111
|
1 /* IPA function body analysis.
|
|
2 Copyright (C) 2003-2017 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 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 {
|
|
90 /* Information about the function body itself. */
|
|
91
|
|
92 /* Estimated stack frame consumption by the function. */
|
|
93 HOST_WIDE_INT estimated_self_stack_size;
|
|
94 /* Size of the function body. */
|
|
95 int self_size;
|
|
96 /* Minimal size increase after inlining. */
|
|
97 int min_size;
|
|
98
|
|
99 /* False when there something makes inlining impossible (such as va_arg). */
|
|
100 unsigned inlinable : 1;
|
|
101 /* True when function contains cilk spawn (and thus we can not inline
|
|
102 into it). */
|
|
103 unsigned contains_cilk_spawn : 1;
|
|
104 /* True wen there is only one caller of the function before small function
|
|
105 inlining. */
|
|
106 unsigned int single_caller : 1;
|
|
107 /* True if function contains any floating point expressions. */
|
|
108 unsigned int fp_expressions : 1;
|
|
109
|
|
110 /* Information about function that will result after applying all the
|
|
111 inline decisions present in the callgraph. Generally kept up to
|
|
112 date only for functions that are not inline clones. */
|
|
113
|
|
114 /* Estimated stack frame consumption by the function. */
|
|
115 HOST_WIDE_INT estimated_stack_size;
|
|
116 /* Expected offset of the stack frame of function. */
|
|
117 HOST_WIDE_INT stack_frame_offset;
|
|
118 /* Estimated size of the function after inlining. */
|
|
119 sreal GTY((skip)) time;
|
|
120 int size;
|
|
121
|
|
122 /* Conditional size/time information. The summaries are being
|
|
123 merged during inlining. */
|
|
124 conditions conds;
|
|
125 vec<size_time_entry, va_gc> *size_time_table;
|
|
126
|
|
127 /* Predicate on when some loop in the function becomes to have known
|
|
128 bounds. */
|
|
129 predicate * GTY((skip)) loop_iterations;
|
|
130 /* Predicate on when some loop in the function becomes to have known
|
|
131 stride. */
|
|
132 predicate * GTY((skip)) loop_stride;
|
|
133 /* Predicate on when some array indexes become constants. */
|
|
134 predicate * GTY((skip)) array_index;
|
|
135 /* Estimated growth for inlining all copies of the function before start
|
|
136 of small functions inlining.
|
|
137 This value will get out of date as the callers are duplicated, but
|
|
138 using up-to-date value in the badness metric mean a lot of extra
|
|
139 expenses. */
|
|
140 int growth;
|
|
141 /* Number of SCC on the beginning of inlining process. */
|
|
142 int scc_no;
|
|
143
|
|
144 /* Keep all field empty so summary dumping works during its computation.
|
|
145 This is useful for debugging. */
|
|
146 ipa_fn_summary ()
|
|
147 : estimated_self_stack_size (0), self_size (0), min_size (0),
|
|
148 inlinable (false), contains_cilk_spawn (false), single_caller (false),
|
|
149 fp_expressions (false), estimated_stack_size (false),
|
|
150 stack_frame_offset (false), time (0), size (0), conds (NULL),
|
|
151 size_time_table (NULL), loop_iterations (NULL), loop_stride (NULL),
|
|
152 array_index (NULL), growth (0), scc_no (0)
|
|
153 {
|
|
154 }
|
|
155
|
|
156 /* Record time and size under given predicates. */
|
|
157 void account_size_time (int, sreal, const predicate &, const predicate &);
|
|
158
|
|
159 /* Reset summary to empty state. */
|
|
160 void reset (struct cgraph_node *node);
|
|
161
|
|
162 /* We keep values scaled up, so fractional sizes can be accounted. */
|
|
163 static const int size_scale = 2;
|
|
164 };
|
|
165
|
|
166 class GTY((user)) ipa_fn_summary_t: public function_summary <ipa_fn_summary *>
|
|
167 {
|
|
168 public:
|
|
169 ipa_fn_summary_t (symbol_table *symtab, bool ggc):
|
|
170 function_summary <ipa_fn_summary *> (symtab, ggc) {}
|
|
171
|
|
172 static ipa_fn_summary_t *create_ggc (symbol_table *symtab)
|
|
173 {
|
|
174 struct ipa_fn_summary_t *summary = new (ggc_alloc <ipa_fn_summary_t> ())
|
|
175 ipa_fn_summary_t(symtab, true);
|
|
176 summary->disable_insertion_hook ();
|
|
177 return summary;
|
|
178 }
|
|
179
|
|
180
|
|
181 virtual void insert (cgraph_node *, ipa_fn_summary *);
|
|
182 virtual void remove (cgraph_node *node, ipa_fn_summary *);
|
|
183 virtual void duplicate (cgraph_node *src, cgraph_node *dst,
|
|
184 ipa_fn_summary *src_data, ipa_fn_summary *dst_data);
|
|
185 };
|
|
186
|
|
187 extern GTY(()) function_summary <ipa_fn_summary *> *ipa_fn_summaries;
|
|
188
|
|
189 /* Information kept about callgraph edges. */
|
|
190 struct ipa_call_summary
|
|
191 {
|
|
192 class predicate *predicate;
|
|
193 /* Vector indexed by parameters. */
|
|
194 vec<inline_param_summary> param;
|
|
195 /* Estimated size and time of the call statement. */
|
|
196 int call_stmt_size;
|
|
197 int call_stmt_time;
|
|
198 /* Depth of loop nest, 0 means no nesting. */
|
|
199 unsigned int loop_depth;
|
|
200
|
|
201 /* Keep all field empty so summary dumping works during its computation.
|
|
202 This is useful for debugging. */
|
|
203 ipa_call_summary ()
|
|
204 : predicate (NULL), param (vNULL), call_stmt_size (0), call_stmt_time (0),
|
|
205 loop_depth (0)
|
|
206 {
|
|
207 }
|
|
208
|
|
209 /* Reset inline summary to empty state. */
|
|
210 void reset ();
|
|
211 };
|
|
212
|
|
213 class ipa_call_summary_t: public call_summary <ipa_call_summary *>
|
|
214 {
|
|
215 public:
|
|
216 ipa_call_summary_t (symbol_table *symtab, bool ggc):
|
|
217 call_summary <ipa_call_summary *> (symtab, ggc) {}
|
|
218
|
|
219 /* Hook that is called by summary when an edge is duplicated. */
|
|
220 virtual void remove (cgraph_edge *cs, ipa_call_summary *);
|
|
221 /* Hook that is called by summary when an edge is duplicated. */
|
|
222 virtual void duplicate (cgraph_edge *src, cgraph_edge *dst,
|
|
223 ipa_call_summary *src_data,
|
|
224 ipa_call_summary *dst_data);
|
|
225 };
|
|
226
|
|
227 extern call_summary <ipa_call_summary *> *ipa_call_summaries;
|
|
228
|
|
229 /* In ipa-fnsummary.c */
|
|
230 void ipa_debug_fn_summary (struct cgraph_node *);
|
|
231 void ipa_dump_fn_summaries (FILE *f);
|
|
232 void ipa_dump_fn_summary (FILE *f, struct cgraph_node *node);
|
|
233 void ipa_dump_hints (FILE *f, ipa_hints);
|
|
234 void ipa_free_fn_summary (void);
|
|
235 void inline_analyze_function (struct cgraph_node *node);
|
|
236 void estimate_ipcp_clone_size_and_time (struct cgraph_node *,
|
|
237 vec<tree>,
|
|
238 vec<ipa_polymorphic_call_context>,
|
|
239 vec<ipa_agg_jump_function_p>,
|
|
240 int *, sreal *, sreal *,
|
|
241 ipa_hints *);
|
|
242 void ipa_merge_fn_summary_after_inlining (struct cgraph_edge *edge);
|
|
243 void ipa_update_overall_fn_summary (struct cgraph_node *node);
|
|
244 void compute_fn_summary (struct cgraph_node *, bool);
|
|
245
|
|
246
|
|
247 void evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p,
|
|
248 clause_t *clause_ptr,
|
|
249 clause_t *nonspec_clause_ptr,
|
|
250 vec<tree> *known_vals_ptr,
|
|
251 vec<ipa_polymorphic_call_context>
|
|
252 *known_contexts_ptr,
|
|
253 vec<ipa_agg_jump_function_p> *);
|
|
254 void estimate_node_size_and_time (struct cgraph_node *node,
|
|
255 clause_t possible_truths,
|
|
256 clause_t nonspec_possible_truths,
|
|
257 vec<tree> known_vals,
|
|
258 vec<ipa_polymorphic_call_context>,
|
|
259 vec<ipa_agg_jump_function_p> known_aggs,
|
|
260 int *ret_size, int *ret_min_size,
|
|
261 sreal *ret_time,
|
|
262 sreal *ret_nonspecialized_time,
|
|
263 ipa_hints *ret_hints,
|
|
264 vec<inline_param_summary>
|
|
265 inline_param_summary);
|
|
266
|
|
267 #endif /* GCC_IPA_FNSUMMARY_H */
|