ |
5609 | |
5610 | static void |
5611 | push_agg_values_from_edge (struct cgraph_edge *cs, |
5612 | ipa_node_params *dest_info, |
5613 | vec<ipa_argagg_value> *res, |
5614 | const ipa_argagg_value_list *interim, |
5615 | bool optimize_self_recursion) |
5616 | { |
5617 | ipa_edge_args *args = ipa_edge_args_sum->get (edge: cs); |
5618 | if (!args) |
5619 | return; |
5620 | |
5621 | int count = MIN (ipa_get_param_count (dest_info), |
5622 | ipa_get_cs_argument_count (args)); |
5623 | |
5624 | unsigned interim_index = 0; |
5625 | for (int index = 0; index < count; index++) |
5626 | { |
5627 | if (interim) |
5628 | { |
5629 | while (interim_index < interim->m_elts.size () |
5630 | && interim->m_elts[interim_index].value |
5631 | && interim->m_elts[interim_index].index < index) |
5632 | interim_index++; |
5633 | if (interim_index >= interim->m_elts.size () |
5634 | || interim->m_elts[interim_index].index > index) |
5635 | continue; |
5636 | } |
5637 | |
5638 | ipcp_param_lattices *plats = ipa_get_parm_lattices (info: dest_info, i: index); |
5639 | if (!ipa_is_param_used (info: dest_info, i: index) |
5640 | || plats->aggs_bottom) |
5641 | continue; |
5642 | push_agg_values_for_index_from_edge (cs, index, res, |
5643 | interim: optimize_self_recursion ? interim |
5644 | : NULL); |
5645 | } |
5646 | } |
5647 | |
5648 | |
5649 | /* Look at edges in CALLERS and collect all known aggregate values that arrive |
5650 | from all of them. Return nullptr if there are none. */ |
5651 | |
5652 | static struct vec<ipa_argagg_value, va_gc> * |
5653 | find_aggregate_values_for_callers_subset (struct cgraph_node *node, |
5654 | const vec<cgraph_edge *> &callers) |
5655 | { |
5656 | ipa_node_params *dest_info = ipa_node_params_sum->get (node); |
5657 | if (dest_info->ipcp_orig_node) |
5658 | dest_info = ipa_node_params_sum->get (node: dest_info->ipcp_orig_node); |
5659 | |
5660 | /* gather_edges_for_value puts a non-recursive call into the first element of |
5661 | callers if it can. */ |
5662 | auto_vec<ipa_argagg_value, 32> interim; |
5663 | push_agg_values_from_edge (cs: callers[0], dest_info, res: &interim, NULL, optimize_self_recursion: true); |
5664 | |
5665 | unsigned valid_entries = interim.length (); |
5666 | if (!valid_entries) |
5667 | return nullptr; |
5668 | |
5669 | unsigned caller_count = callers.length(); |
5670 | for (unsigned i = 1; i < caller_count; i++) |
5671 | { |
5672 | auto_vec<ipa_argagg_value, 32> last; |
5673 | ipa_argagg_value_list avs (&interim); |
5674 | push_agg_values_from_edge (cs: callers[i], dest_info, res: &last, interim: &avs, optimize_self_recursion: true); |
5675 | |
5676 | valid_entries = intersect_argaggs_with (elts&: interim, other: last); |
5677 | if (!valid_entries) |
5678 | return nullptr; |
5679 | } |
5680 | |
5681 | vec<ipa_argagg_value, va_gc> *res = NULL; |
5682 | vec_safe_reserve_exact (v&: res, nelems: valid_entries); |
5683 | for (const ipa_argagg_value &av : interim) |
5684 | if (av.value) |
5685 | res->quick_push(obj: av); |
5686 | gcc_checking_assert (res->length () == valid_entries); |
5687 | return res; |
5688 | } |
5689 | |
5690 | /* Determine whether CS also brings all scalar values that the NODE is |
5691 | specialized for. */ |
5692 | |
5693 | static bool |
5694 | cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs, |
5695 | struct cgraph_node *node) |
5696 | { |
5697 | ipa_node_params *dest_info = ipa_node_params_sum->get (node); |
5698 | int count = ipa_get_param_count (info: dest_info); |
5699 | class ipa_node_params *caller_info; |
5700 | class ipa_edge_args *args; |
5701 | int i; |
5702 | |
5703 | caller_info = ipa_node_params_sum->get (node: cs->caller); |
5704 | args = ipa_edge_args_sum->get (edge: cs); |
5705 | for (i = 0; i < count; i++) |
5706 | { |
5707 | struct ipa_jump_func *jump_func; |
5708 | tree val, t; |
5709 | |
5710 | val = dest_info->known_csts[i]; |
5711 | if (!val) |
5712 | continue; |
5713 | |
5714 | if (i >= ipa_get_cs_argument_count (args)) |
5715 | return false; |
5716 | jump_func = ipa_get_ith_jump_func (args, i); |
5717 | t = ipa_value_from_jfunc (info: caller_info, jfunc: jump_func, |
5718 | parm_type: ipa_get_type (info: dest_info, i)); |
5719 | if (!t || !values_equal_for_ipcp_p (x: val, y: t)) |
5720 | return false; |
5721 | } |
5722 | return true; |
5723 | } |
5724 | |
5725 | /* Determine whether CS also brings all aggregate values that NODE is |
5726 | specialized for. */ |
5727 | |
5728 | static bool |
5729 | cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs, |
5730 | struct cgraph_node *node) |
5731 | { |
5732 | ipcp_transformation *ts = ipcp_get_transformation_summary (node); |
5733 | if (!ts || vec_safe_is_empty (v: ts->m_agg_values)) |
5734 | return true; |
5735 | |
5736 | const ipa_argagg_value_list existing (ts->m_agg_values); |
5737 | auto_vec<ipa_argagg_value, 32> edge_values; |
5738 | ipa_node_params *dest_info = ipa_node_params_sum->get (node); |
5739 | gcc_checking_assert (dest_info->ipcp_orig_node); |
5740 | dest_info = ipa_node_params_sum->get (node: dest_info->ipcp_orig_node); |
5741 | push_agg_values_from_edge (cs, dest_info, res: &edge_values, interim: &existing, optimize_self_recursion: false); |
5742 | const ipa_argagg_value_list avl (&edge_values); |
5743 | return avl.superset_of_p (other: existing); |
5744 | } |
5745 | |
5746 | /* Given an original NODE and a VAL for which we have already created a |
5747 | specialized clone, look whether there are incoming edges that still lead |
5748 | into the old node but now also bring the requested value and also conform to |
5749 | all other criteria such that they can be redirected the special node. |
5750 | This function can therefore redirect the final edge in a SCC. */ |
5751 | |
5752 | template <typename valtype> |
5753 | static void |
5754 | perhaps_add_new_callers (cgraph_node *node, ipcp_value<valtype> *val) |
5755 | { |
5756 | ipcp_value_source<valtype> *src; |
5757 | profile_count redirected_sum = profile_count::zero (); |
5758 | |
5759 | for (src = val->sources; src; src = src->next) |
5760 | { |
5761 | struct cgraph_edge *cs = src->cs; |
5762 | while (cs) |
5763 | { |
5764 | if (cgraph_edge_brings_value_p (cs, src, node, val) |
5765 | && cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node) |
5766 | && cgraph_edge_brings_all_agg_vals_for_node (cs, val->spec_node)) |
5767 | { |
5768 | if (dump_file) |
5769 | fprintf (dump_file, " - adding an extra caller %s of %s\n" , |
5770 | cs->caller->dump_name (), |
5771 | val->spec_node->dump_name ()); |
5772 | |
5773 | cs->redirect_callee_duplicating_thunks (n: val->spec_node); |
5774 | val->spec_node->expand_all_artificial_thunks (); |
5775 | if (cs->count.ipa ().initialized_p ()) |
5776 | redirected_sum = redirected_sum + cs->count.ipa (); |
5777 | } |
5778 | cs = get_next_cgraph_edge_clone (cs); |
5779 | } |
5780 | } |
5781 | |
5782 | if (redirected_sum.nonzero_p ()) |
5783 | update_specialized_profile (val->spec_node, node, redirected_sum); |
5784 | } |
5785 | |
5786 | /* Return true if KNOWN_CONTEXTS contain at least one useful context. */ |
5787 | |
5788 | static bool |
5789 | known_contexts_useful_p (vec<ipa_polymorphic_call_context> known_contexts) |
5790 | { |
5791 | ipa_polymorphic_call_context *ctx; |
5792 | int i; |
5793 | |
5794 | FOR_EACH_VEC_ELT (known_contexts, i, ctx) |
5795 | if (!ctx->useless_p ()) |
5796 | return true; |
5797 | return false; |
5798 | } |
5799 | |
5800 | /* Return a copy of KNOWN_CSTS if it is not empty, otherwise return vNULL. */ |
5801 | |
5802 | static vec<ipa_polymorphic_call_context> |
5803 | copy_useful_known_contexts (const vec<ipa_polymorphic_call_context> &known_contexts) |
5804 | { |
5805 | if (known_contexts_useful_p (known_contexts)) |
5806 | return known_contexts.copy (); |
5807 | else |
5808 | return vNULL; |
5809 | } |
5810 | |
5811 | /* Copy known scalar values from AVALS into KNOWN_CSTS and modify the copy |
5812 | according to VAL and INDEX. If non-empty, replace KNOWN_CONTEXTS with its |
5813 | copy too. */ |
5814 | |
5815 | static void |
5816 | copy_known_vectors_add_val (ipa_auto_call_arg_values *avals, |
5817 | vec<tree> *known_csts, |
5818 | vec<ipa_polymorphic_call_context> *known_contexts, |
5819 | ipcp_value<tree> *val, int index) |
5820 | { |
5821 | *known_csts = avals->m_known_vals.copy (); |
5822 | *known_contexts = copy_useful_known_contexts (known_contexts: avals->m_known_contexts); |
5823 | (*known_csts)[index] = val->value; |
5824 | } |
5825 | |
5826 | /* Copy known scalar values from AVALS into KNOWN_CSTS. Similarly, copy |
5827 | contexts to KNOWN_CONTEXTS and modify the copy according to VAL and |
5828 | INDEX. */ |
5829 | |
5830 | static void |
5831 | copy_known_vectors_add_val (ipa_auto_call_arg_values *avals, |
5832 | vec<tree> *known_csts, |
5833 | vec<ipa_polymorphic_call_context> *known_contexts, |
5834 | ipcp_value<ipa_polymorphic_call_context> *val, |
5835 | int index) |
5836 | { |
5837 | *known_csts = avals->m_known_vals.copy (); |
5838 | *known_contexts = avals->m_known_contexts.copy (); |
5839 | (*known_contexts)[index] = val->value; |
5840 | } |
5841 | |
5842 | /* Return true if OFFSET indicates this was not an aggregate value or there is |
5843 | a replacement equivalent to VALUE, INDEX and OFFSET among those in the |
5844 | AGGVALS list. */ |
5845 | |
5846 | DEBUG_FUNCTION bool |
5847 | ipcp_val_agg_replacement_ok_p (vec<ipa_argagg_value, va_gc> *aggvals, |
5848 | int index, HOST_WIDE_INT offset, tree value) |
5849 | { |
5850 | if (offset == -1) |
5851 | return true; |
5852 | |
5853 | const ipa_argagg_value_list avl (aggvals); |
5854 | tree v = avl.get_value (index, unit_offset: offset / BITS_PER_UNIT); |
5855 | return v && values_equal_for_ipcp_p (x: v, y: value); |
5856 | } |
5857 | |
5858 | /* Return true if offset is minus one because source of a polymorphic context |
5859 | cannot be an aggregate value. */ |
5860 | |
5861 | DEBUG_FUNCTION bool |
5862 | ipcp_val_agg_replacement_ok_p (vec<ipa_argagg_value, va_gc> *, |
5863 | int , HOST_WIDE_INT offset, |
5864 | ipa_polymorphic_call_context) |
5865 | { |
5866 | return offset == -1; |
5867 | } |
5868 | |
5869 | /* Decide whether to create a special version of NODE for value VAL of |
5870 | parameter at the given INDEX. If OFFSET is -1, the value is for the |
5871 | parameter itself, otherwise it is stored at the given OFFSET of the |
5872 | parameter. AVALS describes the other already known values. SELF_GEN_CLONES |
5873 | is a vector which contains clones created for self-recursive calls with an |
5874 | arithmetic pass-through jump function. */ |
5875 | |
5876 | template <typename valtype> |
5877 | static bool |
5878 | decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset, |
5879 | ipcp_value<valtype> *val, ipa_auto_call_arg_values *avals, |
5880 | vec<cgraph_node *> *self_gen_clones) |
5881 | { |
5882 | int caller_count; |
5883 | sreal freq_sum; |
5884 | profile_count count_sum, rec_count_sum; |
5885 | vec<cgraph_edge *> callers; |
5886 | |
5887 | if (val->spec_node) |
5888 | { |
5889 | perhaps_add_new_callers (node, val); |
5890 | return false; |
5891 | } |
5892 | else if (val->local_size_cost + overall_size > get_max_overall_size (node)) |
5893 | { |
5894 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5895 | fprintf (dump_file, " Ignoring candidate value because " |
5896 | "maximum unit size would be reached with %li.\n" , |
5897 | val->local_size_cost + overall_size); |
5898 | return false; |
5899 | } |
5900 | else if (!get_info_about_necessary_edges (val, node, &freq_sum, &caller_count, |
5901 | &rec_count_sum, &count_sum)) |
5902 | return false; |
5903 | |
5904 | if (!dbg_cnt (index: ipa_cp_values)) |
5905 | return false; |
5906 | |
5907 | if (val->self_recursion_generated_p ()) |
5908 | { |
5909 | /* The edge counts in this case might not have been adjusted yet. |
5910 | Nevertleless, even if they were it would be only a guesswork which we |
5911 | can do now. The recursive part of the counts can be derived from the |
5912 | count of the original node anyway. */ |
5913 | if (node->count.ipa ().nonzero_p ()) |
5914 | { |
5915 | unsigned dem = self_gen_clones->length () + 1; |
5916 | rec_count_sum = node->count.ipa () / dem; |
5917 | } |
5918 | else |
5919 | rec_count_sum = profile_count::zero (); |
5920 | } |
5921 | |
5922 | /* get_info_about_necessary_edges only sums up ipa counts. */ |
5923 | count_sum += rec_count_sum; |
5924 | |
5925 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5926 | { |
5927 | fprintf (stream: dump_file, format: " - considering value " ); |
5928 | print_ipcp_constant_value (dump_file, val->value); |
5929 | fprintf (stream: dump_file, format: " for " ); |
5930 | ipa_dump_param (dump_file, info: ipa_node_params_sum->get (node), i: index); |
5931 | if (offset != -1) |
5932 | fprintf (stream: dump_file, format: ", offset: " HOST_WIDE_INT_PRINT_DEC, offset); |
5933 | fprintf (stream: dump_file, format: " (caller_count: %i)\n" , caller_count); |
5934 | } |
5935 | |
5936 | if (!good_cloning_opportunity_p (node, val->local_time_benefit, |
5937 | freq_sum, count_sum, |
5938 | val->local_size_cost) |
5939 | && !good_cloning_opportunity_p (node, val->prop_time_benefit, |
5940 | freq_sum, count_sum, val->prop_size_cost)) |
5941 | return false; |
5942 | |
5943 | if (dump_file) |
5944 | fprintf (stream: dump_file, format: " Creating a specialized node of %s.\n" , |
5945 | node->dump_name ()); |
5946 | |
5947 | vec<tree> known_csts; |
5948 | vec<ipa_polymorphic_call_context> known_contexts; |
5949 | |
5950 | callers = gather_edges_for_value (val, node, caller_count); |
5951 | if (offset == -1) |
5952 | copy_known_vectors_add_val (avals, &known_csts, &known_contexts, val, index); |
5953 | else |
5954 | { |
5955 | known_csts = avals->m_known_vals.copy (); |
5956 | known_contexts = copy_useful_known_contexts (known_contexts: avals->m_known_contexts); |
5957 | } |
5958 | find_more_scalar_values_for_callers_subset (node, known_csts, callers); |
5959 | find_more_contexts_for_caller_subset (node, known_contexts: &known_contexts, callers); |
5960 | vec<ipa_argagg_value, va_gc> *aggvals |
5961 | = find_aggregate_values_for_callers_subset (node, callers); |
5962 | gcc_checking_assert (ipcp_val_agg_replacement_ok_p (aggvals, index, |
5963 | offset, val->value)); |
5964 | val->spec_node = create_specialized_node (node, known_csts, known_contexts, |
5965 | aggvals, callers); |
5966 | |
5967 | if (val->self_recursion_generated_p ()) |
5968 | self_gen_clones->safe_push (obj: val->spec_node); |
5969 | else |
5970 | update_profiling_info (node, val->spec_node); |
5971 | |
5972 | callers.release (); |
5973 | overall_size += val->local_size_cost; |
5974 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5975 | fprintf (stream: dump_file, format: " overall size reached %li\n" , |
5976 | overall_size); |
5977 | |
5978 | /* TODO: If for some lattice there is only one other known value |
5979 | left, make a special node for it too. */ |
5980 | |
5981 | return true; |
5982 | } |
5983 | |
5984 | /* Like irange::contains_p(), but convert VAL to the range of R if |
5985 | necessary. */ |
5986 | |
5987 | static inline bool |
5988 | ipa_range_contains_p (const vrange &r, tree val) |
5989 | { |
5990 | if (r.undefined_p ()) |
5991 | return false; |
5992 | |
5993 | tree type = r.type (); |
5994 | if (!wi::fits_to_tree_p (x: wi::to_wide (t: val), type)) |
5995 | return false; |
5996 | |
5997 | val = fold_convert (type, val); |
5998 | return r.contains_p (cst: val); |
5999 | } |
6000 | |
6001 | /* Decide whether and what specialized clones of NODE should be created. */ |
6002 | |
6003 | static bool |
6004 | decide_whether_version_node (struct cgraph_node *node) |
6005 | { |
6006 | ipa_node_params *info = ipa_node_params_sum->get (node); |
6007 | int i, count = ipa_get_param_count (info); |
6008 | bool ret = false; |
6009 | |
6010 | if (count == 0) |
6011 | return false; |
6012 | |
6013 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6014 | fprintf (stream: dump_file, format: "\nEvaluating opportunities for %s.\n" , |
6015 | node->dump_name ()); |
6016 | |
6017 | auto_vec <cgraph_node *, 9> self_gen_clones; |
6018 | ipa_auto_call_arg_values avals; |
6019 | gather_context_independent_values (info, avals: &avals, calculate_aggs: false, NULL); |
6020 | |
6021 | for (i = 0; i < count;i++) |
6022 | { |
6023 | if (!ipa_is_param_used (info, i)) |
6024 | continue; |
6025 | |
6026 | class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i); |
6027 | ipcp_lattice<tree> *lat = &plats->itself; |
6028 | ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat; |
6029 | |
6030 | if (!lat->bottom |
6031 | && !avals.m_known_vals[i]) |
6032 | { |
6033 | ipcp_value<tree> *val; |
6034 | for (val = lat->values; val; val = val->next) |
6035 | { |
6036 | /* If some values generated for self-recursive calls with |
6037 | arithmetic jump functions fall outside of the known |
6038 | range for the parameter, we can skip them. */ |
6039 | if (TREE_CODE (val->value) == INTEGER_CST |
6040 | && !plats->m_value_range.bottom_p () |
6041 | && !ipa_range_contains_p (r: plats->m_value_range.m_vr, |
6042 | val: val->value)) |
6043 | { |
6044 | /* This can happen also if a constant present in the source |
6045 | code falls outside of the range of parameter's type, so we |
6046 | cannot assert. */ |
6047 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6048 | { |
6049 | fprintf (stream: dump_file, format: " - skipping%s value " , |
6050 | val->self_recursion_generated_p () |
6051 | ? " self_recursion_generated" : "" ); |
6052 | print_ipcp_constant_value (f: dump_file, v: val->value); |
6053 | fprintf (stream: dump_file, format: " because it is outside known " |
6054 | "value range.\n" ); |
6055 | } |
6056 | continue; |
6057 | } |
6058 | ret |= decide_about_value (node, index: i, offset: -1, val, avals: &avals, |
6059 | self_gen_clones: &self_gen_clones); |
6060 | } |
6061 | } |
6062 | |
6063 | if (!plats->aggs_bottom) |
6064 | { |
6065 | struct ipcp_agg_lattice *aglat; |
6066 | ipcp_value<tree> *val; |
6067 | for (aglat = plats->aggs; aglat; aglat = aglat->next) |
6068 | if (!aglat->bottom && aglat->values |
6069 | /* If the following is false, the one value has been considered |
6070 | for cloning for all contexts. */ |
6071 | && (plats->aggs_contain_variable |
6072 | || !aglat->is_single_const ())) |
6073 | for (val = aglat->values; val; val = val->next) |
6074 | ret |= decide_about_value (node, index: i, offset: aglat->offset, val, avals: &avals, |
6075 | self_gen_clones: &self_gen_clones); |
6076 | } |
6077 | |
6078 | if (!ctxlat->bottom |
6079 | && avals.m_known_contexts[i].useless_p ()) |
6080 | { |
6081 | ipcp_value<ipa_polymorphic_call_context> *val; |
6082 | for (val = ctxlat->values; val; val = val->next) |
6083 | ret |= decide_about_value (node, index: i, offset: -1, val, avals: &avals, |
6084 | self_gen_clones: &self_gen_clones); |
6085 | } |
6086 | } |
6087 | |
6088 | if (!self_gen_clones.is_empty ()) |
6089 | { |
6090 | self_gen_clones.safe_push (obj: node); |
6091 | update_counts_for_self_gen_clones (orig_node: node, self_gen_clones); |
6092 | } |
6093 | |
6094 | if (info->do_clone_for_all_contexts) |
6095 | { |
6096 | if (!dbg_cnt (index: ipa_cp_values)) |
6097 | { |
6098 | info->do_clone_for_all_contexts = false; |
6099 | return ret; |
6100 | } |
6101 | |
6102 | struct cgraph_node *clone; |
6103 | auto_vec<cgraph_edge *> callers = node->collect_callers (); |
6104 | |
6105 | for (int i = callers.length () - 1; i >= 0; i--) |
6106 | { |
6107 | cgraph_edge *cs = callers[i]; |
6108 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: cs->caller); |
6109 | |
6110 | if (caller_info && caller_info->node_dead) |
6111 | callers.unordered_remove (ix: i); |
6112 | } |
6113 | |
6114 | if (!adjust_callers_for_value_intersection (callers, node)) |
6115 | { |
6116 | /* If node is not called by anyone, or all its caller edges are |
6117 | self-recursive, the node is not really in use, no need to do |
6118 | cloning. */ |
6119 | info->do_clone_for_all_contexts = false; |
6120 | return ret; |
6121 | } |
6122 | |
6123 | if (dump_file) |
6124 | fprintf (stream: dump_file, format: " - Creating a specialized node of %s " |
6125 | "for all known contexts.\n" , node->dump_name ()); |
6126 | |
6127 | vec<tree> known_csts = avals.m_known_vals.copy (); |
6128 | vec<ipa_polymorphic_call_context> known_contexts |
6129 | = copy_useful_known_contexts (known_contexts: avals.m_known_contexts); |
6130 | find_more_scalar_values_for_callers_subset (node, known_csts, callers); |
6131 | find_more_contexts_for_caller_subset (node, known_contexts: &known_contexts, callers); |
6132 | vec<ipa_argagg_value, va_gc> *aggvals |
6133 | = find_aggregate_values_for_callers_subset (node, callers); |
6134 | |
6135 | if (!known_contexts_useful_p (known_contexts)) |
6136 | { |
6137 | known_contexts.release (); |
6138 | known_contexts = vNULL; |
6139 | } |
6140 | clone = create_specialized_node (node, known_csts, known_contexts, |
6141 | aggvals, callers); |
6142 | info->do_clone_for_all_contexts = false; |
6143 | ipa_node_params_sum->get (node: clone)->is_all_contexts_clone = true; |
6144 | ret = true; |
6145 | } |
6146 | |
6147 | return ret; |
6148 | } |
6149 | |
6150 | /* Transitively mark all callees of NODE within the same SCC as not dead. */ |
6151 | |
6152 | static void |
6153 | spread_undeadness (struct cgraph_node *node) |
6154 | { |
6155 | struct cgraph_edge *cs; |
6156 | |
6157 | for (cs = node->callees; cs; cs = cs->next_callee) |
6158 | if (ipa_edge_within_scc (cs)) |
6159 | { |
6160 | struct cgraph_node *callee; |
6161 | class ipa_node_params *info; |
6162 | |
6163 | callee = cs->callee->function_symbol (NULL); |
6164 | info = ipa_node_params_sum->get (node: callee); |
6165 | |
6166 | if (info && info->node_dead) |
6167 | { |
6168 | info->node_dead = 0; |
6169 | spread_undeadness (node: callee); |
6170 | } |
6171 | } |
6172 | } |
6173 | |
6174 | /* Return true if NODE has a caller from outside of its SCC that is not |
6175 | dead. Worker callback for cgraph_for_node_and_aliases. */ |
6176 | |
6177 | static bool |
6178 | has_undead_caller_from_outside_scc_p (struct cgraph_node *node, |
6179 | void *data ATTRIBUTE_UNUSED) |
6180 | { |
6181 | struct cgraph_edge *cs; |
6182 | |
6183 | for (cs = node->callers; cs; cs = cs->next_caller) |
6184 | if (cs->caller->thunk |
6185 | && cs->caller->call_for_symbol_thunks_and_aliases |
6186 | (callback: has_undead_caller_from_outside_scc_p, NULL, include_overwritable: true)) |
6187 | return true; |
6188 | else if (!ipa_edge_within_scc (cs)) |
6189 | { |
6190 | ipa_node_params *caller_info = ipa_node_params_sum->get (node: cs->caller); |
6191 | if (!caller_info /* Unoptimized caller are like dead ones. */ |
6192 | || !caller_info->node_dead) |
6193 | return true; |
6194 | } |
6195 | return false; |
6196 | } |
6197 | |
6198 | |
6199 | /* Identify nodes within the same SCC as NODE which are no longer needed |
6200 | because of new clones and will be removed as unreachable. */ |
6201 | |
6202 | static void |
6203 | identify_dead_nodes (struct cgraph_node *node) |
6204 | { |
6205 | struct cgraph_node *v; |
6206 | for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle) |
6207 | if (v->local) |
6208 | { |
6209 | ipa_node_params *info = ipa_node_params_sum->get (node: v); |
6210 | if (info |
6211 | && !v->call_for_symbol_thunks_and_aliases |
6212 | (callback: has_undead_caller_from_outside_scc_p, NULL, include_overwritable: true)) |
6213 | info->node_dead = 1; |
6214 | } |
6215 | |
6216 | for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle) |
6217 | { |
6218 | ipa_node_params *info = ipa_node_params_sum->get (node: v); |
6219 | if (info && !info->node_dead) |
6220 | spread_undeadness (node: v); |
6221 | } |
6222 | |
6223 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6224 | { |
6225 | for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle) |
6226 | if (ipa_node_params_sum->get (node: v) |
6227 | && ipa_node_params_sum->get (node: v)->node_dead) |
6228 | fprintf (stream: dump_file, format: " Marking node as dead: %s.\n" , |
6229 | v->dump_name ()); |
6230 | } |
6231 | } |
6232 | |
6233 | /* The decision stage. Iterate over the topological order of call graph nodes |
6234 | TOPO and make specialized clones if deemed beneficial. */ |
6235 | |
6236 | static void |
6237 | ipcp_decision_stage (class ipa_topo_info *topo) |
6238 | { |
6239 | int i; |
6240 | |
6241 | if (dump_file) |
6242 | fprintf (stream: dump_file, format: "\nIPA decision stage:\n\n" ); |
6243 | |
6244 | for (i = topo->nnodes - 1; i >= 0; i--) |
6245 | { |
6246 | struct cgraph_node *node = topo->order[i]; |
6247 | bool change = false, iterate = true; |
6248 | |
6249 | while (iterate) |
6250 | { |
6251 | struct cgraph_node *v; |
6252 | iterate = false; |
6253 | for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle) |
6254 | if (v->has_gimple_body_p () |
6255 | && ipcp_versionable_function_p (node: v)) |
6256 | iterate |= decide_whether_version_node (node: v); |
6257 | |
6258 | change |= iterate; |
6259 | } |
6260 | if (change) |
6261 | identify_dead_nodes (node); |
6262 | } |
6263 | } |
6264 | |
6265 | /* Look up all VR and bits information that we have discovered and copy it |
6266 | over to the transformation summary. */ |
6267 | |
6268 | static void |
6269 | ipcp_store_vr_results (void) |
6270 | { |
6271 | cgraph_node *node; |
6272 | |
6273 | FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node) |
6274 | { |
6275 | ipa_node_params *info = ipa_node_params_sum->get (node); |
6276 | bool dumped_sth = false; |
6277 | bool found_useful_result = false; |
6278 | bool do_vr = true; |
6279 | bool do_bits = true; |
6280 | |
6281 | if (!info || !opt_for_fn (node->decl, flag_ipa_vrp)) |
6282 | { |
6283 | if (dump_file) |
6284 | fprintf (stream: dump_file, format: "Not considering %s for VR discovery " |
6285 | "and propagate; -fipa-ipa-vrp: disabled.\n" , |
6286 | node->dump_name ()); |
6287 | do_vr = false; |
6288 | } |
6289 | if (!info || !opt_for_fn (node->decl, flag_ipa_bit_cp)) |
6290 | { |
6291 | if (dump_file) |
6292 | fprintf (stream: dump_file, format: "Not considering %s for ipa bitwise " |
6293 | "propagation ; -fipa-bit-cp: disabled.\n" , |
6294 | node->dump_name ()); |
6295 | do_bits = false; |
6296 | } |
6297 | if (!do_bits && !do_vr) |
6298 | continue; |
6299 | |
6300 | if (info->ipcp_orig_node) |
6301 | info = ipa_node_params_sum->get (node: info->ipcp_orig_node); |
6302 | if (info->lattices.is_empty ()) |
6303 | /* Newly expanded artificial thunks do not have lattices. */ |
6304 | continue; |
6305 | |
6306 | unsigned count = ipa_get_param_count (info); |
6307 | for (unsigned i = 0; i < count; i++) |
6308 | { |
6309 | ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i); |
6310 | if (do_vr |
6311 | && !plats->m_value_range.bottom_p () |
6312 | && !plats->m_value_range.top_p ()) |
6313 | { |
6314 | found_useful_result = true; |
6315 | break; |
6316 | } |
6317 | if (do_bits && plats->bits_lattice.constant_p ()) |
6318 | { |
6319 | found_useful_result = true; |
6320 | break; |
6321 | } |
6322 | } |
6323 | if (!found_useful_result) |
6324 | continue; |
6325 | |
6326 | ipcp_transformation_initialize (); |
6327 | ipcp_transformation *ts = ipcp_transformation_sum->get_create (node); |
6328 | vec_safe_reserve_exact (v&: ts->m_vr, nelems: count); |
6329 | |
6330 | for (unsigned i = 0; i < count; i++) |
6331 | { |
6332 | ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i); |
6333 | ipcp_bits_lattice *bits = NULL; |
6334 | |
6335 | if (do_bits |
6336 | && plats->bits_lattice.constant_p () |
6337 | && dbg_cnt (index: ipa_cp_bits)) |
6338 | bits = &plats->bits_lattice; |
6339 | |
6340 | if (do_vr |
6341 | && !plats->m_value_range.bottom_p () |
6342 | && !plats->m_value_range.top_p () |
6343 | && dbg_cnt (index: ipa_cp_vr)) |
6344 | { |
6345 | if (bits) |
6346 | { |
6347 | Value_Range tmp = plats->m_value_range.m_vr; |
6348 | tree type = ipa_get_type (info, i); |
6349 | irange &r = as_a<irange> (v&: tmp); |
6350 | irange_bitmask bm (wide_int::from (x: bits->get_value (), |
6351 | TYPE_PRECISION (type), |
6352 | TYPE_SIGN (type)), |
6353 | wide_int::from (x: bits->get_mask (), |
6354 | TYPE_PRECISION (type), |
6355 | TYPE_SIGN (type))); |
6356 | r.update_bitmask (bm); |
6357 | ipa_vr vr (tmp); |
6358 | ts->m_vr->quick_push (obj: vr); |
6359 | } |
6360 | else |
6361 | { |
6362 | ipa_vr vr (plats->m_value_range.m_vr); |
6363 | ts->m_vr->quick_push (obj: vr); |
6364 | } |
6365 | } |
6366 | else if (bits) |
6367 | { |
6368 | tree type = ipa_get_type (info, i); |
6369 | Value_Range tmp; |
6370 | tmp.set_varying (type); |
6371 | irange &r = as_a<irange> (v&: tmp); |
6372 | irange_bitmask bm (wide_int::from (x: bits->get_value (), |
6373 | TYPE_PRECISION (type), |
6374 | TYPE_SIGN (type)), |
6375 | wide_int::from (x: bits->get_mask (), |
6376 | TYPE_PRECISION (type), |
6377 | TYPE_SIGN (type))); |
6378 | r.update_bitmask (bm); |
6379 | ipa_vr vr (tmp); |
6380 | ts->m_vr->quick_push (obj: vr); |
6381 | } |
6382 | else |
6383 | { |
6384 | ipa_vr vr; |
6385 | ts->m_vr->quick_push (obj: vr); |
6386 | } |
6387 | |
6388 | if (!dump_file || !bits) |
6389 | continue; |
6390 | |
6391 | if (!dumped_sth) |
6392 | { |
6393 | fprintf (stream: dump_file, format: "Propagated bits info for function %s:\n" , |
6394 | node->dump_name ()); |
6395 | dumped_sth = true; |
6396 | } |
6397 | fprintf (stream: dump_file, format: " param %i: value = " , i); |
6398 | print_hex (wi: bits->get_value (), file: dump_file); |
6399 | fprintf (stream: dump_file, format: ", mask = " ); |
6400 | print_hex (wi: bits->get_mask (), file: dump_file); |
6401 | fprintf (stream: dump_file, format: "\n" ); |
6402 | } |
6403 | } |
6404 | } |
6405 | |
6406 | /* The IPCP driver. */ |
6407 | |
6408 | static unsigned int |
6409 | ipcp_driver (void) |
6410 | { |
6411 | class ipa_topo_info topo; |
6412 | |
6413 | if (edge_clone_summaries == NULL) |
6414 | edge_clone_summaries = new edge_clone_summary_t (symtab); |
6415 | |
6416 | ipa_check_create_node_params (); |
6417 | ipa_check_create_edge_args (); |
6418 | clone_num_suffixes = new hash_map<const char *, unsigned>; |
6419 | |
6420 | if (dump_file) |
6421 | { |
6422 | fprintf (stream: dump_file, format: "\nIPA structures before propagation:\n" ); |
6423 | if (dump_flags & TDF_DETAILS) |
6424 | ipa_print_all_params (dump_file); |
6425 | ipa_print_all_jump_functions (f: dump_file); |
6426 | } |
6427 | |
6428 | /* Topological sort. */ |
6429 | build_toporder_info (topo: &topo); |
6430 | /* Do the interprocedural propagation. */ |
6431 | ipcp_propagate_stage (topo: &topo); |
6432 | /* Decide what constant propagation and cloning should be performed. */ |
6433 | ipcp_decision_stage (topo: &topo); |
6434 | /* Store results of value range and bits propagation. */ |
6435 | ipcp_store_vr_results (); |
6436 | |
6437 | /* Free all IPCP structures. */ |
6438 | delete clone_num_suffixes; |
6439 | free_toporder_info (topo: &topo); |
6440 | delete edge_clone_summaries; |
6441 | edge_clone_summaries = NULL; |
6442 | ipa_free_all_structures_after_ipa_cp (); |
6443 | if (dump_file) |
6444 | fprintf (stream: dump_file, format: "\nIPA constant propagation end\n" ); |
6445 | return 0; |
6446 | } |
6447 | |
6448 | /* Initialization and computation of IPCP data structures. This is the initial |
6449 | intraprocedural analysis of functions, which gathers information to be |
6450 | propagated later on. */ |
6451 | |
6452 | static void |
6453 | ipcp_generate_summary (void) |
6454 | { |
6455 | struct cgraph_node *node; |
6456 | |
6457 | if (dump_file) |
6458 | fprintf (stream: dump_file, format: "\nIPA constant propagation start:\n" ); |
6459 | ipa_register_cgraph_hooks (); |
6460 | |
6461 | FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node) |
6462 | ipa_analyze_node (node); |
6463 | } |
6464 | |
6465 | namespace { |
6466 | |
6467 | const pass_data pass_data_ipa_cp = |
6468 | { |
6469 | .type: IPA_PASS, /* type */ |
6470 | .name: "cp" , /* name */ |
6471 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
6472 | .tv_id: TV_IPA_CONSTANT_PROP, /* tv_id */ |
6473 | .properties_required: 0, /* properties_required */ |
6474 | .properties_provided: 0, /* properties_provided */ |
6475 | .properties_destroyed: 0, /* properties_destroyed */ |
6476 | .todo_flags_start: 0, /* todo_flags_start */ |
6477 | .todo_flags_finish: ( TODO_dump_symtab | TODO_remove_functions ), /* todo_flags_finish */ |
6478 | }; |
6479 | |
6480 | class pass_ipa_cp : public ipa_opt_pass_d |
6481 | { |
6482 | public: |
6483 | pass_ipa_cp (gcc::context *ctxt) |
6484 | : ipa_opt_pass_d (pass_data_ipa_cp, ctxt, |
6485 | ipcp_generate_summary, /* generate_summary */ |
6486 | NULL, /* write_summary */ |
6487 | NULL, /* read_summary */ |
6488 | ipcp_write_transformation_summaries, /* |
6489 | write_optimization_summary */ |
6490 | ipcp_read_transformation_summaries, /* |
6491 | read_optimization_summary */ |
6492 | NULL, /* stmt_fixup */ |
6493 | 0, /* function_transform_todo_flags_start */ |
6494 | ipcp_transform_function, /* function_transform */ |
6495 | NULL) /* variable_transform */ |
6496 | {} |
6497 | |
6498 | /* opt_pass methods: */ |
6499 | bool gate (function *) final override |
6500 | { |
6501 | /* FIXME: We should remove the optimize check after we ensure we never run |
6502 | IPA passes when not optimizing. */ |
6503 | return (flag_ipa_cp && optimize) || in_lto_p; |
6504 | } |
6505 | |
6506 | unsigned int execute (function *) final override { return ipcp_driver (); } |
6507 | |
6508 | }; // class pass_ipa_cp |
6509 | |
6510 | } // anon namespace |
6511 | |
6512 | ipa_opt_pass_d * |
6513 | make_pass_ipa_cp (gcc::context *ctxt) |
6514 | { |
6515 | return new pass_ipa_cp (ctxt); |
6516 | } |
6517 | |
6518 | /* Reset all state within ipa-cp.cc so that we can rerun the compiler |
6519 | within the same process. For use by toplev::finalize. */ |
6520 | |
6521 | void |
6522 | ipa_cp_cc_finalize (void) |
6523 | { |
6524 | base_count = profile_count::uninitialized (); |
6525 | overall_size = 0; |
6526 | orig_overall_size = 0; |
6527 | ipcp_free_transformation_sum (); |
6528 | } |
6529 | |
6530 | /* Given PARAM which must be a parameter of function FNDECL described by THIS, |
6531 | return its index in the DECL_ARGUMENTS chain, using a pre-computed |
6532 | DECL_UID-sorted vector if available (which is pre-computed only if there are |
6533 | many parameters). Can return -1 if param is static chain not represented |
6534 | among DECL_ARGUMENTS. */ |
6535 | |
6536 | int |
6537 | ipcp_transformation::get_param_index (const_tree fndecl, const_tree param) const |
6538 | { |
6539 | gcc_assert (TREE_CODE (param) == PARM_DECL); |
6540 | if (m_uid_to_idx) |
6541 | { |
6542 | unsigned puid = DECL_UID (param); |
6543 | const ipa_uid_to_idx_map_elt *res |
6544 | = std::lower_bound (first: m_uid_to_idx->begin(), last: m_uid_to_idx->end (), val: puid, |
6545 | comp: [] (const ipa_uid_to_idx_map_elt &elt, unsigned uid) |
6546 | { |
6547 | return elt.uid < uid; |
6548 | }); |
6549 | if (res == m_uid_to_idx->end () |
6550 | || res->uid != puid) |
6551 | { |
6552 | gcc_assert (DECL_STATIC_CHAIN (fndecl)); |
6553 | return -1; |
6554 | } |
6555 | return res->index; |
6556 | } |
6557 | |
6558 | unsigned index = 0; |
6559 | for (tree p = DECL_ARGUMENTS (fndecl); p; p = DECL_CHAIN (p), index++) |
6560 | if (p == param) |
6561 | return (int) index; |
6562 | |
6563 | gcc_assert (DECL_STATIC_CHAIN (fndecl)); |
6564 | return -1; |
6565 | } |
6566 | |
6567 | /* Helper function to qsort a vector of ipa_uid_to_idx_map_elt elements |
6568 | according to the uid. */ |
6569 | |
6570 | static int |
6571 | compare_uids (const void *a, const void *b) |
6572 | { |
6573 | const ipa_uid_to_idx_map_elt *e1 = (const ipa_uid_to_idx_map_elt *) a; |
6574 | const ipa_uid_to_idx_map_elt *e2 = (const ipa_uid_to_idx_map_elt *) b; |
6575 | if (e1->uid < e2->uid) |
6576 | return -1; |
6577 | if (e1->uid > e2->uid) |
6578 | return 1; |
6579 | gcc_unreachable (); |
6580 | } |
6581 | |
6582 | /* Assuming THIS describes FNDECL and it has sufficiently many parameters to |
6583 | justify the overhead, create a DECL_UID-sorted vector to speed up mapping |
6584 | from parameters to their indices in DECL_ARGUMENTS chain. */ |
6585 | |
6586 | void |
6587 | ipcp_transformation::maybe_create_parm_idx_map (tree fndecl) |
6588 | { |
6589 | int c = count_formal_params (fndecl); |
6590 | if (c < 32) |
6591 | return; |
6592 | |
6593 | m_uid_to_idx = NULL; |
6594 | vec_safe_reserve (v&: m_uid_to_idx, nelems: c, exact: true); |
6595 | unsigned index = 0; |
6596 | for (tree p = DECL_ARGUMENTS (fndecl); p; p = DECL_CHAIN (p), index++) |
6597 | { |
6598 | ipa_uid_to_idx_map_elt elt; |
6599 | elt.uid = DECL_UID (p); |
6600 | elt.index = index; |
6601 | m_uid_to_idx->quick_push (obj: elt); |
6602 | } |
6603 | m_uid_to_idx->qsort (compare_uids); |
6604 | } |
6605 | |