File: | build/gcc/haifa-sched.cc |
Warning: | line 3240, column 52 The right operand of '!=' is a garbage value |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | /* Instruction scheduling pass. | |||
2 | Copyright (C) 1992-2023 Free Software Foundation, Inc. | |||
3 | Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by, | |||
4 | and currently maintained by, Jim Wilson (wilson@cygnus.com) | |||
5 | ||||
6 | This file is part of GCC. | |||
7 | ||||
8 | GCC is free software; you can redistribute it and/or modify it under | |||
9 | the terms of the GNU General Public License as published by the Free | |||
10 | Software Foundation; either version 3, or (at your option) any later | |||
11 | version. | |||
12 | ||||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |||
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |||
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |||
16 | for more details. | |||
17 | ||||
18 | You should have received a copy of the GNU General Public License | |||
19 | along with GCC; see the file COPYING3. If not see | |||
20 | <http://www.gnu.org/licenses/>. */ | |||
21 | ||||
22 | /* Instruction scheduling pass. This file, along with sched-deps.cc, | |||
23 | contains the generic parts. The actual entry point for | |||
24 | the normal instruction scheduling pass is found in sched-rgn.cc. | |||
25 | ||||
26 | We compute insn priorities based on data dependencies. Flow | |||
27 | analysis only creates a fraction of the data-dependencies we must | |||
28 | observe: namely, only those dependencies which the combiner can be | |||
29 | expected to use. For this pass, we must therefore create the | |||
30 | remaining dependencies we need to observe: register dependencies, | |||
31 | memory dependencies, dependencies to keep function calls in order, | |||
32 | and the dependence between a conditional branch and the setting of | |||
33 | condition codes are all dealt with here. | |||
34 | ||||
35 | The scheduler first traverses the data flow graph, starting with | |||
36 | the last instruction, and proceeding to the first, assigning values | |||
37 | to insn_priority as it goes. This sorts the instructions | |||
38 | topologically by data dependence. | |||
39 | ||||
40 | Once priorities have been established, we order the insns using | |||
41 | list scheduling. This works as follows: starting with a list of | |||
42 | all the ready insns, and sorted according to priority number, we | |||
43 | schedule the insn from the end of the list by placing its | |||
44 | predecessors in the list according to their priority order. We | |||
45 | consider this insn scheduled by setting the pointer to the "end" of | |||
46 | the list to point to the previous insn. When an insn has no | |||
47 | predecessors, we either queue it until sufficient time has elapsed | |||
48 | or add it to the ready list. As the instructions are scheduled or | |||
49 | when stalls are introduced, the queue advances and dumps insns into | |||
50 | the ready list. When all insns down to the lowest priority have | |||
51 | been scheduled, the critical path of the basic block has been made | |||
52 | as short as possible. The remaining insns are then scheduled in | |||
53 | remaining slots. | |||
54 | ||||
55 | The following list shows the order in which we want to break ties | |||
56 | among insns in the ready list: | |||
57 | ||||
58 | 1. choose insn with the longest path to end of bb, ties | |||
59 | broken by | |||
60 | 2. choose insn with least contribution to register pressure, | |||
61 | ties broken by | |||
62 | 3. prefer in-block upon interblock motion, ties broken by | |||
63 | 4. prefer useful upon speculative motion, ties broken by | |||
64 | 5. choose insn with largest control flow probability, ties | |||
65 | broken by | |||
66 | 6. choose insn with the least dependences upon the previously | |||
67 | scheduled insn, or finally | |||
68 | 7 choose the insn which has the most insns dependent on it. | |||
69 | 8. choose insn with lowest UID. | |||
70 | ||||
71 | Memory references complicate matters. Only if we can be certain | |||
72 | that memory references are not part of the data dependency graph | |||
73 | (via true, anti, or output dependence), can we move operations past | |||
74 | memory references. To first approximation, reads can be done | |||
75 | independently, while writes introduce dependencies. Better | |||
76 | approximations will yield fewer dependencies. | |||
77 | ||||
78 | Before reload, an extended analysis of interblock data dependences | |||
79 | is required for interblock scheduling. This is performed in | |||
80 | compute_block_dependences (). | |||
81 | ||||
82 | Dependencies set up by memory references are treated in exactly the | |||
83 | same way as other dependencies, by using insn backward dependences | |||
84 | INSN_BACK_DEPS. INSN_BACK_DEPS are translated into forward dependences | |||
85 | INSN_FORW_DEPS for the purpose of forward list scheduling. | |||
86 | ||||
87 | Having optimized the critical path, we may have also unduly | |||
88 | extended the lifetimes of some registers. If an operation requires | |||
89 | that constants be loaded into registers, it is certainly desirable | |||
90 | to load those constants as early as necessary, but no earlier. | |||
91 | I.e., it will not do to load up a bunch of registers at the | |||
92 | beginning of a basic block only to use them at the end, if they | |||
93 | could be loaded later, since this may result in excessive register | |||
94 | utilization. | |||
95 | ||||
96 | Note that since branches are never in basic blocks, but only end | |||
97 | basic blocks, this pass will not move branches. But that is ok, | |||
98 | since we can use GNU's delayed branch scheduling pass to take care | |||
99 | of this case. | |||
100 | ||||
101 | Also note that no further optimizations based on algebraic | |||
102 | identities are performed, so this pass would be a good one to | |||
103 | perform instruction splitting, such as breaking up a multiply | |||
104 | instruction into shifts and adds where that is profitable. | |||
105 | ||||
106 | Given the memory aliasing analysis that this pass should perform, | |||
107 | it should be possible to remove redundant stores to memory, and to | |||
108 | load values from registers instead of hitting memory. | |||
109 | ||||
110 | Before reload, speculative insns are moved only if a 'proof' exists | |||
111 | that no exception will be caused by this, and if no live registers | |||
112 | exist that inhibit the motion (live registers constraints are not | |||
113 | represented by data dependence edges). | |||
114 | ||||
115 | This pass must update information that subsequent passes expect to | |||
116 | be correct. Namely: reg_n_refs, reg_n_sets, reg_n_deaths, | |||
117 | reg_n_calls_crossed, and reg_live_length. Also, BB_HEAD, BB_END. | |||
118 | ||||
119 | The information in the line number notes is carefully retained by | |||
120 | this pass. Notes that refer to the starting and ending of | |||
121 | exception regions are also carefully retained by this pass. All | |||
122 | other NOTE insns are grouped in their same relative order at the | |||
123 | beginning of basic blocks and regions that have been scheduled. */ | |||
124 | ||||
125 | #include "config.h" | |||
126 | #include "system.h" | |||
127 | #include "coretypes.h" | |||
128 | #include "backend.h" | |||
129 | #include "target.h" | |||
130 | #include "rtl.h" | |||
131 | #include "cfghooks.h" | |||
132 | #include "df.h" | |||
133 | #include "memmodel.h" | |||
134 | #include "tm_p.h" | |||
135 | #include "insn-config.h" | |||
136 | #include "regs.h" | |||
137 | #include "ira.h" | |||
138 | #include "recog.h" | |||
139 | #include "insn-attr.h" | |||
140 | #include "cfgrtl.h" | |||
141 | #include "cfgbuild.h" | |||
142 | #include "sched-int.h" | |||
143 | #include "common/common-target.h" | |||
144 | #include "dbgcnt.h" | |||
145 | #include "cfgloop.h" | |||
146 | #include "dumpfile.h" | |||
147 | #include "print-rtl.h" | |||
148 | #include "function-abi.h" | |||
149 | ||||
150 | #ifdef INSN_SCHEDULING | |||
151 | ||||
152 | /* True if we do register pressure relief through live-range | |||
153 | shrinkage. */ | |||
154 | static bool live_range_shrinkage_p; | |||
155 | ||||
156 | /* Switch on live range shrinkage. */ | |||
157 | void | |||
158 | initialize_live_range_shrinkage (void) | |||
159 | { | |||
160 | live_range_shrinkage_p = true; | |||
161 | } | |||
162 | ||||
163 | /* Switch off live range shrinkage. */ | |||
164 | void | |||
165 | finish_live_range_shrinkage (void) | |||
166 | { | |||
167 | live_range_shrinkage_p = false; | |||
168 | } | |||
169 | ||||
170 | /* issue_rate is the number of insns that can be scheduled in the same | |||
171 | machine cycle. It can be defined in the config/mach/mach.h file, | |||
172 | otherwise we set it to 1. */ | |||
173 | ||||
174 | int issue_rate; | |||
175 | ||||
176 | /* This can be set to true by a backend if the scheduler should not | |||
177 | enable a DCE pass. */ | |||
178 | bool sched_no_dce; | |||
179 | ||||
180 | /* The current initiation interval used when modulo scheduling. */ | |||
181 | static int modulo_ii; | |||
182 | ||||
183 | /* The maximum number of stages we are prepared to handle. */ | |||
184 | static int modulo_max_stages; | |||
185 | ||||
186 | /* The number of insns that exist in each iteration of the loop. We use this | |||
187 | to detect when we've scheduled all insns from the first iteration. */ | |||
188 | static int modulo_n_insns; | |||
189 | ||||
190 | /* The current count of insns in the first iteration of the loop that have | |||
191 | already been scheduled. */ | |||
192 | static int modulo_insns_scheduled; | |||
193 | ||||
194 | /* The maximum uid of insns from the first iteration of the loop. */ | |||
195 | static int modulo_iter0_max_uid; | |||
196 | ||||
197 | /* The number of times we should attempt to backtrack when modulo scheduling. | |||
198 | Decreased each time we have to backtrack. */ | |||
199 | static int modulo_backtracks_left; | |||
200 | ||||
201 | /* The stage in which the last insn from the original loop was | |||
202 | scheduled. */ | |||
203 | static int modulo_last_stage; | |||
204 | ||||
205 | /* sched-verbose controls the amount of debugging output the | |||
206 | scheduler prints. It is controlled by -fsched-verbose=N: | |||
207 | N=0: no debugging output. | |||
208 | N=1: default value. | |||
209 | N=2: bb's probabilities, detailed ready list info, unit/insn info. | |||
210 | N=3: rtl at abort point, control-flow, regions info. | |||
211 | N=5: dependences info. */ | |||
212 | int sched_verbose = 0; | |||
213 | ||||
214 | /* Debugging file. All printouts are sent to dump. */ | |||
215 | FILE *sched_dump = 0; | |||
216 | ||||
217 | /* This is a placeholder for the scheduler parameters common | |||
218 | to all schedulers. */ | |||
219 | struct common_sched_info_def *common_sched_info; | |||
220 | ||||
221 | #define INSN_TICK(INSN)((&h_i_d[INSN_UID (INSN)])->tick) (HID (INSN)(&h_i_d[INSN_UID (INSN)])->tick) | |||
222 | #define INSN_EXACT_TICK(INSN)((&h_i_d[INSN_UID (INSN)])->exact_tick) (HID (INSN)(&h_i_d[INSN_UID (INSN)])->exact_tick) | |||
223 | #define INSN_TICK_ESTIMATE(INSN)((&h_i_d[INSN_UID (INSN)])->tick_estimate) (HID (INSN)(&h_i_d[INSN_UID (INSN)])->tick_estimate) | |||
224 | #define INTER_TICK(INSN)((&h_i_d[INSN_UID (INSN)])->inter_tick) (HID (INSN)(&h_i_d[INSN_UID (INSN)])->inter_tick) | |||
225 | #define FEEDS_BACKTRACK_INSN(INSN)((&h_i_d[INSN_UID (INSN)])->feeds_backtrack_insn) (HID (INSN)(&h_i_d[INSN_UID (INSN)])->feeds_backtrack_insn) | |||
226 | #define SHADOW_P(INSN)((&h_i_d[INSN_UID (INSN)])->shadow_p) (HID (INSN)(&h_i_d[INSN_UID (INSN)])->shadow_p) | |||
227 | #define MUST_RECOMPUTE_SPEC_P(INSN)((&h_i_d[INSN_UID (INSN)])->must_recompute_spec) (HID (INSN)(&h_i_d[INSN_UID (INSN)])->must_recompute_spec) | |||
228 | /* Cached cost of the instruction. Use insn_sched_cost to get cost of the | |||
229 | insn. -1 here means that the field is not initialized. */ | |||
230 | #define INSN_COST(INSN)((&h_i_d[INSN_UID (INSN)])->cost) (HID (INSN)(&h_i_d[INSN_UID (INSN)])->cost) | |||
231 | ||||
232 | /* If INSN_TICK of an instruction is equal to INVALID_TICK, | |||
233 | then it should be recalculated from scratch. */ | |||
234 | #define INVALID_TICK(-(max_insn_queue_index + 1)) (-(max_insn_queue_index + 1)) | |||
235 | /* The minimal value of the INSN_TICK of an instruction. */ | |||
236 | #define MIN_TICK(-max_insn_queue_index) (-max_insn_queue_index) | |||
237 | ||||
238 | /* Original order of insns in the ready list. | |||
239 | Used to keep order of normal insns while separating DEBUG_INSNs. */ | |||
240 | #define INSN_RFS_DEBUG_ORIG_ORDER(INSN)((&h_i_d[INSN_UID (INSN)])->rfs_debug_orig_order) (HID (INSN)(&h_i_d[INSN_UID (INSN)])->rfs_debug_orig_order) | |||
241 | ||||
242 | /* The deciding reason for INSN's place in the ready list. */ | |||
243 | #define INSN_LAST_RFS_WIN(INSN)((&h_i_d[INSN_UID (INSN)])->last_rfs_win) (HID (INSN)(&h_i_d[INSN_UID (INSN)])->last_rfs_win) | |||
244 | ||||
245 | /* List of important notes we must keep around. This is a pointer to the | |||
246 | last element in the list. */ | |||
247 | rtx_insn *note_list; | |||
248 | ||||
249 | static struct spec_info_def spec_info_var; | |||
250 | /* Description of the speculative part of the scheduling. | |||
251 | If NULL - no speculation. */ | |||
252 | spec_info_t spec_info = NULLnullptr; | |||
253 | ||||
254 | /* True, if recovery block was added during scheduling of current block. | |||
255 | Used to determine, if we need to fix INSN_TICKs. */ | |||
256 | static bool haifa_recovery_bb_recently_added_p; | |||
257 | ||||
258 | /* True, if recovery block was added during this scheduling pass. | |||
259 | Used to determine if we should have empty memory pools of dependencies | |||
260 | after finishing current region. */ | |||
261 | bool haifa_recovery_bb_ever_added_p; | |||
262 | ||||
263 | /* Counters of different types of speculative instructions. */ | |||
264 | static int nr_begin_data, nr_be_in_data, nr_begin_control, nr_be_in_control; | |||
265 | ||||
266 | /* Array used in {unlink, restore}_bb_notes. */ | |||
267 | static rtx_insn **bb_header = 0; | |||
268 | ||||
269 | /* Basic block after which recovery blocks will be created. */ | |||
270 | static basic_block before_recovery; | |||
271 | ||||
272 | /* Basic block just before the EXIT_BLOCK and after recovery, if we have | |||
273 | created it. */ | |||
274 | basic_block after_recovery; | |||
275 | ||||
276 | /* FALSE if we add bb to another region, so we don't need to initialize it. */ | |||
277 | bool adding_bb_to_current_region_p = true; | |||
278 | ||||
279 | /* Queues, etc. */ | |||
280 | ||||
281 | /* An instruction is ready to be scheduled when all insns preceding it | |||
282 | have already been scheduled. It is important to ensure that all | |||
283 | insns which use its result will not be executed until its result | |||
284 | has been computed. An insn is maintained in one of four structures: | |||
285 | ||||
286 | (P) the "Pending" set of insns which cannot be scheduled until | |||
287 | their dependencies have been satisfied. | |||
288 | (Q) the "Queued" set of insns that can be scheduled when sufficient | |||
289 | time has passed. | |||
290 | (R) the "Ready" list of unscheduled, uncommitted insns. | |||
291 | (S) the "Scheduled" list of insns. | |||
292 | ||||
293 | Initially, all insns are either "Pending" or "Ready" depending on | |||
294 | whether their dependencies are satisfied. | |||
295 | ||||
296 | Insns move from the "Ready" list to the "Scheduled" list as they | |||
297 | are committed to the schedule. As this occurs, the insns in the | |||
298 | "Pending" list have their dependencies satisfied and move to either | |||
299 | the "Ready" list or the "Queued" set depending on whether | |||
300 | sufficient time has passed to make them ready. As time passes, | |||
301 | insns move from the "Queued" set to the "Ready" list. | |||
302 | ||||
303 | The "Pending" list (P) are the insns in the INSN_FORW_DEPS of the | |||
304 | unscheduled insns, i.e., those that are ready, queued, and pending. | |||
305 | The "Queued" set (Q) is implemented by the variable `insn_queue'. | |||
306 | The "Ready" list (R) is implemented by the variables `ready' and | |||
307 | `n_ready'. | |||
308 | The "Scheduled" list (S) is the new insn chain built by this pass. | |||
309 | ||||
310 | The transition (R->S) is implemented in the scheduling loop in | |||
311 | `schedule_block' when the best insn to schedule is chosen. | |||
312 | The transitions (P->R and P->Q) are implemented in `schedule_insn' as | |||
313 | insns move from the ready list to the scheduled list. | |||
314 | The transition (Q->R) is implemented in 'queue_to_insn' as time | |||
315 | passes or stalls are introduced. */ | |||
316 | ||||
317 | /* Implement a circular buffer to delay instructions until sufficient | |||
318 | time has passed. For the new pipeline description interface, | |||
319 | MAX_INSN_QUEUE_INDEX is a power of two minus one which is not less | |||
320 | than maximal time of instruction execution computed by genattr.cc on | |||
321 | the base maximal time of functional unit reservations and getting a | |||
322 | result. This is the longest time an insn may be queued. */ | |||
323 | ||||
324 | static rtx_insn_list **insn_queue; | |||
325 | static int q_ptr = 0; | |||
326 | static int q_size = 0; | |||
327 | #define NEXT_Q(X)(((X)+1) & max_insn_queue_index) (((X)+1) & max_insn_queue_index) | |||
328 | #define NEXT_Q_AFTER(X, C)(((X)+C) & max_insn_queue_index) (((X)+C) & max_insn_queue_index) | |||
329 | ||||
330 | #define QUEUE_SCHEDULED(-3) (-3) | |||
331 | #define QUEUE_NOWHERE(-2) (-2) | |||
332 | #define QUEUE_READY(-1) (-1) | |||
333 | /* QUEUE_SCHEDULED - INSN is scheduled. | |||
334 | QUEUE_NOWHERE - INSN isn't scheduled yet and is neither in | |||
335 | queue or ready list. | |||
336 | QUEUE_READY - INSN is in ready list. | |||
337 | N >= 0 - INSN queued for X [where NEXT_Q_AFTER (q_ptr, X) == N] cycles. */ | |||
338 | ||||
339 | #define QUEUE_INDEX(INSN)((&h_i_d[INSN_UID (INSN)])->queue_index) (HID (INSN)(&h_i_d[INSN_UID (INSN)])->queue_index) | |||
340 | ||||
341 | /* The following variable value refers for all current and future | |||
342 | reservations of the processor units. */ | |||
343 | state_t curr_state; | |||
344 | ||||
345 | /* The following variable value is size of memory representing all | |||
346 | current and future reservations of the processor units. */ | |||
347 | size_t dfa_state_size; | |||
348 | ||||
349 | /* The following array is used to find the best insn from ready when | |||
350 | the automaton pipeline interface is used. */ | |||
351 | signed char *ready_try = NULLnullptr; | |||
352 | ||||
353 | /* The ready list. */ | |||
354 | struct ready_list ready = {NULLnullptr, 0, 0, 0, 0}; | |||
355 | ||||
356 | /* The pointer to the ready list (to be removed). */ | |||
357 | static struct ready_list *readyp = &ready; | |||
358 | ||||
359 | /* Scheduling clock. */ | |||
360 | static int clock_var; | |||
361 | ||||
362 | /* Clock at which the previous instruction was issued. */ | |||
363 | static int last_clock_var; | |||
364 | ||||
365 | /* Set to true if, when queuing a shadow insn, we discover that it would be | |||
366 | scheduled too late. */ | |||
367 | static bool must_backtrack; | |||
368 | ||||
369 | /* The following variable value is number of essential insns issued on | |||
370 | the current cycle. An insn is essential one if it changes the | |||
371 | processors state. */ | |||
372 | int cycle_issued_insns; | |||
373 | ||||
374 | /* This records the actual schedule. It is built up during the main phase | |||
375 | of schedule_block, and afterwards used to reorder the insns in the RTL. */ | |||
376 | static vec<rtx_insn *> scheduled_insns; | |||
377 | ||||
378 | static int may_trap_exp (const_rtx, int); | |||
379 | ||||
380 | /* Nonzero iff the address is comprised from at most 1 register. */ | |||
381 | #define CONST_BASED_ADDRESS_P(x)((((enum rtx_code) (x)->code) == REG) || ((((enum rtx_code ) (x)->code) == PLUS || ((enum rtx_code) (x)->code) == MINUS || (((enum rtx_code) (x)->code) == LO_SUM)) && (( (rtx_class[(int) (((enum rtx_code) ((((x)->u.fld[0]).rt_rtx ))->code))]) == RTX_CONST_OBJ) || ((rtx_class[(int) (((enum rtx_code) ((((x)->u.fld[1]).rt_rtx))->code))]) == RTX_CONST_OBJ )))) \ | |||
382 | (REG_P (x)(((enum rtx_code) (x)->code) == REG) \ | |||
383 | || ((GET_CODE (x)((enum rtx_code) (x)->code) == PLUS || GET_CODE (x)((enum rtx_code) (x)->code) == MINUS \ | |||
384 | || (GET_CODE (x)((enum rtx_code) (x)->code) == LO_SUM)) \ | |||
385 | && (CONSTANT_P (XEXP (x, 0))((rtx_class[(int) (((enum rtx_code) ((((x)->u.fld[0]).rt_rtx ))->code))]) == RTX_CONST_OBJ) \ | |||
386 | || CONSTANT_P (XEXP (x, 1))((rtx_class[(int) (((enum rtx_code) ((((x)->u.fld[1]).rt_rtx ))->code))]) == RTX_CONST_OBJ)))) | |||
387 | ||||
388 | /* Returns a class that insn with GET_DEST(insn)=x may belong to, | |||
389 | as found by analyzing insn's expression. */ | |||
390 | ||||
391 | ||||
392 | static int haifa_luid_for_non_insn (rtx x); | |||
393 | ||||
394 | /* Haifa version of sched_info hooks common to all headers. */ | |||
395 | const struct common_sched_info_def haifa_common_sched_info = | |||
396 | { | |||
397 | NULLnullptr, /* fix_recovery_cfg */ | |||
398 | NULLnullptr, /* add_block */ | |||
399 | NULLnullptr, /* estimate_number_of_insns */ | |||
400 | haifa_luid_for_non_insn, /* luid_for_non_insn */ | |||
401 | SCHED_PASS_UNKNOWN /* sched_pass_id */ | |||
402 | }; | |||
403 | ||||
404 | /* Mapping from instruction UID to its Logical UID. */ | |||
405 | vec<int> sched_luids; | |||
406 | ||||
407 | /* Next LUID to assign to an instruction. */ | |||
408 | int sched_max_luid = 1; | |||
409 | ||||
410 | /* Haifa Instruction Data. */ | |||
411 | vec<haifa_insn_data_def> h_i_d; | |||
412 | ||||
413 | void (* sched_init_only_bb) (basic_block, basic_block); | |||
414 | ||||
415 | /* Split block function. Different schedulers might use different functions | |||
416 | to handle their internal data consistent. */ | |||
417 | basic_block (* sched_split_block) (basic_block, rtx); | |||
418 | ||||
419 | /* Create empty basic block after the specified block. */ | |||
420 | basic_block (* sched_create_empty_bb) (basic_block); | |||
421 | ||||
422 | /* Return the number of cycles until INSN is expected to be ready. | |||
423 | Return zero if it already is. */ | |||
424 | static int | |||
425 | insn_delay (rtx_insn *insn) | |||
426 | { | |||
427 | return MAX (INSN_TICK (insn) - clock_var, 0)((((&h_i_d[INSN_UID (insn)])->tick) - clock_var) > ( 0) ? (((&h_i_d[INSN_UID (insn)])->tick) - clock_var) : (0)); | |||
428 | } | |||
429 | ||||
430 | static int | |||
431 | may_trap_exp (const_rtx x, int is_store) | |||
432 | { | |||
433 | enum rtx_code code; | |||
434 | ||||
435 | if (x == 0) | |||
436 | return TRAP_FREE; | |||
437 | code = GET_CODE (x)((enum rtx_code) (x)->code); | |||
438 | if (is_store) | |||
439 | { | |||
440 | if (code == MEM && may_trap_p (x)) | |||
441 | return TRAP_RISKY; | |||
442 | else | |||
443 | return TRAP_FREE; | |||
444 | } | |||
445 | if (code == MEM) | |||
446 | { | |||
447 | /* The insn uses memory: a volatile load. */ | |||
448 | if (MEM_VOLATILE_P (x)(__extension__ ({ __typeof ((x)) const _rtx = ((x)); if (((enum rtx_code) (_rtx)->code) != MEM && ((enum rtx_code ) (_rtx)->code) != ASM_OPERANDS && ((enum rtx_code ) (_rtx)->code) != ASM_INPUT) rtl_check_failed_flag ("MEM_VOLATILE_P" , _rtx, "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 448, __FUNCTION__); _rtx; })->volatil)) | |||
449 | return IRISKY; | |||
450 | /* An exception-free load. */ | |||
451 | if (!may_trap_p (x)) | |||
452 | return IFREE; | |||
453 | /* A load with 1 base register, to be further checked. */ | |||
454 | if (CONST_BASED_ADDRESS_P (XEXP (x, 0))((((enum rtx_code) ((((x)->u.fld[0]).rt_rtx))->code) == REG) || ((((enum rtx_code) ((((x)->u.fld[0]).rt_rtx))-> code) == PLUS || ((enum rtx_code) ((((x)->u.fld[0]).rt_rtx ))->code) == MINUS || (((enum rtx_code) ((((x)->u.fld[0 ]).rt_rtx))->code) == LO_SUM)) && (((rtx_class[(int ) (((enum rtx_code) (((((((x)->u.fld[0]).rt_rtx))->u.fld [0]).rt_rtx))->code))]) == RTX_CONST_OBJ) || ((rtx_class[( int) (((enum rtx_code) (((((((x)->u.fld[0]).rt_rtx))->u .fld[1]).rt_rtx))->code))]) == RTX_CONST_OBJ))))) | |||
455 | return PFREE_CANDIDATE; | |||
456 | /* No info on the load, to be further checked. */ | |||
457 | return PRISKY_CANDIDATE; | |||
458 | } | |||
459 | else | |||
460 | { | |||
461 | const char *fmt; | |||
462 | int i, insn_class = TRAP_FREE; | |||
463 | ||||
464 | /* Neither store nor load, check if it may cause a trap. */ | |||
465 | if (may_trap_p (x)) | |||
466 | return TRAP_RISKY; | |||
467 | /* Recursive step: walk the insn... */ | |||
468 | fmt = GET_RTX_FORMAT (code)(rtx_format[(int) (code)]); | |||
469 | for (i = GET_RTX_LENGTH (code)(rtx_length[(int) (code)]) - 1; i >= 0; i--) | |||
470 | { | |||
471 | if (fmt[i] == 'e') | |||
472 | { | |||
473 | int tmp_class = may_trap_exp (XEXP (x, i)(((x)->u.fld[i]).rt_rtx), is_store); | |||
474 | insn_class = WORST_CLASS (insn_class, tmp_class)((insn_class > tmp_class) ? insn_class : tmp_class); | |||
475 | } | |||
476 | else if (fmt[i] == 'E') | |||
477 | { | |||
478 | int j; | |||
479 | for (j = 0; j < XVECLEN (x, i)(((((x)->u.fld[i]).rt_rtvec))->num_elem); j++) | |||
480 | { | |||
481 | int tmp_class = may_trap_exp (XVECEXP (x, i, j)(((((x)->u.fld[i]).rt_rtvec))->elem[j]), is_store); | |||
482 | insn_class = WORST_CLASS (insn_class, tmp_class)((insn_class > tmp_class) ? insn_class : tmp_class); | |||
483 | if (insn_class == TRAP_RISKY || insn_class == IRISKY) | |||
484 | break; | |||
485 | } | |||
486 | } | |||
487 | if (insn_class == TRAP_RISKY || insn_class == IRISKY) | |||
488 | break; | |||
489 | } | |||
490 | return insn_class; | |||
491 | } | |||
492 | } | |||
493 | ||||
494 | /* Classifies rtx X of an insn for the purpose of verifying that X can be | |||
495 | executed speculatively (and consequently the insn can be moved | |||
496 | speculatively), by examining X, returning: | |||
497 | TRAP_RISKY: store, or risky non-load insn (e.g. division by variable). | |||
498 | TRAP_FREE: non-load insn. | |||
499 | IFREE: load from a globally safe location. | |||
500 | IRISKY: volatile load. | |||
501 | PFREE_CANDIDATE, PRISKY_CANDIDATE: load that need to be checked for | |||
502 | being either PFREE or PRISKY. */ | |||
503 | ||||
504 | static int | |||
505 | haifa_classify_rtx (const_rtx x) | |||
506 | { | |||
507 | int tmp_class = TRAP_FREE; | |||
508 | int insn_class = TRAP_FREE; | |||
509 | enum rtx_code code; | |||
510 | ||||
511 | if (GET_CODE (x)((enum rtx_code) (x)->code) == PARALLEL) | |||
512 | { | |||
513 | int i, len = XVECLEN (x, 0)(((((x)->u.fld[0]).rt_rtvec))->num_elem); | |||
514 | ||||
515 | for (i = len - 1; i >= 0; i--) | |||
516 | { | |||
517 | tmp_class = haifa_classify_rtx (XVECEXP (x, 0, i)(((((x)->u.fld[0]).rt_rtvec))->elem[i])); | |||
518 | insn_class = WORST_CLASS (insn_class, tmp_class)((insn_class > tmp_class) ? insn_class : tmp_class); | |||
519 | if (insn_class == TRAP_RISKY || insn_class == IRISKY) | |||
520 | break; | |||
521 | } | |||
522 | } | |||
523 | else | |||
524 | { | |||
525 | code = GET_CODE (x)((enum rtx_code) (x)->code); | |||
526 | switch (code) | |||
527 | { | |||
528 | case CLOBBER: | |||
529 | /* Test if it is a 'store'. */ | |||
530 | tmp_class = may_trap_exp (XEXP (x, 0)(((x)->u.fld[0]).rt_rtx), 1); | |||
531 | break; | |||
532 | case SET: | |||
533 | /* Test if it is a store. */ | |||
534 | tmp_class = may_trap_exp (SET_DEST (x)(((x)->u.fld[0]).rt_rtx), 1); | |||
535 | if (tmp_class == TRAP_RISKY) | |||
536 | break; | |||
537 | /* Test if it is a load. */ | |||
538 | tmp_class = | |||
539 | WORST_CLASS (tmp_class,((tmp_class > may_trap_exp ((((x)->u.fld[1]).rt_rtx), 0 )) ? tmp_class : may_trap_exp ((((x)->u.fld[1]).rt_rtx), 0 )) | |||
540 | may_trap_exp (SET_SRC (x), 0))((tmp_class > may_trap_exp ((((x)->u.fld[1]).rt_rtx), 0 )) ? tmp_class : may_trap_exp ((((x)->u.fld[1]).rt_rtx), 0 )); | |||
541 | break; | |||
542 | case COND_EXEC: | |||
543 | tmp_class = haifa_classify_rtx (COND_EXEC_CODE (x)(((x)->u.fld[1]).rt_rtx)); | |||
544 | if (tmp_class == TRAP_RISKY) | |||
545 | break; | |||
546 | tmp_class = WORST_CLASS (tmp_class,((tmp_class > may_trap_exp ((((x)->u.fld[0]).rt_rtx), 0 )) ? tmp_class : may_trap_exp ((((x)->u.fld[0]).rt_rtx), 0 )) | |||
547 | may_trap_exp (COND_EXEC_TEST (x), 0))((tmp_class > may_trap_exp ((((x)->u.fld[0]).rt_rtx), 0 )) ? tmp_class : may_trap_exp ((((x)->u.fld[0]).rt_rtx), 0 )); | |||
548 | break; | |||
549 | case TRAP_IF: | |||
550 | tmp_class = TRAP_RISKY; | |||
551 | break; | |||
552 | default:; | |||
553 | } | |||
554 | insn_class = tmp_class; | |||
555 | } | |||
556 | ||||
557 | return insn_class; | |||
558 | } | |||
559 | ||||
560 | int | |||
561 | haifa_classify_insn (const_rtx insn) | |||
562 | { | |||
563 | return haifa_classify_rtx (PATTERN (insn)); | |||
564 | } | |||
565 | ||||
566 | /* After the scheduler initialization function has been called, this function | |||
567 | can be called to enable modulo scheduling. II is the initiation interval | |||
568 | we should use, it affects the delays for delay_pairs that were recorded as | |||
569 | separated by a given number of stages. | |||
570 | ||||
571 | MAX_STAGES provides us with a limit | |||
572 | after which we give up scheduling; the caller must have unrolled at least | |||
573 | as many copies of the loop body and recorded delay_pairs for them. | |||
574 | ||||
575 | INSNS is the number of real (non-debug) insns in one iteration of | |||
576 | the loop. MAX_UID can be used to test whether an insn belongs to | |||
577 | the first iteration of the loop; all of them have a uid lower than | |||
578 | MAX_UID. */ | |||
579 | void | |||
580 | set_modulo_params (int ii, int max_stages, int insns, int max_uid) | |||
581 | { | |||
582 | modulo_ii = ii; | |||
583 | modulo_max_stages = max_stages; | |||
584 | modulo_n_insns = insns; | |||
585 | modulo_iter0_max_uid = max_uid; | |||
586 | modulo_backtracks_left = param_max_modulo_backtrack_attemptsglobal_options.x_param_max_modulo_backtrack_attempts; | |||
587 | } | |||
588 | ||||
589 | /* A structure to record a pair of insns where the first one is a real | |||
590 | insn that has delay slots, and the second is its delayed shadow. | |||
591 | I1 is scheduled normally and will emit an assembly instruction, | |||
592 | while I2 describes the side effect that takes place at the | |||
593 | transition between cycles CYCLES and (CYCLES + 1) after I1. */ | |||
594 | struct delay_pair | |||
595 | { | |||
596 | struct delay_pair *next_same_i1; | |||
597 | rtx_insn *i1, *i2; | |||
598 | int cycles; | |||
599 | /* When doing modulo scheduling, we a delay_pair can also be used to | |||
600 | show that I1 and I2 are the same insn in a different stage. If that | |||
601 | is the case, STAGES will be nonzero. */ | |||
602 | int stages; | |||
603 | }; | |||
604 | ||||
605 | /* Helpers for delay hashing. */ | |||
606 | ||||
607 | struct delay_i1_hasher : nofree_ptr_hash <delay_pair> | |||
608 | { | |||
609 | typedef void *compare_type; | |||
610 | static inline hashval_t hash (const delay_pair *); | |||
611 | static inline bool equal (const delay_pair *, const void *); | |||
612 | }; | |||
613 | ||||
614 | /* Returns a hash value for X, based on hashing just I1. */ | |||
615 | ||||
616 | inline hashval_t | |||
617 | delay_i1_hasher::hash (const delay_pair *x) | |||
618 | { | |||
619 | return htab_hash_pointer (x->i1); | |||
620 | } | |||
621 | ||||
622 | /* Return true if I1 of pair X is the same as that of pair Y. */ | |||
623 | ||||
624 | inline bool | |||
625 | delay_i1_hasher::equal (const delay_pair *x, const void *y) | |||
626 | { | |||
627 | return x->i1 == y; | |||
628 | } | |||
629 | ||||
630 | struct delay_i2_hasher : free_ptr_hash <delay_pair> | |||
631 | { | |||
632 | typedef void *compare_type; | |||
633 | static inline hashval_t hash (const delay_pair *); | |||
634 | static inline bool equal (const delay_pair *, const void *); | |||
635 | }; | |||
636 | ||||
637 | /* Returns a hash value for X, based on hashing just I2. */ | |||
638 | ||||
639 | inline hashval_t | |||
640 | delay_i2_hasher::hash (const delay_pair *x) | |||
641 | { | |||
642 | return htab_hash_pointer (x->i2); | |||
643 | } | |||
644 | ||||
645 | /* Return true if I2 of pair X is the same as that of pair Y. */ | |||
646 | ||||
647 | inline bool | |||
648 | delay_i2_hasher::equal (const delay_pair *x, const void *y) | |||
649 | { | |||
650 | return x->i2 == y; | |||
651 | } | |||
652 | ||||
653 | /* Two hash tables to record delay_pairs, one indexed by I1 and the other | |||
654 | indexed by I2. */ | |||
655 | static hash_table<delay_i1_hasher> *delay_htab; | |||
656 | static hash_table<delay_i2_hasher> *delay_htab_i2; | |||
657 | ||||
658 | /* Called through htab_traverse. Walk the hashtable using I2 as | |||
659 | index, and delete all elements involving an UID higher than | |||
660 | that pointed to by *DATA. */ | |||
661 | int | |||
662 | haifa_htab_i2_traverse (delay_pair **slot, int *data) | |||
663 | { | |||
664 | int maxuid = *data; | |||
665 | struct delay_pair *p = *slot; | |||
666 | if (INSN_UID (p->i2) >= maxuid || INSN_UID (p->i1) >= maxuid) | |||
667 | { | |||
668 | delay_htab_i2->clear_slot (slot); | |||
669 | } | |||
670 | return 1; | |||
671 | } | |||
672 | ||||
673 | /* Called through htab_traverse. Walk the hashtable using I2 as | |||
674 | index, and delete all elements involving an UID higher than | |||
675 | that pointed to by *DATA. */ | |||
676 | int | |||
677 | haifa_htab_i1_traverse (delay_pair **pslot, int *data) | |||
678 | { | |||
679 | int maxuid = *data; | |||
680 | struct delay_pair *p, *first, **pprev; | |||
681 | ||||
682 | if (INSN_UID ((*pslot)->i1) >= maxuid) | |||
683 | { | |||
684 | delay_htab->clear_slot (pslot); | |||
685 | return 1; | |||
686 | } | |||
687 | pprev = &first; | |||
688 | for (p = *pslot; p; p = p->next_same_i1) | |||
689 | { | |||
690 | if (INSN_UID (p->i2) < maxuid) | |||
691 | { | |||
692 | *pprev = p; | |||
693 | pprev = &p->next_same_i1; | |||
694 | } | |||
695 | } | |||
696 | *pprev = NULLnullptr; | |||
697 | if (first == NULLnullptr) | |||
698 | delay_htab->clear_slot (pslot); | |||
699 | else | |||
700 | *pslot = first; | |||
701 | return 1; | |||
702 | } | |||
703 | ||||
704 | /* Discard all delay pairs which involve an insn with an UID higher | |||
705 | than MAX_UID. */ | |||
706 | void | |||
707 | discard_delay_pairs_above (int max_uid) | |||
708 | { | |||
709 | delay_htab->traverse <int *, haifa_htab_i1_traverse> (&max_uid); | |||
710 | delay_htab_i2->traverse <int *, haifa_htab_i2_traverse> (&max_uid); | |||
711 | } | |||
712 | ||||
713 | /* This function can be called by a port just before it starts the final | |||
714 | scheduling pass. It records the fact that an instruction with delay | |||
715 | slots has been split into two insns, I1 and I2. The first one will be | |||
716 | scheduled normally and initiates the operation. The second one is a | |||
717 | shadow which must follow a specific number of cycles after I1; its only | |||
718 | purpose is to show the side effect that occurs at that cycle in the RTL. | |||
719 | If a JUMP_INSN or a CALL_INSN has been split, I1 should be a normal INSN, | |||
720 | while I2 retains the original insn type. | |||
721 | ||||
722 | There are two ways in which the number of cycles can be specified, | |||
723 | involving the CYCLES and STAGES arguments to this function. If STAGES | |||
724 | is zero, we just use the value of CYCLES. Otherwise, STAGES is a factor | |||
725 | which is multiplied by MODULO_II to give the number of cycles. This is | |||
726 | only useful if the caller also calls set_modulo_params to enable modulo | |||
727 | scheduling. */ | |||
728 | ||||
729 | void | |||
730 | record_delay_slot_pair (rtx_insn *i1, rtx_insn *i2, int cycles, int stages) | |||
731 | { | |||
732 | struct delay_pair *p = XNEW (struct delay_pair)((struct delay_pair *) xmalloc (sizeof (struct delay_pair))); | |||
733 | struct delay_pair **slot; | |||
734 | ||||
735 | p->i1 = i1; | |||
736 | p->i2 = i2; | |||
737 | p->cycles = cycles; | |||
738 | p->stages = stages; | |||
739 | ||||
740 | if (!delay_htab) | |||
741 | { | |||
742 | delay_htab = new hash_table<delay_i1_hasher> (10); | |||
743 | delay_htab_i2 = new hash_table<delay_i2_hasher> (10); | |||
744 | } | |||
745 | slot = delay_htab->find_slot_with_hash (i1, htab_hash_pointer (i1), INSERT); | |||
746 | p->next_same_i1 = *slot; | |||
747 | *slot = p; | |||
748 | slot = delay_htab_i2->find_slot (p, INSERT); | |||
749 | *slot = p; | |||
750 | } | |||
751 | ||||
752 | /* Examine the delay pair hashtable to see if INSN is a shadow for another, | |||
753 | and return the other insn if so. Return NULL otherwise. */ | |||
754 | rtx_insn * | |||
755 | real_insn_for_shadow (rtx_insn *insn) | |||
756 | { | |||
757 | struct delay_pair *pair; | |||
758 | ||||
759 | if (!delay_htab) | |||
760 | return NULLnullptr; | |||
761 | ||||
762 | pair = delay_htab_i2->find_with_hash (insn, htab_hash_pointer (insn)); | |||
763 | if (!pair || pair->stages > 0) | |||
764 | return NULLnullptr; | |||
765 | return pair->i1; | |||
766 | } | |||
767 | ||||
768 | /* For a pair P of insns, return the fixed distance in cycles from the first | |||
769 | insn after which the second must be scheduled. */ | |||
770 | static int | |||
771 | pair_delay (struct delay_pair *p) | |||
772 | { | |||
773 | if (p->stages == 0) | |||
774 | return p->cycles; | |||
775 | else | |||
776 | return p->stages * modulo_ii; | |||
777 | } | |||
778 | ||||
779 | /* Given an insn INSN, add a dependence on its delayed shadow if it | |||
780 | has one. Also try to find situations where shadows depend on each other | |||
781 | and add dependencies to the real insns to limit the amount of backtracking | |||
782 | needed. */ | |||
783 | void | |||
784 | add_delay_dependencies (rtx_insn *insn) | |||
785 | { | |||
786 | struct delay_pair *pair; | |||
787 | sd_iterator_def sd_it; | |||
788 | dep_t dep; | |||
789 | ||||
790 | if (!delay_htab) | |||
791 | return; | |||
792 | ||||
793 | pair = delay_htab_i2->find_with_hash (insn, htab_hash_pointer (insn)); | |||
794 | if (!pair) | |||
795 | return; | |||
796 | add_dependence (insn, pair->i1, REG_DEP_ANTI); | |||
797 | if (pair->stages) | |||
798 | return; | |||
799 | ||||
800 | FOR_EACH_DEP (pair->i2, SD_LIST_BACK, sd_it, dep)for ((sd_it) = sd_iterator_start ((pair->i2), (((1) | (2)) )); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
801 | { | |||
802 | rtx_insn *pro = DEP_PRO (dep)((dep)->pro); | |||
803 | struct delay_pair *other_pair | |||
804 | = delay_htab_i2->find_with_hash (pro, htab_hash_pointer (pro)); | |||
805 | if (!other_pair || other_pair->stages) | |||
806 | continue; | |||
807 | if (pair_delay (other_pair) >= pair_delay (pair)) | |||
808 | { | |||
809 | if (sched_verbose >= 4) | |||
810 | { | |||
811 | fprintf (sched_dump, ";;\tadding dependence %d <- %d\n", | |||
812 | INSN_UID (other_pair->i1), | |||
813 | INSN_UID (pair->i1)); | |||
814 | fprintf (sched_dump, ";;\tpair1 %d <- %d, cost %d\n", | |||
815 | INSN_UID (pair->i1), | |||
816 | INSN_UID (pair->i2), | |||
817 | pair_delay (pair)); | |||
818 | fprintf (sched_dump, ";;\tpair2 %d <- %d, cost %d\n", | |||
819 | INSN_UID (other_pair->i1), | |||
820 | INSN_UID (other_pair->i2), | |||
821 | pair_delay (other_pair)); | |||
822 | } | |||
823 | add_dependence (pair->i1, other_pair->i1, REG_DEP_ANTI); | |||
824 | } | |||
825 | } | |||
826 | } | |||
827 | ||||
828 | /* Forward declarations. */ | |||
829 | ||||
830 | static int priority (rtx_insn *, bool force_recompute = false); | |||
831 | static int autopref_rank_for_schedule (const rtx_insn *, const rtx_insn *); | |||
832 | static int rank_for_schedule (const void *, const void *); | |||
833 | static void swap_sort (rtx_insn **, int); | |||
834 | static void queue_insn (rtx_insn *, int, const char *); | |||
835 | static int schedule_insn (rtx_insn *); | |||
836 | static void adjust_priority (rtx_insn *); | |||
837 | static void advance_one_cycle (void); | |||
838 | static void extend_h_i_d (void); | |||
839 | ||||
840 | ||||
841 | /* Notes handling mechanism: | |||
842 | ========================= | |||
843 | Generally, NOTES are saved before scheduling and restored after scheduling. | |||
844 | The scheduler distinguishes between two types of notes: | |||
845 | ||||
846 | (1) LOOP_BEGIN, LOOP_END, SETJMP, EHREGION_BEG, EHREGION_END notes: | |||
847 | Before scheduling a region, a pointer to the note is added to the insn | |||
848 | that follows or precedes it. (This happens as part of the data dependence | |||
849 | computation). After scheduling an insn, the pointer contained in it is | |||
850 | used for regenerating the corresponding note (in reemit_notes). | |||
851 | ||||
852 | (2) All other notes (e.g. INSN_DELETED): Before scheduling a block, | |||
853 | these notes are put in a list (in rm_other_notes() and | |||
854 | unlink_other_notes ()). After scheduling the block, these notes are | |||
855 | inserted at the beginning of the block (in schedule_block()). */ | |||
856 | ||||
857 | static void ready_add (struct ready_list *, rtx_insn *, bool); | |||
858 | static rtx_insn *ready_remove_first (struct ready_list *); | |||
859 | static rtx_insn *ready_remove_first_dispatch (struct ready_list *ready); | |||
860 | ||||
861 | static void queue_to_ready (struct ready_list *); | |||
862 | static int early_queue_to_ready (state_t, struct ready_list *); | |||
863 | ||||
864 | /* The following functions are used to implement multi-pass scheduling | |||
865 | on the first cycle. */ | |||
866 | static rtx_insn *ready_remove (struct ready_list *, int); | |||
867 | static void ready_remove_insn (rtx_insn *); | |||
868 | ||||
869 | static void fix_inter_tick (rtx_insn *, rtx_insn *); | |||
870 | static int fix_tick_ready (rtx_insn *); | |||
871 | static void change_queue_index (rtx_insn *, int); | |||
872 | ||||
873 | /* The following functions are used to implement scheduling of data/control | |||
874 | speculative instructions. */ | |||
875 | ||||
876 | static void extend_h_i_d (void); | |||
877 | static void init_h_i_d (rtx_insn *); | |||
878 | static int haifa_speculate_insn (rtx_insn *, ds_t, rtx *); | |||
879 | static void generate_recovery_code (rtx_insn *); | |||
880 | static void process_insn_forw_deps_be_in_spec (rtx_insn *, rtx_insn *, ds_t); | |||
881 | static void begin_speculative_block (rtx_insn *); | |||
882 | static void add_to_speculative_block (rtx_insn *); | |||
883 | static void init_before_recovery (basic_block *); | |||
884 | static void create_check_block_twin (rtx_insn *, bool); | |||
885 | static void fix_recovery_deps (basic_block); | |||
886 | static bool haifa_change_pattern (rtx_insn *, rtx); | |||
887 | static void dump_new_block_header (int, basic_block, rtx_insn *, rtx_insn *); | |||
888 | static void restore_bb_notes (basic_block); | |||
889 | static void fix_jump_move (rtx_insn *); | |||
890 | static void move_block_after_check (rtx_insn *); | |||
891 | static void move_succs (vec<edge, va_gc> **, basic_block); | |||
892 | static void sched_remove_insn (rtx_insn *); | |||
893 | static void clear_priorities (rtx_insn *, rtx_vec_t *); | |||
894 | static void calc_priorities (const rtx_vec_t &); | |||
895 | static void add_jump_dependencies (rtx_insn *, rtx_insn *); | |||
896 | ||||
897 | #endif /* INSN_SCHEDULING */ | |||
898 | ||||
899 | /* Point to state used for the current scheduling pass. */ | |||
900 | struct haifa_sched_info *current_sched_info; | |||
901 | ||||
902 | #ifndef INSN_SCHEDULING | |||
903 | void | |||
904 | schedule_insns (void) | |||
905 | { | |||
906 | } | |||
907 | #else | |||
908 | ||||
909 | /* Do register pressure sensitive insn scheduling if the flag is set | |||
910 | up. */ | |||
911 | enum sched_pressure_algorithm sched_pressure; | |||
912 | ||||
913 | /* Map regno -> its pressure class. The map defined only when | |||
914 | SCHED_PRESSURE != SCHED_PRESSURE_NONE. */ | |||
915 | enum reg_class *sched_regno_pressure_class; | |||
916 | ||||
917 | /* The current register pressure. Only elements corresponding pressure | |||
918 | classes are defined. */ | |||
919 | static int curr_reg_pressure[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
920 | ||||
921 | /* Saved value of the previous array. */ | |||
922 | static int saved_reg_pressure[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
923 | ||||
924 | /* Register living at given scheduling point. */ | |||
925 | static bitmap curr_reg_live; | |||
926 | ||||
927 | /* Saved value of the previous array. */ | |||
928 | static bitmap saved_reg_live; | |||
929 | ||||
930 | /* Registers mentioned in the current region. */ | |||
931 | static bitmap region_ref_regs; | |||
932 | ||||
933 | /* Temporary bitmap used for SCHED_PRESSURE_MODEL. */ | |||
934 | static bitmap tmp_bitmap; | |||
935 | ||||
936 | /* Effective number of available registers of a given class (see comment | |||
937 | in sched_pressure_start_bb). */ | |||
938 | static int sched_class_regs_num[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
939 | /* The number of registers that the function would need to save before it | |||
940 | uses them, and the number of fixed_regs. Helpers for calculating of | |||
941 | sched_class_regs_num. */ | |||
942 | static int call_saved_regs_num[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
943 | static int fixed_regs_num[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
944 | ||||
945 | /* Initiate register pressure relative info for scheduling the current | |||
946 | region. Currently it is only clearing register mentioned in the | |||
947 | current region. */ | |||
948 | void | |||
949 | sched_init_region_reg_pressure_info (void) | |||
950 | { | |||
951 | bitmap_clear (region_ref_regs); | |||
952 | } | |||
953 | ||||
954 | /* PRESSURE[CL] describes the pressure on register class CL. Update it | |||
955 | for the birth (if BIRTH_P) or death (if !BIRTH_P) of register REGNO. | |||
956 | LIVE tracks the set of live registers; if it is null, assume that | |||
957 | every birth or death is genuine. */ | |||
958 | static inline void | |||
959 | mark_regno_birth_or_death (bitmap live, int *pressure, int regno, bool birth_p) | |||
960 | { | |||
961 | enum reg_class pressure_class; | |||
962 | ||||
963 | pressure_class = sched_regno_pressure_class[regno]; | |||
964 | if (regno >= FIRST_PSEUDO_REGISTER76) | |||
965 | { | |||
966 | if (pressure_class != NO_REGS) | |||
967 | { | |||
968 | if (birth_p) | |||
969 | { | |||
970 | if (!live || bitmap_set_bit (live, regno)) | |||
971 | pressure[pressure_class] | |||
972 | += (ira_reg_class_max_nregs(this_target_ira->x_ira_reg_class_max_nregs) | |||
973 | [pressure_class][PSEUDO_REGNO_MODE (regno)((machine_mode) (regno_reg_rtx[regno])->mode)]); | |||
974 | } | |||
975 | else | |||
976 | { | |||
977 | if (!live || bitmap_clear_bit (live, regno)) | |||
978 | pressure[pressure_class] | |||
979 | -= (ira_reg_class_max_nregs(this_target_ira->x_ira_reg_class_max_nregs) | |||
980 | [pressure_class][PSEUDO_REGNO_MODE (regno)((machine_mode) (regno_reg_rtx[regno])->mode)]); | |||
981 | } | |||
982 | } | |||
983 | } | |||
984 | else if (pressure_class != NO_REGS | |||
985 | && ! TEST_HARD_REG_BIT (ira_no_alloc_regs(this_target_ira->x_ira_no_alloc_regs), regno)) | |||
986 | { | |||
987 | if (birth_p) | |||
988 | { | |||
989 | if (!live || bitmap_set_bit (live, regno)) | |||
990 | pressure[pressure_class]++; | |||
991 | } | |||
992 | else | |||
993 | { | |||
994 | if (!live || bitmap_clear_bit (live, regno)) | |||
995 | pressure[pressure_class]--; | |||
996 | } | |||
997 | } | |||
998 | } | |||
999 | ||||
1000 | /* Initiate current register pressure related info from living | |||
1001 | registers given by LIVE. */ | |||
1002 | static void | |||
1003 | initiate_reg_pressure_info (bitmap live) | |||
1004 | { | |||
1005 | int i; | |||
1006 | unsigned int j; | |||
1007 | bitmap_iterator bi; | |||
1008 | ||||
1009 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
1010 | curr_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]] = 0; | |||
1011 | bitmap_clear (curr_reg_live); | |||
1012 | EXECUTE_IF_SET_IN_BITMAP (live, 0, j, bi)for (bmp_iter_set_init (&(bi), (live), (0), &(j)); bmp_iter_set (&(bi), &(j)); bmp_iter_next (&(bi), &(j))) | |||
1013 | if (sched_pressure == SCHED_PRESSURE_MODEL | |||
1014 | || current_nr_blocks == 1 | |||
1015 | || bitmap_bit_p (region_ref_regs, j)) | |||
1016 | mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure, j, true); | |||
1017 | } | |||
1018 | ||||
1019 | /* Mark registers in X as mentioned in the current region. */ | |||
1020 | static void | |||
1021 | setup_ref_regs (rtx x) | |||
1022 | { | |||
1023 | int i, j; | |||
1024 | const RTX_CODEenum rtx_code code = GET_CODE (x)((enum rtx_code) (x)->code); | |||
1025 | const char *fmt; | |||
1026 | ||||
1027 | if (REG_P (x)(((enum rtx_code) (x)->code) == REG)) | |||
1028 | { | |||
1029 | bitmap_set_range (region_ref_regs, REGNO (x)(rhs_regno(x)), REG_NREGS (x)((&(x)->u.reg)->nregs)); | |||
1030 | return; | |||
1031 | } | |||
1032 | fmt = GET_RTX_FORMAT (code)(rtx_format[(int) (code)]); | |||
1033 | for (i = GET_RTX_LENGTH (code)(rtx_length[(int) (code)]) - 1; i >= 0; i--) | |||
1034 | if (fmt[i] == 'e') | |||
1035 | setup_ref_regs (XEXP (x, i)(((x)->u.fld[i]).rt_rtx)); | |||
1036 | else if (fmt[i] == 'E') | |||
1037 | { | |||
1038 | for (j = 0; j < XVECLEN (x, i)(((((x)->u.fld[i]).rt_rtvec))->num_elem); j++) | |||
1039 | setup_ref_regs (XVECEXP (x, i, j)(((((x)->u.fld[i]).rt_rtvec))->elem[j])); | |||
1040 | } | |||
1041 | } | |||
1042 | ||||
1043 | /* Initiate current register pressure related info at the start of | |||
1044 | basic block BB. */ | |||
1045 | static void | |||
1046 | initiate_bb_reg_pressure_info (basic_block bb) | |||
1047 | { | |||
1048 | unsigned int i ATTRIBUTE_UNUSED__attribute__ ((__unused__)); | |||
1049 | rtx_insn *insn; | |||
1050 | ||||
1051 | if (current_nr_blocks > 1) | |||
1052 | FOR_BB_INSNS (bb, insn)for ((insn) = (bb)->il.x.head_; (insn) && (insn) != NEXT_INSN ((bb)->il.x.rtl->end_); (insn) = NEXT_INSN ( insn)) | |||
1053 | if (NONDEBUG_INSN_P (insn)((((enum rtx_code) (insn)->code) == INSN) || (((enum rtx_code ) (insn)->code) == JUMP_INSN) || (((enum rtx_code) (insn)-> code) == CALL_INSN))) | |||
1054 | setup_ref_regs (PATTERN (insn)); | |||
1055 | initiate_reg_pressure_info (df_get_live_in (bb)); | |||
1056 | if (bb_has_eh_pred (bb)) | |||
1057 | for (i = 0; ; ++i) | |||
1058 | { | |||
1059 | unsigned int regno = EH_RETURN_DATA_REGNO (i)((i) <= 1 ? (i) : (~(unsigned int) 0)); | |||
1060 | ||||
1061 | if (regno == INVALID_REGNUM(~(unsigned int) 0)) | |||
1062 | break; | |||
1063 | if (! bitmap_bit_p (df_get_live_in (bb), regno)) | |||
1064 | mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure, | |||
1065 | regno, true); | |||
1066 | } | |||
1067 | } | |||
1068 | ||||
1069 | /* Save current register pressure related info. */ | |||
1070 | static void | |||
1071 | save_reg_pressure (void) | |||
1072 | { | |||
1073 | int i; | |||
1074 | ||||
1075 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
1076 | saved_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]] | |||
1077 | = curr_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]]; | |||
1078 | bitmap_copy (saved_reg_live, curr_reg_live); | |||
1079 | } | |||
1080 | ||||
1081 | /* Restore saved register pressure related info. */ | |||
1082 | static void | |||
1083 | restore_reg_pressure (void) | |||
1084 | { | |||
1085 | int i; | |||
1086 | ||||
1087 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
1088 | curr_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]] | |||
1089 | = saved_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]]; | |||
1090 | bitmap_copy (curr_reg_live, saved_reg_live); | |||
1091 | } | |||
1092 | ||||
1093 | /* Return TRUE if the register is dying after its USE. */ | |||
1094 | static bool | |||
1095 | dying_use_p (struct reg_use_data *use) | |||
1096 | { | |||
1097 | struct reg_use_data *next; | |||
1098 | ||||
1099 | for (next = use->next_regno_use; next != use; next = next->next_regno_use) | |||
1100 | if (NONDEBUG_INSN_P (next->insn)((((enum rtx_code) (next->insn)->code) == INSN) || (((enum rtx_code) (next->insn)->code) == JUMP_INSN) || (((enum rtx_code) (next->insn)->code) == CALL_INSN)) | |||
1101 | && QUEUE_INDEX (next->insn)((&h_i_d[INSN_UID (next->insn)])->queue_index) != QUEUE_SCHEDULED(-3)) | |||
1102 | return false; | |||
1103 | return true; | |||
1104 | } | |||
1105 | ||||
1106 | /* Print info about the current register pressure and its excess for | |||
1107 | each pressure class. */ | |||
1108 | static void | |||
1109 | print_curr_reg_pressure (void) | |||
1110 | { | |||
1111 | int i; | |||
1112 | enum reg_class cl; | |||
1113 | ||||
1114 | fprintf (sched_dump, ";;\t"); | |||
1115 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
1116 | { | |||
1117 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]; | |||
1118 | gcc_assert (curr_reg_pressure[cl] >= 0)((void)(!(curr_reg_pressure[cl] >= 0) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1118, __FUNCTION__), 0 : 0)); | |||
1119 | fprintf (sched_dump, " %s:%d(%d)", reg_class_names[cl], | |||
1120 | curr_reg_pressure[cl], | |||
1121 | curr_reg_pressure[cl] - sched_class_regs_num[cl]); | |||
1122 | } | |||
1123 | fprintf (sched_dump, "\n"); | |||
1124 | } | |||
1125 | ||||
1126 | /* Determine if INSN has a condition that is clobbered if a register | |||
1127 | in SET_REGS is modified. */ | |||
1128 | static bool | |||
1129 | cond_clobbered_p (rtx_insn *insn, HARD_REG_SET set_regs) | |||
1130 | { | |||
1131 | rtx pat = PATTERN (insn); | |||
1132 | gcc_assert (GET_CODE (pat) == COND_EXEC)((void)(!(((enum rtx_code) (pat)->code) == COND_EXEC) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1132, __FUNCTION__), 0 : 0)); | |||
1133 | if (TEST_HARD_REG_BIT (set_regs, REGNO (XEXP (COND_EXEC_TEST (pat), 0))(rhs_regno(((((((pat)->u.fld[0]).rt_rtx))->u.fld[0]).rt_rtx ))))) | |||
1134 | { | |||
1135 | sd_iterator_def sd_it; | |||
1136 | dep_t dep; | |||
1137 | haifa_change_pattern (insn, ORIG_PAT (insn)((&h_i_d[INSN_UID (insn)])->orig_pat)); | |||
1138 | FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)for ((sd_it) = sd_iterator_start ((insn), (((1) | (2)))); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
1139 | DEP_STATUS (dep)((dep)->status) &= ~DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1); | |||
1140 | TODO_SPEC (insn)((&h_i_d[INSN_UID (insn)])->todo_spec) = HARD_DEP(((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) << 1); | |||
1141 | if (sched_verbose >= 2) | |||
1142 | fprintf (sched_dump, | |||
1143 | ";;\t\tdequeue insn %s because of clobbered condition\n", | |||
1144 | (*current_sched_info->print_insn) (insn, 0)); | |||
1145 | return true; | |||
1146 | } | |||
1147 | ||||
1148 | return false; | |||
1149 | } | |||
1150 | ||||
1151 | /* This function should be called after modifying the pattern of INSN, | |||
1152 | to update scheduler data structures as needed. */ | |||
1153 | static void | |||
1154 | update_insn_after_change (rtx_insn *insn) | |||
1155 | { | |||
1156 | sd_iterator_def sd_it; | |||
1157 | dep_t dep; | |||
1158 | ||||
1159 | dfa_clear_single_insn_cache (insn); | |||
1160 | ||||
1161 | sd_it = sd_iterator_start (insn, | |||
1162 | SD_LIST_FORW(4) | SD_LIST_BACK((1) | (2)) | SD_LIST_RES_BACK(8)); | |||
1163 | while (sd_iterator_cond (&sd_it, &dep)) | |||
1164 | { | |||
1165 | DEP_COST (dep)((dep)->cost) = UNKNOWN_DEP_COST((int) ((unsigned int) -1 << 19)); | |||
1166 | sd_iterator_next (&sd_it); | |||
1167 | } | |||
1168 | ||||
1169 | /* Invalidate INSN_COST, so it'll be recalculated. */ | |||
1170 | INSN_COST (insn)((&h_i_d[INSN_UID (insn)])->cost) = -1; | |||
1171 | /* Invalidate INSN_TICK, so it'll be recalculated. */ | |||
1172 | INSN_TICK (insn)((&h_i_d[INSN_UID (insn)])->tick) = INVALID_TICK(-(max_insn_queue_index + 1)); | |||
1173 | ||||
1174 | /* Invalidate autoprefetch data entry. */ | |||
1175 | INSN_AUTOPREF_MULTIPASS_DATA (insn)((&h_i_d[INSN_UID (insn)])->autopref_multipass_data)[0].status | |||
1176 | = AUTOPREF_MULTIPASS_DATA_UNINITIALIZED; | |||
1177 | INSN_AUTOPREF_MULTIPASS_DATA (insn)((&h_i_d[INSN_UID (insn)])->autopref_multipass_data)[1].status | |||
1178 | = AUTOPREF_MULTIPASS_DATA_UNINITIALIZED; | |||
1179 | } | |||
1180 | ||||
1181 | ||||
1182 | /* Two VECs, one to hold dependencies for which pattern replacements | |||
1183 | need to be applied or restored at the start of the next cycle, and | |||
1184 | another to hold an integer that is either one, to apply the | |||
1185 | corresponding replacement, or zero to restore it. */ | |||
1186 | static vec<dep_t> next_cycle_replace_deps; | |||
1187 | static vec<int> next_cycle_apply; | |||
1188 | ||||
1189 | static void apply_replacement (dep_t, bool); | |||
1190 | static void restore_pattern (dep_t, bool); | |||
1191 | ||||
1192 | /* Look at the remaining dependencies for insn NEXT, and compute and return | |||
1193 | the TODO_SPEC value we should use for it. This is called after one of | |||
1194 | NEXT's dependencies has been resolved. | |||
1195 | We also perform pattern replacements for predication, and for broken | |||
1196 | replacement dependencies. The latter is only done if FOR_BACKTRACK is | |||
1197 | false. */ | |||
1198 | ||||
1199 | static ds_t | |||
1200 | recompute_todo_spec (rtx_insn *next, bool for_backtrack) | |||
1201 | { | |||
1202 | ds_t new_ds; | |||
1203 | sd_iterator_def sd_it; | |||
1204 | dep_t dep, modify_dep = NULLnullptr; | |||
1205 | int n_spec = 0; | |||
1206 | int n_control = 0; | |||
1207 | int n_replace = 0; | |||
1208 | bool first_p = true; | |||
1209 | ||||
1210 | if (sd_lists_empty_p (next, SD_LIST_BACK((1) | (2)))) | |||
1211 | /* NEXT has all its dependencies resolved. */ | |||
1212 | return 0; | |||
1213 | ||||
1214 | if (!sd_lists_empty_p (next, SD_LIST_HARD_BACK(1))) | |||
1215 | return HARD_DEP(((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) << 1); | |||
1216 | ||||
1217 | /* If NEXT is intended to sit adjacent to this instruction, we don't | |||
1218 | want to try to break any dependencies. Treat it as a HARD_DEP. */ | |||
1219 | if (SCHED_GROUP_P (next)(__extension__ ({ __typeof ((next)) const _rtx = ((next)); if (((enum rtx_code) (_rtx)->code) != DEBUG_INSN && ( (enum rtx_code) (_rtx)->code) != INSN && ((enum rtx_code ) (_rtx)->code) != JUMP_INSN && ((enum rtx_code) ( _rtx)->code) != CALL_INSN) rtl_check_failed_flag ("SCHED_GROUP_P" , _rtx, "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1219, __FUNCTION__); _rtx; })->in_struct)) | |||
1220 | return HARD_DEP(((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) << 1); | |||
1221 | ||||
1222 | /* Now we've got NEXT with speculative deps only. | |||
1223 | 1. Look at the deps to see what we have to do. | |||
1224 | 2. Check if we can do 'todo'. */ | |||
1225 | new_ds = 0; | |||
1226 | ||||
1227 | FOR_EACH_DEP (next, SD_LIST_BACK, sd_it, dep)for ((sd_it) = sd_iterator_start ((next), (((1) | (2)))); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
1228 | { | |||
1229 | rtx_insn *pro = DEP_PRO (dep)((dep)->pro); | |||
1230 | ds_t ds = DEP_STATUS (dep)((dep)->status) & SPECULATIVE(((((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BEGIN_DATA_BITS_OFFSET ) | (((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BE_IN_DATA_BITS_OFFSET)) | ((((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BEGIN_CONTROL_BITS_OFFSET) | (((ds_t ) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BE_IN_CONTROL_BITS_OFFSET ))); | |||
1231 | ||||
1232 | if (DEBUG_INSN_P (pro)(((enum rtx_code) (pro)->code) == DEBUG_INSN) && !DEBUG_INSN_P (next)(((enum rtx_code) (next)->code) == DEBUG_INSN)) | |||
1233 | continue; | |||
1234 | ||||
1235 | if (ds) | |||
1236 | { | |||
1237 | n_spec++; | |||
1238 | if (first_p) | |||
1239 | { | |||
1240 | first_p = false; | |||
1241 | ||||
1242 | new_ds = ds; | |||
1243 | } | |||
1244 | else | |||
1245 | new_ds = ds_merge (new_ds, ds); | |||
1246 | } | |||
1247 | else if (DEP_TYPE (dep)((dep)->type) == REG_DEP_CONTROL) | |||
1248 | { | |||
1249 | if (QUEUE_INDEX (pro)((&h_i_d[INSN_UID (pro)])->queue_index) != QUEUE_SCHEDULED(-3)) | |||
1250 | { | |||
1251 | n_control++; | |||
1252 | modify_dep = dep; | |||
1253 | } | |||
1254 | DEP_STATUS (dep)((dep)->status) &= ~DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1); | |||
1255 | } | |||
1256 | else if (DEP_REPLACE (dep)((dep)->replace) != NULLnullptr) | |||
1257 | { | |||
1258 | if (QUEUE_INDEX (pro)((&h_i_d[INSN_UID (pro)])->queue_index) != QUEUE_SCHEDULED(-3)) | |||
1259 | { | |||
1260 | n_replace++; | |||
1261 | modify_dep = dep; | |||
1262 | } | |||
1263 | DEP_STATUS (dep)((dep)->status) &= ~DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1); | |||
1264 | } | |||
1265 | } | |||
1266 | ||||
1267 | if (n_replace > 0 && n_control == 0 && n_spec == 0) | |||
1268 | { | |||
1269 | if (!dbg_cnt (sched_breakdep)) | |||
1270 | return HARD_DEP(((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) << 1); | |||
1271 | FOR_EACH_DEP (next, SD_LIST_BACK, sd_it, dep)for ((sd_it) = sd_iterator_start ((next), (((1) | (2)))); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
1272 | { | |||
1273 | struct dep_replacement *desc = DEP_REPLACE (dep)((dep)->replace); | |||
1274 | if (desc != NULLnullptr) | |||
1275 | { | |||
1276 | if (desc->insn == next && !for_backtrack) | |||
1277 | { | |||
1278 | gcc_assert (n_replace == 1)((void)(!(n_replace == 1) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1278, __FUNCTION__), 0 : 0)); | |||
1279 | apply_replacement (dep, true); | |||
1280 | } | |||
1281 | DEP_STATUS (dep)((dep)->status) |= DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1); | |||
1282 | } | |||
1283 | } | |||
1284 | return 0; | |||
1285 | } | |||
1286 | ||||
1287 | else if (n_control == 1 && n_replace == 0 && n_spec == 0) | |||
1288 | { | |||
1289 | rtx_insn *pro, *other; | |||
1290 | rtx new_pat; | |||
1291 | rtx cond = NULL_RTX(rtx) 0; | |||
1292 | bool success; | |||
1293 | rtx_insn *prev = NULLnullptr; | |||
1294 | int i; | |||
1295 | unsigned regno; | |||
1296 | ||||
1297 | if ((current_sched_info->flags & DO_PREDICATION) == 0 | |||
1298 | || (ORIG_PAT (next)((&h_i_d[INSN_UID (next)])->orig_pat) != NULL_RTX(rtx) 0 | |||
1299 | && PREDICATED_PAT (next)((&h_i_d[INSN_UID (next)])->predicated_pat) == NULL_RTX(rtx) 0)) | |||
1300 | return HARD_DEP(((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) << 1); | |||
1301 | ||||
1302 | pro = DEP_PRO (modify_dep)((modify_dep)->pro); | |||
1303 | other = real_insn_for_shadow (pro); | |||
1304 | if (other != NULL_RTX(rtx) 0) | |||
1305 | pro = other; | |||
1306 | ||||
1307 | cond = sched_get_reverse_condition_uncached (pro); | |||
1308 | regno = REGNO (XEXP (cond, 0))(rhs_regno((((cond)->u.fld[0]).rt_rtx))); | |||
1309 | ||||
1310 | /* Find the last scheduled insn that modifies the condition register. | |||
1311 | We can stop looking once we find the insn we depend on through the | |||
1312 | REG_DEP_CONTROL; if the condition register isn't modified after it, | |||
1313 | we know that it still has the right value. */ | |||
1314 | if (QUEUE_INDEX (pro)((&h_i_d[INSN_UID (pro)])->queue_index) == QUEUE_SCHEDULED(-3)) | |||
1315 | FOR_EACH_VEC_ELT_REVERSE (scheduled_insns, i, prev)for (i = (scheduled_insns).length () - 1; (scheduled_insns).iterate ((i), &(prev)); (i)--) | |||
1316 | { | |||
1317 | HARD_REG_SET t; | |||
1318 | ||||
1319 | find_all_hard_reg_sets (prev, &t, true); | |||
1320 | if (TEST_HARD_REG_BIT (t, regno)) | |||
1321 | return HARD_DEP(((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) << 1); | |||
1322 | if (prev == pro) | |||
1323 | break; | |||
1324 | } | |||
1325 | if (ORIG_PAT (next)((&h_i_d[INSN_UID (next)])->orig_pat) == NULL_RTX(rtx) 0) | |||
1326 | { | |||
1327 | ORIG_PAT (next)((&h_i_d[INSN_UID (next)])->orig_pat) = PATTERN (next); | |||
1328 | ||||
1329 | new_pat = gen_rtx_COND_EXEC (VOIDmode, cond, PATTERN (next))gen_rtx_fmt_ee_stat ((COND_EXEC), ((((void) 0, E_VOIDmode))), ((cond)), ((PATTERN (next))) ); | |||
1330 | success = haifa_change_pattern (next, new_pat); | |||
1331 | if (!success) | |||
1332 | return HARD_DEP(((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) << 1); | |||
1333 | PREDICATED_PAT (next)((&h_i_d[INSN_UID (next)])->predicated_pat) = new_pat; | |||
1334 | } | |||
1335 | else if (PATTERN (next) != PREDICATED_PAT (next)((&h_i_d[INSN_UID (next)])->predicated_pat)) | |||
1336 | { | |||
1337 | bool success = haifa_change_pattern (next, | |||
1338 | PREDICATED_PAT (next)((&h_i_d[INSN_UID (next)])->predicated_pat)); | |||
1339 | gcc_assert (success)((void)(!(success) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1339, __FUNCTION__), 0 : 0)); | |||
1340 | } | |||
1341 | DEP_STATUS (modify_dep)((modify_dep)->status) |= DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1); | |||
1342 | return DEP_CONTROL((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1); | |||
1343 | } | |||
1344 | ||||
1345 | if (PREDICATED_PAT (next)((&h_i_d[INSN_UID (next)])->predicated_pat) != NULL_RTX(rtx) 0) | |||
1346 | { | |||
1347 | int tick = INSN_TICK (next)((&h_i_d[INSN_UID (next)])->tick); | |||
1348 | bool success = haifa_change_pattern (next, | |||
1349 | ORIG_PAT (next)((&h_i_d[INSN_UID (next)])->orig_pat)); | |||
1350 | INSN_TICK (next)((&h_i_d[INSN_UID (next)])->tick) = tick; | |||
1351 | gcc_assert (success)((void)(!(success) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1351, __FUNCTION__), 0 : 0)); | |||
1352 | } | |||
1353 | ||||
1354 | /* We can't handle the case where there are both speculative and control | |||
1355 | dependencies, so we return HARD_DEP in such a case. Also fail if | |||
1356 | we have speculative dependencies with not enough points, or more than | |||
1357 | one control dependency. */ | |||
1358 | if ((n_spec > 0 && (n_control > 0 || n_replace > 0)) | |||
1359 | || (n_spec > 0 | |||
1360 | /* Too few points? */ | |||
1361 | && ds_weak (new_ds) < spec_info->data_weakness_cutoff) | |||
1362 | || n_control > 0 | |||
1363 | || n_replace > 0) | |||
1364 | return HARD_DEP(((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) << 1); | |||
1365 | ||||
1366 | return new_ds; | |||
1367 | } | |||
1368 | ||||
1369 | /* Pointer to the last instruction scheduled. */ | |||
1370 | static rtx_insn *last_scheduled_insn; | |||
1371 | ||||
1372 | /* Pointer to the last nondebug instruction scheduled within the | |||
1373 | block, or the prev_head of the scheduling block. Used by | |||
1374 | rank_for_schedule, so that insns independent of the last scheduled | |||
1375 | insn will be preferred over dependent instructions. */ | |||
1376 | static rtx_insn *last_nondebug_scheduled_insn; | |||
1377 | ||||
1378 | /* Pointer that iterates through the list of unscheduled insns if we | |||
1379 | have a dbg_cnt enabled. It always points at an insn prior to the | |||
1380 | first unscheduled one. */ | |||
1381 | static rtx_insn *nonscheduled_insns_begin; | |||
1382 | ||||
1383 | /* Compute cost of executing INSN. | |||
1384 | This is the number of cycles between instruction issue and | |||
1385 | instruction results. */ | |||
1386 | int | |||
1387 | insn_sched_cost (rtx_insn *insn) | |||
1388 | { | |||
1389 | int cost; | |||
1390 | ||||
1391 | if (sched_fusion) | |||
1392 | return 0; | |||
1393 | ||||
1394 | if (sel_sched_p ()) | |||
1395 | { | |||
1396 | if (recog_memoized (insn) < 0) | |||
1397 | return 0; | |||
1398 | ||||
1399 | cost = insn_default_latency (insn); | |||
1400 | if (cost < 0) | |||
1401 | cost = 0; | |||
1402 | ||||
1403 | return cost; | |||
1404 | } | |||
1405 | ||||
1406 | cost = INSN_COST (insn)((&h_i_d[INSN_UID (insn)])->cost); | |||
1407 | ||||
1408 | if (cost < 0) | |||
1409 | { | |||
1410 | /* A USE insn, or something else we don't need to | |||
1411 | understand. We can't pass these directly to | |||
1412 | result_ready_cost or insn_default_latency because it will | |||
1413 | trigger a fatal error for unrecognizable insns. */ | |||
1414 | if (recog_memoized (insn) < 0) | |||
1415 | { | |||
1416 | INSN_COST (insn)((&h_i_d[INSN_UID (insn)])->cost) = 0; | |||
1417 | return 0; | |||
1418 | } | |||
1419 | else | |||
1420 | { | |||
1421 | cost = insn_default_latency (insn); | |||
1422 | if (cost < 0) | |||
1423 | cost = 0; | |||
1424 | ||||
1425 | INSN_COST (insn)((&h_i_d[INSN_UID (insn)])->cost) = cost; | |||
1426 | } | |||
1427 | } | |||
1428 | ||||
1429 | return cost; | |||
1430 | } | |||
1431 | ||||
1432 | /* Compute cost of dependence LINK. | |||
1433 | This is the number of cycles between instruction issue and | |||
1434 | instruction results. | |||
1435 | ??? We also use this function to call recog_memoized on all insns. */ | |||
1436 | int | |||
1437 | dep_cost_1 (dep_t link, dw_t dw) | |||
1438 | { | |||
1439 | rtx_insn *insn = DEP_PRO (link)((link)->pro); | |||
1440 | rtx_insn *used = DEP_CON (link)((link)->con); | |||
1441 | int cost; | |||
1442 | ||||
1443 | if (DEP_COST (link)((link)->cost) != UNKNOWN_DEP_COST((int) ((unsigned int) -1 << 19))) | |||
1444 | return DEP_COST (link)((link)->cost); | |||
1445 | ||||
1446 | if (delay_htab) | |||
1447 | { | |||
1448 | struct delay_pair *delay_entry; | |||
1449 | delay_entry | |||
1450 | = delay_htab_i2->find_with_hash (used, htab_hash_pointer (used)); | |||
1451 | if (delay_entry) | |||
1452 | { | |||
1453 | if (delay_entry->i1 == insn) | |||
1454 | { | |||
1455 | DEP_COST (link)((link)->cost) = pair_delay (delay_entry); | |||
1456 | return DEP_COST (link)((link)->cost); | |||
1457 | } | |||
1458 | } | |||
1459 | } | |||
1460 | ||||
1461 | /* A USE insn should never require the value used to be computed. | |||
1462 | This allows the computation of a function's result and parameter | |||
1463 | values to overlap the return and call. We don't care about the | |||
1464 | dependence cost when only decreasing register pressure. */ | |||
1465 | if (recog_memoized (used) < 0) | |||
1466 | { | |||
1467 | cost = 0; | |||
1468 | recog_memoized (insn); | |||
1469 | } | |||
1470 | else | |||
1471 | { | |||
1472 | enum reg_note dep_type = DEP_TYPE (link)((link)->type); | |||
1473 | ||||
1474 | cost = insn_sched_cost (insn); | |||
1475 | ||||
1476 | if (INSN_CODE (insn)(((insn)->u.fld[5]).rt_int) >= 0) | |||
1477 | { | |||
1478 | if (dep_type == REG_DEP_ANTI) | |||
1479 | cost = 0; | |||
1480 | else if (dep_type == REG_DEP_OUTPUT) | |||
1481 | { | |||
1482 | cost = (insn_default_latency (insn) | |||
1483 | - insn_default_latency (used)); | |||
1484 | if (cost <= 0) | |||
1485 | cost = 1; | |||
1486 | } | |||
1487 | else if (bypass_p (insn)) | |||
1488 | cost = insn_latency (insn, used); | |||
1489 | } | |||
1490 | ||||
1491 | ||||
1492 | if (targetm.sched.adjust_cost) | |||
1493 | cost = targetm.sched.adjust_cost (used, (int) dep_type, insn, cost, | |||
1494 | dw); | |||
1495 | ||||
1496 | if (cost < 0) | |||
1497 | cost = 0; | |||
1498 | } | |||
1499 | ||||
1500 | DEP_COST (link)((link)->cost) = cost; | |||
1501 | return cost; | |||
1502 | } | |||
1503 | ||||
1504 | /* Compute cost of dependence LINK. | |||
1505 | This is the number of cycles between instruction issue and | |||
1506 | instruction results. */ | |||
1507 | int | |||
1508 | dep_cost (dep_t link) | |||
1509 | { | |||
1510 | return dep_cost_1 (link, 0); | |||
1511 | } | |||
1512 | ||||
1513 | /* Use this sel-sched.cc friendly function in reorder2 instead of increasing | |||
1514 | INSN_PRIORITY explicitly. */ | |||
1515 | void | |||
1516 | increase_insn_priority (rtx_insn *insn, int amount) | |||
1517 | { | |||
1518 | if (!sel_sched_p ()) | |||
1519 | { | |||
1520 | /* We're dealing with haifa-sched.cc INSN_PRIORITY. */ | |||
1521 | if (INSN_PRIORITY_KNOWN (insn)(((&h_i_d[INSN_UID (insn)])->priority_status) > 0)) | |||
1522 | INSN_PRIORITY (insn)((&h_i_d[INSN_UID (insn)])->priority) += amount; | |||
1523 | } | |||
1524 | else | |||
1525 | { | |||
1526 | /* In sel-sched.cc INSN_PRIORITY is not kept up to date. | |||
1527 | Use EXPR_PRIORITY instead. */ | |||
1528 | sel_add_to_insn_priority (insn, amount); | |||
1529 | } | |||
1530 | } | |||
1531 | ||||
1532 | /* Return 'true' if DEP should be included in priority calculations. */ | |||
1533 | static bool | |||
1534 | contributes_to_priority_p (dep_t dep) | |||
1535 | { | |||
1536 | if (DEBUG_INSN_P (DEP_CON (dep))(((enum rtx_code) (((dep)->con))->code) == DEBUG_INSN) | |||
1537 | || DEBUG_INSN_P (DEP_PRO (dep))(((enum rtx_code) (((dep)->pro))->code) == DEBUG_INSN)) | |||
1538 | return false; | |||
1539 | ||||
1540 | /* Critical path is meaningful in block boundaries only. */ | |||
1541 | if (!current_sched_info->contributes_to_priority (DEP_CON (dep)((dep)->con), | |||
1542 | DEP_PRO (dep)((dep)->pro))) | |||
1543 | return false; | |||
1544 | ||||
1545 | if (DEP_REPLACE (dep)((dep)->replace) != NULLnullptr) | |||
1546 | return false; | |||
1547 | ||||
1548 | /* If flag COUNT_SPEC_IN_CRITICAL_PATH is set, | |||
1549 | then speculative instructions will less likely be | |||
1550 | scheduled. That is because the priority of | |||
1551 | their producers will increase, and, thus, the | |||
1552 | producers will more likely be scheduled, thus, | |||
1553 | resolving the dependence. */ | |||
1554 | if (sched_deps_info->generate_spec_deps | |||
1555 | && !(spec_info->flags & COUNT_SPEC_IN_CRITICAL_PATH) | |||
1556 | && (DEP_STATUS (dep)((dep)->status) & SPECULATIVE(((((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BEGIN_DATA_BITS_OFFSET ) | (((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BE_IN_DATA_BITS_OFFSET)) | ((((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BEGIN_CONTROL_BITS_OFFSET) | (((ds_t ) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BE_IN_CONTROL_BITS_OFFSET ))))) | |||
1557 | return false; | |||
1558 | ||||
1559 | return true; | |||
1560 | } | |||
1561 | ||||
1562 | /* Compute the number of nondebug deps in list LIST for INSN. */ | |||
1563 | ||||
1564 | static int | |||
1565 | dep_list_size (rtx_insn *insn, sd_list_types_def list) | |||
1566 | { | |||
1567 | sd_iterator_def sd_it; | |||
1568 | dep_t dep; | |||
1569 | int dbgcount = 0, nodbgcount = 0; | |||
1570 | ||||
1571 | if (!MAY_HAVE_DEBUG_INSNS(global_options.x_debug_nonbind_markers_p || global_options.x_flag_var_tracking_assignments )) | |||
1572 | return sd_lists_size (insn, list); | |||
1573 | ||||
1574 | FOR_EACH_DEP (insn, list, sd_it, dep)for ((sd_it) = sd_iterator_start ((insn), (list)); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
1575 | { | |||
1576 | if (DEBUG_INSN_P (DEP_CON (dep))(((enum rtx_code) (((dep)->con))->code) == DEBUG_INSN)) | |||
1577 | dbgcount++; | |||
1578 | else if (!DEBUG_INSN_P (DEP_PRO (dep))(((enum rtx_code) (((dep)->pro))->code) == DEBUG_INSN)) | |||
1579 | nodbgcount++; | |||
1580 | } | |||
1581 | ||||
1582 | gcc_assert (dbgcount + nodbgcount == sd_lists_size (insn, list))((void)(!(dbgcount + nodbgcount == sd_lists_size (insn, list) ) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1582, __FUNCTION__), 0 : 0)); | |||
1583 | ||||
1584 | return nodbgcount; | |||
1585 | } | |||
1586 | ||||
1587 | bool sched_fusion; | |||
1588 | ||||
1589 | /* Compute the priority number for INSN. */ | |||
1590 | static int | |||
1591 | priority (rtx_insn *insn, bool force_recompute) | |||
1592 | { | |||
1593 | if (! INSN_P (insn)(((((enum rtx_code) (insn)->code) == INSN) || (((enum rtx_code ) (insn)->code) == JUMP_INSN) || (((enum rtx_code) (insn)-> code) == CALL_INSN)) || (((enum rtx_code) (insn)->code) == DEBUG_INSN))) | |||
1594 | return 0; | |||
1595 | ||||
1596 | /* We should not be interested in priority of an already scheduled insn. */ | |||
1597 | gcc_assert (QUEUE_INDEX (insn) != QUEUE_SCHEDULED)((void)(!(((&h_i_d[INSN_UID (insn)])->queue_index) != ( -3)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1597, __FUNCTION__), 0 : 0)); | |||
1598 | ||||
1599 | if (force_recompute || !INSN_PRIORITY_KNOWN (insn)(((&h_i_d[INSN_UID (insn)])->priority_status) > 0)) | |||
1600 | { | |||
1601 | int this_priority = -1; | |||
1602 | ||||
1603 | if (sched_fusion) | |||
1604 | { | |||
1605 | int this_fusion_priority; | |||
1606 | ||||
1607 | targetm.sched.fusion_priority (insn, FUSION_MAX_PRIORITY(2147483647), | |||
1608 | &this_fusion_priority, &this_priority); | |||
1609 | INSN_FUSION_PRIORITY (insn)((&h_i_d[INSN_UID (insn)])->fusion_priority) = this_fusion_priority; | |||
1610 | } | |||
1611 | else if (dep_list_size (insn, SD_LIST_FORW(4)) == 0) | |||
1612 | /* ??? We should set INSN_PRIORITY to insn_sched_cost when and insn | |||
1613 | has some forward deps but all of them are ignored by | |||
1614 | contributes_to_priority hook. At the moment we set priority of | |||
1615 | such insn to 0. */ | |||
1616 | this_priority = insn_sched_cost (insn); | |||
1617 | else | |||
1618 | { | |||
1619 | rtx_insn *prev_first, *twin; | |||
1620 | basic_block rec; | |||
1621 | ||||
1622 | /* For recovery check instructions we calculate priority slightly | |||
1623 | different than that of normal instructions. Instead of walking | |||
1624 | through INSN_FORW_DEPS (check) list, we walk through | |||
1625 | INSN_FORW_DEPS list of each instruction in the corresponding | |||
1626 | recovery block. */ | |||
1627 | ||||
1628 | /* Selective scheduling does not define RECOVERY_BLOCK macro. */ | |||
1629 | rec = sel_sched_p () ? NULLnullptr : RECOVERY_BLOCK (insn)((&h_i_d[INSN_UID (insn)])->recovery_block); | |||
1630 | if (!rec || rec == EXIT_BLOCK_PTR_FOR_FN (cfun)(((cfun + 0))->cfg->x_exit_block_ptr)) | |||
1631 | { | |||
1632 | prev_first = PREV_INSN (insn); | |||
1633 | twin = insn; | |||
1634 | } | |||
1635 | else | |||
1636 | { | |||
1637 | prev_first = NEXT_INSN (BB_HEAD (rec)(rec)->il.x.head_); | |||
1638 | twin = PREV_INSN (BB_END (rec)(rec)->il.x.rtl->end_); | |||
1639 | } | |||
1640 | ||||
1641 | do | |||
1642 | { | |||
1643 | sd_iterator_def sd_it; | |||
1644 | dep_t dep; | |||
1645 | ||||
1646 | FOR_EACH_DEP (twin, SD_LIST_FORW, sd_it, dep)for ((sd_it) = sd_iterator_start ((twin), ((4))); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
1647 | { | |||
1648 | rtx_insn *next; | |||
1649 | int next_priority; | |||
1650 | ||||
1651 | next = DEP_CON (dep)((dep)->con); | |||
1652 | ||||
1653 | if (BLOCK_FOR_INSN (next) != rec) | |||
1654 | { | |||
1655 | int cost; | |||
1656 | ||||
1657 | if (!contributes_to_priority_p (dep)) | |||
1658 | continue; | |||
1659 | ||||
1660 | if (twin == insn) | |||
1661 | cost = dep_cost (dep); | |||
1662 | else | |||
1663 | { | |||
1664 | struct _dep _dep1, *dep1 = &_dep1; | |||
1665 | ||||
1666 | init_dep (dep1, insn, next, REG_DEP_ANTI); | |||
1667 | ||||
1668 | cost = dep_cost (dep1); | |||
1669 | } | |||
1670 | ||||
1671 | next_priority = cost + priority (next); | |||
1672 | ||||
1673 | if (next_priority > this_priority) | |||
1674 | this_priority = next_priority; | |||
1675 | } | |||
1676 | } | |||
1677 | ||||
1678 | twin = PREV_INSN (twin); | |||
1679 | } | |||
1680 | while (twin != prev_first); | |||
1681 | } | |||
1682 | ||||
1683 | if (this_priority < 0) | |||
1684 | { | |||
1685 | gcc_assert (this_priority == -1)((void)(!(this_priority == -1) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1685, __FUNCTION__), 0 : 0)); | |||
1686 | ||||
1687 | this_priority = insn_sched_cost (insn); | |||
1688 | } | |||
1689 | ||||
1690 | INSN_PRIORITY (insn)((&h_i_d[INSN_UID (insn)])->priority) = this_priority; | |||
1691 | INSN_PRIORITY_STATUS (insn)((&h_i_d[INSN_UID (insn)])->priority_status) = 1; | |||
1692 | } | |||
1693 | ||||
1694 | return INSN_PRIORITY (insn)((&h_i_d[INSN_UID (insn)])->priority); | |||
1695 | } | |||
1696 | ||||
1697 | /* Macros and functions for keeping the priority queue sorted, and | |||
1698 | dealing with queuing and dequeuing of instructions. */ | |||
1699 | ||||
1700 | /* For each pressure class CL, set DEATH[CL] to the number of registers | |||
1701 | in that class that die in INSN. */ | |||
1702 | ||||
1703 | static void | |||
1704 | calculate_reg_deaths (rtx_insn *insn, int *death) | |||
1705 | { | |||
1706 | int i; | |||
1707 | struct reg_use_data *use; | |||
1708 | ||||
1709 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
1710 | death[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]] = 0; | |||
1711 | for (use = INSN_REG_USE_LIST (insn)((&h_i_d[INSN_UID (insn)])->reg_use_list); use != NULLnullptr; use = use->next_insn_use) | |||
1712 | if (dying_use_p (use)) | |||
1713 | mark_regno_birth_or_death (0, death, use->regno, true); | |||
1714 | } | |||
1715 | ||||
1716 | /* Setup info about the current register pressure impact of scheduling | |||
1717 | INSN at the current scheduling point. */ | |||
1718 | static void | |||
1719 | setup_insn_reg_pressure_info (rtx_insn *insn) | |||
1720 | { | |||
1721 | int i, change, before, after, hard_regno; | |||
1722 | int excess_cost_change; | |||
1723 | machine_mode mode; | |||
1724 | enum reg_class cl; | |||
1725 | struct reg_pressure_data *pressure_info; | |||
1726 | int *max_reg_pressure; | |||
1727 | static int death[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
1728 | ||||
1729 | gcc_checking_assert (!DEBUG_INSN_P (insn))((void)(!(!(((enum rtx_code) (insn)->code) == DEBUG_INSN)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1729, __FUNCTION__), 0 : 0)); | |||
1730 | ||||
1731 | excess_cost_change = 0; | |||
1732 | calculate_reg_deaths (insn, death); | |||
1733 | pressure_info = INSN_REG_PRESSURE (insn)((&h_i_d[INSN_UID (insn)])->reg_pressure); | |||
1734 | max_reg_pressure = INSN_MAX_REG_PRESSURE (insn)((&h_i_d[INSN_UID (insn)])->max_reg_pressure); | |||
1735 | gcc_assert (pressure_info != NULL && max_reg_pressure != NULL)((void)(!(pressure_info != nullptr && max_reg_pressure != nullptr) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1735, __FUNCTION__), 0 : 0)); | |||
1736 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
1737 | { | |||
1738 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]; | |||
1739 | gcc_assert (curr_reg_pressure[cl] >= 0)((void)(!(curr_reg_pressure[cl] >= 0) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1739, __FUNCTION__), 0 : 0)); | |||
1740 | change = (int) pressure_info[i].set_increase - death[cl]; | |||
1741 | before = MAX (0, max_reg_pressure[i] - sched_class_regs_num[cl])((0) > (max_reg_pressure[i] - sched_class_regs_num[cl]) ? ( 0) : (max_reg_pressure[i] - sched_class_regs_num[cl])); | |||
1742 | after = MAX (0, max_reg_pressure[i] + change((0) > (max_reg_pressure[i] + change - sched_class_regs_num [cl]) ? (0) : (max_reg_pressure[i] + change - sched_class_regs_num [cl])) | |||
1743 | - sched_class_regs_num[cl])((0) > (max_reg_pressure[i] + change - sched_class_regs_num [cl]) ? (0) : (max_reg_pressure[i] + change - sched_class_regs_num [cl])); | |||
1744 | hard_regno = ira_class_hard_regs(this_target_ira->x_ira_class_hard_regs)[cl][0]; | |||
1745 | gcc_assert (hard_regno >= 0)((void)(!(hard_regno >= 0) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1745, __FUNCTION__), 0 : 0)); | |||
1746 | mode = reg_raw_mode(this_target_regs->x_reg_raw_mode)[hard_regno]; | |||
1747 | excess_cost_change += ((after - before) | |||
1748 | * (ira_memory_move_cost(this_target_ira->x_ira_memory_move_cost)[mode][cl][0] | |||
1749 | + ira_memory_move_cost(this_target_ira->x_ira_memory_move_cost)[mode][cl][1])); | |||
1750 | } | |||
1751 | INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insn)((&h_i_d[INSN_UID (insn)])->reg_pressure_excess_cost_change ) = excess_cost_change; | |||
1752 | } | |||
1753 | ||||
1754 | /* This is the first page of code related to SCHED_PRESSURE_MODEL. | |||
1755 | It tries to make the scheduler take register pressure into account | |||
1756 | without introducing too many unnecessary stalls. It hooks into the | |||
1757 | main scheduling algorithm at several points: | |||
1758 | ||||
1759 | - Before scheduling starts, model_start_schedule constructs a | |||
1760 | "model schedule" for the current block. This model schedule is | |||
1761 | chosen solely to keep register pressure down. It does not take the | |||
1762 | target's pipeline or the original instruction order into account, | |||
1763 | except as a tie-breaker. It also doesn't work to a particular | |||
1764 | pressure limit. | |||
1765 | ||||
1766 | This model schedule gives us an idea of what pressure can be | |||
1767 | achieved for the block and gives us an example of a schedule that | |||
1768 | keeps to that pressure. It also makes the final schedule less | |||
1769 | dependent on the original instruction order. This is important | |||
1770 | because the original order can either be "wide" (many values live | |||
1771 | at once, such as in user-scheduled code) or "narrow" (few values | |||
1772 | live at once, such as after loop unrolling, where several | |||
1773 | iterations are executed sequentially). | |||
1774 | ||||
1775 | We do not apply this model schedule to the rtx stream. We simply | |||
1776 | record it in model_schedule. We also compute the maximum pressure, | |||
1777 | MP, that was seen during this schedule. | |||
1778 | ||||
1779 | - Instructions are added to the ready queue even if they require | |||
1780 | a stall. The length of the stall is instead computed as: | |||
1781 | ||||
1782 | MAX (INSN_TICK (INSN) - clock_var, 0) | |||
1783 | ||||
1784 | (= insn_delay). This allows rank_for_schedule to choose between | |||
1785 | introducing a deliberate stall or increasing pressure. | |||
1786 | ||||
1787 | - Before sorting the ready queue, model_set_excess_costs assigns | |||
1788 | a pressure-based cost to each ready instruction in the queue. | |||
1789 | This is the instruction's INSN_REG_PRESSURE_EXCESS_COST_CHANGE | |||
1790 | (ECC for short) and is effectively measured in cycles. | |||
1791 | ||||
1792 | - rank_for_schedule ranks instructions based on: | |||
1793 | ||||
1794 | ECC (insn) + insn_delay (insn) | |||
1795 | ||||
1796 | then as: | |||
1797 | ||||
1798 | insn_delay (insn) | |||
1799 | ||||
1800 | So, for example, an instruction X1 with an ECC of 1 that can issue | |||
1801 | now will win over an instruction X0 with an ECC of zero that would | |||
1802 | introduce a stall of one cycle. However, an instruction X2 with an | |||
1803 | ECC of 2 that can issue now will lose to both X0 and X1. | |||
1804 | ||||
1805 | - When an instruction is scheduled, model_recompute updates the model | |||
1806 | schedule with the new pressures (some of which might now exceed the | |||
1807 | original maximum pressure MP). model_update_limit_points then searches | |||
1808 | for the new point of maximum pressure, if not already known. */ | |||
1809 | ||||
1810 | /* Used to separate high-verbosity debug information for SCHED_PRESSURE_MODEL | |||
1811 | from surrounding debug information. */ | |||
1812 | #define MODEL_BAR";;\t\t+------------------------------------------------------\n" \ | |||
1813 | ";;\t\t+------------------------------------------------------\n" | |||
1814 | ||||
1815 | /* Information about the pressure on a particular register class at a | |||
1816 | particular point of the model schedule. */ | |||
1817 | struct model_pressure_data { | |||
1818 | /* The pressure at this point of the model schedule, or -1 if the | |||
1819 | point is associated with an instruction that has already been | |||
1820 | scheduled. */ | |||
1821 | int ref_pressure; | |||
1822 | ||||
1823 | /* The maximum pressure during or after this point of the model schedule. */ | |||
1824 | int max_pressure; | |||
1825 | }; | |||
1826 | ||||
1827 | /* Per-instruction information that is used while building the model | |||
1828 | schedule. Here, "schedule" refers to the model schedule rather | |||
1829 | than the main schedule. */ | |||
1830 | struct model_insn_info { | |||
1831 | /* The instruction itself. */ | |||
1832 | rtx_insn *insn; | |||
1833 | ||||
1834 | /* If this instruction is in model_worklist, these fields link to the | |||
1835 | previous (higher-priority) and next (lower-priority) instructions | |||
1836 | in the list. */ | |||
1837 | struct model_insn_info *prev; | |||
1838 | struct model_insn_info *next; | |||
1839 | ||||
1840 | /* While constructing the schedule, QUEUE_INDEX describes whether an | |||
1841 | instruction has already been added to the schedule (QUEUE_SCHEDULED), | |||
1842 | is in model_worklist (QUEUE_READY), or neither (QUEUE_NOWHERE). | |||
1843 | old_queue records the value that QUEUE_INDEX had before scheduling | |||
1844 | started, so that we can restore it once the schedule is complete. */ | |||
1845 | int old_queue; | |||
1846 | ||||
1847 | /* The relative importance of an unscheduled instruction. Higher | |||
1848 | values indicate greater importance. */ | |||
1849 | unsigned int model_priority; | |||
1850 | ||||
1851 | /* The length of the longest path of satisfied true dependencies | |||
1852 | that leads to this instruction. */ | |||
1853 | unsigned int depth; | |||
1854 | ||||
1855 | /* The length of the longest path of dependencies of any kind | |||
1856 | that leads from this instruction. */ | |||
1857 | unsigned int alap; | |||
1858 | ||||
1859 | /* The number of predecessor nodes that must still be scheduled. */ | |||
1860 | int unscheduled_preds; | |||
1861 | }; | |||
1862 | ||||
1863 | /* Information about the pressure limit for a particular register class. | |||
1864 | This structure is used when applying a model schedule to the main | |||
1865 | schedule. */ | |||
1866 | struct model_pressure_limit { | |||
1867 | /* The maximum register pressure seen in the original model schedule. */ | |||
1868 | int orig_pressure; | |||
1869 | ||||
1870 | /* The maximum register pressure seen in the current model schedule | |||
1871 | (which excludes instructions that have already been scheduled). */ | |||
1872 | int pressure; | |||
1873 | ||||
1874 | /* The point of the current model schedule at which PRESSURE is first | |||
1875 | reached. It is set to -1 if the value needs to be recomputed. */ | |||
1876 | int point; | |||
1877 | }; | |||
1878 | ||||
1879 | /* Describes a particular way of measuring register pressure. */ | |||
1880 | struct model_pressure_group { | |||
1881 | /* Index PCI describes the maximum pressure on ira_pressure_classes[PCI]. */ | |||
1882 | struct model_pressure_limit limits[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
1883 | ||||
1884 | /* Index (POINT * ira_num_pressure_classes + PCI) describes the pressure | |||
1885 | on register class ira_pressure_classes[PCI] at point POINT of the | |||
1886 | current model schedule. A POINT of model_num_insns describes the | |||
1887 | pressure at the end of the schedule. */ | |||
1888 | struct model_pressure_data *model; | |||
1889 | }; | |||
1890 | ||||
1891 | /* Index POINT gives the instruction at point POINT of the model schedule. | |||
1892 | This array doesn't change during main scheduling. */ | |||
1893 | static vec<rtx_insn *> model_schedule; | |||
1894 | ||||
1895 | /* The list of instructions in the model worklist, sorted in order of | |||
1896 | decreasing priority. */ | |||
1897 | static struct model_insn_info *model_worklist; | |||
1898 | ||||
1899 | /* Index I describes the instruction with INSN_LUID I. */ | |||
1900 | static struct model_insn_info *model_insns; | |||
1901 | ||||
1902 | /* The number of instructions in the model schedule. */ | |||
1903 | static int model_num_insns; | |||
1904 | ||||
1905 | /* The index of the first instruction in model_schedule that hasn't yet been | |||
1906 | added to the main schedule, or model_num_insns if all of them have. */ | |||
1907 | static int model_curr_point; | |||
1908 | ||||
1909 | /* Describes the pressure before each instruction in the model schedule. */ | |||
1910 | static struct model_pressure_group model_before_pressure; | |||
1911 | ||||
1912 | /* The first unused model_priority value (as used in model_insn_info). */ | |||
1913 | static unsigned int model_next_priority; | |||
1914 | ||||
1915 | ||||
1916 | /* The model_pressure_data for ira_pressure_classes[PCI] in GROUP | |||
1917 | at point POINT of the model schedule. */ | |||
1918 | #define MODEL_PRESSURE_DATA(GROUP, POINT, PCI)(&(GROUP)->model[(POINT) * (this_target_ira->x_ira_pressure_classes_num ) + (PCI)]) \ | |||
1919 | (&(GROUP)->model[(POINT) * ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num) + (PCI)]) | |||
1920 | ||||
1921 | /* The maximum pressure on ira_pressure_classes[PCI] in GROUP at or | |||
1922 | after point POINT of the model schedule. */ | |||
1923 | #define MODEL_MAX_PRESSURE(GROUP, POINT, PCI)((&(GROUP)->model[(POINT) * (this_target_ira->x_ira_pressure_classes_num ) + (PCI)])->max_pressure) \ | |||
1924 | (MODEL_PRESSURE_DATA (GROUP, POINT, PCI)(&(GROUP)->model[(POINT) * (this_target_ira->x_ira_pressure_classes_num ) + (PCI)])->max_pressure) | |||
1925 | ||||
1926 | /* The pressure on ira_pressure_classes[PCI] in GROUP at point POINT | |||
1927 | of the model schedule. */ | |||
1928 | #define MODEL_REF_PRESSURE(GROUP, POINT, PCI)((&(GROUP)->model[(POINT) * (this_target_ira->x_ira_pressure_classes_num ) + (PCI)])->ref_pressure) \ | |||
1929 | (MODEL_PRESSURE_DATA (GROUP, POINT, PCI)(&(GROUP)->model[(POINT) * (this_target_ira->x_ira_pressure_classes_num ) + (PCI)])->ref_pressure) | |||
1930 | ||||
1931 | /* Information about INSN that is used when creating the model schedule. */ | |||
1932 | #define MODEL_INSN_INFO(INSN)(&model_insns[(sched_luids[INSN_UID (INSN)])]) \ | |||
1933 | (&model_insns[INSN_LUID (INSN)(sched_luids[INSN_UID (INSN)])]) | |||
1934 | ||||
1935 | /* The instruction at point POINT of the model schedule. */ | |||
1936 | #define MODEL_INSN(POINT)(model_schedule[POINT]) \ | |||
1937 | (model_schedule[POINT]) | |||
1938 | ||||
1939 | ||||
1940 | /* Return INSN's index in the model schedule, or model_num_insns if it | |||
1941 | doesn't belong to that schedule. */ | |||
1942 | ||||
1943 | static int | |||
1944 | model_index (rtx_insn *insn) | |||
1945 | { | |||
1946 | if (INSN_MODEL_INDEX (insn)((&h_i_d[INSN_UID (insn)])->model_index) == 0) | |||
1947 | return model_num_insns; | |||
1948 | return INSN_MODEL_INDEX (insn)((&h_i_d[INSN_UID (insn)])->model_index) - 1; | |||
1949 | } | |||
1950 | ||||
1951 | /* Make sure that GROUP->limits is up-to-date for the current point | |||
1952 | of the model schedule. */ | |||
1953 | ||||
1954 | static void | |||
1955 | model_update_limit_points_in_group (struct model_pressure_group *group) | |||
1956 | { | |||
1957 | int pci, max_pressure, point; | |||
1958 | ||||
1959 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
1960 | { | |||
1961 | /* We may have passed the final point at which the pressure in | |||
1962 | group->limits[pci].pressure was reached. Update the limit if so. */ | |||
1963 | max_pressure = MODEL_MAX_PRESSURE (group, model_curr_point, pci)((&(group)->model[(model_curr_point) * (this_target_ira ->x_ira_pressure_classes_num) + (pci)])->max_pressure); | |||
1964 | group->limits[pci].pressure = max_pressure; | |||
1965 | ||||
1966 | /* Find the point at which MAX_PRESSURE is first reached. We need | |||
1967 | to search in three cases: | |||
1968 | ||||
1969 | - We've already moved past the previous pressure point. | |||
1970 | In this case we search forward from model_curr_point. | |||
1971 | ||||
1972 | - We scheduled the previous point of maximum pressure ahead of | |||
1973 | its position in the model schedule, but doing so didn't bring | |||
1974 | the pressure point earlier. In this case we search forward | |||
1975 | from that previous pressure point. | |||
1976 | ||||
1977 | - Scheduling an instruction early caused the maximum pressure | |||
1978 | to decrease. In this case we will have set the pressure | |||
1979 | point to -1, and we search forward from model_curr_point. */ | |||
1980 | point = MAX (group->limits[pci].point, model_curr_point)((group->limits[pci].point) > (model_curr_point) ? (group ->limits[pci].point) : (model_curr_point)); | |||
1981 | while (point < model_num_insns | |||
1982 | && MODEL_REF_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->ref_pressure) < max_pressure) | |||
1983 | point++; | |||
1984 | group->limits[pci].point = point; | |||
1985 | ||||
1986 | gcc_assert (MODEL_REF_PRESSURE (group, point, pci) == max_pressure)((void)(!(((&(group)->model[(point) * (this_target_ira ->x_ira_pressure_classes_num) + (pci)])->ref_pressure) == max_pressure) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1986, __FUNCTION__), 0 : 0)); | |||
1987 | gcc_assert (MODEL_MAX_PRESSURE (group, point, pci) == max_pressure)((void)(!(((&(group)->model[(point) * (this_target_ira ->x_ira_pressure_classes_num) + (pci)])->max_pressure) == max_pressure) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 1987, __FUNCTION__), 0 : 0)); | |||
1988 | } | |||
1989 | } | |||
1990 | ||||
1991 | /* Make sure that all register-pressure limits are up-to-date for the | |||
1992 | current position in the model schedule. */ | |||
1993 | ||||
1994 | static void | |||
1995 | model_update_limit_points (void) | |||
1996 | { | |||
1997 | model_update_limit_points_in_group (&model_before_pressure); | |||
1998 | } | |||
1999 | ||||
2000 | /* Return the model_index of the last unscheduled use in chain USE | |||
2001 | outside of USE's instruction. Return -1 if there are no other uses, | |||
2002 | or model_num_insns if the register is live at the end of the block. */ | |||
2003 | ||||
2004 | static int | |||
2005 | model_last_use_except (struct reg_use_data *use) | |||
2006 | { | |||
2007 | struct reg_use_data *next; | |||
2008 | int last, index; | |||
2009 | ||||
2010 | last = -1; | |||
2011 | for (next = use->next_regno_use; next != use; next = next->next_regno_use) | |||
2012 | if (NONDEBUG_INSN_P (next->insn)((((enum rtx_code) (next->insn)->code) == INSN) || (((enum rtx_code) (next->insn)->code) == JUMP_INSN) || (((enum rtx_code) (next->insn)->code) == CALL_INSN)) | |||
2013 | && QUEUE_INDEX (next->insn)((&h_i_d[INSN_UID (next->insn)])->queue_index) != QUEUE_SCHEDULED(-3)) | |||
2014 | { | |||
2015 | index = model_index (next->insn); | |||
2016 | if (index == model_num_insns) | |||
2017 | return model_num_insns; | |||
2018 | if (last < index) | |||
2019 | last = index; | |||
2020 | } | |||
2021 | return last; | |||
2022 | } | |||
2023 | ||||
2024 | /* An instruction with model_index POINT has just been scheduled, and it | |||
2025 | adds DELTA to the pressure on ira_pressure_classes[PCI] after POINT - 1. | |||
2026 | Update MODEL_REF_PRESSURE (GROUP, POINT, PCI) and | |||
2027 | MODEL_MAX_PRESSURE (GROUP, POINT, PCI) accordingly. */ | |||
2028 | ||||
2029 | static void | |||
2030 | model_start_update_pressure (struct model_pressure_group *group, | |||
2031 | int point, int pci, int delta) | |||
2032 | { | |||
2033 | int next_max_pressure; | |||
2034 | ||||
2035 | if (point == model_num_insns) | |||
2036 | { | |||
2037 | /* The instruction wasn't part of the model schedule; it was moved | |||
2038 | from a different block. Update the pressure for the end of | |||
2039 | the model schedule. */ | |||
2040 | MODEL_REF_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->ref_pressure) += delta; | |||
2041 | MODEL_MAX_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->max_pressure) += delta; | |||
2042 | } | |||
2043 | else | |||
2044 | { | |||
2045 | /* Record that this instruction has been scheduled. Nothing now | |||
2046 | changes between POINT and POINT + 1, so get the maximum pressure | |||
2047 | from the latter. If the maximum pressure decreases, the new | |||
2048 | pressure point may be before POINT. */ | |||
2049 | MODEL_REF_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->ref_pressure) = -1; | |||
2050 | next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci)((&(group)->model[(point + 1) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->max_pressure); | |||
2051 | if (MODEL_MAX_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->max_pressure) > next_max_pressure) | |||
2052 | { | |||
2053 | MODEL_MAX_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->max_pressure) = next_max_pressure; | |||
2054 | if (group->limits[pci].point == point) | |||
2055 | group->limits[pci].point = -1; | |||
2056 | } | |||
2057 | } | |||
2058 | } | |||
2059 | ||||
2060 | /* Record that scheduling a later instruction has changed the pressure | |||
2061 | at point POINT of the model schedule by DELTA (which might be 0). | |||
2062 | Update GROUP accordingly. Return nonzero if these changes might | |||
2063 | trigger changes to previous points as well. */ | |||
2064 | ||||
2065 | static int | |||
2066 | model_update_pressure (struct model_pressure_group *group, | |||
2067 | int point, int pci, int delta) | |||
2068 | { | |||
2069 | int ref_pressure, max_pressure, next_max_pressure; | |||
2070 | ||||
2071 | /* If POINT hasn't yet been scheduled, update its pressure. */ | |||
2072 | ref_pressure = MODEL_REF_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->ref_pressure); | |||
2073 | if (ref_pressure >= 0 && delta != 0) | |||
2074 | { | |||
2075 | ref_pressure += delta; | |||
2076 | MODEL_REF_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->ref_pressure) = ref_pressure; | |||
2077 | ||||
2078 | /* Check whether the maximum pressure in the overall schedule | |||
2079 | has increased. (This means that the MODEL_MAX_PRESSURE of | |||
2080 | every point <= POINT will need to increase too; see below.) */ | |||
2081 | if (group->limits[pci].pressure < ref_pressure) | |||
2082 | group->limits[pci].pressure = ref_pressure; | |||
2083 | ||||
2084 | /* If we are at maximum pressure, and the maximum pressure | |||
2085 | point was previously unknown or later than POINT, | |||
2086 | bring it forward. */ | |||
2087 | if (group->limits[pci].pressure == ref_pressure | |||
2088 | && !IN_RANGE (group->limits[pci].point, 0, point)((unsigned long) (group->limits[pci].point) - (unsigned long ) (0) <= (unsigned long) (point) - (unsigned long) (0))) | |||
2089 | group->limits[pci].point = point; | |||
2090 | ||||
2091 | /* If POINT used to be the point of maximum pressure, but isn't | |||
2092 | any longer, we need to recalculate it using a forward walk. */ | |||
2093 | if (group->limits[pci].pressure > ref_pressure | |||
2094 | && group->limits[pci].point == point) | |||
2095 | group->limits[pci].point = -1; | |||
2096 | } | |||
2097 | ||||
2098 | /* Update the maximum pressure at POINT. Changes here might also | |||
2099 | affect the maximum pressure at POINT - 1. */ | |||
2100 | next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci)((&(group)->model[(point + 1) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->max_pressure); | |||
2101 | max_pressure = MAX (ref_pressure, next_max_pressure)((ref_pressure) > (next_max_pressure) ? (ref_pressure) : ( next_max_pressure)); | |||
2102 | if (MODEL_MAX_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->max_pressure) != max_pressure) | |||
2103 | { | |||
2104 | MODEL_MAX_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->max_pressure) = max_pressure; | |||
2105 | return 1; | |||
2106 | } | |||
2107 | return 0; | |||
2108 | } | |||
2109 | ||||
2110 | /* INSN has just been scheduled. Update the model schedule accordingly. */ | |||
2111 | ||||
2112 | static void | |||
2113 | model_recompute (rtx_insn *insn) | |||
2114 | { | |||
2115 | struct { | |||
2116 | int last_use; | |||
2117 | int regno; | |||
2118 | } uses[FIRST_PSEUDO_REGISTER76 + MAX_RECOG_OPERANDS30]; | |||
2119 | struct reg_use_data *use; | |||
2120 | struct reg_pressure_data *reg_pressure; | |||
2121 | int delta[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
2122 | int pci, point, mix, new_last, cl, ref_pressure, queue; | |||
2123 | unsigned int i, num_uses, num_pending_births; | |||
2124 | bool print_p; | |||
2125 | ||||
2126 | /* The destinations of INSN were previously live from POINT onwards, but are | |||
2127 | now live from model_curr_point onwards. Set up DELTA accordingly. */ | |||
2128 | point = model_index (insn); | |||
2129 | reg_pressure = INSN_REG_PRESSURE (insn)((&h_i_d[INSN_UID (insn)])->reg_pressure); | |||
2130 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
2131 | { | |||
2132 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
2133 | delta[cl] = reg_pressure[pci].set_increase; | |||
2134 | } | |||
2135 | ||||
2136 | /* Record which registers previously died at POINT, but which now die | |||
2137 | before POINT. Adjust DELTA so that it represents the effect of | |||
2138 | this change after POINT - 1. Set NUM_PENDING_BIRTHS to the number of | |||
2139 | registers that will be born in the range [model_curr_point, POINT). */ | |||
2140 | num_uses = 0; | |||
2141 | num_pending_births = 0; | |||
2142 | bitmap_clear (tmp_bitmap); | |||
2143 | for (use = INSN_REG_USE_LIST (insn)((&h_i_d[INSN_UID (insn)])->reg_use_list); use != NULLnullptr; use = use->next_insn_use) | |||
2144 | { | |||
2145 | new_last = model_last_use_except (use); | |||
2146 | if (new_last < point && bitmap_set_bit (tmp_bitmap, use->regno)) | |||
2147 | { | |||
2148 | gcc_assert (num_uses < ARRAY_SIZE (uses))((void)(!(num_uses < (sizeof (uses) / sizeof ((uses)[0]))) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2148, __FUNCTION__), 0 : 0)); | |||
2149 | uses[num_uses].last_use = new_last; | |||
2150 | uses[num_uses].regno = use->regno; | |||
2151 | /* This register is no longer live after POINT - 1. */ | |||
2152 | mark_regno_birth_or_death (NULLnullptr, delta, use->regno, false); | |||
2153 | num_uses++; | |||
2154 | if (new_last >= 0) | |||
2155 | num_pending_births++; | |||
2156 | } | |||
2157 | } | |||
2158 | ||||
2159 | /* Update the MODEL_REF_PRESSURE and MODEL_MAX_PRESSURE for POINT. | |||
2160 | Also set each group pressure limit for POINT. */ | |||
2161 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
2162 | { | |||
2163 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
2164 | model_start_update_pressure (&model_before_pressure, | |||
2165 | point, pci, delta[cl]); | |||
2166 | } | |||
2167 | ||||
2168 | /* Walk the model schedule backwards, starting immediately before POINT. */ | |||
2169 | print_p = false; | |||
2170 | if (point != model_curr_point) | |||
2171 | do | |||
2172 | { | |||
2173 | point--; | |||
2174 | insn = MODEL_INSN (point)(model_schedule[point]); | |||
2175 | queue = QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index); | |||
2176 | ||||
2177 | if (queue != QUEUE_SCHEDULED(-3)) | |||
2178 | { | |||
2179 | /* DELTA describes the effect of the move on the register pressure | |||
2180 | after POINT. Make it describe the effect on the pressure | |||
2181 | before POINT. */ | |||
2182 | i = 0; | |||
2183 | while (i < num_uses) | |||
2184 | { | |||
2185 | if (uses[i].last_use == point) | |||
2186 | { | |||
2187 | /* This register is now live again. */ | |||
2188 | mark_regno_birth_or_death (NULLnullptr, delta, | |||
2189 | uses[i].regno, true); | |||
2190 | ||||
2191 | /* Remove this use from the array. */ | |||
2192 | uses[i] = uses[num_uses - 1]; | |||
2193 | num_uses--; | |||
2194 | num_pending_births--; | |||
2195 | } | |||
2196 | else | |||
2197 | i++; | |||
2198 | } | |||
2199 | ||||
2200 | if (sched_verbose >= 5) | |||
2201 | { | |||
2202 | if (!print_p) | |||
2203 | { | |||
2204 | fprintf (sched_dump, MODEL_BAR";;\t\t+------------------------------------------------------\n"); | |||
2205 | fprintf (sched_dump, ";;\t\t| New pressure for model" | |||
2206 | " schedule\n"); | |||
2207 | fprintf (sched_dump, MODEL_BAR";;\t\t+------------------------------------------------------\n"); | |||
2208 | print_p = true; | |||
2209 | } | |||
2210 | ||||
2211 | fprintf (sched_dump, ";;\t\t| %3d %4d %-30s ", | |||
2212 | point, INSN_UID (insn), | |||
2213 | str_pattern_slim (PATTERN (insn))); | |||
2214 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
2215 | { | |||
2216 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
2217 | ref_pressure = MODEL_REF_PRESSURE (&model_before_pressure,((&(&model_before_pressure)->model[(point) * (this_target_ira ->x_ira_pressure_classes_num) + (pci)])->ref_pressure) | |||
2218 | point, pci)((&(&model_before_pressure)->model[(point) * (this_target_ira ->x_ira_pressure_classes_num) + (pci)])->ref_pressure); | |||
2219 | fprintf (sched_dump, " %s:[%d->%d]", | |||
2220 | reg_class_names[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]], | |||
2221 | ref_pressure, ref_pressure + delta[cl]); | |||
2222 | } | |||
2223 | fprintf (sched_dump, "\n"); | |||
2224 | } | |||
2225 | } | |||
2226 | ||||
2227 | /* Adjust the pressure at POINT. Set MIX to nonzero if POINT - 1 | |||
2228 | might have changed as well. */ | |||
2229 | mix = num_pending_births; | |||
2230 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
2231 | { | |||
2232 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
2233 | mix |= delta[cl]; | |||
2234 | mix |= model_update_pressure (&model_before_pressure, | |||
2235 | point, pci, delta[cl]); | |||
2236 | } | |||
2237 | } | |||
2238 | while (mix && point > model_curr_point); | |||
2239 | ||||
2240 | if (print_p) | |||
2241 | fprintf (sched_dump, MODEL_BAR";;\t\t+------------------------------------------------------\n"); | |||
2242 | } | |||
2243 | ||||
2244 | /* After DEP, which was cancelled, has been resolved for insn NEXT, | |||
2245 | check whether the insn's pattern needs restoring. */ | |||
2246 | static bool | |||
2247 | must_restore_pattern_p (rtx_insn *next, dep_t dep) | |||
2248 | { | |||
2249 | if (QUEUE_INDEX (next)((&h_i_d[INSN_UID (next)])->queue_index) == QUEUE_SCHEDULED(-3)) | |||
2250 | return false; | |||
2251 | ||||
2252 | if (DEP_TYPE (dep)((dep)->type) == REG_DEP_CONTROL) | |||
2253 | { | |||
2254 | gcc_assert (ORIG_PAT (next) != NULL_RTX)((void)(!(((&h_i_d[INSN_UID (next)])->orig_pat) != (rtx ) 0) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2254, __FUNCTION__), 0 : 0)); | |||
2255 | gcc_assert (next == DEP_CON (dep))((void)(!(next == ((dep)->con)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2255, __FUNCTION__), 0 : 0)); | |||
2256 | } | |||
2257 | else | |||
2258 | { | |||
2259 | struct dep_replacement *desc = DEP_REPLACE (dep)((dep)->replace); | |||
2260 | if (desc->insn != next) | |||
2261 | { | |||
2262 | gcc_assert (*desc->loc == desc->orig)((void)(!(*desc->loc == desc->orig) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2262, __FUNCTION__), 0 : 0)); | |||
2263 | return false; | |||
2264 | } | |||
2265 | } | |||
2266 | return true; | |||
2267 | } | |||
2268 | ||||
2269 | /* model_spill_cost (CL, P, P') returns the cost of increasing the | |||
2270 | pressure on CL from P to P'. We use this to calculate a "base ECC", | |||
2271 | baseECC (CL, X), for each pressure class CL and each instruction X. | |||
2272 | Supposing X changes the pressure on CL from P to P', and that the | |||
2273 | maximum pressure on CL in the current model schedule is MP', then: | |||
2274 | ||||
2275 | * if X occurs before or at the next point of maximum pressure in | |||
2276 | the model schedule and P' > MP', then: | |||
2277 | ||||
2278 | baseECC (CL, X) = model_spill_cost (CL, MP, P') | |||
2279 | ||||
2280 | The idea is that the pressure after scheduling a fixed set of | |||
2281 | instructions -- in this case, the set up to and including the | |||
2282 | next maximum pressure point -- is going to be the same regardless | |||
2283 | of the order; we simply want to keep the intermediate pressure | |||
2284 | under control. Thus X has a cost of zero unless scheduling it | |||
2285 | now would exceed MP'. | |||
2286 | ||||
2287 | If all increases in the set are by the same amount, no zero-cost | |||
2288 | instruction will ever cause the pressure to exceed MP'. However, | |||
2289 | if X is instead moved past an instruction X' with pressure in the | |||
2290 | range (MP' - (P' - P), MP'), the pressure at X' will increase | |||
2291 | beyond MP'. Since baseECC is very much a heuristic anyway, | |||
2292 | it doesn't seem worth the overhead of tracking cases like these. | |||
2293 | ||||
2294 | The cost of exceeding MP' is always based on the original maximum | |||
2295 | pressure MP. This is so that going 2 registers over the original | |||
2296 | limit has the same cost regardless of whether it comes from two | |||
2297 | separate +1 deltas or from a single +2 delta. | |||
2298 | ||||
2299 | * if X occurs after the next point of maximum pressure in the model | |||
2300 | schedule and P' > P, then: | |||
2301 | ||||
2302 | baseECC (CL, X) = model_spill_cost (CL, MP, MP' + (P' - P)) | |||
2303 | ||||
2304 | That is, if we move X forward across a point of maximum pressure, | |||
2305 | and if X increases the pressure by P' - P, then we conservatively | |||
2306 | assume that scheduling X next would increase the maximum pressure | |||
2307 | by P' - P. Again, the cost of doing this is based on the original | |||
2308 | maximum pressure MP, for the same reason as above. | |||
2309 | ||||
2310 | * if P' < P, P > MP, and X occurs at or after the next point of | |||
2311 | maximum pressure, then: | |||
2312 | ||||
2313 | baseECC (CL, X) = -model_spill_cost (CL, MAX (MP, P'), P) | |||
2314 | ||||
2315 | That is, if we have already exceeded the original maximum pressure MP, | |||
2316 | and if X might reduce the maximum pressure again -- or at least push | |||
2317 | it further back, and thus allow more scheduling freedom -- it is given | |||
2318 | a negative cost to reflect the improvement. | |||
2319 | ||||
2320 | * otherwise, | |||
2321 | ||||
2322 | baseECC (CL, X) = 0 | |||
2323 | ||||
2324 | In this case, X is not expected to affect the maximum pressure MP', | |||
2325 | so it has zero cost. | |||
2326 | ||||
2327 | We then create a combined value baseECC (X) that is the sum of | |||
2328 | baseECC (CL, X) for each pressure class CL. | |||
2329 | ||||
2330 | baseECC (X) could itself be used as the ECC value described above. | |||
2331 | However, this is often too conservative, in the sense that it | |||
2332 | tends to make high-priority instructions that increase pressure | |||
2333 | wait too long in cases where introducing a spill would be better. | |||
2334 | For this reason the final ECC is a priority-adjusted form of | |||
2335 | baseECC (X). Specifically, we calculate: | |||
2336 | ||||
2337 | P (X) = INSN_PRIORITY (X) - insn_delay (X) - baseECC (X) | |||
2338 | baseP = MAX { P (X) | baseECC (X) <= 0 } | |||
2339 | ||||
2340 | Then: | |||
2341 | ||||
2342 | ECC (X) = MAX (MIN (baseP - P (X), baseECC (X)), 0) | |||
2343 | ||||
2344 | Thus an instruction's effect on pressure is ignored if it has a high | |||
2345 | enough priority relative to the ones that don't increase pressure. | |||
2346 | Negative values of baseECC (X) do not increase the priority of X | |||
2347 | itself, but they do make it harder for other instructions to | |||
2348 | increase the pressure further. | |||
2349 | ||||
2350 | This pressure cost is deliberately timid. The intention has been | |||
2351 | to choose a heuristic that rarely interferes with the normal list | |||
2352 | scheduler in cases where that scheduler would produce good code. | |||
2353 | We simply want to curb some of its worst excesses. */ | |||
2354 | ||||
2355 | /* Return the cost of increasing the pressure in class CL from FROM to TO. | |||
2356 | ||||
2357 | Here we use the very simplistic cost model that every register above | |||
2358 | sched_class_regs_num[CL] has a spill cost of 1. We could use other | |||
2359 | measures instead, such as one based on MEMORY_MOVE_COST. However: | |||
2360 | ||||
2361 | (1) In order for an instruction to be scheduled, the higher cost | |||
2362 | would need to be justified in a single saving of that many stalls. | |||
2363 | This is overly pessimistic, because the benefit of spilling is | |||
2364 | often to avoid a sequence of several short stalls rather than | |||
2365 | a single long one. | |||
2366 | ||||
2367 | (2) The cost is still arbitrary. Because we are not allocating | |||
2368 | registers during scheduling, we have no way of knowing for | |||
2369 | sure how many memory accesses will be required by each spill, | |||
2370 | where the spills will be placed within the block, or even | |||
2371 | which block(s) will contain the spills. | |||
2372 | ||||
2373 | So a higher cost than 1 is often too conservative in practice, | |||
2374 | forcing blocks to contain unnecessary stalls instead of spill code. | |||
2375 | The simple cost below seems to be the best compromise. It reduces | |||
2376 | the interference with the normal list scheduler, which helps make | |||
2377 | it more suitable for a default-on option. */ | |||
2378 | ||||
2379 | static int | |||
2380 | model_spill_cost (int cl, int from, int to) | |||
2381 | { | |||
2382 | from = MAX (from, sched_class_regs_num[cl])((from) > (sched_class_regs_num[cl]) ? (from) : (sched_class_regs_num [cl])); | |||
2383 | return MAX (to, from)((to) > (from) ? (to) : (from)) - from; | |||
2384 | } | |||
2385 | ||||
2386 | /* Return baseECC (ira_pressure_classes[PCI], POINT), given that | |||
2387 | P = curr_reg_pressure[ira_pressure_classes[PCI]] and that | |||
2388 | P' = P + DELTA. */ | |||
2389 | ||||
2390 | static int | |||
2391 | model_excess_group_cost (struct model_pressure_group *group, | |||
2392 | int point, int pci, int delta) | |||
2393 | { | |||
2394 | int pressure, cl; | |||
2395 | ||||
2396 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
2397 | if (delta < 0 && point >= group->limits[pci].point) | |||
2398 | { | |||
2399 | pressure = MAX (group->limits[pci].orig_pressure,((group->limits[pci].orig_pressure) > (curr_reg_pressure [cl] + delta) ? (group->limits[pci].orig_pressure) : (curr_reg_pressure [cl] + delta)) | |||
2400 | curr_reg_pressure[cl] + delta)((group->limits[pci].orig_pressure) > (curr_reg_pressure [cl] + delta) ? (group->limits[pci].orig_pressure) : (curr_reg_pressure [cl] + delta)); | |||
2401 | return -model_spill_cost (cl, pressure, curr_reg_pressure[cl]); | |||
2402 | } | |||
2403 | ||||
2404 | if (delta > 0) | |||
2405 | { | |||
2406 | if (point > group->limits[pci].point) | |||
2407 | pressure = group->limits[pci].pressure + delta; | |||
2408 | else | |||
2409 | pressure = curr_reg_pressure[cl] + delta; | |||
2410 | ||||
2411 | if (pressure > group->limits[pci].pressure) | |||
2412 | return model_spill_cost (cl, group->limits[pci].orig_pressure, | |||
2413 | pressure); | |||
2414 | } | |||
2415 | ||||
2416 | return 0; | |||
2417 | } | |||
2418 | ||||
2419 | /* Return baseECC (MODEL_INSN (INSN)). Dump the costs to sched_dump | |||
2420 | if PRINT_P. */ | |||
2421 | ||||
2422 | static int | |||
2423 | model_excess_cost (rtx_insn *insn, bool print_p) | |||
2424 | { | |||
2425 | int point, pci, cl, cost, this_cost, delta; | |||
2426 | struct reg_pressure_data *insn_reg_pressure; | |||
2427 | int insn_death[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
2428 | ||||
2429 | calculate_reg_deaths (insn, insn_death); | |||
2430 | point = model_index (insn); | |||
2431 | insn_reg_pressure = INSN_REG_PRESSURE (insn)((&h_i_d[INSN_UID (insn)])->reg_pressure); | |||
2432 | cost = 0; | |||
2433 | ||||
2434 | if (print_p) | |||
2435 | fprintf (sched_dump, ";;\t\t| %3d %4d | %4d %+3d |", point, | |||
2436 | INSN_UID (insn), INSN_PRIORITY (insn)((&h_i_d[INSN_UID (insn)])->priority), insn_delay (insn)); | |||
2437 | ||||
2438 | /* Sum up the individual costs for each register class. */ | |||
2439 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
2440 | { | |||
2441 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
2442 | delta = insn_reg_pressure[pci].set_increase - insn_death[cl]; | |||
2443 | this_cost = model_excess_group_cost (&model_before_pressure, | |||
2444 | point, pci, delta); | |||
2445 | cost += this_cost; | |||
2446 | if (print_p) | |||
2447 | fprintf (sched_dump, " %s:[%d base cost %d]", | |||
2448 | reg_class_names[cl], delta, this_cost); | |||
2449 | } | |||
2450 | ||||
2451 | if (print_p) | |||
2452 | fprintf (sched_dump, "\n"); | |||
2453 | ||||
2454 | return cost; | |||
2455 | } | |||
2456 | ||||
2457 | /* Dump the next points of maximum pressure for GROUP. */ | |||
2458 | ||||
2459 | static void | |||
2460 | model_dump_pressure_points (struct model_pressure_group *group) | |||
2461 | { | |||
2462 | int pci, cl; | |||
2463 | ||||
2464 | fprintf (sched_dump, ";;\t\t| pressure points"); | |||
2465 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
2466 | { | |||
2467 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
2468 | fprintf (sched_dump, " %s:[%d->%d at ", reg_class_names[cl], | |||
2469 | curr_reg_pressure[cl], group->limits[pci].pressure); | |||
2470 | if (group->limits[pci].point < model_num_insns) | |||
2471 | fprintf (sched_dump, "%d:%d]", group->limits[pci].point, | |||
2472 | INSN_UID (MODEL_INSN (group->limits[pci].point)(model_schedule[group->limits[pci].point]))); | |||
2473 | else | |||
2474 | fprintf (sched_dump, "end]"); | |||
2475 | } | |||
2476 | fprintf (sched_dump, "\n"); | |||
2477 | } | |||
2478 | ||||
2479 | /* Set INSN_REG_PRESSURE_EXCESS_COST_CHANGE for INSNS[0...COUNT-1]. */ | |||
2480 | ||||
2481 | static void | |||
2482 | model_set_excess_costs (rtx_insn **insns, int count) | |||
2483 | { | |||
2484 | int i, cost, priority_base, priority; | |||
2485 | bool print_p; | |||
2486 | ||||
2487 | /* Record the baseECC value for each instruction in the model schedule, | |||
2488 | except that negative costs are converted to zero ones now rather than | |||
2489 | later. Do not assign a cost to debug instructions, since they must | |||
2490 | not change code-generation decisions. Experiments suggest we also | |||
2491 | get better results by not assigning a cost to instructions from | |||
2492 | a different block. | |||
2493 | ||||
2494 | Set PRIORITY_BASE to baseP in the block comment above. This is the | |||
2495 | maximum priority of the "cheap" instructions, which should always | |||
2496 | include the next model instruction. */ | |||
2497 | priority_base = 0; | |||
2498 | print_p = false; | |||
2499 | for (i = 0; i < count; i++) | |||
2500 | if (INSN_MODEL_INDEX (insns[i])((&h_i_d[INSN_UID (insns[i])])->model_index)) | |||
2501 | { | |||
2502 | if (sched_verbose >= 6 && !print_p) | |||
2503 | { | |||
2504 | fprintf (sched_dump, MODEL_BAR";;\t\t+------------------------------------------------------\n"); | |||
2505 | fprintf (sched_dump, ";;\t\t| Pressure costs for ready queue\n"); | |||
2506 | model_dump_pressure_points (&model_before_pressure); | |||
2507 | fprintf (sched_dump, MODEL_BAR";;\t\t+------------------------------------------------------\n"); | |||
2508 | print_p = true; | |||
2509 | } | |||
2510 | cost = model_excess_cost (insns[i], print_p); | |||
2511 | if (cost <= 0) | |||
2512 | { | |||
2513 | priority = INSN_PRIORITY (insns[i])((&h_i_d[INSN_UID (insns[i])])->priority) - insn_delay (insns[i]) - cost; | |||
2514 | priority_base = MAX (priority_base, priority)((priority_base) > (priority) ? (priority_base) : (priority )); | |||
2515 | cost = 0; | |||
2516 | } | |||
2517 | INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i])((&h_i_d[INSN_UID (insns[i])])->reg_pressure_excess_cost_change ) = cost; | |||
2518 | } | |||
2519 | if (print_p) | |||
2520 | fprintf (sched_dump, MODEL_BAR";;\t\t+------------------------------------------------------\n"); | |||
2521 | ||||
2522 | /* Use MAX (baseECC, 0) and baseP to calculcate ECC for each | |||
2523 | instruction. */ | |||
2524 | for (i = 0; i < count; i++) | |||
2525 | { | |||
2526 | cost = INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i])((&h_i_d[INSN_UID (insns[i])])->reg_pressure_excess_cost_change ); | |||
2527 | priority = INSN_PRIORITY (insns[i])((&h_i_d[INSN_UID (insns[i])])->priority) - insn_delay (insns[i]); | |||
2528 | if (cost > 0 && priority > priority_base) | |||
2529 | { | |||
2530 | cost += priority_base - priority; | |||
2531 | INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i])((&h_i_d[INSN_UID (insns[i])])->reg_pressure_excess_cost_change ) = MAX (cost, 0)((cost) > (0) ? (cost) : (0)); | |||
2532 | } | |||
2533 | } | |||
2534 | } | |||
2535 | ||||
2536 | ||||
2537 | /* Enum of rank_for_schedule heuristic decisions. */ | |||
2538 | enum rfs_decision { | |||
2539 | RFS_LIVE_RANGE_SHRINK1, RFS_LIVE_RANGE_SHRINK2, | |||
2540 | RFS_SCHED_GROUP, RFS_PRESSURE_DELAY, RFS_PRESSURE_TICK, | |||
2541 | RFS_FEEDS_BACKTRACK_INSN, RFS_PRIORITY, RFS_AUTOPREF, RFS_SPECULATION, | |||
2542 | RFS_SCHED_RANK, RFS_LAST_INSN, RFS_PRESSURE_INDEX, | |||
2543 | RFS_DEP_COUNT, RFS_TIE, RFS_FUSION, RFS_COST, RFS_N }; | |||
2544 | ||||
2545 | /* Corresponding strings for print outs. */ | |||
2546 | static const char *rfs_str[RFS_N] = { | |||
2547 | "RFS_LIVE_RANGE_SHRINK1", "RFS_LIVE_RANGE_SHRINK2", | |||
2548 | "RFS_SCHED_GROUP", "RFS_PRESSURE_DELAY", "RFS_PRESSURE_TICK", | |||
2549 | "RFS_FEEDS_BACKTRACK_INSN", "RFS_PRIORITY", "RFS_AUTOPREF", "RFS_SPECULATION", | |||
2550 | "RFS_SCHED_RANK", "RFS_LAST_INSN", "RFS_PRESSURE_INDEX", | |||
2551 | "RFS_DEP_COUNT", "RFS_TIE", "RFS_FUSION", "RFS_COST" }; | |||
2552 | ||||
2553 | /* Statistical breakdown of rank_for_schedule decisions. */ | |||
2554 | struct rank_for_schedule_stats_t { unsigned stats[RFS_N]; }; | |||
2555 | static rank_for_schedule_stats_t rank_for_schedule_stats; | |||
2556 | ||||
2557 | /* Return the result of comparing insns TMP and TMP2 and update | |||
2558 | Rank_For_Schedule statistics. */ | |||
2559 | static int | |||
2560 | rfs_result (enum rfs_decision decision, int result, rtx tmp, rtx tmp2) | |||
2561 | { | |||
2562 | ++rank_for_schedule_stats.stats[decision]; | |||
2563 | if (result < 0) | |||
2564 | INSN_LAST_RFS_WIN (tmp)((&h_i_d[INSN_UID (tmp)])->last_rfs_win) = decision; | |||
2565 | else if (result > 0) | |||
2566 | INSN_LAST_RFS_WIN (tmp2)((&h_i_d[INSN_UID (tmp2)])->last_rfs_win) = decision; | |||
2567 | else | |||
2568 | gcc_unreachable ()(fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2568, __FUNCTION__)); | |||
2569 | return result; | |||
2570 | } | |||
2571 | ||||
2572 | /* Sorting predicate to move DEBUG_INSNs to the top of ready list, while | |||
2573 | keeping normal insns in original order. */ | |||
2574 | ||||
2575 | static int | |||
2576 | rank_for_schedule_debug (const void *x, const void *y) | |||
2577 | { | |||
2578 | rtx_insn *tmp = *(rtx_insn * const *) y; | |||
2579 | rtx_insn *tmp2 = *(rtx_insn * const *) x; | |||
2580 | ||||
2581 | /* Schedule debug insns as early as possible. */ | |||
2582 | if (DEBUG_INSN_P (tmp)(((enum rtx_code) (tmp)->code) == DEBUG_INSN) && !DEBUG_INSN_P (tmp2)(((enum rtx_code) (tmp2)->code) == DEBUG_INSN)) | |||
2583 | return -1; | |||
2584 | else if (!DEBUG_INSN_P (tmp)(((enum rtx_code) (tmp)->code) == DEBUG_INSN) && DEBUG_INSN_P (tmp2)(((enum rtx_code) (tmp2)->code) == DEBUG_INSN)) | |||
2585 | return 1; | |||
2586 | else if (DEBUG_INSN_P (tmp)(((enum rtx_code) (tmp)->code) == DEBUG_INSN) && DEBUG_INSN_P (tmp2)(((enum rtx_code) (tmp2)->code) == DEBUG_INSN)) | |||
2587 | return INSN_LUID (tmp)(sched_luids[INSN_UID (tmp)]) - INSN_LUID (tmp2)(sched_luids[INSN_UID (tmp2)]); | |||
2588 | else | |||
2589 | return INSN_RFS_DEBUG_ORIG_ORDER (tmp2)((&h_i_d[INSN_UID (tmp2)])->rfs_debug_orig_order) - INSN_RFS_DEBUG_ORIG_ORDER (tmp)((&h_i_d[INSN_UID (tmp)])->rfs_debug_orig_order); | |||
2590 | } | |||
2591 | ||||
2592 | /* Returns a positive value if x is preferred; returns a negative value if | |||
2593 | y is preferred. Should never return 0, since that will make the sort | |||
2594 | unstable. */ | |||
2595 | ||||
2596 | static int | |||
2597 | rank_for_schedule (const void *x, const void *y) | |||
2598 | { | |||
2599 | rtx_insn *tmp = *(rtx_insn * const *) y; | |||
2600 | rtx_insn *tmp2 = *(rtx_insn * const *) x; | |||
2601 | int tmp_class, tmp2_class; | |||
2602 | int val, priority_val, info_val, diff; | |||
2603 | ||||
2604 | if (live_range_shrinkage_p) | |||
2605 | { | |||
2606 | /* Don't use SCHED_PRESSURE_MODEL -- it results in much worse | |||
2607 | code. */ | |||
2608 | gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED)((void)(!(sched_pressure == SCHED_PRESSURE_WEIGHTED) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2608, __FUNCTION__), 0 : 0)); | |||
2609 | if ((INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)((&h_i_d[INSN_UID (tmp)])->reg_pressure_excess_cost_change ) < 0 | |||
2610 | || INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2)((&h_i_d[INSN_UID (tmp2)])->reg_pressure_excess_cost_change ) < 0) | |||
2611 | && (diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)((&h_i_d[INSN_UID (tmp)])->reg_pressure_excess_cost_change ) | |||
2612 | - INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2)((&h_i_d[INSN_UID (tmp2)])->reg_pressure_excess_cost_change ))) != 0) | |||
2613 | return rfs_result (RFS_LIVE_RANGE_SHRINK1, diff, tmp, tmp2); | |||
2614 | /* Sort by INSN_LUID (original insn order), so that we make the | |||
2615 | sort stable. This minimizes instruction movement, thus | |||
2616 | minimizing sched's effect on debugging and cross-jumping. */ | |||
2617 | return rfs_result (RFS_LIVE_RANGE_SHRINK2, | |||
2618 | INSN_LUID (tmp)(sched_luids[INSN_UID (tmp)]) - INSN_LUID (tmp2)(sched_luids[INSN_UID (tmp2)]), tmp, tmp2); | |||
2619 | } | |||
2620 | ||||
2621 | /* The insn in a schedule group should be issued the first. */ | |||
2622 | if (flag_sched_group_heuristicglobal_options.x_flag_sched_group_heuristic && | |||
2623 | SCHED_GROUP_P (tmp)(__extension__ ({ __typeof ((tmp)) const _rtx = ((tmp)); if ( ((enum rtx_code) (_rtx)->code) != DEBUG_INSN && (( enum rtx_code) (_rtx)->code) != INSN && ((enum rtx_code ) (_rtx)->code) != JUMP_INSN && ((enum rtx_code) ( _rtx)->code) != CALL_INSN) rtl_check_failed_flag ("SCHED_GROUP_P" , _rtx, "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2623, __FUNCTION__); _rtx; })->in_struct) != SCHED_GROUP_P (tmp2)(__extension__ ({ __typeof ((tmp2)) const _rtx = ((tmp2)); if (((enum rtx_code) (_rtx)->code) != DEBUG_INSN && ( (enum rtx_code) (_rtx)->code) != INSN && ((enum rtx_code ) (_rtx)->code) != JUMP_INSN && ((enum rtx_code) ( _rtx)->code) != CALL_INSN) rtl_check_failed_flag ("SCHED_GROUP_P" , _rtx, "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2623, __FUNCTION__); _rtx; })->in_struct)) | |||
2624 | return rfs_result (RFS_SCHED_GROUP, SCHED_GROUP_P (tmp2)(__extension__ ({ __typeof ((tmp2)) const _rtx = ((tmp2)); if (((enum rtx_code) (_rtx)->code) != DEBUG_INSN && ( (enum rtx_code) (_rtx)->code) != INSN && ((enum rtx_code ) (_rtx)->code) != JUMP_INSN && ((enum rtx_code) ( _rtx)->code) != CALL_INSN) rtl_check_failed_flag ("SCHED_GROUP_P" , _rtx, "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2624, __FUNCTION__); _rtx; })->in_struct) ? 1 : -1, | |||
2625 | tmp, tmp2); | |||
2626 | ||||
2627 | /* Make sure that priority of TMP and TMP2 are initialized. */ | |||
2628 | gcc_assert (INSN_PRIORITY_KNOWN (tmp) && INSN_PRIORITY_KNOWN (tmp2))((void)(!((((&h_i_d[INSN_UID (tmp)])->priority_status) > 0) && (((&h_i_d[INSN_UID (tmp2)])->priority_status ) > 0)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2628, __FUNCTION__), 0 : 0)); | |||
2629 | ||||
2630 | if (sched_fusion) | |||
2631 | { | |||
2632 | /* The instruction that has the same fusion priority as the last | |||
2633 | instruction is the instruction we picked next. If that is not | |||
2634 | the case, we sort ready list firstly by fusion priority, then | |||
2635 | by priority, and at last by INSN_LUID. */ | |||
2636 | int a = INSN_FUSION_PRIORITY (tmp)((&h_i_d[INSN_UID (tmp)])->fusion_priority); | |||
2637 | int b = INSN_FUSION_PRIORITY (tmp2)((&h_i_d[INSN_UID (tmp2)])->fusion_priority); | |||
2638 | int last = -1; | |||
2639 | ||||
2640 | if (last_nondebug_scheduled_insn | |||
2641 | && !NOTE_P (last_nondebug_scheduled_insn)(((enum rtx_code) (last_nondebug_scheduled_insn)->code) == NOTE) | |||
2642 | && BLOCK_FOR_INSN (tmp) | |||
2643 | == BLOCK_FOR_INSN (last_nondebug_scheduled_insn)) | |||
2644 | last = INSN_FUSION_PRIORITY (last_nondebug_scheduled_insn)((&h_i_d[INSN_UID (last_nondebug_scheduled_insn)])->fusion_priority ); | |||
2645 | ||||
2646 | if (a != last && b != last) | |||
2647 | { | |||
2648 | if (a == b) | |||
2649 | { | |||
2650 | a = INSN_PRIORITY (tmp)((&h_i_d[INSN_UID (tmp)])->priority); | |||
2651 | b = INSN_PRIORITY (tmp2)((&h_i_d[INSN_UID (tmp2)])->priority); | |||
2652 | } | |||
2653 | if (a != b) | |||
2654 | return rfs_result (RFS_FUSION, b - a, tmp, tmp2); | |||
2655 | else | |||
2656 | return rfs_result (RFS_FUSION, | |||
2657 | INSN_LUID (tmp)(sched_luids[INSN_UID (tmp)]) - INSN_LUID (tmp2)(sched_luids[INSN_UID (tmp2)]), tmp, tmp2); | |||
2658 | } | |||
2659 | else if (a == b) | |||
2660 | { | |||
2661 | gcc_assert (last_nondebug_scheduled_insn((void)(!(last_nondebug_scheduled_insn && !(((enum rtx_code ) (last_nondebug_scheduled_insn)->code) == NOTE)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2662, __FUNCTION__), 0 : 0)) | |||
2662 | && !NOTE_P (last_nondebug_scheduled_insn))((void)(!(last_nondebug_scheduled_insn && !(((enum rtx_code ) (last_nondebug_scheduled_insn)->code) == NOTE)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2662, __FUNCTION__), 0 : 0)); | |||
2663 | last = INSN_PRIORITY (last_nondebug_scheduled_insn)((&h_i_d[INSN_UID (last_nondebug_scheduled_insn)])->priority ); | |||
2664 | ||||
2665 | a = abs (INSN_PRIORITY (tmp)((&h_i_d[INSN_UID (tmp)])->priority) - last); | |||
2666 | b = abs (INSN_PRIORITY (tmp2)((&h_i_d[INSN_UID (tmp2)])->priority) - last); | |||
2667 | if (a != b) | |||
2668 | return rfs_result (RFS_FUSION, a - b, tmp, tmp2); | |||
2669 | else | |||
2670 | return rfs_result (RFS_FUSION, | |||
2671 | INSN_LUID (tmp)(sched_luids[INSN_UID (tmp)]) - INSN_LUID (tmp2)(sched_luids[INSN_UID (tmp2)]), tmp, tmp2); | |||
2672 | } | |||
2673 | else if (a == last) | |||
2674 | return rfs_result (RFS_FUSION, -1, tmp, tmp2); | |||
2675 | else | |||
2676 | return rfs_result (RFS_FUSION, 1, tmp, tmp2); | |||
2677 | } | |||
2678 | ||||
2679 | if (sched_pressure != SCHED_PRESSURE_NONE) | |||
2680 | { | |||
2681 | /* Prefer insn whose scheduling results in the smallest register | |||
2682 | pressure excess. */ | |||
2683 | if ((diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)((&h_i_d[INSN_UID (tmp)])->reg_pressure_excess_cost_change ) | |||
2684 | + insn_delay (tmp) | |||
2685 | - INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2)((&h_i_d[INSN_UID (tmp2)])->reg_pressure_excess_cost_change ) | |||
2686 | - insn_delay (tmp2)))) | |||
2687 | return rfs_result (RFS_PRESSURE_DELAY, diff, tmp, tmp2); | |||
2688 | } | |||
2689 | ||||
2690 | if (sched_pressure != SCHED_PRESSURE_NONE | |||
2691 | && (INSN_TICK (tmp2)((&h_i_d[INSN_UID (tmp2)])->tick) > clock_var || INSN_TICK (tmp)((&h_i_d[INSN_UID (tmp)])->tick) > clock_var) | |||
2692 | && INSN_TICK (tmp2)((&h_i_d[INSN_UID (tmp2)])->tick) != INSN_TICK (tmp)((&h_i_d[INSN_UID (tmp)])->tick)) | |||
2693 | { | |||
2694 | diff = INSN_TICK (tmp)((&h_i_d[INSN_UID (tmp)])->tick) - INSN_TICK (tmp2)((&h_i_d[INSN_UID (tmp2)])->tick); | |||
2695 | return rfs_result (RFS_PRESSURE_TICK, diff, tmp, tmp2); | |||
2696 | } | |||
2697 | ||||
2698 | /* If we are doing backtracking in this schedule, prefer insns that | |||
2699 | have forward dependencies with negative cost against an insn that | |||
2700 | was already scheduled. */ | |||
2701 | if (current_sched_info->flags & DO_BACKTRACKING) | |||
2702 | { | |||
2703 | priority_val = FEEDS_BACKTRACK_INSN (tmp2)((&h_i_d[INSN_UID (tmp2)])->feeds_backtrack_insn) - FEEDS_BACKTRACK_INSN (tmp)((&h_i_d[INSN_UID (tmp)])->feeds_backtrack_insn); | |||
2704 | if (priority_val) | |||
2705 | return rfs_result (RFS_FEEDS_BACKTRACK_INSN, priority_val, tmp, tmp2); | |||
2706 | } | |||
2707 | ||||
2708 | /* Prefer insn with higher priority. */ | |||
2709 | priority_val = INSN_PRIORITY (tmp2)((&h_i_d[INSN_UID (tmp2)])->priority) - INSN_PRIORITY (tmp)((&h_i_d[INSN_UID (tmp)])->priority); | |||
2710 | ||||
2711 | if (flag_sched_critical_path_heuristicglobal_options.x_flag_sched_critical_path_heuristic && priority_val) | |||
2712 | return rfs_result (RFS_PRIORITY, priority_val, tmp, tmp2); | |||
2713 | ||||
2714 | if (param_sched_autopref_queue_depthglobal_options.x_param_sched_autopref_queue_depth >= 0) | |||
2715 | { | |||
2716 | int autopref = autopref_rank_for_schedule (tmp, tmp2); | |||
2717 | if (autopref != 0) | |||
2718 | return rfs_result (RFS_AUTOPREF, autopref, tmp, tmp2); | |||
2719 | } | |||
2720 | ||||
2721 | /* Prefer speculative insn with greater dependencies weakness. */ | |||
2722 | if (flag_sched_spec_insn_heuristicglobal_options.x_flag_sched_spec_insn_heuristic && spec_info) | |||
2723 | { | |||
2724 | ds_t ds1, ds2; | |||
2725 | dw_t dw1, dw2; | |||
2726 | int dw; | |||
2727 | ||||
2728 | ds1 = TODO_SPEC (tmp)((&h_i_d[INSN_UID (tmp)])->todo_spec) & SPECULATIVE(((((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BEGIN_DATA_BITS_OFFSET ) | (((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BE_IN_DATA_BITS_OFFSET)) | ((((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BEGIN_CONTROL_BITS_OFFSET) | (((ds_t ) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BE_IN_CONTROL_BITS_OFFSET ))); | |||
2729 | if (ds1) | |||
2730 | dw1 = ds_weak (ds1); | |||
2731 | else | |||
2732 | dw1 = NO_DEP_WEAK((((1 << (((8 * 4) - 8) / 4)) - 1) - 1) + 1); | |||
2733 | ||||
2734 | ds2 = TODO_SPEC (tmp2)((&h_i_d[INSN_UID (tmp2)])->todo_spec) & SPECULATIVE(((((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BEGIN_DATA_BITS_OFFSET ) | (((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BE_IN_DATA_BITS_OFFSET)) | ((((ds_t) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BEGIN_CONTROL_BITS_OFFSET) | (((ds_t ) ((1 << (((8 * 4) - 8) / 4)) - 1)) << BE_IN_CONTROL_BITS_OFFSET ))); | |||
2735 | if (ds2) | |||
2736 | dw2 = ds_weak (ds2); | |||
2737 | else | |||
2738 | dw2 = NO_DEP_WEAK((((1 << (((8 * 4) - 8) / 4)) - 1) - 1) + 1); | |||
2739 | ||||
2740 | dw = dw2 - dw1; | |||
2741 | if (dw > (NO_DEP_WEAK((((1 << (((8 * 4) - 8) / 4)) - 1) - 1) + 1) / 8) || dw < -(NO_DEP_WEAK((((1 << (((8 * 4) - 8) / 4)) - 1) - 1) + 1) / 8)) | |||
2742 | return rfs_result (RFS_SPECULATION, dw, tmp, tmp2); | |||
2743 | } | |||
2744 | ||||
2745 | info_val = (*current_sched_info->rank) (tmp, tmp2); | |||
2746 | if (flag_sched_rank_heuristicglobal_options.x_flag_sched_rank_heuristic && info_val) | |||
2747 | return rfs_result (RFS_SCHED_RANK, info_val, tmp, tmp2); | |||
2748 | ||||
2749 | /* Compare insns based on their relation to the last scheduled | |||
2750 | non-debug insn. */ | |||
2751 | if (flag_sched_last_insn_heuristicglobal_options.x_flag_sched_last_insn_heuristic && last_nondebug_scheduled_insn) | |||
2752 | { | |||
2753 | dep_t dep1; | |||
2754 | dep_t dep2; | |||
2755 | rtx_insn *last = last_nondebug_scheduled_insn; | |||
2756 | ||||
2757 | /* Classify the instructions into three classes: | |||
2758 | 1) Data dependent on last schedule insn. | |||
2759 | 2) Anti/Output dependent on last scheduled insn. | |||
2760 | 3) Independent of last scheduled insn, or has latency of one. | |||
2761 | Choose the insn from the highest numbered class if different. */ | |||
2762 | dep1 = sd_find_dep_between (last, tmp, true); | |||
2763 | ||||
2764 | if (dep1 == NULLnullptr || dep_cost (dep1) == 1) | |||
2765 | tmp_class = 3; | |||
2766 | else if (/* Data dependence. */ | |||
2767 | DEP_TYPE (dep1)((dep1)->type) == REG_DEP_TRUE) | |||
2768 | tmp_class = 1; | |||
2769 | else | |||
2770 | tmp_class = 2; | |||
2771 | ||||
2772 | dep2 = sd_find_dep_between (last, tmp2, true); | |||
2773 | ||||
2774 | if (dep2 == NULLnullptr || dep_cost (dep2) == 1) | |||
2775 | tmp2_class = 3; | |||
2776 | else if (/* Data dependence. */ | |||
2777 | DEP_TYPE (dep2)((dep2)->type) == REG_DEP_TRUE) | |||
2778 | tmp2_class = 1; | |||
2779 | else | |||
2780 | tmp2_class = 2; | |||
2781 | ||||
2782 | if ((val = tmp2_class - tmp_class)) | |||
2783 | return rfs_result (RFS_LAST_INSN, val, tmp, tmp2); | |||
2784 | } | |||
2785 | ||||
2786 | /* Prefer instructions that occur earlier in the model schedule. */ | |||
2787 | if (sched_pressure == SCHED_PRESSURE_MODEL) | |||
2788 | { | |||
2789 | diff = model_index (tmp) - model_index (tmp2); | |||
2790 | if (diff != 0) | |||
2791 | return rfs_result (RFS_PRESSURE_INDEX, diff, tmp, tmp2); | |||
2792 | } | |||
2793 | ||||
2794 | /* Prefer the insn which has more later insns that depend on it. | |||
2795 | This gives the scheduler more freedom when scheduling later | |||
2796 | instructions at the expense of added register pressure. */ | |||
2797 | ||||
2798 | val = (dep_list_size (tmp2, SD_LIST_FORW(4)) | |||
2799 | - dep_list_size (tmp, SD_LIST_FORW(4))); | |||
2800 | ||||
2801 | if (flag_sched_dep_count_heuristicglobal_options.x_flag_sched_dep_count_heuristic && val != 0) | |||
2802 | return rfs_result (RFS_DEP_COUNT, val, tmp, tmp2); | |||
2803 | ||||
2804 | /* Sort by INSN_COST rather than INSN_LUID. This means that instructions | |||
2805 | which take longer to execute are prioritised and it leads to more | |||
2806 | dual-issue opportunities on in-order cores which have this feature. */ | |||
2807 | ||||
2808 | if (INSN_COST (tmp)((&h_i_d[INSN_UID (tmp)])->cost) != INSN_COST (tmp2)((&h_i_d[INSN_UID (tmp2)])->cost)) | |||
2809 | return rfs_result (RFS_COST, INSN_COST (tmp2)((&h_i_d[INSN_UID (tmp2)])->cost) - INSN_COST (tmp)((&h_i_d[INSN_UID (tmp)])->cost), | |||
2810 | tmp, tmp2); | |||
2811 | ||||
2812 | /* If insns are equally good, sort by INSN_LUID (original insn order), | |||
2813 | so that we make the sort stable. This minimizes instruction movement, | |||
2814 | thus minimizing sched's effect on debugging and cross-jumping. */ | |||
2815 | return rfs_result (RFS_TIE, INSN_LUID (tmp)(sched_luids[INSN_UID (tmp)]) - INSN_LUID (tmp2)(sched_luids[INSN_UID (tmp2)]), tmp, tmp2); | |||
2816 | } | |||
2817 | ||||
2818 | /* Resort the array A in which only element at index N may be out of order. */ | |||
2819 | ||||
2820 | HAIFA_INLINE__inline static void | |||
2821 | swap_sort (rtx_insn **a, int n) | |||
2822 | { | |||
2823 | rtx_insn *insn = a[n - 1]; | |||
2824 | int i = n - 2; | |||
2825 | ||||
2826 | while (i >= 0 && rank_for_schedule (a + i, &insn) >= 0) | |||
2827 | { | |||
2828 | a[i + 1] = a[i]; | |||
2829 | i -= 1; | |||
2830 | } | |||
2831 | a[i + 1] = insn; | |||
2832 | } | |||
2833 | ||||
2834 | /* Add INSN to the insn queue so that it can be executed at least | |||
2835 | N_CYCLES after the currently executing insn. Preserve insns | |||
2836 | chain for debugging purposes. REASON will be printed in debugging | |||
2837 | output. */ | |||
2838 | ||||
2839 | HAIFA_INLINE__inline static void | |||
2840 | queue_insn (rtx_insn *insn, int n_cycles, const char *reason) | |||
2841 | { | |||
2842 | int next_q = NEXT_Q_AFTER (q_ptr, n_cycles)(((q_ptr)+n_cycles) & max_insn_queue_index); | |||
2843 | rtx_insn_list *link = alloc_INSN_LIST (insn, insn_queue[next_q]); | |||
2844 | int new_tick; | |||
2845 | ||||
2846 | gcc_assert (n_cycles <= max_insn_queue_index)((void)(!(n_cycles <= max_insn_queue_index) ? fancy_abort ( "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2846, __FUNCTION__), 0 : 0)); | |||
2847 | gcc_assert (!DEBUG_INSN_P (insn))((void)(!(!(((enum rtx_code) (insn)->code) == DEBUG_INSN)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2847, __FUNCTION__), 0 : 0)); | |||
2848 | ||||
2849 | insn_queue[next_q] = link; | |||
2850 | q_size += 1; | |||
2851 | ||||
2852 | if (sched_verbose >= 2) | |||
2853 | { | |||
2854 | fprintf (sched_dump, ";;\t\tReady-->Q: insn %s: ", | |||
2855 | (*current_sched_info->print_insn) (insn, 0)); | |||
2856 | ||||
2857 | fprintf (sched_dump, "queued for %d cycles (%s).\n", n_cycles, reason); | |||
2858 | } | |||
2859 | ||||
2860 | QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index) = next_q; | |||
2861 | ||||
2862 | if (current_sched_info->flags & DO_BACKTRACKING) | |||
2863 | { | |||
2864 | new_tick = clock_var + n_cycles; | |||
2865 | if (INSN_TICK (insn)((&h_i_d[INSN_UID (insn)])->tick) == INVALID_TICK(-(max_insn_queue_index + 1)) || INSN_TICK (insn)((&h_i_d[INSN_UID (insn)])->tick) < new_tick) | |||
2866 | INSN_TICK (insn)((&h_i_d[INSN_UID (insn)])->tick) = new_tick; | |||
2867 | ||||
2868 | if (INSN_EXACT_TICK (insn)((&h_i_d[INSN_UID (insn)])->exact_tick) != INVALID_TICK(-(max_insn_queue_index + 1)) | |||
2869 | && INSN_EXACT_TICK (insn)((&h_i_d[INSN_UID (insn)])->exact_tick) < clock_var + n_cycles) | |||
2870 | { | |||
2871 | must_backtrack = true; | |||
2872 | if (sched_verbose >= 2) | |||
2873 | fprintf (sched_dump, ";;\t\tcausing a backtrack.\n"); | |||
2874 | } | |||
2875 | } | |||
2876 | } | |||
2877 | ||||
2878 | /* Remove INSN from queue. */ | |||
2879 | static void | |||
2880 | queue_remove (rtx_insn *insn) | |||
2881 | { | |||
2882 | gcc_assert (QUEUE_INDEX (insn) >= 0)((void)(!(((&h_i_d[INSN_UID (insn)])->queue_index) >= 0) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2882, __FUNCTION__), 0 : 0)); | |||
2883 | remove_free_INSN_LIST_elem (insn, &insn_queue[QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index)]); | |||
2884 | q_size--; | |||
2885 | QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index) = QUEUE_NOWHERE(-2); | |||
2886 | } | |||
2887 | ||||
2888 | /* Return a pointer to the bottom of the ready list, i.e. the insn | |||
2889 | with the lowest priority. */ | |||
2890 | ||||
2891 | rtx_insn ** | |||
2892 | ready_lastpos (struct ready_list *ready) | |||
2893 | { | |||
2894 | gcc_assert (ready->n_ready >= 1)((void)(!(ready->n_ready >= 1) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2894, __FUNCTION__), 0 : 0)); | |||
2895 | return ready->vec + ready->first - ready->n_ready + 1; | |||
2896 | } | |||
2897 | ||||
2898 | /* Add an element INSN to the ready list so that it ends up with the | |||
2899 | lowest/highest priority depending on FIRST_P. */ | |||
2900 | ||||
2901 | HAIFA_INLINE__inline static void | |||
2902 | ready_add (struct ready_list *ready, rtx_insn *insn, bool first_p) | |||
2903 | { | |||
2904 | if (!first_p) | |||
2905 | { | |||
2906 | if (ready->first == ready->n_ready) | |||
2907 | { | |||
2908 | memmove (ready->vec + ready->veclen - ready->n_ready, | |||
2909 | ready_lastpos (ready), | |||
2910 | ready->n_ready * sizeof (rtx)); | |||
2911 | ready->first = ready->veclen - 1; | |||
2912 | } | |||
2913 | ready->vec[ready->first - ready->n_ready] = insn; | |||
2914 | } | |||
2915 | else | |||
2916 | { | |||
2917 | if (ready->first == ready->veclen - 1) | |||
2918 | { | |||
2919 | if (ready->n_ready) | |||
2920 | /* ready_lastpos() fails when called with (ready->n_ready == 0). */ | |||
2921 | memmove (ready->vec + ready->veclen - ready->n_ready - 1, | |||
2922 | ready_lastpos (ready), | |||
2923 | ready->n_ready * sizeof (rtx)); | |||
2924 | ready->first = ready->veclen - 2; | |||
2925 | } | |||
2926 | ready->vec[++(ready->first)] = insn; | |||
2927 | } | |||
2928 | ||||
2929 | ready->n_ready++; | |||
2930 | if (DEBUG_INSN_P (insn)(((enum rtx_code) (insn)->code) == DEBUG_INSN)) | |||
2931 | ready->n_debug++; | |||
2932 | ||||
2933 | gcc_assert (QUEUE_INDEX (insn) != QUEUE_READY)((void)(!(((&h_i_d[INSN_UID (insn)])->queue_index) != ( -1)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2933, __FUNCTION__), 0 : 0)); | |||
2934 | QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index) = QUEUE_READY(-1); | |||
2935 | ||||
2936 | if (INSN_EXACT_TICK (insn)((&h_i_d[INSN_UID (insn)])->exact_tick) != INVALID_TICK(-(max_insn_queue_index + 1)) | |||
2937 | && INSN_EXACT_TICK (insn)((&h_i_d[INSN_UID (insn)])->exact_tick) < clock_var) | |||
2938 | { | |||
2939 | must_backtrack = true; | |||
2940 | } | |||
2941 | } | |||
2942 | ||||
2943 | /* Remove the element with the highest priority from the ready list and | |||
2944 | return it. */ | |||
2945 | ||||
2946 | HAIFA_INLINE__inline static rtx_insn * | |||
2947 | ready_remove_first (struct ready_list *ready) | |||
2948 | { | |||
2949 | rtx_insn *t; | |||
2950 | ||||
2951 | gcc_assert (ready->n_ready)((void)(!(ready->n_ready) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2951, __FUNCTION__), 0 : 0)); | |||
2952 | t = ready->vec[ready->first--]; | |||
2953 | ready->n_ready--; | |||
2954 | if (DEBUG_INSN_P (t)(((enum rtx_code) (t)->code) == DEBUG_INSN)) | |||
2955 | ready->n_debug--; | |||
2956 | /* If the queue becomes empty, reset it. */ | |||
2957 | if (ready->n_ready == 0) | |||
2958 | ready->first = ready->veclen - 1; | |||
2959 | ||||
2960 | gcc_assert (QUEUE_INDEX (t) == QUEUE_READY)((void)(!(((&h_i_d[INSN_UID (t)])->queue_index) == (-1 )) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2960, __FUNCTION__), 0 : 0)); | |||
2961 | QUEUE_INDEX (t)((&h_i_d[INSN_UID (t)])->queue_index) = QUEUE_NOWHERE(-2); | |||
2962 | ||||
2963 | return t; | |||
2964 | } | |||
2965 | ||||
2966 | /* The following code implements multi-pass scheduling for the first | |||
2967 | cycle. In other words, we will try to choose ready insn which | |||
2968 | permits to start maximum number of insns on the same cycle. */ | |||
2969 | ||||
2970 | /* Return a pointer to the element INDEX from the ready. INDEX for | |||
2971 | insn with the highest priority is 0, and the lowest priority has | |||
2972 | N_READY - 1. */ | |||
2973 | ||||
2974 | rtx_insn * | |||
2975 | ready_element (struct ready_list *ready, int index) | |||
2976 | { | |||
2977 | gcc_assert (ready->n_ready && index < ready->n_ready)((void)(!(ready->n_ready && index < ready->n_ready ) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2977, __FUNCTION__), 0 : 0)); | |||
2978 | ||||
2979 | return ready->vec[ready->first - index]; | |||
2980 | } | |||
2981 | ||||
2982 | /* Remove the element INDEX from the ready list and return it. INDEX | |||
2983 | for insn with the highest priority is 0, and the lowest priority | |||
2984 | has N_READY - 1. */ | |||
2985 | ||||
2986 | HAIFA_INLINE__inline static rtx_insn * | |||
2987 | ready_remove (struct ready_list *ready, int index) | |||
2988 | { | |||
2989 | rtx_insn *t; | |||
2990 | int i; | |||
2991 | ||||
2992 | if (index == 0) | |||
2993 | return ready_remove_first (ready); | |||
2994 | gcc_assert (ready->n_ready && index < ready->n_ready)((void)(!(ready->n_ready && index < ready->n_ready ) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 2994, __FUNCTION__), 0 : 0)); | |||
2995 | t = ready->vec[ready->first - index]; | |||
2996 | ready->n_ready--; | |||
2997 | if (DEBUG_INSN_P (t)(((enum rtx_code) (t)->code) == DEBUG_INSN)) | |||
2998 | ready->n_debug--; | |||
2999 | for (i = index; i < ready->n_ready; i++) | |||
3000 | ready->vec[ready->first - i] = ready->vec[ready->first - i - 1]; | |||
3001 | QUEUE_INDEX (t)((&h_i_d[INSN_UID (t)])->queue_index) = QUEUE_NOWHERE(-2); | |||
3002 | return t; | |||
3003 | } | |||
3004 | ||||
3005 | /* Remove INSN from the ready list. */ | |||
3006 | static void | |||
3007 | ready_remove_insn (rtx_insn *insn) | |||
3008 | { | |||
3009 | int i; | |||
3010 | ||||
3011 | for (i = 0; i < readyp->n_ready; i++) | |||
3012 | if (ready_element (readyp, i) == insn) | |||
3013 | { | |||
3014 | ready_remove (readyp, i); | |||
3015 | return; | |||
3016 | } | |||
3017 | gcc_unreachable ()(fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 3017, __FUNCTION__)); | |||
3018 | } | |||
3019 | ||||
3020 | /* Calculate difference of two statistics set WAS and NOW. | |||
3021 | Result returned in WAS. */ | |||
3022 | static void | |||
3023 | rank_for_schedule_stats_diff (rank_for_schedule_stats_t *was, | |||
3024 | const rank_for_schedule_stats_t *now) | |||
3025 | { | |||
3026 | for (int i = 0; i < RFS_N; ++i) | |||
3027 | was->stats[i] = now->stats[i] - was->stats[i]; | |||
3028 | } | |||
3029 | ||||
3030 | /* Print rank_for_schedule statistics. */ | |||
3031 | static void | |||
3032 | print_rank_for_schedule_stats (const char *prefix, | |||
3033 | const rank_for_schedule_stats_t *stats, | |||
3034 | struct ready_list *ready) | |||
3035 | { | |||
3036 | for (int i = 0; i < RFS_N; ++i) | |||
3037 | if (stats->stats[i]) | |||
3038 | { | |||
3039 | fprintf (sched_dump, "%s%20s: %u", prefix, rfs_str[i], stats->stats[i]); | |||
3040 | ||||
3041 | if (ready != NULLnullptr) | |||
3042 | /* Print out insns that won due to RFS_<I>. */ | |||
3043 | { | |||
3044 | rtx_insn **p = ready_lastpos (ready); | |||
3045 | ||||
3046 | fprintf (sched_dump, ":"); | |||
3047 | /* Start with 1 since least-priority insn didn't have any wins. */ | |||
3048 | for (int j = 1; j < ready->n_ready; ++j) | |||
3049 | if (INSN_LAST_RFS_WIN (p[j])((&h_i_d[INSN_UID (p[j])])->last_rfs_win) == i) | |||
3050 | fprintf (sched_dump, " %s", | |||
3051 | (*current_sched_info->print_insn) (p[j], 0)); | |||
3052 | } | |||
3053 | fprintf (sched_dump, "\n"); | |||
3054 | } | |||
3055 | } | |||
3056 | ||||
3057 | /* Separate DEBUG_INSNS from normal insns. DEBUG_INSNs go to the end | |||
3058 | of array. */ | |||
3059 | static void | |||
3060 | ready_sort_debug (struct ready_list *ready) | |||
3061 | { | |||
3062 | int i; | |||
3063 | rtx_insn **first = ready_lastpos (ready); | |||
3064 | ||||
3065 | for (i = 0; i < ready->n_ready; ++i) | |||
3066 | if (!DEBUG_INSN_P (first[i])(((enum rtx_code) (first[i])->code) == DEBUG_INSN)) | |||
3067 | INSN_RFS_DEBUG_ORIG_ORDER (first[i])((&h_i_d[INSN_UID (first[i])])->rfs_debug_orig_order) = i; | |||
3068 | ||||
3069 | qsort (first, ready->n_ready, sizeof (rtx), rank_for_schedule_debug)gcc_qsort (first, ready->n_ready, sizeof (rtx), rank_for_schedule_debug ); | |||
3070 | } | |||
3071 | ||||
3072 | /* Sort non-debug insns in the ready list READY by ascending priority. | |||
3073 | Assumes that all debug insns are separated from the real insns. */ | |||
3074 | static void | |||
3075 | ready_sort_real (struct ready_list *ready) | |||
3076 | { | |||
3077 | int i; | |||
3078 | rtx_insn **first = ready_lastpos (ready); | |||
3079 | int n_ready_real = ready->n_ready - ready->n_debug; | |||
3080 | ||||
3081 | if (sched_pressure == SCHED_PRESSURE_WEIGHTED) | |||
3082 | for (i = 0; i < n_ready_real; ++i) | |||
3083 | setup_insn_reg_pressure_info (first[i]); | |||
3084 | else if (sched_pressure == SCHED_PRESSURE_MODEL | |||
3085 | && model_curr_point < model_num_insns) | |||
3086 | model_set_excess_costs (first, n_ready_real); | |||
3087 | ||||
3088 | rank_for_schedule_stats_t stats1; | |||
3089 | if (sched_verbose >= 4) | |||
3090 | stats1 = rank_for_schedule_stats; | |||
3091 | ||||
3092 | if (n_ready_real == 2) | |||
3093 | swap_sort (first, n_ready_real); | |||
3094 | else if (n_ready_real > 2) | |||
3095 | qsort (first, n_ready_real, sizeof (rtx), rank_for_schedule)gcc_qsort (first, n_ready_real, sizeof (rtx), rank_for_schedule ); | |||
3096 | ||||
3097 | if (sched_verbose >= 4) | |||
3098 | { | |||
3099 | rank_for_schedule_stats_diff (&stats1, &rank_for_schedule_stats); | |||
3100 | print_rank_for_schedule_stats (";;\t\t", &stats1, ready); | |||
3101 | } | |||
3102 | } | |||
3103 | ||||
3104 | /* Sort the ready list READY by ascending priority. */ | |||
3105 | static void | |||
3106 | ready_sort (struct ready_list *ready) | |||
3107 | { | |||
3108 | if (ready->n_debug > 0) | |||
3109 | ready_sort_debug (ready); | |||
3110 | else | |||
3111 | ready_sort_real (ready); | |||
3112 | } | |||
3113 | ||||
3114 | /* PREV is an insn that is ready to execute. Adjust its priority if that | |||
3115 | will help shorten or lengthen register lifetimes as appropriate. Also | |||
3116 | provide a hook for the target to tweak itself. */ | |||
3117 | ||||
3118 | HAIFA_INLINE__inline static void | |||
3119 | adjust_priority (rtx_insn *prev) | |||
3120 | { | |||
3121 | /* ??? There used to be code here to try and estimate how an insn | |||
3122 | affected register lifetimes, but it did it by looking at REG_DEAD | |||
3123 | notes, which we removed in schedule_region. Nor did it try to | |||
3124 | take into account register pressure or anything useful like that. | |||
3125 | ||||
3126 | Revisit when we have a machine model to work with and not before. */ | |||
3127 | ||||
3128 | if (targetm.sched.adjust_priority) | |||
3129 | INSN_PRIORITY (prev)((&h_i_d[INSN_UID (prev)])->priority) = | |||
3130 | targetm.sched.adjust_priority (prev, INSN_PRIORITY (prev)((&h_i_d[INSN_UID (prev)])->priority)); | |||
3131 | } | |||
3132 | ||||
3133 | /* Advance DFA state STATE on one cycle. */ | |||
3134 | void | |||
3135 | advance_state (state_t state) | |||
3136 | { | |||
3137 | if (targetm.sched.dfa_pre_advance_cycle) | |||
3138 | targetm.sched.dfa_pre_advance_cycle (); | |||
3139 | ||||
3140 | if (targetm.sched.dfa_pre_cycle_insn) | |||
3141 | state_transition (state, | |||
3142 | targetm.sched.dfa_pre_cycle_insn ()); | |||
3143 | ||||
3144 | state_transition (state, NULLnullptr); | |||
3145 | ||||
3146 | if (targetm.sched.dfa_post_cycle_insn) | |||
3147 | state_transition (state, | |||
3148 | targetm.sched.dfa_post_cycle_insn ()); | |||
3149 | ||||
3150 | if (targetm.sched.dfa_post_advance_cycle) | |||
3151 | targetm.sched.dfa_post_advance_cycle (); | |||
3152 | } | |||
3153 | ||||
3154 | /* Advance time on one cycle. */ | |||
3155 | HAIFA_INLINE__inline static void | |||
3156 | advance_one_cycle (void) | |||
3157 | { | |||
3158 | int i; | |||
3159 | ||||
3160 | advance_state (curr_state); | |||
3161 | for (i = 4; i <= sched_verbose; ++i) | |||
3162 | fprintf (sched_dump, ";;\tAdvance the current state: %d.\n", clock_var); | |||
3163 | } | |||
3164 | ||||
3165 | /* Update register pressure after scheduling INSN. */ | |||
3166 | static void | |||
3167 | update_register_pressure (rtx_insn *insn) | |||
3168 | { | |||
3169 | struct reg_use_data *use; | |||
3170 | struct reg_set_data *set; | |||
3171 | ||||
3172 | gcc_checking_assert (!DEBUG_INSN_P (insn))((void)(!(!(((enum rtx_code) (insn)->code) == DEBUG_INSN)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 3172, __FUNCTION__), 0 : 0)); | |||
3173 | ||||
3174 | for (use = INSN_REG_USE_LIST (insn)((&h_i_d[INSN_UID (insn)])->reg_use_list); use != NULLnullptr; use = use->next_insn_use) | |||
3175 | if (dying_use_p (use)) | |||
3176 | mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure, | |||
3177 | use->regno, false); | |||
3178 | for (set = INSN_REG_SET_LIST (insn)((&h_i_d[INSN_UID (insn)])->reg_set_list); set != NULLnullptr; set = set->next_insn_set) | |||
3179 | mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure, | |||
3180 | set->regno, true); | |||
3181 | } | |||
3182 | ||||
3183 | /* Set up or update (if UPDATE_P) max register pressure (see its | |||
3184 | meaning in sched-int.h::_haifa_insn_data) for all current BB insns | |||
3185 | after insn AFTER. */ | |||
3186 | static void | |||
3187 | setup_insn_max_reg_pressure (rtx_insn *after, bool update_p) | |||
3188 | { | |||
3189 | int i, p; | |||
3190 | bool eq_p; | |||
3191 | rtx_insn *insn; | |||
3192 | static int max_reg_pressure[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
3193 | ||||
3194 | save_reg_pressure (); | |||
3195 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
3196 | max_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]] | |||
3197 | = curr_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]]; | |||
3198 | for (insn = NEXT_INSN (after); | |||
3199 | insn != NULL_RTX(rtx) 0 && ! BARRIER_P (insn)(((enum rtx_code) (insn)->code) == BARRIER) | |||
3200 | && BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (after); | |||
3201 | insn = NEXT_INSN (insn)) | |||
3202 | if (NONDEBUG_INSN_P (insn)((((enum rtx_code) (insn)->code) == INSN) || (((enum rtx_code ) (insn)->code) == JUMP_INSN) || (((enum rtx_code) (insn)-> code) == CALL_INSN))) | |||
3203 | { | |||
3204 | eq_p = true; | |||
3205 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
3206 | { | |||
3207 | p = max_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]]; | |||
3208 | if (INSN_MAX_REG_PRESSURE (insn)((&h_i_d[INSN_UID (insn)])->max_reg_pressure)[i] != p) | |||
3209 | { | |||
3210 | eq_p = false; | |||
3211 | INSN_MAX_REG_PRESSURE (insn)((&h_i_d[INSN_UID (insn)])->max_reg_pressure)[i] | |||
3212 | = max_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]]; | |||
3213 | } | |||
3214 | } | |||
3215 | if (update_p && eq_p) | |||
3216 | break; | |||
3217 | update_register_pressure (insn); | |||
3218 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
3219 | if (max_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]] | |||
3220 | < curr_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]]) | |||
3221 | max_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]] | |||
3222 | = curr_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]]; | |||
3223 | } | |||
3224 | restore_reg_pressure (); | |||
3225 | } | |||
3226 | ||||
3227 | /* Update the current register pressure after scheduling INSN. Update | |||
3228 | also max register pressure for unscheduled insns of the current | |||
3229 | BB. */ | |||
3230 | static void | |||
3231 | update_reg_and_insn_max_reg_pressure (rtx_insn *insn) | |||
3232 | { | |||
3233 | int i; | |||
3234 | int before[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
3235 | ||||
3236 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
3237 | before[i] = curr_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]]; | |||
3238 | update_register_pressure (insn); | |||
3239 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
3240 | if (curr_reg_pressure[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]] != before[i]) | |||
| ||||
3241 | break; | |||
3242 | if (i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num)) | |||
3243 | setup_insn_max_reg_pressure (insn, true); | |||
3244 | } | |||
3245 | ||||
3246 | /* Set up register pressure at the beginning of basic block BB whose | |||
3247 | insns starting after insn AFTER. Set up also max register pressure | |||
3248 | for all insns of the basic block. */ | |||
3249 | void | |||
3250 | sched_setup_bb_reg_pressure_info (basic_block bb, rtx_insn *after) | |||
3251 | { | |||
3252 | gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED)((void)(!(sched_pressure == SCHED_PRESSURE_WEIGHTED) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 3252, __FUNCTION__), 0 : 0)); | |||
3253 | initiate_bb_reg_pressure_info (bb); | |||
3254 | setup_insn_max_reg_pressure (after, false); | |||
3255 | } | |||
3256 | ||||
3257 | /* If doing predication while scheduling, verify whether INSN, which | |||
3258 | has just been scheduled, clobbers the conditions of any | |||
3259 | instructions that must be predicated in order to break their | |||
3260 | dependencies. If so, remove them from the queues so that they will | |||
3261 | only be scheduled once their control dependency is resolved. */ | |||
3262 | ||||
3263 | static void | |||
3264 | check_clobbered_conditions (rtx_insn *insn) | |||
3265 | { | |||
3266 | HARD_REG_SET t; | |||
3267 | int i; | |||
3268 | ||||
3269 | if ((current_sched_info->flags & DO_PREDICATION) == 0) | |||
3270 | return; | |||
3271 | ||||
3272 | find_all_hard_reg_sets (insn, &t, true); | |||
3273 | ||||
3274 | restart: | |||
3275 | for (i = 0; i < ready.n_ready; i++) | |||
3276 | { | |||
3277 | rtx_insn *x = ready_element (&ready, i); | |||
3278 | if (TODO_SPEC (x)((&h_i_d[INSN_UID (x)])->todo_spec) == DEP_CONTROL((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) && cond_clobbered_p (x, t)) | |||
3279 | { | |||
3280 | ready_remove_insn (x); | |||
3281 | goto restart; | |||
3282 | } | |||
3283 | } | |||
3284 | for (i = 0; i <= max_insn_queue_index; i++) | |||
3285 | { | |||
3286 | rtx_insn_list *link; | |||
3287 | int q = NEXT_Q_AFTER (q_ptr, i)(((q_ptr)+i) & max_insn_queue_index); | |||
3288 | ||||
3289 | restart_queue: | |||
3290 | for (link = insn_queue[q]; link; link = link->next ()) | |||
3291 | { | |||
3292 | rtx_insn *x = link->insn (); | |||
3293 | if (TODO_SPEC (x)((&h_i_d[INSN_UID (x)])->todo_spec) == DEP_CONTROL((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) && cond_clobbered_p (x, t)) | |||
3294 | { | |||
3295 | queue_remove (x); | |||
3296 | goto restart_queue; | |||
3297 | } | |||
3298 | } | |||
3299 | } | |||
3300 | } | |||
3301 | ||||
3302 | /* Return (in order): | |||
3303 | ||||
3304 | - positive if INSN adversely affects the pressure on one | |||
3305 | register class | |||
3306 | ||||
3307 | - negative if INSN reduces the pressure on one register class | |||
3308 | ||||
3309 | - 0 if INSN doesn't affect the pressure on any register class. */ | |||
3310 | ||||
3311 | static int | |||
3312 | model_classify_pressure (struct model_insn_info *insn) | |||
3313 | { | |||
3314 | struct reg_pressure_data *reg_pressure; | |||
3315 | int death[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
3316 | int pci, cl, sum; | |||
3317 | ||||
3318 | calculate_reg_deaths (insn->insn, death); | |||
3319 | reg_pressure = INSN_REG_PRESSURE (insn->insn)((&h_i_d[INSN_UID (insn->insn)])->reg_pressure); | |||
3320 | sum = 0; | |||
3321 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
3322 | { | |||
3323 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
3324 | if (death[cl] < reg_pressure[pci].set_increase) | |||
3325 | return 1; | |||
3326 | sum += reg_pressure[pci].set_increase - death[cl]; | |||
3327 | } | |||
3328 | return sum; | |||
3329 | } | |||
3330 | ||||
3331 | /* Return true if INSN1 should come before INSN2 in the model schedule. */ | |||
3332 | ||||
3333 | static int | |||
3334 | model_order_p (struct model_insn_info *insn1, struct model_insn_info *insn2) | |||
3335 | { | |||
3336 | unsigned int height1, height2; | |||
3337 | unsigned int priority1, priority2; | |||
3338 | ||||
3339 | /* Prefer instructions with a higher model priority. */ | |||
3340 | if (insn1->model_priority != insn2->model_priority) | |||
3341 | return insn1->model_priority > insn2->model_priority; | |||
3342 | ||||
3343 | /* Combine the length of the longest path of satisfied true dependencies | |||
3344 | that leads to each instruction (depth) with the length of the longest | |||
3345 | path of any dependencies that leads from the instruction (alap). | |||
3346 | Prefer instructions with the greatest combined length. If the combined | |||
3347 | lengths are equal, prefer instructions with the greatest depth. | |||
3348 | ||||
3349 | The idea is that, if we have a set S of "equal" instructions that each | |||
3350 | have ALAP value X, and we pick one such instruction I, any true-dependent | |||
3351 | successors of I that have ALAP value X - 1 should be preferred over S. | |||
3352 | This encourages the schedule to be "narrow" rather than "wide". | |||
3353 | However, if I is a low-priority instruction that we decided to | |||
3354 | schedule because of its model_classify_pressure, and if there | |||
3355 | is a set of higher-priority instructions T, the aforementioned | |||
3356 | successors of I should not have the edge over T. */ | |||
3357 | height1 = insn1->depth + insn1->alap; | |||
3358 | height2 = insn2->depth + insn2->alap; | |||
3359 | if (height1 != height2) | |||
3360 | return height1 > height2; | |||
3361 | if (insn1->depth != insn2->depth) | |||
3362 | return insn1->depth > insn2->depth; | |||
3363 | ||||
3364 | /* We have no real preference between INSN1 an INSN2 as far as attempts | |||
3365 | to reduce pressure go. Prefer instructions with higher priorities. */ | |||
3366 | priority1 = INSN_PRIORITY (insn1->insn)((&h_i_d[INSN_UID (insn1->insn)])->priority); | |||
3367 | priority2 = INSN_PRIORITY (insn2->insn)((&h_i_d[INSN_UID (insn2->insn)])->priority); | |||
3368 | if (priority1 != priority2) | |||
3369 | return priority1 > priority2; | |||
3370 | ||||
3371 | /* Use the original rtl sequence as a tie-breaker. */ | |||
3372 | return insn1 < insn2; | |||
3373 | } | |||
3374 | ||||
3375 | /* Add INSN to the model worklist immediately after PREV. Add it to the | |||
3376 | beginning of the list if PREV is null. */ | |||
3377 | ||||
3378 | static void | |||
3379 | model_add_to_worklist_at (struct model_insn_info *insn, | |||
3380 | struct model_insn_info *prev) | |||
3381 | { | |||
3382 | gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_NOWHERE)((void)(!(((&h_i_d[INSN_UID (insn->insn)])->queue_index ) == (-2)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 3382, __FUNCTION__), 0 : 0)); | |||
3383 | QUEUE_INDEX (insn->insn)((&h_i_d[INSN_UID (insn->insn)])->queue_index) = QUEUE_READY(-1); | |||
3384 | ||||
3385 | insn->prev = prev; | |||
3386 | if (prev) | |||
3387 | { | |||
3388 | insn->next = prev->next; | |||
3389 | prev->next = insn; | |||
3390 | } | |||
3391 | else | |||
3392 | { | |||
3393 | insn->next = model_worklist; | |||
3394 | model_worklist = insn; | |||
3395 | } | |||
3396 | if (insn->next) | |||
3397 | insn->next->prev = insn; | |||
3398 | } | |||
3399 | ||||
3400 | /* Remove INSN from the model worklist. */ | |||
3401 | ||||
3402 | static void | |||
3403 | model_remove_from_worklist (struct model_insn_info *insn) | |||
3404 | { | |||
3405 | gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_READY)((void)(!(((&h_i_d[INSN_UID (insn->insn)])->queue_index ) == (-1)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 3405, __FUNCTION__), 0 : 0)); | |||
3406 | QUEUE_INDEX (insn->insn)((&h_i_d[INSN_UID (insn->insn)])->queue_index) = QUEUE_NOWHERE(-2); | |||
3407 | ||||
3408 | if (insn->prev) | |||
3409 | insn->prev->next = insn->next; | |||
3410 | else | |||
3411 | model_worklist = insn->next; | |||
3412 | if (insn->next) | |||
3413 | insn->next->prev = insn->prev; | |||
3414 | } | |||
3415 | ||||
3416 | /* Add INSN to the model worklist. Start looking for a suitable position | |||
3417 | between neighbors PREV and NEXT, testing at most param_max_sched_ready_insns | |||
3418 | insns either side. A null PREV indicates the beginning of the list and | |||
3419 | a null NEXT indicates the end. */ | |||
3420 | ||||
3421 | static void | |||
3422 | model_add_to_worklist (struct model_insn_info *insn, | |||
3423 | struct model_insn_info *prev, | |||
3424 | struct model_insn_info *next) | |||
3425 | { | |||
3426 | int count; | |||
3427 | ||||
3428 | count = param_max_sched_ready_insnsglobal_options.x_param_max_sched_ready_insns; | |||
3429 | if (count > 0 && prev && model_order_p (insn, prev)) | |||
3430 | do | |||
3431 | { | |||
3432 | count--; | |||
3433 | prev = prev->prev; | |||
3434 | } | |||
3435 | while (count > 0 && prev && model_order_p (insn, prev)); | |||
3436 | else | |||
3437 | while (count > 0 && next && model_order_p (next, insn)) | |||
3438 | { | |||
3439 | count--; | |||
3440 | prev = next; | |||
3441 | next = next->next; | |||
3442 | } | |||
3443 | model_add_to_worklist_at (insn, prev); | |||
3444 | } | |||
3445 | ||||
3446 | /* INSN may now have a higher priority (in the model_order_p sense) | |||
3447 | than before. Move it up the worklist if necessary. */ | |||
3448 | ||||
3449 | static void | |||
3450 | model_promote_insn (struct model_insn_info *insn) | |||
3451 | { | |||
3452 | struct model_insn_info *prev; | |||
3453 | int count; | |||
3454 | ||||
3455 | prev = insn->prev; | |||
3456 | count = param_max_sched_ready_insnsglobal_options.x_param_max_sched_ready_insns; | |||
3457 | while (count > 0 && prev && model_order_p (insn, prev)) | |||
3458 | { | |||
3459 | count--; | |||
3460 | prev = prev->prev; | |||
3461 | } | |||
3462 | if (prev != insn->prev) | |||
3463 | { | |||
3464 | model_remove_from_worklist (insn); | |||
3465 | model_add_to_worklist_at (insn, prev); | |||
3466 | } | |||
3467 | } | |||
3468 | ||||
3469 | /* Add INSN to the end of the model schedule. */ | |||
3470 | ||||
3471 | static void | |||
3472 | model_add_to_schedule (rtx_insn *insn) | |||
3473 | { | |||
3474 | unsigned int point; | |||
3475 | ||||
3476 | gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE)((void)(!(((&h_i_d[INSN_UID (insn)])->queue_index) == ( -2)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 3476, __FUNCTION__), 0 : 0)); | |||
3477 | QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index) = QUEUE_SCHEDULED(-3); | |||
3478 | ||||
3479 | point = model_schedule.length (); | |||
3480 | model_schedule.quick_push (insn); | |||
3481 | INSN_MODEL_INDEX (insn)((&h_i_d[INSN_UID (insn)])->model_index) = point + 1; | |||
3482 | } | |||
3483 | ||||
3484 | /* Analyze the instructions that are to be scheduled, setting up | |||
3485 | MODEL_INSN_INFO (...) and model_num_insns accordingly. Add ready | |||
3486 | instructions to model_worklist. */ | |||
3487 | ||||
3488 | static void | |||
3489 | model_analyze_insns (void) | |||
3490 | { | |||
3491 | rtx_insn *start, *end, *iter; | |||
3492 | sd_iterator_def sd_it; | |||
3493 | dep_t dep; | |||
3494 | struct model_insn_info *insn, *con; | |||
3495 | ||||
3496 | model_num_insns = 0; | |||
3497 | start = PREV_INSN (current_sched_info->next_tail); | |||
3498 | end = current_sched_info->prev_head; | |||
3499 | for (iter = start; iter != end; iter = PREV_INSN (iter)) | |||
3500 | if (NONDEBUG_INSN_P (iter)((((enum rtx_code) (iter)->code) == INSN) || (((enum rtx_code ) (iter)->code) == JUMP_INSN) || (((enum rtx_code) (iter)-> code) == CALL_INSN))) | |||
3501 | { | |||
3502 | insn = MODEL_INSN_INFO (iter)(&model_insns[(sched_luids[INSN_UID (iter)])]); | |||
3503 | insn->insn = iter; | |||
3504 | FOR_EACH_DEP (iter, SD_LIST_FORW, sd_it, dep)for ((sd_it) = sd_iterator_start ((iter), ((4))); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
3505 | { | |||
3506 | con = MODEL_INSN_INFO (DEP_CON (dep))(&model_insns[(sched_luids[INSN_UID (((dep)->con))])]); | |||
3507 | if (con->insn && insn->alap < con->alap + 1) | |||
3508 | insn->alap = con->alap + 1; | |||
3509 | } | |||
3510 | ||||
3511 | insn->old_queue = QUEUE_INDEX (iter)((&h_i_d[INSN_UID (iter)])->queue_index); | |||
3512 | QUEUE_INDEX (iter)((&h_i_d[INSN_UID (iter)])->queue_index) = QUEUE_NOWHERE(-2); | |||
3513 | ||||
3514 | insn->unscheduled_preds = dep_list_size (iter, SD_LIST_HARD_BACK(1)); | |||
3515 | if (insn->unscheduled_preds == 0) | |||
3516 | model_add_to_worklist (insn, NULLnullptr, model_worklist); | |||
3517 | ||||
3518 | model_num_insns++; | |||
3519 | } | |||
3520 | } | |||
3521 | ||||
3522 | /* The global state describes the register pressure at the start of the | |||
3523 | model schedule. Initialize GROUP accordingly. */ | |||
3524 | ||||
3525 | static void | |||
3526 | model_init_pressure_group (struct model_pressure_group *group) | |||
3527 | { | |||
3528 | int pci, cl; | |||
3529 | ||||
3530 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
3531 | { | |||
3532 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
3533 | group->limits[pci].pressure = curr_reg_pressure[cl]; | |||
3534 | group->limits[pci].point = 0; | |||
3535 | } | |||
3536 | /* Use index model_num_insns to record the state after the last | |||
3537 | instruction in the model schedule. */ | |||
3538 | group->model = XNEWVEC (struct model_pressure_data,((struct model_pressure_data *) xmalloc (sizeof (struct model_pressure_data ) * ((model_num_insns + 1) * (this_target_ira->x_ira_pressure_classes_num )))) | |||
3539 | (model_num_insns + 1) * ira_pressure_classes_num)((struct model_pressure_data *) xmalloc (sizeof (struct model_pressure_data ) * ((model_num_insns + 1) * (this_target_ira->x_ira_pressure_classes_num )))); | |||
3540 | } | |||
3541 | ||||
3542 | /* Record that MODEL_REF_PRESSURE (GROUP, POINT, PCI) is PRESSURE. | |||
3543 | Update the maximum pressure for the whole schedule. */ | |||
3544 | ||||
3545 | static void | |||
3546 | model_record_pressure (struct model_pressure_group *group, | |||
3547 | int point, int pci, int pressure) | |||
3548 | { | |||
3549 | MODEL_REF_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->ref_pressure) = pressure; | |||
3550 | if (group->limits[pci].pressure < pressure) | |||
3551 | { | |||
3552 | group->limits[pci].pressure = pressure; | |||
3553 | group->limits[pci].point = point; | |||
3554 | } | |||
3555 | } | |||
3556 | ||||
3557 | /* INSN has just been added to the end of the model schedule. Record its | |||
3558 | register-pressure information. */ | |||
3559 | ||||
3560 | static void | |||
3561 | model_record_pressures (struct model_insn_info *insn) | |||
3562 | { | |||
3563 | struct reg_pressure_data *reg_pressure; | |||
3564 | int point, pci, cl, delta; | |||
3565 | int death[N_REG_CLASSES((int) LIM_REG_CLASSES)]; | |||
3566 | ||||
3567 | point = model_index (insn->insn); | |||
3568 | if (sched_verbose >= 2) | |||
3569 | { | |||
3570 | if (point == 0) | |||
3571 | { | |||
3572 | fprintf (sched_dump, "\n;;\tModel schedule:\n;;\n"); | |||
3573 | fprintf (sched_dump, ";;\t| idx insn | mpri hght dpth prio |\n"); | |||
3574 | } | |||
3575 | fprintf (sched_dump, ";;\t| %3d %4d | %4d %4d %4d %4d | %-30s ", | |||
3576 | point, INSN_UID (insn->insn), insn->model_priority, | |||
3577 | insn->depth + insn->alap, insn->depth, | |||
3578 | INSN_PRIORITY (insn->insn)((&h_i_d[INSN_UID (insn->insn)])->priority), | |||
3579 | str_pattern_slim (PATTERN (insn->insn))); | |||
3580 | } | |||
3581 | calculate_reg_deaths (insn->insn, death); | |||
3582 | reg_pressure = INSN_REG_PRESSURE (insn->insn)((&h_i_d[INSN_UID (insn->insn)])->reg_pressure); | |||
3583 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
3584 | { | |||
3585 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
3586 | delta = reg_pressure[pci].set_increase - death[cl]; | |||
3587 | if (sched_verbose >= 2) | |||
3588 | fprintf (sched_dump, " %s:[%d,%+d]", reg_class_names[cl], | |||
3589 | curr_reg_pressure[cl], delta); | |||
3590 | model_record_pressure (&model_before_pressure, point, pci, | |||
3591 | curr_reg_pressure[cl]); | |||
3592 | } | |||
3593 | if (sched_verbose >= 2) | |||
3594 | fprintf (sched_dump, "\n"); | |||
3595 | } | |||
3596 | ||||
3597 | /* All instructions have been added to the model schedule. Record the | |||
3598 | final register pressure in GROUP and set up all MODEL_MAX_PRESSUREs. */ | |||
3599 | ||||
3600 | static void | |||
3601 | model_record_final_pressures (struct model_pressure_group *group) | |||
3602 | { | |||
3603 | int point, pci, max_pressure, ref_pressure, cl; | |||
3604 | ||||
3605 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
3606 | { | |||
3607 | /* Record the final pressure for this class. */ | |||
3608 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
3609 | point = model_num_insns; | |||
3610 | ref_pressure = curr_reg_pressure[cl]; | |||
3611 | model_record_pressure (group, point, pci, ref_pressure); | |||
3612 | ||||
3613 | /* Record the original maximum pressure. */ | |||
3614 | group->limits[pci].orig_pressure = group->limits[pci].pressure; | |||
3615 | ||||
3616 | /* Update the MODEL_MAX_PRESSURE for every point of the schedule. */ | |||
3617 | max_pressure = ref_pressure; | |||
3618 | MODEL_MAX_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->max_pressure) = max_pressure; | |||
3619 | while (point > 0) | |||
3620 | { | |||
3621 | point--; | |||
3622 | ref_pressure = MODEL_REF_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->ref_pressure); | |||
3623 | max_pressure = MAX (max_pressure, ref_pressure)((max_pressure) > (ref_pressure) ? (max_pressure) : (ref_pressure )); | |||
3624 | MODEL_MAX_PRESSURE (group, point, pci)((&(group)->model[(point) * (this_target_ira->x_ira_pressure_classes_num ) + (pci)])->max_pressure) = max_pressure; | |||
3625 | } | |||
3626 | } | |||
3627 | } | |||
3628 | ||||
3629 | /* Update all successors of INSN, given that INSN has just been scheduled. */ | |||
3630 | ||||
3631 | static void | |||
3632 | model_add_successors_to_worklist (struct model_insn_info *insn) | |||
3633 | { | |||
3634 | sd_iterator_def sd_it; | |||
3635 | struct model_insn_info *con; | |||
3636 | dep_t dep; | |||
3637 | ||||
3638 | FOR_EACH_DEP (insn->insn, SD_LIST_FORW, sd_it, dep)for ((sd_it) = sd_iterator_start ((insn->insn), ((4))); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
3639 | { | |||
3640 | con = MODEL_INSN_INFO (DEP_CON (dep))(&model_insns[(sched_luids[INSN_UID (((dep)->con))])]); | |||
3641 | /* Ignore debug instructions, and instructions from other blocks. */ | |||
3642 | if (con->insn) | |||
3643 | { | |||
3644 | con->unscheduled_preds--; | |||
3645 | ||||
3646 | /* Update the depth field of each true-dependent successor. | |||
3647 | Increasing the depth gives them a higher priority than | |||
3648 | before. */ | |||
3649 | if (DEP_TYPE (dep)((dep)->type) == REG_DEP_TRUE && con->depth < insn->depth + 1) | |||
3650 | { | |||
3651 | con->depth = insn->depth + 1; | |||
3652 | if (QUEUE_INDEX (con->insn)((&h_i_d[INSN_UID (con->insn)])->queue_index) == QUEUE_READY(-1)) | |||
3653 | model_promote_insn (con); | |||
3654 | } | |||
3655 | ||||
3656 | /* If this is a true dependency, or if there are no remaining | |||
3657 | dependencies for CON (meaning that CON only had non-true | |||
3658 | dependencies), make sure that CON is on the worklist. | |||
3659 | We don't bother otherwise because it would tend to fill the | |||
3660 | worklist with a lot of low-priority instructions that are not | |||
3661 | yet ready to issue. */ | |||
3662 | if ((con->depth > 0 || con->unscheduled_preds == 0) | |||
3663 | && QUEUE_INDEX (con->insn)((&h_i_d[INSN_UID (con->insn)])->queue_index) == QUEUE_NOWHERE(-2)) | |||
3664 | model_add_to_worklist (con, insn, insn->next); | |||
3665 | } | |||
3666 | } | |||
3667 | } | |||
3668 | ||||
3669 | /* Give INSN a higher priority than any current instruction, then give | |||
3670 | unscheduled predecessors of INSN a higher priority still. If any of | |||
3671 | those predecessors are not on the model worklist, do the same for its | |||
3672 | predecessors, and so on. */ | |||
3673 | ||||
3674 | static void | |||
3675 | model_promote_predecessors (struct model_insn_info *insn) | |||
3676 | { | |||
3677 | struct model_insn_info *pro, *first; | |||
3678 | sd_iterator_def sd_it; | |||
3679 | dep_t dep; | |||
3680 | ||||
3681 | if (sched_verbose >= 7) | |||
3682 | fprintf (sched_dump, ";;\t+--- priority of %d = %d, priority of", | |||
3683 | INSN_UID (insn->insn), model_next_priority); | |||
3684 | insn->model_priority = model_next_priority++; | |||
3685 | model_remove_from_worklist (insn); | |||
3686 | model_add_to_worklist_at (insn, NULLnullptr); | |||
3687 | ||||
3688 | first = NULLnullptr; | |||
3689 | for (;;) | |||
3690 | { | |||
3691 | FOR_EACH_DEP (insn->insn, SD_LIST_HARD_BACK, sd_it, dep)for ((sd_it) = sd_iterator_start ((insn->insn), ((1))); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
3692 | { | |||
3693 | pro = MODEL_INSN_INFO (DEP_PRO (dep))(&model_insns[(sched_luids[INSN_UID (((dep)->pro))])]); | |||
3694 | /* The first test is to ignore debug instructions, and instructions | |||
3695 | from other blocks. */ | |||
3696 | if (pro->insn | |||
3697 | && pro->model_priority != model_next_priority | |||
3698 | && QUEUE_INDEX (pro->insn)((&h_i_d[INSN_UID (pro->insn)])->queue_index) != QUEUE_SCHEDULED(-3)) | |||
3699 | { | |||
3700 | pro->model_priority = model_next_priority; | |||
3701 | if (sched_verbose >= 7) | |||
3702 | fprintf (sched_dump, " %d", INSN_UID (pro->insn)); | |||
3703 | if (QUEUE_INDEX (pro->insn)((&h_i_d[INSN_UID (pro->insn)])->queue_index) == QUEUE_READY(-1)) | |||
3704 | { | |||
3705 | /* PRO is already in the worklist, but it now has | |||
3706 | a higher priority than before. Move it at the | |||
3707 | appropriate place. */ | |||
3708 | model_remove_from_worklist (pro); | |||
3709 | model_add_to_worklist (pro, NULLnullptr, model_worklist); | |||
3710 | } | |||
3711 | else | |||
3712 | { | |||
3713 | /* PRO isn't in the worklist. Recursively process | |||
3714 | its predecessors until we find one that is. */ | |||
3715 | pro->next = first; | |||
3716 | first = pro; | |||
3717 | } | |||
3718 | } | |||
3719 | } | |||
3720 | if (!first) | |||
3721 | break; | |||
3722 | insn = first; | |||
3723 | first = insn->next; | |||
3724 | } | |||
3725 | if (sched_verbose >= 7) | |||
3726 | fprintf (sched_dump, " = %d\n", model_next_priority); | |||
3727 | model_next_priority++; | |||
3728 | } | |||
3729 | ||||
3730 | /* Pick one instruction from model_worklist and process it. */ | |||
3731 | ||||
3732 | static void | |||
3733 | model_choose_insn (void) | |||
3734 | { | |||
3735 | struct model_insn_info *insn, *fallback; | |||
3736 | int count; | |||
3737 | ||||
3738 | if (sched_verbose >= 7) | |||
3739 | { | |||
3740 | fprintf (sched_dump, ";;\t+--- worklist:\n"); | |||
3741 | insn = model_worklist; | |||
3742 | count = param_max_sched_ready_insnsglobal_options.x_param_max_sched_ready_insns; | |||
3743 | while (count > 0 && insn) | |||
3744 | { | |||
3745 | fprintf (sched_dump, ";;\t+--- %d [%d, %d, %d, %d]\n", | |||
3746 | INSN_UID (insn->insn), insn->model_priority, | |||
3747 | insn->depth + insn->alap, insn->depth, | |||
3748 | INSN_PRIORITY (insn->insn)((&h_i_d[INSN_UID (insn->insn)])->priority)); | |||
3749 | count--; | |||
3750 | insn = insn->next; | |||
3751 | } | |||
3752 | } | |||
3753 | ||||
3754 | /* Look for a ready instruction whose model_classify_priority is zero | |||
3755 | or negative, picking the highest-priority one. Adding such an | |||
3756 | instruction to the schedule now should do no harm, and may actually | |||
3757 | do some good. | |||
3758 | ||||
3759 | Failing that, see whether there is an instruction with the highest | |||
3760 | extant model_priority that is not yet ready, but which would reduce | |||
3761 | pressure if it became ready. This is designed to catch cases like: | |||
3762 | ||||
3763 | (set (mem (reg R1)) (reg R2)) | |||
3764 | ||||
3765 | where the instruction is the last remaining use of R1 and where the | |||
3766 | value of R2 is not yet available (or vice versa). The death of R1 | |||
3767 | means that this instruction already reduces pressure. It is of | |||
3768 | course possible that the computation of R2 involves other registers | |||
3769 | that are hard to kill, but such cases are rare enough for this | |||
3770 | heuristic to be a win in general. | |||
3771 | ||||
3772 | Failing that, just pick the highest-priority instruction in the | |||
3773 | worklist. */ | |||
3774 | count = param_max_sched_ready_insnsglobal_options.x_param_max_sched_ready_insns; | |||
3775 | insn = model_worklist; | |||
3776 | fallback = 0; | |||
3777 | for (;;) | |||
3778 | { | |||
3779 | if (count == 0 || !insn) | |||
3780 | { | |||
3781 | insn = fallback ? fallback : model_worklist; | |||
3782 | break; | |||
3783 | } | |||
3784 | if (insn->unscheduled_preds) | |||
3785 | { | |||
3786 | if (model_worklist->model_priority == insn->model_priority | |||
3787 | && !fallback | |||
3788 | && model_classify_pressure (insn) < 0) | |||
3789 | fallback = insn; | |||
3790 | } | |||
3791 | else | |||
3792 | { | |||
3793 | if (model_classify_pressure (insn) <= 0) | |||
3794 | break; | |||
3795 | } | |||
3796 | count--; | |||
3797 | insn = insn->next; | |||
3798 | } | |||
3799 | ||||
3800 | if (sched_verbose >= 7 && insn != model_worklist) | |||
3801 | { | |||
3802 | if (insn->unscheduled_preds) | |||
3803 | fprintf (sched_dump, ";;\t+--- promoting insn %d, with dependencies\n", | |||
3804 | INSN_UID (insn->insn)); | |||
3805 | else | |||
3806 | fprintf (sched_dump, ";;\t+--- promoting insn %d, which is ready\n", | |||
3807 | INSN_UID (insn->insn)); | |||
3808 | } | |||
3809 | if (insn->unscheduled_preds) | |||
3810 | /* INSN isn't yet ready to issue. Give all its predecessors the | |||
3811 | highest priority. */ | |||
3812 | model_promote_predecessors (insn); | |||
3813 | else | |||
3814 | { | |||
3815 | /* INSN is ready. Add it to the end of model_schedule and | |||
3816 | process its successors. */ | |||
3817 | model_add_successors_to_worklist (insn); | |||
3818 | model_remove_from_worklist (insn); | |||
3819 | model_add_to_schedule (insn->insn); | |||
3820 | model_record_pressures (insn); | |||
3821 | update_register_pressure (insn->insn); | |||
3822 | } | |||
3823 | } | |||
3824 | ||||
3825 | /* Restore all QUEUE_INDEXs to the values that they had before | |||
3826 | model_start_schedule was called. */ | |||
3827 | ||||
3828 | static void | |||
3829 | model_reset_queue_indices (void) | |||
3830 | { | |||
3831 | unsigned int i; | |||
3832 | rtx_insn *insn; | |||
3833 | ||||
3834 | FOR_EACH_VEC_ELT (model_schedule, i, insn)for (i = 0; (model_schedule).iterate ((i), &(insn)); ++(i )) | |||
3835 | QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index) = MODEL_INSN_INFO (insn)(&model_insns[(sched_luids[INSN_UID (insn)])])->old_queue; | |||
3836 | } | |||
3837 | ||||
3838 | /* We have calculated the model schedule and spill costs. Print a summary | |||
3839 | to sched_dump. */ | |||
3840 | ||||
3841 | static void | |||
3842 | model_dump_pressure_summary (void) | |||
3843 | { | |||
3844 | int pci, cl; | |||
3845 | ||||
3846 | fprintf (sched_dump, ";; Pressure summary:"); | |||
3847 | for (pci = 0; pci < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); pci++) | |||
3848 | { | |||
3849 | cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[pci]; | |||
3850 | fprintf (sched_dump, " %s:%d", reg_class_names[cl], | |||
3851 | model_before_pressure.limits[pci].pressure); | |||
3852 | } | |||
3853 | fprintf (sched_dump, "\n\n"); | |||
3854 | } | |||
3855 | ||||
3856 | /* Initialize the SCHED_PRESSURE_MODEL information for the current | |||
3857 | scheduling region. */ | |||
3858 | ||||
3859 | static void | |||
3860 | model_start_schedule (basic_block bb) | |||
3861 | { | |||
3862 | model_next_priority = 1; | |||
3863 | model_schedule.create (sched_max_luid); | |||
3864 | model_insns = XCNEWVEC (struct model_insn_info, sched_max_luid)((struct model_insn_info *) xcalloc ((sched_max_luid), sizeof (struct model_insn_info))); | |||
3865 | ||||
3866 | gcc_assert (bb == BLOCK_FOR_INSN (NEXT_INSN (current_sched_info->prev_head)))((void)(!(bb == BLOCK_FOR_INSN (NEXT_INSN (current_sched_info ->prev_head))) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 3866, __FUNCTION__), 0 : 0)); | |||
3867 | initiate_reg_pressure_info (df_get_live_in (bb)); | |||
3868 | ||||
3869 | model_analyze_insns (); | |||
3870 | model_init_pressure_group (&model_before_pressure); | |||
3871 | while (model_worklist) | |||
3872 | model_choose_insn (); | |||
3873 | gcc_assert (model_num_insns == (int) model_schedule.length ())((void)(!(model_num_insns == (int) model_schedule.length ()) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 3873, __FUNCTION__), 0 : 0)); | |||
3874 | if (sched_verbose >= 2) | |||
3875 | fprintf (sched_dump, "\n"); | |||
3876 | ||||
3877 | model_record_final_pressures (&model_before_pressure); | |||
3878 | model_reset_queue_indices (); | |||
3879 | ||||
3880 | XDELETEVEC (model_insns)free ((void*) (model_insns)); | |||
3881 | ||||
3882 | model_curr_point = 0; | |||
3883 | initiate_reg_pressure_info (df_get_live_in (bb)); | |||
3884 | if (sched_verbose >= 1) | |||
3885 | model_dump_pressure_summary (); | |||
3886 | } | |||
3887 | ||||
3888 | /* Free the information associated with GROUP. */ | |||
3889 | ||||
3890 | static void | |||
3891 | model_finalize_pressure_group (struct model_pressure_group *group) | |||
3892 | { | |||
3893 | XDELETEVEC (group->model)free ((void*) (group->model)); | |||
3894 | } | |||
3895 | ||||
3896 | /* Free the information created by model_start_schedule. */ | |||
3897 | ||||
3898 | static void | |||
3899 | model_end_schedule (void) | |||
3900 | { | |||
3901 | model_finalize_pressure_group (&model_before_pressure); | |||
3902 | model_schedule.release (); | |||
3903 | } | |||
3904 | ||||
3905 | /* Prepare reg pressure scheduling for basic block BB. */ | |||
3906 | static void | |||
3907 | sched_pressure_start_bb (basic_block bb) | |||
3908 | { | |||
3909 | /* Set the number of available registers for each class taking into account | |||
3910 | relative probability of current basic block versus function prologue and | |||
3911 | epilogue. | |||
3912 | * If the basic block executes much more often than the prologue/epilogue | |||
3913 | (e.g., inside a hot loop), then cost of spill in the prologue is close to | |||
3914 | nil, so the effective number of available registers is | |||
3915 | (ira_class_hard_regs_num[cl] - fixed_regs_num[cl] - 0). | |||
3916 | * If the basic block executes as often as the prologue/epilogue, | |||
3917 | then spill in the block is as costly as in the prologue, so the effective | |||
3918 | number of available registers is | |||
3919 | (ira_class_hard_regs_num[cl] - fixed_regs_num[cl] | |||
3920 | - call_saved_regs_num[cl]). | |||
3921 | Note that all-else-equal, we prefer to spill in the prologue, since that | |||
3922 | allows "extra" registers for other basic blocks of the function. | |||
3923 | * If the basic block is on the cold path of the function and executes | |||
3924 | rarely, then we should always prefer to spill in the block, rather than | |||
3925 | in the prologue/epilogue. The effective number of available register is | |||
3926 | (ira_class_hard_regs_num[cl] - fixed_regs_num[cl] | |||
3927 | - call_saved_regs_num[cl]). */ | |||
3928 | { | |||
3929 | int i; | |||
3930 | int entry_freq = ENTRY_BLOCK_PTR_FOR_FN (cfun)(((cfun + 0))->cfg->x_entry_block_ptr)->count.to_frequency (cfun(cfun + 0)); | |||
3931 | int bb_freq = bb->count.to_frequency (cfun(cfun + 0)); | |||
3932 | ||||
3933 | if (bb_freq == 0) | |||
3934 | { | |||
3935 | if (entry_freq == 0) | |||
3936 | entry_freq = bb_freq = 1; | |||
3937 | } | |||
3938 | if (bb_freq < entry_freq) | |||
3939 | bb_freq = entry_freq; | |||
3940 | ||||
3941 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); ++i) | |||
3942 | { | |||
3943 | enum reg_class cl = ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]; | |||
3944 | sched_class_regs_num[cl] = ira_class_hard_regs_num(this_target_ira->x_ira_class_hard_regs_num)[cl] | |||
3945 | - fixed_regs_num[cl]; | |||
3946 | sched_class_regs_num[cl] | |||
3947 | -= (call_saved_regs_num[cl] * entry_freq) / bb_freq; | |||
3948 | } | |||
3949 | } | |||
3950 | ||||
3951 | if (sched_pressure == SCHED_PRESSURE_MODEL) | |||
3952 | model_start_schedule (bb); | |||
3953 | } | |||
3954 | ||||
3955 | /* A structure that holds local state for the loop in schedule_block. */ | |||
3956 | struct sched_block_state | |||
3957 | { | |||
3958 | /* True if no real insns have been scheduled in the current cycle. */ | |||
3959 | bool first_cycle_insn_p; | |||
3960 | /* True if a shadow insn has been scheduled in the current cycle, which | |||
3961 | means that no more normal insns can be issued. */ | |||
3962 | bool shadows_only_p; | |||
3963 | /* True if we're winding down a modulo schedule, which means that we only | |||
3964 | issue insns with INSN_EXACT_TICK set. */ | |||
3965 | bool modulo_epilogue; | |||
3966 | /* Initialized with the machine's issue rate every cycle, and updated | |||
3967 | by calls to the variable_issue hook. */ | |||
3968 | int can_issue_more; | |||
3969 | }; | |||
3970 | ||||
3971 | /* INSN is the "currently executing insn". Launch each insn which was | |||
3972 | waiting on INSN. READY is the ready list which contains the insns | |||
3973 | that are ready to fire. CLOCK is the current cycle. The function | |||
3974 | returns necessary cycle advance after issuing the insn (it is not | |||
3975 | zero for insns in a schedule group). */ | |||
3976 | ||||
3977 | static int | |||
3978 | schedule_insn (rtx_insn *insn) | |||
3979 | { | |||
3980 | sd_iterator_def sd_it; | |||
3981 | dep_t dep; | |||
3982 | int i; | |||
3983 | int advance = 0; | |||
3984 | ||||
3985 | if (sched_verbose >= 1) | |||
| ||||
3986 | { | |||
3987 | struct reg_pressure_data *pressure_info; | |||
3988 | fprintf (sched_dump, ";;\t%3i--> %s %-40s:", | |||
3989 | clock_var, (*current_sched_info->print_insn) (insn, 1), | |||
3990 | str_pattern_slim (PATTERN (insn))); | |||
3991 | ||||
3992 | if (recog_memoized (insn) < 0) | |||
3993 | fprintf (sched_dump, "nothing"); | |||
3994 | else | |||
3995 | print_reservation (sched_dump, insn); | |||
3996 | pressure_info = INSN_REG_PRESSURE (insn)((&h_i_d[INSN_UID (insn)])->reg_pressure); | |||
3997 | if (pressure_info != NULLnullptr) | |||
3998 | { | |||
3999 | fputc (':', sched_dump); | |||
4000 | for (i = 0; i < ira_pressure_classes_num(this_target_ira->x_ira_pressure_classes_num); i++) | |||
4001 | fprintf (sched_dump, "%s%s%+d(%d)", | |||
4002 | scheduled_insns.length () > 1 | |||
4003 | && INSN_LUID (insn)(sched_luids[INSN_UID (insn)]) | |||
4004 | < INSN_LUID (scheduled_insns[scheduled_insns.length () - 2])(sched_luids[INSN_UID (scheduled_insns[scheduled_insns.length () - 2])]) ? "@" : "", | |||
4005 | reg_class_names[ira_pressure_classes(this_target_ira->x_ira_pressure_classes)[i]], | |||
4006 | pressure_info[i].set_increase, pressure_info[i].change); | |||
4007 | } | |||
4008 | if (sched_pressure == SCHED_PRESSURE_MODEL | |||
4009 | && model_curr_point < model_num_insns | |||
4010 | && model_index (insn) == model_curr_point) | |||
4011 | fprintf (sched_dump, ":model %d", model_curr_point); | |||
4012 | fputc ('\n', sched_dump); | |||
4013 | } | |||
4014 | ||||
4015 | if (sched_pressure == SCHED_PRESSURE_WEIGHTED && !DEBUG_INSN_P (insn)(((enum rtx_code) (insn)->code) == DEBUG_INSN)) | |||
4016 | update_reg_and_insn_max_reg_pressure (insn); | |||
4017 | ||||
4018 | /* Scheduling instruction should have all its dependencies resolved and | |||
4019 | should have been removed from the ready list. */ | |||
4020 | gcc_assert (sd_lists_empty_p (insn, SD_LIST_HARD_BACK))((void)(!(sd_lists_empty_p (insn, (1))) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4020, __FUNCTION__), 0 : 0)); | |||
4021 | ||||
4022 | /* Reset debug insns invalidated by moving this insn. */ | |||
4023 | if (MAY_HAVE_DEBUG_BIND_INSNSglobal_options.x_flag_var_tracking_assignments && !DEBUG_INSN_P (insn)(((enum rtx_code) (insn)->code) == DEBUG_INSN)) | |||
4024 | for (sd_it = sd_iterator_start (insn, SD_LIST_BACK((1) | (2))); | |||
4025 | sd_iterator_cond (&sd_it, &dep);) | |||
4026 | { | |||
4027 | rtx_insn *dbg = DEP_PRO (dep)((dep)->pro); | |||
4028 | struct reg_use_data *use, *next; | |||
4029 | ||||
4030 | if (DEP_STATUS (dep)((dep)->status) & DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1)) | |||
4031 | { | |||
4032 | sd_iterator_next (&sd_it); | |||
4033 | continue; | |||
4034 | } | |||
4035 | ||||
4036 | gcc_assert (DEBUG_BIND_INSN_P (dbg))((void)(!(((((enum rtx_code) (dbg)->code) == DEBUG_INSN) && (((enum rtx_code) (PATTERN (dbg))->code) == VAR_LOCATION) )) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4036, __FUNCTION__), 0 : 0)); | |||
4037 | ||||
4038 | if (sched_verbose >= 6) | |||
4039 | fprintf (sched_dump, ";;\t\tresetting: debug insn %d\n", | |||
4040 | INSN_UID (dbg)); | |||
4041 | ||||
4042 | /* ??? Rather than resetting the debug insn, we might be able | |||
4043 | to emit a debug temp before the just-scheduled insn, but | |||
4044 | this would involve checking that the expression at the | |||
4045 | point of the debug insn is equivalent to the expression | |||
4046 | before the just-scheduled insn. They might not be: the | |||
4047 | expression in the debug insn may depend on other insns not | |||
4048 | yet scheduled that set MEMs, REGs or even other debug | |||
4049 | insns. It's not clear that attempting to preserve debug | |||
4050 | information in these cases is worth the effort, given how | |||
4051 | uncommon these resets are and the likelihood that the debug | |||
4052 | temps introduced won't survive the schedule change. */ | |||
4053 | INSN_VAR_LOCATION_LOC (dbg)((((((__extension__ ({ __typeof (PATTERN (dbg)) const _rtx = ( PATTERN (dbg)); if (((enum rtx_code) (_rtx)->code) != VAR_LOCATION ) rtl_check_failed_flag ("INSN_VAR_LOCATION", _rtx, "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4053, __FUNCTION__); _rtx; }))))->u.fld[1]).rt_rtx)) = gen_rtx_UNKNOWN_VAR_LOC ()(gen_rtx_fmt_e_stat ((CLOBBER), ((((void) 0, E_VOIDmode))), ( ((const_int_rtx[64]))) )); | |||
4054 | df_insn_rescan (dbg); | |||
4055 | ||||
4056 | /* Unknown location doesn't use any registers. */ | |||
4057 | for (use = INSN_REG_USE_LIST (dbg)((&h_i_d[INSN_UID (dbg)])->reg_use_list); use != NULLnullptr; use = next) | |||
4058 | { | |||
4059 | struct reg_use_data *prev = use; | |||
4060 | ||||
4061 | /* Remove use from the cyclic next_regno_use chain first. */ | |||
4062 | while (prev->next_regno_use != use) | |||
4063 | prev = prev->next_regno_use; | |||
4064 | prev->next_regno_use = use->next_regno_use; | |||
4065 | next = use->next_insn_use; | |||
4066 | free (use); | |||
4067 | } | |||
4068 | INSN_REG_USE_LIST (dbg)((&h_i_d[INSN_UID (dbg)])->reg_use_list) = NULLnullptr; | |||
4069 | ||||
4070 | /* We delete rather than resolve these deps, otherwise we | |||
4071 | crash in sched_free_deps(), because forward deps are | |||
4072 | expected to be released before backward deps. */ | |||
4073 | sd_delete_dep (sd_it); | |||
4074 | } | |||
4075 | ||||
4076 | gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE)((void)(!(((&h_i_d[INSN_UID (insn)])->queue_index) == ( -2)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4076, __FUNCTION__), 0 : 0)); | |||
4077 | QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index) = QUEUE_SCHEDULED(-3); | |||
4078 | ||||
4079 | if (sched_pressure == SCHED_PRESSURE_MODEL | |||
4080 | && model_curr_point < model_num_insns | |||
4081 | && NONDEBUG_INSN_P (insn)((((enum rtx_code) (insn)->code) == INSN) || (((enum rtx_code ) (insn)->code) == JUMP_INSN) || (((enum rtx_code) (insn)-> code) == CALL_INSN))) | |||
4082 | { | |||
4083 | if (model_index (insn) == model_curr_point) | |||
4084 | do | |||
4085 | model_curr_point++; | |||
4086 | while (model_curr_point < model_num_insns | |||
4087 | && (QUEUE_INDEX (MODEL_INSN (model_curr_point))((&h_i_d[INSN_UID ((model_schedule[model_curr_point]))])-> queue_index) | |||
4088 | == QUEUE_SCHEDULED(-3))); | |||
4089 | else | |||
4090 | model_recompute (insn); | |||
4091 | model_update_limit_points (); | |||
4092 | update_register_pressure (insn); | |||
4093 | if (sched_verbose >= 2) | |||
4094 | print_curr_reg_pressure (); | |||
4095 | } | |||
4096 | ||||
4097 | gcc_assert (INSN_TICK (insn) >= MIN_TICK)((void)(!(((&h_i_d[INSN_UID (insn)])->tick) >= (-max_insn_queue_index )) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4097, __FUNCTION__), 0 : 0)); | |||
4098 | if (INSN_TICK (insn)((&h_i_d[INSN_UID (insn)])->tick) > clock_var) | |||
4099 | /* INSN has been prematurely moved from the queue to the ready list. | |||
4100 | This is possible only if following flags are set. */ | |||
4101 | gcc_assert (flag_sched_stalled_insns || sched_fusion)((void)(!(global_options.x_flag_sched_stalled_insns || sched_fusion ) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4101, __FUNCTION__), 0 : 0)); | |||
4102 | ||||
4103 | /* ??? Probably, if INSN is scheduled prematurely, we should leave | |||
4104 | INSN_TICK untouched. This is a machine-dependent issue, actually. */ | |||
4105 | INSN_TICK (insn)((&h_i_d[INSN_UID (insn)])->tick) = clock_var; | |||
4106 | ||||
4107 | check_clobbered_conditions (insn); | |||
4108 | ||||
4109 | /* Update dependent instructions. First, see if by scheduling this insn | |||
4110 | now we broke a dependence in a way that requires us to change another | |||
4111 | insn. */ | |||
4112 | for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK(2)); | |||
4113 | sd_iterator_cond (&sd_it, &dep); sd_iterator_next (&sd_it)) | |||
4114 | { | |||
4115 | struct dep_replacement *desc = DEP_REPLACE (dep)((dep)->replace); | |||
4116 | rtx_insn *pro = DEP_PRO (dep)((dep)->pro); | |||
4117 | if (QUEUE_INDEX (pro)((&h_i_d[INSN_UID (pro)])->queue_index) != QUEUE_SCHEDULED(-3) | |||
4118 | && desc != NULLnullptr && desc->insn == pro) | |||
4119 | apply_replacement (dep, false); | |||
4120 | } | |||
4121 | ||||
4122 | /* Go through and resolve forward dependencies. */ | |||
4123 | for (sd_it = sd_iterator_start (insn, SD_LIST_FORW(4)); | |||
4124 | sd_iterator_cond (&sd_it, &dep);) | |||
4125 | { | |||
4126 | rtx_insn *next = DEP_CON (dep)((dep)->con); | |||
4127 | bool cancelled = (DEP_STATUS (dep)((dep)->status) & DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1)) != 0; | |||
4128 | ||||
4129 | /* Resolve the dependence between INSN and NEXT. | |||
4130 | sd_resolve_dep () moves current dep to another list thus | |||
4131 | advancing the iterator. */ | |||
4132 | sd_resolve_dep (sd_it); | |||
4133 | ||||
4134 | if (cancelled) | |||
4135 | { | |||
4136 | if (must_restore_pattern_p (next, dep)) | |||
4137 | restore_pattern (dep, false); | |||
4138 | continue; | |||
4139 | } | |||
4140 | ||||
4141 | /* Don't bother trying to mark next as ready if insn is a debug | |||
4142 | insn. If insn is the last hard dependency, it will have | |||
4143 | already been discounted. */ | |||
4144 | if (DEBUG_INSN_P (insn)(((enum rtx_code) (insn)->code) == DEBUG_INSN) && !DEBUG_INSN_P (next)(((enum rtx_code) (next)->code) == DEBUG_INSN)) | |||
4145 | continue; | |||
4146 | ||||
4147 | if (!IS_SPECULATION_BRANCHY_CHECK_P (insn)(((&h_i_d[INSN_UID (insn)])->recovery_block) != nullptr && ((&h_i_d[INSN_UID (insn)])->recovery_block ) != (((cfun + 0))->cfg->x_exit_block_ptr))) | |||
4148 | { | |||
4149 | int effective_cost; | |||
4150 | ||||
4151 | effective_cost = try_ready (next); | |||
4152 | ||||
4153 | if (effective_cost >= 0 | |||
4154 | && SCHED_GROUP_P (next)(__extension__ ({ __typeof ((next)) const _rtx = ((next)); if (((enum rtx_code) (_rtx)->code) != DEBUG_INSN && ( (enum rtx_code) (_rtx)->code) != INSN && ((enum rtx_code ) (_rtx)->code) != JUMP_INSN && ((enum rtx_code) ( _rtx)->code) != CALL_INSN) rtl_check_failed_flag ("SCHED_GROUP_P" , _rtx, "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4154, __FUNCTION__); _rtx; })->in_struct) | |||
4155 | && advance < effective_cost) | |||
4156 | advance = effective_cost; | |||
4157 | } | |||
4158 | else | |||
4159 | /* Check always has only one forward dependence (to the first insn in | |||
4160 | the recovery block), therefore, this will be executed only once. */ | |||
4161 | { | |||
4162 | gcc_assert (sd_lists_empty_p (insn, SD_LIST_FORW))((void)(!(sd_lists_empty_p (insn, (4))) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4162, __FUNCTION__), 0 : 0)); | |||
4163 | fix_recovery_deps (RECOVERY_BLOCK (insn)((&h_i_d[INSN_UID (insn)])->recovery_block)); | |||
4164 | } | |||
4165 | } | |||
4166 | ||||
4167 | /* Annotate the instruction with issue information -- TImode | |||
4168 | indicates that the instruction is expected not to be able | |||
4169 | to issue on the same cycle as the previous insn. A machine | |||
4170 | may use this information to decide how the instruction should | |||
4171 | be aligned. */ | |||
4172 | if (issue_rate > 1 | |||
4173 | && GET_CODE (PATTERN (insn))((enum rtx_code) (PATTERN (insn))->code) != USE | |||
4174 | && GET_CODE (PATTERN (insn))((enum rtx_code) (PATTERN (insn))->code) != CLOBBER | |||
4175 | && !DEBUG_INSN_P (insn)(((enum rtx_code) (insn)->code) == DEBUG_INSN)) | |||
4176 | { | |||
4177 | if (reload_completed) | |||
4178 | PUT_MODE (insn, clock_var > last_clock_var ? TImode(scalar_int_mode ((scalar_int_mode::from_int) E_TImode)) : VOIDmode((void) 0, E_VOIDmode)); | |||
4179 | last_clock_var = clock_var; | |||
4180 | } | |||
4181 | ||||
4182 | if (nonscheduled_insns_begin != NULL_RTX(rtx) 0) | |||
4183 | /* Indicate to debug counters that INSN is scheduled. */ | |||
4184 | nonscheduled_insns_begin = insn; | |||
4185 | ||||
4186 | return advance; | |||
4187 | } | |||
4188 | ||||
4189 | /* Functions for handling of notes. */ | |||
4190 | ||||
4191 | /* Add note list that ends on FROM_END to the end of TO_ENDP. */ | |||
4192 | void | |||
4193 | concat_note_lists (rtx_insn *from_end, rtx_insn **to_endp) | |||
4194 | { | |||
4195 | rtx_insn *from_start; | |||
4196 | ||||
4197 | /* It's easy when have nothing to concat. */ | |||
4198 | if (from_end == NULLnullptr) | |||
4199 | return; | |||
4200 | ||||
4201 | /* It's also easy when destination is empty. */ | |||
4202 | if (*to_endp == NULLnullptr) | |||
4203 | { | |||
4204 | *to_endp = from_end; | |||
4205 | return; | |||
4206 | } | |||
4207 | ||||
4208 | from_start = from_end; | |||
4209 | while (PREV_INSN (from_start) != NULLnullptr) | |||
4210 | from_start = PREV_INSN (from_start); | |||
4211 | ||||
4212 | SET_PREV_INSN (from_start) = *to_endp; | |||
4213 | SET_NEXT_INSN (*to_endp) = from_start; | |||
4214 | *to_endp = from_end; | |||
4215 | } | |||
4216 | ||||
4217 | /* Delete notes between HEAD and TAIL and put them in the chain | |||
4218 | of notes ended by NOTE_LIST. */ | |||
4219 | void | |||
4220 | remove_notes (rtx_insn *head, rtx_insn *tail) | |||
4221 | { | |||
4222 | rtx_insn *next_tail, *insn, *next; | |||
4223 | ||||
4224 | note_list = 0; | |||
4225 | if (head == tail && !INSN_P (head)(((((enum rtx_code) (head)->code) == INSN) || (((enum rtx_code ) (head)->code) == JUMP_INSN) || (((enum rtx_code) (head)-> code) == CALL_INSN)) || (((enum rtx_code) (head)->code) == DEBUG_INSN))) | |||
4226 | return; | |||
4227 | ||||
4228 | next_tail = NEXT_INSN (tail); | |||
4229 | for (insn = head; insn != next_tail; insn = next) | |||
4230 | { | |||
4231 | next = NEXT_INSN (insn); | |||
4232 | if (!NOTE_P (insn)(((enum rtx_code) (insn)->code) == NOTE)) | |||
4233 | continue; | |||
4234 | ||||
4235 | switch (NOTE_KIND (insn)(((insn)->u.fld[4]).rt_int)) | |||
4236 | { | |||
4237 | case NOTE_INSN_BASIC_BLOCK: | |||
4238 | continue; | |||
4239 | ||||
4240 | case NOTE_INSN_EPILOGUE_BEG: | |||
4241 | if (insn != tail) | |||
4242 | { | |||
4243 | remove_insn (insn); | |||
4244 | /* If an insn was split just before the EPILOGUE_BEG note and | |||
4245 | that split created new basic blocks, we could have a | |||
4246 | BASIC_BLOCK note here. Safely advance over it in that case | |||
4247 | and assert that we land on a real insn. */ | |||
4248 | if (NOTE_P (next)(((enum rtx_code) (next)->code) == NOTE) | |||
4249 | && NOTE_KIND (next)(((next)->u.fld[4]).rt_int) == NOTE_INSN_BASIC_BLOCK | |||
4250 | && next != next_tail) | |||
4251 | next = NEXT_INSN (next); | |||
4252 | gcc_assert (INSN_P (next))((void)(!((((((enum rtx_code) (next)->code) == INSN) || (( (enum rtx_code) (next)->code) == JUMP_INSN) || (((enum rtx_code ) (next)->code) == CALL_INSN)) || (((enum rtx_code) (next) ->code) == DEBUG_INSN))) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4252, __FUNCTION__), 0 : 0)); | |||
4253 | add_reg_note (next, REG_SAVE_NOTE, | |||
4254 | GEN_INT (NOTE_INSN_EPILOGUE_BEG)gen_rtx_CONST_INT (((void) 0, E_VOIDmode), (NOTE_INSN_EPILOGUE_BEG ))); | |||
4255 | break; | |||
4256 | } | |||
4257 | /* FALLTHRU */ | |||
4258 | ||||
4259 | default: | |||
4260 | remove_insn (insn); | |||
4261 | ||||
4262 | /* Add the note to list that ends at NOTE_LIST. */ | |||
4263 | SET_PREV_INSN (insn) = note_list; | |||
4264 | SET_NEXT_INSN (insn) = NULL_RTX(rtx) 0; | |||
4265 | if (note_list) | |||
4266 | SET_NEXT_INSN (note_list) = insn; | |||
4267 | note_list = insn; | |||
4268 | break; | |||
4269 | } | |||
4270 | ||||
4271 | gcc_assert ((sel_sched_p () || insn != tail) && insn != head)((void)(!((sel_sched_p () || insn != tail) && insn != head) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4271, __FUNCTION__), 0 : 0)); | |||
4272 | } | |||
4273 | } | |||
4274 | ||||
4275 | /* A structure to record enough data to allow us to backtrack the scheduler to | |||
4276 | a previous state. */ | |||
4277 | struct haifa_saved_data | |||
4278 | { | |||
4279 | /* Next entry on the list. */ | |||
4280 | struct haifa_saved_data *next; | |||
4281 | ||||
4282 | /* Backtracking is associated with scheduling insns that have delay slots. | |||
4283 | DELAY_PAIR points to the structure that contains the insns involved, and | |||
4284 | the number of cycles between them. */ | |||
4285 | struct delay_pair *delay_pair; | |||
4286 | ||||
4287 | /* Data used by the frontend (e.g. sched-ebb or sched-rgn). */ | |||
4288 | void *fe_saved_data; | |||
4289 | /* Data used by the backend. */ | |||
4290 | void *be_saved_data; | |||
4291 | ||||
4292 | /* Copies of global state. */ | |||
4293 | int clock_var, last_clock_var; | |||
4294 | struct ready_list ready; | |||
4295 | state_t curr_state; | |||
4296 | ||||
4297 | rtx_insn *last_scheduled_insn; | |||
4298 | rtx_insn *last_nondebug_scheduled_insn; | |||
4299 | rtx_insn *nonscheduled_insns_begin; | |||
4300 | int cycle_issued_insns; | |||
4301 | ||||
4302 | /* Copies of state used in the inner loop of schedule_block. */ | |||
4303 | struct sched_block_state sched_block; | |||
4304 | ||||
4305 | /* We don't need to save q_ptr, as its value is arbitrary and we can set it | |||
4306 | to 0 when restoring. */ | |||
4307 | int q_size; | |||
4308 | rtx_insn_list **insn_queue; | |||
4309 | ||||
4310 | /* Describe pattern replacements that occurred since this backtrack point | |||
4311 | was queued. */ | |||
4312 | vec<dep_t> replacement_deps; | |||
4313 | vec<int> replace_apply; | |||
4314 | ||||
4315 | /* A copy of the next-cycle replacement vectors at the time of the backtrack | |||
4316 | point. */ | |||
4317 | vec<dep_t> next_cycle_deps; | |||
4318 | vec<int> next_cycle_apply; | |||
4319 | }; | |||
4320 | ||||
4321 | /* A record, in reverse order, of all scheduled insns which have delay slots | |||
4322 | and may require backtracking. */ | |||
4323 | static struct haifa_saved_data *backtrack_queue; | |||
4324 | ||||
4325 | /* For every dependency of INSN, set the FEEDS_BACKTRACK_INSN bit according | |||
4326 | to SET_P. */ | |||
4327 | static void | |||
4328 | mark_backtrack_feeds (rtx_insn *insn, int set_p) | |||
4329 | { | |||
4330 | sd_iterator_def sd_it; | |||
4331 | dep_t dep; | |||
4332 | FOR_EACH_DEP (insn, SD_LIST_HARD_BACK, sd_it, dep)for ((sd_it) = sd_iterator_start ((insn), ((1))); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
4333 | { | |||
4334 | FEEDS_BACKTRACK_INSN (DEP_PRO (dep))((&h_i_d[INSN_UID (((dep)->pro))])->feeds_backtrack_insn ) = set_p; | |||
4335 | } | |||
4336 | } | |||
4337 | ||||
4338 | /* Save the current scheduler state so that we can backtrack to it | |||
4339 | later if necessary. PAIR gives the insns that make it necessary to | |||
4340 | save this point. SCHED_BLOCK is the local state of schedule_block | |||
4341 | that need to be saved. */ | |||
4342 | static void | |||
4343 | save_backtrack_point (struct delay_pair *pair, | |||
4344 | struct sched_block_state sched_block) | |||
4345 | { | |||
4346 | int i; | |||
4347 | struct haifa_saved_data *save = XNEW (struct haifa_saved_data)((struct haifa_saved_data *) xmalloc (sizeof (struct haifa_saved_data ))); | |||
4348 | ||||
4349 | save->curr_state = xmalloc (dfa_state_size); | |||
4350 | memcpy (save->curr_state, curr_state, dfa_state_size); | |||
4351 | ||||
4352 | save->ready.first = ready.first; | |||
4353 | save->ready.n_ready = ready.n_ready; | |||
4354 | save->ready.n_debug = ready.n_debug; | |||
4355 | save->ready.veclen = ready.veclen; | |||
4356 | save->ready.vec = XNEWVEC (rtx_insn *, ready.veclen)((rtx_insn * *) xmalloc (sizeof (rtx_insn *) * (ready.veclen) )); | |||
4357 | memcpy (save->ready.vec, ready.vec, ready.veclen * sizeof (rtx)); | |||
4358 | ||||
4359 | save->insn_queue = XNEWVEC (rtx_insn_list *, max_insn_queue_index + 1)((rtx_insn_list * *) xmalloc (sizeof (rtx_insn_list *) * (max_insn_queue_index + 1))); | |||
4360 | save->q_size = q_size; | |||
4361 | for (i = 0; i <= max_insn_queue_index; i++) | |||
4362 | { | |||
4363 | int q = NEXT_Q_AFTER (q_ptr, i)(((q_ptr)+i) & max_insn_queue_index); | |||
4364 | save->insn_queue[i] = copy_INSN_LIST (insn_queue[q]); | |||
4365 | } | |||
4366 | ||||
4367 | save->clock_var = clock_var; | |||
4368 | save->last_clock_var = last_clock_var; | |||
4369 | save->cycle_issued_insns = cycle_issued_insns; | |||
4370 | save->last_scheduled_insn = last_scheduled_insn; | |||
4371 | save->last_nondebug_scheduled_insn = last_nondebug_scheduled_insn; | |||
4372 | save->nonscheduled_insns_begin = nonscheduled_insns_begin; | |||
4373 | ||||
4374 | save->sched_block = sched_block; | |||
4375 | ||||
4376 | save->replacement_deps.create (0); | |||
4377 | save->replace_apply.create (0); | |||
4378 | save->next_cycle_deps = next_cycle_replace_deps.copy (); | |||
4379 | save->next_cycle_apply = next_cycle_apply.copy (); | |||
4380 | ||||
4381 | if (current_sched_info->save_state) | |||
4382 | save->fe_saved_data = (*current_sched_info->save_state) (); | |||
4383 | ||||
4384 | if (targetm.sched.alloc_sched_context) | |||
4385 | { | |||
4386 | save->be_saved_data = targetm.sched.alloc_sched_context (); | |||
4387 | targetm.sched.init_sched_context (save->be_saved_data, false); | |||
4388 | } | |||
4389 | else | |||
4390 | save->be_saved_data = NULLnullptr; | |||
4391 | ||||
4392 | save->delay_pair = pair; | |||
4393 | ||||
4394 | save->next = backtrack_queue; | |||
4395 | backtrack_queue = save; | |||
4396 | ||||
4397 | while (pair) | |||
4398 | { | |||
4399 | mark_backtrack_feeds (pair->i2, 1); | |||
4400 | INSN_TICK (pair->i2)((&h_i_d[INSN_UID (pair->i2)])->tick) = INVALID_TICK(-(max_insn_queue_index + 1)); | |||
4401 | INSN_EXACT_TICK (pair->i2)((&h_i_d[INSN_UID (pair->i2)])->exact_tick) = clock_var + pair_delay (pair); | |||
4402 | SHADOW_P (pair->i2)((&h_i_d[INSN_UID (pair->i2)])->shadow_p) = pair->stages == 0; | |||
4403 | pair = pair->next_same_i1; | |||
4404 | } | |||
4405 | } | |||
4406 | ||||
4407 | /* Walk the ready list and all queues. If any insns have unresolved backwards | |||
4408 | dependencies, these must be cancelled deps, broken by predication. Set or | |||
4409 | clear (depending on SET) the DEP_CANCELLED bit in DEP_STATUS. */ | |||
4410 | ||||
4411 | static void | |||
4412 | toggle_cancelled_flags (bool set) | |||
4413 | { | |||
4414 | int i; | |||
4415 | sd_iterator_def sd_it; | |||
4416 | dep_t dep; | |||
4417 | ||||
4418 | if (ready.n_ready > 0) | |||
4419 | { | |||
4420 | rtx_insn **first = ready_lastpos (&ready); | |||
4421 | for (i = 0; i < ready.n_ready; i++) | |||
4422 | FOR_EACH_DEP (first[i], SD_LIST_BACK, sd_it, dep)for ((sd_it) = sd_iterator_start ((first[i]), (((1) | (2)))); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
4423 | if (!DEBUG_INSN_P (DEP_PRO (dep))(((enum rtx_code) (((dep)->pro))->code) == DEBUG_INSN)) | |||
4424 | { | |||
4425 | if (set) | |||
4426 | DEP_STATUS (dep)((dep)->status) |= DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1); | |||
4427 | else | |||
4428 | DEP_STATUS (dep)((dep)->status) &= ~DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1); | |||
4429 | } | |||
4430 | } | |||
4431 | for (i = 0; i <= max_insn_queue_index; i++) | |||
4432 | { | |||
4433 | int q = NEXT_Q_AFTER (q_ptr, i)(((q_ptr)+i) & max_insn_queue_index); | |||
4434 | rtx_insn_list *link; | |||
4435 | for (link = insn_queue[q]; link; link = link->next ()) | |||
4436 | { | |||
4437 | rtx_insn *insn = link->insn (); | |||
4438 | FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)for ((sd_it) = sd_iterator_start ((insn), (((1) | (2)))); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
4439 | if (!DEBUG_INSN_P (DEP_PRO (dep))(((enum rtx_code) (((dep)->pro))->code) == DEBUG_INSN)) | |||
4440 | { | |||
4441 | if (set) | |||
4442 | DEP_STATUS (dep)((dep)->status) |= DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1); | |||
4443 | else | |||
4444 | DEP_STATUS (dep)((dep)->status) &= ~DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1); | |||
4445 | } | |||
4446 | } | |||
4447 | } | |||
4448 | } | |||
4449 | ||||
4450 | /* Undo the replacements that have occurred after backtrack point SAVE | |||
4451 | was placed. */ | |||
4452 | static void | |||
4453 | undo_replacements_for_backtrack (struct haifa_saved_data *save) | |||
4454 | { | |||
4455 | while (!save->replacement_deps.is_empty ()) | |||
4456 | { | |||
4457 | dep_t dep = save->replacement_deps.pop (); | |||
4458 | int apply_p = save->replace_apply.pop (); | |||
4459 | ||||
4460 | if (apply_p) | |||
4461 | restore_pattern (dep, true); | |||
4462 | else | |||
4463 | apply_replacement (dep, true); | |||
4464 | } | |||
4465 | save->replacement_deps.release (); | |||
4466 | save->replace_apply.release (); | |||
4467 | } | |||
4468 | ||||
4469 | /* Pop entries from the SCHEDULED_INSNS vector up to and including INSN. | |||
4470 | Restore their dependencies to an unresolved state, and mark them as | |||
4471 | queued nowhere. */ | |||
4472 | ||||
4473 | static void | |||
4474 | unschedule_insns_until (rtx_insn *insn) | |||
4475 | { | |||
4476 | auto_vec<rtx_insn *> recompute_vec; | |||
4477 | ||||
4478 | /* Make two passes over the insns to be unscheduled. First, we clear out | |||
4479 | dependencies and other trivial bookkeeping. */ | |||
4480 | for (;;) | |||
4481 | { | |||
4482 | rtx_insn *last; | |||
4483 | sd_iterator_def sd_it; | |||
4484 | dep_t dep; | |||
4485 | ||||
4486 | last = scheduled_insns.pop (); | |||
4487 | ||||
4488 | /* This will be changed by restore_backtrack_point if the insn is in | |||
4489 | any queue. */ | |||
4490 | QUEUE_INDEX (last)((&h_i_d[INSN_UID (last)])->queue_index) = QUEUE_NOWHERE(-2); | |||
4491 | if (last != insn) | |||
4492 | INSN_TICK (last)((&h_i_d[INSN_UID (last)])->tick) = INVALID_TICK(-(max_insn_queue_index + 1)); | |||
4493 | ||||
4494 | if (modulo_ii > 0 && INSN_UID (last) < modulo_iter0_max_uid) | |||
4495 | modulo_insns_scheduled--; | |||
4496 | ||||
4497 | for (sd_it = sd_iterator_start (last, SD_LIST_RES_FORW(16)); | |||
4498 | sd_iterator_cond (&sd_it, &dep);) | |||
4499 | { | |||
4500 | rtx_insn *con = DEP_CON (dep)((dep)->con); | |||
4501 | sd_unresolve_dep (sd_it); | |||
4502 | if (!MUST_RECOMPUTE_SPEC_P (con)((&h_i_d[INSN_UID (con)])->must_recompute_spec)) | |||
4503 | { | |||
4504 | MUST_RECOMPUTE_SPEC_P (con)((&h_i_d[INSN_UID (con)])->must_recompute_spec) = 1; | |||
4505 | recompute_vec.safe_push (con); | |||
4506 | } | |||
4507 | } | |||
4508 | ||||
4509 | if (last == insn) | |||
4510 | break; | |||
4511 | } | |||
4512 | ||||
4513 | /* A second pass, to update ready and speculation status for insns | |||
4514 | depending on the unscheduled ones. The first pass must have | |||
4515 | popped the scheduled_insns vector up to the point where we | |||
4516 | restart scheduling, as recompute_todo_spec requires it to be | |||
4517 | up-to-date. */ | |||
4518 | while (!recompute_vec.is_empty ()) | |||
4519 | { | |||
4520 | rtx_insn *con; | |||
4521 | ||||
4522 | con = recompute_vec.pop (); | |||
4523 | MUST_RECOMPUTE_SPEC_P (con)((&h_i_d[INSN_UID (con)])->must_recompute_spec) = 0; | |||
4524 | if (!sd_lists_empty_p (con, SD_LIST_HARD_BACK(1))) | |||
4525 | { | |||
4526 | TODO_SPEC (con)((&h_i_d[INSN_UID (con)])->todo_spec) = HARD_DEP(((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) << 1); | |||
4527 | INSN_TICK (con)((&h_i_d[INSN_UID (con)])->tick) = INVALID_TICK(-(max_insn_queue_index + 1)); | |||
4528 | if (PREDICATED_PAT (con)((&h_i_d[INSN_UID (con)])->predicated_pat) != NULL_RTX(rtx) 0) | |||
4529 | haifa_change_pattern (con, ORIG_PAT (con)((&h_i_d[INSN_UID (con)])->orig_pat)); | |||
4530 | } | |||
4531 | else if (QUEUE_INDEX (con)((&h_i_d[INSN_UID (con)])->queue_index) != QUEUE_SCHEDULED(-3)) | |||
4532 | TODO_SPEC (con)((&h_i_d[INSN_UID (con)])->todo_spec) = recompute_todo_spec (con, true); | |||
4533 | } | |||
4534 | } | |||
4535 | ||||
4536 | /* Restore scheduler state from the topmost entry on the backtracking queue. | |||
4537 | PSCHED_BLOCK_P points to the local data of schedule_block that we must | |||
4538 | overwrite with the saved data. | |||
4539 | The caller must already have called unschedule_insns_until. */ | |||
4540 | ||||
4541 | static void | |||
4542 | restore_last_backtrack_point (struct sched_block_state *psched_block) | |||
4543 | { | |||
4544 | int i; | |||
4545 | struct haifa_saved_data *save = backtrack_queue; | |||
4546 | ||||
4547 | backtrack_queue = save->next; | |||
4548 | ||||
4549 | if (current_sched_info->restore_state) | |||
4550 | (*current_sched_info->restore_state) (save->fe_saved_data); | |||
4551 | ||||
4552 | if (targetm.sched.alloc_sched_context) | |||
4553 | { | |||
4554 | targetm.sched.set_sched_context (save->be_saved_data); | |||
4555 | targetm.sched.free_sched_context (save->be_saved_data); | |||
4556 | } | |||
4557 | ||||
4558 | /* Do this first since it clobbers INSN_TICK of the involved | |||
4559 | instructions. */ | |||
4560 | undo_replacements_for_backtrack (save); | |||
4561 | ||||
4562 | /* Clear the QUEUE_INDEX of everything in the ready list or one | |||
4563 | of the queues. */ | |||
4564 | if (ready.n_ready > 0) | |||
4565 | { | |||
4566 | rtx_insn **first = ready_lastpos (&ready); | |||
4567 | for (i = 0; i < ready.n_ready; i++) | |||
4568 | { | |||
4569 | rtx_insn *insn = first[i]; | |||
4570 | QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index) = QUEUE_NOWHERE(-2); | |||
4571 | INSN_TICK (insn)((&h_i_d[INSN_UID (insn)])->tick) = INVALID_TICK(-(max_insn_queue_index + 1)); | |||
4572 | } | |||
4573 | } | |||
4574 | for (i = 0; i <= max_insn_queue_index; i++) | |||
4575 | { | |||
4576 | int q = NEXT_Q_AFTER (q_ptr, i)(((q_ptr)+i) & max_insn_queue_index); | |||
4577 | ||||
4578 | for (rtx_insn_list *link = insn_queue[q]; link; link = link->next ()) | |||
4579 | { | |||
4580 | rtx_insn *x = link->insn (); | |||
4581 | QUEUE_INDEX (x)((&h_i_d[INSN_UID (x)])->queue_index) = QUEUE_NOWHERE(-2); | |||
4582 | INSN_TICK (x)((&h_i_d[INSN_UID (x)])->tick) = INVALID_TICK(-(max_insn_queue_index + 1)); | |||
4583 | } | |||
4584 | free_INSN_LIST_list (&insn_queue[q]); | |||
4585 | } | |||
4586 | ||||
4587 | free (ready.vec); | |||
4588 | ready = save->ready; | |||
4589 | ||||
4590 | if (ready.n_ready > 0) | |||
4591 | { | |||
4592 | rtx_insn **first = ready_lastpos (&ready); | |||
4593 | for (i = 0; i < ready.n_ready; i++) | |||
4594 | { | |||
4595 | rtx_insn *insn = first[i]; | |||
4596 | QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index) = QUEUE_READY(-1); | |||
4597 | TODO_SPEC (insn)((&h_i_d[INSN_UID (insn)])->todo_spec) = recompute_todo_spec (insn, true); | |||
4598 | INSN_TICK (insn)((&h_i_d[INSN_UID (insn)])->tick) = save->clock_var; | |||
4599 | } | |||
4600 | } | |||
4601 | ||||
4602 | q_ptr = 0; | |||
4603 | q_size = save->q_size; | |||
4604 | for (i = 0; i <= max_insn_queue_index; i++) | |||
4605 | { | |||
4606 | int q = NEXT_Q_AFTER (q_ptr, i)(((q_ptr)+i) & max_insn_queue_index); | |||
4607 | ||||
4608 | insn_queue[q] = save->insn_queue[q]; | |||
4609 | ||||
4610 | for (rtx_insn_list *link = insn_queue[q]; link; link = link->next ()) | |||
4611 | { | |||
4612 | rtx_insn *x = link->insn (); | |||
4613 | QUEUE_INDEX (x)((&h_i_d[INSN_UID (x)])->queue_index) = i; | |||
4614 | TODO_SPEC (x)((&h_i_d[INSN_UID (x)])->todo_spec) = recompute_todo_spec (x, true); | |||
4615 | INSN_TICK (x)((&h_i_d[INSN_UID (x)])->tick) = save->clock_var + i; | |||
4616 | } | |||
4617 | } | |||
4618 | free (save->insn_queue); | |||
4619 | ||||
4620 | toggle_cancelled_flags (true); | |||
4621 | ||||
4622 | clock_var = save->clock_var; | |||
4623 | last_clock_var = save->last_clock_var; | |||
4624 | cycle_issued_insns = save->cycle_issued_insns; | |||
4625 | last_scheduled_insn = save->last_scheduled_insn; | |||
4626 | last_nondebug_scheduled_insn = save->last_nondebug_scheduled_insn; | |||
4627 | nonscheduled_insns_begin = save->nonscheduled_insns_begin; | |||
4628 | ||||
4629 | *psched_block = save->sched_block; | |||
4630 | ||||
4631 | memcpy (curr_state, save->curr_state, dfa_state_size); | |||
4632 | free (save->curr_state); | |||
4633 | ||||
4634 | mark_backtrack_feeds (save->delay_pair->i2, 0); | |||
4635 | ||||
4636 | gcc_assert (next_cycle_replace_deps.is_empty ())((void)(!(next_cycle_replace_deps.is_empty ()) ? fancy_abort ( "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4636, __FUNCTION__), 0 : 0)); | |||
4637 | next_cycle_replace_deps = save->next_cycle_deps.copy (); | |||
4638 | next_cycle_apply = save->next_cycle_apply.copy (); | |||
4639 | ||||
4640 | free (save); | |||
4641 | ||||
4642 | for (save = backtrack_queue; save; save = save->next) | |||
4643 | { | |||
4644 | mark_backtrack_feeds (save->delay_pair->i2, 1); | |||
4645 | } | |||
4646 | } | |||
4647 | ||||
4648 | /* Discard all data associated with the topmost entry in the backtrack | |||
4649 | queue. If RESET_TICK is false, we just want to free the data. If true, | |||
4650 | we are doing this because we discovered a reason to backtrack. In the | |||
4651 | latter case, also reset the INSN_TICK for the shadow insn. */ | |||
4652 | static void | |||
4653 | free_topmost_backtrack_point (bool reset_tick) | |||
4654 | { | |||
4655 | struct haifa_saved_data *save = backtrack_queue; | |||
4656 | int i; | |||
4657 | ||||
4658 | backtrack_queue = save->next; | |||
4659 | ||||
4660 | if (reset_tick) | |||
4661 | { | |||
4662 | struct delay_pair *pair = save->delay_pair; | |||
4663 | while (pair) | |||
4664 | { | |||
4665 | INSN_TICK (pair->i2)((&h_i_d[INSN_UID (pair->i2)])->tick) = INVALID_TICK(-(max_insn_queue_index + 1)); | |||
4666 | INSN_EXACT_TICK (pair->i2)((&h_i_d[INSN_UID (pair->i2)])->exact_tick) = INVALID_TICK(-(max_insn_queue_index + 1)); | |||
4667 | pair = pair->next_same_i1; | |||
4668 | } | |||
4669 | undo_replacements_for_backtrack (save); | |||
4670 | } | |||
4671 | else | |||
4672 | { | |||
4673 | save->replacement_deps.release (); | |||
4674 | save->replace_apply.release (); | |||
4675 | } | |||
4676 | ||||
4677 | if (targetm.sched.free_sched_context) | |||
4678 | targetm.sched.free_sched_context (save->be_saved_data); | |||
4679 | if (current_sched_info->restore_state) | |||
4680 | free (save->fe_saved_data); | |||
4681 | for (i = 0; i <= max_insn_queue_index; i++) | |||
4682 | free_INSN_LIST_list (&save->insn_queue[i]); | |||
4683 | free (save->insn_queue); | |||
4684 | free (save->curr_state); | |||
4685 | free (save->ready.vec); | |||
4686 | free (save); | |||
4687 | } | |||
4688 | ||||
4689 | /* Free the entire backtrack queue. */ | |||
4690 | static void | |||
4691 | free_backtrack_queue (void) | |||
4692 | { | |||
4693 | while (backtrack_queue) | |||
4694 | free_topmost_backtrack_point (false); | |||
4695 | } | |||
4696 | ||||
4697 | /* Apply a replacement described by DESC. If IMMEDIATELY is false, we | |||
4698 | may have to postpone the replacement until the start of the next cycle, | |||
4699 | at which point we will be called again with IMMEDIATELY true. This is | |||
4700 | only done for machines which have instruction packets with explicit | |||
4701 | parallelism however. */ | |||
4702 | static void | |||
4703 | apply_replacement (dep_t dep, bool immediately) | |||
4704 | { | |||
4705 | struct dep_replacement *desc = DEP_REPLACE (dep)((dep)->replace); | |||
4706 | if (!immediately && targetm.sched.exposed_pipeline && reload_completed) | |||
4707 | { | |||
4708 | next_cycle_replace_deps.safe_push (dep); | |||
4709 | next_cycle_apply.safe_push (1); | |||
4710 | } | |||
4711 | else | |||
4712 | { | |||
4713 | bool success; | |||
4714 | ||||
4715 | if (QUEUE_INDEX (desc->insn)((&h_i_d[INSN_UID (desc->insn)])->queue_index) == QUEUE_SCHEDULED(-3)) | |||
4716 | return; | |||
4717 | ||||
4718 | if (sched_verbose >= 5) | |||
4719 | fprintf (sched_dump, "applying replacement for insn %d\n", | |||
4720 | INSN_UID (desc->insn)); | |||
4721 | ||||
4722 | success = validate_change (desc->insn, desc->loc, desc->newval, 0); | |||
4723 | gcc_assert (success)((void)(!(success) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4723, __FUNCTION__), 0 : 0)); | |||
4724 | ||||
4725 | rtx_insn *insn = DEP_PRO (dep)((dep)->pro); | |||
4726 | ||||
4727 | /* Recompute priority since dependent priorities may have changed. */ | |||
4728 | priority (insn, true); | |||
4729 | update_insn_after_change (desc->insn); | |||
4730 | ||||
4731 | if ((TODO_SPEC (desc->insn)((&h_i_d[INSN_UID (desc->insn)])->todo_spec) & (HARD_DEP(((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) << 1) | DEP_POSTPONED((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1))) == 0) | |||
4732 | fix_tick_ready (desc->insn); | |||
4733 | ||||
4734 | if (backtrack_queue != NULLnullptr) | |||
4735 | { | |||
4736 | backtrack_queue->replacement_deps.safe_push (dep); | |||
4737 | backtrack_queue->replace_apply.safe_push (1); | |||
4738 | } | |||
4739 | } | |||
4740 | } | |||
4741 | ||||
4742 | /* We have determined that a pattern involved in DEP must be restored. | |||
4743 | If IMMEDIATELY is false, we may have to postpone the replacement | |||
4744 | until the start of the next cycle, at which point we will be called | |||
4745 | again with IMMEDIATELY true. */ | |||
4746 | static void | |||
4747 | restore_pattern (dep_t dep, bool immediately) | |||
4748 | { | |||
4749 | rtx_insn *next = DEP_CON (dep)((dep)->con); | |||
4750 | int tick = INSN_TICK (next)((&h_i_d[INSN_UID (next)])->tick); | |||
4751 | ||||
4752 | /* If we already scheduled the insn, the modified version is | |||
4753 | correct. */ | |||
4754 | if (QUEUE_INDEX (next)((&h_i_d[INSN_UID (next)])->queue_index) == QUEUE_SCHEDULED(-3)) | |||
4755 | return; | |||
4756 | ||||
4757 | if (!immediately && targetm.sched.exposed_pipeline && reload_completed) | |||
4758 | { | |||
4759 | next_cycle_replace_deps.safe_push (dep); | |||
4760 | next_cycle_apply.safe_push (0); | |||
4761 | return; | |||
4762 | } | |||
4763 | ||||
4764 | ||||
4765 | if (DEP_TYPE (dep)((dep)->type) == REG_DEP_CONTROL) | |||
4766 | { | |||
4767 | if (sched_verbose >= 5) | |||
4768 | fprintf (sched_dump, "restoring pattern for insn %d\n", | |||
4769 | INSN_UID (next)); | |||
4770 | haifa_change_pattern (next, ORIG_PAT (next)((&h_i_d[INSN_UID (next)])->orig_pat)); | |||
4771 | } | |||
4772 | else | |||
4773 | { | |||
4774 | struct dep_replacement *desc = DEP_REPLACE (dep)((dep)->replace); | |||
4775 | bool success; | |||
4776 | ||||
4777 | if (sched_verbose >= 5) | |||
4778 | fprintf (sched_dump, "restoring pattern for insn %d\n", | |||
4779 | INSN_UID (desc->insn)); | |||
4780 | tick = INSN_TICK (desc->insn)((&h_i_d[INSN_UID (desc->insn)])->tick); | |||
4781 | ||||
4782 | success = validate_change (desc->insn, desc->loc, desc->orig, 0); | |||
4783 | gcc_assert (success)((void)(!(success) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4783, __FUNCTION__), 0 : 0)); | |||
4784 | ||||
4785 | rtx_insn *insn = DEP_PRO (dep)((dep)->pro); | |||
4786 | ||||
4787 | if (QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index) != QUEUE_SCHEDULED(-3)) | |||
4788 | { | |||
4789 | /* Recompute priority since dependent priorities may have changed. */ | |||
4790 | priority (insn, true); | |||
4791 | } | |||
4792 | ||||
4793 | update_insn_after_change (desc->insn); | |||
4794 | ||||
4795 | if (backtrack_queue != NULLnullptr) | |||
4796 | { | |||
4797 | backtrack_queue->replacement_deps.safe_push (dep); | |||
4798 | backtrack_queue->replace_apply.safe_push (0); | |||
4799 | } | |||
4800 | } | |||
4801 | INSN_TICK (next)((&h_i_d[INSN_UID (next)])->tick) = tick; | |||
4802 | if (TODO_SPEC (next)((&h_i_d[INSN_UID (next)])->todo_spec) == DEP_POSTPONED((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1)) | |||
4803 | return; | |||
4804 | ||||
4805 | if (sd_lists_empty_p (next, SD_LIST_BACK((1) | (2)))) | |||
4806 | TODO_SPEC (next)((&h_i_d[INSN_UID (next)])->todo_spec) = 0; | |||
4807 | else if (!sd_lists_empty_p (next, SD_LIST_HARD_BACK(1))) | |||
4808 | TODO_SPEC (next)((&h_i_d[INSN_UID (next)])->todo_spec) = HARD_DEP(((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4 ) - 8) / 4))) << 1) << 1) << 1) << 1); | |||
4809 | } | |||
4810 | ||||
4811 | /* Perform pattern replacements that were queued up until the next | |||
4812 | cycle. */ | |||
4813 | static void | |||
4814 | perform_replacements_new_cycle (void) | |||
4815 | { | |||
4816 | int i; | |||
4817 | dep_t dep; | |||
4818 | FOR_EACH_VEC_ELT (next_cycle_replace_deps, i, dep)for (i = 0; (next_cycle_replace_deps).iterate ((i), &(dep )); ++(i)) | |||
4819 | { | |||
4820 | int apply_p = next_cycle_apply[i]; | |||
4821 | if (apply_p) | |||
4822 | apply_replacement (dep, true); | |||
4823 | else | |||
4824 | restore_pattern (dep, true); | |||
4825 | } | |||
4826 | next_cycle_replace_deps.truncate (0); | |||
4827 | next_cycle_apply.truncate (0); | |||
4828 | } | |||
4829 | ||||
4830 | /* Compute INSN_TICK_ESTIMATE for INSN. PROCESSED is a bitmap of | |||
4831 | instructions we've previously encountered, a set bit prevents | |||
4832 | recursion. BUDGET is a limit on how far ahead we look, it is | |||
4833 | reduced on recursive calls. Return true if we produced a good | |||
4834 | estimate, or false if we exceeded the budget. */ | |||
4835 | static bool | |||
4836 | estimate_insn_tick (bitmap processed, rtx_insn *insn, int budget) | |||
4837 | { | |||
4838 | sd_iterator_def sd_it; | |||
4839 | dep_t dep; | |||
4840 | int earliest = INSN_TICK (insn)((&h_i_d[INSN_UID (insn)])->tick); | |||
4841 | ||||
4842 | FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)for ((sd_it) = sd_iterator_start ((insn), (((1) | (2)))); sd_iterator_cond (&(sd_it), &(dep)); sd_iterator_next (&(sd_it))) | |||
4843 | { | |||
4844 | rtx_insn *pro = DEP_PRO (dep)((dep)->pro); | |||
4845 | int t; | |||
4846 | ||||
4847 | if (DEP_STATUS (dep)((dep)->status) & DEP_CANCELLED(((((((((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + (((8 * 4) - 8) / 4))) << 1) << 1) << 1) << 1 ) << 1) << 1)) | |||
4848 | continue; | |||
4849 | ||||
4850 | if (QUEUE_INDEX (pro)((&h_i_d[INSN_UID (pro)])->queue_index) == QUEUE_SCHEDULED(-3)) | |||
4851 | gcc_assert (INSN_TICK (pro) + dep_cost (dep) <= INSN_TICK (insn))((void)(!(((&h_i_d[INSN_UID (pro)])->tick) + dep_cost ( dep) <= ((&h_i_d[INSN_UID (insn)])->tick)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4851, __FUNCTION__), 0 : 0)); | |||
4852 | else | |||
4853 | { | |||
4854 | int cost = dep_cost (dep); | |||
4855 | if (cost >= budget) | |||
4856 | return false; | |||
4857 | if (!bitmap_bit_p (processed, INSN_LUID (pro)(sched_luids[INSN_UID (pro)]))) | |||
4858 | { | |||
4859 | if (!estimate_insn_tick (processed, pro, budget - cost)) | |||
4860 | return false; | |||
4861 | } | |||
4862 | gcc_assert (INSN_TICK_ESTIMATE (pro) != INVALID_TICK)((void)(!(((&h_i_d[INSN_UID (pro)])->tick_estimate) != (-(max_insn_queue_index + 1))) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4862, __FUNCTION__), 0 : 0)); | |||
4863 | t = INSN_TICK_ESTIMATE (pro)((&h_i_d[INSN_UID (pro)])->tick_estimate) + cost; | |||
4864 | if (earliest == INVALID_TICK(-(max_insn_queue_index + 1)) || t > earliest) | |||
4865 | earliest = t; | |||
4866 | } | |||
4867 | } | |||
4868 | bitmap_set_bit (processed, INSN_LUID (insn)(sched_luids[INSN_UID (insn)])); | |||
4869 | INSN_TICK_ESTIMATE (insn)((&h_i_d[INSN_UID (insn)])->tick_estimate) = earliest; | |||
4870 | return true; | |||
4871 | } | |||
4872 | ||||
4873 | /* Examine the pair of insns in P, and estimate (optimistically, assuming | |||
4874 | infinite resources) the cycle in which the delayed shadow can be issued. | |||
4875 | Return the number of cycles that must pass before the real insn can be | |||
4876 | issued in order to meet this constraint. */ | |||
4877 | static int | |||
4878 | estimate_shadow_tick (struct delay_pair *p) | |||
4879 | { | |||
4880 | auto_bitmap processed; | |||
4881 | int t; | |||
4882 | bool cutoff; | |||
4883 | ||||
4884 | cutoff = !estimate_insn_tick (processed, p->i2, | |||
4885 | max_insn_queue_index + pair_delay (p)); | |||
4886 | if (cutoff) | |||
4887 | return max_insn_queue_index; | |||
4888 | t = INSN_TICK_ESTIMATE (p->i2)((&h_i_d[INSN_UID (p->i2)])->tick_estimate) - (clock_var + pair_delay (p) + 1); | |||
4889 | if (t > 0) | |||
4890 | return t; | |||
4891 | return 0; | |||
4892 | } | |||
4893 | ||||
4894 | /* If INSN has no unresolved backwards dependencies, add it to the schedule and | |||
4895 | recursively resolve all its forward dependencies. */ | |||
4896 | static void | |||
4897 | resolve_dependencies (rtx_insn *insn) | |||
4898 | { | |||
4899 | sd_iterator_def sd_it; | |||
4900 | dep_t dep; | |||
4901 | ||||
4902 | /* Don't use sd_lists_empty_p; it ignores debug insns. */ | |||
4903 | if (DEPS_LIST_FIRST (INSN_HARD_BACK_DEPS (insn))((((&h_d_i_d[(sched_luids[INSN_UID (insn)])])->hard_back_deps ))->first) != NULLnullptr | |||
4904 | || DEPS_LIST_FIRST (INSN_SPEC_BACK_DEPS (insn))((((&h_d_i_d[(sched_luids[INSN_UID (insn)])])->spec_back_deps ))->first) != NULLnullptr) | |||
4905 | return; | |||
4906 | ||||
4907 | if (sched_verbose >= 4) | |||
4908 | fprintf (sched_dump, ";;\tquickly resolving %d\n", INSN_UID (insn)); | |||
4909 | ||||
4910 | if (QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index) >= 0) | |||
4911 | queue_remove (insn); | |||
4912 | ||||
4913 | scheduled_insns.safe_push (insn); | |||
4914 | ||||
4915 | /* Update dependent instructions. */ | |||
4916 | for (sd_it = sd_iterator_start (insn, SD_LIST_FORW(4)); | |||
4917 | sd_iterator_cond (&sd_it, &dep);) | |||
4918 | { | |||
4919 | rtx_insn *next = DEP_CON (dep)((dep)->con); | |||
4920 | ||||
4921 | if (sched_verbose >= 4) | |||
4922 | fprintf (sched_dump, ";;\t\tdep %d against %d\n", INSN_UID (insn), | |||
4923 | INSN_UID (next)); | |||
4924 | ||||
4925 | /* Resolve the dependence between INSN and NEXT. | |||
4926 | sd_resolve_dep () moves current dep to another list thus | |||
4927 | advancing the iterator. */ | |||
4928 | sd_resolve_dep (sd_it); | |||
4929 | ||||
4930 | if (!IS_SPECULATION_BRANCHY_CHECK_P (insn)(((&h_i_d[INSN_UID (insn)])->recovery_block) != nullptr && ((&h_i_d[INSN_UID (insn)])->recovery_block ) != (((cfun + 0))->cfg->x_exit_block_ptr))) | |||
4931 | { | |||
4932 | resolve_dependencies (next); | |||
4933 | } | |||
4934 | else | |||
4935 | /* Check always has only one forward dependence (to the first insn in | |||
4936 | the recovery block), therefore, this will be executed only once. */ | |||
4937 | { | |||
4938 | gcc_assert (sd_lists_empty_p (insn, SD_LIST_FORW))((void)(!(sd_lists_empty_p (insn, (4))) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 4938, __FUNCTION__), 0 : 0)); | |||
4939 | } | |||
4940 | } | |||
4941 | } | |||
4942 | ||||
4943 | ||||
4944 | /* Return the head and tail pointers of ebb starting at BEG and ending | |||
4945 | at END. */ | |||
4946 | void | |||
4947 | get_ebb_head_tail (basic_block beg, basic_block end, | |||
4948 | rtx_insn **headp, rtx_insn **tailp) | |||
4949 | { | |||
4950 | rtx_insn *beg_head = BB_HEAD (beg)(beg)->il.x.head_; | |||
4951 | rtx_insn * beg_tail = BB_END (beg)(beg)->il.x.rtl->end_; | |||
4952 | rtx_insn * end_head = BB_HEAD (end)(end)->il.x.head_; | |||
4953 | rtx_insn * end_tail = BB_END (end)(end)->il.x.rtl->end_; | |||
4954 | ||||
4955 | /* Don't include any notes or labels at the beginning of the BEG | |||
4956 | basic block, or notes at the end of the END basic blocks. */ | |||
4957 | ||||
4958 | if (LABEL_P (beg_head)(((enum rtx_code) (beg_head)->code) == CODE_LABEL)) | |||
4959 | beg_head = NEXT_INSN (beg_head); | |||
4960 | ||||
4961 | while (beg_head != beg_tail) | |||
4962 | if (NOTE_P (beg_head)(((enum rtx_code) (beg_head)->code) == NOTE)) | |||
4963 | beg_head = NEXT_INSN (beg_head); | |||
4964 | else if (DEBUG_INSN_P (beg_head)(((enum rtx_code) (beg_head)->code) == DEBUG_INSN)) | |||
4965 | { | |||
4966 | rtx_insn * note, *next; | |||
4967 | ||||
4968 | for (note = NEXT_INSN (beg_head); | |||
4969 | note != beg_tail; | |||
4970 | note = next) | |||
4971 | { | |||
4972 | next = NEXT_INSN (note); | |||
4973 | if (NOTE_P (note)(((enum rtx_code) (note)->code) == NOTE)) | |||
4974 | { | |||
4975 | if (sched_verbose >= 9) | |||
4976 | fprintf (sched_dump, "reorder %i\n", INSN_UID (note)); | |||
4977 | ||||
4978 | reorder_insns_nobb (note, note, PREV_INSN (beg_head)); | |||
4979 | ||||
4980 | if (BLOCK_FOR_INSN (note) != beg) | |||
4981 | df_insn_change_bb (note, beg); | |||
4982 | } | |||
4983 | else if (!DEBUG_INSN_P (note)(((enum rtx_code) (note)->code) == DEBUG_INSN)) | |||
4984 | break; | |||
4985 | } | |||
4986 | ||||
4987 | break; | |||
4988 | } | |||
4989 | else | |||
4990 | break; | |||
4991 | ||||
4992 | *headp = beg_head; | |||
4993 | ||||
4994 | if (beg == end) | |||
4995 | end_head = beg_head; | |||
4996 | else if (LABEL_P (end_head)(((enum rtx_code) (end_head)->code) == CODE_LABEL)) | |||
4997 | end_head = NEXT_INSN (end_head); | |||
4998 | ||||
4999 | while (end_head != end_tail) | |||
5000 | if (NOTE_P (end_tail)(((enum rtx_code) (end_tail)->code) == NOTE)) | |||
5001 | end_tail = PREV_INSN (end_tail); | |||
5002 | else if (DEBUG_INSN_P (end_tail)(((enum rtx_code) (end_tail)->code) == DEBUG_INSN)) | |||
5003 | { | |||
5004 | rtx_insn * note, *prev; | |||
5005 | ||||
5006 | for (note = PREV_INSN (end_tail); | |||
5007 | note != end_head; | |||
5008 | note = prev) | |||
5009 | { | |||
5010 | prev = PREV_INSN (note); | |||
5011 | if (NOTE_P (note)(((enum rtx_code) (note)->code) == NOTE)) | |||
5012 | { | |||
5013 | if (sched_verbose >= 9) | |||
5014 | fprintf (sched_dump, "reorder %i\n", INSN_UID (note)); | |||
5015 | ||||
5016 | reorder_insns_nobb (note, note, end_tail); | |||
5017 | ||||
5018 | if (end_tail == BB_END (end)(end)->il.x.rtl->end_) | |||
5019 | BB_END (end)(end)->il.x.rtl->end_ = note; | |||
5020 | ||||
5021 | if (BLOCK_FOR_INSN (note) != end) | |||
5022 | df_insn_change_bb (note, end); | |||
5023 | } | |||
5024 | else if (!DEBUG_INSN_P (note)(((enum rtx_code) (note)->code) == DEBUG_INSN)) | |||
5025 | break; | |||
5026 | } | |||
5027 | ||||
5028 | break; | |||
5029 | } | |||
5030 | else | |||
5031 | break; | |||
5032 | ||||
5033 | *tailp = end_tail; | |||
5034 | } | |||
5035 | ||||
5036 | /* Return nonzero if there are no real insns in the range [ HEAD, TAIL ]. */ | |||
5037 | ||||
5038 | int | |||
5039 | no_real_insns_p (const rtx_insn *head, const rtx_insn *tail) | |||
5040 | { | |||
5041 | while (head != NEXT_INSN (tail)) | |||
5042 | { | |||
5043 | if (!NOTE_P (head)(((enum rtx_code) (head)->code) == NOTE) && !LABEL_P (head)(((enum rtx_code) (head)->code) == CODE_LABEL)) | |||
5044 | return 0; | |||
5045 | head = NEXT_INSN (head); | |||
5046 | } | |||
5047 | return 1; | |||
5048 | } | |||
5049 | ||||
5050 | /* Restore-other-notes: NOTE_LIST is the end of a chain of notes | |||
5051 | previously found among the insns. Insert them just before HEAD. */ | |||
5052 | rtx_insn * | |||
5053 | restore_other_notes (rtx_insn *head, basic_block head_bb) | |||
5054 | { | |||
5055 | if (note_list != 0) | |||
5056 | { | |||
5057 | rtx_insn *note_head = note_list; | |||
5058 | ||||
5059 | if (head) | |||
5060 | head_bb = BLOCK_FOR_INSN (head); | |||
5061 | else | |||
5062 | head = NEXT_INSN (bb_note (head_bb)); | |||
5063 | ||||
5064 | while (PREV_INSN (note_head)) | |||
5065 | { | |||
5066 | set_block_for_insn (note_head, head_bb); | |||
5067 | note_head = PREV_INSN (note_head); | |||
5068 | } | |||
5069 | /* In the above cycle we've missed this note. */ | |||
5070 | set_block_for_insn (note_head, head_bb); | |||
5071 | ||||
5072 | SET_PREV_INSN (note_head) = PREV_INSN (head); | |||
5073 | SET_NEXT_INSN (PREV_INSN (head)) = note_head; | |||
5074 | SET_PREV_INSN (head) = note_list; | |||
5075 | SET_NEXT_INSN (note_list) = head; | |||
5076 | ||||
5077 | if (BLOCK_FOR_INSN (head) != head_bb) | |||
5078 | BB_END (head_bb)(head_bb)->il.x.rtl->end_ = note_list; | |||
5079 | ||||
5080 | head = note_head; | |||
5081 | } | |||
5082 | ||||
5083 | return head; | |||
5084 | } | |||
5085 | ||||
5086 | /* When we know we are going to discard the schedule due to a failed attempt | |||
5087 | at modulo scheduling, undo all replacements. */ | |||
5088 | static void | |||
5089 | undo_all_replacements (void) | |||
5090 | { | |||
5091 | rtx_insn *insn; | |||
5092 | int i; | |||
5093 | ||||
5094 | FOR_EACH_VEC_ELT (scheduled_insns, i, insn)for (i = 0; (scheduled_insns).iterate ((i), &(insn)); ++( i)) | |||
5095 | { | |||
5096 | sd_iterator_def sd_it; | |||
5097 | dep_t dep; | |||
5098 | ||||
5099 | /* See if we must undo a replacement. */ | |||
5100 | for (sd_it = sd_iterator_start (insn, SD_LIST_RES_FORW(16)); | |||
5101 | sd_iterator_cond (&sd_it, &dep); sd_iterator_next (&sd_it)) | |||
5102 | { | |||
5103 | struct dep_replacement *desc = DEP_REPLACE (dep)((dep)->replace); | |||
5104 | if (desc != NULLnullptr) | |||
5105 | validate_change (desc->insn, desc->loc, desc->orig, 0); | |||
5106 | } | |||
5107 | } | |||
5108 | } | |||
5109 | ||||
5110 | /* Return first non-scheduled insn in the current scheduling block. | |||
5111 | This is mostly used for debug-counter purposes. */ | |||
5112 | static rtx_insn * | |||
5113 | first_nonscheduled_insn (void) | |||
5114 | { | |||
5115 | rtx_insn *insn = (nonscheduled_insns_begin != NULL_RTX(rtx) 0 | |||
5116 | ? nonscheduled_insns_begin | |||
5117 | : current_sched_info->prev_head); | |||
5118 | ||||
5119 | do | |||
5120 | { | |||
5121 | insn = next_nonnote_nondebug_insn (insn); | |||
5122 | } | |||
5123 | while (QUEUE_INDEX (insn)((&h_i_d[INSN_UID (insn)])->queue_index) == QUEUE_SCHEDULED(-3)); | |||
5124 | ||||
5125 | return insn; | |||
5126 | } | |||
5127 | ||||
5128 | /* Move insns that became ready to fire from queue to ready list. */ | |||
5129 | ||||
5130 | static void | |||
5131 | queue_to_ready (struct ready_list *ready) | |||
5132 | { | |||
5133 | rtx_insn *insn; | |||
5134 | rtx_insn_list *link; | |||
5135 | rtx_insn *skip_insn; | |||
5136 | ||||
5137 | q_ptr = NEXT_Q (q_ptr)(((q_ptr)+1) & max_insn_queue_index); | |||
5138 | ||||
5139 | if (dbg_cnt (sched_insn) == false) | |||
5140 | /* If debug counter is activated do not requeue the first | |||
5141 | nonscheduled insn. */ | |||
5142 | skip_insn = first_nonscheduled_insn (); | |||
5143 | else | |||
5144 | skip_insn = NULLnullptr; | |||
5145 | ||||
5146 | /* Add all pending insns that can be scheduled without stalls to the | |||
5147 | ready list. */ | |||
5148 | for (link = insn_queue[q_ptr]; link; link = link->next ()) | |||
5149 | { | |||
5150 | insn = link->insn (); | |||
5151 | q_size -= 1; | |||
5152 | ||||
5153 | if (sched_verbose >= 2) | |||
5154 | fprintf (sched_dump, ";;\t\tQ-->Ready: insn %s: ", | |||
5155 | (*current_sched_info->print_insn) (insn, 0)); | |||
5156 | ||||
5157 | /* If the ready list is full, delay the insn for 1 cycle. | |||
5158 | See the comment in schedule_block for the rationale. */ | |||
5159 | if (!reload_completed | |||
5160 | && (ready->n_ready - ready->n_debug > param_max_sched_ready_insnsglobal_options.x_param_max_sched_ready_insns | |||
5161 | || (sched_pressure == SCHED_PRESSURE_MODEL | |||
5162 | /* Limit pressure recalculations to | |||
5163 | param_max_sched_ready_insns instructions too. */ | |||
5164 | && model_index (insn) > (model_curr_point | |||
5165 | + param_max_sched_ready_insnsglobal_options.x_param_max_sched_ready_insns))) | |||
5166 | && !(sched_pressure == SCHED_PRESSURE_MODEL | |||
5167 | && model_curr_point < model_num_insns | |||
5168 | /* Always allow the next model instruction to issue. */ | |||
5169 | && model_index (insn) == model_curr_point) | |||
5170 | && !SCHED_GROUP_P (insn)(__extension__ ({ __typeof ((insn)) const _rtx = ((insn)); if (((enum rtx_code) (_rtx)->code) != DEBUG_INSN && ( (enum rtx_code) (_rtx)->code) != INSN && ((enum rtx_code ) (_rtx)->code) != JUMP_INSN && ((enum rtx_code) ( _rtx)->code) != CALL_INSN) rtl_check_failed_flag ("SCHED_GROUP_P" , _rtx, "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5170, __FUNCTION__); _rtx; })->in_struct) | |||
5171 | && insn != skip_insn) | |||
5172 | { | |||
5173 | if (sched_verbose >= 2) | |||
5174 | fprintf (sched_dump, "keeping in queue, ready full\n"); | |||
5175 | queue_insn (insn, 1, "ready full"); | |||
5176 | } | |||
5177 | else | |||
5178 | { | |||
5179 | ready_add (ready, insn, false); | |||
5180 | if (sched_verbose >= 2) | |||
5181 | fprintf (sched_dump, "moving to ready without stalls\n"); | |||
5182 | } | |||
5183 | } | |||
5184 | free_INSN_LIST_list (&insn_queue[q_ptr]); | |||
5185 | ||||
5186 | /* If there are no ready insns, stall until one is ready and add all | |||
5187 | of the pending insns at that point to the ready list. */ | |||
5188 | if (ready->n_ready == 0) | |||
5189 | { | |||
5190 | int stalls; | |||
5191 | ||||
5192 | for (stalls = 1; stalls <= max_insn_queue_index; stalls++) | |||
5193 | { | |||
5194 | if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)(((q_ptr)+stalls) & max_insn_queue_index)])) | |||
5195 | { | |||
5196 | for (; link; link = link->next ()) | |||
5197 | { | |||
5198 | insn = link->insn (); | |||
5199 | q_size -= 1; | |||
5200 | ||||
5201 | if (sched_verbose >= 2) | |||
5202 | fprintf (sched_dump, ";;\t\tQ-->Ready: insn %s: ", | |||
5203 | (*current_sched_info->print_insn) (insn, 0)); | |||
5204 | ||||
5205 | ready_add (ready, insn, false); | |||
5206 | if (sched_verbose >= 2) | |||
5207 | fprintf (sched_dump, "moving to ready with %d stalls\n", stalls); | |||
5208 | } | |||
5209 | free_INSN_LIST_list (&insn_queue[NEXT_Q_AFTER (q_ptr, stalls)(((q_ptr)+stalls) & max_insn_queue_index)]); | |||
5210 | ||||
5211 | advance_one_cycle (); | |||
5212 | ||||
5213 | break; | |||
5214 | } | |||
5215 | ||||
5216 | advance_one_cycle (); | |||
5217 | } | |||
5218 | ||||
5219 | q_ptr = NEXT_Q_AFTER (q_ptr, stalls)(((q_ptr)+stalls) & max_insn_queue_index); | |||
5220 | clock_var += stalls; | |||
5221 | if (sched_verbose >= 2) | |||
5222 | fprintf (sched_dump, ";;\tAdvancing clock by %d cycle[s] to %d\n", | |||
5223 | stalls, clock_var); | |||
5224 | } | |||
5225 | } | |||
5226 | ||||
5227 | /* Used by early_queue_to_ready. Determines whether it is "ok" to | |||
5228 | prematurely move INSN from the queue to the ready list. Currently, | |||
5229 | if a target defines the hook 'is_costly_dependence', this function | |||
5230 | uses the hook to check whether there exist any dependences which are | |||
5231 | considered costly by the target, between INSN and other insns that | |||
5232 | have already been scheduled. Dependences are checked up to Y cycles | |||
5233 | back, with default Y=1; The flag -fsched-stalled-insns-dep=Y allows | |||
5234 | controlling this value. | |||
5235 | (Other considerations could be taken into account instead (or in | |||
5236 | addition) depending on user flags and target hooks. */ | |||
5237 | ||||
5238 | static bool | |||
5239 | ok_for_early_queue_removal (rtx_insn *insn) | |||
5240 | { | |||
5241 | if (targetm.sched.is_costly_dependence) | |||
5242 | { | |||
5243 | int n_cycles; | |||
5244 | int i = scheduled_insns.length (); | |||
5245 | for (n_cycles = flag_sched_stalled_insns_depglobal_options.x_flag_sched_stalled_insns_dep; n_cycles; n_cycles--) | |||
5246 | { | |||
5247 | while (i-- > 0) | |||
5248 | { | |||
5249 | int cost; | |||
5250 | ||||
5251 | rtx_insn *prev_insn = scheduled_insns[i]; | |||
5252 | ||||
5253 | if (!NOTE_P (prev_insn)(((enum rtx_code) (prev_insn)->code) == NOTE)) | |||
5254 | { | |||
5255 | dep_t dep; | |||
5256 | ||||
5257 | dep = sd_find_dep_between (prev_insn, insn, true); | |||
5258 | ||||
5259 | if (dep != NULLnullptr) | |||
5260 | { | |||
5261 | cost = dep_cost (dep); | |||
5262 | ||||
5263 | if (targetm.sched.is_costly_dependence (dep, cost, | |||
5264 | flag_sched_stalled_insns_depglobal_options.x_flag_sched_stalled_insns_dep - n_cycles)) | |||
5265 | return false; | |||
5266 | } | |||
5267 | } | |||
5268 | ||||
5269 | if (GET_MODE (prev_insn)((machine_mode) (prev_insn)->mode) == TImode(scalar_int_mode ((scalar_int_mode::from_int) E_TImode))) /* end of dispatch group */ | |||
5270 | break; | |||
5271 | } | |||
5272 | ||||
5273 | if (i == 0) | |||
5274 | break; | |||
5275 | } | |||
5276 | } | |||
5277 | ||||
5278 | return true; | |||
5279 | } | |||
5280 | ||||
5281 | ||||
5282 | /* Remove insns from the queue, before they become "ready" with respect | |||
5283 | to FU latency considerations. */ | |||
5284 | ||||
5285 | static int | |||
5286 | early_queue_to_ready (state_t state, struct ready_list *ready) | |||
5287 | { | |||
5288 | rtx_insn *insn; | |||
5289 | rtx_insn_list *link; | |||
5290 | rtx_insn_list *next_link; | |||
5291 | rtx_insn_list *prev_link; | |||
5292 | bool move_to_ready; | |||
5293 | int cost; | |||
5294 | state_t temp_state = alloca (dfa_state_size)__builtin_alloca(dfa_state_size); | |||
5295 | int stalls; | |||
5296 | int insns_removed = 0; | |||
5297 | ||||
5298 | /* | |||
5299 | Flag '-fsched-stalled-insns=X' determines the aggressiveness of this | |||
5300 | function: | |||
5301 | ||||
5302 | X == 0: There is no limit on how many queued insns can be removed | |||
5303 | prematurely. (flag_sched_stalled_insns = -1). | |||
5304 | ||||
5305 | X >= 1: Only X queued insns can be removed prematurely in each | |||
5306 | invocation. (flag_sched_stalled_insns = X). | |||
5307 | ||||
5308 | Otherwise: Early queue removal is disabled. | |||
5309 | (flag_sched_stalled_insns = 0) | |||
5310 | */ | |||
5311 | ||||
5312 | if (! flag_sched_stalled_insnsglobal_options.x_flag_sched_stalled_insns) | |||
5313 | return 0; | |||
5314 | ||||
5315 | for (stalls = 0; stalls <= max_insn_queue_index; stalls++) | |||
5316 | { | |||
5317 | if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)(((q_ptr)+stalls) & max_insn_queue_index)])) | |||
5318 | { | |||
5319 | if (sched_verbose > 6) | |||
5320 | fprintf (sched_dump, ";; look at index %d + %d\n", q_ptr, stalls); | |||
5321 | ||||
5322 | prev_link = 0; | |||
5323 | while (link) | |||
5324 | { | |||
5325 | next_link = link->next (); | |||
5326 | insn = link->insn (); | |||
5327 | if (insn && sched_verbose > 6) | |||
5328 | print_rtl_single (sched_dump, insn); | |||
5329 | ||||
5330 | memcpy (temp_state, state, dfa_state_size); | |||
5331 | if (recog_memoized (insn) < 0) | |||
5332 | /* non-negative to indicate that it's not ready | |||
5333 | to avoid infinite Q->R->Q->R... */ | |||
5334 | cost = 0; | |||
5335 | else | |||
5336 | cost = state_transition (temp_state, insn); | |||
5337 | ||||
5338 | if (sched_verbose >= 6) | |||
5339 | fprintf (sched_dump, "transition cost = %d\n", cost); | |||
5340 | ||||
5341 | move_to_ready = false; | |||
5342 | if (cost < 0) | |||
5343 | { | |||
5344 | move_to_ready = ok_for_early_queue_removal (insn); | |||
5345 | if (move_to_ready == true) | |||
5346 | { | |||
5347 | /* move from Q to R */ | |||
5348 | q_size -= 1; | |||
5349 | ready_add (ready, insn, false); | |||
5350 | ||||
5351 | if (prev_link) | |||
5352 | XEXP (prev_link, 1)(((prev_link)->u.fld[1]).rt_rtx) = next_link; | |||
5353 | else | |||
5354 | insn_queue[NEXT_Q_AFTER (q_ptr, stalls)(((q_ptr)+stalls) & max_insn_queue_index)] = next_link; | |||
5355 | ||||
5356 | free_INSN_LIST_node (link); | |||
5357 | ||||
5358 | if (sched_verbose >= 2) | |||
5359 | fprintf (sched_dump, ";;\t\tEarly Q-->Ready: insn %s\n", | |||
5360 | (*current_sched_info->print_insn) (insn, 0)); | |||
5361 | ||||
5362 | insns_removed++; | |||
5363 | if (insns_removed == flag_sched_stalled_insnsglobal_options.x_flag_sched_stalled_insns) | |||
5364 | /* Remove no more than flag_sched_stalled_insns insns | |||
5365 | from Q at a time. */ | |||
5366 | return insns_removed; | |||
5367 | } | |||
5368 | } | |||
5369 | ||||
5370 | if (move_to_ready == false) | |||
5371 | prev_link = link; | |||
5372 | ||||
5373 | link = next_link; | |||
5374 | } /* while link */ | |||
5375 | } /* if link */ | |||
5376 | ||||
5377 | } /* for stalls.. */ | |||
5378 | ||||
5379 | return insns_removed; | |||
5380 | } | |||
5381 | ||||
5382 | ||||
5383 | /* Print the ready list for debugging purposes. | |||
5384 | If READY_TRY is non-zero then only print insns that max_issue | |||
5385 | will consider. */ | |||
5386 | static void | |||
5387 | debug_ready_list_1 (struct ready_list *ready, signed char *ready_try) | |||
5388 | { | |||
5389 | rtx_insn **p; | |||
5390 | int i; | |||
5391 | ||||
5392 | if (ready->n_ready == 0) | |||
5393 | { | |||
5394 | fprintf (sched_dump, "\n"); | |||
5395 | return; | |||
5396 | } | |||
5397 | ||||
5398 | p = ready_lastpos (ready); | |||
5399 | for (i = 0; i < ready->n_ready; i++) | |||
5400 | { | |||
5401 | if (ready_try != NULLnullptr && ready_try[ready->n_ready - i - 1]) | |||
5402 | continue; | |||
5403 | ||||
5404 | fprintf (sched_dump, " %s:%d", | |||
5405 | (*current_sched_info->print_insn) (p[i], 0), | |||
5406 | INSN_LUID (p[i])(sched_luids[INSN_UID (p[i])])); | |||
5407 | if (sched_pressure != SCHED_PRESSURE_NONE) | |||
5408 | fprintf (sched_dump, "(cost=%d", | |||
5409 | INSN_REG_PRESSURE_EXCESS_COST_CHANGE (p[i])((&h_i_d[INSN_UID (p[i])])->reg_pressure_excess_cost_change )); | |||
5410 | fprintf (sched_dump, ":prio=%d", INSN_PRIORITY (p[i])((&h_i_d[INSN_UID (p[i])])->priority)); | |||
5411 | if (INSN_TICK (p[i])((&h_i_d[INSN_UID (p[i])])->tick) > clock_var) | |||
5412 | fprintf (sched_dump, ":delay=%d", INSN_TICK (p[i])((&h_i_d[INSN_UID (p[i])])->tick) - clock_var); | |||
5413 | if (sched_pressure == SCHED_PRESSURE_MODEL) | |||
5414 | fprintf (sched_dump, ":idx=%d", | |||
5415 | model_index (p[i])); | |||
5416 | if (sched_pressure != SCHED_PRESSURE_NONE) | |||
5417 | fprintf (sched_dump, ")"); | |||
5418 | } | |||
5419 | fprintf (sched_dump, "\n"); | |||
5420 | } | |||
5421 | ||||
5422 | /* Print the ready list. Callable from debugger. */ | |||
5423 | static void | |||
5424 | debug_ready_list (struct ready_list *ready) | |||
5425 | { | |||
5426 | debug_ready_list_1 (ready, NULLnullptr); | |||
5427 | } | |||
5428 | ||||
5429 | /* Search INSN for REG_SAVE_NOTE notes and convert them back into insn | |||
5430 | NOTEs. This is used for NOTE_INSN_EPILOGUE_BEG, so that sched-ebb | |||
5431 | replaces the epilogue note in the correct basic block. */ | |||
5432 | void | |||
5433 | reemit_notes (rtx_insn *insn) | |||
5434 | { | |||
5435 | rtx note; | |||
5436 | rtx_insn *last = insn; | |||
5437 | ||||
5438 | for (note = REG_NOTES (insn)(((insn)->u.fld[6]).rt_rtx); note; note = XEXP (note, 1)(((note)->u.fld[1]).rt_rtx)) | |||
5439 | { | |||
5440 | if (REG_NOTE_KIND (note)((enum reg_note) ((machine_mode) (note)->mode)) == REG_SAVE_NOTE) | |||
5441 | { | |||
5442 | enum insn_note note_type = (enum insn_note) INTVAL (XEXP (note, 0))(((((note)->u.fld[0]).rt_rtx))->u.hwint[0]); | |||
5443 | ||||
5444 | last = emit_note_before (note_type, last); | |||
5445 | remove_note (insn, note); | |||
5446 | df_insn_create_insn_record (last); | |||
5447 | } | |||
5448 | } | |||
5449 | } | |||
5450 | ||||
5451 | /* Move INSN. Reemit notes if needed. Update CFG, if needed. */ | |||
5452 | static void | |||
5453 | move_insn (rtx_insn *insn, rtx_insn *last, rtx nt) | |||
5454 | { | |||
5455 | if (PREV_INSN (insn) != last) | |||
5456 | { | |||
5457 | basic_block bb; | |||
5458 | rtx_insn *note; | |||
5459 | int jump_p = 0; | |||
5460 | ||||
5461 | bb = BLOCK_FOR_INSN (insn); | |||
5462 | ||||
5463 | /* BB_HEAD is either LABEL or NOTE. */ | |||
5464 | gcc_assert (BB_HEAD (bb) != insn)((void)(!((bb)->il.x.head_ != insn) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5464, __FUNCTION__), 0 : 0)); | |||
5465 | ||||
5466 | if (BB_END (bb)(bb)->il.x.rtl->end_ == insn) | |||
5467 | /* If this is last instruction in BB, move end marker one | |||
5468 | instruction up. */ | |||
5469 | { | |||
5470 | /* Jumps are always placed at the end of basic block. */ | |||
5471 | jump_p = control_flow_insn_p (insn); | |||
5472 | ||||
5473 | gcc_assert (!jump_p((void)(!(!jump_p || ((common_sched_info->sched_pass_id == SCHED_RGN_PASS) && (((&h_i_d[INSN_UID (insn)])-> recovery_block) != nullptr && ((&h_i_d[INSN_UID ( insn)])->recovery_block) != (((cfun + 0))->cfg->x_exit_block_ptr ))) || (common_sched_info->sched_pass_id == SCHED_EBB_PASS )) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5477, __FUNCTION__), 0 : 0)) | |||
5474 | || ((common_sched_info->sched_pass_id == SCHED_RGN_PASS)((void)(!(!jump_p || ((common_sched_info->sched_pass_id == SCHED_RGN_PASS) && (((&h_i_d[INSN_UID (insn)])-> recovery_block) != nullptr && ((&h_i_d[INSN_UID ( insn)])->recovery_block) != (((cfun + 0))->cfg->x_exit_block_ptr ))) || (common_sched_info->sched_pass_id == SCHED_EBB_PASS )) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5477, __FUNCTION__), 0 : 0)) | |||
5475 | && IS_SPECULATION_BRANCHY_CHECK_P (insn))((void)(!(!jump_p || ((common_sched_info->sched_pass_id == SCHED_RGN_PASS) && (((&h_i_d[INSN_UID (insn)])-> recovery_block) != nullptr && ((&h_i_d[INSN_UID ( insn)])->recovery_block) != (((cfun + 0))->cfg->x_exit_block_ptr ))) || (common_sched_info->sched_pass_id == SCHED_EBB_PASS )) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5477, __FUNCTION__), 0 : 0)) | |||
5476 | || (common_sched_info->sched_pass_id((void)(!(!jump_p || ((common_sched_info->sched_pass_id == SCHED_RGN_PASS) && (((&h_i_d[INSN_UID (insn)])-> recovery_block) != nullptr && ((&h_i_d[INSN_UID ( insn)])->recovery_block) != (((cfun + 0))->cfg->x_exit_block_ptr ))) || (common_sched_info->sched_pass_id == SCHED_EBB_PASS )) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5477, __FUNCTION__), 0 : 0)) | |||
5477 | == SCHED_EBB_PASS))((void)(!(!jump_p || ((common_sched_info->sched_pass_id == SCHED_RGN_PASS) && (((&h_i_d[INSN_UID (insn)])-> recovery_block) != nullptr && ((&h_i_d[INSN_UID ( insn)])->recovery_block) != (((cfun + 0))->cfg->x_exit_block_ptr ))) || (common_sched_info->sched_pass_id == SCHED_EBB_PASS )) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5477, __FUNCTION__), 0 : 0)); | |||
5478 | ||||
5479 | gcc_assert (BLOCK_FOR_INSN (PREV_INSN (insn)) == bb)((void)(!(BLOCK_FOR_INSN (PREV_INSN (insn)) == bb) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5479, __FUNCTION__), 0 : 0)); | |||
5480 | ||||
5481 | BB_END (bb)(bb)->il.x.rtl->end_ = PREV_INSN (insn); | |||
5482 | } | |||
5483 | ||||
5484 | gcc_assert (BB_END (bb) != last)((void)(!((bb)->il.x.rtl->end_ != last) ? fancy_abort ( "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5484, __FUNCTION__), 0 : 0)); | |||
5485 | ||||
5486 | if (jump_p) | |||
5487 | /* We move the block note along with jump. */ | |||
5488 | { | |||
5489 | gcc_assert (nt)((void)(!(nt) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5489, __FUNCTION__), 0 : 0)); | |||
5490 | ||||
5491 | note = NEXT_INSN (insn); | |||
5492 | while (NOTE_NOT_BB_P (note)((((enum rtx_code) (note)->code) == NOTE) && ((((note )->u.fld[4]).rt_int) != NOTE_INSN_BASIC_BLOCK)) && note != nt) | |||
5493 | note = NEXT_INSN (note); | |||
5494 | ||||
5495 | if (note != nt | |||
5496 | && (LABEL_P (note)(((enum rtx_code) (note)->code) == CODE_LABEL) | |||
5497 | || BARRIER_P (note)(((enum rtx_code) (note)->code) == BARRIER))) | |||
5498 | note = NEXT_INSN (note); | |||
5499 | ||||
5500 | gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note))((void)(!(((((enum rtx_code) (note)->code) == NOTE) && (((note)->u.fld[4]).rt_int) == NOTE_INSN_BASIC_BLOCK)) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5500, __FUNCTION__), 0 : 0)); | |||
5501 | } | |||
5502 | else | |||
5503 | note = insn; | |||
5504 | ||||
5505 | SET_NEXT_INSN (PREV_INSN (insn)) = NEXT_INSN (note); | |||
5506 | SET_PREV_INSN (NEXT_INSN (note)) = PREV_INSN (insn); | |||
5507 | ||||
5508 | SET_NEXT_INSN (note) = NEXT_INSN (last); | |||
5509 | SET_PREV_INSN (NEXT_INSN (last)) = note; | |||
5510 | ||||
5511 | SET_NEXT_INSN (last) = insn; | |||
5512 | SET_PREV_INSN (insn) = last; | |||
5513 | ||||
5514 | bb = BLOCK_FOR_INSN (last); | |||
5515 | ||||
5516 | if (jump_p) | |||
5517 | { | |||
5518 | fix_jump_move (insn); | |||
5519 | ||||
5520 | if (BLOCK_FOR_INSN (insn) != bb) | |||
5521 | move_block_after_check (insn); | |||
5522 | ||||
5523 | gcc_assert (BB_END (bb) == last)((void)(!((bb)->il.x.rtl->end_ == last) ? fancy_abort ( "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5523, __FUNCTION__), 0 : 0)); | |||
5524 | } | |||
5525 | ||||
5526 | df_insn_change_bb (insn, bb); | |||
5527 | ||||
5528 | /* Update BB_END, if needed. */ | |||
5529 | if (BB_END (bb)(bb)->il.x.rtl->end_ == last) | |||
5530 | BB_END (bb)(bb)->il.x.rtl->end_ = insn; | |||
5531 | } | |||
5532 | ||||
5533 | SCHED_GROUP_P (insn)(__extension__ ({ __typeof ((insn)) const _rtx = ((insn)); if (((enum rtx_code) (_rtx)->code) != DEBUG_INSN && ( (enum rtx_code) (_rtx)->code) != INSN && ((enum rtx_code ) (_rtx)->code) != JUMP_INSN && ((enum rtx_code) ( _rtx)->code) != CALL_INSN) rtl_check_failed_flag ("SCHED_GROUP_P" , _rtx, "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5533, __FUNCTION__); _rtx; })->in_struct) = 0; | |||
5534 | } | |||
5535 | ||||
5536 | /* Return true if scheduling INSN will finish current clock cycle. */ | |||
5537 | static bool | |||
5538 | insn_finishes_cycle_p (rtx_insn *insn) | |||
5539 | { | |||
5540 | if (SCHED_GROUP_P (insn)(__extension__ ({ __typeof ((insn)) const _rtx = ((insn)); if (((enum rtx_code) (_rtx)->code) != DEBUG_INSN && ( (enum rtx_code) (_rtx)->code) != INSN && ((enum rtx_code ) (_rtx)->code) != JUMP_INSN && ((enum rtx_code) ( _rtx)->code) != CALL_INSN) rtl_check_failed_flag ("SCHED_GROUP_P" , _rtx, "/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5540, __FUNCTION__); _rtx; })->in_struct)) | |||
5541 | /* After issuing INSN, rest of the sched_group will be forced to issue | |||
5542 | in order. Don't make any plans for the rest of cycle. */ | |||
5543 | return true; | |||
5544 | ||||
5545 | /* Finishing the block will, apparently, finish the cycle. */ | |||
5546 | if (current_sched_info->insn_finishes_block_p | |||
5547 | && current_sched_info->insn_finishes_block_p (insn)) | |||
5548 | return true; | |||
5549 | ||||
5550 | return false; | |||
5551 | } | |||
5552 | ||||
5553 | /* Helper for autopref_multipass_init. Given a SET in PAT and whether | |||
5554 | we're expecting a memory WRITE or not, check that the insn is relevant to | |||
5555 | the autoprefetcher modelling code. Return true iff that is the case. | |||
5556 | If it is relevant, record the base register of the memory op in BASE and | |||
5557 | the offset in OFFSET. */ | |||
5558 | ||||
5559 | static bool | |||
5560 | analyze_set_insn_for_autopref (rtx pat, bool write, rtx *base, int *offset) | |||
5561 | { | |||
5562 | if (GET_CODE (pat)((enum rtx_code) (pat)->code) != SET) | |||
5563 | return false; | |||
5564 | ||||
5565 | rtx mem = write ? SET_DEST (pat)(((pat)->u.fld[0]).rt_rtx) : SET_SRC (pat)(((pat)->u.fld[1]).rt_rtx); | |||
5566 | if (!MEM_P (mem)(((enum rtx_code) (mem)->code) == MEM)) | |||
5567 | return false; | |||
5568 | ||||
5569 | struct address_info info; | |||
5570 | decompose_mem_address (&info, mem); | |||
5571 | ||||
5572 | /* TODO: Currently only (base+const) addressing is supported. */ | |||
5573 | if (info.base == NULLnullptr || !REG_P (*info.base)(((enum rtx_code) (*info.base)->code) == REG) | |||
5574 | || (info.disp != NULLnullptr && !CONST_INT_P (*info.disp)(((enum rtx_code) (*info.disp)->code) == CONST_INT))) | |||
5575 | return false; | |||
5576 | ||||
5577 | *base = *info.base; | |||
5578 | *offset = info.disp ? INTVAL (*info.disp)((*info.disp)->u.hwint[0]) : 0; | |||
5579 | return true; | |||
5580 | } | |||
5581 | ||||
5582 | /* Functions to model cache auto-prefetcher. | |||
5583 | ||||
5584 | Some of the CPUs have cache auto-prefetcher, which /seems/ to initiate | |||
5585 | memory prefetches if it sees instructions with consequitive memory accesses | |||
5586 | in the instruction stream. Details of such hardware units are not published, | |||
5587 | so we can only guess what exactly is going on there. | |||
5588 | In the scheduler, we model abstract auto-prefetcher. If there are memory | |||
5589 | insns in the ready list (or the queue) that have same memory base, but | |||
5590 | different offsets, then we delay the insns with larger offsets until insns | |||
5591 | with smaller offsets get scheduled. If PARAM_SCHED_AUTOPREF_QUEUE_DEPTH | |||
5592 | is "1", then we look at the ready list; if it is N>1, then we also look | |||
5593 | through N-1 queue entries. | |||
5594 | If the param is N>=0, then rank_for_schedule will consider auto-prefetching | |||
5595 | among its heuristics. | |||
5596 | Param value of "-1" disables modelling of the auto-prefetcher. */ | |||
5597 | ||||
5598 | /* Initialize autoprefetcher model data for INSN. */ | |||
5599 | static void | |||
5600 | autopref_multipass_init (const rtx_insn *insn, int write) | |||
5601 | { | |||
5602 | autopref_multipass_data_t data = &INSN_AUTOPREF_MULTIPASS_DATA (insn)((&h_i_d[INSN_UID (insn)])->autopref_multipass_data)[write]; | |||
5603 | ||||
5604 | gcc_assert (data->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED)((void)(!(data->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED ) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5604, __FUNCTION__), 0 : 0)); | |||
5605 | data->base = NULL_RTX(rtx) 0; | |||
5606 | data->offset = 0; | |||
5607 | /* Set insn entry initialized, but not relevant for auto-prefetcher. */ | |||
5608 | data->status = AUTOPREF_MULTIPASS_DATA_IRRELEVANT; | |||
5609 | ||||
5610 | rtx pat = PATTERN (insn); | |||
5611 | ||||
5612 | /* We have a multi-set insn like a load-multiple or store-multiple. | |||
5613 | We care about these as long as all the memory ops inside the PARALLEL | |||
5614 | have the same base register. We care about the minimum and maximum | |||
5615 | offsets from that base but don't check for the order of those offsets | |||
5616 | within the PARALLEL insn itself. */ | |||
5617 | if (GET_CODE (pat)((enum rtx_code) (pat)->code) == PARALLEL) | |||
5618 | { | |||
5619 | int n_elems = XVECLEN (pat, 0)(((((pat)->u.fld[0]).rt_rtvec))->num_elem); | |||
5620 | ||||
5621 | int i, offset; | |||
5622 | rtx base, prev_base = NULL_RTX(rtx) 0; | |||
5623 | int min_offset = INT_MAX2147483647; | |||
5624 | ||||
5625 | for (i = 0; i < n_elems; i++) | |||
5626 | { | |||
5627 | rtx set = XVECEXP (pat, 0, i)(((((pat)->u.fld[0]).rt_rtvec))->elem[i]); | |||
5628 | if (GET_CODE (set)((enum rtx_code) (set)->code) != SET) | |||
5629 | return; | |||
5630 | ||||
5631 | if (!analyze_set_insn_for_autopref (set, write, &base, &offset)) | |||
5632 | return; | |||
5633 | ||||
5634 | /* Ensure that all memory operations in the PARALLEL use the same | |||
5635 | base register. */ | |||
5636 | if (i > 0 && REGNO (base)(rhs_regno(base)) != REGNO (prev_base)(rhs_regno(prev_base))) | |||
5637 | return; | |||
5638 | prev_base = base; | |||
5639 | min_offset = MIN (min_offset, offset)((min_offset) < (offset) ? (min_offset) : (offset)); | |||
5640 | } | |||
5641 | ||||
5642 | /* If we reached here then we have a valid PARALLEL of multiple memory ops | |||
5643 | with prev_base as the base and min_offset containing the offset. */ | |||
5644 | gcc_assert (prev_base)((void)(!(prev_base) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5644, __FUNCTION__), 0 : 0)); | |||
5645 | data->base = prev_base; | |||
5646 | data->offset = min_offset; | |||
5647 | data->status = AUTOPREF_MULTIPASS_DATA_NORMAL; | |||
5648 | return; | |||
5649 | } | |||
5650 | ||||
5651 | /* Otherwise this is a single set memory operation. */ | |||
5652 | rtx set = single_set (insn); | |||
5653 | if (set == NULL_RTX(rtx) 0) | |||
5654 | return; | |||
5655 | ||||
5656 | if (!analyze_set_insn_for_autopref (set, write, &data->base, | |||
5657 | &data->offset)) | |||
5658 | return; | |||
5659 | ||||
5660 | /* This insn is relevant for the auto-prefetcher. | |||
5661 | The base and offset fields will have been filled in the | |||
5662 | analyze_set_insn_for_autopref call above. */ | |||
5663 | data->status = AUTOPREF_MULTIPASS_DATA_NORMAL; | |||
5664 | } | |||
5665 | ||||
5666 | /* Helper function for rank_for_schedule sorting. */ | |||
5667 | static int | |||
5668 | autopref_rank_for_schedule (const rtx_insn *insn1, const rtx_insn *insn2) | |||
5669 | { | |||
5670 | int r = 0; | |||
5671 | for (int write = 0; write < 2 && !r; ++write) | |||
5672 | { | |||
5673 | autopref_multipass_data_t data1 | |||
5674 | = &INSN_AUTOPREF_MULTIPASS_DATA (insn1)((&h_i_d[INSN_UID (insn1)])->autopref_multipass_data)[write]; | |||
5675 | autopref_multipass_data_t data2 | |||
5676 | = &INSN_AUTOPREF_MULTIPASS_DATA (insn2)((&h_i_d[INSN_UID (insn2)])->autopref_multipass_data)[write]; | |||
5677 | ||||
5678 | if (data1->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED) | |||
5679 | autopref_multipass_init (insn1, write); | |||
5680 | ||||
5681 | if (data2->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED) | |||
5682 | autopref_multipass_init (insn2, write); | |||
5683 | ||||
5684 | int irrel1 = data1->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT; | |||
5685 | int irrel2 = data2->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT; | |||
5686 | ||||
5687 | if (!irrel1 && !irrel2) | |||
5688 | /* Sort memory references from lowest offset to the largest. */ | |||
5689 | r = data1->offset - data2->offset; | |||
5690 | else if (write) | |||
5691 | /* Schedule "irrelevant" insns before memory stores to resolve | |||
5692 | as many producer dependencies of stores as possible. */ | |||
5693 | r = irrel2 - irrel1; | |||
5694 | else | |||
5695 | /* Schedule "irrelevant" insns after memory reads to avoid breaking | |||
5696 | memory read sequences. */ | |||
5697 | r = irrel1 - irrel2; | |||
5698 | } | |||
5699 | ||||
5700 | return r; | |||
5701 | } | |||
5702 | ||||
5703 | /* True if header of debug dump was printed. */ | |||
5704 | static bool autopref_multipass_dfa_lookahead_guard_started_dump_p; | |||
5705 | ||||
5706 | /* Helper for autopref_multipass_dfa_lookahead_guard. | |||
5707 | Return "1" if INSN1 should be delayed in favor of INSN2. */ | |||
5708 | static int | |||
5709 | autopref_multipass_dfa_lookahead_guard_1 (const rtx_insn *insn1, | |||
5710 | const rtx_insn *insn2, int write) | |||
5711 | { | |||
5712 | autopref_multipass_data_t data1 | |||
5713 | = &INSN_AUTOPREF_MULTIPASS_DATA (insn1)((&h_i_d[INSN_UID (insn1)])->autopref_multipass_data)[write]; | |||
5714 | autopref_multipass_data_t data2 | |||
5715 | = &INSN_AUTOPREF_MULTIPASS_DATA (insn2)((&h_i_d[INSN_UID (insn2)])->autopref_multipass_data)[write]; | |||
5716 | ||||
5717 | if (data2->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED) | |||
5718 | autopref_multipass_init (insn2, write); | |||
5719 | if (data2->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT) | |||
5720 | return 0; | |||
5721 | ||||
5722 | if (rtx_equal_p (data1->base, data2->base) | |||
5723 | && data1->offset > data2->offset) | |||
5724 | { | |||
5725 | if (sched_verbose >= 2) | |||
5726 | { | |||
5727 | if (!autopref_multipass_dfa_lookahead_guard_started_dump_p) | |||
5728 | { | |||
5729 | fprintf (sched_dump, | |||
5730 | ";;\t\tnot trying in max_issue due to autoprefetch " | |||
5731 | "model: "); | |||
5732 | autopref_multipass_dfa_lookahead_guard_started_dump_p = true; | |||
5733 | } | |||
5734 | ||||
5735 | fprintf (sched_dump, " %d(%d)", INSN_UID (insn1), INSN_UID (insn2)); | |||
5736 | } | |||
5737 | ||||
5738 | return 1; | |||
5739 | } | |||
5740 | ||||
5741 | return 0; | |||
5742 | } | |||
5743 | ||||
5744 | /* General note: | |||
5745 | ||||
5746 | We could have also hooked autoprefetcher model into | |||
5747 | first_cycle_multipass_backtrack / first_cycle_multipass_issue hooks | |||
5748 | to enable intelligent selection of "[r1+0]=r2; [r1+4]=r3" on the same cycle | |||
5749 | (e.g., once "[r1+0]=r2" is issued in max_issue(), "[r1+4]=r3" gets | |||
5750 | unblocked). We don't bother about this yet because target of interest | |||
5751 | (ARM Cortex-A15) can issue only 1 memory operation per cycle. */ | |||
5752 | ||||
5753 | /* Implementation of first_cycle_multipass_dfa_lookahead_guard hook. | |||
5754 | Return "1" if INSN1 should not be considered in max_issue due to | |||
5755 | auto-prefetcher considerations. */ | |||
5756 | int | |||
5757 | autopref_multipass_dfa_lookahead_guard (rtx_insn *insn1, int ready_index) | |||
5758 | { | |||
5759 | int r = 0; | |||
5760 | ||||
5761 | /* Exit early if the param forbids this or if we're not entering here through | |||
5762 | normal haifa scheduling. This can happen if selective scheduling is | |||
5763 | explicitly enabled. */ | |||
5764 | if (!insn_queue || param_sched_autopref_queue_depthglobal_options.x_param_sched_autopref_queue_depth <= 0) | |||
5765 | return 0; | |||
5766 | ||||
5767 | if (sched_verbose >= 2 && ready_index == 0) | |||
5768 | autopref_multipass_dfa_lookahead_guard_started_dump_p = false; | |||
5769 | ||||
5770 | for (int write = 0; write < 2; ++write) | |||
5771 | { | |||
5772 | autopref_multipass_data_t data1 | |||
5773 | = &INSN_AUTOPREF_MULTIPASS_DATA (insn1)((&h_i_d[INSN_UID (insn1)])->autopref_multipass_data)[write]; | |||
5774 | ||||
5775 | if (data1->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED) | |||
5776 | autopref_multipass_init (insn1, write); | |||
5777 | if (data1->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT) | |||
5778 | continue; | |||
5779 | ||||
5780 | if (ready_index == 0 | |||
5781 | && data1->status == AUTOPREF_MULTIPASS_DATA_DONT_DELAY) | |||
5782 | /* We allow only a single delay on priviledged instructions. | |||
5783 | Doing otherwise would cause infinite loop. */ | |||
5784 | { | |||
5785 | if (sched_verbose >= 2) | |||
5786 | { | |||
5787 | if (!autopref_multipass_dfa_lookahead_guard_started_dump_p) | |||
5788 | { | |||
5789 | fprintf (sched_dump, | |||
5790 | ";;\t\tnot trying in max_issue due to autoprefetch " | |||
5791 | "model: "); | |||
5792 | autopref_multipass_dfa_lookahead_guard_started_dump_p = true; | |||
5793 | } | |||
5794 | ||||
5795 | fprintf (sched_dump, " *%d*", INSN_UID (insn1)); | |||
5796 | } | |||
5797 | continue; | |||
5798 | } | |||
5799 | ||||
5800 | for (int i2 = 0; i2 < ready.n_ready; ++i2) | |||
5801 | { | |||
5802 | rtx_insn *insn2 = get_ready_element (i2); | |||
5803 | if (insn1 == insn2) | |||
5804 | continue; | |||
5805 | r = autopref_multipass_dfa_lookahead_guard_1 (insn1, insn2, write); | |||
5806 | if (r) | |||
5807 | { | |||
5808 | if (ready_index == 0) | |||
5809 | { | |||
5810 | r = -1; | |||
5811 | data1->status = AUTOPREF_MULTIPASS_DATA_DONT_DELAY; | |||
5812 | } | |||
5813 | goto finish; | |||
5814 | } | |||
5815 | } | |||
5816 | ||||
5817 | if (param_sched_autopref_queue_depthglobal_options.x_param_sched_autopref_queue_depth == 1) | |||
5818 | continue; | |||
5819 | ||||
5820 | /* Everything from the current queue slot should have been moved to | |||
5821 | the ready list. */ | |||
5822 | gcc_assert (insn_queue[NEXT_Q_AFTER (q_ptr, 0)] == NULL_RTX)((void)(!(insn_queue[(((q_ptr)+0) & max_insn_queue_index) ] == (rtx) 0) ? fancy_abort ("/buildworker/marxinbox-gcc-clang-static-analyzer/build/gcc/haifa-sched.cc" , 5822, __FUNCTION__), 0 : 0)); | |||
5823 | ||||
5824 | int n_stalls = param_sched_autopref_queue_depthglobal_options.x_param_sched_autopref_queue_depth - 1; | |||
5825 | if (n_stalls > max_insn_queue_index) | |||
5826 | n_stalls = max_insn_queue_index; | |||
5827 | ||||
5828 | for (int stalls = 1; stalls <= n_stalls; ++stalls) | |||
5829 | { | |||
5830 | for (rtx_insn_list *link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)(((q_ptr)+stalls) & max_insn_queue_index)]; | |||
5831 | link != NULL_RTX(rtx) 0; | |||
5832 | link = link->next ()) | |||
5833 | { | |||
5834 | rtx_insn *insn2 = link->insn (); | |||
5835 | r = autopref_multipass_dfa_lookahead_guard_1 (insn1, insn2, | |||
5836 | write); | |||
5837 | if (r) | |||
5838 | { | |||
5839 | /* Queue INSN1 until INSN2 can issue. */ | |||
5840 | r = -stalls; | |||