1 |
38 |
julius |
/* Instruction scheduling pass. This file contains definitions used
|
2 |
|
|
internally in the scheduler.
|
3 |
|
|
Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998,
|
4 |
|
|
1999, 2000, 2001, 2003, 2004, 2005, 2006, 2007
|
5 |
|
|
Free Software Foundation, Inc.
|
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 |
|
|
#ifndef GCC_SCHED_INT_H
|
24 |
|
|
#define GCC_SCHED_INT_H
|
25 |
|
|
|
26 |
|
|
/* For state_t. */
|
27 |
|
|
#include "insn-attr.h"
|
28 |
|
|
/* For regset_head. */
|
29 |
|
|
#include "basic-block.h"
|
30 |
|
|
/* For reg_note. */
|
31 |
|
|
#include "rtl.h"
|
32 |
|
|
|
33 |
|
|
/* Pointer to data describing the current DFA state. */
|
34 |
|
|
extern state_t curr_state;
|
35 |
|
|
|
36 |
|
|
/* Forward declaration. */
|
37 |
|
|
struct ready_list;
|
38 |
|
|
|
39 |
|
|
/* Type to represent status of a dependence. */
|
40 |
|
|
typedef int ds_t;
|
41 |
|
|
|
42 |
|
|
/* Type to represent weakness of speculative dependence. */
|
43 |
|
|
typedef int dw_t;
|
44 |
|
|
|
45 |
|
|
/* Describe state of dependencies used during sched_analyze phase. */
|
46 |
|
|
struct deps
|
47 |
|
|
{
|
48 |
|
|
/* The *_insns and *_mems are paired lists. Each pending memory operation
|
49 |
|
|
will have a pointer to the MEM rtx on one list and a pointer to the
|
50 |
|
|
containing insn on the other list in the same place in the list. */
|
51 |
|
|
|
52 |
|
|
/* We can't use add_dependence like the old code did, because a single insn
|
53 |
|
|
may have multiple memory accesses, and hence needs to be on the list
|
54 |
|
|
once for each memory access. Add_dependence won't let you add an insn
|
55 |
|
|
to a list more than once. */
|
56 |
|
|
|
57 |
|
|
/* An INSN_LIST containing all insns with pending read operations. */
|
58 |
|
|
rtx pending_read_insns;
|
59 |
|
|
|
60 |
|
|
/* An EXPR_LIST containing all MEM rtx's which are pending reads. */
|
61 |
|
|
rtx pending_read_mems;
|
62 |
|
|
|
63 |
|
|
/* An INSN_LIST containing all insns with pending write operations. */
|
64 |
|
|
rtx pending_write_insns;
|
65 |
|
|
|
66 |
|
|
/* An EXPR_LIST containing all MEM rtx's which are pending writes. */
|
67 |
|
|
rtx pending_write_mems;
|
68 |
|
|
|
69 |
|
|
/* Indicates the combined length of the two pending lists. We must prevent
|
70 |
|
|
these lists from ever growing too large since the number of dependencies
|
71 |
|
|
produced is at least O(N*N), and execution time is at least O(4*N*N), as
|
72 |
|
|
a function of the length of these pending lists. */
|
73 |
|
|
int pending_lists_length;
|
74 |
|
|
|
75 |
|
|
/* Length of the pending memory flush list. Large functions with no
|
76 |
|
|
calls may build up extremely large lists. */
|
77 |
|
|
int pending_flush_length;
|
78 |
|
|
|
79 |
|
|
/* The last insn upon which all memory references must depend.
|
80 |
|
|
This is an insn which flushed the pending lists, creating a dependency
|
81 |
|
|
between it and all previously pending memory references. This creates
|
82 |
|
|
a barrier (or a checkpoint) which no memory reference is allowed to cross.
|
83 |
|
|
|
84 |
|
|
This includes all non constant CALL_INSNs. When we do interprocedural
|
85 |
|
|
alias analysis, this restriction can be relaxed.
|
86 |
|
|
This may also be an INSN that writes memory if the pending lists grow
|
87 |
|
|
too large. */
|
88 |
|
|
rtx last_pending_memory_flush;
|
89 |
|
|
|
90 |
|
|
/* A list of the last function calls we have seen. We use a list to
|
91 |
|
|
represent last function calls from multiple predecessor blocks.
|
92 |
|
|
Used to prevent register lifetimes from expanding unnecessarily. */
|
93 |
|
|
rtx last_function_call;
|
94 |
|
|
|
95 |
|
|
/* A list of insns which use a pseudo register that does not already
|
96 |
|
|
cross a call. We create dependencies between each of those insn
|
97 |
|
|
and the next call insn, to ensure that they won't cross a call after
|
98 |
|
|
scheduling is done. */
|
99 |
|
|
rtx sched_before_next_call;
|
100 |
|
|
|
101 |
|
|
/* Used to keep post-call pseudo/hard reg movements together with
|
102 |
|
|
the call. */
|
103 |
|
|
enum { not_post_call, post_call, post_call_initial } in_post_call_group_p;
|
104 |
|
|
|
105 |
|
|
/* Set to the tail insn of the outermost libcall block.
|
106 |
|
|
|
107 |
|
|
When nonzero, we will mark each insn processed by sched_analyze_insn
|
108 |
|
|
with SCHED_GROUP_P to ensure libcalls are scheduled as a unit. */
|
109 |
|
|
rtx libcall_block_tail_insn;
|
110 |
|
|
|
111 |
|
|
/* The maximum register number for the following arrays. Before reload
|
112 |
|
|
this is max_reg_num; after reload it is FIRST_PSEUDO_REGISTER. */
|
113 |
|
|
int max_reg;
|
114 |
|
|
|
115 |
|
|
/* Element N is the next insn that sets (hard or pseudo) register
|
116 |
|
|
N within the current basic block; or zero, if there is no
|
117 |
|
|
such insn. Needed for new registers which may be introduced
|
118 |
|
|
by splitting insns. */
|
119 |
|
|
struct deps_reg
|
120 |
|
|
{
|
121 |
|
|
rtx uses;
|
122 |
|
|
rtx sets;
|
123 |
|
|
rtx clobbers;
|
124 |
|
|
int uses_length;
|
125 |
|
|
int clobbers_length;
|
126 |
|
|
} *reg_last;
|
127 |
|
|
|
128 |
|
|
/* Element N is set for each register that has any nonzero element
|
129 |
|
|
in reg_last[N].{uses,sets,clobbers}. */
|
130 |
|
|
regset_head reg_last_in_use;
|
131 |
|
|
|
132 |
|
|
/* Element N is set for each register that is conditionally set. */
|
133 |
|
|
regset_head reg_conditional_sets;
|
134 |
|
|
};
|
135 |
|
|
|
136 |
|
|
/* This structure holds some state of the current scheduling pass, and
|
137 |
|
|
contains some function pointers that abstract out some of the non-generic
|
138 |
|
|
functionality from functions such as schedule_block or schedule_insn.
|
139 |
|
|
There is one global variable, current_sched_info, which points to the
|
140 |
|
|
sched_info structure currently in use. */
|
141 |
|
|
struct sched_info
|
142 |
|
|
{
|
143 |
|
|
/* Add all insns that are initially ready to the ready list. Called once
|
144 |
|
|
before scheduling a set of insns. */
|
145 |
|
|
void (*init_ready_list) (void);
|
146 |
|
|
/* Called after taking an insn from the ready list. Returns nonzero if
|
147 |
|
|
this insn can be scheduled, nonzero if we should silently discard it. */
|
148 |
|
|
int (*can_schedule_ready_p) (rtx);
|
149 |
|
|
/* Return nonzero if there are more insns that should be scheduled. */
|
150 |
|
|
int (*schedule_more_p) (void);
|
151 |
|
|
/* Called after an insn has all its hard dependencies resolved.
|
152 |
|
|
Adjusts status of instruction (which is passed through second parameter)
|
153 |
|
|
to indicate if instruction should be moved to the ready list or the
|
154 |
|
|
queue, or if it should silently discard it (until next resolved
|
155 |
|
|
dependence). */
|
156 |
|
|
ds_t (*new_ready) (rtx, ds_t);
|
157 |
|
|
/* Compare priority of two insns. Return a positive number if the second
|
158 |
|
|
insn is to be preferred for scheduling, and a negative one if the first
|
159 |
|
|
is to be preferred. Zero if they are equally good. */
|
160 |
|
|
int (*rank) (rtx, rtx);
|
161 |
|
|
/* Return a string that contains the insn uid and optionally anything else
|
162 |
|
|
necessary to identify this insn in an output. It's valid to use a
|
163 |
|
|
static buffer for this. The ALIGNED parameter should cause the string
|
164 |
|
|
to be formatted so that multiple output lines will line up nicely. */
|
165 |
|
|
const char *(*print_insn) (rtx, int);
|
166 |
|
|
/* Return nonzero if an insn should be included in priority
|
167 |
|
|
calculations. */
|
168 |
|
|
int (*contributes_to_priority) (rtx, rtx);
|
169 |
|
|
/* Called when computing dependencies for a JUMP_INSN. This function
|
170 |
|
|
should store the set of registers that must be considered as set by
|
171 |
|
|
the jump in the regset. */
|
172 |
|
|
void (*compute_jump_reg_dependencies) (rtx, regset, regset, regset);
|
173 |
|
|
|
174 |
|
|
/* The boundaries of the set of insns to be scheduled. */
|
175 |
|
|
rtx prev_head, next_tail;
|
176 |
|
|
|
177 |
|
|
/* Filled in after the schedule is finished; the first and last scheduled
|
178 |
|
|
insns. */
|
179 |
|
|
rtx head, tail;
|
180 |
|
|
|
181 |
|
|
/* If nonzero, enables an additional sanity check in schedule_block. */
|
182 |
|
|
unsigned int queue_must_finish_empty:1;
|
183 |
|
|
/* Nonzero if we should use cselib for better alias analysis. This
|
184 |
|
|
must be 0 if the dependency information is used after sched_analyze
|
185 |
|
|
has completed, e.g. if we're using it to initialize state for successor
|
186 |
|
|
blocks in region scheduling. */
|
187 |
|
|
unsigned int use_cselib:1;
|
188 |
|
|
|
189 |
|
|
/* Maximum priority that has been assigned to an insn. */
|
190 |
|
|
int sched_max_insns_priority;
|
191 |
|
|
|
192 |
|
|
/* Hooks to support speculative scheduling. */
|
193 |
|
|
|
194 |
|
|
/* Called to notify frontend that instruction is being added (second
|
195 |
|
|
parameter == 0) or removed (second parameter == 1). */
|
196 |
|
|
void (*add_remove_insn) (rtx, int);
|
197 |
|
|
|
198 |
|
|
/* Called to notify frontend that instruction is being scheduled.
|
199 |
|
|
The first parameter - instruction to scheduled, the second parameter -
|
200 |
|
|
last scheduled instruction. */
|
201 |
|
|
void (*begin_schedule_ready) (rtx, rtx);
|
202 |
|
|
|
203 |
|
|
/* Called to notify frontend, that new basic block is being added.
|
204 |
|
|
The first parameter - new basic block.
|
205 |
|
|
The second parameter - block, after which new basic block is being added,
|
206 |
|
|
or EXIT_BLOCK_PTR, if recovery block is being added,
|
207 |
|
|
or NULL, if standalone block is being added. */
|
208 |
|
|
void (*add_block) (basic_block, basic_block);
|
209 |
|
|
|
210 |
|
|
/* If the second parameter is not NULL, return nonnull value, if the
|
211 |
|
|
basic block should be advanced.
|
212 |
|
|
If the second parameter is NULL, return the next basic block in EBB.
|
213 |
|
|
The first parameter is the current basic block in EBB. */
|
214 |
|
|
basic_block (*advance_target_bb) (basic_block, rtx);
|
215 |
|
|
|
216 |
|
|
/* Called after blocks were rearranged due to movement of jump instruction.
|
217 |
|
|
The first parameter - index of basic block, in which jump currently is.
|
218 |
|
|
The second parameter - index of basic block, in which jump used
|
219 |
|
|
to be.
|
220 |
|
|
The third parameter - index of basic block, that follows the second
|
221 |
|
|
parameter. */
|
222 |
|
|
void (*fix_recovery_cfg) (int, int, int);
|
223 |
|
|
|
224 |
|
|
#ifdef ENABLE_CHECKING
|
225 |
|
|
/* If the second parameter is zero, return nonzero, if block is head of the
|
226 |
|
|
region.
|
227 |
|
|
If the second parameter is nonzero, return nonzero, if block is leaf of
|
228 |
|
|
the region.
|
229 |
|
|
global_live_at_start should not change in region heads and
|
230 |
|
|
global_live_at_end should not change in region leafs due to scheduling. */
|
231 |
|
|
int (*region_head_or_leaf_p) (basic_block, int);
|
232 |
|
|
#endif
|
233 |
|
|
|
234 |
|
|
/* ??? FIXME: should use straight bitfields inside sched_info instead of
|
235 |
|
|
this flag field. */
|
236 |
|
|
unsigned int flags;
|
237 |
|
|
};
|
238 |
|
|
|
239 |
|
|
/* This structure holds description of the properties for speculative
|
240 |
|
|
scheduling. */
|
241 |
|
|
struct spec_info_def
|
242 |
|
|
{
|
243 |
|
|
/* Holds types of allowed speculations: BEGIN_{DATA|CONTROL},
|
244 |
|
|
BE_IN_{DATA_CONTROL}. */
|
245 |
|
|
int mask;
|
246 |
|
|
|
247 |
|
|
/* A dump file for additional information on speculative scheduling. */
|
248 |
|
|
FILE *dump;
|
249 |
|
|
|
250 |
|
|
/* Minimal cumulative weakness of speculative instruction's
|
251 |
|
|
dependencies, so that insn will be scheduled. */
|
252 |
|
|
dw_t weakness_cutoff;
|
253 |
|
|
|
254 |
|
|
/* Flags from the enum SPEC_SCHED_FLAGS. */
|
255 |
|
|
int flags;
|
256 |
|
|
};
|
257 |
|
|
typedef struct spec_info_def *spec_info_t;
|
258 |
|
|
|
259 |
|
|
extern struct sched_info *current_sched_info;
|
260 |
|
|
|
261 |
|
|
/* Indexed by INSN_UID, the collection of all data associated with
|
262 |
|
|
a single instruction. */
|
263 |
|
|
|
264 |
|
|
struct haifa_insn_data
|
265 |
|
|
{
|
266 |
|
|
/* A list of insns which depend on the instruction. Unlike LOG_LINKS,
|
267 |
|
|
it represents forward dependencies. */
|
268 |
|
|
rtx depend;
|
269 |
|
|
|
270 |
|
|
/* A list of scheduled producers of the instruction. Links are being moved
|
271 |
|
|
from LOG_LINKS to RESOLVED_DEPS during scheduling. */
|
272 |
|
|
rtx resolved_deps;
|
273 |
|
|
|
274 |
|
|
/* The line number note in effect for each insn. For line number
|
275 |
|
|
notes, this indicates whether the note may be reused. */
|
276 |
|
|
rtx line_note;
|
277 |
|
|
|
278 |
|
|
/* Logical uid gives the original ordering of the insns. */
|
279 |
|
|
int luid;
|
280 |
|
|
|
281 |
|
|
/* A priority for each insn. */
|
282 |
|
|
int priority;
|
283 |
|
|
|
284 |
|
|
/* The number of incoming edges in the forward dependency graph.
|
285 |
|
|
As scheduling proceeds, counts are decreased. An insn moves to
|
286 |
|
|
the ready queue when its counter reaches zero. */
|
287 |
|
|
int dep_count;
|
288 |
|
|
|
289 |
|
|
/* Number of instructions referring to this insn. */
|
290 |
|
|
int ref_count;
|
291 |
|
|
|
292 |
|
|
/* The minimum clock tick at which the insn becomes ready. This is
|
293 |
|
|
used to note timing constraints for the insns in the pending list. */
|
294 |
|
|
int tick;
|
295 |
|
|
|
296 |
|
|
/* INTER_TICK is used to adjust INSN_TICKs of instructions from the
|
297 |
|
|
subsequent blocks in a region. */
|
298 |
|
|
int inter_tick;
|
299 |
|
|
|
300 |
|
|
/* See comment on QUEUE_INDEX macro in haifa-sched.c. */
|
301 |
|
|
int queue_index;
|
302 |
|
|
|
303 |
|
|
short cost;
|
304 |
|
|
|
305 |
|
|
/* This weight is an estimation of the insn's contribution to
|
306 |
|
|
register pressure. */
|
307 |
|
|
short reg_weight;
|
308 |
|
|
|
309 |
|
|
/* Some insns (e.g. call) are not allowed to move across blocks. */
|
310 |
|
|
unsigned int cant_move : 1;
|
311 |
|
|
|
312 |
|
|
/* Set if there's DEF-USE dependence between some speculatively
|
313 |
|
|
moved load insn and this one. */
|
314 |
|
|
unsigned int fed_by_spec_load : 1;
|
315 |
|
|
unsigned int is_load_insn : 1;
|
316 |
|
|
|
317 |
|
|
/* Nonzero if priority has been computed already. */
|
318 |
|
|
unsigned int priority_known : 1;
|
319 |
|
|
|
320 |
|
|
/* Nonzero if instruction has internal dependence
|
321 |
|
|
(e.g. add_dependence was invoked with (insn == elem)). */
|
322 |
|
|
unsigned int has_internal_dep : 1;
|
323 |
|
|
|
324 |
|
|
/* What speculations are necessary to apply to schedule the instruction. */
|
325 |
|
|
ds_t todo_spec;
|
326 |
|
|
/* What speculations were already applied. */
|
327 |
|
|
ds_t done_spec;
|
328 |
|
|
/* What speculations are checked by this instruction. */
|
329 |
|
|
ds_t check_spec;
|
330 |
|
|
|
331 |
|
|
/* Recovery block for speculation checks. */
|
332 |
|
|
basic_block recovery_block;
|
333 |
|
|
|
334 |
|
|
/* Original pattern of the instruction. */
|
335 |
|
|
rtx orig_pat;
|
336 |
|
|
};
|
337 |
|
|
|
338 |
|
|
extern struct haifa_insn_data *h_i_d;
|
339 |
|
|
/* Used only if (current_sched_info->flags & USE_GLAT) != 0.
|
340 |
|
|
These regsets store global_live_at_{start, end} information
|
341 |
|
|
for each basic block. */
|
342 |
|
|
extern regset *glat_start, *glat_end;
|
343 |
|
|
|
344 |
|
|
/* Accessor macros for h_i_d. There are more in haifa-sched.c and
|
345 |
|
|
sched-rgn.c. */
|
346 |
|
|
#define INSN_DEPEND(INSN) (h_i_d[INSN_UID (INSN)].depend)
|
347 |
|
|
#define RESOLVED_DEPS(INSN) (h_i_d[INSN_UID (INSN)].resolved_deps)
|
348 |
|
|
#define INSN_LUID(INSN) (h_i_d[INSN_UID (INSN)].luid)
|
349 |
|
|
#define CANT_MOVE(insn) (h_i_d[INSN_UID (insn)].cant_move)
|
350 |
|
|
#define INSN_DEP_COUNT(INSN) (h_i_d[INSN_UID (INSN)].dep_count)
|
351 |
|
|
#define INSN_PRIORITY(INSN) (h_i_d[INSN_UID (INSN)].priority)
|
352 |
|
|
#define INSN_PRIORITY_KNOWN(INSN) (h_i_d[INSN_UID (INSN)].priority_known)
|
353 |
|
|
#define INSN_COST(INSN) (h_i_d[INSN_UID (INSN)].cost)
|
354 |
|
|
#define INSN_REG_WEIGHT(INSN) (h_i_d[INSN_UID (INSN)].reg_weight)
|
355 |
|
|
#define HAS_INTERNAL_DEP(INSN) (h_i_d[INSN_UID (INSN)].has_internal_dep)
|
356 |
|
|
#define TODO_SPEC(INSN) (h_i_d[INSN_UID (INSN)].todo_spec)
|
357 |
|
|
#define DONE_SPEC(INSN) (h_i_d[INSN_UID (INSN)].done_spec)
|
358 |
|
|
#define CHECK_SPEC(INSN) (h_i_d[INSN_UID (INSN)].check_spec)
|
359 |
|
|
#define RECOVERY_BLOCK(INSN) (h_i_d[INSN_UID (INSN)].recovery_block)
|
360 |
|
|
#define ORIG_PAT(INSN) (h_i_d[INSN_UID (INSN)].orig_pat)
|
361 |
|
|
|
362 |
|
|
/* INSN is either a simple or a branchy speculation check. */
|
363 |
|
|
#define IS_SPECULATION_CHECK_P(INSN) (RECOVERY_BLOCK (INSN) != NULL)
|
364 |
|
|
|
365 |
|
|
/* INSN is a speculation check that will simply reexecute the speculatively
|
366 |
|
|
scheduled instruction if the speculation fails. */
|
367 |
|
|
#define IS_SPECULATION_SIMPLE_CHECK_P(INSN) \
|
368 |
|
|
(RECOVERY_BLOCK (INSN) == EXIT_BLOCK_PTR)
|
369 |
|
|
|
370 |
|
|
/* INSN is a speculation check that will branch to RECOVERY_BLOCK if the
|
371 |
|
|
speculation fails. Insns in that block will reexecute the speculatively
|
372 |
|
|
scheduled code and then will return immediately after INSN thus preserving
|
373 |
|
|
semantics of the program. */
|
374 |
|
|
#define IS_SPECULATION_BRANCHY_CHECK_P(INSN) \
|
375 |
|
|
(RECOVERY_BLOCK (INSN) != NULL && RECOVERY_BLOCK (INSN) != EXIT_BLOCK_PTR)
|
376 |
|
|
|
377 |
|
|
/* DEP_STATUS of the link encapsulates information, that is needed for
|
378 |
|
|
speculative scheduling. Namely, it is 4 integers in the range
|
379 |
|
|
[0, MAX_DEP_WEAK] and 3 bits.
|
380 |
|
|
The integers correspond to the probability of the dependence to *not*
|
381 |
|
|
exist, it is the probability, that overcoming of this dependence will
|
382 |
|
|
not be followed by execution of the recovery code. Nevertheless,
|
383 |
|
|
whatever high the probability of success is, recovery code should still
|
384 |
|
|
be generated to preserve semantics of the program. To find a way to
|
385 |
|
|
get/set these integers, please refer to the {get, set}_dep_weak ()
|
386 |
|
|
functions in sched-deps.c .
|
387 |
|
|
The 3 bits in the DEP_STATUS correspond to 3 dependence types: true-,
|
388 |
|
|
output- and anti- dependence. It is not enough for speculative scheduling
|
389 |
|
|
to know just the major type of all the dependence between two instructions,
|
390 |
|
|
as only true dependence can be overcome.
|
391 |
|
|
There also is the 4-th bit in the DEP_STATUS (HARD_DEP), that is reserved
|
392 |
|
|
for using to describe instruction's status. It is set whenever instruction
|
393 |
|
|
has at least one dependence, that cannot be overcome.
|
394 |
|
|
See also: check_dep_status () in sched-deps.c . */
|
395 |
|
|
#define DEP_STATUS(LINK) XINT (LINK, 2)
|
396 |
|
|
|
397 |
|
|
/* We exclude sign bit. */
|
398 |
|
|
#define BITS_PER_DEP_STATUS (HOST_BITS_PER_INT - 1)
|
399 |
|
|
|
400 |
|
|
/* First '4' stands for 3 dep type bits and HARD_DEP bit.
|
401 |
|
|
Second '4' stands for BEGIN_{DATA, CONTROL}, BE_IN_{DATA, CONTROL}
|
402 |
|
|
dep weakness. */
|
403 |
|
|
#define BITS_PER_DEP_WEAK ((BITS_PER_DEP_STATUS - 4) / 4)
|
404 |
|
|
|
405 |
|
|
/* Mask of speculative weakness in dep_status. */
|
406 |
|
|
#define DEP_WEAK_MASK ((1 << BITS_PER_DEP_WEAK) - 1)
|
407 |
|
|
|
408 |
|
|
/* This constant means that dependence is fake with 99.999...% probability.
|
409 |
|
|
This is the maximum value, that can appear in dep_status.
|
410 |
|
|
Note, that we don't want MAX_DEP_WEAK to be the same as DEP_WEAK_MASK for
|
411 |
|
|
debugging reasons. Though, it can be set to DEP_WEAK_MASK, and, when
|
412 |
|
|
done so, we'll get fast (mul for)/(div by) NO_DEP_WEAK. */
|
413 |
|
|
#define MAX_DEP_WEAK (DEP_WEAK_MASK - 1)
|
414 |
|
|
|
415 |
|
|
/* This constant means that dependence is 99.999...% real and it is a really
|
416 |
|
|
bad idea to overcome it (though this can be done, preserving program
|
417 |
|
|
semantics). */
|
418 |
|
|
#define MIN_DEP_WEAK 1
|
419 |
|
|
|
420 |
|
|
/* This constant represents 100% probability.
|
421 |
|
|
E.g. it is used to represent weakness of dependence, that doesn't exist. */
|
422 |
|
|
#define NO_DEP_WEAK (MAX_DEP_WEAK + MIN_DEP_WEAK)
|
423 |
|
|
|
424 |
|
|
/* Default weakness of speculative dependence. Used when we can't say
|
425 |
|
|
neither bad nor good about the dependence. */
|
426 |
|
|
#define UNCERTAIN_DEP_WEAK (MAX_DEP_WEAK - MAX_DEP_WEAK / 4)
|
427 |
|
|
|
428 |
|
|
/* Offset for speculative weaknesses in dep_status. */
|
429 |
|
|
enum SPEC_TYPES_OFFSETS {
|
430 |
|
|
BEGIN_DATA_BITS_OFFSET = 0,
|
431 |
|
|
BE_IN_DATA_BITS_OFFSET = BEGIN_DATA_BITS_OFFSET + BITS_PER_DEP_WEAK,
|
432 |
|
|
BEGIN_CONTROL_BITS_OFFSET = BE_IN_DATA_BITS_OFFSET + BITS_PER_DEP_WEAK,
|
433 |
|
|
BE_IN_CONTROL_BITS_OFFSET = BEGIN_CONTROL_BITS_OFFSET + BITS_PER_DEP_WEAK
|
434 |
|
|
};
|
435 |
|
|
|
436 |
|
|
/* The following defines provide numerous constants used to distinguish between
|
437 |
|
|
different types of speculative dependencies. */
|
438 |
|
|
|
439 |
|
|
/* Dependence can be overcome with generation of new data speculative
|
440 |
|
|
instruction. */
|
441 |
|
|
#define BEGIN_DATA (((ds_t) DEP_WEAK_MASK) << BEGIN_DATA_BITS_OFFSET)
|
442 |
|
|
|
443 |
|
|
/* This dependence is to the instruction in the recovery block, that was
|
444 |
|
|
formed to recover after data-speculation failure.
|
445 |
|
|
Thus, this dependence can overcome with generating of the copy of
|
446 |
|
|
this instruction in the recovery block. */
|
447 |
|
|
#define BE_IN_DATA (((ds_t) DEP_WEAK_MASK) << BE_IN_DATA_BITS_OFFSET)
|
448 |
|
|
|
449 |
|
|
/* Dependence can be overcome with generation of new control speculative
|
450 |
|
|
instruction. */
|
451 |
|
|
#define BEGIN_CONTROL (((ds_t) DEP_WEAK_MASK) << BEGIN_CONTROL_BITS_OFFSET)
|
452 |
|
|
|
453 |
|
|
/* This dependence is to the instruction in the recovery block, that was
|
454 |
|
|
formed to recover after control-speculation failure.
|
455 |
|
|
Thus, this dependence can be overcome with generating of the copy of
|
456 |
|
|
this instruction in the recovery block. */
|
457 |
|
|
#define BE_IN_CONTROL (((ds_t) DEP_WEAK_MASK) << BE_IN_CONTROL_BITS_OFFSET)
|
458 |
|
|
|
459 |
|
|
/* A few convenient combinations. */
|
460 |
|
|
#define BEGIN_SPEC (BEGIN_DATA | BEGIN_CONTROL)
|
461 |
|
|
#define DATA_SPEC (BEGIN_DATA | BE_IN_DATA)
|
462 |
|
|
#define CONTROL_SPEC (BEGIN_CONTROL | BE_IN_CONTROL)
|
463 |
|
|
#define SPECULATIVE (DATA_SPEC | CONTROL_SPEC)
|
464 |
|
|
#define BE_IN_SPEC (BE_IN_DATA | BE_IN_CONTROL)
|
465 |
|
|
|
466 |
|
|
/* Constants, that are helpful in iterating through dep_status. */
|
467 |
|
|
#define FIRST_SPEC_TYPE BEGIN_DATA
|
468 |
|
|
#define LAST_SPEC_TYPE BE_IN_CONTROL
|
469 |
|
|
#define SPEC_TYPE_SHIFT BITS_PER_DEP_WEAK
|
470 |
|
|
|
471 |
|
|
/* Dependence on instruction can be of multiple types
|
472 |
|
|
(e.g. true and output). This fields enhance REG_NOTE_KIND information
|
473 |
|
|
of the dependence. */
|
474 |
|
|
#define DEP_TRUE (((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + BITS_PER_DEP_WEAK))
|
475 |
|
|
#define DEP_OUTPUT (DEP_TRUE << 1)
|
476 |
|
|
#define DEP_ANTI (DEP_OUTPUT << 1)
|
477 |
|
|
|
478 |
|
|
#define DEP_TYPES (DEP_TRUE | DEP_OUTPUT | DEP_ANTI)
|
479 |
|
|
|
480 |
|
|
/* Instruction has non-speculative dependence. This bit represents the
|
481 |
|
|
property of an instruction - not the one of a dependence.
|
482 |
|
|
Therefore, it can appear only in TODO_SPEC field of an instruction. */
|
483 |
|
|
#define HARD_DEP (DEP_ANTI << 1)
|
484 |
|
|
|
485 |
|
|
/* This represents the results of calling sched-deps.c functions,
|
486 |
|
|
which modify dependencies. Possible choices are: a dependence
|
487 |
|
|
is already present and nothing has been changed; a dependence type
|
488 |
|
|
has been changed; brand new dependence has been created. */
|
489 |
|
|
enum DEPS_ADJUST_RESULT {
|
490 |
|
|
DEP_PRESENT = 1,
|
491 |
|
|
DEP_CHANGED = 2,
|
492 |
|
|
DEP_CREATED = 3
|
493 |
|
|
};
|
494 |
|
|
|
495 |
|
|
/* Represents the bits that can be set in the flags field of the
|
496 |
|
|
sched_info structure. */
|
497 |
|
|
enum SCHED_FLAGS {
|
498 |
|
|
/* If set, generate links between instruction as DEPS_LIST.
|
499 |
|
|
Otherwise, generate usual INSN_LIST links. */
|
500 |
|
|
USE_DEPS_LIST = 1,
|
501 |
|
|
/* Perform data or control (or both) speculation.
|
502 |
|
|
Results in generation of data and control speculative dependencies.
|
503 |
|
|
Requires USE_DEPS_LIST set. */
|
504 |
|
|
DO_SPECULATION = USE_DEPS_LIST << 1,
|
505 |
|
|
SCHED_RGN = DO_SPECULATION << 1,
|
506 |
|
|
SCHED_EBB = SCHED_RGN << 1,
|
507 |
|
|
/* Detach register live information from basic block headers.
|
508 |
|
|
This is necessary to invoke functions, that change CFG (e.g. split_edge).
|
509 |
|
|
Requires USE_GLAT. */
|
510 |
|
|
DETACH_LIFE_INFO = SCHED_EBB << 1,
|
511 |
|
|
/* Save register live information from basic block headers to
|
512 |
|
|
glat_{start, end} arrays. */
|
513 |
|
|
USE_GLAT = DETACH_LIFE_INFO << 1
|
514 |
|
|
};
|
515 |
|
|
|
516 |
|
|
enum SPEC_SCHED_FLAGS {
|
517 |
|
|
COUNT_SPEC_IN_CRITICAL_PATH = 1,
|
518 |
|
|
PREFER_NON_DATA_SPEC = COUNT_SPEC_IN_CRITICAL_PATH << 1,
|
519 |
|
|
PREFER_NON_CONTROL_SPEC = PREFER_NON_DATA_SPEC << 1
|
520 |
|
|
};
|
521 |
|
|
|
522 |
|
|
#define NOTE_NOT_BB_P(NOTE) (NOTE_P (NOTE) && (NOTE_LINE_NUMBER (NOTE) \
|
523 |
|
|
!= NOTE_INSN_BASIC_BLOCK))
|
524 |
|
|
|
525 |
|
|
extern FILE *sched_dump;
|
526 |
|
|
extern int sched_verbose;
|
527 |
|
|
|
528 |
|
|
/* Exception Free Loads:
|
529 |
|
|
|
530 |
|
|
We define five classes of speculative loads: IFREE, IRISKY,
|
531 |
|
|
PFREE, PRISKY, and MFREE.
|
532 |
|
|
|
533 |
|
|
IFREE loads are loads that are proved to be exception-free, just
|
534 |
|
|
by examining the load insn. Examples for such loads are loads
|
535 |
|
|
from TOC and loads of global data.
|
536 |
|
|
|
537 |
|
|
IRISKY loads are loads that are proved to be exception-risky,
|
538 |
|
|
just by examining the load insn. Examples for such loads are
|
539 |
|
|
volatile loads and loads from shared memory.
|
540 |
|
|
|
541 |
|
|
PFREE loads are loads for which we can prove, by examining other
|
542 |
|
|
insns, that they are exception-free. Currently, this class consists
|
543 |
|
|
of loads for which we are able to find a "similar load", either in
|
544 |
|
|
the target block, or, if only one split-block exists, in that split
|
545 |
|
|
block. Load2 is similar to load1 if both have same single base
|
546 |
|
|
register. We identify only part of the similar loads, by finding
|
547 |
|
|
an insn upon which both load1 and load2 have a DEF-USE dependence.
|
548 |
|
|
|
549 |
|
|
PRISKY loads are loads for which we can prove, by examining other
|
550 |
|
|
insns, that they are exception-risky. Currently we have two proofs for
|
551 |
|
|
such loads. The first proof detects loads that are probably guarded by a
|
552 |
|
|
test on the memory address. This proof is based on the
|
553 |
|
|
backward and forward data dependence information for the region.
|
554 |
|
|
Let load-insn be the examined load.
|
555 |
|
|
Load-insn is PRISKY iff ALL the following hold:
|
556 |
|
|
|
557 |
|
|
- insn1 is not in the same block as load-insn
|
558 |
|
|
- there is a DEF-USE dependence chain (insn1, ..., load-insn)
|
559 |
|
|
- test-insn is either a compare or a branch, not in the same block
|
560 |
|
|
as load-insn
|
561 |
|
|
- load-insn is reachable from test-insn
|
562 |
|
|
- there is a DEF-USE dependence chain (insn1, ..., test-insn)
|
563 |
|
|
|
564 |
|
|
This proof might fail when the compare and the load are fed
|
565 |
|
|
by an insn not in the region. To solve this, we will add to this
|
566 |
|
|
group all loads that have no input DEF-USE dependence.
|
567 |
|
|
|
568 |
|
|
The second proof detects loads that are directly or indirectly
|
569 |
|
|
fed by a speculative load. This proof is affected by the
|
570 |
|
|
scheduling process. We will use the flag fed_by_spec_load.
|
571 |
|
|
Initially, all insns have this flag reset. After a speculative
|
572 |
|
|
motion of an insn, if insn is either a load, or marked as
|
573 |
|
|
fed_by_spec_load, we will also mark as fed_by_spec_load every
|
574 |
|
|
insn1 for which a DEF-USE dependence (insn, insn1) exists. A
|
575 |
|
|
load which is fed_by_spec_load is also PRISKY.
|
576 |
|
|
|
577 |
|
|
MFREE (maybe-free) loads are all the remaining loads. They may be
|
578 |
|
|
exception-free, but we cannot prove it.
|
579 |
|
|
|
580 |
|
|
Now, all loads in IFREE and PFREE classes are considered
|
581 |
|
|
exception-free, while all loads in IRISKY and PRISKY classes are
|
582 |
|
|
considered exception-risky. As for loads in the MFREE class,
|
583 |
|
|
these are considered either exception-free or exception-risky,
|
584 |
|
|
depending on whether we are pessimistic or optimistic. We have
|
585 |
|
|
to take the pessimistic approach to assure the safety of
|
586 |
|
|
speculative scheduling, but we can take the optimistic approach
|
587 |
|
|
by invoking the -fsched_spec_load_dangerous option. */
|
588 |
|
|
|
589 |
|
|
enum INSN_TRAP_CLASS
|
590 |
|
|
{
|
591 |
|
|
TRAP_FREE = 0, IFREE = 1, PFREE_CANDIDATE = 2,
|
592 |
|
|
PRISKY_CANDIDATE = 3, IRISKY = 4, TRAP_RISKY = 5
|
593 |
|
|
};
|
594 |
|
|
|
595 |
|
|
#define WORST_CLASS(class1, class2) \
|
596 |
|
|
((class1 > class2) ? class1 : class2)
|
597 |
|
|
|
598 |
|
|
#ifndef __GNUC__
|
599 |
|
|
#define __inline
|
600 |
|
|
#endif
|
601 |
|
|
|
602 |
|
|
#ifndef HAIFA_INLINE
|
603 |
|
|
#define HAIFA_INLINE __inline
|
604 |
|
|
#endif
|
605 |
|
|
|
606 |
|
|
/* Functions in sched-vis.c. */
|
607 |
|
|
extern void print_insn (char *, rtx, int);
|
608 |
|
|
|
609 |
|
|
/* Functions in sched-deps.c. */
|
610 |
|
|
extern bool sched_insns_conditions_mutex_p (rtx, rtx);
|
611 |
|
|
extern void add_dependence (rtx, rtx, enum reg_note);
|
612 |
|
|
extern void sched_analyze (struct deps *, rtx, rtx);
|
613 |
|
|
extern void init_deps (struct deps *);
|
614 |
|
|
extern void free_deps (struct deps *);
|
615 |
|
|
extern void init_deps_global (void);
|
616 |
|
|
extern void finish_deps_global (void);
|
617 |
|
|
extern void add_forw_dep (rtx, rtx);
|
618 |
|
|
extern void compute_forward_dependences (rtx, rtx);
|
619 |
|
|
extern rtx find_insn_list (rtx, rtx);
|
620 |
|
|
extern void init_dependency_caches (int);
|
621 |
|
|
extern void free_dependency_caches (void);
|
622 |
|
|
extern void extend_dependency_caches (int, bool);
|
623 |
|
|
extern enum DEPS_ADJUST_RESULT add_or_update_back_dep (rtx, rtx,
|
624 |
|
|
enum reg_note, ds_t);
|
625 |
|
|
extern void add_or_update_back_forw_dep (rtx, rtx, enum reg_note, ds_t);
|
626 |
|
|
extern void add_back_forw_dep (rtx, rtx, enum reg_note, ds_t);
|
627 |
|
|
extern void delete_back_forw_dep (rtx, rtx);
|
628 |
|
|
extern dw_t get_dep_weak (ds_t, ds_t);
|
629 |
|
|
extern ds_t set_dep_weak (ds_t, ds_t, dw_t);
|
630 |
|
|
extern ds_t ds_merge (ds_t, ds_t);
|
631 |
|
|
|
632 |
|
|
/* Functions in haifa-sched.c. */
|
633 |
|
|
extern int haifa_classify_insn (rtx);
|
634 |
|
|
extern void get_ebb_head_tail (basic_block, basic_block, rtx *, rtx *);
|
635 |
|
|
extern int no_real_insns_p (rtx, rtx);
|
636 |
|
|
|
637 |
|
|
extern void rm_line_notes (rtx, rtx);
|
638 |
|
|
extern void save_line_notes (int, rtx, rtx);
|
639 |
|
|
extern void restore_line_notes (rtx, rtx);
|
640 |
|
|
extern void rm_redundant_line_notes (void);
|
641 |
|
|
extern void rm_other_notes (rtx, rtx);
|
642 |
|
|
|
643 |
|
|
extern int insn_cost (rtx, rtx, rtx);
|
644 |
|
|
extern int set_priorities (rtx, rtx);
|
645 |
|
|
|
646 |
|
|
extern void schedule_block (basic_block *, int);
|
647 |
|
|
extern void sched_init (void);
|
648 |
|
|
extern void sched_finish (void);
|
649 |
|
|
|
650 |
|
|
extern int try_ready (rtx);
|
651 |
|
|
extern void * xrecalloc (void *, size_t, size_t, size_t);
|
652 |
|
|
extern void unlink_bb_notes (basic_block, basic_block);
|
653 |
|
|
extern void add_block (basic_block, basic_block);
|
654 |
|
|
extern void attach_life_info (void);
|
655 |
|
|
extern rtx bb_note (basic_block);
|
656 |
|
|
|
657 |
|
|
#ifdef ENABLE_CHECKING
|
658 |
|
|
extern void check_reg_live (bool);
|
659 |
|
|
#endif
|
660 |
|
|
|
661 |
|
|
#endif /* GCC_SCHED_INT_H */
|