1 |
12 |
jlechner |
/* Calculate branch probabilities, and basic block execution counts.
|
2 |
|
|
Copyright (C) 1990, 1991, 1992, 1993, 1994, 1996, 1997, 1998, 1999,
|
3 |
|
|
2000, 2001, 2002, 2003, 2005 Free Software Foundation, Inc.
|
4 |
|
|
Contributed by James E. Wilson, UC Berkeley/Cygnus Support;
|
5 |
|
|
based on some ideas from Dain Samples of UC Berkeley.
|
6 |
|
|
Further mangling by Bob Manson, Cygnus Support.
|
7 |
|
|
|
8 |
|
|
This file is part of GCC.
|
9 |
|
|
|
10 |
|
|
GCC is free software; you can redistribute it and/or modify it under
|
11 |
|
|
the terms of the GNU General Public License as published by the Free
|
12 |
|
|
Software Foundation; either version 2, or (at your option) any later
|
13 |
|
|
version.
|
14 |
|
|
|
15 |
|
|
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
|
16 |
|
|
WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
17 |
|
|
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
18 |
|
|
for more details.
|
19 |
|
|
|
20 |
|
|
You should have received a copy of the GNU General Public License
|
21 |
|
|
along with GCC; see the file COPYING. If not, write to the Free
|
22 |
|
|
Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
|
23 |
|
|
02110-1301, USA. */
|
24 |
|
|
|
25 |
|
|
/* Generate basic block profile instrumentation and auxiliary files.
|
26 |
|
|
RTL-based version. See profile.c for overview. */
|
27 |
|
|
|
28 |
|
|
#include "config.h"
|
29 |
|
|
#include "system.h"
|
30 |
|
|
#include "coretypes.h"
|
31 |
|
|
#include "tm.h"
|
32 |
|
|
#include "rtl.h"
|
33 |
|
|
#include "flags.h"
|
34 |
|
|
#include "output.h"
|
35 |
|
|
#include "regs.h"
|
36 |
|
|
#include "expr.h"
|
37 |
|
|
#include "function.h"
|
38 |
|
|
#include "toplev.h"
|
39 |
|
|
#include "coverage.h"
|
40 |
|
|
#include "value-prof.h"
|
41 |
|
|
#include "tree.h"
|
42 |
|
|
#include "ggc.h"
|
43 |
|
|
|
44 |
|
|
/* Do initialization work for the edge profiler. */
|
45 |
|
|
|
46 |
|
|
static void
|
47 |
|
|
rtl_init_edge_profiler (void)
|
48 |
|
|
{
|
49 |
|
|
/* gen_edge_profiler calls safe_insert_insn_on_edge which needs
|
50 |
|
|
register liveness data to be available. */
|
51 |
|
|
life_analysis (NULL, 0);
|
52 |
|
|
}
|
53 |
|
|
|
54 |
|
|
/* Output instructions as RTL to increment the edge execution count. */
|
55 |
|
|
|
56 |
|
|
static void
|
57 |
|
|
rtl_gen_edge_profiler (int edgeno, edge e)
|
58 |
|
|
{
|
59 |
|
|
rtx ref = rtl_coverage_counter_ref (GCOV_COUNTER_ARCS, edgeno);
|
60 |
|
|
rtx tmp;
|
61 |
|
|
enum machine_mode mode = GET_MODE (ref);
|
62 |
|
|
rtx sequence;
|
63 |
|
|
|
64 |
|
|
start_sequence ();
|
65 |
|
|
ref = validize_mem (ref);
|
66 |
|
|
|
67 |
|
|
tmp = expand_simple_binop (mode, PLUS, ref, const1_rtx,
|
68 |
|
|
ref, 0, OPTAB_WIDEN);
|
69 |
|
|
|
70 |
|
|
if (tmp != ref)
|
71 |
|
|
emit_move_insn (copy_rtx (ref), tmp);
|
72 |
|
|
|
73 |
|
|
sequence = get_insns ();
|
74 |
|
|
end_sequence ();
|
75 |
|
|
safe_insert_insn_on_edge (sequence, e);
|
76 |
|
|
rebuild_jump_labels (e->insns.r);
|
77 |
|
|
}
|
78 |
|
|
|
79 |
|
|
/* Output instructions as RTL to increment the interval histogram counter.
|
80 |
|
|
VALUE is the expression whose value is profiled. TAG is the tag of the
|
81 |
|
|
section for counters, BASE is offset of the counter position. */
|
82 |
|
|
|
83 |
|
|
static void
|
84 |
|
|
rtl_gen_interval_profiler (histogram_value value, unsigned tag, unsigned base)
|
85 |
|
|
{
|
86 |
|
|
enum machine_mode mode = mode_for_size (GCOV_TYPE_SIZE, MODE_INT, 0);
|
87 |
|
|
rtx mem_ref, tmp, tmp1, mr, val;
|
88 |
|
|
rtx sequence;
|
89 |
|
|
rtx more_label = gen_label_rtx ();
|
90 |
|
|
rtx less_label = gen_label_rtx ();
|
91 |
|
|
rtx end_of_code_label = gen_label_rtx ();
|
92 |
|
|
int per_counter = GCOV_TYPE_SIZE / BITS_PER_UNIT;
|
93 |
|
|
edge e = split_block (BLOCK_FOR_INSN (value->hvalue.rtl.insn),
|
94 |
|
|
PREV_INSN (value->hvalue.rtl.insn));
|
95 |
|
|
|
96 |
|
|
start_sequence ();
|
97 |
|
|
|
98 |
|
|
if (value->hvalue.rtl.seq)
|
99 |
|
|
emit_insn (value->hvalue.rtl.seq);
|
100 |
|
|
|
101 |
|
|
mr = gen_reg_rtx (Pmode);
|
102 |
|
|
|
103 |
|
|
tmp = rtl_coverage_counter_ref (tag, base);
|
104 |
|
|
tmp = force_reg (Pmode, XEXP (tmp, 0));
|
105 |
|
|
|
106 |
|
|
val = expand_simple_binop (value->hvalue.rtl.mode, MINUS,
|
107 |
|
|
copy_rtx (value->hvalue.rtl.value),
|
108 |
|
|
GEN_INT (value->hdata.intvl.int_start),
|
109 |
|
|
NULL_RTX, 0, OPTAB_WIDEN);
|
110 |
|
|
|
111 |
|
|
do_compare_rtx_and_jump (copy_rtx (val), GEN_INT (value->hdata.intvl.steps),
|
112 |
|
|
GE, 0, value->hvalue.rtl.mode, NULL_RTX, NULL_RTX,
|
113 |
|
|
more_label);
|
114 |
|
|
do_compare_rtx_and_jump (copy_rtx (val), const0_rtx, LT, 0,
|
115 |
|
|
value->hvalue.rtl.mode,
|
116 |
|
|
NULL_RTX, NULL_RTX, less_label);
|
117 |
|
|
|
118 |
|
|
/* We are in range. */
|
119 |
|
|
tmp1 = expand_simple_binop (value->hvalue.rtl.mode, MULT,
|
120 |
|
|
copy_rtx (val), GEN_INT (per_counter),
|
121 |
|
|
NULL_RTX, 0, OPTAB_WIDEN);
|
122 |
|
|
tmp1 = expand_simple_binop (Pmode, PLUS, copy_rtx (tmp), tmp1, mr,
|
123 |
|
|
0, OPTAB_WIDEN);
|
124 |
|
|
if (tmp1 != mr)
|
125 |
|
|
emit_move_insn (copy_rtx (mr), tmp1);
|
126 |
|
|
|
127 |
|
|
emit_jump_insn (gen_jump (end_of_code_label));
|
128 |
|
|
emit_barrier ();
|
129 |
|
|
|
130 |
|
|
/* Above the interval. */
|
131 |
|
|
emit_label (more_label);
|
132 |
|
|
tmp1 = expand_simple_binop (Pmode, PLUS, copy_rtx (tmp),
|
133 |
|
|
GEN_INT (per_counter * value->hdata.intvl.steps),
|
134 |
|
|
mr, 0, OPTAB_WIDEN);
|
135 |
|
|
if (tmp1 != mr)
|
136 |
|
|
emit_move_insn (copy_rtx (mr), tmp1);
|
137 |
|
|
emit_jump_insn (gen_jump (end_of_code_label));
|
138 |
|
|
emit_barrier ();
|
139 |
|
|
|
140 |
|
|
/* Below the interval. */
|
141 |
|
|
emit_label (less_label);
|
142 |
|
|
tmp1 = expand_simple_binop (Pmode, PLUS, copy_rtx (tmp),
|
143 |
|
|
GEN_INT (per_counter * (value->hdata.intvl.steps +1)),
|
144 |
|
|
mr, 0, OPTAB_WIDEN);
|
145 |
|
|
if (tmp1 != mr)
|
146 |
|
|
emit_move_insn (copy_rtx (mr), tmp1);
|
147 |
|
|
|
148 |
|
|
emit_label (end_of_code_label);
|
149 |
|
|
|
150 |
|
|
mem_ref = validize_mem (gen_rtx_MEM (mode, mr));
|
151 |
|
|
|
152 |
|
|
tmp = expand_simple_binop (mode, PLUS, copy_rtx (mem_ref), const1_rtx,
|
153 |
|
|
mem_ref, 0, OPTAB_WIDEN);
|
154 |
|
|
|
155 |
|
|
if (tmp != mem_ref)
|
156 |
|
|
emit_move_insn (copy_rtx (mem_ref), tmp);
|
157 |
|
|
|
158 |
|
|
sequence = get_insns ();
|
159 |
|
|
end_sequence ();
|
160 |
|
|
rebuild_jump_labels (sequence);
|
161 |
|
|
safe_insert_insn_on_edge (sequence, e);
|
162 |
|
|
}
|
163 |
|
|
|
164 |
|
|
/* Output instructions as RTL to increment the power of two histogram counter.
|
165 |
|
|
VALUE is the expression whose value is profiled. TAG is the tag of the
|
166 |
|
|
section for counters, BASE is offset of the counter position. */
|
167 |
|
|
|
168 |
|
|
static void
|
169 |
|
|
rtl_gen_pow2_profiler (histogram_value value, unsigned tag, unsigned base)
|
170 |
|
|
{
|
171 |
|
|
enum machine_mode mode = mode_for_size (GCOV_TYPE_SIZE, MODE_INT, 0);
|
172 |
|
|
rtx mem_ref, tmp, mr, uval;
|
173 |
|
|
rtx sequence;
|
174 |
|
|
rtx end_of_code_label = gen_label_rtx ();
|
175 |
|
|
int per_counter = GCOV_TYPE_SIZE / BITS_PER_UNIT;
|
176 |
|
|
edge e = split_block (BLOCK_FOR_INSN (value->hvalue.rtl.insn),
|
177 |
|
|
PREV_INSN (value->hvalue.rtl.insn));
|
178 |
|
|
|
179 |
|
|
start_sequence ();
|
180 |
|
|
|
181 |
|
|
if (value->hvalue.rtl.seq)
|
182 |
|
|
emit_insn (value->hvalue.rtl.seq);
|
183 |
|
|
|
184 |
|
|
mr = gen_reg_rtx (Pmode);
|
185 |
|
|
tmp = rtl_coverage_counter_ref (tag, base);
|
186 |
|
|
tmp = force_reg (Pmode, XEXP (tmp, 0));
|
187 |
|
|
emit_move_insn (mr, tmp);
|
188 |
|
|
|
189 |
|
|
uval = gen_reg_rtx (value->hvalue.rtl.mode);
|
190 |
|
|
emit_move_insn (uval, copy_rtx (value->hvalue.rtl.value));
|
191 |
|
|
|
192 |
|
|
/* Check for non-power of 2. */
|
193 |
|
|
do_compare_rtx_and_jump (copy_rtx (uval), const0_rtx, LE, 0, value->hvalue.rtl.mode,
|
194 |
|
|
NULL_RTX, NULL_RTX, end_of_code_label);
|
195 |
|
|
tmp = expand_simple_binop (value->hvalue.rtl.mode, PLUS, copy_rtx (uval),
|
196 |
|
|
constm1_rtx, NULL_RTX, 0, OPTAB_WIDEN);
|
197 |
|
|
tmp = expand_simple_binop (value->hvalue.rtl.mode, AND, copy_rtx (uval), tmp,
|
198 |
|
|
NULL_RTX, 0, OPTAB_WIDEN);
|
199 |
|
|
do_compare_rtx_and_jump (tmp, const0_rtx, NE, 0, value->hvalue.rtl.mode, NULL_RTX,
|
200 |
|
|
NULL_RTX, end_of_code_label);
|
201 |
|
|
|
202 |
|
|
tmp = expand_simple_binop (Pmode, PLUS, copy_rtx (mr), GEN_INT (per_counter),
|
203 |
|
|
mr, 0, OPTAB_WIDEN);
|
204 |
|
|
if (tmp != mr)
|
205 |
|
|
emit_move_insn (copy_rtx (mr), tmp);
|
206 |
|
|
|
207 |
|
|
/* Increase the counter. */
|
208 |
|
|
emit_label (end_of_code_label);
|
209 |
|
|
|
210 |
|
|
mem_ref = validize_mem (gen_rtx_MEM (mode, mr));
|
211 |
|
|
|
212 |
|
|
tmp = expand_simple_binop (mode, PLUS, copy_rtx (mem_ref), const1_rtx,
|
213 |
|
|
mem_ref, 0, OPTAB_WIDEN);
|
214 |
|
|
|
215 |
|
|
if (tmp != mem_ref)
|
216 |
|
|
emit_move_insn (copy_rtx (mem_ref), tmp);
|
217 |
|
|
|
218 |
|
|
sequence = get_insns ();
|
219 |
|
|
end_sequence ();
|
220 |
|
|
rebuild_jump_labels (sequence);
|
221 |
|
|
safe_insert_insn_on_edge (sequence, e);
|
222 |
|
|
}
|
223 |
|
|
|
224 |
|
|
/* Output instructions as RTL for code to find the most common value.
|
225 |
|
|
VALUE is the expression whose value is profiled. TAG is the tag of the
|
226 |
|
|
section for counters, BASE is offset of the counter position. */
|
227 |
|
|
|
228 |
|
|
static rtx
|
229 |
|
|
rtl_gen_one_value_profiler_no_edge_manipulation (histogram_value value,
|
230 |
|
|
unsigned tag, unsigned base)
|
231 |
|
|
{
|
232 |
|
|
enum machine_mode mode = mode_for_size (GCOV_TYPE_SIZE, MODE_INT, 0);
|
233 |
|
|
rtx stored_value_ref, counter_ref, all_ref, stored_value, counter, all;
|
234 |
|
|
rtx tmp, uval;
|
235 |
|
|
rtx sequence;
|
236 |
|
|
rtx same_label = gen_label_rtx ();
|
237 |
|
|
rtx zero_label = gen_label_rtx ();
|
238 |
|
|
rtx end_of_code_label = gen_label_rtx ();
|
239 |
|
|
|
240 |
|
|
start_sequence ();
|
241 |
|
|
|
242 |
|
|
if (value->hvalue.rtl.seq)
|
243 |
|
|
emit_insn (value->hvalue.rtl.seq);
|
244 |
|
|
|
245 |
|
|
stored_value_ref = rtl_coverage_counter_ref (tag, base);
|
246 |
|
|
counter_ref = rtl_coverage_counter_ref (tag, base + 1);
|
247 |
|
|
all_ref = rtl_coverage_counter_ref (tag, base + 2);
|
248 |
|
|
stored_value = validize_mem (stored_value_ref);
|
249 |
|
|
counter = validize_mem (counter_ref);
|
250 |
|
|
all = validize_mem (all_ref);
|
251 |
|
|
|
252 |
|
|
uval = gen_reg_rtx (mode);
|
253 |
|
|
convert_move (uval, copy_rtx (value->hvalue.rtl.value), 0);
|
254 |
|
|
|
255 |
|
|
/* Check if the stored value matches. */
|
256 |
|
|
do_compare_rtx_and_jump (copy_rtx (uval), copy_rtx (stored_value), EQ,
|
257 |
|
|
0, mode, NULL_RTX, NULL_RTX, same_label);
|
258 |
|
|
|
259 |
|
|
/* Does not match; check whether the counter is zero. */
|
260 |
|
|
do_compare_rtx_and_jump (copy_rtx (counter), const0_rtx, EQ, 0, mode,
|
261 |
|
|
NULL_RTX, NULL_RTX, zero_label);
|
262 |
|
|
|
263 |
|
|
/* The counter is not zero yet. */
|
264 |
|
|
tmp = expand_simple_binop (mode, PLUS, copy_rtx (counter), constm1_rtx,
|
265 |
|
|
counter, 0, OPTAB_WIDEN);
|
266 |
|
|
|
267 |
|
|
if (tmp != counter)
|
268 |
|
|
emit_move_insn (copy_rtx (counter), tmp);
|
269 |
|
|
|
270 |
|
|
emit_jump_insn (gen_jump (end_of_code_label));
|
271 |
|
|
emit_barrier ();
|
272 |
|
|
|
273 |
|
|
emit_label (zero_label);
|
274 |
|
|
/* Set new value. */
|
275 |
|
|
emit_move_insn (copy_rtx (stored_value), copy_rtx (uval));
|
276 |
|
|
|
277 |
|
|
emit_label (same_label);
|
278 |
|
|
/* Increase the counter. */
|
279 |
|
|
tmp = expand_simple_binop (mode, PLUS, copy_rtx (counter), const1_rtx,
|
280 |
|
|
counter, 0, OPTAB_WIDEN);
|
281 |
|
|
|
282 |
|
|
if (tmp != counter)
|
283 |
|
|
emit_move_insn (copy_rtx (counter), tmp);
|
284 |
|
|
|
285 |
|
|
emit_label (end_of_code_label);
|
286 |
|
|
|
287 |
|
|
/* Increase the counter of all executions; this seems redundant given
|
288 |
|
|
that ve have counts for edges in cfg, but it may happen that some
|
289 |
|
|
optimization will change the counts for the block (either because
|
290 |
|
|
it is unable to update them correctly, or because it will duplicate
|
291 |
|
|
the block or its part). */
|
292 |
|
|
tmp = expand_simple_binop (mode, PLUS, copy_rtx (all), const1_rtx,
|
293 |
|
|
all, 0, OPTAB_WIDEN);
|
294 |
|
|
|
295 |
|
|
if (tmp != all)
|
296 |
|
|
emit_move_insn (copy_rtx (all), tmp);
|
297 |
|
|
sequence = get_insns ();
|
298 |
|
|
end_sequence ();
|
299 |
|
|
return sequence;
|
300 |
|
|
}
|
301 |
|
|
|
302 |
|
|
/* Output instructions as RTL for code to find the most common value.
|
303 |
|
|
VALUE is the expression whose value is profiled. TAG is the tag of the
|
304 |
|
|
section for counters, BASE is offset of the counter position. */
|
305 |
|
|
|
306 |
|
|
static void
|
307 |
|
|
rtl_gen_one_value_profiler (histogram_value value, unsigned tag, unsigned base)
|
308 |
|
|
{
|
309 |
|
|
edge e = split_block (BLOCK_FOR_INSN (value->hvalue.rtl.insn),
|
310 |
|
|
PREV_INSN (value->hvalue.rtl.insn));
|
311 |
|
|
rtx sequence = rtl_gen_one_value_profiler_no_edge_manipulation (value,
|
312 |
|
|
tag, base);
|
313 |
|
|
rebuild_jump_labels (sequence);
|
314 |
|
|
safe_insert_insn_on_edge (sequence, e);
|
315 |
|
|
}
|
316 |
|
|
|
317 |
|
|
/* Output instructions as RTL for code to find the most common value of
|
318 |
|
|
a difference between two evaluations of an expression.
|
319 |
|
|
VALUE is the expression whose value is profiled. TAG is the tag of the
|
320 |
|
|
section for counters, BASE is offset of the counter position. */
|
321 |
|
|
|
322 |
|
|
static void
|
323 |
|
|
rtl_gen_const_delta_profiler (histogram_value value, unsigned tag, unsigned base)
|
324 |
|
|
{
|
325 |
|
|
histogram_value one_value_delta;
|
326 |
|
|
enum machine_mode mode = mode_for_size (GCOV_TYPE_SIZE, MODE_INT, 0);
|
327 |
|
|
rtx stored_value_ref, stored_value, tmp, uval;
|
328 |
|
|
rtx sequence;
|
329 |
|
|
edge e = split_block (BLOCK_FOR_INSN (value->hvalue.rtl.insn),
|
330 |
|
|
PREV_INSN (value->hvalue.rtl.insn));
|
331 |
|
|
|
332 |
|
|
start_sequence ();
|
333 |
|
|
|
334 |
|
|
if (value->hvalue.rtl.seq)
|
335 |
|
|
emit_insn (value->hvalue.rtl.seq);
|
336 |
|
|
|
337 |
|
|
stored_value_ref = rtl_coverage_counter_ref (tag, base);
|
338 |
|
|
stored_value = validize_mem (stored_value_ref);
|
339 |
|
|
|
340 |
|
|
uval = gen_reg_rtx (mode);
|
341 |
|
|
convert_move (uval, copy_rtx (value->hvalue.rtl.value), 0);
|
342 |
|
|
tmp = expand_simple_binop (mode, MINUS,
|
343 |
|
|
copy_rtx (uval), copy_rtx (stored_value),
|
344 |
|
|
NULL_RTX, 0, OPTAB_WIDEN);
|
345 |
|
|
|
346 |
|
|
one_value_delta = ggc_alloc (sizeof (*one_value_delta));
|
347 |
|
|
one_value_delta->hvalue.rtl.value = tmp;
|
348 |
|
|
one_value_delta->hvalue.rtl.mode = mode;
|
349 |
|
|
one_value_delta->hvalue.rtl.seq = NULL_RTX;
|
350 |
|
|
one_value_delta->hvalue.rtl.insn = value->hvalue.rtl.insn;
|
351 |
|
|
one_value_delta->type = HIST_TYPE_SINGLE_VALUE;
|
352 |
|
|
emit_insn (rtl_gen_one_value_profiler_no_edge_manipulation (one_value_delta,
|
353 |
|
|
tag, base + 1));
|
354 |
|
|
emit_move_insn (copy_rtx (stored_value), uval);
|
355 |
|
|
sequence = get_insns ();
|
356 |
|
|
end_sequence ();
|
357 |
|
|
rebuild_jump_labels (sequence);
|
358 |
|
|
safe_insert_insn_on_edge (sequence, e);
|
359 |
|
|
}
|
360 |
|
|
|
361 |
|
|
/* Return the file on which profile dump output goes, if any. */
|
362 |
|
|
|
363 |
|
|
static FILE *rtl_profile_dump_file (void) {
|
364 |
|
|
return dump_file;
|
365 |
|
|
}
|
366 |
|
|
|
367 |
|
|
struct profile_hooks rtl_profile_hooks =
|
368 |
|
|
{
|
369 |
|
|
rtl_init_edge_profiler,
|
370 |
|
|
rtl_gen_edge_profiler,
|
371 |
|
|
rtl_gen_interval_profiler,
|
372 |
|
|
rtl_gen_pow2_profiler,
|
373 |
|
|
rtl_gen_one_value_profiler,
|
374 |
|
|
rtl_gen_const_delta_profiler,
|
375 |
|
|
rtl_profile_dump_file
|
376 |
|
|
};
|