Mercurial > hg > CbC > CbC_gcc
comparison gcc/sel-sched.c @ 0:a06113de4d67
first commit
author | kent <kent@cr.ie.u-ryukyu.ac.jp> |
---|---|
date | Fri, 17 Jul 2009 14:47:48 +0900 |
parents | |
children | 77e2b8dfacca |
comparison
equal
deleted
inserted
replaced
-1:000000000000 | 0:a06113de4d67 |
---|---|
1 /* Instruction scheduling pass. Selective scheduler and pipeliner. | |
2 Copyright (C) 2006, 2007, 2008, 2009 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 "tm.h" | |
24 #include "toplev.h" | |
25 #include "rtl.h" | |
26 #include "tm_p.h" | |
27 #include "hard-reg-set.h" | |
28 #include "regs.h" | |
29 #include "function.h" | |
30 #include "flags.h" | |
31 #include "insn-config.h" | |
32 #include "insn-attr.h" | |
33 #include "except.h" | |
34 #include "toplev.h" | |
35 #include "recog.h" | |
36 #include "params.h" | |
37 #include "target.h" | |
38 #include "output.h" | |
39 #include "timevar.h" | |
40 #include "tree-pass.h" | |
41 #include "sched-int.h" | |
42 #include "ggc.h" | |
43 #include "tree.h" | |
44 #include "vec.h" | |
45 #include "langhooks.h" | |
46 #include "rtlhooks-def.h" | |
47 #include "output.h" | |
48 | |
49 #ifdef INSN_SCHEDULING | |
50 #include "sel-sched-ir.h" | |
51 #include "sel-sched-dump.h" | |
52 #include "sel-sched.h" | |
53 #include "dbgcnt.h" | |
54 | |
55 /* Implementation of selective scheduling approach. | |
56 The below implementation follows the original approach with the following | |
57 changes: | |
58 | |
59 o the scheduler works after register allocation (but can be also tuned | |
60 to work before RA); | |
61 o some instructions are not copied or register renamed; | |
62 o conditional jumps are not moved with code duplication; | |
63 o several jumps in one parallel group are not supported; | |
64 o when pipelining outer loops, code motion through inner loops | |
65 is not supported; | |
66 o control and data speculation are supported; | |
67 o some improvements for better compile time/performance were made. | |
68 | |
69 Terminology | |
70 =========== | |
71 | |
72 A vinsn, or virtual insn, is an insn with additional data characterizing | |
73 insn pattern, such as LHS, RHS, register sets used/set/clobbered, etc. | |
74 Vinsns also act as smart pointers to save memory by reusing them in | |
75 different expressions. A vinsn is described by vinsn_t type. | |
76 | |
77 An expression is a vinsn with additional data characterizing its properties | |
78 at some point in the control flow graph. The data may be its usefulness, | |
79 priority, speculative status, whether it was renamed/subsituted, etc. | |
80 An expression is described by expr_t type. | |
81 | |
82 Availability set (av_set) is a set of expressions at a given control flow | |
83 point. It is represented as av_set_t. The expressions in av sets are kept | |
84 sorted in the terms of expr_greater_p function. It allows to truncate | |
85 the set while leaving the best expressions. | |
86 | |
87 A fence is a point through which code motion is prohibited. On each step, | |
88 we gather a parallel group of insns at a fence. It is possible to have | |
89 multiple fences. A fence is represented via fence_t. | |
90 | |
91 A boundary is the border between the fence group and the rest of the code. | |
92 Currently, we never have more than one boundary per fence, as we finalize | |
93 the fence group when a jump is scheduled. A boundary is represented | |
94 via bnd_t. | |
95 | |
96 High-level overview | |
97 =================== | |
98 | |
99 The scheduler finds regions to schedule, schedules each one, and finalizes. | |
100 The regions are formed starting from innermost loops, so that when the inner | |
101 loop is pipelined, its prologue can be scheduled together with yet unprocessed | |
102 outer loop. The rest of acyclic regions are found using extend_rgns: | |
103 the blocks that are not yet allocated to any regions are traversed in top-down | |
104 order, and a block is added to a region to which all its predecessors belong; | |
105 otherwise, the block starts its own region. | |
106 | |
107 The main scheduling loop (sel_sched_region_2) consists of just | |
108 scheduling on each fence and updating fences. For each fence, | |
109 we fill a parallel group of insns (fill_insns) until some insns can be added. | |
110 First, we compute available exprs (av-set) at the boundary of the current | |
111 group. Second, we choose the best expression from it. If the stall is | |
112 required to schedule any of the expressions, we advance the current cycle | |
113 appropriately. So, the final group does not exactly correspond to a VLIW | |
114 word. Third, we move the chosen expression to the boundary (move_op) | |
115 and update the intermediate av sets and liveness sets. We quit fill_insns | |
116 when either no insns left for scheduling or we have scheduled enough insns | |
117 so we feel like advancing a scheduling point. | |
118 | |
119 Computing available expressions | |
120 =============================== | |
121 | |
122 The computation (compute_av_set) is a bottom-up traversal. At each insn, | |
123 we're moving the union of its successors' sets through it via | |
124 moveup_expr_set. The dependent expressions are removed. Local | |
125 transformations (substitution, speculation) are applied to move more | |
126 exprs. Then the expr corresponding to the current insn is added. | |
127 The result is saved on each basic block header. | |
128 | |
129 When traversing the CFG, we're moving down for no more than max_ws insns. | |
130 Also, we do not move down to ineligible successors (is_ineligible_successor), | |
131 which include moving along a back-edge, moving to already scheduled code, | |
132 and moving to another fence. The first two restrictions are lifted during | |
133 pipelining, which allows us to move insns along a back-edge. We always have | |
134 an acyclic region for scheduling because we forbid motion through fences. | |
135 | |
136 Choosing the best expression | |
137 ============================ | |
138 | |
139 We sort the final availability set via sel_rank_for_schedule, then we remove | |
140 expressions which are not yet ready (tick_check_p) or which dest registers | |
141 cannot be used. For some of them, we choose another register via | |
142 find_best_reg. To do this, we run find_used_regs to calculate the set of | |
143 registers which cannot be used. The find_used_regs function performs | |
144 a traversal of code motion paths for an expr. We consider for renaming | |
145 only registers which are from the same regclass as the original one and | |
146 using which does not interfere with any live ranges. Finally, we convert | |
147 the resulting set to the ready list format and use max_issue and reorder* | |
148 hooks similarly to the Haifa scheduler. | |
149 | |
150 Scheduling the best expression | |
151 ============================== | |
152 | |
153 We run the move_op routine to perform the same type of code motion paths | |
154 traversal as in find_used_regs. (These are working via the same driver, | |
155 code_motion_path_driver.) When moving down the CFG, we look for original | |
156 instruction that gave birth to a chosen expression. We undo | |
157 the transformations performed on an expression via the history saved in it. | |
158 When found, we remove the instruction or leave a reg-reg copy/speculation | |
159 check if needed. On a way up, we insert bookkeeping copies at each join | |
160 point. If a copy is not needed, it will be removed later during this | |
161 traversal. We update the saved av sets and liveness sets on the way up, too. | |
162 | |
163 Finalizing the schedule | |
164 ======================= | |
165 | |
166 When pipelining, we reschedule the blocks from which insns were pipelined | |
167 to get a tighter schedule. On Itanium, we also perform bundling via | |
168 the same routine from ia64.c. | |
169 | |
170 Dependence analysis changes | |
171 =========================== | |
172 | |
173 We augmented the sched-deps.c with hooks that get called when a particular | |
174 dependence is found in a particular part of an insn. Using these hooks, we | |
175 can do several actions such as: determine whether an insn can be moved through | |
176 another (has_dependence_p, moveup_expr); find out whether an insn can be | |
177 scheduled on the current cycle (tick_check_p); find out registers that | |
178 are set/used/clobbered by an insn and find out all the strange stuff that | |
179 restrict its movement, like SCHED_GROUP_P or CANT_MOVE (done in | |
180 init_global_and_expr_for_insn). | |
181 | |
182 Initialization changes | |
183 ====================== | |
184 | |
185 There are parts of haifa-sched.c, sched-deps.c, and sched-rgn.c that are | |
186 reused in all of the schedulers. We have split up the initialization of data | |
187 of such parts into different functions prefixed with scheduler type and | |
188 postfixed with the type of data initialized: {,sel_,haifa_}sched_{init,finish}, | |
189 sched_rgn_init/finish, sched_deps_init/finish, sched_init_{luids/bbs}, etc. | |
190 The same splitting is done with current_sched_info structure: | |
191 dependence-related parts are in sched_deps_info, common part is in | |
192 common_sched_info, and haifa/sel/etc part is in current_sched_info. | |
193 | |
194 Target contexts | |
195 =============== | |
196 | |
197 As we now have multiple-point scheduling, this would not work with backends | |
198 which save some of the scheduler state to use it in the target hooks. | |
199 For this purpose, we introduce a concept of target contexts, which | |
200 encapsulate such information. The backend should implement simple routines | |
201 of allocating/freeing/setting such a context. The scheduler calls these | |
202 as target hooks and handles the target context as an opaque pointer (similar | |
203 to the DFA state type, state_t). | |
204 | |
205 Various speedups | |
206 ================ | |
207 | |
208 As the correct data dependence graph is not supported during scheduling (which | |
209 is to be changed in mid-term), we cache as much of the dependence analysis | |
210 results as possible to avoid reanalyzing. This includes: bitmap caches on | |
211 each insn in stream of the region saying yes/no for a query with a pair of | |
212 UIDs; hashtables with the previously done transformations on each insn in | |
213 stream; a vector keeping a history of transformations on each expr. | |
214 | |
215 Also, we try to minimize the dependence context used on each fence to check | |
216 whether the given expression is ready for scheduling by removing from it | |
217 insns that are definitely completed the execution. The results of | |
218 tick_check_p checks are also cached in a vector on each fence. | |
219 | |
220 We keep a valid liveness set on each insn in a region to avoid the high | |
221 cost of recomputation on large basic blocks. | |
222 | |
223 Finally, we try to minimize the number of needed updates to the availability | |
224 sets. The updates happen in two cases: when fill_insns terminates, | |
225 we advance all fences and increase the stage number to show that the region | |
226 has changed and the sets are to be recomputed; and when the next iteration | |
227 of a loop in fill_insns happens (but this one reuses the saved av sets | |
228 on bb headers.) Thus, we try to break the fill_insns loop only when | |
229 "significant" number of insns from the current scheduling window was | |
230 scheduled. This should be made a target param. | |
231 | |
232 | |
233 TODO: correctly support the data dependence graph at all stages and get rid | |
234 of all caches. This should speed up the scheduler. | |
235 TODO: implement moving cond jumps with bookkeeping copies on both targets. | |
236 TODO: tune the scheduler before RA so it does not create too much pseudos. | |
237 | |
238 | |
239 References: | |
240 S.-M. Moon and K. Ebcioglu. Parallelizing nonnumerical code with | |
241 selective scheduling and software pipelining. | |
242 ACM TOPLAS, Vol 19, No. 6, pages 853--898, Nov. 1997. | |
243 | |
244 Andrey Belevantsev, Maxim Kuvyrkov, Vladimir Makarov, Dmitry Melnik, | |
245 and Dmitry Zhurikhin. An interblock VLIW-targeted instruction scheduler | |
246 for GCC. In Proceedings of GCC Developers' Summit 2006. | |
247 | |
248 Arutyun Avetisyan, Andrey Belevantsev, and Dmitry Melnik. GCC Instruction | |
249 Scheduler and Software Pipeliner on the Itanium Platform. EPIC-7 Workshop. | |
250 http://rogue.colorado.edu/EPIC7/. | |
251 | |
252 */ | |
253 | |
254 /* True when pipelining is enabled. */ | |
255 bool pipelining_p; | |
256 | |
257 /* True if bookkeeping is enabled. */ | |
258 bool bookkeeping_p; | |
259 | |
260 /* Maximum number of insns that are eligible for renaming. */ | |
261 int max_insns_to_rename; | |
262 | |
263 | |
264 /* Definitions of local types and macros. */ | |
265 | |
266 /* Represents possible outcomes of moving an expression through an insn. */ | |
267 enum MOVEUP_EXPR_CODE | |
268 { | |
269 /* The expression is not changed. */ | |
270 MOVEUP_EXPR_SAME, | |
271 | |
272 /* Not changed, but requires a new destination register. */ | |
273 MOVEUP_EXPR_AS_RHS, | |
274 | |
275 /* Cannot be moved. */ | |
276 MOVEUP_EXPR_NULL, | |
277 | |
278 /* Changed (substituted or speculated). */ | |
279 MOVEUP_EXPR_CHANGED | |
280 }; | |
281 | |
282 /* The container to be passed into rtx search & replace functions. */ | |
283 struct rtx_search_arg | |
284 { | |
285 /* What we are searching for. */ | |
286 rtx x; | |
287 | |
288 /* The occurence counter. */ | |
289 int n; | |
290 }; | |
291 | |
292 typedef struct rtx_search_arg *rtx_search_arg_p; | |
293 | |
294 /* This struct contains precomputed hard reg sets that are needed when | |
295 computing registers available for renaming. */ | |
296 struct hard_regs_data | |
297 { | |
298 /* For every mode, this stores registers available for use with | |
299 that mode. */ | |
300 HARD_REG_SET regs_for_mode[NUM_MACHINE_MODES]; | |
301 | |
302 /* True when regs_for_mode[mode] is initialized. */ | |
303 bool regs_for_mode_ok[NUM_MACHINE_MODES]; | |
304 | |
305 /* For every register, it has regs that are ok to rename into it. | |
306 The register in question is always set. If not, this means | |
307 that the whole set is not computed yet. */ | |
308 HARD_REG_SET regs_for_rename[FIRST_PSEUDO_REGISTER]; | |
309 | |
310 /* For every mode, this stores registers not available due to | |
311 call clobbering. */ | |
312 HARD_REG_SET regs_for_call_clobbered[NUM_MACHINE_MODES]; | |
313 | |
314 /* All registers that are used or call used. */ | |
315 HARD_REG_SET regs_ever_used; | |
316 | |
317 #ifdef STACK_REGS | |
318 /* Stack registers. */ | |
319 HARD_REG_SET stack_regs; | |
320 #endif | |
321 }; | |
322 | |
323 /* Holds the results of computation of available for renaming and | |
324 unavailable hard registers. */ | |
325 struct reg_rename | |
326 { | |
327 /* These are unavailable due to calls crossing, globalness, etc. */ | |
328 HARD_REG_SET unavailable_hard_regs; | |
329 | |
330 /* These are *available* for renaming. */ | |
331 HARD_REG_SET available_for_renaming; | |
332 | |
333 /* Whether this code motion path crosses a call. */ | |
334 bool crosses_call; | |
335 }; | |
336 | |
337 /* A global structure that contains the needed information about harg | |
338 regs. */ | |
339 static struct hard_regs_data sel_hrd; | |
340 | |
341 | |
342 /* This structure holds local data used in code_motion_path_driver hooks on | |
343 the same or adjacent levels of recursion. Here we keep those parameters | |
344 that are not used in code_motion_path_driver routine itself, but only in | |
345 its hooks. Moreover, all parameters that can be modified in hooks are | |
346 in this structure, so all other parameters passed explicitly to hooks are | |
347 read-only. */ | |
348 struct cmpd_local_params | |
349 { | |
350 /* Local params used in move_op_* functions. */ | |
351 | |
352 /* Edges for bookkeeping generation. */ | |
353 edge e1, e2; | |
354 | |
355 /* C_EXPR merged from all successors and locally allocated temporary C_EXPR. */ | |
356 expr_t c_expr_merged, c_expr_local; | |
357 | |
358 /* Local params used in fur_* functions. */ | |
359 /* Copy of the ORIGINAL_INSN list, stores the original insns already | |
360 found before entering the current level of code_motion_path_driver. */ | |
361 def_list_t old_original_insns; | |
362 | |
363 /* Local params used in move_op_* functions. */ | |
364 /* True when we have removed last insn in the block which was | |
365 also a boundary. Do not update anything or create bookkeeping copies. */ | |
366 BOOL_BITFIELD removed_last_insn : 1; | |
367 }; | |
368 | |
369 /* Stores the static parameters for move_op_* calls. */ | |
370 struct moveop_static_params | |
371 { | |
372 /* Destination register. */ | |
373 rtx dest; | |
374 | |
375 /* Current C_EXPR. */ | |
376 expr_t c_expr; | |
377 | |
378 /* An UID of expr_vliw which is to be moved up. If we find other exprs, | |
379 they are to be removed. */ | |
380 int uid; | |
381 | |
382 #ifdef ENABLE_CHECKING | |
383 /* This is initialized to the insn on which the driver stopped its traversal. */ | |
384 insn_t failed_insn; | |
385 #endif | |
386 | |
387 /* True if we scheduled an insn with different register. */ | |
388 bool was_renamed; | |
389 }; | |
390 | |
391 /* Stores the static parameters for fur_* calls. */ | |
392 struct fur_static_params | |
393 { | |
394 /* Set of registers unavailable on the code motion path. */ | |
395 regset used_regs; | |
396 | |
397 /* Pointer to the list of original insns definitions. */ | |
398 def_list_t *original_insns; | |
399 | |
400 /* True if a code motion path contains a CALL insn. */ | |
401 bool crosses_call; | |
402 }; | |
403 | |
404 typedef struct fur_static_params *fur_static_params_p; | |
405 typedef struct cmpd_local_params *cmpd_local_params_p; | |
406 typedef struct moveop_static_params *moveop_static_params_p; | |
407 | |
408 /* Set of hooks and parameters that determine behaviour specific to | |
409 move_op or find_used_regs functions. */ | |
410 struct code_motion_path_driver_info_def | |
411 { | |
412 /* Called on enter to the basic block. */ | |
413 int (*on_enter) (insn_t, cmpd_local_params_p, void *, bool); | |
414 | |
415 /* Called when original expr is found. */ | |
416 void (*orig_expr_found) (insn_t, expr_t, cmpd_local_params_p, void *); | |
417 | |
418 /* Called while descending current basic block if current insn is not | |
419 the original EXPR we're searching for. */ | |
420 bool (*orig_expr_not_found) (insn_t, av_set_t, void *); | |
421 | |
422 /* Function to merge C_EXPRes from different successors. */ | |
423 void (*merge_succs) (insn_t, insn_t, int, cmpd_local_params_p, void *); | |
424 | |
425 /* Function to finalize merge from different successors and possibly | |
426 deallocate temporary data structures used for merging. */ | |
427 void (*after_merge_succs) (cmpd_local_params_p, void *); | |
428 | |
429 /* Called on the backward stage of recursion to do moveup_expr. | |
430 Used only with move_op_*. */ | |
431 void (*ascend) (insn_t, void *); | |
432 | |
433 /* Called on the ascending pass, before returning from the current basic | |
434 block or from the whole traversal. */ | |
435 void (*at_first_insn) (insn_t, cmpd_local_params_p, void *); | |
436 | |
437 /* When processing successors in move_op we need only descend into | |
438 SUCCS_NORMAL successors, while in find_used_regs we need SUCCS_ALL. */ | |
439 int succ_flags; | |
440 | |
441 /* The routine name to print in dumps ("move_op" of "find_used_regs"). */ | |
442 const char *routine_name; | |
443 }; | |
444 | |
445 /* Global pointer to current hooks, either points to MOVE_OP_HOOKS or | |
446 FUR_HOOKS. */ | |
447 struct code_motion_path_driver_info_def *code_motion_path_driver_info; | |
448 | |
449 /* Set of hooks for performing move_op and find_used_regs routines with | |
450 code_motion_path_driver. */ | |
451 struct code_motion_path_driver_info_def move_op_hooks, fur_hooks; | |
452 | |
453 /* True if/when we want to emulate Haifa scheduler in the common code. | |
454 This is used in sched_rgn_local_init and in various places in | |
455 sched-deps.c. */ | |
456 int sched_emulate_haifa_p; | |
457 | |
458 /* GLOBAL_LEVEL is used to discard information stored in basic block headers | |
459 av_sets. Av_set of bb header is valid if its (bb header's) level is equal | |
460 to GLOBAL_LEVEL. And invalid if lesser. This is primarily used to advance | |
461 scheduling window. */ | |
462 int global_level; | |
463 | |
464 /* Current fences. */ | |
465 flist_t fences; | |
466 | |
467 /* True when separable insns should be scheduled as RHSes. */ | |
468 static bool enable_schedule_as_rhs_p; | |
469 | |
470 /* Used in verify_target_availability to assert that target reg is reported | |
471 unavailabile by both TARGET_UNAVAILABLE and find_used_regs only if | |
472 we haven't scheduled anything on the previous fence. | |
473 if scheduled_something_on_previous_fence is true, TARGET_UNAVAILABLE can | |
474 have more conservative value than the one returned by the | |
475 find_used_regs, thus we shouldn't assert that these values are equal. */ | |
476 static bool scheduled_something_on_previous_fence; | |
477 | |
478 /* All newly emitted insns will have their uids greater than this value. */ | |
479 static int first_emitted_uid; | |
480 | |
481 /* Set of basic blocks that are forced to start new ebbs. This is a subset | |
482 of all the ebb heads. */ | |
483 static bitmap_head _forced_ebb_heads; | |
484 bitmap_head *forced_ebb_heads = &_forced_ebb_heads; | |
485 | |
486 /* Blocks that need to be rescheduled after pipelining. */ | |
487 bitmap blocks_to_reschedule = NULL; | |
488 | |
489 /* True when the first lv set should be ignored when updating liveness. */ | |
490 static bool ignore_first = false; | |
491 | |
492 /* Number of insns max_issue has initialized data structures for. */ | |
493 static int max_issue_size = 0; | |
494 | |
495 /* Whether we can issue more instructions. */ | |
496 static int can_issue_more; | |
497 | |
498 /* Maximum software lookahead window size, reduced when rescheduling after | |
499 pipelining. */ | |
500 static int max_ws; | |
501 | |
502 /* Number of insns scheduled in current region. */ | |
503 static int num_insns_scheduled; | |
504 | |
505 /* A vector of expressions is used to be able to sort them. */ | |
506 DEF_VEC_P(expr_t); | |
507 DEF_VEC_ALLOC_P(expr_t,heap); | |
508 static VEC(expr_t, heap) *vec_av_set = NULL; | |
509 | |
510 /* A vector of vinsns is used to hold temporary lists of vinsns. */ | |
511 DEF_VEC_P(vinsn_t); | |
512 DEF_VEC_ALLOC_P(vinsn_t,heap); | |
513 typedef VEC(vinsn_t, heap) *vinsn_vec_t; | |
514 | |
515 /* This vector has the exprs which may still present in av_sets, but actually | |
516 can't be moved up due to bookkeeping created during code motion to another | |
517 fence. See comment near the call to update_and_record_unavailable_insns | |
518 for the detailed explanations. */ | |
519 static vinsn_vec_t vec_bookkeeping_blocked_vinsns = NULL; | |
520 | |
521 /* This vector has vinsns which are scheduled with renaming on the first fence | |
522 and then seen on the second. For expressions with such vinsns, target | |
523 availability information may be wrong. */ | |
524 static vinsn_vec_t vec_target_unavailable_vinsns = NULL; | |
525 | |
526 /* Vector to store temporary nops inserted in move_op to prevent removal | |
527 of empty bbs. */ | |
528 DEF_VEC_P(insn_t); | |
529 DEF_VEC_ALLOC_P(insn_t,heap); | |
530 static VEC(insn_t, heap) *vec_temp_moveop_nops = NULL; | |
531 | |
532 /* These bitmaps record original instructions scheduled on the current | |
533 iteration and bookkeeping copies created by them. */ | |
534 static bitmap current_originators = NULL; | |
535 static bitmap current_copies = NULL; | |
536 | |
537 /* This bitmap marks the blocks visited by code_motion_path_driver so we don't | |
538 visit them afterwards. */ | |
539 static bitmap code_motion_visited_blocks = NULL; | |
540 | |
541 /* Variables to accumulate different statistics. */ | |
542 | |
543 /* The number of bookkeeping copies created. */ | |
544 static int stat_bookkeeping_copies; | |
545 | |
546 /* The number of insns that required bookkeeiping for their scheduling. */ | |
547 static int stat_insns_needed_bookkeeping; | |
548 | |
549 /* The number of insns that got renamed. */ | |
550 static int stat_renamed_scheduled; | |
551 | |
552 /* The number of substitutions made during scheduling. */ | |
553 static int stat_substitutions_total; | |
554 | |
555 | |
556 /* Forward declarations of static functions. */ | |
557 static bool rtx_ok_for_substitution_p (rtx, rtx); | |
558 static int sel_rank_for_schedule (const void *, const void *); | |
559 static av_set_t find_sequential_best_exprs (bnd_t, expr_t, bool); | |
560 | |
561 static rtx get_dest_from_orig_ops (av_set_t); | |
562 static basic_block generate_bookkeeping_insn (expr_t, edge, edge); | |
563 static bool find_used_regs (insn_t, av_set_t, regset, struct reg_rename *, | |
564 def_list_t *); | |
565 static bool move_op (insn_t, av_set_t, expr_t, rtx, expr_t, bool*); | |
566 static int code_motion_path_driver (insn_t, av_set_t, ilist_t, | |
567 cmpd_local_params_p, void *); | |
568 static void sel_sched_region_1 (void); | |
569 static void sel_sched_region_2 (int); | |
570 static av_set_t compute_av_set_inside_bb (insn_t, ilist_t, int, bool); | |
571 | |
572 static void debug_state (state_t); | |
573 | |
574 | |
575 /* Functions that work with fences. */ | |
576 | |
577 /* Advance one cycle on FENCE. */ | |
578 static void | |
579 advance_one_cycle (fence_t fence) | |
580 { | |
581 unsigned i; | |
582 int cycle; | |
583 rtx insn; | |
584 | |
585 advance_state (FENCE_STATE (fence)); | |
586 cycle = ++FENCE_CYCLE (fence); | |
587 FENCE_ISSUED_INSNS (fence) = 0; | |
588 FENCE_STARTS_CYCLE_P (fence) = 1; | |
589 can_issue_more = issue_rate; | |
590 | |
591 for (i = 0; VEC_iterate (rtx, FENCE_EXECUTING_INSNS (fence), i, insn); ) | |
592 { | |
593 if (INSN_READY_CYCLE (insn) < cycle) | |
594 { | |
595 remove_from_deps (FENCE_DC (fence), insn); | |
596 VEC_unordered_remove (rtx, FENCE_EXECUTING_INSNS (fence), i); | |
597 continue; | |
598 } | |
599 i++; | |
600 } | |
601 if (sched_verbose >= 2) | |
602 { | |
603 sel_print ("Finished a cycle. Current cycle = %d\n", FENCE_CYCLE (fence)); | |
604 debug_state (FENCE_STATE (fence)); | |
605 } | |
606 } | |
607 | |
608 /* Returns true when SUCC in a fallthru bb of INSN, possibly | |
609 skipping empty basic blocks. */ | |
610 static bool | |
611 in_fallthru_bb_p (rtx insn, rtx succ) | |
612 { | |
613 basic_block bb = BLOCK_FOR_INSN (insn); | |
614 | |
615 if (bb == BLOCK_FOR_INSN (succ)) | |
616 return true; | |
617 | |
618 if (find_fallthru_edge (bb)) | |
619 bb = find_fallthru_edge (bb)->dest; | |
620 else | |
621 return false; | |
622 | |
623 while (sel_bb_empty_p (bb)) | |
624 bb = bb->next_bb; | |
625 | |
626 return bb == BLOCK_FOR_INSN (succ); | |
627 } | |
628 | |
629 /* Construct successor fences from OLD_FENCEs and put them in NEW_FENCES. | |
630 When a successor will continue a ebb, transfer all parameters of a fence | |
631 to the new fence. ORIG_MAX_SEQNO is the maximal seqno before this round | |
632 of scheduling helping to distinguish between the old and the new code. */ | |
633 static void | |
634 extract_new_fences_from (flist_t old_fences, flist_tail_t new_fences, | |
635 int orig_max_seqno) | |
636 { | |
637 bool was_here_p = false; | |
638 insn_t insn = NULL_RTX; | |
639 insn_t succ; | |
640 succ_iterator si; | |
641 ilist_iterator ii; | |
642 fence_t fence = FLIST_FENCE (old_fences); | |
643 basic_block bb; | |
644 | |
645 /* Get the only element of FENCE_BNDS (fence). */ | |
646 FOR_EACH_INSN (insn, ii, FENCE_BNDS (fence)) | |
647 { | |
648 gcc_assert (!was_here_p); | |
649 was_here_p = true; | |
650 } | |
651 gcc_assert (was_here_p && insn != NULL_RTX); | |
652 | |
653 /* When in the "middle" of the block, just move this fence | |
654 to the new list. */ | |
655 bb = BLOCK_FOR_INSN (insn); | |
656 if (! sel_bb_end_p (insn) | |
657 || (single_succ_p (bb) | |
658 && single_pred_p (single_succ (bb)))) | |
659 { | |
660 insn_t succ; | |
661 | |
662 succ = (sel_bb_end_p (insn) | |
663 ? sel_bb_head (single_succ (bb)) | |
664 : NEXT_INSN (insn)); | |
665 | |
666 if (INSN_SEQNO (succ) > 0 | |
667 && INSN_SEQNO (succ) <= orig_max_seqno | |
668 && INSN_SCHED_TIMES (succ) <= 0) | |
669 { | |
670 FENCE_INSN (fence) = succ; | |
671 move_fence_to_fences (old_fences, new_fences); | |
672 | |
673 if (sched_verbose >= 1) | |
674 sel_print ("Fence %d continues as %d[%d] (state continue)\n", | |
675 INSN_UID (insn), INSN_UID (succ), BLOCK_NUM (succ)); | |
676 } | |
677 return; | |
678 } | |
679 | |
680 /* Otherwise copy fence's structures to (possibly) multiple successors. */ | |
681 FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
682 { | |
683 int seqno = INSN_SEQNO (succ); | |
684 | |
685 if (0 < seqno && seqno <= orig_max_seqno | |
686 && (pipelining_p || INSN_SCHED_TIMES (succ) <= 0)) | |
687 { | |
688 bool b = (in_same_ebb_p (insn, succ) | |
689 || in_fallthru_bb_p (insn, succ)); | |
690 | |
691 if (sched_verbose >= 1) | |
692 sel_print ("Fence %d continues as %d[%d] (state %s)\n", | |
693 INSN_UID (insn), INSN_UID (succ), | |
694 BLOCK_NUM (succ), b ? "continue" : "reset"); | |
695 | |
696 if (b) | |
697 add_dirty_fence_to_fences (new_fences, succ, fence); | |
698 else | |
699 { | |
700 /* Mark block of the SUCC as head of the new ebb. */ | |
701 bitmap_set_bit (forced_ebb_heads, BLOCK_NUM (succ)); | |
702 add_clean_fence_to_fences (new_fences, succ, fence); | |
703 } | |
704 } | |
705 } | |
706 } | |
707 | |
708 | |
709 /* Functions to support substitution. */ | |
710 | |
711 /* Returns whether INSN with dependence status DS is eligible for | |
712 substitution, i.e. it's a copy operation x := y, and RHS that is | |
713 moved up through this insn should be substituted. */ | |
714 static bool | |
715 can_substitute_through_p (insn_t insn, ds_t ds) | |
716 { | |
717 /* We can substitute only true dependencies. */ | |
718 if ((ds & DEP_OUTPUT) | |
719 || (ds & DEP_ANTI) | |
720 || ! INSN_RHS (insn) | |
721 || ! INSN_LHS (insn)) | |
722 return false; | |
723 | |
724 /* Now we just need to make sure the INSN_RHS consists of only one | |
725 simple REG rtx. */ | |
726 if (REG_P (INSN_LHS (insn)) | |
727 && REG_P (INSN_RHS (insn))) | |
728 return true; | |
729 return false; | |
730 } | |
731 | |
732 /* Substitute all occurences of INSN's destination in EXPR' vinsn with INSN's | |
733 source (if INSN is eligible for substitution). Returns TRUE if | |
734 substitution was actually performed, FALSE otherwise. Substitution might | |
735 be not performed because it's either EXPR' vinsn doesn't contain INSN's | |
736 destination or the resulting insn is invalid for the target machine. | |
737 When UNDO is true, perform unsubstitution instead (the difference is in | |
738 the part of rtx on which validate_replace_rtx is called). */ | |
739 static bool | |
740 substitute_reg_in_expr (expr_t expr, insn_t insn, bool undo) | |
741 { | |
742 rtx *where; | |
743 bool new_insn_valid; | |
744 vinsn_t *vi = &EXPR_VINSN (expr); | |
745 bool has_rhs = VINSN_RHS (*vi) != NULL; | |
746 rtx old, new_rtx; | |
747 | |
748 /* Do not try to replace in SET_DEST. Although we'll choose new | |
749 register for the RHS, we don't want to change RHS' original reg. | |
750 If the insn is not SET, we may still be able to substitute something | |
751 in it, and if we're here (don't have deps), it doesn't write INSN's | |
752 dest. */ | |
753 where = (has_rhs | |
754 ? &VINSN_RHS (*vi) | |
755 : &PATTERN (VINSN_INSN_RTX (*vi))); | |
756 old = undo ? INSN_RHS (insn) : INSN_LHS (insn); | |
757 | |
758 /* Substitute if INSN has a form of x:=y and LHS(INSN) occurs in *VI. */ | |
759 if (rtx_ok_for_substitution_p (old, *where)) | |
760 { | |
761 rtx new_insn; | |
762 rtx *where_replace; | |
763 | |
764 /* We should copy these rtxes before substitution. */ | |
765 new_rtx = copy_rtx (undo ? INSN_LHS (insn) : INSN_RHS (insn)); | |
766 new_insn = create_copy_of_insn_rtx (VINSN_INSN_RTX (*vi)); | |
767 | |
768 /* Where we'll replace. | |
769 WHERE_REPLACE should point inside NEW_INSN, so INSN_RHS couldn't be | |
770 used instead of SET_SRC. */ | |
771 where_replace = (has_rhs | |
772 ? &SET_SRC (PATTERN (new_insn)) | |
773 : &PATTERN (new_insn)); | |
774 | |
775 new_insn_valid | |
776 = validate_replace_rtx_part_nosimplify (old, new_rtx, where_replace, | |
777 new_insn); | |
778 | |
779 /* ??? Actually, constrain_operands result depends upon choice of | |
780 destination register. E.g. if we allow single register to be an rhs, | |
781 and if we try to move dx=ax(as rhs) through ax=dx, we'll result | |
782 in invalid insn dx=dx, so we'll loose this rhs here. | |
783 Just can't come up with significant testcase for this, so just | |
784 leaving it for now. */ | |
785 if (new_insn_valid) | |
786 { | |
787 change_vinsn_in_expr (expr, | |
788 create_vinsn_from_insn_rtx (new_insn, false)); | |
789 | |
790 /* Do not allow clobbering the address register of speculative | |
791 insns. */ | |
792 if ((EXPR_SPEC_DONE_DS (expr) & SPECULATIVE) | |
793 && bitmap_bit_p (VINSN_REG_USES (EXPR_VINSN (expr)), | |
794 expr_dest_regno (expr))) | |
795 EXPR_TARGET_AVAILABLE (expr) = false; | |
796 | |
797 return true; | |
798 } | |
799 else | |
800 return false; | |
801 } | |
802 else | |
803 return false; | |
804 } | |
805 | |
806 /* Helper function for count_occurences_equiv. */ | |
807 static int | |
808 count_occurrences_1 (rtx *cur_rtx, void *arg) | |
809 { | |
810 rtx_search_arg_p p = (rtx_search_arg_p) arg; | |
811 | |
812 /* The last param FOR_GCSE is true, because otherwise it performs excessive | |
813 substitutions like | |
814 r8 = r33 | |
815 r16 = r33 | |
816 for the last insn it presumes r33 equivalent to r8, so it changes it to | |
817 r33. Actually, there's no change, but it spoils debugging. */ | |
818 if (exp_equiv_p (*cur_rtx, p->x, 0, true)) | |
819 { | |
820 /* Bail out if we occupy more than one register. */ | |
821 if (REG_P (*cur_rtx) | |
822 && HARD_REGISTER_P (*cur_rtx) | |
823 && hard_regno_nregs[REGNO(*cur_rtx)][GET_MODE (*cur_rtx)] > 1) | |
824 { | |
825 p->n = 0; | |
826 return 1; | |
827 } | |
828 | |
829 p->n++; | |
830 | |
831 /* Do not traverse subexprs. */ | |
832 return -1; | |
833 } | |
834 | |
835 if (GET_CODE (*cur_rtx) == SUBREG | |
836 && REG_P (p->x) | |
837 && REGNO (SUBREG_REG (*cur_rtx)) == REGNO (p->x)) | |
838 { | |
839 /* ??? Do not support substituting regs inside subregs. In that case, | |
840 simplify_subreg will be called by validate_replace_rtx, and | |
841 unsubstitution will fail later. */ | |
842 p->n = 0; | |
843 return 1; | |
844 } | |
845 | |
846 /* Continue search. */ | |
847 return 0; | |
848 } | |
849 | |
850 /* Return the number of places WHAT appears within WHERE. | |
851 Bail out when we found a reference occupying several hard registers. */ | |
852 static int | |
853 count_occurrences_equiv (rtx what, rtx where) | |
854 { | |
855 struct rtx_search_arg arg; | |
856 | |
857 arg.x = what; | |
858 arg.n = 0; | |
859 | |
860 for_each_rtx (&where, &count_occurrences_1, (void *) &arg); | |
861 | |
862 return arg.n; | |
863 } | |
864 | |
865 /* Returns TRUE if WHAT is found in WHERE rtx tree. */ | |
866 static bool | |
867 rtx_ok_for_substitution_p (rtx what, rtx where) | |
868 { | |
869 return (count_occurrences_equiv (what, where) > 0); | |
870 } | |
871 | |
872 | |
873 /* Functions to support register renaming. */ | |
874 | |
875 /* Substitute VI's set source with REGNO. Returns newly created pattern | |
876 that has REGNO as its source. */ | |
877 static rtx | |
878 create_insn_rtx_with_rhs (vinsn_t vi, rtx rhs_rtx) | |
879 { | |
880 rtx lhs_rtx; | |
881 rtx pattern; | |
882 rtx insn_rtx; | |
883 | |
884 lhs_rtx = copy_rtx (VINSN_LHS (vi)); | |
885 | |
886 pattern = gen_rtx_SET (VOIDmode, lhs_rtx, rhs_rtx); | |
887 insn_rtx = create_insn_rtx_from_pattern (pattern, NULL_RTX); | |
888 | |
889 return insn_rtx; | |
890 } | |
891 | |
892 /* Returns whether INSN's src can be replaced with register number | |
893 NEW_SRC_REG. E.g. the following insn is valid for i386: | |
894 | |
895 (insn:HI 2205 6585 2207 727 ../../gcc/libiberty/regex.c:3337 | |
896 (set (mem/s:QI (plus:SI (plus:SI (reg/f:SI 7 sp) | |
897 (reg:SI 0 ax [orig:770 c1 ] [770])) | |
898 (const_int 288 [0x120])) [0 str S1 A8]) | |
899 (const_int 0 [0x0])) 43 {*movqi_1} (nil) | |
900 (nil)) | |
901 | |
902 But if we change (const_int 0 [0x0]) to (reg:QI 4 si), it will be invalid | |
903 because of operand constraints: | |
904 | |
905 (define_insn "*movqi_1" | |
906 [(set (match_operand:QI 0 "nonimmediate_operand" "=q,q ,q ,r,r ,?r,m") | |
907 (match_operand:QI 1 "general_operand" " q,qn,qm,q,rn,qm,qn") | |
908 )] | |
909 | |
910 So do constrain_operands here, before choosing NEW_SRC_REG as best | |
911 reg for rhs. */ | |
912 | |
913 static bool | |
914 replace_src_with_reg_ok_p (insn_t insn, rtx new_src_reg) | |
915 { | |
916 vinsn_t vi = INSN_VINSN (insn); | |
917 enum machine_mode mode; | |
918 rtx dst_loc; | |
919 bool res; | |
920 | |
921 gcc_assert (VINSN_SEPARABLE_P (vi)); | |
922 | |
923 get_dest_and_mode (insn, &dst_loc, &mode); | |
924 gcc_assert (mode == GET_MODE (new_src_reg)); | |
925 | |
926 if (REG_P (dst_loc) && REGNO (new_src_reg) == REGNO (dst_loc)) | |
927 return true; | |
928 | |
929 /* See whether SET_SRC can be replaced with this register. */ | |
930 validate_change (insn, &SET_SRC (PATTERN (insn)), new_src_reg, 1); | |
931 res = verify_changes (0); | |
932 cancel_changes (0); | |
933 | |
934 return res; | |
935 } | |
936 | |
937 /* Returns whether INSN still be valid after replacing it's DEST with | |
938 register NEW_REG. */ | |
939 static bool | |
940 replace_dest_with_reg_ok_p (insn_t insn, rtx new_reg) | |
941 { | |
942 vinsn_t vi = INSN_VINSN (insn); | |
943 bool res; | |
944 | |
945 /* We should deal here only with separable insns. */ | |
946 gcc_assert (VINSN_SEPARABLE_P (vi)); | |
947 gcc_assert (GET_MODE (VINSN_LHS (vi)) == GET_MODE (new_reg)); | |
948 | |
949 /* See whether SET_DEST can be replaced with this register. */ | |
950 validate_change (insn, &SET_DEST (PATTERN (insn)), new_reg, 1); | |
951 res = verify_changes (0); | |
952 cancel_changes (0); | |
953 | |
954 return res; | |
955 } | |
956 | |
957 /* Create a pattern with rhs of VI and lhs of LHS_RTX. */ | |
958 static rtx | |
959 create_insn_rtx_with_lhs (vinsn_t vi, rtx lhs_rtx) | |
960 { | |
961 rtx rhs_rtx; | |
962 rtx pattern; | |
963 rtx insn_rtx; | |
964 | |
965 rhs_rtx = copy_rtx (VINSN_RHS (vi)); | |
966 | |
967 pattern = gen_rtx_SET (VOIDmode, lhs_rtx, rhs_rtx); | |
968 insn_rtx = create_insn_rtx_from_pattern (pattern, NULL_RTX); | |
969 | |
970 return insn_rtx; | |
971 } | |
972 | |
973 /* Substitute lhs in the given expression EXPR for the register with number | |
974 NEW_REGNO. SET_DEST may be arbitrary rtx, not only register. */ | |
975 static void | |
976 replace_dest_with_reg_in_expr (expr_t expr, rtx new_reg) | |
977 { | |
978 rtx insn_rtx; | |
979 vinsn_t vinsn; | |
980 | |
981 insn_rtx = create_insn_rtx_with_lhs (EXPR_VINSN (expr), new_reg); | |
982 vinsn = create_vinsn_from_insn_rtx (insn_rtx, false); | |
983 | |
984 change_vinsn_in_expr (expr, vinsn); | |
985 EXPR_WAS_RENAMED (expr) = 1; | |
986 EXPR_TARGET_AVAILABLE (expr) = 1; | |
987 } | |
988 | |
989 /* Returns whether VI writes either one of the USED_REGS registers or, | |
990 if a register is a hard one, one of the UNAVAILABLE_HARD_REGS registers. */ | |
991 static bool | |
992 vinsn_writes_one_of_regs_p (vinsn_t vi, regset used_regs, | |
993 HARD_REG_SET unavailable_hard_regs) | |
994 { | |
995 unsigned regno; | |
996 reg_set_iterator rsi; | |
997 | |
998 EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (vi), 0, regno, rsi) | |
999 { | |
1000 if (REGNO_REG_SET_P (used_regs, regno)) | |
1001 return true; | |
1002 if (HARD_REGISTER_NUM_P (regno) | |
1003 && TEST_HARD_REG_BIT (unavailable_hard_regs, regno)) | |
1004 return true; | |
1005 } | |
1006 | |
1007 EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (vi), 0, regno, rsi) | |
1008 { | |
1009 if (REGNO_REG_SET_P (used_regs, regno)) | |
1010 return true; | |
1011 if (HARD_REGISTER_NUM_P (regno) | |
1012 && TEST_HARD_REG_BIT (unavailable_hard_regs, regno)) | |
1013 return true; | |
1014 } | |
1015 | |
1016 return false; | |
1017 } | |
1018 | |
1019 /* Returns register class of the output register in INSN. | |
1020 Returns NO_REGS for call insns because some targets have constraints on | |
1021 destination register of a call insn. | |
1022 | |
1023 Code adopted from regrename.c::build_def_use. */ | |
1024 static enum reg_class | |
1025 get_reg_class (rtx insn) | |
1026 { | |
1027 int alt, i, n_ops; | |
1028 | |
1029 extract_insn (insn); | |
1030 if (! constrain_operands (1)) | |
1031 fatal_insn_not_found (insn); | |
1032 preprocess_constraints (); | |
1033 alt = which_alternative; | |
1034 n_ops = recog_data.n_operands; | |
1035 | |
1036 for (i = 0; i < n_ops; ++i) | |
1037 { | |
1038 int matches = recog_op_alt[i][alt].matches; | |
1039 if (matches >= 0) | |
1040 recog_op_alt[i][alt].cl = recog_op_alt[matches][alt].cl; | |
1041 } | |
1042 | |
1043 if (asm_noperands (PATTERN (insn)) > 0) | |
1044 { | |
1045 for (i = 0; i < n_ops; i++) | |
1046 if (recog_data.operand_type[i] == OP_OUT) | |
1047 { | |
1048 rtx *loc = recog_data.operand_loc[i]; | |
1049 rtx op = *loc; | |
1050 enum reg_class cl = recog_op_alt[i][alt].cl; | |
1051 | |
1052 if (REG_P (op) | |
1053 && REGNO (op) == ORIGINAL_REGNO (op)) | |
1054 continue; | |
1055 | |
1056 return cl; | |
1057 } | |
1058 } | |
1059 else if (!CALL_P (insn)) | |
1060 { | |
1061 for (i = 0; i < n_ops + recog_data.n_dups; i++) | |
1062 { | |
1063 int opn = i < n_ops ? i : recog_data.dup_num[i - n_ops]; | |
1064 enum reg_class cl = recog_op_alt[opn][alt].cl; | |
1065 | |
1066 if (recog_data.operand_type[opn] == OP_OUT || | |
1067 recog_data.operand_type[opn] == OP_INOUT) | |
1068 return cl; | |
1069 } | |
1070 } | |
1071 | |
1072 /* Insns like | |
1073 (insn (set (reg:CCZ 17 flags) (compare:CCZ ...))) | |
1074 may result in returning NO_REGS, cause flags is written implicitly through | |
1075 CMP insn, which has no OP_OUT | OP_INOUT operands. */ | |
1076 return NO_REGS; | |
1077 } | |
1078 | |
1079 #ifdef HARD_REGNO_RENAME_OK | |
1080 /* Calculate HARD_REGNO_RENAME_OK data for REGNO. */ | |
1081 static void | |
1082 init_hard_regno_rename (int regno) | |
1083 { | |
1084 int cur_reg; | |
1085 | |
1086 SET_HARD_REG_BIT (sel_hrd.regs_for_rename[regno], regno); | |
1087 | |
1088 for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) | |
1089 { | |
1090 /* We are not interested in renaming in other regs. */ | |
1091 if (!TEST_HARD_REG_BIT (sel_hrd.regs_ever_used, cur_reg)) | |
1092 continue; | |
1093 | |
1094 if (HARD_REGNO_RENAME_OK (regno, cur_reg)) | |
1095 SET_HARD_REG_BIT (sel_hrd.regs_for_rename[regno], cur_reg); | |
1096 } | |
1097 } | |
1098 #endif | |
1099 | |
1100 /* A wrapper around HARD_REGNO_RENAME_OK that will look into the hard regs | |
1101 data first. */ | |
1102 static inline bool | |
1103 sel_hard_regno_rename_ok (int from ATTRIBUTE_UNUSED, int to ATTRIBUTE_UNUSED) | |
1104 { | |
1105 #ifdef HARD_REGNO_RENAME_OK | |
1106 /* Check whether this is all calculated. */ | |
1107 if (TEST_HARD_REG_BIT (sel_hrd.regs_for_rename[from], from)) | |
1108 return TEST_HARD_REG_BIT (sel_hrd.regs_for_rename[from], to); | |
1109 | |
1110 init_hard_regno_rename (from); | |
1111 | |
1112 return TEST_HARD_REG_BIT (sel_hrd.regs_for_rename[from], to); | |
1113 #else | |
1114 return true; | |
1115 #endif | |
1116 } | |
1117 | |
1118 /* Calculate set of registers that are capable of holding MODE. */ | |
1119 static void | |
1120 init_regs_for_mode (enum machine_mode mode) | |
1121 { | |
1122 int cur_reg; | |
1123 | |
1124 CLEAR_HARD_REG_SET (sel_hrd.regs_for_mode[mode]); | |
1125 CLEAR_HARD_REG_SET (sel_hrd.regs_for_call_clobbered[mode]); | |
1126 | |
1127 for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) | |
1128 { | |
1129 int nregs = hard_regno_nregs[cur_reg][mode]; | |
1130 int i; | |
1131 | |
1132 for (i = nregs - 1; i >= 0; --i) | |
1133 if (fixed_regs[cur_reg + i] | |
1134 || global_regs[cur_reg + i] | |
1135 /* Can't use regs which aren't saved by | |
1136 the prologue. */ | |
1137 || !TEST_HARD_REG_BIT (sel_hrd.regs_ever_used, cur_reg + i) | |
1138 #ifdef LEAF_REGISTERS | |
1139 /* We can't use a non-leaf register if we're in a | |
1140 leaf function. */ | |
1141 || (current_function_is_leaf | |
1142 && !LEAF_REGISTERS[cur_reg + i]) | |
1143 #endif | |
1144 ) | |
1145 break; | |
1146 | |
1147 if (i >= 0) | |
1148 continue; | |
1149 | |
1150 /* See whether it accepts all modes that occur in | |
1151 original insns. */ | |
1152 if (! HARD_REGNO_MODE_OK (cur_reg, mode)) | |
1153 continue; | |
1154 | |
1155 if (HARD_REGNO_CALL_PART_CLOBBERED (cur_reg, mode)) | |
1156 SET_HARD_REG_BIT (sel_hrd.regs_for_call_clobbered[mode], | |
1157 cur_reg); | |
1158 | |
1159 /* If the CUR_REG passed all the checks above, | |
1160 then it's ok. */ | |
1161 SET_HARD_REG_BIT (sel_hrd.regs_for_mode[mode], cur_reg); | |
1162 } | |
1163 | |
1164 sel_hrd.regs_for_mode_ok[mode] = true; | |
1165 } | |
1166 | |
1167 /* Init all register sets gathered in HRD. */ | |
1168 static void | |
1169 init_hard_regs_data (void) | |
1170 { | |
1171 int cur_reg = 0; | |
1172 enum machine_mode cur_mode = 0; | |
1173 | |
1174 CLEAR_HARD_REG_SET (sel_hrd.regs_ever_used); | |
1175 for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) | |
1176 if (df_regs_ever_live_p (cur_reg) || call_used_regs[cur_reg]) | |
1177 SET_HARD_REG_BIT (sel_hrd.regs_ever_used, cur_reg); | |
1178 | |
1179 /* Initialize registers that are valid based on mode when this is | |
1180 really needed. */ | |
1181 for (cur_mode = 0; cur_mode < NUM_MACHINE_MODES; cur_mode++) | |
1182 sel_hrd.regs_for_mode_ok[cur_mode] = false; | |
1183 | |
1184 /* Mark that all HARD_REGNO_RENAME_OK is not calculated. */ | |
1185 for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) | |
1186 CLEAR_HARD_REG_SET (sel_hrd.regs_for_rename[cur_reg]); | |
1187 | |
1188 #ifdef STACK_REGS | |
1189 CLEAR_HARD_REG_SET (sel_hrd.stack_regs); | |
1190 | |
1191 for (cur_reg = FIRST_STACK_REG; cur_reg <= LAST_STACK_REG; cur_reg++) | |
1192 SET_HARD_REG_BIT (sel_hrd.stack_regs, cur_reg); | |
1193 #endif | |
1194 } | |
1195 | |
1196 /* Mark hardware regs in REG_RENAME_P that are not suitable | |
1197 for renaming rhs in INSN due to hardware restrictions (register class, | |
1198 modes compatibility etc). This doesn't affect original insn's dest reg, | |
1199 if it isn't in USED_REGS. DEF is a definition insn of rhs for which the | |
1200 destination register is sought. LHS (DEF->ORIG_INSN) may be REG or MEM. | |
1201 Registers that are in used_regs are always marked in | |
1202 unavailable_hard_regs as well. */ | |
1203 | |
1204 static void | |
1205 mark_unavailable_hard_regs (def_t def, struct reg_rename *reg_rename_p, | |
1206 regset used_regs ATTRIBUTE_UNUSED) | |
1207 { | |
1208 enum machine_mode mode; | |
1209 enum reg_class cl = NO_REGS; | |
1210 rtx orig_dest; | |
1211 unsigned cur_reg, regno; | |
1212 hard_reg_set_iterator hrsi; | |
1213 | |
1214 gcc_assert (GET_CODE (PATTERN (def->orig_insn)) == SET); | |
1215 gcc_assert (reg_rename_p); | |
1216 | |
1217 orig_dest = SET_DEST (PATTERN (def->orig_insn)); | |
1218 | |
1219 /* We have decided not to rename 'mem = something;' insns, as 'something' | |
1220 is usually a register. */ | |
1221 if (!REG_P (orig_dest)) | |
1222 return; | |
1223 | |
1224 regno = REGNO (orig_dest); | |
1225 | |
1226 /* If before reload, don't try to work with pseudos. */ | |
1227 if (!reload_completed && !HARD_REGISTER_NUM_P (regno)) | |
1228 return; | |
1229 | |
1230 mode = GET_MODE (orig_dest); | |
1231 | |
1232 /* Stop when mode is not supported for renaming. Also can't proceed | |
1233 if the original register is one of the fixed_regs, global_regs or | |
1234 frame pointer. */ | |
1235 if (fixed_regs[regno] | |
1236 || global_regs[regno] | |
1237 #if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM | |
1238 || (frame_pointer_needed && regno == HARD_FRAME_POINTER_REGNUM) | |
1239 #else | |
1240 || (frame_pointer_needed && regno == FRAME_POINTER_REGNUM) | |
1241 #endif | |
1242 ) | |
1243 { | |
1244 SET_HARD_REG_SET (reg_rename_p->unavailable_hard_regs); | |
1245 | |
1246 /* Give a chance for original register, if it isn't in used_regs. */ | |
1247 if (!def->crosses_call) | |
1248 CLEAR_HARD_REG_BIT (reg_rename_p->unavailable_hard_regs, regno); | |
1249 | |
1250 return; | |
1251 } | |
1252 | |
1253 /* If something allocated on stack in this function, mark frame pointer | |
1254 register unavailable, considering also modes. | |
1255 FIXME: it is enough to do this once per all original defs. */ | |
1256 if (frame_pointer_needed) | |
1257 { | |
1258 int i; | |
1259 | |
1260 for (i = hard_regno_nregs[FRAME_POINTER_REGNUM][Pmode]; i--;) | |
1261 SET_HARD_REG_BIT (reg_rename_p->unavailable_hard_regs, | |
1262 FRAME_POINTER_REGNUM + i); | |
1263 | |
1264 #if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM | |
1265 for (i = hard_regno_nregs[HARD_FRAME_POINTER_REGNUM][Pmode]; i--;) | |
1266 SET_HARD_REG_BIT (reg_rename_p->unavailable_hard_regs, | |
1267 HARD_FRAME_POINTER_REGNUM + i); | |
1268 #endif | |
1269 } | |
1270 | |
1271 #ifdef STACK_REGS | |
1272 /* For the stack registers the presence of FIRST_STACK_REG in USED_REGS | |
1273 is equivalent to as if all stack regs were in this set. | |
1274 I.e. no stack register can be renamed, and even if it's an original | |
1275 register here we make sure it won't be lifted over it's previous def | |
1276 (it's previous def will appear as if it's a FIRST_STACK_REG def. | |
1277 The HARD_REGNO_RENAME_OK covers other cases in condition below. */ | |
1278 if (IN_RANGE (REGNO (orig_dest), FIRST_STACK_REG, LAST_STACK_REG) | |
1279 && REGNO_REG_SET_P (used_regs, FIRST_STACK_REG)) | |
1280 IOR_HARD_REG_SET (reg_rename_p->unavailable_hard_regs, | |
1281 sel_hrd.stack_regs); | |
1282 #endif | |
1283 | |
1284 /* If there's a call on this path, make regs from call_used_reg_set | |
1285 unavailable. */ | |
1286 if (def->crosses_call) | |
1287 IOR_HARD_REG_SET (reg_rename_p->unavailable_hard_regs, | |
1288 call_used_reg_set); | |
1289 | |
1290 /* Stop here before reload: we need FRAME_REGS, STACK_REGS, and crosses_call, | |
1291 but not register classes. */ | |
1292 if (!reload_completed) | |
1293 return; | |
1294 | |
1295 /* Leave regs as 'available' only from the current | |
1296 register class. */ | |
1297 cl = get_reg_class (def->orig_insn); | |
1298 gcc_assert (cl != NO_REGS); | |
1299 COPY_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1300 reg_class_contents[cl]); | |
1301 | |
1302 /* Leave only registers available for this mode. */ | |
1303 if (!sel_hrd.regs_for_mode_ok[mode]) | |
1304 init_regs_for_mode (mode); | |
1305 AND_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1306 sel_hrd.regs_for_mode[mode]); | |
1307 | |
1308 /* Exclude registers that are partially call clobbered. */ | |
1309 if (def->crosses_call | |
1310 && ! HARD_REGNO_CALL_PART_CLOBBERED (regno, mode)) | |
1311 AND_COMPL_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1312 sel_hrd.regs_for_call_clobbered[mode]); | |
1313 | |
1314 /* Leave only those that are ok to rename. */ | |
1315 EXECUTE_IF_SET_IN_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1316 0, cur_reg, hrsi) | |
1317 { | |
1318 int nregs; | |
1319 int i; | |
1320 | |
1321 nregs = hard_regno_nregs[cur_reg][mode]; | |
1322 gcc_assert (nregs > 0); | |
1323 | |
1324 for (i = nregs - 1; i >= 0; --i) | |
1325 if (! sel_hard_regno_rename_ok (regno + i, cur_reg + i)) | |
1326 break; | |
1327 | |
1328 if (i >= 0) | |
1329 CLEAR_HARD_REG_BIT (reg_rename_p->available_for_renaming, | |
1330 cur_reg); | |
1331 } | |
1332 | |
1333 AND_COMPL_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1334 reg_rename_p->unavailable_hard_regs); | |
1335 | |
1336 /* Regno is always ok from the renaming part of view, but it really | |
1337 could be in *unavailable_hard_regs already, so set it here instead | |
1338 of there. */ | |
1339 SET_HARD_REG_BIT (reg_rename_p->available_for_renaming, regno); | |
1340 } | |
1341 | |
1342 /* reg_rename_tick[REG1] > reg_rename_tick[REG2] if REG1 was chosen as the | |
1343 best register more recently than REG2. */ | |
1344 static int reg_rename_tick[FIRST_PSEUDO_REGISTER]; | |
1345 | |
1346 /* Indicates the number of times renaming happened before the current one. */ | |
1347 static int reg_rename_this_tick; | |
1348 | |
1349 /* Choose the register among free, that is suitable for storing | |
1350 the rhs value. | |
1351 | |
1352 ORIGINAL_INSNS is the list of insns where the operation (rhs) | |
1353 originally appears. There could be multiple original operations | |
1354 for single rhs since we moving it up and merging along different | |
1355 paths. | |
1356 | |
1357 Some code is adapted from regrename.c (regrename_optimize). | |
1358 If original register is available, function returns it. | |
1359 Otherwise it performs the checks, so the new register should | |
1360 comply with the following: | |
1361 - it should not violate any live ranges (such registers are in | |
1362 REG_RENAME_P->available_for_renaming set); | |
1363 - it should not be in the HARD_REGS_USED regset; | |
1364 - it should be in the class compatible with original uses; | |
1365 - it should not be clobbered through reference with different mode; | |
1366 - if we're in the leaf function, then the new register should | |
1367 not be in the LEAF_REGISTERS; | |
1368 - etc. | |
1369 | |
1370 If several registers meet the conditions, the register with smallest | |
1371 tick is returned to achieve more even register allocation. | |
1372 | |
1373 If original register seems to be ok, we set *IS_ORIG_REG_P_PTR to true. | |
1374 | |
1375 If no register satisfies the above conditions, NULL_RTX is returned. */ | |
1376 static rtx | |
1377 choose_best_reg_1 (HARD_REG_SET hard_regs_used, | |
1378 struct reg_rename *reg_rename_p, | |
1379 def_list_t original_insns, bool *is_orig_reg_p_ptr) | |
1380 { | |
1381 int best_new_reg; | |
1382 unsigned cur_reg; | |
1383 enum machine_mode mode = VOIDmode; | |
1384 unsigned regno, i, n; | |
1385 hard_reg_set_iterator hrsi; | |
1386 def_list_iterator di; | |
1387 def_t def; | |
1388 | |
1389 /* If original register is available, return it. */ | |
1390 *is_orig_reg_p_ptr = true; | |
1391 | |
1392 FOR_EACH_DEF (def, di, original_insns) | |
1393 { | |
1394 rtx orig_dest = SET_DEST (PATTERN (def->orig_insn)); | |
1395 | |
1396 gcc_assert (REG_P (orig_dest)); | |
1397 | |
1398 /* Check that all original operations have the same mode. | |
1399 This is done for the next loop; if we'd return from this | |
1400 loop, we'd check only part of them, but in this case | |
1401 it doesn't matter. */ | |
1402 if (mode == VOIDmode) | |
1403 mode = GET_MODE (orig_dest); | |
1404 gcc_assert (mode == GET_MODE (orig_dest)); | |
1405 | |
1406 regno = REGNO (orig_dest); | |
1407 for (i = 0, n = hard_regno_nregs[regno][mode]; i < n; i++) | |
1408 if (TEST_HARD_REG_BIT (hard_regs_used, regno + i)) | |
1409 break; | |
1410 | |
1411 /* All hard registers are available. */ | |
1412 if (i == n) | |
1413 { | |
1414 gcc_assert (mode != VOIDmode); | |
1415 | |
1416 /* Hard registers should not be shared. */ | |
1417 return gen_rtx_REG (mode, regno); | |
1418 } | |
1419 } | |
1420 | |
1421 *is_orig_reg_p_ptr = false; | |
1422 best_new_reg = -1; | |
1423 | |
1424 /* Among all available regs choose the register that was | |
1425 allocated earliest. */ | |
1426 EXECUTE_IF_SET_IN_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1427 0, cur_reg, hrsi) | |
1428 if (! TEST_HARD_REG_BIT (hard_regs_used, cur_reg)) | |
1429 { | |
1430 /* All hard registers are available. */ | |
1431 if (best_new_reg < 0 | |
1432 || reg_rename_tick[cur_reg] < reg_rename_tick[best_new_reg]) | |
1433 { | |
1434 best_new_reg = cur_reg; | |
1435 | |
1436 /* Return immediately when we know there's no better reg. */ | |
1437 if (! reg_rename_tick[best_new_reg]) | |
1438 break; | |
1439 } | |
1440 } | |
1441 | |
1442 if (best_new_reg >= 0) | |
1443 { | |
1444 /* Use the check from the above loop. */ | |
1445 gcc_assert (mode != VOIDmode); | |
1446 return gen_rtx_REG (mode, best_new_reg); | |
1447 } | |
1448 | |
1449 return NULL_RTX; | |
1450 } | |
1451 | |
1452 /* A wrapper around choose_best_reg_1 () to verify that we make correct | |
1453 assumptions about available registers in the function. */ | |
1454 static rtx | |
1455 choose_best_reg (HARD_REG_SET hard_regs_used, struct reg_rename *reg_rename_p, | |
1456 def_list_t original_insns, bool *is_orig_reg_p_ptr) | |
1457 { | |
1458 rtx best_reg = choose_best_reg_1 (hard_regs_used, reg_rename_p, | |
1459 original_insns, is_orig_reg_p_ptr); | |
1460 | |
1461 gcc_assert (best_reg == NULL_RTX | |
1462 || TEST_HARD_REG_BIT (sel_hrd.regs_ever_used, REGNO (best_reg))); | |
1463 | |
1464 return best_reg; | |
1465 } | |
1466 | |
1467 /* Choose the pseudo register for storing rhs value. As this is supposed | |
1468 to work before reload, we return either the original register or make | |
1469 the new one. The parameters are the same that in choose_nest_reg_1 | |
1470 functions, except that USED_REGS may contain pseudos. | |
1471 If we work with hard regs, check also REG_RENAME_P->UNAVAILABLE_HARD_REGS. | |
1472 | |
1473 TODO: take into account register pressure while doing this. Up to this | |
1474 moment, this function would never return NULL for pseudos, but we should | |
1475 not rely on this. */ | |
1476 static rtx | |
1477 choose_best_pseudo_reg (regset used_regs, | |
1478 struct reg_rename *reg_rename_p, | |
1479 def_list_t original_insns, bool *is_orig_reg_p_ptr) | |
1480 { | |
1481 def_list_iterator i; | |
1482 def_t def; | |
1483 enum machine_mode mode = VOIDmode; | |
1484 bool bad_hard_regs = false; | |
1485 | |
1486 /* We should not use this after reload. */ | |
1487 gcc_assert (!reload_completed); | |
1488 | |
1489 /* If original register is available, return it. */ | |
1490 *is_orig_reg_p_ptr = true; | |
1491 | |
1492 FOR_EACH_DEF (def, i, original_insns) | |
1493 { | |
1494 rtx dest = SET_DEST (PATTERN (def->orig_insn)); | |
1495 int orig_regno; | |
1496 | |
1497 gcc_assert (REG_P (dest)); | |
1498 | |
1499 /* Check that all original operations have the same mode. */ | |
1500 if (mode == VOIDmode) | |
1501 mode = GET_MODE (dest); | |
1502 else | |
1503 gcc_assert (mode == GET_MODE (dest)); | |
1504 orig_regno = REGNO (dest); | |
1505 | |
1506 if (!REGNO_REG_SET_P (used_regs, orig_regno)) | |
1507 { | |
1508 if (orig_regno < FIRST_PSEUDO_REGISTER) | |
1509 { | |
1510 gcc_assert (df_regs_ever_live_p (orig_regno)); | |
1511 | |
1512 /* For hard registers, we have to check hardware imposed | |
1513 limitations (frame/stack registers, calls crossed). */ | |
1514 if (!TEST_HARD_REG_BIT (reg_rename_p->unavailable_hard_regs, | |
1515 orig_regno)) | |
1516 { | |
1517 /* Don't let register cross a call if it doesn't already | |
1518 cross one. This condition is written in accordance with | |
1519 that in sched-deps.c sched_analyze_reg(). */ | |
1520 if (!reg_rename_p->crosses_call | |
1521 || REG_N_CALLS_CROSSED (orig_regno) > 0) | |
1522 return gen_rtx_REG (mode, orig_regno); | |
1523 } | |
1524 | |
1525 bad_hard_regs = true; | |
1526 } | |
1527 else | |
1528 return dest; | |
1529 } | |
1530 } | |
1531 | |
1532 *is_orig_reg_p_ptr = false; | |
1533 | |
1534 /* We had some original hard registers that couldn't be used. | |
1535 Those were likely special. Don't try to create a pseudo. */ | |
1536 if (bad_hard_regs) | |
1537 return NULL_RTX; | |
1538 | |
1539 /* We haven't found a register from original operations. Get a new one. | |
1540 FIXME: control register pressure somehow. */ | |
1541 { | |
1542 rtx new_reg = gen_reg_rtx (mode); | |
1543 | |
1544 gcc_assert (mode != VOIDmode); | |
1545 | |
1546 max_regno = max_reg_num (); | |
1547 maybe_extend_reg_info_p (); | |
1548 REG_N_CALLS_CROSSED (REGNO (new_reg)) = reg_rename_p->crosses_call ? 1 : 0; | |
1549 | |
1550 return new_reg; | |
1551 } | |
1552 } | |
1553 | |
1554 /* True when target of EXPR is available due to EXPR_TARGET_AVAILABLE, | |
1555 USED_REGS and REG_RENAME_P->UNAVAILABLE_HARD_REGS. */ | |
1556 static void | |
1557 verify_target_availability (expr_t expr, regset used_regs, | |
1558 struct reg_rename *reg_rename_p) | |
1559 { | |
1560 unsigned n, i, regno; | |
1561 enum machine_mode mode; | |
1562 bool target_available, live_available, hard_available; | |
1563 | |
1564 if (!REG_P (EXPR_LHS (expr)) || EXPR_TARGET_AVAILABLE (expr) < 0) | |
1565 return; | |
1566 | |
1567 regno = expr_dest_regno (expr); | |
1568 mode = GET_MODE (EXPR_LHS (expr)); | |
1569 target_available = EXPR_TARGET_AVAILABLE (expr) == 1; | |
1570 n = reload_completed ? hard_regno_nregs[regno][mode] : 1; | |
1571 | |
1572 live_available = hard_available = true; | |
1573 for (i = 0; i < n; i++) | |
1574 { | |
1575 if (bitmap_bit_p (used_regs, regno + i)) | |
1576 live_available = false; | |
1577 if (TEST_HARD_REG_BIT (reg_rename_p->unavailable_hard_regs, regno + i)) | |
1578 hard_available = false; | |
1579 } | |
1580 | |
1581 /* When target is not available, it may be due to hard register | |
1582 restrictions, e.g. crosses calls, so we check hard_available too. */ | |
1583 if (target_available) | |
1584 gcc_assert (live_available); | |
1585 else | |
1586 /* Check only if we haven't scheduled something on the previous fence, | |
1587 cause due to MAX_SOFTWARE_LOOKAHEAD_WINDOW_SIZE issues | |
1588 and having more than one fence, we may end having targ_un in a block | |
1589 in which successors target register is actually available. | |
1590 | |
1591 The last condition handles the case when a dependence from a call insn | |
1592 was created in sched-deps.c for insns with destination registers that | |
1593 never crossed a call before, but do cross one after our code motion. | |
1594 | |
1595 FIXME: in the latter case, we just uselessly called find_used_regs, | |
1596 because we can't move this expression with any other register | |
1597 as well. */ | |
1598 gcc_assert (scheduled_something_on_previous_fence || !live_available | |
1599 || !hard_available | |
1600 || (!reload_completed && reg_rename_p->crosses_call | |
1601 && REG_N_CALLS_CROSSED (regno) == 0)); | |
1602 } | |
1603 | |
1604 /* Collect unavailable registers due to liveness for EXPR from BNDS | |
1605 into USED_REGS. Save additional information about available | |
1606 registers and unavailable due to hardware restriction registers | |
1607 into REG_RENAME_P structure. Save original insns into ORIGINAL_INSNS | |
1608 list. */ | |
1609 static void | |
1610 collect_unavailable_regs_from_bnds (expr_t expr, blist_t bnds, regset used_regs, | |
1611 struct reg_rename *reg_rename_p, | |
1612 def_list_t *original_insns) | |
1613 { | |
1614 for (; bnds; bnds = BLIST_NEXT (bnds)) | |
1615 { | |
1616 bool res; | |
1617 av_set_t orig_ops = NULL; | |
1618 bnd_t bnd = BLIST_BND (bnds); | |
1619 | |
1620 /* If the chosen best expr doesn't belong to current boundary, | |
1621 skip it. */ | |
1622 if (!av_set_is_in_p (BND_AV1 (bnd), EXPR_VINSN (expr))) | |
1623 continue; | |
1624 | |
1625 /* Put in ORIG_OPS all exprs from this boundary that became | |
1626 RES on top. */ | |
1627 orig_ops = find_sequential_best_exprs (bnd, expr, false); | |
1628 | |
1629 /* Compute used regs and OR it into the USED_REGS. */ | |
1630 res = find_used_regs (BND_TO (bnd), orig_ops, used_regs, | |
1631 reg_rename_p, original_insns); | |
1632 | |
1633 /* FIXME: the assert is true until we'd have several boundaries. */ | |
1634 gcc_assert (res); | |
1635 av_set_clear (&orig_ops); | |
1636 } | |
1637 } | |
1638 | |
1639 /* Return TRUE if it is possible to replace LHSes of ORIG_INSNS with BEST_REG. | |
1640 If BEST_REG is valid, replace LHS of EXPR with it. */ | |
1641 static bool | |
1642 try_replace_dest_reg (ilist_t orig_insns, rtx best_reg, expr_t expr) | |
1643 { | |
1644 if (expr_dest_regno (expr) == REGNO (best_reg)) | |
1645 { | |
1646 EXPR_TARGET_AVAILABLE (expr) = 1; | |
1647 return true; | |
1648 } | |
1649 | |
1650 gcc_assert (orig_insns); | |
1651 | |
1652 /* Try whether we'll be able to generate the insn | |
1653 'dest := best_reg' at the place of the original operation. */ | |
1654 for (; orig_insns; orig_insns = ILIST_NEXT (orig_insns)) | |
1655 { | |
1656 insn_t orig_insn = DEF_LIST_DEF (orig_insns)->orig_insn; | |
1657 | |
1658 gcc_assert (EXPR_SEPARABLE_P (INSN_EXPR (orig_insn))); | |
1659 | |
1660 if (!replace_src_with_reg_ok_p (orig_insn, best_reg) | |
1661 || !replace_dest_with_reg_ok_p (orig_insn, best_reg)) | |
1662 return false; | |
1663 } | |
1664 | |
1665 /* Make sure that EXPR has the right destination | |
1666 register. */ | |
1667 replace_dest_with_reg_in_expr (expr, best_reg); | |
1668 return true; | |
1669 } | |
1670 | |
1671 /* Select and assign best register to EXPR searching from BNDS. | |
1672 Set *IS_ORIG_REG_P to TRUE if original register was selected. | |
1673 Return FALSE if no register can be chosen, which could happen when: | |
1674 * EXPR_SEPARABLE_P is true but we were unable to find suitable register; | |
1675 * EXPR_SEPARABLE_P is false but the insn sets/clobbers one of the registers | |
1676 that are used on the moving path. */ | |
1677 static bool | |
1678 find_best_reg_for_expr (expr_t expr, blist_t bnds, bool *is_orig_reg_p) | |
1679 { | |
1680 static struct reg_rename reg_rename_data; | |
1681 | |
1682 regset used_regs; | |
1683 def_list_t original_insns = NULL; | |
1684 bool reg_ok; | |
1685 | |
1686 *is_orig_reg_p = false; | |
1687 | |
1688 /* Don't bother to do anything if this insn doesn't set any registers. */ | |
1689 if (bitmap_empty_p (VINSN_REG_SETS (EXPR_VINSN (expr))) | |
1690 && bitmap_empty_p (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)))) | |
1691 return true; | |
1692 | |
1693 used_regs = get_clear_regset_from_pool (); | |
1694 CLEAR_HARD_REG_SET (reg_rename_data.unavailable_hard_regs); | |
1695 | |
1696 collect_unavailable_regs_from_bnds (expr, bnds, used_regs, ®_rename_data, | |
1697 &original_insns); | |
1698 | |
1699 #ifdef ENABLE_CHECKING | |
1700 /* If after reload, make sure we're working with hard regs here. */ | |
1701 if (reload_completed) | |
1702 { | |
1703 reg_set_iterator rsi; | |
1704 unsigned i; | |
1705 | |
1706 EXECUTE_IF_SET_IN_REG_SET (used_regs, FIRST_PSEUDO_REGISTER, i, rsi) | |
1707 gcc_unreachable (); | |
1708 } | |
1709 #endif | |
1710 | |
1711 if (EXPR_SEPARABLE_P (expr)) | |
1712 { | |
1713 rtx best_reg = NULL_RTX; | |
1714 /* Check that we have computed availability of a target register | |
1715 correctly. */ | |
1716 verify_target_availability (expr, used_regs, ®_rename_data); | |
1717 | |
1718 /* Turn everything in hard regs after reload. */ | |
1719 if (reload_completed) | |
1720 { | |
1721 HARD_REG_SET hard_regs_used; | |
1722 REG_SET_TO_HARD_REG_SET (hard_regs_used, used_regs); | |
1723 | |
1724 /* Join hard registers unavailable due to register class | |
1725 restrictions and live range intersection. */ | |
1726 IOR_HARD_REG_SET (hard_regs_used, | |
1727 reg_rename_data.unavailable_hard_regs); | |
1728 | |
1729 best_reg = choose_best_reg (hard_regs_used, ®_rename_data, | |
1730 original_insns, is_orig_reg_p); | |
1731 } | |
1732 else | |
1733 best_reg = choose_best_pseudo_reg (used_regs, ®_rename_data, | |
1734 original_insns, is_orig_reg_p); | |
1735 | |
1736 if (!best_reg) | |
1737 reg_ok = false; | |
1738 else if (*is_orig_reg_p) | |
1739 { | |
1740 /* In case of unification BEST_REG may be different from EXPR's LHS | |
1741 when EXPR's LHS is unavailable, and there is another LHS among | |
1742 ORIGINAL_INSNS. */ | |
1743 reg_ok = try_replace_dest_reg (original_insns, best_reg, expr); | |
1744 } | |
1745 else | |
1746 { | |
1747 /* Forbid renaming of low-cost insns. */ | |
1748 if (sel_vinsn_cost (EXPR_VINSN (expr)) < 2) | |
1749 reg_ok = false; | |
1750 else | |
1751 reg_ok = try_replace_dest_reg (original_insns, best_reg, expr); | |
1752 } | |
1753 } | |
1754 else | |
1755 { | |
1756 /* If !EXPR_SCHEDULE_AS_RHS (EXPR), just make sure INSN doesn't set | |
1757 any of the HARD_REGS_USED set. */ | |
1758 if (vinsn_writes_one_of_regs_p (EXPR_VINSN (expr), used_regs, | |
1759 reg_rename_data.unavailable_hard_regs)) | |
1760 { | |
1761 reg_ok = false; | |
1762 gcc_assert (EXPR_TARGET_AVAILABLE (expr) <= 0); | |
1763 } | |
1764 else | |
1765 { | |
1766 reg_ok = true; | |
1767 gcc_assert (EXPR_TARGET_AVAILABLE (expr) != 0); | |
1768 } | |
1769 } | |
1770 | |
1771 ilist_clear (&original_insns); | |
1772 return_regset_to_pool (used_regs); | |
1773 | |
1774 return reg_ok; | |
1775 } | |
1776 | |
1777 | |
1778 /* Return true if dependence described by DS can be overcomed. */ | |
1779 static bool | |
1780 can_speculate_dep_p (ds_t ds) | |
1781 { | |
1782 if (spec_info == NULL) | |
1783 return false; | |
1784 | |
1785 /* Leave only speculative data. */ | |
1786 ds &= SPECULATIVE; | |
1787 | |
1788 if (ds == 0) | |
1789 return false; | |
1790 | |
1791 { | |
1792 /* FIXME: make sched-deps.c produce only those non-hard dependencies, | |
1793 that we can overcome. */ | |
1794 ds_t spec_mask = spec_info->mask; | |
1795 | |
1796 if ((ds & spec_mask) != ds) | |
1797 return false; | |
1798 } | |
1799 | |
1800 if (ds_weak (ds) < spec_info->data_weakness_cutoff) | |
1801 return false; | |
1802 | |
1803 return true; | |
1804 } | |
1805 | |
1806 /* Get a speculation check instruction. | |
1807 C_EXPR is a speculative expression, | |
1808 CHECK_DS describes speculations that should be checked, | |
1809 ORIG_INSN is the original non-speculative insn in the stream. */ | |
1810 static insn_t | |
1811 create_speculation_check (expr_t c_expr, ds_t check_ds, insn_t orig_insn) | |
1812 { | |
1813 rtx check_pattern; | |
1814 rtx insn_rtx; | |
1815 insn_t insn; | |
1816 basic_block recovery_block; | |
1817 rtx label; | |
1818 | |
1819 /* Create a recovery block if target is going to emit branchy check, or if | |
1820 ORIG_INSN was speculative already. */ | |
1821 if (targetm.sched.needs_block_p (check_ds) | |
1822 || EXPR_SPEC_DONE_DS (INSN_EXPR (orig_insn)) != 0) | |
1823 { | |
1824 recovery_block = sel_create_recovery_block (orig_insn); | |
1825 label = BB_HEAD (recovery_block); | |
1826 } | |
1827 else | |
1828 { | |
1829 recovery_block = NULL; | |
1830 label = NULL_RTX; | |
1831 } | |
1832 | |
1833 /* Get pattern of the check. */ | |
1834 check_pattern = targetm.sched.gen_spec_check (EXPR_INSN_RTX (c_expr), label, | |
1835 check_ds); | |
1836 | |
1837 gcc_assert (check_pattern != NULL); | |
1838 | |
1839 /* Emit check. */ | |
1840 insn_rtx = create_insn_rtx_from_pattern (check_pattern, label); | |
1841 | |
1842 insn = sel_gen_insn_from_rtx_after (insn_rtx, INSN_EXPR (orig_insn), | |
1843 INSN_SEQNO (orig_insn), orig_insn); | |
1844 | |
1845 /* Make check to be non-speculative. */ | |
1846 EXPR_SPEC_DONE_DS (INSN_EXPR (insn)) = 0; | |
1847 INSN_SPEC_CHECKED_DS (insn) = check_ds; | |
1848 | |
1849 /* Decrease priority of check by difference of load/check instruction | |
1850 latencies. */ | |
1851 EXPR_PRIORITY (INSN_EXPR (insn)) -= (sel_vinsn_cost (INSN_VINSN (orig_insn)) | |
1852 - sel_vinsn_cost (INSN_VINSN (insn))); | |
1853 | |
1854 /* Emit copy of original insn (though with replaced target register, | |
1855 if needed) to the recovery block. */ | |
1856 if (recovery_block != NULL) | |
1857 { | |
1858 rtx twin_rtx; | |
1859 insn_t twin; | |
1860 | |
1861 twin_rtx = copy_rtx (PATTERN (EXPR_INSN_RTX (c_expr))); | |
1862 twin_rtx = create_insn_rtx_from_pattern (twin_rtx, NULL_RTX); | |
1863 twin = sel_gen_recovery_insn_from_rtx_after (twin_rtx, | |
1864 INSN_EXPR (orig_insn), | |
1865 INSN_SEQNO (insn), | |
1866 bb_note (recovery_block)); | |
1867 } | |
1868 | |
1869 /* If we've generated a data speculation check, make sure | |
1870 that all the bookkeeping instruction we'll create during | |
1871 this move_op () will allocate an ALAT entry so that the | |
1872 check won't fail. | |
1873 In case of control speculation we must convert C_EXPR to control | |
1874 speculative mode, because failing to do so will bring us an exception | |
1875 thrown by the non-control-speculative load. */ | |
1876 check_ds = ds_get_max_dep_weak (check_ds); | |
1877 speculate_expr (c_expr, check_ds); | |
1878 | |
1879 return insn; | |
1880 } | |
1881 | |
1882 /* True when INSN is a "regN = regN" copy. */ | |
1883 static bool | |
1884 identical_copy_p (rtx insn) | |
1885 { | |
1886 rtx lhs, rhs, pat; | |
1887 | |
1888 pat = PATTERN (insn); | |
1889 | |
1890 if (GET_CODE (pat) != SET) | |
1891 return false; | |
1892 | |
1893 lhs = SET_DEST (pat); | |
1894 if (!REG_P (lhs)) | |
1895 return false; | |
1896 | |
1897 rhs = SET_SRC (pat); | |
1898 if (!REG_P (rhs)) | |
1899 return false; | |
1900 | |
1901 return REGNO (lhs) == REGNO (rhs); | |
1902 } | |
1903 | |
1904 /* Undo all transformations on *AV_PTR that were done when | |
1905 moving through INSN. */ | |
1906 static void | |
1907 undo_transformations (av_set_t *av_ptr, rtx insn) | |
1908 { | |
1909 av_set_iterator av_iter; | |
1910 expr_t expr; | |
1911 av_set_t new_set = NULL; | |
1912 | |
1913 /* First, kill any EXPR that uses registers set by an insn. This is | |
1914 required for correctness. */ | |
1915 FOR_EACH_EXPR_1 (expr, av_iter, av_ptr) | |
1916 if (!sched_insns_conditions_mutex_p (insn, EXPR_INSN_RTX (expr)) | |
1917 && bitmap_intersect_p (INSN_REG_SETS (insn), | |
1918 VINSN_REG_USES (EXPR_VINSN (expr))) | |
1919 /* When an insn looks like 'r1 = r1', we could substitute through | |
1920 it, but the above condition will still hold. This happened with | |
1921 gcc.c-torture/execute/961125-1.c. */ | |
1922 && !identical_copy_p (insn)) | |
1923 { | |
1924 if (sched_verbose >= 6) | |
1925 sel_print ("Expr %d removed due to use/set conflict\n", | |
1926 INSN_UID (EXPR_INSN_RTX (expr))); | |
1927 av_set_iter_remove (&av_iter); | |
1928 } | |
1929 | |
1930 /* Undo transformations looking at the history vector. */ | |
1931 FOR_EACH_EXPR (expr, av_iter, *av_ptr) | |
1932 { | |
1933 int index = find_in_history_vect (EXPR_HISTORY_OF_CHANGES (expr), | |
1934 insn, EXPR_VINSN (expr), true); | |
1935 | |
1936 if (index >= 0) | |
1937 { | |
1938 expr_history_def *phist; | |
1939 | |
1940 phist = VEC_index (expr_history_def, | |
1941 EXPR_HISTORY_OF_CHANGES (expr), | |
1942 index); | |
1943 | |
1944 switch (phist->type) | |
1945 { | |
1946 case TRANS_SPECULATION: | |
1947 { | |
1948 ds_t old_ds, new_ds; | |
1949 | |
1950 /* Compute the difference between old and new speculative | |
1951 statuses: that's what we need to check. | |
1952 Earlier we used to assert that the status will really | |
1953 change. This no longer works because only the probability | |
1954 bits in the status may have changed during compute_av_set, | |
1955 and in the case of merging different probabilities of the | |
1956 same speculative status along different paths we do not | |
1957 record this in the history vector. */ | |
1958 old_ds = phist->spec_ds; | |
1959 new_ds = EXPR_SPEC_DONE_DS (expr); | |
1960 | |
1961 old_ds &= SPECULATIVE; | |
1962 new_ds &= SPECULATIVE; | |
1963 new_ds &= ~old_ds; | |
1964 | |
1965 EXPR_SPEC_TO_CHECK_DS (expr) |= new_ds; | |
1966 break; | |
1967 } | |
1968 case TRANS_SUBSTITUTION: | |
1969 { | |
1970 expr_def _tmp_expr, *tmp_expr = &_tmp_expr; | |
1971 vinsn_t new_vi; | |
1972 bool add = true; | |
1973 | |
1974 new_vi = phist->old_expr_vinsn; | |
1975 | |
1976 gcc_assert (VINSN_SEPARABLE_P (new_vi) | |
1977 == EXPR_SEPARABLE_P (expr)); | |
1978 copy_expr (tmp_expr, expr); | |
1979 | |
1980 if (vinsn_equal_p (phist->new_expr_vinsn, | |
1981 EXPR_VINSN (tmp_expr))) | |
1982 change_vinsn_in_expr (tmp_expr, new_vi); | |
1983 else | |
1984 /* This happens when we're unsubstituting on a bookkeeping | |
1985 copy, which was in turn substituted. The history is wrong | |
1986 in this case. Do it the hard way. */ | |
1987 add = substitute_reg_in_expr (tmp_expr, insn, true); | |
1988 if (add) | |
1989 av_set_add (&new_set, tmp_expr); | |
1990 clear_expr (tmp_expr); | |
1991 break; | |
1992 } | |
1993 default: | |
1994 gcc_unreachable (); | |
1995 } | |
1996 } | |
1997 | |
1998 } | |
1999 | |
2000 av_set_union_and_clear (av_ptr, &new_set, NULL); | |
2001 } | |
2002 | |
2003 | |
2004 /* Moveup_* helpers for code motion and computing av sets. */ | |
2005 | |
2006 /* Propagates EXPR inside an insn group through THROUGH_INSN. | |
2007 The difference from the below function is that only substitution is | |
2008 performed. */ | |
2009 static enum MOVEUP_EXPR_CODE | |
2010 moveup_expr_inside_insn_group (expr_t expr, insn_t through_insn) | |
2011 { | |
2012 vinsn_t vi = EXPR_VINSN (expr); | |
2013 ds_t *has_dep_p; | |
2014 ds_t full_ds; | |
2015 | |
2016 /* Do this only inside insn group. */ | |
2017 gcc_assert (INSN_SCHED_CYCLE (through_insn) > 0); | |
2018 | |
2019 full_ds = has_dependence_p (expr, through_insn, &has_dep_p); | |
2020 if (full_ds == 0) | |
2021 return MOVEUP_EXPR_SAME; | |
2022 | |
2023 /* Substitution is the possible choice in this case. */ | |
2024 if (has_dep_p[DEPS_IN_RHS]) | |
2025 { | |
2026 /* Can't substitute UNIQUE VINSNs. */ | |
2027 gcc_assert (!VINSN_UNIQUE_P (vi)); | |
2028 | |
2029 if (can_substitute_through_p (through_insn, | |
2030 has_dep_p[DEPS_IN_RHS]) | |
2031 && substitute_reg_in_expr (expr, through_insn, false)) | |
2032 { | |
2033 EXPR_WAS_SUBSTITUTED (expr) = true; | |
2034 return MOVEUP_EXPR_CHANGED; | |
2035 } | |
2036 | |
2037 /* Don't care about this, as even true dependencies may be allowed | |
2038 in an insn group. */ | |
2039 return MOVEUP_EXPR_SAME; | |
2040 } | |
2041 | |
2042 /* This can catch output dependencies in COND_EXECs. */ | |
2043 if (has_dep_p[DEPS_IN_INSN]) | |
2044 return MOVEUP_EXPR_NULL; | |
2045 | |
2046 /* This is either an output or an anti dependence, which usually have | |
2047 a zero latency. Allow this here, if we'd be wrong, tick_check_p | |
2048 will fix this. */ | |
2049 gcc_assert (has_dep_p[DEPS_IN_LHS]); | |
2050 return MOVEUP_EXPR_AS_RHS; | |
2051 } | |
2052 | |
2053 /* True when a trapping EXPR cannot be moved through THROUGH_INSN. */ | |
2054 #define CANT_MOVE_TRAPPING(expr, through_insn) \ | |
2055 (VINSN_MAY_TRAP_P (EXPR_VINSN (expr)) \ | |
2056 && !sel_insn_has_single_succ_p ((through_insn), SUCCS_ALL) \ | |
2057 && !sel_insn_is_speculation_check (through_insn)) | |
2058 | |
2059 /* True when a conflict on a target register was found during moveup_expr. */ | |
2060 static bool was_target_conflict = false; | |
2061 | |
2062 /* Modifies EXPR so it can be moved through the THROUGH_INSN, | |
2063 performing necessary transformations. Record the type of transformation | |
2064 made in PTRANS_TYPE, when it is not NULL. When INSIDE_INSN_GROUP, | |
2065 permit all dependencies except true ones, and try to remove those | |
2066 too via forward substitution. All cases when a non-eliminable | |
2067 non-zero cost dependency exists inside an insn group will be fixed | |
2068 in tick_check_p instead. */ | |
2069 static enum MOVEUP_EXPR_CODE | |
2070 moveup_expr (expr_t expr, insn_t through_insn, bool inside_insn_group, | |
2071 enum local_trans_type *ptrans_type) | |
2072 { | |
2073 vinsn_t vi = EXPR_VINSN (expr); | |
2074 insn_t insn = VINSN_INSN_RTX (vi); | |
2075 bool was_changed = false; | |
2076 bool as_rhs = false; | |
2077 ds_t *has_dep_p; | |
2078 ds_t full_ds; | |
2079 | |
2080 /* When inside_insn_group, delegate to the helper. */ | |
2081 if (inside_insn_group) | |
2082 return moveup_expr_inside_insn_group (expr, through_insn); | |
2083 | |
2084 /* Deal with unique insns and control dependencies. */ | |
2085 if (VINSN_UNIQUE_P (vi)) | |
2086 { | |
2087 /* We can move jumps without side-effects or jumps that are | |
2088 mutually exclusive with instruction THROUGH_INSN (all in cases | |
2089 dependencies allow to do so and jump is not speculative). */ | |
2090 if (control_flow_insn_p (insn)) | |
2091 { | |
2092 basic_block fallthru_bb; | |
2093 | |
2094 /* Do not move checks and do not move jumps through other | |
2095 jumps. */ | |
2096 if (control_flow_insn_p (through_insn) | |
2097 || sel_insn_is_speculation_check (insn)) | |
2098 return MOVEUP_EXPR_NULL; | |
2099 | |
2100 /* Don't move jumps through CFG joins. */ | |
2101 if (bookkeeping_can_be_created_if_moved_through_p (through_insn)) | |
2102 return MOVEUP_EXPR_NULL; | |
2103 | |
2104 /* The jump should have a clear fallthru block, and | |
2105 this block should be in the current region. */ | |
2106 if ((fallthru_bb = fallthru_bb_of_jump (insn)) == NULL | |
2107 || ! in_current_region_p (fallthru_bb)) | |
2108 return MOVEUP_EXPR_NULL; | |
2109 | |
2110 /* And it should be mutually exclusive with through_insn, or | |
2111 be an unconditional jump. */ | |
2112 if (! any_uncondjump_p (insn) | |
2113 && ! sched_insns_conditions_mutex_p (insn, through_insn)) | |
2114 return MOVEUP_EXPR_NULL; | |
2115 } | |
2116 | |
2117 /* Don't move what we can't move. */ | |
2118 if (EXPR_CANT_MOVE (expr) | |
2119 && BLOCK_FOR_INSN (through_insn) != BLOCK_FOR_INSN (insn)) | |
2120 return MOVEUP_EXPR_NULL; | |
2121 | |
2122 /* Don't move SCHED_GROUP instruction through anything. | |
2123 If we don't force this, then it will be possible to start | |
2124 scheduling a sched_group before all its dependencies are | |
2125 resolved. | |
2126 ??? Haifa deals with this issue by delaying the SCHED_GROUP | |
2127 as late as possible through rank_for_schedule. */ | |
2128 if (SCHED_GROUP_P (insn)) | |
2129 return MOVEUP_EXPR_NULL; | |
2130 } | |
2131 else | |
2132 gcc_assert (!control_flow_insn_p (insn)); | |
2133 | |
2134 /* Deal with data dependencies. */ | |
2135 was_target_conflict = false; | |
2136 full_ds = has_dependence_p (expr, through_insn, &has_dep_p); | |
2137 if (full_ds == 0) | |
2138 { | |
2139 if (!CANT_MOVE_TRAPPING (expr, through_insn)) | |
2140 return MOVEUP_EXPR_SAME; | |
2141 } | |
2142 else | |
2143 { | |
2144 /* We can move UNIQUE insn up only as a whole and unchanged, | |
2145 so it shouldn't have any dependencies. */ | |
2146 if (VINSN_UNIQUE_P (vi)) | |
2147 return MOVEUP_EXPR_NULL; | |
2148 } | |
2149 | |
2150 if (full_ds != 0 && can_speculate_dep_p (full_ds)) | |
2151 { | |
2152 int res; | |
2153 | |
2154 res = speculate_expr (expr, full_ds); | |
2155 if (res >= 0) | |
2156 { | |
2157 /* Speculation was successful. */ | |
2158 full_ds = 0; | |
2159 was_changed = (res > 0); | |
2160 if (res == 2) | |
2161 was_target_conflict = true; | |
2162 if (ptrans_type) | |
2163 *ptrans_type = TRANS_SPECULATION; | |
2164 sel_clear_has_dependence (); | |
2165 } | |
2166 } | |
2167 | |
2168 if (has_dep_p[DEPS_IN_INSN]) | |
2169 /* We have some dependency that cannot be discarded. */ | |
2170 return MOVEUP_EXPR_NULL; | |
2171 | |
2172 if (has_dep_p[DEPS_IN_LHS]) | |
2173 { | |
2174 /* Only separable insns can be moved up with the new register. | |
2175 Anyways, we should mark that the original register is | |
2176 unavailable. */ | |
2177 if (!enable_schedule_as_rhs_p || !EXPR_SEPARABLE_P (expr)) | |
2178 return MOVEUP_EXPR_NULL; | |
2179 | |
2180 EXPR_TARGET_AVAILABLE (expr) = false; | |
2181 was_target_conflict = true; | |
2182 as_rhs = true; | |
2183 } | |
2184 | |
2185 /* At this point we have either separable insns, that will be lifted | |
2186 up only as RHSes, or non-separable insns with no dependency in lhs. | |
2187 If dependency is in RHS, then try to perform substitution and move up | |
2188 substituted RHS: | |
2189 | |
2190 Ex. 1: Ex.2 | |
2191 y = x; y = x; | |
2192 z = y*2; y = y*2; | |
2193 | |
2194 In Ex.1 y*2 can be substituted for x*2 and the whole operation can be | |
2195 moved above y=x assignment as z=x*2. | |
2196 | |
2197 In Ex.2 y*2 also can be substituted for x*2, but only the right hand | |
2198 side can be moved because of the output dependency. The operation was | |
2199 cropped to its rhs above. */ | |
2200 if (has_dep_p[DEPS_IN_RHS]) | |
2201 { | |
2202 ds_t *rhs_dsp = &has_dep_p[DEPS_IN_RHS]; | |
2203 | |
2204 /* Can't substitute UNIQUE VINSNs. */ | |
2205 gcc_assert (!VINSN_UNIQUE_P (vi)); | |
2206 | |
2207 if (can_speculate_dep_p (*rhs_dsp)) | |
2208 { | |
2209 int res; | |
2210 | |
2211 res = speculate_expr (expr, *rhs_dsp); | |
2212 if (res >= 0) | |
2213 { | |
2214 /* Speculation was successful. */ | |
2215 *rhs_dsp = 0; | |
2216 was_changed = (res > 0); | |
2217 if (res == 2) | |
2218 was_target_conflict = true; | |
2219 if (ptrans_type) | |
2220 *ptrans_type = TRANS_SPECULATION; | |
2221 } | |
2222 else | |
2223 return MOVEUP_EXPR_NULL; | |
2224 } | |
2225 else if (can_substitute_through_p (through_insn, | |
2226 *rhs_dsp) | |
2227 && substitute_reg_in_expr (expr, through_insn, false)) | |
2228 { | |
2229 /* ??? We cannot perform substitution AND speculation on the same | |
2230 insn. */ | |
2231 gcc_assert (!was_changed); | |
2232 was_changed = true; | |
2233 if (ptrans_type) | |
2234 *ptrans_type = TRANS_SUBSTITUTION; | |
2235 EXPR_WAS_SUBSTITUTED (expr) = true; | |
2236 } | |
2237 else | |
2238 return MOVEUP_EXPR_NULL; | |
2239 } | |
2240 | |
2241 /* Don't move trapping insns through jumps. | |
2242 This check should be at the end to give a chance to control speculation | |
2243 to perform its duties. */ | |
2244 if (CANT_MOVE_TRAPPING (expr, through_insn)) | |
2245 return MOVEUP_EXPR_NULL; | |
2246 | |
2247 return (was_changed | |
2248 ? MOVEUP_EXPR_CHANGED | |
2249 : (as_rhs | |
2250 ? MOVEUP_EXPR_AS_RHS | |
2251 : MOVEUP_EXPR_SAME)); | |
2252 } | |
2253 | |
2254 /* Try to look at bitmap caches for EXPR and INSN pair, return true | |
2255 if successful. When INSIDE_INSN_GROUP, also try ignore dependencies | |
2256 that can exist within a parallel group. Write to RES the resulting | |
2257 code for moveup_expr. */ | |
2258 static bool | |
2259 try_bitmap_cache (expr_t expr, insn_t insn, | |
2260 bool inside_insn_group, | |
2261 enum MOVEUP_EXPR_CODE *res) | |
2262 { | |
2263 int expr_uid = INSN_UID (EXPR_INSN_RTX (expr)); | |
2264 | |
2265 /* First check whether we've analyzed this situation already. */ | |
2266 if (bitmap_bit_p (INSN_ANALYZED_DEPS (insn), expr_uid)) | |
2267 { | |
2268 if (bitmap_bit_p (INSN_FOUND_DEPS (insn), expr_uid)) | |
2269 { | |
2270 if (sched_verbose >= 6) | |
2271 sel_print ("removed (cached)\n"); | |
2272 *res = MOVEUP_EXPR_NULL; | |
2273 return true; | |
2274 } | |
2275 else | |
2276 { | |
2277 if (sched_verbose >= 6) | |
2278 sel_print ("unchanged (cached)\n"); | |
2279 *res = MOVEUP_EXPR_SAME; | |
2280 return true; | |
2281 } | |
2282 } | |
2283 else if (bitmap_bit_p (INSN_FOUND_DEPS (insn), expr_uid)) | |
2284 { | |
2285 if (inside_insn_group) | |
2286 { | |
2287 if (sched_verbose >= 6) | |
2288 sel_print ("unchanged (as RHS, cached, inside insn group)\n"); | |
2289 *res = MOVEUP_EXPR_SAME; | |
2290 return true; | |
2291 | |
2292 } | |
2293 else | |
2294 EXPR_TARGET_AVAILABLE (expr) = false; | |
2295 | |
2296 /* This is the only case when propagation result can change over time, | |
2297 as we can dynamically switch off scheduling as RHS. In this case, | |
2298 just check the flag to reach the correct decision. */ | |
2299 if (enable_schedule_as_rhs_p) | |
2300 { | |
2301 if (sched_verbose >= 6) | |
2302 sel_print ("unchanged (as RHS, cached)\n"); | |
2303 *res = MOVEUP_EXPR_AS_RHS; | |
2304 return true; | |
2305 } | |
2306 else | |
2307 { | |
2308 if (sched_verbose >= 6) | |
2309 sel_print ("removed (cached as RHS, but renaming" | |
2310 " is now disabled)\n"); | |
2311 *res = MOVEUP_EXPR_NULL; | |
2312 return true; | |
2313 } | |
2314 } | |
2315 | |
2316 return false; | |
2317 } | |
2318 | |
2319 /* Try to look at bitmap caches for EXPR and INSN pair, return true | |
2320 if successful. Write to RES the resulting code for moveup_expr. */ | |
2321 static bool | |
2322 try_transformation_cache (expr_t expr, insn_t insn, | |
2323 enum MOVEUP_EXPR_CODE *res) | |
2324 { | |
2325 struct transformed_insns *pti | |
2326 = (struct transformed_insns *) | |
2327 htab_find_with_hash (INSN_TRANSFORMED_INSNS (insn), | |
2328 &EXPR_VINSN (expr), | |
2329 VINSN_HASH_RTX (EXPR_VINSN (expr))); | |
2330 if (pti) | |
2331 { | |
2332 /* This EXPR was already moved through this insn and was | |
2333 changed as a result. Fetch the proper data from | |
2334 the hashtable. */ | |
2335 insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (expr), | |
2336 INSN_UID (insn), pti->type, | |
2337 pti->vinsn_old, pti->vinsn_new, | |
2338 EXPR_SPEC_DONE_DS (expr)); | |
2339 | |
2340 if (INSN_IN_STREAM_P (VINSN_INSN_RTX (pti->vinsn_new))) | |
2341 pti->vinsn_new = vinsn_copy (pti->vinsn_new, true); | |
2342 change_vinsn_in_expr (expr, pti->vinsn_new); | |
2343 if (pti->was_target_conflict) | |
2344 EXPR_TARGET_AVAILABLE (expr) = false; | |
2345 if (pti->type == TRANS_SPECULATION) | |
2346 { | |
2347 ds_t ds; | |
2348 | |
2349 ds = EXPR_SPEC_DONE_DS (expr); | |
2350 | |
2351 EXPR_SPEC_DONE_DS (expr) = pti->ds; | |
2352 EXPR_NEEDS_SPEC_CHECK_P (expr) |= pti->needs_check; | |
2353 } | |
2354 | |
2355 if (sched_verbose >= 6) | |
2356 { | |
2357 sel_print ("changed (cached): "); | |
2358 dump_expr (expr); | |
2359 sel_print ("\n"); | |
2360 } | |
2361 | |
2362 *res = MOVEUP_EXPR_CHANGED; | |
2363 return true; | |
2364 } | |
2365 | |
2366 return false; | |
2367 } | |
2368 | |
2369 /* Update bitmap caches on INSN with result RES of propagating EXPR. */ | |
2370 static void | |
2371 update_bitmap_cache (expr_t expr, insn_t insn, bool inside_insn_group, | |
2372 enum MOVEUP_EXPR_CODE res) | |
2373 { | |
2374 int expr_uid = INSN_UID (EXPR_INSN_RTX (expr)); | |
2375 | |
2376 /* Do not cache result of propagating jumps through an insn group, | |
2377 as it is always true, which is not useful outside the group. */ | |
2378 if (inside_insn_group) | |
2379 return; | |
2380 | |
2381 if (res == MOVEUP_EXPR_NULL) | |
2382 { | |
2383 bitmap_set_bit (INSN_ANALYZED_DEPS (insn), expr_uid); | |
2384 bitmap_set_bit (INSN_FOUND_DEPS (insn), expr_uid); | |
2385 } | |
2386 else if (res == MOVEUP_EXPR_SAME) | |
2387 { | |
2388 bitmap_set_bit (INSN_ANALYZED_DEPS (insn), expr_uid); | |
2389 bitmap_clear_bit (INSN_FOUND_DEPS (insn), expr_uid); | |
2390 } | |
2391 else if (res == MOVEUP_EXPR_AS_RHS) | |
2392 { | |
2393 bitmap_clear_bit (INSN_ANALYZED_DEPS (insn), expr_uid); | |
2394 bitmap_set_bit (INSN_FOUND_DEPS (insn), expr_uid); | |
2395 } | |
2396 else | |
2397 gcc_unreachable (); | |
2398 } | |
2399 | |
2400 /* Update hashtable on INSN with changed EXPR, old EXPR_OLD_VINSN | |
2401 and transformation type TRANS_TYPE. */ | |
2402 static void | |
2403 update_transformation_cache (expr_t expr, insn_t insn, | |
2404 bool inside_insn_group, | |
2405 enum local_trans_type trans_type, | |
2406 vinsn_t expr_old_vinsn) | |
2407 { | |
2408 struct transformed_insns *pti; | |
2409 | |
2410 if (inside_insn_group) | |
2411 return; | |
2412 | |
2413 pti = XNEW (struct transformed_insns); | |
2414 pti->vinsn_old = expr_old_vinsn; | |
2415 pti->vinsn_new = EXPR_VINSN (expr); | |
2416 pti->type = trans_type; | |
2417 pti->was_target_conflict = was_target_conflict; | |
2418 pti->ds = EXPR_SPEC_DONE_DS (expr); | |
2419 pti->needs_check = EXPR_NEEDS_SPEC_CHECK_P (expr); | |
2420 vinsn_attach (pti->vinsn_old); | |
2421 vinsn_attach (pti->vinsn_new); | |
2422 *((struct transformed_insns **) | |
2423 htab_find_slot_with_hash (INSN_TRANSFORMED_INSNS (insn), | |
2424 pti, VINSN_HASH_RTX (expr_old_vinsn), | |
2425 INSERT)) = pti; | |
2426 } | |
2427 | |
2428 /* Same as moveup_expr, but first looks up the result of | |
2429 transformation in caches. */ | |
2430 static enum MOVEUP_EXPR_CODE | |
2431 moveup_expr_cached (expr_t expr, insn_t insn, bool inside_insn_group) | |
2432 { | |
2433 enum MOVEUP_EXPR_CODE res; | |
2434 bool got_answer = false; | |
2435 | |
2436 if (sched_verbose >= 6) | |
2437 { | |
2438 sel_print ("Moving "); | |
2439 dump_expr (expr); | |
2440 sel_print (" through %d: ", INSN_UID (insn)); | |
2441 } | |
2442 | |
2443 if (try_bitmap_cache (expr, insn, inside_insn_group, &res)) | |
2444 /* When inside insn group, we do not want remove stores conflicting | |
2445 with previosly issued loads. */ | |
2446 got_answer = ! inside_insn_group || res != MOVEUP_EXPR_NULL; | |
2447 else if (try_transformation_cache (expr, insn, &res)) | |
2448 got_answer = true; | |
2449 | |
2450 if (! got_answer) | |
2451 { | |
2452 /* Invoke moveup_expr and record the results. */ | |
2453 vinsn_t expr_old_vinsn = EXPR_VINSN (expr); | |
2454 ds_t expr_old_spec_ds = EXPR_SPEC_DONE_DS (expr); | |
2455 int expr_uid = INSN_UID (VINSN_INSN_RTX (expr_old_vinsn)); | |
2456 bool unique_p = VINSN_UNIQUE_P (expr_old_vinsn); | |
2457 enum local_trans_type trans_type = TRANS_SUBSTITUTION; | |
2458 | |
2459 /* ??? Invent something better than this. We can't allow old_vinsn | |
2460 to go, we need it for the history vector. */ | |
2461 vinsn_attach (expr_old_vinsn); | |
2462 | |
2463 res = moveup_expr (expr, insn, inside_insn_group, | |
2464 &trans_type); | |
2465 switch (res) | |
2466 { | |
2467 case MOVEUP_EXPR_NULL: | |
2468 update_bitmap_cache (expr, insn, inside_insn_group, res); | |
2469 if (sched_verbose >= 6) | |
2470 sel_print ("removed\n"); | |
2471 break; | |
2472 | |
2473 case MOVEUP_EXPR_SAME: | |
2474 update_bitmap_cache (expr, insn, inside_insn_group, res); | |
2475 if (sched_verbose >= 6) | |
2476 sel_print ("unchanged\n"); | |
2477 break; | |
2478 | |
2479 case MOVEUP_EXPR_AS_RHS: | |
2480 gcc_assert (!unique_p || inside_insn_group); | |
2481 update_bitmap_cache (expr, insn, inside_insn_group, res); | |
2482 if (sched_verbose >= 6) | |
2483 sel_print ("unchanged (as RHS)\n"); | |
2484 break; | |
2485 | |
2486 case MOVEUP_EXPR_CHANGED: | |
2487 gcc_assert (INSN_UID (EXPR_INSN_RTX (expr)) != expr_uid | |
2488 || EXPR_SPEC_DONE_DS (expr) != expr_old_spec_ds); | |
2489 insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (expr), | |
2490 INSN_UID (insn), trans_type, | |
2491 expr_old_vinsn, EXPR_VINSN (expr), | |
2492 expr_old_spec_ds); | |
2493 update_transformation_cache (expr, insn, inside_insn_group, | |
2494 trans_type, expr_old_vinsn); | |
2495 if (sched_verbose >= 6) | |
2496 { | |
2497 sel_print ("changed: "); | |
2498 dump_expr (expr); | |
2499 sel_print ("\n"); | |
2500 } | |
2501 break; | |
2502 default: | |
2503 gcc_unreachable (); | |
2504 } | |
2505 | |
2506 vinsn_detach (expr_old_vinsn); | |
2507 } | |
2508 | |
2509 return res; | |
2510 } | |
2511 | |
2512 /* Moves an av set AVP up through INSN, performing necessary | |
2513 transformations. */ | |
2514 static void | |
2515 moveup_set_expr (av_set_t *avp, insn_t insn, bool inside_insn_group) | |
2516 { | |
2517 av_set_iterator i; | |
2518 expr_t expr; | |
2519 | |
2520 FOR_EACH_EXPR_1 (expr, i, avp) | |
2521 { | |
2522 | |
2523 switch (moveup_expr_cached (expr, insn, inside_insn_group)) | |
2524 { | |
2525 case MOVEUP_EXPR_SAME: | |
2526 case MOVEUP_EXPR_AS_RHS: | |
2527 break; | |
2528 | |
2529 case MOVEUP_EXPR_NULL: | |
2530 av_set_iter_remove (&i); | |
2531 break; | |
2532 | |
2533 case MOVEUP_EXPR_CHANGED: | |
2534 expr = merge_with_other_exprs (avp, &i, expr); | |
2535 break; | |
2536 | |
2537 default: | |
2538 gcc_unreachable (); | |
2539 } | |
2540 } | |
2541 } | |
2542 | |
2543 /* Moves AVP set along PATH. */ | |
2544 static void | |
2545 moveup_set_inside_insn_group (av_set_t *avp, ilist_t path) | |
2546 { | |
2547 int last_cycle; | |
2548 | |
2549 if (sched_verbose >= 6) | |
2550 sel_print ("Moving expressions up in the insn group...\n"); | |
2551 if (! path) | |
2552 return; | |
2553 last_cycle = INSN_SCHED_CYCLE (ILIST_INSN (path)); | |
2554 while (path | |
2555 && INSN_SCHED_CYCLE (ILIST_INSN (path)) == last_cycle) | |
2556 { | |
2557 moveup_set_expr (avp, ILIST_INSN (path), true); | |
2558 path = ILIST_NEXT (path); | |
2559 } | |
2560 } | |
2561 | |
2562 /* Returns true if after moving EXPR along PATH it equals to EXPR_VLIW. */ | |
2563 static bool | |
2564 equal_after_moveup_path_p (expr_t expr, ilist_t path, expr_t expr_vliw) | |
2565 { | |
2566 expr_def _tmp, *tmp = &_tmp; | |
2567 int last_cycle; | |
2568 bool res = true; | |
2569 | |
2570 copy_expr_onside (tmp, expr); | |
2571 last_cycle = path ? INSN_SCHED_CYCLE (ILIST_INSN (path)) : 0; | |
2572 while (path | |
2573 && res | |
2574 && INSN_SCHED_CYCLE (ILIST_INSN (path)) == last_cycle) | |
2575 { | |
2576 res = (moveup_expr_cached (tmp, ILIST_INSN (path), true) | |
2577 != MOVEUP_EXPR_NULL); | |
2578 path = ILIST_NEXT (path); | |
2579 } | |
2580 | |
2581 if (res) | |
2582 { | |
2583 vinsn_t tmp_vinsn = EXPR_VINSN (tmp); | |
2584 vinsn_t expr_vliw_vinsn = EXPR_VINSN (expr_vliw); | |
2585 | |
2586 if (tmp_vinsn != expr_vliw_vinsn) | |
2587 res = vinsn_equal_p (tmp_vinsn, expr_vliw_vinsn); | |
2588 } | |
2589 | |
2590 clear_expr (tmp); | |
2591 return res; | |
2592 } | |
2593 | |
2594 | |
2595 /* Functions that compute av and lv sets. */ | |
2596 | |
2597 /* Returns true if INSN is not a downward continuation of the given path P in | |
2598 the current stage. */ | |
2599 static bool | |
2600 is_ineligible_successor (insn_t insn, ilist_t p) | |
2601 { | |
2602 insn_t prev_insn; | |
2603 | |
2604 /* Check if insn is not deleted. */ | |
2605 if (PREV_INSN (insn) && NEXT_INSN (PREV_INSN (insn)) != insn) | |
2606 gcc_unreachable (); | |
2607 else if (NEXT_INSN (insn) && PREV_INSN (NEXT_INSN (insn)) != insn) | |
2608 gcc_unreachable (); | |
2609 | |
2610 /* If it's the first insn visited, then the successor is ok. */ | |
2611 if (!p) | |
2612 return false; | |
2613 | |
2614 prev_insn = ILIST_INSN (p); | |
2615 | |
2616 if (/* a backward edge. */ | |
2617 INSN_SEQNO (insn) < INSN_SEQNO (prev_insn) | |
2618 /* is already visited. */ | |
2619 || (INSN_SEQNO (insn) == INSN_SEQNO (prev_insn) | |
2620 && (ilist_is_in_p (p, insn) | |
2621 /* We can reach another fence here and still seqno of insn | |
2622 would be equal to seqno of prev_insn. This is possible | |
2623 when prev_insn is a previously created bookkeeping copy. | |
2624 In that case it'd get a seqno of insn. Thus, check here | |
2625 whether insn is in current fence too. */ | |
2626 || IN_CURRENT_FENCE_P (insn))) | |
2627 /* Was already scheduled on this round. */ | |
2628 || (INSN_SEQNO (insn) > INSN_SEQNO (prev_insn) | |
2629 && IN_CURRENT_FENCE_P (insn)) | |
2630 /* An insn from another fence could also be | |
2631 scheduled earlier even if this insn is not in | |
2632 a fence list right now. Check INSN_SCHED_CYCLE instead. */ | |
2633 || (!pipelining_p | |
2634 && INSN_SCHED_TIMES (insn) > 0)) | |
2635 return true; | |
2636 else | |
2637 return false; | |
2638 } | |
2639 | |
2640 /* Computes the av_set below the last bb insn INSN, doing all the 'dirty work' | |
2641 of handling multiple successors and properly merging its av_sets. P is | |
2642 the current path traversed. WS is the size of lookahead window. | |
2643 Return the av set computed. */ | |
2644 static av_set_t | |
2645 compute_av_set_at_bb_end (insn_t insn, ilist_t p, int ws) | |
2646 { | |
2647 struct succs_info *sinfo; | |
2648 av_set_t expr_in_all_succ_branches = NULL; | |
2649 int is; | |
2650 insn_t succ, zero_succ = NULL; | |
2651 av_set_t av1 = NULL; | |
2652 | |
2653 gcc_assert (sel_bb_end_p (insn)); | |
2654 | |
2655 /* Find different kind of successors needed for correct computing of | |
2656 SPEC and TARGET_AVAILABLE attributes. */ | |
2657 sinfo = compute_succs_info (insn, SUCCS_NORMAL); | |
2658 | |
2659 /* Debug output. */ | |
2660 if (sched_verbose >= 6) | |
2661 { | |
2662 sel_print ("successors of bb end (%d): ", INSN_UID (insn)); | |
2663 dump_insn_vector (sinfo->succs_ok); | |
2664 sel_print ("\n"); | |
2665 if (sinfo->succs_ok_n != sinfo->all_succs_n) | |
2666 sel_print ("real successors num: %d\n", sinfo->all_succs_n); | |
2667 } | |
2668 | |
2669 /* Add insn to to the tail of current path. */ | |
2670 ilist_add (&p, insn); | |
2671 | |
2672 for (is = 0; VEC_iterate (rtx, sinfo->succs_ok, is, succ); is++) | |
2673 { | |
2674 av_set_t succ_set; | |
2675 | |
2676 /* We will edit SUCC_SET and EXPR_SPEC field of its elements. */ | |
2677 succ_set = compute_av_set_inside_bb (succ, p, ws, true); | |
2678 | |
2679 av_set_split_usefulness (succ_set, | |
2680 VEC_index (int, sinfo->probs_ok, is), | |
2681 sinfo->all_prob); | |
2682 | |
2683 if (sinfo->all_succs_n > 1 | |
2684 && sinfo->all_succs_n == sinfo->succs_ok_n) | |
2685 { | |
2686 /* Find EXPR'es that came from *all* successors and save them | |
2687 into expr_in_all_succ_branches. This set will be used later | |
2688 for calculating speculation attributes of EXPR'es. */ | |
2689 if (is == 0) | |
2690 { | |
2691 expr_in_all_succ_branches = av_set_copy (succ_set); | |
2692 | |
2693 /* Remember the first successor for later. */ | |
2694 zero_succ = succ; | |
2695 } | |
2696 else | |
2697 { | |
2698 av_set_iterator i; | |
2699 expr_t expr; | |
2700 | |
2701 FOR_EACH_EXPR_1 (expr, i, &expr_in_all_succ_branches) | |
2702 if (!av_set_is_in_p (succ_set, EXPR_VINSN (expr))) | |
2703 av_set_iter_remove (&i); | |
2704 } | |
2705 } | |
2706 | |
2707 /* Union the av_sets. Check liveness restrictions on target registers | |
2708 in special case of two successors. */ | |
2709 if (sinfo->succs_ok_n == 2 && is == 1) | |
2710 { | |
2711 basic_block bb0 = BLOCK_FOR_INSN (zero_succ); | |
2712 basic_block bb1 = BLOCK_FOR_INSN (succ); | |
2713 | |
2714 gcc_assert (BB_LV_SET_VALID_P (bb0) && BB_LV_SET_VALID_P (bb1)); | |
2715 av_set_union_and_live (&av1, &succ_set, | |
2716 BB_LV_SET (bb0), | |
2717 BB_LV_SET (bb1), | |
2718 insn); | |
2719 } | |
2720 else | |
2721 av_set_union_and_clear (&av1, &succ_set, insn); | |
2722 } | |
2723 | |
2724 /* Check liveness restrictions via hard way when there are more than | |
2725 two successors. */ | |
2726 if (sinfo->succs_ok_n > 2) | |
2727 for (is = 0; VEC_iterate (rtx, sinfo->succs_ok, is, succ); is++) | |
2728 { | |
2729 basic_block succ_bb = BLOCK_FOR_INSN (succ); | |
2730 | |
2731 gcc_assert (BB_LV_SET_VALID_P (succ_bb)); | |
2732 mark_unavailable_targets (av1, BB_AV_SET (succ_bb), | |
2733 BB_LV_SET (succ_bb)); | |
2734 } | |
2735 | |
2736 /* Finally, check liveness restrictions on paths leaving the region. */ | |
2737 if (sinfo->all_succs_n > sinfo->succs_ok_n) | |
2738 for (is = 0; VEC_iterate (rtx, sinfo->succs_other, is, succ); is++) | |
2739 mark_unavailable_targets | |
2740 (av1, NULL, BB_LV_SET (BLOCK_FOR_INSN (succ))); | |
2741 | |
2742 if (sinfo->all_succs_n > 1) | |
2743 { | |
2744 av_set_iterator i; | |
2745 expr_t expr; | |
2746 | |
2747 /* Increase the spec attribute of all EXPR'es that didn't come | |
2748 from all successors. */ | |
2749 FOR_EACH_EXPR (expr, i, av1) | |
2750 if (!av_set_is_in_p (expr_in_all_succ_branches, EXPR_VINSN (expr))) | |
2751 EXPR_SPEC (expr)++; | |
2752 | |
2753 av_set_clear (&expr_in_all_succ_branches); | |
2754 | |
2755 /* Do not move conditional branches through other | |
2756 conditional branches. So, remove all conditional | |
2757 branches from av_set if current operator is a conditional | |
2758 branch. */ | |
2759 av_set_substract_cond_branches (&av1); | |
2760 } | |
2761 | |
2762 ilist_remove (&p); | |
2763 free_succs_info (sinfo); | |
2764 | |
2765 if (sched_verbose >= 6) | |
2766 { | |
2767 sel_print ("av_succs (%d): ", INSN_UID (insn)); | |
2768 dump_av_set (av1); | |
2769 sel_print ("\n"); | |
2770 } | |
2771 | |
2772 return av1; | |
2773 } | |
2774 | |
2775 /* This function computes av_set for the FIRST_INSN by dragging valid | |
2776 av_set through all basic block insns either from the end of basic block | |
2777 (computed using compute_av_set_at_bb_end) or from the insn on which | |
2778 MAX_WS was exceeded. It uses compute_av_set_at_bb_end to compute av_set | |
2779 below the basic block and handling conditional branches. | |
2780 FIRST_INSN - the basic block head, P - path consisting of the insns | |
2781 traversed on the way to the FIRST_INSN (the path is sparse, only bb heads | |
2782 and bb ends are added to the path), WS - current window size, | |
2783 NEED_COPY_P - true if we'll make a copy of av_set before returning it. */ | |
2784 static av_set_t | |
2785 compute_av_set_inside_bb (insn_t first_insn, ilist_t p, int ws, | |
2786 bool need_copy_p) | |
2787 { | |
2788 insn_t cur_insn; | |
2789 int end_ws = ws; | |
2790 insn_t bb_end = sel_bb_end (BLOCK_FOR_INSN (first_insn)); | |
2791 insn_t after_bb_end = NEXT_INSN (bb_end); | |
2792 insn_t last_insn; | |
2793 av_set_t av = NULL; | |
2794 basic_block cur_bb = BLOCK_FOR_INSN (first_insn); | |
2795 | |
2796 /* Return NULL if insn is not on the legitimate downward path. */ | |
2797 if (is_ineligible_successor (first_insn, p)) | |
2798 { | |
2799 if (sched_verbose >= 6) | |
2800 sel_print ("Insn %d is ineligible_successor\n", INSN_UID (first_insn)); | |
2801 | |
2802 return NULL; | |
2803 } | |
2804 | |
2805 /* If insn already has valid av(insn) computed, just return it. */ | |
2806 if (AV_SET_VALID_P (first_insn)) | |
2807 { | |
2808 av_set_t av_set; | |
2809 | |
2810 if (sel_bb_head_p (first_insn)) | |
2811 av_set = BB_AV_SET (BLOCK_FOR_INSN (first_insn)); | |
2812 else | |
2813 av_set = NULL; | |
2814 | |
2815 if (sched_verbose >= 6) | |
2816 { | |
2817 sel_print ("Insn %d has a valid av set: ", INSN_UID (first_insn)); | |
2818 dump_av_set (av_set); | |
2819 sel_print ("\n"); | |
2820 } | |
2821 | |
2822 return need_copy_p ? av_set_copy (av_set) : av_set; | |
2823 } | |
2824 | |
2825 ilist_add (&p, first_insn); | |
2826 | |
2827 /* As the result after this loop have completed, in LAST_INSN we'll | |
2828 have the insn which has valid av_set to start backward computation | |
2829 from: it either will be NULL because on it the window size was exceeded | |
2830 or other valid av_set as returned by compute_av_set for the last insn | |
2831 of the basic block. */ | |
2832 for (last_insn = first_insn; last_insn != after_bb_end; | |
2833 last_insn = NEXT_INSN (last_insn)) | |
2834 { | |
2835 /* We may encounter valid av_set not only on bb_head, but also on | |
2836 those insns on which previously MAX_WS was exceeded. */ | |
2837 if (AV_SET_VALID_P (last_insn)) | |
2838 { | |
2839 if (sched_verbose >= 6) | |
2840 sel_print ("Insn %d has a valid empty av set\n", INSN_UID (last_insn)); | |
2841 break; | |
2842 } | |
2843 | |
2844 /* The special case: the last insn of the BB may be an | |
2845 ineligible_successor due to its SEQ_NO that was set on | |
2846 it as a bookkeeping. */ | |
2847 if (last_insn != first_insn | |
2848 && is_ineligible_successor (last_insn, p)) | |
2849 { | |
2850 if (sched_verbose >= 6) | |
2851 sel_print ("Insn %d is ineligible_successor\n", INSN_UID (last_insn)); | |
2852 break; | |
2853 } | |
2854 | |
2855 if (end_ws > max_ws) | |
2856 { | |
2857 /* We can reach max lookahead size at bb_header, so clean av_set | |
2858 first. */ | |
2859 INSN_WS_LEVEL (last_insn) = global_level; | |
2860 | |
2861 if (sched_verbose >= 6) | |
2862 sel_print ("Insn %d is beyond the software lookahead window size\n", | |
2863 INSN_UID (last_insn)); | |
2864 break; | |
2865 } | |
2866 | |
2867 end_ws++; | |
2868 } | |
2869 | |
2870 /* Get the valid av_set into AV above the LAST_INSN to start backward | |
2871 computation from. It either will be empty av_set or av_set computed from | |
2872 the successors on the last insn of the current bb. */ | |
2873 if (last_insn != after_bb_end) | |
2874 { | |
2875 av = NULL; | |
2876 | |
2877 /* This is needed only to obtain av_sets that are identical to | |
2878 those computed by the old compute_av_set version. */ | |
2879 if (last_insn == first_insn && !INSN_NOP_P (last_insn)) | |
2880 av_set_add (&av, INSN_EXPR (last_insn)); | |
2881 } | |
2882 else | |
2883 /* END_WS is always already increased by 1 if LAST_INSN == AFTER_BB_END. */ | |
2884 av = compute_av_set_at_bb_end (bb_end, p, end_ws); | |
2885 | |
2886 /* Compute av_set in AV starting from below the LAST_INSN up to | |
2887 location above the FIRST_INSN. */ | |
2888 for (cur_insn = PREV_INSN (last_insn); cur_insn != PREV_INSN (first_insn); | |
2889 cur_insn = PREV_INSN (cur_insn)) | |
2890 if (!INSN_NOP_P (cur_insn)) | |
2891 { | |
2892 expr_t expr; | |
2893 | |
2894 moveup_set_expr (&av, cur_insn, false); | |
2895 | |
2896 /* If the expression for CUR_INSN is already in the set, | |
2897 replace it by the new one. */ | |
2898 expr = av_set_lookup (av, INSN_VINSN (cur_insn)); | |
2899 if (expr != NULL) | |
2900 { | |
2901 clear_expr (expr); | |
2902 copy_expr (expr, INSN_EXPR (cur_insn)); | |
2903 } | |
2904 else | |
2905 av_set_add (&av, INSN_EXPR (cur_insn)); | |
2906 } | |
2907 | |
2908 /* Clear stale bb_av_set. */ | |
2909 if (sel_bb_head_p (first_insn)) | |
2910 { | |
2911 av_set_clear (&BB_AV_SET (cur_bb)); | |
2912 BB_AV_SET (cur_bb) = need_copy_p ? av_set_copy (av) : av; | |
2913 BB_AV_LEVEL (cur_bb) = global_level; | |
2914 } | |
2915 | |
2916 if (sched_verbose >= 6) | |
2917 { | |
2918 sel_print ("Computed av set for insn %d: ", INSN_UID (first_insn)); | |
2919 dump_av_set (av); | |
2920 sel_print ("\n"); | |
2921 } | |
2922 | |
2923 ilist_remove (&p); | |
2924 return av; | |
2925 } | |
2926 | |
2927 /* Compute av set before INSN. | |
2928 INSN - the current operation (actual rtx INSN) | |
2929 P - the current path, which is list of insns visited so far | |
2930 WS - software lookahead window size. | |
2931 UNIQUE_P - TRUE, if returned av_set will be changed, hence | |
2932 if we want to save computed av_set in s_i_d, we should make a copy of it. | |
2933 | |
2934 In the resulting set we will have only expressions that don't have delay | |
2935 stalls and nonsubstitutable dependences. */ | |
2936 static av_set_t | |
2937 compute_av_set (insn_t insn, ilist_t p, int ws, bool unique_p) | |
2938 { | |
2939 return compute_av_set_inside_bb (insn, p, ws, unique_p); | |
2940 } | |
2941 | |
2942 /* Propagate a liveness set LV through INSN. */ | |
2943 static void | |
2944 propagate_lv_set (regset lv, insn_t insn) | |
2945 { | |
2946 gcc_assert (INSN_P (insn)); | |
2947 | |
2948 if (INSN_NOP_P (insn)) | |
2949 return; | |
2950 | |
2951 df_simulate_one_insn_backwards (BLOCK_FOR_INSN (insn), insn, lv); | |
2952 } | |
2953 | |
2954 /* Return livness set at the end of BB. */ | |
2955 static regset | |
2956 compute_live_after_bb (basic_block bb) | |
2957 { | |
2958 edge e; | |
2959 edge_iterator ei; | |
2960 regset lv = get_clear_regset_from_pool (); | |
2961 | |
2962 gcc_assert (!ignore_first); | |
2963 | |
2964 FOR_EACH_EDGE (e, ei, bb->succs) | |
2965 if (sel_bb_empty_p (e->dest)) | |
2966 { | |
2967 if (! BB_LV_SET_VALID_P (e->dest)) | |
2968 { | |
2969 gcc_unreachable (); | |
2970 gcc_assert (BB_LV_SET (e->dest) == NULL); | |
2971 BB_LV_SET (e->dest) = compute_live_after_bb (e->dest); | |
2972 BB_LV_SET_VALID_P (e->dest) = true; | |
2973 } | |
2974 IOR_REG_SET (lv, BB_LV_SET (e->dest)); | |
2975 } | |
2976 else | |
2977 IOR_REG_SET (lv, compute_live (sel_bb_head (e->dest))); | |
2978 | |
2979 return lv; | |
2980 } | |
2981 | |
2982 /* Compute the set of all live registers at the point before INSN and save | |
2983 it at INSN if INSN is bb header. */ | |
2984 regset | |
2985 compute_live (insn_t insn) | |
2986 { | |
2987 basic_block bb = BLOCK_FOR_INSN (insn); | |
2988 insn_t final, temp; | |
2989 regset lv; | |
2990 | |
2991 /* Return the valid set if we're already on it. */ | |
2992 if (!ignore_first) | |
2993 { | |
2994 regset src = NULL; | |
2995 | |
2996 if (sel_bb_head_p (insn) && BB_LV_SET_VALID_P (bb)) | |
2997 src = BB_LV_SET (bb); | |
2998 else | |
2999 { | |
3000 gcc_assert (in_current_region_p (bb)); | |
3001 if (INSN_LIVE_VALID_P (insn)) | |
3002 src = INSN_LIVE (insn); | |
3003 } | |
3004 | |
3005 if (src) | |
3006 { | |
3007 lv = get_regset_from_pool (); | |
3008 COPY_REG_SET (lv, src); | |
3009 | |
3010 if (sel_bb_head_p (insn) && ! BB_LV_SET_VALID_P (bb)) | |
3011 { | |
3012 COPY_REG_SET (BB_LV_SET (bb), lv); | |
3013 BB_LV_SET_VALID_P (bb) = true; | |
3014 } | |
3015 | |
3016 return_regset_to_pool (lv); | |
3017 return lv; | |
3018 } | |
3019 } | |
3020 | |
3021 /* We've skipped the wrong lv_set. Don't skip the right one. */ | |
3022 ignore_first = false; | |
3023 gcc_assert (in_current_region_p (bb)); | |
3024 | |
3025 /* Find a valid LV set in this block or below, if needed. | |
3026 Start searching from the next insn: either ignore_first is true, or | |
3027 INSN doesn't have a correct live set. */ | |
3028 temp = NEXT_INSN (insn); | |
3029 final = NEXT_INSN (BB_END (bb)); | |
3030 while (temp != final && ! INSN_LIVE_VALID_P (temp)) | |
3031 temp = NEXT_INSN (temp); | |
3032 if (temp == final) | |
3033 { | |
3034 lv = compute_live_after_bb (bb); | |
3035 temp = PREV_INSN (temp); | |
3036 } | |
3037 else | |
3038 { | |
3039 lv = get_regset_from_pool (); | |
3040 COPY_REG_SET (lv, INSN_LIVE (temp)); | |
3041 } | |
3042 | |
3043 /* Put correct lv sets on the insns which have bad sets. */ | |
3044 final = PREV_INSN (insn); | |
3045 while (temp != final) | |
3046 { | |
3047 propagate_lv_set (lv, temp); | |
3048 COPY_REG_SET (INSN_LIVE (temp), lv); | |
3049 INSN_LIVE_VALID_P (temp) = true; | |
3050 temp = PREV_INSN (temp); | |
3051 } | |
3052 | |
3053 /* Also put it in a BB. */ | |
3054 if (sel_bb_head_p (insn)) | |
3055 { | |
3056 basic_block bb = BLOCK_FOR_INSN (insn); | |
3057 | |
3058 COPY_REG_SET (BB_LV_SET (bb), lv); | |
3059 BB_LV_SET_VALID_P (bb) = true; | |
3060 } | |
3061 | |
3062 /* We return LV to the pool, but will not clear it there. Thus we can | |
3063 legimatelly use LV till the next use of regset_pool_get (). */ | |
3064 return_regset_to_pool (lv); | |
3065 return lv; | |
3066 } | |
3067 | |
3068 /* Update liveness sets for INSN. */ | |
3069 static inline void | |
3070 update_liveness_on_insn (rtx insn) | |
3071 { | |
3072 ignore_first = true; | |
3073 compute_live (insn); | |
3074 } | |
3075 | |
3076 /* Compute liveness below INSN and write it into REGS. */ | |
3077 static inline void | |
3078 compute_live_below_insn (rtx insn, regset regs) | |
3079 { | |
3080 rtx succ; | |
3081 succ_iterator si; | |
3082 | |
3083 FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL) | |
3084 IOR_REG_SET (regs, compute_live (succ)); | |
3085 } | |
3086 | |
3087 /* Update the data gathered in av and lv sets starting from INSN. */ | |
3088 static void | |
3089 update_data_sets (rtx insn) | |
3090 { | |
3091 update_liveness_on_insn (insn); | |
3092 if (sel_bb_head_p (insn)) | |
3093 { | |
3094 gcc_assert (AV_LEVEL (insn) != 0); | |
3095 BB_AV_LEVEL (BLOCK_FOR_INSN (insn)) = -1; | |
3096 compute_av_set (insn, NULL, 0, 0); | |
3097 } | |
3098 } | |
3099 | |
3100 | |
3101 /* Helper for move_op () and find_used_regs (). | |
3102 Return speculation type for which a check should be created on the place | |
3103 of INSN. EXPR is one of the original ops we are searching for. */ | |
3104 static ds_t | |
3105 get_spec_check_type_for_insn (insn_t insn, expr_t expr) | |
3106 { | |
3107 ds_t to_check_ds; | |
3108 ds_t already_checked_ds = EXPR_SPEC_DONE_DS (INSN_EXPR (insn)); | |
3109 | |
3110 to_check_ds = EXPR_SPEC_TO_CHECK_DS (expr); | |
3111 | |
3112 if (targetm.sched.get_insn_checked_ds) | |
3113 already_checked_ds |= targetm.sched.get_insn_checked_ds (insn); | |
3114 | |
3115 if (spec_info != NULL | |
3116 && (spec_info->flags & SEL_SCHED_SPEC_DONT_CHECK_CONTROL)) | |
3117 already_checked_ds |= BEGIN_CONTROL; | |
3118 | |
3119 already_checked_ds = ds_get_speculation_types (already_checked_ds); | |
3120 | |
3121 to_check_ds &= ~already_checked_ds; | |
3122 | |
3123 return to_check_ds; | |
3124 } | |
3125 | |
3126 /* Find the set of registers that are unavailable for storing expres | |
3127 while moving ORIG_OPS up on the path starting from INSN due to | |
3128 liveness (USED_REGS) or hardware restrictions (REG_RENAME_P). | |
3129 | |
3130 All the original operations found during the traversal are saved in the | |
3131 ORIGINAL_INSNS list. | |
3132 | |
3133 REG_RENAME_P denotes the set of hardware registers that | |
3134 can not be used with renaming due to the register class restrictions, | |
3135 mode restrictions and other (the register we'll choose should be | |
3136 compatible class with the original uses, shouldn't be in call_used_regs, | |
3137 should be HARD_REGNO_RENAME_OK etc). | |
3138 | |
3139 Returns TRUE if we've found all original insns, FALSE otherwise. | |
3140 | |
3141 This function utilizes code_motion_path_driver (formerly find_used_regs_1) | |
3142 to traverse the code motion paths. This helper function finds registers | |
3143 that are not available for storing expres while moving ORIG_OPS up on the | |
3144 path starting from INSN. A register considered as used on the moving path, | |
3145 if one of the following conditions is not satisfied: | |
3146 | |
3147 (1) a register not set or read on any path from xi to an instance of | |
3148 the original operation, | |
3149 (2) not among the live registers of the point immediately following the | |
3150 first original operation on a given downward path, except for the | |
3151 original target register of the operation, | |
3152 (3) not live on the other path of any conditional branch that is passed | |
3153 by the operation, in case original operations are not present on | |
3154 both paths of the conditional branch. | |
3155 | |
3156 All the original operations found during the traversal are saved in the | |
3157 ORIGINAL_INSNS list. | |
3158 | |
3159 REG_RENAME_P->CROSSES_CALL is true, if there is a call insn on the path | |
3160 from INSN to original insn. In this case CALL_USED_REG_SET will be added | |
3161 to unavailable hard regs at the point original operation is found. */ | |
3162 | |
3163 static bool | |
3164 find_used_regs (insn_t insn, av_set_t orig_ops, regset used_regs, | |
3165 struct reg_rename *reg_rename_p, def_list_t *original_insns) | |
3166 { | |
3167 def_list_iterator i; | |
3168 def_t def; | |
3169 int res; | |
3170 bool needs_spec_check_p = false; | |
3171 expr_t expr; | |
3172 av_set_iterator expr_iter; | |
3173 struct fur_static_params sparams; | |
3174 struct cmpd_local_params lparams; | |
3175 | |
3176 /* We haven't visited any blocks yet. */ | |
3177 bitmap_clear (code_motion_visited_blocks); | |
3178 | |
3179 /* Init parameters for code_motion_path_driver. */ | |
3180 sparams.crosses_call = false; | |
3181 sparams.original_insns = original_insns; | |
3182 sparams.used_regs = used_regs; | |
3183 | |
3184 /* Set the appropriate hooks and data. */ | |
3185 code_motion_path_driver_info = &fur_hooks; | |
3186 | |
3187 res = code_motion_path_driver (insn, orig_ops, NULL, &lparams, &sparams); | |
3188 | |
3189 reg_rename_p->crosses_call |= sparams.crosses_call; | |
3190 | |
3191 gcc_assert (res == 1); | |
3192 gcc_assert (original_insns && *original_insns); | |
3193 | |
3194 /* ??? We calculate whether an expression needs a check when computing | |
3195 av sets. This information is not as precise as it could be due to | |
3196 merging this bit in merge_expr. We can do better in find_used_regs, | |
3197 but we want to avoid multiple traversals of the same code motion | |
3198 paths. */ | |
3199 FOR_EACH_EXPR (expr, expr_iter, orig_ops) | |
3200 needs_spec_check_p |= EXPR_NEEDS_SPEC_CHECK_P (expr); | |
3201 | |
3202 /* Mark hardware regs in REG_RENAME_P that are not suitable | |
3203 for renaming expr in INSN due to hardware restrictions (register class, | |
3204 modes compatibility etc). */ | |
3205 FOR_EACH_DEF (def, i, *original_insns) | |
3206 { | |
3207 vinsn_t vinsn = INSN_VINSN (def->orig_insn); | |
3208 | |
3209 if (VINSN_SEPARABLE_P (vinsn)) | |
3210 mark_unavailable_hard_regs (def, reg_rename_p, used_regs); | |
3211 | |
3212 /* Do not allow clobbering of ld.[sa] address in case some of the | |
3213 original operations need a check. */ | |
3214 if (needs_spec_check_p) | |
3215 IOR_REG_SET (used_regs, VINSN_REG_USES (vinsn)); | |
3216 } | |
3217 | |
3218 return true; | |
3219 } | |
3220 | |
3221 | |
3222 /* Functions to choose the best insn from available ones. */ | |
3223 | |
3224 /* Adjusts the priority for EXPR using the backend *_adjust_priority hook. */ | |
3225 static int | |
3226 sel_target_adjust_priority (expr_t expr) | |
3227 { | |
3228 int priority = EXPR_PRIORITY (expr); | |
3229 int new_priority; | |
3230 | |
3231 if (targetm.sched.adjust_priority) | |
3232 new_priority = targetm.sched.adjust_priority (EXPR_INSN_RTX (expr), priority); | |
3233 else | |
3234 new_priority = priority; | |
3235 | |
3236 /* If the priority has changed, adjust EXPR_PRIORITY_ADJ accordingly. */ | |
3237 EXPR_PRIORITY_ADJ (expr) = new_priority - EXPR_PRIORITY (expr); | |
3238 | |
3239 gcc_assert (EXPR_PRIORITY_ADJ (expr) >= 0); | |
3240 | |
3241 if (sched_verbose >= 2) | |
3242 sel_print ("sel_target_adjust_priority: insn %d, %d +%d = %d.\n", | |
3243 INSN_UID (EXPR_INSN_RTX (expr)), EXPR_PRIORITY (expr), | |
3244 EXPR_PRIORITY_ADJ (expr), new_priority); | |
3245 | |
3246 return new_priority; | |
3247 } | |
3248 | |
3249 /* Rank two available exprs for schedule. Never return 0 here. */ | |
3250 static int | |
3251 sel_rank_for_schedule (const void *x, const void *y) | |
3252 { | |
3253 expr_t tmp = *(const expr_t *) y; | |
3254 expr_t tmp2 = *(const expr_t *) x; | |
3255 insn_t tmp_insn, tmp2_insn; | |
3256 vinsn_t tmp_vinsn, tmp2_vinsn; | |
3257 int val; | |
3258 | |
3259 tmp_vinsn = EXPR_VINSN (tmp); | |
3260 tmp2_vinsn = EXPR_VINSN (tmp2); | |
3261 tmp_insn = EXPR_INSN_RTX (tmp); | |
3262 tmp2_insn = EXPR_INSN_RTX (tmp2); | |
3263 | |
3264 /* Prefer SCHED_GROUP_P insns to any others. */ | |
3265 if (SCHED_GROUP_P (tmp_insn) != SCHED_GROUP_P (tmp2_insn)) | |
3266 { | |
3267 if (VINSN_UNIQUE_P (tmp_vinsn) && VINSN_UNIQUE_P (tmp2_vinsn)) | |
3268 return SCHED_GROUP_P (tmp2_insn) ? 1 : -1; | |
3269 | |
3270 /* Now uniqueness means SCHED_GROUP_P is set, because schedule groups | |
3271 cannot be cloned. */ | |
3272 if (VINSN_UNIQUE_P (tmp2_vinsn)) | |
3273 return 1; | |
3274 return -1; | |
3275 } | |
3276 | |
3277 /* Discourage scheduling of speculative checks. */ | |
3278 val = (sel_insn_is_speculation_check (tmp_insn) | |
3279 - sel_insn_is_speculation_check (tmp2_insn)); | |
3280 if (val) | |
3281 return val; | |
3282 | |
3283 /* Prefer not scheduled insn over scheduled one. */ | |
3284 if (EXPR_SCHED_TIMES (tmp) > 0 || EXPR_SCHED_TIMES (tmp2) > 0) | |
3285 { | |
3286 val = EXPR_SCHED_TIMES (tmp) - EXPR_SCHED_TIMES (tmp2); | |
3287 if (val) | |
3288 return val; | |
3289 } | |
3290 | |
3291 /* Prefer jump over non-jump instruction. */ | |
3292 if (control_flow_insn_p (tmp_insn) && !control_flow_insn_p (tmp2_insn)) | |
3293 return -1; | |
3294 else if (control_flow_insn_p (tmp2_insn) && !control_flow_insn_p (tmp_insn)) | |
3295 return 1; | |
3296 | |
3297 /* Prefer an expr with greater priority. */ | |
3298 if (EXPR_USEFULNESS (tmp) != 0 && EXPR_USEFULNESS (tmp2) != 0) | |
3299 { | |
3300 int p2 = EXPR_PRIORITY (tmp2) + EXPR_PRIORITY_ADJ (tmp2), | |
3301 p1 = EXPR_PRIORITY (tmp) + EXPR_PRIORITY_ADJ (tmp); | |
3302 | |
3303 val = p2 * EXPR_USEFULNESS (tmp2) - p1 * EXPR_USEFULNESS (tmp); | |
3304 } | |
3305 else | |
3306 val = EXPR_PRIORITY (tmp2) - EXPR_PRIORITY (tmp) | |
3307 + EXPR_PRIORITY_ADJ (tmp2) - EXPR_PRIORITY_ADJ (tmp); | |
3308 if (val) | |
3309 return val; | |
3310 | |
3311 if (spec_info != NULL && spec_info->mask != 0) | |
3312 /* This code was taken from haifa-sched.c: rank_for_schedule (). */ | |
3313 { | |
3314 ds_t ds1, ds2; | |
3315 dw_t dw1, dw2; | |
3316 int dw; | |
3317 | |
3318 ds1 = EXPR_SPEC_DONE_DS (tmp); | |
3319 if (ds1) | |
3320 dw1 = ds_weak (ds1); | |
3321 else | |
3322 dw1 = NO_DEP_WEAK; | |
3323 | |
3324 ds2 = EXPR_SPEC_DONE_DS (tmp2); | |
3325 if (ds2) | |
3326 dw2 = ds_weak (ds2); | |
3327 else | |
3328 dw2 = NO_DEP_WEAK; | |
3329 | |
3330 dw = dw2 - dw1; | |
3331 if (dw > (NO_DEP_WEAK / 8) || dw < -(NO_DEP_WEAK / 8)) | |
3332 return dw; | |
3333 } | |
3334 | |
3335 tmp_insn = EXPR_INSN_RTX (tmp); | |
3336 tmp2_insn = EXPR_INSN_RTX (tmp2); | |
3337 | |
3338 /* Prefer an old insn to a bookkeeping insn. */ | |
3339 if (INSN_UID (tmp_insn) < first_emitted_uid | |
3340 && INSN_UID (tmp2_insn) >= first_emitted_uid) | |
3341 return -1; | |
3342 if (INSN_UID (tmp_insn) >= first_emitted_uid | |
3343 && INSN_UID (tmp2_insn) < first_emitted_uid) | |
3344 return 1; | |
3345 | |
3346 /* Prefer an insn with smaller UID, as a last resort. | |
3347 We can't safely use INSN_LUID as it is defined only for those insns | |
3348 that are in the stream. */ | |
3349 return INSN_UID (tmp_insn) - INSN_UID (tmp2_insn); | |
3350 } | |
3351 | |
3352 /* Filter out expressions from av set pointed to by AV_PTR | |
3353 that are pipelined too many times. */ | |
3354 static void | |
3355 process_pipelined_exprs (av_set_t *av_ptr) | |
3356 { | |
3357 expr_t expr; | |
3358 av_set_iterator si; | |
3359 | |
3360 /* Don't pipeline already pipelined code as that would increase | |
3361 number of unnecessary register moves. */ | |
3362 FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3363 { | |
3364 if (EXPR_SCHED_TIMES (expr) | |
3365 >= PARAM_VALUE (PARAM_SELSCHED_MAX_SCHED_TIMES)) | |
3366 av_set_iter_remove (&si); | |
3367 } | |
3368 } | |
3369 | |
3370 /* Filter speculative insns from AV_PTR if we don't want them. */ | |
3371 static void | |
3372 process_spec_exprs (av_set_t *av_ptr) | |
3373 { | |
3374 bool try_data_p = true; | |
3375 bool try_control_p = true; | |
3376 expr_t expr; | |
3377 av_set_iterator si; | |
3378 | |
3379 if (spec_info == NULL) | |
3380 return; | |
3381 | |
3382 /* Scan *AV_PTR to find out if we want to consider speculative | |
3383 instructions for scheduling. */ | |
3384 FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3385 { | |
3386 ds_t ds; | |
3387 | |
3388 ds = EXPR_SPEC_DONE_DS (expr); | |
3389 | |
3390 /* The probability of a success is too low - don't speculate. */ | |
3391 if ((ds & SPECULATIVE) | |
3392 && (ds_weak (ds) < spec_info->data_weakness_cutoff | |
3393 || EXPR_USEFULNESS (expr) < spec_info->control_weakness_cutoff | |
3394 || (pipelining_p && false | |
3395 && (ds & DATA_SPEC) | |
3396 && (ds & CONTROL_SPEC)))) | |
3397 { | |
3398 av_set_iter_remove (&si); | |
3399 continue; | |
3400 } | |
3401 | |
3402 if ((spec_info->flags & PREFER_NON_DATA_SPEC) | |
3403 && !(ds & BEGIN_DATA)) | |
3404 try_data_p = false; | |
3405 | |
3406 if ((spec_info->flags & PREFER_NON_CONTROL_SPEC) | |
3407 && !(ds & BEGIN_CONTROL)) | |
3408 try_control_p = false; | |
3409 } | |
3410 | |
3411 FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3412 { | |
3413 ds_t ds; | |
3414 | |
3415 ds = EXPR_SPEC_DONE_DS (expr); | |
3416 | |
3417 if (ds & SPECULATIVE) | |
3418 { | |
3419 if ((ds & BEGIN_DATA) && !try_data_p) | |
3420 /* We don't want any data speculative instructions right | |
3421 now. */ | |
3422 av_set_iter_remove (&si); | |
3423 | |
3424 if ((ds & BEGIN_CONTROL) && !try_control_p) | |
3425 /* We don't want any control speculative instructions right | |
3426 now. */ | |
3427 av_set_iter_remove (&si); | |
3428 } | |
3429 } | |
3430 } | |
3431 | |
3432 /* Search for any use-like insns in AV_PTR and decide on scheduling | |
3433 them. Return one when found, and NULL otherwise. | |
3434 Note that we check here whether a USE could be scheduled to avoid | |
3435 an infinite loop later. */ | |
3436 static expr_t | |
3437 process_use_exprs (av_set_t *av_ptr) | |
3438 { | |
3439 expr_t expr; | |
3440 av_set_iterator si; | |
3441 bool uses_present_p = false; | |
3442 bool try_uses_p = true; | |
3443 | |
3444 FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3445 { | |
3446 /* This will also initialize INSN_CODE for later use. */ | |
3447 if (recog_memoized (EXPR_INSN_RTX (expr)) < 0) | |
3448 { | |
3449 /* If we have a USE in *AV_PTR that was not scheduled yet, | |
3450 do so because it will do good only. */ | |
3451 if (EXPR_SCHED_TIMES (expr) <= 0) | |
3452 { | |
3453 if (EXPR_TARGET_AVAILABLE (expr) == 1) | |
3454 return expr; | |
3455 | |
3456 av_set_iter_remove (&si); | |
3457 } | |
3458 else | |
3459 { | |
3460 gcc_assert (pipelining_p); | |
3461 | |
3462 uses_present_p = true; | |
3463 } | |
3464 } | |
3465 else | |
3466 try_uses_p = false; | |
3467 } | |
3468 | |
3469 if (uses_present_p) | |
3470 { | |
3471 /* If we don't want to schedule any USEs right now and we have some | |
3472 in *AV_PTR, remove them, else just return the first one found. */ | |
3473 if (!try_uses_p) | |
3474 { | |
3475 FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3476 if (INSN_CODE (EXPR_INSN_RTX (expr)) < 0) | |
3477 av_set_iter_remove (&si); | |
3478 } | |
3479 else | |
3480 { | |
3481 FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3482 { | |
3483 gcc_assert (INSN_CODE (EXPR_INSN_RTX (expr)) < 0); | |
3484 | |
3485 if (EXPR_TARGET_AVAILABLE (expr) == 1) | |
3486 return expr; | |
3487 | |
3488 av_set_iter_remove (&si); | |
3489 } | |
3490 } | |
3491 } | |
3492 | |
3493 return NULL; | |
3494 } | |
3495 | |
3496 /* Lookup EXPR in VINSN_VEC and return TRUE if found. */ | |
3497 static bool | |
3498 vinsn_vec_has_expr_p (vinsn_vec_t vinsn_vec, expr_t expr) | |
3499 { | |
3500 vinsn_t vinsn; | |
3501 int n; | |
3502 | |
3503 for (n = 0; VEC_iterate (vinsn_t, vinsn_vec, n, vinsn); n++) | |
3504 if (VINSN_SEPARABLE_P (vinsn)) | |
3505 { | |
3506 if (vinsn_equal_p (vinsn, EXPR_VINSN (expr))) | |
3507 return true; | |
3508 } | |
3509 else | |
3510 { | |
3511 /* For non-separable instructions, the blocking insn can have | |
3512 another pattern due to substitution, and we can't choose | |
3513 different register as in the above case. Check all registers | |
3514 being written instead. */ | |
3515 if (bitmap_intersect_p (VINSN_REG_SETS (vinsn), | |
3516 VINSN_REG_SETS (EXPR_VINSN (expr)))) | |
3517 return true; | |
3518 } | |
3519 | |
3520 return false; | |
3521 } | |
3522 | |
3523 #ifdef ENABLE_CHECKING | |
3524 /* Return true if either of expressions from ORIG_OPS can be blocked | |
3525 by previously created bookkeeping code. STATIC_PARAMS points to static | |
3526 parameters of move_op. */ | |
3527 static bool | |
3528 av_set_could_be_blocked_by_bookkeeping_p (av_set_t orig_ops, void *static_params) | |
3529 { | |
3530 expr_t expr; | |
3531 av_set_iterator iter; | |
3532 moveop_static_params_p sparams; | |
3533 | |
3534 /* This checks that expressions in ORIG_OPS are not blocked by bookkeeping | |
3535 created while scheduling on another fence. */ | |
3536 FOR_EACH_EXPR (expr, iter, orig_ops) | |
3537 if (vinsn_vec_has_expr_p (vec_bookkeeping_blocked_vinsns, expr)) | |
3538 return true; | |
3539 | |
3540 gcc_assert (code_motion_path_driver_info == &move_op_hooks); | |
3541 sparams = (moveop_static_params_p) static_params; | |
3542 | |
3543 /* Expressions can be also blocked by bookkeeping created during current | |
3544 move_op. */ | |
3545 if (bitmap_bit_p (current_copies, INSN_UID (sparams->failed_insn))) | |
3546 FOR_EACH_EXPR (expr, iter, orig_ops) | |
3547 if (moveup_expr_cached (expr, sparams->failed_insn, false) != MOVEUP_EXPR_NULL) | |
3548 return true; | |
3549 | |
3550 /* Expressions in ORIG_OPS may have wrong destination register due to | |
3551 renaming. Check with the right register instead. */ | |
3552 if (sparams->dest && REG_P (sparams->dest)) | |
3553 { | |
3554 unsigned regno = REGNO (sparams->dest); | |
3555 vinsn_t failed_vinsn = INSN_VINSN (sparams->failed_insn); | |
3556 | |
3557 if (bitmap_bit_p (VINSN_REG_SETS (failed_vinsn), regno) | |
3558 || bitmap_bit_p (VINSN_REG_USES (failed_vinsn), regno) | |
3559 || bitmap_bit_p (VINSN_REG_CLOBBERS (failed_vinsn), regno)) | |
3560 return true; | |
3561 } | |
3562 | |
3563 return false; | |
3564 } | |
3565 #endif | |
3566 | |
3567 /* Clear VINSN_VEC and detach vinsns. */ | |
3568 static void | |
3569 vinsn_vec_clear (vinsn_vec_t *vinsn_vec) | |
3570 { | |
3571 unsigned len = VEC_length (vinsn_t, *vinsn_vec); | |
3572 if (len > 0) | |
3573 { | |
3574 vinsn_t vinsn; | |
3575 int n; | |
3576 | |
3577 for (n = 0; VEC_iterate (vinsn_t, *vinsn_vec, n, vinsn); n++) | |
3578 vinsn_detach (vinsn); | |
3579 VEC_block_remove (vinsn_t, *vinsn_vec, 0, len); | |
3580 } | |
3581 } | |
3582 | |
3583 /* Add the vinsn of EXPR to the VINSN_VEC. */ | |
3584 static void | |
3585 vinsn_vec_add (vinsn_vec_t *vinsn_vec, expr_t expr) | |
3586 { | |
3587 vinsn_attach (EXPR_VINSN (expr)); | |
3588 VEC_safe_push (vinsn_t, heap, *vinsn_vec, EXPR_VINSN (expr)); | |
3589 } | |
3590 | |
3591 /* Free the vector representing blocked expressions. */ | |
3592 static void | |
3593 vinsn_vec_free (vinsn_vec_t *vinsn_vec) | |
3594 { | |
3595 if (*vinsn_vec) | |
3596 VEC_free (vinsn_t, heap, *vinsn_vec); | |
3597 } | |
3598 | |
3599 /* Increase EXPR_PRIORITY_ADJ for INSN by AMOUNT. */ | |
3600 | |
3601 void sel_add_to_insn_priority (rtx insn, int amount) | |
3602 { | |
3603 EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) += amount; | |
3604 | |
3605 if (sched_verbose >= 2) | |
3606 sel_print ("sel_add_to_insn_priority: insn %d, by %d (now %d+%d).\n", | |
3607 INSN_UID (insn), amount, EXPR_PRIORITY (INSN_EXPR (insn)), | |
3608 EXPR_PRIORITY_ADJ (INSN_EXPR (insn))); | |
3609 } | |
3610 | |
3611 /* Turn AV into a vector, filter inappropriate insns and sort it. Return | |
3612 true if there is something to schedule. BNDS and FENCE are current | |
3613 boundaries and fence, respectively. If we need to stall for some cycles | |
3614 before an expr from AV would become available, write this number to | |
3615 *PNEED_STALL. */ | |
3616 static bool | |
3617 fill_vec_av_set (av_set_t av, blist_t bnds, fence_t fence, | |
3618 int *pneed_stall) | |
3619 { | |
3620 av_set_iterator si; | |
3621 expr_t expr; | |
3622 int sched_next_worked = 0, stalled, n; | |
3623 static int av_max_prio, est_ticks_till_branch; | |
3624 int min_need_stall = -1; | |
3625 deps_t dc = BND_DC (BLIST_BND (bnds)); | |
3626 | |
3627 /* Bail out early when the ready list contained only USEs/CLOBBERs that are | |
3628 already scheduled. */ | |
3629 if (av == NULL) | |
3630 return false; | |
3631 | |
3632 /* Empty vector from the previous stuff. */ | |
3633 if (VEC_length (expr_t, vec_av_set) > 0) | |
3634 VEC_block_remove (expr_t, vec_av_set, 0, VEC_length (expr_t, vec_av_set)); | |
3635 | |
3636 /* Turn the set into a vector for sorting and call sel_target_adjust_priority | |
3637 for each insn. */ | |
3638 gcc_assert (VEC_empty (expr_t, vec_av_set)); | |
3639 FOR_EACH_EXPR (expr, si, av) | |
3640 { | |
3641 VEC_safe_push (expr_t, heap, vec_av_set, expr); | |
3642 | |
3643 gcc_assert (EXPR_PRIORITY_ADJ (expr) == 0 || *pneed_stall); | |
3644 | |
3645 /* Adjust priority using target backend hook. */ | |
3646 sel_target_adjust_priority (expr); | |
3647 } | |
3648 | |
3649 /* Sort the vector. */ | |
3650 qsort (VEC_address (expr_t, vec_av_set), VEC_length (expr_t, vec_av_set), | |
3651 sizeof (expr_t), sel_rank_for_schedule); | |
3652 | |
3653 /* We record maximal priority of insns in av set for current instruction | |
3654 group. */ | |
3655 if (FENCE_STARTS_CYCLE_P (fence)) | |
3656 av_max_prio = est_ticks_till_branch = INT_MIN; | |
3657 | |
3658 /* Filter out inappropriate expressions. Loop's direction is reversed to | |
3659 visit "best" instructions first. We assume that VEC_unordered_remove | |
3660 moves last element in place of one being deleted. */ | |
3661 for (n = VEC_length (expr_t, vec_av_set) - 1, stalled = 0; n >= 0; n--) | |
3662 { | |
3663 expr_t expr = VEC_index (expr_t, vec_av_set, n); | |
3664 insn_t insn = EXPR_INSN_RTX (expr); | |
3665 char target_available; | |
3666 bool is_orig_reg_p = true; | |
3667 int need_cycles, new_prio; | |
3668 | |
3669 /* Don't allow any insns other than from SCHED_GROUP if we have one. */ | |
3670 if (FENCE_SCHED_NEXT (fence) && insn != FENCE_SCHED_NEXT (fence)) | |
3671 { | |
3672 VEC_unordered_remove (expr_t, vec_av_set, n); | |
3673 continue; | |
3674 } | |
3675 | |
3676 /* Set number of sched_next insns (just in case there | |
3677 could be several). */ | |
3678 if (FENCE_SCHED_NEXT (fence)) | |
3679 sched_next_worked++; | |
3680 | |
3681 /* Check all liveness requirements and try renaming. | |
3682 FIXME: try to minimize calls to this. */ | |
3683 target_available = EXPR_TARGET_AVAILABLE (expr); | |
3684 | |
3685 /* If insn was already scheduled on the current fence, | |
3686 set TARGET_AVAILABLE to -1 no matter what expr's attribute says. */ | |
3687 if (vinsn_vec_has_expr_p (vec_target_unavailable_vinsns, expr)) | |
3688 target_available = -1; | |
3689 | |
3690 /* If the availability of the EXPR is invalidated by the insertion of | |
3691 bookkeeping earlier, make sure that we won't choose this expr for | |
3692 scheduling if it's not separable, and if it is separable, then | |
3693 we have to recompute the set of available registers for it. */ | |
3694 if (vinsn_vec_has_expr_p (vec_bookkeeping_blocked_vinsns, expr)) | |
3695 { | |
3696 VEC_unordered_remove (expr_t, vec_av_set, n); | |
3697 if (sched_verbose >= 4) | |
3698 sel_print ("Expr %d is blocked by bookkeeping inserted earlier\n", | |
3699 INSN_UID (insn)); | |
3700 continue; | |
3701 } | |
3702 | |
3703 if (target_available == true) | |
3704 { | |
3705 /* Do nothing -- we can use an existing register. */ | |
3706 is_orig_reg_p = EXPR_SEPARABLE_P (expr); | |
3707 } | |
3708 else if (/* Non-separable instruction will never | |
3709 get another register. */ | |
3710 (target_available == false | |
3711 && !EXPR_SEPARABLE_P (expr)) | |
3712 /* Don't try to find a register for low-priority expression. */ | |
3713 || (int) VEC_length (expr_t, vec_av_set) - 1 - n >= max_insns_to_rename | |
3714 /* ??? FIXME: Don't try to rename data speculation. */ | |
3715 || (EXPR_SPEC_DONE_DS (expr) & BEGIN_DATA) | |
3716 || ! find_best_reg_for_expr (expr, bnds, &is_orig_reg_p)) | |
3717 { | |
3718 VEC_unordered_remove (expr_t, vec_av_set, n); | |
3719 if (sched_verbose >= 4) | |
3720 sel_print ("Expr %d has no suitable target register\n", | |
3721 INSN_UID (insn)); | |
3722 continue; | |
3723 } | |
3724 | |
3725 /* Filter expressions that need to be renamed or speculated when | |
3726 pipelining, because compensating register copies or speculation | |
3727 checks are likely to be placed near the beginning of the loop, | |
3728 causing a stall. */ | |
3729 if (pipelining_p && EXPR_ORIG_SCHED_CYCLE (expr) > 0 | |
3730 && (!is_orig_reg_p || EXPR_SPEC_DONE_DS (expr) != 0)) | |
3731 { | |
3732 /* Estimation of number of cycles until loop branch for | |
3733 renaming/speculation to be successful. */ | |
3734 int need_n_ticks_till_branch = sel_vinsn_cost (EXPR_VINSN (expr)); | |
3735 | |
3736 if ((int) current_loop_nest->ninsns < 9) | |
3737 { | |
3738 VEC_unordered_remove (expr_t, vec_av_set, n); | |
3739 if (sched_verbose >= 4) | |
3740 sel_print ("Pipelining expr %d will likely cause stall\n", | |
3741 INSN_UID (insn)); | |
3742 continue; | |
3743 } | |
3744 | |
3745 if ((int) current_loop_nest->ninsns - num_insns_scheduled | |
3746 < need_n_ticks_till_branch * issue_rate / 2 | |
3747 && est_ticks_till_branch < need_n_ticks_till_branch) | |
3748 { | |
3749 VEC_unordered_remove (expr_t, vec_av_set, n); | |
3750 if (sched_verbose >= 4) | |
3751 sel_print ("Pipelining expr %d will likely cause stall\n", | |
3752 INSN_UID (insn)); | |
3753 continue; | |
3754 } | |
3755 } | |
3756 | |
3757 /* We want to schedule speculation checks as late as possible. Discard | |
3758 them from av set if there are instructions with higher priority. */ | |
3759 if (sel_insn_is_speculation_check (insn) | |
3760 && EXPR_PRIORITY (expr) < av_max_prio) | |
3761 { | |
3762 stalled++; | |
3763 min_need_stall = min_need_stall < 0 ? 1 : MIN (min_need_stall, 1); | |
3764 VEC_unordered_remove (expr_t, vec_av_set, n); | |
3765 if (sched_verbose >= 4) | |
3766 sel_print ("Delaying speculation check %d until its first use\n", | |
3767 INSN_UID (insn)); | |
3768 continue; | |
3769 } | |
3770 | |
3771 /* Ignore EXPRs available from pipelining to update AV_MAX_PRIO. */ | |
3772 if (EXPR_ORIG_SCHED_CYCLE (expr) <= 0) | |
3773 av_max_prio = MAX (av_max_prio, EXPR_PRIORITY (expr)); | |
3774 | |
3775 /* Don't allow any insns whose data is not yet ready. | |
3776 Check first whether we've already tried them and failed. */ | |
3777 if (INSN_UID (insn) < FENCE_READY_TICKS_SIZE (fence)) | |
3778 { | |
3779 need_cycles = (FENCE_READY_TICKS (fence)[INSN_UID (insn)] | |
3780 - FENCE_CYCLE (fence)); | |
3781 if (EXPR_ORIG_SCHED_CYCLE (expr) <= 0) | |
3782 est_ticks_till_branch = MAX (est_ticks_till_branch, | |
3783 EXPR_PRIORITY (expr) + need_cycles); | |
3784 | |
3785 if (need_cycles > 0) | |
3786 { | |
3787 stalled++; | |
3788 min_need_stall = (min_need_stall < 0 | |
3789 ? need_cycles | |
3790 : MIN (min_need_stall, need_cycles)); | |
3791 VEC_unordered_remove (expr_t, vec_av_set, n); | |
3792 | |
3793 if (sched_verbose >= 4) | |
3794 sel_print ("Expr %d is not ready until cycle %d (cached)\n", | |
3795 INSN_UID (insn), | |
3796 FENCE_READY_TICKS (fence)[INSN_UID (insn)]); | |
3797 continue; | |
3798 } | |
3799 } | |
3800 | |
3801 /* Now resort to dependence analysis to find whether EXPR might be | |
3802 stalled due to dependencies from FENCE's context. */ | |
3803 need_cycles = tick_check_p (expr, dc, fence); | |
3804 new_prio = EXPR_PRIORITY (expr) + EXPR_PRIORITY_ADJ (expr) + need_cycles; | |
3805 | |
3806 if (EXPR_ORIG_SCHED_CYCLE (expr) <= 0) | |
3807 est_ticks_till_branch = MAX (est_ticks_till_branch, | |
3808 new_prio); | |
3809 | |
3810 if (need_cycles > 0) | |
3811 { | |
3812 if (INSN_UID (insn) >= FENCE_READY_TICKS_SIZE (fence)) | |
3813 { | |
3814 int new_size = INSN_UID (insn) * 3 / 2; | |
3815 | |
3816 FENCE_READY_TICKS (fence) | |
3817 = (int *) xrecalloc (FENCE_READY_TICKS (fence), | |
3818 new_size, FENCE_READY_TICKS_SIZE (fence), | |
3819 sizeof (int)); | |
3820 } | |
3821 FENCE_READY_TICKS (fence)[INSN_UID (insn)] | |
3822 = FENCE_CYCLE (fence) + need_cycles; | |
3823 | |
3824 stalled++; | |
3825 min_need_stall = (min_need_stall < 0 | |
3826 ? need_cycles | |
3827 : MIN (min_need_stall, need_cycles)); | |
3828 | |
3829 VEC_unordered_remove (expr_t, vec_av_set, n); | |
3830 | |
3831 if (sched_verbose >= 4) | |
3832 sel_print ("Expr %d is not ready yet until cycle %d\n", | |
3833 INSN_UID (insn), | |
3834 FENCE_READY_TICKS (fence)[INSN_UID (insn)]); | |
3835 continue; | |
3836 } | |
3837 | |
3838 if (sched_verbose >= 4) | |
3839 sel_print ("Expr %d is ok\n", INSN_UID (insn)); | |
3840 min_need_stall = 0; | |
3841 } | |
3842 | |
3843 /* Clear SCHED_NEXT. */ | |
3844 if (FENCE_SCHED_NEXT (fence)) | |
3845 { | |
3846 gcc_assert (sched_next_worked == 1); | |
3847 FENCE_SCHED_NEXT (fence) = NULL_RTX; | |
3848 } | |
3849 | |
3850 /* No need to stall if this variable was not initialized. */ | |
3851 if (min_need_stall < 0) | |
3852 min_need_stall = 0; | |
3853 | |
3854 if (VEC_empty (expr_t, vec_av_set)) | |
3855 { | |
3856 /* We need to set *pneed_stall here, because later we skip this code | |
3857 when ready list is empty. */ | |
3858 *pneed_stall = min_need_stall; | |
3859 return false; | |
3860 } | |
3861 else | |
3862 gcc_assert (min_need_stall == 0); | |
3863 | |
3864 /* Sort the vector. */ | |
3865 qsort (VEC_address (expr_t, vec_av_set), VEC_length (expr_t, vec_av_set), | |
3866 sizeof (expr_t), sel_rank_for_schedule); | |
3867 | |
3868 if (sched_verbose >= 4) | |
3869 { | |
3870 sel_print ("Total ready exprs: %d, stalled: %d\n", | |
3871 VEC_length (expr_t, vec_av_set), stalled); | |
3872 sel_print ("Sorted av set (%d): ", VEC_length (expr_t, vec_av_set)); | |
3873 for (n = 0; VEC_iterate (expr_t, vec_av_set, n, expr); n++) | |
3874 dump_expr (expr); | |
3875 sel_print ("\n"); | |
3876 } | |
3877 | |
3878 *pneed_stall = 0; | |
3879 return true; | |
3880 } | |
3881 | |
3882 /* Convert a vectored and sorted av set to the ready list that | |
3883 the rest of the backend wants to see. */ | |
3884 static void | |
3885 convert_vec_av_set_to_ready (void) | |
3886 { | |
3887 int n; | |
3888 expr_t expr; | |
3889 | |
3890 /* Allocate and fill the ready list from the sorted vector. */ | |
3891 ready.n_ready = VEC_length (expr_t, vec_av_set); | |
3892 ready.first = ready.n_ready - 1; | |
3893 | |
3894 gcc_assert (ready.n_ready > 0); | |
3895 | |
3896 if (ready.n_ready > max_issue_size) | |
3897 { | |
3898 max_issue_size = ready.n_ready; | |
3899 sched_extend_ready_list (ready.n_ready); | |
3900 } | |
3901 | |
3902 for (n = 0; VEC_iterate (expr_t, vec_av_set, n, expr); n++) | |
3903 { | |
3904 vinsn_t vi = EXPR_VINSN (expr); | |
3905 insn_t insn = VINSN_INSN_RTX (vi); | |
3906 | |
3907 ready_try[n] = 0; | |
3908 ready.vec[n] = insn; | |
3909 } | |
3910 } | |
3911 | |
3912 /* Initialize ready list from *AV_PTR for the max_issue () call. | |
3913 If any unrecognizable insn found in *AV_PTR, return it (and skip | |
3914 max_issue). BND and FENCE are current boundary and fence, | |
3915 respectively. If we need to stall for some cycles before an expr | |
3916 from *AV_PTR would become available, write this number to *PNEED_STALL. */ | |
3917 static expr_t | |
3918 fill_ready_list (av_set_t *av_ptr, blist_t bnds, fence_t fence, | |
3919 int *pneed_stall) | |
3920 { | |
3921 expr_t expr; | |
3922 | |
3923 /* We do not support multiple boundaries per fence. */ | |
3924 gcc_assert (BLIST_NEXT (bnds) == NULL); | |
3925 | |
3926 /* Process expressions required special handling, i.e. pipelined, | |
3927 speculative and recog() < 0 expressions first. */ | |
3928 process_pipelined_exprs (av_ptr); | |
3929 process_spec_exprs (av_ptr); | |
3930 | |
3931 /* A USE could be scheduled immediately. */ | |
3932 expr = process_use_exprs (av_ptr); | |
3933 if (expr) | |
3934 { | |
3935 *pneed_stall = 0; | |
3936 return expr; | |
3937 } | |
3938 | |
3939 /* Turn the av set to a vector for sorting. */ | |
3940 if (! fill_vec_av_set (*av_ptr, bnds, fence, pneed_stall)) | |
3941 { | |
3942 ready.n_ready = 0; | |
3943 return NULL; | |
3944 } | |
3945 | |
3946 /* Build the final ready list. */ | |
3947 convert_vec_av_set_to_ready (); | |
3948 return NULL; | |
3949 } | |
3950 | |
3951 /* Wrapper for dfa_new_cycle (). Returns TRUE if cycle was advanced. */ | |
3952 static bool | |
3953 sel_dfa_new_cycle (insn_t insn, fence_t fence) | |
3954 { | |
3955 int last_scheduled_cycle = FENCE_LAST_SCHEDULED_INSN (fence) | |
3956 ? INSN_SCHED_CYCLE (FENCE_LAST_SCHEDULED_INSN (fence)) | |
3957 : FENCE_CYCLE (fence) - 1; | |
3958 bool res = false; | |
3959 int sort_p = 0; | |
3960 | |
3961 if (!targetm.sched.dfa_new_cycle) | |
3962 return false; | |
3963 | |
3964 memcpy (curr_state, FENCE_STATE (fence), dfa_state_size); | |
3965 | |
3966 while (!sort_p && targetm.sched.dfa_new_cycle (sched_dump, sched_verbose, | |
3967 insn, last_scheduled_cycle, | |
3968 FENCE_CYCLE (fence), &sort_p)) | |
3969 { | |
3970 memcpy (FENCE_STATE (fence), curr_state, dfa_state_size); | |
3971 advance_one_cycle (fence); | |
3972 memcpy (curr_state, FENCE_STATE (fence), dfa_state_size); | |
3973 res = true; | |
3974 } | |
3975 | |
3976 return res; | |
3977 } | |
3978 | |
3979 /* Invoke reorder* target hooks on the ready list. Return the number of insns | |
3980 we can issue. FENCE is the current fence. */ | |
3981 static int | |
3982 invoke_reorder_hooks (fence_t fence) | |
3983 { | |
3984 int issue_more; | |
3985 bool ran_hook = false; | |
3986 | |
3987 /* Call the reorder hook at the beginning of the cycle, and call | |
3988 the reorder2 hook in the middle of the cycle. */ | |
3989 if (FENCE_ISSUED_INSNS (fence) == 0) | |
3990 { | |
3991 if (targetm.sched.reorder | |
3992 && !SCHED_GROUP_P (ready_element (&ready, 0)) | |
3993 && ready.n_ready > 1) | |
3994 { | |
3995 /* Don't give reorder the most prioritized insn as it can break | |
3996 pipelining. */ | |
3997 if (pipelining_p) | |
3998 --ready.n_ready; | |
3999 | |
4000 issue_more | |
4001 = targetm.sched.reorder (sched_dump, sched_verbose, | |
4002 ready_lastpos (&ready), | |
4003 &ready.n_ready, FENCE_CYCLE (fence)); | |
4004 | |
4005 if (pipelining_p) | |
4006 ++ready.n_ready; | |
4007 | |
4008 ran_hook = true; | |
4009 } | |
4010 else | |
4011 /* Initialize can_issue_more for variable_issue. */ | |
4012 issue_more = issue_rate; | |
4013 } | |
4014 else if (targetm.sched.reorder2 | |
4015 && !SCHED_GROUP_P (ready_element (&ready, 0))) | |
4016 { | |
4017 if (ready.n_ready == 1) | |
4018 issue_more = | |
4019 targetm.sched.reorder2 (sched_dump, sched_verbose, | |
4020 ready_lastpos (&ready), | |
4021 &ready.n_ready, FENCE_CYCLE (fence)); | |
4022 else | |
4023 { | |
4024 if (pipelining_p) | |
4025 --ready.n_ready; | |
4026 | |
4027 issue_more = | |
4028 targetm.sched.reorder2 (sched_dump, sched_verbose, | |
4029 ready.n_ready | |
4030 ? ready_lastpos (&ready) : NULL, | |
4031 &ready.n_ready, FENCE_CYCLE (fence)); | |
4032 | |
4033 if (pipelining_p) | |
4034 ++ready.n_ready; | |
4035 } | |
4036 | |
4037 ran_hook = true; | |
4038 } | |
4039 else | |
4040 issue_more = issue_rate; | |
4041 | |
4042 /* Ensure that ready list and vec_av_set are in line with each other, | |
4043 i.e. vec_av_set[i] == ready_element (&ready, i). */ | |
4044 if (issue_more && ran_hook) | |
4045 { | |
4046 int i, j, n; | |
4047 rtx *arr = ready.vec; | |
4048 expr_t *vec = VEC_address (expr_t, vec_av_set); | |
4049 | |
4050 for (i = 0, n = ready.n_ready; i < n; i++) | |
4051 if (EXPR_INSN_RTX (vec[i]) != arr[i]) | |
4052 { | |
4053 expr_t tmp; | |
4054 | |
4055 for (j = i; j < n; j++) | |
4056 if (EXPR_INSN_RTX (vec[j]) == arr[i]) | |
4057 break; | |
4058 gcc_assert (j < n); | |
4059 | |
4060 tmp = vec[i]; | |
4061 vec[i] = vec[j]; | |
4062 vec[j] = tmp; | |
4063 } | |
4064 } | |
4065 | |
4066 return issue_more; | |
4067 } | |
4068 | |
4069 /* Return an EXPR correponding to INDEX element of ready list, if | |
4070 FOLLOW_READY_ELEMENT is true (i.e., an expr of | |
4071 ready_element (&ready, INDEX) will be returned), and to INDEX element of | |
4072 ready.vec otherwise. */ | |
4073 static inline expr_t | |
4074 find_expr_for_ready (int index, bool follow_ready_element) | |
4075 { | |
4076 expr_t expr; | |
4077 int real_index; | |
4078 | |
4079 real_index = follow_ready_element ? ready.first - index : index; | |
4080 | |
4081 expr = VEC_index (expr_t, vec_av_set, real_index); | |
4082 gcc_assert (ready.vec[real_index] == EXPR_INSN_RTX (expr)); | |
4083 | |
4084 return expr; | |
4085 } | |
4086 | |
4087 /* Calculate insns worth trying via lookahead_guard hook. Return a number | |
4088 of such insns found. */ | |
4089 static int | |
4090 invoke_dfa_lookahead_guard (void) | |
4091 { | |
4092 int i, n; | |
4093 bool have_hook | |
4094 = targetm.sched.first_cycle_multipass_dfa_lookahead_guard != NULL; | |
4095 | |
4096 if (sched_verbose >= 2) | |
4097 sel_print ("ready after reorder: "); | |
4098 | |
4099 for (i = 0, n = 0; i < ready.n_ready; i++) | |
4100 { | |
4101 expr_t expr; | |
4102 insn_t insn; | |
4103 int r; | |
4104 | |
4105 /* In this loop insn is Ith element of the ready list given by | |
4106 ready_element, not Ith element of ready.vec. */ | |
4107 insn = ready_element (&ready, i); | |
4108 | |
4109 if (! have_hook || i == 0) | |
4110 r = 0; | |
4111 else | |
4112 r = !targetm.sched.first_cycle_multipass_dfa_lookahead_guard (insn); | |
4113 | |
4114 gcc_assert (INSN_CODE (insn) >= 0); | |
4115 | |
4116 /* Only insns with ready_try = 0 can get here | |
4117 from fill_ready_list. */ | |
4118 gcc_assert (ready_try [i] == 0); | |
4119 ready_try[i] = r; | |
4120 if (!r) | |
4121 n++; | |
4122 | |
4123 expr = find_expr_for_ready (i, true); | |
4124 | |
4125 if (sched_verbose >= 2) | |
4126 { | |
4127 dump_vinsn (EXPR_VINSN (expr)); | |
4128 sel_print (":%d; ", ready_try[i]); | |
4129 } | |
4130 } | |
4131 | |
4132 if (sched_verbose >= 2) | |
4133 sel_print ("\n"); | |
4134 return n; | |
4135 } | |
4136 | |
4137 /* Calculate the number of privileged insns and return it. */ | |
4138 static int | |
4139 calculate_privileged_insns (void) | |
4140 { | |
4141 expr_t cur_expr, min_spec_expr = NULL; | |
4142 insn_t cur_insn, min_spec_insn; | |
4143 int privileged_n = 0, i; | |
4144 | |
4145 for (i = 0; i < ready.n_ready; i++) | |
4146 { | |
4147 if (ready_try[i]) | |
4148 continue; | |
4149 | |
4150 if (! min_spec_expr) | |
4151 { | |
4152 min_spec_insn = ready_element (&ready, i); | |
4153 min_spec_expr = find_expr_for_ready (i, true); | |
4154 } | |
4155 | |
4156 cur_insn = ready_element (&ready, i); | |
4157 cur_expr = find_expr_for_ready (i, true); | |
4158 | |
4159 if (EXPR_SPEC (cur_expr) > EXPR_SPEC (min_spec_expr)) | |
4160 break; | |
4161 | |
4162 ++privileged_n; | |
4163 } | |
4164 | |
4165 if (i == ready.n_ready) | |
4166 privileged_n = 0; | |
4167 | |
4168 if (sched_verbose >= 2) | |
4169 sel_print ("privileged_n: %d insns with SPEC %d\n", | |
4170 privileged_n, privileged_n ? EXPR_SPEC (min_spec_expr) : -1); | |
4171 return privileged_n; | |
4172 } | |
4173 | |
4174 /* Call the rest of the hooks after the choice was made. Return | |
4175 the number of insns that still can be issued given that the current | |
4176 number is ISSUE_MORE. FENCE and BEST_INSN are the current fence | |
4177 and the insn chosen for scheduling, respectively. */ | |
4178 static int | |
4179 invoke_aftermath_hooks (fence_t fence, rtx best_insn, int issue_more) | |
4180 { | |
4181 gcc_assert (INSN_P (best_insn)); | |
4182 | |
4183 /* First, call dfa_new_cycle, and then variable_issue, if available. */ | |
4184 sel_dfa_new_cycle (best_insn, fence); | |
4185 | |
4186 if (targetm.sched.variable_issue) | |
4187 { | |
4188 memcpy (curr_state, FENCE_STATE (fence), dfa_state_size); | |
4189 issue_more = | |
4190 targetm.sched.variable_issue (sched_dump, sched_verbose, best_insn, | |
4191 issue_more); | |
4192 memcpy (FENCE_STATE (fence), curr_state, dfa_state_size); | |
4193 } | |
4194 else if (GET_CODE (PATTERN (best_insn)) != USE | |
4195 && GET_CODE (PATTERN (best_insn)) != CLOBBER) | |
4196 issue_more--; | |
4197 | |
4198 return issue_more; | |
4199 } | |
4200 | |
4201 /* Estimate the cost of issuing INSN on DFA state STATE. */ | |
4202 static int | |
4203 estimate_insn_cost (rtx insn, state_t state) | |
4204 { | |
4205 static state_t temp = NULL; | |
4206 int cost; | |
4207 | |
4208 if (!temp) | |
4209 temp = xmalloc (dfa_state_size); | |
4210 | |
4211 memcpy (temp, state, dfa_state_size); | |
4212 cost = state_transition (temp, insn); | |
4213 | |
4214 if (cost < 0) | |
4215 return 0; | |
4216 else if (cost == 0) | |
4217 return 1; | |
4218 return cost; | |
4219 } | |
4220 | |
4221 /* Return the cost of issuing EXPR on the FENCE as estimated by DFA. | |
4222 This function properly handles ASMs, USEs etc. */ | |
4223 static int | |
4224 get_expr_cost (expr_t expr, fence_t fence) | |
4225 { | |
4226 rtx insn = EXPR_INSN_RTX (expr); | |
4227 | |
4228 if (recog_memoized (insn) < 0) | |
4229 { | |
4230 if (!FENCE_STARTS_CYCLE_P (fence) | |
4231 /* FIXME: Is this condition necessary? */ | |
4232 && VINSN_UNIQUE_P (EXPR_VINSN (expr)) | |
4233 && INSN_ASM_P (insn)) | |
4234 /* This is asm insn which is tryed to be issued on the | |
4235 cycle not first. Issue it on the next cycle. */ | |
4236 return 1; | |
4237 else | |
4238 /* A USE insn, or something else we don't need to | |
4239 understand. We can't pass these directly to | |
4240 state_transition because it will trigger a | |
4241 fatal error for unrecognizable insns. */ | |
4242 return 0; | |
4243 } | |
4244 else | |
4245 return estimate_insn_cost (insn, FENCE_STATE (fence)); | |
4246 } | |
4247 | |
4248 /* Find the best insn for scheduling, either via max_issue or just take | |
4249 the most prioritized available. */ | |
4250 static int | |
4251 choose_best_insn (fence_t fence, int privileged_n, int *index) | |
4252 { | |
4253 int can_issue = 0; | |
4254 | |
4255 if (dfa_lookahead > 0) | |
4256 { | |
4257 cycle_issued_insns = FENCE_ISSUED_INSNS (fence); | |
4258 can_issue = max_issue (&ready, privileged_n, | |
4259 FENCE_STATE (fence), index); | |
4260 if (sched_verbose >= 2) | |
4261 sel_print ("max_issue: we can issue %d insns, already did %d insns\n", | |
4262 can_issue, FENCE_ISSUED_INSNS (fence)); | |
4263 } | |
4264 else | |
4265 { | |
4266 /* We can't use max_issue; just return the first available element. */ | |
4267 int i; | |
4268 | |
4269 for (i = 0; i < ready.n_ready; i++) | |
4270 { | |
4271 expr_t expr = find_expr_for_ready (i, true); | |
4272 | |
4273 if (get_expr_cost (expr, fence) < 1) | |
4274 { | |
4275 can_issue = can_issue_more; | |
4276 *index = i; | |
4277 | |
4278 if (sched_verbose >= 2) | |
4279 sel_print ("using %dth insn from the ready list\n", i + 1); | |
4280 | |
4281 break; | |
4282 } | |
4283 } | |
4284 | |
4285 if (i == ready.n_ready) | |
4286 { | |
4287 can_issue = 0; | |
4288 *index = -1; | |
4289 } | |
4290 } | |
4291 | |
4292 return can_issue; | |
4293 } | |
4294 | |
4295 /* Choose the best expr from *AV_VLIW_PTR and a suitable register for it. | |
4296 BNDS and FENCE are current boundaries and scheduling fence respectively. | |
4297 Return the expr found and NULL if nothing can be issued atm. | |
4298 Write to PNEED_STALL the number of cycles to stall if no expr was found. */ | |
4299 static expr_t | |
4300 find_best_expr (av_set_t *av_vliw_ptr, blist_t bnds, fence_t fence, | |
4301 int *pneed_stall) | |
4302 { | |
4303 expr_t best; | |
4304 | |
4305 /* Choose the best insn for scheduling via: | |
4306 1) sorting the ready list based on priority; | |
4307 2) calling the reorder hook; | |
4308 3) calling max_issue. */ | |
4309 best = fill_ready_list (av_vliw_ptr, bnds, fence, pneed_stall); | |
4310 if (best == NULL && ready.n_ready > 0) | |
4311 { | |
4312 int privileged_n, index, avail_n; | |
4313 | |
4314 can_issue_more = invoke_reorder_hooks (fence); | |
4315 if (can_issue_more > 0) | |
4316 { | |
4317 /* Try choosing the best insn until we find one that is could be | |
4318 scheduled due to liveness restrictions on its destination register. | |
4319 In the future, we'd like to choose once and then just probe insns | |
4320 in the order of their priority. */ | |
4321 avail_n = invoke_dfa_lookahead_guard (); | |
4322 privileged_n = calculate_privileged_insns (); | |
4323 can_issue_more = choose_best_insn (fence, privileged_n, &index); | |
4324 if (can_issue_more) | |
4325 best = find_expr_for_ready (index, true); | |
4326 } | |
4327 /* We had some available insns, so if we can't issue them, | |
4328 we have a stall. */ | |
4329 if (can_issue_more == 0) | |
4330 { | |
4331 best = NULL; | |
4332 *pneed_stall = 1; | |
4333 } | |
4334 } | |
4335 | |
4336 if (best != NULL) | |
4337 { | |
4338 can_issue_more = invoke_aftermath_hooks (fence, EXPR_INSN_RTX (best), | |
4339 can_issue_more); | |
4340 if (can_issue_more == 0) | |
4341 *pneed_stall = 1; | |
4342 } | |
4343 | |
4344 if (sched_verbose >= 2) | |
4345 { | |
4346 if (best != NULL) | |
4347 { | |
4348 sel_print ("Best expression (vliw form): "); | |
4349 dump_expr (best); | |
4350 sel_print ("; cycle %d\n", FENCE_CYCLE (fence)); | |
4351 } | |
4352 else | |
4353 sel_print ("No best expr found!\n"); | |
4354 } | |
4355 | |
4356 return best; | |
4357 } | |
4358 | |
4359 | |
4360 /* Functions that implement the core of the scheduler. */ | |
4361 | |
4362 | |
4363 /* Emit an instruction from EXPR with SEQNO and VINSN after | |
4364 PLACE_TO_INSERT. */ | |
4365 static insn_t | |
4366 emit_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno, | |
4367 insn_t place_to_insert) | |
4368 { | |
4369 /* This assert fails when we have identical instructions | |
4370 one of which dominates the other. In this case move_op () | |
4371 finds the first instruction and doesn't search for second one. | |
4372 The solution would be to compute av_set after the first found | |
4373 insn and, if insn present in that set, continue searching. | |
4374 For now we workaround this issue in move_op. */ | |
4375 gcc_assert (!INSN_IN_STREAM_P (EXPR_INSN_RTX (expr))); | |
4376 | |
4377 if (EXPR_WAS_RENAMED (expr)) | |
4378 { | |
4379 unsigned regno = expr_dest_regno (expr); | |
4380 | |
4381 if (HARD_REGISTER_NUM_P (regno)) | |
4382 { | |
4383 df_set_regs_ever_live (regno, true); | |
4384 reg_rename_tick[regno] = ++reg_rename_this_tick; | |
4385 } | |
4386 } | |
4387 | |
4388 return sel_gen_insn_from_expr_after (expr, vinsn, seqno, | |
4389 place_to_insert); | |
4390 } | |
4391 | |
4392 /* Return TRUE if BB can hold bookkeeping code. */ | |
4393 static bool | |
4394 block_valid_for_bookkeeping_p (basic_block bb) | |
4395 { | |
4396 insn_t bb_end = BB_END (bb); | |
4397 | |
4398 if (!in_current_region_p (bb) || EDGE_COUNT (bb->succs) > 1) | |
4399 return false; | |
4400 | |
4401 if (INSN_P (bb_end)) | |
4402 { | |
4403 if (INSN_SCHED_TIMES (bb_end) > 0) | |
4404 return false; | |
4405 } | |
4406 else | |
4407 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (bb_end)); | |
4408 | |
4409 return true; | |
4410 } | |
4411 | |
4412 /* Attempt to find a block that can hold bookkeeping code for path(s) incoming | |
4413 into E2->dest, except from E1->src (there may be a sequence of empty basic | |
4414 blocks between E1->src and E2->dest). Return found block, or NULL if new | |
4415 one must be created. */ | |
4416 static basic_block | |
4417 find_block_for_bookkeeping (edge e1, edge e2) | |
4418 { | |
4419 basic_block candidate_block = NULL; | |
4420 edge e; | |
4421 | |
4422 /* Loop over edges from E1 to E2, inclusive. */ | |
4423 for (e = e1; ; e = EDGE_SUCC (e->dest, 0)) | |
4424 { | |
4425 if (EDGE_COUNT (e->dest->preds) == 2) | |
4426 { | |
4427 if (candidate_block == NULL) | |
4428 candidate_block = (EDGE_PRED (e->dest, 0) == e | |
4429 ? EDGE_PRED (e->dest, 1)->src | |
4430 : EDGE_PRED (e->dest, 0)->src); | |
4431 else | |
4432 /* Found additional edge leading to path from e1 to e2 | |
4433 from aside. */ | |
4434 return NULL; | |
4435 } | |
4436 else if (EDGE_COUNT (e->dest->preds) > 2) | |
4437 /* Several edges leading to path from e1 to e2 from aside. */ | |
4438 return NULL; | |
4439 | |
4440 if (e == e2) | |
4441 return (block_valid_for_bookkeeping_p (candidate_block) | |
4442 ? candidate_block | |
4443 : NULL); | |
4444 } | |
4445 gcc_unreachable (); | |
4446 } | |
4447 | |
4448 /* Create new basic block for bookkeeping code for path(s) incoming into | |
4449 E2->dest, except from E1->src. Return created block. */ | |
4450 static basic_block | |
4451 create_block_for_bookkeeping (edge e1, edge e2) | |
4452 { | |
4453 basic_block new_bb, bb = e2->dest; | |
4454 | |
4455 /* Check that we don't spoil the loop structure. */ | |
4456 if (current_loop_nest) | |
4457 { | |
4458 basic_block latch = current_loop_nest->latch; | |
4459 | |
4460 /* We do not split header. */ | |
4461 gcc_assert (e2->dest != current_loop_nest->header); | |
4462 | |
4463 /* We do not redirect the only edge to the latch block. */ | |
4464 gcc_assert (e1->dest != latch | |
4465 || !single_pred_p (latch) | |
4466 || e1 != single_pred_edge (latch)); | |
4467 } | |
4468 | |
4469 /* Split BB to insert BOOK_INSN there. */ | |
4470 new_bb = sched_split_block (bb, NULL); | |
4471 | |
4472 /* Move note_list from the upper bb. */ | |
4473 gcc_assert (BB_NOTE_LIST (new_bb) == NULL_RTX); | |
4474 BB_NOTE_LIST (new_bb) = BB_NOTE_LIST (bb); | |
4475 BB_NOTE_LIST (bb) = NULL_RTX; | |
4476 | |
4477 gcc_assert (e2->dest == bb); | |
4478 | |
4479 /* Skip block for bookkeeping copy when leaving E1->src. */ | |
4480 if (e1->flags & EDGE_FALLTHRU) | |
4481 sel_redirect_edge_and_branch_force (e1, new_bb); | |
4482 else | |
4483 sel_redirect_edge_and_branch (e1, new_bb); | |
4484 | |
4485 gcc_assert (e1->dest == new_bb); | |
4486 gcc_assert (sel_bb_empty_p (bb)); | |
4487 | |
4488 return bb; | |
4489 } | |
4490 | |
4491 /* Return insn after which we must insert bookkeeping code for path(s) incoming | |
4492 into E2->dest, except from E1->src. */ | |
4493 static insn_t | |
4494 find_place_for_bookkeeping (edge e1, edge e2) | |
4495 { | |
4496 insn_t place_to_insert; | |
4497 /* Find a basic block that can hold bookkeeping. If it can be found, do not | |
4498 create new basic block, but insert bookkeeping there. */ | |
4499 basic_block book_block = find_block_for_bookkeeping (e1, e2); | |
4500 | |
4501 if (!book_block) | |
4502 book_block = create_block_for_bookkeeping (e1, e2); | |
4503 | |
4504 place_to_insert = BB_END (book_block); | |
4505 | |
4506 /* If basic block ends with a jump, insert bookkeeping code right before it. */ | |
4507 if (INSN_P (place_to_insert) && control_flow_insn_p (place_to_insert)) | |
4508 place_to_insert = PREV_INSN (place_to_insert); | |
4509 | |
4510 return place_to_insert; | |
4511 } | |
4512 | |
4513 /* Find a proper seqno for bookkeeing insn inserted at PLACE_TO_INSERT | |
4514 for JOIN_POINT. */ | |
4515 static int | |
4516 find_seqno_for_bookkeeping (insn_t place_to_insert, insn_t join_point) | |
4517 { | |
4518 int seqno; | |
4519 rtx next; | |
4520 | |
4521 /* Check if we are about to insert bookkeeping copy before a jump, and use | |
4522 jump's seqno for the copy; otherwise, use JOIN_POINT's seqno. */ | |
4523 next = NEXT_INSN (place_to_insert); | |
4524 if (INSN_P (next) | |
4525 && JUMP_P (next) | |
4526 && BLOCK_FOR_INSN (next) == BLOCK_FOR_INSN (place_to_insert)) | |
4527 seqno = INSN_SEQNO (next); | |
4528 else if (INSN_SEQNO (join_point) > 0) | |
4529 seqno = INSN_SEQNO (join_point); | |
4530 else | |
4531 seqno = get_seqno_by_preds (place_to_insert); | |
4532 | |
4533 gcc_assert (seqno > 0); | |
4534 return seqno; | |
4535 } | |
4536 | |
4537 /* Insert bookkeeping copy of C_EXPS's insn after PLACE_TO_INSERT, assigning | |
4538 NEW_SEQNO to it. Return created insn. */ | |
4539 static insn_t | |
4540 emit_bookkeeping_insn (insn_t place_to_insert, expr_t c_expr, int new_seqno) | |
4541 { | |
4542 rtx new_insn_rtx = create_copy_of_insn_rtx (EXPR_INSN_RTX (c_expr)); | |
4543 | |
4544 vinsn_t new_vinsn | |
4545 = create_vinsn_from_insn_rtx (new_insn_rtx, | |
4546 VINSN_UNIQUE_P (EXPR_VINSN (c_expr))); | |
4547 | |
4548 insn_t new_insn = emit_insn_from_expr_after (c_expr, new_vinsn, new_seqno, | |
4549 place_to_insert); | |
4550 | |
4551 INSN_SCHED_TIMES (new_insn) = 0; | |
4552 bitmap_set_bit (current_copies, INSN_UID (new_insn)); | |
4553 | |
4554 return new_insn; | |
4555 } | |
4556 | |
4557 /* Generate a bookkeeping copy of C_EXPR's insn for path(s) incoming into to | |
4558 E2->dest, except from E1->src (there may be a sequence of empty blocks | |
4559 between E1->src and E2->dest). Return block containing the copy. | |
4560 All scheduler data is initialized for the newly created insn. */ | |
4561 static basic_block | |
4562 generate_bookkeeping_insn (expr_t c_expr, edge e1, edge e2) | |
4563 { | |
4564 insn_t join_point, place_to_insert, new_insn; | |
4565 int new_seqno; | |
4566 bool need_to_exchange_data_sets; | |
4567 | |
4568 if (sched_verbose >= 4) | |
4569 sel_print ("Generating bookkeeping insn (%d->%d)\n", e1->src->index, | |
4570 e2->dest->index); | |
4571 | |
4572 join_point = sel_bb_head (e2->dest); | |
4573 place_to_insert = find_place_for_bookkeeping (e1, e2); | |
4574 new_seqno = find_seqno_for_bookkeeping (place_to_insert, join_point); | |
4575 need_to_exchange_data_sets | |
4576 = sel_bb_empty_p (BLOCK_FOR_INSN (place_to_insert)); | |
4577 | |
4578 new_insn = emit_bookkeeping_insn (place_to_insert, c_expr, new_seqno); | |
4579 | |
4580 /* When inserting bookkeeping insn in new block, av sets should be | |
4581 following: old basic block (that now holds bookkeeping) data sets are | |
4582 the same as was before generation of bookkeeping, and new basic block | |
4583 (that now hold all other insns of old basic block) data sets are | |
4584 invalid. So exchange data sets for these basic blocks as sel_split_block | |
4585 mistakenly exchanges them in this case. Cannot do it earlier because | |
4586 when single instruction is added to new basic block it should hold NULL | |
4587 lv_set. */ | |
4588 if (need_to_exchange_data_sets) | |
4589 exchange_data_sets (BLOCK_FOR_INSN (new_insn), | |
4590 BLOCK_FOR_INSN (join_point)); | |
4591 | |
4592 stat_bookkeeping_copies++; | |
4593 return BLOCK_FOR_INSN (new_insn); | |
4594 } | |
4595 | |
4596 /* Remove from AV_PTR all insns that may need bookkeeping when scheduling | |
4597 on FENCE, but we are unable to copy them. */ | |
4598 static void | |
4599 remove_insns_that_need_bookkeeping (fence_t fence, av_set_t *av_ptr) | |
4600 { | |
4601 expr_t expr; | |
4602 av_set_iterator i; | |
4603 | |
4604 /* An expression does not need bookkeeping if it is available on all paths | |
4605 from current block to original block and current block dominates | |
4606 original block. We check availability on all paths by examining | |
4607 EXPR_SPEC; this is not equivalent, because it may be positive even | |
4608 if expr is available on all paths (but if expr is not available on | |
4609 any path, EXPR_SPEC will be positive). */ | |
4610 | |
4611 FOR_EACH_EXPR_1 (expr, i, av_ptr) | |
4612 { | |
4613 if (!control_flow_insn_p (EXPR_INSN_RTX (expr)) | |
4614 && (!bookkeeping_p || VINSN_UNIQUE_P (EXPR_VINSN (expr))) | |
4615 && (EXPR_SPEC (expr) | |
4616 || !EXPR_ORIG_BB_INDEX (expr) | |
4617 || !dominated_by_p (CDI_DOMINATORS, | |
4618 BASIC_BLOCK (EXPR_ORIG_BB_INDEX (expr)), | |
4619 BLOCK_FOR_INSN (FENCE_INSN (fence))))) | |
4620 { | |
4621 if (sched_verbose >= 4) | |
4622 sel_print ("Expr %d removed because it would need bookkeeping, which " | |
4623 "cannot be created\n", INSN_UID (EXPR_INSN_RTX (expr))); | |
4624 av_set_iter_remove (&i); | |
4625 } | |
4626 } | |
4627 } | |
4628 | |
4629 /* Moving conditional jump through some instructions. | |
4630 | |
4631 Consider example: | |
4632 | |
4633 ... <- current scheduling point | |
4634 NOTE BASIC BLOCK: <- bb header | |
4635 (p8) add r14=r14+0x9;; | |
4636 (p8) mov [r14]=r23 | |
4637 (!p8) jump L1;; | |
4638 NOTE BASIC BLOCK: | |
4639 ... | |
4640 | |
4641 We can schedule jump one cycle earlier, than mov, because they cannot be | |
4642 executed together as their predicates are mutually exclusive. | |
4643 | |
4644 This is done in this way: first, new fallthrough basic block is created | |
4645 after jump (it is always can be done, because there already should be a | |
4646 fallthrough block, where control flow goes in case of predicate being true - | |
4647 in our example; otherwise there should be a dependence between those | |
4648 instructions and jump and we cannot schedule jump right now); | |
4649 next, all instructions between jump and current scheduling point are moved | |
4650 to this new block. And the result is this: | |
4651 | |
4652 NOTE BASIC BLOCK: | |
4653 (!p8) jump L1 <- current scheduling point | |
4654 NOTE BASIC BLOCK: <- bb header | |
4655 (p8) add r14=r14+0x9;; | |
4656 (p8) mov [r14]=r23 | |
4657 NOTE BASIC BLOCK: | |
4658 ... | |
4659 */ | |
4660 static void | |
4661 move_cond_jump (rtx insn, bnd_t bnd) | |
4662 { | |
4663 edge ft_edge; | |
4664 basic_block block_from, block_next, block_new; | |
4665 rtx next, prev, link; | |
4666 | |
4667 /* BLOCK_FROM holds basic block of the jump. */ | |
4668 block_from = BLOCK_FOR_INSN (insn); | |
4669 | |
4670 /* Moving of jump should not cross any other jumps or | |
4671 beginnings of new basic blocks. */ | |
4672 gcc_assert (block_from == BLOCK_FOR_INSN (BND_TO (bnd))); | |
4673 | |
4674 /* Jump is moved to the boundary. */ | |
4675 prev = BND_TO (bnd); | |
4676 next = PREV_INSN (insn); | |
4677 BND_TO (bnd) = insn; | |
4678 | |
4679 ft_edge = find_fallthru_edge (block_from); | |
4680 block_next = ft_edge->dest; | |
4681 /* There must be a fallthrough block (or where should go | |
4682 control flow in case of false jump predicate otherwise?). */ | |
4683 gcc_assert (block_next); | |
4684 | |
4685 /* Create new empty basic block after source block. */ | |
4686 block_new = sel_split_edge (ft_edge); | |
4687 gcc_assert (block_new->next_bb == block_next | |
4688 && block_from->next_bb == block_new); | |
4689 | |
4690 gcc_assert (BB_END (block_from) == insn); | |
4691 | |
4692 /* Move all instructions except INSN from BLOCK_FROM to | |
4693 BLOCK_NEW. */ | |
4694 for (link = prev; link != insn; link = NEXT_INSN (link)) | |
4695 { | |
4696 EXPR_ORIG_BB_INDEX (INSN_EXPR (link)) = block_new->index; | |
4697 df_insn_change_bb (link, block_new); | |
4698 } | |
4699 | |
4700 /* Set correct basic block and instructions properties. */ | |
4701 BB_END (block_new) = PREV_INSN (insn); | |
4702 | |
4703 NEXT_INSN (PREV_INSN (prev)) = insn; | |
4704 PREV_INSN (insn) = PREV_INSN (prev); | |
4705 | |
4706 /* Assert there is no jump to BLOCK_NEW, only fallthrough edge. */ | |
4707 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (BB_HEAD (block_new))); | |
4708 PREV_INSN (prev) = BB_HEAD (block_new); | |
4709 NEXT_INSN (next) = NEXT_INSN (BB_HEAD (block_new)); | |
4710 NEXT_INSN (BB_HEAD (block_new)) = prev; | |
4711 PREV_INSN (NEXT_INSN (next)) = next; | |
4712 | |
4713 gcc_assert (!sel_bb_empty_p (block_from) | |
4714 && !sel_bb_empty_p (block_new)); | |
4715 | |
4716 /* Update data sets for BLOCK_NEW to represent that INSN and | |
4717 instructions from the other branch of INSN is no longer | |
4718 available at BLOCK_NEW. */ | |
4719 BB_AV_LEVEL (block_new) = global_level; | |
4720 gcc_assert (BB_LV_SET (block_new) == NULL); | |
4721 BB_LV_SET (block_new) = get_clear_regset_from_pool (); | |
4722 update_data_sets (sel_bb_head (block_new)); | |
4723 | |
4724 /* INSN is a new basic block header - so prepare its data | |
4725 structures and update availability and liveness sets. */ | |
4726 update_data_sets (insn); | |
4727 | |
4728 if (sched_verbose >= 4) | |
4729 sel_print ("Moving jump %d\n", INSN_UID (insn)); | |
4730 } | |
4731 | |
4732 /* Remove nops generated during move_op for preventing removal of empty | |
4733 basic blocks. */ | |
4734 static void | |
4735 remove_temp_moveop_nops (void) | |
4736 { | |
4737 int i; | |
4738 insn_t insn; | |
4739 | |
4740 for (i = 0; VEC_iterate (insn_t, vec_temp_moveop_nops, i, insn); i++) | |
4741 { | |
4742 gcc_assert (INSN_NOP_P (insn)); | |
4743 return_nop_to_pool (insn); | |
4744 } | |
4745 | |
4746 /* Empty the vector. */ | |
4747 if (VEC_length (insn_t, vec_temp_moveop_nops) > 0) | |
4748 VEC_block_remove (insn_t, vec_temp_moveop_nops, 0, | |
4749 VEC_length (insn_t, vec_temp_moveop_nops)); | |
4750 } | |
4751 | |
4752 /* Records the maximal UID before moving up an instruction. Used for | |
4753 distinguishing between bookkeeping copies and original insns. */ | |
4754 static int max_uid_before_move_op = 0; | |
4755 | |
4756 /* Remove from AV_VLIW_P all instructions but next when debug counter | |
4757 tells us so. Next instruction is fetched from BNDS. */ | |
4758 static void | |
4759 remove_insns_for_debug (blist_t bnds, av_set_t *av_vliw_p) | |
4760 { | |
4761 if (! dbg_cnt (sel_sched_insn_cnt)) | |
4762 /* Leave only the next insn in av_vliw. */ | |
4763 { | |
4764 av_set_iterator av_it; | |
4765 expr_t expr; | |
4766 bnd_t bnd = BLIST_BND (bnds); | |
4767 insn_t next = BND_TO (bnd); | |
4768 | |
4769 gcc_assert (BLIST_NEXT (bnds) == NULL); | |
4770 | |
4771 FOR_EACH_EXPR_1 (expr, av_it, av_vliw_p) | |
4772 if (EXPR_INSN_RTX (expr) != next) | |
4773 av_set_iter_remove (&av_it); | |
4774 } | |
4775 } | |
4776 | |
4777 /* Compute available instructions on BNDS. FENCE is the current fence. Write | |
4778 the computed set to *AV_VLIW_P. */ | |
4779 static void | |
4780 compute_av_set_on_boundaries (fence_t fence, blist_t bnds, av_set_t *av_vliw_p) | |
4781 { | |
4782 if (sched_verbose >= 2) | |
4783 { | |
4784 sel_print ("Boundaries: "); | |
4785 dump_blist (bnds); | |
4786 sel_print ("\n"); | |
4787 } | |
4788 | |
4789 for (; bnds; bnds = BLIST_NEXT (bnds)) | |
4790 { | |
4791 bnd_t bnd = BLIST_BND (bnds); | |
4792 av_set_t av1_copy; | |
4793 insn_t bnd_to = BND_TO (bnd); | |
4794 | |
4795 /* Rewind BND->TO to the basic block header in case some bookkeeping | |
4796 instructions were inserted before BND->TO and it needs to be | |
4797 adjusted. */ | |
4798 if (sel_bb_head_p (bnd_to)) | |
4799 gcc_assert (INSN_SCHED_TIMES (bnd_to) == 0); | |
4800 else | |
4801 while (INSN_SCHED_TIMES (PREV_INSN (bnd_to)) == 0) | |
4802 { | |
4803 bnd_to = PREV_INSN (bnd_to); | |
4804 if (sel_bb_head_p (bnd_to)) | |
4805 break; | |
4806 } | |
4807 | |
4808 if (BND_TO (bnd) != bnd_to) | |
4809 { | |
4810 gcc_assert (FENCE_INSN (fence) == BND_TO (bnd)); | |
4811 FENCE_INSN (fence) = bnd_to; | |
4812 BND_TO (bnd) = bnd_to; | |
4813 } | |
4814 | |
4815 av_set_clear (&BND_AV (bnd)); | |
4816 BND_AV (bnd) = compute_av_set (BND_TO (bnd), NULL, 0, true); | |
4817 | |
4818 av_set_clear (&BND_AV1 (bnd)); | |
4819 BND_AV1 (bnd) = av_set_copy (BND_AV (bnd)); | |
4820 | |
4821 moveup_set_inside_insn_group (&BND_AV1 (bnd), NULL); | |
4822 | |
4823 av1_copy = av_set_copy (BND_AV1 (bnd)); | |
4824 av_set_union_and_clear (av_vliw_p, &av1_copy, NULL); | |
4825 } | |
4826 | |
4827 if (sched_verbose >= 2) | |
4828 { | |
4829 sel_print ("Available exprs (vliw form): "); | |
4830 dump_av_set (*av_vliw_p); | |
4831 sel_print ("\n"); | |
4832 } | |
4833 } | |
4834 | |
4835 /* Calculate the sequential av set on BND corresponding to the EXPR_VLIW | |
4836 expression. When FOR_MOVEOP is true, also replace the register of | |
4837 expressions found with the register from EXPR_VLIW. */ | |
4838 static av_set_t | |
4839 find_sequential_best_exprs (bnd_t bnd, expr_t expr_vliw, bool for_moveop) | |
4840 { | |
4841 av_set_t expr_seq = NULL; | |
4842 expr_t expr; | |
4843 av_set_iterator i; | |
4844 | |
4845 FOR_EACH_EXPR (expr, i, BND_AV (bnd)) | |
4846 { | |
4847 if (equal_after_moveup_path_p (expr, NULL, expr_vliw)) | |
4848 { | |
4849 if (for_moveop) | |
4850 { | |
4851 /* The sequential expression has the right form to pass | |
4852 to move_op except when renaming happened. Put the | |
4853 correct register in EXPR then. */ | |
4854 if (EXPR_SEPARABLE_P (expr) && REG_P (EXPR_LHS (expr))) | |
4855 { | |
4856 if (expr_dest_regno (expr) != expr_dest_regno (expr_vliw)) | |
4857 { | |
4858 replace_dest_with_reg_in_expr (expr, EXPR_LHS (expr_vliw)); | |
4859 stat_renamed_scheduled++; | |
4860 } | |
4861 /* Also put the correct TARGET_AVAILABLE bit on the expr. | |
4862 This is needed when renaming came up with original | |
4863 register. */ | |
4864 else if (EXPR_TARGET_AVAILABLE (expr) | |
4865 != EXPR_TARGET_AVAILABLE (expr_vliw)) | |
4866 { | |
4867 gcc_assert (EXPR_TARGET_AVAILABLE (expr_vliw) == 1); | |
4868 EXPR_TARGET_AVAILABLE (expr) = 1; | |
4869 } | |
4870 } | |
4871 if (EXPR_WAS_SUBSTITUTED (expr)) | |
4872 stat_substitutions_total++; | |
4873 } | |
4874 | |
4875 av_set_add (&expr_seq, expr); | |
4876 | |
4877 /* With substitution inside insn group, it is possible | |
4878 that more than one expression in expr_seq will correspond | |
4879 to expr_vliw. In this case, choose one as the attempt to | |
4880 move both leads to miscompiles. */ | |
4881 break; | |
4882 } | |
4883 } | |
4884 | |
4885 if (for_moveop && sched_verbose >= 2) | |
4886 { | |
4887 sel_print ("Best expression(s) (sequential form): "); | |
4888 dump_av_set (expr_seq); | |
4889 sel_print ("\n"); | |
4890 } | |
4891 | |
4892 return expr_seq; | |
4893 } | |
4894 | |
4895 | |
4896 /* Move nop to previous block. */ | |
4897 static void ATTRIBUTE_UNUSED | |
4898 move_nop_to_previous_block (insn_t nop, basic_block prev_bb) | |
4899 { | |
4900 insn_t prev_insn, next_insn, note; | |
4901 | |
4902 gcc_assert (sel_bb_head_p (nop) | |
4903 && prev_bb == BLOCK_FOR_INSN (nop)->prev_bb); | |
4904 note = bb_note (BLOCK_FOR_INSN (nop)); | |
4905 prev_insn = sel_bb_end (prev_bb); | |
4906 next_insn = NEXT_INSN (nop); | |
4907 gcc_assert (prev_insn != NULL_RTX | |
4908 && PREV_INSN (note) == prev_insn); | |
4909 | |
4910 NEXT_INSN (prev_insn) = nop; | |
4911 PREV_INSN (nop) = prev_insn; | |
4912 | |
4913 PREV_INSN (note) = nop; | |
4914 NEXT_INSN (note) = next_insn; | |
4915 | |
4916 NEXT_INSN (nop) = note; | |
4917 PREV_INSN (next_insn) = note; | |
4918 | |
4919 BB_END (prev_bb) = nop; | |
4920 BLOCK_FOR_INSN (nop) = prev_bb; | |
4921 } | |
4922 | |
4923 /* Prepare a place to insert the chosen expression on BND. */ | |
4924 static insn_t | |
4925 prepare_place_to_insert (bnd_t bnd) | |
4926 { | |
4927 insn_t place_to_insert; | |
4928 | |
4929 /* Init place_to_insert before calling move_op, as the later | |
4930 can possibly remove BND_TO (bnd). */ | |
4931 if (/* If this is not the first insn scheduled. */ | |
4932 BND_PTR (bnd)) | |
4933 { | |
4934 /* Add it after last scheduled. */ | |
4935 place_to_insert = ILIST_INSN (BND_PTR (bnd)); | |
4936 } | |
4937 else | |
4938 { | |
4939 /* Add it before BND_TO. The difference is in the | |
4940 basic block, where INSN will be added. */ | |
4941 place_to_insert = get_nop_from_pool (BND_TO (bnd)); | |
4942 gcc_assert (BLOCK_FOR_INSN (place_to_insert) | |
4943 == BLOCK_FOR_INSN (BND_TO (bnd))); | |
4944 } | |
4945 | |
4946 return place_to_insert; | |
4947 } | |
4948 | |
4949 /* Find original instructions for EXPR_SEQ and move it to BND boundary. | |
4950 Return the expression to emit in C_EXPR. */ | |
4951 static bool | |
4952 move_exprs_to_boundary (bnd_t bnd, expr_t expr_vliw, | |
4953 av_set_t expr_seq, expr_t c_expr) | |
4954 { | |
4955 bool b, should_move; | |
4956 unsigned book_uid; | |
4957 bitmap_iterator bi; | |
4958 int n_bookkeeping_copies_before_moveop; | |
4959 | |
4960 /* Make a move. This call will remove the original operation, | |
4961 insert all necessary bookkeeping instructions and update the | |
4962 data sets. After that all we have to do is add the operation | |
4963 at before BND_TO (BND). */ | |
4964 n_bookkeeping_copies_before_moveop = stat_bookkeeping_copies; | |
4965 max_uid_before_move_op = get_max_uid (); | |
4966 bitmap_clear (current_copies); | |
4967 bitmap_clear (current_originators); | |
4968 | |
4969 b = move_op (BND_TO (bnd), expr_seq, expr_vliw, | |
4970 get_dest_from_orig_ops (expr_seq), c_expr, &should_move); | |
4971 | |
4972 /* We should be able to find the expression we've chosen for | |
4973 scheduling. */ | |
4974 gcc_assert (b); | |
4975 | |
4976 if (stat_bookkeeping_copies > n_bookkeeping_copies_before_moveop) | |
4977 stat_insns_needed_bookkeeping++; | |
4978 | |
4979 EXECUTE_IF_SET_IN_BITMAP (current_copies, 0, book_uid, bi) | |
4980 { | |
4981 /* We allocate these bitmaps lazily. */ | |
4982 if (! INSN_ORIGINATORS_BY_UID (book_uid)) | |
4983 INSN_ORIGINATORS_BY_UID (book_uid) = BITMAP_ALLOC (NULL); | |
4984 | |
4985 bitmap_copy (INSN_ORIGINATORS_BY_UID (book_uid), | |
4986 current_originators); | |
4987 } | |
4988 | |
4989 return should_move; | |
4990 } | |
4991 | |
4992 | |
4993 /* Debug a DFA state as an array of bytes. */ | |
4994 static void | |
4995 debug_state (state_t state) | |
4996 { | |
4997 unsigned char *p; | |
4998 unsigned int i, size = dfa_state_size; | |
4999 | |
5000 sel_print ("state (%u):", size); | |
5001 for (i = 0, p = (unsigned char *) state; i < size; i++) | |
5002 sel_print (" %d", p[i]); | |
5003 sel_print ("\n"); | |
5004 } | |
5005 | |
5006 /* Advance state on FENCE with INSN. Return true if INSN is | |
5007 an ASM, and we should advance state once more. */ | |
5008 static bool | |
5009 advance_state_on_fence (fence_t fence, insn_t insn) | |
5010 { | |
5011 bool asm_p; | |
5012 | |
5013 if (recog_memoized (insn) >= 0) | |
5014 { | |
5015 int res; | |
5016 state_t temp_state = alloca (dfa_state_size); | |
5017 | |
5018 gcc_assert (!INSN_ASM_P (insn)); | |
5019 asm_p = false; | |
5020 | |
5021 memcpy (temp_state, FENCE_STATE (fence), dfa_state_size); | |
5022 res = state_transition (FENCE_STATE (fence), insn); | |
5023 gcc_assert (res < 0); | |
5024 | |
5025 if (memcmp (temp_state, FENCE_STATE (fence), dfa_state_size)) | |
5026 { | |
5027 FENCE_ISSUED_INSNS (fence)++; | |
5028 | |
5029 /* We should never issue more than issue_rate insns. */ | |
5030 if (FENCE_ISSUED_INSNS (fence) > issue_rate) | |
5031 gcc_unreachable (); | |
5032 } | |
5033 } | |
5034 else | |
5035 { | |
5036 /* This could be an ASM insn which we'd like to schedule | |
5037 on the next cycle. */ | |
5038 asm_p = INSN_ASM_P (insn); | |
5039 if (!FENCE_STARTS_CYCLE_P (fence) && asm_p) | |
5040 advance_one_cycle (fence); | |
5041 } | |
5042 | |
5043 if (sched_verbose >= 2) | |
5044 debug_state (FENCE_STATE (fence)); | |
5045 FENCE_STARTS_CYCLE_P (fence) = 0; | |
5046 return asm_p; | |
5047 } | |
5048 | |
5049 /* Update FENCE on which INSN was scheduled and this INSN, too. NEED_STALL | |
5050 is nonzero if we need to stall after issuing INSN. */ | |
5051 static void | |
5052 update_fence_and_insn (fence_t fence, insn_t insn, int need_stall) | |
5053 { | |
5054 bool asm_p; | |
5055 | |
5056 /* First, reflect that something is scheduled on this fence. */ | |
5057 asm_p = advance_state_on_fence (fence, insn); | |
5058 FENCE_LAST_SCHEDULED_INSN (fence) = insn; | |
5059 VEC_safe_push (rtx, gc, FENCE_EXECUTING_INSNS (fence), insn); | |
5060 if (SCHED_GROUP_P (insn)) | |
5061 { | |
5062 FENCE_SCHED_NEXT (fence) = INSN_SCHED_NEXT (insn); | |
5063 SCHED_GROUP_P (insn) = 0; | |
5064 } | |
5065 else | |
5066 FENCE_SCHED_NEXT (fence) = NULL_RTX; | |
5067 if (INSN_UID (insn) < FENCE_READY_TICKS_SIZE (fence)) | |
5068 FENCE_READY_TICKS (fence) [INSN_UID (insn)] = 0; | |
5069 | |
5070 /* Set instruction scheduling info. This will be used in bundling, | |
5071 pipelining, tick computations etc. */ | |
5072 ++INSN_SCHED_TIMES (insn); | |
5073 EXPR_TARGET_AVAILABLE (INSN_EXPR (insn)) = true; | |
5074 EXPR_ORIG_SCHED_CYCLE (INSN_EXPR (insn)) = FENCE_CYCLE (fence); | |
5075 INSN_AFTER_STALL_P (insn) = FENCE_AFTER_STALL_P (fence); | |
5076 INSN_SCHED_CYCLE (insn) = FENCE_CYCLE (fence); | |
5077 | |
5078 /* This does not account for adjust_cost hooks, just add the biggest | |
5079 constant the hook may add to the latency. TODO: make this | |
5080 a target dependent constant. */ | |
5081 INSN_READY_CYCLE (insn) | |
5082 = INSN_SCHED_CYCLE (insn) + (INSN_CODE (insn) < 0 | |
5083 ? 1 | |
5084 : maximal_insn_latency (insn) + 1); | |
5085 | |
5086 /* Change these fields last, as they're used above. */ | |
5087 FENCE_AFTER_STALL_P (fence) = 0; | |
5088 if (asm_p || need_stall) | |
5089 advance_one_cycle (fence); | |
5090 | |
5091 /* Indicate that we've scheduled something on this fence. */ | |
5092 FENCE_SCHEDULED_P (fence) = true; | |
5093 scheduled_something_on_previous_fence = true; | |
5094 | |
5095 /* Print debug information when insn's fields are updated. */ | |
5096 if (sched_verbose >= 2) | |
5097 { | |
5098 sel_print ("Scheduling insn: "); | |
5099 dump_insn_1 (insn, 1); | |
5100 sel_print ("\n"); | |
5101 } | |
5102 } | |
5103 | |
5104 /* Update boundary BND with INSN, remove the old boundary from | |
5105 BNDSP, add new boundaries to BNDS_TAIL_P and return it. */ | |
5106 static blist_t * | |
5107 update_boundaries (bnd_t bnd, insn_t insn, blist_t *bndsp, | |
5108 blist_t *bnds_tailp) | |
5109 { | |
5110 succ_iterator si; | |
5111 insn_t succ; | |
5112 | |
5113 advance_deps_context (BND_DC (bnd), insn); | |
5114 FOR_EACH_SUCC_1 (succ, si, insn, | |
5115 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
5116 { | |
5117 ilist_t ptr = ilist_copy (BND_PTR (bnd)); | |
5118 | |
5119 ilist_add (&ptr, insn); | |
5120 blist_add (bnds_tailp, succ, ptr, BND_DC (bnd)); | |
5121 bnds_tailp = &BLIST_NEXT (*bnds_tailp); | |
5122 } | |
5123 | |
5124 blist_remove (bndsp); | |
5125 return bnds_tailp; | |
5126 } | |
5127 | |
5128 /* Schedule EXPR_VLIW on BND. Return the insn emitted. */ | |
5129 static insn_t | |
5130 schedule_expr_on_boundary (bnd_t bnd, expr_t expr_vliw, int seqno) | |
5131 { | |
5132 av_set_t expr_seq; | |
5133 expr_t c_expr = XALLOCA (expr_def); | |
5134 insn_t place_to_insert; | |
5135 insn_t insn; | |
5136 bool should_move; | |
5137 | |
5138 expr_seq = find_sequential_best_exprs (bnd, expr_vliw, true); | |
5139 | |
5140 /* In case of scheduling a jump skipping some other instructions, | |
5141 prepare CFG. After this, jump is at the boundary and can be | |
5142 scheduled as usual insn by MOVE_OP. */ | |
5143 if (vinsn_cond_branch_p (EXPR_VINSN (expr_vliw))) | |
5144 { | |
5145 insn = EXPR_INSN_RTX (expr_vliw); | |
5146 | |
5147 /* Speculative jumps are not handled. */ | |
5148 if (insn != BND_TO (bnd) | |
5149 && !sel_insn_is_speculation_check (insn)) | |
5150 move_cond_jump (insn, bnd); | |
5151 } | |
5152 | |
5153 /* Find a place for C_EXPR to schedule. */ | |
5154 place_to_insert = prepare_place_to_insert (bnd); | |
5155 should_move = move_exprs_to_boundary (bnd, expr_vliw, expr_seq, c_expr); | |
5156 clear_expr (c_expr); | |
5157 | |
5158 /* Add the instruction. The corner case to care about is when | |
5159 the expr_seq set has more than one expr, and we chose the one that | |
5160 is not equal to expr_vliw. Then expr_vliw may be insn in stream, and | |
5161 we can't use it. Generate the new vinsn. */ | |
5162 if (INSN_IN_STREAM_P (EXPR_INSN_RTX (expr_vliw))) | |
5163 { | |
5164 vinsn_t vinsn_new; | |
5165 | |
5166 vinsn_new = vinsn_copy (EXPR_VINSN (expr_vliw), false); | |
5167 change_vinsn_in_expr (expr_vliw, vinsn_new); | |
5168 should_move = false; | |
5169 } | |
5170 if (should_move) | |
5171 insn = sel_move_insn (expr_vliw, seqno, place_to_insert); | |
5172 else | |
5173 insn = emit_insn_from_expr_after (expr_vliw, NULL, seqno, | |
5174 place_to_insert); | |
5175 | |
5176 /* Return the nops generated for preserving of data sets back | |
5177 into pool. */ | |
5178 if (INSN_NOP_P (place_to_insert)) | |
5179 return_nop_to_pool (place_to_insert); | |
5180 remove_temp_moveop_nops (); | |
5181 | |
5182 av_set_clear (&expr_seq); | |
5183 | |
5184 /* Save the expression scheduled so to reset target availability if we'll | |
5185 meet it later on the same fence. */ | |
5186 if (EXPR_WAS_RENAMED (expr_vliw)) | |
5187 vinsn_vec_add (&vec_target_unavailable_vinsns, INSN_EXPR (insn)); | |
5188 | |
5189 /* Check that the recent movement didn't destroyed loop | |
5190 structure. */ | |
5191 gcc_assert (!pipelining_p | |
5192 || current_loop_nest == NULL | |
5193 || loop_latch_edge (current_loop_nest)); | |
5194 return insn; | |
5195 } | |
5196 | |
5197 /* Stall for N cycles on FENCE. */ | |
5198 static void | |
5199 stall_for_cycles (fence_t fence, int n) | |
5200 { | |
5201 int could_more; | |
5202 | |
5203 could_more = n > 1 || FENCE_ISSUED_INSNS (fence) < issue_rate; | |
5204 while (n--) | |
5205 advance_one_cycle (fence); | |
5206 if (could_more) | |
5207 FENCE_AFTER_STALL_P (fence) = 1; | |
5208 } | |
5209 | |
5210 /* Gather a parallel group of insns at FENCE and assign their seqno | |
5211 to SEQNO. All scheduled insns are gathered in SCHEDULED_INSNS_TAILPP | |
5212 list for later recalculation of seqnos. */ | |
5213 static void | |
5214 fill_insns (fence_t fence, int seqno, ilist_t **scheduled_insns_tailpp) | |
5215 { | |
5216 blist_t bnds = NULL, *bnds_tailp; | |
5217 av_set_t av_vliw = NULL; | |
5218 insn_t insn = FENCE_INSN (fence); | |
5219 | |
5220 if (sched_verbose >= 2) | |
5221 sel_print ("Starting fill_insns for insn %d, cycle %d\n", | |
5222 INSN_UID (insn), FENCE_CYCLE (fence)); | |
5223 | |
5224 blist_add (&bnds, insn, NULL, FENCE_DC (fence)); | |
5225 bnds_tailp = &BLIST_NEXT (bnds); | |
5226 set_target_context (FENCE_TC (fence)); | |
5227 target_bb = INSN_BB (insn); | |
5228 | |
5229 /* Do while we can add any operation to the current group. */ | |
5230 do | |
5231 { | |
5232 blist_t *bnds_tailp1, *bndsp; | |
5233 expr_t expr_vliw; | |
5234 int need_stall; | |
5235 int was_stall = 0, scheduled_insns = 0, stall_iterations = 0; | |
5236 int max_insns = pipelining_p ? issue_rate : 2 * issue_rate; | |
5237 int max_stall = pipelining_p ? 1 : 3; | |
5238 | |
5239 compute_av_set_on_boundaries (fence, bnds, &av_vliw); | |
5240 remove_insns_that_need_bookkeeping (fence, &av_vliw); | |
5241 remove_insns_for_debug (bnds, &av_vliw); | |
5242 | |
5243 /* Return early if we have nothing to schedule. */ | |
5244 if (av_vliw == NULL) | |
5245 break; | |
5246 | |
5247 /* Choose the best expression and, if needed, destination register | |
5248 for it. */ | |
5249 do | |
5250 { | |
5251 expr_vliw = find_best_expr (&av_vliw, bnds, fence, &need_stall); | |
5252 if (!expr_vliw && need_stall) | |
5253 { | |
5254 /* All expressions required a stall. Do not recompute av sets | |
5255 as we'll get the same answer (modulo the insns between | |
5256 the fence and its boundary, which will not be available for | |
5257 pipelining). */ | |
5258 gcc_assert (! expr_vliw && stall_iterations < 2); | |
5259 was_stall++; | |
5260 /* If we are going to stall for too long, break to recompute av | |
5261 sets and bring more insns for pipelining. */ | |
5262 if (need_stall <= 3) | |
5263 stall_for_cycles (fence, need_stall); | |
5264 else | |
5265 { | |
5266 stall_for_cycles (fence, 1); | |
5267 break; | |
5268 } | |
5269 } | |
5270 } | |
5271 while (! expr_vliw && need_stall); | |
5272 | |
5273 /* Now either we've selected expr_vliw or we have nothing to schedule. */ | |
5274 if (!expr_vliw) | |
5275 { | |
5276 av_set_clear (&av_vliw); | |
5277 break; | |
5278 } | |
5279 | |
5280 bndsp = &bnds; | |
5281 bnds_tailp1 = bnds_tailp; | |
5282 | |
5283 do | |
5284 /* This code will be executed only once until we'd have several | |
5285 boundaries per fence. */ | |
5286 { | |
5287 bnd_t bnd = BLIST_BND (*bndsp); | |
5288 | |
5289 if (!av_set_is_in_p (BND_AV1 (bnd), EXPR_VINSN (expr_vliw))) | |
5290 { | |
5291 bndsp = &BLIST_NEXT (*bndsp); | |
5292 continue; | |
5293 } | |
5294 | |
5295 insn = schedule_expr_on_boundary (bnd, expr_vliw, seqno); | |
5296 update_fence_and_insn (fence, insn, need_stall); | |
5297 bnds_tailp = update_boundaries (bnd, insn, bndsp, bnds_tailp); | |
5298 | |
5299 /* Add insn to the list of scheduled on this cycle instructions. */ | |
5300 ilist_add (*scheduled_insns_tailpp, insn); | |
5301 *scheduled_insns_tailpp = &ILIST_NEXT (**scheduled_insns_tailpp); | |
5302 } | |
5303 while (*bndsp != *bnds_tailp1); | |
5304 | |
5305 av_set_clear (&av_vliw); | |
5306 scheduled_insns++; | |
5307 | |
5308 /* We currently support information about candidate blocks only for | |
5309 one 'target_bb' block. Hence we can't schedule after jump insn, | |
5310 as this will bring two boundaries and, hence, necessity to handle | |
5311 information for two or more blocks concurrently. */ | |
5312 if (sel_bb_end_p (insn) | |
5313 || (was_stall | |
5314 && (was_stall >= max_stall | |
5315 || scheduled_insns >= max_insns))) | |
5316 break; | |
5317 } | |
5318 while (bnds); | |
5319 | |
5320 gcc_assert (!FENCE_BNDS (fence)); | |
5321 | |
5322 /* Update boundaries of the FENCE. */ | |
5323 while (bnds) | |
5324 { | |
5325 ilist_t ptr = BND_PTR (BLIST_BND (bnds)); | |
5326 | |
5327 if (ptr) | |
5328 { | |
5329 insn = ILIST_INSN (ptr); | |
5330 | |
5331 if (!ilist_is_in_p (FENCE_BNDS (fence), insn)) | |
5332 ilist_add (&FENCE_BNDS (fence), insn); | |
5333 } | |
5334 | |
5335 blist_remove (&bnds); | |
5336 } | |
5337 | |
5338 /* Update target context on the fence. */ | |
5339 reset_target_context (FENCE_TC (fence), false); | |
5340 } | |
5341 | |
5342 /* All exprs in ORIG_OPS must have the same destination register or memory. | |
5343 Return that destination. */ | |
5344 static rtx | |
5345 get_dest_from_orig_ops (av_set_t orig_ops) | |
5346 { | |
5347 rtx dest = NULL_RTX; | |
5348 av_set_iterator av_it; | |
5349 expr_t expr; | |
5350 bool first_p = true; | |
5351 | |
5352 FOR_EACH_EXPR (expr, av_it, orig_ops) | |
5353 { | |
5354 rtx x = EXPR_LHS (expr); | |
5355 | |
5356 if (first_p) | |
5357 { | |
5358 first_p = false; | |
5359 dest = x; | |
5360 } | |
5361 else | |
5362 gcc_assert (dest == x | |
5363 || (dest != NULL_RTX && x != NULL_RTX | |
5364 && rtx_equal_p (dest, x))); | |
5365 } | |
5366 | |
5367 return dest; | |
5368 } | |
5369 | |
5370 /* Update data sets for the bookkeeping block and record those expressions | |
5371 which become no longer available after inserting this bookkeeping. */ | |
5372 static void | |
5373 update_and_record_unavailable_insns (basic_block book_block) | |
5374 { | |
5375 av_set_iterator i; | |
5376 av_set_t old_av_set = NULL; | |
5377 expr_t cur_expr; | |
5378 rtx bb_end = sel_bb_end (book_block); | |
5379 | |
5380 /* First, get correct liveness in the bookkeeping block. The problem is | |
5381 the range between the bookeeping insn and the end of block. */ | |
5382 update_liveness_on_insn (bb_end); | |
5383 if (control_flow_insn_p (bb_end)) | |
5384 update_liveness_on_insn (PREV_INSN (bb_end)); | |
5385 | |
5386 /* If there's valid av_set on BOOK_BLOCK, then there might exist another | |
5387 fence above, where we may choose to schedule an insn which is | |
5388 actually blocked from moving up with the bookkeeping we create here. */ | |
5389 if (AV_SET_VALID_P (sel_bb_head (book_block))) | |
5390 { | |
5391 old_av_set = av_set_copy (BB_AV_SET (book_block)); | |
5392 update_data_sets (sel_bb_head (book_block)); | |
5393 | |
5394 /* Traverse all the expressions in the old av_set and check whether | |
5395 CUR_EXPR is in new AV_SET. */ | |
5396 FOR_EACH_EXPR (cur_expr, i, old_av_set) | |
5397 { | |
5398 expr_t new_expr = av_set_lookup (BB_AV_SET (book_block), | |
5399 EXPR_VINSN (cur_expr)); | |
5400 | |
5401 if (! new_expr | |
5402 /* In this case, we can just turn off the E_T_A bit, but we can't | |
5403 represent this information with the current vector. */ | |
5404 || EXPR_TARGET_AVAILABLE (new_expr) | |
5405 != EXPR_TARGET_AVAILABLE (cur_expr)) | |
5406 /* Unfortunately, the below code could be also fired up on | |
5407 separable insns. | |
5408 FIXME: add an example of how this could happen. */ | |
5409 vinsn_vec_add (&vec_bookkeeping_blocked_vinsns, cur_expr); | |
5410 } | |
5411 | |
5412 av_set_clear (&old_av_set); | |
5413 } | |
5414 } | |
5415 | |
5416 /* The main effect of this function is that sparams->c_expr is merged | |
5417 with (or copied to) lparams->c_expr_merged. If there's only one successor, | |
5418 we avoid merging anything by copying sparams->c_expr to lparams->c_expr_merged. | |
5419 lparams->c_expr_merged is copied back to sparams->c_expr after all | |
5420 successors has been traversed. lparams->c_expr_local is an expr allocated | |
5421 on stack in the caller function, and is used if there is more than one | |
5422 successor. | |
5423 | |
5424 SUCC is one of the SUCCS_NORMAL successors of INSN, | |
5425 MOVEOP_DRV_CALL_RES is the result of call code_motion_path_driver on succ, | |
5426 LPARAMS and STATIC_PARAMS contain the parameters described above. */ | |
5427 static void | |
5428 move_op_merge_succs (insn_t insn ATTRIBUTE_UNUSED, | |
5429 insn_t succ ATTRIBUTE_UNUSED, | |
5430 int moveop_drv_call_res, | |
5431 cmpd_local_params_p lparams, void *static_params) | |
5432 { | |
5433 moveop_static_params_p sparams = (moveop_static_params_p) static_params; | |
5434 | |
5435 /* Nothing to do, if original expr wasn't found below. */ | |
5436 if (moveop_drv_call_res != 1) | |
5437 return; | |
5438 | |
5439 /* If this is a first successor. */ | |
5440 if (!lparams->c_expr_merged) | |
5441 { | |
5442 lparams->c_expr_merged = sparams->c_expr; | |
5443 sparams->c_expr = lparams->c_expr_local; | |
5444 } | |
5445 else | |
5446 { | |
5447 /* We must merge all found expressions to get reasonable | |
5448 EXPR_SPEC_DONE_DS for the resulting insn. If we don't | |
5449 do so then we can first find the expr with epsilon | |
5450 speculation success probability and only then with the | |
5451 good probability. As a result the insn will get epsilon | |
5452 probability and will never be scheduled because of | |
5453 weakness_cutoff in find_best_expr. | |
5454 | |
5455 We call merge_expr_data here instead of merge_expr | |
5456 because due to speculation C_EXPR and X may have the | |
5457 same insns with different speculation types. And as of | |
5458 now such insns are considered non-equal. | |
5459 | |
5460 However, EXPR_SCHED_TIMES is different -- we must get | |
5461 SCHED_TIMES from a real insn, not a bookkeeping copy. | |
5462 We force this here. Instead, we may consider merging | |
5463 SCHED_TIMES to the maximum instead of minimum in the | |
5464 below function. */ | |
5465 int old_times = EXPR_SCHED_TIMES (lparams->c_expr_merged); | |
5466 | |
5467 merge_expr_data (lparams->c_expr_merged, sparams->c_expr, NULL); | |
5468 if (EXPR_SCHED_TIMES (sparams->c_expr) == 0) | |
5469 EXPR_SCHED_TIMES (lparams->c_expr_merged) = old_times; | |
5470 | |
5471 clear_expr (sparams->c_expr); | |
5472 } | |
5473 } | |
5474 | |
5475 /* Add used regs for the successor SUCC into SPARAMS->USED_REGS. | |
5476 | |
5477 SUCC is one of the SUCCS_NORMAL successors of INSN, | |
5478 MOVEOP_DRV_CALL_RES is the result of call code_motion_path_driver on succ or 0, | |
5479 if SUCC is one of SUCCS_BACK or SUCCS_OUT. | |
5480 STATIC_PARAMS contain USED_REGS set. */ | |
5481 static void | |
5482 fur_merge_succs (insn_t insn ATTRIBUTE_UNUSED, insn_t succ, | |
5483 int moveop_drv_call_res, | |
5484 cmpd_local_params_p lparams ATTRIBUTE_UNUSED, | |
5485 void *static_params) | |
5486 { | |
5487 regset succ_live; | |
5488 fur_static_params_p sparams = (fur_static_params_p) static_params; | |
5489 | |
5490 /* Here we compute live regsets only for branches that do not lie | |
5491 on the code motion paths. These branches correspond to value | |
5492 MOVEOP_DRV_CALL_RES==0 and include SUCCS_BACK and SUCCS_OUT, though | |
5493 for such branches code_motion_path_driver is not called. */ | |
5494 if (moveop_drv_call_res != 0) | |
5495 return; | |
5496 | |
5497 /* Mark all registers that do not meet the following condition: | |
5498 (3) not live on the other path of any conditional branch | |
5499 that is passed by the operation, in case original | |
5500 operations are not present on both paths of the | |
5501 conditional branch. */ | |
5502 succ_live = compute_live (succ); | |
5503 IOR_REG_SET (sparams->used_regs, succ_live); | |
5504 } | |
5505 | |
5506 /* This function is called after the last successor. Copies LP->C_EXPR_MERGED | |
5507 into SP->CEXPR. */ | |
5508 static void | |
5509 move_op_after_merge_succs (cmpd_local_params_p lp, void *sparams) | |
5510 { | |
5511 moveop_static_params_p sp = (moveop_static_params_p) sparams; | |
5512 | |
5513 sp->c_expr = lp->c_expr_merged; | |
5514 } | |
5515 | |
5516 /* Track bookkeeping copies created, insns scheduled, and blocks for | |
5517 rescheduling when INSN is found by move_op. */ | |
5518 static void | |
5519 track_scheduled_insns_and_blocks (rtx insn) | |
5520 { | |
5521 /* Even if this insn can be a copy that will be removed during current move_op, | |
5522 we still need to count it as an originator. */ | |
5523 bitmap_set_bit (current_originators, INSN_UID (insn)); | |
5524 | |
5525 if (!bitmap_bit_p (current_copies, INSN_UID (insn))) | |
5526 { | |
5527 /* Note that original block needs to be rescheduled, as we pulled an | |
5528 instruction out of it. */ | |
5529 if (INSN_SCHED_TIMES (insn) > 0) | |
5530 bitmap_set_bit (blocks_to_reschedule, BLOCK_FOR_INSN (insn)->index); | |
5531 else if (INSN_UID (insn) < first_emitted_uid) | |
5532 num_insns_scheduled++; | |
5533 } | |
5534 else | |
5535 bitmap_clear_bit (current_copies, INSN_UID (insn)); | |
5536 | |
5537 /* For instructions we must immediately remove insn from the | |
5538 stream, so subsequent update_data_sets () won't include this | |
5539 insn into av_set. | |
5540 For expr we must make insn look like "INSN_REG (insn) := c_expr". */ | |
5541 if (INSN_UID (insn) > max_uid_before_move_op) | |
5542 stat_bookkeeping_copies--; | |
5543 } | |
5544 | |
5545 /* Emit a register-register copy for INSN if needed. Return true if | |
5546 emitted one. PARAMS is the move_op static parameters. */ | |
5547 static bool | |
5548 maybe_emit_renaming_copy (rtx insn, | |
5549 moveop_static_params_p params) | |
5550 { | |
5551 bool insn_emitted = false; | |
5552 rtx cur_reg = expr_dest_reg (params->c_expr); | |
5553 | |
5554 gcc_assert (!cur_reg || (params->dest && REG_P (params->dest))); | |
5555 | |
5556 /* If original operation has expr and the register chosen for | |
5557 that expr is not original operation's dest reg, substitute | |
5558 operation's right hand side with the register chosen. */ | |
5559 if (cur_reg != NULL_RTX && REGNO (params->dest) != REGNO (cur_reg)) | |
5560 { | |
5561 insn_t reg_move_insn, reg_move_insn_rtx; | |
5562 | |
5563 reg_move_insn_rtx = create_insn_rtx_with_rhs (INSN_VINSN (insn), | |
5564 params->dest); | |
5565 reg_move_insn = sel_gen_insn_from_rtx_after (reg_move_insn_rtx, | |
5566 INSN_EXPR (insn), | |
5567 INSN_SEQNO (insn), | |
5568 insn); | |
5569 EXPR_SPEC_DONE_DS (INSN_EXPR (reg_move_insn)) = 0; | |
5570 replace_dest_with_reg_in_expr (params->c_expr, params->dest); | |
5571 | |
5572 insn_emitted = true; | |
5573 params->was_renamed = true; | |
5574 } | |
5575 | |
5576 return insn_emitted; | |
5577 } | |
5578 | |
5579 /* Emit a speculative check for INSN speculated as EXPR if needed. | |
5580 Return true if we've emitted one. PARAMS is the move_op static | |
5581 parameters. */ | |
5582 static bool | |
5583 maybe_emit_speculative_check (rtx insn, expr_t expr, | |
5584 moveop_static_params_p params) | |
5585 { | |
5586 bool insn_emitted = false; | |
5587 insn_t x; | |
5588 ds_t check_ds; | |
5589 | |
5590 check_ds = get_spec_check_type_for_insn (insn, expr); | |
5591 if (check_ds != 0) | |
5592 { | |
5593 /* A speculation check should be inserted. */ | |
5594 x = create_speculation_check (params->c_expr, check_ds, insn); | |
5595 insn_emitted = true; | |
5596 } | |
5597 else | |
5598 { | |
5599 EXPR_SPEC_DONE_DS (INSN_EXPR (insn)) = 0; | |
5600 x = insn; | |
5601 } | |
5602 | |
5603 gcc_assert (EXPR_SPEC_DONE_DS (INSN_EXPR (x)) == 0 | |
5604 && EXPR_SPEC_TO_CHECK_DS (INSN_EXPR (x)) == 0); | |
5605 return insn_emitted; | |
5606 } | |
5607 | |
5608 /* Handle transformations that leave an insn in place of original | |
5609 insn such as renaming/speculation. Return true if one of such | |
5610 transformations actually happened, and we have emitted this insn. */ | |
5611 static bool | |
5612 handle_emitting_transformations (rtx insn, expr_t expr, | |
5613 moveop_static_params_p params) | |
5614 { | |
5615 bool insn_emitted = false; | |
5616 | |
5617 insn_emitted = maybe_emit_renaming_copy (insn, params); | |
5618 insn_emitted |= maybe_emit_speculative_check (insn, expr, params); | |
5619 | |
5620 return insn_emitted; | |
5621 } | |
5622 | |
5623 /* Remove INSN from stream. When ONLY_DISCONNECT is true, its data | |
5624 is not removed but reused when INSN is re-emitted. */ | |
5625 static void | |
5626 remove_insn_from_stream (rtx insn, bool only_disconnect) | |
5627 { | |
5628 insn_t nop, bb_head, bb_end; | |
5629 bool need_nop_to_preserve_bb; | |
5630 basic_block bb = BLOCK_FOR_INSN (insn); | |
5631 | |
5632 /* If INSN is the only insn in the basic block (not counting JUMP, | |
5633 which may be a jump to next insn), leave NOP there till the | |
5634 return to fill_insns. */ | |
5635 bb_head = sel_bb_head (bb); | |
5636 bb_end = sel_bb_end (bb); | |
5637 need_nop_to_preserve_bb = ((bb_head == bb_end) | |
5638 || (NEXT_INSN (bb_head) == bb_end | |
5639 && JUMP_P (bb_end)) | |
5640 || IN_CURRENT_FENCE_P (NEXT_INSN (insn))); | |
5641 | |
5642 /* If there's only one insn in the BB, make sure that a nop is | |
5643 inserted into it, so the basic block won't disappear when we'll | |
5644 delete INSN below with sel_remove_insn. It should also survive | |
5645 till the return to fill_insns. */ | |
5646 if (need_nop_to_preserve_bb) | |
5647 { | |
5648 nop = get_nop_from_pool (insn); | |
5649 gcc_assert (INSN_NOP_P (nop)); | |
5650 VEC_safe_push (insn_t, heap, vec_temp_moveop_nops, nop); | |
5651 } | |
5652 | |
5653 sel_remove_insn (insn, only_disconnect, false); | |
5654 } | |
5655 | |
5656 /* This function is called when original expr is found. | |
5657 INSN - current insn traversed, EXPR - the corresponding expr found. | |
5658 LPARAMS is the local parameters of code modion driver, STATIC_PARAMS | |
5659 is static parameters of move_op. */ | |
5660 static void | |
5661 move_op_orig_expr_found (insn_t insn, expr_t expr, | |
5662 cmpd_local_params_p lparams ATTRIBUTE_UNUSED, | |
5663 void *static_params) | |
5664 { | |
5665 bool only_disconnect, insn_emitted; | |
5666 moveop_static_params_p params = (moveop_static_params_p) static_params; | |
5667 | |
5668 copy_expr_onside (params->c_expr, INSN_EXPR (insn)); | |
5669 track_scheduled_insns_and_blocks (insn); | |
5670 insn_emitted = handle_emitting_transformations (insn, expr, params); | |
5671 only_disconnect = (params->uid == INSN_UID (insn) | |
5672 && ! insn_emitted && ! EXPR_WAS_CHANGED (expr)); | |
5673 | |
5674 /* Mark that we've disconnected an insn. */ | |
5675 if (only_disconnect) | |
5676 params->uid = -1; | |
5677 remove_insn_from_stream (insn, only_disconnect); | |
5678 } | |
5679 | |
5680 /* The function is called when original expr is found. | |
5681 INSN - current insn traversed, EXPR - the corresponding expr found, | |
5682 crosses_call and original_insns in STATIC_PARAMS are updated. */ | |
5683 static void | |
5684 fur_orig_expr_found (insn_t insn, expr_t expr ATTRIBUTE_UNUSED, | |
5685 cmpd_local_params_p lparams ATTRIBUTE_UNUSED, | |
5686 void *static_params) | |
5687 { | |
5688 fur_static_params_p params = (fur_static_params_p) static_params; | |
5689 regset tmp; | |
5690 | |
5691 if (CALL_P (insn)) | |
5692 params->crosses_call = true; | |
5693 | |
5694 def_list_add (params->original_insns, insn, params->crosses_call); | |
5695 | |
5696 /* Mark the registers that do not meet the following condition: | |
5697 (2) not among the live registers of the point | |
5698 immediately following the first original operation on | |
5699 a given downward path, except for the original target | |
5700 register of the operation. */ | |
5701 tmp = get_clear_regset_from_pool (); | |
5702 compute_live_below_insn (insn, tmp); | |
5703 AND_COMPL_REG_SET (tmp, INSN_REG_SETS (insn)); | |
5704 AND_COMPL_REG_SET (tmp, INSN_REG_CLOBBERS (insn)); | |
5705 IOR_REG_SET (params->used_regs, tmp); | |
5706 return_regset_to_pool (tmp); | |
5707 | |
5708 /* (*1) We need to add to USED_REGS registers that are read by | |
5709 INSN's lhs. This may lead to choosing wrong src register. | |
5710 E.g. (scheduling const expr enabled): | |
5711 | |
5712 429: ax=0x0 <- Can't use AX for this expr (0x0) | |
5713 433: dx=[bp-0x18] | |
5714 427: [ax+dx+0x1]=ax | |
5715 REG_DEAD: ax | |
5716 168: di=dx | |
5717 REG_DEAD: dx | |
5718 */ | |
5719 /* FIXME: see comment above and enable MEM_P | |
5720 in vinsn_separable_p. */ | |
5721 gcc_assert (!VINSN_SEPARABLE_P (INSN_VINSN (insn)) | |
5722 || !MEM_P (INSN_LHS (insn))); | |
5723 } | |
5724 | |
5725 /* This function is called on the ascending pass, before returning from | |
5726 current basic block. */ | |
5727 static void | |
5728 move_op_at_first_insn (insn_t insn, cmpd_local_params_p lparams, | |
5729 void *static_params) | |
5730 { | |
5731 moveop_static_params_p sparams = (moveop_static_params_p) static_params; | |
5732 basic_block book_block = NULL; | |
5733 | |
5734 /* When we have removed the boundary insn for scheduling, which also | |
5735 happened to be the end insn in its bb, we don't need to update sets. */ | |
5736 if (!lparams->removed_last_insn | |
5737 && lparams->e1 | |
5738 && sel_bb_head_p (insn)) | |
5739 { | |
5740 /* We should generate bookkeeping code only if we are not at the | |
5741 top level of the move_op. */ | |
5742 if (sel_num_cfg_preds_gt_1 (insn)) | |
5743 book_block = generate_bookkeeping_insn (sparams->c_expr, | |
5744 lparams->e1, lparams->e2); | |
5745 /* Update data sets for the current insn. */ | |
5746 update_data_sets (insn); | |
5747 } | |
5748 | |
5749 /* If bookkeeping code was inserted, we need to update av sets of basic | |
5750 block that received bookkeeping. After generation of bookkeeping insn, | |
5751 bookkeeping block does not contain valid av set because we are not following | |
5752 the original algorithm in every detail with regards to e.g. renaming | |
5753 simple reg-reg copies. Consider example: | |
5754 | |
5755 bookkeeping block scheduling fence | |
5756 \ / | |
5757 \ join / | |
5758 ---------- | |
5759 | | | |
5760 ---------- | |
5761 / \ | |
5762 / \ | |
5763 r1 := r2 r1 := r3 | |
5764 | |
5765 We try to schedule insn "r1 := r3" on the current | |
5766 scheduling fence. Also, note that av set of bookkeeping block | |
5767 contain both insns "r1 := r2" and "r1 := r3". When the insn has | |
5768 been scheduled, the CFG is as follows: | |
5769 | |
5770 r1 := r3 r1 := r3 | |
5771 bookkeeping block scheduling fence | |
5772 \ / | |
5773 \ join / | |
5774 ---------- | |
5775 | | | |
5776 ---------- | |
5777 / \ | |
5778 / \ | |
5779 r1 := r2 | |
5780 | |
5781 Here, insn "r1 := r3" was scheduled at the current scheduling point | |
5782 and bookkeeping code was generated at the bookeeping block. This | |
5783 way insn "r1 := r2" is no longer available as a whole instruction | |
5784 (but only as expr) ahead of insn "r1 := r3" in bookkeeping block. | |
5785 This situation is handled by calling update_data_sets. | |
5786 | |
5787 Since update_data_sets is called only on the bookkeeping block, and | |
5788 it also may have predecessors with av_sets, containing instructions that | |
5789 are no longer available, we save all such expressions that become | |
5790 unavailable during data sets update on the bookkeeping block in | |
5791 VEC_BOOKKEEPING_BLOCKED_VINSNS. Later we avoid selecting such | |
5792 expressions for scheduling. This allows us to avoid recomputation of | |
5793 av_sets outside the code motion path. */ | |
5794 | |
5795 if (book_block) | |
5796 update_and_record_unavailable_insns (book_block); | |
5797 | |
5798 /* If INSN was previously marked for deletion, it's time to do it. */ | |
5799 if (lparams->removed_last_insn) | |
5800 insn = PREV_INSN (insn); | |
5801 | |
5802 /* Do not tidy control flow at the topmost moveop, as we can erroneously | |
5803 kill a block with a single nop in which the insn should be emitted. */ | |
5804 if (lparams->e1) | |
5805 tidy_control_flow (BLOCK_FOR_INSN (insn), true); | |
5806 } | |
5807 | |
5808 /* This function is called on the ascending pass, before returning from the | |
5809 current basic block. */ | |
5810 static void | |
5811 fur_at_first_insn (insn_t insn, | |
5812 cmpd_local_params_p lparams ATTRIBUTE_UNUSED, | |
5813 void *static_params ATTRIBUTE_UNUSED) | |
5814 { | |
5815 gcc_assert (!sel_bb_head_p (insn) || AV_SET_VALID_P (insn) | |
5816 || AV_LEVEL (insn) == -1); | |
5817 } | |
5818 | |
5819 /* Called on the backward stage of recursion to call moveup_expr for insn | |
5820 and sparams->c_expr. */ | |
5821 static void | |
5822 move_op_ascend (insn_t insn, void *static_params) | |
5823 { | |
5824 enum MOVEUP_EXPR_CODE res; | |
5825 moveop_static_params_p sparams = (moveop_static_params_p) static_params; | |
5826 | |
5827 if (! INSN_NOP_P (insn)) | |
5828 { | |
5829 res = moveup_expr_cached (sparams->c_expr, insn, false); | |
5830 gcc_assert (res != MOVEUP_EXPR_NULL); | |
5831 } | |
5832 | |
5833 /* Update liveness for this insn as it was invalidated. */ | |
5834 update_liveness_on_insn (insn); | |
5835 } | |
5836 | |
5837 /* This function is called on enter to the basic block. | |
5838 Returns TRUE if this block already have been visited and | |
5839 code_motion_path_driver should return 1, FALSE otherwise. */ | |
5840 static int | |
5841 fur_on_enter (insn_t insn ATTRIBUTE_UNUSED, cmpd_local_params_p local_params, | |
5842 void *static_params, bool visited_p) | |
5843 { | |
5844 fur_static_params_p sparams = (fur_static_params_p) static_params; | |
5845 | |
5846 if (visited_p) | |
5847 { | |
5848 /* If we have found something below this block, there should be at | |
5849 least one insn in ORIGINAL_INSNS. */ | |
5850 gcc_assert (*sparams->original_insns); | |
5851 | |
5852 /* Adjust CROSSES_CALL, since we may have come to this block along | |
5853 different path. */ | |
5854 DEF_LIST_DEF (*sparams->original_insns)->crosses_call | |
5855 |= sparams->crosses_call; | |
5856 } | |
5857 else | |
5858 local_params->old_original_insns = *sparams->original_insns; | |
5859 | |
5860 return 1; | |
5861 } | |
5862 | |
5863 /* Same as above but for move_op. */ | |
5864 static int | |
5865 move_op_on_enter (insn_t insn ATTRIBUTE_UNUSED, | |
5866 cmpd_local_params_p local_params ATTRIBUTE_UNUSED, | |
5867 void *static_params ATTRIBUTE_UNUSED, bool visited_p) | |
5868 { | |
5869 if (visited_p) | |
5870 return -1; | |
5871 return 1; | |
5872 } | |
5873 | |
5874 /* This function is called while descending current basic block if current | |
5875 insn is not the original EXPR we're searching for. | |
5876 | |
5877 Return value: FALSE, if code_motion_path_driver should perform a local | |
5878 cleanup and return 0 itself; | |
5879 TRUE, if code_motion_path_driver should continue. */ | |
5880 static bool | |
5881 move_op_orig_expr_not_found (insn_t insn, av_set_t orig_ops ATTRIBUTE_UNUSED, | |
5882 void *static_params) | |
5883 { | |
5884 moveop_static_params_p sparams = (moveop_static_params_p) static_params; | |
5885 | |
5886 #ifdef ENABLE_CHECKING | |
5887 sparams->failed_insn = insn; | |
5888 #endif | |
5889 | |
5890 /* If we're scheduling separate expr, in order to generate correct code | |
5891 we need to stop the search at bookkeeping code generated with the | |
5892 same destination register or memory. */ | |
5893 if (lhs_of_insn_equals_to_dest_p (insn, sparams->dest)) | |
5894 return false; | |
5895 return true; | |
5896 } | |
5897 | |
5898 /* This function is called while descending current basic block if current | |
5899 insn is not the original EXPR we're searching for. | |
5900 | |
5901 Return value: TRUE (code_motion_path_driver should continue). */ | |
5902 static bool | |
5903 fur_orig_expr_not_found (insn_t insn, av_set_t orig_ops, void *static_params) | |
5904 { | |
5905 bool mutexed; | |
5906 expr_t r; | |
5907 av_set_iterator avi; | |
5908 fur_static_params_p sparams = (fur_static_params_p) static_params; | |
5909 | |
5910 if (CALL_P (insn)) | |
5911 sparams->crosses_call = true; | |
5912 | |
5913 /* If current insn we are looking at cannot be executed together | |
5914 with original insn, then we can skip it safely. | |
5915 | |
5916 Example: ORIG_OPS = { (p6) r14 = sign_extend (r15); } | |
5917 INSN = (!p6) r14 = r14 + 1; | |
5918 | |
5919 Here we can schedule ORIG_OP with lhs = r14, though only | |
5920 looking at the set of used and set registers of INSN we must | |
5921 forbid it. So, add set/used in INSN registers to the | |
5922 untouchable set only if there is an insn in ORIG_OPS that can | |
5923 affect INSN. */ | |
5924 mutexed = true; | |
5925 FOR_EACH_EXPR (r, avi, orig_ops) | |
5926 if (!sched_insns_conditions_mutex_p (insn, EXPR_INSN_RTX (r))) | |
5927 { | |
5928 mutexed = false; | |
5929 break; | |
5930 } | |
5931 | |
5932 /* Mark all registers that do not meet the following condition: | |
5933 (1) Not set or read on any path from xi to an instance of the | |
5934 original operation. */ | |
5935 if (!mutexed) | |
5936 { | |
5937 IOR_REG_SET (sparams->used_regs, INSN_REG_SETS (insn)); | |
5938 IOR_REG_SET (sparams->used_regs, INSN_REG_USES (insn)); | |
5939 IOR_REG_SET (sparams->used_regs, INSN_REG_CLOBBERS (insn)); | |
5940 } | |
5941 | |
5942 return true; | |
5943 } | |
5944 | |
5945 /* Hooks and data to perform move_op operations with code_motion_path_driver. */ | |
5946 struct code_motion_path_driver_info_def move_op_hooks = { | |
5947 move_op_on_enter, | |
5948 move_op_orig_expr_found, | |
5949 move_op_orig_expr_not_found, | |
5950 move_op_merge_succs, | |
5951 move_op_after_merge_succs, | |
5952 move_op_ascend, | |
5953 move_op_at_first_insn, | |
5954 SUCCS_NORMAL, | |
5955 "move_op" | |
5956 }; | |
5957 | |
5958 /* Hooks and data to perform find_used_regs operations | |
5959 with code_motion_path_driver. */ | |
5960 struct code_motion_path_driver_info_def fur_hooks = { | |
5961 fur_on_enter, | |
5962 fur_orig_expr_found, | |
5963 fur_orig_expr_not_found, | |
5964 fur_merge_succs, | |
5965 NULL, /* fur_after_merge_succs */ | |
5966 NULL, /* fur_ascend */ | |
5967 fur_at_first_insn, | |
5968 SUCCS_ALL, | |
5969 "find_used_regs" | |
5970 }; | |
5971 | |
5972 /* Traverse all successors of INSN. For each successor that is SUCCS_NORMAL | |
5973 code_motion_path_driver is called recursively. Original operation | |
5974 was found at least on one path that is starting with one of INSN's | |
5975 successors (this fact is asserted). ORIG_OPS is expressions we're looking | |
5976 for, PATH is the path we've traversed, STATIC_PARAMS is the parameters | |
5977 of either move_op or find_used_regs depending on the caller. | |
5978 | |
5979 Return 0 if we haven't found expression, 1 if we found it, -1 if we don't | |
5980 know for sure at this point. */ | |
5981 static int | |
5982 code_motion_process_successors (insn_t insn, av_set_t orig_ops, | |
5983 ilist_t path, void *static_params) | |
5984 { | |
5985 int res = 0; | |
5986 succ_iterator succ_i; | |
5987 rtx succ; | |
5988 basic_block bb; | |
5989 int old_index; | |
5990 unsigned old_succs; | |
5991 | |
5992 struct cmpd_local_params lparams; | |
5993 expr_def _x; | |
5994 | |
5995 lparams.c_expr_local = &_x; | |
5996 lparams.c_expr_merged = NULL; | |
5997 | |
5998 /* We need to process only NORMAL succs for move_op, and collect live | |
5999 registers from ALL branches (including those leading out of the | |
6000 region) for find_used_regs. | |
6001 | |
6002 In move_op, there can be a case when insn's bb number has changed | |
6003 due to created bookkeeping. This happens very rare, as we need to | |
6004 move expression from the beginning to the end of the same block. | |
6005 Rescan successors in this case. */ | |
6006 | |
6007 rescan: | |
6008 bb = BLOCK_FOR_INSN (insn); | |
6009 old_index = bb->index; | |
6010 old_succs = EDGE_COUNT (bb->succs); | |
6011 | |
6012 FOR_EACH_SUCC_1 (succ, succ_i, insn, code_motion_path_driver_info->succ_flags) | |
6013 { | |
6014 int b; | |
6015 | |
6016 lparams.e1 = succ_i.e1; | |
6017 lparams.e2 = succ_i.e2; | |
6018 | |
6019 /* Go deep into recursion only for NORMAL edges (non-backedges within the | |
6020 current region). */ | |
6021 if (succ_i.current_flags == SUCCS_NORMAL) | |
6022 b = code_motion_path_driver (succ, orig_ops, path, &lparams, | |
6023 static_params); | |
6024 else | |
6025 b = 0; | |
6026 | |
6027 /* Merge c_expres found or unify live register sets from different | |
6028 successors. */ | |
6029 code_motion_path_driver_info->merge_succs (insn, succ, b, &lparams, | |
6030 static_params); | |
6031 if (b == 1) | |
6032 res = b; | |
6033 else if (b == -1 && res != 1) | |
6034 res = b; | |
6035 | |
6036 /* We have simplified the control flow below this point. In this case, | |
6037 the iterator becomes invalid. We need to try again. */ | |
6038 if (BLOCK_FOR_INSN (insn)->index != old_index | |
6039 || EDGE_COUNT (bb->succs) != old_succs) | |
6040 goto rescan; | |
6041 } | |
6042 | |
6043 #ifdef ENABLE_CHECKING | |
6044 /* Here, RES==1 if original expr was found at least for one of the | |
6045 successors. After the loop, RES may happen to have zero value | |
6046 only if at some point the expr searched is present in av_set, but is | |
6047 not found below. In most cases, this situation is an error. | |
6048 The exception is when the original operation is blocked by | |
6049 bookkeeping generated for another fence or for another path in current | |
6050 move_op. */ | |
6051 gcc_assert (res == 1 | |
6052 || (res == 0 | |
6053 && av_set_could_be_blocked_by_bookkeeping_p (orig_ops, | |
6054 static_params)) | |
6055 || res == -1); | |
6056 #endif | |
6057 | |
6058 /* Merge data, clean up, etc. */ | |
6059 if (res != -1 && code_motion_path_driver_info->after_merge_succs) | |
6060 code_motion_path_driver_info->after_merge_succs (&lparams, static_params); | |
6061 | |
6062 return res; | |
6063 } | |
6064 | |
6065 | |
6066 /* Perform a cleanup when the driver is about to terminate. ORIG_OPS_P | |
6067 is the pointer to the av set with expressions we were looking for, | |
6068 PATH_P is the pointer to the traversed path. */ | |
6069 static inline void | |
6070 code_motion_path_driver_cleanup (av_set_t *orig_ops_p, ilist_t *path_p) | |
6071 { | |
6072 ilist_remove (path_p); | |
6073 av_set_clear (orig_ops_p); | |
6074 } | |
6075 | |
6076 /* The driver function that implements move_op or find_used_regs | |
6077 functionality dependent whether code_motion_path_driver_INFO is set to | |
6078 &MOVE_OP_HOOKS or &FUR_HOOKS. This function implements the common parts | |
6079 of code (CFG traversal etc) that are shared among both functions. INSN | |
6080 is the insn we're starting the search from, ORIG_OPS are the expressions | |
6081 we're searching for, PATH is traversed path, LOCAL_PARAMS_IN are local | |
6082 parameters of the driver, and STATIC_PARAMS are static parameters of | |
6083 the caller. | |
6084 | |
6085 Returns whether original instructions were found. Note that top-level | |
6086 code_motion_path_driver always returns true. */ | |
6087 static int | |
6088 code_motion_path_driver (insn_t insn, av_set_t orig_ops, ilist_t path, | |
6089 cmpd_local_params_p local_params_in, | |
6090 void *static_params) | |
6091 { | |
6092 expr_t expr = NULL; | |
6093 basic_block bb = BLOCK_FOR_INSN (insn); | |
6094 insn_t first_insn, bb_tail, before_first; | |
6095 bool removed_last_insn = false; | |
6096 | |
6097 if (sched_verbose >= 6) | |
6098 { | |
6099 sel_print ("%s (", code_motion_path_driver_info->routine_name); | |
6100 dump_insn (insn); | |
6101 sel_print (","); | |
6102 dump_av_set (orig_ops); | |
6103 sel_print (")\n"); | |
6104 } | |
6105 | |
6106 gcc_assert (orig_ops); | |
6107 | |
6108 /* If no original operations exist below this insn, return immediately. */ | |
6109 if (is_ineligible_successor (insn, path)) | |
6110 { | |
6111 if (sched_verbose >= 6) | |
6112 sel_print ("Insn %d is ineligible successor\n", INSN_UID (insn)); | |
6113 return false; | |
6114 } | |
6115 | |
6116 /* The block can have invalid av set, in which case it was created earlier | |
6117 during move_op. Return immediately. */ | |
6118 if (sel_bb_head_p (insn)) | |
6119 { | |
6120 if (! AV_SET_VALID_P (insn)) | |
6121 { | |
6122 if (sched_verbose >= 6) | |
6123 sel_print ("Returned from block %d as it had invalid av set\n", | |
6124 bb->index); | |
6125 return false; | |
6126 } | |
6127 | |
6128 if (bitmap_bit_p (code_motion_visited_blocks, bb->index)) | |
6129 { | |
6130 /* We have already found an original operation on this branch, do not | |
6131 go any further and just return TRUE here. If we don't stop here, | |
6132 function can have exponential behaviour even on the small code | |
6133 with many different paths (e.g. with data speculation and | |
6134 recovery blocks). */ | |
6135 if (sched_verbose >= 6) | |
6136 sel_print ("Block %d already visited in this traversal\n", bb->index); | |
6137 if (code_motion_path_driver_info->on_enter) | |
6138 return code_motion_path_driver_info->on_enter (insn, | |
6139 local_params_in, | |
6140 static_params, | |
6141 true); | |
6142 } | |
6143 } | |
6144 | |
6145 if (code_motion_path_driver_info->on_enter) | |
6146 code_motion_path_driver_info->on_enter (insn, local_params_in, | |
6147 static_params, false); | |
6148 orig_ops = av_set_copy (orig_ops); | |
6149 | |
6150 /* Filter the orig_ops set. */ | |
6151 if (AV_SET_VALID_P (insn)) | |
6152 av_set_intersect (&orig_ops, AV_SET (insn)); | |
6153 | |
6154 /* If no more original ops, return immediately. */ | |
6155 if (!orig_ops) | |
6156 { | |
6157 if (sched_verbose >= 6) | |
6158 sel_print ("No intersection with av set of block %d\n", bb->index); | |
6159 return false; | |
6160 } | |
6161 | |
6162 /* For non-speculative insns we have to leave only one form of the | |
6163 original operation, because if we don't, we may end up with | |
6164 different C_EXPRes and, consequently, with bookkeepings for different | |
6165 expression forms along the same code motion path. That may lead to | |
6166 generation of incorrect code. So for each code motion we stick to | |
6167 the single form of the instruction, except for speculative insns | |
6168 which we need to keep in different forms with all speculation | |
6169 types. */ | |
6170 av_set_leave_one_nonspec (&orig_ops); | |
6171 | |
6172 /* It is not possible that all ORIG_OPS are filtered out. */ | |
6173 gcc_assert (orig_ops); | |
6174 | |
6175 /* It is enough to place only heads and tails of visited basic blocks into | |
6176 the PATH. */ | |
6177 ilist_add (&path, insn); | |
6178 first_insn = insn; | |
6179 bb_tail = sel_bb_end (bb); | |
6180 | |
6181 /* Descend the basic block in search of the original expr; this part | |
6182 corresponds to the part of the original move_op procedure executed | |
6183 before the recursive call. */ | |
6184 for (;;) | |
6185 { | |
6186 /* Look at the insn and decide if it could be an ancestor of currently | |
6187 scheduling operation. If it is so, then the insn "dest = op" could | |
6188 either be replaced with "dest = reg", because REG now holds the result | |
6189 of OP, or just removed, if we've scheduled the insn as a whole. | |
6190 | |
6191 If this insn doesn't contain currently scheduling OP, then proceed | |
6192 with searching and look at its successors. Operations we're searching | |
6193 for could have changed when moving up through this insn via | |
6194 substituting. In this case, perform unsubstitution on them first. | |
6195 | |
6196 When traversing the DAG below this insn is finished, insert | |
6197 bookkeeping code, if the insn is a joint point, and remove | |
6198 leftovers. */ | |
6199 | |
6200 expr = av_set_lookup (orig_ops, INSN_VINSN (insn)); | |
6201 if (expr) | |
6202 { | |
6203 insn_t last_insn = PREV_INSN (insn); | |
6204 | |
6205 /* We have found the original operation. */ | |
6206 if (sched_verbose >= 6) | |
6207 sel_print ("Found original operation at insn %d\n", INSN_UID (insn)); | |
6208 | |
6209 code_motion_path_driver_info->orig_expr_found | |
6210 (insn, expr, local_params_in, static_params); | |
6211 | |
6212 /* Step back, so on the way back we'll start traversing from the | |
6213 previous insn (or we'll see that it's bb_note and skip that | |
6214 loop). */ | |
6215 if (insn == first_insn) | |
6216 { | |
6217 first_insn = NEXT_INSN (last_insn); | |
6218 removed_last_insn = sel_bb_end_p (last_insn); | |
6219 } | |
6220 insn = last_insn; | |
6221 break; | |
6222 } | |
6223 else | |
6224 { | |
6225 /* We haven't found the original expr, continue descending the basic | |
6226 block. */ | |
6227 if (code_motion_path_driver_info->orig_expr_not_found | |
6228 (insn, orig_ops, static_params)) | |
6229 { | |
6230 /* Av set ops could have been changed when moving through this | |
6231 insn. To find them below it, we have to un-substitute them. */ | |
6232 undo_transformations (&orig_ops, insn); | |
6233 } | |
6234 else | |
6235 { | |
6236 /* Clean up and return, if the hook tells us to do so. It may | |
6237 happen if we've encountered the previously created | |
6238 bookkeeping. */ | |
6239 code_motion_path_driver_cleanup (&orig_ops, &path); | |
6240 return -1; | |
6241 } | |
6242 | |
6243 gcc_assert (orig_ops); | |
6244 } | |
6245 | |
6246 /* Stop at insn if we got to the end of BB. */ | |
6247 if (insn == bb_tail) | |
6248 break; | |
6249 | |
6250 insn = NEXT_INSN (insn); | |
6251 } | |
6252 | |
6253 /* Here INSN either points to the insn before the original insn (may be | |
6254 bb_note, if original insn was a bb_head) or to the bb_end. */ | |
6255 if (!expr) | |
6256 { | |
6257 int res; | |
6258 | |
6259 gcc_assert (insn == sel_bb_end (bb)); | |
6260 | |
6261 /* Add bb tail to PATH (but it doesn't make any sense if it's a bb_head - | |
6262 it's already in PATH then). */ | |
6263 if (insn != first_insn) | |
6264 ilist_add (&path, insn); | |
6265 | |
6266 /* Process_successors should be able to find at least one | |
6267 successor for which code_motion_path_driver returns TRUE. */ | |
6268 res = code_motion_process_successors (insn, orig_ops, | |
6269 path, static_params); | |
6270 | |
6271 /* Remove bb tail from path. */ | |
6272 if (insn != first_insn) | |
6273 ilist_remove (&path); | |
6274 | |
6275 if (res != 1) | |
6276 { | |
6277 /* This is the case when one of the original expr is no longer available | |
6278 due to bookkeeping created on this branch with the same register. | |
6279 In the original algorithm, which doesn't have update_data_sets call | |
6280 on a bookkeeping block, it would simply result in returning | |
6281 FALSE when we've encountered a previously generated bookkeeping | |
6282 insn in moveop_orig_expr_not_found. */ | |
6283 code_motion_path_driver_cleanup (&orig_ops, &path); | |
6284 return res; | |
6285 } | |
6286 } | |
6287 | |
6288 /* Don't need it any more. */ | |
6289 av_set_clear (&orig_ops); | |
6290 | |
6291 /* Backward pass: now, when we have C_EXPR computed, we'll drag it to | |
6292 the beginning of the basic block. */ | |
6293 before_first = PREV_INSN (first_insn); | |
6294 while (insn != before_first) | |
6295 { | |
6296 if (code_motion_path_driver_info->ascend) | |
6297 code_motion_path_driver_info->ascend (insn, static_params); | |
6298 | |
6299 insn = PREV_INSN (insn); | |
6300 } | |
6301 | |
6302 /* Now we're at the bb head. */ | |
6303 insn = first_insn; | |
6304 ilist_remove (&path); | |
6305 local_params_in->removed_last_insn = removed_last_insn; | |
6306 code_motion_path_driver_info->at_first_insn (insn, local_params_in, static_params); | |
6307 | |
6308 /* This should be the very last operation as at bb head we could change | |
6309 the numbering by creating bookkeeping blocks. */ | |
6310 if (removed_last_insn) | |
6311 insn = PREV_INSN (insn); | |
6312 bitmap_set_bit (code_motion_visited_blocks, BLOCK_FOR_INSN (insn)->index); | |
6313 return true; | |
6314 } | |
6315 | |
6316 /* Move up the operations from ORIG_OPS set traversing the dag starting | |
6317 from INSN. PATH represents the edges traversed so far. | |
6318 DEST is the register chosen for scheduling the current expr. Insert | |
6319 bookkeeping code in the join points. EXPR_VLIW is the chosen expression, | |
6320 C_EXPR is how it looks like at the given cfg point. | |
6321 Set *SHOULD_MOVE to indicate whether we have only disconnected | |
6322 one of the insns found. | |
6323 | |
6324 Returns whether original instructions were found, which is asserted | |
6325 to be true in the caller. */ | |
6326 static bool | |
6327 move_op (insn_t insn, av_set_t orig_ops, expr_t expr_vliw, | |
6328 rtx dest, expr_t c_expr, bool *should_move) | |
6329 { | |
6330 struct moveop_static_params sparams; | |
6331 struct cmpd_local_params lparams; | |
6332 bool res; | |
6333 | |
6334 /* Init params for code_motion_path_driver. */ | |
6335 sparams.dest = dest; | |
6336 sparams.c_expr = c_expr; | |
6337 sparams.uid = INSN_UID (EXPR_INSN_RTX (expr_vliw)); | |
6338 #ifdef ENABLE_CHECKING | |
6339 sparams.failed_insn = NULL; | |
6340 #endif | |
6341 sparams.was_renamed = false; | |
6342 lparams.e1 = NULL; | |
6343 | |
6344 /* We haven't visited any blocks yet. */ | |
6345 bitmap_clear (code_motion_visited_blocks); | |
6346 | |
6347 /* Set appropriate hooks and data. */ | |
6348 code_motion_path_driver_info = &move_op_hooks; | |
6349 res = code_motion_path_driver (insn, orig_ops, NULL, &lparams, &sparams); | |
6350 | |
6351 if (sparams.was_renamed) | |
6352 EXPR_WAS_RENAMED (expr_vliw) = true; | |
6353 | |
6354 *should_move = (sparams.uid == -1); | |
6355 | |
6356 return res; | |
6357 } | |
6358 | |
6359 | |
6360 /* Functions that work with regions. */ | |
6361 | |
6362 /* Current number of seqno used in init_seqno and init_seqno_1. */ | |
6363 static int cur_seqno; | |
6364 | |
6365 /* A helper for init_seqno. Traverse the region starting from BB and | |
6366 compute seqnos for visited insns, marking visited bbs in VISITED_BBS. | |
6367 Clear visited blocks from BLOCKS_TO_RESCHEDULE. */ | |
6368 static void | |
6369 init_seqno_1 (basic_block bb, sbitmap visited_bbs, bitmap blocks_to_reschedule) | |
6370 { | |
6371 int bbi = BLOCK_TO_BB (bb->index); | |
6372 insn_t insn, note = bb_note (bb); | |
6373 insn_t succ_insn; | |
6374 succ_iterator si; | |
6375 | |
6376 SET_BIT (visited_bbs, bbi); | |
6377 if (blocks_to_reschedule) | |
6378 bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
6379 | |
6380 FOR_EACH_SUCC_1 (succ_insn, si, BB_END (bb), | |
6381 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
6382 { | |
6383 basic_block succ = BLOCK_FOR_INSN (succ_insn); | |
6384 int succ_bbi = BLOCK_TO_BB (succ->index); | |
6385 | |
6386 gcc_assert (in_current_region_p (succ)); | |
6387 | |
6388 if (!TEST_BIT (visited_bbs, succ_bbi)) | |
6389 { | |
6390 gcc_assert (succ_bbi > bbi); | |
6391 | |
6392 init_seqno_1 (succ, visited_bbs, blocks_to_reschedule); | |
6393 } | |
6394 } | |
6395 | |
6396 for (insn = BB_END (bb); insn != note; insn = PREV_INSN (insn)) | |
6397 INSN_SEQNO (insn) = cur_seqno--; | |
6398 } | |
6399 | |
6400 /* Initialize seqnos for the current region. NUMBER_OF_INSNS is the number | |
6401 of instructions in the region, BLOCKS_TO_RESCHEDULE contains blocks on | |
6402 which we're rescheduling when pipelining, FROM is the block where | |
6403 traversing region begins (it may not be the head of the region when | |
6404 pipelining, but the head of the loop instead). | |
6405 | |
6406 Returns the maximal seqno found. */ | |
6407 static int | |
6408 init_seqno (int number_of_insns, bitmap blocks_to_reschedule, basic_block from) | |
6409 { | |
6410 sbitmap visited_bbs; | |
6411 bitmap_iterator bi; | |
6412 unsigned bbi; | |
6413 | |
6414 visited_bbs = sbitmap_alloc (current_nr_blocks); | |
6415 | |
6416 if (blocks_to_reschedule) | |
6417 { | |
6418 sbitmap_ones (visited_bbs); | |
6419 EXECUTE_IF_SET_IN_BITMAP (blocks_to_reschedule, 0, bbi, bi) | |
6420 { | |
6421 gcc_assert (BLOCK_TO_BB (bbi) < current_nr_blocks); | |
6422 RESET_BIT (visited_bbs, BLOCK_TO_BB (bbi)); | |
6423 } | |
6424 } | |
6425 else | |
6426 { | |
6427 sbitmap_zero (visited_bbs); | |
6428 from = EBB_FIRST_BB (0); | |
6429 } | |
6430 | |
6431 cur_seqno = number_of_insns > 0 ? number_of_insns : sched_max_luid - 1; | |
6432 init_seqno_1 (from, visited_bbs, blocks_to_reschedule); | |
6433 gcc_assert (cur_seqno == 0 || number_of_insns == 0); | |
6434 | |
6435 sbitmap_free (visited_bbs); | |
6436 return sched_max_luid - 1; | |
6437 } | |
6438 | |
6439 /* Initialize scheduling parameters for current region. */ | |
6440 static void | |
6441 sel_setup_region_sched_flags (void) | |
6442 { | |
6443 enable_schedule_as_rhs_p = 1; | |
6444 bookkeeping_p = 1; | |
6445 pipelining_p = (bookkeeping_p | |
6446 && (flag_sel_sched_pipelining != 0) | |
6447 && current_loop_nest != NULL); | |
6448 max_insns_to_rename = PARAM_VALUE (PARAM_SELSCHED_INSNS_TO_RENAME); | |
6449 max_ws = MAX_WS; | |
6450 } | |
6451 | |
6452 /* Return true if all basic blocks of current region are empty. */ | |
6453 static bool | |
6454 current_region_empty_p (void) | |
6455 { | |
6456 int i; | |
6457 for (i = 0; i < current_nr_blocks; i++) | |
6458 if (! sel_bb_empty_p (BASIC_BLOCK (BB_TO_BLOCK (i)))) | |
6459 return false; | |
6460 | |
6461 return true; | |
6462 } | |
6463 | |
6464 /* Prepare and verify loop nest for pipelining. */ | |
6465 static void | |
6466 setup_current_loop_nest (int rgn) | |
6467 { | |
6468 current_loop_nest = get_loop_nest_for_rgn (rgn); | |
6469 | |
6470 if (!current_loop_nest) | |
6471 return; | |
6472 | |
6473 /* If this loop has any saved loop preheaders from nested loops, | |
6474 add these basic blocks to the current region. */ | |
6475 sel_add_loop_preheaders (); | |
6476 | |
6477 /* Check that we're starting with a valid information. */ | |
6478 gcc_assert (loop_latch_edge (current_loop_nest)); | |
6479 gcc_assert (LOOP_MARKED_FOR_PIPELINING_P (current_loop_nest)); | |
6480 } | |
6481 | |
6482 /* Purge meaningless empty blocks in the middle of a region. */ | |
6483 static void | |
6484 purge_empty_blocks (void) | |
6485 { | |
6486 int i ; | |
6487 | |
6488 for (i = 1; i < current_nr_blocks; ) | |
6489 { | |
6490 basic_block b = BASIC_BLOCK (BB_TO_BLOCK (i)); | |
6491 | |
6492 if (maybe_tidy_empty_bb (b)) | |
6493 continue; | |
6494 | |
6495 i++; | |
6496 } | |
6497 } | |
6498 | |
6499 /* Compute instruction priorities for current region. */ | |
6500 static void | |
6501 sel_compute_priorities (int rgn) | |
6502 { | |
6503 sched_rgn_compute_dependencies (rgn); | |
6504 | |
6505 /* Compute insn priorities in haifa style. Then free haifa style | |
6506 dependencies that we've calculated for this. */ | |
6507 compute_priorities (); | |
6508 | |
6509 if (sched_verbose >= 5) | |
6510 debug_rgn_dependencies (0); | |
6511 | |
6512 free_rgn_deps (); | |
6513 } | |
6514 | |
6515 /* Init scheduling data for RGN. Returns true when this region should not | |
6516 be scheduled. */ | |
6517 static bool | |
6518 sel_region_init (int rgn) | |
6519 { | |
6520 int i; | |
6521 bb_vec_t bbs; | |
6522 | |
6523 rgn_setup_region (rgn); | |
6524 | |
6525 /* Even if sched_is_disabled_for_current_region_p() is true, we still | |
6526 do region initialization here so the region can be bundled correctly, | |
6527 but we'll skip the scheduling in sel_sched_region (). */ | |
6528 if (current_region_empty_p ()) | |
6529 return true; | |
6530 | |
6531 if (flag_sel_sched_pipelining) | |
6532 setup_current_loop_nest (rgn); | |
6533 | |
6534 sel_setup_region_sched_flags (); | |
6535 | |
6536 bbs = VEC_alloc (basic_block, heap, current_nr_blocks); | |
6537 | |
6538 for (i = 0; i < current_nr_blocks; i++) | |
6539 VEC_quick_push (basic_block, bbs, BASIC_BLOCK (BB_TO_BLOCK (i))); | |
6540 | |
6541 sel_init_bbs (bbs, NULL); | |
6542 | |
6543 /* Initialize luids and dependence analysis which both sel-sched and haifa | |
6544 need. */ | |
6545 sched_init_luids (bbs, NULL, NULL, NULL); | |
6546 sched_deps_init (false); | |
6547 | |
6548 /* Initialize haifa data. */ | |
6549 rgn_setup_sched_infos (); | |
6550 sel_set_sched_flags (); | |
6551 haifa_init_h_i_d (bbs, NULL, NULL, NULL); | |
6552 | |
6553 sel_compute_priorities (rgn); | |
6554 init_deps_global (); | |
6555 | |
6556 /* Main initialization. */ | |
6557 sel_setup_sched_infos (); | |
6558 sel_init_global_and_expr (bbs); | |
6559 | |
6560 VEC_free (basic_block, heap, bbs); | |
6561 | |
6562 blocks_to_reschedule = BITMAP_ALLOC (NULL); | |
6563 | |
6564 /* Init correct liveness sets on each instruction of a single-block loop. | |
6565 This is the only situation when we can't update liveness when calling | |
6566 compute_live for the first insn of the loop. */ | |
6567 if (current_loop_nest) | |
6568 { | |
6569 int header = (sel_is_loop_preheader_p (BASIC_BLOCK (BB_TO_BLOCK (0))) | |
6570 ? 1 | |
6571 : 0); | |
6572 | |
6573 if (current_nr_blocks == header + 1) | |
6574 update_liveness_on_insn | |
6575 (sel_bb_head (BASIC_BLOCK (BB_TO_BLOCK (header)))); | |
6576 } | |
6577 | |
6578 /* Set hooks so that no newly generated insn will go out unnoticed. */ | |
6579 sel_register_cfg_hooks (); | |
6580 | |
6581 /* !!! We call target.sched.md_init () for the whole region, but we invoke | |
6582 targetm.sched.md_finish () for every ebb. */ | |
6583 if (targetm.sched.md_init) | |
6584 /* None of the arguments are actually used in any target. */ | |
6585 targetm.sched.md_init (sched_dump, sched_verbose, -1); | |
6586 | |
6587 first_emitted_uid = get_max_uid () + 1; | |
6588 preheader_removed = false; | |
6589 | |
6590 /* Reset register allocation ticks array. */ | |
6591 memset (reg_rename_tick, 0, sizeof reg_rename_tick); | |
6592 reg_rename_this_tick = 0; | |
6593 | |
6594 bitmap_initialize (forced_ebb_heads, 0); | |
6595 bitmap_clear (forced_ebb_heads); | |
6596 | |
6597 setup_nop_vinsn (); | |
6598 current_copies = BITMAP_ALLOC (NULL); | |
6599 current_originators = BITMAP_ALLOC (NULL); | |
6600 code_motion_visited_blocks = BITMAP_ALLOC (NULL); | |
6601 | |
6602 return false; | |
6603 } | |
6604 | |
6605 /* Simplify insns after the scheduling. */ | |
6606 static void | |
6607 simplify_changed_insns (void) | |
6608 { | |
6609 int i; | |
6610 | |
6611 for (i = 0; i < current_nr_blocks; i++) | |
6612 { | |
6613 basic_block bb = BASIC_BLOCK (BB_TO_BLOCK (i)); | |
6614 rtx insn; | |
6615 | |
6616 FOR_BB_INSNS (bb, insn) | |
6617 if (INSN_P (insn)) | |
6618 { | |
6619 expr_t expr = INSN_EXPR (insn); | |
6620 | |
6621 if (EXPR_WAS_SUBSTITUTED (expr)) | |
6622 validate_simplify_insn (insn); | |
6623 } | |
6624 } | |
6625 } | |
6626 | |
6627 /* Find boundaries of the EBB starting from basic block BB, marking blocks of | |
6628 this EBB in SCHEDULED_BLOCKS and appropriately filling in HEAD, TAIL, | |
6629 PREV_HEAD, and NEXT_TAIL fields of CURRENT_SCHED_INFO structure. */ | |
6630 static void | |
6631 find_ebb_boundaries (basic_block bb, bitmap scheduled_blocks) | |
6632 { | |
6633 insn_t head, tail; | |
6634 basic_block bb1 = bb; | |
6635 if (sched_verbose >= 2) | |
6636 sel_print ("Finishing schedule in bbs: "); | |
6637 | |
6638 do | |
6639 { | |
6640 bitmap_set_bit (scheduled_blocks, BLOCK_TO_BB (bb1->index)); | |
6641 | |
6642 if (sched_verbose >= 2) | |
6643 sel_print ("%d; ", bb1->index); | |
6644 } | |
6645 while (!bb_ends_ebb_p (bb1) && (bb1 = bb_next_bb (bb1))); | |
6646 | |
6647 if (sched_verbose >= 2) | |
6648 sel_print ("\n"); | |
6649 | |
6650 get_ebb_head_tail (bb, bb1, &head, &tail); | |
6651 | |
6652 current_sched_info->head = head; | |
6653 current_sched_info->tail = tail; | |
6654 current_sched_info->prev_head = PREV_INSN (head); | |
6655 current_sched_info->next_tail = NEXT_INSN (tail); | |
6656 } | |
6657 | |
6658 /* Regenerate INSN_SCHED_CYCLEs for insns of current EBB. */ | |
6659 static void | |
6660 reset_sched_cycles_in_current_ebb (void) | |
6661 { | |
6662 int last_clock = 0; | |
6663 int haifa_last_clock = -1; | |
6664 int haifa_clock = 0; | |
6665 insn_t insn; | |
6666 | |
6667 if (targetm.sched.md_init) | |
6668 { | |
6669 /* None of the arguments are actually used in any target. | |
6670 NB: We should have md_reset () hook for cases like this. */ | |
6671 targetm.sched.md_init (sched_dump, sched_verbose, -1); | |
6672 } | |
6673 | |
6674 state_reset (curr_state); | |
6675 advance_state (curr_state); | |
6676 | |
6677 for (insn = current_sched_info->head; | |
6678 insn != current_sched_info->next_tail; | |
6679 insn = NEXT_INSN (insn)) | |
6680 { | |
6681 int cost, haifa_cost; | |
6682 int sort_p; | |
6683 bool asm_p, real_insn, after_stall; | |
6684 int clock; | |
6685 | |
6686 if (!INSN_P (insn)) | |
6687 continue; | |
6688 | |
6689 asm_p = false; | |
6690 real_insn = recog_memoized (insn) >= 0; | |
6691 clock = INSN_SCHED_CYCLE (insn); | |
6692 | |
6693 cost = clock - last_clock; | |
6694 | |
6695 /* Initialize HAIFA_COST. */ | |
6696 if (! real_insn) | |
6697 { | |
6698 asm_p = INSN_ASM_P (insn); | |
6699 | |
6700 if (asm_p) | |
6701 /* This is asm insn which *had* to be scheduled first | |
6702 on the cycle. */ | |
6703 haifa_cost = 1; | |
6704 else | |
6705 /* This is a use/clobber insn. It should not change | |
6706 cost. */ | |
6707 haifa_cost = 0; | |
6708 } | |
6709 else | |
6710 haifa_cost = estimate_insn_cost (insn, curr_state); | |
6711 | |
6712 /* Stall for whatever cycles we've stalled before. */ | |
6713 after_stall = 0; | |
6714 if (INSN_AFTER_STALL_P (insn) && cost > haifa_cost) | |
6715 { | |
6716 haifa_cost = cost; | |
6717 after_stall = 1; | |
6718 } | |
6719 | |
6720 if (haifa_cost > 0) | |
6721 { | |
6722 int i = 0; | |
6723 | |
6724 while (haifa_cost--) | |
6725 { | |
6726 advance_state (curr_state); | |
6727 i++; | |
6728 | |
6729 if (sched_verbose >= 2) | |
6730 { | |
6731 sel_print ("advance_state (state_transition)\n"); | |
6732 debug_state (curr_state); | |
6733 } | |
6734 | |
6735 /* The DFA may report that e.g. insn requires 2 cycles to be | |
6736 issued, but on the next cycle it says that insn is ready | |
6737 to go. Check this here. */ | |
6738 if (!after_stall | |
6739 && real_insn | |
6740 && haifa_cost > 0 | |
6741 && estimate_insn_cost (insn, curr_state) == 0) | |
6742 break; | |
6743 } | |
6744 | |
6745 haifa_clock += i; | |
6746 } | |
6747 else | |
6748 gcc_assert (haifa_cost == 0); | |
6749 | |
6750 if (sched_verbose >= 2) | |
6751 sel_print ("Haifa cost for insn %d: %d\n", INSN_UID (insn), haifa_cost); | |
6752 | |
6753 if (targetm.sched.dfa_new_cycle) | |
6754 while (targetm.sched.dfa_new_cycle (sched_dump, sched_verbose, insn, | |
6755 haifa_last_clock, haifa_clock, | |
6756 &sort_p)) | |
6757 { | |
6758 advance_state (curr_state); | |
6759 haifa_clock++; | |
6760 if (sched_verbose >= 2) | |
6761 { | |
6762 sel_print ("advance_state (dfa_new_cycle)\n"); | |
6763 debug_state (curr_state); | |
6764 } | |
6765 } | |
6766 | |
6767 if (real_insn) | |
6768 { | |
6769 cost = state_transition (curr_state, insn); | |
6770 | |
6771 if (sched_verbose >= 2) | |
6772 debug_state (curr_state); | |
6773 | |
6774 gcc_assert (cost < 0); | |
6775 } | |
6776 | |
6777 if (targetm.sched.variable_issue) | |
6778 targetm.sched.variable_issue (sched_dump, sched_verbose, insn, 0); | |
6779 | |
6780 INSN_SCHED_CYCLE (insn) = haifa_clock; | |
6781 | |
6782 last_clock = clock; | |
6783 haifa_last_clock = haifa_clock; | |
6784 } | |
6785 } | |
6786 | |
6787 /* Put TImode markers on insns starting a new issue group. */ | |
6788 static void | |
6789 put_TImodes (void) | |
6790 { | |
6791 int last_clock = -1; | |
6792 insn_t insn; | |
6793 | |
6794 for (insn = current_sched_info->head; insn != current_sched_info->next_tail; | |
6795 insn = NEXT_INSN (insn)) | |
6796 { | |
6797 int cost, clock; | |
6798 | |
6799 if (!INSN_P (insn)) | |
6800 continue; | |
6801 | |
6802 clock = INSN_SCHED_CYCLE (insn); | |
6803 cost = (last_clock == -1) ? 1 : clock - last_clock; | |
6804 | |
6805 gcc_assert (cost >= 0); | |
6806 | |
6807 if (issue_rate > 1 | |
6808 && GET_CODE (PATTERN (insn)) != USE | |
6809 && GET_CODE (PATTERN (insn)) != CLOBBER) | |
6810 { | |
6811 if (reload_completed && cost > 0) | |
6812 PUT_MODE (insn, TImode); | |
6813 | |
6814 last_clock = clock; | |
6815 } | |
6816 | |
6817 if (sched_verbose >= 2) | |
6818 sel_print ("Cost for insn %d is %d\n", INSN_UID (insn), cost); | |
6819 } | |
6820 } | |
6821 | |
6822 /* Perform MD_FINISH on EBBs comprising current region. When | |
6823 RESET_SCHED_CYCLES_P is true, run a pass emulating the scheduler | |
6824 to produce correct sched cycles on insns. */ | |
6825 static void | |
6826 sel_region_target_finish (bool reset_sched_cycles_p) | |
6827 { | |
6828 int i; | |
6829 bitmap scheduled_blocks = BITMAP_ALLOC (NULL); | |
6830 | |
6831 for (i = 0; i < current_nr_blocks; i++) | |
6832 { | |
6833 if (bitmap_bit_p (scheduled_blocks, i)) | |
6834 continue; | |
6835 | |
6836 /* While pipelining outer loops, skip bundling for loop | |
6837 preheaders. Those will be rescheduled in the outer loop. */ | |
6838 if (sel_is_loop_preheader_p (EBB_FIRST_BB (i))) | |
6839 continue; | |
6840 | |
6841 find_ebb_boundaries (EBB_FIRST_BB (i), scheduled_blocks); | |
6842 | |
6843 if (no_real_insns_p (current_sched_info->head, current_sched_info->tail)) | |
6844 continue; | |
6845 | |
6846 if (reset_sched_cycles_p) | |
6847 reset_sched_cycles_in_current_ebb (); | |
6848 | |
6849 if (targetm.sched.md_init) | |
6850 targetm.sched.md_init (sched_dump, sched_verbose, -1); | |
6851 | |
6852 put_TImodes (); | |
6853 | |
6854 if (targetm.sched.md_finish) | |
6855 { | |
6856 targetm.sched.md_finish (sched_dump, sched_verbose); | |
6857 | |
6858 /* Extend luids so that insns generated by the target will | |
6859 get zero luid. */ | |
6860 sched_init_luids (NULL, NULL, NULL, NULL); | |
6861 } | |
6862 } | |
6863 | |
6864 BITMAP_FREE (scheduled_blocks); | |
6865 } | |
6866 | |
6867 /* Free the scheduling data for the current region. When RESET_SCHED_CYCLES_P | |
6868 is true, make an additional pass emulating scheduler to get correct insn | |
6869 cycles for md_finish calls. */ | |
6870 static void | |
6871 sel_region_finish (bool reset_sched_cycles_p) | |
6872 { | |
6873 simplify_changed_insns (); | |
6874 sched_finish_ready_list (); | |
6875 free_nop_pool (); | |
6876 | |
6877 /* Free the vectors. */ | |
6878 if (vec_av_set) | |
6879 VEC_free (expr_t, heap, vec_av_set); | |
6880 BITMAP_FREE (current_copies); | |
6881 BITMAP_FREE (current_originators); | |
6882 BITMAP_FREE (code_motion_visited_blocks); | |
6883 vinsn_vec_free (&vec_bookkeeping_blocked_vinsns); | |
6884 vinsn_vec_free (&vec_target_unavailable_vinsns); | |
6885 | |
6886 /* If LV_SET of the region head should be updated, do it now because | |
6887 there will be no other chance. */ | |
6888 { | |
6889 succ_iterator si; | |
6890 insn_t insn; | |
6891 | |
6892 FOR_EACH_SUCC_1 (insn, si, bb_note (EBB_FIRST_BB (0)), | |
6893 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
6894 { | |
6895 basic_block bb = BLOCK_FOR_INSN (insn); | |
6896 | |
6897 if (!BB_LV_SET_VALID_P (bb)) | |
6898 compute_live (insn); | |
6899 } | |
6900 } | |
6901 | |
6902 /* Emulate the Haifa scheduler for bundling. */ | |
6903 if (reload_completed) | |
6904 sel_region_target_finish (reset_sched_cycles_p); | |
6905 | |
6906 sel_finish_global_and_expr (); | |
6907 | |
6908 bitmap_clear (forced_ebb_heads); | |
6909 | |
6910 free_nop_vinsn (); | |
6911 | |
6912 finish_deps_global (); | |
6913 sched_finish_luids (); | |
6914 | |
6915 sel_finish_bbs (); | |
6916 BITMAP_FREE (blocks_to_reschedule); | |
6917 | |
6918 sel_unregister_cfg_hooks (); | |
6919 | |
6920 max_issue_size = 0; | |
6921 } | |
6922 | |
6923 | |
6924 /* Functions that implement the scheduler driver. */ | |
6925 | |
6926 /* Schedule a parallel instruction group on each of FENCES. MAX_SEQNO | |
6927 is the current maximum seqno. SCHEDULED_INSNS_TAILPP is the list | |
6928 of insns scheduled -- these would be postprocessed later. */ | |
6929 static void | |
6930 schedule_on_fences (flist_t fences, int max_seqno, | |
6931 ilist_t **scheduled_insns_tailpp) | |
6932 { | |
6933 flist_t old_fences = fences; | |
6934 | |
6935 if (sched_verbose >= 1) | |
6936 { | |
6937 sel_print ("\nScheduling on fences: "); | |
6938 dump_flist (fences); | |
6939 sel_print ("\n"); | |
6940 } | |
6941 | |
6942 scheduled_something_on_previous_fence = false; | |
6943 for (; fences; fences = FLIST_NEXT (fences)) | |
6944 { | |
6945 fence_t fence = NULL; | |
6946 int seqno = 0; | |
6947 flist_t fences2; | |
6948 bool first_p = true; | |
6949 | |
6950 /* Choose the next fence group to schedule. | |
6951 The fact that insn can be scheduled only once | |
6952 on the cycle is guaranteed by two properties: | |
6953 1. seqnos of parallel groups decrease with each iteration. | |
6954 2. If is_ineligible_successor () sees the larger seqno, it | |
6955 checks if candidate insn is_in_current_fence_p (). */ | |
6956 for (fences2 = old_fences; fences2; fences2 = FLIST_NEXT (fences2)) | |
6957 { | |
6958 fence_t f = FLIST_FENCE (fences2); | |
6959 | |
6960 if (!FENCE_PROCESSED_P (f)) | |
6961 { | |
6962 int i = INSN_SEQNO (FENCE_INSN (f)); | |
6963 | |
6964 if (first_p || i > seqno) | |
6965 { | |
6966 seqno = i; | |
6967 fence = f; | |
6968 first_p = false; | |
6969 } | |
6970 else | |
6971 /* ??? Seqnos of different groups should be different. */ | |
6972 gcc_assert (1 || i != seqno); | |
6973 } | |
6974 } | |
6975 | |
6976 gcc_assert (fence); | |
6977 | |
6978 /* As FENCE is nonnull, SEQNO is initialized. */ | |
6979 seqno -= max_seqno + 1; | |
6980 fill_insns (fence, seqno, scheduled_insns_tailpp); | |
6981 FENCE_PROCESSED_P (fence) = true; | |
6982 } | |
6983 | |
6984 /* All av_sets are invalidated by GLOBAL_LEVEL increase, thus we | |
6985 don't need to keep bookkeeping-invalidated and target-unavailable | |
6986 vinsns any more. */ | |
6987 vinsn_vec_clear (&vec_bookkeeping_blocked_vinsns); | |
6988 vinsn_vec_clear (&vec_target_unavailable_vinsns); | |
6989 } | |
6990 | |
6991 /* Calculate MIN_SEQNO and MAX_SEQNO. */ | |
6992 static void | |
6993 find_min_max_seqno (flist_t fences, int *min_seqno, int *max_seqno) | |
6994 { | |
6995 *min_seqno = *max_seqno = INSN_SEQNO (FENCE_INSN (FLIST_FENCE (fences))); | |
6996 | |
6997 /* The first element is already processed. */ | |
6998 while ((fences = FLIST_NEXT (fences))) | |
6999 { | |
7000 int seqno = INSN_SEQNO (FENCE_INSN (FLIST_FENCE (fences))); | |
7001 | |
7002 if (*min_seqno > seqno) | |
7003 *min_seqno = seqno; | |
7004 else if (*max_seqno < seqno) | |
7005 *max_seqno = seqno; | |
7006 } | |
7007 } | |
7008 | |
7009 /* Calculate new fences from FENCES. */ | |
7010 static flist_t | |
7011 calculate_new_fences (flist_t fences, int orig_max_seqno) | |
7012 { | |
7013 flist_t old_fences = fences; | |
7014 struct flist_tail_def _new_fences, *new_fences = &_new_fences; | |
7015 | |
7016 flist_tail_init (new_fences); | |
7017 for (; fences; fences = FLIST_NEXT (fences)) | |
7018 { | |
7019 fence_t fence = FLIST_FENCE (fences); | |
7020 insn_t insn; | |
7021 | |
7022 if (!FENCE_BNDS (fence)) | |
7023 { | |
7024 /* This fence doesn't have any successors. */ | |
7025 if (!FENCE_SCHEDULED_P (fence)) | |
7026 { | |
7027 /* Nothing was scheduled on this fence. */ | |
7028 int seqno; | |
7029 | |
7030 insn = FENCE_INSN (fence); | |
7031 seqno = INSN_SEQNO (insn); | |
7032 gcc_assert (seqno > 0 && seqno <= orig_max_seqno); | |
7033 | |
7034 if (sched_verbose >= 1) | |
7035 sel_print ("Fence %d[%d] has not changed\n", | |
7036 INSN_UID (insn), | |
7037 BLOCK_NUM (insn)); | |
7038 move_fence_to_fences (fences, new_fences); | |
7039 } | |
7040 } | |
7041 else | |
7042 extract_new_fences_from (fences, new_fences, orig_max_seqno); | |
7043 } | |
7044 | |
7045 flist_clear (&old_fences); | |
7046 return FLIST_TAIL_HEAD (new_fences); | |
7047 } | |
7048 | |
7049 /* Update seqnos of insns given by PSCHEDULED_INSNS. MIN_SEQNO and MAX_SEQNO | |
7050 are the miminum and maximum seqnos of the group, HIGHEST_SEQNO_IN_USE is | |
7051 the highest seqno used in a region. Return the updated highest seqno. */ | |
7052 static int | |
7053 update_seqnos_and_stage (int min_seqno, int max_seqno, | |
7054 int highest_seqno_in_use, | |
7055 ilist_t *pscheduled_insns) | |
7056 { | |
7057 int new_hs; | |
7058 ilist_iterator ii; | |
7059 insn_t insn; | |
7060 | |
7061 /* Actually, new_hs is the seqno of the instruction, that was | |
7062 scheduled first (i.e. it is the first one in SCHEDULED_INSNS). */ | |
7063 if (*pscheduled_insns) | |
7064 { | |
7065 new_hs = (INSN_SEQNO (ILIST_INSN (*pscheduled_insns)) | |
7066 + highest_seqno_in_use + max_seqno - min_seqno + 2); | |
7067 gcc_assert (new_hs > highest_seqno_in_use); | |
7068 } | |
7069 else | |
7070 new_hs = highest_seqno_in_use; | |
7071 | |
7072 FOR_EACH_INSN (insn, ii, *pscheduled_insns) | |
7073 { | |
7074 gcc_assert (INSN_SEQNO (insn) < 0); | |
7075 INSN_SEQNO (insn) += highest_seqno_in_use + max_seqno - min_seqno + 2; | |
7076 gcc_assert (INSN_SEQNO (insn) <= new_hs); | |
7077 } | |
7078 | |
7079 ilist_clear (pscheduled_insns); | |
7080 global_level++; | |
7081 | |
7082 return new_hs; | |
7083 } | |
7084 | |
7085 /* The main driver for scheduling a region. This function is responsible | |
7086 for correct propagation of fences (i.e. scheduling points) and creating | |
7087 a group of parallel insns at each of them. It also supports | |
7088 pipelining. ORIG_MAX_SEQNO is the maximal seqno before this pass | |
7089 of scheduling. */ | |
7090 static void | |
7091 sel_sched_region_2 (int orig_max_seqno) | |
7092 { | |
7093 int highest_seqno_in_use = orig_max_seqno; | |
7094 | |
7095 stat_bookkeeping_copies = 0; | |
7096 stat_insns_needed_bookkeeping = 0; | |
7097 stat_renamed_scheduled = 0; | |
7098 stat_substitutions_total = 0; | |
7099 num_insns_scheduled = 0; | |
7100 | |
7101 while (fences) | |
7102 { | |
7103 int min_seqno, max_seqno; | |
7104 ilist_t scheduled_insns = NULL; | |
7105 ilist_t *scheduled_insns_tailp = &scheduled_insns; | |
7106 | |
7107 find_min_max_seqno (fences, &min_seqno, &max_seqno); | |
7108 schedule_on_fences (fences, max_seqno, &scheduled_insns_tailp); | |
7109 fences = calculate_new_fences (fences, orig_max_seqno); | |
7110 highest_seqno_in_use = update_seqnos_and_stage (min_seqno, max_seqno, | |
7111 highest_seqno_in_use, | |
7112 &scheduled_insns); | |
7113 } | |
7114 | |
7115 if (sched_verbose >= 1) | |
7116 sel_print ("Scheduled %d bookkeeping copies, %d insns needed " | |
7117 "bookkeeping, %d insns renamed, %d insns substituted\n", | |
7118 stat_bookkeeping_copies, | |
7119 stat_insns_needed_bookkeeping, | |
7120 stat_renamed_scheduled, | |
7121 stat_substitutions_total); | |
7122 } | |
7123 | |
7124 /* Schedule a region. When pipelining, search for possibly never scheduled | |
7125 bookkeeping code and schedule it. Reschedule pipelined code without | |
7126 pipelining after. */ | |
7127 static void | |
7128 sel_sched_region_1 (void) | |
7129 { | |
7130 int number_of_insns; | |
7131 int orig_max_seqno; | |
7132 | |
7133 /* Remove empty blocks that might be in the region from the beginning. | |
7134 We need to do save sched_max_luid before that, as it actually shows | |
7135 the number of insns in the region, and purge_empty_blocks can | |
7136 alter it. */ | |
7137 number_of_insns = sched_max_luid - 1; | |
7138 purge_empty_blocks (); | |
7139 | |
7140 orig_max_seqno = init_seqno (number_of_insns, NULL, NULL); | |
7141 gcc_assert (orig_max_seqno >= 1); | |
7142 | |
7143 /* When pipelining outer loops, create fences on the loop header, | |
7144 not preheader. */ | |
7145 fences = NULL; | |
7146 if (current_loop_nest) | |
7147 init_fences (BB_END (EBB_FIRST_BB (0))); | |
7148 else | |
7149 init_fences (bb_note (EBB_FIRST_BB (0))); | |
7150 global_level = 1; | |
7151 | |
7152 sel_sched_region_2 (orig_max_seqno); | |
7153 | |
7154 gcc_assert (fences == NULL); | |
7155 | |
7156 if (pipelining_p) | |
7157 { | |
7158 int i; | |
7159 basic_block bb; | |
7160 struct flist_tail_def _new_fences; | |
7161 flist_tail_t new_fences = &_new_fences; | |
7162 bool do_p = true; | |
7163 | |
7164 pipelining_p = false; | |
7165 max_ws = MIN (max_ws, issue_rate * 3 / 2); | |
7166 bookkeeping_p = false; | |
7167 enable_schedule_as_rhs_p = false; | |
7168 | |
7169 /* Schedule newly created code, that has not been scheduled yet. */ | |
7170 do_p = true; | |
7171 | |
7172 while (do_p) | |
7173 { | |
7174 do_p = false; | |
7175 | |
7176 for (i = 0; i < current_nr_blocks; i++) | |
7177 { | |
7178 basic_block bb = EBB_FIRST_BB (i); | |
7179 | |
7180 if (sel_bb_empty_p (bb)) | |
7181 { | |
7182 bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
7183 continue; | |
7184 } | |
7185 | |
7186 if (bitmap_bit_p (blocks_to_reschedule, bb->index)) | |
7187 { | |
7188 clear_outdated_rtx_info (bb); | |
7189 if (sel_insn_is_speculation_check (BB_END (bb)) | |
7190 && JUMP_P (BB_END (bb))) | |
7191 bitmap_set_bit (blocks_to_reschedule, | |
7192 BRANCH_EDGE (bb)->dest->index); | |
7193 } | |
7194 else if (INSN_SCHED_TIMES (sel_bb_head (bb)) <= 0) | |
7195 bitmap_set_bit (blocks_to_reschedule, bb->index); | |
7196 } | |
7197 | |
7198 for (i = 0; i < current_nr_blocks; i++) | |
7199 { | |
7200 bb = EBB_FIRST_BB (i); | |
7201 | |
7202 /* While pipelining outer loops, skip bundling for loop | |
7203 preheaders. Those will be rescheduled in the outer | |
7204 loop. */ | |
7205 if (sel_is_loop_preheader_p (bb)) | |
7206 { | |
7207 clear_outdated_rtx_info (bb); | |
7208 continue; | |
7209 } | |
7210 | |
7211 if (bitmap_bit_p (blocks_to_reschedule, bb->index)) | |
7212 { | |
7213 flist_tail_init (new_fences); | |
7214 | |
7215 orig_max_seqno = init_seqno (0, blocks_to_reschedule, bb); | |
7216 | |
7217 /* Mark BB as head of the new ebb. */ | |
7218 bitmap_set_bit (forced_ebb_heads, bb->index); | |
7219 | |
7220 bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
7221 | |
7222 gcc_assert (fences == NULL); | |
7223 | |
7224 init_fences (bb_note (bb)); | |
7225 | |
7226 sel_sched_region_2 (orig_max_seqno); | |
7227 | |
7228 do_p = true; | |
7229 break; | |
7230 } | |
7231 } | |
7232 } | |
7233 } | |
7234 } | |
7235 | |
7236 /* Schedule the RGN region. */ | |
7237 void | |
7238 sel_sched_region (int rgn) | |
7239 { | |
7240 bool schedule_p; | |
7241 bool reset_sched_cycles_p; | |
7242 | |
7243 if (sel_region_init (rgn)) | |
7244 return; | |
7245 | |
7246 if (sched_verbose >= 1) | |
7247 sel_print ("Scheduling region %d\n", rgn); | |
7248 | |
7249 schedule_p = (!sched_is_disabled_for_current_region_p () | |
7250 && dbg_cnt (sel_sched_region_cnt)); | |
7251 reset_sched_cycles_p = pipelining_p; | |
7252 if (schedule_p) | |
7253 sel_sched_region_1 (); | |
7254 else | |
7255 /* Force initialization of INSN_SCHED_CYCLEs for correct bundling. */ | |
7256 reset_sched_cycles_p = true; | |
7257 | |
7258 sel_region_finish (reset_sched_cycles_p); | |
7259 } | |
7260 | |
7261 /* Perform global init for the scheduler. */ | |
7262 static void | |
7263 sel_global_init (void) | |
7264 { | |
7265 calculate_dominance_info (CDI_DOMINATORS); | |
7266 alloc_sched_pools (); | |
7267 | |
7268 /* Setup the infos for sched_init. */ | |
7269 sel_setup_sched_infos (); | |
7270 setup_sched_dump (); | |
7271 | |
7272 sched_rgn_init (false); | |
7273 sched_init (); | |
7274 | |
7275 sched_init_bbs (); | |
7276 /* Reset AFTER_RECOVERY if it has been set by the 1st scheduler pass. */ | |
7277 after_recovery = 0; | |
7278 can_issue_more = issue_rate; | |
7279 | |
7280 sched_extend_target (); | |
7281 sched_deps_init (true); | |
7282 setup_nop_and_exit_insns (); | |
7283 sel_extend_global_bb_info (); | |
7284 init_lv_sets (); | |
7285 init_hard_regs_data (); | |
7286 } | |
7287 | |
7288 /* Free the global data of the scheduler. */ | |
7289 static void | |
7290 sel_global_finish (void) | |
7291 { | |
7292 free_bb_note_pool (); | |
7293 free_lv_sets (); | |
7294 sel_finish_global_bb_info (); | |
7295 | |
7296 free_regset_pool (); | |
7297 free_nop_and_exit_insns (); | |
7298 | |
7299 sched_rgn_finish (); | |
7300 sched_deps_finish (); | |
7301 sched_finish (); | |
7302 | |
7303 if (current_loops) | |
7304 sel_finish_pipelining (); | |
7305 | |
7306 free_sched_pools (); | |
7307 free_dominance_info (CDI_DOMINATORS); | |
7308 } | |
7309 | |
7310 /* Return true when we need to skip selective scheduling. Used for debugging. */ | |
7311 bool | |
7312 maybe_skip_selective_scheduling (void) | |
7313 { | |
7314 return ! dbg_cnt (sel_sched_cnt); | |
7315 } | |
7316 | |
7317 /* The entry point. */ | |
7318 void | |
7319 run_selective_scheduling (void) | |
7320 { | |
7321 int rgn; | |
7322 | |
7323 if (n_basic_blocks == NUM_FIXED_BLOCKS) | |
7324 return; | |
7325 | |
7326 sel_global_init (); | |
7327 | |
7328 for (rgn = 0; rgn < nr_regions; rgn++) | |
7329 sel_sched_region (rgn); | |
7330 | |
7331 sel_global_finish (); | |
7332 } | |
7333 | |
7334 #endif |