Daily bump.
[gcc.git] / gcc / ipa-fnsummary.h
1 /* IPA function body analysis.
2 Copyright (C) 2003-2021 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
4
5 This file is part of GCC.
6
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
11
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
16
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
20
21 #ifndef GCC_IPA_SUMMARY_H
22 #define GCC_IPA_SUMMARY_H
23
24 #include "sreal.h"
25 #include "ipa-predicate.h"
26
27
28 /* Hints are reasons why IPA heuristics should prefer specializing given
29 function. They are represented as bitmap of the following values. */
30 enum ipa_hints_vals {
31 /* When specialization turns indirect call into a direct call,
32 it is good idea to do so. */
33 INLINE_HINT_indirect_call = 1,
34 /* Inlining may make loop iterations or loop stride known. It is good idea
35 to do so because it enables loop optimizations. */
36 INLINE_HINT_loop_iterations = 2,
37 INLINE_HINT_loop_stride = 4,
38 /* Inlining within same strongly connected component of callgraph is often
39 a loss due to increased stack frame usage and prologue setup costs. */
40 INLINE_HINT_same_scc = 8,
41 /* Inlining functions in strongly connected component is not such a great
42 win. */
43 INLINE_HINT_in_scc = 16,
44 /* If function is declared inline by user, it may be good idea to inline
45 it. Set by simple_edge_hints in ipa-inline-analysis.c. */
46 INLINE_HINT_declared_inline = 32,
47 /* Programs are usually still organized for non-LTO compilation and thus
48 if functions are in different modules, inlining may not be so important.
49 Set by simple_edge_hints in ipa-inline-analysis.c. */
50 INLINE_HINT_cross_module = 64,
51 /* We know that the callee is hot by profile. */
52 INLINE_HINT_known_hot = 128,
53 /* There is builtin_constant_p dependent on parameter which is usually
54 a strong hint to inline. */
55 INLINE_HINT_builtin_constant_p = 256
56 };
57
58 typedef int ipa_hints;
59
60 /* Simple description of whether a memory load or a condition refers to a load
61 from an aggregate and if so, how and where from in the aggregate.
62 Individual fields have the same meaning like fields with the same name in
63 struct condition. */
64
65 struct agg_position_info
66 {
67 HOST_WIDE_INT offset;
68 bool agg_contents;
69 bool by_ref;
70 };
71
72 /* Representation of function body size and time depending on the call
73 context. We keep simple array of record, every containing of predicate
74 and time/size to account. */
75 class size_time_entry
76 {
77 public:
78 /* Predicate for code to be executed. */
79 predicate exec_predicate;
80 /* Predicate for value to be constant and optimized out in a specialized copy.
81 When deciding on specialization this makes it possible to see how much
82 the executed code paths will simplify. */
83 predicate nonconst_predicate;
84 int size;
85 sreal time;
86 };
87
88 /* Summary about function and stack frame sizes. We keep this info
89 for inline clones and also for WPA streaming. For this reason this is not
90 part of ipa_fn_summary which exists only for offline functions. */
91 class ipa_size_summary
92 {
93 public:
94 /* Estimated stack frame consumption by the function. */
95 HOST_WIDE_INT estimated_self_stack_size;
96 /* Size of the function body. */
97 int self_size;
98 /* Estimated size of the function after inlining. */
99 int size;
100
101 ipa_size_summary ()
102 : estimated_self_stack_size (0), self_size (0), size (0)
103 {
104 }
105 };
106
107 /* Structure to capture how frequently some interesting events occur given a
108 particular predicate. The structure is used to estimate how often we
109 encounter loops with known iteration count or stride in various
110 contexts. */
111
112 struct GTY(()) ipa_freqcounting_predicate
113 {
114 /* The described event happens with this frequency... */
115 sreal freq;
116 /* ...when this predicate evaluates to false. */
117 class predicate * GTY((skip)) predicate;
118 };
119
120 /* Function inlining information. */
121 class GTY(()) ipa_fn_summary
122 {
123 public:
124 /* Keep all field empty so summary dumping works during its computation.
125 This is useful for debugging. */
126 ipa_fn_summary ()
127 : min_size (0),
128 inlinable (false), single_caller (false),
129 fp_expressions (false),
130 estimated_stack_size (false),
131 time (0), conds (NULL),
132 size_time_table (), call_size_time_table (vNULL),
133 loop_iterations (NULL), loop_strides (NULL),
134 builtin_constant_p_parms (vNULL),
135 growth (0), scc_no (0)
136 {
137 }
138
139 /* Copy constructor. */
140 ipa_fn_summary (const ipa_fn_summary &s)
141 : min_size (s.min_size),
142 inlinable (s.inlinable), single_caller (s.single_caller),
143 fp_expressions (s.fp_expressions),
144 estimated_stack_size (s.estimated_stack_size),
145 time (s.time), conds (s.conds), size_time_table (),
146 call_size_time_table (vNULL),
147 loop_iterations (s.loop_iterations), loop_strides (s.loop_strides),
148 builtin_constant_p_parms (s.builtin_constant_p_parms),
149 growth (s.growth), scc_no (s.scc_no)
150 {}
151
152 /* Default constructor. */
153 ~ipa_fn_summary ();
154
155 /* Information about the function body itself. */
156
157 /* Minimal size increase after inlining. */
158 int min_size;
159
160 /* False when there something makes inlining impossible (such as va_arg). */
161 unsigned inlinable : 1;
162 /* True wen there is only one caller of the function before small function
163 inlining. */
164 unsigned int single_caller : 1;
165 /* True if function contains any floating point expressions. */
166 unsigned int fp_expressions : 1;
167
168 /* Information about function that will result after applying all the
169 inline decisions present in the callgraph. Generally kept up to
170 date only for functions that are not inline clones. */
171
172 /* Estimated stack frame consumption by the function. */
173 HOST_WIDE_INT estimated_stack_size;
174 /* Estimated runtime of function after inlining. */
175 sreal GTY((skip)) time;
176
177 /* Conditional size/time information. The summaries are being
178 merged during inlining. */
179 conditions conds;
180 /* Normal code is accounted in size_time_table, while calls are
181 accounted in call_size_time_table. This is because calls
182 are often adjusted by IPA optimizations and thus this summary
183 is generated from call summary information when needed. */
184 auto_vec<size_time_entry> GTY((skip)) size_time_table;
185 /* Unlike size_time_table that is initialized for all summaries
186 call_size_time_table is allocated only for functions with
187 many calls. Use effecient vl_ptr storage. */
188 vec<size_time_entry, va_heap, vl_ptr> GTY((skip)) call_size_time_table;
189
190 /* Predicates on when some loops in the function can have known bounds. */
191 vec<ipa_freqcounting_predicate, va_gc> *loop_iterations;
192 /* Predicates on when some loops in the function can have known strides. */
193 vec<ipa_freqcounting_predicate, va_gc> *loop_strides;
194 /* Parameters tested by builtin_constant_p. */
195 vec<int, va_heap, vl_ptr> GTY((skip)) builtin_constant_p_parms;
196 /* Estimated growth for inlining all copies of the function before start
197 of small functions inlining.
198 This value will get out of date as the callers are duplicated, but
199 using up-to-date value in the badness metric mean a lot of extra
200 expenses. */
201 int growth;
202 /* Number of SCC on the beginning of inlining process. */
203 int scc_no;
204
205 /* Record time and size under given predicates. */
206 void account_size_time (int, sreal, const predicate &, const predicate &,
207 bool call = false);
208
209 /* We keep values scaled up, so fractional sizes can be accounted. */
210 static const int size_scale = 2;
211 /* Maximal size of size_time_table before we start to be conservative. */
212 static const int max_size_time_table_size = 256;
213 };
214
215 class GTY((user)) ipa_fn_summary_t:
216 public fast_function_summary <ipa_fn_summary *, va_gc>
217 {
218 public:
219 ipa_fn_summary_t (symbol_table *symtab):
220 fast_function_summary <ipa_fn_summary *, va_gc> (symtab) {}
221
222 static ipa_fn_summary_t *create_ggc (symbol_table *symtab)
223 {
224 class ipa_fn_summary_t *summary
225 = new (ggc_alloc_no_dtor<ipa_fn_summary_t> ()) ipa_fn_summary_t (symtab);
226 summary->disable_insertion_hook ();
227 return summary;
228 }
229
230 /* Remove ipa_fn_summary for all callees of NODE. */
231 void remove_callees (cgraph_node *node);
232
233 virtual void insert (cgraph_node *, ipa_fn_summary *);
234 virtual void remove (cgraph_node *node, ipa_fn_summary *)
235 {
236 remove_callees (node);
237 }
238
239 virtual void duplicate (cgraph_node *src, cgraph_node *dst,
240 ipa_fn_summary *src_data, ipa_fn_summary *dst_data);
241 };
242
243 extern GTY(()) fast_function_summary <ipa_fn_summary *, va_gc>
244 *ipa_fn_summaries;
245
246 class ipa_size_summary_t:
247 public fast_function_summary <ipa_size_summary *, va_heap>
248 {
249 public:
250 ipa_size_summary_t (symbol_table *symtab):
251 fast_function_summary <ipa_size_summary *, va_heap> (symtab)
252 {
253 disable_insertion_hook ();
254 }
255
256 virtual void duplicate (cgraph_node *, cgraph_node *,
257 ipa_size_summary *src_data,
258 ipa_size_summary *dst_data)
259 {
260 *dst_data = *src_data;
261 }
262 };
263 extern fast_function_summary <ipa_size_summary *, va_heap>
264 *ipa_size_summaries;
265
266 /* Information kept about callgraph edges. */
267 class ipa_call_summary
268 {
269 public:
270 /* Keep all field empty so summary dumping works during its computation.
271 This is useful for debugging. */
272 ipa_call_summary ()
273 : predicate (NULL), param (vNULL), call_stmt_size (0), call_stmt_time (0),
274 loop_depth (0), is_return_callee_uncaptured (false)
275 {
276 }
277
278 /* Copy constructor. */
279 ipa_call_summary (const ipa_call_summary &s):
280 predicate (s.predicate), param (s.param), call_stmt_size (s.call_stmt_size),
281 call_stmt_time (s.call_stmt_time), loop_depth (s.loop_depth),
282 is_return_callee_uncaptured (s.is_return_callee_uncaptured)
283 {
284 }
285
286 /* Default destructor. */
287 ~ipa_call_summary ();
288
289 class predicate *predicate;
290 /* Vector indexed by parameters. */
291 vec<inline_param_summary> param;
292 /* Estimated size and time of the call statement. */
293 int call_stmt_size;
294 int call_stmt_time;
295 /* Depth of loop nest, 0 means no nesting. */
296 unsigned int loop_depth;
297 /* Indicates whether the caller returns the value of it's callee. */
298 bool is_return_callee_uncaptured;
299 };
300
301 class ipa_call_summary_t: public fast_call_summary <ipa_call_summary *, va_heap>
302 {
303 public:
304 ipa_call_summary_t (symbol_table *symtab):
305 fast_call_summary <ipa_call_summary *, va_heap> (symtab) {}
306
307 /* Hook that is called by summary when an edge is duplicated. */
308 virtual void duplicate (cgraph_edge *src, cgraph_edge *dst,
309 ipa_call_summary *src_data,
310 ipa_call_summary *dst_data);
311 };
312
313 /* Estimated execution times, code sizes and other information about the
314 code executing a call described by ipa_call_context. */
315
316 struct ipa_call_estimates
317 {
318 /* Estimated size needed to execute call in the given context. */
319 int size;
320
321 /* Minimal size needed for the call that is + independent on the call context
322 and can be used for fast estimates. */
323 int min_size;
324
325 /* Estimated time needed to execute call in the given context. */
326 sreal time;
327
328 /* Estimated time needed to execute the function when not ignoring
329 computations known to be constant in this context. */
330 sreal nonspecialized_time;
331
332 /* Further discovered reasons why to inline or specialize the give calls. */
333 ipa_hints hints;
334
335 /* Frequency how often a loop with known number of iterations is encountered.
336 Calculated with hints. */
337 sreal loops_with_known_iterations;
338
339 /* Frequency how often a loop with known strides is encountered. Calculated
340 with hints. */
341 sreal loops_with_known_strides;
342 };
343
344 class ipa_cached_call_context;
345
346 /* This object describe a context of call. That is a summary of known
347 information about its parameters. Main purpose of this context is
348 to give more realistic estimations of function runtime, size and
349 inline hints. */
350 class ipa_call_context
351 {
352 public:
353 ipa_call_context (cgraph_node *node,
354 clause_t possible_truths,
355 clause_t nonspec_possible_truths,
356 vec<inline_param_summary> inline_param_summary,
357 ipa_auto_call_arg_values *arg_values);
358 ipa_call_context ()
359 : m_node(NULL)
360 {
361 }
362 void estimate_size_and_time (ipa_call_estimates *estimates,
363 bool est_times = true, bool est_hints = true);
364 bool equal_to (const ipa_call_context &);
365 bool exists_p ()
366 {
367 return m_node != NULL;
368 }
369 private:
370 /* Called function. */
371 cgraph_node *m_node;
372 /* Clause describing what predicate conditionals can be satisfied
373 in this context if function is inlined/specialized. */
374 clause_t m_possible_truths;
375 /* Clause describing what predicate conditionals can be satisfied
376 in this context if function is kept offline. */
377 clause_t m_nonspec_possible_truths;
378 /* Inline summary maintains info about change probabilities. */
379 vec<inline_param_summary> m_inline_param_summary;
380
381 /* Even after having calculated clauses, the information about argument
382 values is used to resolve indirect calls. */
383 ipa_call_arg_values m_avals;
384
385 friend ipa_cached_call_context;
386 };
387
388 /* Variant of ipa_call_context that is stored in a cache over a longer period
389 of time. */
390
391 class ipa_cached_call_context : public ipa_call_context
392 {
393 public:
394 void duplicate_from (const ipa_call_context &ctx);
395 void release ();
396 };
397
398 extern fast_call_summary <ipa_call_summary *, va_heap> *ipa_call_summaries;
399
400 /* In ipa-fnsummary.c */
401 void ipa_debug_fn_summary (struct cgraph_node *);
402 void ipa_dump_fn_summaries (FILE *f);
403 void ipa_dump_fn_summary (FILE *f, struct cgraph_node *node);
404 void ipa_dump_hints (FILE *f, ipa_hints);
405 void ipa_free_fn_summary (void);
406 void ipa_free_size_summary (void);
407 void inline_analyze_function (struct cgraph_node *node);
408 void estimate_ipcp_clone_size_and_time (struct cgraph_node *node,
409 ipa_auto_call_arg_values *avals,
410 ipa_call_estimates *estimates);
411 void ipa_merge_fn_summary_after_inlining (struct cgraph_edge *edge);
412 void ipa_update_overall_fn_summary (struct cgraph_node *node, bool reset = true);
413 void compute_fn_summary (struct cgraph_node *, bool);
414 bool refs_local_or_readonly_memory_p (tree);
415 bool points_to_local_or_readonly_memory_p (tree);
416
417
418 void evaluate_properties_for_edge (struct cgraph_edge *e,
419 bool inline_p,
420 clause_t *clause_ptr,
421 clause_t *nonspec_clause_ptr,
422 ipa_auto_call_arg_values *avals,
423 bool compute_contexts);
424
425 void ipa_fnsummary_c_finalize (void);
426 HOST_WIDE_INT ipa_get_stack_frame_offset (struct cgraph_node *node);
427 void ipa_remove_from_growth_caches (struct cgraph_edge *edge);
428
429 /* Return true if EDGE is a cross module call. */
430
431 static inline bool
432 cross_module_call_p (struct cgraph_edge *edge)
433 {
434 /* Here we do not want to walk to alias target becuase ICF may create
435 cross-unit aliases. */
436 if (edge->caller->unit_id == edge->callee->unit_id)
437 return false;
438 /* If the call is to a (former) comdat function or s symbol with mutiple
439 extern inline definitions then treat is as in-module call. */
440 if (edge->callee->merged_extern_inline || edge->callee->merged_comdat
441 || DECL_COMDAT (edge->callee->decl))
442 return false;
443 return true;
444 }
445
446 #endif /* GCC_IPA_FNSUMMARY_H */