111
|
1 /* IPA function body analysis.
|
131
|
2 Copyright (C) 2003-2018 Free Software Foundation, Inc.
|
111
|
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 preffer specializing given
|
|
29 function. They are represtented 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 optimizatoins. */
|
|
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 /* If array indexes of loads/stores become known there may be room for
|
|
52 further optimization. */
|
|
53 INLINE_HINT_array_index = 128,
|
|
54 /* We know that the callee is hot by profile. */
|
|
55 INLINE_HINT_known_hot = 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 struct GTY(()) size_time_entry
|
|
76 {
|
|
77 /* Predicate for code to be executed. */
|
|
78 predicate exec_predicate;
|
|
79 /* Predicate for value to be constant and optimized out in a specialized copy.
|
|
80 When deciding on specialization this makes it possible to see how much
|
|
81 the executed code paths will simplify. */
|
|
82 predicate nonconst_predicate;
|
|
83 int size;
|
|
84 sreal GTY((skip)) time;
|
|
85 };
|
|
86
|
|
87 /* Function inlining information. */
|
|
88 struct GTY(()) ipa_fn_summary
|
|
89 {
|
131
|
90 /* Keep all field empty so summary dumping works during its computation.
|
|
91 This is useful for debugging. */
|
|
92 ipa_fn_summary ()
|
|
93 : estimated_self_stack_size (0), self_size (0), min_size (0),
|
|
94 inlinable (false), single_caller (false),
|
|
95 fp_expressions (false), estimated_stack_size (false),
|
|
96 stack_frame_offset (false), time (0), size (0), conds (NULL),
|
|
97 size_time_table (NULL), loop_iterations (NULL), loop_stride (NULL),
|
|
98 array_index (NULL), growth (0), scc_no (0)
|
|
99 {
|
|
100 }
|
|
101
|
|
102 /* Copy constructor. */
|
|
103 ipa_fn_summary (const ipa_fn_summary &s)
|
|
104 : estimated_self_stack_size (s.estimated_self_stack_size),
|
|
105 self_size (s.self_size), min_size (s.min_size),
|
|
106 inlinable (s.inlinable), single_caller (s.single_caller),
|
|
107 fp_expressions (s.fp_expressions),
|
|
108 estimated_stack_size (s.estimated_stack_size),
|
|
109 stack_frame_offset (s.stack_frame_offset), time (s.time), size (s.size),
|
|
110 conds (s.conds), size_time_table (s.size_time_table),
|
|
111 loop_iterations (s.loop_iterations), loop_stride (s.loop_stride),
|
|
112 array_index (s.array_index), growth (s.growth), scc_no (s.scc_no)
|
|
113 {}
|
|
114
|
|
115 /* Default constructor. */
|
|
116 ~ipa_fn_summary ();
|
|
117
|
111
|
118 /* Information about the function body itself. */
|
|
119
|
|
120 /* Estimated stack frame consumption by the function. */
|
|
121 HOST_WIDE_INT estimated_self_stack_size;
|
|
122 /* Size of the function body. */
|
|
123 int self_size;
|
|
124 /* Minimal size increase after inlining. */
|
|
125 int min_size;
|
|
126
|
|
127 /* False when there something makes inlining impossible (such as va_arg). */
|
|
128 unsigned inlinable : 1;
|
|
129 /* True wen there is only one caller of the function before small function
|
|
130 inlining. */
|
|
131 unsigned int single_caller : 1;
|
|
132 /* True if function contains any floating point expressions. */
|
|
133 unsigned int fp_expressions : 1;
|
|
134
|
|
135 /* Information about function that will result after applying all the
|
|
136 inline decisions present in the callgraph. Generally kept up to
|
|
137 date only for functions that are not inline clones. */
|
|
138
|
|
139 /* Estimated stack frame consumption by the function. */
|
|
140 HOST_WIDE_INT estimated_stack_size;
|
|
141 /* Expected offset of the stack frame of function. */
|
|
142 HOST_WIDE_INT stack_frame_offset;
|
|
143 /* Estimated size of the function after inlining. */
|
|
144 sreal GTY((skip)) time;
|
|
145 int size;
|
|
146
|
|
147 /* Conditional size/time information. The summaries are being
|
|
148 merged during inlining. */
|
|
149 conditions conds;
|
|
150 vec<size_time_entry, va_gc> *size_time_table;
|
|
151
|
|
152 /* Predicate on when some loop in the function becomes to have known
|
|
153 bounds. */
|
|
154 predicate * GTY((skip)) loop_iterations;
|
|
155 /* Predicate on when some loop in the function becomes to have known
|
|
156 stride. */
|
|
157 predicate * GTY((skip)) loop_stride;
|
|
158 /* Predicate on when some array indexes become constants. */
|
|
159 predicate * GTY((skip)) array_index;
|
|
160 /* Estimated growth for inlining all copies of the function before start
|
|
161 of small functions inlining.
|
|
162 This value will get out of date as the callers are duplicated, but
|
|
163 using up-to-date value in the badness metric mean a lot of extra
|
|
164 expenses. */
|
|
165 int growth;
|
|
166 /* Number of SCC on the beginning of inlining process. */
|
|
167 int scc_no;
|
|
168
|
|
169 /* Record time and size under given predicates. */
|
|
170 void account_size_time (int, sreal, const predicate &, const predicate &);
|
|
171
|
|
172 /* We keep values scaled up, so fractional sizes can be accounted. */
|
|
173 static const int size_scale = 2;
|
|
174 };
|
|
175
|
|
176 class GTY((user)) ipa_fn_summary_t: public function_summary <ipa_fn_summary *>
|
|
177 {
|
|
178 public:
|
|
179 ipa_fn_summary_t (symbol_table *symtab, bool ggc):
|
|
180 function_summary <ipa_fn_summary *> (symtab, ggc) {}
|
|
181
|
|
182 static ipa_fn_summary_t *create_ggc (symbol_table *symtab)
|
|
183 {
|
|
184 struct ipa_fn_summary_t *summary = new (ggc_alloc <ipa_fn_summary_t> ())
|
|
185 ipa_fn_summary_t(symtab, true);
|
|
186 summary->disable_insertion_hook ();
|
|
187 return summary;
|
|
188 }
|
|
189
|
131
|
190 /* Remove ipa_fn_summary for all callees of NODE. */
|
|
191 void remove_callees (cgraph_node *node);
|
111
|
192
|
|
193 virtual void insert (cgraph_node *, ipa_fn_summary *);
|
131
|
194 virtual void remove (cgraph_node *node, ipa_fn_summary *)
|
|
195 {
|
|
196 remove_callees (node);
|
|
197 }
|
|
198
|
111
|
199 virtual void duplicate (cgraph_node *src, cgraph_node *dst,
|
|
200 ipa_fn_summary *src_data, ipa_fn_summary *dst_data);
|
|
201 };
|
|
202
|
|
203 extern GTY(()) function_summary <ipa_fn_summary *> *ipa_fn_summaries;
|
|
204
|
|
205 /* Information kept about callgraph edges. */
|
|
206 struct ipa_call_summary
|
|
207 {
|
131
|
208 /* Keep all field empty so summary dumping works during its computation.
|
|
209 This is useful for debugging. */
|
|
210 ipa_call_summary ()
|
|
211 : predicate (NULL), param (vNULL), call_stmt_size (0), call_stmt_time (0),
|
|
212 loop_depth (0), is_return_callee_uncaptured (false)
|
|
213 {
|
|
214 }
|
|
215
|
|
216 /* Copy constructor. */
|
|
217 ipa_call_summary (const ipa_call_summary &s):
|
|
218 predicate (s.predicate), param (s.param), call_stmt_size (s.call_stmt_size),
|
|
219 call_stmt_time (s.call_stmt_time), loop_depth (s.loop_depth),
|
|
220 is_return_callee_uncaptured (s.is_return_callee_uncaptured)
|
|
221 {
|
|
222 }
|
|
223
|
|
224 /* Default destructor. */
|
|
225 ~ipa_call_summary ();
|
|
226
|
111
|
227 class predicate *predicate;
|
|
228 /* Vector indexed by parameters. */
|
|
229 vec<inline_param_summary> param;
|
|
230 /* Estimated size and time of the call statement. */
|
|
231 int call_stmt_size;
|
|
232 int call_stmt_time;
|
|
233 /* Depth of loop nest, 0 means no nesting. */
|
|
234 unsigned int loop_depth;
|
131
|
235 /* Indicates whether the caller returns the value of it's callee. */
|
|
236 bool is_return_callee_uncaptured;
|
111
|
237 };
|
|
238
|
|
239 class ipa_call_summary_t: public call_summary <ipa_call_summary *>
|
|
240 {
|
|
241 public:
|
|
242 ipa_call_summary_t (symbol_table *symtab, bool ggc):
|
|
243 call_summary <ipa_call_summary *> (symtab, ggc) {}
|
|
244
|
|
245 /* Hook that is called by summary when an edge is duplicated. */
|
|
246 virtual void duplicate (cgraph_edge *src, cgraph_edge *dst,
|
|
247 ipa_call_summary *src_data,
|
|
248 ipa_call_summary *dst_data);
|
|
249 };
|
|
250
|
|
251 extern call_summary <ipa_call_summary *> *ipa_call_summaries;
|
|
252
|
|
253 /* In ipa-fnsummary.c */
|
|
254 void ipa_debug_fn_summary (struct cgraph_node *);
|
|
255 void ipa_dump_fn_summaries (FILE *f);
|
|
256 void ipa_dump_fn_summary (FILE *f, struct cgraph_node *node);
|
|
257 void ipa_dump_hints (FILE *f, ipa_hints);
|
|
258 void ipa_free_fn_summary (void);
|
|
259 void inline_analyze_function (struct cgraph_node *node);
|
|
260 void estimate_ipcp_clone_size_and_time (struct cgraph_node *,
|
|
261 vec<tree>,
|
|
262 vec<ipa_polymorphic_call_context>,
|
|
263 vec<ipa_agg_jump_function_p>,
|
|
264 int *, sreal *, sreal *,
|
|
265 ipa_hints *);
|
|
266 void ipa_merge_fn_summary_after_inlining (struct cgraph_edge *edge);
|
|
267 void ipa_update_overall_fn_summary (struct cgraph_node *node);
|
|
268 void compute_fn_summary (struct cgraph_node *, bool);
|
|
269
|
|
270
|
|
271 void evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p,
|
|
272 clause_t *clause_ptr,
|
|
273 clause_t *nonspec_clause_ptr,
|
|
274 vec<tree> *known_vals_ptr,
|
|
275 vec<ipa_polymorphic_call_context>
|
|
276 *known_contexts_ptr,
|
|
277 vec<ipa_agg_jump_function_p> *);
|
|
278 void estimate_node_size_and_time (struct cgraph_node *node,
|
|
279 clause_t possible_truths,
|
|
280 clause_t nonspec_possible_truths,
|
|
281 vec<tree> known_vals,
|
|
282 vec<ipa_polymorphic_call_context>,
|
|
283 vec<ipa_agg_jump_function_p> known_aggs,
|
|
284 int *ret_size, int *ret_min_size,
|
|
285 sreal *ret_time,
|
|
286 sreal *ret_nonspecialized_time,
|
|
287 ipa_hints *ret_hints,
|
|
288 vec<inline_param_summary>
|
|
289 inline_param_summary);
|
|
290
|
131
|
291 void ipa_fnsummary_c_finalize (void);
|
|
292
|
111
|
293 #endif /* GCC_IPA_FNSUMMARY_H */
|