1 | /* Instruction scheduling pass. Selective scheduler and pipeliner. |
2 | Copyright (C) 2006-2024 Free Software Foundation, Inc. |
3 | |
4 | This file is part of GCC. |
5 | |
6 | GCC is free software; you can redistribute it and/or modify it under |
7 | the terms of the GNU General Public License as published by the Free |
8 | Software Foundation; either version 3, or (at your option) any later |
9 | version. |
10 | |
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
14 | for more details. |
15 | |
16 | You should have received a copy of the GNU General Public License |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | #include "config.h" |
21 | #include "system.h" |
22 | #include "coretypes.h" |
23 | #include "backend.h" |
24 | #include "tree.h" |
25 | #include "rtl.h" |
26 | #include "df.h" |
27 | #include "memmodel.h" |
28 | #include "tm_p.h" |
29 | #include "regs.h" |
30 | #include "cfgbuild.h" |
31 | #include "cfgcleanup.h" |
32 | #include "insn-config.h" |
33 | #include "insn-attr.h" |
34 | #include "target.h" |
35 | #include "sched-int.h" |
36 | #include "rtlhooks-def.h" |
37 | #include "ira.h" |
38 | #include "ira-int.h" |
39 | #include "rtl-iter.h" |
40 | |
41 | #ifdef INSN_SCHEDULING |
42 | #include "regset.h" |
43 | #include "cfgloop.h" |
44 | #include "sel-sched-ir.h" |
45 | #include "sel-sched-dump.h" |
46 | #include "sel-sched.h" |
47 | #include "dbgcnt.h" |
48 | #include "function-abi.h" |
49 | |
50 | /* Implementation of selective scheduling approach. |
51 | The below implementation follows the original approach with the following |
52 | changes: |
53 | |
54 | o the scheduler works after register allocation (but can be also tuned |
55 | to work before RA); |
56 | o some instructions are not copied or register renamed; |
57 | o conditional jumps are not moved with code duplication; |
58 | o several jumps in one parallel group are not supported; |
59 | o when pipelining outer loops, code motion through inner loops |
60 | is not supported; |
61 | o control and data speculation are supported; |
62 | o some improvements for better compile time/performance were made. |
63 | |
64 | Terminology |
65 | =========== |
66 | |
67 | A vinsn, or virtual insn, is an insn with additional data characterizing |
68 | insn pattern, such as LHS, RHS, register sets used/set/clobbered, etc. |
69 | Vinsns also act as smart pointers to save memory by reusing them in |
70 | different expressions. A vinsn is described by vinsn_t type. |
71 | |
72 | An expression is a vinsn with additional data characterizing its properties |
73 | at some point in the control flow graph. The data may be its usefulness, |
74 | priority, speculative status, whether it was renamed/subsituted, etc. |
75 | An expression is described by expr_t type. |
76 | |
77 | Availability set (av_set) is a set of expressions at a given control flow |
78 | point. It is represented as av_set_t. The expressions in av sets are kept |
79 | sorted in the terms of expr_greater_p function. It allows to truncate |
80 | the set while leaving the best expressions. |
81 | |
82 | A fence is a point through which code motion is prohibited. On each step, |
83 | we gather a parallel group of insns at a fence. It is possible to have |
84 | multiple fences. A fence is represented via fence_t. |
85 | |
86 | A boundary is the border between the fence group and the rest of the code. |
87 | Currently, we never have more than one boundary per fence, as we finalize |
88 | the fence group when a jump is scheduled. A boundary is represented |
89 | via bnd_t. |
90 | |
91 | High-level overview |
92 | =================== |
93 | |
94 | The scheduler finds regions to schedule, schedules each one, and finalizes. |
95 | The regions are formed starting from innermost loops, so that when the inner |
96 | loop is pipelined, its prologue can be scheduled together with yet unprocessed |
97 | outer loop. The rest of acyclic regions are found using extend_rgns: |
98 | the blocks that are not yet allocated to any regions are traversed in top-down |
99 | order, and a block is added to a region to which all its predecessors belong; |
100 | otherwise, the block starts its own region. |
101 | |
102 | The main scheduling loop (sel_sched_region_2) consists of just |
103 | scheduling on each fence and updating fences. For each fence, |
104 | we fill a parallel group of insns (fill_insns) until some insns can be added. |
105 | First, we compute available exprs (av-set) at the boundary of the current |
106 | group. Second, we choose the best expression from it. If the stall is |
107 | required to schedule any of the expressions, we advance the current cycle |
108 | appropriately. So, the final group does not exactly correspond to a VLIW |
109 | word. Third, we move the chosen expression to the boundary (move_op) |
110 | and update the intermediate av sets and liveness sets. We quit fill_insns |
111 | when either no insns left for scheduling or we have scheduled enough insns |
112 | so we feel like advancing a scheduling point. |
113 | |
114 | Computing available expressions |
115 | =============================== |
116 | |
117 | The computation (compute_av_set) is a bottom-up traversal. At each insn, |
118 | we're moving the union of its successors' sets through it via |
119 | moveup_expr_set. The dependent expressions are removed. Local |
120 | transformations (substitution, speculation) are applied to move more |
121 | exprs. Then the expr corresponding to the current insn is added. |
122 | The result is saved on each basic block header. |
123 | |
124 | When traversing the CFG, we're moving down for no more than max_ws insns. |
125 | Also, we do not move down to ineligible successors (is_ineligible_successor), |
126 | which include moving along a back-edge, moving to already scheduled code, |
127 | and moving to another fence. The first two restrictions are lifted during |
128 | pipelining, which allows us to move insns along a back-edge. We always have |
129 | an acyclic region for scheduling because we forbid motion through fences. |
130 | |
131 | Choosing the best expression |
132 | ============================ |
133 | |
134 | We sort the final availability set via sel_rank_for_schedule, then we remove |
135 | expressions which are not yet ready (tick_check_p) or which dest registers |
136 | cannot be used. For some of them, we choose another register via |
137 | find_best_reg. To do this, we run find_used_regs to calculate the set of |
138 | registers which cannot be used. The find_used_regs function performs |
139 | a traversal of code motion paths for an expr. We consider for renaming |
140 | only registers which are from the same regclass as the original one and |
141 | using which does not interfere with any live ranges. Finally, we convert |
142 | the resulting set to the ready list format and use max_issue and reorder* |
143 | hooks similarly to the Haifa scheduler. |
144 | |
145 | Scheduling the best expression |
146 | ============================== |
147 | |
148 | We run the move_op routine to perform the same type of code motion paths |
149 | traversal as in find_used_regs. (These are working via the same driver, |
150 | code_motion_path_driver.) When moving down the CFG, we look for original |
151 | instruction that gave birth to a chosen expression. We undo |
152 | the transformations performed on an expression via the history saved in it. |
153 | When found, we remove the instruction or leave a reg-reg copy/speculation |
154 | check if needed. On a way up, we insert bookkeeping copies at each join |
155 | point. If a copy is not needed, it will be removed later during this |
156 | traversal. We update the saved av sets and liveness sets on the way up, too. |
157 | |
158 | Finalizing the schedule |
159 | ======================= |
160 | |
161 | When pipelining, we reschedule the blocks from which insns were pipelined |
162 | to get a tighter schedule. On Itanium, we also perform bundling via |
163 | the same routine from ia64.cc. |
164 | |
165 | Dependence analysis changes |
166 | =========================== |
167 | |
168 | We augmented the sched-deps.cc with hooks that get called when a particular |
169 | dependence is found in a particular part of an insn. Using these hooks, we |
170 | can do several actions such as: determine whether an insn can be moved through |
171 | another (has_dependence_p, moveup_expr); find out whether an insn can be |
172 | scheduled on the current cycle (tick_check_p); find out registers that |
173 | are set/used/clobbered by an insn and find out all the strange stuff that |
174 | restrict its movement, like SCHED_GROUP_P or CANT_MOVE (done in |
175 | init_global_and_expr_for_insn). |
176 | |
177 | Initialization changes |
178 | ====================== |
179 | |
180 | There are parts of haifa-sched.cc, sched-deps.cc, and sched-rgn.cc that are |
181 | reused in all of the schedulers. We have split up the initialization of data |
182 | of such parts into different functions prefixed with scheduler type and |
183 | postfixed with the type of data initialized: {,sel_,haifa_}sched_{init,finish}, |
184 | sched_rgn_init/finish, sched_deps_init/finish, sched_init_{luids/bbs}, etc. |
185 | The same splitting is done with current_sched_info structure: |
186 | dependence-related parts are in sched_deps_info, common part is in |
187 | common_sched_info, and haifa/sel/etc part is in current_sched_info. |
188 | |
189 | Target contexts |
190 | =============== |
191 | |
192 | As we now have multiple-point scheduling, this would not work with backends |
193 | which save some of the scheduler state to use it in the target hooks. |
194 | For this purpose, we introduce a concept of target contexts, which |
195 | encapsulate such information. The backend should implement simple routines |
196 | of allocating/freeing/setting such a context. The scheduler calls these |
197 | as target hooks and handles the target context as an opaque pointer (similar |
198 | to the DFA state type, state_t). |
199 | |
200 | Various speedups |
201 | ================ |
202 | |
203 | As the correct data dependence graph is not supported during scheduling (which |
204 | is to be changed in mid-term), we cache as much of the dependence analysis |
205 | results as possible to avoid reanalyzing. This includes: bitmap caches on |
206 | each insn in stream of the region saying yes/no for a query with a pair of |
207 | UIDs; hashtables with the previously done transformations on each insn in |
208 | stream; a vector keeping a history of transformations on each expr. |
209 | |
210 | Also, we try to minimize the dependence context used on each fence to check |
211 | whether the given expression is ready for scheduling by removing from it |
212 | insns that are definitely completed the execution. The results of |
213 | tick_check_p checks are also cached in a vector on each fence. |
214 | |
215 | We keep a valid liveness set on each insn in a region to avoid the high |
216 | cost of recomputation on large basic blocks. |
217 | |
218 | Finally, we try to minimize the number of needed updates to the availability |
219 | sets. The updates happen in two cases: when fill_insns terminates, |
220 | we advance all fences and increase the stage number to show that the region |
221 | has changed and the sets are to be recomputed; and when the next iteration |
222 | of a loop in fill_insns happens (but this one reuses the saved av sets |
223 | on bb headers.) Thus, we try to break the fill_insns loop only when |
224 | "significant" number of insns from the current scheduling window was |
225 | scheduled. This should be made a target param. |
226 | |
227 | |
228 | TODO: correctly support the data dependence graph at all stages and get rid |
229 | of all caches. This should speed up the scheduler. |
230 | TODO: implement moving cond jumps with bookkeeping copies on both targets. |
231 | TODO: tune the scheduler before RA so it does not create too much pseudos. |
232 | |
233 | |
234 | References: |
235 | S.-M. Moon and K. Ebcioglu. Parallelizing nonnumerical code with |
236 | selective scheduling and software pipelining. |
237 | ACM TOPLAS, Vol 19, No. 6, pages 853--898, Nov. 1997. |
238 | |
239 | Andrey Belevantsev, Maxim Kuvyrkov, Vladimir Makarov, Dmitry Melnik, |
240 | and Dmitry Zhurikhin. An interblock VLIW-targeted instruction scheduler |
241 | for GCC. In Proceedings of GCC Developers' Summit 2006. |
242 | |
243 | Arutyun Avetisyan, Andrey Belevantsev, and Dmitry Melnik. GCC Instruction |
244 | Scheduler and Software Pipeliner on the Itanium Platform. EPIC-7 Workshop. |
245 | http://rogue.colorado.edu/EPIC7/. |
246 | |
247 | */ |
248 | |
249 | /* True when pipelining is enabled. */ |
250 | bool pipelining_p; |
251 | |
252 | /* True if bookkeeping is enabled. */ |
253 | bool bookkeeping_p; |
254 | |
255 | /* Maximum number of insns that are eligible for renaming. */ |
256 | int max_insns_to_rename; |
257 | |
258 | |
259 | /* Definitions of local types and macros. */ |
260 | |
261 | /* Represents possible outcomes of moving an expression through an insn. */ |
262 | enum MOVEUP_EXPR_CODE |
263 | { |
264 | /* The expression is not changed. */ |
265 | MOVEUP_EXPR_SAME, |
266 | |
267 | /* Not changed, but requires a new destination register. */ |
268 | MOVEUP_EXPR_AS_RHS, |
269 | |
270 | /* Cannot be moved. */ |
271 | MOVEUP_EXPR_NULL, |
272 | |
273 | /* Changed (substituted or speculated). */ |
274 | MOVEUP_EXPR_CHANGED |
275 | }; |
276 | |
277 | /* The container to be passed into rtx search & replace functions. */ |
278 | struct rtx_search_arg |
279 | { |
280 | /* What we are searching for. */ |
281 | rtx x; |
282 | |
283 | /* The occurrence counter. */ |
284 | int n; |
285 | }; |
286 | |
287 | typedef struct rtx_search_arg *rtx_search_arg_p; |
288 | |
289 | /* This struct contains precomputed hard reg sets that are needed when |
290 | computing registers available for renaming. */ |
291 | struct hard_regs_data |
292 | { |
293 | /* For every mode, this stores registers available for use with |
294 | that mode. */ |
295 | HARD_REG_SET regs_for_mode[NUM_MACHINE_MODES]; |
296 | |
297 | /* True when regs_for_mode[mode] is initialized. */ |
298 | bool regs_for_mode_ok[NUM_MACHINE_MODES]; |
299 | |
300 | /* For every register, it has regs that are ok to rename into it. |
301 | The register in question is always set. If not, this means |
302 | that the whole set is not computed yet. */ |
303 | HARD_REG_SET regs_for_rename[FIRST_PSEUDO_REGISTER]; |
304 | |
305 | /* All registers that are used or call used. */ |
306 | HARD_REG_SET regs_ever_used; |
307 | |
308 | #ifdef STACK_REGS |
309 | /* Stack registers. */ |
310 | HARD_REG_SET stack_regs; |
311 | #endif |
312 | }; |
313 | |
314 | /* Holds the results of computation of available for renaming and |
315 | unavailable hard registers. */ |
316 | struct reg_rename |
317 | { |
318 | /* These are unavailable due to calls crossing, globalness, etc. */ |
319 | HARD_REG_SET unavailable_hard_regs; |
320 | |
321 | /* These are *available* for renaming. */ |
322 | HARD_REG_SET available_for_renaming; |
323 | |
324 | /* The set of ABIs used by calls that the code motion path crosses. */ |
325 | unsigned int crossed_call_abis : NUM_ABI_IDS; |
326 | }; |
327 | |
328 | /* A global structure that contains the needed information about harg |
329 | regs. */ |
330 | static struct hard_regs_data sel_hrd; |
331 | |
332 | |
333 | /* This structure holds local data used in code_motion_path_driver hooks on |
334 | the same or adjacent levels of recursion. Here we keep those parameters |
335 | that are not used in code_motion_path_driver routine itself, but only in |
336 | its hooks. Moreover, all parameters that can be modified in hooks are |
337 | in this structure, so all other parameters passed explicitly to hooks are |
338 | read-only. */ |
339 | struct cmpd_local_params |
340 | { |
341 | /* Local params used in move_op_* functions. */ |
342 | |
343 | /* Edges for bookkeeping generation. */ |
344 | edge e1, e2; |
345 | |
346 | /* C_EXPR merged from all successors and locally allocated temporary C_EXPR. */ |
347 | expr_t c_expr_merged, c_expr_local; |
348 | |
349 | /* Local params used in fur_* functions. */ |
350 | /* Copy of the ORIGINAL_INSN list, stores the original insns already |
351 | found before entering the current level of code_motion_path_driver. */ |
352 | def_list_t old_original_insns; |
353 | |
354 | /* Local params used in move_op_* functions. */ |
355 | /* True when we have removed last insn in the block which was |
356 | also a boundary. Do not update anything or create bookkeeping copies. */ |
357 | BOOL_BITFIELD removed_last_insn : 1; |
358 | }; |
359 | |
360 | /* Stores the static parameters for move_op_* calls. */ |
361 | struct moveop_static_params |
362 | { |
363 | /* Destination register. */ |
364 | rtx dest; |
365 | |
366 | /* Current C_EXPR. */ |
367 | expr_t c_expr; |
368 | |
369 | /* An UID of expr_vliw which is to be moved up. If we find other exprs, |
370 | they are to be removed. */ |
371 | int uid; |
372 | |
373 | /* This is initialized to the insn on which the driver stopped its traversal. */ |
374 | insn_t failed_insn; |
375 | |
376 | /* True if we scheduled an insn with different register. */ |
377 | bool was_renamed; |
378 | }; |
379 | |
380 | /* Stores the static parameters for fur_* calls. */ |
381 | struct fur_static_params |
382 | { |
383 | /* Set of registers unavailable on the code motion path. */ |
384 | regset used_regs; |
385 | |
386 | /* Pointer to the list of original insns definitions. */ |
387 | def_list_t *original_insns; |
388 | |
389 | /* The set of ABIs used by calls that the code motion path crosses. */ |
390 | unsigned int crossed_call_abis : NUM_ABI_IDS; |
391 | }; |
392 | |
393 | typedef struct fur_static_params *fur_static_params_p; |
394 | typedef struct cmpd_local_params *cmpd_local_params_p; |
395 | typedef struct moveop_static_params *moveop_static_params_p; |
396 | |
397 | /* Set of hooks and parameters that determine behavior specific to |
398 | move_op or find_used_regs functions. */ |
399 | struct code_motion_path_driver_info_def |
400 | { |
401 | /* Called on enter to the basic block. */ |
402 | int (*on_enter) (insn_t, cmpd_local_params_p, void *, bool); |
403 | |
404 | /* Called when original expr is found. */ |
405 | void (*orig_expr_found) (insn_t, expr_t, cmpd_local_params_p, void *); |
406 | |
407 | /* Called while descending current basic block if current insn is not |
408 | the original EXPR we're searching for. */ |
409 | bool (*orig_expr_not_found) (insn_t, av_set_t, void *); |
410 | |
411 | /* Function to merge C_EXPRes from different successors. */ |
412 | void (*merge_succs) (insn_t, insn_t, int, cmpd_local_params_p, void *); |
413 | |
414 | /* Function to finalize merge from different successors and possibly |
415 | deallocate temporary data structures used for merging. */ |
416 | void (*after_merge_succs) (cmpd_local_params_p, void *); |
417 | |
418 | /* Called on the backward stage of recursion to do moveup_expr. |
419 | Used only with move_op_*. */ |
420 | void (*ascend) (insn_t, void *); |
421 | |
422 | /* Called on the ascending pass, before returning from the current basic |
423 | block or from the whole traversal. */ |
424 | void (*at_first_insn) (insn_t, cmpd_local_params_p, void *); |
425 | |
426 | /* When processing successors in move_op we need only descend into |
427 | SUCCS_NORMAL successors, while in find_used_regs we need SUCCS_ALL. */ |
428 | int succ_flags; |
429 | |
430 | /* The routine name to print in dumps ("move_op" of "find_used_regs"). */ |
431 | const char *routine_name; |
432 | }; |
433 | |
434 | /* Global pointer to current hooks, either points to MOVE_OP_HOOKS or |
435 | FUR_HOOKS. */ |
436 | struct code_motion_path_driver_info_def *code_motion_path_driver_info; |
437 | |
438 | /* Set of hooks for performing move_op and find_used_regs routines with |
439 | code_motion_path_driver. */ |
440 | extern struct code_motion_path_driver_info_def move_op_hooks, fur_hooks; |
441 | |
442 | /* True if/when we want to emulate Haifa scheduler in the common code. |
443 | This is used in sched_rgn_local_init and in various places in |
444 | sched-deps.cc. */ |
445 | int sched_emulate_haifa_p; |
446 | |
447 | /* GLOBAL_LEVEL is used to discard information stored in basic block headers |
448 | av_sets. Av_set of bb header is valid if its (bb header's) level is equal |
449 | to GLOBAL_LEVEL. And invalid if lesser. This is primarily used to advance |
450 | scheduling window. */ |
451 | int global_level; |
452 | |
453 | /* Current fences. */ |
454 | flist_t fences; |
455 | |
456 | /* True when separable insns should be scheduled as RHSes. */ |
457 | static bool enable_schedule_as_rhs_p; |
458 | |
459 | /* Used in verify_target_availability to assert that target reg is reported |
460 | unavailabile by both TARGET_UNAVAILABLE and find_used_regs only if |
461 | we haven't scheduled anything on the previous fence. |
462 | if scheduled_something_on_previous_fence is true, TARGET_UNAVAILABLE can |
463 | have more conservative value than the one returned by the |
464 | find_used_regs, thus we shouldn't assert that these values are equal. */ |
465 | static bool scheduled_something_on_previous_fence; |
466 | |
467 | /* All newly emitted insns will have their uids greater than this value. */ |
468 | static int first_emitted_uid; |
469 | |
470 | /* Set of basic blocks that are forced to start new ebbs. This is a subset |
471 | of all the ebb heads. */ |
472 | bitmap forced_ebb_heads; |
473 | |
474 | /* Blocks that need to be rescheduled after pipelining. */ |
475 | bitmap blocks_to_reschedule = NULL; |
476 | |
477 | /* True when the first lv set should be ignored when updating liveness. */ |
478 | static bool ignore_first = false; |
479 | |
480 | /* Number of insns max_issue has initialized data structures for. */ |
481 | static int max_issue_size = 0; |
482 | |
483 | /* Whether we can issue more instructions. */ |
484 | static int can_issue_more; |
485 | |
486 | /* Maximum software lookahead window size, reduced when rescheduling after |
487 | pipelining. */ |
488 | static int max_ws; |
489 | |
490 | /* Number of insns scheduled in current region. */ |
491 | static int num_insns_scheduled; |
492 | |
493 | /* A vector of expressions is used to be able to sort them. */ |
494 | static vec<expr_t> vec_av_set; |
495 | |
496 | /* A vector of vinsns is used to hold temporary lists of vinsns. */ |
497 | typedef vec<vinsn_t> vinsn_vec_t; |
498 | |
499 | /* This vector has the exprs which may still present in av_sets, but actually |
500 | can't be moved up due to bookkeeping created during code motion to another |
501 | fence. See comment near the call to update_and_record_unavailable_insns |
502 | for the detailed explanations. */ |
503 | static vinsn_vec_t vec_bookkeeping_blocked_vinsns = vinsn_vec_t (); |
504 | |
505 | /* This vector has vinsns which are scheduled with renaming on the first fence |
506 | and then seen on the second. For expressions with such vinsns, target |
507 | availability information may be wrong. */ |
508 | static vinsn_vec_t vec_target_unavailable_vinsns = vinsn_vec_t (); |
509 | |
510 | /* Vector to store temporary nops inserted in move_op to prevent removal |
511 | of empty bbs. */ |
512 | static vec<insn_t> vec_temp_moveop_nops; |
513 | |
514 | /* These bitmaps record original instructions scheduled on the current |
515 | iteration and bookkeeping copies created by them. */ |
516 | static bitmap current_originators = NULL; |
517 | static bitmap current_copies = NULL; |
518 | |
519 | /* This bitmap marks the blocks visited by code_motion_path_driver so we don't |
520 | visit them afterwards. */ |
521 | static bitmap code_motion_visited_blocks = NULL; |
522 | |
523 | /* Variables to accumulate different statistics. */ |
524 | |
525 | /* The number of bookkeeping copies created. */ |
526 | static int stat_bookkeeping_copies; |
527 | |
528 | /* The number of insns that required bookkeeiping for their scheduling. */ |
529 | static int stat_insns_needed_bookkeeping; |
530 | |
531 | /* The number of insns that got renamed. */ |
532 | static int stat_renamed_scheduled; |
533 | |
534 | /* The number of substitutions made during scheduling. */ |
535 | static int stat_substitutions_total; |
536 | |
537 | |
538 | /* Forward declarations of static functions. */ |
539 | static bool rtx_ok_for_substitution_p (rtx, rtx); |
540 | static int sel_rank_for_schedule (const void *, const void *); |
541 | static av_set_t find_sequential_best_exprs (bnd_t, expr_t, bool); |
542 | static basic_block find_block_for_bookkeeping (edge e1, edge e2, bool lax); |
543 | |
544 | static rtx get_dest_from_orig_ops (av_set_t); |
545 | static basic_block generate_bookkeeping_insn (expr_t, edge, edge); |
546 | static bool find_used_regs (insn_t, av_set_t, regset, struct reg_rename *, |
547 | def_list_t *); |
548 | static bool move_op (insn_t, av_set_t, expr_t, rtx, expr_t, bool*); |
549 | static int code_motion_path_driver (insn_t, av_set_t, ilist_t, |
550 | cmpd_local_params_p, void *); |
551 | static void sel_sched_region_1 (void); |
552 | static void sel_sched_region_2 (int); |
553 | static av_set_t compute_av_set_inside_bb (insn_t, ilist_t, int, bool); |
554 | |
555 | static void debug_state (state_t); |
556 | |
557 | |
558 | /* Functions that work with fences. */ |
559 | |
560 | /* Advance one cycle on FENCE. */ |
561 | static void |
562 | advance_one_cycle (fence_t fence) |
563 | { |
564 | unsigned i; |
565 | int cycle; |
566 | rtx_insn *insn; |
567 | |
568 | advance_state (FENCE_STATE (fence)); |
569 | cycle = ++FENCE_CYCLE (fence); |
570 | FENCE_ISSUED_INSNS (fence) = 0; |
571 | FENCE_STARTS_CYCLE_P (fence) = 1; |
572 | can_issue_more = issue_rate; |
573 | FENCE_ISSUE_MORE (fence) = can_issue_more; |
574 | |
575 | for (i = 0; vec_safe_iterate (FENCE_EXECUTING_INSNS (fence), ix: i, ptr: &insn); ) |
576 | { |
577 | if (INSN_READY_CYCLE (insn) < cycle) |
578 | { |
579 | remove_from_deps (FENCE_DC (fence), insn); |
580 | FENCE_EXECUTING_INSNS (fence)->unordered_remove (ix: i); |
581 | continue; |
582 | } |
583 | i++; |
584 | } |
585 | if (sched_verbose >= 2) |
586 | { |
587 | sel_print (fmt: "Finished a cycle. Current cycle = %d\n" , FENCE_CYCLE (fence)); |
588 | debug_state (FENCE_STATE (fence)); |
589 | } |
590 | } |
591 | |
592 | /* Returns true when SUCC in a fallthru bb of INSN, possibly |
593 | skipping empty basic blocks. */ |
594 | static bool |
595 | in_fallthru_bb_p (rtx_insn *insn, rtx succ) |
596 | { |
597 | basic_block bb = BLOCK_FOR_INSN (insn); |
598 | edge e; |
599 | |
600 | if (bb == BLOCK_FOR_INSN (insn: succ)) |
601 | return true; |
602 | |
603 | e = find_fallthru_edge_from (bb); |
604 | if (e) |
605 | bb = e->dest; |
606 | else |
607 | return false; |
608 | |
609 | while (sel_bb_empty_p (bb)) |
610 | bb = bb->next_bb; |
611 | |
612 | return bb == BLOCK_FOR_INSN (insn: succ); |
613 | } |
614 | |
615 | /* Construct successor fences from OLD_FENCEs and put them in NEW_FENCES. |
616 | When a successor will continue a ebb, transfer all parameters of a fence |
617 | to the new fence. ORIG_MAX_SEQNO is the maximal seqno before this round |
618 | of scheduling helping to distinguish between the old and the new code. */ |
619 | static void |
620 | (flist_t old_fences, flist_tail_t new_fences, |
621 | int orig_max_seqno) |
622 | { |
623 | bool was_here_p = false; |
624 | insn_t insn = NULL; |
625 | insn_t succ; |
626 | succ_iterator si; |
627 | ilist_iterator ii; |
628 | fence_t fence = FLIST_FENCE (old_fences); |
629 | basic_block bb; |
630 | |
631 | /* Get the only element of FENCE_BNDS (fence). */ |
632 | FOR_EACH_INSN (insn, ii, FENCE_BNDS (fence)) |
633 | { |
634 | gcc_assert (!was_here_p); |
635 | was_here_p = true; |
636 | } |
637 | gcc_assert (was_here_p && insn != NULL_RTX); |
638 | |
639 | /* When in the "middle" of the block, just move this fence |
640 | to the new list. */ |
641 | bb = BLOCK_FOR_INSN (insn); |
642 | if (! sel_bb_end_p (insn) |
643 | || (single_succ_p (bb) |
644 | && single_pred_p (bb: single_succ (bb)))) |
645 | { |
646 | insn_t succ; |
647 | |
648 | succ = (sel_bb_end_p (insn) |
649 | ? sel_bb_head (single_succ (bb)) |
650 | : NEXT_INSN (insn)); |
651 | |
652 | if (INSN_SEQNO (succ) > 0 |
653 | && INSN_SEQNO (succ) <= orig_max_seqno |
654 | && INSN_SCHED_TIMES (succ) <= 0) |
655 | { |
656 | FENCE_INSN (fence) = succ; |
657 | move_fence_to_fences (old_fences, new_fences); |
658 | |
659 | if (sched_verbose >= 1) |
660 | sel_print (fmt: "Fence %d continues as %d[%d] (state continue)\n" , |
661 | INSN_UID (insn), INSN_UID (insn: succ), BLOCK_NUM (succ)); |
662 | } |
663 | return; |
664 | } |
665 | |
666 | /* Otherwise copy fence's structures to (possibly) multiple successors. */ |
667 | FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
668 | { |
669 | int seqno = INSN_SEQNO (succ); |
670 | |
671 | if (seqno > 0 && seqno <= orig_max_seqno |
672 | && (pipelining_p || INSN_SCHED_TIMES (succ) <= 0)) |
673 | { |
674 | bool b = (in_same_ebb_p (insn, succ) |
675 | || in_fallthru_bb_p (insn, succ)); |
676 | |
677 | if (sched_verbose >= 1) |
678 | sel_print (fmt: "Fence %d continues as %d[%d] (state %s)\n" , |
679 | INSN_UID (insn), INSN_UID (insn: succ), |
680 | BLOCK_NUM (succ), b ? "continue" : "reset" ); |
681 | |
682 | if (b) |
683 | add_dirty_fence_to_fences (new_fences, succ, fence); |
684 | else |
685 | { |
686 | /* Mark block of the SUCC as head of the new ebb. */ |
687 | bitmap_set_bit (forced_ebb_heads, BLOCK_NUM (succ)); |
688 | add_clean_fence_to_fences (new_fences, succ, fence); |
689 | } |
690 | } |
691 | } |
692 | } |
693 | |
694 | |
695 | /* Functions to support substitution. */ |
696 | |
697 | /* Returns whether INSN with dependence status DS is eligible for |
698 | substitution, i.e. it's a copy operation x := y, and RHS that is |
699 | moved up through this insn should be substituted. */ |
700 | static bool |
701 | can_substitute_through_p (insn_t insn, ds_t ds) |
702 | { |
703 | /* We can substitute only true dependencies. */ |
704 | if ((ds & DEP_OUTPUT) |
705 | || (ds & DEP_ANTI) |
706 | || ! INSN_RHS (insn) |
707 | || ! INSN_LHS (insn)) |
708 | return false; |
709 | |
710 | /* Now we just need to make sure the INSN_RHS consists of only one |
711 | simple REG rtx. */ |
712 | if (REG_P (INSN_LHS (insn)) |
713 | && REG_P (INSN_RHS (insn))) |
714 | return true; |
715 | return false; |
716 | } |
717 | |
718 | /* Substitute all occurrences of INSN's destination in EXPR' vinsn with INSN's |
719 | source (if INSN is eligible for substitution). Returns TRUE if |
720 | substitution was actually performed, FALSE otherwise. Substitution might |
721 | be not performed because it's either EXPR' vinsn doesn't contain INSN's |
722 | destination or the resulting insn is invalid for the target machine. |
723 | When UNDO is true, perform unsubstitution instead (the difference is in |
724 | the part of rtx on which validate_replace_rtx is called). */ |
725 | static bool |
726 | substitute_reg_in_expr (expr_t expr, insn_t insn, bool undo) |
727 | { |
728 | rtx *where; |
729 | bool new_insn_valid; |
730 | vinsn_t *vi = &EXPR_VINSN (expr); |
731 | bool has_rhs = VINSN_RHS (*vi) != NULL; |
732 | rtx old, new_rtx; |
733 | |
734 | /* Do not try to replace in SET_DEST. Although we'll choose new |
735 | register for the RHS, we don't want to change RHS' original reg. |
736 | If the insn is not SET, we may still be able to substitute something |
737 | in it, and if we're here (don't have deps), it doesn't write INSN's |
738 | dest. */ |
739 | where = (has_rhs |
740 | ? &VINSN_RHS (*vi) |
741 | : &PATTERN (VINSN_INSN_RTX (*vi))); |
742 | old = undo ? INSN_RHS (insn) : INSN_LHS (insn); |
743 | |
744 | /* Substitute if INSN has a form of x:=y and LHS(INSN) occurs in *VI. */ |
745 | if (rtx_ok_for_substitution_p (old, *where)) |
746 | { |
747 | rtx_insn *new_insn; |
748 | rtx *where_replace; |
749 | |
750 | /* We should copy these rtxes before substitution. */ |
751 | new_rtx = copy_rtx (undo ? INSN_LHS (insn) : INSN_RHS (insn)); |
752 | new_insn = create_copy_of_insn_rtx (VINSN_INSN_RTX (*vi)); |
753 | |
754 | /* Where we'll replace. |
755 | WHERE_REPLACE should point inside NEW_INSN, so INSN_RHS couldn't be |
756 | used instead of SET_SRC. */ |
757 | where_replace = (has_rhs |
758 | ? &SET_SRC (PATTERN (new_insn)) |
759 | : &PATTERN (insn: new_insn)); |
760 | |
761 | new_insn_valid |
762 | = validate_replace_rtx_part_nosimplify (old, new_rtx, where_replace, |
763 | new_insn); |
764 | |
765 | /* ??? Actually, constrain_operands result depends upon choice of |
766 | destination register. E.g. if we allow single register to be an rhs, |
767 | and if we try to move dx=ax(as rhs) through ax=dx, we'll result |
768 | in invalid insn dx=dx, so we'll loose this rhs here. |
769 | Just can't come up with significant testcase for this, so just |
770 | leaving it for now. */ |
771 | if (new_insn_valid) |
772 | { |
773 | change_vinsn_in_expr (expr, |
774 | create_vinsn_from_insn_rtx (new_insn, false)); |
775 | |
776 | /* Do not allow clobbering the address register of speculative |
777 | insns. */ |
778 | if ((EXPR_SPEC_DONE_DS (expr) & SPECULATIVE) |
779 | && register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)), |
780 | expr_dest_reg (expr))) |
781 | EXPR_TARGET_AVAILABLE (expr) = false; |
782 | |
783 | return true; |
784 | } |
785 | else |
786 | return false; |
787 | } |
788 | else |
789 | return false; |
790 | } |
791 | |
792 | /* Return the number of places WHAT appears within WHERE. |
793 | Bail out when we found a reference occupying several hard registers. */ |
794 | static int |
795 | count_occurrences_equiv (const_rtx what, const_rtx where) |
796 | { |
797 | int count = 0; |
798 | subrtx_iterator::array_type array; |
799 | FOR_EACH_SUBRTX (iter, array, where, NONCONST) |
800 | { |
801 | const_rtx x = *iter; |
802 | if (REG_P (x) && REGNO (x) == REGNO (what)) |
803 | { |
804 | /* Bail out if mode is different or more than one register is |
805 | used. */ |
806 | if (GET_MODE (x) != GET_MODE (what) || REG_NREGS (x) > 1) |
807 | return 0; |
808 | count += 1; |
809 | } |
810 | else if (GET_CODE (x) == SUBREG |
811 | && (!REG_P (SUBREG_REG (x)) |
812 | || REGNO (SUBREG_REG (x)) == REGNO (what))) |
813 | /* ??? Do not support substituting regs inside subregs. In that case, |
814 | simplify_subreg will be called by validate_replace_rtx, and |
815 | unsubstitution will fail later. */ |
816 | return 0; |
817 | } |
818 | return count; |
819 | } |
820 | |
821 | /* Returns TRUE if WHAT is found in WHERE rtx tree. */ |
822 | static bool |
823 | rtx_ok_for_substitution_p (rtx what, rtx where) |
824 | { |
825 | return (count_occurrences_equiv (what, where) > 0); |
826 | } |
827 | |
828 | |
829 | /* Functions to support register renaming. */ |
830 | |
831 | /* Substitute VI's set source with REGNO. Returns newly created pattern |
832 | that has REGNO as its source. */ |
833 | static rtx_insn * |
834 | create_insn_rtx_with_rhs (vinsn_t vi, rtx rhs_rtx) |
835 | { |
836 | rtx lhs_rtx; |
837 | rtx pattern; |
838 | rtx_insn *insn_rtx; |
839 | |
840 | lhs_rtx = copy_rtx (VINSN_LHS (vi)); |
841 | |
842 | pattern = gen_rtx_SET (lhs_rtx, rhs_rtx); |
843 | insn_rtx = create_insn_rtx_from_pattern (pattern, NULL_RTX); |
844 | |
845 | return insn_rtx; |
846 | } |
847 | |
848 | /* Returns whether INSN's src can be replaced with register number |
849 | NEW_SRC_REG. E.g. the following insn is valid for i386: |
850 | |
851 | (insn:HI 2205 6585 2207 727 ../../gcc/libiberty/regex.c:3337 |
852 | (set (mem/s:QI (plus:SI (plus:SI (reg/f:SI 7 sp) |
853 | (reg:SI 0 ax [orig:770 c1 ] [770])) |
854 | (const_int 288 [0x120])) [0 str S1 A8]) |
855 | (const_int 0 [0x0])) 43 {*movqi_1} (nil) |
856 | (nil)) |
857 | |
858 | But if we change (const_int 0 [0x0]) to (reg:QI 4 si), it will be invalid |
859 | because of operand constraints: |
860 | |
861 | (define_insn "*movqi_1" |
862 | [(set (match_operand:QI 0 "nonimmediate_operand" "=q,q ,q ,r,r ,?r,m") |
863 | (match_operand:QI 1 "general_operand" " q,qn,qm,q,rn,qm,qn") |
864 | )] |
865 | |
866 | So do constrain_operands here, before choosing NEW_SRC_REG as best |
867 | reg for rhs. */ |
868 | |
869 | static bool |
870 | replace_src_with_reg_ok_p (insn_t insn, rtx new_src_reg) |
871 | { |
872 | vinsn_t vi = INSN_VINSN (insn); |
873 | machine_mode mode; |
874 | rtx dst_loc; |
875 | bool res; |
876 | |
877 | gcc_assert (VINSN_SEPARABLE_P (vi)); |
878 | |
879 | get_dest_and_mode (insn, &dst_loc, &mode); |
880 | gcc_assert (mode == GET_MODE (new_src_reg)); |
881 | |
882 | if (REG_P (dst_loc) && REGNO (new_src_reg) == REGNO (dst_loc)) |
883 | return true; |
884 | |
885 | /* See whether SET_SRC can be replaced with this register. */ |
886 | validate_change (insn, &SET_SRC (PATTERN (insn)), new_src_reg, 1); |
887 | res = verify_changes (0); |
888 | cancel_changes (0); |
889 | |
890 | return res; |
891 | } |
892 | |
893 | /* Returns whether INSN still be valid after replacing it's DEST with |
894 | register NEW_REG. */ |
895 | static bool |
896 | replace_dest_with_reg_ok_p (insn_t insn, rtx new_reg) |
897 | { |
898 | vinsn_t vi = INSN_VINSN (insn); |
899 | bool res; |
900 | |
901 | /* We should deal here only with separable insns. */ |
902 | gcc_assert (VINSN_SEPARABLE_P (vi)); |
903 | gcc_assert (GET_MODE (VINSN_LHS (vi)) == GET_MODE (new_reg)); |
904 | |
905 | /* See whether SET_DEST can be replaced with this register. */ |
906 | validate_change (insn, &SET_DEST (PATTERN (insn)), new_reg, 1); |
907 | res = verify_changes (0); |
908 | cancel_changes (0); |
909 | |
910 | return res; |
911 | } |
912 | |
913 | /* Create a pattern with rhs of VI and lhs of LHS_RTX. */ |
914 | static rtx_insn * |
915 | create_insn_rtx_with_lhs (vinsn_t vi, rtx lhs_rtx) |
916 | { |
917 | rtx rhs_rtx; |
918 | rtx pattern; |
919 | rtx_insn *insn_rtx; |
920 | |
921 | rhs_rtx = copy_rtx (VINSN_RHS (vi)); |
922 | |
923 | pattern = gen_rtx_SET (lhs_rtx, rhs_rtx); |
924 | insn_rtx = create_insn_rtx_from_pattern (pattern, NULL_RTX); |
925 | |
926 | return insn_rtx; |
927 | } |
928 | |
929 | /* Substitute lhs in the given expression EXPR for the register with number |
930 | NEW_REGNO. SET_DEST may be arbitrary rtx, not only register. */ |
931 | static void |
932 | replace_dest_with_reg_in_expr (expr_t expr, rtx new_reg) |
933 | { |
934 | rtx_insn *insn_rtx; |
935 | vinsn_t vinsn; |
936 | |
937 | insn_rtx = create_insn_rtx_with_lhs (EXPR_VINSN (expr), lhs_rtx: new_reg); |
938 | vinsn = create_vinsn_from_insn_rtx (insn_rtx, false); |
939 | |
940 | change_vinsn_in_expr (expr, vinsn); |
941 | EXPR_WAS_RENAMED (expr) = 1; |
942 | EXPR_TARGET_AVAILABLE (expr) = 1; |
943 | } |
944 | |
945 | /* Returns whether VI writes either one of the USED_REGS registers or, |
946 | if a register is a hard one, one of the UNAVAILABLE_HARD_REGS registers. */ |
947 | static bool |
948 | vinsn_writes_one_of_regs_p (vinsn_t vi, regset used_regs, |
949 | HARD_REG_SET unavailable_hard_regs) |
950 | { |
951 | unsigned regno; |
952 | reg_set_iterator rsi; |
953 | |
954 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (vi), 0, regno, rsi) |
955 | { |
956 | if (REGNO_REG_SET_P (used_regs, regno)) |
957 | return true; |
958 | if (HARD_REGISTER_NUM_P (regno) |
959 | && TEST_HARD_REG_BIT (set: unavailable_hard_regs, bit: regno)) |
960 | return true; |
961 | } |
962 | |
963 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (vi), 0, regno, rsi) |
964 | { |
965 | if (REGNO_REG_SET_P (used_regs, regno)) |
966 | return true; |
967 | if (HARD_REGISTER_NUM_P (regno) |
968 | && TEST_HARD_REG_BIT (set: unavailable_hard_regs, bit: regno)) |
969 | return true; |
970 | } |
971 | |
972 | return false; |
973 | } |
974 | |
975 | /* Returns register class of the output register in INSN. |
976 | Returns NO_REGS for call insns because some targets have constraints on |
977 | destination register of a call insn. |
978 | |
979 | Code adopted from regrename.cc::build_def_use. */ |
980 | static enum reg_class |
981 | get_reg_class (rtx_insn *insn) |
982 | { |
983 | int i, n_ops; |
984 | |
985 | extract_constrain_insn (insn); |
986 | preprocess_constraints (insn); |
987 | n_ops = recog_data.n_operands; |
988 | |
989 | const operand_alternative *op_alt = which_op_alt (); |
990 | if (asm_noperands (PATTERN (insn)) > 0) |
991 | { |
992 | for (i = 0; i < n_ops; i++) |
993 | if (recog_data.operand_type[i] == OP_OUT) |
994 | { |
995 | rtx *loc = recog_data.operand_loc[i]; |
996 | rtx op = *loc; |
997 | enum reg_class cl = alternative_class (alt: op_alt, i); |
998 | |
999 | if (REG_P (op) |
1000 | && REGNO (op) == ORIGINAL_REGNO (op)) |
1001 | continue; |
1002 | |
1003 | return cl; |
1004 | } |
1005 | } |
1006 | else if (!CALL_P (insn)) |
1007 | { |
1008 | for (i = 0; i < n_ops + recog_data.n_dups; i++) |
1009 | { |
1010 | int opn = i < n_ops ? i : recog_data.dup_num[i - n_ops]; |
1011 | enum reg_class cl = alternative_class (alt: op_alt, i: opn); |
1012 | |
1013 | if (recog_data.operand_type[opn] == OP_OUT || |
1014 | recog_data.operand_type[opn] == OP_INOUT) |
1015 | return cl; |
1016 | } |
1017 | } |
1018 | |
1019 | /* Insns like |
1020 | (insn (set (reg:CCZ 17 flags) (compare:CCZ ...))) |
1021 | may result in returning NO_REGS, cause flags is written implicitly through |
1022 | CMP insn, which has no OP_OUT | OP_INOUT operands. */ |
1023 | return NO_REGS; |
1024 | } |
1025 | |
1026 | /* Calculate HARD_REGNO_RENAME_OK data for REGNO. */ |
1027 | static void |
1028 | init_hard_regno_rename (int regno) |
1029 | { |
1030 | int cur_reg; |
1031 | |
1032 | SET_HARD_REG_BIT (set&: sel_hrd.regs_for_rename[regno], bit: regno); |
1033 | |
1034 | for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) |
1035 | { |
1036 | /* We are not interested in renaming in other regs. */ |
1037 | if (!TEST_HARD_REG_BIT (set: sel_hrd.regs_ever_used, bit: cur_reg)) |
1038 | continue; |
1039 | |
1040 | if (HARD_REGNO_RENAME_OK (regno, cur_reg)) |
1041 | SET_HARD_REG_BIT (set&: sel_hrd.regs_for_rename[regno], bit: cur_reg); |
1042 | } |
1043 | } |
1044 | |
1045 | /* A wrapper around HARD_REGNO_RENAME_OK that will look into the hard regs |
1046 | data first. */ |
1047 | static inline bool |
1048 | sel_hard_regno_rename_ok (int from ATTRIBUTE_UNUSED, int to ATTRIBUTE_UNUSED) |
1049 | { |
1050 | /* Check whether this is all calculated. */ |
1051 | if (TEST_HARD_REG_BIT (set: sel_hrd.regs_for_rename[from], bit: from)) |
1052 | return TEST_HARD_REG_BIT (set: sel_hrd.regs_for_rename[from], bit: to); |
1053 | |
1054 | init_hard_regno_rename (regno: from); |
1055 | |
1056 | return TEST_HARD_REG_BIT (set: sel_hrd.regs_for_rename[from], bit: to); |
1057 | } |
1058 | |
1059 | /* Calculate set of registers that are capable of holding MODE. */ |
1060 | static void |
1061 | init_regs_for_mode (machine_mode mode) |
1062 | { |
1063 | int cur_reg; |
1064 | |
1065 | CLEAR_HARD_REG_SET (set&: sel_hrd.regs_for_mode[mode]); |
1066 | |
1067 | for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) |
1068 | { |
1069 | int nregs; |
1070 | int i; |
1071 | |
1072 | /* See whether it accepts all modes that occur in |
1073 | original insns. */ |
1074 | if (!targetm.hard_regno_mode_ok (cur_reg, mode)) |
1075 | continue; |
1076 | |
1077 | nregs = hard_regno_nregs (regno: cur_reg, mode); |
1078 | |
1079 | for (i = nregs - 1; i >= 0; --i) |
1080 | if (fixed_regs[cur_reg + i] |
1081 | || global_regs[cur_reg + i] |
1082 | /* Can't use regs which aren't saved by |
1083 | the prologue. */ |
1084 | || !TEST_HARD_REG_BIT (set: sel_hrd.regs_ever_used, bit: cur_reg + i) |
1085 | /* Can't use regs with non-null REG_BASE_VALUE, because adjusting |
1086 | it affects aliasing globally and invalidates all AV sets. */ |
1087 | || get_reg_base_value (cur_reg + i) |
1088 | #ifdef LEAF_REGISTERS |
1089 | /* We can't use a non-leaf register if we're in a |
1090 | leaf function. */ |
1091 | || (crtl->is_leaf |
1092 | && !LEAF_REGISTERS[cur_reg + i]) |
1093 | #endif |
1094 | ) |
1095 | break; |
1096 | |
1097 | if (i >= 0) |
1098 | continue; |
1099 | |
1100 | /* If the CUR_REG passed all the checks above, |
1101 | then it's ok. */ |
1102 | SET_HARD_REG_BIT (set&: sel_hrd.regs_for_mode[mode], bit: cur_reg); |
1103 | } |
1104 | |
1105 | sel_hrd.regs_for_mode_ok[mode] = true; |
1106 | } |
1107 | |
1108 | /* Init all register sets gathered in HRD. */ |
1109 | static void |
1110 | init_hard_regs_data (void) |
1111 | { |
1112 | int cur_reg = 0; |
1113 | int cur_mode = 0; |
1114 | |
1115 | CLEAR_HARD_REG_SET (set&: sel_hrd.regs_ever_used); |
1116 | for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) |
1117 | if (df_regs_ever_live_p (cur_reg) |
1118 | || crtl->abi->clobbers_full_reg_p (regno: cur_reg)) |
1119 | SET_HARD_REG_BIT (set&: sel_hrd.regs_ever_used, bit: cur_reg); |
1120 | |
1121 | /* Initialize registers that are valid based on mode when this is |
1122 | really needed. */ |
1123 | for (cur_mode = 0; cur_mode < NUM_MACHINE_MODES; cur_mode++) |
1124 | sel_hrd.regs_for_mode_ok[cur_mode] = false; |
1125 | |
1126 | /* Mark that all HARD_REGNO_RENAME_OK is not calculated. */ |
1127 | for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) |
1128 | CLEAR_HARD_REG_SET (set&: sel_hrd.regs_for_rename[cur_reg]); |
1129 | |
1130 | #ifdef STACK_REGS |
1131 | CLEAR_HARD_REG_SET (set&: sel_hrd.stack_regs); |
1132 | |
1133 | for (cur_reg = FIRST_STACK_REG; cur_reg <= LAST_STACK_REG; cur_reg++) |
1134 | SET_HARD_REG_BIT (set&: sel_hrd.stack_regs, bit: cur_reg); |
1135 | #endif |
1136 | } |
1137 | |
1138 | /* Mark hardware regs in REG_RENAME_P that are not suitable |
1139 | for renaming rhs in INSN due to hardware restrictions (register class, |
1140 | modes compatibility etc). This doesn't affect original insn's dest reg, |
1141 | if it isn't in USED_REGS. DEF is a definition insn of rhs for which the |
1142 | destination register is sought. LHS (DEF->ORIG_INSN) may be REG or MEM. |
1143 | Registers that are in used_regs are always marked in |
1144 | unavailable_hard_regs as well. */ |
1145 | |
1146 | static void |
1147 | mark_unavailable_hard_regs (def_t def, struct reg_rename *reg_rename_p, |
1148 | regset used_regs ATTRIBUTE_UNUSED) |
1149 | { |
1150 | machine_mode mode; |
1151 | enum reg_class cl = NO_REGS; |
1152 | rtx orig_dest; |
1153 | unsigned cur_reg, regno; |
1154 | hard_reg_set_iterator hrsi; |
1155 | |
1156 | gcc_assert (GET_CODE (PATTERN (def->orig_insn)) == SET); |
1157 | gcc_assert (reg_rename_p); |
1158 | |
1159 | orig_dest = SET_DEST (PATTERN (def->orig_insn)); |
1160 | |
1161 | /* We have decided not to rename 'mem = something;' insns, as 'something' |
1162 | is usually a register. */ |
1163 | if (!REG_P (orig_dest)) |
1164 | return; |
1165 | |
1166 | regno = REGNO (orig_dest); |
1167 | |
1168 | /* If before reload, don't try to work with pseudos. */ |
1169 | if (!reload_completed && !HARD_REGISTER_NUM_P (regno)) |
1170 | return; |
1171 | |
1172 | if (reload_completed) |
1173 | cl = get_reg_class (insn: def->orig_insn); |
1174 | |
1175 | /* Stop if the original register is one of the fixed_regs, global_regs or |
1176 | frame pointer, or we could not discover its class. */ |
1177 | if (fixed_regs[regno] |
1178 | || global_regs[regno] |
1179 | || (!HARD_FRAME_POINTER_IS_FRAME_POINTER && frame_pointer_needed |
1180 | && regno == HARD_FRAME_POINTER_REGNUM) |
1181 | || (HARD_FRAME_POINTER_IS_FRAME_POINTER && frame_pointer_needed |
1182 | && regno == FRAME_POINTER_REGNUM) |
1183 | || (reload_completed && cl == NO_REGS)) |
1184 | { |
1185 | SET_HARD_REG_SET (reg_rename_p->unavailable_hard_regs); |
1186 | |
1187 | /* Give a chance for original register, if it isn't in used_regs. */ |
1188 | if (!def->crossed_call_abis) |
1189 | CLEAR_HARD_REG_BIT (set&: reg_rename_p->unavailable_hard_regs, bit: regno); |
1190 | |
1191 | return; |
1192 | } |
1193 | |
1194 | /* If something allocated on stack in this function, mark frame pointer |
1195 | register unavailable, considering also modes. |
1196 | FIXME: it is enough to do this once per all original defs. */ |
1197 | if (frame_pointer_needed) |
1198 | { |
1199 | add_to_hard_reg_set (regs: ®_rename_p->unavailable_hard_regs, |
1200 | Pmode, FRAME_POINTER_REGNUM); |
1201 | |
1202 | if (!HARD_FRAME_POINTER_IS_FRAME_POINTER) |
1203 | add_to_hard_reg_set (regs: ®_rename_p->unavailable_hard_regs, |
1204 | Pmode, HARD_FRAME_POINTER_REGNUM); |
1205 | } |
1206 | |
1207 | #ifdef STACK_REGS |
1208 | /* For the stack registers the presence of FIRST_STACK_REG in USED_REGS |
1209 | is equivalent to as if all stack regs were in this set. |
1210 | I.e. no stack register can be renamed, and even if it's an original |
1211 | register here we make sure it won't be lifted over it's previous def |
1212 | (it's previous def will appear as if it's a FIRST_STACK_REG def. |
1213 | The HARD_REGNO_RENAME_OK covers other cases in condition below. */ |
1214 | if (IN_RANGE (REGNO (orig_dest), FIRST_STACK_REG, LAST_STACK_REG) |
1215 | && REGNO_REG_SET_P (used_regs, FIRST_STACK_REG)) |
1216 | reg_rename_p->unavailable_hard_regs |= sel_hrd.stack_regs; |
1217 | #endif |
1218 | |
1219 | mode = GET_MODE (orig_dest); |
1220 | |
1221 | /* If there's a call on this path, make regs from full_reg_clobbers |
1222 | unavailable. |
1223 | |
1224 | ??? It would be better to track the set of clobbered registers |
1225 | directly, but that would be quite expensive in a def_t. */ |
1226 | if (def->crossed_call_abis) |
1227 | reg_rename_p->unavailable_hard_regs |
1228 | |= call_clobbers_in_region (def->crossed_call_abis, |
1229 | reg_class_contents[ALL_REGS], mode); |
1230 | |
1231 | /* Stop here before reload: we need FRAME_REGS, STACK_REGS, and |
1232 | crossed_call_abis, but not register classes. */ |
1233 | if (!reload_completed) |
1234 | return; |
1235 | |
1236 | /* Leave regs as 'available' only from the current |
1237 | register class. */ |
1238 | reg_rename_p->available_for_renaming = reg_class_contents[cl]; |
1239 | |
1240 | /* Leave only registers available for this mode. */ |
1241 | if (!sel_hrd.regs_for_mode_ok[mode]) |
1242 | init_regs_for_mode (mode); |
1243 | reg_rename_p->available_for_renaming &= sel_hrd.regs_for_mode[mode]; |
1244 | |
1245 | /* Leave only those that are ok to rename. */ |
1246 | EXECUTE_IF_SET_IN_HARD_REG_SET (reg_rename_p->available_for_renaming, |
1247 | 0, cur_reg, hrsi) |
1248 | { |
1249 | int nregs; |
1250 | int i; |
1251 | |
1252 | nregs = hard_regno_nregs (regno: cur_reg, mode); |
1253 | gcc_assert (nregs > 0); |
1254 | |
1255 | for (i = nregs - 1; i >= 0; --i) |
1256 | if (! sel_hard_regno_rename_ok (from: regno + i, to: cur_reg + i)) |
1257 | break; |
1258 | |
1259 | if (i >= 0) |
1260 | CLEAR_HARD_REG_BIT (set&: reg_rename_p->available_for_renaming, |
1261 | bit: cur_reg); |
1262 | } |
1263 | |
1264 | reg_rename_p->available_for_renaming &= ~reg_rename_p->unavailable_hard_regs; |
1265 | |
1266 | /* Regno is always ok from the renaming part of view, but it really |
1267 | could be in *unavailable_hard_regs already, so set it here instead |
1268 | of there. */ |
1269 | SET_HARD_REG_BIT (set&: reg_rename_p->available_for_renaming, bit: regno); |
1270 | } |
1271 | |
1272 | /* reg_rename_tick[REG1] > reg_rename_tick[REG2] if REG1 was chosen as the |
1273 | best register more recently than REG2. */ |
1274 | static int reg_rename_tick[FIRST_PSEUDO_REGISTER]; |
1275 | |
1276 | /* Indicates the number of times renaming happened before the current one. */ |
1277 | static int reg_rename_this_tick; |
1278 | |
1279 | /* Choose the register among free, that is suitable for storing |
1280 | the rhs value. |
1281 | |
1282 | ORIGINAL_INSNS is the list of insns where the operation (rhs) |
1283 | originally appears. There could be multiple original operations |
1284 | for single rhs since we moving it up and merging along different |
1285 | paths. |
1286 | |
1287 | Some code is adapted from regrename.cc (regrename_optimize). |
1288 | If original register is available, function returns it. |
1289 | Otherwise it performs the checks, so the new register should |
1290 | comply with the following: |
1291 | - it should not violate any live ranges (such registers are in |
1292 | REG_RENAME_P->available_for_renaming set); |
1293 | - it should not be in the HARD_REGS_USED regset; |
1294 | - it should be in the class compatible with original uses; |
1295 | - it should not be clobbered through reference with different mode; |
1296 | - if we're in the leaf function, then the new register should |
1297 | not be in the LEAF_REGISTERS; |
1298 | - etc. |
1299 | |
1300 | If several registers meet the conditions, the register with smallest |
1301 | tick is returned to achieve more even register allocation. |
1302 | |
1303 | If original register seems to be ok, we set *IS_ORIG_REG_P_PTR to true. |
1304 | |
1305 | If no register satisfies the above conditions, NULL_RTX is returned. */ |
1306 | static rtx |
1307 | choose_best_reg_1 (HARD_REG_SET hard_regs_used, |
1308 | struct reg_rename *reg_rename_p, |
1309 | def_list_t original_insns, bool *is_orig_reg_p_ptr) |
1310 | { |
1311 | int best_new_reg; |
1312 | unsigned cur_reg; |
1313 | machine_mode mode = VOIDmode; |
1314 | unsigned regno, i, n; |
1315 | hard_reg_set_iterator hrsi; |
1316 | def_list_iterator di; |
1317 | def_t def; |
1318 | |
1319 | /* If original register is available, return it. */ |
1320 | *is_orig_reg_p_ptr = true; |
1321 | |
1322 | FOR_EACH_DEF (def, di, original_insns) |
1323 | { |
1324 | rtx orig_dest = SET_DEST (PATTERN (def->orig_insn)); |
1325 | |
1326 | gcc_assert (REG_P (orig_dest)); |
1327 | |
1328 | /* Check that all original operations have the same mode. |
1329 | This is done for the next loop; if we'd return from this |
1330 | loop, we'd check only part of them, but in this case |
1331 | it doesn't matter. */ |
1332 | if (mode == VOIDmode) |
1333 | mode = GET_MODE (orig_dest); |
1334 | gcc_assert (mode == GET_MODE (orig_dest)); |
1335 | |
1336 | regno = REGNO (orig_dest); |
1337 | for (i = 0, n = REG_NREGS (orig_dest); i < n; i++) |
1338 | if (TEST_HARD_REG_BIT (set: hard_regs_used, bit: regno + i)) |
1339 | break; |
1340 | |
1341 | /* All hard registers are available. */ |
1342 | if (i == n) |
1343 | { |
1344 | gcc_assert (mode != VOIDmode); |
1345 | |
1346 | /* Hard registers should not be shared. */ |
1347 | return gen_rtx_REG (mode, regno); |
1348 | } |
1349 | } |
1350 | |
1351 | *is_orig_reg_p_ptr = false; |
1352 | best_new_reg = -1; |
1353 | |
1354 | /* Among all available regs choose the register that was |
1355 | allocated earliest. */ |
1356 | EXECUTE_IF_SET_IN_HARD_REG_SET (reg_rename_p->available_for_renaming, |
1357 | 0, cur_reg, hrsi) |
1358 | if (! TEST_HARD_REG_BIT (set: hard_regs_used, bit: cur_reg)) |
1359 | { |
1360 | /* Check that all hard regs for mode are available. */ |
1361 | for (i = 1, n = hard_regno_nregs (regno: cur_reg, mode); i < n; i++) |
1362 | if (TEST_HARD_REG_BIT (set: hard_regs_used, bit: cur_reg + i) |
1363 | || !TEST_HARD_REG_BIT (set: reg_rename_p->available_for_renaming, |
1364 | bit: cur_reg + i)) |
1365 | break; |
1366 | |
1367 | if (i < n) |
1368 | continue; |
1369 | |
1370 | /* All hard registers are available. */ |
1371 | if (best_new_reg < 0 |
1372 | || reg_rename_tick[cur_reg] < reg_rename_tick[best_new_reg]) |
1373 | { |
1374 | best_new_reg = cur_reg; |
1375 | |
1376 | /* Return immediately when we know there's no better reg. */ |
1377 | if (! reg_rename_tick[best_new_reg]) |
1378 | break; |
1379 | } |
1380 | } |
1381 | |
1382 | if (best_new_reg >= 0) |
1383 | { |
1384 | /* Use the check from the above loop. */ |
1385 | gcc_assert (mode != VOIDmode); |
1386 | return gen_rtx_REG (mode, best_new_reg); |
1387 | } |
1388 | |
1389 | return NULL_RTX; |
1390 | } |
1391 | |
1392 | /* A wrapper around choose_best_reg_1 () to verify that we make correct |
1393 | assumptions about available registers in the function. */ |
1394 | static rtx |
1395 | choose_best_reg (HARD_REG_SET hard_regs_used, struct reg_rename *reg_rename_p, |
1396 | def_list_t original_insns, bool *is_orig_reg_p_ptr) |
1397 | { |
1398 | rtx best_reg = choose_best_reg_1 (hard_regs_used, reg_rename_p, |
1399 | original_insns, is_orig_reg_p_ptr); |
1400 | |
1401 | /* FIXME loop over hard_regno_nregs here. */ |
1402 | gcc_assert (best_reg == NULL_RTX |
1403 | || TEST_HARD_REG_BIT (sel_hrd.regs_ever_used, REGNO (best_reg))); |
1404 | |
1405 | return best_reg; |
1406 | } |
1407 | |
1408 | /* Choose the pseudo register for storing rhs value. As this is supposed |
1409 | to work before reload, we return either the original register or make |
1410 | the new one. The parameters are the same that in choose_nest_reg_1 |
1411 | functions, except that USED_REGS may contain pseudos. |
1412 | If we work with hard regs, check also REG_RENAME_P->UNAVAILABLE_HARD_REGS. |
1413 | |
1414 | TODO: take into account register pressure while doing this. Up to this |
1415 | moment, this function would never return NULL for pseudos, but we should |
1416 | not rely on this. */ |
1417 | static rtx |
1418 | choose_best_pseudo_reg (regset used_regs, |
1419 | struct reg_rename *reg_rename_p, |
1420 | def_list_t original_insns, bool *is_orig_reg_p_ptr) |
1421 | { |
1422 | def_list_iterator i; |
1423 | def_t def; |
1424 | machine_mode mode = VOIDmode; |
1425 | bool bad_hard_regs = false; |
1426 | |
1427 | /* We should not use this after reload. */ |
1428 | gcc_assert (!reload_completed); |
1429 | |
1430 | /* If original register is available, return it. */ |
1431 | *is_orig_reg_p_ptr = true; |
1432 | |
1433 | FOR_EACH_DEF (def, i, original_insns) |
1434 | { |
1435 | rtx dest = SET_DEST (PATTERN (def->orig_insn)); |
1436 | int orig_regno; |
1437 | |
1438 | gcc_assert (REG_P (dest)); |
1439 | |
1440 | /* Check that all original operations have the same mode. */ |
1441 | if (mode == VOIDmode) |
1442 | mode = GET_MODE (dest); |
1443 | else |
1444 | gcc_assert (mode == GET_MODE (dest)); |
1445 | orig_regno = REGNO (dest); |
1446 | |
1447 | /* Check that nothing in used_regs intersects with orig_regno. When |
1448 | we have a hard reg here, still loop over hard_regno_nregs. */ |
1449 | if (HARD_REGISTER_NUM_P (orig_regno)) |
1450 | { |
1451 | int j, n; |
1452 | for (j = 0, n = REG_NREGS (dest); j < n; j++) |
1453 | if (REGNO_REG_SET_P (used_regs, orig_regno + j)) |
1454 | break; |
1455 | if (j < n) |
1456 | continue; |
1457 | } |
1458 | else |
1459 | { |
1460 | if (REGNO_REG_SET_P (used_regs, orig_regno)) |
1461 | continue; |
1462 | } |
1463 | if (HARD_REGISTER_NUM_P (orig_regno)) |
1464 | { |
1465 | gcc_assert (df_regs_ever_live_p (orig_regno)); |
1466 | |
1467 | /* For hard registers, we have to check hardware imposed |
1468 | limitations (frame/stack registers, calls crossed). */ |
1469 | if (!TEST_HARD_REG_BIT (set: reg_rename_p->unavailable_hard_regs, |
1470 | bit: orig_regno)) |
1471 | { |
1472 | /* Don't let register cross a call if it doesn't already |
1473 | cross one. This condition is written in accordance with |
1474 | that in sched-deps.cc sched_analyze_reg(). */ |
1475 | if (!reg_rename_p->crossed_call_abis |
1476 | || REG_N_CALLS_CROSSED (orig_regno) > 0) |
1477 | return gen_rtx_REG (mode, orig_regno); |
1478 | } |
1479 | |
1480 | bad_hard_regs = true; |
1481 | } |
1482 | else |
1483 | return dest; |
1484 | } |
1485 | |
1486 | *is_orig_reg_p_ptr = false; |
1487 | |
1488 | /* We had some original hard registers that couldn't be used. |
1489 | Those were likely special. Don't try to create a pseudo. */ |
1490 | if (bad_hard_regs) |
1491 | return NULL_RTX; |
1492 | |
1493 | /* We haven't found a register from original operations. Get a new one. |
1494 | FIXME: control register pressure somehow. */ |
1495 | { |
1496 | rtx new_reg = gen_reg_rtx (mode); |
1497 | |
1498 | gcc_assert (mode != VOIDmode); |
1499 | |
1500 | max_regno = max_reg_num (); |
1501 | maybe_extend_reg_info_p (); |
1502 | REG_N_CALLS_CROSSED (REGNO (new_reg)) |
1503 | = reg_rename_p->crossed_call_abis ? 1 : 0; |
1504 | |
1505 | return new_reg; |
1506 | } |
1507 | } |
1508 | |
1509 | /* True when target of EXPR is available due to EXPR_TARGET_AVAILABLE, |
1510 | USED_REGS and REG_RENAME_P->UNAVAILABLE_HARD_REGS. */ |
1511 | static void |
1512 | verify_target_availability (expr_t expr, regset used_regs, |
1513 | struct reg_rename *reg_rename_p) |
1514 | { |
1515 | unsigned n, i, regno; |
1516 | machine_mode mode; |
1517 | bool target_available, live_available, hard_available; |
1518 | |
1519 | if (!REG_P (EXPR_LHS (expr)) || EXPR_TARGET_AVAILABLE (expr) < 0) |
1520 | return; |
1521 | |
1522 | regno = expr_dest_regno (expr); |
1523 | mode = GET_MODE (EXPR_LHS (expr)); |
1524 | target_available = EXPR_TARGET_AVAILABLE (expr) == 1; |
1525 | n = HARD_REGISTER_NUM_P (regno) ? hard_regno_nregs (regno, mode) : 1; |
1526 | |
1527 | live_available = hard_available = true; |
1528 | for (i = 0; i < n; i++) |
1529 | { |
1530 | if (bitmap_bit_p (used_regs, regno + i)) |
1531 | live_available = false; |
1532 | if (TEST_HARD_REG_BIT (set: reg_rename_p->unavailable_hard_regs, bit: regno + i)) |
1533 | hard_available = false; |
1534 | } |
1535 | |
1536 | /* When target is not available, it may be due to hard register |
1537 | restrictions, e.g. crosses calls, so we check hard_available too. */ |
1538 | if (target_available) |
1539 | gcc_assert (live_available); |
1540 | else |
1541 | /* Check only if we haven't scheduled something on the previous fence, |
1542 | cause due to MAX_SOFTWARE_LOOKAHEAD_WINDOW_SIZE issues |
1543 | and having more than one fence, we may end having targ_un in a block |
1544 | in which successors target register is actually available. |
1545 | |
1546 | The last condition handles the case when a dependence from a call insn |
1547 | was created in sched-deps.cc for insns with destination registers that |
1548 | never crossed a call before, but do cross one after our code motion. |
1549 | |
1550 | FIXME: in the latter case, we just uselessly called find_used_regs, |
1551 | because we can't move this expression with any other register |
1552 | as well. */ |
1553 | gcc_assert (scheduled_something_on_previous_fence || !live_available |
1554 | || !hard_available |
1555 | || (!reload_completed |
1556 | && reg_rename_p->crossed_call_abis |
1557 | && REG_N_CALLS_CROSSED (regno) == 0)); |
1558 | } |
1559 | |
1560 | /* Collect unavailable registers due to liveness for EXPR from BNDS |
1561 | into USED_REGS. Save additional information about available |
1562 | registers and unavailable due to hardware restriction registers |
1563 | into REG_RENAME_P structure. Save original insns into ORIGINAL_INSNS |
1564 | list. */ |
1565 | static void |
1566 | collect_unavailable_regs_from_bnds (expr_t expr, blist_t bnds, regset used_regs, |
1567 | struct reg_rename *reg_rename_p, |
1568 | def_list_t *original_insns) |
1569 | { |
1570 | for (; bnds; bnds = BLIST_NEXT (bnds)) |
1571 | { |
1572 | bool res; |
1573 | av_set_t orig_ops = NULL; |
1574 | bnd_t bnd = BLIST_BND (bnds); |
1575 | |
1576 | /* If the chosen best expr doesn't belong to current boundary, |
1577 | skip it. */ |
1578 | if (!av_set_is_in_p (BND_AV1 (bnd), EXPR_VINSN (expr))) |
1579 | continue; |
1580 | |
1581 | /* Put in ORIG_OPS all exprs from this boundary that became |
1582 | RES on top. */ |
1583 | orig_ops = find_sequential_best_exprs (bnd, expr, false); |
1584 | |
1585 | /* Compute used regs and OR it into the USED_REGS. */ |
1586 | res = find_used_regs (BND_TO (bnd), orig_ops, used_regs, |
1587 | reg_rename_p, original_insns); |
1588 | |
1589 | /* FIXME: the assert is true until we'd have several boundaries. */ |
1590 | gcc_assert (res); |
1591 | av_set_clear (&orig_ops); |
1592 | } |
1593 | } |
1594 | |
1595 | /* Return TRUE if it is possible to replace LHSes of ORIG_INSNS with BEST_REG. |
1596 | If BEST_REG is valid, replace LHS of EXPR with it. */ |
1597 | static bool |
1598 | try_replace_dest_reg (ilist_t orig_insns, rtx best_reg, expr_t expr) |
1599 | { |
1600 | /* Try whether we'll be able to generate the insn |
1601 | 'dest := best_reg' at the place of the original operation. */ |
1602 | for (; orig_insns; orig_insns = ILIST_NEXT (orig_insns)) |
1603 | { |
1604 | insn_t orig_insn = DEF_LIST_DEF (orig_insns)->orig_insn; |
1605 | |
1606 | gcc_assert (EXPR_SEPARABLE_P (INSN_EXPR (orig_insn))); |
1607 | |
1608 | if (REGNO (best_reg) != REGNO (INSN_LHS (orig_insn)) |
1609 | && (! replace_src_with_reg_ok_p (insn: orig_insn, new_src_reg: best_reg) |
1610 | || ! replace_dest_with_reg_ok_p (insn: orig_insn, new_reg: best_reg))) |
1611 | return false; |
1612 | } |
1613 | |
1614 | /* Make sure that EXPR has the right destination |
1615 | register. */ |
1616 | if (expr_dest_regno (expr) != REGNO (best_reg)) |
1617 | { |
1618 | rtx_insn *vinsn = EXPR_INSN_RTX (expr); |
1619 | validate_change (vinsn, &SET_DEST (PATTERN (vinsn)), best_reg, 1); |
1620 | bool res = verify_changes (0); |
1621 | cancel_changes (0); |
1622 | if (!res) |
1623 | return false; |
1624 | replace_dest_with_reg_in_expr (expr, new_reg: best_reg); |
1625 | } |
1626 | else |
1627 | EXPR_TARGET_AVAILABLE (expr) = 1; |
1628 | |
1629 | return true; |
1630 | } |
1631 | |
1632 | /* Select and assign best register to EXPR searching from BNDS. |
1633 | Set *IS_ORIG_REG_P to TRUE if original register was selected. |
1634 | Return FALSE if no register can be chosen, which could happen when: |
1635 | * EXPR_SEPARABLE_P is true but we were unable to find suitable register; |
1636 | * EXPR_SEPARABLE_P is false but the insn sets/clobbers one of the registers |
1637 | that are used on the moving path. */ |
1638 | static bool |
1639 | find_best_reg_for_expr (expr_t expr, blist_t bnds, bool *is_orig_reg_p) |
1640 | { |
1641 | static struct reg_rename reg_rename_data; |
1642 | |
1643 | regset used_regs; |
1644 | def_list_t original_insns = NULL; |
1645 | bool reg_ok; |
1646 | |
1647 | *is_orig_reg_p = false; |
1648 | |
1649 | /* Don't bother to do anything if this insn doesn't set any registers. */ |
1650 | if (bitmap_empty_p (VINSN_REG_SETS (EXPR_VINSN (expr))) |
1651 | && bitmap_empty_p (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)))) |
1652 | return true; |
1653 | |
1654 | used_regs = get_clear_regset_from_pool (); |
1655 | CLEAR_HARD_REG_SET (set&: reg_rename_data.unavailable_hard_regs); |
1656 | |
1657 | collect_unavailable_regs_from_bnds (expr, bnds, used_regs, reg_rename_p: ®_rename_data, |
1658 | original_insns: &original_insns); |
1659 | |
1660 | /* If after reload, make sure we're working with hard regs here. */ |
1661 | if (flag_checking && reload_completed) |
1662 | { |
1663 | reg_set_iterator rsi; |
1664 | unsigned i; |
1665 | |
1666 | EXECUTE_IF_SET_IN_REG_SET (used_regs, FIRST_PSEUDO_REGISTER, i, rsi) |
1667 | gcc_unreachable (); |
1668 | } |
1669 | |
1670 | if (EXPR_SEPARABLE_P (expr)) |
1671 | { |
1672 | rtx best_reg = NULL_RTX; |
1673 | /* Check that we have computed availability of a target register |
1674 | correctly. */ |
1675 | verify_target_availability (expr, used_regs, reg_rename_p: ®_rename_data); |
1676 | |
1677 | /* Turn everything in hard regs after reload. */ |
1678 | if (reload_completed) |
1679 | { |
1680 | HARD_REG_SET hard_regs_used; |
1681 | REG_SET_TO_HARD_REG_SET (hard_regs_used, used_regs); |
1682 | |
1683 | /* Join hard registers unavailable due to register class |
1684 | restrictions and live range intersection. */ |
1685 | hard_regs_used |= reg_rename_data.unavailable_hard_regs; |
1686 | |
1687 | best_reg = choose_best_reg (hard_regs_used, reg_rename_p: ®_rename_data, |
1688 | original_insns, is_orig_reg_p_ptr: is_orig_reg_p); |
1689 | } |
1690 | else |
1691 | best_reg = choose_best_pseudo_reg (used_regs, reg_rename_p: ®_rename_data, |
1692 | original_insns, is_orig_reg_p_ptr: is_orig_reg_p); |
1693 | |
1694 | if (!best_reg) |
1695 | reg_ok = false; |
1696 | else if (*is_orig_reg_p) |
1697 | { |
1698 | /* In case of unification BEST_REG may be different from EXPR's LHS |
1699 | when EXPR's LHS is unavailable, and there is another LHS among |
1700 | ORIGINAL_INSNS. */ |
1701 | reg_ok = try_replace_dest_reg (orig_insns: original_insns, best_reg, expr); |
1702 | } |
1703 | else |
1704 | { |
1705 | /* Forbid renaming of low-cost insns. */ |
1706 | if (sel_vinsn_cost (EXPR_VINSN (expr)) < 2) |
1707 | reg_ok = false; |
1708 | else |
1709 | reg_ok = try_replace_dest_reg (orig_insns: original_insns, best_reg, expr); |
1710 | } |
1711 | } |
1712 | else |
1713 | { |
1714 | /* If !EXPR_SCHEDULE_AS_RHS (EXPR), just make sure INSN doesn't set |
1715 | any of the HARD_REGS_USED set. */ |
1716 | if (vinsn_writes_one_of_regs_p (EXPR_VINSN (expr), used_regs, |
1717 | unavailable_hard_regs: reg_rename_data.unavailable_hard_regs)) |
1718 | { |
1719 | reg_ok = false; |
1720 | gcc_assert (EXPR_TARGET_AVAILABLE (expr) <= 0); |
1721 | } |
1722 | else |
1723 | { |
1724 | reg_ok = true; |
1725 | gcc_assert (EXPR_TARGET_AVAILABLE (expr) != 0); |
1726 | } |
1727 | } |
1728 | |
1729 | ilist_clear (&original_insns); |
1730 | return_regset_to_pool (used_regs); |
1731 | |
1732 | return reg_ok; |
1733 | } |
1734 | |
1735 | |
1736 | /* Return true if dependence described by DS can be overcomed. */ |
1737 | static bool |
1738 | can_speculate_dep_p (ds_t ds) |
1739 | { |
1740 | if (spec_info == NULL) |
1741 | return false; |
1742 | |
1743 | /* Leave only speculative data. */ |
1744 | ds &= SPECULATIVE; |
1745 | |
1746 | if (ds == 0) |
1747 | return false; |
1748 | |
1749 | { |
1750 | /* FIXME: make sched-deps.cc produce only those non-hard dependencies, |
1751 | that we can overcome. */ |
1752 | ds_t spec_mask = spec_info->mask; |
1753 | |
1754 | if ((ds & spec_mask) != ds) |
1755 | return false; |
1756 | } |
1757 | |
1758 | if (ds_weak (ds) < spec_info->data_weakness_cutoff) |
1759 | return false; |
1760 | |
1761 | return true; |
1762 | } |
1763 | |
1764 | /* Get a speculation check instruction. |
1765 | C_EXPR is a speculative expression, |
1766 | CHECK_DS describes speculations that should be checked, |
1767 | ORIG_INSN is the original non-speculative insn in the stream. */ |
1768 | static insn_t |
1769 | create_speculation_check (expr_t c_expr, ds_t check_ds, insn_t orig_insn) |
1770 | { |
1771 | rtx check_pattern; |
1772 | rtx_insn *insn_rtx; |
1773 | insn_t insn; |
1774 | basic_block recovery_block; |
1775 | rtx_insn *label; |
1776 | |
1777 | /* Create a recovery block if target is going to emit branchy check, or if |
1778 | ORIG_INSN was speculative already. */ |
1779 | if (targetm.sched.needs_block_p (check_ds) |
1780 | || EXPR_SPEC_DONE_DS (INSN_EXPR (orig_insn)) != 0) |
1781 | { |
1782 | recovery_block = sel_create_recovery_block (orig_insn); |
1783 | label = BB_HEAD (recovery_block); |
1784 | } |
1785 | else |
1786 | { |
1787 | recovery_block = NULL; |
1788 | label = NULL; |
1789 | } |
1790 | |
1791 | /* Get pattern of the check. */ |
1792 | check_pattern = targetm.sched.gen_spec_check (EXPR_INSN_RTX (c_expr), label, |
1793 | check_ds); |
1794 | |
1795 | gcc_assert (check_pattern != NULL); |
1796 | |
1797 | /* Emit check. */ |
1798 | insn_rtx = create_insn_rtx_from_pattern (check_pattern, label); |
1799 | |
1800 | insn = sel_gen_insn_from_rtx_after (insn_rtx, INSN_EXPR (orig_insn), |
1801 | INSN_SEQNO (orig_insn), orig_insn); |
1802 | |
1803 | /* Make check to be non-speculative. */ |
1804 | EXPR_SPEC_DONE_DS (INSN_EXPR (insn)) = 0; |
1805 | INSN_SPEC_CHECKED_DS (insn) = check_ds; |
1806 | |
1807 | /* Decrease priority of check by difference of load/check instruction |
1808 | latencies. */ |
1809 | EXPR_PRIORITY (INSN_EXPR (insn)) -= (sel_vinsn_cost (INSN_VINSN (orig_insn)) |
1810 | - sel_vinsn_cost (INSN_VINSN (insn))); |
1811 | |
1812 | /* Emit copy of original insn (though with replaced target register, |
1813 | if needed) to the recovery block. */ |
1814 | if (recovery_block != NULL) |
1815 | { |
1816 | rtx twin_rtx; |
1817 | |
1818 | twin_rtx = copy_rtx (PATTERN (EXPR_INSN_RTX (c_expr))); |
1819 | twin_rtx = create_insn_rtx_from_pattern (twin_rtx, NULL_RTX); |
1820 | sel_gen_recovery_insn_from_rtx_after (twin_rtx, |
1821 | INSN_EXPR (orig_insn), |
1822 | INSN_SEQNO (insn), |
1823 | bb_note (recovery_block)); |
1824 | } |
1825 | |
1826 | /* If we've generated a data speculation check, make sure |
1827 | that all the bookkeeping instruction we'll create during |
1828 | this move_op () will allocate an ALAT entry so that the |
1829 | check won't fail. |
1830 | In case of control speculation we must convert C_EXPR to control |
1831 | speculative mode, because failing to do so will bring us an exception |
1832 | thrown by the non-control-speculative load. */ |
1833 | check_ds = ds_get_max_dep_weak (check_ds); |
1834 | speculate_expr (c_expr, check_ds); |
1835 | |
1836 | return insn; |
1837 | } |
1838 | |
1839 | /* True when INSN is a "regN = regN" copy. */ |
1840 | static bool |
1841 | identical_copy_p (rtx_insn *insn) |
1842 | { |
1843 | rtx lhs, rhs, pat; |
1844 | |
1845 | pat = PATTERN (insn); |
1846 | |
1847 | if (GET_CODE (pat) != SET) |
1848 | return false; |
1849 | |
1850 | lhs = SET_DEST (pat); |
1851 | if (!REG_P (lhs)) |
1852 | return false; |
1853 | |
1854 | rhs = SET_SRC (pat); |
1855 | if (!REG_P (rhs)) |
1856 | return false; |
1857 | |
1858 | return REGNO (lhs) == REGNO (rhs); |
1859 | } |
1860 | |
1861 | /* Undo all transformations on *AV_PTR that were done when |
1862 | moving through INSN. */ |
1863 | static void |
1864 | undo_transformations (av_set_t *av_ptr, rtx_insn *insn) |
1865 | { |
1866 | av_set_iterator av_iter; |
1867 | expr_t expr; |
1868 | av_set_t new_set = NULL; |
1869 | |
1870 | /* First, kill any EXPR that uses registers set by an insn. This is |
1871 | required for correctness. */ |
1872 | FOR_EACH_EXPR_1 (expr, av_iter, av_ptr) |
1873 | if (!sched_insns_conditions_mutex_p (insn, EXPR_INSN_RTX (expr)) |
1874 | && bitmap_intersect_p (INSN_REG_SETS (insn), |
1875 | VINSN_REG_USES (EXPR_VINSN (expr))) |
1876 | /* When an insn looks like 'r1 = r1', we could substitute through |
1877 | it, but the above condition will still hold. This happened with |
1878 | gcc.c-torture/execute/961125-1.c. */ |
1879 | && !identical_copy_p (insn)) |
1880 | { |
1881 | if (sched_verbose >= 6) |
1882 | sel_print (fmt: "Expr %d removed due to use/set conflict\n" , |
1883 | INSN_UID (EXPR_INSN_RTX (expr))); |
1884 | av_set_iter_remove (&av_iter); |
1885 | } |
1886 | |
1887 | /* Undo transformations looking at the history vector. */ |
1888 | FOR_EACH_EXPR (expr, av_iter, *av_ptr) |
1889 | { |
1890 | int index = find_in_history_vect (EXPR_HISTORY_OF_CHANGES (expr), |
1891 | insn, EXPR_VINSN (expr), true); |
1892 | |
1893 | if (index >= 0) |
1894 | { |
1895 | expr_history_def *phist; |
1896 | |
1897 | phist = &EXPR_HISTORY_OF_CHANGES (expr)[index]; |
1898 | |
1899 | switch (phist->type) |
1900 | { |
1901 | case TRANS_SPECULATION: |
1902 | { |
1903 | ds_t old_ds, new_ds; |
1904 | |
1905 | /* Compute the difference between old and new speculative |
1906 | statuses: that's what we need to check. |
1907 | Earlier we used to assert that the status will really |
1908 | change. This no longer works because only the probability |
1909 | bits in the status may have changed during compute_av_set, |
1910 | and in the case of merging different probabilities of the |
1911 | same speculative status along different paths we do not |
1912 | record this in the history vector. */ |
1913 | old_ds = phist->spec_ds; |
1914 | new_ds = EXPR_SPEC_DONE_DS (expr); |
1915 | |
1916 | old_ds &= SPECULATIVE; |
1917 | new_ds &= SPECULATIVE; |
1918 | new_ds &= ~old_ds; |
1919 | |
1920 | EXPR_SPEC_TO_CHECK_DS (expr) |= new_ds; |
1921 | break; |
1922 | } |
1923 | case TRANS_SUBSTITUTION: |
1924 | { |
1925 | expr_def _tmp_expr, *tmp_expr = &_tmp_expr; |
1926 | vinsn_t new_vi; |
1927 | bool add = true; |
1928 | |
1929 | new_vi = phist->old_expr_vinsn; |
1930 | |
1931 | gcc_assert (VINSN_SEPARABLE_P (new_vi) |
1932 | == EXPR_SEPARABLE_P (expr)); |
1933 | copy_expr (tmp_expr, expr); |
1934 | |
1935 | if (vinsn_equal_p (phist->new_expr_vinsn, |
1936 | EXPR_VINSN (tmp_expr))) |
1937 | change_vinsn_in_expr (tmp_expr, new_vi); |
1938 | else |
1939 | /* This happens when we're unsubstituting on a bookkeeping |
1940 | copy, which was in turn substituted. The history is wrong |
1941 | in this case. Do it the hard way. */ |
1942 | add = substitute_reg_in_expr (expr: tmp_expr, insn, undo: true); |
1943 | if (add) |
1944 | av_set_add (&new_set, tmp_expr); |
1945 | clear_expr (tmp_expr); |
1946 | break; |
1947 | } |
1948 | default: |
1949 | gcc_unreachable (); |
1950 | } |
1951 | } |
1952 | |
1953 | } |
1954 | |
1955 | av_set_union_and_clear (av_ptr, &new_set, NULL); |
1956 | } |
1957 | |
1958 | |
1959 | /* Moveup_* helpers for code motion and computing av sets. */ |
1960 | |
1961 | /* Propagates EXPR inside an insn group through THROUGH_INSN. |
1962 | The difference from the below function is that only substitution is |
1963 | performed. */ |
1964 | static enum MOVEUP_EXPR_CODE |
1965 | moveup_expr_inside_insn_group (expr_t expr, insn_t through_insn) |
1966 | { |
1967 | vinsn_t vi = EXPR_VINSN (expr); |
1968 | ds_t *has_dep_p; |
1969 | ds_t full_ds; |
1970 | |
1971 | /* Do this only inside insn group. */ |
1972 | gcc_assert (INSN_SCHED_CYCLE (through_insn) > 0); |
1973 | |
1974 | full_ds = has_dependence_p (expr, through_insn, &has_dep_p); |
1975 | if (full_ds == 0) |
1976 | return MOVEUP_EXPR_SAME; |
1977 | |
1978 | /* Substitution is the possible choice in this case. */ |
1979 | if (has_dep_p[DEPS_IN_RHS]) |
1980 | { |
1981 | /* Can't substitute UNIQUE VINSNs. */ |
1982 | gcc_assert (!VINSN_UNIQUE_P (vi)); |
1983 | |
1984 | if (can_substitute_through_p (insn: through_insn, |
1985 | ds: has_dep_p[DEPS_IN_RHS]) |
1986 | && substitute_reg_in_expr (expr, insn: through_insn, undo: false)) |
1987 | { |
1988 | EXPR_WAS_SUBSTITUTED (expr) = true; |
1989 | return MOVEUP_EXPR_CHANGED; |
1990 | } |
1991 | |
1992 | /* Don't care about this, as even true dependencies may be allowed |
1993 | in an insn group. */ |
1994 | return MOVEUP_EXPR_SAME; |
1995 | } |
1996 | |
1997 | /* This can catch output dependencies in COND_EXECs. */ |
1998 | if (has_dep_p[DEPS_IN_INSN]) |
1999 | return MOVEUP_EXPR_NULL; |
2000 | |
2001 | /* This is either an output or an anti dependence, which usually have |
2002 | a zero latency. Allow this here, if we'd be wrong, tick_check_p |
2003 | will fix this. */ |
2004 | gcc_assert (has_dep_p[DEPS_IN_LHS]); |
2005 | return MOVEUP_EXPR_AS_RHS; |
2006 | } |
2007 | |
2008 | /* True when a trapping EXPR cannot be moved through THROUGH_INSN. */ |
2009 | #define CANT_MOVE_TRAPPING(expr, through_insn) \ |
2010 | (VINSN_MAY_TRAP_P (EXPR_VINSN (expr)) \ |
2011 | && !sel_insn_has_single_succ_p ((through_insn), SUCCS_ALL) \ |
2012 | && !sel_insn_is_speculation_check (through_insn)) |
2013 | |
2014 | /* True when a conflict on a target register was found during moveup_expr. */ |
2015 | static bool was_target_conflict = false; |
2016 | |
2017 | /* Return true when moving a debug INSN across THROUGH_INSN will |
2018 | create a bookkeeping block. We don't want to create such blocks, |
2019 | for they would cause codegen differences between compilations with |
2020 | and without debug info. */ |
2021 | |
2022 | static bool |
2023 | moving_insn_creates_bookkeeping_block_p (insn_t insn, |
2024 | insn_t through_insn) |
2025 | { |
2026 | basic_block bbi, bbt; |
2027 | edge e1, e2; |
2028 | edge_iterator ei1, ei2; |
2029 | |
2030 | if (!bookkeeping_can_be_created_if_moved_through_p (through_insn)) |
2031 | { |
2032 | if (sched_verbose >= 9) |
2033 | sel_print (fmt: "no bookkeeping required: " ); |
2034 | return false; |
2035 | } |
2036 | |
2037 | bbi = BLOCK_FOR_INSN (insn); |
2038 | |
2039 | if (EDGE_COUNT (bbi->preds) == 1) |
2040 | { |
2041 | if (sched_verbose >= 9) |
2042 | sel_print (fmt: "only one pred edge: " ); |
2043 | return true; |
2044 | } |
2045 | |
2046 | bbt = BLOCK_FOR_INSN (insn: through_insn); |
2047 | |
2048 | FOR_EACH_EDGE (e1, ei1, bbt->succs) |
2049 | { |
2050 | FOR_EACH_EDGE (e2, ei2, bbi->preds) |
2051 | { |
2052 | if (find_block_for_bookkeeping (e1, e2, lax: true)) |
2053 | { |
2054 | if (sched_verbose >= 9) |
2055 | sel_print (fmt: "found existing block: " ); |
2056 | return false; |
2057 | } |
2058 | } |
2059 | } |
2060 | |
2061 | if (sched_verbose >= 9) |
2062 | sel_print (fmt: "would create bookkeeping block: " ); |
2063 | |
2064 | return true; |
2065 | } |
2066 | |
2067 | /* Return true when the conflict with newly created implicit clobbers |
2068 | between EXPR and THROUGH_INSN is found because of renaming. */ |
2069 | static bool |
2070 | implicit_clobber_conflict_p (insn_t through_insn, expr_t expr) |
2071 | { |
2072 | HARD_REG_SET temp; |
2073 | rtx_insn *insn; |
2074 | rtx reg, rhs, pat; |
2075 | hard_reg_set_iterator hrsi; |
2076 | unsigned regno; |
2077 | bool valid; |
2078 | |
2079 | /* Make a new pseudo register. */ |
2080 | reg = gen_reg_rtx (GET_MODE (EXPR_LHS (expr))); |
2081 | max_regno = max_reg_num (); |
2082 | maybe_extend_reg_info_p (); |
2083 | |
2084 | /* Validate a change and bail out early. */ |
2085 | insn = EXPR_INSN_RTX (expr); |
2086 | validate_change (insn, &SET_DEST (PATTERN (insn)), reg, true); |
2087 | valid = verify_changes (0); |
2088 | cancel_changes (0); |
2089 | if (!valid) |
2090 | { |
2091 | if (sched_verbose >= 6) |
2092 | sel_print (fmt: "implicit clobbers failed validation, " ); |
2093 | return true; |
2094 | } |
2095 | |
2096 | /* Make a new insn with it. */ |
2097 | rhs = copy_rtx (VINSN_RHS (EXPR_VINSN (expr))); |
2098 | pat = gen_rtx_SET (reg, rhs); |
2099 | start_sequence (); |
2100 | insn = emit_insn (pat); |
2101 | end_sequence (); |
2102 | |
2103 | /* Calculate implicit clobbers. */ |
2104 | extract_insn (insn); |
2105 | preprocess_constraints (insn); |
2106 | alternative_mask prefrred = get_preferred_alternatives (insn); |
2107 | ira_implicitly_set_insn_hard_regs (&temp, prefrred); |
2108 | temp &= ~ira_no_alloc_regs; |
2109 | |
2110 | /* If any implicit clobber registers intersect with regular ones in |
2111 | through_insn, we have a dependency and thus bail out. */ |
2112 | EXECUTE_IF_SET_IN_HARD_REG_SET (temp, 0, regno, hrsi) |
2113 | { |
2114 | vinsn_t vi = INSN_VINSN (through_insn); |
2115 | if (bitmap_bit_p (VINSN_REG_SETS (vi), regno) |
2116 | || bitmap_bit_p (VINSN_REG_CLOBBERS (vi), regno) |
2117 | || bitmap_bit_p (VINSN_REG_USES (vi), regno)) |
2118 | return true; |
2119 | } |
2120 | |
2121 | return false; |
2122 | } |
2123 | |
2124 | /* Modifies EXPR so it can be moved through the THROUGH_INSN, |
2125 | performing necessary transformations. Record the type of transformation |
2126 | made in PTRANS_TYPE, when it is not NULL. When INSIDE_INSN_GROUP, |
2127 | permit all dependencies except true ones, and try to remove those |
2128 | too via forward substitution. All cases when a non-eliminable |
2129 | non-zero cost dependency exists inside an insn group will be fixed |
2130 | in tick_check_p instead. */ |
2131 | static enum MOVEUP_EXPR_CODE |
2132 | moveup_expr (expr_t expr, insn_t through_insn, bool inside_insn_group, |
2133 | enum local_trans_type *ptrans_type) |
2134 | { |
2135 | vinsn_t vi = EXPR_VINSN (expr); |
2136 | insn_t insn = VINSN_INSN_RTX (vi); |
2137 | bool was_changed = false; |
2138 | bool as_rhs = false; |
2139 | ds_t *has_dep_p; |
2140 | ds_t full_ds; |
2141 | |
2142 | /* ??? We use dependencies of non-debug insns on debug insns to |
2143 | indicate that the debug insns need to be reset if the non-debug |
2144 | insn is pulled ahead of it. It's hard to figure out how to |
2145 | introduce such a notion in sel-sched, but it already fails to |
2146 | support debug insns in other ways, so we just go ahead and |
2147 | let the deug insns go corrupt for now. */ |
2148 | if (DEBUG_INSN_P (through_insn) && !DEBUG_INSN_P (insn)) |
2149 | return MOVEUP_EXPR_SAME; |
2150 | |
2151 | /* When inside_insn_group, delegate to the helper. */ |
2152 | if (inside_insn_group) |
2153 | return moveup_expr_inside_insn_group (expr, through_insn); |
2154 | |
2155 | /* Deal with unique insns and control dependencies. */ |
2156 | if (VINSN_UNIQUE_P (vi)) |
2157 | { |
2158 | /* We can move jumps without side-effects or jumps that are |
2159 | mutually exclusive with instruction THROUGH_INSN (all in cases |
2160 | dependencies allow to do so and jump is not speculative). */ |
2161 | if (control_flow_insn_p (insn)) |
2162 | { |
2163 | basic_block fallthru_bb; |
2164 | |
2165 | /* Do not move checks and do not move jumps through other |
2166 | jumps. */ |
2167 | if (control_flow_insn_p (through_insn) |
2168 | || sel_insn_is_speculation_check (insn)) |
2169 | return MOVEUP_EXPR_NULL; |
2170 | |
2171 | /* Don't move jumps through CFG joins. */ |
2172 | if (bookkeeping_can_be_created_if_moved_through_p (through_insn)) |
2173 | return MOVEUP_EXPR_NULL; |
2174 | |
2175 | /* The jump should have a clear fallthru block, and |
2176 | this block should be in the current region. */ |
2177 | if ((fallthru_bb = fallthru_bb_of_jump (insn)) == NULL |
2178 | || ! in_current_region_p (fallthru_bb)) |
2179 | return MOVEUP_EXPR_NULL; |
2180 | |
2181 | /* And it should be mutually exclusive with through_insn. */ |
2182 | if (! sched_insns_conditions_mutex_p (insn, through_insn) |
2183 | && ! DEBUG_INSN_P (through_insn)) |
2184 | return MOVEUP_EXPR_NULL; |
2185 | } |
2186 | |
2187 | /* Don't move what we can't move. */ |
2188 | if (EXPR_CANT_MOVE (expr) |
2189 | && BLOCK_FOR_INSN (insn: through_insn) != BLOCK_FOR_INSN (insn)) |
2190 | return MOVEUP_EXPR_NULL; |
2191 | |
2192 | /* Don't move SCHED_GROUP instruction through anything. |
2193 | If we don't force this, then it will be possible to start |
2194 | scheduling a sched_group before all its dependencies are |
2195 | resolved. |
2196 | ??? Haifa deals with this issue by delaying the SCHED_GROUP |
2197 | as late as possible through rank_for_schedule. */ |
2198 | if (SCHED_GROUP_P (insn)) |
2199 | return MOVEUP_EXPR_NULL; |
2200 | } |
2201 | else |
2202 | gcc_assert (!control_flow_insn_p (insn)); |
2203 | |
2204 | /* Don't move debug insns if this would require bookkeeping. */ |
2205 | if (DEBUG_INSN_P (insn) |
2206 | && BLOCK_FOR_INSN (insn: through_insn) != BLOCK_FOR_INSN (insn) |
2207 | && moving_insn_creates_bookkeeping_block_p (insn, through_insn)) |
2208 | return MOVEUP_EXPR_NULL; |
2209 | |
2210 | /* Deal with data dependencies. */ |
2211 | was_target_conflict = false; |
2212 | full_ds = has_dependence_p (expr, through_insn, &has_dep_p); |
2213 | if (full_ds == 0) |
2214 | { |
2215 | if (!CANT_MOVE_TRAPPING (expr, through_insn)) |
2216 | return MOVEUP_EXPR_SAME; |
2217 | } |
2218 | else |
2219 | { |
2220 | /* We can move UNIQUE insn up only as a whole and unchanged, |
2221 | so it shouldn't have any dependencies. */ |
2222 | if (VINSN_UNIQUE_P (vi)) |
2223 | return MOVEUP_EXPR_NULL; |
2224 | } |
2225 | |
2226 | if (full_ds != 0 && can_speculate_dep_p (ds: full_ds)) |
2227 | { |
2228 | int res; |
2229 | |
2230 | res = speculate_expr (expr, full_ds); |
2231 | if (res >= 0) |
2232 | { |
2233 | /* Speculation was successful. */ |
2234 | full_ds = 0; |
2235 | was_changed = (res > 0); |
2236 | if (res == 2) |
2237 | was_target_conflict = true; |
2238 | if (ptrans_type) |
2239 | *ptrans_type = TRANS_SPECULATION; |
2240 | sel_clear_has_dependence (); |
2241 | } |
2242 | } |
2243 | |
2244 | if (has_dep_p[DEPS_IN_INSN]) |
2245 | /* We have some dependency that cannot be discarded. */ |
2246 | return MOVEUP_EXPR_NULL; |
2247 | |
2248 | if (has_dep_p[DEPS_IN_LHS]) |
2249 | { |
2250 | /* Only separable insns can be moved up with the new register. |
2251 | Anyways, we should mark that the original register is |
2252 | unavailable. */ |
2253 | if (!enable_schedule_as_rhs_p || !EXPR_SEPARABLE_P (expr)) |
2254 | return MOVEUP_EXPR_NULL; |
2255 | |
2256 | /* When renaming a hard register to a pseudo before reload, extra |
2257 | dependencies can occur from the implicit clobbers of the insn. |
2258 | Filter out such cases here. */ |
2259 | if (!reload_completed && REG_P (EXPR_LHS (expr)) |
2260 | && HARD_REGISTER_P (EXPR_LHS (expr)) |
2261 | && implicit_clobber_conflict_p (through_insn, expr)) |
2262 | { |
2263 | if (sched_verbose >= 6) |
2264 | sel_print (fmt: "implicit clobbers conflict detected, " ); |
2265 | return MOVEUP_EXPR_NULL; |
2266 | } |
2267 | EXPR_TARGET_AVAILABLE (expr) = false; |
2268 | was_target_conflict = true; |
2269 | as_rhs = true; |
2270 | } |
2271 | |
2272 | /* At this point we have either separable insns, that will be lifted |
2273 | up only as RHSes, or non-separable insns with no dependency in lhs. |
2274 | If dependency is in RHS, then try to perform substitution and move up |
2275 | substituted RHS: |
2276 | |
2277 | Ex. 1: Ex.2 |
2278 | y = x; y = x; |
2279 | z = y*2; y = y*2; |
2280 | |
2281 | In Ex.1 y*2 can be substituted for x*2 and the whole operation can be |
2282 | moved above y=x assignment as z=x*2. |
2283 | |
2284 | In Ex.2 y*2 also can be substituted for x*2, but only the right hand |
2285 | side can be moved because of the output dependency. The operation was |
2286 | cropped to its rhs above. */ |
2287 | if (has_dep_p[DEPS_IN_RHS]) |
2288 | { |
2289 | ds_t *rhs_dsp = &has_dep_p[DEPS_IN_RHS]; |
2290 | |
2291 | /* Can't substitute UNIQUE VINSNs. */ |
2292 | gcc_assert (!VINSN_UNIQUE_P (vi)); |
2293 | |
2294 | if (can_speculate_dep_p (ds: *rhs_dsp)) |
2295 | { |
2296 | int res; |
2297 | |
2298 | res = speculate_expr (expr, *rhs_dsp); |
2299 | if (res >= 0) |
2300 | { |
2301 | /* Speculation was successful. */ |
2302 | *rhs_dsp = 0; |
2303 | was_changed = (res > 0); |
2304 | if (res == 2) |
2305 | was_target_conflict = true; |
2306 | if (ptrans_type) |
2307 | *ptrans_type = TRANS_SPECULATION; |
2308 | } |
2309 | else |
2310 | return MOVEUP_EXPR_NULL; |
2311 | } |
2312 | else if (can_substitute_through_p (insn: through_insn, |
2313 | ds: *rhs_dsp) |
2314 | && substitute_reg_in_expr (expr, insn: through_insn, undo: false)) |
2315 | { |
2316 | /* ??? We cannot perform substitution AND speculation on the same |
2317 | insn. */ |
2318 | gcc_assert (!was_changed); |
2319 | was_changed = true; |
2320 | if (ptrans_type) |
2321 | *ptrans_type = TRANS_SUBSTITUTION; |
2322 | EXPR_WAS_SUBSTITUTED (expr) = true; |
2323 | } |
2324 | else |
2325 | return MOVEUP_EXPR_NULL; |
2326 | } |
2327 | |
2328 | /* Don't move trapping insns through jumps. |
2329 | This check should be at the end to give a chance to control speculation |
2330 | to perform its duties. */ |
2331 | if (CANT_MOVE_TRAPPING (expr, through_insn)) |
2332 | return MOVEUP_EXPR_NULL; |
2333 | |
2334 | return (was_changed |
2335 | ? MOVEUP_EXPR_CHANGED |
2336 | : (as_rhs |
2337 | ? MOVEUP_EXPR_AS_RHS |
2338 | : MOVEUP_EXPR_SAME)); |
2339 | } |
2340 | |
2341 | /* Try to look at bitmap caches for EXPR and INSN pair, return true |
2342 | if successful. When INSIDE_INSN_GROUP, also try ignore dependencies |
2343 | that can exist within a parallel group. Write to RES the resulting |
2344 | code for moveup_expr. */ |
2345 | static bool |
2346 | try_bitmap_cache (expr_t expr, insn_t insn, |
2347 | bool inside_insn_group, |
2348 | enum MOVEUP_EXPR_CODE *res) |
2349 | { |
2350 | int expr_uid = INSN_UID (EXPR_INSN_RTX (expr)); |
2351 | |
2352 | /* First check whether we've analyzed this situation already. */ |
2353 | if (bitmap_bit_p (INSN_ANALYZED_DEPS (insn), expr_uid)) |
2354 | { |
2355 | if (bitmap_bit_p (INSN_FOUND_DEPS (insn), expr_uid)) |
2356 | { |
2357 | if (sched_verbose >= 6) |
2358 | sel_print (fmt: "removed (cached)\n" ); |
2359 | *res = MOVEUP_EXPR_NULL; |
2360 | return true; |
2361 | } |
2362 | else |
2363 | { |
2364 | if (sched_verbose >= 6) |
2365 | sel_print (fmt: "unchanged (cached)\n" ); |
2366 | *res = MOVEUP_EXPR_SAME; |
2367 | return true; |
2368 | } |
2369 | } |
2370 | else if (bitmap_bit_p (INSN_FOUND_DEPS (insn), expr_uid)) |
2371 | { |
2372 | if (inside_insn_group) |
2373 | { |
2374 | if (sched_verbose >= 6) |
2375 | sel_print (fmt: "unchanged (as RHS, cached, inside insn group)\n" ); |
2376 | *res = MOVEUP_EXPR_SAME; |
2377 | return true; |
2378 | |
2379 | } |
2380 | else |
2381 | EXPR_TARGET_AVAILABLE (expr) = false; |
2382 | |
2383 | /* This is the only case when propagation result can change over time, |
2384 | as we can dynamically switch off scheduling as RHS. In this case, |
2385 | just check the flag to reach the correct decision. */ |
2386 | if (enable_schedule_as_rhs_p) |
2387 | { |
2388 | if (sched_verbose >= 6) |
2389 | sel_print (fmt: "unchanged (as RHS, cached)\n" ); |
2390 | *res = MOVEUP_EXPR_AS_RHS; |
2391 | return true; |
2392 | } |
2393 | else |
2394 | { |
2395 | if (sched_verbose >= 6) |
2396 | sel_print (fmt: "removed (cached as RHS, but renaming" |
2397 | " is now disabled)\n" ); |
2398 | *res = MOVEUP_EXPR_NULL; |
2399 | return true; |
2400 | } |
2401 | } |
2402 | |
2403 | return false; |
2404 | } |
2405 | |
2406 | /* Try to look at bitmap caches for EXPR and INSN pair, return true |
2407 | if successful. Write to RES the resulting code for moveup_expr. */ |
2408 | static bool |
2409 | try_transformation_cache (expr_t expr, insn_t insn, |
2410 | enum MOVEUP_EXPR_CODE *res) |
2411 | { |
2412 | struct transformed_insns *pti |
2413 | = (struct transformed_insns *) |
2414 | htab_find_with_hash (INSN_TRANSFORMED_INSNS (insn), |
2415 | &EXPR_VINSN (expr), |
2416 | VINSN_HASH_RTX (EXPR_VINSN (expr))); |
2417 | if (pti) |
2418 | { |
2419 | /* This EXPR was already moved through this insn and was |
2420 | changed as a result. Fetch the proper data from |
2421 | the hashtable. */ |
2422 | insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (expr), |
2423 | INSN_UID (insn), pti->type, |
2424 | pti->vinsn_old, pti->vinsn_new, |
2425 | EXPR_SPEC_DONE_DS (expr)); |
2426 | |
2427 | if (INSN_IN_STREAM_P (VINSN_INSN_RTX (pti->vinsn_new))) |
2428 | pti->vinsn_new = vinsn_copy (pti->vinsn_new, true); |
2429 | change_vinsn_in_expr (expr, pti->vinsn_new); |
2430 | if (pti->was_target_conflict) |
2431 | EXPR_TARGET_AVAILABLE (expr) = false; |
2432 | if (pti->type == TRANS_SPECULATION) |
2433 | { |
2434 | EXPR_SPEC_DONE_DS (expr) = pti->ds; |
2435 | EXPR_NEEDS_SPEC_CHECK_P (expr) |= pti->needs_check; |
2436 | } |
2437 | |
2438 | if (sched_verbose >= 6) |
2439 | { |
2440 | sel_print (fmt: "changed (cached): " ); |
2441 | dump_expr (expr); |
2442 | sel_print (fmt: "\n" ); |
2443 | } |
2444 | |
2445 | *res = MOVEUP_EXPR_CHANGED; |
2446 | return true; |
2447 | } |
2448 | |
2449 | return false; |
2450 | } |
2451 | |
2452 | /* Update bitmap caches on INSN with result RES of propagating EXPR. */ |
2453 | static void |
2454 | update_bitmap_cache (expr_t expr, insn_t insn, bool inside_insn_group, |
2455 | enum MOVEUP_EXPR_CODE res) |
2456 | { |
2457 | int expr_uid = INSN_UID (EXPR_INSN_RTX (expr)); |
2458 | |
2459 | /* Do not cache result of propagating jumps through an insn group, |
2460 | as it is always true, which is not useful outside the group. */ |
2461 | if (inside_insn_group) |
2462 | return; |
2463 | |
2464 | if (res == MOVEUP_EXPR_NULL) |
2465 | { |
2466 | bitmap_set_bit (INSN_ANALYZED_DEPS (insn), expr_uid); |
2467 | bitmap_set_bit (INSN_FOUND_DEPS (insn), expr_uid); |
2468 | } |
2469 | else if (res == MOVEUP_EXPR_SAME) |
2470 | { |
2471 | bitmap_set_bit (INSN_ANALYZED_DEPS (insn), expr_uid); |
2472 | bitmap_clear_bit (INSN_FOUND_DEPS (insn), expr_uid); |
2473 | } |
2474 | else if (res == MOVEUP_EXPR_AS_RHS) |
2475 | { |
2476 | bitmap_clear_bit (INSN_ANALYZED_DEPS (insn), expr_uid); |
2477 | bitmap_set_bit (INSN_FOUND_DEPS (insn), expr_uid); |
2478 | } |
2479 | else |
2480 | gcc_unreachable (); |
2481 | } |
2482 | |
2483 | /* Update hashtable on INSN with changed EXPR, old EXPR_OLD_VINSN |
2484 | and transformation type TRANS_TYPE. */ |
2485 | static void |
2486 | update_transformation_cache (expr_t expr, insn_t insn, |
2487 | bool inside_insn_group, |
2488 | enum local_trans_type trans_type, |
2489 | vinsn_t expr_old_vinsn) |
2490 | { |
2491 | struct transformed_insns *pti; |
2492 | |
2493 | if (inside_insn_group) |
2494 | return; |
2495 | |
2496 | pti = XNEW (struct transformed_insns); |
2497 | pti->vinsn_old = expr_old_vinsn; |
2498 | pti->vinsn_new = EXPR_VINSN (expr); |
2499 | pti->type = trans_type; |
2500 | pti->was_target_conflict = was_target_conflict; |
2501 | pti->ds = EXPR_SPEC_DONE_DS (expr); |
2502 | pti->needs_check = EXPR_NEEDS_SPEC_CHECK_P (expr); |
2503 | vinsn_attach (pti->vinsn_old); |
2504 | vinsn_attach (pti->vinsn_new); |
2505 | *((struct transformed_insns **) |
2506 | htab_find_slot_with_hash (INSN_TRANSFORMED_INSNS (insn), |
2507 | pti, VINSN_HASH_RTX (expr_old_vinsn), |
2508 | INSERT)) = pti; |
2509 | } |
2510 | |
2511 | /* Same as moveup_expr, but first looks up the result of |
2512 | transformation in caches. */ |
2513 | static enum MOVEUP_EXPR_CODE |
2514 | moveup_expr_cached (expr_t expr, insn_t insn, bool inside_insn_group) |
2515 | { |
2516 | enum MOVEUP_EXPR_CODE res; |
2517 | bool got_answer = false; |
2518 | |
2519 | if (sched_verbose >= 6) |
2520 | { |
2521 | sel_print (fmt: "Moving " ); |
2522 | dump_expr (expr); |
2523 | sel_print (fmt: " through %d: " , INSN_UID (insn)); |
2524 | } |
2525 | |
2526 | if (DEBUG_INSN_P (EXPR_INSN_RTX (expr)) |
2527 | && BLOCK_FOR_INSN (EXPR_INSN_RTX (expr)) |
2528 | && (sel_bb_head (BLOCK_FOR_INSN (EXPR_INSN_RTX (expr))) |
2529 | == EXPR_INSN_RTX (expr))) |
2530 | /* Don't use cached information for debug insns that are heads of |
2531 | basic blocks. */; |
2532 | else if (try_bitmap_cache (expr, insn, inside_insn_group, res: &res)) |
2533 | /* When inside insn group, we do not want remove stores conflicting |
2534 | with previosly issued loads. */ |
2535 | got_answer = ! inside_insn_group || res != MOVEUP_EXPR_NULL; |
2536 | else if (try_transformation_cache (expr, insn, res: &res)) |
2537 | got_answer = true; |
2538 | |
2539 | if (! got_answer) |
2540 | { |
2541 | /* Invoke moveup_expr and record the results. */ |
2542 | vinsn_t expr_old_vinsn = EXPR_VINSN (expr); |
2543 | ds_t expr_old_spec_ds = EXPR_SPEC_DONE_DS (expr); |
2544 | int expr_uid = INSN_UID (VINSN_INSN_RTX (expr_old_vinsn)); |
2545 | bool unique_p = VINSN_UNIQUE_P (expr_old_vinsn); |
2546 | enum local_trans_type trans_type = TRANS_SUBSTITUTION; |
2547 | |
2548 | /* ??? Invent something better than this. We can't allow old_vinsn |
2549 | to go, we need it for the history vector. */ |
2550 | vinsn_attach (expr_old_vinsn); |
2551 | |
2552 | res = moveup_expr (expr, through_insn: insn, inside_insn_group, |
2553 | ptrans_type: &trans_type); |
2554 | switch (res) |
2555 | { |
2556 | case MOVEUP_EXPR_NULL: |
2557 | update_bitmap_cache (expr, insn, inside_insn_group, res); |
2558 | if (sched_verbose >= 6) |
2559 | sel_print (fmt: "removed\n" ); |
2560 | break; |
2561 | |
2562 | case MOVEUP_EXPR_SAME: |
2563 | update_bitmap_cache (expr, insn, inside_insn_group, res); |
2564 | if (sched_verbose >= 6) |
2565 | sel_print (fmt: "unchanged\n" ); |
2566 | break; |
2567 | |
2568 | case MOVEUP_EXPR_AS_RHS: |
2569 | gcc_assert (!unique_p || inside_insn_group); |
2570 | update_bitmap_cache (expr, insn, inside_insn_group, res); |
2571 | if (sched_verbose >= 6) |
2572 | sel_print (fmt: "unchanged (as RHS)\n" ); |
2573 | break; |
2574 | |
2575 | case MOVEUP_EXPR_CHANGED: |
2576 | gcc_assert (INSN_UID (EXPR_INSN_RTX (expr)) != expr_uid |
2577 | || EXPR_SPEC_DONE_DS (expr) != expr_old_spec_ds); |
2578 | insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (expr), |
2579 | INSN_UID (insn), trans_type, |
2580 | expr_old_vinsn, EXPR_VINSN (expr), |
2581 | expr_old_spec_ds); |
2582 | update_transformation_cache (expr, insn, inside_insn_group, |
2583 | trans_type, expr_old_vinsn); |
2584 | if (sched_verbose >= 6) |
2585 | { |
2586 | sel_print (fmt: "changed: " ); |
2587 | dump_expr (expr); |
2588 | sel_print (fmt: "\n" ); |
2589 | } |
2590 | break; |
2591 | default: |
2592 | gcc_unreachable (); |
2593 | } |
2594 | |
2595 | vinsn_detach (expr_old_vinsn); |
2596 | } |
2597 | |
2598 | return res; |
2599 | } |
2600 | |
2601 | /* Moves an av set AVP up through INSN, performing necessary |
2602 | transformations. */ |
2603 | static void |
2604 | moveup_set_expr (av_set_t *avp, insn_t insn, bool inside_insn_group) |
2605 | { |
2606 | av_set_iterator i; |
2607 | expr_t expr; |
2608 | |
2609 | FOR_EACH_EXPR_1 (expr, i, avp) |
2610 | { |
2611 | |
2612 | switch (moveup_expr_cached (expr, insn, inside_insn_group)) |
2613 | { |
2614 | case MOVEUP_EXPR_SAME: |
2615 | case MOVEUP_EXPR_AS_RHS: |
2616 | break; |
2617 | |
2618 | case MOVEUP_EXPR_NULL: |
2619 | av_set_iter_remove (&i); |
2620 | break; |
2621 | |
2622 | case MOVEUP_EXPR_CHANGED: |
2623 | expr = merge_with_other_exprs (avp, &i, expr); |
2624 | break; |
2625 | |
2626 | default: |
2627 | gcc_unreachable (); |
2628 | } |
2629 | } |
2630 | } |
2631 | |
2632 | /* Moves AVP set along PATH. */ |
2633 | static void |
2634 | moveup_set_inside_insn_group (av_set_t *avp, ilist_t path) |
2635 | { |
2636 | int last_cycle; |
2637 | |
2638 | if (sched_verbose >= 6) |
2639 | sel_print (fmt: "Moving expressions up in the insn group...\n" ); |
2640 | if (! path) |
2641 | return; |
2642 | last_cycle = INSN_SCHED_CYCLE (ILIST_INSN (path)); |
2643 | while (path |
2644 | && INSN_SCHED_CYCLE (ILIST_INSN (path)) == last_cycle) |
2645 | { |
2646 | moveup_set_expr (avp, ILIST_INSN (path), inside_insn_group: true); |
2647 | path = ILIST_NEXT (path); |
2648 | } |
2649 | } |
2650 | |
2651 | /* Returns true if after moving EXPR along PATH it equals to EXPR_VLIW. */ |
2652 | static bool |
2653 | equal_after_moveup_path_p (expr_t expr, ilist_t path, expr_t expr_vliw) |
2654 | { |
2655 | expr_def _tmp, *tmp = &_tmp; |
2656 | int last_cycle; |
2657 | bool res = true; |
2658 | |
2659 | copy_expr_onside (tmp, expr); |
2660 | last_cycle = path ? INSN_SCHED_CYCLE (ILIST_INSN (path)) : 0; |
2661 | while (path |
2662 | && res |
2663 | && INSN_SCHED_CYCLE (ILIST_INSN (path)) == last_cycle) |
2664 | { |
2665 | res = (moveup_expr_cached (expr: tmp, ILIST_INSN (path), inside_insn_group: true) |
2666 | != MOVEUP_EXPR_NULL); |
2667 | path = ILIST_NEXT (path); |
2668 | } |
2669 | |
2670 | if (res) |
2671 | { |
2672 | vinsn_t tmp_vinsn = EXPR_VINSN (tmp); |
2673 | vinsn_t expr_vliw_vinsn = EXPR_VINSN (expr_vliw); |
2674 | |
2675 | if (tmp_vinsn != expr_vliw_vinsn) |
2676 | res = vinsn_equal_p (tmp_vinsn, expr_vliw_vinsn); |
2677 | } |
2678 | |
2679 | clear_expr (tmp); |
2680 | return res; |
2681 | } |
2682 | |
2683 | |
2684 | /* Functions that compute av and lv sets. */ |
2685 | |
2686 | /* Returns true if INSN is not a downward continuation of the given path P in |
2687 | the current stage. */ |
2688 | static bool |
2689 | is_ineligible_successor (insn_t insn, ilist_t p) |
2690 | { |
2691 | insn_t prev_insn; |
2692 | |
2693 | /* Check if insn is not deleted. */ |
2694 | if (PREV_INSN (insn) && NEXT_INSN (insn: PREV_INSN (insn)) != insn) |
2695 | gcc_unreachable (); |
2696 | else if (NEXT_INSN (insn) && PREV_INSN (insn: NEXT_INSN (insn)) != insn) |
2697 | gcc_unreachable (); |
2698 | |
2699 | /* If it's the first insn visited, then the successor is ok. */ |
2700 | if (!p) |
2701 | return false; |
2702 | |
2703 | prev_insn = ILIST_INSN (p); |
2704 | |
2705 | if (/* a backward edge. */ |
2706 | INSN_SEQNO (insn) < INSN_SEQNO (prev_insn) |
2707 | /* is already visited. */ |
2708 | || (INSN_SEQNO (insn) == INSN_SEQNO (prev_insn) |
2709 | && (ilist_is_in_p (l: p, insn) |
2710 | /* We can reach another fence here and still seqno of insn |
2711 | would be equal to seqno of prev_insn. This is possible |
2712 | when prev_insn is a previously created bookkeeping copy. |
2713 | In that case it'd get a seqno of insn. Thus, check here |
2714 | whether insn is in current fence too. */ |
2715 | || IN_CURRENT_FENCE_P (insn))) |
2716 | /* Was already scheduled on this round. */ |
2717 | || (INSN_SEQNO (insn) > INSN_SEQNO (prev_insn) |
2718 | && IN_CURRENT_FENCE_P (insn)) |
2719 | /* An insn from another fence could also be |
2720 | scheduled earlier even if this insn is not in |
2721 | a fence list right now. Check INSN_SCHED_CYCLE instead. */ |
2722 | || (!pipelining_p |
2723 | && INSN_SCHED_TIMES (insn) > 0)) |
2724 | return true; |
2725 | else |
2726 | return false; |
2727 | } |
2728 | |
2729 | /* Computes the av_set below the last bb insn INSN, doing all the 'dirty work' |
2730 | of handling multiple successors and properly merging its av_sets. P is |
2731 | the current path traversed. WS is the size of lookahead window. |
2732 | Return the av set computed. */ |
2733 | static av_set_t |
2734 | compute_av_set_at_bb_end (insn_t insn, ilist_t p, int ws) |
2735 | { |
2736 | struct succs_info *sinfo; |
2737 | av_set_t expr_in_all_succ_branches = NULL; |
2738 | int is; |
2739 | insn_t succ, zero_succ = NULL; |
2740 | av_set_t av1 = NULL; |
2741 | |
2742 | gcc_assert (sel_bb_end_p (insn)); |
2743 | |
2744 | /* Find different kind of successors needed for correct computing of |
2745 | SPEC and TARGET_AVAILABLE attributes. */ |
2746 | sinfo = compute_succs_info (insn, SUCCS_NORMAL); |
2747 | |
2748 | /* Debug output. */ |
2749 | if (sched_verbose >= 6) |
2750 | { |
2751 | sel_print (fmt: "successors of bb end (%d): " , INSN_UID (insn)); |
2752 | dump_insn_vector (sinfo->succs_ok); |
2753 | sel_print (fmt: "\n" ); |
2754 | if (sinfo->succs_ok_n != sinfo->all_succs_n) |
2755 | sel_print (fmt: "real successors num: %d\n" , sinfo->all_succs_n); |
2756 | } |
2757 | |
2758 | /* Add insn to the tail of current path. */ |
2759 | ilist_add (lp: &p, insn); |
2760 | |
2761 | FOR_EACH_VEC_ELT (sinfo->succs_ok, is, succ) |
2762 | { |
2763 | av_set_t succ_set; |
2764 | |
2765 | /* We will edit SUCC_SET and EXPR_SPEC field of its elements. */ |
2766 | succ_set = compute_av_set_inside_bb (succ, p, ws, true); |
2767 | |
2768 | av_set_split_usefulness (succ_set, |
2769 | sinfo->probs_ok[is], |
2770 | sinfo->all_prob); |
2771 | |
2772 | if (sinfo->all_succs_n > 1) |
2773 | { |
2774 | /* Find EXPR'es that came from *all* successors and save them |
2775 | into expr_in_all_succ_branches. This set will be used later |
2776 | for calculating speculation attributes of EXPR'es. */ |
2777 | if (is == 0) |
2778 | { |
2779 | expr_in_all_succ_branches = av_set_copy (succ_set); |
2780 | |
2781 | /* Remember the first successor for later. */ |
2782 | zero_succ = succ; |
2783 | } |
2784 | else |
2785 | { |
2786 | av_set_iterator i; |
2787 | expr_t expr; |
2788 | |
2789 | FOR_EACH_EXPR_1 (expr, i, &expr_in_all_succ_branches) |
2790 | if (!av_set_is_in_p (succ_set, EXPR_VINSN (expr))) |
2791 | av_set_iter_remove (&i); |
2792 | } |
2793 | } |
2794 | |
2795 | /* Union the av_sets. Check liveness restrictions on target registers |
2796 | in special case of two successors. */ |
2797 | if (sinfo->succs_ok_n == 2 && is == 1) |
2798 | { |
2799 | basic_block bb0 = BLOCK_FOR_INSN (insn: zero_succ); |
2800 | basic_block bb1 = BLOCK_FOR_INSN (insn: succ); |
2801 | |
2802 | gcc_assert (BB_LV_SET_VALID_P (bb0) && BB_LV_SET_VALID_P (bb1)); |
2803 | av_set_union_and_live (&av1, &succ_set, |
2804 | BB_LV_SET (bb0), |
2805 | BB_LV_SET (bb1), |
2806 | insn); |
2807 | } |
2808 | else |
2809 | av_set_union_and_clear (&av1, &succ_set, insn); |
2810 | } |
2811 | |
2812 | /* Check liveness restrictions via hard way when there are more than |
2813 | two successors. */ |
2814 | if (sinfo->succs_ok_n > 2) |
2815 | FOR_EACH_VEC_ELT (sinfo->succs_ok, is, succ) |
2816 | { |
2817 | basic_block succ_bb = BLOCK_FOR_INSN (insn: succ); |
2818 | av_set_t av_succ = (is_ineligible_successor (insn: succ, p) |
2819 | ? NULL |
2820 | : BB_AV_SET (succ_bb)); |
2821 | |
2822 | gcc_assert (BB_LV_SET_VALID_P (succ_bb)); |
2823 | mark_unavailable_targets (av1, av_succ, BB_LV_SET (succ_bb)); |
2824 | } |
2825 | |
2826 | /* Finally, check liveness restrictions on paths leaving the region. */ |
2827 | if (sinfo->all_succs_n > sinfo->succs_ok_n) |
2828 | FOR_EACH_VEC_ELT (sinfo->succs_other, is, succ) |
2829 | mark_unavailable_targets |
2830 | (av1, NULL, BB_LV_SET (BLOCK_FOR_INSN (succ))); |
2831 | |
2832 | if (sinfo->all_succs_n > 1) |
2833 | { |
2834 | av_set_iterator i; |
2835 | expr_t expr; |
2836 | |
2837 | /* Increase the spec attribute of all EXPR'es that didn't come |
2838 | from all successors. */ |
2839 | FOR_EACH_EXPR (expr, i, av1) |
2840 | if (!av_set_is_in_p (expr_in_all_succ_branches, EXPR_VINSN (expr))) |
2841 | EXPR_SPEC (expr)++; |
2842 | |
2843 | av_set_clear (&expr_in_all_succ_branches); |
2844 | |
2845 | /* Do not move conditional branches through other |
2846 | conditional branches. So, remove all conditional |
2847 | branches from av_set if current operator is a conditional |
2848 | branch. */ |
2849 | av_set_substract_cond_branches (&av1); |
2850 | } |
2851 | |
2852 | ilist_remove (&p); |
2853 | free_succs_info (sinfo); |
2854 | |
2855 | if (sched_verbose >= 6) |
2856 | { |
2857 | sel_print (fmt: "av_succs (%d): " , INSN_UID (insn)); |
2858 | dump_av_set (av1); |
2859 | sel_print (fmt: "\n" ); |
2860 | } |
2861 | |
2862 | return av1; |
2863 | } |
2864 | |
2865 | /* This function computes av_set for the FIRST_INSN by dragging valid |
2866 | av_set through all basic block insns either from the end of basic block |
2867 | (computed using compute_av_set_at_bb_end) or from the insn on which |
2868 | MAX_WS was exceeded. It uses compute_av_set_at_bb_end to compute av_set |
2869 | below the basic block and handling conditional branches. |
2870 | FIRST_INSN - the basic block head, P - path consisting of the insns |
2871 | traversed on the way to the FIRST_INSN (the path is sparse, only bb heads |
2872 | and bb ends are added to the path), WS - current window size, |
2873 | NEED_COPY_P - true if we'll make a copy of av_set before returning it. */ |
2874 | static av_set_t |
2875 | compute_av_set_inside_bb (insn_t first_insn, ilist_t p, int ws, |
2876 | bool need_copy_p) |
2877 | { |
2878 | insn_t cur_insn; |
2879 | int end_ws = ws; |
2880 | insn_t bb_end = sel_bb_end (BLOCK_FOR_INSN (insn: first_insn)); |
2881 | insn_t after_bb_end = NEXT_INSN (insn: bb_end); |
2882 | insn_t last_insn; |
2883 | av_set_t av = NULL; |
2884 | basic_block cur_bb = BLOCK_FOR_INSN (insn: first_insn); |
2885 | |
2886 | /* Return NULL if insn is not on the legitimate downward path. */ |
2887 | if (is_ineligible_successor (insn: first_insn, p)) |
2888 | { |
2889 | if (sched_verbose >= 6) |
2890 | sel_print (fmt: "Insn %d is ineligible_successor\n" , INSN_UID (insn: first_insn)); |
2891 | |
2892 | return NULL; |
2893 | } |
2894 | |
2895 | /* If insn already has valid av(insn) computed, just return it. */ |
2896 | if (AV_SET_VALID_P (first_insn)) |
2897 | { |
2898 | av_set_t av_set; |
2899 | |
2900 | if (sel_bb_head_p (first_insn)) |
2901 | av_set = BB_AV_SET (BLOCK_FOR_INSN (first_insn)); |
2902 | else |
2903 | av_set = NULL; |
2904 | |
2905 | if (sched_verbose >= 6) |
2906 | { |
2907 | sel_print (fmt: "Insn %d has a valid av set: " , INSN_UID (insn: first_insn)); |
2908 | dump_av_set (av_set); |
2909 | sel_print (fmt: "\n" ); |
2910 | } |
2911 | |
2912 | return need_copy_p ? av_set_copy (av_set) : av_set; |
2913 | } |
2914 | |
2915 | ilist_add (lp: &p, insn: first_insn); |
2916 | |
2917 | /* As the result after this loop have completed, in LAST_INSN we'll |
2918 | have the insn which has valid av_set to start backward computation |
2919 | from: it either will be NULL because on it the window size was exceeded |
2920 | or other valid av_set as returned by compute_av_set for the last insn |
2921 | of the basic block. */ |
2922 | for (last_insn = first_insn; last_insn != after_bb_end; |
2923 | last_insn = NEXT_INSN (insn: last_insn)) |
2924 | { |
2925 | /* We may encounter valid av_set not only on bb_head, but also on |
2926 | those insns on which previously MAX_WS was exceeded. */ |
2927 | if (AV_SET_VALID_P (last_insn)) |
2928 | { |
2929 | if (sched_verbose >= 6) |
2930 | sel_print (fmt: "Insn %d has a valid empty av set\n" , INSN_UID (insn: last_insn)); |
2931 | break; |
2932 | } |
2933 | |
2934 | /* The special case: the last insn of the BB may be an |
2935 | ineligible_successor due to its SEQ_NO that was set on |
2936 | it as a bookkeeping. */ |
2937 | if (last_insn != first_insn |
2938 | && is_ineligible_successor (insn: last_insn, p)) |
2939 | { |
2940 | if (sched_verbose >= 6) |
2941 | sel_print (fmt: "Insn %d is ineligible_successor\n" , INSN_UID (insn: last_insn)); |
2942 | break; |
2943 | } |
2944 | |
2945 | if (DEBUG_INSN_P (last_insn)) |
2946 | continue; |
2947 | |
2948 | if (end_ws > max_ws) |
2949 | { |
2950 | /* We can reach max lookahead size at bb_header, so clean av_set |
2951 | first. */ |
2952 | INSN_WS_LEVEL (last_insn) = global_level; |
2953 | |
2954 | if (sched_verbose >= 6) |
2955 | sel_print (fmt: "Insn %d is beyond the software lookahead window size\n" , |
2956 | INSN_UID (insn: last_insn)); |
2957 | break; |
2958 | } |
2959 | |
2960 | end_ws++; |
2961 | } |
2962 | |
2963 | /* Get the valid av_set into AV above the LAST_INSN to start backward |
2964 | computation from. It either will be empty av_set or av_set computed from |
2965 | the successors on the last insn of the current bb. */ |
2966 | if (last_insn != after_bb_end) |
2967 | { |
2968 | av = NULL; |
2969 | |
2970 | /* This is needed only to obtain av_sets that are identical to |
2971 | those computed by the old compute_av_set version. */ |
2972 | if (last_insn == first_insn && !INSN_NOP_P (last_insn)) |
2973 | av_set_add (&av, INSN_EXPR (last_insn)); |
2974 | } |
2975 | else |
2976 | /* END_WS is always already increased by 1 if LAST_INSN == AFTER_BB_END. */ |
2977 | av = compute_av_set_at_bb_end (insn: bb_end, p, ws: end_ws); |
2978 | |
2979 | /* Compute av_set in AV starting from below the LAST_INSN up to |
2980 | location above the FIRST_INSN. */ |
2981 | for (cur_insn = PREV_INSN (insn: last_insn); cur_insn != PREV_INSN (insn: first_insn); |
2982 | cur_insn = PREV_INSN (insn: cur_insn)) |
2983 | if (!INSN_NOP_P (cur_insn)) |
2984 | { |
2985 | expr_t expr; |
2986 | |
2987 | moveup_set_expr (avp: &av, insn: cur_insn, inside_insn_group: false); |
2988 | |
2989 | /* If the expression for CUR_INSN is already in the set, |
2990 | replace it by the new one. */ |
2991 | expr = av_set_lookup (av, INSN_VINSN (cur_insn)); |
2992 | if (expr != NULL) |
2993 | { |
2994 | clear_expr (expr); |
2995 | copy_expr (expr, INSN_EXPR (cur_insn)); |
2996 | } |
2997 | else |
2998 | av_set_add (&av, INSN_EXPR (cur_insn)); |
2999 | } |
3000 | |
3001 | /* Clear stale bb_av_set. */ |
3002 | if (sel_bb_head_p (first_insn)) |
3003 | { |
3004 | av_set_clear (&BB_AV_SET (cur_bb)); |
3005 | BB_AV_SET (cur_bb) = need_copy_p ? av_set_copy (av) : av; |
3006 | BB_AV_LEVEL (cur_bb) = global_level; |
3007 | } |
3008 | |
3009 | if (sched_verbose >= 6) |
3010 | { |
3011 | sel_print (fmt: "Computed av set for insn %d: " , INSN_UID (insn: first_insn)); |
3012 | dump_av_set (av); |
3013 | sel_print (fmt: "\n" ); |
3014 | } |
3015 | |
3016 | ilist_remove (&p); |
3017 | return av; |
3018 | } |
3019 | |
3020 | /* Compute av set before INSN. |
3021 | INSN - the current operation (actual rtx INSN) |
3022 | P - the current path, which is list of insns visited so far |
3023 | WS - software lookahead window size. |
3024 | UNIQUE_P - TRUE, if returned av_set will be changed, hence |
3025 | if we want to save computed av_set in s_i_d, we should make a copy of it. |
3026 | |
3027 | In the resulting set we will have only expressions that don't have delay |
3028 | stalls and nonsubstitutable dependences. */ |
3029 | static av_set_t |
3030 | compute_av_set (insn_t insn, ilist_t p, int ws, bool unique_p) |
3031 | { |
3032 | return compute_av_set_inside_bb (first_insn: insn, p, ws, need_copy_p: unique_p); |
3033 | } |
3034 | |
3035 | /* Propagate a liveness set LV through INSN. */ |
3036 | static void |
3037 | propagate_lv_set (regset lv, insn_t insn) |
3038 | { |
3039 | gcc_assert (INSN_P (insn)); |
3040 | |
3041 | if (INSN_NOP_P (insn)) |
3042 | return; |
3043 | |
3044 | df_simulate_one_insn_backwards (BLOCK_FOR_INSN (insn), insn, lv); |
3045 | } |
3046 | |
3047 | /* Return livness set at the end of BB. */ |
3048 | static regset |
3049 | compute_live_after_bb (basic_block bb) |
3050 | { |
3051 | edge e; |
3052 | edge_iterator ei; |
3053 | regset lv = get_clear_regset_from_pool (); |
3054 | |
3055 | gcc_assert (!ignore_first); |
3056 | |
3057 | FOR_EACH_EDGE (e, ei, bb->succs) |
3058 | if (sel_bb_empty_p (e->dest)) |
3059 | { |
3060 | if (! BB_LV_SET_VALID_P (e->dest)) |
3061 | { |
3062 | gcc_unreachable (); |
3063 | gcc_assert (BB_LV_SET (e->dest) == NULL); |
3064 | BB_LV_SET (e->dest) = compute_live_after_bb (bb: e->dest); |
3065 | BB_LV_SET_VALID_P (e->dest) = true; |
3066 | } |
3067 | IOR_REG_SET (lv, BB_LV_SET (e->dest)); |
3068 | } |
3069 | else |
3070 | IOR_REG_SET (lv, compute_live (sel_bb_head (e->dest))); |
3071 | |
3072 | return lv; |
3073 | } |
3074 | |
3075 | /* Compute the set of all live registers at the point before INSN and save |
3076 | it at INSN if INSN is bb header. */ |
3077 | regset |
3078 | compute_live (insn_t insn) |
3079 | { |
3080 | basic_block bb = BLOCK_FOR_INSN (insn); |
3081 | insn_t final, temp; |
3082 | regset lv; |
3083 | |
3084 | /* Return the valid set if we're already on it. */ |
3085 | if (!ignore_first) |
3086 | { |
3087 | regset src = NULL; |
3088 | |
3089 | if (sel_bb_head_p (insn) && BB_LV_SET_VALID_P (bb)) |
3090 | src = BB_LV_SET (bb); |
3091 | else |
3092 | { |
3093 | gcc_assert (in_current_region_p (bb)); |
3094 | if (INSN_LIVE_VALID_P (insn)) |
3095 | src = INSN_LIVE (insn); |
3096 | } |
3097 | |
3098 | if (src) |
3099 | { |
3100 | lv = get_regset_from_pool (); |
3101 | COPY_REG_SET (lv, src); |
3102 | |
3103 | if (sel_bb_head_p (insn) && ! BB_LV_SET_VALID_P (bb)) |
3104 | { |
3105 | COPY_REG_SET (BB_LV_SET (bb), lv); |
3106 | BB_LV_SET_VALID_P (bb) = true; |
3107 | } |
3108 | |
3109 | return_regset_to_pool (lv); |
3110 | return lv; |
3111 | } |
3112 | } |
3113 | |
3114 | /* We've skipped the wrong lv_set. Don't skip the right one. */ |
3115 | ignore_first = false; |
3116 | gcc_assert (in_current_region_p (bb)); |
3117 | |
3118 | /* Find a valid LV set in this block or below, if needed. |
3119 | Start searching from the next insn: either ignore_first is true, or |
3120 | INSN doesn't have a correct live set. */ |
3121 | temp = NEXT_INSN (insn); |
3122 | final = NEXT_INSN (BB_END (bb)); |
3123 | while (temp != final && ! INSN_LIVE_VALID_P (temp)) |
3124 | temp = NEXT_INSN (insn: temp); |
3125 | if (temp == final) |
3126 | { |
3127 | lv = compute_live_after_bb (bb); |
3128 | temp = PREV_INSN (insn: temp); |
3129 | } |
3130 | else |
3131 | { |
3132 | lv = get_regset_from_pool (); |
3133 | COPY_REG_SET (lv, INSN_LIVE (temp)); |
3134 | } |
3135 | |
3136 | /* Put correct lv sets on the insns which have bad sets. */ |
3137 | final = PREV_INSN (insn); |
3138 | while (temp != final) |
3139 | { |
3140 | propagate_lv_set (lv, insn: temp); |
3141 | COPY_REG_SET (INSN_LIVE (temp), lv); |
3142 | INSN_LIVE_VALID_P (temp) = true; |
3143 | temp = PREV_INSN (insn: temp); |
3144 | } |
3145 | |
3146 | /* Also put it in a BB. */ |
3147 | if (sel_bb_head_p (insn)) |
3148 | { |
3149 | basic_block bb = BLOCK_FOR_INSN (insn); |
3150 | |
3151 | COPY_REG_SET (BB_LV_SET (bb), lv); |
3152 | BB_LV_SET_VALID_P (bb) = true; |
3153 | } |
3154 | |
3155 | /* We return LV to the pool, but will not clear it there. Thus we can |
3156 | legimatelly use LV till the next use of regset_pool_get (). */ |
3157 | return_regset_to_pool (lv); |
3158 | return lv; |
3159 | } |
3160 | |
3161 | /* Update liveness sets for INSN. */ |
3162 | static inline void |
3163 | update_liveness_on_insn (rtx_insn *insn) |
3164 | { |
3165 | ignore_first = true; |
3166 | compute_live (insn); |
3167 | } |
3168 | |
3169 | /* Compute liveness below INSN and write it into REGS. */ |
3170 | static inline void |
3171 | compute_live_below_insn (rtx_insn *insn, regset regs) |
3172 | { |
3173 | rtx_insn *succ; |
3174 | succ_iterator si; |
3175 | |
3176 | FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL) |
3177 | IOR_REG_SET (regs, compute_live (succ)); |
3178 | } |
3179 | |
3180 | /* Update the data gathered in av and lv sets starting from INSN. */ |
3181 | static void |
3182 | update_data_sets (rtx_insn *insn) |
3183 | { |
3184 | update_liveness_on_insn (insn); |
3185 | if (sel_bb_head_p (insn)) |
3186 | { |
3187 | gcc_assert (AV_LEVEL (insn) != 0); |
3188 | BB_AV_LEVEL (BLOCK_FOR_INSN (insn)) = -1; |
3189 | compute_av_set (insn, NULL, ws: 0, unique_p: 0); |
3190 | } |
3191 | } |
3192 | |
3193 | |
3194 | /* Helper for move_op () and find_used_regs (). |
3195 | Return speculation type for which a check should be created on the place |
3196 | of INSN. EXPR is one of the original ops we are searching for. */ |
3197 | static ds_t |
3198 | get_spec_check_type_for_insn (insn_t insn, expr_t expr) |
3199 | { |
3200 | ds_t to_check_ds; |
3201 | ds_t already_checked_ds = EXPR_SPEC_DONE_DS (INSN_EXPR (insn)); |
3202 | |
3203 | to_check_ds = EXPR_SPEC_TO_CHECK_DS (expr); |
3204 | |
3205 | if (targetm.sched.get_insn_checked_ds) |
3206 | already_checked_ds |= targetm.sched.get_insn_checked_ds (insn); |
3207 | |
3208 | if (spec_info != NULL |
3209 | && (spec_info->flags & SEL_SCHED_SPEC_DONT_CHECK_CONTROL)) |
3210 | already_checked_ds |= BEGIN_CONTROL; |
3211 | |
3212 | already_checked_ds = ds_get_speculation_types (already_checked_ds); |
3213 | |
3214 | to_check_ds &= ~already_checked_ds; |
3215 | |
3216 | return to_check_ds; |
3217 | } |
3218 | |
3219 | /* Find the set of registers that are unavailable for storing expres |
3220 | while moving ORIG_OPS up on the path starting from INSN due to |
3221 | liveness (USED_REGS) or hardware restrictions (REG_RENAME_P). |
3222 | |
3223 | All the original operations found during the traversal are saved in the |
3224 | ORIGINAL_INSNS list. |
3225 | |
3226 | REG_RENAME_P denotes the set of hardware registers that |
3227 | cannot be used with renaming due to the register class restrictions, |
3228 | mode restrictions and other (the register we'll choose should be |
3229 | compatible class with the original uses, shouldn't be in call_used_regs, |
3230 | should be HARD_REGNO_RENAME_OK etc). |
3231 | |
3232 | Returns TRUE if we've found all original insns, FALSE otherwise. |
3233 | |
3234 | This function utilizes code_motion_path_driver (formerly find_used_regs_1) |
3235 | to traverse the code motion paths. This helper function finds registers |
3236 | that are not available for storing expres while moving ORIG_OPS up on the |
3237 | path starting from INSN. A register considered as used on the moving path, |
3238 | if one of the following conditions is not satisfied: |
3239 | |
3240 | (1) a register not set or read on any path from xi to an instance of |
3241 | the original operation, |
3242 | (2) not among the live registers of the point immediately following the |
3243 | first original operation on a given downward path, except for the |
3244 | original target register of the operation, |
3245 | (3) not live on the other path of any conditional branch that is passed |
3246 | by the operation, in case original operations are not present on |
3247 | both paths of the conditional branch. |
3248 | |
3249 | All the original operations found during the traversal are saved in the |
3250 | ORIGINAL_INSNS list. |
3251 | |
3252 | REG_RENAME_P->CROSSED_CALL_ABIS is true, if there is a call insn on the path |
3253 | from INSN to original insn. In this case CALL_USED_REG_SET will be added |
3254 | to unavailable hard regs at the point original operation is found. */ |
3255 | |
3256 | static bool |
3257 | find_used_regs (insn_t insn, av_set_t orig_ops, regset used_regs, |
3258 | struct reg_rename *reg_rename_p, def_list_t *original_insns) |
3259 | { |
3260 | def_list_iterator i; |
3261 | def_t def; |
3262 | int res; |
3263 | bool needs_spec_check_p = false; |
3264 | expr_t expr; |
3265 | av_set_iterator expr_iter; |
3266 | struct fur_static_params sparams; |
3267 | struct cmpd_local_params lparams; |
3268 | |
3269 | /* We haven't visited any blocks yet. */ |
3270 | bitmap_clear (code_motion_visited_blocks); |
3271 | |
3272 | /* Init parameters for code_motion_path_driver. */ |
3273 | sparams.crossed_call_abis = 0; |
3274 | sparams.original_insns = original_insns; |
3275 | sparams.used_regs = used_regs; |
3276 | |
3277 | /* Set the appropriate hooks and data. */ |
3278 | code_motion_path_driver_info = &fur_hooks; |
3279 | |
3280 | res = code_motion_path_driver (insn, orig_ops, NULL, &lparams, &sparams); |
3281 | |
3282 | reg_rename_p->crossed_call_abis |= sparams.crossed_call_abis; |
3283 | |
3284 | gcc_assert (res == 1); |
3285 | gcc_assert (original_insns && *original_insns); |
3286 | |
3287 | /* ??? We calculate whether an expression needs a check when computing |
3288 | av sets. This information is not as precise as it could be due to |
3289 | merging this bit in merge_expr. We can do better in find_used_regs, |
3290 | but we want to avoid multiple traversals of the same code motion |
3291 | paths. */ |
3292 | FOR_EACH_EXPR (expr, expr_iter, orig_ops) |
3293 | needs_spec_check_p |= EXPR_NEEDS_SPEC_CHECK_P (expr); |
3294 | |
3295 | /* Mark hardware regs in REG_RENAME_P that are not suitable |
3296 | for renaming expr in INSN due to hardware restrictions (register class, |
3297 | modes compatibility etc). */ |
3298 | FOR_EACH_DEF (def, i, *original_insns) |
3299 | { |
3300 | vinsn_t vinsn = INSN_VINSN (def->orig_insn); |
3301 | |
3302 | if (VINSN_SEPARABLE_P (vinsn)) |
3303 | mark_unavailable_hard_regs (def, reg_rename_p, used_regs); |
3304 | |
3305 | /* Do not allow clobbering of ld.[sa] address in case some of the |
3306 | original operations need a check. */ |
3307 | if (needs_spec_check_p) |
3308 | IOR_REG_SET (used_regs, VINSN_REG_USES (vinsn)); |
3309 | } |
3310 | |
3311 | return true; |
3312 | } |
3313 | |
3314 | |
3315 | /* Functions to choose the best insn from available ones. */ |
3316 | |
3317 | /* Adjusts the priority for EXPR using the backend *_adjust_priority hook. */ |
3318 | static int |
3319 | sel_target_adjust_priority (expr_t expr) |
3320 | { |
3321 | int priority = EXPR_PRIORITY (expr); |
3322 | int new_priority; |
3323 | |
3324 | if (targetm.sched.adjust_priority) |
3325 | new_priority = targetm.sched.adjust_priority (EXPR_INSN_RTX (expr), priority); |
3326 | else |
3327 | new_priority = priority; |
3328 | |
3329 | /* If the priority has changed, adjust EXPR_PRIORITY_ADJ accordingly. */ |
3330 | EXPR_PRIORITY_ADJ (expr) = new_priority - EXPR_PRIORITY (expr); |
3331 | |
3332 | if (sched_verbose >= 4) |
3333 | sel_print (fmt: "sel_target_adjust_priority: insn %d, %d+%d = %d.\n" , |
3334 | INSN_UID (EXPR_INSN_RTX (expr)), EXPR_PRIORITY (expr), |
3335 | EXPR_PRIORITY_ADJ (expr), new_priority); |
3336 | |
3337 | return new_priority; |
3338 | } |
3339 | |
3340 | /* Rank two available exprs for schedule. Never return 0 here. */ |
3341 | static int |
3342 | sel_rank_for_schedule (const void *x, const void *y) |
3343 | { |
3344 | expr_t tmp = *(const expr_t *) y; |
3345 | expr_t tmp2 = *(const expr_t *) x; |
3346 | insn_t tmp_insn, tmp2_insn; |
3347 | vinsn_t tmp_vinsn, tmp2_vinsn; |
3348 | int val; |
3349 | |
3350 | tmp_vinsn = EXPR_VINSN (tmp); |
3351 | tmp2_vinsn = EXPR_VINSN (tmp2); |
3352 | tmp_insn = EXPR_INSN_RTX (tmp); |
3353 | tmp2_insn = EXPR_INSN_RTX (tmp2); |
3354 | |
3355 | /* Schedule debug insns as early as possible. */ |
3356 | if (DEBUG_INSN_P (tmp_insn) && !DEBUG_INSN_P (tmp2_insn)) |
3357 | return -1; |
3358 | else if (DEBUG_INSN_P (tmp2_insn)) |
3359 | return 1; |
3360 | |
3361 | /* Prefer SCHED_GROUP_P insns to any others. */ |
3362 | if (SCHED_GROUP_P (tmp_insn) != SCHED_GROUP_P (tmp2_insn)) |
3363 | { |
3364 | if (VINSN_UNIQUE_P (tmp_vinsn) && VINSN_UNIQUE_P (tmp2_vinsn)) |
3365 | return SCHED_GROUP_P (tmp2_insn) ? 1 : -1; |
3366 | |
3367 | /* Now uniqueness means SCHED_GROUP_P is set, because schedule groups |
3368 | cannot be cloned. */ |
3369 | if (VINSN_UNIQUE_P (tmp2_vinsn)) |
3370 | return 1; |
3371 | return -1; |
3372 | } |
3373 | |
3374 | /* Discourage scheduling of speculative checks. */ |
3375 | val = (sel_insn_is_speculation_check (tmp_insn) |
3376 | - sel_insn_is_speculation_check (tmp2_insn)); |
3377 | if (val) |
3378 | return val; |
3379 | |
3380 | /* Prefer not scheduled insn over scheduled one. */ |
3381 | if (EXPR_SCHED_TIMES (tmp) > 0 || EXPR_SCHED_TIMES (tmp2) > 0) |
3382 | { |
3383 | val = EXPR_SCHED_TIMES (tmp) - EXPR_SCHED_TIMES (tmp2); |
3384 | if (val) |
3385 | return val; |
3386 | } |
3387 | |
3388 | /* Prefer jump over non-jump instruction. */ |
3389 | if (control_flow_insn_p (tmp_insn) && !control_flow_insn_p (tmp2_insn)) |
3390 | return -1; |
3391 | else if (control_flow_insn_p (tmp2_insn) && !control_flow_insn_p (tmp_insn)) |
3392 | return 1; |
3393 | |
3394 | /* Prefer an expr with non-zero usefulness. */ |
3395 | int u1 = EXPR_USEFULNESS (tmp), u2 = EXPR_USEFULNESS (tmp2); |
3396 | |
3397 | if (u1 == 0) |
3398 | { |
3399 | if (u2 == 0) |
3400 | u1 = u2 = 1; |
3401 | else |
3402 | return 1; |
3403 | } |
3404 | else if (u2 == 0) |
3405 | return -1; |
3406 | |
3407 | /* Prefer an expr with greater priority. */ |
3408 | val = (u2 * (EXPR_PRIORITY (tmp2) + EXPR_PRIORITY_ADJ (tmp2)) |
3409 | - u1 * (EXPR_PRIORITY (tmp) + EXPR_PRIORITY_ADJ (tmp))); |
3410 | if (val) |
3411 | return val; |
3412 | |
3413 | if (spec_info != NULL && spec_info->mask != 0) |
3414 | /* This code was taken from haifa-sched.cc: rank_for_schedule (). */ |
3415 | { |
3416 | ds_t ds1, ds2; |
3417 | dw_t dw1, dw2; |
3418 | int dw; |
3419 | |
3420 | ds1 = EXPR_SPEC_DONE_DS (tmp); |
3421 | if (ds1) |
3422 | dw1 = ds_weak (ds1); |
3423 | else |
3424 | dw1 = NO_DEP_WEAK; |
3425 | |
3426 | ds2 = EXPR_SPEC_DONE_DS (tmp2); |
3427 | if (ds2) |
3428 | dw2 = ds_weak (ds2); |
3429 | else |
3430 | dw2 = NO_DEP_WEAK; |
3431 | |
3432 | dw = dw2 - dw1; |
3433 | if (dw > (NO_DEP_WEAK / 8) || dw < -(NO_DEP_WEAK / 8)) |
3434 | return dw; |
3435 | } |
3436 | |
3437 | /* Prefer an old insn to a bookkeeping insn. */ |
3438 | if (INSN_UID (insn: tmp_insn) < first_emitted_uid |
3439 | && INSN_UID (insn: tmp2_insn) >= first_emitted_uid) |
3440 | return -1; |
3441 | if (INSN_UID (insn: tmp_insn) >= first_emitted_uid |
3442 | && INSN_UID (insn: tmp2_insn) < first_emitted_uid) |
3443 | return 1; |
3444 | |
3445 | /* Prefer an insn with smaller UID, as a last resort. |
3446 | We can't safely use INSN_LUID as it is defined only for those insns |
3447 | that are in the stream. */ |
3448 | return INSN_UID (insn: tmp_insn) - INSN_UID (insn: tmp2_insn); |
3449 | } |
3450 | |
3451 | /* Filter out expressions from av set pointed to by AV_PTR |
3452 | that are pipelined too many times. */ |
3453 | static void |
3454 | process_pipelined_exprs (av_set_t *av_ptr) |
3455 | { |
3456 | expr_t expr; |
3457 | av_set_iterator si; |
3458 | |
3459 | /* Don't pipeline already pipelined code as that would increase |
3460 | number of unnecessary register moves. */ |
3461 | FOR_EACH_EXPR_1 (expr, si, av_ptr) |
3462 | { |
3463 | if (EXPR_SCHED_TIMES (expr) |
3464 | >= param_selsched_max_sched_times) |
3465 | av_set_iter_remove (&si); |
3466 | } |
3467 | } |
3468 | |
3469 | /* Filter speculative insns from AV_PTR if we don't want them. */ |
3470 | static void |
3471 | process_spec_exprs (av_set_t *av_ptr) |
3472 | { |
3473 | expr_t expr; |
3474 | av_set_iterator si; |
3475 | |
3476 | if (spec_info == NULL) |
3477 | return; |
3478 | |
3479 | /* Scan *AV_PTR to find out if we want to consider speculative |
3480 | instructions for scheduling. */ |
3481 | FOR_EACH_EXPR_1 (expr, si, av_ptr) |
3482 | { |
3483 | ds_t ds; |
3484 | |
3485 | ds = EXPR_SPEC_DONE_DS (expr); |
3486 | |
3487 | /* The probability of a success is too low - don't speculate. */ |
3488 | if ((ds & SPECULATIVE) |
3489 | && (ds_weak (ds) < spec_info->data_weakness_cutoff |
3490 | || EXPR_USEFULNESS (expr) < spec_info->control_weakness_cutoff |
3491 | || (pipelining_p && false |
3492 | && (ds & DATA_SPEC) |
3493 | && (ds & CONTROL_SPEC)))) |
3494 | { |
3495 | av_set_iter_remove (&si); |
3496 | continue; |
3497 | } |
3498 | } |
3499 | } |
3500 | |
3501 | /* Search for any use-like insns in AV_PTR and decide on scheduling |
3502 | them. Return one when found, and NULL otherwise. |
3503 | Note that we check here whether a USE could be scheduled to avoid |
3504 | an infinite loop later. */ |
3505 | static expr_t |
3506 | process_use_exprs (av_set_t *av_ptr) |
3507 | { |
3508 | expr_t expr; |
3509 | av_set_iterator si; |
3510 | bool uses_present_p = false; |
3511 | bool try_uses_p = true; |
3512 | |
3513 | FOR_EACH_EXPR_1 (expr, si, av_ptr) |
3514 | { |
3515 | /* This will also initialize INSN_CODE for later use. */ |
3516 | if (recog_memoized (EXPR_INSN_RTX (expr)) < 0) |
3517 | { |
3518 | /* If we have a USE in *AV_PTR that was not scheduled yet, |
3519 | do so because it will do good only. */ |
3520 | if (EXPR_SCHED_TIMES (expr) <= 0) |
3521 | { |
3522 | if (EXPR_TARGET_AVAILABLE (expr) == 1) |
3523 | return expr; |
3524 | |
3525 | av_set_iter_remove (&si); |
3526 | } |
3527 | else |
3528 | { |
3529 | gcc_assert (pipelining_p); |
3530 | |
3531 | uses_present_p = true; |
3532 | } |
3533 | } |
3534 | else |
3535 | try_uses_p = false; |
3536 | } |
3537 | |
3538 | if (uses_present_p) |
3539 | { |
3540 | /* If we don't want to schedule any USEs right now and we have some |
3541 | in *AV_PTR, remove them, else just return the first one found. */ |
3542 | if (!try_uses_p) |
3543 | { |
3544 | FOR_EACH_EXPR_1 (expr, si, av_ptr) |
3545 | if (INSN_CODE (EXPR_INSN_RTX (expr)) < 0) |
3546 | av_set_iter_remove (&si); |
3547 | } |
3548 | else |
3549 | { |
3550 | FOR_EACH_EXPR_1 (expr, si, av_ptr) |
3551 | { |
3552 | gcc_assert (INSN_CODE (EXPR_INSN_RTX (expr)) < 0); |
3553 | |
3554 | if (EXPR_TARGET_AVAILABLE (expr) == 1) |
3555 | return expr; |
3556 | |
3557 | av_set_iter_remove (&si); |
3558 | } |
3559 | } |
3560 | } |
3561 | |
3562 | return NULL; |
3563 | } |
3564 | |
3565 | /* Lookup EXPR in VINSN_VEC and return TRUE if found. Also check patterns from |
3566 | EXPR's history of changes. */ |
3567 | static bool |
3568 | vinsn_vec_has_expr_p (vinsn_vec_t vinsn_vec, expr_t expr) |
3569 | { |
3570 | vinsn_t vinsn, expr_vinsn; |
3571 | int n; |
3572 | unsigned i; |
3573 | |
3574 | /* Start with checking expr itself and then proceed with all the old forms |
3575 | of expr taken from its history vector. */ |
3576 | for (i = 0, expr_vinsn = EXPR_VINSN (expr); |
3577 | expr_vinsn; |
3578 | expr_vinsn = (i < EXPR_HISTORY_OF_CHANGES (expr).length () |
3579 | ? EXPR_HISTORY_OF_CHANGES (expr)[i++].old_expr_vinsn |
3580 | : NULL)) |
3581 | FOR_EACH_VEC_ELT (vinsn_vec, n, vinsn) |
3582 | if (VINSN_SEPARABLE_P (vinsn)) |
3583 | { |
3584 | if (vinsn_equal_p (vinsn, expr_vinsn)) |
3585 | return true; |
3586 | } |
3587 | else |
3588 | { |
3589 | /* For non-separable instructions, the blocking insn can have |
3590 | another pattern due to substitution, and we can't choose |
3591 | different register as in the above case. Check all registers |
3592 | being written instead. */ |
3593 | if (bitmap_intersect_p (VINSN_REG_SETS (vinsn), |
3594 | VINSN_REG_SETS (expr_vinsn))) |
3595 | return true; |
3596 | } |
3597 | |
3598 | return false; |
3599 | } |
3600 | |
3601 | /* Return true if either of expressions from ORIG_OPS can be blocked |
3602 | by previously created bookkeeping code. STATIC_PARAMS points to static |
3603 | parameters of move_op. */ |
3604 | static bool |
3605 | av_set_could_be_blocked_by_bookkeeping_p (av_set_t orig_ops, void *static_params) |
3606 | { |
3607 | expr_t expr; |
3608 | av_set_iterator iter; |
3609 | moveop_static_params_p sparams; |
3610 | |
3611 | /* This checks that expressions in ORIG_OPS are not blocked by bookkeeping |
3612 | created while scheduling on another fence. */ |
3613 | FOR_EACH_EXPR (expr, iter, orig_ops) |
3614 | if (vinsn_vec_has_expr_p (vinsn_vec: vec_bookkeeping_blocked_vinsns, expr)) |
3615 | return true; |
3616 | |
3617 | gcc_assert (code_motion_path_driver_info == &move_op_hooks); |
3618 | sparams = (moveop_static_params_p) static_params; |
3619 | |
3620 | /* Expressions can be also blocked by bookkeeping created during current |
3621 | move_op. */ |
3622 | if (bitmap_bit_p (current_copies, INSN_UID (insn: sparams->failed_insn))) |
3623 | FOR_EACH_EXPR (expr, iter, orig_ops) |
3624 | if (moveup_expr_cached (expr, insn: sparams->failed_insn, inside_insn_group: false) != MOVEUP_EXPR_NULL) |
3625 | return true; |
3626 | |
3627 | /* Expressions in ORIG_OPS may have wrong destination register due to |
3628 | renaming. Check with the right register instead. */ |
3629 | if (sparams->dest && REG_P (sparams->dest)) |
3630 | { |
3631 | rtx reg = sparams->dest; |
3632 | vinsn_t failed_vinsn = INSN_VINSN (sparams->failed_insn); |
3633 | |
3634 | if (register_unavailable_p (VINSN_REG_SETS (failed_vinsn), reg) |
3635 | || register_unavailable_p (VINSN_REG_USES (failed_vinsn), reg) |
3636 | || register_unavailable_p (VINSN_REG_CLOBBERS (failed_vinsn), reg)) |
3637 | return true; |
3638 | } |
3639 | |
3640 | return false; |
3641 | } |
3642 | |
3643 | /* Clear VINSN_VEC and detach vinsns. */ |
3644 | static void |
3645 | vinsn_vec_clear (vinsn_vec_t *vinsn_vec) |
3646 | { |
3647 | unsigned len = vinsn_vec->length (); |
3648 | if (len > 0) |
3649 | { |
3650 | vinsn_t vinsn; |
3651 | int n; |
3652 | |
3653 | FOR_EACH_VEC_ELT (*vinsn_vec, n, vinsn) |
3654 | vinsn_detach (vinsn); |
3655 | vinsn_vec->block_remove (ix: 0, len); |
3656 | } |
3657 | } |
3658 | |
3659 | /* Add the vinsn of EXPR to the VINSN_VEC. */ |
3660 | static void |
3661 | vinsn_vec_add (vinsn_vec_t *vinsn_vec, expr_t expr) |
3662 | { |
3663 | vinsn_attach (EXPR_VINSN (expr)); |
3664 | vinsn_vec->safe_push (EXPR_VINSN (expr)); |
3665 | } |
3666 | |
3667 | /* Free the vector representing blocked expressions. */ |
3668 | static void |
3669 | vinsn_vec_free (vinsn_vec_t &vinsn_vec) |
3670 | { |
3671 | vinsn_vec.release (); |
3672 | } |
3673 | |
3674 | /* Increase EXPR_PRIORITY_ADJ for INSN by AMOUNT. */ |
3675 | |
3676 | void sel_add_to_insn_priority (rtx insn, int amount) |
3677 | { |
3678 | EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) += amount; |
3679 | |
3680 | if (sched_verbose >= 2) |
3681 | sel_print (fmt: "sel_add_to_insn_priority: insn %d, by %d (now %d+%d).\n" , |
3682 | INSN_UID (insn), amount, EXPR_PRIORITY (INSN_EXPR (insn)), |
3683 | EXPR_PRIORITY_ADJ (INSN_EXPR (insn))); |
3684 | } |
3685 | |
3686 | /* Turn AV into a vector, filter inappropriate insns and sort it. Return |
3687 | true if there is something to schedule. BNDS and FENCE are current |
3688 | boundaries and fence, respectively. If we need to stall for some cycles |
3689 | before an expr from AV would become available, write this number to |
3690 | *PNEED_STALL. */ |
3691 | static bool |
3692 | fill_vec_av_set (av_set_t av, blist_t bnds, fence_t fence, |
3693 | int *pneed_stall) |
3694 | { |
3695 | av_set_iterator si; |
3696 | expr_t expr; |
3697 | int sched_next_worked = 0, stalled, n; |
3698 | static int av_max_prio, est_ticks_till_branch; |
3699 | int min_need_stall = -1; |
3700 | deps_t dc = BND_DC (BLIST_BND (bnds)); |
3701 | |
3702 | /* Bail out early when the ready list contained only USEs/CLOBBERs that are |
3703 | already scheduled. */ |
3704 | if (av == NULL) |
3705 | return false; |
3706 | |
3707 | /* Empty vector from the previous stuff. */ |
3708 | if (vec_av_set.length () > 0) |
3709 | vec_av_set.block_remove (ix: 0, len: vec_av_set.length ()); |
3710 | |
3711 | /* Turn the set into a vector for sorting and call sel_target_adjust_priority |
3712 | for each insn. */ |
3713 | gcc_assert (vec_av_set.is_empty ()); |
3714 | FOR_EACH_EXPR (expr, si, av) |
3715 | { |
3716 | vec_av_set.safe_push (obj: expr); |
3717 | |
3718 | gcc_assert (EXPR_PRIORITY_ADJ (expr) == 0 || *pneed_stall); |
3719 | |
3720 | /* Adjust priority using target backend hook. */ |
3721 | sel_target_adjust_priority (expr); |
3722 | } |
3723 | |
3724 | /* Sort the vector. */ |
3725 | vec_av_set.qsort (sel_rank_for_schedule); |
3726 | |
3727 | /* We record maximal priority of insns in av set for current instruction |
3728 | group. */ |
3729 | if (FENCE_STARTS_CYCLE_P (fence)) |
3730 | av_max_prio = est_ticks_till_branch = INT_MIN; |
3731 | |
3732 | /* Filter out inappropriate expressions. Loop's direction is reversed to |
3733 | visit "best" instructions first. We assume that vec::unordered_remove |
3734 | moves last element in place of one being deleted. */ |
3735 | for (n = vec_av_set.length () - 1, stalled = 0; n >= 0; n--) |
3736 | { |
3737 | expr_t expr = vec_av_set[n]; |
3738 | insn_t insn = EXPR_INSN_RTX (expr); |
3739 | signed char target_available; |
3740 | bool is_orig_reg_p = true; |
3741 | int need_cycles, new_prio; |
3742 | bool fence_insn_p = INSN_UID (insn) == INSN_UID (FENCE_INSN (fence)); |
3743 | |
3744 | /* Don't allow any insns other than from SCHED_GROUP if we have one. */ |
3745 | if (FENCE_SCHED_NEXT (fence) && insn != FENCE_SCHED_NEXT (fence)) |
3746 | { |
3747 | vec_av_set.unordered_remove (ix: n); |
3748 | continue; |
3749 | } |
3750 | |
3751 | /* Set number of sched_next insns (just in case there |
3752 | could be several). */ |
3753 | if (FENCE_SCHED_NEXT (fence)) |
3754 | sched_next_worked++; |
3755 | |
3756 | /* Check all liveness requirements and try renaming. |
3757 | FIXME: try to minimize calls to this. */ |
3758 | target_available = EXPR_TARGET_AVAILABLE (expr); |
3759 | |
3760 | /* If insn was already scheduled on the current fence, |
3761 | set TARGET_AVAILABLE to -1 no matter what expr's attribute says. */ |
3762 | if (vinsn_vec_has_expr_p (vinsn_vec: vec_target_unavailable_vinsns, expr) |
3763 | && !fence_insn_p) |
3764 | target_available = -1; |
3765 | |
3766 | /* If the availability of the EXPR is invalidated by the insertion of |
3767 | bookkeeping earlier, make sure that we won't choose this expr for |
3768 | scheduling if it's not separable, and if it is separable, then |
3769 | we have to recompute the set of available registers for it. */ |
3770 | if (vinsn_vec_has_expr_p (vinsn_vec: vec_bookkeeping_blocked_vinsns, expr)) |
3771 | { |
3772 | vec_av_set.unordered_remove (ix: n); |
3773 | if (sched_verbose >= 4) |
3774 | sel_print (fmt: "Expr %d is blocked by bookkeeping inserted earlier\n" , |
3775 | INSN_UID (insn)); |
3776 | continue; |
3777 | } |
3778 | |
3779 | if (target_available == true) |
3780 | { |
3781 | /* Do nothing -- we can use an existing register. */ |
3782 | is_orig_reg_p = EXPR_SEPARABLE_P (expr); |
3783 | } |
3784 | else if (/* Non-separable instruction will never |
3785 | get another register. */ |
3786 | (target_available == false |
3787 | && !EXPR_SEPARABLE_P (expr)) |
3788 | /* Don't try to find a register for low-priority expression. */ |
3789 | || (int) vec_av_set.length () - 1 - n >= max_insns_to_rename |
3790 | /* ??? FIXME: Don't try to rename data speculation. */ |
3791 | || (EXPR_SPEC_DONE_DS (expr) & BEGIN_DATA) |
3792 | || ! find_best_reg_for_expr (expr, bnds, is_orig_reg_p: &is_orig_reg_p)) |
3793 | { |
3794 | vec_av_set.unordered_remove (ix: n); |
3795 | if (sched_verbose >= 4) |
3796 | sel_print (fmt: "Expr %d has no suitable target register\n" , |
3797 | INSN_UID (insn)); |
3798 | |
3799 | /* A fence insn should not get here. */ |
3800 | gcc_assert (!fence_insn_p); |
3801 | continue; |
3802 | } |
3803 | |
3804 | /* At this point a fence insn should always be available. */ |
3805 | gcc_assert (!fence_insn_p |
3806 | || INSN_UID (FENCE_INSN (fence)) == INSN_UID (EXPR_INSN_RTX (expr))); |
3807 | |
3808 | /* Filter expressions that need to be renamed or speculated when |
3809 | pipelining, because compensating register copies or speculation |
3810 | checks are likely to be placed near the beginning of the loop, |
3811 | causing a stall. */ |
3812 | if (pipelining_p && EXPR_ORIG_SCHED_CYCLE (expr) > 0 |
3813 | && (!is_orig_reg_p || EXPR_SPEC_DONE_DS (expr) != 0)) |
3814 | { |
3815 | /* Estimation of number of cycles until loop branch for |
3816 | renaming/speculation to be successful. */ |
3817 | int need_n_ticks_till_branch = sel_vinsn_cost (EXPR_VINSN (expr)); |
3818 | |
3819 | if ((int) current_loop_nest->ninsns < 9) |
3820 | { |
3821 | vec_av_set.unordered_remove (ix: n); |
3822 | if (sched_verbose >= 4) |
3823 | sel_print (fmt: "Pipelining expr %d will likely cause stall\n" , |
3824 | INSN_UID (insn)); |
3825 | continue; |
3826 | } |
3827 | |
3828 | if ((int) current_loop_nest->ninsns - num_insns_scheduled |
3829 | < need_n_ticks_till_branch * issue_rate / 2 |
3830 | && est_ticks_till_branch < need_n_ticks_till_branch) |
3831 | { |
3832 | vec_av_set.unordered_remove (ix: n); |
3833 | if (sched_verbose >= 4) |
3834 | sel_print (fmt: "Pipelining expr %d will likely cause stall\n" , |
3835 | INSN_UID (insn)); |
3836 | continue; |
3837 | } |
3838 | } |
3839 | |
3840 | /* We want to schedule speculation checks as late as possible. Discard |
3841 | them from av set if there are instructions with higher priority. */ |
3842 | if (sel_insn_is_speculation_check (insn) |
3843 | && EXPR_PRIORITY (expr) < av_max_prio) |
3844 | { |
3845 | stalled++; |
3846 | min_need_stall = min_need_stall < 0 ? 1 : MIN (min_need_stall, 1); |
3847 | vec_av_set.unordered_remove (ix: n); |
3848 | if (sched_verbose >= 4) |
3849 | sel_print (fmt: "Delaying speculation check %d until its first use\n" , |
3850 | INSN_UID (insn)); |
3851 | continue; |
3852 | } |
3853 | |
3854 | /* Ignore EXPRs available from pipelining to update AV_MAX_PRIO. */ |
3855 | if (EXPR_ORIG_SCHED_CYCLE (expr) <= 0) |
3856 | av_max_prio = MAX (av_max_prio, EXPR_PRIORITY (expr)); |
3857 | |
3858 | /* Don't allow any insns whose data is not yet ready. |
3859 | Check first whether we've already tried them and failed. */ |
3860 | if (INSN_UID (insn) < FENCE_READY_TICKS_SIZE (fence)) |
3861 | { |
3862 | need_cycles = (FENCE_READY_TICKS (fence)[INSN_UID (insn)] |
3863 | - FENCE_CYCLE (fence)); |
3864 | if (EXPR_ORIG_SCHED_CYCLE (expr) <= 0) |
3865 | est_ticks_till_branch = MAX (est_ticks_till_branch, |
3866 | EXPR_PRIORITY (expr) + need_cycles); |
3867 | |
3868 | if (need_cycles > 0) |
3869 | { |
3870 | stalled++; |
3871 | min_need_stall = (min_need_stall < 0 |
3872 | ? need_cycles |
3873 | : MIN (min_need_stall, need_cycles)); |
3874 | vec_av_set.unordered_remove (ix: n); |
3875 | |
3876 | if (sched_verbose >= 4) |
3877 | sel_print (fmt: "Expr %d is not ready until cycle %d (cached)\n" , |
3878 | INSN_UID (insn), |
3879 | FENCE_READY_TICKS (fence)[INSN_UID (insn)]); |
3880 | continue; |
3881 | } |
3882 | } |
3883 | |
3884 | /* Now resort to dependence analysis to find whether EXPR might be |
3885 | stalled due to dependencies from FENCE's context. */ |
3886 | need_cycles = tick_check_p (expr, dc, fence); |
3887 | new_prio = EXPR_PRIORITY (expr) + EXPR_PRIORITY_ADJ (expr) + need_cycles; |
3888 | |
3889 | if (EXPR_ORIG_SCHED_CYCLE (expr) <= 0) |
3890 | est_ticks_till_branch = MAX (est_ticks_till_branch, |
3891 | new_prio); |
3892 | |
3893 | if (need_cycles > 0) |
3894 | { |
3895 | if (INSN_UID (insn) >= FENCE_READY_TICKS_SIZE (fence)) |
3896 | { |
3897 | int new_size = INSN_UID (insn) * 3 / 2; |
3898 | |
3899 | FENCE_READY_TICKS (fence) |
3900 | = (int *) xrecalloc (FENCE_READY_TICKS (fence), |
3901 | new_size, FENCE_READY_TICKS_SIZE (fence), |
3902 | sizeof (int)); |
3903 | } |
3904 | FENCE_READY_TICKS (fence)[INSN_UID (insn)] |
3905 | = FENCE_CYCLE (fence) + need_cycles; |
3906 | |
3907 | stalled++; |
3908 | min_need_stall = (min_need_stall < 0 |
3909 | ? need_cycles |
3910 | : MIN (min_need_stall, need_cycles)); |
3911 | |
3912 | vec_av_set.unordered_remove (ix: n); |
3913 | |
3914 | if (sched_verbose >= 4) |
3915 | sel_print (fmt: "Expr %d is not ready yet until cycle %d\n" , |
3916 | INSN_UID (insn), |
3917 | FENCE_READY_TICKS (fence)[INSN_UID (insn)]); |
3918 | continue; |
3919 | } |
3920 | |
3921 | if (sched_verbose >= 4) |
3922 | sel_print (fmt: "Expr %d is ok\n" , INSN_UID (insn)); |
3923 | min_need_stall = 0; |
3924 | } |
3925 | |
3926 | /* Clear SCHED_NEXT. */ |
3927 | if (FENCE_SCHED_NEXT (fence)) |
3928 | { |
3929 | gcc_assert (sched_next_worked == 1); |
3930 | FENCE_SCHED_NEXT (fence) = NULL; |
3931 | } |
3932 | |
3933 | /* No need to stall if this variable was not initialized. */ |
3934 | if (min_need_stall < 0) |
3935 | min_need_stall = 0; |
3936 | |
3937 | if (vec_av_set.is_empty ()) |
3938 | { |
3939 | /* We need to set *pneed_stall here, because later we skip this code |
3940 | when ready list is empty. */ |
3941 | *pneed_stall = min_need_stall; |
3942 | return false; |
3943 | } |
3944 | else |
3945 | gcc_assert (min_need_stall == 0); |
3946 | |
3947 | /* Sort the vector. */ |
3948 | vec_av_set.qsort (sel_rank_for_schedule); |
3949 | |
3950 | if (sched_verbose >= 4) |
3951 | { |
3952 | sel_print (fmt: "Total ready exprs: %d, stalled: %d\n" , |
3953 | vec_av_set.length (), stalled); |
3954 | sel_print (fmt: "Sorted av set (%d): " , vec_av_set.length ()); |
3955 | FOR_EACH_VEC_ELT (vec_av_set, n, expr) |
3956 | dump_expr (expr); |
3957 | sel_print (fmt: "\n" ); |
3958 | } |
3959 | |
3960 | *pneed_stall = 0; |
3961 | return true; |
3962 | } |
3963 | |
3964 | /* Convert a vectored and sorted av set to the ready list that |
3965 | the rest of the backend wants to see. */ |
3966 | static void |
3967 | convert_vec_av_set_to_ready (void) |
3968 | { |
3969 | int n; |
3970 | expr_t expr; |
3971 | |
3972 | /* Allocate and fill the ready list from the sorted vector. */ |
3973 | ready.n_ready = vec_av_set.length (); |
3974 | ready.first = ready.n_ready - 1; |
3975 | |
3976 | gcc_assert (ready.n_ready > 0); |
3977 | |
3978 | if (ready.n_ready > max_issue_size) |
3979 | { |
3980 | max_issue_size = ready.n_ready; |
3981 | sched_extend_ready_list (ready.n_ready); |
3982 | } |
3983 | |
3984 | FOR_EACH_VEC_ELT (vec_av_set, n, expr) |
3985 | { |
3986 | vinsn_t vi = EXPR_VINSN (expr); |
3987 | insn_t insn = VINSN_INSN_RTX (vi); |
3988 | |
3989 | ready_try[n] = 0; |
3990 | ready.vec[n] = insn; |
3991 | } |
3992 | } |
3993 | |
3994 | /* Initialize ready list from *AV_PTR for the max_issue () call. |
3995 | If any unrecognizable insn found in *AV_PTR, return it (and skip |
3996 | max_issue). BND and FENCE are current boundary and fence, |
3997 | respectively. If we need to stall for some cycles before an expr |
3998 | from *AV_PTR would become available, write this number to *PNEED_STALL. */ |
3999 | static expr_t |
4000 | fill_ready_list (av_set_t *av_ptr, blist_t bnds, fence_t fence, |
4001 | int *pneed_stall) |
4002 | { |
4003 | expr_t expr; |
4004 | |
4005 | /* We do not support multiple boundaries per fence. */ |
4006 | gcc_assert (BLIST_NEXT (bnds) == NULL); |
4007 | |
4008 | /* Process expressions required special handling, i.e. pipelined, |
4009 | speculative and recog() < 0 expressions first. */ |
4010 | process_pipelined_exprs (av_ptr); |
4011 | process_spec_exprs (av_ptr); |
4012 | |
4013 | /* A USE could be scheduled immediately. */ |
4014 | expr = process_use_exprs (av_ptr); |
4015 | if (expr) |
4016 | { |
4017 | *pneed_stall = 0; |
4018 | return expr; |
4019 | } |
4020 | |
4021 | /* Turn the av set to a vector for sorting. */ |
4022 | if (! fill_vec_av_set (av: *av_ptr, bnds, fence, pneed_stall)) |
4023 | { |
4024 | ready.n_ready = 0; |
4025 | return NULL; |
4026 | } |
4027 | |
4028 | /* Build the final ready list. */ |
4029 | convert_vec_av_set_to_ready (); |
4030 | return NULL; |
4031 | } |
4032 | |
4033 | /* Wrapper for dfa_new_cycle (). Returns TRUE if cycle was advanced. */ |
4034 | static bool |
4035 | sel_dfa_new_cycle (insn_t insn, fence_t fence) |
4036 | { |
4037 | int last_scheduled_cycle = FENCE_LAST_SCHEDULED_INSN (fence) |
4038 | ? INSN_SCHED_CYCLE (FENCE_LAST_SCHEDULED_INSN (fence)) |
4039 | : FENCE_CYCLE (fence) - 1; |
4040 | bool res = false; |
4041 | int sort_p = 0; |
4042 | |
4043 | if (!targetm.sched.dfa_new_cycle) |
4044 | return false; |
4045 | |
4046 | memcpy (dest: curr_state, FENCE_STATE (fence), n: dfa_state_size); |
4047 | |
4048 | while (!sort_p && targetm.sched.dfa_new_cycle (sched_dump, sched_verbose, |
4049 | insn, last_scheduled_cycle, |
4050 | FENCE_CYCLE (fence), &sort_p)) |
4051 | { |
4052 | memcpy (FENCE_STATE (fence), src: curr_state, n: dfa_state_size); |
4053 | advance_one_cycle (fence); |
4054 | memcpy (dest: curr_state, FENCE_STATE (fence), n: dfa_state_size); |
4055 | res = true; |
4056 | } |
4057 | |
4058 | return res; |
4059 | } |
4060 | |
4061 | /* Invoke reorder* target hooks on the ready list. Return the number of insns |
4062 | we can issue. FENCE is the current fence. */ |
4063 | static int |
4064 | invoke_reorder_hooks (fence_t fence) |
4065 | { |
4066 | int issue_more; |
4067 | bool ran_hook = false; |
4068 | |
4069 | /* Call the reorder hook at the beginning of the cycle, and call |
4070 | the reorder2 hook in the middle of the cycle. */ |
4071 | if (FENCE_ISSUED_INSNS (fence) == 0) |
4072 | { |
4073 | if (targetm.sched.reorder |
4074 | && !SCHED_GROUP_P (ready_element (&ready, 0)) |
4075 | && ready.n_ready > 1) |
4076 | { |
4077 | /* Don't give reorder the most prioritized insn as it can break |
4078 | pipelining. */ |
4079 | if (pipelining_p) |
4080 | --ready.n_ready; |
4081 | |
4082 | issue_more |
4083 | = targetm.sched.reorder (sched_dump, sched_verbose, |
4084 | ready_lastpos (&ready), |
4085 | &ready.n_ready, FENCE_CYCLE (fence)); |
4086 | |
4087 | if (pipelining_p) |
4088 | ++ready.n_ready; |
4089 | |
4090 | ran_hook = true; |
4091 | } |
4092 | else |
4093 | /* Initialize can_issue_more for variable_issue. */ |
4094 | issue_more = issue_rate; |
4095 | } |
4096 | else if (targetm.sched.reorder2 |
4097 | && !SCHED_GROUP_P (ready_element (&ready, 0))) |
4098 | { |
4099 | if (ready.n_ready == 1) |
4100 | issue_more = |
4101 | targetm.sched.reorder2 (sched_dump, sched_verbose, |
4102 | ready_lastpos (&ready), |
4103 | &ready.n_ready, FENCE_CYCLE (fence)); |
4104 | else |
4105 | { |
4106 | if (pipelining_p) |
4107 | --ready.n_ready; |
4108 | |
4109 | issue_more = |
4110 | targetm.sched.reorder2 (sched_dump, sched_verbose, |
4111 | ready.n_ready |
4112 | ? ready_lastpos (&ready) : NULL, |
4113 | &ready.n_ready, FENCE_CYCLE (fence)); |
4114 | |
4115 | if (pipelining_p) |
4116 | ++ready.n_ready; |
4117 | } |
4118 | |
4119 | ran_hook = true; |
4120 | } |
4121 | else |
4122 | issue_more = FENCE_ISSUE_MORE (fence); |
4123 | |
4124 | /* Ensure that ready list and vec_av_set are in line with each other, |
4125 | i.e. vec_av_set[i] == ready_element (&ready, i). */ |
4126 | if (issue_more && ran_hook) |
4127 | { |
4128 | int i, j, n; |
4129 | rtx_insn **arr = ready.vec; |
4130 | expr_t *vec = vec_av_set.address (); |
4131 | |
4132 | for (i = 0, n = ready.n_ready; i < n; i++) |
4133 | if (EXPR_INSN_RTX (vec[i]) != arr[i]) |
4134 | { |
4135 | for (j = i; j < n; j++) |
4136 | if (EXPR_INSN_RTX (vec[j]) == arr[i]) |
4137 | break; |
4138 | gcc_assert (j < n); |
4139 | |
4140 | std::swap (a&: vec[i], b&: vec[j]); |
4141 | } |
4142 | } |
4143 | |
4144 | return issue_more; |
4145 | } |
4146 | |
4147 | /* Return an EXPR corresponding to INDEX element of ready list, if |
4148 | FOLLOW_READY_ELEMENT is true (i.e., an expr of |
4149 | ready_element (&ready, INDEX) will be returned), and to INDEX element of |
4150 | ready.vec otherwise. */ |
4151 | static inline expr_t |
4152 | find_expr_for_ready (int index, bool follow_ready_element) |
4153 | { |
4154 | expr_t expr; |
4155 | int real_index; |
4156 | |
4157 | real_index = follow_ready_element ? ready.first - index : index; |
4158 | |
4159 | expr = vec_av_set[real_index]; |
4160 | gcc_assert (ready.vec[real_index] == EXPR_INSN_RTX (expr)); |
4161 | |
4162 | return expr; |
4163 | } |
4164 | |
4165 | /* Calculate insns worth trying via lookahead_guard hook. Return a number |
4166 | of such insns found. */ |
4167 | static int |
4168 | invoke_dfa_lookahead_guard (void) |
4169 | { |
4170 | int i, n; |
4171 | bool have_hook |
4172 | = targetm.sched.first_cycle_multipass_dfa_lookahead_guard != NULL; |
4173 | |
4174 | if (sched_verbose >= 2) |
4175 | sel_print (fmt: "ready after reorder: " ); |
4176 | |
4177 | for (i = 0, n = 0; i < ready.n_ready; i++) |
4178 | { |
4179 | expr_t expr; |
4180 | insn_t insn; |
4181 | int r; |
4182 | |
4183 | /* In this loop insn is Ith element of the ready list given by |
4184 | ready_element, not Ith element of ready.vec. */ |
4185 | insn = ready_element (&ready, i); |
4186 | |
4187 | if (! have_hook || i == 0) |
4188 | r = 0; |
4189 | else |
4190 | r = targetm.sched.first_cycle_multipass_dfa_lookahead_guard (insn, i); |
4191 | |
4192 | gcc_assert (INSN_CODE (insn) >= 0); |
4193 | |
4194 | /* Only insns with ready_try = 0 can get here |
4195 | from fill_ready_list. */ |
4196 | gcc_assert (ready_try [i] == 0); |
4197 | ready_try[i] = r; |
4198 | if (!r) |
4199 | n++; |
4200 | |
4201 | expr = find_expr_for_ready (index: i, follow_ready_element: true); |
4202 | |
4203 | if (sched_verbose >= 2) |
4204 | { |
4205 | dump_vinsn (EXPR_VINSN (expr)); |
4206 | sel_print (fmt: ":%d; " , ready_try[i]); |
4207 | } |
4208 | } |
4209 | |
4210 | if (sched_verbose >= 2) |
4211 | sel_print (fmt: "\n" ); |
4212 | return n; |
4213 | } |
4214 | |
4215 | /* Calculate the number of privileged insns and return it. */ |
4216 | static int |
4217 | calculate_privileged_insns (void) |
4218 | { |
4219 | expr_t cur_expr, min_spec_expr = NULL; |
4220 | int privileged_n = 0, i; |
4221 | |
4222 | for (i = 0; i < ready.n_ready; i++) |
4223 | { |
4224 | if (ready_try[i]) |
4225 | continue; |
4226 | |
4227 | if (! min_spec_expr) |
4228 | min_spec_expr = find_expr_for_ready (index: i, follow_ready_element: true); |
4229 | |
4230 | cur_expr = find_expr_for_ready (index: i, follow_ready_element: true); |
4231 | |
4232 | if (EXPR_SPEC (cur_expr) > EXPR_SPEC (min_spec_expr)) |
4233 | break; |
4234 | |
4235 | ++privileged_n; |
4236 | } |
4237 | |
4238 | if (i == ready.n_ready) |
4239 | privileged_n = 0; |
4240 | |
4241 | if (sched_verbose >= 2) |
4242 | sel_print (fmt: "privileged_n: %d insns with SPEC %d\n" , |
4243 | privileged_n, privileged_n ? EXPR_SPEC (min_spec_expr) : -1); |
4244 | return privileged_n; |
4245 | } |
4246 | |
4247 | /* Call the rest of the hooks after the choice was made. Return |
4248 | the number of insns that still can be issued given that the current |
4249 | number is ISSUE_MORE. FENCE and BEST_INSN are the current fence |
4250 | and the insn chosen for scheduling, respectively. */ |
4251 | static int |
4252 | invoke_aftermath_hooks (fence_t fence, rtx_insn *best_insn, int issue_more) |
4253 | { |
4254 | gcc_assert (INSN_P (best_insn)); |
4255 | |
4256 | /* First, call dfa_new_cycle, and then variable_issue, if available. */ |
4257 | sel_dfa_new_cycle (insn: best_insn, fence); |
4258 | |
4259 | if (targetm.sched.variable_issue) |
4260 | { |
4261 | memcpy (dest: curr_state, FENCE_STATE (fence), n: dfa_state_size); |
4262 | issue_more = |
4263 | targetm.sched.variable_issue (sched_dump, sched_verbose, best_insn, |
4264 | issue_more); |
4265 | memcpy (FENCE_STATE (fence), src: curr_state, n: dfa_state_size); |
4266 | } |
4267 | else if (!DEBUG_INSN_P (best_insn) |
4268 | && GET_CODE (PATTERN (best_insn)) != USE |
4269 | && GET_CODE (PATTERN (best_insn)) != CLOBBER) |
4270 | issue_more--; |
4271 | |
4272 | return issue_more; |
4273 | } |
4274 | |
4275 | /* Estimate the cost of issuing INSN on DFA state STATE. */ |
4276 | static int |
4277 | estimate_insn_cost (rtx_insn *insn, state_t state) |
4278 | { |
4279 | static state_t temp = NULL; |
4280 | int cost; |
4281 | |
4282 | if (!temp) |
4283 | temp = xmalloc (dfa_state_size); |
4284 | |
4285 | memcpy (dest: temp, src: state, n: dfa_state_size); |
4286 | cost = state_transition (temp, insn); |
4287 | |
4288 | if (cost < 0) |
4289 | return 0; |
4290 | else if (cost == 0) |
4291 | return 1; |
4292 | return cost; |
4293 | } |
4294 | |
4295 | /* Return the cost of issuing EXPR on the FENCE as estimated by DFA. |
4296 | This function properly handles ASMs, USEs etc. */ |
4297 | static int |
4298 | get_expr_cost (expr_t expr, fence_t fence) |
4299 | { |
4300 | rtx_insn *insn = EXPR_INSN_RTX (expr); |
4301 | |
4302 | if (recog_memoized (insn) < 0) |
4303 | { |
4304 | if (!FENCE_STARTS_CYCLE_P (fence) |
4305 | && INSN_ASM_P (insn)) |
4306 | /* This is asm insn which is tryed to be issued on the |
4307 | cycle not first. Issue it on the next cycle. */ |
4308 | return 1; |
4309 | else |
4310 | /* A USE insn, or something else we don't need to |
4311 | understand. We can't pass these directly to |
4312 | state_transition because it will trigger a |
4313 | fatal error for unrecognizable insns. */ |
4314 | return 0; |
4315 | } |
4316 | else |
4317 | return estimate_insn_cost (insn, FENCE_STATE (fence)); |
4318 | } |
4319 | |
4320 | /* Find the best insn for scheduling, either via max_issue or just take |
4321 | the most prioritized available. */ |
4322 | static int |
4323 | choose_best_insn (fence_t fence, int privileged_n, int *index) |
4324 | { |
4325 | int can_issue = 0; |
4326 | |
4327 | if (dfa_lookahead > 0) |
4328 | { |
4329 | cycle_issued_insns = FENCE_ISSUED_INSNS (fence); |
4330 | /* TODO: pass equivalent of first_cycle_insn_p to max_issue (). */ |
4331 | can_issue = max_issue (&ready, privileged_n, |
4332 | FENCE_STATE (fence), true, index); |
4333 | if (sched_verbose >= 2) |
4334 | sel_print (fmt: "max_issue: we can issue %d insns, already did %d insns\n" , |
4335 | can_issue, FENCE_ISSUED_INSNS (fence)); |
4336 | } |
4337 | else |
4338 | { |
4339 | /* We can't use max_issue; just return the first available element. */ |
4340 | int i; |
4341 | |
4342 | for (i = 0; i < ready.n_ready; i++) |
4343 | { |
4344 | expr_t expr = find_expr_for_ready (index: i, follow_ready_element: true); |
4345 | |
4346 | if (get_expr_cost (expr, fence) < 1) |
4347 | { |
4348 | can_issue = can_issue_more; |
4349 | *index = i; |
4350 | |
4351 | if (sched_verbose >= 2) |
4352 | sel_print (fmt: "using %dth insn from the ready list\n" , i + 1); |
4353 | |
4354 | break; |
4355 | } |
4356 | } |
4357 | |
4358 | if (i == ready.n_ready) |
4359 | { |
4360 | can_issue = 0; |
4361 | *index = -1; |
4362 | } |
4363 | } |
4364 | |
4365 | return can_issue; |
4366 | } |
4367 | |
4368 | /* Choose the best expr from *AV_VLIW_PTR and a suitable register for it. |
4369 | BNDS and FENCE are current boundaries and scheduling fence respectively. |
4370 | Return the expr found and NULL if nothing can be issued atm. |
4371 | Write to PNEED_STALL the number of cycles to stall if no expr was found. */ |
4372 | static expr_t |
4373 | find_best_expr (av_set_t *av_vliw_ptr, blist_t bnds, fence_t fence, |
4374 | int *pneed_stall) |
4375 | { |
4376 | expr_t best; |
4377 | |
4378 | /* Choose the best insn for scheduling via: |
4379 | 1) sorting the ready list based on priority; |
4380 | 2) calling the reorder hook; |
4381 | 3) calling max_issue. */ |
4382 | best = fill_ready_list (av_ptr: av_vliw_ptr, bnds, fence, pneed_stall); |
4383 | if (best == NULL && ready.n_ready > 0) |
4384 | { |
4385 | int privileged_n, index; |
4386 | |
4387 | can_issue_more = invoke_reorder_hooks (fence); |
4388 | if (can_issue_more > 0) |
4389 | { |
4390 | /* Try choosing the best insn until we find one that is could be |
4391 | scheduled due to liveness restrictions on its destination register. |
4392 | In the future, we'd like to choose once and then just probe insns |
4393 | in the order of their priority. */ |
4394 | invoke_dfa_lookahead_guard (); |
4395 | privileged_n = calculate_privileged_insns (); |
4396 | can_issue_more = choose_best_insn (fence, privileged_n, index: &index); |
4397 | if (can_issue_more) |
4398 | best = find_expr_for_ready (index, follow_ready_element: true); |
4399 | } |
4400 | /* We had some available insns, so if we can't issue them, |
4401 | we have a stall. */ |
4402 | if (can_issue_more == 0) |
4403 | { |
4404 | best = NULL; |
4405 | *pneed_stall = 1; |
4406 | } |
4407 | } |
4408 | |
4409 | if (best != NULL) |
4410 | { |
4411 | can_issue_more = invoke_aftermath_hooks (fence, EXPR_INSN_RTX (best), |
4412 | issue_more: can_issue_more); |
4413 | if (targetm.sched.variable_issue |
4414 | && can_issue_more == 0) |
4415 | *pneed_stall = 1; |
4416 | } |
4417 | |
4418 | if (sched_verbose >= 2) |
4419 | { |
4420 | if (best != NULL) |
4421 | { |
4422 | sel_print (fmt: "Best expression (vliw form): " ); |
4423 | dump_expr (best); |
4424 | sel_print (fmt: "; cycle %d\n" , FENCE_CYCLE (fence)); |
4425 | } |
4426 | else |
4427 | sel_print (fmt: "No best expr found!\n" ); |
4428 | } |
4429 | |
4430 | return best; |
4431 | } |
4432 | |
4433 | |
4434 | /* Functions that implement the core of the scheduler. */ |
4435 | |
4436 | |
4437 | /* Emit an instruction from EXPR with SEQNO and VINSN after |
4438 | PLACE_TO_INSERT. */ |
4439 | static insn_t |
4440 | emit_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno, |
4441 | insn_t place_to_insert) |
4442 | { |
4443 | /* This assert fails when we have identical instructions |
4444 | one of which dominates the other. In this case move_op () |
4445 | finds the first instruction and doesn't search for second one. |
4446 | The solution would be to compute av_set after the first found |
4447 | insn and, if insn present in that set, continue searching. |
4448 | For now we workaround this issue in move_op. */ |
4449 | gcc_assert (!INSN_IN_STREAM_P (EXPR_INSN_RTX (expr))); |
4450 | |
4451 | if (EXPR_WAS_RENAMED (expr)) |
4452 | { |
4453 | unsigned regno = expr_dest_regno (expr); |
4454 | |
4455 | if (HARD_REGISTER_NUM_P (regno)) |
4456 | { |
4457 | df_set_regs_ever_live (regno, true); |
4458 | reg_rename_tick[regno] = ++reg_rename_this_tick; |
4459 | } |
4460 | } |
4461 | |
4462 | return sel_gen_insn_from_expr_after (expr, vinsn, seqno, |
4463 | place_to_insert); |
4464 | } |
4465 | |
4466 | /* Return TRUE if BB can hold bookkeeping code. */ |
4467 | static bool |
4468 | block_valid_for_bookkeeping_p (basic_block bb) |
4469 | { |
4470 | insn_t bb_end = BB_END (bb); |
4471 | |
4472 | if (!in_current_region_p (bb) || EDGE_COUNT (bb->succs) > 1) |
4473 | return false; |
4474 | |
4475 | if (INSN_P (bb_end)) |
4476 | { |
4477 | if (INSN_SCHED_TIMES (bb_end) > 0) |
4478 | return false; |
4479 | } |
4480 | else |
4481 | gcc_assert (NOTE_INSN_BASIC_BLOCK_P (bb_end)); |
4482 | |
4483 | return true; |
4484 | } |
4485 | |
4486 | /* Attempt to find a block that can hold bookkeeping code for path(s) incoming |
4487 | into E2->dest, except from E1->src (there may be a sequence of empty basic |
4488 | blocks between E1->src and E2->dest). Return found block, or NULL if new |
4489 | one must be created. If LAX holds, don't assume there is a simple path |
4490 | from E1->src to E2->dest. */ |
4491 | static basic_block |
4492 | find_block_for_bookkeeping (edge e1, edge e2, bool lax) |
4493 | { |
4494 | basic_block candidate_block = NULL; |
4495 | edge e; |
4496 | |
4497 | /* Loop over edges from E1 to E2, inclusive. */ |
4498 | for (e = e1; !lax || e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun); e = |
4499 | EDGE_SUCC (e->dest, 0)) |
4500 | { |
4501 | if (EDGE_COUNT (e->dest->preds) == 2) |
4502 | { |
4503 | if (candidate_block == NULL) |
4504 | candidate_block = (EDGE_PRED (e->dest, 0) == e |
4505 | ? EDGE_PRED (e->dest, 1)->src |
4506 | : EDGE_PRED (e->dest, 0)->src); |
4507 | else |
4508 | /* Found additional edge leading to path from e1 to e2 |
4509 | from aside. */ |
4510 | return NULL; |
4511 | } |
4512 | else if (EDGE_COUNT (e->dest->preds) > 2) |
4513 | /* Several edges leading to path from e1 to e2 from aside. */ |
4514 | return NULL; |
4515 | |
4516 | if (e == e2) |
4517 | return ((!lax || candidate_block) |
4518 | && block_valid_for_bookkeeping_p (bb: candidate_block) |
4519 | ? candidate_block |
4520 | : NULL); |
4521 | |
4522 | if (lax && EDGE_COUNT (e->dest->succs) != 1) |
4523 | return NULL; |
4524 | } |
4525 | |
4526 | if (lax) |
4527 | return NULL; |
4528 | |
4529 | gcc_unreachable (); |
4530 | } |
4531 | |
4532 | /* Create new basic block for bookkeeping code for path(s) incoming into |
4533 | E2->dest, except from E1->src. Return created block. */ |
4534 | static basic_block |
4535 | create_block_for_bookkeeping (edge e1, edge e2) |
4536 | { |
4537 | basic_block new_bb, bb = e2->dest; |
4538 | |
4539 | /* Check that we don't spoil the loop structure. */ |
4540 | if (current_loop_nest) |
4541 | { |
4542 | basic_block latch = current_loop_nest->latch; |
4543 | |
4544 | /* We do not split header. */ |
4545 | gcc_assert (e2->dest != current_loop_nest->header); |
4546 | |
4547 | /* We do not redirect the only edge to the latch block. */ |
4548 | gcc_assert (e1->dest != latch |
4549 | || !single_pred_p (latch) |
4550 | || e1 != single_pred_edge (latch)); |
4551 | } |
4552 | |
4553 | /* Split BB to insert BOOK_INSN there. */ |
4554 | new_bb = sched_split_block (bb, NULL); |
4555 | |
4556 | /* Move note_list from the upper bb. */ |
4557 | gcc_assert (BB_NOTE_LIST (new_bb) == NULL_RTX); |
4558 | BB_NOTE_LIST (new_bb) = BB_NOTE_LIST (bb); |
4559 | BB_NOTE_LIST (bb) = NULL; |
4560 | |
4561 | gcc_assert (e2->dest == bb); |
4562 | |
4563 | /* Skip block for bookkeeping copy when leaving E1->src. */ |
4564 | if (e1->flags & EDGE_FALLTHRU) |
4565 | sel_redirect_edge_and_branch_force (e1, new_bb); |
4566 | else |
4567 | sel_redirect_edge_and_branch (e1, new_bb); |
4568 | |
4569 | gcc_assert (e1->dest == new_bb); |
4570 | gcc_assert (sel_bb_empty_p (bb)); |
4571 | |
4572 | /* To keep basic block numbers in sync between debug and non-debug |
4573 | compilations, we have to rotate blocks here. Consider that we |
4574 | started from (a,b)->d, (c,d)->e, and d contained only debug |
4575 | insns. It would have been removed before if the debug insns |
4576 | weren't there, so we'd have split e rather than d. So what we do |
4577 | now is to swap the block numbers of new_bb and |
4578 | single_succ(new_bb) == e, so that the insns that were in e before |
4579 | get the new block number. */ |
4580 | |
4581 | if (MAY_HAVE_DEBUG_INSNS) |
4582 | { |
4583 | basic_block succ; |
4584 | insn_t insn = sel_bb_head (new_bb); |
4585 | insn_t last; |
4586 | |
4587 | if (DEBUG_INSN_P (insn) |
4588 | && single_succ_p (bb: new_bb) |
4589 | && (succ = single_succ (bb: new_bb)) |
4590 | && succ != EXIT_BLOCK_PTR_FOR_FN (cfun) |
4591 | && DEBUG_INSN_P ((last = sel_bb_end (new_bb)))) |
4592 | { |
4593 | while (insn != last && (DEBUG_INSN_P (insn) || NOTE_P (insn))) |
4594 | insn = NEXT_INSN (insn); |
4595 | |
4596 | if (insn == last) |
4597 | { |
4598 | sel_global_bb_info_def gbi; |
4599 | sel_region_bb_info_def rbi; |
4600 | |
4601 | if (sched_verbose >= 2) |
4602 | sel_print (fmt: "Swapping block ids %i and %i\n" , |
4603 | new_bb->index, succ->index); |
4604 | |
4605 | std::swap (a&: new_bb->index, b&: succ->index); |
4606 | |
4607 | SET_BASIC_BLOCK_FOR_FN (cfun, new_bb->index, new_bb); |
4608 | SET_BASIC_BLOCK_FOR_FN (cfun, succ->index, succ); |
4609 | |
4610 | memcpy (dest: &gbi, SEL_GLOBAL_BB_INFO (new_bb), n: sizeof (gbi)); |
4611 | memcpy (SEL_GLOBAL_BB_INFO (new_bb), SEL_GLOBAL_BB_INFO (succ), |
4612 | n: sizeof (gbi)); |
4613 | memcpy (SEL_GLOBAL_BB_INFO (succ), src: &gbi, n: sizeof (gbi)); |
4614 | |
4615 | memcpy (dest: &rbi, SEL_REGION_BB_INFO (new_bb), n: sizeof (rbi)); |
4616 | memcpy (SEL_REGION_BB_INFO (new_bb), SEL_REGION_BB_INFO (succ), |
4617 | n: sizeof (rbi)); |
4618 | memcpy (SEL_REGION_BB_INFO (succ), src: &rbi, n: sizeof (rbi)); |
4619 | |
4620 | std::swap (BLOCK_TO_BB (new_bb->index), |
4621 | BLOCK_TO_BB (succ->index)); |
4622 | |
4623 | std::swap (CONTAINING_RGN (new_bb->index), |
4624 | CONTAINING_RGN (succ->index)); |
4625 | |
4626 | for (int i = 0; i < current_nr_blocks; i++) |
4627 | if (BB_TO_BLOCK (i) == succ->index) |
4628 | BB_TO_BLOCK (i) = new_bb->index; |
4629 | else if (BB_TO_BLOCK (i) == new_bb->index) |
4630 | BB_TO_BLOCK (i) = succ->index; |
4631 | |
4632 | FOR_BB_INSNS (new_bb, insn) |
4633 | if (INSN_P (insn)) |
4634 | EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = new_bb->index; |
4635 | |
4636 | FOR_BB_INSNS (succ, insn) |
4637 | if (INSN_P (insn)) |
4638 | EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = succ->index; |
4639 | |
4640 | if (bitmap_clear_bit (code_motion_visited_blocks, new_bb->index)) |
4641 | bitmap_set_bit (code_motion_visited_blocks, succ->index); |
4642 | |
4643 | gcc_assert (LABEL_P (BB_HEAD (new_bb)) |
4644 | && LABEL_P (BB_HEAD (succ))); |
4645 | |
4646 | if (sched_verbose >= 4) |
4647 | sel_print (fmt: "Swapping code labels %i and %i\n" , |
4648 | CODE_LABEL_NUMBER (BB_HEAD (new_bb)), |
4649 | CODE_LABEL_NUMBER (BB_HEAD (succ))); |
4650 | |
4651 | std::swap (CODE_LABEL_NUMBER (BB_HEAD (new_bb)), |
4652 | CODE_LABEL_NUMBER (BB_HEAD (succ))); |
4653 | } |
4654 | } |
4655 | } |
4656 | |
4657 | return bb; |
4658 | } |
4659 | |
4660 | /* Return insn after which we must insert bookkeeping code for path(s) incoming |
4661 | into E2->dest, except from E1->src. If the returned insn immediately |
4662 | precedes a fence, assign that fence to *FENCE_TO_REWIND. */ |
4663 | static insn_t |
4664 | find_place_for_bookkeeping (edge e1, edge e2, fence_t *fence_to_rewind) |
4665 | { |
4666 | insn_t place_to_insert; |
4667 | /* Find a basic block that can hold bookkeeping. If it can be found, do not |
4668 | create new basic block, but insert bookkeeping there. */ |
4669 | basic_block book_block = find_block_for_bookkeeping (e1, e2, lax: false); |
4670 | |
4671 | if (book_block) |
4672 | { |
4673 | place_to_insert = BB_END (book_block); |
4674 | |
4675 | /* Don't use a block containing only debug insns for |
4676 | bookkeeping, this causes scheduling differences between debug |
4677 | and non-debug compilations, for the block would have been |
4678 | removed already. */ |
4679 | if (DEBUG_INSN_P (place_to_insert)) |
4680 | { |
4681 | rtx_insn *insn = sel_bb_head (book_block); |
4682 | |
4683 | while (insn != place_to_insert && |
4684 | (DEBUG_INSN_P (insn) || NOTE_P (insn))) |
4685 | insn = NEXT_INSN (insn); |
4686 | |
4687 | if (insn == place_to_insert) |
4688 | book_block = NULL; |
4689 | } |
4690 | } |
4691 | |
4692 | if (!book_block) |
4693 | { |
4694 | book_block = create_block_for_bookkeeping (e1, e2); |
4695 | place_to_insert = BB_END (book_block); |
4696 | if (sched_verbose >= 9) |
4697 | sel_print (fmt: "New block is %i, split from bookkeeping block %i\n" , |
4698 | EDGE_SUCC (book_block, 0)->dest->index, book_block->index); |
4699 | } |
4700 | else |
4701 | { |
4702 | if (sched_verbose >= 9) |
4703 | sel_print (fmt: "Pre-existing bookkeeping block is %i\n" , book_block->index); |
4704 | } |
4705 | |
4706 | *fence_to_rewind = NULL; |
4707 | /* If basic block ends with a jump, insert bookkeeping code right before it. |
4708 | Notice if we are crossing a fence when taking PREV_INSN. */ |
4709 | if (INSN_P (place_to_insert) && control_flow_insn_p (place_to_insert)) |
4710 | { |
4711 | *fence_to_rewind = flist_lookup (fences, place_to_insert); |
4712 | place_to_insert = PREV_INSN (insn: place_to_insert); |
4713 | } |
4714 | |
4715 | return place_to_insert; |
4716 | } |
4717 | |
4718 | /* Find a proper seqno for bookkeeing insn inserted at PLACE_TO_INSERT |
4719 | for JOIN_POINT. */ |
4720 | static int |
4721 | find_seqno_for_bookkeeping (insn_t place_to_insert, insn_t join_point) |
4722 | { |
4723 | int seqno; |
4724 | |
4725 | /* Check if we are about to insert bookkeeping copy before a jump, and use |
4726 | jump's seqno for the copy; otherwise, use JOIN_POINT's seqno. */ |
4727 | rtx_insn *next = NEXT_INSN (insn: place_to_insert); |
4728 | if (INSN_P (next) |
4729 | && JUMP_P (next) |
4730 | && BLOCK_FOR_INSN (insn: next) == BLOCK_FOR_INSN (insn: place_to_insert)) |
4731 | { |
4732 | gcc_assert (INSN_SCHED_TIMES (next) == 0); |
4733 | seqno = INSN_SEQNO (next); |
4734 | } |
4735 | else if (INSN_SEQNO (join_point) > 0) |
4736 | seqno = INSN_SEQNO (join_point); |
4737 | else |
4738 | { |
4739 | seqno = get_seqno_by_preds (place_to_insert); |
4740 | |
4741 | /* Sometimes the fences can move in such a way that there will be |
4742 | no instructions with positive seqno around this bookkeeping. |
4743 | This means that there will be no way to get to it by a regular |
4744 | fence movement. Never mind because we pick up such pieces for |
4745 | rescheduling anyways, so any positive value will do for now. */ |
4746 | if (seqno < 0) |
4747 | { |
4748 | gcc_assert (pipelining_p); |
4749 | seqno = 1; |
4750 | } |
4751 | } |
4752 | |
4753 | gcc_assert (seqno > 0); |
4754 | return seqno; |
4755 | } |
4756 | |
4757 | /* Insert bookkeeping copy of C_EXPS's insn after PLACE_TO_INSERT, assigning |
4758 | NEW_SEQNO to it. Return created insn. */ |
4759 | static insn_t |
4760 | emit_bookkeeping_insn (insn_t place_to_insert, expr_t c_expr, int new_seqno) |
4761 | { |
4762 | rtx_insn *new_insn_rtx = create_copy_of_insn_rtx (EXPR_INSN_RTX (c_expr)); |
4763 | |
4764 | vinsn_t new_vinsn |
4765 | = create_vinsn_from_insn_rtx (new_insn_rtx, |
4766 | VINSN_UNIQUE_P (EXPR_VINSN (c_expr))); |
4767 | |
4768 | insn_t new_insn = emit_insn_from_expr_after (expr: c_expr, vinsn: new_vinsn, seqno: new_seqno, |
4769 | place_to_insert); |
4770 | |
4771 | INSN_SCHED_TIMES (new_insn) = 0; |
4772 | bitmap_set_bit (current_copies, INSN_UID (insn: new_insn)); |
4773 | |
4774 | return new_insn; |
4775 | } |
4776 | |
4777 | /* Generate a bookkeeping copy of C_EXPR's insn for path(s) incoming into to |
4778 | E2->dest, except from E1->src (there may be a sequence of empty blocks |
4779 | between E1->src and E2->dest). Return block containing the copy. |
4780 | All scheduler data is initialized for the newly created insn. */ |
4781 | static basic_block |
4782 | generate_bookkeeping_insn (expr_t c_expr, edge e1, edge e2) |
4783 | { |
4784 | insn_t join_point, place_to_insert, new_insn; |
4785 | int new_seqno; |
4786 | bool need_to_exchange_data_sets; |
4787 | fence_t fence_to_rewind; |
4788 | |
4789 | if (sched_verbose >= 4) |
4790 | sel_print (fmt: "Generating bookkeeping insn (%d->%d)\n" , e1->src->index, |
4791 | e2->dest->index); |
4792 | |
4793 | join_point = sel_bb_head (e2->dest); |
4794 | place_to_insert = find_place_for_bookkeeping (e1, e2, fence_to_rewind: &fence_to_rewind); |
4795 | new_seqno = find_seqno_for_bookkeeping (place_to_insert, join_point); |
4796 | need_to_exchange_data_sets |
4797 | = sel_bb_empty_p (BLOCK_FOR_INSN (insn: place_to_insert)); |
4798 | |
4799 | new_insn = emit_bookkeeping_insn (place_to_insert, c_expr, new_seqno); |
4800 | |
4801 | if (fence_to_rewind) |
4802 | FENCE_INSN (fence_to_rewind) = new_insn; |
4803 | |
4804 | /* When inserting bookkeeping insn in new block, av sets should be |
4805 | following: old basic block (that now holds bookkeeping) data sets are |
4806 | the same as was before generation of bookkeeping, and new basic block |
4807 | (that now hold all other insns of old basic block) data sets are |
4808 | invalid. So exchange data sets for these basic blocks as sel_split_block |
4809 | mistakenly exchanges them in this case. Cannot do it earlier because |
4810 | when single instruction is added to new basic block it should hold NULL |
4811 | lv_set. */ |
4812 | if (need_to_exchange_data_sets) |
4813 | exchange_data_sets (BLOCK_FOR_INSN (insn: new_insn), |
4814 | BLOCK_FOR_INSN (insn: join_point)); |
4815 | |
4816 | stat_bookkeeping_copies++; |
4817 | return BLOCK_FOR_INSN (insn: new_insn); |
4818 | } |
4819 | |
4820 | /* Remove from AV_PTR all insns that may need bookkeeping when scheduling |
4821 | on FENCE, but we are unable to copy them. */ |
4822 | static void |
4823 | remove_insns_that_need_bookkeeping (fence_t fence, av_set_t *av_ptr) |
4824 | { |
4825 | expr_t expr; |
4826 | av_set_iterator i; |
4827 | |
4828 | /* An expression does not need bookkeeping if it is available on all paths |
4829 | from current block to original block and current block dominates |
4830 | original block. We check availability on all paths by examining |
4831 | EXPR_SPEC; this is not equivalent, because it may be positive even |
4832 | if expr is available on all paths (but if expr is not available on |
4833 | any path, EXPR_SPEC will be positive). */ |
4834 | |
4835 | FOR_EACH_EXPR_1 (expr, i, av_ptr) |
4836 | { |
4837 | if (!control_flow_insn_p (EXPR_INSN_RTX (expr)) |
4838 | && (!bookkeeping_p || VINSN_UNIQUE_P (EXPR_VINSN (expr))) |
4839 | && (EXPR_SPEC (expr) |
4840 | || !EXPR_ORIG_BB_INDEX (expr) |
4841 | || !dominated_by_p (CDI_DOMINATORS, |
4842 | BASIC_BLOCK_FOR_FN (cfun, |
4843 | EXPR_ORIG_BB_INDEX (expr)), |
4844 | BLOCK_FOR_INSN (FENCE_INSN (fence))))) |
4845 | { |
4846 | if (sched_verbose >= 4) |
4847 | sel_print (fmt: "Expr %d removed because it would need bookkeeping, which " |
4848 | "cannot be created\n" , INSN_UID (EXPR_INSN_RTX (expr))); |
4849 | av_set_iter_remove (&i); |
4850 | } |
4851 | } |
4852 | } |
4853 | |
4854 | /* Moving conditional jump through some instructions. |
4855 | |
4856 | Consider example: |
4857 | |
4858 | ... <- current scheduling point |
4859 | NOTE BASIC BLOCK: <- bb header |
4860 | (p8) add r14=r14+0x9;; |
4861 | (p8) mov [r14]=r23 |
4862 | (!p8) jump L1;; |
4863 | NOTE BASIC BLOCK: |
4864 | ... |
4865 | |
4866 | We can schedule jump one cycle earlier, than mov, because they cannot be |
4867 | executed together as their predicates are mutually exclusive. |
4868 | |
4869 | This is done in this way: first, new fallthrough basic block is created |
4870 | after jump (it is always can be done, because there already should be a |
4871 | fallthrough block, where control flow goes in case of predicate being true - |
4872 | in our example; otherwise there should be a dependence between those |
4873 | instructions and jump and we cannot schedule jump right now); |
4874 | next, all instructions between jump and current scheduling point are moved |
4875 | to this new block. And the result is this: |
4876 | |
4877 | NOTE BASIC BLOCK: |
4878 | (!p8) jump L1 <- current scheduling point |
4879 | NOTE BASIC BLOCK: <- bb header |
4880 | (p8) add r14=r14+0x9;; |
4881 | (p8) mov [r14]=r23 |
4882 | NOTE BASIC BLOCK: |
4883 | ... |
4884 | */ |
4885 | static void |
4886 | move_cond_jump (rtx_insn *insn, bnd_t bnd) |
4887 | { |
4888 | edge ft_edge; |
4889 | basic_block block_from, block_next, block_new, block_bnd, bb; |
4890 | rtx_insn *next, *prev, *link, *head; |
4891 | |
4892 | block_from = BLOCK_FOR_INSN (insn); |
4893 | block_bnd = BLOCK_FOR_INSN (BND_TO (bnd)); |
4894 | prev = BND_TO (bnd); |
4895 | |
4896 | /* Moving of jump should not cross any other jumps or beginnings of new |
4897 | basic blocks. The only exception is when we move a jump through |
4898 | mutually exclusive insns along fallthru edges. */ |
4899 | if (flag_checking && block_from != block_bnd) |
4900 | { |
4901 | bb = block_from; |
4902 | for (link = PREV_INSN (insn); link != PREV_INSN (insn: prev); |
4903 | link = PREV_INSN (insn: link)) |
4904 | { |
4905 | if (INSN_P (link)) |
4906 | gcc_assert (sched_insns_conditions_mutex_p (insn, link)); |
4907 | if (BLOCK_FOR_INSN (insn: link) && BLOCK_FOR_INSN (insn: link) != bb) |
4908 | { |
4909 | gcc_assert (single_pred (bb) == BLOCK_FOR_INSN (link)); |
4910 | bb = BLOCK_FOR_INSN (insn: link); |
4911 | } |
4912 | } |
4913 | } |
4914 | |
4915 | /* Jump is moved to the boundary. */ |
4916 | next = PREV_INSN (insn); |
4917 | BND_TO (bnd) = insn; |
4918 | |
4919 | ft_edge = find_fallthru_edge_from (block_from); |
4920 | block_next = ft_edge->dest; |
4921 | /* There must be a fallthrough block (or where should go |
4922 | control flow in case of false jump predicate otherwise?). */ |
4923 | gcc_assert (block_next); |
4924 | |
4925 | /* Create new empty basic block after source block. */ |
4926 | block_new = sel_split_edge (ft_edge); |
4927 | gcc_assert (block_new->next_bb == block_next |
4928 | && block_from->next_bb == block_new); |
4929 | |
4930 | /* Move all instructions except INSN to BLOCK_NEW. */ |
4931 | bb = block_bnd; |
4932 | head = BB_HEAD (block_new); |
4933 | while (bb != block_from->next_bb) |
4934 | { |
4935 | rtx_insn *from, *to; |
4936 | from = bb == block_bnd ? prev : sel_bb_head (bb); |
4937 | to = bb == block_from ? next : sel_bb_end (bb); |
4938 | |
4939 | /* The jump being moved can be the first insn in the block. |
4940 | In this case we don't have to move anything in this block. */ |
4941 | if (NEXT_INSN (insn: to) != from) |
4942 | { |
4943 | reorder_insns (from, to, head); |
4944 | |
4945 | for (link = to; link != head; link = PREV_INSN (insn: link)) |
4946 | EXPR_ORIG_BB_INDEX (INSN_EXPR (link)) = block_new->index; |
4947 | head = to; |
4948 | } |
4949 | |
4950 | /* Cleanup possibly empty blocks left. */ |
4951 | block_next = bb->next_bb; |
4952 | if (bb != block_from) |
4953 | tidy_control_flow (bb, false); |
4954 | bb = block_next; |
4955 | } |
4956 | |
4957 | /* Assert there is no jump to BLOCK_NEW, only fallthrough edge. */ |
4958 | gcc_assert (NOTE_INSN_BASIC_BLOCK_P (BB_HEAD (block_new))); |
4959 | |
4960 | gcc_assert (!sel_bb_empty_p (block_from) |
4961 | && !sel_bb_empty_p (block_new)); |
4962 | |
4963 | /* Update data sets for BLOCK_NEW to represent that INSN and |
4964 | instructions from the other branch of INSN is no longer |
4965 | available at BLOCK_NEW. */ |
4966 | BB_AV_LEVEL (block_new) = global_level; |
4967 | gcc_assert (BB_LV_SET (block_new) == NULL); |
4968 | BB_LV_SET (block_new) = get_clear_regset_from_pool (); |
4969 | update_data_sets (insn: sel_bb_head (block_new)); |
4970 | |
4971 | /* INSN is a new basic block header - so prepare its data |
4972 | structures and update availability and liveness sets. */ |
4973 | update_data_sets (insn); |
4974 | |
4975 | if (sched_verbose >= 4) |
4976 | sel_print (fmt: "Moving jump %d\n" , INSN_UID (insn)); |
4977 | } |
4978 | |
4979 | /* Remove nops generated during move_op for preventing removal of empty |
4980 | basic blocks. */ |
4981 | static void |
4982 | remove_temp_moveop_nops (bool full_tidying) |
4983 | { |
4984 | int i; |
4985 | insn_t insn; |
4986 | |
4987 | FOR_EACH_VEC_ELT (vec_temp_moveop_nops, i, insn) |
4988 | { |
4989 | gcc_assert (INSN_NOP_P (insn)); |
4990 | return_nop_to_pool (insn, full_tidying); |
4991 | } |
4992 | |
4993 | /* Empty the vector. */ |
4994 | if (vec_temp_moveop_nops.length () > 0) |
4995 | vec_temp_moveop_nops.block_remove (ix: 0, len: vec_temp_moveop_nops.length ()); |
4996 | } |
4997 | |
4998 | /* Records the maximal UID before moving up an instruction. Used for |
4999 | distinguishing between bookkeeping copies and original insns. */ |
5000 | static int max_uid_before_move_op = 0; |
5001 | |
5002 | /* When true, we're always scheduling next insn on the already scheduled code |
5003 | to get the right insn data for the following bundling or other passes. */ |
5004 | static int force_next_insn = 0; |
5005 | |
5006 | /* Remove from AV_VLIW_P all instructions but next when debug counter |
5007 | tells us so. Next instruction is fetched from BNDS. */ |
5008 | static void |
5009 | remove_insns_for_debug (blist_t bnds, av_set_t *av_vliw_p) |
5010 | { |
5011 | if (! dbg_cnt (index: sel_sched_insn_cnt) || force_next_insn) |
5012 | /* Leave only the next insn in av_vliw. */ |
5013 | { |
5014 | av_set_iterator av_it; |
5015 | expr_t expr; |
5016 | bnd_t bnd = BLIST_BND (bnds); |
5017 | insn_t next = BND_TO (bnd); |
5018 | |
5019 | gcc_assert (BLIST_NEXT (bnds) == NULL); |
5020 | |
5021 | FOR_EACH_EXPR_1 (expr, av_it, av_vliw_p) |
5022 | if (EXPR_INSN_RTX (expr) != next) |
5023 | av_set_iter_remove (&av_it); |
5024 | } |
5025 | } |
5026 | |
5027 | /* Compute available instructions on BNDS. FENCE is the current fence. Write |
5028 | the computed set to *AV_VLIW_P. */ |
5029 | static void |
5030 | compute_av_set_on_boundaries (fence_t fence, blist_t bnds, av_set_t *av_vliw_p) |
5031 | { |
5032 | if (sched_verbose >= 2) |
5033 | { |
5034 | sel_print (fmt: "Boundaries: " ); |
5035 | dump_blist (bnds); |
5036 | sel_print (fmt: "\n" ); |
5037 | } |
5038 | |
5039 | for (; bnds; bnds = BLIST_NEXT (bnds)) |
5040 | { |
5041 | bnd_t bnd = BLIST_BND (bnds); |
5042 | av_set_t av1_copy; |
5043 | insn_t bnd_to = BND_TO (bnd); |
5044 | |
5045 | /* Rewind BND->TO to the basic block header in case some bookkeeping |
5046 | instructions were inserted before BND->TO and it needs to be |
5047 | adjusted. */ |
5048 | if (sel_bb_head_p (bnd_to)) |
5049 | gcc_assert (INSN_SCHED_TIMES (bnd_to) == 0); |
5050 | else |
5051 | while (INSN_SCHED_TIMES (PREV_INSN (bnd_to)) == 0) |
5052 | { |
5053 | bnd_to = PREV_INSN (insn: bnd_to); |
5054 | if (sel_bb_head_p (bnd_to)) |
5055 | break; |
5056 | } |
5057 | |
5058 | if (BND_TO (bnd) != bnd_to) |
5059 | { |
5060 | gcc_assert (FENCE_INSN (fence) == BND_TO (bnd)); |
5061 | FENCE_INSN (fence) = bnd_to; |
5062 | BND_TO (bnd) = bnd_to; |
5063 | } |
5064 | |
5065 | av_set_clear (&BND_AV (bnd)); |
5066 | BND_AV (bnd) = compute_av_set (BND_TO (bnd), NULL, ws: 0, unique_p: true); |
5067 | |
5068 | av_set_clear (&BND_AV1 (bnd)); |
5069 | BND_AV1 (bnd) = av_set_copy (BND_AV (bnd)); |
5070 | |
5071 | moveup_set_inside_insn_group (avp: &BND_AV1 (bnd), NULL); |
5072 | |
5073 | av1_copy = av_set_copy (BND_AV1 (bnd)); |
5074 | av_set_union_and_clear (av_vliw_p, &av1_copy, NULL); |
5075 | } |
5076 | |
5077 | if (sched_verbose >= 2) |
5078 | { |
5079 | sel_print (fmt: "Available exprs (vliw form): " ); |
5080 | dump_av_set (*av_vliw_p); |
5081 | sel_print (fmt: "\n" ); |
5082 | } |
5083 | } |
5084 | |
5085 | /* Calculate the sequential av set on BND corresponding to the EXPR_VLIW |
5086 | expression. When FOR_MOVEOP is true, also replace the register of |
5087 | expressions found with the register from EXPR_VLIW. */ |
5088 | static av_set_t |
5089 | find_sequential_best_exprs (bnd_t bnd, expr_t expr_vliw, bool for_moveop) |
5090 | { |
5091 | av_set_t expr_seq = NULL; |
5092 | expr_t expr; |
5093 | av_set_iterator i; |
5094 | |
5095 | FOR_EACH_EXPR (expr, i, BND_AV (bnd)) |
5096 | { |
5097 | if (equal_after_moveup_path_p (expr, NULL, expr_vliw)) |
5098 | { |
5099 | if (for_moveop) |
5100 | { |
5101 | /* The sequential expression has the right form to pass |
5102 | to move_op except when renaming happened. Put the |
5103 | correct register in EXPR then. */ |
5104 | if (EXPR_SEPARABLE_P (expr) && REG_P (EXPR_LHS (expr))) |
5105 | { |
5106 | if (expr_dest_regno (expr) != expr_dest_regno (expr_vliw)) |
5107 | { |
5108 | replace_dest_with_reg_in_expr (expr, EXPR_LHS (expr_vliw)); |
5109 | stat_renamed_scheduled++; |
5110 | } |
5111 | /* Also put the correct TARGET_AVAILABLE bit on the expr. |
5112 | This is needed when renaming came up with original |
5113 | register. */ |
5114 | else if (EXPR_TARGET_AVAILABLE (expr) |
5115 | != EXPR_TARGET_AVAILABLE (expr_vliw)) |
5116 | { |
5117 | gcc_assert (EXPR_TARGET_AVAILABLE (expr_vliw) == 1); |
5118 | EXPR_TARGET_AVAILABLE (expr) = 1; |
5119 | } |
5120 | } |
5121 | if (EXPR_WAS_SUBSTITUTED (expr)) |
5122 | stat_substitutions_total++; |
5123 | } |
5124 | |
5125 | av_set_add (&expr_seq, expr); |
5126 | |
5127 | /* With substitution inside insn group, it is possible |
5128 | that more than one expression in expr_seq will correspond |
5129 | to expr_vliw. In this case, choose one as the attempt to |
5130 | move both leads to miscompiles. */ |
5131 | break; |
5132 | } |
5133 | } |
5134 | |
5135 | if (for_moveop && sched_verbose >= 2) |
5136 | { |
5137 | sel_print (fmt: "Best expression(s) (sequential form): " ); |
5138 | dump_av_set (expr_seq); |
5139 | sel_print (fmt: "\n" ); |
5140 | } |
5141 | |
5142 | return expr_seq; |
5143 | } |
5144 | |
5145 | |
5146 | /* Move nop to previous block. */ |
5147 | static void ATTRIBUTE_UNUSED |
5148 | move_nop_to_previous_block (insn_t nop, basic_block prev_bb) |
5149 | { |
5150 | insn_t prev_insn, next_insn; |
5151 | |
5152 | gcc_assert (sel_bb_head_p (nop) |
5153 | && prev_bb == BLOCK_FOR_INSN (nop)->prev_bb); |
5154 | rtx_note *note = bb_note (BLOCK_FOR_INSN (insn: nop)); |
5155 | prev_insn = sel_bb_end (prev_bb); |
5156 | next_insn = NEXT_INSN (insn: nop); |
5157 | gcc_assert (prev_insn != NULL_RTX |
5158 | && PREV_INSN (note) == prev_insn); |
5159 | |
5160 | SET_NEXT_INSN (prev_insn) = nop; |
5161 | SET_PREV_INSN (nop) = prev_insn; |
5162 | |
5163 | SET_PREV_INSN (note) = nop; |
5164 | SET_NEXT_INSN (note) = next_insn; |
5165 | |
5166 | SET_NEXT_INSN (nop) = note; |
5167 | SET_PREV_INSN (next_insn) = note; |
5168 | |
5169 | BB_END (prev_bb) = nop; |
5170 | BLOCK_FOR_INSN (insn: nop) = prev_bb; |
5171 | } |
5172 | |
5173 | /* Prepare a place to insert the chosen expression on BND. */ |
5174 | static insn_t |
5175 | prepare_place_to_insert (bnd_t bnd) |
5176 | { |
5177 | insn_t place_to_insert; |
5178 | |
5179 | /* Init place_to_insert before calling move_op, as the later |
5180 | can possibly remove BND_TO (bnd). */ |
5181 | if (/* If this is not the first insn scheduled. */ |
5182 | BND_PTR (bnd)) |
5183 | { |
5184 | /* Add it after last scheduled. */ |
5185 | place_to_insert = ILIST_INSN (BND_PTR (bnd)); |
5186 | if (DEBUG_INSN_P (place_to_insert)) |
5187 | { |
5188 | ilist_t l = BND_PTR (bnd); |
5189 | while ((l = ILIST_NEXT (l)) && |
5190 | DEBUG_INSN_P (ILIST_INSN (l))) |
5191 | ; |
5192 | if (!l) |
5193 | place_to_insert = NULL; |
5194 | } |
5195 | } |
5196 | else |
5197 | place_to_insert = NULL; |
5198 | |
5199 | if (!place_to_insert) |
5200 | { |
5201 | /* Add it before BND_TO. The difference is in the |
5202 | basic block, where INSN will be added. */ |
5203 | place_to_insert = get_nop_from_pool (BND_TO (bnd)); |
5204 | gcc_assert (BLOCK_FOR_INSN (place_to_insert) |
5205 | == BLOCK_FOR_INSN (BND_TO (bnd))); |
5206 | } |
5207 | |
5208 | return place_to_insert; |
5209 | } |
5210 | |
5211 | /* Find original instructions for EXPR_SEQ and move it to BND boundary. |
5212 | Return the expression to emit in C_EXPR. */ |
5213 | static bool |
5214 | move_exprs_to_boundary (bnd_t bnd, expr_t expr_vliw, |
5215 | av_set_t expr_seq, expr_t c_expr) |
5216 | { |
5217 | bool b, should_move; |
5218 | unsigned book_uid; |
5219 | bitmap_iterator bi; |
5220 | int n_bookkeeping_copies_before_moveop; |
5221 | |
5222 | /* Make a move. This call will remove the original operation, |
5223 | insert all necessary bookkeeping instructions and update the |
5224 | data sets. After that all we have to do is add the operation |
5225 | at before BND_TO (BND). */ |
5226 | n_bookkeeping_copies_before_moveop = stat_bookkeeping_copies; |
5227 | max_uid_before_move_op = get_max_uid (); |
5228 | bitmap_clear (current_copies); |
5229 | bitmap_clear (current_originators); |
5230 | |
5231 | b = move_op (BND_TO (bnd), expr_seq, expr_vliw, |
5232 | get_dest_from_orig_ops (expr_seq), c_expr, &should_move); |
5233 | |
5234 | /* We should be able to find the expression we've chosen for |
5235 | scheduling. */ |
5236 | gcc_assert (b); |
5237 | |
5238 | if (stat_bookkeeping_copies > n_bookkeeping_copies_before_moveop) |
5239 | stat_insns_needed_bookkeeping++; |
5240 | |
5241 | EXECUTE_IF_SET_IN_BITMAP (current_copies, 0, book_uid, bi) |
5242 | { |
5243 | unsigned uid; |
5244 | bitmap_iterator bi; |
5245 | |
5246 | /* We allocate these bitmaps lazily. */ |
5247 | if (! INSN_ORIGINATORS_BY_UID (book_uid)) |
5248 | INSN_ORIGINATORS_BY_UID (book_uid) = BITMAP_ALLOC (NULL); |
5249 | |
5250 | bitmap_copy (INSN_ORIGINATORS_BY_UID (book_uid), |
5251 | current_originators); |
5252 | |
5253 | /* Transitively add all originators' originators. */ |
5254 | EXECUTE_IF_SET_IN_BITMAP (current_originators, 0, uid, bi) |
5255 | if (INSN_ORIGINATORS_BY_UID (uid)) |
5256 | bitmap_ior_into (INSN_ORIGINATORS_BY_UID (book_uid), |
5257 | INSN_ORIGINATORS_BY_UID (uid)); |
5258 | } |
5259 | |
5260 | return should_move; |
5261 | } |
5262 | |
5263 | |
5264 | /* Debug a DFA state as an array of bytes. */ |
5265 | static void |
5266 | debug_state (state_t state) |
5267 | { |
5268 | unsigned char *p; |
5269 | unsigned int i, size = dfa_state_size; |
5270 | |
5271 | sel_print (fmt: "state (%u):" , size); |
5272 | for (i = 0, p = (unsigned char *) state; i < size; i++) |
5273 | sel_print (fmt: " %d" , p[i]); |
5274 | sel_print (fmt: "\n" ); |
5275 | } |
5276 | |
5277 | /* Advance state on FENCE with INSN. Return true if INSN is |
5278 | an ASM, and we should advance state once more. */ |
5279 | static bool |
5280 | advance_state_on_fence (fence_t fence, insn_t insn) |
5281 | { |
5282 | bool asm_p; |
5283 | |
5284 | if (recog_memoized (insn) >= 0) |
5285 | { |
5286 | int res; |
5287 | state_t temp_state = alloca (dfa_state_size); |
5288 | |
5289 | gcc_assert (!INSN_ASM_P (insn)); |
5290 | asm_p = false; |
5291 | |
5292 | memcpy (dest: temp_state, FENCE_STATE (fence), n: dfa_state_size); |
5293 | res = state_transition (FENCE_STATE (fence), insn); |
5294 | gcc_assert (res < 0); |
5295 | |
5296 | if (memcmp (s1: temp_state, FENCE_STATE (fence), n: dfa_state_size)) |
5297 | { |
5298 | FENCE_ISSUED_INSNS (fence)++; |
5299 | |
5300 | /* We should never issue more than issue_rate insns. */ |
5301 | if (FENCE_ISSUED_INSNS (fence) > issue_rate) |
5302 | gcc_unreachable (); |
5303 | } |
5304 | } |
5305 | else |
5306 | { |
5307 | /* This could be an ASM insn which we'd like to schedule |
5308 | on the next cycle. */ |
5309 | asm_p = INSN_ASM_P (insn); |
5310 | if (!FENCE_STARTS_CYCLE_P (fence) && asm_p) |
5311 | advance_one_cycle (fence); |
5312 | } |
5313 | |
5314 | if (sched_verbose >= 2) |
5315 | debug_state (FENCE_STATE (fence)); |
5316 | if (!DEBUG_INSN_P (insn)) |
5317 | FENCE_STARTS_CYCLE_P (fence) = 0; |
5318 | FENCE_ISSUE_MORE (fence) = can_issue_more; |
5319 | return asm_p; |
5320 | } |
5321 | |
5322 | /* Update FENCE on which INSN was scheduled and this INSN, too. NEED_STALL |
5323 | is nonzero if we need to stall after issuing INSN. */ |
5324 | static void |
5325 | update_fence_and_insn (fence_t fence, insn_t insn, int need_stall) |
5326 | { |
5327 | bool asm_p; |
5328 | |
5329 | /* First, reflect that something is scheduled on this fence. */ |
5330 | asm_p = advance_state_on_fence (fence, insn); |
5331 | FENCE_LAST_SCHEDULED_INSN (fence) = insn; |
5332 | vec_safe_push (FENCE_EXECUTING_INSNS (fence), obj: insn); |
5333 | if (SCHED_GROUP_P (insn)) |
5334 | { |
5335 | FENCE_SCHED_NEXT (fence) = INSN_SCHED_NEXT (insn); |
5336 | SCHED_GROUP_P (insn) = 0; |
5337 | } |
5338 | else |
5339 | FENCE_SCHED_NEXT (fence) = NULL; |
5340 | if (INSN_UID (insn) < FENCE_READY_TICKS_SIZE (fence)) |
5341 | FENCE_READY_TICKS (fence) [INSN_UID (insn)] = 0; |
5342 | |
5343 | /* Set instruction scheduling info. This will be used in bundling, |
5344 | pipelining, tick computations etc. */ |
5345 | ++INSN_SCHED_TIMES (insn); |
5346 | EXPR_TARGET_AVAILABLE (INSN_EXPR (insn)) = true; |
5347 | EXPR_ORIG_SCHED_CYCLE (INSN_EXPR (insn)) = FENCE_CYCLE (fence); |
5348 | INSN_AFTER_STALL_P (insn) = FENCE_AFTER_STALL_P (fence); |
5349 | INSN_SCHED_CYCLE (insn) = FENCE_CYCLE (fence); |
5350 | |
5351 | /* This does not account for adjust_cost hooks, just add the biggest |
5352 | constant the hook may add to the latency. TODO: make this |
5353 | a target dependent constant. */ |
5354 | INSN_READY_CYCLE (insn) |
5355 | = INSN_SCHED_CYCLE (insn) + (INSN_CODE (insn) < 0 |
5356 | ? 1 |
5357 | : maximal_insn_latency (insn) + 1); |
5358 | |
5359 | /* Change these fields last, as they're used above. */ |
5360 | FENCE_AFTER_STALL_P (fence) = 0; |
5361 | if (asm_p || need_stall) |
5362 | advance_one_cycle (fence); |
5363 | |
5364 | /* Indicate that we've scheduled something on this fence. */ |
5365 | FENCE_SCHEDULED_P (fence) = true; |
5366 | scheduled_something_on_previous_fence = true; |
5367 | |
5368 | /* Print debug information when insn's fields are updated. */ |
5369 | if (sched_verbose >= 2) |
5370 | { |
5371 | sel_print (fmt: "Scheduling insn: " ); |
5372 | dump_insn_1 (insn, 1); |
5373 | sel_print (fmt: "\n" ); |
5374 | } |
5375 | } |
5376 | |
5377 | /* Update boundary BND (and, if needed, FENCE) with INSN, remove the |
5378 | old boundary from BNDSP, add new boundaries to BNDS_TAIL_P and |
5379 | return it. */ |
5380 | static blist_t * |
5381 | update_boundaries (fence_t fence, bnd_t bnd, insn_t insn, blist_t *bndsp, |
5382 | blist_t *bnds_tailp) |
5383 | { |
5384 | succ_iterator si; |
5385 | insn_t succ; |
5386 | |
5387 | advance_deps_context (BND_DC (bnd), insn); |
5388 | FOR_EACH_SUCC_1 (succ, si, insn, |
5389 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
5390 | { |
5391 | ilist_t ptr = ilist_copy (BND_PTR (bnd)); |
5392 | |
5393 | ilist_add (lp: &ptr, insn); |
5394 | |
5395 | if (DEBUG_INSN_P (insn) && sel_bb_end_p (insn) |
5396 | && is_ineligible_successor (insn: succ, p: ptr)) |
5397 | { |
5398 | ilist_clear (&ptr); |
5399 | continue; |
5400 | } |
5401 | |
5402 | if (FENCE_INSN (fence) == insn && !sel_bb_end_p (insn)) |
5403 | { |
5404 | if (sched_verbose >= 9) |
5405 | sel_print (fmt: "Updating fence insn from %i to %i\n" , |
5406 | INSN_UID (insn), INSN_UID (insn: succ)); |
5407 | FENCE_INSN (fence) = succ; |
5408 | } |
5409 | blist_add (bnds_tailp, succ, ptr, BND_DC (bnd)); |
5410 | bnds_tailp = &BLIST_NEXT (*bnds_tailp); |
5411 | } |
5412 | |
5413 | blist_remove (bndsp); |
5414 | return bnds_tailp; |
5415 | } |
5416 | |
5417 | /* Schedule EXPR_VLIW on BND. Return the insn emitted. */ |
5418 | static insn_t |
5419 | schedule_expr_on_boundary (bnd_t bnd, expr_t expr_vliw, int seqno) |
5420 | { |
5421 | av_set_t expr_seq; |
5422 | expr_t c_expr = XALLOCA (expr_def); |
5423 | insn_t place_to_insert; |
5424 | insn_t insn; |
5425 | bool should_move; |
5426 | |
5427 | expr_seq = find_sequential_best_exprs (bnd, expr_vliw, for_moveop: true); |
5428 | |
5429 | /* In case of scheduling a jump skipping some other instructions, |
5430 | prepare CFG. After this, jump is at the boundary and can be |
5431 | scheduled as usual insn by MOVE_OP. */ |
5432 | if (vinsn_cond_branch_p (EXPR_VINSN (expr_vliw))) |
5433 | { |
5434 | insn = EXPR_INSN_RTX (expr_vliw); |
5435 | |
5436 | /* Speculative jumps are not handled. */ |
5437 | if (insn != BND_TO (bnd) |
5438 | && !sel_insn_is_speculation_check (insn)) |
5439 | move_cond_jump (insn, bnd); |
5440 | } |
5441 | |
5442 | /* Find a place for C_EXPR to schedule. */ |
5443 | place_to_insert = prepare_place_to_insert (bnd); |
5444 | should_move = move_exprs_to_boundary (bnd, expr_vliw, expr_seq, c_expr); |
5445 | clear_expr (c_expr); |
5446 | |
5447 | /* Add the instruction. The corner case to care about is when |
5448 | the expr_seq set has more than one expr, and we chose the one that |
5449 | is not equal to expr_vliw. Then expr_vliw may be insn in stream, and |
5450 | we can't use it. Generate the new vinsn. */ |
5451 | if (INSN_IN_STREAM_P (EXPR_INSN_RTX (expr_vliw))) |
5452 | { |
5453 | vinsn_t vinsn_new; |
5454 | |
5455 | vinsn_new = vinsn_copy (EXPR_VINSN (expr_vliw), false); |
5456 | change_vinsn_in_expr (expr_vliw, vinsn_new); |
5457 | should_move = false; |
5458 | } |
5459 | if (should_move) |
5460 | insn = sel_move_insn (expr_vliw, seqno, place_to_insert); |
5461 | else |
5462 | insn = emit_insn_from_expr_after (expr: expr_vliw, NULL, seqno, |
5463 | place_to_insert); |
5464 | |
5465 | /* Return the nops generated for preserving of data sets back |
5466 | into pool. */ |
5467 | if (INSN_NOP_P (place_to_insert)) |
5468 | return_nop_to_pool (place_to_insert, !DEBUG_INSN_P (insn)); |
5469 | remove_temp_moveop_nops (full_tidying: !DEBUG_INSN_P (insn)); |
5470 | |
5471 | av_set_clear (&expr_seq); |
5472 | |
5473 | /* Save the expression scheduled so to reset target availability if we'll |
5474 | meet it later on the same fence. */ |
5475 | if (EXPR_WAS_RENAMED (expr_vliw)) |
5476 | vinsn_vec_add (vinsn_vec: &vec_target_unavailable_vinsns, INSN_EXPR (insn)); |
5477 | |
5478 | /* Check that the recent movement didn't destroyed loop |
5479 | structure. */ |
5480 | gcc_assert (!pipelining_p |
5481 | || current_loop_nest == NULL |
5482 | || loop_latch_edge (current_loop_nest)); |
5483 | return insn; |
5484 | } |
5485 | |
5486 | /* Stall for N cycles on FENCE. */ |
5487 | static void |
5488 | stall_for_cycles (fence_t fence, int n) |
5489 | { |
5490 | int could_more; |
5491 | |
5492 | could_more = n > 1 || FENCE_ISSUED_INSNS (fence) < issue_rate; |
5493 | while (n--) |
5494 | advance_one_cycle (fence); |
5495 | if (could_more) |
5496 | FENCE_AFTER_STALL_P (fence) = 1; |
5497 | } |
5498 | |
5499 | /* Gather a parallel group of insns at FENCE and assign their seqno |
5500 | to SEQNO. All scheduled insns are gathered in SCHEDULED_INSNS_TAILPP |
5501 | list for later recalculation of seqnos. */ |
5502 | static void |
5503 | fill_insns (fence_t fence, int seqno, ilist_t **scheduled_insns_tailpp) |
5504 | { |
5505 | blist_t bnds = NULL, *bnds_tailp; |
5506 | av_set_t av_vliw = NULL; |
5507 | insn_t insn = FENCE_INSN (fence); |
5508 | |
5509 | if (sched_verbose >= 2) |
5510 | sel_print (fmt: "Starting fill_insns for insn %d, cycle %d\n" , |
5511 | INSN_UID (insn), FENCE_CYCLE (fence)); |
5512 | |
5513 | blist_add (&bnds, insn, NULL, FENCE_DC (fence)); |
5514 | bnds_tailp = &BLIST_NEXT (bnds); |
5515 | set_target_context (FENCE_TC (fence)); |
5516 | can_issue_more = FENCE_ISSUE_MORE (fence); |
5517 | target_bb = INSN_BB (insn); |
5518 | |
5519 | /* Do while we can add any operation to the current group. */ |
5520 | do |
5521 | { |
5522 | blist_t *bnds_tailp1, *bndsp; |
5523 | expr_t expr_vliw; |
5524 | int need_stall = false; |
5525 | int was_stall = 0, scheduled_insns = 0; |
5526 | int max_insns = pipelining_p ? issue_rate : 2 * issue_rate; |
5527 | int max_stall = pipelining_p ? 1 : 3; |
5528 | bool last_insn_was_debug = false; |
5529 | bool was_debug_bb_end_p = false; |
5530 | |
5531 | compute_av_set_on_boundaries (fence, bnds, av_vliw_p: &av_vliw); |
5532 | remove_insns_that_need_bookkeeping (fence, av_ptr: &av_vliw); |
5533 | remove_insns_for_debug (bnds, av_vliw_p: &av_vliw); |
5534 | |
5535 | /* Return early if we have nothing to schedule. */ |
5536 | if (av_vliw == NULL) |
5537 | break; |
5538 | |
5539 | /* Choose the best expression and, if needed, destination register |
5540 | for it. */ |
5541 | do |
5542 | { |
5543 | expr_vliw = find_best_expr (av_vliw_ptr: &av_vliw, bnds, fence, pneed_stall: &need_stall); |
5544 | if (! expr_vliw && need_stall) |
5545 | { |
5546 | /* All expressions required a stall. Do not recompute av sets |
5547 | as we'll get the same answer (modulo the insns between |
5548 | the fence and its boundary, which will not be available for |
5549 | pipelining). |
5550 | If we are going to stall for too long, break to recompute av |
5551 | sets and bring more insns for pipelining. */ |
5552 | was_stall++; |
5553 | if (need_stall <= 3) |
5554 | stall_for_cycles (fence, n: need_stall); |
5555 | else |
5556 | { |
5557 | stall_for_cycles (fence, n: 1); |
5558 | break; |
5559 | } |
5560 | } |
5561 | } |
5562 | while (! expr_vliw && need_stall); |
5563 | |
5564 | /* Now either we've selected expr_vliw or we have nothing to schedule. */ |
5565 | if (!expr_vliw) |
5566 | { |
5567 | av_set_clear (&av_vliw); |
5568 | break; |
5569 | } |
5570 | |
5571 | bndsp = &bnds; |
5572 | bnds_tailp1 = bnds_tailp; |
5573 | |
5574 | do |
5575 | /* This code will be executed only once until we'd have several |
5576 | boundaries per fence. */ |
5577 | { |
5578 | bnd_t bnd = BLIST_BND (*bndsp); |
5579 | |
5580 | if (!av_set_is_in_p (BND_AV1 (bnd), EXPR_VINSN (expr_vliw))) |
5581 | { |
5582 | bndsp = &BLIST_NEXT (*bndsp); |
5583 | continue; |
5584 | } |
5585 | |
5586 | insn = schedule_expr_on_boundary (bnd, expr_vliw, seqno); |
5587 | last_insn_was_debug = DEBUG_INSN_P (insn); |
5588 | if (last_insn_was_debug) |
5589 | was_debug_bb_end_p = (insn == BND_TO (bnd) && sel_bb_end_p (insn)); |
5590 | update_fence_and_insn (fence, insn, need_stall); |
5591 | bnds_tailp = update_boundaries (fence, bnd, insn, bndsp, bnds_tailp); |
5592 | |
5593 | /* Add insn to the list of scheduled on this cycle instructions. */ |
5594 | ilist_add (lp: *scheduled_insns_tailpp, insn); |
5595 | *scheduled_insns_tailpp = &ILIST_NEXT (**scheduled_insns_tailpp); |
5596 | } |
5597 | while (*bndsp != *bnds_tailp1); |
5598 | |
5599 | av_set_clear (&av_vliw); |
5600 | if (!last_insn_was_debug) |
5601 | scheduled_insns++; |
5602 | |
5603 | /* We currently support information about candidate blocks only for |
5604 | one 'target_bb' block. Hence we can't schedule after jump insn, |
5605 | as this will bring two boundaries and, hence, necessity to handle |
5606 | information for two or more blocks concurrently. */ |
5607 | if ((last_insn_was_debug ? was_debug_bb_end_p : sel_bb_end_p (insn)) |
5608 | || (was_stall |
5609 | && (was_stall >= max_stall |
5610 | || scheduled_insns >= max_insns))) |
5611 | break; |
5612 | } |
5613 | while (bnds); |
5614 | |
5615 | gcc_assert (!FENCE_BNDS (fence)); |
5616 | |
5617 | /* Update boundaries of the FENCE. */ |
5618 | while (bnds) |
5619 | { |
5620 | ilist_t ptr = BND_PTR (BLIST_BND (bnds)); |
5621 | |
5622 | if (ptr) |
5623 | { |
5624 | insn = ILIST_INSN (ptr); |
5625 | |
5626 | if (!ilist_is_in_p (FENCE_BNDS (fence), insn)) |
5627 | ilist_add (lp: &FENCE_BNDS (fence), insn); |
5628 | } |
5629 | |
5630 | blist_remove (&bnds); |
5631 | } |
5632 | |
5633 | /* Update target context on the fence. */ |
5634 | reset_target_context (FENCE_TC (fence), false); |
5635 | } |
5636 | |
5637 | /* All exprs in ORIG_OPS must have the same destination register or memory. |
5638 | Return that destination. */ |
5639 | static rtx |
5640 | get_dest_from_orig_ops (av_set_t orig_ops) |
5641 | { |
5642 | rtx dest = NULL_RTX; |
5643 | av_set_iterator av_it; |
5644 | expr_t expr; |
5645 | bool first_p = true; |
5646 | |
5647 | FOR_EACH_EXPR (expr, av_it, orig_ops) |
5648 | { |
5649 | rtx x = EXPR_LHS (expr); |
5650 | |
5651 | if (first_p) |
5652 | { |
5653 | first_p = false; |
5654 | dest = x; |
5655 | } |
5656 | else |
5657 | gcc_assert (dest == x |
5658 | || (dest != NULL_RTX && x != NULL_RTX |
5659 | && rtx_equal_p (dest, x))); |
5660 | } |
5661 | |
5662 | return dest; |
5663 | } |
5664 | |
5665 | /* Update data sets for the bookkeeping block and record those expressions |
5666 | which become no longer available after inserting this bookkeeping. */ |
5667 | static void |
5668 | update_and_record_unavailable_insns (basic_block book_block) |
5669 | { |
5670 | av_set_iterator i; |
5671 | av_set_t old_av_set = NULL; |
5672 | expr_t cur_expr; |
5673 | rtx_insn *bb_end = sel_bb_end (book_block); |
5674 | |
5675 | /* First, get correct liveness in the bookkeeping block. The problem is |
5676 | the range between the bookeeping insn and the end of block. */ |
5677 | update_liveness_on_insn (insn: bb_end); |
5678 | if (control_flow_insn_p (bb_end)) |
5679 | update_liveness_on_insn (insn: PREV_INSN (insn: bb_end)); |
5680 | |
5681 | /* If there's valid av_set on BOOK_BLOCK, then there might exist another |
5682 | fence above, where we may choose to schedule an insn which is |
5683 | actually blocked from moving up with the bookkeeping we create here. */ |
5684 | if (AV_SET_VALID_P (sel_bb_head (book_block))) |
5685 | { |
5686 | old_av_set = av_set_copy (BB_AV_SET (book_block)); |
5687 | update_data_sets (insn: sel_bb_head (book_block)); |
5688 | |
5689 | /* Traverse all the expressions in the old av_set and check whether |
5690 | CUR_EXPR is in new AV_SET. */ |
5691 | FOR_EACH_EXPR (cur_expr, i, old_av_set) |
5692 | { |
5693 | expr_t new_expr = av_set_lookup (BB_AV_SET (book_block), |
5694 | EXPR_VINSN (cur_expr)); |
5695 | |
5696 | if (! new_expr |
5697 | /* In this case, we can just turn off the E_T_A bit, but we can't |
5698 | represent this information with the current vector. */ |
5699 | || EXPR_TARGET_AVAILABLE (new_expr) |
5700 | != EXPR_TARGET_AVAILABLE (cur_expr)) |
5701 | /* Unfortunately, the below code could be also fired up on |
5702 | separable insns, e.g. when moving insns through the new |
5703 | speculation check as in PR 53701. */ |
5704 | vinsn_vec_add (vinsn_vec: &vec_bookkeeping_blocked_vinsns, expr: cur_expr); |
5705 | } |
5706 | |
5707 | av_set_clear (&old_av_set); |
5708 | } |
5709 | } |
5710 | |
5711 | /* The main effect of this function is that sparams->c_expr is merged |
5712 | with (or copied to) lparams->c_expr_merged. If there's only one successor, |
5713 | we avoid merging anything by copying sparams->c_expr to lparams->c_expr_merged. |
5714 | lparams->c_expr_merged is copied back to sparams->c_expr after all |
5715 | successors has been traversed. lparams->c_expr_local is an expr allocated |
5716 | on stack in the caller function, and is used if there is more than one |
5717 | successor. |
5718 | |
5719 | SUCC is one of the SUCCS_NORMAL successors of INSN, |
5720 | MOVEOP_DRV_CALL_RES is the result of call code_motion_path_driver on succ, |
5721 | LPARAMS and STATIC_PARAMS contain the parameters described above. */ |
5722 | static void |
5723 | move_op_merge_succs (insn_t insn ATTRIBUTE_UNUSED, |
5724 | insn_t succ ATTRIBUTE_UNUSED, |
5725 | int moveop_drv_call_res, |
5726 | cmpd_local_params_p lparams, void *static_params) |
5727 | { |
5728 | moveop_static_params_p sparams = (moveop_static_params_p) static_params; |
5729 | |
5730 | /* Nothing to do, if original expr wasn't found below. */ |
5731 | if (moveop_drv_call_res != 1) |
5732 | return; |
5733 | |
5734 | /* If this is a first successor. */ |
5735 | if (!lparams->c_expr_merged) |
5736 | { |
5737 | lparams->c_expr_merged = sparams->c_expr; |
5738 | sparams->c_expr = lparams->c_expr_local; |
5739 | } |
5740 | else |
5741 | { |
5742 | /* We must merge all found expressions to get reasonable |
5743 | EXPR_SPEC_DONE_DS for the resulting insn. If we don't |
5744 | do so then we can first find the expr with epsilon |
5745 | speculation success probability and only then with the |
5746 | good probability. As a result the insn will get epsilon |
5747 | probability and will never be scheduled because of |
5748 | weakness_cutoff in find_best_expr. |
5749 | |
5750 | We call merge_expr_data here instead of merge_expr |
5751 | because due to speculation C_EXPR and X may have the |
5752 | same insns with different speculation types. And as of |
5753 | now such insns are considered non-equal. |
5754 | |
5755 | However, EXPR_SCHED_TIMES is different -- we must get |
5756 | SCHED_TIMES from a real insn, not a bookkeeping copy. |
5757 | We force this here. Instead, we may consider merging |
5758 | SCHED_TIMES to the maximum instead of minimum in the |
5759 | below function. */ |
5760 | int old_times = EXPR_SCHED_TIMES (lparams->c_expr_merged); |
5761 | |
5762 | merge_expr_data (lparams->c_expr_merged, sparams->c_expr, NULL); |
5763 | if (EXPR_SCHED_TIMES (sparams->c_expr) == 0) |
5764 | EXPR_SCHED_TIMES (lparams->c_expr_merged) = old_times; |
5765 | |
5766 | clear_expr (sparams->c_expr); |
5767 | } |
5768 | } |
5769 | |
5770 | /* Add used regs for the successor SUCC into SPARAMS->USED_REGS. |
5771 | |
5772 | SUCC is one of the SUCCS_NORMAL successors of INSN, |
5773 | MOVEOP_DRV_CALL_RES is the result of call code_motion_path_driver on succ or 0, |
5774 | if SUCC is one of SUCCS_BACK or SUCCS_OUT. |
5775 | STATIC_PARAMS contain USED_REGS set. */ |
5776 | static void |
5777 | fur_merge_succs (insn_t insn ATTRIBUTE_UNUSED, insn_t succ, |
5778 | int moveop_drv_call_res, |
5779 | cmpd_local_params_p lparams ATTRIBUTE_UNUSED, |
5780 | void *static_params) |
5781 | { |
5782 | regset succ_live; |
5783 | fur_static_params_p sparams = (fur_static_params_p) static_params; |
5784 | |
5785 | /* Here we compute live regsets only for branches that do not lie |
5786 | on the code motion paths. These branches correspond to value |
5787 | MOVEOP_DRV_CALL_RES==0 and include SUCCS_BACK and SUCCS_OUT, though |
5788 | for such branches code_motion_path_driver is not called. */ |
5789 | if (moveop_drv_call_res != 0) |
5790 | return; |
5791 | |
5792 | /* Mark all registers that do not meet the following condition: |
5793 | (3) not live on the other path of any conditional branch |
5794 | that is passed by the operation, in case original |
5795 | operations are not present on both paths of the |
5796 | conditional branch. */ |
5797 | succ_live = compute_live (insn: succ); |
5798 | IOR_REG_SET (sparams->used_regs, succ_live); |
5799 | } |
5800 | |
5801 | /* This function is called after the last successor. Copies LP->C_EXPR_MERGED |
5802 | into SP->CEXPR. */ |
5803 | static void |
5804 | move_op_after_merge_succs (cmpd_local_params_p lp, void *sparams) |
5805 | { |
5806 | moveop_static_params_p sp = (moveop_static_params_p) sparams; |
5807 | |
5808 | sp->c_expr = lp->c_expr_merged; |
5809 | } |
5810 | |
5811 | /* Track bookkeeping copies created, insns scheduled, and blocks for |
5812 | rescheduling when INSN is found by move_op. */ |
5813 | static void |
5814 | track_scheduled_insns_and_blocks (rtx_insn *insn) |
5815 | { |
5816 | /* Even if this insn can be a copy that will be removed during current move_op, |
5817 | we still need to count it as an originator. */ |
5818 | bitmap_set_bit (current_originators, INSN_UID (insn)); |
5819 | |
5820 | if (!bitmap_clear_bit (current_copies, INSN_UID (insn))) |
5821 | { |
5822 | /* Note that original block needs to be rescheduled, as we pulled an |
5823 | instruction out of it. */ |
5824 | if (INSN_SCHED_TIMES (insn) > 0) |
5825 | bitmap_set_bit (blocks_to_reschedule, BLOCK_FOR_INSN (insn)->index); |
5826 | else if (INSN_UID (insn) < first_emitted_uid && !DEBUG_INSN_P (insn)) |
5827 | num_insns_scheduled++; |
5828 | } |
5829 | |
5830 | /* For instructions we must immediately remove insn from the |
5831 | stream, so subsequent update_data_sets () won't include this |
5832 | insn into av_set. |
5833 | For expr we must make insn look like "INSN_REG (insn) := c_expr". */ |
5834 | if (INSN_UID (insn) > max_uid_before_move_op) |
5835 | stat_bookkeeping_copies--; |
5836 | } |
5837 | |
5838 | /* Emit a register-register copy for INSN if needed. Return true if |
5839 | emitted one. PARAMS is the move_op static parameters. */ |
5840 | static bool |
5841 | maybe_emit_renaming_copy (rtx_insn *insn, |
5842 | moveop_static_params_p params) |
5843 | { |
5844 | bool insn_emitted = false; |
5845 | rtx cur_reg; |
5846 | |
5847 | /* Bail out early when expression cannot be renamed at all. */ |
5848 | if (!EXPR_SEPARABLE_P (params->c_expr)) |
5849 | return false; |
5850 | |
5851 | cur_reg = expr_dest_reg (params->c_expr); |
5852 | gcc_assert (cur_reg && params->dest && REG_P (params->dest)); |
5853 | |
5854 | /* If original operation has expr and the register chosen for |
5855 | that expr is not original operation's dest reg, substitute |
5856 | operation's right hand side with the register chosen. */ |
5857 | if (REGNO (params->dest) != REGNO (cur_reg)) |
5858 | { |
5859 | insn_t reg_move_insn, reg_move_insn_rtx; |
5860 | |
5861 | reg_move_insn_rtx = create_insn_rtx_with_rhs (INSN_VINSN (insn), |
5862 | rhs_rtx: params->dest); |
5863 | reg_move_insn = sel_gen_insn_from_rtx_after (reg_move_insn_rtx, |
5864 | INSN_EXPR (insn), |
5865 | INSN_SEQNO (insn), |
5866 | insn); |
5867 | EXPR_SPEC_DONE_DS (INSN_EXPR (reg_move_insn)) = 0; |
5868 | replace_dest_with_reg_in_expr (expr: params->c_expr, new_reg: params->dest); |
5869 | |
5870 | insn_emitted = true; |
5871 | params->was_renamed = true; |
5872 | } |
5873 | |
5874 | return insn_emitted; |
5875 | } |
5876 | |
5877 | /* Emit a speculative check for INSN speculated as EXPR if needed. |
5878 | Return true if we've emitted one. PARAMS is the move_op static |
5879 | parameters. */ |
5880 | static bool |
5881 | maybe_emit_speculative_check (rtx_insn *insn, expr_t expr, |
5882 | moveop_static_params_p params) |
5883 | { |
5884 | bool insn_emitted = false; |
5885 | insn_t x; |
5886 | ds_t check_ds; |
5887 | |
5888 | check_ds = get_spec_check_type_for_insn (insn, expr); |
5889 | if (check_ds != 0) |
5890 | { |
5891 | /* A speculation check should be inserted. */ |
5892 | x = create_speculation_check (c_expr: params->c_expr, check_ds, orig_insn: insn); |
5893 | insn_emitted = true; |
5894 | } |
5895 | else |
5896 | { |
5897 | EXPR_SPEC_DONE_DS (INSN_EXPR (insn)) = 0; |
5898 | x = insn; |
5899 | } |
5900 | |
5901 | gcc_assert (EXPR_SPEC_DONE_DS (INSN_EXPR (x)) == 0 |
5902 | && EXPR_SPEC_TO_CHECK_DS (INSN_EXPR (x)) == 0); |
5903 | return insn_emitted; |
5904 | } |
5905 | |
5906 | /* Handle transformations that leave an insn in place of original |
5907 | insn such as renaming/speculation. Return true if one of such |
5908 | transformations actually happened, and we have emitted this insn. */ |
5909 | static bool |
5910 | handle_emitting_transformations (rtx_insn *insn, expr_t expr, |
5911 | moveop_static_params_p params) |
5912 | { |
5913 | bool insn_emitted = false; |
5914 | |
5915 | insn_emitted = maybe_emit_renaming_copy (insn, params); |
5916 | insn_emitted |= maybe_emit_speculative_check (insn, expr, params); |
5917 | |
5918 | return insn_emitted; |
5919 | } |
5920 | |
5921 | /* If INSN is the only insn in the basic block (not counting JUMP, |
5922 | which may be a jump to next insn, and DEBUG_INSNs), we want to |
5923 | leave a NOP there till the return to fill_insns. */ |
5924 | |
5925 | static bool |
5926 | need_nop_to_preserve_insn_bb (rtx_insn *insn) |
5927 | { |
5928 | insn_t bb_head, bb_end, bb_next, in_next; |
5929 | basic_block bb = BLOCK_FOR_INSN (insn); |
5930 | |
5931 | bb_head = sel_bb_head (bb); |
5932 | bb_end = sel_bb_end (bb); |
5933 | |
5934 | if (bb_head == bb_end) |
5935 | return true; |
5936 | |
5937 | while (bb_head != bb_end && DEBUG_INSN_P (bb_head)) |
5938 | bb_head = NEXT_INSN (insn: bb_head); |
5939 | |
5940 | if (bb_head == bb_end) |
5941 | return true; |
5942 | |
5943 | while (bb_head != bb_end && DEBUG_INSN_P (bb_end)) |
5944 | bb_end = PREV_INSN (insn: bb_end); |
5945 | |
5946 | if (bb_head == bb_end) |
5947 | return true; |
5948 | |
5949 | bb_next = NEXT_INSN (insn: bb_head); |
5950 | while (bb_next != bb_end && DEBUG_INSN_P (bb_next)) |
5951 | bb_next = NEXT_INSN (insn: bb_next); |
5952 | |
5953 | if (bb_next == bb_end && JUMP_P (bb_end)) |
5954 | return true; |
5955 | |
5956 | in_next = NEXT_INSN (insn); |
5957 | while (DEBUG_INSN_P (in_next)) |
5958 | in_next = NEXT_INSN (insn: in_next); |
5959 | |
5960 | if (IN_CURRENT_FENCE_P (in_next)) |
5961 | return true; |
5962 | |
5963 | return false; |
5964 | } |
5965 | |
5966 | /* Remove INSN from stream. When ONLY_DISCONNECT is true, its data |
5967 | is not removed but reused when INSN is re-emitted. */ |
5968 | static void |
5969 | remove_insn_from_stream (rtx_insn *insn, bool only_disconnect) |
5970 | { |
5971 | /* If there's only one insn in the BB, make sure that a nop is |
5972 | inserted into it, so the basic block won't disappear when we'll |
5973 | delete INSN below with sel_remove_insn. It should also survive |
5974 | till the return to fill_insns. */ |
5975 | if (need_nop_to_preserve_insn_bb (insn)) |
5976 | { |
5977 | insn_t nop = get_nop_from_pool (insn); |
5978 | gcc_assert (INSN_NOP_P (nop)); |
5979 | vec_temp_moveop_nops.safe_push (obj: nop); |
5980 | } |
5981 | |
5982 | sel_remove_insn (insn, only_disconnect, false); |
5983 | } |
5984 | |
5985 | /* This function is called when original expr is found. |
5986 | INSN - current insn traversed, EXPR - the corresponding expr found. |
5987 | LPARAMS is the local parameters of code modion driver, STATIC_PARAMS |
5988 | is static parameters of move_op. */ |
5989 | static void |
5990 | move_op_orig_expr_found (insn_t insn, expr_t expr, |
5991 | cmpd_local_params_p lparams ATTRIBUTE_UNUSED, |
5992 | void *static_params) |
5993 | { |
5994 | bool only_disconnect; |
5995 | moveop_static_params_p params = (moveop_static_params_p) static_params; |
5996 | |
5997 | copy_expr_onside (params->c_expr, INSN_EXPR (insn)); |
5998 | track_scheduled_insns_and_blocks (insn); |
5999 | handle_emitting_transformations (insn, expr, params); |
6000 | only_disconnect = params->uid == INSN_UID (insn); |
6001 | |
6002 | /* Mark that we've disconnected an insn. */ |
6003 | if (only_disconnect) |
6004 | params->uid = -1; |
6005 | remove_insn_from_stream (insn, only_disconnect); |
6006 | } |
6007 | |
6008 | /* The function is called when original expr is found. |
6009 | INSN - current insn traversed, EXPR - the corresponding expr found, |
6010 | crossed_call_abis and original_insns in STATIC_PARAMS are updated. */ |
6011 | static void |
6012 | fur_orig_expr_found (insn_t insn, expr_t expr ATTRIBUTE_UNUSED, |
6013 | cmpd_local_params_p lparams ATTRIBUTE_UNUSED, |
6014 | void *static_params) |
6015 | { |
6016 | fur_static_params_p params = (fur_static_params_p) static_params; |
6017 | regset tmp; |
6018 | |
6019 | if (CALL_P (insn)) |
6020 | params->crossed_call_abis |= 1 << insn_callee_abi (insn).id (); |
6021 | |
6022 | def_list_add (params->original_insns, insn, params->crossed_call_abis); |
6023 | |
6024 | /* Mark the registers that do not meet the following condition: |
6025 | (2) not among the live registers of the point |
6026 | immediately following the first original operation on |
6027 | a given downward path, except for the original target |
6028 | register of the operation. */ |
6029 | tmp = get_clear_regset_from_pool (); |
6030 | compute_live_below_insn (insn, regs: tmp); |
6031 | AND_COMPL_REG_SET (tmp, INSN_REG_SETS (insn)); |
6032 | AND_COMPL_REG_SET (tmp, INSN_REG_CLOBBERS (insn)); |
6033 | IOR_REG_SET (params->used_regs, tmp); |
6034 | return_regset_to_pool (tmp); |
6035 | |
6036 | /* (*1) We need to add to USED_REGS registers that are read by |
6037 | INSN's lhs. This may lead to choosing wrong src register. |
6038 | E.g. (scheduling const expr enabled): |
6039 | |
6040 | 429: ax=0x0 <- Can't use AX for this expr (0x0) |
6041 | 433: dx=[bp-0x18] |
6042 | 427: [ax+dx+0x1]=ax |
6043 | REG_DEAD: ax |
6044 | 168: di=dx |
6045 | REG_DEAD: dx |
6046 | */ |
6047 | /* FIXME: see comment above and enable MEM_P |
6048 | in vinsn_separable_p. */ |
6049 | gcc_assert (!VINSN_SEPARABLE_P (INSN_VINSN (insn)) |
6050 | || !MEM_P (INSN_LHS (insn))); |
6051 | } |
6052 | |
6053 | /* This function is called on the ascending pass, before returning from |
6054 | current basic block. */ |
6055 | static void |
6056 | move_op_at_first_insn (insn_t insn, cmpd_local_params_p lparams, |
6057 | void *static_params) |
6058 | { |
6059 | moveop_static_params_p sparams = (moveop_static_params_p) static_params; |
6060 | basic_block book_block = NULL; |
6061 | |
6062 | /* When we have removed the boundary insn for scheduling, which also |
6063 | happened to be the end insn in its bb, we don't need to update sets. */ |
6064 | if (!lparams->removed_last_insn |
6065 | && lparams->e1 |
6066 | && sel_bb_head_p (insn)) |
6067 | { |
6068 | /* We should generate bookkeeping code only if we are not at the |
6069 | top level of the move_op. */ |
6070 | if (sel_num_cfg_preds_gt_1 (insn)) |
6071 | book_block = generate_bookkeeping_insn (c_expr: sparams->c_expr, |
6072 | e1: lparams->e1, e2: lparams->e2); |
6073 | /* Update data sets for the current insn. */ |
6074 | update_data_sets (insn); |
6075 | } |
6076 | |
6077 | /* If bookkeeping code was inserted, we need to update av sets of basic |
6078 | block that received bookkeeping. After generation of bookkeeping insn, |
6079 | bookkeeping block does not contain valid av set because we are not following |
6080 | the original algorithm in every detail with regards to e.g. renaming |
6081 | simple reg-reg copies. Consider example: |
6082 | |
6083 | bookkeeping block scheduling fence |
6084 | \ / |
6085 | \ join / |
6086 | ---------- |
6087 | | | |
6088 | ---------- |
6089 | / \ |
6090 | / \ |
6091 | r1 := r2 r1 := r3 |
6092 | |
6093 | We try to schedule insn "r1 := r3" on the current |
6094 | scheduling fence. Also, note that av set of bookkeeping block |
6095 | contain both insns "r1 := r2" and "r1 := r3". When the insn has |
6096 | been scheduled, the CFG is as follows: |
6097 | |
6098 | r1 := r3 r1 := r3 |
6099 | bookkeeping block scheduling fence |
6100 | \ / |
6101 | \ join / |
6102 | ---------- |
6103 | | | |
6104 | ---------- |
6105 | / \ |
6106 | / \ |
6107 | r1 := r2 |
6108 | |
6109 | Here, insn "r1 := r3" was scheduled at the current scheduling point |
6110 | and bookkeeping code was generated at the bookeeping block. This |
6111 | way insn "r1 := r2" is no longer available as a whole instruction |
6112 | (but only as expr) ahead of insn "r1 := r3" in bookkeeping block. |
6113 | This situation is handled by calling update_data_sets. |
6114 | |
6115 | Since update_data_sets is called only on the bookkeeping block, and |
6116 | it also may have predecessors with av_sets, containing instructions that |
6117 | are no longer available, we save all such expressions that become |
6118 | unavailable during data sets update on the bookkeeping block in |
6119 | VEC_BOOKKEEPING_BLOCKED_VINSNS. Later we avoid selecting such |
6120 | expressions for scheduling. This allows us to avoid recomputation of |
6121 | av_sets outside the code motion path. */ |
6122 | |
6123 | if (book_block) |
6124 | update_and_record_unavailable_insns (book_block); |
6125 | |
6126 | /* If INSN was previously marked for deletion, it's time to do it. */ |
6127 | if (lparams->removed_last_insn) |
6128 | insn = PREV_INSN (insn); |
6129 | |
6130 | /* Do not tidy control flow at the topmost moveop, as we can erroneously |
6131 | kill a block with a single nop in which the insn should be emitted. */ |
6132 | if (lparams->e1) |
6133 | tidy_control_flow (BLOCK_FOR_INSN (insn), true); |
6134 | } |
6135 | |
6136 | /* This function is called on the ascending pass, before returning from the |
6137 | current basic block. */ |
6138 | static void |
6139 | fur_at_first_insn (insn_t insn, |
6140 | cmpd_local_params_p lparams ATTRIBUTE_UNUSED, |
6141 | void *static_params ATTRIBUTE_UNUSED) |
6142 | { |
6143 | gcc_assert (!sel_bb_head_p (insn) || AV_SET_VALID_P (insn) |
6144 | || AV_LEVEL (insn) == -1); |
6145 | } |
6146 | |
6147 | /* Called on the backward stage of recursion to call moveup_expr for insn |
6148 | and sparams->c_expr. */ |
6149 | static void |
6150 | move_op_ascend (insn_t insn, void *static_params) |
6151 | { |
6152 | enum MOVEUP_EXPR_CODE res; |
6153 | moveop_static_params_p sparams = (moveop_static_params_p) static_params; |
6154 | |
6155 | if (! INSN_NOP_P (insn)) |
6156 | { |
6157 | res = moveup_expr_cached (expr: sparams->c_expr, insn, inside_insn_group: false); |
6158 | gcc_assert (res != MOVEUP_EXPR_NULL); |
6159 | } |
6160 | |
6161 | /* Update liveness for this insn as it was invalidated. */ |
6162 | update_liveness_on_insn (insn); |
6163 | } |
6164 | |
6165 | /* This function is called on enter to the basic block. |
6166 | Returns TRUE if this block already have been visited and |
6167 | code_motion_path_driver should return 1, FALSE otherwise. */ |
6168 | static int |
6169 | fur_on_enter (insn_t insn ATTRIBUTE_UNUSED, cmpd_local_params_p local_params, |
6170 | void *static_params, bool visited_p) |
6171 | { |
6172 | fur_static_params_p sparams = (fur_static_params_p) static_params; |
6173 | |
6174 | if (visited_p) |
6175 | { |
6176 | /* If we have found something below this block, there should be at |
6177 | least one insn in ORIGINAL_INSNS. */ |
6178 | gcc_assert (*sparams->original_insns); |
6179 | |
6180 | /* Adjust CROSSED_CALL_ABIS, since we may have come to this block along |
6181 | different path. */ |
6182 | DEF_LIST_DEF (*sparams->original_insns)->crossed_call_abis |
6183 | |= sparams->crossed_call_abis; |
6184 | } |
6185 | else |
6186 | local_params->old_original_insns = *sparams->original_insns; |
6187 | |
6188 | return 1; |
6189 | } |
6190 | |
6191 | /* Same as above but for move_op. */ |
6192 | static int |
6193 | move_op_on_enter (insn_t insn ATTRIBUTE_UNUSED, |
6194 | cmpd_local_params_p local_params ATTRIBUTE_UNUSED, |
6195 | void *static_params ATTRIBUTE_UNUSED, bool visited_p) |
6196 | { |
6197 | if (visited_p) |
6198 | return -1; |
6199 | return 1; |
6200 | } |
6201 | |
6202 | /* This function is called while descending current basic block if current |
6203 | insn is not the original EXPR we're searching for. |
6204 | |
6205 | Return value: FALSE, if code_motion_path_driver should perform a local |
6206 | cleanup and return 0 itself; |
6207 | TRUE, if code_motion_path_driver should continue. */ |
6208 | static bool |
6209 | move_op_orig_expr_not_found (insn_t insn, av_set_t orig_ops ATTRIBUTE_UNUSED, |
6210 | void *static_params) |
6211 | { |
6212 | moveop_static_params_p sparams = (moveop_static_params_p) static_params; |
6213 | |
6214 | sparams->failed_insn = insn; |
6215 | |
6216 | /* If we're scheduling separate expr, in order to generate correct code |
6217 | we need to stop the search at bookkeeping code generated with the |
6218 | same destination register or memory. */ |
6219 | if (lhs_of_insn_equals_to_dest_p (insn, sparams->dest)) |
6220 | return false; |
6221 | return true; |
6222 | } |
6223 | |
6224 | /* This function is called while descending current basic block if current |
6225 | insn is not the original EXPR we're searching for. |
6226 | |
6227 | Return value: TRUE (code_motion_path_driver should continue). */ |
6228 | static bool |
6229 | fur_orig_expr_not_found (insn_t insn, av_set_t orig_ops, void *static_params) |
6230 | { |
6231 | bool mutexed; |
6232 | expr_t r; |
6233 | av_set_iterator avi; |
6234 | fur_static_params_p sparams = (fur_static_params_p) static_params; |
6235 | |
6236 | if (CALL_P (insn)) |
6237 | sparams->crossed_call_abis |= 1 << insn_callee_abi (insn).id (); |
6238 | else if (DEBUG_INSN_P (insn)) |
6239 | return true; |
6240 | |
6241 | /* If current insn we are looking at cannot be executed together |
6242 | with original insn, then we can skip it safely. |
6243 | |
6244 | Example: ORIG_OPS = { (p6) r14 = sign_extend (r15); } |
6245 | INSN = (!p6) r14 = r14 + 1; |
6246 | |
6247 | Here we can schedule ORIG_OP with lhs = r14, though only |
6248 | looking at the set of used and set registers of INSN we must |
6249 | forbid it. So, add set/used in INSN registers to the |
6250 | untouchable set only if there is an insn in ORIG_OPS that can |
6251 | affect INSN. */ |
6252 | mutexed = true; |
6253 | FOR_EACH_EXPR (r, avi, orig_ops) |
6254 | if (!sched_insns_conditions_mutex_p (insn, EXPR_INSN_RTX (r))) |
6255 | { |
6256 | mutexed = false; |
6257 | break; |
6258 | } |
6259 | |
6260 | /* Mark all registers that do not meet the following condition: |
6261 | (1) Not set or read on any path from xi to an instance of the |
6262 | original operation. */ |
6263 | if (!mutexed) |
6264 | { |
6265 | IOR_REG_SET (sparams->used_regs, INSN_REG_SETS (insn)); |
6266 | IOR_REG_SET (sparams->used_regs, INSN_REG_USES (insn)); |
6267 | IOR_REG_SET (sparams->used_regs, INSN_REG_CLOBBERS (insn)); |
6268 | } |
6269 | |
6270 | return true; |
6271 | } |
6272 | |
6273 | /* Hooks and data to perform move_op operations with code_motion_path_driver. */ |
6274 | struct code_motion_path_driver_info_def move_op_hooks = { |
6275 | .on_enter: move_op_on_enter, |
6276 | .orig_expr_found: move_op_orig_expr_found, |
6277 | .orig_expr_not_found: move_op_orig_expr_not_found, |
6278 | .merge_succs: move_op_merge_succs, |
6279 | .after_merge_succs: move_op_after_merge_succs, |
6280 | .ascend: move_op_ascend, |
6281 | .at_first_insn: move_op_at_first_insn, |
6282 | SUCCS_NORMAL, |
6283 | .routine_name: "move_op" |
6284 | }; |
6285 | |
6286 | /* Hooks and data to perform find_used_regs operations |
6287 | with code_motion_path_driver. */ |
6288 | struct code_motion_path_driver_info_def fur_hooks = { |
6289 | .on_enter: fur_on_enter, |
6290 | .orig_expr_found: fur_orig_expr_found, |
6291 | .orig_expr_not_found: fur_orig_expr_not_found, |
6292 | .merge_succs: fur_merge_succs, |
6293 | NULL, /* fur_after_merge_succs */ |
6294 | NULL, /* fur_ascend */ |
6295 | .at_first_insn: fur_at_first_insn, |
6296 | SUCCS_ALL, |
6297 | .routine_name: "find_used_regs" |
6298 | }; |
6299 | |
6300 | /* Traverse all successors of INSN. For each successor that is SUCCS_NORMAL |
6301 | code_motion_path_driver is called recursively. Original operation |
6302 | was found at least on one path that is starting with one of INSN's |
6303 | successors (this fact is asserted). ORIG_OPS is expressions we're looking |
6304 | for, PATH is the path we've traversed, STATIC_PARAMS is the parameters |
6305 | of either move_op or find_used_regs depending on the caller. |
6306 | |
6307 | Return 0 if we haven't found expression, 1 if we found it, -1 if we don't |
6308 | know for sure at this point. */ |
6309 | static int |
6310 | code_motion_process_successors (insn_t insn, av_set_t orig_ops, |
6311 | ilist_t path, void *static_params) |
6312 | { |
6313 | int res = 0; |
6314 | succ_iterator succ_i; |
6315 | insn_t succ; |
6316 | basic_block bb; |
6317 | int old_index; |
6318 | unsigned old_succs; |
6319 | |
6320 | struct cmpd_local_params lparams; |
6321 | expr_def _x; |
6322 | |
6323 | lparams.c_expr_local = &_x; |
6324 | lparams.c_expr_merged = NULL; |
6325 | |
6326 | /* We need to process only NORMAL succs for move_op, and collect live |
6327 | registers from ALL branches (including those leading out of the |
6328 | region) for find_used_regs. |
6329 | |
6330 | In move_op, there can be a case when insn's bb number has changed |
6331 | due to created bookkeeping. This happens very rare, as we need to |
6332 | move expression from the beginning to the end of the same block. |
6333 | Rescan successors in this case. */ |
6334 | |
6335 | rescan: |
6336 | bb = BLOCK_FOR_INSN (insn); |
6337 | old_index = bb->index; |
6338 | old_succs = EDGE_COUNT (bb->succs); |
6339 | |
6340 | FOR_EACH_SUCC_1 (succ, succ_i, insn, code_motion_path_driver_info->succ_flags) |
6341 | { |
6342 | int b; |
6343 | |
6344 | lparams.e1 = succ_i.e1; |
6345 | lparams.e2 = succ_i.e2; |
6346 | |
6347 | /* Go deep into recursion only for NORMAL edges (non-backedges within the |
6348 | current region). */ |
6349 | if (succ_i.current_flags == SUCCS_NORMAL) |
6350 | b = code_motion_path_driver (succ, orig_ops, path, &lparams, |
6351 | static_params); |
6352 | else |
6353 | b = 0; |
6354 | |
6355 | /* Merge c_expres found or unify live register sets from different |
6356 | successors. */ |
6357 | code_motion_path_driver_info->merge_succs (insn, succ, b, &lparams, |
6358 | static_params); |
6359 | if (b == 1) |
6360 | res = b; |
6361 | else if (b == -1 && res != 1) |
6362 | res = b; |
6363 | |
6364 | /* We have simplified the control flow below this point. In this case, |
6365 | the iterator becomes invalid. We need to try again. |
6366 | If we have removed the insn itself, it could be only an |
6367 | unconditional jump. Thus, do not rescan but break immediately -- |
6368 | we have already visited the only successor block. */ |
6369 | if (!BLOCK_FOR_INSN (insn)) |
6370 | { |
6371 | if (sched_verbose >= 6) |
6372 | sel_print (fmt: "Not doing rescan: already visited the only successor" |
6373 | " of block %d\n" , old_index); |
6374 | break; |
6375 | } |
6376 | if (BLOCK_FOR_INSN (insn)->index != old_index |
6377 | || EDGE_COUNT (bb->succs) != old_succs) |
6378 | { |
6379 | if (sched_verbose >= 6) |
6380 | sel_print (fmt: "Rescan: CFG was simplified below insn %d, block %d\n" , |
6381 | INSN_UID (insn), BLOCK_FOR_INSN (insn)->index); |
6382 | insn = sel_bb_end (BLOCK_FOR_INSN (insn)); |
6383 | goto rescan; |
6384 | } |
6385 | } |
6386 | |
6387 | /* Here, RES==1 if original expr was found at least for one of the |
6388 | successors. After the loop, RES may happen to have zero value |
6389 | only if at some point the expr searched is present in av_set, but is |
6390 | not found below. In most cases, this situation is an error. |
6391 | The exception is when the original operation is blocked by |
6392 | bookkeeping generated for another fence or for another path in current |
6393 | move_op. */ |
6394 | gcc_checking_assert (res == 1 |
6395 | || (res == 0 |
6396 | && av_set_could_be_blocked_by_bookkeeping_p (orig_ops, static_params)) |
6397 | || res == -1); |
6398 | |
6399 | /* Merge data, clean up, etc. */ |
6400 | if (res != -1 && code_motion_path_driver_info->after_merge_succs) |
6401 | code_motion_path_driver_info->after_merge_succs (&lparams, static_params); |
6402 | |
6403 | return res; |
6404 | } |
6405 | |
6406 | |
6407 | /* Perform a cleanup when the driver is about to terminate. ORIG_OPS_P |
6408 | is the pointer to the av set with expressions we were looking for, |
6409 | PATH_P is the pointer to the traversed path. */ |
6410 | static inline void |
6411 | code_motion_path_driver_cleanup (av_set_t *orig_ops_p, ilist_t *path_p) |
6412 | { |
6413 | ilist_remove (path_p); |
6414 | av_set_clear (orig_ops_p); |
6415 | } |
6416 | |
6417 | /* The driver function that implements move_op or find_used_regs |
6418 | functionality dependent whether code_motion_path_driver_INFO is set to |
6419 | &MOVE_OP_HOOKS or &FUR_HOOKS. This function implements the common parts |
6420 | of code (CFG traversal etc) that are shared among both functions. INSN |
6421 | is the insn we're starting the search from, ORIG_OPS are the expressions |
6422 | we're searching for, PATH is traversed path, LOCAL_PARAMS_IN are local |
6423 | parameters of the driver, and STATIC_PARAMS are static parameters of |
6424 | the caller. |
6425 | |
6426 | Returns whether original instructions were found. Note that top-level |
6427 | code_motion_path_driver always returns true. */ |
6428 | static int |
6429 | code_motion_path_driver (insn_t insn, av_set_t orig_ops, ilist_t path, |
6430 | cmpd_local_params_p local_params_in, |
6431 | void *static_params) |
6432 | { |
6433 | expr_t expr = NULL; |
6434 | basic_block bb = BLOCK_FOR_INSN (insn); |
6435 | insn_t first_insn, original_insn, bb_tail, before_first; |
6436 | bool removed_last_insn = false; |
6437 | |
6438 | if (sched_verbose >= 6) |
6439 | { |
6440 | sel_print (fmt: "%s (" , code_motion_path_driver_info->routine_name); |
6441 | dump_insn (insn); |
6442 | sel_print (fmt: "," ); |
6443 | dump_av_set (orig_ops); |
6444 | sel_print (fmt: ")\n" ); |
6445 | } |
6446 | |
6447 | gcc_assert (orig_ops); |
6448 | |
6449 | /* If no original operations exist below this insn, return immediately. */ |
6450 | if (is_ineligible_successor (insn, p: path)) |
6451 | { |
6452 | if (sched_verbose >= 6) |
6453 | sel_print (fmt: "Insn %d is ineligible successor\n" , INSN_UID (insn)); |
6454 | return false; |
6455 | } |
6456 | |
6457 | /* The block can have invalid av set, in which case it was created earlier |
6458 | during move_op. Return immediately. */ |
6459 | if (sel_bb_head_p (insn)) |
6460 | { |
6461 | if (! AV_SET_VALID_P (insn)) |
6462 | { |
6463 | if (sched_verbose >= 6) |
6464 | sel_print (fmt: "Returned from block %d as it had invalid av set\n" , |
6465 | bb->index); |
6466 | return false; |
6467 | } |
6468 | |
6469 | if (bitmap_bit_p (code_motion_visited_blocks, bb->index)) |
6470 | { |
6471 | /* We have already found an original operation on this branch, do not |
6472 | go any further and just return TRUE here. If we don't stop here, |
6473 | function can have exponential behavior even on the small code |
6474 | with many different paths (e.g. with data speculation and |
6475 | recovery blocks). */ |
6476 | if (sched_verbose >= 6) |
6477 | sel_print (fmt: "Block %d already visited in this traversal\n" , bb->index); |
6478 | if (code_motion_path_driver_info->on_enter) |
6479 | return code_motion_path_driver_info->on_enter (insn, |
6480 | local_params_in, |
6481 | static_params, |
6482 | true); |
6483 | } |
6484 | } |
6485 | |
6486 | if (code_motion_path_driver_info->on_enter) |
6487 | code_motion_path_driver_info->on_enter (insn, local_params_in, |
6488 | static_params, false); |
6489 | orig_ops = av_set_copy (orig_ops); |
6490 | |
6491 | /* Filter the orig_ops set. */ |
6492 | if (AV_SET_VALID_P (insn)) |
6493 | av_set_code_motion_filter (&orig_ops, AV_SET (insn)); |
6494 | |
6495 | /* If no more original ops, return immediately. */ |
6496 | if (!orig_ops) |
6497 | { |
6498 | if (sched_verbose >= 6) |
6499 | sel_print (fmt: "No intersection with av set of block %d\n" , bb->index); |
6500 | return false; |
6501 | } |
6502 | |
6503 | /* For non-speculative insns we have to leave only one form of the |
6504 | original operation, because if we don't, we may end up with |
6505 | different C_EXPRes and, consequently, with bookkeepings for different |
6506 | expression forms along the same code motion path. That may lead to |
6507 | generation of incorrect code. So for each code motion we stick to |
6508 | the single form of the instruction, except for speculative insns |
6509 | which we need to keep in different forms with all speculation |
6510 | types. */ |
6511 | av_set_leave_one_nonspec (&orig_ops); |
6512 | |
6513 | /* It is not possible that all ORIG_OPS are filtered out. */ |
6514 | gcc_assert (orig_ops); |
6515 | |
6516 | /* It is enough to place only heads and tails of visited basic blocks into |
6517 | the PATH. */ |
6518 | ilist_add (lp: &path, insn); |
6519 | first_insn = original_insn = insn; |
6520 | bb_tail = sel_bb_end (bb); |
6521 | |
6522 | /* Descend the basic block in search of the original expr; this part |
6523 | corresponds to the part of the original move_op procedure executed |
6524 | before the recursive call. */ |
6525 | for (;;) |
6526 | { |
6527 | /* Look at the insn and decide if it could be an ancestor of currently |
6528 | scheduling operation. If it is so, then the insn "dest = op" could |
6529 | either be replaced with "dest = reg", because REG now holds the result |
6530 | of OP, or just removed, if we've scheduled the insn as a whole. |
6531 | |
6532 | If this insn doesn't contain currently scheduling OP, then proceed |
6533 | with searching and look at its successors. Operations we're searching |
6534 | for could have changed when moving up through this insn via |
6535 | substituting. In this case, perform unsubstitution on them first. |
6536 | |
6537 | When traversing the DAG below this insn is finished, insert |
6538 | bookkeeping code, if the insn is a joint point, and remove |
6539 | leftovers. */ |
6540 | |
6541 | expr = av_set_lookup (orig_ops, INSN_VINSN (insn)); |
6542 | if (expr) |
6543 | { |
6544 | insn_t last_insn = PREV_INSN (insn); |
6545 | |
6546 | /* We have found the original operation. */ |
6547 | if (sched_verbose >= 6) |
6548 | sel_print (fmt: "Found original operation at insn %d\n" , INSN_UID (insn)); |
6549 | |
6550 | code_motion_path_driver_info->orig_expr_found |
6551 | (insn, expr, local_params_in, static_params); |
6552 | |
6553 | /* Step back, so on the way back we'll start traversing from the |
6554 | previous insn (or we'll see that it's bb_note and skip that |
6555 | loop). */ |
6556 | if (insn == first_insn) |
6557 | { |
6558 | first_insn = NEXT_INSN (insn: last_insn); |
6559 | removed_last_insn = sel_bb_end_p (last_insn); |
6560 | } |
6561 | insn = last_insn; |
6562 | break; |
6563 | } |
6564 | else |
6565 | { |
6566 | /* We haven't found the original expr, continue descending the basic |
6567 | block. */ |
6568 | if (code_motion_path_driver_info->orig_expr_not_found |
6569 | (insn, orig_ops, static_params)) |
6570 | { |
6571 | /* Av set ops could have been changed when moving through this |
6572 | insn. To find them below it, we have to un-substitute them. */ |
6573 | undo_transformations (av_ptr: &orig_ops, insn); |
6574 | } |
6575 | else |
6576 | { |
6577 | /* Clean up and return, if the hook tells us to do so. It may |
6578 | happen if we've encountered the previously created |
6579 | bookkeeping. */ |
6580 | code_motion_path_driver_cleanup (orig_ops_p: &orig_ops, path_p: &path); |
6581 | return -1; |
6582 | } |
6583 | |
6584 | gcc_assert (orig_ops); |
6585 | } |
6586 | |
6587 | /* Stop at insn if we got to the end of BB. */ |
6588 | if (insn == bb_tail) |
6589 | break; |
6590 | |
6591 | insn = NEXT_INSN (insn); |
6592 | } |
6593 | |
6594 | /* Here INSN either points to the insn before the original insn (may be |
6595 | bb_note, if original insn was a bb_head) or to the bb_end. */ |
6596 | if (!expr) |
6597 | { |
6598 | int res; |
6599 | rtx_insn *last_insn = PREV_INSN (insn); |
6600 | bool added_to_path; |
6601 | |
6602 | gcc_assert (insn == sel_bb_end (bb)); |
6603 | |
6604 | /* Add bb tail to PATH (but it doesn't make any sense if it's a bb_head - |
6605 | it's already in PATH then). */ |
6606 | if (insn != first_insn) |
6607 | { |
6608 | ilist_add (lp: &path, insn); |
6609 | added_to_path = true; |
6610 | } |
6611 | else |
6612 | added_to_path = false; |
6613 | |
6614 | /* Process_successors should be able to find at least one |
6615 | successor for which code_motion_path_driver returns TRUE. */ |
6616 | res = code_motion_process_successors (insn, orig_ops, |
6617 | path, static_params); |
6618 | |
6619 | /* Jump in the end of basic block could have been removed or replaced |
6620 | during code_motion_process_successors, so recompute insn as the |
6621 | last insn in bb. */ |
6622 | if (NEXT_INSN (insn: last_insn) != insn) |
6623 | { |
6624 | insn = sel_bb_end (bb); |
6625 | first_insn = sel_bb_head (bb); |
6626 | if (first_insn != original_insn) |
6627 | first_insn = original_insn; |
6628 | } |
6629 | |
6630 | /* Remove bb tail from path. */ |
6631 | if (added_to_path) |
6632 | ilist_remove (&path); |
6633 | |
6634 | if (res != 1) |
6635 | { |
6636 | /* This is the case when one of the original expr is no longer available |
6637 | due to bookkeeping created on this branch with the same register. |
6638 | In the original algorithm, which doesn't have update_data_sets call |
6639 | on a bookkeeping block, it would simply result in returning |
6640 | FALSE when we've encountered a previously generated bookkeeping |
6641 | insn in moveop_orig_expr_not_found. */ |
6642 | code_motion_path_driver_cleanup (orig_ops_p: &orig_ops, path_p: &path); |
6643 | return res; |
6644 | } |
6645 | } |
6646 | |
6647 | /* Don't need it any more. */ |
6648 | av_set_clear (&orig_ops); |
6649 | |
6650 | /* Backward pass: now, when we have C_EXPR computed, we'll drag it to |
6651 | the beginning of the basic block. */ |
6652 | before_first = PREV_INSN (insn: first_insn); |
6653 | while (insn != before_first) |
6654 | { |
6655 | if (code_motion_path_driver_info->ascend) |
6656 | code_motion_path_driver_info->ascend (insn, static_params); |
6657 | |
6658 | insn = PREV_INSN (insn); |
6659 | } |
6660 | |
6661 | /* Now we're at the bb head. */ |
6662 | insn = first_insn; |
6663 | ilist_remove (&path); |
6664 | local_params_in->removed_last_insn = removed_last_insn; |
6665 | code_motion_path_driver_info->at_first_insn (insn, local_params_in, static_params); |
6666 | |
6667 | /* This should be the very last operation as at bb head we could change |
6668 | the numbering by creating bookkeeping blocks. */ |
6669 | if (removed_last_insn) |
6670 | insn = PREV_INSN (insn); |
6671 | |
6672 | /* If we have simplified the control flow and removed the first jump insn, |
6673 | there's no point in marking this block in the visited blocks bitmap. */ |
6674 | if (BLOCK_FOR_INSN (insn)) |
6675 | bitmap_set_bit (code_motion_visited_blocks, BLOCK_FOR_INSN (insn)->index); |
6676 | return true; |
6677 | } |
6678 | |
6679 | /* Move up the operations from ORIG_OPS set traversing the dag starting |
6680 | from INSN. PATH represents the edges traversed so far. |
6681 | DEST is the register chosen for scheduling the current expr. Insert |
6682 | bookkeeping code in the join points. EXPR_VLIW is the chosen expression, |
6683 | C_EXPR is how it looks like at the given cfg point. |
6684 | Set *SHOULD_MOVE to indicate whether we have only disconnected |
6685 | one of the insns found. |
6686 | |
6687 | Returns whether original instructions were found, which is asserted |
6688 | to be true in the caller. */ |
6689 | static bool |
6690 | move_op (insn_t insn, av_set_t orig_ops, expr_t expr_vliw, |
6691 | rtx dest, expr_t c_expr, bool *should_move) |
6692 | { |
6693 | struct moveop_static_params sparams; |
6694 | struct cmpd_local_params lparams; |
6695 | int res; |
6696 | |
6697 | /* Init params for code_motion_path_driver. */ |
6698 | sparams.dest = dest; |
6699 | sparams.c_expr = c_expr; |
6700 | sparams.uid = INSN_UID (EXPR_INSN_RTX (expr_vliw)); |
6701 | sparams.failed_insn = NULL; |
6702 | sparams.was_renamed = false; |
6703 | lparams.e1 = NULL; |
6704 | |
6705 | /* We haven't visited any blocks yet. */ |
6706 | bitmap_clear (code_motion_visited_blocks); |
6707 | |
6708 | /* Set appropriate hooks and data. */ |
6709 | code_motion_path_driver_info = &move_op_hooks; |
6710 | res = code_motion_path_driver (insn, orig_ops, NULL, local_params_in: &lparams, static_params: &sparams); |
6711 | |
6712 | gcc_assert (res != -1); |
6713 | |
6714 | if (sparams.was_renamed) |
6715 | EXPR_WAS_RENAMED (expr_vliw) = true; |
6716 | |
6717 | *should_move = (sparams.uid == -1); |
6718 | |
6719 | return res; |
6720 | } |
6721 | |
6722 | |
6723 | /* Functions that work with regions. */ |
6724 | |
6725 | /* Current number of seqno used in init_seqno and init_seqno_1. */ |
6726 | static int cur_seqno; |
6727 | |
6728 | /* A helper for init_seqno. Traverse the region starting from BB and |
6729 | compute seqnos for visited insns, marking visited bbs in VISITED_BBS. |
6730 | Clear visited blocks from BLOCKS_TO_RESCHEDULE. */ |
6731 | static void |
6732 | init_seqno_1 (basic_block bb, sbitmap visited_bbs, bitmap blocks_to_reschedule) |
6733 | { |
6734 | int bbi = BLOCK_TO_BB (bb->index); |
6735 | insn_t insn; |
6736 | insn_t succ_insn; |
6737 | succ_iterator si; |
6738 | |
6739 | rtx_note *note = bb_note (bb); |
6740 | bitmap_set_bit (map: visited_bbs, bitno: bbi); |
6741 | if (blocks_to_reschedule) |
6742 | bitmap_clear_bit (blocks_to_reschedule, bb->index); |
6743 | |
6744 | FOR_EACH_SUCC_1 (succ_insn, si, BB_END (bb), |
6745 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
6746 | { |
6747 | basic_block succ = BLOCK_FOR_INSN (insn: succ_insn); |
6748 | int succ_bbi = BLOCK_TO_BB (succ->index); |
6749 | |
6750 | gcc_assert (in_current_region_p (succ)); |
6751 | |
6752 | if (!bitmap_bit_p (map: visited_bbs, bitno: succ_bbi)) |
6753 | { |
6754 | gcc_assert (succ_bbi > bbi); |
6755 | |
6756 | init_seqno_1 (bb: succ, visited_bbs, blocks_to_reschedule); |
6757 | } |
6758 | else if (blocks_to_reschedule) |
6759 | bitmap_set_bit (forced_ebb_heads, succ->index); |
6760 | } |
6761 | |
6762 | for (insn = BB_END (bb); insn != note; insn = PREV_INSN (insn)) |
6763 | INSN_SEQNO (insn) = cur_seqno--; |
6764 | } |
6765 | |
6766 | /* Initialize seqnos for the current region. BLOCKS_TO_RESCHEDULE contains |
6767 | blocks on which we're rescheduling when pipelining, FROM is the block where |
6768 | traversing region begins (it may not be the head of the region when |
6769 | pipelining, but the head of the loop instead). |
6770 | |
6771 | Returns the maximal seqno found. */ |
6772 | static int |
6773 | init_seqno (bitmap blocks_to_reschedule, basic_block from) |
6774 | { |
6775 | bitmap_iterator bi; |
6776 | unsigned bbi; |
6777 | |
6778 | auto_sbitmap visited_bbs (current_nr_blocks); |
6779 | |
6780 | if (blocks_to_reschedule) |
6781 | { |
6782 | bitmap_ones (visited_bbs); |
6783 | EXECUTE_IF_SET_IN_BITMAP (blocks_to_reschedule, 0, bbi, bi) |
6784 | { |
6785 | gcc_assert (BLOCK_TO_BB (bbi) < current_nr_blocks); |
6786 | bitmap_clear_bit (map: visited_bbs, BLOCK_TO_BB (bbi)); |
6787 | } |
6788 | } |
6789 | else |
6790 | { |
6791 | bitmap_clear (visited_bbs); |
6792 | from = EBB_FIRST_BB (0); |
6793 | } |
6794 | |
6795 | cur_seqno = sched_max_luid - 1; |
6796 | init_seqno_1 (bb: from, visited_bbs, blocks_to_reschedule); |
6797 | |
6798 | /* cur_seqno may be positive if the number of instructions is less than |
6799 | sched_max_luid - 1 (when rescheduling or if some instructions have been |
6800 | removed by the call to purge_empty_blocks in sel_sched_region_1). */ |
6801 | gcc_assert (cur_seqno >= 0); |
6802 | |
6803 | return sched_max_luid - 1; |
6804 | } |
6805 | |
6806 | /* Initialize scheduling parameters for current region. */ |
6807 | static void |
6808 | sel_setup_region_sched_flags (void) |
6809 | { |
6810 | enable_schedule_as_rhs_p = 1; |
6811 | bookkeeping_p = 1; |
6812 | pipelining_p = (bookkeeping_p |
6813 | && (flag_sel_sched_pipelining != 0) |
6814 | && current_loop_nest != NULL |
6815 | && loop_has_exit_edges (loop: current_loop_nest)); |
6816 | max_insns_to_rename = param_selsched_insns_to_rename; |
6817 | max_ws = MAX_WS; |
6818 | } |
6819 | |
6820 | /* Return true if all basic blocks of current region are empty. */ |
6821 | static bool |
6822 | current_region_empty_p (void) |
6823 | { |
6824 | int i; |
6825 | for (i = 0; i < current_nr_blocks; i++) |
6826 | if (! sel_bb_empty_p (BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i)))) |
6827 | return false; |
6828 | |
6829 | return true; |
6830 | } |
6831 | |
6832 | /* Prepare and verify loop nest for pipelining. */ |
6833 | static void |
6834 | setup_current_loop_nest (int rgn, bb_vec_t *bbs) |
6835 | { |
6836 | current_loop_nest = get_loop_nest_for_rgn (rgn); |
6837 | |
6838 | if (!current_loop_nest) |
6839 | return; |
6840 | |
6841 | /* If this loop has any saved loop preheaders from nested loops, |
6842 | add these basic blocks to the current region. */ |
6843 | sel_add_loop_preheaders (bbs); |
6844 | |
6845 | /* Check that we're starting with a valid information. */ |
6846 | gcc_assert (loop_latch_edge (current_loop_nest)); |
6847 | gcc_assert (LOOP_MARKED_FOR_PIPELINING_P (current_loop_nest)); |
6848 | } |
6849 | |
6850 | /* Compute instruction priorities for current region. */ |
6851 | static void |
6852 | sel_compute_priorities (int rgn) |
6853 | { |
6854 | sched_rgn_compute_dependencies (rgn); |
6855 | |
6856 | /* Compute insn priorities in haifa style. Then free haifa style |
6857 | dependencies that we've calculated for this. */ |
6858 | compute_priorities (); |
6859 | |
6860 | if (sched_verbose >= 5) |
6861 | debug_rgn_dependencies (0); |
6862 | |
6863 | free_rgn_deps (); |
6864 | } |
6865 | |
6866 | /* Init scheduling data for RGN. Returns true when this region should not |
6867 | be scheduled. */ |
6868 | static bool |
6869 | sel_region_init (int rgn) |
6870 | { |
6871 | int i; |
6872 | bb_vec_t bbs; |
6873 | |
6874 | rgn_setup_region (rgn); |
6875 | |
6876 | /* Even if sched_is_disabled_for_current_region_p() is true, we still |
6877 | do region initialization here so the region can be bundled correctly, |
6878 | but we'll skip the scheduling in sel_sched_region (). */ |
6879 | if (current_region_empty_p ()) |
6880 | return true; |
6881 | |
6882 | bbs.create (nelems: current_nr_blocks); |
6883 | |
6884 | for (i = 0; i < current_nr_blocks; i++) |
6885 | bbs.quick_push (BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i))); |
6886 | |
6887 | sel_init_bbs (bbs); |
6888 | |
6889 | if (flag_sel_sched_pipelining) |
6890 | setup_current_loop_nest (rgn, bbs: &bbs); |
6891 | |
6892 | sel_setup_region_sched_flags (); |
6893 | |
6894 | /* Initialize luids and dependence analysis which both sel-sched and haifa |
6895 | need. */ |
6896 | sched_init_luids (bbs); |
6897 | sched_deps_init (false); |
6898 | |
6899 | /* Initialize haifa data. */ |
6900 | rgn_setup_sched_infos (); |
6901 | sel_set_sched_flags (); |
6902 | haifa_init_h_i_d (bbs); |
6903 | |
6904 | sel_compute_priorities (rgn); |
6905 | init_deps_global (); |
6906 | |
6907 | /* Main initialization. */ |
6908 | sel_setup_sched_infos (); |
6909 | sel_init_global_and_expr (bbs); |
6910 | |
6911 | bbs.release (); |
6912 | |
6913 | blocks_to_reschedule = BITMAP_ALLOC (NULL); |
6914 | |
6915 | /* Init correct liveness sets on each instruction of a single-block loop. |
6916 | This is the only situation when we can't update liveness when calling |
6917 | compute_live for the first insn of the loop. */ |
6918 | if (current_loop_nest) |
6919 | { |
6920 | int = |
6921 | (sel_is_loop_preheader_p (BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (0))) |
6922 | ? 1 |
6923 | : 0); |
6924 | |
6925 | if (current_nr_blocks == header + 1) |
6926 | update_liveness_on_insn |
6927 | (insn: sel_bb_head (BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (header)))); |
6928 | } |
6929 | |
6930 | /* Set hooks so that no newly generated insn will go out unnoticed. */ |
6931 | sel_register_cfg_hooks (); |
6932 | |
6933 | /* !!! We call target.sched.init () for the whole region, but we invoke |
6934 | targetm.sched.finish () for every ebb. */ |
6935 | if (targetm.sched.init) |
6936 | /* None of the arguments are actually used in any target. */ |
6937 | targetm.sched.init (sched_dump, sched_verbose, -1); |
6938 | |
6939 | first_emitted_uid = get_max_uid () + 1; |
6940 | preheader_removed = false; |
6941 | |
6942 | /* Reset register allocation ticks array. */ |
6943 | memset (s: reg_rename_tick, c: 0, n: sizeof reg_rename_tick); |
6944 | reg_rename_this_tick = 0; |
6945 | |
6946 | forced_ebb_heads = BITMAP_ALLOC (NULL); |
6947 | |
6948 | setup_nop_vinsn (); |
6949 | current_copies = BITMAP_ALLOC (NULL); |
6950 | current_originators = BITMAP_ALLOC (NULL); |
6951 | code_motion_visited_blocks = BITMAP_ALLOC (NULL); |
6952 | |
6953 | return false; |
6954 | } |
6955 | |
6956 | /* Simplify insns after the scheduling. */ |
6957 | static void |
6958 | simplify_changed_insns (void) |
6959 | { |
6960 | int i; |
6961 | |
6962 | for (i = 0; i < current_nr_blocks; i++) |
6963 | { |
6964 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i)); |
6965 | rtx_insn *insn; |
6966 | |
6967 | FOR_BB_INSNS (bb, insn) |
6968 | if (INSN_P (insn)) |
6969 | { |
6970 | expr_t expr = INSN_EXPR (insn); |
6971 | |
6972 | if (EXPR_WAS_SUBSTITUTED (expr)) |
6973 | validate_simplify_insn (insn); |
6974 | } |
6975 | } |
6976 | } |
6977 | |
6978 | /* Find boundaries of the EBB starting from basic block BB, marking blocks of |
6979 | this EBB in SCHEDULED_BLOCKS and appropriately filling in HEAD, TAIL, |
6980 | PREV_HEAD, and NEXT_TAIL fields of CURRENT_SCHED_INFO structure. */ |
6981 | static void |
6982 | find_ebb_boundaries (basic_block bb, bitmap scheduled_blocks) |
6983 | { |
6984 | rtx_insn *head, *tail; |
6985 | basic_block bb1 = bb; |
6986 | if (sched_verbose >= 2) |
6987 | sel_print (fmt: "Finishing schedule in bbs: " ); |
6988 | |
6989 | do |
6990 | { |
6991 | bitmap_set_bit (scheduled_blocks, BLOCK_TO_BB (bb1->index)); |
6992 | |
6993 | if (sched_verbose >= 2) |
6994 | sel_print (fmt: "%d; " , bb1->index); |
6995 | } |
6996 | while (!bb_ends_ebb_p (bb1) && (bb1 = bb_next_bb (bb: bb1))); |
6997 | |
6998 | if (sched_verbose >= 2) |
6999 | sel_print (fmt: "\n" ); |
7000 | |
7001 | get_ebb_head_tail (bb, bb1, &head, &tail); |
7002 | |
7003 | current_sched_info->head = head; |
7004 | current_sched_info->tail = tail; |
7005 | current_sched_info->prev_head = PREV_INSN (insn: head); |
7006 | current_sched_info->next_tail = NEXT_INSN (insn: tail); |
7007 | } |
7008 | |
7009 | /* Regenerate INSN_SCHED_CYCLEs for insns of current EBB. */ |
7010 | static void |
7011 | reset_sched_cycles_in_current_ebb (void) |
7012 | { |
7013 | int last_clock = 0; |
7014 | int haifa_last_clock = -1; |
7015 | int haifa_clock = 0; |
7016 | int issued_insns = 0; |
7017 | insn_t insn; |
7018 | |
7019 | if (targetm.sched.init) |
7020 | { |
7021 | /* None of the arguments are actually used in any target. |
7022 | NB: We should have md_reset () hook for cases like this. */ |
7023 | targetm.sched.init (sched_dump, sched_verbose, -1); |
7024 | } |
7025 | |
7026 | state_reset (curr_state); |
7027 | advance_state (curr_state); |
7028 | |
7029 | for (insn = current_sched_info->head; |
7030 | insn != current_sched_info->next_tail; |
7031 | insn = NEXT_INSN (insn)) |
7032 | { |
7033 | int cost, haifa_cost; |
7034 | int sort_p; |
7035 | bool asm_p, real_insn, after_stall, all_issued; |
7036 | int clock; |
7037 | |
7038 | if (!INSN_P (insn)) |
7039 | continue; |
7040 | |
7041 | asm_p = false; |
7042 | real_insn = recog_memoized (insn) >= 0; |
7043 | clock = INSN_SCHED_CYCLE (insn); |
7044 | |
7045 | cost = clock - last_clock; |
7046 | |
7047 | /* Initialize HAIFA_COST. */ |
7048 | if (! real_insn) |
7049 | { |
7050 | asm_p = INSN_ASM_P (insn); |
7051 | |
7052 | if (asm_p) |
7053 | /* This is asm insn which *had* to be scheduled first |
7054 | on the cycle. */ |
7055 | haifa_cost = 1; |
7056 | else |
7057 | /* This is a use/clobber insn. It should not change |
7058 | cost. */ |
7059 | haifa_cost = 0; |
7060 | } |
7061 | else |
7062 | haifa_cost = estimate_insn_cost (insn, state: curr_state); |
7063 | |
7064 | /* Stall for whatever cycles we've stalled before. */ |
7065 | after_stall = 0; |
7066 | if (INSN_AFTER_STALL_P (insn) && cost > haifa_cost) |
7067 | { |
7068 | haifa_cost = cost; |
7069 | after_stall = 1; |
7070 | } |
7071 | all_issued = issued_insns == issue_rate; |
7072 | if (haifa_cost == 0 && all_issued) |
7073 | haifa_cost = 1; |
7074 | if (haifa_cost > 0) |
7075 | { |
7076 | int i = 0; |
7077 | |
7078 | while (haifa_cost--) |
7079 | { |
7080 | advance_state (curr_state); |
7081 | issued_insns = 0; |
7082 | i++; |
7083 | |
7084 | if (sched_verbose >= 2) |
7085 | { |
7086 | sel_print (fmt: "advance_state (state_transition)\n" ); |
7087 | debug_state (state: curr_state); |
7088 | } |
7089 | |
7090 | /* The DFA may report that e.g. insn requires 2 cycles to be |
7091 | issued, but on the next cycle it says that insn is ready |
7092 | to go. Check this here. */ |
7093 | if (!after_stall |
7094 | && real_insn |
7095 | && haifa_cost > 0 |
7096 | && estimate_insn_cost (insn, state: curr_state) == 0) |
7097 | break; |
7098 | |
7099 | /* When the data dependency stall is longer than the DFA stall, |
7100 | and when we have issued exactly issue_rate insns and stalled, |
7101 | it could be that after this longer stall the insn will again |
7102 | become unavailable to the DFA restrictions. Looks strange |
7103 | but happens e.g. on x86-64. So recheck DFA on the last |
7104 | iteration. */ |
7105 | if ((after_stall || all_issued) |
7106 | && real_insn |
7107 | && haifa_cost == 0) |
7108 | haifa_cost = estimate_insn_cost (insn, state: curr_state); |
7109 | } |
7110 | |
7111 | haifa_clock += i; |
7112 | if (sched_verbose >= 2) |
7113 | sel_print (fmt: "haifa clock: %d\n" , haifa_clock); |
7114 | } |
7115 | else |
7116 | gcc_assert (haifa_cost == 0); |
7117 | |
7118 | if (sched_verbose >= 2) |
7119 | sel_print (fmt: "Haifa cost for insn %d: %d\n" , INSN_UID (insn), haifa_cost); |
7120 | |
7121 | if (targetm.sched.dfa_new_cycle) |
7122 | while (targetm.sched.dfa_new_cycle (sched_dump, sched_verbose, insn, |
7123 | haifa_last_clock, haifa_clock, |
7124 | &sort_p)) |
7125 | { |
7126 | advance_state (curr_state); |
7127 | issued_insns = 0; |
7128 | haifa_clock++; |
7129 | if (sched_verbose >= 2) |
7130 | { |
7131 | sel_print (fmt: "advance_state (dfa_new_cycle)\n" ); |
7132 | debug_state (state: curr_state); |
7133 | sel_print (fmt: "haifa clock: %d\n" , haifa_clock + 1); |
7134 | } |
7135 | } |
7136 | |
7137 | if (real_insn) |
7138 | { |
7139 | static state_t temp = NULL; |
7140 | |
7141 | if (!temp) |
7142 | temp = xmalloc (dfa_state_size); |
7143 | memcpy (dest: temp, src: curr_state, n: dfa_state_size); |
7144 | |
7145 | cost = state_transition (curr_state, insn); |
7146 | if (memcmp (s1: temp, s2: curr_state, n: dfa_state_size)) |
7147 | issued_insns++; |
7148 | |
7149 | if (sched_verbose >= 2) |
7150 | { |
7151 | sel_print (fmt: "scheduled insn %d, clock %d\n" , INSN_UID (insn), |
7152 | haifa_clock + 1); |
7153 | debug_state (state: curr_state); |
7154 | } |
7155 | gcc_assert (cost < 0); |
7156 | } |
7157 | |
7158 | if (targetm.sched.variable_issue) |
7159 | targetm.sched.variable_issue (sched_dump, sched_verbose, insn, 0); |
7160 | |
7161 | INSN_SCHED_CYCLE (insn) = haifa_clock; |
7162 | |
7163 | last_clock = clock; |
7164 | haifa_last_clock = haifa_clock; |
7165 | } |
7166 | } |
7167 | |
7168 | /* Put TImode markers on insns starting a new issue group. */ |
7169 | static void |
7170 | put_TImodes (void) |
7171 | { |
7172 | int last_clock = -1; |
7173 | insn_t insn; |
7174 | |
7175 | for (insn = current_sched_info->head; insn != current_sched_info->next_tail; |
7176 | insn = NEXT_INSN (insn)) |
7177 | { |
7178 | int cost, clock; |
7179 | |
7180 | if (!INSN_P (insn)) |
7181 | continue; |
7182 | |
7183 | clock = INSN_SCHED_CYCLE (insn); |
7184 | cost = (last_clock == -1) ? 1 : clock - last_clock; |
7185 | |
7186 | gcc_assert (cost >= 0); |
7187 | |
7188 | if (issue_rate > 1 |
7189 | && GET_CODE (PATTERN (insn)) != USE |
7190 | && GET_CODE (PATTERN (insn)) != CLOBBER) |
7191 | { |
7192 | if (reload_completed && cost > 0) |
7193 | PUT_MODE (x: insn, TImode); |
7194 | |
7195 | last_clock = clock; |
7196 | } |
7197 | |
7198 | if (sched_verbose >= 2) |
7199 | sel_print (fmt: "Cost for insn %d is %d\n" , INSN_UID (insn), cost); |
7200 | } |
7201 | } |
7202 | |
7203 | /* Perform MD_FINISH on EBBs comprising current region. When |
7204 | RESET_SCHED_CYCLES_P is true, run a pass emulating the scheduler |
7205 | to produce correct sched cycles on insns. */ |
7206 | static void |
7207 | sel_region_target_finish (bool reset_sched_cycles_p) |
7208 | { |
7209 | int i; |
7210 | bitmap scheduled_blocks = BITMAP_ALLOC (NULL); |
7211 | |
7212 | for (i = 0; i < current_nr_blocks; i++) |
7213 | { |
7214 | if (bitmap_bit_p (scheduled_blocks, i)) |
7215 | continue; |
7216 | |
7217 | /* While pipelining outer loops, skip bundling for loop |
7218 | preheaders. Those will be rescheduled in the outer loop. */ |
7219 | if (sel_is_loop_preheader_p (EBB_FIRST_BB (i))) |
7220 | continue; |
7221 | |
7222 | find_ebb_boundaries (EBB_FIRST_BB (i), scheduled_blocks); |
7223 | |
7224 | if (no_real_insns_p (current_sched_info->head, current_sched_info->tail)) |
7225 | continue; |
7226 | |
7227 | if (reset_sched_cycles_p) |
7228 | reset_sched_cycles_in_current_ebb (); |
7229 | |
7230 | if (targetm.sched.init) |
7231 | targetm.sched.init (sched_dump, sched_verbose, -1); |
7232 | |
7233 | put_TImodes (); |
7234 | |
7235 | if (targetm.sched.finish) |
7236 | { |
7237 | targetm.sched.finish (sched_dump, sched_verbose); |
7238 | |
7239 | /* Extend luids so that insns generated by the target will |
7240 | get zero luid. */ |
7241 | sched_extend_luids (); |
7242 | } |
7243 | } |
7244 | |
7245 | BITMAP_FREE (scheduled_blocks); |
7246 | } |
7247 | |
7248 | /* Free the scheduling data for the current region. When RESET_SCHED_CYCLES_P |
7249 | is true, make an additional pass emulating scheduler to get correct insn |
7250 | cycles for md_finish calls. */ |
7251 | static void |
7252 | sel_region_finish (bool reset_sched_cycles_p) |
7253 | { |
7254 | simplify_changed_insns (); |
7255 | sched_finish_ready_list (); |
7256 | free_nop_pool (); |
7257 | |
7258 | /* Free the vectors. */ |
7259 | vec_av_set.release (); |
7260 | BITMAP_FREE (current_copies); |
7261 | BITMAP_FREE (current_originators); |
7262 | BITMAP_FREE (code_motion_visited_blocks); |
7263 | vinsn_vec_free (vinsn_vec&: vec_bookkeeping_blocked_vinsns); |
7264 | vinsn_vec_free (vinsn_vec&: vec_target_unavailable_vinsns); |
7265 | |
7266 | /* If LV_SET of the region head should be updated, do it now because |
7267 | there will be no other chance. */ |
7268 | { |
7269 | succ_iterator si; |
7270 | insn_t insn; |
7271 | |
7272 | FOR_EACH_SUCC_1 (insn, si, bb_note (EBB_FIRST_BB (0)), |
7273 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) |
7274 | { |
7275 | basic_block bb = BLOCK_FOR_INSN (insn); |
7276 | |
7277 | if (!BB_LV_SET_VALID_P (bb)) |
7278 | compute_live (insn); |
7279 | } |
7280 | } |
7281 | |
7282 | /* Emulate the Haifa scheduler for bundling. */ |
7283 | if (reload_completed) |
7284 | sel_region_target_finish (reset_sched_cycles_p); |
7285 | |
7286 | sel_finish_global_and_expr (); |
7287 | |
7288 | BITMAP_FREE (forced_ebb_heads); |
7289 | |
7290 | free_nop_vinsn (); |
7291 | |
7292 | finish_deps_global (); |
7293 | sched_finish_luids (); |
7294 | h_d_i_d.release (); |
7295 | |
7296 | sel_finish_bbs (); |
7297 | BITMAP_FREE (blocks_to_reschedule); |
7298 | |
7299 | sel_unregister_cfg_hooks (); |
7300 | |
7301 | max_issue_size = 0; |
7302 | } |
7303 | |
7304 | |
7305 | /* Functions that implement the scheduler driver. */ |
7306 | |
7307 | /* Schedule a parallel instruction group on each of FENCES. MAX_SEQNO |
7308 | is the current maximum seqno. SCHEDULED_INSNS_TAILPP is the list |
7309 | of insns scheduled -- these would be postprocessed later. */ |
7310 | static void |
7311 | schedule_on_fences (flist_t fences, int max_seqno, |
7312 | ilist_t **scheduled_insns_tailpp) |
7313 | { |
7314 | flist_t old_fences = fences; |
7315 | |
7316 | if (sched_verbose >= 1) |
7317 | { |
7318 | sel_print (fmt: "\nScheduling on fences: " ); |
7319 | dump_flist (fences); |
7320 | sel_print (fmt: "\n" ); |
7321 | } |
7322 | |
7323 | scheduled_something_on_previous_fence = false; |
7324 | for (; fences; fences = FLIST_NEXT (fences)) |
7325 | { |
7326 | fence_t fence = NULL; |
7327 | int seqno = 0; |
7328 | flist_t fences2; |
7329 | bool first_p = true; |
7330 | |
7331 | /* Choose the next fence group to schedule. |
7332 | The fact that insn can be scheduled only once |
7333 | on the cycle is guaranteed by two properties: |
7334 | 1. seqnos of parallel groups decrease with each iteration. |
7335 | 2. If is_ineligible_successor () sees the larger seqno, it |
7336 | checks if candidate insn is_in_current_fence_p (). */ |
7337 | for (fences2 = old_fences; fences2; fences2 = FLIST_NEXT (fences2)) |
7338 | { |
7339 | fence_t f = FLIST_FENCE (fences2); |
7340 | |
7341 | if (!FENCE_PROCESSED_P (f)) |
7342 | { |
7343 | int i = INSN_SEQNO (FENCE_INSN (f)); |
7344 | |
7345 | if (first_p || i > seqno) |
7346 | { |
7347 | seqno = i; |
7348 | fence = f; |
7349 | first_p = false; |
7350 | } |
7351 | else |
7352 | /* ??? Seqnos of different groups should be different. */ |
7353 | gcc_assert (1 || i != seqno); |
7354 | } |
7355 | } |
7356 | |
7357 | gcc_assert (fence); |
7358 | |
7359 | /* As FENCE is nonnull, SEQNO is initialized. */ |
7360 | seqno -= max_seqno + 1; |
7361 | fill_insns (fence, seqno, scheduled_insns_tailpp); |
7362 | FENCE_PROCESSED_P (fence) = true; |
7363 | } |
7364 | |
7365 | /* All av_sets are invalidated by GLOBAL_LEVEL increase, thus we |
7366 | don't need to keep bookkeeping-invalidated and target-unavailable |
7367 | vinsns any more. */ |
7368 | vinsn_vec_clear (vinsn_vec: &vec_bookkeeping_blocked_vinsns); |
7369 | vinsn_vec_clear (vinsn_vec: &vec_target_unavailable_vinsns); |
7370 | } |
7371 | |
7372 | /* Calculate MIN_SEQNO and MAX_SEQNO. */ |
7373 | static void |
7374 | find_min_max_seqno (flist_t fences, int *min_seqno, int *max_seqno) |
7375 | { |
7376 | *min_seqno = *max_seqno = INSN_SEQNO (FENCE_INSN (FLIST_FENCE (fences))); |
7377 | |
7378 | /* The first element is already processed. */ |
7379 | while ((fences = FLIST_NEXT (fences))) |
7380 | { |
7381 | int seqno = INSN_SEQNO (FENCE_INSN (FLIST_FENCE (fences))); |
7382 | |
7383 | if (*min_seqno > seqno) |
7384 | *min_seqno = seqno; |
7385 | else if (*max_seqno < seqno) |
7386 | *max_seqno = seqno; |
7387 | } |
7388 | } |
7389 | |
7390 | /* Calculate new fences from FENCES. Write the current time to PTIME. */ |
7391 | static flist_t |
7392 | calculate_new_fences (flist_t fences, int orig_max_seqno, int *ptime) |
7393 | { |
7394 | flist_t old_fences = fences; |
7395 | struct flist_tail_def _new_fences, *new_fences = &_new_fences; |
7396 | int max_time = 0; |
7397 | |
7398 | flist_tail_init (new_fences); |
7399 | for (; fences; fences = FLIST_NEXT (fences)) |
7400 | { |
7401 | fence_t fence = FLIST_FENCE (fences); |
7402 | insn_t insn; |
7403 | |
7404 | if (!FENCE_BNDS (fence)) |
7405 | { |
7406 | /* This fence doesn't have any successors. */ |
7407 | if (!FENCE_SCHEDULED_P (fence)) |
7408 | { |
7409 | /* Nothing was scheduled on this fence. */ |
7410 | int seqno; |
7411 | |
7412 | insn = FENCE_INSN (fence); |
7413 | seqno = INSN_SEQNO (insn); |
7414 | gcc_assert (seqno > 0 && seqno <= orig_max_seqno); |
7415 | |
7416 | if (sched_verbose >= 1) |
7417 | sel_print (fmt: "Fence %d[%d] has not changed\n" , |
7418 | INSN_UID (insn), |
7419 | BLOCK_NUM (insn)); |
7420 | move_fence_to_fences (fences, new_fences); |
7421 | } |
7422 | } |
7423 | else |
7424 | extract_new_fences_from (old_fences: fences, new_fences, orig_max_seqno); |
7425 | max_time = MAX (max_time, FENCE_CYCLE (fence)); |
7426 | } |
7427 | |
7428 | flist_clear (&old_fences); |
7429 | *ptime = max_time; |
7430 | return FLIST_TAIL_HEAD (new_fences); |
7431 | } |
7432 | |
7433 | /* Update seqnos of insns given by PSCHEDULED_INSNS. MIN_SEQNO and MAX_SEQNO |
7434 | are the miminum and maximum seqnos of the group, HIGHEST_SEQNO_IN_USE is |
7435 | the highest seqno used in a region. Return the updated highest seqno. */ |
7436 | static int |
7437 | update_seqnos_and_stage (int min_seqno, int max_seqno, |
7438 | int highest_seqno_in_use, |
7439 | ilist_t *pscheduled_insns) |
7440 | { |
7441 | int new_hs; |
7442 | ilist_iterator ii; |
7443 | insn_t insn; |
7444 | |
7445 | /* Actually, new_hs is the seqno of the instruction, that was |
7446 | scheduled first (i.e. it is the first one in SCHEDULED_INSNS). */ |
7447 | if (*pscheduled_insns) |
7448 | { |
7449 | new_hs = (INSN_SEQNO (ILIST_INSN (*pscheduled_insns)) |
7450 | + highest_seqno_in_use + max_seqno - min_seqno + 2); |
7451 | gcc_assert (new_hs > highest_seqno_in_use); |
7452 | } |
7453 | else |
7454 | new_hs = highest_seqno_in_use; |
7455 | |
7456 | FOR_EACH_INSN (insn, ii, *pscheduled_insns) |
7457 | { |
7458 | gcc_assert (INSN_SEQNO (insn) < 0); |
7459 | INSN_SEQNO (insn) += highest_seqno_in_use + max_seqno - min_seqno + 2; |
7460 | gcc_assert (INSN_SEQNO (insn) <= new_hs); |
7461 | |
7462 | /* When not pipelining, purge unneeded insn info on the scheduled insns. |
7463 | For example, having reg_last array of INSN_DEPS_CONTEXT in memory may |
7464 | require > 1GB of memory e.g. on limit-fnargs.c. */ |
7465 | if (! pipelining_p) |
7466 | free_data_for_scheduled_insn (insn); |
7467 | } |
7468 | |
7469 | ilist_clear (pscheduled_insns); |
7470 | global_level++; |
7471 | |
7472 | return new_hs; |
7473 | } |
7474 | |
7475 | /* The main driver for scheduling a region. This function is responsible |
7476 | for correct propagation of fences (i.e. scheduling points) and creating |
7477 | a group of parallel insns at each of them. It also supports |
7478 | pipelining. ORIG_MAX_SEQNO is the maximal seqno before this pass |
7479 | of scheduling. */ |
7480 | static void |
7481 | sel_sched_region_2 (int orig_max_seqno) |
7482 | { |
7483 | int highest_seqno_in_use = orig_max_seqno; |
7484 | int max_time = 0; |
7485 | |
7486 | stat_bookkeeping_copies = 0; |
7487 | stat_insns_needed_bookkeeping = 0; |
7488 | stat_renamed_scheduled = 0; |
7489 | stat_substitutions_total = 0; |
7490 | num_insns_scheduled = 0; |
7491 | |
7492 | while (fences) |
7493 | { |
7494 | int min_seqno, max_seqno; |
7495 | ilist_t scheduled_insns = NULL; |
7496 | ilist_t *scheduled_insns_tailp = &scheduled_insns; |
7497 | |
7498 | find_min_max_seqno (fences, min_seqno: &min_seqno, max_seqno: &max_seqno); |
7499 | schedule_on_fences (fences, max_seqno, scheduled_insns_tailpp: &scheduled_insns_tailp); |
7500 | fences = calculate_new_fences (fences, orig_max_seqno, ptime: &max_time); |
7501 | highest_seqno_in_use = update_seqnos_and_stage (min_seqno, max_seqno, |
7502 | highest_seqno_in_use, |
7503 | pscheduled_insns: &scheduled_insns); |
7504 | } |
7505 | |
7506 | if (sched_verbose >= 1) |
7507 | { |
7508 | sel_print (fmt: "Total scheduling time: %d cycles\n" , max_time); |
7509 | sel_print (fmt: "Scheduled %d bookkeeping copies, %d insns needed " |
7510 | "bookkeeping, %d insns renamed, %d insns substituted\n" , |
7511 | stat_bookkeeping_copies, |
7512 | stat_insns_needed_bookkeeping, |
7513 | stat_renamed_scheduled, |
7514 | stat_substitutions_total); |
7515 | } |
7516 | } |
7517 | |
7518 | /* Schedule a region. When pipelining, search for possibly never scheduled |
7519 | bookkeeping code and schedule it. Reschedule pipelined code without |
7520 | pipelining after. */ |
7521 | static void |
7522 | sel_sched_region_1 (void) |
7523 | { |
7524 | int orig_max_seqno; |
7525 | |
7526 | /* Remove empty blocks that might be in the region from the beginning. */ |
7527 | purge_empty_blocks (); |
7528 | |
7529 | orig_max_seqno = init_seqno (NULL, NULL); |
7530 | gcc_assert (orig_max_seqno >= 1); |
7531 | |
7532 | /* When pipelining outer loops, create fences on the loop header, |
7533 | not preheader. */ |
7534 | fences = NULL; |
7535 | if (current_loop_nest) |
7536 | init_fences (BB_END (EBB_FIRST_BB (0))); |
7537 | else |
7538 | init_fences (bb_note (EBB_FIRST_BB (0))); |
7539 | global_level = 1; |
7540 | |
7541 | sel_sched_region_2 (orig_max_seqno); |
7542 | |
7543 | gcc_assert (fences == NULL); |
7544 | |
7545 | if (pipelining_p) |
7546 | { |
7547 | int i; |
7548 | basic_block bb; |
7549 | struct flist_tail_def _new_fences; |
7550 | flist_tail_t new_fences = &_new_fences; |
7551 | bool do_p = true; |
7552 | |
7553 | pipelining_p = false; |
7554 | max_ws = MIN (max_ws, issue_rate * 3 / 2); |
7555 | bookkeeping_p = false; |
7556 | enable_schedule_as_rhs_p = false; |
7557 | |
7558 | /* Schedule newly created code, that has not been scheduled yet. */ |
7559 | do_p = true; |
7560 | |
7561 | while (do_p) |
7562 | { |
7563 | do_p = false; |
7564 | |
7565 | for (i = 0; i < current_nr_blocks; i++) |
7566 | { |
7567 | basic_block bb = EBB_FIRST_BB (i); |
7568 | |
7569 | if (bitmap_bit_p (blocks_to_reschedule, bb->index)) |
7570 | { |
7571 | if (! bb_ends_ebb_p (bb)) |
7572 | bitmap_set_bit (blocks_to_reschedule, bb_next_bb (bb)->index); |
7573 | if (sel_bb_empty_p (bb)) |
7574 | { |
7575 | bitmap_clear_bit (blocks_to_reschedule, bb->index); |
7576 | continue; |
7577 | } |
7578 | clear_outdated_rtx_info (bb); |
7579 | if (sel_insn_is_speculation_check (BB_END (bb)) |
7580 | && JUMP_P (BB_END (bb))) |
7581 | bitmap_set_bit (blocks_to_reschedule, |
7582 | BRANCH_EDGE (bb)->dest->index); |
7583 | } |
7584 | else if (! sel_bb_empty_p (bb) |
7585 | && INSN_SCHED_TIMES (sel_bb_head (bb)) <= 0) |
7586 | bitmap_set_bit (blocks_to_reschedule, bb->index); |
7587 | } |
7588 | |
7589 | for (i = 0; i < current_nr_blocks; i++) |
7590 | { |
7591 | bb = EBB_FIRST_BB (i); |
7592 | |
7593 | /* While pipelining outer loops, skip bundling for loop |
7594 | preheaders. Those will be rescheduled in the outer |
7595 | loop. */ |
7596 | if (sel_is_loop_preheader_p (bb)) |
7597 | { |
7598 | clear_outdated_rtx_info (bb); |
7599 | continue; |
7600 | } |
7601 | |
7602 | if (bitmap_bit_p (blocks_to_reschedule, bb->index)) |
7603 | { |
7604 | flist_tail_init (new_fences); |
7605 | |
7606 | orig_max_seqno = init_seqno (blocks_to_reschedule, from: bb); |
7607 | |
7608 | /* Mark BB as head of the new ebb. */ |
7609 | bitmap_set_bit (forced_ebb_heads, bb->index); |
7610 | |
7611 | gcc_assert (fences == NULL); |
7612 | |
7613 | init_fences (bb_note (bb)); |
7614 | |
7615 | sel_sched_region_2 (orig_max_seqno); |
7616 | |
7617 | do_p = true; |
7618 | break; |
7619 | } |
7620 | } |
7621 | } |
7622 | } |
7623 | } |
7624 | |
7625 | /* Schedule the RGN region. */ |
7626 | void |
7627 | sel_sched_region (int rgn) |
7628 | { |
7629 | bool schedule_p; |
7630 | bool reset_sched_cycles_p; |
7631 | |
7632 | if (sel_region_init (rgn)) |
7633 | return; |
7634 | |
7635 | if (sched_verbose >= 1) |
7636 | sel_print (fmt: "Scheduling region %d\n" , rgn); |
7637 | |
7638 | schedule_p = (!sched_is_disabled_for_current_region_p () |
7639 | && dbg_cnt (index: sel_sched_region_cnt)); |
7640 | reset_sched_cycles_p = pipelining_p; |
7641 | if (schedule_p) |
7642 | sel_sched_region_1 (); |
7643 | else |
7644 | { |
7645 | /* Schedule always selecting the next insn to make the correct data |
7646 | for bundling or other later passes. */ |
7647 | pipelining_p = false; |
7648 | reset_sched_cycles_p = false; |
7649 | force_next_insn = 1; |
7650 | sel_sched_region_1 (); |
7651 | force_next_insn = 0; |
7652 | } |
7653 | sel_region_finish (reset_sched_cycles_p); |
7654 | } |
7655 | |
7656 | /* Perform global init for the scheduler. */ |
7657 | static void |
7658 | sel_global_init (void) |
7659 | { |
7660 | /* Remove empty blocks: their presence can break assumptions elsewhere, |
7661 | e.g. the logic to invoke update_liveness_on_insn in sel_region_init. */ |
7662 | cleanup_cfg (0); |
7663 | |
7664 | calculate_dominance_info (CDI_DOMINATORS); |
7665 | alloc_sched_pools (); |
7666 | |
7667 | /* Setup the infos for sched_init. */ |
7668 | sel_setup_sched_infos (); |
7669 | setup_sched_dump (); |
7670 | |
7671 | sched_rgn_init (false); |
7672 | sched_init (); |
7673 | |
7674 | sched_init_bbs (); |
7675 | /* Reset AFTER_RECOVERY if it has been set by the 1st scheduler pass. */ |
7676 | after_recovery = 0; |
7677 | can_issue_more = issue_rate; |
7678 | |
7679 | sched_extend_target (); |
7680 | sched_deps_init (true); |
7681 | setup_nop_and_exit_insns (); |
7682 | sel_extend_global_bb_info (); |
7683 | init_lv_sets (); |
7684 | init_hard_regs_data (); |
7685 | } |
7686 | |
7687 | /* Free the global data of the scheduler. */ |
7688 | static void |
7689 | sel_global_finish (void) |
7690 | { |
7691 | free_bb_note_pool (); |
7692 | free_lv_sets (); |
7693 | sel_finish_global_bb_info (); |
7694 | |
7695 | free_regset_pool (); |
7696 | free_nop_and_exit_insns (); |
7697 | |
7698 | sched_rgn_finish (); |
7699 | sched_deps_finish (); |
7700 | sched_finish (); |
7701 | |
7702 | if (current_loops) |
7703 | sel_finish_pipelining (); |
7704 | |
7705 | free_sched_pools (); |
7706 | free_dominance_info (CDI_DOMINATORS); |
7707 | } |
7708 | |
7709 | /* Return true when we need to skip selective scheduling. Used for debugging. */ |
7710 | bool |
7711 | maybe_skip_selective_scheduling (void) |
7712 | { |
7713 | return ! dbg_cnt (index: sel_sched_cnt); |
7714 | } |
7715 | |
7716 | /* The entry point. */ |
7717 | void |
7718 | run_selective_scheduling (void) |
7719 | { |
7720 | int rgn; |
7721 | |
7722 | if (n_basic_blocks_for_fn (cfun) == NUM_FIXED_BLOCKS) |
7723 | return; |
7724 | |
7725 | sel_global_init (); |
7726 | |
7727 | for (rgn = 0; rgn < nr_regions; rgn++) |
7728 | sel_sched_region (rgn); |
7729 | |
7730 | sel_global_finish (); |
7731 | } |
7732 | |
7733 | #endif |
7734 | |