]>
Commit | Line | Data |
---|---|---|
e855c69d | 1 | /* Instruction scheduling pass. Selective scheduler and pipeliner. |
66647d44 | 2 | Copyright (C) 2006, 2007, 2008, 2009 Free Software Foundation, Inc. |
e855c69d AB |
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 | \f | |
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 | \f | |
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. */ | |
c32e2175 | 451 | extern struct code_motion_path_driver_info_def move_op_hooks, fur_hooks; |
e855c69d AB |
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 | \f | |
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); | |
b5b8b0ac | 560 | static basic_block find_block_for_bookkeeping (edge e1, edge e2, bool lax); |
e855c69d AB |
561 | |
562 | static rtx get_dest_from_orig_ops (av_set_t); | |
563 | static basic_block generate_bookkeeping_insn (expr_t, edge, edge); | |
564 | static bool find_used_regs (insn_t, av_set_t, regset, struct reg_rename *, | |
565 | def_list_t *); | |
72a54528 AM |
566 | static bool move_op (insn_t, av_set_t, expr_t, rtx, expr_t, bool*); |
567 | static int code_motion_path_driver (insn_t, av_set_t, ilist_t, | |
568 | cmpd_local_params_p, void *); | |
e855c69d AB |
569 | static void sel_sched_region_1 (void); |
570 | static void sel_sched_region_2 (int); | |
571 | static av_set_t compute_av_set_inside_bb (insn_t, ilist_t, int, bool); | |
572 | ||
573 | static void debug_state (state_t); | |
574 | \f | |
575 | ||
576 | /* Functions that work with fences. */ | |
577 | ||
578 | /* Advance one cycle on FENCE. */ | |
579 | static void | |
580 | advance_one_cycle (fence_t fence) | |
581 | { | |
582 | unsigned i; | |
583 | int cycle; | |
584 | rtx insn; | |
585 | ||
586 | advance_state (FENCE_STATE (fence)); | |
587 | cycle = ++FENCE_CYCLE (fence); | |
588 | FENCE_ISSUED_INSNS (fence) = 0; | |
589 | FENCE_STARTS_CYCLE_P (fence) = 1; | |
590 | can_issue_more = issue_rate; | |
591 | ||
592 | for (i = 0; VEC_iterate (rtx, FENCE_EXECUTING_INSNS (fence), i, insn); ) | |
593 | { | |
594 | if (INSN_READY_CYCLE (insn) < cycle) | |
595 | { | |
596 | remove_from_deps (FENCE_DC (fence), insn); | |
597 | VEC_unordered_remove (rtx, FENCE_EXECUTING_INSNS (fence), i); | |
598 | continue; | |
599 | } | |
600 | i++; | |
601 | } | |
602 | if (sched_verbose >= 2) | |
603 | { | |
604 | sel_print ("Finished a cycle. Current cycle = %d\n", FENCE_CYCLE (fence)); | |
605 | debug_state (FENCE_STATE (fence)); | |
606 | } | |
607 | } | |
608 | ||
609 | /* Returns true when SUCC in a fallthru bb of INSN, possibly | |
610 | skipping empty basic blocks. */ | |
611 | static bool | |
612 | in_fallthru_bb_p (rtx insn, rtx succ) | |
613 | { | |
614 | basic_block bb = BLOCK_FOR_INSN (insn); | |
615 | ||
616 | if (bb == BLOCK_FOR_INSN (succ)) | |
617 | return true; | |
618 | ||
619 | if (find_fallthru_edge (bb)) | |
620 | bb = find_fallthru_edge (bb)->dest; | |
621 | else | |
622 | return false; | |
623 | ||
624 | while (sel_bb_empty_p (bb)) | |
625 | bb = bb->next_bb; | |
626 | ||
627 | return bb == BLOCK_FOR_INSN (succ); | |
628 | } | |
629 | ||
630 | /* Construct successor fences from OLD_FENCEs and put them in NEW_FENCES. | |
631 | When a successor will continue a ebb, transfer all parameters of a fence | |
632 | to the new fence. ORIG_MAX_SEQNO is the maximal seqno before this round | |
633 | of scheduling helping to distinguish between the old and the new code. */ | |
634 | static void | |
635 | extract_new_fences_from (flist_t old_fences, flist_tail_t new_fences, | |
636 | int orig_max_seqno) | |
637 | { | |
638 | bool was_here_p = false; | |
639 | insn_t insn = NULL_RTX; | |
640 | insn_t succ; | |
641 | succ_iterator si; | |
642 | ilist_iterator ii; | |
643 | fence_t fence = FLIST_FENCE (old_fences); | |
644 | basic_block bb; | |
645 | ||
646 | /* Get the only element of FENCE_BNDS (fence). */ | |
647 | FOR_EACH_INSN (insn, ii, FENCE_BNDS (fence)) | |
648 | { | |
649 | gcc_assert (!was_here_p); | |
650 | was_here_p = true; | |
651 | } | |
652 | gcc_assert (was_here_p && insn != NULL_RTX); | |
653 | ||
654 | /* When in the "middle" of the block, just move this fence | |
655 | to the new list. */ | |
656 | bb = BLOCK_FOR_INSN (insn); | |
657 | if (! sel_bb_end_p (insn) | |
658 | || (single_succ_p (bb) | |
659 | && single_pred_p (single_succ (bb)))) | |
660 | { | |
661 | insn_t succ; | |
662 | ||
663 | succ = (sel_bb_end_p (insn) | |
664 | ? sel_bb_head (single_succ (bb)) | |
665 | : NEXT_INSN (insn)); | |
666 | ||
667 | if (INSN_SEQNO (succ) > 0 | |
668 | && INSN_SEQNO (succ) <= orig_max_seqno | |
669 | && INSN_SCHED_TIMES (succ) <= 0) | |
670 | { | |
671 | FENCE_INSN (fence) = succ; | |
672 | move_fence_to_fences (old_fences, new_fences); | |
673 | ||
674 | if (sched_verbose >= 1) | |
675 | sel_print ("Fence %d continues as %d[%d] (state continue)\n", | |
676 | INSN_UID (insn), INSN_UID (succ), BLOCK_NUM (succ)); | |
677 | } | |
678 | return; | |
679 | } | |
680 | ||
681 | /* Otherwise copy fence's structures to (possibly) multiple successors. */ | |
682 | FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
683 | { | |
684 | int seqno = INSN_SEQNO (succ); | |
685 | ||
686 | if (0 < seqno && seqno <= orig_max_seqno | |
687 | && (pipelining_p || INSN_SCHED_TIMES (succ) <= 0)) | |
688 | { | |
689 | bool b = (in_same_ebb_p (insn, succ) | |
690 | || in_fallthru_bb_p (insn, succ)); | |
691 | ||
692 | if (sched_verbose >= 1) | |
693 | sel_print ("Fence %d continues as %d[%d] (state %s)\n", | |
694 | INSN_UID (insn), INSN_UID (succ), | |
695 | BLOCK_NUM (succ), b ? "continue" : "reset"); | |
696 | ||
697 | if (b) | |
698 | add_dirty_fence_to_fences (new_fences, succ, fence); | |
699 | else | |
700 | { | |
701 | /* Mark block of the SUCC as head of the new ebb. */ | |
702 | bitmap_set_bit (forced_ebb_heads, BLOCK_NUM (succ)); | |
703 | add_clean_fence_to_fences (new_fences, succ, fence); | |
704 | } | |
705 | } | |
706 | } | |
707 | } | |
708 | \f | |
709 | ||
710 | /* Functions to support substitution. */ | |
711 | ||
712 | /* Returns whether INSN with dependence status DS is eligible for | |
713 | substitution, i.e. it's a copy operation x := y, and RHS that is | |
714 | moved up through this insn should be substituted. */ | |
715 | static bool | |
716 | can_substitute_through_p (insn_t insn, ds_t ds) | |
717 | { | |
718 | /* We can substitute only true dependencies. */ | |
719 | if ((ds & DEP_OUTPUT) | |
720 | || (ds & DEP_ANTI) | |
721 | || ! INSN_RHS (insn) | |
722 | || ! INSN_LHS (insn)) | |
723 | return false; | |
724 | ||
725 | /* Now we just need to make sure the INSN_RHS consists of only one | |
726 | simple REG rtx. */ | |
727 | if (REG_P (INSN_LHS (insn)) | |
728 | && REG_P (INSN_RHS (insn))) | |
729 | return true; | |
730 | return false; | |
731 | } | |
732 | ||
733 | /* Substitute all occurences of INSN's destination in EXPR' vinsn with INSN's | |
734 | source (if INSN is eligible for substitution). Returns TRUE if | |
735 | substitution was actually performed, FALSE otherwise. Substitution might | |
736 | be not performed because it's either EXPR' vinsn doesn't contain INSN's | |
737 | destination or the resulting insn is invalid for the target machine. | |
738 | When UNDO is true, perform unsubstitution instead (the difference is in | |
739 | the part of rtx on which validate_replace_rtx is called). */ | |
740 | static bool | |
741 | substitute_reg_in_expr (expr_t expr, insn_t insn, bool undo) | |
742 | { | |
743 | rtx *where; | |
744 | bool new_insn_valid; | |
745 | vinsn_t *vi = &EXPR_VINSN (expr); | |
746 | bool has_rhs = VINSN_RHS (*vi) != NULL; | |
747 | rtx old, new_rtx; | |
748 | ||
749 | /* Do not try to replace in SET_DEST. Although we'll choose new | |
750 | register for the RHS, we don't want to change RHS' original reg. | |
751 | If the insn is not SET, we may still be able to substitute something | |
752 | in it, and if we're here (don't have deps), it doesn't write INSN's | |
753 | dest. */ | |
754 | where = (has_rhs | |
755 | ? &VINSN_RHS (*vi) | |
756 | : &PATTERN (VINSN_INSN_RTX (*vi))); | |
757 | old = undo ? INSN_RHS (insn) : INSN_LHS (insn); | |
758 | ||
759 | /* Substitute if INSN has a form of x:=y and LHS(INSN) occurs in *VI. */ | |
760 | if (rtx_ok_for_substitution_p (old, *where)) | |
761 | { | |
762 | rtx new_insn; | |
763 | rtx *where_replace; | |
764 | ||
765 | /* We should copy these rtxes before substitution. */ | |
766 | new_rtx = copy_rtx (undo ? INSN_LHS (insn) : INSN_RHS (insn)); | |
767 | new_insn = create_copy_of_insn_rtx (VINSN_INSN_RTX (*vi)); | |
768 | ||
769 | /* Where we'll replace. | |
770 | WHERE_REPLACE should point inside NEW_INSN, so INSN_RHS couldn't be | |
771 | used instead of SET_SRC. */ | |
772 | where_replace = (has_rhs | |
773 | ? &SET_SRC (PATTERN (new_insn)) | |
774 | : &PATTERN (new_insn)); | |
775 | ||
776 | new_insn_valid | |
777 | = validate_replace_rtx_part_nosimplify (old, new_rtx, where_replace, | |
778 | new_insn); | |
779 | ||
780 | /* ??? Actually, constrain_operands result depends upon choice of | |
781 | destination register. E.g. if we allow single register to be an rhs, | |
782 | and if we try to move dx=ax(as rhs) through ax=dx, we'll result | |
783 | in invalid insn dx=dx, so we'll loose this rhs here. | |
784 | Just can't come up with significant testcase for this, so just | |
785 | leaving it for now. */ | |
786 | if (new_insn_valid) | |
787 | { | |
788 | change_vinsn_in_expr (expr, | |
789 | create_vinsn_from_insn_rtx (new_insn, false)); | |
790 | ||
791 | /* Do not allow clobbering the address register of speculative | |
792 | insns. */ | |
793 | if ((EXPR_SPEC_DONE_DS (expr) & SPECULATIVE) | |
794 | && bitmap_bit_p (VINSN_REG_USES (EXPR_VINSN (expr)), | |
795 | expr_dest_regno (expr))) | |
796 | EXPR_TARGET_AVAILABLE (expr) = false; | |
797 | ||
798 | return true; | |
799 | } | |
800 | else | |
801 | return false; | |
802 | } | |
803 | else | |
804 | return false; | |
805 | } | |
806 | ||
807 | /* Helper function for count_occurences_equiv. */ | |
808 | static int | |
809 | count_occurrences_1 (rtx *cur_rtx, void *arg) | |
810 | { | |
811 | rtx_search_arg_p p = (rtx_search_arg_p) arg; | |
812 | ||
813 | /* The last param FOR_GCSE is true, because otherwise it performs excessive | |
814 | substitutions like | |
815 | r8 = r33 | |
816 | r16 = r33 | |
817 | for the last insn it presumes r33 equivalent to r8, so it changes it to | |
818 | r33. Actually, there's no change, but it spoils debugging. */ | |
819 | if (exp_equiv_p (*cur_rtx, p->x, 0, true)) | |
820 | { | |
821 | /* Bail out if we occupy more than one register. */ | |
822 | if (REG_P (*cur_rtx) | |
72a54528 | 823 | && HARD_REGISTER_P (*cur_rtx) |
e855c69d AB |
824 | && hard_regno_nregs[REGNO(*cur_rtx)][GET_MODE (*cur_rtx)] > 1) |
825 | { | |
826 | p->n = 0; | |
827 | return 1; | |
828 | } | |
829 | ||
830 | p->n++; | |
831 | ||
832 | /* Do not traverse subexprs. */ | |
833 | return -1; | |
834 | } | |
835 | ||
836 | if (GET_CODE (*cur_rtx) == SUBREG | |
837 | && REG_P (p->x) | |
838 | && REGNO (SUBREG_REG (*cur_rtx)) == REGNO (p->x)) | |
839 | { | |
840 | /* ??? Do not support substituting regs inside subregs. In that case, | |
841 | simplify_subreg will be called by validate_replace_rtx, and | |
842 | unsubstitution will fail later. */ | |
843 | p->n = 0; | |
844 | return 1; | |
845 | } | |
846 | ||
847 | /* Continue search. */ | |
848 | return 0; | |
849 | } | |
850 | ||
851 | /* Return the number of places WHAT appears within WHERE. | |
852 | Bail out when we found a reference occupying several hard registers. */ | |
853 | static int | |
854 | count_occurrences_equiv (rtx what, rtx where) | |
855 | { | |
856 | struct rtx_search_arg arg; | |
857 | ||
858 | arg.x = what; | |
859 | arg.n = 0; | |
860 | ||
861 | for_each_rtx (&where, &count_occurrences_1, (void *) &arg); | |
862 | ||
863 | return arg.n; | |
864 | } | |
865 | ||
866 | /* Returns TRUE if WHAT is found in WHERE rtx tree. */ | |
867 | static bool | |
868 | rtx_ok_for_substitution_p (rtx what, rtx where) | |
869 | { | |
870 | return (count_occurrences_equiv (what, where) > 0); | |
871 | } | |
872 | \f | |
873 | ||
874 | /* Functions to support register renaming. */ | |
875 | ||
876 | /* Substitute VI's set source with REGNO. Returns newly created pattern | |
877 | that has REGNO as its source. */ | |
878 | static rtx | |
879 | create_insn_rtx_with_rhs (vinsn_t vi, rtx rhs_rtx) | |
880 | { | |
881 | rtx lhs_rtx; | |
882 | rtx pattern; | |
883 | rtx insn_rtx; | |
884 | ||
885 | lhs_rtx = copy_rtx (VINSN_LHS (vi)); | |
886 | ||
887 | pattern = gen_rtx_SET (VOIDmode, lhs_rtx, rhs_rtx); | |
888 | insn_rtx = create_insn_rtx_from_pattern (pattern, NULL_RTX); | |
889 | ||
890 | return insn_rtx; | |
891 | } | |
892 | ||
893 | /* Returns whether INSN's src can be replaced with register number | |
894 | NEW_SRC_REG. E.g. the following insn is valid for i386: | |
895 | ||
896 | (insn:HI 2205 6585 2207 727 ../../gcc/libiberty/regex.c:3337 | |
897 | (set (mem/s:QI (plus:SI (plus:SI (reg/f:SI 7 sp) | |
898 | (reg:SI 0 ax [orig:770 c1 ] [770])) | |
899 | (const_int 288 [0x120])) [0 str S1 A8]) | |
900 | (const_int 0 [0x0])) 43 {*movqi_1} (nil) | |
901 | (nil)) | |
902 | ||
903 | But if we change (const_int 0 [0x0]) to (reg:QI 4 si), it will be invalid | |
904 | because of operand constraints: | |
905 | ||
906 | (define_insn "*movqi_1" | |
907 | [(set (match_operand:QI 0 "nonimmediate_operand" "=q,q ,q ,r,r ,?r,m") | |
908 | (match_operand:QI 1 "general_operand" " q,qn,qm,q,rn,qm,qn") | |
909 | )] | |
910 | ||
911 | So do constrain_operands here, before choosing NEW_SRC_REG as best | |
912 | reg for rhs. */ | |
913 | ||
914 | static bool | |
915 | replace_src_with_reg_ok_p (insn_t insn, rtx new_src_reg) | |
916 | { | |
917 | vinsn_t vi = INSN_VINSN (insn); | |
918 | enum machine_mode mode; | |
919 | rtx dst_loc; | |
920 | bool res; | |
921 | ||
922 | gcc_assert (VINSN_SEPARABLE_P (vi)); | |
923 | ||
924 | get_dest_and_mode (insn, &dst_loc, &mode); | |
925 | gcc_assert (mode == GET_MODE (new_src_reg)); | |
926 | ||
927 | if (REG_P (dst_loc) && REGNO (new_src_reg) == REGNO (dst_loc)) | |
928 | return true; | |
929 | ||
930 | /* See whether SET_SRC can be replaced with this register. */ | |
931 | validate_change (insn, &SET_SRC (PATTERN (insn)), new_src_reg, 1); | |
932 | res = verify_changes (0); | |
933 | cancel_changes (0); | |
934 | ||
935 | return res; | |
936 | } | |
937 | ||
938 | /* Returns whether INSN still be valid after replacing it's DEST with | |
939 | register NEW_REG. */ | |
940 | static bool | |
941 | replace_dest_with_reg_ok_p (insn_t insn, rtx new_reg) | |
942 | { | |
943 | vinsn_t vi = INSN_VINSN (insn); | |
944 | bool res; | |
945 | ||
946 | /* We should deal here only with separable insns. */ | |
947 | gcc_assert (VINSN_SEPARABLE_P (vi)); | |
948 | gcc_assert (GET_MODE (VINSN_LHS (vi)) == GET_MODE (new_reg)); | |
949 | ||
950 | /* See whether SET_DEST can be replaced with this register. */ | |
951 | validate_change (insn, &SET_DEST (PATTERN (insn)), new_reg, 1); | |
952 | res = verify_changes (0); | |
953 | cancel_changes (0); | |
954 | ||
955 | return res; | |
956 | } | |
957 | ||
958 | /* Create a pattern with rhs of VI and lhs of LHS_RTX. */ | |
959 | static rtx | |
960 | create_insn_rtx_with_lhs (vinsn_t vi, rtx lhs_rtx) | |
961 | { | |
962 | rtx rhs_rtx; | |
963 | rtx pattern; | |
964 | rtx insn_rtx; | |
965 | ||
966 | rhs_rtx = copy_rtx (VINSN_RHS (vi)); | |
967 | ||
968 | pattern = gen_rtx_SET (VOIDmode, lhs_rtx, rhs_rtx); | |
969 | insn_rtx = create_insn_rtx_from_pattern (pattern, NULL_RTX); | |
970 | ||
971 | return insn_rtx; | |
972 | } | |
973 | ||
974 | /* Substitute lhs in the given expression EXPR for the register with number | |
975 | NEW_REGNO. SET_DEST may be arbitrary rtx, not only register. */ | |
976 | static void | |
977 | replace_dest_with_reg_in_expr (expr_t expr, rtx new_reg) | |
978 | { | |
979 | rtx insn_rtx; | |
980 | vinsn_t vinsn; | |
981 | ||
982 | insn_rtx = create_insn_rtx_with_lhs (EXPR_VINSN (expr), new_reg); | |
983 | vinsn = create_vinsn_from_insn_rtx (insn_rtx, false); | |
984 | ||
985 | change_vinsn_in_expr (expr, vinsn); | |
986 | EXPR_WAS_RENAMED (expr) = 1; | |
987 | EXPR_TARGET_AVAILABLE (expr) = 1; | |
988 | } | |
989 | ||
990 | /* Returns whether VI writes either one of the USED_REGS registers or, | |
991 | if a register is a hard one, one of the UNAVAILABLE_HARD_REGS registers. */ | |
992 | static bool | |
993 | vinsn_writes_one_of_regs_p (vinsn_t vi, regset used_regs, | |
994 | HARD_REG_SET unavailable_hard_regs) | |
995 | { | |
996 | unsigned regno; | |
997 | reg_set_iterator rsi; | |
998 | ||
999 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (vi), 0, regno, rsi) | |
1000 | { | |
1001 | if (REGNO_REG_SET_P (used_regs, regno)) | |
1002 | return true; | |
1003 | if (HARD_REGISTER_NUM_P (regno) | |
1004 | && TEST_HARD_REG_BIT (unavailable_hard_regs, regno)) | |
1005 | return true; | |
1006 | } | |
1007 | ||
1008 | EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (vi), 0, regno, rsi) | |
1009 | { | |
1010 | if (REGNO_REG_SET_P (used_regs, regno)) | |
1011 | return true; | |
1012 | if (HARD_REGISTER_NUM_P (regno) | |
1013 | && TEST_HARD_REG_BIT (unavailable_hard_regs, regno)) | |
1014 | return true; | |
1015 | } | |
1016 | ||
1017 | return false; | |
1018 | } | |
1019 | ||
1020 | /* Returns register class of the output register in INSN. | |
1021 | Returns NO_REGS for call insns because some targets have constraints on | |
1022 | destination register of a call insn. | |
1023 | ||
1024 | Code adopted from regrename.c::build_def_use. */ | |
1025 | static enum reg_class | |
1026 | get_reg_class (rtx insn) | |
1027 | { | |
1028 | int alt, i, n_ops; | |
1029 | ||
1030 | extract_insn (insn); | |
1031 | if (! constrain_operands (1)) | |
1032 | fatal_insn_not_found (insn); | |
1033 | preprocess_constraints (); | |
1034 | alt = which_alternative; | |
1035 | n_ops = recog_data.n_operands; | |
1036 | ||
1037 | for (i = 0; i < n_ops; ++i) | |
1038 | { | |
1039 | int matches = recog_op_alt[i][alt].matches; | |
1040 | if (matches >= 0) | |
1041 | recog_op_alt[i][alt].cl = recog_op_alt[matches][alt].cl; | |
1042 | } | |
1043 | ||
1044 | if (asm_noperands (PATTERN (insn)) > 0) | |
1045 | { | |
1046 | for (i = 0; i < n_ops; i++) | |
1047 | if (recog_data.operand_type[i] == OP_OUT) | |
1048 | { | |
1049 | rtx *loc = recog_data.operand_loc[i]; | |
1050 | rtx op = *loc; | |
1051 | enum reg_class cl = recog_op_alt[i][alt].cl; | |
1052 | ||
1053 | if (REG_P (op) | |
1054 | && REGNO (op) == ORIGINAL_REGNO (op)) | |
1055 | continue; | |
1056 | ||
1057 | return cl; | |
1058 | } | |
1059 | } | |
1060 | else if (!CALL_P (insn)) | |
1061 | { | |
1062 | for (i = 0; i < n_ops + recog_data.n_dups; i++) | |
1063 | { | |
1064 | int opn = i < n_ops ? i : recog_data.dup_num[i - n_ops]; | |
1065 | enum reg_class cl = recog_op_alt[opn][alt].cl; | |
1066 | ||
1067 | if (recog_data.operand_type[opn] == OP_OUT || | |
1068 | recog_data.operand_type[opn] == OP_INOUT) | |
1069 | return cl; | |
1070 | } | |
1071 | } | |
1072 | ||
1073 | /* Insns like | |
1074 | (insn (set (reg:CCZ 17 flags) (compare:CCZ ...))) | |
1075 | may result in returning NO_REGS, cause flags is written implicitly through | |
1076 | CMP insn, which has no OP_OUT | OP_INOUT operands. */ | |
1077 | return NO_REGS; | |
1078 | } | |
1079 | ||
1080 | #ifdef HARD_REGNO_RENAME_OK | |
1081 | /* Calculate HARD_REGNO_RENAME_OK data for REGNO. */ | |
1082 | static void | |
1083 | init_hard_regno_rename (int regno) | |
1084 | { | |
1085 | int cur_reg; | |
1086 | ||
1087 | SET_HARD_REG_BIT (sel_hrd.regs_for_rename[regno], regno); | |
1088 | ||
1089 | for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) | |
1090 | { | |
1091 | /* We are not interested in renaming in other regs. */ | |
1092 | if (!TEST_HARD_REG_BIT (sel_hrd.regs_ever_used, cur_reg)) | |
1093 | continue; | |
1094 | ||
1095 | if (HARD_REGNO_RENAME_OK (regno, cur_reg)) | |
1096 | SET_HARD_REG_BIT (sel_hrd.regs_for_rename[regno], cur_reg); | |
1097 | } | |
1098 | } | |
1099 | #endif | |
1100 | ||
1101 | /* A wrapper around HARD_REGNO_RENAME_OK that will look into the hard regs | |
1102 | data first. */ | |
1103 | static inline bool | |
a20d7130 | 1104 | sel_hard_regno_rename_ok (int from ATTRIBUTE_UNUSED, int to ATTRIBUTE_UNUSED) |
e855c69d AB |
1105 | { |
1106 | #ifdef HARD_REGNO_RENAME_OK | |
1107 | /* Check whether this is all calculated. */ | |
1108 | if (TEST_HARD_REG_BIT (sel_hrd.regs_for_rename[from], from)) | |
1109 | return TEST_HARD_REG_BIT (sel_hrd.regs_for_rename[from], to); | |
1110 | ||
1111 | init_hard_regno_rename (from); | |
1112 | ||
1113 | return TEST_HARD_REG_BIT (sel_hrd.regs_for_rename[from], to); | |
1114 | #else | |
1115 | return true; | |
1116 | #endif | |
1117 | } | |
1118 | ||
1119 | /* Calculate set of registers that are capable of holding MODE. */ | |
1120 | static void | |
1121 | init_regs_for_mode (enum machine_mode mode) | |
1122 | { | |
1123 | int cur_reg; | |
1124 | ||
1125 | CLEAR_HARD_REG_SET (sel_hrd.regs_for_mode[mode]); | |
1126 | CLEAR_HARD_REG_SET (sel_hrd.regs_for_call_clobbered[mode]); | |
1127 | ||
1128 | for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) | |
1129 | { | |
1130 | int nregs = hard_regno_nregs[cur_reg][mode]; | |
1131 | int i; | |
1132 | ||
1133 | for (i = nregs - 1; i >= 0; --i) | |
1134 | if (fixed_regs[cur_reg + i] | |
1135 | || global_regs[cur_reg + i] | |
1136 | /* Can't use regs which aren't saved by | |
1137 | the prologue. */ | |
1138 | || !TEST_HARD_REG_BIT (sel_hrd.regs_ever_used, cur_reg + i) | |
1139 | #ifdef LEAF_REGISTERS | |
1140 | /* We can't use a non-leaf register if we're in a | |
1141 | leaf function. */ | |
1142 | || (current_function_is_leaf | |
1143 | && !LEAF_REGISTERS[cur_reg + i]) | |
1144 | #endif | |
1145 | ) | |
1146 | break; | |
1147 | ||
1148 | if (i >= 0) | |
1149 | continue; | |
1150 | ||
1151 | /* See whether it accepts all modes that occur in | |
1152 | original insns. */ | |
1153 | if (! HARD_REGNO_MODE_OK (cur_reg, mode)) | |
1154 | continue; | |
1155 | ||
1156 | if (HARD_REGNO_CALL_PART_CLOBBERED (cur_reg, mode)) | |
1157 | SET_HARD_REG_BIT (sel_hrd.regs_for_call_clobbered[mode], | |
1158 | cur_reg); | |
1159 | ||
1160 | /* If the CUR_REG passed all the checks above, | |
1161 | then it's ok. */ | |
1162 | SET_HARD_REG_BIT (sel_hrd.regs_for_mode[mode], cur_reg); | |
1163 | } | |
1164 | ||
1165 | sel_hrd.regs_for_mode_ok[mode] = true; | |
1166 | } | |
1167 | ||
1168 | /* Init all register sets gathered in HRD. */ | |
1169 | static void | |
1170 | init_hard_regs_data (void) | |
1171 | { | |
1172 | int cur_reg = 0; | |
32e8bb8e | 1173 | int cur_mode = 0; |
e855c69d AB |
1174 | |
1175 | CLEAR_HARD_REG_SET (sel_hrd.regs_ever_used); | |
1176 | for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) | |
1177 | if (df_regs_ever_live_p (cur_reg) || call_used_regs[cur_reg]) | |
1178 | SET_HARD_REG_BIT (sel_hrd.regs_ever_used, cur_reg); | |
1179 | ||
1180 | /* Initialize registers that are valid based on mode when this is | |
1181 | really needed. */ | |
1182 | for (cur_mode = 0; cur_mode < NUM_MACHINE_MODES; cur_mode++) | |
1183 | sel_hrd.regs_for_mode_ok[cur_mode] = false; | |
1184 | ||
1185 | /* Mark that all HARD_REGNO_RENAME_OK is not calculated. */ | |
1186 | for (cur_reg = 0; cur_reg < FIRST_PSEUDO_REGISTER; cur_reg++) | |
1187 | CLEAR_HARD_REG_SET (sel_hrd.regs_for_rename[cur_reg]); | |
1188 | ||
1189 | #ifdef STACK_REGS | |
1190 | CLEAR_HARD_REG_SET (sel_hrd.stack_regs); | |
1191 | ||
1192 | for (cur_reg = FIRST_STACK_REG; cur_reg <= LAST_STACK_REG; cur_reg++) | |
1193 | SET_HARD_REG_BIT (sel_hrd.stack_regs, cur_reg); | |
1194 | #endif | |
1195 | } | |
1196 | ||
1197 | /* Mark hardware regs in REG_RENAME_P that are not suitable | |
1198 | for renaming rhs in INSN due to hardware restrictions (register class, | |
1199 | modes compatibility etc). This doesn't affect original insn's dest reg, | |
1200 | if it isn't in USED_REGS. DEF is a definition insn of rhs for which the | |
1201 | destination register is sought. LHS (DEF->ORIG_INSN) may be REG or MEM. | |
1202 | Registers that are in used_regs are always marked in | |
1203 | unavailable_hard_regs as well. */ | |
1204 | ||
1205 | static void | |
1206 | mark_unavailable_hard_regs (def_t def, struct reg_rename *reg_rename_p, | |
1207 | regset used_regs ATTRIBUTE_UNUSED) | |
1208 | { | |
1209 | enum machine_mode mode; | |
1210 | enum reg_class cl = NO_REGS; | |
1211 | rtx orig_dest; | |
1212 | unsigned cur_reg, regno; | |
1213 | hard_reg_set_iterator hrsi; | |
1214 | ||
1215 | gcc_assert (GET_CODE (PATTERN (def->orig_insn)) == SET); | |
1216 | gcc_assert (reg_rename_p); | |
1217 | ||
1218 | orig_dest = SET_DEST (PATTERN (def->orig_insn)); | |
1219 | ||
1220 | /* We have decided not to rename 'mem = something;' insns, as 'something' | |
1221 | is usually a register. */ | |
1222 | if (!REG_P (orig_dest)) | |
1223 | return; | |
1224 | ||
1225 | regno = REGNO (orig_dest); | |
1226 | ||
1227 | /* If before reload, don't try to work with pseudos. */ | |
1228 | if (!reload_completed && !HARD_REGISTER_NUM_P (regno)) | |
1229 | return; | |
1230 | ||
1231 | mode = GET_MODE (orig_dest); | |
1232 | ||
1233 | /* Stop when mode is not supported for renaming. Also can't proceed | |
1234 | if the original register is one of the fixed_regs, global_regs or | |
1235 | frame pointer. */ | |
1236 | if (fixed_regs[regno] | |
1237 | || global_regs[regno] | |
1238 | #if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM | |
1239 | || (frame_pointer_needed && regno == HARD_FRAME_POINTER_REGNUM) | |
1240 | #else | |
1241 | || (frame_pointer_needed && regno == FRAME_POINTER_REGNUM) | |
1242 | #endif | |
1243 | ) | |
1244 | { | |
1245 | SET_HARD_REG_SET (reg_rename_p->unavailable_hard_regs); | |
1246 | ||
1247 | /* Give a chance for original register, if it isn't in used_regs. */ | |
1248 | if (!def->crosses_call) | |
1249 | CLEAR_HARD_REG_BIT (reg_rename_p->unavailable_hard_regs, regno); | |
1250 | ||
1251 | return; | |
1252 | } | |
1253 | ||
1254 | /* If something allocated on stack in this function, mark frame pointer | |
1255 | register unavailable, considering also modes. | |
1256 | FIXME: it is enough to do this once per all original defs. */ | |
1257 | if (frame_pointer_needed) | |
1258 | { | |
1259 | int i; | |
1260 | ||
1261 | for (i = hard_regno_nregs[FRAME_POINTER_REGNUM][Pmode]; i--;) | |
1262 | SET_HARD_REG_BIT (reg_rename_p->unavailable_hard_regs, | |
1263 | FRAME_POINTER_REGNUM + i); | |
1264 | ||
1265 | #if FRAME_POINTER_REGNUM != HARD_FRAME_POINTER_REGNUM | |
1266 | for (i = hard_regno_nregs[HARD_FRAME_POINTER_REGNUM][Pmode]; i--;) | |
1267 | SET_HARD_REG_BIT (reg_rename_p->unavailable_hard_regs, | |
1268 | HARD_FRAME_POINTER_REGNUM + i); | |
1269 | #endif | |
1270 | } | |
1271 | ||
1272 | #ifdef STACK_REGS | |
1273 | /* For the stack registers the presence of FIRST_STACK_REG in USED_REGS | |
1274 | is equivalent to as if all stack regs were in this set. | |
1275 | I.e. no stack register can be renamed, and even if it's an original | |
1276 | register here we make sure it won't be lifted over it's previous def | |
1277 | (it's previous def will appear as if it's a FIRST_STACK_REG def. | |
1278 | The HARD_REGNO_RENAME_OK covers other cases in condition below. */ | |
1279 | if (IN_RANGE (REGNO (orig_dest), FIRST_STACK_REG, LAST_STACK_REG) | |
1280 | && REGNO_REG_SET_P (used_regs, FIRST_STACK_REG)) | |
1281 | IOR_HARD_REG_SET (reg_rename_p->unavailable_hard_regs, | |
1282 | sel_hrd.stack_regs); | |
1283 | #endif | |
1284 | ||
1285 | /* If there's a call on this path, make regs from call_used_reg_set | |
1286 | unavailable. */ | |
1287 | if (def->crosses_call) | |
1288 | IOR_HARD_REG_SET (reg_rename_p->unavailable_hard_regs, | |
1289 | call_used_reg_set); | |
1290 | ||
1291 | /* Stop here before reload: we need FRAME_REGS, STACK_REGS, and crosses_call, | |
1292 | but not register classes. */ | |
1293 | if (!reload_completed) | |
1294 | return; | |
1295 | ||
1296 | /* Leave regs as 'available' only from the current | |
1297 | register class. */ | |
1298 | cl = get_reg_class (def->orig_insn); | |
1299 | gcc_assert (cl != NO_REGS); | |
1300 | COPY_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1301 | reg_class_contents[cl]); | |
1302 | ||
1303 | /* Leave only registers available for this mode. */ | |
1304 | if (!sel_hrd.regs_for_mode_ok[mode]) | |
1305 | init_regs_for_mode (mode); | |
1306 | AND_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1307 | sel_hrd.regs_for_mode[mode]); | |
1308 | ||
1309 | /* Exclude registers that are partially call clobbered. */ | |
1310 | if (def->crosses_call | |
1311 | && ! HARD_REGNO_CALL_PART_CLOBBERED (regno, mode)) | |
1312 | AND_COMPL_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1313 | sel_hrd.regs_for_call_clobbered[mode]); | |
1314 | ||
1315 | /* Leave only those that are ok to rename. */ | |
1316 | EXECUTE_IF_SET_IN_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1317 | 0, cur_reg, hrsi) | |
1318 | { | |
1319 | int nregs; | |
1320 | int i; | |
1321 | ||
1322 | nregs = hard_regno_nregs[cur_reg][mode]; | |
1323 | gcc_assert (nregs > 0); | |
1324 | ||
1325 | for (i = nregs - 1; i >= 0; --i) | |
1326 | if (! sel_hard_regno_rename_ok (regno + i, cur_reg + i)) | |
1327 | break; | |
1328 | ||
1329 | if (i >= 0) | |
1330 | CLEAR_HARD_REG_BIT (reg_rename_p->available_for_renaming, | |
1331 | cur_reg); | |
1332 | } | |
1333 | ||
1334 | AND_COMPL_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1335 | reg_rename_p->unavailable_hard_regs); | |
1336 | ||
1337 | /* Regno is always ok from the renaming part of view, but it really | |
1338 | could be in *unavailable_hard_regs already, so set it here instead | |
1339 | of there. */ | |
1340 | SET_HARD_REG_BIT (reg_rename_p->available_for_renaming, regno); | |
1341 | } | |
1342 | ||
1343 | /* reg_rename_tick[REG1] > reg_rename_tick[REG2] if REG1 was chosen as the | |
1344 | best register more recently than REG2. */ | |
1345 | static int reg_rename_tick[FIRST_PSEUDO_REGISTER]; | |
1346 | ||
1347 | /* Indicates the number of times renaming happened before the current one. */ | |
1348 | static int reg_rename_this_tick; | |
1349 | ||
1350 | /* Choose the register among free, that is suitable for storing | |
1351 | the rhs value. | |
1352 | ||
1353 | ORIGINAL_INSNS is the list of insns where the operation (rhs) | |
1354 | originally appears. There could be multiple original operations | |
1355 | for single rhs since we moving it up and merging along different | |
1356 | paths. | |
1357 | ||
1358 | Some code is adapted from regrename.c (regrename_optimize). | |
1359 | If original register is available, function returns it. | |
1360 | Otherwise it performs the checks, so the new register should | |
1361 | comply with the following: | |
1362 | - it should not violate any live ranges (such registers are in | |
1363 | REG_RENAME_P->available_for_renaming set); | |
1364 | - it should not be in the HARD_REGS_USED regset; | |
1365 | - it should be in the class compatible with original uses; | |
1366 | - it should not be clobbered through reference with different mode; | |
1367 | - if we're in the leaf function, then the new register should | |
1368 | not be in the LEAF_REGISTERS; | |
1369 | - etc. | |
1370 | ||
1371 | If several registers meet the conditions, the register with smallest | |
1372 | tick is returned to achieve more even register allocation. | |
1373 | ||
1374 | If original register seems to be ok, we set *IS_ORIG_REG_P_PTR to true. | |
1375 | ||
1376 | If no register satisfies the above conditions, NULL_RTX is returned. */ | |
1377 | static rtx | |
1378 | choose_best_reg_1 (HARD_REG_SET hard_regs_used, | |
1379 | struct reg_rename *reg_rename_p, | |
1380 | def_list_t original_insns, bool *is_orig_reg_p_ptr) | |
1381 | { | |
1382 | int best_new_reg; | |
1383 | unsigned cur_reg; | |
1384 | enum machine_mode mode = VOIDmode; | |
1385 | unsigned regno, i, n; | |
1386 | hard_reg_set_iterator hrsi; | |
1387 | def_list_iterator di; | |
1388 | def_t def; | |
1389 | ||
1390 | /* If original register is available, return it. */ | |
1391 | *is_orig_reg_p_ptr = true; | |
1392 | ||
1393 | FOR_EACH_DEF (def, di, original_insns) | |
1394 | { | |
1395 | rtx orig_dest = SET_DEST (PATTERN (def->orig_insn)); | |
1396 | ||
1397 | gcc_assert (REG_P (orig_dest)); | |
1398 | ||
1399 | /* Check that all original operations have the same mode. | |
1400 | This is done for the next loop; if we'd return from this | |
1401 | loop, we'd check only part of them, but in this case | |
1402 | it doesn't matter. */ | |
1403 | if (mode == VOIDmode) | |
1404 | mode = GET_MODE (orig_dest); | |
1405 | gcc_assert (mode == GET_MODE (orig_dest)); | |
1406 | ||
1407 | regno = REGNO (orig_dest); | |
1408 | for (i = 0, n = hard_regno_nregs[regno][mode]; i < n; i++) | |
1409 | if (TEST_HARD_REG_BIT (hard_regs_used, regno + i)) | |
1410 | break; | |
1411 | ||
1412 | /* All hard registers are available. */ | |
1413 | if (i == n) | |
1414 | { | |
1415 | gcc_assert (mode != VOIDmode); | |
1416 | ||
1417 | /* Hard registers should not be shared. */ | |
1418 | return gen_rtx_REG (mode, regno); | |
1419 | } | |
1420 | } | |
1421 | ||
1422 | *is_orig_reg_p_ptr = false; | |
1423 | best_new_reg = -1; | |
1424 | ||
1425 | /* Among all available regs choose the register that was | |
1426 | allocated earliest. */ | |
1427 | EXECUTE_IF_SET_IN_HARD_REG_SET (reg_rename_p->available_for_renaming, | |
1428 | 0, cur_reg, hrsi) | |
1429 | if (! TEST_HARD_REG_BIT (hard_regs_used, cur_reg)) | |
1430 | { | |
1431 | /* All hard registers are available. */ | |
1432 | if (best_new_reg < 0 | |
1433 | || reg_rename_tick[cur_reg] < reg_rename_tick[best_new_reg]) | |
1434 | { | |
1435 | best_new_reg = cur_reg; | |
1436 | ||
1437 | /* Return immediately when we know there's no better reg. */ | |
1438 | if (! reg_rename_tick[best_new_reg]) | |
1439 | break; | |
1440 | } | |
1441 | } | |
1442 | ||
1443 | if (best_new_reg >= 0) | |
1444 | { | |
1445 | /* Use the check from the above loop. */ | |
1446 | gcc_assert (mode != VOIDmode); | |
1447 | return gen_rtx_REG (mode, best_new_reg); | |
1448 | } | |
1449 | ||
1450 | return NULL_RTX; | |
1451 | } | |
1452 | ||
1453 | /* A wrapper around choose_best_reg_1 () to verify that we make correct | |
1454 | assumptions about available registers in the function. */ | |
1455 | static rtx | |
1456 | choose_best_reg (HARD_REG_SET hard_regs_used, struct reg_rename *reg_rename_p, | |
1457 | def_list_t original_insns, bool *is_orig_reg_p_ptr) | |
1458 | { | |
1459 | rtx best_reg = choose_best_reg_1 (hard_regs_used, reg_rename_p, | |
1460 | original_insns, is_orig_reg_p_ptr); | |
1461 | ||
1462 | gcc_assert (best_reg == NULL_RTX | |
1463 | || TEST_HARD_REG_BIT (sel_hrd.regs_ever_used, REGNO (best_reg))); | |
1464 | ||
1465 | return best_reg; | |
1466 | } | |
1467 | ||
1468 | /* Choose the pseudo register for storing rhs value. As this is supposed | |
1469 | to work before reload, we return either the original register or make | |
1470 | the new one. The parameters are the same that in choose_nest_reg_1 | |
1471 | functions, except that USED_REGS may contain pseudos. | |
1472 | If we work with hard regs, check also REG_RENAME_P->UNAVAILABLE_HARD_REGS. | |
1473 | ||
1474 | TODO: take into account register pressure while doing this. Up to this | |
1475 | moment, this function would never return NULL for pseudos, but we should | |
1476 | not rely on this. */ | |
1477 | static rtx | |
1478 | choose_best_pseudo_reg (regset used_regs, | |
1479 | struct reg_rename *reg_rename_p, | |
1480 | def_list_t original_insns, bool *is_orig_reg_p_ptr) | |
1481 | { | |
1482 | def_list_iterator i; | |
1483 | def_t def; | |
1484 | enum machine_mode mode = VOIDmode; | |
1485 | bool bad_hard_regs = false; | |
1486 | ||
1487 | /* We should not use this after reload. */ | |
1488 | gcc_assert (!reload_completed); | |
1489 | ||
1490 | /* If original register is available, return it. */ | |
1491 | *is_orig_reg_p_ptr = true; | |
1492 | ||
1493 | FOR_EACH_DEF (def, i, original_insns) | |
1494 | { | |
1495 | rtx dest = SET_DEST (PATTERN (def->orig_insn)); | |
1496 | int orig_regno; | |
1497 | ||
1498 | gcc_assert (REG_P (dest)); | |
1499 | ||
1500 | /* Check that all original operations have the same mode. */ | |
1501 | if (mode == VOIDmode) | |
1502 | mode = GET_MODE (dest); | |
1503 | else | |
1504 | gcc_assert (mode == GET_MODE (dest)); | |
1505 | orig_regno = REGNO (dest); | |
1506 | ||
1507 | if (!REGNO_REG_SET_P (used_regs, orig_regno)) | |
1508 | { | |
1509 | if (orig_regno < FIRST_PSEUDO_REGISTER) | |
1510 | { | |
1511 | gcc_assert (df_regs_ever_live_p (orig_regno)); | |
1512 | ||
1513 | /* For hard registers, we have to check hardware imposed | |
1514 | limitations (frame/stack registers, calls crossed). */ | |
1515 | if (!TEST_HARD_REG_BIT (reg_rename_p->unavailable_hard_regs, | |
1516 | orig_regno)) | |
1517 | { | |
1518 | /* Don't let register cross a call if it doesn't already | |
1519 | cross one. This condition is written in accordance with | |
1520 | that in sched-deps.c sched_analyze_reg(). */ | |
1521 | if (!reg_rename_p->crosses_call | |
1522 | || REG_N_CALLS_CROSSED (orig_regno) > 0) | |
1523 | return gen_rtx_REG (mode, orig_regno); | |
1524 | } | |
1525 | ||
1526 | bad_hard_regs = true; | |
1527 | } | |
1528 | else | |
1529 | return dest; | |
1530 | } | |
1531 | } | |
1532 | ||
1533 | *is_orig_reg_p_ptr = false; | |
1534 | ||
1535 | /* We had some original hard registers that couldn't be used. | |
1536 | Those were likely special. Don't try to create a pseudo. */ | |
1537 | if (bad_hard_regs) | |
1538 | return NULL_RTX; | |
1539 | ||
1540 | /* We haven't found a register from original operations. Get a new one. | |
1541 | FIXME: control register pressure somehow. */ | |
1542 | { | |
1543 | rtx new_reg = gen_reg_rtx (mode); | |
1544 | ||
1545 | gcc_assert (mode != VOIDmode); | |
1546 | ||
1547 | max_regno = max_reg_num (); | |
1548 | maybe_extend_reg_info_p (); | |
1549 | REG_N_CALLS_CROSSED (REGNO (new_reg)) = reg_rename_p->crosses_call ? 1 : 0; | |
1550 | ||
1551 | return new_reg; | |
1552 | } | |
1553 | } | |
1554 | ||
1555 | /* True when target of EXPR is available due to EXPR_TARGET_AVAILABLE, | |
1556 | USED_REGS and REG_RENAME_P->UNAVAILABLE_HARD_REGS. */ | |
1557 | static void | |
1558 | verify_target_availability (expr_t expr, regset used_regs, | |
1559 | struct reg_rename *reg_rename_p) | |
1560 | { | |
1561 | unsigned n, i, regno; | |
1562 | enum machine_mode mode; | |
1563 | bool target_available, live_available, hard_available; | |
1564 | ||
1565 | if (!REG_P (EXPR_LHS (expr)) || EXPR_TARGET_AVAILABLE (expr) < 0) | |
1566 | return; | |
1567 | ||
1568 | regno = expr_dest_regno (expr); | |
1569 | mode = GET_MODE (EXPR_LHS (expr)); | |
1570 | target_available = EXPR_TARGET_AVAILABLE (expr) == 1; | |
1571 | n = reload_completed ? hard_regno_nregs[regno][mode] : 1; | |
1572 | ||
1573 | live_available = hard_available = true; | |
1574 | for (i = 0; i < n; i++) | |
1575 | { | |
1576 | if (bitmap_bit_p (used_regs, regno + i)) | |
1577 | live_available = false; | |
1578 | if (TEST_HARD_REG_BIT (reg_rename_p->unavailable_hard_regs, regno + i)) | |
1579 | hard_available = false; | |
1580 | } | |
1581 | ||
1582 | /* When target is not available, it may be due to hard register | |
1583 | restrictions, e.g. crosses calls, so we check hard_available too. */ | |
1584 | if (target_available) | |
1585 | gcc_assert (live_available); | |
1586 | else | |
1587 | /* Check only if we haven't scheduled something on the previous fence, | |
1588 | cause due to MAX_SOFTWARE_LOOKAHEAD_WINDOW_SIZE issues | |
1589 | and having more than one fence, we may end having targ_un in a block | |
1590 | in which successors target register is actually available. | |
1591 | ||
1592 | The last condition handles the case when a dependence from a call insn | |
1593 | was created in sched-deps.c for insns with destination registers that | |
1594 | never crossed a call before, but do cross one after our code motion. | |
1595 | ||
1596 | FIXME: in the latter case, we just uselessly called find_used_regs, | |
1597 | because we can't move this expression with any other register | |
1598 | as well. */ | |
1599 | gcc_assert (scheduled_something_on_previous_fence || !live_available | |
1600 | || !hard_available | |
1601 | || (!reload_completed && reg_rename_p->crosses_call | |
1602 | && REG_N_CALLS_CROSSED (regno) == 0)); | |
1603 | } | |
1604 | ||
1605 | /* Collect unavailable registers due to liveness for EXPR from BNDS | |
1606 | into USED_REGS. Save additional information about available | |
1607 | registers and unavailable due to hardware restriction registers | |
1608 | into REG_RENAME_P structure. Save original insns into ORIGINAL_INSNS | |
1609 | list. */ | |
1610 | static void | |
1611 | collect_unavailable_regs_from_bnds (expr_t expr, blist_t bnds, regset used_regs, | |
1612 | struct reg_rename *reg_rename_p, | |
1613 | def_list_t *original_insns) | |
1614 | { | |
1615 | for (; bnds; bnds = BLIST_NEXT (bnds)) | |
1616 | { | |
1617 | bool res; | |
1618 | av_set_t orig_ops = NULL; | |
1619 | bnd_t bnd = BLIST_BND (bnds); | |
1620 | ||
1621 | /* If the chosen best expr doesn't belong to current boundary, | |
1622 | skip it. */ | |
1623 | if (!av_set_is_in_p (BND_AV1 (bnd), EXPR_VINSN (expr))) | |
1624 | continue; | |
1625 | ||
1626 | /* Put in ORIG_OPS all exprs from this boundary that became | |
1627 | RES on top. */ | |
1628 | orig_ops = find_sequential_best_exprs (bnd, expr, false); | |
1629 | ||
1630 | /* Compute used regs and OR it into the USED_REGS. */ | |
1631 | res = find_used_regs (BND_TO (bnd), orig_ops, used_regs, | |
1632 | reg_rename_p, original_insns); | |
1633 | ||
1634 | /* FIXME: the assert is true until we'd have several boundaries. */ | |
1635 | gcc_assert (res); | |
1636 | av_set_clear (&orig_ops); | |
1637 | } | |
1638 | } | |
1639 | ||
1640 | /* Return TRUE if it is possible to replace LHSes of ORIG_INSNS with BEST_REG. | |
1641 | If BEST_REG is valid, replace LHS of EXPR with it. */ | |
1642 | static bool | |
1643 | try_replace_dest_reg (ilist_t orig_insns, rtx best_reg, expr_t expr) | |
1644 | { | |
1645 | if (expr_dest_regno (expr) == REGNO (best_reg)) | |
1646 | { | |
1647 | EXPR_TARGET_AVAILABLE (expr) = 1; | |
1648 | return true; | |
1649 | } | |
1650 | ||
1651 | gcc_assert (orig_insns); | |
1652 | ||
1653 | /* Try whether we'll be able to generate the insn | |
1654 | 'dest := best_reg' at the place of the original operation. */ | |
1655 | for (; orig_insns; orig_insns = ILIST_NEXT (orig_insns)) | |
1656 | { | |
1657 | insn_t orig_insn = DEF_LIST_DEF (orig_insns)->orig_insn; | |
1658 | ||
1659 | gcc_assert (EXPR_SEPARABLE_P (INSN_EXPR (orig_insn))); | |
1660 | ||
1661 | if (!replace_src_with_reg_ok_p (orig_insn, best_reg) | |
1662 | || !replace_dest_with_reg_ok_p (orig_insn, best_reg)) | |
1663 | return false; | |
1664 | } | |
1665 | ||
1666 | /* Make sure that EXPR has the right destination | |
1667 | register. */ | |
1668 | replace_dest_with_reg_in_expr (expr, best_reg); | |
1669 | return true; | |
1670 | } | |
1671 | ||
1672 | /* Select and assign best register to EXPR searching from BNDS. | |
1673 | Set *IS_ORIG_REG_P to TRUE if original register was selected. | |
1674 | Return FALSE if no register can be chosen, which could happen when: | |
1675 | * EXPR_SEPARABLE_P is true but we were unable to find suitable register; | |
1676 | * EXPR_SEPARABLE_P is false but the insn sets/clobbers one of the registers | |
1677 | that are used on the moving path. */ | |
1678 | static bool | |
1679 | find_best_reg_for_expr (expr_t expr, blist_t bnds, bool *is_orig_reg_p) | |
1680 | { | |
1681 | static struct reg_rename reg_rename_data; | |
1682 | ||
1683 | regset used_regs; | |
1684 | def_list_t original_insns = NULL; | |
1685 | bool reg_ok; | |
1686 | ||
1687 | *is_orig_reg_p = false; | |
1688 | ||
1689 | /* Don't bother to do anything if this insn doesn't set any registers. */ | |
1690 | if (bitmap_empty_p (VINSN_REG_SETS (EXPR_VINSN (expr))) | |
1691 | && bitmap_empty_p (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)))) | |
1692 | return true; | |
1693 | ||
1694 | used_regs = get_clear_regset_from_pool (); | |
1695 | CLEAR_HARD_REG_SET (reg_rename_data.unavailable_hard_regs); | |
1696 | ||
1697 | collect_unavailable_regs_from_bnds (expr, bnds, used_regs, ®_rename_data, | |
1698 | &original_insns); | |
1699 | ||
1700 | #ifdef ENABLE_CHECKING | |
1701 | /* If after reload, make sure we're working with hard regs here. */ | |
1702 | if (reload_completed) | |
1703 | { | |
1704 | reg_set_iterator rsi; | |
1705 | unsigned i; | |
1706 | ||
1707 | EXECUTE_IF_SET_IN_REG_SET (used_regs, FIRST_PSEUDO_REGISTER, i, rsi) | |
1708 | gcc_unreachable (); | |
1709 | } | |
1710 | #endif | |
1711 | ||
1712 | if (EXPR_SEPARABLE_P (expr)) | |
1713 | { | |
1714 | rtx best_reg = NULL_RTX; | |
1715 | /* Check that we have computed availability of a target register | |
1716 | correctly. */ | |
1717 | verify_target_availability (expr, used_regs, ®_rename_data); | |
1718 | ||
1719 | /* Turn everything in hard regs after reload. */ | |
1720 | if (reload_completed) | |
1721 | { | |
1722 | HARD_REG_SET hard_regs_used; | |
1723 | REG_SET_TO_HARD_REG_SET (hard_regs_used, used_regs); | |
1724 | ||
1725 | /* Join hard registers unavailable due to register class | |
1726 | restrictions and live range intersection. */ | |
1727 | IOR_HARD_REG_SET (hard_regs_used, | |
1728 | reg_rename_data.unavailable_hard_regs); | |
1729 | ||
1730 | best_reg = choose_best_reg (hard_regs_used, ®_rename_data, | |
1731 | original_insns, is_orig_reg_p); | |
1732 | } | |
1733 | else | |
1734 | best_reg = choose_best_pseudo_reg (used_regs, ®_rename_data, | |
1735 | original_insns, is_orig_reg_p); | |
1736 | ||
1737 | if (!best_reg) | |
1738 | reg_ok = false; | |
1739 | else if (*is_orig_reg_p) | |
1740 | { | |
1741 | /* In case of unification BEST_REG may be different from EXPR's LHS | |
1742 | when EXPR's LHS is unavailable, and there is another LHS among | |
1743 | ORIGINAL_INSNS. */ | |
1744 | reg_ok = try_replace_dest_reg (original_insns, best_reg, expr); | |
1745 | } | |
1746 | else | |
1747 | { | |
1748 | /* Forbid renaming of low-cost insns. */ | |
1749 | if (sel_vinsn_cost (EXPR_VINSN (expr)) < 2) | |
1750 | reg_ok = false; | |
1751 | else | |
1752 | reg_ok = try_replace_dest_reg (original_insns, best_reg, expr); | |
1753 | } | |
1754 | } | |
1755 | else | |
1756 | { | |
1757 | /* If !EXPR_SCHEDULE_AS_RHS (EXPR), just make sure INSN doesn't set | |
1758 | any of the HARD_REGS_USED set. */ | |
1759 | if (vinsn_writes_one_of_regs_p (EXPR_VINSN (expr), used_regs, | |
1760 | reg_rename_data.unavailable_hard_regs)) | |
1761 | { | |
1762 | reg_ok = false; | |
1763 | gcc_assert (EXPR_TARGET_AVAILABLE (expr) <= 0); | |
1764 | } | |
1765 | else | |
1766 | { | |
1767 | reg_ok = true; | |
1768 | gcc_assert (EXPR_TARGET_AVAILABLE (expr) != 0); | |
1769 | } | |
1770 | } | |
1771 | ||
1772 | ilist_clear (&original_insns); | |
1773 | return_regset_to_pool (used_regs); | |
1774 | ||
1775 | return reg_ok; | |
1776 | } | |
1777 | \f | |
1778 | ||
1779 | /* Return true if dependence described by DS can be overcomed. */ | |
1780 | static bool | |
1781 | can_speculate_dep_p (ds_t ds) | |
1782 | { | |
1783 | if (spec_info == NULL) | |
1784 | return false; | |
1785 | ||
1786 | /* Leave only speculative data. */ | |
1787 | ds &= SPECULATIVE; | |
1788 | ||
1789 | if (ds == 0) | |
1790 | return false; | |
1791 | ||
1792 | { | |
1793 | /* FIXME: make sched-deps.c produce only those non-hard dependencies, | |
1794 | that we can overcome. */ | |
1795 | ds_t spec_mask = spec_info->mask; | |
1796 | ||
1797 | if ((ds & spec_mask) != ds) | |
1798 | return false; | |
1799 | } | |
1800 | ||
1801 | if (ds_weak (ds) < spec_info->data_weakness_cutoff) | |
1802 | return false; | |
1803 | ||
1804 | return true; | |
1805 | } | |
1806 | ||
1807 | /* Get a speculation check instruction. | |
1808 | C_EXPR is a speculative expression, | |
1809 | CHECK_DS describes speculations that should be checked, | |
1810 | ORIG_INSN is the original non-speculative insn in the stream. */ | |
1811 | static insn_t | |
1812 | create_speculation_check (expr_t c_expr, ds_t check_ds, insn_t orig_insn) | |
1813 | { | |
1814 | rtx check_pattern; | |
1815 | rtx insn_rtx; | |
1816 | insn_t insn; | |
1817 | basic_block recovery_block; | |
1818 | rtx label; | |
1819 | ||
1820 | /* Create a recovery block if target is going to emit branchy check, or if | |
1821 | ORIG_INSN was speculative already. */ | |
388092d5 | 1822 | if (targetm.sched.needs_block_p (check_ds) |
e855c69d AB |
1823 | || EXPR_SPEC_DONE_DS (INSN_EXPR (orig_insn)) != 0) |
1824 | { | |
1825 | recovery_block = sel_create_recovery_block (orig_insn); | |
1826 | label = BB_HEAD (recovery_block); | |
1827 | } | |
1828 | else | |
1829 | { | |
1830 | recovery_block = NULL; | |
1831 | label = NULL_RTX; | |
1832 | } | |
1833 | ||
1834 | /* Get pattern of the check. */ | |
1835 | check_pattern = targetm.sched.gen_spec_check (EXPR_INSN_RTX (c_expr), label, | |
1836 | check_ds); | |
1837 | ||
1838 | gcc_assert (check_pattern != NULL); | |
1839 | ||
1840 | /* Emit check. */ | |
1841 | insn_rtx = create_insn_rtx_from_pattern (check_pattern, label); | |
1842 | ||
1843 | insn = sel_gen_insn_from_rtx_after (insn_rtx, INSN_EXPR (orig_insn), | |
1844 | INSN_SEQNO (orig_insn), orig_insn); | |
1845 | ||
1846 | /* Make check to be non-speculative. */ | |
1847 | EXPR_SPEC_DONE_DS (INSN_EXPR (insn)) = 0; | |
1848 | INSN_SPEC_CHECKED_DS (insn) = check_ds; | |
1849 | ||
1850 | /* Decrease priority of check by difference of load/check instruction | |
1851 | latencies. */ | |
1852 | EXPR_PRIORITY (INSN_EXPR (insn)) -= (sel_vinsn_cost (INSN_VINSN (orig_insn)) | |
1853 | - sel_vinsn_cost (INSN_VINSN (insn))); | |
1854 | ||
1855 | /* Emit copy of original insn (though with replaced target register, | |
1856 | if needed) to the recovery block. */ | |
1857 | if (recovery_block != NULL) | |
1858 | { | |
1859 | rtx twin_rtx; | |
1860 | insn_t twin; | |
1861 | ||
1862 | twin_rtx = copy_rtx (PATTERN (EXPR_INSN_RTX (c_expr))); | |
1863 | twin_rtx = create_insn_rtx_from_pattern (twin_rtx, NULL_RTX); | |
1864 | twin = sel_gen_recovery_insn_from_rtx_after (twin_rtx, | |
1865 | INSN_EXPR (orig_insn), | |
1866 | INSN_SEQNO (insn), | |
1867 | bb_note (recovery_block)); | |
1868 | } | |
1869 | ||
1870 | /* If we've generated a data speculation check, make sure | |
1871 | that all the bookkeeping instruction we'll create during | |
1872 | this move_op () will allocate an ALAT entry so that the | |
1873 | check won't fail. | |
1874 | In case of control speculation we must convert C_EXPR to control | |
1875 | speculative mode, because failing to do so will bring us an exception | |
1876 | thrown by the non-control-speculative load. */ | |
1877 | check_ds = ds_get_max_dep_weak (check_ds); | |
1878 | speculate_expr (c_expr, check_ds); | |
1879 | ||
1880 | return insn; | |
1881 | } | |
1882 | ||
1883 | /* True when INSN is a "regN = regN" copy. */ | |
1884 | static bool | |
1885 | identical_copy_p (rtx insn) | |
1886 | { | |
1887 | rtx lhs, rhs, pat; | |
1888 | ||
1889 | pat = PATTERN (insn); | |
1890 | ||
1891 | if (GET_CODE (pat) != SET) | |
1892 | return false; | |
1893 | ||
1894 | lhs = SET_DEST (pat); | |
1895 | if (!REG_P (lhs)) | |
1896 | return false; | |
1897 | ||
1898 | rhs = SET_SRC (pat); | |
1899 | if (!REG_P (rhs)) | |
1900 | return false; | |
1901 | ||
1902 | return REGNO (lhs) == REGNO (rhs); | |
1903 | } | |
1904 | ||
1905 | /* Undo all transformations on *AV_PTR that were done when | |
1906 | moving through INSN. */ | |
1907 | static void | |
1908 | undo_transformations (av_set_t *av_ptr, rtx insn) | |
1909 | { | |
1910 | av_set_iterator av_iter; | |
1911 | expr_t expr; | |
1912 | av_set_t new_set = NULL; | |
1913 | ||
1914 | /* First, kill any EXPR that uses registers set by an insn. This is | |
1915 | required for correctness. */ | |
1916 | FOR_EACH_EXPR_1 (expr, av_iter, av_ptr) | |
1917 | if (!sched_insns_conditions_mutex_p (insn, EXPR_INSN_RTX (expr)) | |
1918 | && bitmap_intersect_p (INSN_REG_SETS (insn), | |
1919 | VINSN_REG_USES (EXPR_VINSN (expr))) | |
1920 | /* When an insn looks like 'r1 = r1', we could substitute through | |
1921 | it, but the above condition will still hold. This happened with | |
1922 | gcc.c-torture/execute/961125-1.c. */ | |
1923 | && !identical_copy_p (insn)) | |
1924 | { | |
1925 | if (sched_verbose >= 6) | |
1926 | sel_print ("Expr %d removed due to use/set conflict\n", | |
1927 | INSN_UID (EXPR_INSN_RTX (expr))); | |
1928 | av_set_iter_remove (&av_iter); | |
1929 | } | |
1930 | ||
1931 | /* Undo transformations looking at the history vector. */ | |
1932 | FOR_EACH_EXPR (expr, av_iter, *av_ptr) | |
1933 | { | |
1934 | int index = find_in_history_vect (EXPR_HISTORY_OF_CHANGES (expr), | |
1935 | insn, EXPR_VINSN (expr), true); | |
1936 | ||
1937 | if (index >= 0) | |
1938 | { | |
1939 | expr_history_def *phist; | |
1940 | ||
1941 | phist = VEC_index (expr_history_def, | |
1942 | EXPR_HISTORY_OF_CHANGES (expr), | |
1943 | index); | |
1944 | ||
1945 | switch (phist->type) | |
1946 | { | |
1947 | case TRANS_SPECULATION: | |
1948 | { | |
1949 | ds_t old_ds, new_ds; | |
1950 | ||
1951 | /* Compute the difference between old and new speculative | |
1952 | statuses: that's what we need to check. | |
1953 | Earlier we used to assert that the status will really | |
1954 | change. This no longer works because only the probability | |
1955 | bits in the status may have changed during compute_av_set, | |
1956 | and in the case of merging different probabilities of the | |
1957 | same speculative status along different paths we do not | |
1958 | record this in the history vector. */ | |
1959 | old_ds = phist->spec_ds; | |
1960 | new_ds = EXPR_SPEC_DONE_DS (expr); | |
1961 | ||
1962 | old_ds &= SPECULATIVE; | |
1963 | new_ds &= SPECULATIVE; | |
1964 | new_ds &= ~old_ds; | |
1965 | ||
1966 | EXPR_SPEC_TO_CHECK_DS (expr) |= new_ds; | |
1967 | break; | |
1968 | } | |
1969 | case TRANS_SUBSTITUTION: | |
1970 | { | |
1971 | expr_def _tmp_expr, *tmp_expr = &_tmp_expr; | |
1972 | vinsn_t new_vi; | |
1973 | bool add = true; | |
1974 | ||
1975 | new_vi = phist->old_expr_vinsn; | |
1976 | ||
1977 | gcc_assert (VINSN_SEPARABLE_P (new_vi) | |
1978 | == EXPR_SEPARABLE_P (expr)); | |
1979 | copy_expr (tmp_expr, expr); | |
1980 | ||
1981 | if (vinsn_equal_p (phist->new_expr_vinsn, | |
1982 | EXPR_VINSN (tmp_expr))) | |
1983 | change_vinsn_in_expr (tmp_expr, new_vi); | |
1984 | else | |
1985 | /* This happens when we're unsubstituting on a bookkeeping | |
1986 | copy, which was in turn substituted. The history is wrong | |
1987 | in this case. Do it the hard way. */ | |
1988 | add = substitute_reg_in_expr (tmp_expr, insn, true); | |
1989 | if (add) | |
1990 | av_set_add (&new_set, tmp_expr); | |
1991 | clear_expr (tmp_expr); | |
1992 | break; | |
1993 | } | |
1994 | default: | |
1995 | gcc_unreachable (); | |
1996 | } | |
1997 | } | |
1998 | ||
1999 | } | |
2000 | ||
2001 | av_set_union_and_clear (av_ptr, &new_set, NULL); | |
2002 | } | |
2003 | \f | |
2004 | ||
2005 | /* Moveup_* helpers for code motion and computing av sets. */ | |
2006 | ||
2007 | /* Propagates EXPR inside an insn group through THROUGH_INSN. | |
2008 | The difference from the below function is that only substitution is | |
2009 | performed. */ | |
2010 | static enum MOVEUP_EXPR_CODE | |
2011 | moveup_expr_inside_insn_group (expr_t expr, insn_t through_insn) | |
2012 | { | |
2013 | vinsn_t vi = EXPR_VINSN (expr); | |
2014 | ds_t *has_dep_p; | |
2015 | ds_t full_ds; | |
2016 | ||
2017 | /* Do this only inside insn group. */ | |
2018 | gcc_assert (INSN_SCHED_CYCLE (through_insn) > 0); | |
2019 | ||
2020 | full_ds = has_dependence_p (expr, through_insn, &has_dep_p); | |
2021 | if (full_ds == 0) | |
2022 | return MOVEUP_EXPR_SAME; | |
2023 | ||
2024 | /* Substitution is the possible choice in this case. */ | |
2025 | if (has_dep_p[DEPS_IN_RHS]) | |
2026 | { | |
2027 | /* Can't substitute UNIQUE VINSNs. */ | |
2028 | gcc_assert (!VINSN_UNIQUE_P (vi)); | |
2029 | ||
2030 | if (can_substitute_through_p (through_insn, | |
2031 | has_dep_p[DEPS_IN_RHS]) | |
2032 | && substitute_reg_in_expr (expr, through_insn, false)) | |
2033 | { | |
2034 | EXPR_WAS_SUBSTITUTED (expr) = true; | |
2035 | return MOVEUP_EXPR_CHANGED; | |
2036 | } | |
2037 | ||
2038 | /* Don't care about this, as even true dependencies may be allowed | |
2039 | in an insn group. */ | |
2040 | return MOVEUP_EXPR_SAME; | |
2041 | } | |
2042 | ||
2043 | /* This can catch output dependencies in COND_EXECs. */ | |
2044 | if (has_dep_p[DEPS_IN_INSN]) | |
2045 | return MOVEUP_EXPR_NULL; | |
2046 | ||
2047 | /* This is either an output or an anti dependence, which usually have | |
2048 | a zero latency. Allow this here, if we'd be wrong, tick_check_p | |
2049 | will fix this. */ | |
2050 | gcc_assert (has_dep_p[DEPS_IN_LHS]); | |
2051 | return MOVEUP_EXPR_AS_RHS; | |
2052 | } | |
2053 | ||
2054 | /* True when a trapping EXPR cannot be moved through THROUGH_INSN. */ | |
2055 | #define CANT_MOVE_TRAPPING(expr, through_insn) \ | |
2056 | (VINSN_MAY_TRAP_P (EXPR_VINSN (expr)) \ | |
2057 | && !sel_insn_has_single_succ_p ((through_insn), SUCCS_ALL) \ | |
2058 | && !sel_insn_is_speculation_check (through_insn)) | |
2059 | ||
2060 | /* True when a conflict on a target register was found during moveup_expr. */ | |
2061 | static bool was_target_conflict = false; | |
2062 | ||
b5b8b0ac AO |
2063 | /* Return true when moving a debug INSN across THROUGH_INSN will |
2064 | create a bookkeeping block. We don't want to create such blocks, | |
2065 | for they would cause codegen differences between compilations with | |
2066 | and without debug info. */ | |
2067 | ||
2068 | static bool | |
2069 | moving_insn_creates_bookkeeping_block_p (insn_t insn, | |
2070 | insn_t through_insn) | |
2071 | { | |
2072 | basic_block bbi, bbt; | |
2073 | edge e1, e2; | |
2074 | edge_iterator ei1, ei2; | |
2075 | ||
2076 | if (!bookkeeping_can_be_created_if_moved_through_p (through_insn)) | |
2077 | { | |
2078 | if (sched_verbose >= 9) | |
2079 | sel_print ("no bookkeeping required: "); | |
2080 | return FALSE; | |
2081 | } | |
2082 | ||
2083 | bbi = BLOCK_FOR_INSN (insn); | |
2084 | ||
2085 | if (EDGE_COUNT (bbi->preds) == 1) | |
2086 | { | |
2087 | if (sched_verbose >= 9) | |
2088 | sel_print ("only one pred edge: "); | |
2089 | return TRUE; | |
2090 | } | |
2091 | ||
2092 | bbt = BLOCK_FOR_INSN (through_insn); | |
2093 | ||
2094 | FOR_EACH_EDGE (e1, ei1, bbt->succs) | |
2095 | { | |
2096 | FOR_EACH_EDGE (e2, ei2, bbi->preds) | |
2097 | { | |
2098 | if (find_block_for_bookkeeping (e1, e2, TRUE)) | |
2099 | { | |
2100 | if (sched_verbose >= 9) | |
2101 | sel_print ("found existing block: "); | |
2102 | return FALSE; | |
2103 | } | |
2104 | } | |
2105 | } | |
2106 | ||
2107 | if (sched_verbose >= 9) | |
2108 | sel_print ("would create bookkeeping block: "); | |
2109 | ||
2110 | return TRUE; | |
2111 | } | |
2112 | ||
e855c69d AB |
2113 | /* Modifies EXPR so it can be moved through the THROUGH_INSN, |
2114 | performing necessary transformations. Record the type of transformation | |
2115 | made in PTRANS_TYPE, when it is not NULL. When INSIDE_INSN_GROUP, | |
2116 | permit all dependencies except true ones, and try to remove those | |
2117 | too via forward substitution. All cases when a non-eliminable | |
2118 | non-zero cost dependency exists inside an insn group will be fixed | |
2119 | in tick_check_p instead. */ | |
2120 | static enum MOVEUP_EXPR_CODE | |
2121 | moveup_expr (expr_t expr, insn_t through_insn, bool inside_insn_group, | |
2122 | enum local_trans_type *ptrans_type) | |
2123 | { | |
2124 | vinsn_t vi = EXPR_VINSN (expr); | |
2125 | insn_t insn = VINSN_INSN_RTX (vi); | |
2126 | bool was_changed = false; | |
2127 | bool as_rhs = false; | |
2128 | ds_t *has_dep_p; | |
2129 | ds_t full_ds; | |
2130 | ||
2131 | /* When inside_insn_group, delegate to the helper. */ | |
2132 | if (inside_insn_group) | |
2133 | return moveup_expr_inside_insn_group (expr, through_insn); | |
2134 | ||
2135 | /* Deal with unique insns and control dependencies. */ | |
2136 | if (VINSN_UNIQUE_P (vi)) | |
2137 | { | |
2138 | /* We can move jumps without side-effects or jumps that are | |
2139 | mutually exclusive with instruction THROUGH_INSN (all in cases | |
2140 | dependencies allow to do so and jump is not speculative). */ | |
2141 | if (control_flow_insn_p (insn)) | |
2142 | { | |
2143 | basic_block fallthru_bb; | |
2144 | ||
2145 | /* Do not move checks and do not move jumps through other | |
2146 | jumps. */ | |
2147 | if (control_flow_insn_p (through_insn) | |
2148 | || sel_insn_is_speculation_check (insn)) | |
2149 | return MOVEUP_EXPR_NULL; | |
2150 | ||
2151 | /* Don't move jumps through CFG joins. */ | |
2152 | if (bookkeeping_can_be_created_if_moved_through_p (through_insn)) | |
2153 | return MOVEUP_EXPR_NULL; | |
2154 | ||
2155 | /* The jump should have a clear fallthru block, and | |
2156 | this block should be in the current region. */ | |
2157 | if ((fallthru_bb = fallthru_bb_of_jump (insn)) == NULL | |
2158 | || ! in_current_region_p (fallthru_bb)) | |
2159 | return MOVEUP_EXPR_NULL; | |
2160 | ||
2161 | /* And it should be mutually exclusive with through_insn, or | |
2162 | be an unconditional jump. */ | |
2163 | if (! any_uncondjump_p (insn) | |
b5b8b0ac AO |
2164 | && ! sched_insns_conditions_mutex_p (insn, through_insn) |
2165 | && ! DEBUG_INSN_P (through_insn)) | |
e855c69d AB |
2166 | return MOVEUP_EXPR_NULL; |
2167 | } | |
2168 | ||
2169 | /* Don't move what we can't move. */ | |
2170 | if (EXPR_CANT_MOVE (expr) | |
2171 | && BLOCK_FOR_INSN (through_insn) != BLOCK_FOR_INSN (insn)) | |
2172 | return MOVEUP_EXPR_NULL; | |
2173 | ||
2174 | /* Don't move SCHED_GROUP instruction through anything. | |
2175 | If we don't force this, then it will be possible to start | |
2176 | scheduling a sched_group before all its dependencies are | |
2177 | resolved. | |
2178 | ??? Haifa deals with this issue by delaying the SCHED_GROUP | |
2179 | as late as possible through rank_for_schedule. */ | |
2180 | if (SCHED_GROUP_P (insn)) | |
2181 | return MOVEUP_EXPR_NULL; | |
2182 | } | |
2183 | else | |
2184 | gcc_assert (!control_flow_insn_p (insn)); | |
2185 | ||
b5b8b0ac AO |
2186 | /* Don't move debug insns if this would require bookkeeping. */ |
2187 | if (DEBUG_INSN_P (insn) | |
2188 | && BLOCK_FOR_INSN (through_insn) != BLOCK_FOR_INSN (insn) | |
2189 | && moving_insn_creates_bookkeeping_block_p (insn, through_insn)) | |
2190 | return MOVEUP_EXPR_NULL; | |
2191 | ||
e855c69d AB |
2192 | /* Deal with data dependencies. */ |
2193 | was_target_conflict = false; | |
2194 | full_ds = has_dependence_p (expr, through_insn, &has_dep_p); | |
2195 | if (full_ds == 0) | |
2196 | { | |
2197 | if (!CANT_MOVE_TRAPPING (expr, through_insn)) | |
2198 | return MOVEUP_EXPR_SAME; | |
2199 | } | |
2200 | else | |
2201 | { | |
2202 | /* We can move UNIQUE insn up only as a whole and unchanged, | |
2203 | so it shouldn't have any dependencies. */ | |
2204 | if (VINSN_UNIQUE_P (vi)) | |
2205 | return MOVEUP_EXPR_NULL; | |
2206 | } | |
2207 | ||
2208 | if (full_ds != 0 && can_speculate_dep_p (full_ds)) | |
2209 | { | |
2210 | int res; | |
2211 | ||
2212 | res = speculate_expr (expr, full_ds); | |
2213 | if (res >= 0) | |
2214 | { | |
2215 | /* Speculation was successful. */ | |
2216 | full_ds = 0; | |
2217 | was_changed = (res > 0); | |
2218 | if (res == 2) | |
2219 | was_target_conflict = true; | |
2220 | if (ptrans_type) | |
2221 | *ptrans_type = TRANS_SPECULATION; | |
2222 | sel_clear_has_dependence (); | |
2223 | } | |
2224 | } | |
2225 | ||
2226 | if (has_dep_p[DEPS_IN_INSN]) | |
2227 | /* We have some dependency that cannot be discarded. */ | |
2228 | return MOVEUP_EXPR_NULL; | |
2229 | ||
2230 | if (has_dep_p[DEPS_IN_LHS]) | |
2231 | { | |
2232 | /* Only separable insns can be moved up with the new register. | |
2233 | Anyways, we should mark that the original register is | |
2234 | unavailable. */ | |
2235 | if (!enable_schedule_as_rhs_p || !EXPR_SEPARABLE_P (expr)) | |
2236 | return MOVEUP_EXPR_NULL; | |
2237 | ||
2238 | EXPR_TARGET_AVAILABLE (expr) = false; | |
2239 | was_target_conflict = true; | |
2240 | as_rhs = true; | |
2241 | } | |
2242 | ||
2243 | /* At this point we have either separable insns, that will be lifted | |
2244 | up only as RHSes, or non-separable insns with no dependency in lhs. | |
2245 | If dependency is in RHS, then try to perform substitution and move up | |
2246 | substituted RHS: | |
2247 | ||
2248 | Ex. 1: Ex.2 | |
2249 | y = x; y = x; | |
2250 | z = y*2; y = y*2; | |
2251 | ||
2252 | In Ex.1 y*2 can be substituted for x*2 and the whole operation can be | |
2253 | moved above y=x assignment as z=x*2. | |
2254 | ||
2255 | In Ex.2 y*2 also can be substituted for x*2, but only the right hand | |
2256 | side can be moved because of the output dependency. The operation was | |
2257 | cropped to its rhs above. */ | |
2258 | if (has_dep_p[DEPS_IN_RHS]) | |
2259 | { | |
2260 | ds_t *rhs_dsp = &has_dep_p[DEPS_IN_RHS]; | |
2261 | ||
2262 | /* Can't substitute UNIQUE VINSNs. */ | |
2263 | gcc_assert (!VINSN_UNIQUE_P (vi)); | |
2264 | ||
2265 | if (can_speculate_dep_p (*rhs_dsp)) | |
2266 | { | |
2267 | int res; | |
2268 | ||
2269 | res = speculate_expr (expr, *rhs_dsp); | |
2270 | if (res >= 0) | |
2271 | { | |
2272 | /* Speculation was successful. */ | |
2273 | *rhs_dsp = 0; | |
2274 | was_changed = (res > 0); | |
2275 | if (res == 2) | |
2276 | was_target_conflict = true; | |
2277 | if (ptrans_type) | |
2278 | *ptrans_type = TRANS_SPECULATION; | |
2279 | } | |
2280 | else | |
2281 | return MOVEUP_EXPR_NULL; | |
2282 | } | |
2283 | else if (can_substitute_through_p (through_insn, | |
2284 | *rhs_dsp) | |
2285 | && substitute_reg_in_expr (expr, through_insn, false)) | |
2286 | { | |
2287 | /* ??? We cannot perform substitution AND speculation on the same | |
2288 | insn. */ | |
2289 | gcc_assert (!was_changed); | |
2290 | was_changed = true; | |
2291 | if (ptrans_type) | |
2292 | *ptrans_type = TRANS_SUBSTITUTION; | |
2293 | EXPR_WAS_SUBSTITUTED (expr) = true; | |
2294 | } | |
2295 | else | |
2296 | return MOVEUP_EXPR_NULL; | |
2297 | } | |
2298 | ||
2299 | /* Don't move trapping insns through jumps. | |
2300 | This check should be at the end to give a chance to control speculation | |
2301 | to perform its duties. */ | |
2302 | if (CANT_MOVE_TRAPPING (expr, through_insn)) | |
2303 | return MOVEUP_EXPR_NULL; | |
2304 | ||
2305 | return (was_changed | |
2306 | ? MOVEUP_EXPR_CHANGED | |
2307 | : (as_rhs | |
2308 | ? MOVEUP_EXPR_AS_RHS | |
2309 | : MOVEUP_EXPR_SAME)); | |
2310 | } | |
2311 | ||
2312 | /* Try to look at bitmap caches for EXPR and INSN pair, return true | |
2313 | if successful. When INSIDE_INSN_GROUP, also try ignore dependencies | |
2314 | that can exist within a parallel group. Write to RES the resulting | |
2315 | code for moveup_expr. */ | |
2316 | static bool | |
2317 | try_bitmap_cache (expr_t expr, insn_t insn, | |
2318 | bool inside_insn_group, | |
2319 | enum MOVEUP_EXPR_CODE *res) | |
2320 | { | |
2321 | int expr_uid = INSN_UID (EXPR_INSN_RTX (expr)); | |
2322 | ||
2323 | /* First check whether we've analyzed this situation already. */ | |
2324 | if (bitmap_bit_p (INSN_ANALYZED_DEPS (insn), expr_uid)) | |
2325 | { | |
2326 | if (bitmap_bit_p (INSN_FOUND_DEPS (insn), expr_uid)) | |
2327 | { | |
2328 | if (sched_verbose >= 6) | |
2329 | sel_print ("removed (cached)\n"); | |
2330 | *res = MOVEUP_EXPR_NULL; | |
2331 | return true; | |
2332 | } | |
2333 | else | |
2334 | { | |
2335 | if (sched_verbose >= 6) | |
2336 | sel_print ("unchanged (cached)\n"); | |
2337 | *res = MOVEUP_EXPR_SAME; | |
2338 | return true; | |
2339 | } | |
2340 | } | |
2341 | else if (bitmap_bit_p (INSN_FOUND_DEPS (insn), expr_uid)) | |
2342 | { | |
2343 | if (inside_insn_group) | |
2344 | { | |
2345 | if (sched_verbose >= 6) | |
2346 | sel_print ("unchanged (as RHS, cached, inside insn group)\n"); | |
2347 | *res = MOVEUP_EXPR_SAME; | |
2348 | return true; | |
2349 | ||
2350 | } | |
2351 | else | |
2352 | EXPR_TARGET_AVAILABLE (expr) = false; | |
2353 | ||
2354 | /* This is the only case when propagation result can change over time, | |
2355 | as we can dynamically switch off scheduling as RHS. In this case, | |
2356 | just check the flag to reach the correct decision. */ | |
2357 | if (enable_schedule_as_rhs_p) | |
2358 | { | |
2359 | if (sched_verbose >= 6) | |
2360 | sel_print ("unchanged (as RHS, cached)\n"); | |
2361 | *res = MOVEUP_EXPR_AS_RHS; | |
2362 | return true; | |
2363 | } | |
2364 | else | |
2365 | { | |
2366 | if (sched_verbose >= 6) | |
2367 | sel_print ("removed (cached as RHS, but renaming" | |
2368 | " is now disabled)\n"); | |
2369 | *res = MOVEUP_EXPR_NULL; | |
2370 | return true; | |
2371 | } | |
2372 | } | |
2373 | ||
2374 | return false; | |
2375 | } | |
2376 | ||
2377 | /* Try to look at bitmap caches for EXPR and INSN pair, return true | |
2378 | if successful. Write to RES the resulting code for moveup_expr. */ | |
2379 | static bool | |
2380 | try_transformation_cache (expr_t expr, insn_t insn, | |
2381 | enum MOVEUP_EXPR_CODE *res) | |
2382 | { | |
2383 | struct transformed_insns *pti | |
2384 | = (struct transformed_insns *) | |
2385 | htab_find_with_hash (INSN_TRANSFORMED_INSNS (insn), | |
2386 | &EXPR_VINSN (expr), | |
2387 | VINSN_HASH_RTX (EXPR_VINSN (expr))); | |
2388 | if (pti) | |
2389 | { | |
2390 | /* This EXPR was already moved through this insn and was | |
2391 | changed as a result. Fetch the proper data from | |
2392 | the hashtable. */ | |
2393 | insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (expr), | |
2394 | INSN_UID (insn), pti->type, | |
2395 | pti->vinsn_old, pti->vinsn_new, | |
2396 | EXPR_SPEC_DONE_DS (expr)); | |
2397 | ||
2398 | if (INSN_IN_STREAM_P (VINSN_INSN_RTX (pti->vinsn_new))) | |
2399 | pti->vinsn_new = vinsn_copy (pti->vinsn_new, true); | |
2400 | change_vinsn_in_expr (expr, pti->vinsn_new); | |
2401 | if (pti->was_target_conflict) | |
2402 | EXPR_TARGET_AVAILABLE (expr) = false; | |
2403 | if (pti->type == TRANS_SPECULATION) | |
2404 | { | |
2405 | ds_t ds; | |
2406 | ||
2407 | ds = EXPR_SPEC_DONE_DS (expr); | |
2408 | ||
2409 | EXPR_SPEC_DONE_DS (expr) = pti->ds; | |
2410 | EXPR_NEEDS_SPEC_CHECK_P (expr) |= pti->needs_check; | |
2411 | } | |
2412 | ||
2413 | if (sched_verbose >= 6) | |
2414 | { | |
2415 | sel_print ("changed (cached): "); | |
2416 | dump_expr (expr); | |
2417 | sel_print ("\n"); | |
2418 | } | |
2419 | ||
2420 | *res = MOVEUP_EXPR_CHANGED; | |
2421 | return true; | |
2422 | } | |
2423 | ||
2424 | return false; | |
2425 | } | |
2426 | ||
2427 | /* Update bitmap caches on INSN with result RES of propagating EXPR. */ | |
2428 | static void | |
2429 | update_bitmap_cache (expr_t expr, insn_t insn, bool inside_insn_group, | |
2430 | enum MOVEUP_EXPR_CODE res) | |
2431 | { | |
2432 | int expr_uid = INSN_UID (EXPR_INSN_RTX (expr)); | |
2433 | ||
2434 | /* Do not cache result of propagating jumps through an insn group, | |
2435 | as it is always true, which is not useful outside the group. */ | |
2436 | if (inside_insn_group) | |
2437 | return; | |
2438 | ||
2439 | if (res == MOVEUP_EXPR_NULL) | |
2440 | { | |
2441 | bitmap_set_bit (INSN_ANALYZED_DEPS (insn), expr_uid); | |
2442 | bitmap_set_bit (INSN_FOUND_DEPS (insn), expr_uid); | |
2443 | } | |
2444 | else if (res == MOVEUP_EXPR_SAME) | |
2445 | { | |
2446 | bitmap_set_bit (INSN_ANALYZED_DEPS (insn), expr_uid); | |
2447 | bitmap_clear_bit (INSN_FOUND_DEPS (insn), expr_uid); | |
2448 | } | |
2449 | else if (res == MOVEUP_EXPR_AS_RHS) | |
2450 | { | |
2451 | bitmap_clear_bit (INSN_ANALYZED_DEPS (insn), expr_uid); | |
2452 | bitmap_set_bit (INSN_FOUND_DEPS (insn), expr_uid); | |
2453 | } | |
2454 | else | |
2455 | gcc_unreachable (); | |
2456 | } | |
2457 | ||
2458 | /* Update hashtable on INSN with changed EXPR, old EXPR_OLD_VINSN | |
2459 | and transformation type TRANS_TYPE. */ | |
2460 | static void | |
2461 | update_transformation_cache (expr_t expr, insn_t insn, | |
2462 | bool inside_insn_group, | |
2463 | enum local_trans_type trans_type, | |
2464 | vinsn_t expr_old_vinsn) | |
2465 | { | |
2466 | struct transformed_insns *pti; | |
2467 | ||
2468 | if (inside_insn_group) | |
2469 | return; | |
2470 | ||
2471 | pti = XNEW (struct transformed_insns); | |
2472 | pti->vinsn_old = expr_old_vinsn; | |
2473 | pti->vinsn_new = EXPR_VINSN (expr); | |
2474 | pti->type = trans_type; | |
2475 | pti->was_target_conflict = was_target_conflict; | |
2476 | pti->ds = EXPR_SPEC_DONE_DS (expr); | |
2477 | pti->needs_check = EXPR_NEEDS_SPEC_CHECK_P (expr); | |
2478 | vinsn_attach (pti->vinsn_old); | |
2479 | vinsn_attach (pti->vinsn_new); | |
2480 | *((struct transformed_insns **) | |
2481 | htab_find_slot_with_hash (INSN_TRANSFORMED_INSNS (insn), | |
2482 | pti, VINSN_HASH_RTX (expr_old_vinsn), | |
2483 | INSERT)) = pti; | |
2484 | } | |
2485 | ||
2486 | /* Same as moveup_expr, but first looks up the result of | |
2487 | transformation in caches. */ | |
2488 | static enum MOVEUP_EXPR_CODE | |
2489 | moveup_expr_cached (expr_t expr, insn_t insn, bool inside_insn_group) | |
2490 | { | |
2491 | enum MOVEUP_EXPR_CODE res; | |
2492 | bool got_answer = false; | |
2493 | ||
2494 | if (sched_verbose >= 6) | |
2495 | { | |
2496 | sel_print ("Moving "); | |
2497 | dump_expr (expr); | |
2498 | sel_print (" through %d: ", INSN_UID (insn)); | |
2499 | } | |
2500 | ||
b5b8b0ac AO |
2501 | if (DEBUG_INSN_P (EXPR_INSN_RTX (expr)) |
2502 | && (sel_bb_head (BLOCK_FOR_INSN (EXPR_INSN_RTX (expr))) | |
2503 | == EXPR_INSN_RTX (expr))) | |
2504 | /* Don't use cached information for debug insns that are heads of | |
2505 | basic blocks. */; | |
2506 | else if (try_bitmap_cache (expr, insn, inside_insn_group, &res)) | |
e855c69d AB |
2507 | /* When inside insn group, we do not want remove stores conflicting |
2508 | with previosly issued loads. */ | |
2509 | got_answer = ! inside_insn_group || res != MOVEUP_EXPR_NULL; | |
2510 | else if (try_transformation_cache (expr, insn, &res)) | |
2511 | got_answer = true; | |
2512 | ||
2513 | if (! got_answer) | |
2514 | { | |
2515 | /* Invoke moveup_expr and record the results. */ | |
2516 | vinsn_t expr_old_vinsn = EXPR_VINSN (expr); | |
2517 | ds_t expr_old_spec_ds = EXPR_SPEC_DONE_DS (expr); | |
2518 | int expr_uid = INSN_UID (VINSN_INSN_RTX (expr_old_vinsn)); | |
2519 | bool unique_p = VINSN_UNIQUE_P (expr_old_vinsn); | |
2520 | enum local_trans_type trans_type = TRANS_SUBSTITUTION; | |
2521 | ||
2522 | /* ??? Invent something better than this. We can't allow old_vinsn | |
2523 | to go, we need it for the history vector. */ | |
2524 | vinsn_attach (expr_old_vinsn); | |
2525 | ||
2526 | res = moveup_expr (expr, insn, inside_insn_group, | |
2527 | &trans_type); | |
2528 | switch (res) | |
2529 | { | |
2530 | case MOVEUP_EXPR_NULL: | |
2531 | update_bitmap_cache (expr, insn, inside_insn_group, res); | |
2532 | if (sched_verbose >= 6) | |
2533 | sel_print ("removed\n"); | |
2534 | break; | |
2535 | ||
2536 | case MOVEUP_EXPR_SAME: | |
2537 | update_bitmap_cache (expr, insn, inside_insn_group, res); | |
2538 | if (sched_verbose >= 6) | |
2539 | sel_print ("unchanged\n"); | |
2540 | break; | |
2541 | ||
2542 | case MOVEUP_EXPR_AS_RHS: | |
2543 | gcc_assert (!unique_p || inside_insn_group); | |
2544 | update_bitmap_cache (expr, insn, inside_insn_group, res); | |
2545 | if (sched_verbose >= 6) | |
2546 | sel_print ("unchanged (as RHS)\n"); | |
2547 | break; | |
2548 | ||
2549 | case MOVEUP_EXPR_CHANGED: | |
2550 | gcc_assert (INSN_UID (EXPR_INSN_RTX (expr)) != expr_uid | |
2551 | || EXPR_SPEC_DONE_DS (expr) != expr_old_spec_ds); | |
2552 | insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (expr), | |
2553 | INSN_UID (insn), trans_type, | |
2554 | expr_old_vinsn, EXPR_VINSN (expr), | |
2555 | expr_old_spec_ds); | |
2556 | update_transformation_cache (expr, insn, inside_insn_group, | |
2557 | trans_type, expr_old_vinsn); | |
2558 | if (sched_verbose >= 6) | |
2559 | { | |
2560 | sel_print ("changed: "); | |
2561 | dump_expr (expr); | |
2562 | sel_print ("\n"); | |
2563 | } | |
2564 | break; | |
2565 | default: | |
2566 | gcc_unreachable (); | |
2567 | } | |
2568 | ||
2569 | vinsn_detach (expr_old_vinsn); | |
2570 | } | |
2571 | ||
2572 | return res; | |
2573 | } | |
2574 | ||
2575 | /* Moves an av set AVP up through INSN, performing necessary | |
2576 | transformations. */ | |
2577 | static void | |
2578 | moveup_set_expr (av_set_t *avp, insn_t insn, bool inside_insn_group) | |
2579 | { | |
2580 | av_set_iterator i; | |
2581 | expr_t expr; | |
2582 | ||
2583 | FOR_EACH_EXPR_1 (expr, i, avp) | |
2584 | { | |
2585 | ||
2586 | switch (moveup_expr_cached (expr, insn, inside_insn_group)) | |
2587 | { | |
2588 | case MOVEUP_EXPR_SAME: | |
2589 | case MOVEUP_EXPR_AS_RHS: | |
2590 | break; | |
2591 | ||
2592 | case MOVEUP_EXPR_NULL: | |
2593 | av_set_iter_remove (&i); | |
2594 | break; | |
2595 | ||
2596 | case MOVEUP_EXPR_CHANGED: | |
2597 | expr = merge_with_other_exprs (avp, &i, expr); | |
2598 | break; | |
2599 | ||
2600 | default: | |
2601 | gcc_unreachable (); | |
2602 | } | |
2603 | } | |
2604 | } | |
2605 | ||
2606 | /* Moves AVP set along PATH. */ | |
2607 | static void | |
2608 | moveup_set_inside_insn_group (av_set_t *avp, ilist_t path) | |
2609 | { | |
2610 | int last_cycle; | |
2611 | ||
2612 | if (sched_verbose >= 6) | |
2613 | sel_print ("Moving expressions up in the insn group...\n"); | |
2614 | if (! path) | |
2615 | return; | |
2616 | last_cycle = INSN_SCHED_CYCLE (ILIST_INSN (path)); | |
2617 | while (path | |
2618 | && INSN_SCHED_CYCLE (ILIST_INSN (path)) == last_cycle) | |
2619 | { | |
2620 | moveup_set_expr (avp, ILIST_INSN (path), true); | |
2621 | path = ILIST_NEXT (path); | |
2622 | } | |
2623 | } | |
2624 | ||
2625 | /* Returns true if after moving EXPR along PATH it equals to EXPR_VLIW. */ | |
2626 | static bool | |
2627 | equal_after_moveup_path_p (expr_t expr, ilist_t path, expr_t expr_vliw) | |
2628 | { | |
2629 | expr_def _tmp, *tmp = &_tmp; | |
2630 | int last_cycle; | |
2631 | bool res = true; | |
2632 | ||
2633 | copy_expr_onside (tmp, expr); | |
2634 | last_cycle = path ? INSN_SCHED_CYCLE (ILIST_INSN (path)) : 0; | |
2635 | while (path | |
2636 | && res | |
2637 | && INSN_SCHED_CYCLE (ILIST_INSN (path)) == last_cycle) | |
2638 | { | |
2639 | res = (moveup_expr_cached (tmp, ILIST_INSN (path), true) | |
2640 | != MOVEUP_EXPR_NULL); | |
2641 | path = ILIST_NEXT (path); | |
2642 | } | |
2643 | ||
2644 | if (res) | |
2645 | { | |
2646 | vinsn_t tmp_vinsn = EXPR_VINSN (tmp); | |
2647 | vinsn_t expr_vliw_vinsn = EXPR_VINSN (expr_vliw); | |
2648 | ||
2649 | if (tmp_vinsn != expr_vliw_vinsn) | |
2650 | res = vinsn_equal_p (tmp_vinsn, expr_vliw_vinsn); | |
2651 | } | |
2652 | ||
2653 | clear_expr (tmp); | |
2654 | return res; | |
2655 | } | |
2656 | \f | |
2657 | ||
2658 | /* Functions that compute av and lv sets. */ | |
2659 | ||
2660 | /* Returns true if INSN is not a downward continuation of the given path P in | |
2661 | the current stage. */ | |
2662 | static bool | |
2663 | is_ineligible_successor (insn_t insn, ilist_t p) | |
2664 | { | |
2665 | insn_t prev_insn; | |
2666 | ||
2667 | /* Check if insn is not deleted. */ | |
2668 | if (PREV_INSN (insn) && NEXT_INSN (PREV_INSN (insn)) != insn) | |
2669 | gcc_unreachable (); | |
2670 | else if (NEXT_INSN (insn) && PREV_INSN (NEXT_INSN (insn)) != insn) | |
2671 | gcc_unreachable (); | |
2672 | ||
2673 | /* If it's the first insn visited, then the successor is ok. */ | |
2674 | if (!p) | |
2675 | return false; | |
2676 | ||
2677 | prev_insn = ILIST_INSN (p); | |
2678 | ||
2679 | if (/* a backward edge. */ | |
2680 | INSN_SEQNO (insn) < INSN_SEQNO (prev_insn) | |
2681 | /* is already visited. */ | |
2682 | || (INSN_SEQNO (insn) == INSN_SEQNO (prev_insn) | |
2683 | && (ilist_is_in_p (p, insn) | |
2684 | /* We can reach another fence here and still seqno of insn | |
2685 | would be equal to seqno of prev_insn. This is possible | |
2686 | when prev_insn is a previously created bookkeeping copy. | |
2687 | In that case it'd get a seqno of insn. Thus, check here | |
2688 | whether insn is in current fence too. */ | |
2689 | || IN_CURRENT_FENCE_P (insn))) | |
2690 | /* Was already scheduled on this round. */ | |
2691 | || (INSN_SEQNO (insn) > INSN_SEQNO (prev_insn) | |
2692 | && IN_CURRENT_FENCE_P (insn)) | |
2693 | /* An insn from another fence could also be | |
2694 | scheduled earlier even if this insn is not in | |
2695 | a fence list right now. Check INSN_SCHED_CYCLE instead. */ | |
2696 | || (!pipelining_p | |
2697 | && INSN_SCHED_TIMES (insn) > 0)) | |
2698 | return true; | |
2699 | else | |
2700 | return false; | |
2701 | } | |
2702 | ||
2703 | /* Computes the av_set below the last bb insn INSN, doing all the 'dirty work' | |
2704 | of handling multiple successors and properly merging its av_sets. P is | |
2705 | the current path traversed. WS is the size of lookahead window. | |
2706 | Return the av set computed. */ | |
2707 | static av_set_t | |
2708 | compute_av_set_at_bb_end (insn_t insn, ilist_t p, int ws) | |
2709 | { | |
2710 | struct succs_info *sinfo; | |
2711 | av_set_t expr_in_all_succ_branches = NULL; | |
2712 | int is; | |
2713 | insn_t succ, zero_succ = NULL; | |
2714 | av_set_t av1 = NULL; | |
2715 | ||
2716 | gcc_assert (sel_bb_end_p (insn)); | |
2717 | ||
2718 | /* Find different kind of successors needed for correct computing of | |
2719 | SPEC and TARGET_AVAILABLE attributes. */ | |
2720 | sinfo = compute_succs_info (insn, SUCCS_NORMAL); | |
2721 | ||
2722 | /* Debug output. */ | |
2723 | if (sched_verbose >= 6) | |
2724 | { | |
2725 | sel_print ("successors of bb end (%d): ", INSN_UID (insn)); | |
2726 | dump_insn_vector (sinfo->succs_ok); | |
2727 | sel_print ("\n"); | |
2728 | if (sinfo->succs_ok_n != sinfo->all_succs_n) | |
2729 | sel_print ("real successors num: %d\n", sinfo->all_succs_n); | |
2730 | } | |
2731 | ||
2732 | /* Add insn to to the tail of current path. */ | |
2733 | ilist_add (&p, insn); | |
2734 | ||
2735 | for (is = 0; VEC_iterate (rtx, sinfo->succs_ok, is, succ); is++) | |
2736 | { | |
2737 | av_set_t succ_set; | |
2738 | ||
2739 | /* We will edit SUCC_SET and EXPR_SPEC field of its elements. */ | |
2740 | succ_set = compute_av_set_inside_bb (succ, p, ws, true); | |
2741 | ||
2742 | av_set_split_usefulness (succ_set, | |
2743 | VEC_index (int, sinfo->probs_ok, is), | |
2744 | sinfo->all_prob); | |
2745 | ||
2746 | if (sinfo->all_succs_n > 1 | |
2747 | && sinfo->all_succs_n == sinfo->succs_ok_n) | |
2748 | { | |
2749 | /* Find EXPR'es that came from *all* successors and save them | |
2750 | into expr_in_all_succ_branches. This set will be used later | |
2751 | for calculating speculation attributes of EXPR'es. */ | |
2752 | if (is == 0) | |
2753 | { | |
2754 | expr_in_all_succ_branches = av_set_copy (succ_set); | |
2755 | ||
2756 | /* Remember the first successor for later. */ | |
2757 | zero_succ = succ; | |
2758 | } | |
2759 | else | |
2760 | { | |
2761 | av_set_iterator i; | |
2762 | expr_t expr; | |
2763 | ||
2764 | FOR_EACH_EXPR_1 (expr, i, &expr_in_all_succ_branches) | |
2765 | if (!av_set_is_in_p (succ_set, EXPR_VINSN (expr))) | |
2766 | av_set_iter_remove (&i); | |
2767 | } | |
2768 | } | |
2769 | ||
2770 | /* Union the av_sets. Check liveness restrictions on target registers | |
2771 | in special case of two successors. */ | |
2772 | if (sinfo->succs_ok_n == 2 && is == 1) | |
2773 | { | |
2774 | basic_block bb0 = BLOCK_FOR_INSN (zero_succ); | |
2775 | basic_block bb1 = BLOCK_FOR_INSN (succ); | |
2776 | ||
2777 | gcc_assert (BB_LV_SET_VALID_P (bb0) && BB_LV_SET_VALID_P (bb1)); | |
2778 | av_set_union_and_live (&av1, &succ_set, | |
2779 | BB_LV_SET (bb0), | |
2780 | BB_LV_SET (bb1), | |
2781 | insn); | |
2782 | } | |
2783 | else | |
2784 | av_set_union_and_clear (&av1, &succ_set, insn); | |
2785 | } | |
2786 | ||
2787 | /* Check liveness restrictions via hard way when there are more than | |
2788 | two successors. */ | |
2789 | if (sinfo->succs_ok_n > 2) | |
2790 | for (is = 0; VEC_iterate (rtx, sinfo->succs_ok, is, succ); is++) | |
2791 | { | |
2792 | basic_block succ_bb = BLOCK_FOR_INSN (succ); | |
2793 | ||
2794 | gcc_assert (BB_LV_SET_VALID_P (succ_bb)); | |
2795 | mark_unavailable_targets (av1, BB_AV_SET (succ_bb), | |
2796 | BB_LV_SET (succ_bb)); | |
2797 | } | |
2798 | ||
2799 | /* Finally, check liveness restrictions on paths leaving the region. */ | |
2800 | if (sinfo->all_succs_n > sinfo->succs_ok_n) | |
2801 | for (is = 0; VEC_iterate (rtx, sinfo->succs_other, is, succ); is++) | |
2802 | mark_unavailable_targets | |
2803 | (av1, NULL, BB_LV_SET (BLOCK_FOR_INSN (succ))); | |
2804 | ||
2805 | if (sinfo->all_succs_n > 1) | |
2806 | { | |
2807 | av_set_iterator i; | |
2808 | expr_t expr; | |
2809 | ||
2810 | /* Increase the spec attribute of all EXPR'es that didn't come | |
2811 | from all successors. */ | |
2812 | FOR_EACH_EXPR (expr, i, av1) | |
2813 | if (!av_set_is_in_p (expr_in_all_succ_branches, EXPR_VINSN (expr))) | |
2814 | EXPR_SPEC (expr)++; | |
2815 | ||
2816 | av_set_clear (&expr_in_all_succ_branches); | |
2817 | ||
2818 | /* Do not move conditional branches through other | |
2819 | conditional branches. So, remove all conditional | |
2820 | branches from av_set if current operator is a conditional | |
2821 | branch. */ | |
2822 | av_set_substract_cond_branches (&av1); | |
2823 | } | |
2824 | ||
2825 | ilist_remove (&p); | |
2826 | free_succs_info (sinfo); | |
2827 | ||
2828 | if (sched_verbose >= 6) | |
2829 | { | |
2830 | sel_print ("av_succs (%d): ", INSN_UID (insn)); | |
2831 | dump_av_set (av1); | |
2832 | sel_print ("\n"); | |
2833 | } | |
2834 | ||
2835 | return av1; | |
2836 | } | |
2837 | ||
2838 | /* This function computes av_set for the FIRST_INSN by dragging valid | |
2839 | av_set through all basic block insns either from the end of basic block | |
2840 | (computed using compute_av_set_at_bb_end) or from the insn on which | |
2841 | MAX_WS was exceeded. It uses compute_av_set_at_bb_end to compute av_set | |
2842 | below the basic block and handling conditional branches. | |
2843 | FIRST_INSN - the basic block head, P - path consisting of the insns | |
2844 | traversed on the way to the FIRST_INSN (the path is sparse, only bb heads | |
2845 | and bb ends are added to the path), WS - current window size, | |
2846 | NEED_COPY_P - true if we'll make a copy of av_set before returning it. */ | |
2847 | static av_set_t | |
2848 | compute_av_set_inside_bb (insn_t first_insn, ilist_t p, int ws, | |
2849 | bool need_copy_p) | |
2850 | { | |
2851 | insn_t cur_insn; | |
2852 | int end_ws = ws; | |
2853 | insn_t bb_end = sel_bb_end (BLOCK_FOR_INSN (first_insn)); | |
2854 | insn_t after_bb_end = NEXT_INSN (bb_end); | |
2855 | insn_t last_insn; | |
2856 | av_set_t av = NULL; | |
2857 | basic_block cur_bb = BLOCK_FOR_INSN (first_insn); | |
2858 | ||
2859 | /* Return NULL if insn is not on the legitimate downward path. */ | |
2860 | if (is_ineligible_successor (first_insn, p)) | |
2861 | { | |
2862 | if (sched_verbose >= 6) | |
2863 | sel_print ("Insn %d is ineligible_successor\n", INSN_UID (first_insn)); | |
2864 | ||
2865 | return NULL; | |
2866 | } | |
2867 | ||
2868 | /* If insn already has valid av(insn) computed, just return it. */ | |
2869 | if (AV_SET_VALID_P (first_insn)) | |
2870 | { | |
2871 | av_set_t av_set; | |
2872 | ||
2873 | if (sel_bb_head_p (first_insn)) | |
2874 | av_set = BB_AV_SET (BLOCK_FOR_INSN (first_insn)); | |
2875 | else | |
2876 | av_set = NULL; | |
2877 | ||
2878 | if (sched_verbose >= 6) | |
2879 | { | |
2880 | sel_print ("Insn %d has a valid av set: ", INSN_UID (first_insn)); | |
2881 | dump_av_set (av_set); | |
2882 | sel_print ("\n"); | |
2883 | } | |
2884 | ||
2885 | return need_copy_p ? av_set_copy (av_set) : av_set; | |
2886 | } | |
2887 | ||
2888 | ilist_add (&p, first_insn); | |
2889 | ||
2890 | /* As the result after this loop have completed, in LAST_INSN we'll | |
2891 | have the insn which has valid av_set to start backward computation | |
2892 | from: it either will be NULL because on it the window size was exceeded | |
2893 | or other valid av_set as returned by compute_av_set for the last insn | |
2894 | of the basic block. */ | |
2895 | for (last_insn = first_insn; last_insn != after_bb_end; | |
2896 | last_insn = NEXT_INSN (last_insn)) | |
2897 | { | |
2898 | /* We may encounter valid av_set not only on bb_head, but also on | |
2899 | those insns on which previously MAX_WS was exceeded. */ | |
2900 | if (AV_SET_VALID_P (last_insn)) | |
2901 | { | |
2902 | if (sched_verbose >= 6) | |
2903 | sel_print ("Insn %d has a valid empty av set\n", INSN_UID (last_insn)); | |
2904 | break; | |
2905 | } | |
2906 | ||
2907 | /* The special case: the last insn of the BB may be an | |
2908 | ineligible_successor due to its SEQ_NO that was set on | |
2909 | it as a bookkeeping. */ | |
2910 | if (last_insn != first_insn | |
2911 | && is_ineligible_successor (last_insn, p)) | |
2912 | { | |
2913 | if (sched_verbose >= 6) | |
2914 | sel_print ("Insn %d is ineligible_successor\n", INSN_UID (last_insn)); | |
2915 | break; | |
2916 | } | |
2917 | ||
b5b8b0ac AO |
2918 | if (DEBUG_INSN_P (last_insn)) |
2919 | continue; | |
2920 | ||
e855c69d AB |
2921 | if (end_ws > max_ws) |
2922 | { | |
2923 | /* We can reach max lookahead size at bb_header, so clean av_set | |
2924 | first. */ | |
2925 | INSN_WS_LEVEL (last_insn) = global_level; | |
2926 | ||
2927 | if (sched_verbose >= 6) | |
2928 | sel_print ("Insn %d is beyond the software lookahead window size\n", | |
2929 | INSN_UID (last_insn)); | |
2930 | break; | |
2931 | } | |
2932 | ||
2933 | end_ws++; | |
2934 | } | |
2935 | ||
2936 | /* Get the valid av_set into AV above the LAST_INSN to start backward | |
2937 | computation from. It either will be empty av_set or av_set computed from | |
2938 | the successors on the last insn of the current bb. */ | |
2939 | if (last_insn != after_bb_end) | |
2940 | { | |
2941 | av = NULL; | |
2942 | ||
2943 | /* This is needed only to obtain av_sets that are identical to | |
2944 | those computed by the old compute_av_set version. */ | |
2945 | if (last_insn == first_insn && !INSN_NOP_P (last_insn)) | |
2946 | av_set_add (&av, INSN_EXPR (last_insn)); | |
2947 | } | |
2948 | else | |
2949 | /* END_WS is always already increased by 1 if LAST_INSN == AFTER_BB_END. */ | |
2950 | av = compute_av_set_at_bb_end (bb_end, p, end_ws); | |
2951 | ||
2952 | /* Compute av_set in AV starting from below the LAST_INSN up to | |
2953 | location above the FIRST_INSN. */ | |
2954 | for (cur_insn = PREV_INSN (last_insn); cur_insn != PREV_INSN (first_insn); | |
2955 | cur_insn = PREV_INSN (cur_insn)) | |
2956 | if (!INSN_NOP_P (cur_insn)) | |
2957 | { | |
2958 | expr_t expr; | |
2959 | ||
2960 | moveup_set_expr (&av, cur_insn, false); | |
2961 | ||
2962 | /* If the expression for CUR_INSN is already in the set, | |
2963 | replace it by the new one. */ | |
2964 | expr = av_set_lookup (av, INSN_VINSN (cur_insn)); | |
2965 | if (expr != NULL) | |
2966 | { | |
2967 | clear_expr (expr); | |
2968 | copy_expr (expr, INSN_EXPR (cur_insn)); | |
2969 | } | |
2970 | else | |
2971 | av_set_add (&av, INSN_EXPR (cur_insn)); | |
2972 | } | |
2973 | ||
2974 | /* Clear stale bb_av_set. */ | |
2975 | if (sel_bb_head_p (first_insn)) | |
2976 | { | |
2977 | av_set_clear (&BB_AV_SET (cur_bb)); | |
2978 | BB_AV_SET (cur_bb) = need_copy_p ? av_set_copy (av) : av; | |
2979 | BB_AV_LEVEL (cur_bb) = global_level; | |
2980 | } | |
2981 | ||
2982 | if (sched_verbose >= 6) | |
2983 | { | |
2984 | sel_print ("Computed av set for insn %d: ", INSN_UID (first_insn)); | |
2985 | dump_av_set (av); | |
2986 | sel_print ("\n"); | |
2987 | } | |
2988 | ||
2989 | ilist_remove (&p); | |
2990 | return av; | |
2991 | } | |
2992 | ||
2993 | /* Compute av set before INSN. | |
2994 | INSN - the current operation (actual rtx INSN) | |
2995 | P - the current path, which is list of insns visited so far | |
2996 | WS - software lookahead window size. | |
2997 | UNIQUE_P - TRUE, if returned av_set will be changed, hence | |
2998 | if we want to save computed av_set in s_i_d, we should make a copy of it. | |
2999 | ||
3000 | In the resulting set we will have only expressions that don't have delay | |
3001 | stalls and nonsubstitutable dependences. */ | |
3002 | static av_set_t | |
3003 | compute_av_set (insn_t insn, ilist_t p, int ws, bool unique_p) | |
3004 | { | |
3005 | return compute_av_set_inside_bb (insn, p, ws, unique_p); | |
3006 | } | |
3007 | ||
3008 | /* Propagate a liveness set LV through INSN. */ | |
3009 | static void | |
3010 | propagate_lv_set (regset lv, insn_t insn) | |
3011 | { | |
3012 | gcc_assert (INSN_P (insn)); | |
3013 | ||
3014 | if (INSN_NOP_P (insn)) | |
3015 | return; | |
3016 | ||
02b47899 | 3017 | df_simulate_one_insn_backwards (BLOCK_FOR_INSN (insn), insn, lv); |
e855c69d AB |
3018 | } |
3019 | ||
3020 | /* Return livness set at the end of BB. */ | |
3021 | static regset | |
3022 | compute_live_after_bb (basic_block bb) | |
3023 | { | |
3024 | edge e; | |
3025 | edge_iterator ei; | |
3026 | regset lv = get_clear_regset_from_pool (); | |
3027 | ||
3028 | gcc_assert (!ignore_first); | |
3029 | ||
3030 | FOR_EACH_EDGE (e, ei, bb->succs) | |
3031 | if (sel_bb_empty_p (e->dest)) | |
3032 | { | |
3033 | if (! BB_LV_SET_VALID_P (e->dest)) | |
3034 | { | |
3035 | gcc_unreachable (); | |
3036 | gcc_assert (BB_LV_SET (e->dest) == NULL); | |
3037 | BB_LV_SET (e->dest) = compute_live_after_bb (e->dest); | |
3038 | BB_LV_SET_VALID_P (e->dest) = true; | |
3039 | } | |
3040 | IOR_REG_SET (lv, BB_LV_SET (e->dest)); | |
3041 | } | |
3042 | else | |
3043 | IOR_REG_SET (lv, compute_live (sel_bb_head (e->dest))); | |
3044 | ||
3045 | return lv; | |
3046 | } | |
3047 | ||
3048 | /* Compute the set of all live registers at the point before INSN and save | |
3049 | it at INSN if INSN is bb header. */ | |
3050 | regset | |
3051 | compute_live (insn_t insn) | |
3052 | { | |
3053 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3054 | insn_t final, temp; | |
3055 | regset lv; | |
3056 | ||
3057 | /* Return the valid set if we're already on it. */ | |
3058 | if (!ignore_first) | |
3059 | { | |
3060 | regset src = NULL; | |
3061 | ||
3062 | if (sel_bb_head_p (insn) && BB_LV_SET_VALID_P (bb)) | |
3063 | src = BB_LV_SET (bb); | |
3064 | else | |
3065 | { | |
3066 | gcc_assert (in_current_region_p (bb)); | |
3067 | if (INSN_LIVE_VALID_P (insn)) | |
3068 | src = INSN_LIVE (insn); | |
3069 | } | |
3070 | ||
3071 | if (src) | |
3072 | { | |
3073 | lv = get_regset_from_pool (); | |
3074 | COPY_REG_SET (lv, src); | |
3075 | ||
3076 | if (sel_bb_head_p (insn) && ! BB_LV_SET_VALID_P (bb)) | |
3077 | { | |
3078 | COPY_REG_SET (BB_LV_SET (bb), lv); | |
3079 | BB_LV_SET_VALID_P (bb) = true; | |
3080 | } | |
3081 | ||
3082 | return_regset_to_pool (lv); | |
3083 | return lv; | |
3084 | } | |
3085 | } | |
3086 | ||
3087 | /* We've skipped the wrong lv_set. Don't skip the right one. */ | |
3088 | ignore_first = false; | |
3089 | gcc_assert (in_current_region_p (bb)); | |
3090 | ||
3091 | /* Find a valid LV set in this block or below, if needed. | |
3092 | Start searching from the next insn: either ignore_first is true, or | |
3093 | INSN doesn't have a correct live set. */ | |
3094 | temp = NEXT_INSN (insn); | |
3095 | final = NEXT_INSN (BB_END (bb)); | |
3096 | while (temp != final && ! INSN_LIVE_VALID_P (temp)) | |
3097 | temp = NEXT_INSN (temp); | |
3098 | if (temp == final) | |
3099 | { | |
3100 | lv = compute_live_after_bb (bb); | |
3101 | temp = PREV_INSN (temp); | |
3102 | } | |
3103 | else | |
3104 | { | |
3105 | lv = get_regset_from_pool (); | |
3106 | COPY_REG_SET (lv, INSN_LIVE (temp)); | |
3107 | } | |
3108 | ||
3109 | /* Put correct lv sets on the insns which have bad sets. */ | |
3110 | final = PREV_INSN (insn); | |
3111 | while (temp != final) | |
3112 | { | |
3113 | propagate_lv_set (lv, temp); | |
3114 | COPY_REG_SET (INSN_LIVE (temp), lv); | |
3115 | INSN_LIVE_VALID_P (temp) = true; | |
3116 | temp = PREV_INSN (temp); | |
3117 | } | |
3118 | ||
3119 | /* Also put it in a BB. */ | |
3120 | if (sel_bb_head_p (insn)) | |
3121 | { | |
3122 | basic_block bb = BLOCK_FOR_INSN (insn); | |
3123 | ||
3124 | COPY_REG_SET (BB_LV_SET (bb), lv); | |
3125 | BB_LV_SET_VALID_P (bb) = true; | |
3126 | } | |
3127 | ||
3128 | /* We return LV to the pool, but will not clear it there. Thus we can | |
3129 | legimatelly use LV till the next use of regset_pool_get (). */ | |
3130 | return_regset_to_pool (lv); | |
3131 | return lv; | |
3132 | } | |
3133 | ||
3134 | /* Update liveness sets for INSN. */ | |
3135 | static inline void | |
3136 | update_liveness_on_insn (rtx insn) | |
3137 | { | |
3138 | ignore_first = true; | |
3139 | compute_live (insn); | |
3140 | } | |
3141 | ||
3142 | /* Compute liveness below INSN and write it into REGS. */ | |
3143 | static inline void | |
3144 | compute_live_below_insn (rtx insn, regset regs) | |
3145 | { | |
3146 | rtx succ; | |
3147 | succ_iterator si; | |
3148 | ||
3149 | FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL) | |
3150 | IOR_REG_SET (regs, compute_live (succ)); | |
3151 | } | |
3152 | ||
3153 | /* Update the data gathered in av and lv sets starting from INSN. */ | |
3154 | static void | |
3155 | update_data_sets (rtx insn) | |
3156 | { | |
3157 | update_liveness_on_insn (insn); | |
3158 | if (sel_bb_head_p (insn)) | |
3159 | { | |
3160 | gcc_assert (AV_LEVEL (insn) != 0); | |
3161 | BB_AV_LEVEL (BLOCK_FOR_INSN (insn)) = -1; | |
3162 | compute_av_set (insn, NULL, 0, 0); | |
3163 | } | |
3164 | } | |
3165 | \f | |
3166 | ||
3167 | /* Helper for move_op () and find_used_regs (). | |
3168 | Return speculation type for which a check should be created on the place | |
3169 | of INSN. EXPR is one of the original ops we are searching for. */ | |
3170 | static ds_t | |
3171 | get_spec_check_type_for_insn (insn_t insn, expr_t expr) | |
3172 | { | |
3173 | ds_t to_check_ds; | |
3174 | ds_t already_checked_ds = EXPR_SPEC_DONE_DS (INSN_EXPR (insn)); | |
3175 | ||
3176 | to_check_ds = EXPR_SPEC_TO_CHECK_DS (expr); | |
3177 | ||
3178 | if (targetm.sched.get_insn_checked_ds) | |
3179 | already_checked_ds |= targetm.sched.get_insn_checked_ds (insn); | |
3180 | ||
3181 | if (spec_info != NULL | |
3182 | && (spec_info->flags & SEL_SCHED_SPEC_DONT_CHECK_CONTROL)) | |
3183 | already_checked_ds |= BEGIN_CONTROL; | |
3184 | ||
3185 | already_checked_ds = ds_get_speculation_types (already_checked_ds); | |
3186 | ||
3187 | to_check_ds &= ~already_checked_ds; | |
3188 | ||
3189 | return to_check_ds; | |
3190 | } | |
3191 | ||
3192 | /* Find the set of registers that are unavailable for storing expres | |
3193 | while moving ORIG_OPS up on the path starting from INSN due to | |
3194 | liveness (USED_REGS) or hardware restrictions (REG_RENAME_P). | |
3195 | ||
3196 | All the original operations found during the traversal are saved in the | |
3197 | ORIGINAL_INSNS list. | |
3198 | ||
3199 | REG_RENAME_P denotes the set of hardware registers that | |
3200 | can not be used with renaming due to the register class restrictions, | |
3201 | mode restrictions and other (the register we'll choose should be | |
3202 | compatible class with the original uses, shouldn't be in call_used_regs, | |
3203 | should be HARD_REGNO_RENAME_OK etc). | |
3204 | ||
3205 | Returns TRUE if we've found all original insns, FALSE otherwise. | |
3206 | ||
3207 | This function utilizes code_motion_path_driver (formerly find_used_regs_1) | |
3208 | to traverse the code motion paths. This helper function finds registers | |
3209 | that are not available for storing expres while moving ORIG_OPS up on the | |
3210 | path starting from INSN. A register considered as used on the moving path, | |
3211 | if one of the following conditions is not satisfied: | |
3212 | ||
3213 | (1) a register not set or read on any path from xi to an instance of | |
3214 | the original operation, | |
3215 | (2) not among the live registers of the point immediately following the | |
3216 | first original operation on a given downward path, except for the | |
3217 | original target register of the operation, | |
3218 | (3) not live on the other path of any conditional branch that is passed | |
3219 | by the operation, in case original operations are not present on | |
3220 | both paths of the conditional branch. | |
3221 | ||
3222 | All the original operations found during the traversal are saved in the | |
3223 | ORIGINAL_INSNS list. | |
3224 | ||
3225 | REG_RENAME_P->CROSSES_CALL is true, if there is a call insn on the path | |
3226 | from INSN to original insn. In this case CALL_USED_REG_SET will be added | |
3227 | to unavailable hard regs at the point original operation is found. */ | |
3228 | ||
3229 | static bool | |
3230 | find_used_regs (insn_t insn, av_set_t orig_ops, regset used_regs, | |
3231 | struct reg_rename *reg_rename_p, def_list_t *original_insns) | |
3232 | { | |
3233 | def_list_iterator i; | |
3234 | def_t def; | |
3235 | int res; | |
3236 | bool needs_spec_check_p = false; | |
3237 | expr_t expr; | |
3238 | av_set_iterator expr_iter; | |
3239 | struct fur_static_params sparams; | |
3240 | struct cmpd_local_params lparams; | |
3241 | ||
3242 | /* We haven't visited any blocks yet. */ | |
3243 | bitmap_clear (code_motion_visited_blocks); | |
3244 | ||
3245 | /* Init parameters for code_motion_path_driver. */ | |
3246 | sparams.crosses_call = false; | |
3247 | sparams.original_insns = original_insns; | |
3248 | sparams.used_regs = used_regs; | |
3249 | ||
3250 | /* Set the appropriate hooks and data. */ | |
3251 | code_motion_path_driver_info = &fur_hooks; | |
3252 | ||
3253 | res = code_motion_path_driver (insn, orig_ops, NULL, &lparams, &sparams); | |
3254 | ||
3255 | reg_rename_p->crosses_call |= sparams.crosses_call; | |
3256 | ||
3257 | gcc_assert (res == 1); | |
3258 | gcc_assert (original_insns && *original_insns); | |
3259 | ||
3260 | /* ??? We calculate whether an expression needs a check when computing | |
3261 | av sets. This information is not as precise as it could be due to | |
3262 | merging this bit in merge_expr. We can do better in find_used_regs, | |
3263 | but we want to avoid multiple traversals of the same code motion | |
3264 | paths. */ | |
3265 | FOR_EACH_EXPR (expr, expr_iter, orig_ops) | |
3266 | needs_spec_check_p |= EXPR_NEEDS_SPEC_CHECK_P (expr); | |
3267 | ||
3268 | /* Mark hardware regs in REG_RENAME_P that are not suitable | |
3269 | for renaming expr in INSN due to hardware restrictions (register class, | |
3270 | modes compatibility etc). */ | |
3271 | FOR_EACH_DEF (def, i, *original_insns) | |
3272 | { | |
3273 | vinsn_t vinsn = INSN_VINSN (def->orig_insn); | |
3274 | ||
3275 | if (VINSN_SEPARABLE_P (vinsn)) | |
3276 | mark_unavailable_hard_regs (def, reg_rename_p, used_regs); | |
3277 | ||
3278 | /* Do not allow clobbering of ld.[sa] address in case some of the | |
3279 | original operations need a check. */ | |
3280 | if (needs_spec_check_p) | |
3281 | IOR_REG_SET (used_regs, VINSN_REG_USES (vinsn)); | |
3282 | } | |
3283 | ||
3284 | return true; | |
3285 | } | |
3286 | \f | |
3287 | ||
3288 | /* Functions to choose the best insn from available ones. */ | |
3289 | ||
3290 | /* Adjusts the priority for EXPR using the backend *_adjust_priority hook. */ | |
3291 | static int | |
3292 | sel_target_adjust_priority (expr_t expr) | |
3293 | { | |
3294 | int priority = EXPR_PRIORITY (expr); | |
3295 | int new_priority; | |
3296 | ||
3297 | if (targetm.sched.adjust_priority) | |
3298 | new_priority = targetm.sched.adjust_priority (EXPR_INSN_RTX (expr), priority); | |
3299 | else | |
3300 | new_priority = priority; | |
3301 | ||
3302 | /* If the priority has changed, adjust EXPR_PRIORITY_ADJ accordingly. */ | |
3303 | EXPR_PRIORITY_ADJ (expr) = new_priority - EXPR_PRIORITY (expr); | |
3304 | ||
3305 | gcc_assert (EXPR_PRIORITY_ADJ (expr) >= 0); | |
3306 | ||
3307 | if (sched_verbose >= 2) | |
3308 | sel_print ("sel_target_adjust_priority: insn %d, %d +%d = %d.\n", | |
3309 | INSN_UID (EXPR_INSN_RTX (expr)), EXPR_PRIORITY (expr), | |
3310 | EXPR_PRIORITY_ADJ (expr), new_priority); | |
3311 | ||
3312 | return new_priority; | |
3313 | } | |
3314 | ||
3315 | /* Rank two available exprs for schedule. Never return 0 here. */ | |
3316 | static int | |
3317 | sel_rank_for_schedule (const void *x, const void *y) | |
3318 | { | |
3319 | expr_t tmp = *(const expr_t *) y; | |
3320 | expr_t tmp2 = *(const expr_t *) x; | |
3321 | insn_t tmp_insn, tmp2_insn; | |
3322 | vinsn_t tmp_vinsn, tmp2_vinsn; | |
3323 | int val; | |
3324 | ||
3325 | tmp_vinsn = EXPR_VINSN (tmp); | |
3326 | tmp2_vinsn = EXPR_VINSN (tmp2); | |
3327 | tmp_insn = EXPR_INSN_RTX (tmp); | |
3328 | tmp2_insn = EXPR_INSN_RTX (tmp2); | |
3329 | ||
b5b8b0ac AO |
3330 | /* Schedule debug insns as early as possible. */ |
3331 | if (DEBUG_INSN_P (tmp_insn) && !DEBUG_INSN_P (tmp2_insn)) | |
3332 | return -1; | |
3333 | else if (DEBUG_INSN_P (tmp2_insn)) | |
3334 | return 1; | |
3335 | ||
e855c69d AB |
3336 | /* Prefer SCHED_GROUP_P insns to any others. */ |
3337 | if (SCHED_GROUP_P (tmp_insn) != SCHED_GROUP_P (tmp2_insn)) | |
3338 | { | |
3339 | if (VINSN_UNIQUE_P (tmp_vinsn) && VINSN_UNIQUE_P (tmp2_vinsn)) | |
3340 | return SCHED_GROUP_P (tmp2_insn) ? 1 : -1; | |
3341 | ||
3342 | /* Now uniqueness means SCHED_GROUP_P is set, because schedule groups | |
3343 | cannot be cloned. */ | |
3344 | if (VINSN_UNIQUE_P (tmp2_vinsn)) | |
3345 | return 1; | |
3346 | return -1; | |
3347 | } | |
3348 | ||
3349 | /* Discourage scheduling of speculative checks. */ | |
3350 | val = (sel_insn_is_speculation_check (tmp_insn) | |
3351 | - sel_insn_is_speculation_check (tmp2_insn)); | |
3352 | if (val) | |
3353 | return val; | |
3354 | ||
3355 | /* Prefer not scheduled insn over scheduled one. */ | |
3356 | if (EXPR_SCHED_TIMES (tmp) > 0 || EXPR_SCHED_TIMES (tmp2) > 0) | |
3357 | { | |
3358 | val = EXPR_SCHED_TIMES (tmp) - EXPR_SCHED_TIMES (tmp2); | |
3359 | if (val) | |
3360 | return val; | |
3361 | } | |
3362 | ||
3363 | /* Prefer jump over non-jump instruction. */ | |
3364 | if (control_flow_insn_p (tmp_insn) && !control_flow_insn_p (tmp2_insn)) | |
3365 | return -1; | |
3366 | else if (control_flow_insn_p (tmp2_insn) && !control_flow_insn_p (tmp_insn)) | |
3367 | return 1; | |
3368 | ||
3369 | /* Prefer an expr with greater priority. */ | |
3370 | if (EXPR_USEFULNESS (tmp) != 0 && EXPR_USEFULNESS (tmp2) != 0) | |
3371 | { | |
3372 | int p2 = EXPR_PRIORITY (tmp2) + EXPR_PRIORITY_ADJ (tmp2), | |
3373 | p1 = EXPR_PRIORITY (tmp) + EXPR_PRIORITY_ADJ (tmp); | |
3374 | ||
3375 | val = p2 * EXPR_USEFULNESS (tmp2) - p1 * EXPR_USEFULNESS (tmp); | |
3376 | } | |
3377 | else | |
3378 | val = EXPR_PRIORITY (tmp2) - EXPR_PRIORITY (tmp) | |
3379 | + EXPR_PRIORITY_ADJ (tmp2) - EXPR_PRIORITY_ADJ (tmp); | |
3380 | if (val) | |
3381 | return val; | |
3382 | ||
3383 | if (spec_info != NULL && spec_info->mask != 0) | |
3384 | /* This code was taken from haifa-sched.c: rank_for_schedule (). */ | |
3385 | { | |
3386 | ds_t ds1, ds2; | |
3387 | dw_t dw1, dw2; | |
3388 | int dw; | |
3389 | ||
3390 | ds1 = EXPR_SPEC_DONE_DS (tmp); | |
3391 | if (ds1) | |
3392 | dw1 = ds_weak (ds1); | |
3393 | else | |
3394 | dw1 = NO_DEP_WEAK; | |
3395 | ||
3396 | ds2 = EXPR_SPEC_DONE_DS (tmp2); | |
3397 | if (ds2) | |
3398 | dw2 = ds_weak (ds2); | |
3399 | else | |
3400 | dw2 = NO_DEP_WEAK; | |
3401 | ||
3402 | dw = dw2 - dw1; | |
3403 | if (dw > (NO_DEP_WEAK / 8) || dw < -(NO_DEP_WEAK / 8)) | |
3404 | return dw; | |
3405 | } | |
3406 | ||
e855c69d AB |
3407 | /* Prefer an old insn to a bookkeeping insn. */ |
3408 | if (INSN_UID (tmp_insn) < first_emitted_uid | |
3409 | && INSN_UID (tmp2_insn) >= first_emitted_uid) | |
3410 | return -1; | |
3411 | if (INSN_UID (tmp_insn) >= first_emitted_uid | |
3412 | && INSN_UID (tmp2_insn) < first_emitted_uid) | |
3413 | return 1; | |
3414 | ||
3415 | /* Prefer an insn with smaller UID, as a last resort. | |
3416 | We can't safely use INSN_LUID as it is defined only for those insns | |
3417 | that are in the stream. */ | |
3418 | return INSN_UID (tmp_insn) - INSN_UID (tmp2_insn); | |
3419 | } | |
3420 | ||
3421 | /* Filter out expressions from av set pointed to by AV_PTR | |
3422 | that are pipelined too many times. */ | |
3423 | static void | |
3424 | process_pipelined_exprs (av_set_t *av_ptr) | |
3425 | { | |
3426 | expr_t expr; | |
3427 | av_set_iterator si; | |
3428 | ||
3429 | /* Don't pipeline already pipelined code as that would increase | |
3430 | number of unnecessary register moves. */ | |
3431 | FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3432 | { | |
3433 | if (EXPR_SCHED_TIMES (expr) | |
3434 | >= PARAM_VALUE (PARAM_SELSCHED_MAX_SCHED_TIMES)) | |
3435 | av_set_iter_remove (&si); | |
3436 | } | |
3437 | } | |
3438 | ||
3439 | /* Filter speculative insns from AV_PTR if we don't want them. */ | |
3440 | static void | |
3441 | process_spec_exprs (av_set_t *av_ptr) | |
3442 | { | |
3443 | bool try_data_p = true; | |
3444 | bool try_control_p = true; | |
3445 | expr_t expr; | |
3446 | av_set_iterator si; | |
3447 | ||
3448 | if (spec_info == NULL) | |
3449 | return; | |
3450 | ||
3451 | /* Scan *AV_PTR to find out if we want to consider speculative | |
3452 | instructions for scheduling. */ | |
3453 | FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3454 | { | |
3455 | ds_t ds; | |
3456 | ||
3457 | ds = EXPR_SPEC_DONE_DS (expr); | |
3458 | ||
3459 | /* The probability of a success is too low - don't speculate. */ | |
3460 | if ((ds & SPECULATIVE) | |
3461 | && (ds_weak (ds) < spec_info->data_weakness_cutoff | |
3462 | || EXPR_USEFULNESS (expr) < spec_info->control_weakness_cutoff | |
3463 | || (pipelining_p && false | |
3464 | && (ds & DATA_SPEC) | |
3465 | && (ds & CONTROL_SPEC)))) | |
3466 | { | |
3467 | av_set_iter_remove (&si); | |
3468 | continue; | |
3469 | } | |
3470 | ||
3471 | if ((spec_info->flags & PREFER_NON_DATA_SPEC) | |
3472 | && !(ds & BEGIN_DATA)) | |
3473 | try_data_p = false; | |
3474 | ||
3475 | if ((spec_info->flags & PREFER_NON_CONTROL_SPEC) | |
3476 | && !(ds & BEGIN_CONTROL)) | |
3477 | try_control_p = false; | |
3478 | } | |
3479 | ||
3480 | FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3481 | { | |
3482 | ds_t ds; | |
3483 | ||
3484 | ds = EXPR_SPEC_DONE_DS (expr); | |
3485 | ||
3486 | if (ds & SPECULATIVE) | |
3487 | { | |
3488 | if ((ds & BEGIN_DATA) && !try_data_p) | |
3489 | /* We don't want any data speculative instructions right | |
3490 | now. */ | |
3491 | av_set_iter_remove (&si); | |
3492 | ||
3493 | if ((ds & BEGIN_CONTROL) && !try_control_p) | |
3494 | /* We don't want any control speculative instructions right | |
3495 | now. */ | |
3496 | av_set_iter_remove (&si); | |
3497 | } | |
3498 | } | |
3499 | } | |
3500 | ||
3501 | /* Search for any use-like insns in AV_PTR and decide on scheduling | |
3502 | them. Return one when found, and NULL otherwise. | |
3503 | Note that we check here whether a USE could be scheduled to avoid | |
3504 | an infinite loop later. */ | |
3505 | static expr_t | |
3506 | process_use_exprs (av_set_t *av_ptr) | |
3507 | { | |
3508 | expr_t expr; | |
3509 | av_set_iterator si; | |
3510 | bool uses_present_p = false; | |
3511 | bool try_uses_p = true; | |
3512 | ||
3513 | FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3514 | { | |
3515 | /* This will also initialize INSN_CODE for later use. */ | |
3516 | if (recog_memoized (EXPR_INSN_RTX (expr)) < 0) | |
3517 | { | |
3518 | /* If we have a USE in *AV_PTR that was not scheduled yet, | |
3519 | do so because it will do good only. */ | |
3520 | if (EXPR_SCHED_TIMES (expr) <= 0) | |
3521 | { | |
3522 | if (EXPR_TARGET_AVAILABLE (expr) == 1) | |
3523 | return expr; | |
3524 | ||
3525 | av_set_iter_remove (&si); | |
3526 | } | |
3527 | else | |
3528 | { | |
3529 | gcc_assert (pipelining_p); | |
3530 | ||
3531 | uses_present_p = true; | |
3532 | } | |
3533 | } | |
3534 | else | |
3535 | try_uses_p = false; | |
3536 | } | |
3537 | ||
3538 | if (uses_present_p) | |
3539 | { | |
3540 | /* If we don't want to schedule any USEs right now and we have some | |
3541 | in *AV_PTR, remove them, else just return the first one found. */ | |
3542 | if (!try_uses_p) | |
3543 | { | |
3544 | FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3545 | if (INSN_CODE (EXPR_INSN_RTX (expr)) < 0) | |
3546 | av_set_iter_remove (&si); | |
3547 | } | |
3548 | else | |
3549 | { | |
3550 | FOR_EACH_EXPR_1 (expr, si, av_ptr) | |
3551 | { | |
3552 | gcc_assert (INSN_CODE (EXPR_INSN_RTX (expr)) < 0); | |
3553 | ||
3554 | if (EXPR_TARGET_AVAILABLE (expr) == 1) | |
3555 | return expr; | |
3556 | ||
3557 | av_set_iter_remove (&si); | |
3558 | } | |
3559 | } | |
3560 | } | |
3561 | ||
3562 | return NULL; | |
3563 | } | |
3564 | ||
3565 | /* Lookup EXPR in VINSN_VEC and return TRUE if found. */ | |
3566 | static bool | |
3567 | vinsn_vec_has_expr_p (vinsn_vec_t vinsn_vec, expr_t expr) | |
3568 | { | |
3569 | vinsn_t vinsn; | |
3570 | int n; | |
3571 | ||
3572 | for (n = 0; VEC_iterate (vinsn_t, vinsn_vec, n, vinsn); n++) | |
3573 | if (VINSN_SEPARABLE_P (vinsn)) | |
3574 | { | |
3575 | if (vinsn_equal_p (vinsn, EXPR_VINSN (expr))) | |
3576 | return true; | |
3577 | } | |
3578 | else | |
3579 | { | |
3580 | /* For non-separable instructions, the blocking insn can have | |
3581 | another pattern due to substitution, and we can't choose | |
3582 | different register as in the above case. Check all registers | |
3583 | being written instead. */ | |
3584 | if (bitmap_intersect_p (VINSN_REG_SETS (vinsn), | |
3585 | VINSN_REG_SETS (EXPR_VINSN (expr)))) | |
3586 | return true; | |
3587 | } | |
3588 | ||
3589 | return false; | |
3590 | } | |
3591 | ||
3592 | #ifdef ENABLE_CHECKING | |
3593 | /* Return true if either of expressions from ORIG_OPS can be blocked | |
3594 | by previously created bookkeeping code. STATIC_PARAMS points to static | |
3595 | parameters of move_op. */ | |
3596 | static bool | |
3597 | av_set_could_be_blocked_by_bookkeeping_p (av_set_t orig_ops, void *static_params) | |
3598 | { | |
3599 | expr_t expr; | |
3600 | av_set_iterator iter; | |
3601 | moveop_static_params_p sparams; | |
3602 | ||
3603 | /* This checks that expressions in ORIG_OPS are not blocked by bookkeeping | |
3604 | created while scheduling on another fence. */ | |
3605 | FOR_EACH_EXPR (expr, iter, orig_ops) | |
3606 | if (vinsn_vec_has_expr_p (vec_bookkeeping_blocked_vinsns, expr)) | |
3607 | return true; | |
3608 | ||
3609 | gcc_assert (code_motion_path_driver_info == &move_op_hooks); | |
3610 | sparams = (moveop_static_params_p) static_params; | |
3611 | ||
3612 | /* Expressions can be also blocked by bookkeeping created during current | |
3613 | move_op. */ | |
3614 | if (bitmap_bit_p (current_copies, INSN_UID (sparams->failed_insn))) | |
3615 | FOR_EACH_EXPR (expr, iter, orig_ops) | |
3616 | if (moveup_expr_cached (expr, sparams->failed_insn, false) != MOVEUP_EXPR_NULL) | |
3617 | return true; | |
3618 | ||
3619 | /* Expressions in ORIG_OPS may have wrong destination register due to | |
3620 | renaming. Check with the right register instead. */ | |
3621 | if (sparams->dest && REG_P (sparams->dest)) | |
3622 | { | |
3623 | unsigned regno = REGNO (sparams->dest); | |
3624 | vinsn_t failed_vinsn = INSN_VINSN (sparams->failed_insn); | |
3625 | ||
3626 | if (bitmap_bit_p (VINSN_REG_SETS (failed_vinsn), regno) | |
3627 | || bitmap_bit_p (VINSN_REG_USES (failed_vinsn), regno) | |
3628 | || bitmap_bit_p (VINSN_REG_CLOBBERS (failed_vinsn), regno)) | |
3629 | return true; | |
3630 | } | |
3631 | ||
3632 | return false; | |
3633 | } | |
3634 | #endif | |
3635 | ||
3636 | /* Clear VINSN_VEC and detach vinsns. */ | |
3637 | static void | |
3638 | vinsn_vec_clear (vinsn_vec_t *vinsn_vec) | |
3639 | { | |
3640 | unsigned len = VEC_length (vinsn_t, *vinsn_vec); | |
3641 | if (len > 0) | |
3642 | { | |
3643 | vinsn_t vinsn; | |
3644 | int n; | |
3645 | ||
3646 | for (n = 0; VEC_iterate (vinsn_t, *vinsn_vec, n, vinsn); n++) | |
3647 | vinsn_detach (vinsn); | |
3648 | VEC_block_remove (vinsn_t, *vinsn_vec, 0, len); | |
3649 | } | |
3650 | } | |
3651 | ||
3652 | /* Add the vinsn of EXPR to the VINSN_VEC. */ | |
3653 | static void | |
3654 | vinsn_vec_add (vinsn_vec_t *vinsn_vec, expr_t expr) | |
3655 | { | |
3656 | vinsn_attach (EXPR_VINSN (expr)); | |
3657 | VEC_safe_push (vinsn_t, heap, *vinsn_vec, EXPR_VINSN (expr)); | |
3658 | } | |
3659 | ||
3660 | /* Free the vector representing blocked expressions. */ | |
3661 | static void | |
3662 | vinsn_vec_free (vinsn_vec_t *vinsn_vec) | |
3663 | { | |
3664 | if (*vinsn_vec) | |
3665 | VEC_free (vinsn_t, heap, *vinsn_vec); | |
3666 | } | |
3667 | ||
3668 | /* Increase EXPR_PRIORITY_ADJ for INSN by AMOUNT. */ | |
3669 | ||
3670 | void sel_add_to_insn_priority (rtx insn, int amount) | |
3671 | { | |
3672 | EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) += amount; | |
3673 | ||
3674 | if (sched_verbose >= 2) | |
3675 | sel_print ("sel_add_to_insn_priority: insn %d, by %d (now %d+%d).\n", | |
3676 | INSN_UID (insn), amount, EXPR_PRIORITY (INSN_EXPR (insn)), | |
3677 | EXPR_PRIORITY_ADJ (INSN_EXPR (insn))); | |
3678 | } | |
3679 | ||
3680 | /* Turn AV into a vector, filter inappropriate insns and sort it. Return | |
3681 | true if there is something to schedule. BNDS and FENCE are current | |
3682 | boundaries and fence, respectively. If we need to stall for some cycles | |
3683 | before an expr from AV would become available, write this number to | |
3684 | *PNEED_STALL. */ | |
3685 | static bool | |
3686 | fill_vec_av_set (av_set_t av, blist_t bnds, fence_t fence, | |
3687 | int *pneed_stall) | |
3688 | { | |
3689 | av_set_iterator si; | |
3690 | expr_t expr; | |
3691 | int sched_next_worked = 0, stalled, n; | |
3692 | static int av_max_prio, est_ticks_till_branch; | |
3693 | int min_need_stall = -1; | |
3694 | deps_t dc = BND_DC (BLIST_BND (bnds)); | |
3695 | ||
3696 | /* Bail out early when the ready list contained only USEs/CLOBBERs that are | |
3697 | already scheduled. */ | |
3698 | if (av == NULL) | |
3699 | return false; | |
3700 | ||
3701 | /* Empty vector from the previous stuff. */ | |
3702 | if (VEC_length (expr_t, vec_av_set) > 0) | |
3703 | VEC_block_remove (expr_t, vec_av_set, 0, VEC_length (expr_t, vec_av_set)); | |
3704 | ||
3705 | /* Turn the set into a vector for sorting and call sel_target_adjust_priority | |
3706 | for each insn. */ | |
3707 | gcc_assert (VEC_empty (expr_t, vec_av_set)); | |
3708 | FOR_EACH_EXPR (expr, si, av) | |
3709 | { | |
3710 | VEC_safe_push (expr_t, heap, vec_av_set, expr); | |
3711 | ||
3712 | gcc_assert (EXPR_PRIORITY_ADJ (expr) == 0 || *pneed_stall); | |
3713 | ||
3714 | /* Adjust priority using target backend hook. */ | |
3715 | sel_target_adjust_priority (expr); | |
3716 | } | |
3717 | ||
3718 | /* Sort the vector. */ | |
3719 | qsort (VEC_address (expr_t, vec_av_set), VEC_length (expr_t, vec_av_set), | |
3720 | sizeof (expr_t), sel_rank_for_schedule); | |
3721 | ||
3722 | /* We record maximal priority of insns in av set for current instruction | |
3723 | group. */ | |
3724 | if (FENCE_STARTS_CYCLE_P (fence)) | |
3725 | av_max_prio = est_ticks_till_branch = INT_MIN; | |
3726 | ||
3727 | /* Filter out inappropriate expressions. Loop's direction is reversed to | |
3728 | visit "best" instructions first. We assume that VEC_unordered_remove | |
3729 | moves last element in place of one being deleted. */ | |
3730 | for (n = VEC_length (expr_t, vec_av_set) - 1, stalled = 0; n >= 0; n--) | |
3731 | { | |
3732 | expr_t expr = VEC_index (expr_t, vec_av_set, n); | |
3733 | insn_t insn = EXPR_INSN_RTX (expr); | |
3734 | char target_available; | |
3735 | bool is_orig_reg_p = true; | |
3736 | int need_cycles, new_prio; | |
3737 | ||
3738 | /* Don't allow any insns other than from SCHED_GROUP if we have one. */ | |
3739 | if (FENCE_SCHED_NEXT (fence) && insn != FENCE_SCHED_NEXT (fence)) | |
3740 | { | |
3741 | VEC_unordered_remove (expr_t, vec_av_set, n); | |
3742 | continue; | |
3743 | } | |
3744 | ||
3745 | /* Set number of sched_next insns (just in case there | |
3746 | could be several). */ | |
3747 | if (FENCE_SCHED_NEXT (fence)) | |
3748 | sched_next_worked++; | |
3749 | ||
3750 | /* Check all liveness requirements and try renaming. | |
3751 | FIXME: try to minimize calls to this. */ | |
3752 | target_available = EXPR_TARGET_AVAILABLE (expr); | |
3753 | ||
3754 | /* If insn was already scheduled on the current fence, | |
3755 | set TARGET_AVAILABLE to -1 no matter what expr's attribute says. */ | |
3756 | if (vinsn_vec_has_expr_p (vec_target_unavailable_vinsns, expr)) | |
3757 | target_available = -1; | |
3758 | ||
3759 | /* If the availability of the EXPR is invalidated by the insertion of | |
3760 | bookkeeping earlier, make sure that we won't choose this expr for | |
3761 | scheduling if it's not separable, and if it is separable, then | |
3762 | we have to recompute the set of available registers for it. */ | |
3763 | if (vinsn_vec_has_expr_p (vec_bookkeeping_blocked_vinsns, expr)) | |
3764 | { | |
3765 | VEC_unordered_remove (expr_t, vec_av_set, n); | |
3766 | if (sched_verbose >= 4) | |
3767 | sel_print ("Expr %d is blocked by bookkeeping inserted earlier\n", | |
3768 | INSN_UID (insn)); | |
3769 | continue; | |
3770 | } | |
3771 | ||
3772 | if (target_available == true) | |
3773 | { | |
3774 | /* Do nothing -- we can use an existing register. */ | |
3775 | is_orig_reg_p = EXPR_SEPARABLE_P (expr); | |
3776 | } | |
3777 | else if (/* Non-separable instruction will never | |
3778 | get another register. */ | |
3779 | (target_available == false | |
3780 | && !EXPR_SEPARABLE_P (expr)) | |
3781 | /* Don't try to find a register for low-priority expression. */ | |
3782 | || (int) VEC_length (expr_t, vec_av_set) - 1 - n >= max_insns_to_rename | |
3783 | /* ??? FIXME: Don't try to rename data speculation. */ | |
3784 | || (EXPR_SPEC_DONE_DS (expr) & BEGIN_DATA) | |
3785 | || ! find_best_reg_for_expr (expr, bnds, &is_orig_reg_p)) | |
3786 | { | |
3787 | VEC_unordered_remove (expr_t, vec_av_set, n); | |
3788 | if (sched_verbose >= 4) | |
3789 | sel_print ("Expr %d has no suitable target register\n", | |
3790 | INSN_UID (insn)); | |
3791 | continue; | |
3792 | } | |
3793 | ||
3794 | /* Filter expressions that need to be renamed or speculated when | |
3795 | pipelining, because compensating register copies or speculation | |
3796 | checks are likely to be placed near the beginning of the loop, | |
3797 | causing a stall. */ | |
3798 | if (pipelining_p && EXPR_ORIG_SCHED_CYCLE (expr) > 0 | |
3799 | && (!is_orig_reg_p || EXPR_SPEC_DONE_DS (expr) != 0)) | |
3800 | { | |
3801 | /* Estimation of number of cycles until loop branch for | |
3802 | renaming/speculation to be successful. */ | |
3803 | int need_n_ticks_till_branch = sel_vinsn_cost (EXPR_VINSN (expr)); | |
3804 | ||
3805 | if ((int) current_loop_nest->ninsns < 9) | |
3806 | { | |
3807 | VEC_unordered_remove (expr_t, vec_av_set, n); | |
3808 | if (sched_verbose >= 4) | |
3809 | sel_print ("Pipelining expr %d will likely cause stall\n", | |
3810 | INSN_UID (insn)); | |
3811 | continue; | |
3812 | } | |
3813 | ||
3814 | if ((int) current_loop_nest->ninsns - num_insns_scheduled | |
3815 | < need_n_ticks_till_branch * issue_rate / 2 | |
3816 | && est_ticks_till_branch < need_n_ticks_till_branch) | |
3817 | { | |
3818 | VEC_unordered_remove (expr_t, vec_av_set, n); | |
3819 | if (sched_verbose >= 4) | |
3820 | sel_print ("Pipelining expr %d will likely cause stall\n", | |
3821 | INSN_UID (insn)); | |
3822 | continue; | |
3823 | } | |
3824 | } | |
3825 | ||
3826 | /* We want to schedule speculation checks as late as possible. Discard | |
3827 | them from av set if there are instructions with higher priority. */ | |
3828 | if (sel_insn_is_speculation_check (insn) | |
3829 | && EXPR_PRIORITY (expr) < av_max_prio) | |
3830 | { | |
3831 | stalled++; | |
3832 | min_need_stall = min_need_stall < 0 ? 1 : MIN (min_need_stall, 1); | |
3833 | VEC_unordered_remove (expr_t, vec_av_set, n); | |
3834 | if (sched_verbose >= 4) | |
3835 | sel_print ("Delaying speculation check %d until its first use\n", | |
3836 | INSN_UID (insn)); | |
3837 | continue; | |
3838 | } | |
3839 | ||
3840 | /* Ignore EXPRs available from pipelining to update AV_MAX_PRIO. */ | |
3841 | if (EXPR_ORIG_SCHED_CYCLE (expr) <= 0) | |
3842 | av_max_prio = MAX (av_max_prio, EXPR_PRIORITY (expr)); | |
3843 | ||
3844 | /* Don't allow any insns whose data is not yet ready. | |
3845 | Check first whether we've already tried them and failed. */ | |
3846 | if (INSN_UID (insn) < FENCE_READY_TICKS_SIZE (fence)) | |
3847 | { | |
3848 | need_cycles = (FENCE_READY_TICKS (fence)[INSN_UID (insn)] | |
3849 | - FENCE_CYCLE (fence)); | |
3850 | if (EXPR_ORIG_SCHED_CYCLE (expr) <= 0) | |
3851 | est_ticks_till_branch = MAX (est_ticks_till_branch, | |
3852 | EXPR_PRIORITY (expr) + need_cycles); | |
3853 | ||
3854 | if (need_cycles > 0) | |
3855 | { | |
3856 | stalled++; | |
3857 | min_need_stall = (min_need_stall < 0 | |
3858 | ? need_cycles | |
3859 | : MIN (min_need_stall, need_cycles)); | |
3860 | VEC_unordered_remove (expr_t, vec_av_set, n); | |
3861 | ||
3862 | if (sched_verbose >= 4) | |
3863 | sel_print ("Expr %d is not ready until cycle %d (cached)\n", | |
3864 | INSN_UID (insn), | |
3865 | FENCE_READY_TICKS (fence)[INSN_UID (insn)]); | |
3866 | continue; | |
3867 | } | |
3868 | } | |
3869 | ||
3870 | /* Now resort to dependence analysis to find whether EXPR might be | |
3871 | stalled due to dependencies from FENCE's context. */ | |
3872 | need_cycles = tick_check_p (expr, dc, fence); | |
3873 | new_prio = EXPR_PRIORITY (expr) + EXPR_PRIORITY_ADJ (expr) + need_cycles; | |
3874 | ||
3875 | if (EXPR_ORIG_SCHED_CYCLE (expr) <= 0) | |
3876 | est_ticks_till_branch = MAX (est_ticks_till_branch, | |
3877 | new_prio); | |
3878 | ||
3879 | if (need_cycles > 0) | |
3880 | { | |
3881 | if (INSN_UID (insn) >= FENCE_READY_TICKS_SIZE (fence)) | |
3882 | { | |
3883 | int new_size = INSN_UID (insn) * 3 / 2; | |
3884 | ||
3885 | FENCE_READY_TICKS (fence) | |
3886 | = (int *) xrecalloc (FENCE_READY_TICKS (fence), | |
3887 | new_size, FENCE_READY_TICKS_SIZE (fence), | |
3888 | sizeof (int)); | |
3889 | } | |
3890 | FENCE_READY_TICKS (fence)[INSN_UID (insn)] | |
3891 | = FENCE_CYCLE (fence) + need_cycles; | |
3892 | ||
3893 | stalled++; | |
3894 | min_need_stall = (min_need_stall < 0 | |
3895 | ? need_cycles | |
3896 | : MIN (min_need_stall, need_cycles)); | |
3897 | ||
3898 | VEC_unordered_remove (expr_t, vec_av_set, n); | |
3899 | ||
3900 | if (sched_verbose >= 4) | |
3901 | sel_print ("Expr %d is not ready yet until cycle %d\n", | |
3902 | INSN_UID (insn), | |
3903 | FENCE_READY_TICKS (fence)[INSN_UID (insn)]); | |
3904 | continue; | |
3905 | } | |
3906 | ||
3907 | if (sched_verbose >= 4) | |
3908 | sel_print ("Expr %d is ok\n", INSN_UID (insn)); | |
3909 | min_need_stall = 0; | |
3910 | } | |
3911 | ||
3912 | /* Clear SCHED_NEXT. */ | |
3913 | if (FENCE_SCHED_NEXT (fence)) | |
3914 | { | |
3915 | gcc_assert (sched_next_worked == 1); | |
3916 | FENCE_SCHED_NEXT (fence) = NULL_RTX; | |
3917 | } | |
3918 | ||
3919 | /* No need to stall if this variable was not initialized. */ | |
3920 | if (min_need_stall < 0) | |
3921 | min_need_stall = 0; | |
3922 | ||
3923 | if (VEC_empty (expr_t, vec_av_set)) | |
3924 | { | |
3925 | /* We need to set *pneed_stall here, because later we skip this code | |
3926 | when ready list is empty. */ | |
3927 | *pneed_stall = min_need_stall; | |
3928 | return false; | |
3929 | } | |
3930 | else | |
3931 | gcc_assert (min_need_stall == 0); | |
3932 | ||
3933 | /* Sort the vector. */ | |
3934 | qsort (VEC_address (expr_t, vec_av_set), VEC_length (expr_t, vec_av_set), | |
3935 | sizeof (expr_t), sel_rank_for_schedule); | |
3936 | ||
3937 | if (sched_verbose >= 4) | |
3938 | { | |
3939 | sel_print ("Total ready exprs: %d, stalled: %d\n", | |
3940 | VEC_length (expr_t, vec_av_set), stalled); | |
3941 | sel_print ("Sorted av set (%d): ", VEC_length (expr_t, vec_av_set)); | |
3942 | for (n = 0; VEC_iterate (expr_t, vec_av_set, n, expr); n++) | |
3943 | dump_expr (expr); | |
3944 | sel_print ("\n"); | |
3945 | } | |
3946 | ||
3947 | *pneed_stall = 0; | |
3948 | return true; | |
3949 | } | |
3950 | ||
3951 | /* Convert a vectored and sorted av set to the ready list that | |
3952 | the rest of the backend wants to see. */ | |
3953 | static void | |
3954 | convert_vec_av_set_to_ready (void) | |
3955 | { | |
3956 | int n; | |
3957 | expr_t expr; | |
3958 | ||
3959 | /* Allocate and fill the ready list from the sorted vector. */ | |
3960 | ready.n_ready = VEC_length (expr_t, vec_av_set); | |
3961 | ready.first = ready.n_ready - 1; | |
3962 | ||
3963 | gcc_assert (ready.n_ready > 0); | |
3964 | ||
3965 | if (ready.n_ready > max_issue_size) | |
3966 | { | |
3967 | max_issue_size = ready.n_ready; | |
3968 | sched_extend_ready_list (ready.n_ready); | |
3969 | } | |
3970 | ||
3971 | for (n = 0; VEC_iterate (expr_t, vec_av_set, n, expr); n++) | |
3972 | { | |
3973 | vinsn_t vi = EXPR_VINSN (expr); | |
3974 | insn_t insn = VINSN_INSN_RTX (vi); | |
3975 | ||
3976 | ready_try[n] = 0; | |
3977 | ready.vec[n] = insn; | |
3978 | } | |
3979 | } | |
3980 | ||
3981 | /* Initialize ready list from *AV_PTR for the max_issue () call. | |
3982 | If any unrecognizable insn found in *AV_PTR, return it (and skip | |
3983 | max_issue). BND and FENCE are current boundary and fence, | |
3984 | respectively. If we need to stall for some cycles before an expr | |
3985 | from *AV_PTR would become available, write this number to *PNEED_STALL. */ | |
3986 | static expr_t | |
3987 | fill_ready_list (av_set_t *av_ptr, blist_t bnds, fence_t fence, | |
3988 | int *pneed_stall) | |
3989 | { | |
3990 | expr_t expr; | |
3991 | ||
3992 | /* We do not support multiple boundaries per fence. */ | |
3993 | gcc_assert (BLIST_NEXT (bnds) == NULL); | |
3994 | ||
3995 | /* Process expressions required special handling, i.e. pipelined, | |
3996 | speculative and recog() < 0 expressions first. */ | |
3997 | process_pipelined_exprs (av_ptr); | |
3998 | process_spec_exprs (av_ptr); | |
3999 | ||
4000 | /* A USE could be scheduled immediately. */ | |
4001 | expr = process_use_exprs (av_ptr); | |
4002 | if (expr) | |
4003 | { | |
4004 | *pneed_stall = 0; | |
4005 | return expr; | |
4006 | } | |
4007 | ||
4008 | /* Turn the av set to a vector for sorting. */ | |
4009 | if (! fill_vec_av_set (*av_ptr, bnds, fence, pneed_stall)) | |
4010 | { | |
4011 | ready.n_ready = 0; | |
4012 | return NULL; | |
4013 | } | |
4014 | ||
4015 | /* Build the final ready list. */ | |
4016 | convert_vec_av_set_to_ready (); | |
4017 | return NULL; | |
4018 | } | |
4019 | ||
4020 | /* Wrapper for dfa_new_cycle (). Returns TRUE if cycle was advanced. */ | |
4021 | static bool | |
4022 | sel_dfa_new_cycle (insn_t insn, fence_t fence) | |
4023 | { | |
4024 | int last_scheduled_cycle = FENCE_LAST_SCHEDULED_INSN (fence) | |
4025 | ? INSN_SCHED_CYCLE (FENCE_LAST_SCHEDULED_INSN (fence)) | |
4026 | : FENCE_CYCLE (fence) - 1; | |
4027 | bool res = false; | |
4028 | int sort_p = 0; | |
4029 | ||
4030 | if (!targetm.sched.dfa_new_cycle) | |
4031 | return false; | |
4032 | ||
4033 | memcpy (curr_state, FENCE_STATE (fence), dfa_state_size); | |
4034 | ||
4035 | while (!sort_p && targetm.sched.dfa_new_cycle (sched_dump, sched_verbose, | |
4036 | insn, last_scheduled_cycle, | |
4037 | FENCE_CYCLE (fence), &sort_p)) | |
4038 | { | |
4039 | memcpy (FENCE_STATE (fence), curr_state, dfa_state_size); | |
4040 | advance_one_cycle (fence); | |
4041 | memcpy (curr_state, FENCE_STATE (fence), dfa_state_size); | |
4042 | res = true; | |
4043 | } | |
4044 | ||
4045 | return res; | |
4046 | } | |
4047 | ||
4048 | /* Invoke reorder* target hooks on the ready list. Return the number of insns | |
4049 | we can issue. FENCE is the current fence. */ | |
4050 | static int | |
4051 | invoke_reorder_hooks (fence_t fence) | |
4052 | { | |
4053 | int issue_more; | |
4054 | bool ran_hook = false; | |
4055 | ||
4056 | /* Call the reorder hook at the beginning of the cycle, and call | |
4057 | the reorder2 hook in the middle of the cycle. */ | |
4058 | if (FENCE_ISSUED_INSNS (fence) == 0) | |
4059 | { | |
4060 | if (targetm.sched.reorder | |
4061 | && !SCHED_GROUP_P (ready_element (&ready, 0)) | |
4062 | && ready.n_ready > 1) | |
4063 | { | |
4064 | /* Don't give reorder the most prioritized insn as it can break | |
4065 | pipelining. */ | |
4066 | if (pipelining_p) | |
4067 | --ready.n_ready; | |
4068 | ||
4069 | issue_more | |
4070 | = targetm.sched.reorder (sched_dump, sched_verbose, | |
4071 | ready_lastpos (&ready), | |
4072 | &ready.n_ready, FENCE_CYCLE (fence)); | |
4073 | ||
4074 | if (pipelining_p) | |
4075 | ++ready.n_ready; | |
4076 | ||
4077 | ran_hook = true; | |
4078 | } | |
4079 | else | |
4080 | /* Initialize can_issue_more for variable_issue. */ | |
4081 | issue_more = issue_rate; | |
4082 | } | |
4083 | else if (targetm.sched.reorder2 | |
4084 | && !SCHED_GROUP_P (ready_element (&ready, 0))) | |
4085 | { | |
4086 | if (ready.n_ready == 1) | |
4087 | issue_more = | |
4088 | targetm.sched.reorder2 (sched_dump, sched_verbose, | |
4089 | ready_lastpos (&ready), | |
4090 | &ready.n_ready, FENCE_CYCLE (fence)); | |
4091 | else | |
4092 | { | |
4093 | if (pipelining_p) | |
4094 | --ready.n_ready; | |
4095 | ||
4096 | issue_more = | |
4097 | targetm.sched.reorder2 (sched_dump, sched_verbose, | |
4098 | ready.n_ready | |
4099 | ? ready_lastpos (&ready) : NULL, | |
4100 | &ready.n_ready, FENCE_CYCLE (fence)); | |
4101 | ||
4102 | if (pipelining_p) | |
4103 | ++ready.n_ready; | |
4104 | } | |
4105 | ||
4106 | ran_hook = true; | |
4107 | } | |
4108 | else | |
4109 | issue_more = issue_rate; | |
4110 | ||
4111 | /* Ensure that ready list and vec_av_set are in line with each other, | |
4112 | i.e. vec_av_set[i] == ready_element (&ready, i). */ | |
4113 | if (issue_more && ran_hook) | |
4114 | { | |
4115 | int i, j, n; | |
4116 | rtx *arr = ready.vec; | |
4117 | expr_t *vec = VEC_address (expr_t, vec_av_set); | |
4118 | ||
4119 | for (i = 0, n = ready.n_ready; i < n; i++) | |
4120 | if (EXPR_INSN_RTX (vec[i]) != arr[i]) | |
4121 | { | |
4122 | expr_t tmp; | |
4123 | ||
4124 | for (j = i; j < n; j++) | |
4125 | if (EXPR_INSN_RTX (vec[j]) == arr[i]) | |
4126 | break; | |
4127 | gcc_assert (j < n); | |
4128 | ||
4129 | tmp = vec[i]; | |
4130 | vec[i] = vec[j]; | |
4131 | vec[j] = tmp; | |
4132 | } | |
4133 | } | |
4134 | ||
4135 | return issue_more; | |
4136 | } | |
4137 | ||
4138 | /* Return an EXPR correponding to INDEX element of ready list, if | |
4139 | FOLLOW_READY_ELEMENT is true (i.e., an expr of | |
4140 | ready_element (&ready, INDEX) will be returned), and to INDEX element of | |
4141 | ready.vec otherwise. */ | |
4142 | static inline expr_t | |
4143 | find_expr_for_ready (int index, bool follow_ready_element) | |
4144 | { | |
4145 | expr_t expr; | |
4146 | int real_index; | |
4147 | ||
4148 | real_index = follow_ready_element ? ready.first - index : index; | |
4149 | ||
4150 | expr = VEC_index (expr_t, vec_av_set, real_index); | |
4151 | gcc_assert (ready.vec[real_index] == EXPR_INSN_RTX (expr)); | |
4152 | ||
4153 | return expr; | |
4154 | } | |
4155 | ||
4156 | /* Calculate insns worth trying via lookahead_guard hook. Return a number | |
4157 | of such insns found. */ | |
4158 | static int | |
4159 | invoke_dfa_lookahead_guard (void) | |
4160 | { | |
4161 | int i, n; | |
4162 | bool have_hook | |
4163 | = targetm.sched.first_cycle_multipass_dfa_lookahead_guard != NULL; | |
4164 | ||
4165 | if (sched_verbose >= 2) | |
4166 | sel_print ("ready after reorder: "); | |
4167 | ||
4168 | for (i = 0, n = 0; i < ready.n_ready; i++) | |
4169 | { | |
4170 | expr_t expr; | |
4171 | insn_t insn; | |
4172 | int r; | |
4173 | ||
4174 | /* In this loop insn is Ith element of the ready list given by | |
4175 | ready_element, not Ith element of ready.vec. */ | |
4176 | insn = ready_element (&ready, i); | |
4177 | ||
4178 | if (! have_hook || i == 0) | |
4179 | r = 0; | |
4180 | else | |
4181 | r = !targetm.sched.first_cycle_multipass_dfa_lookahead_guard (insn); | |
4182 | ||
4183 | gcc_assert (INSN_CODE (insn) >= 0); | |
4184 | ||
4185 | /* Only insns with ready_try = 0 can get here | |
4186 | from fill_ready_list. */ | |
4187 | gcc_assert (ready_try [i] == 0); | |
4188 | ready_try[i] = r; | |
4189 | if (!r) | |
4190 | n++; | |
4191 | ||
4192 | expr = find_expr_for_ready (i, true); | |
4193 | ||
4194 | if (sched_verbose >= 2) | |
4195 | { | |
4196 | dump_vinsn (EXPR_VINSN (expr)); | |
4197 | sel_print (":%d; ", ready_try[i]); | |
4198 | } | |
4199 | } | |
4200 | ||
4201 | if (sched_verbose >= 2) | |
4202 | sel_print ("\n"); | |
4203 | return n; | |
4204 | } | |
4205 | ||
4206 | /* Calculate the number of privileged insns and return it. */ | |
4207 | static int | |
4208 | calculate_privileged_insns (void) | |
4209 | { | |
4210 | expr_t cur_expr, min_spec_expr = NULL; | |
4211 | insn_t cur_insn, min_spec_insn; | |
4212 | int privileged_n = 0, i; | |
4213 | ||
4214 | for (i = 0; i < ready.n_ready; i++) | |
4215 | { | |
4216 | if (ready_try[i]) | |
4217 | continue; | |
4218 | ||
4219 | if (! min_spec_expr) | |
4220 | { | |
4221 | min_spec_insn = ready_element (&ready, i); | |
4222 | min_spec_expr = find_expr_for_ready (i, true); | |
4223 | } | |
4224 | ||
4225 | cur_insn = ready_element (&ready, i); | |
4226 | cur_expr = find_expr_for_ready (i, true); | |
4227 | ||
4228 | if (EXPR_SPEC (cur_expr) > EXPR_SPEC (min_spec_expr)) | |
4229 | break; | |
4230 | ||
4231 | ++privileged_n; | |
4232 | } | |
4233 | ||
4234 | if (i == ready.n_ready) | |
4235 | privileged_n = 0; | |
4236 | ||
4237 | if (sched_verbose >= 2) | |
4238 | sel_print ("privileged_n: %d insns with SPEC %d\n", | |
4239 | privileged_n, privileged_n ? EXPR_SPEC (min_spec_expr) : -1); | |
4240 | return privileged_n; | |
4241 | } | |
4242 | ||
4243 | /* Call the rest of the hooks after the choice was made. Return | |
4244 | the number of insns that still can be issued given that the current | |
4245 | number is ISSUE_MORE. FENCE and BEST_INSN are the current fence | |
4246 | and the insn chosen for scheduling, respectively. */ | |
4247 | static int | |
4248 | invoke_aftermath_hooks (fence_t fence, rtx best_insn, int issue_more) | |
4249 | { | |
4250 | gcc_assert (INSN_P (best_insn)); | |
4251 | ||
4252 | /* First, call dfa_new_cycle, and then variable_issue, if available. */ | |
4253 | sel_dfa_new_cycle (best_insn, fence); | |
4254 | ||
4255 | if (targetm.sched.variable_issue) | |
4256 | { | |
4257 | memcpy (curr_state, FENCE_STATE (fence), dfa_state_size); | |
4258 | issue_more = | |
4259 | targetm.sched.variable_issue (sched_dump, sched_verbose, best_insn, | |
4260 | issue_more); | |
4261 | memcpy (FENCE_STATE (fence), curr_state, dfa_state_size); | |
4262 | } | |
4263 | else if (GET_CODE (PATTERN (best_insn)) != USE | |
4264 | && GET_CODE (PATTERN (best_insn)) != CLOBBER) | |
4265 | issue_more--; | |
4266 | ||
4267 | return issue_more; | |
4268 | } | |
4269 | ||
4270 | /* Estimate the cost of issuing INSN on DFA state STATE. */ | |
4271 | static int | |
4272 | estimate_insn_cost (rtx insn, state_t state) | |
4273 | { | |
4274 | static state_t temp = NULL; | |
4275 | int cost; | |
4276 | ||
4277 | if (!temp) | |
4278 | temp = xmalloc (dfa_state_size); | |
4279 | ||
4280 | memcpy (temp, state, dfa_state_size); | |
4281 | cost = state_transition (temp, insn); | |
4282 | ||
4283 | if (cost < 0) | |
4284 | return 0; | |
4285 | else if (cost == 0) | |
4286 | return 1; | |
4287 | return cost; | |
4288 | } | |
4289 | ||
4290 | /* Return the cost of issuing EXPR on the FENCE as estimated by DFA. | |
4291 | This function properly handles ASMs, USEs etc. */ | |
4292 | static int | |
4293 | get_expr_cost (expr_t expr, fence_t fence) | |
4294 | { | |
4295 | rtx insn = EXPR_INSN_RTX (expr); | |
4296 | ||
4297 | if (recog_memoized (insn) < 0) | |
4298 | { | |
4299 | if (!FENCE_STARTS_CYCLE_P (fence) | |
4300 | /* FIXME: Is this condition necessary? */ | |
4301 | && VINSN_UNIQUE_P (EXPR_VINSN (expr)) | |
4302 | && INSN_ASM_P (insn)) | |
4303 | /* This is asm insn which is tryed to be issued on the | |
4304 | cycle not first. Issue it on the next cycle. */ | |
4305 | return 1; | |
4306 | else | |
4307 | /* A USE insn, or something else we don't need to | |
4308 | understand. We can't pass these directly to | |
4309 | state_transition because it will trigger a | |
4310 | fatal error for unrecognizable insns. */ | |
4311 | return 0; | |
4312 | } | |
4313 | else | |
4314 | return estimate_insn_cost (insn, FENCE_STATE (fence)); | |
4315 | } | |
4316 | ||
4317 | /* Find the best insn for scheduling, either via max_issue or just take | |
4318 | the most prioritized available. */ | |
4319 | static int | |
4320 | choose_best_insn (fence_t fence, int privileged_n, int *index) | |
4321 | { | |
4322 | int can_issue = 0; | |
4323 | ||
4324 | if (dfa_lookahead > 0) | |
4325 | { | |
4326 | cycle_issued_insns = FENCE_ISSUED_INSNS (fence); | |
4327 | can_issue = max_issue (&ready, privileged_n, | |
4328 | FENCE_STATE (fence), index); | |
4329 | if (sched_verbose >= 2) | |
4330 | sel_print ("max_issue: we can issue %d insns, already did %d insns\n", | |
4331 | can_issue, FENCE_ISSUED_INSNS (fence)); | |
4332 | } | |
4333 | else | |
4334 | { | |
4335 | /* We can't use max_issue; just return the first available element. */ | |
4336 | int i; | |
4337 | ||
4338 | for (i = 0; i < ready.n_ready; i++) | |
4339 | { | |
4340 | expr_t expr = find_expr_for_ready (i, true); | |
4341 | ||
4342 | if (get_expr_cost (expr, fence) < 1) | |
4343 | { | |
4344 | can_issue = can_issue_more; | |
4345 | *index = i; | |
4346 | ||
4347 | if (sched_verbose >= 2) | |
4348 | sel_print ("using %dth insn from the ready list\n", i + 1); | |
4349 | ||
4350 | break; | |
4351 | } | |
4352 | } | |
4353 | ||
4354 | if (i == ready.n_ready) | |
4355 | { | |
4356 | can_issue = 0; | |
4357 | *index = -1; | |
4358 | } | |
4359 | } | |
4360 | ||
4361 | return can_issue; | |
4362 | } | |
4363 | ||
4364 | /* Choose the best expr from *AV_VLIW_PTR and a suitable register for it. | |
4365 | BNDS and FENCE are current boundaries and scheduling fence respectively. | |
4366 | Return the expr found and NULL if nothing can be issued atm. | |
4367 | Write to PNEED_STALL the number of cycles to stall if no expr was found. */ | |
4368 | static expr_t | |
4369 | find_best_expr (av_set_t *av_vliw_ptr, blist_t bnds, fence_t fence, | |
4370 | int *pneed_stall) | |
4371 | { | |
4372 | expr_t best; | |
4373 | ||
4374 | /* Choose the best insn for scheduling via: | |
4375 | 1) sorting the ready list based on priority; | |
4376 | 2) calling the reorder hook; | |
4377 | 3) calling max_issue. */ | |
4378 | best = fill_ready_list (av_vliw_ptr, bnds, fence, pneed_stall); | |
4379 | if (best == NULL && ready.n_ready > 0) | |
4380 | { | |
4381 | int privileged_n, index, avail_n; | |
4382 | ||
4383 | can_issue_more = invoke_reorder_hooks (fence); | |
4384 | if (can_issue_more > 0) | |
4385 | { | |
4386 | /* Try choosing the best insn until we find one that is could be | |
4387 | scheduled due to liveness restrictions on its destination register. | |
4388 | In the future, we'd like to choose once and then just probe insns | |
4389 | in the order of their priority. */ | |
4390 | avail_n = invoke_dfa_lookahead_guard (); | |
4391 | privileged_n = calculate_privileged_insns (); | |
4392 | can_issue_more = choose_best_insn (fence, privileged_n, &index); | |
4393 | if (can_issue_more) | |
4394 | best = find_expr_for_ready (index, true); | |
4395 | } | |
4396 | /* We had some available insns, so if we can't issue them, | |
4397 | we have a stall. */ | |
4398 | if (can_issue_more == 0) | |
4399 | { | |
4400 | best = NULL; | |
4401 | *pneed_stall = 1; | |
4402 | } | |
4403 | } | |
4404 | ||
4405 | if (best != NULL) | |
4406 | { | |
4407 | can_issue_more = invoke_aftermath_hooks (fence, EXPR_INSN_RTX (best), | |
4408 | can_issue_more); | |
4409 | if (can_issue_more == 0) | |
4410 | *pneed_stall = 1; | |
4411 | } | |
4412 | ||
4413 | if (sched_verbose >= 2) | |
4414 | { | |
4415 | if (best != NULL) | |
4416 | { | |
4417 | sel_print ("Best expression (vliw form): "); | |
4418 | dump_expr (best); | |
4419 | sel_print ("; cycle %d\n", FENCE_CYCLE (fence)); | |
4420 | } | |
4421 | else | |
4422 | sel_print ("No best expr found!\n"); | |
4423 | } | |
4424 | ||
4425 | return best; | |
4426 | } | |
4427 | \f | |
4428 | ||
4429 | /* Functions that implement the core of the scheduler. */ | |
4430 | ||
4431 | ||
4432 | /* Emit an instruction from EXPR with SEQNO and VINSN after | |
4433 | PLACE_TO_INSERT. */ | |
4434 | static insn_t | |
4435 | emit_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno, | |
4436 | insn_t place_to_insert) | |
4437 | { | |
4438 | /* This assert fails when we have identical instructions | |
4439 | one of which dominates the other. In this case move_op () | |
4440 | finds the first instruction and doesn't search for second one. | |
4441 | The solution would be to compute av_set after the first found | |
4442 | insn and, if insn present in that set, continue searching. | |
4443 | For now we workaround this issue in move_op. */ | |
4444 | gcc_assert (!INSN_IN_STREAM_P (EXPR_INSN_RTX (expr))); | |
4445 | ||
4446 | if (EXPR_WAS_RENAMED (expr)) | |
4447 | { | |
4448 | unsigned regno = expr_dest_regno (expr); | |
4449 | ||
4450 | if (HARD_REGISTER_NUM_P (regno)) | |
4451 | { | |
4452 | df_set_regs_ever_live (regno, true); | |
4453 | reg_rename_tick[regno] = ++reg_rename_this_tick; | |
4454 | } | |
4455 | } | |
4456 | ||
4457 | return sel_gen_insn_from_expr_after (expr, vinsn, seqno, | |
4458 | place_to_insert); | |
4459 | } | |
4460 | ||
4461 | /* Return TRUE if BB can hold bookkeeping code. */ | |
4462 | static bool | |
4463 | block_valid_for_bookkeeping_p (basic_block bb) | |
4464 | { | |
4465 | insn_t bb_end = BB_END (bb); | |
4466 | ||
4467 | if (!in_current_region_p (bb) || EDGE_COUNT (bb->succs) > 1) | |
4468 | return false; | |
4469 | ||
4470 | if (INSN_P (bb_end)) | |
4471 | { | |
4472 | if (INSN_SCHED_TIMES (bb_end) > 0) | |
4473 | return false; | |
4474 | } | |
4475 | else | |
4476 | gcc_assert (NOTE_INSN_BASIC_BLOCK_P (bb_end)); | |
4477 | ||
4478 | return true; | |
4479 | } | |
4480 | ||
4481 | /* Attempt to find a block that can hold bookkeeping code for path(s) incoming | |
4482 | into E2->dest, except from E1->src (there may be a sequence of empty basic | |
4483 | blocks between E1->src and E2->dest). Return found block, or NULL if new | |
b5b8b0ac AO |
4484 | one must be created. If LAX holds, don't assume there is a simple path |
4485 | from E1->src to E2->dest. */ | |
e855c69d | 4486 | static basic_block |
b5b8b0ac | 4487 | find_block_for_bookkeeping (edge e1, edge e2, bool lax) |
e855c69d AB |
4488 | { |
4489 | basic_block candidate_block = NULL; | |
4490 | edge e; | |
4491 | ||
4492 | /* Loop over edges from E1 to E2, inclusive. */ | |
b5b8b0ac | 4493 | for (e = e1; !lax || e->dest != EXIT_BLOCK_PTR; e = EDGE_SUCC (e->dest, 0)) |
e855c69d AB |
4494 | { |
4495 | if (EDGE_COUNT (e->dest->preds) == 2) | |
4496 | { | |
4497 | if (candidate_block == NULL) | |
4498 | candidate_block = (EDGE_PRED (e->dest, 0) == e | |
4499 | ? EDGE_PRED (e->dest, 1)->src | |
4500 | : EDGE_PRED (e->dest, 0)->src); | |
4501 | else | |
4502 | /* Found additional edge leading to path from e1 to e2 | |
4503 | from aside. */ | |
4504 | return NULL; | |
4505 | } | |
4506 | else if (EDGE_COUNT (e->dest->preds) > 2) | |
4507 | /* Several edges leading to path from e1 to e2 from aside. */ | |
4508 | return NULL; | |
4509 | ||
4510 | if (e == e2) | |
b5b8b0ac AO |
4511 | return ((!lax || candidate_block) |
4512 | && block_valid_for_bookkeeping_p (candidate_block) | |
e855c69d AB |
4513 | ? candidate_block |
4514 | : NULL); | |
b5b8b0ac AO |
4515 | |
4516 | if (lax && EDGE_COUNT (e->dest->succs) != 1) | |
4517 | return NULL; | |
e855c69d | 4518 | } |
b5b8b0ac AO |
4519 | |
4520 | if (lax) | |
4521 | return NULL; | |
4522 | ||
e855c69d AB |
4523 | gcc_unreachable (); |
4524 | } | |
4525 | ||
4526 | /* Create new basic block for bookkeeping code for path(s) incoming into | |
4527 | E2->dest, except from E1->src. Return created block. */ | |
4528 | static basic_block | |
4529 | create_block_for_bookkeeping (edge e1, edge e2) | |
4530 | { | |
4531 | basic_block new_bb, bb = e2->dest; | |
4532 | ||
4533 | /* Check that we don't spoil the loop structure. */ | |
4534 | if (current_loop_nest) | |
4535 | { | |
4536 | basic_block latch = current_loop_nest->latch; | |
4537 | ||
4538 | /* We do not split header. */ | |
4539 | gcc_assert (e2->dest != current_loop_nest->header); | |
4540 | ||
4541 | /* We do not redirect the only edge to the latch block. */ | |
4542 | gcc_assert (e1->dest != latch | |
4543 | || !single_pred_p (latch) | |
4544 | || e1 != single_pred_edge (latch)); | |
4545 | } | |
4546 | ||
4547 | /* Split BB to insert BOOK_INSN there. */ | |
4548 | new_bb = sched_split_block (bb, NULL); | |
4549 | ||
4550 | /* Move note_list from the upper bb. */ | |
4551 | gcc_assert (BB_NOTE_LIST (new_bb) == NULL_RTX); | |
4552 | BB_NOTE_LIST (new_bb) = BB_NOTE_LIST (bb); | |
4553 | BB_NOTE_LIST (bb) = NULL_RTX; | |
4554 | ||
4555 | gcc_assert (e2->dest == bb); | |
4556 | ||
4557 | /* Skip block for bookkeeping copy when leaving E1->src. */ | |
4558 | if (e1->flags & EDGE_FALLTHRU) | |
4559 | sel_redirect_edge_and_branch_force (e1, new_bb); | |
4560 | else | |
4561 | sel_redirect_edge_and_branch (e1, new_bb); | |
4562 | ||
4563 | gcc_assert (e1->dest == new_bb); | |
4564 | gcc_assert (sel_bb_empty_p (bb)); | |
4565 | ||
b5b8b0ac AO |
4566 | /* To keep basic block numbers in sync between debug and non-debug |
4567 | compilations, we have to rotate blocks here. Consider that we | |
4568 | started from (a,b)->d, (c,d)->e, and d contained only debug | |
4569 | insns. It would have been removed before if the debug insns | |
4570 | weren't there, so we'd have split e rather than d. So what we do | |
4571 | now is to swap the block numbers of new_bb and | |
4572 | single_succ(new_bb) == e, so that the insns that were in e before | |
4573 | get the new block number. */ | |
4574 | ||
4575 | if (MAY_HAVE_DEBUG_INSNS) | |
4576 | { | |
4577 | basic_block succ; | |
4578 | insn_t insn = sel_bb_head (new_bb); | |
4579 | insn_t last; | |
4580 | ||
4581 | if (DEBUG_INSN_P (insn) | |
4582 | && single_succ_p (new_bb) | |
4583 | && (succ = single_succ (new_bb)) | |
4584 | && succ != EXIT_BLOCK_PTR | |
4585 | && DEBUG_INSN_P ((last = sel_bb_end (new_bb)))) | |
4586 | { | |
4587 | while (insn != last && (DEBUG_INSN_P (insn) || NOTE_P (insn))) | |
4588 | insn = NEXT_INSN (insn); | |
4589 | ||
4590 | if (insn == last) | |
4591 | { | |
4592 | sel_global_bb_info_def gbi; | |
4593 | sel_region_bb_info_def rbi; | |
4594 | int i; | |
4595 | ||
4596 | if (sched_verbose >= 2) | |
4597 | sel_print ("Swapping block ids %i and %i\n", | |
4598 | new_bb->index, succ->index); | |
4599 | ||
4600 | i = new_bb->index; | |
4601 | new_bb->index = succ->index; | |
4602 | succ->index = i; | |
4603 | ||
4604 | SET_BASIC_BLOCK (new_bb->index, new_bb); | |
4605 | SET_BASIC_BLOCK (succ->index, succ); | |
4606 | ||
4607 | memcpy (&gbi, SEL_GLOBAL_BB_INFO (new_bb), sizeof (gbi)); | |
4608 | memcpy (SEL_GLOBAL_BB_INFO (new_bb), SEL_GLOBAL_BB_INFO (succ), | |
4609 | sizeof (gbi)); | |
4610 | memcpy (SEL_GLOBAL_BB_INFO (succ), &gbi, sizeof (gbi)); | |
4611 | ||
4612 | memcpy (&rbi, SEL_REGION_BB_INFO (new_bb), sizeof (rbi)); | |
4613 | memcpy (SEL_REGION_BB_INFO (new_bb), SEL_REGION_BB_INFO (succ), | |
4614 | sizeof (rbi)); | |
4615 | memcpy (SEL_REGION_BB_INFO (succ), &rbi, sizeof (rbi)); | |
4616 | ||
4617 | i = BLOCK_TO_BB (new_bb->index); | |
4618 | BLOCK_TO_BB (new_bb->index) = BLOCK_TO_BB (succ->index); | |
4619 | BLOCK_TO_BB (succ->index) = i; | |
4620 | ||
4621 | i = CONTAINING_RGN (new_bb->index); | |
4622 | CONTAINING_RGN (new_bb->index) = CONTAINING_RGN (succ->index); | |
4623 | CONTAINING_RGN (succ->index) = i; | |
4624 | ||
4625 | for (i = 0; i < current_nr_blocks; i++) | |
4626 | if (BB_TO_BLOCK (i) == succ->index) | |
4627 | BB_TO_BLOCK (i) = new_bb->index; | |
4628 | else if (BB_TO_BLOCK (i) == new_bb->index) | |
4629 | BB_TO_BLOCK (i) = succ->index; | |
4630 | ||
4631 | FOR_BB_INSNS (new_bb, insn) | |
4632 | if (INSN_P (insn)) | |
4633 | EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = new_bb->index; | |
4634 | ||
4635 | FOR_BB_INSNS (succ, insn) | |
4636 | if (INSN_P (insn)) | |
4637 | EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = succ->index; | |
4638 | ||
4639 | if (bitmap_bit_p (code_motion_visited_blocks, new_bb->index)) | |
4640 | { | |
4641 | bitmap_set_bit (code_motion_visited_blocks, succ->index); | |
4642 | bitmap_clear_bit (code_motion_visited_blocks, new_bb->index); | |
4643 | } | |
4644 | ||
4645 | gcc_assert (LABEL_P (BB_HEAD (new_bb)) | |
4646 | && LABEL_P (BB_HEAD (succ))); | |
4647 | ||
4648 | if (sched_verbose >= 4) | |
4649 | sel_print ("Swapping code labels %i and %i\n", | |
4650 | CODE_LABEL_NUMBER (BB_HEAD (new_bb)), | |
4651 | CODE_LABEL_NUMBER (BB_HEAD (succ))); | |
4652 | ||
4653 | i = CODE_LABEL_NUMBER (BB_HEAD (new_bb)); | |
4654 | CODE_LABEL_NUMBER (BB_HEAD (new_bb)) | |
4655 | = CODE_LABEL_NUMBER (BB_HEAD (succ)); | |
4656 | CODE_LABEL_NUMBER (BB_HEAD (succ)) = i; | |
4657 | } | |
4658 | } | |
4659 | } | |
4660 | ||
e855c69d AB |
4661 | return bb; |
4662 | } | |
4663 | ||
4664 | /* Return insn after which we must insert bookkeeping code for path(s) incoming | |
4665 | into E2->dest, except from E1->src. */ | |
4666 | static insn_t | |
4667 | find_place_for_bookkeeping (edge e1, edge e2) | |
4668 | { | |
4669 | insn_t place_to_insert; | |
4670 | /* Find a basic block that can hold bookkeeping. If it can be found, do not | |
4671 | create new basic block, but insert bookkeeping there. */ | |
b5b8b0ac | 4672 | basic_block book_block = find_block_for_bookkeeping (e1, e2, FALSE); |
e855c69d | 4673 | |
b5b8b0ac AO |
4674 | if (book_block) |
4675 | { | |
4676 | place_to_insert = BB_END (book_block); | |
4677 | ||
4678 | /* Don't use a block containing only debug insns for | |
4679 | bookkeeping, this causes scheduling differences between debug | |
4680 | and non-debug compilations, for the block would have been | |
4681 | removed already. */ | |
4682 | if (DEBUG_INSN_P (place_to_insert)) | |
4683 | { | |
4684 | rtx insn = sel_bb_head (book_block); | |
e855c69d | 4685 | |
b5b8b0ac AO |
4686 | while (insn != place_to_insert && |
4687 | (DEBUG_INSN_P (insn) || NOTE_P (insn))) | |
4688 | insn = NEXT_INSN (insn); | |
4689 | ||
4690 | if (insn == place_to_insert) | |
4691 | book_block = NULL; | |
4692 | } | |
4693 | } | |
4694 | ||
4695 | if (!book_block) | |
4696 | { | |
4697 | book_block = create_block_for_bookkeeping (e1, e2); | |
4698 | place_to_insert = BB_END (book_block); | |
4699 | if (sched_verbose >= 9) | |
4700 | sel_print ("New block is %i, split from bookkeeping block %i\n", | |
4701 | EDGE_SUCC (book_block, 0)->dest->index, book_block->index); | |
4702 | } | |
4703 | else | |
4704 | { | |
4705 | if (sched_verbose >= 9) | |
4706 | sel_print ("Pre-existing bookkeeping block is %i\n", book_block->index); | |
4707 | } | |
e855c69d AB |
4708 | |
4709 | /* If basic block ends with a jump, insert bookkeeping code right before it. */ | |
4710 | if (INSN_P (place_to_insert) && control_flow_insn_p (place_to_insert)) | |
4711 | place_to_insert = PREV_INSN (place_to_insert); | |
4712 | ||
4713 | return place_to_insert; | |
4714 | } | |
4715 | ||
4716 | /* Find a proper seqno for bookkeeing insn inserted at PLACE_TO_INSERT | |
4717 | for JOIN_POINT. */ | |
4718 | static int | |
4719 | find_seqno_for_bookkeeping (insn_t place_to_insert, insn_t join_point) | |
4720 | { | |
4721 | int seqno; | |
4722 | rtx next; | |
4723 | ||
4724 | /* Check if we are about to insert bookkeeping copy before a jump, and use | |
4725 | jump's seqno for the copy; otherwise, use JOIN_POINT's seqno. */ | |
4726 | next = NEXT_INSN (place_to_insert); | |
4727 | if (INSN_P (next) | |
4728 | && JUMP_P (next) | |
4729 | && BLOCK_FOR_INSN (next) == BLOCK_FOR_INSN (place_to_insert)) | |
da7ba240 AB |
4730 | { |
4731 | gcc_assert (INSN_SCHED_TIMES (next) == 0); | |
4732 | seqno = INSN_SEQNO (next); | |
4733 | } | |
e855c69d AB |
4734 | else if (INSN_SEQNO (join_point) > 0) |
4735 | seqno = INSN_SEQNO (join_point); | |
4736 | else | |
da7ba240 AB |
4737 | { |
4738 | seqno = get_seqno_by_preds (place_to_insert); | |
4739 | ||
4740 | /* Sometimes the fences can move in such a way that there will be | |
4741 | no instructions with positive seqno around this bookkeeping. | |
4742 | This means that there will be no way to get to it by a regular | |
4743 | fence movement. Never mind because we pick up such pieces for | |
4744 | rescheduling anyways, so any positive value will do for now. */ | |
4745 | if (seqno < 0) | |
4746 | { | |
4747 | gcc_assert (pipelining_p); | |
4748 | seqno = 1; | |
4749 | } | |
4750 | } | |
e855c69d AB |
4751 | |
4752 | gcc_assert (seqno > 0); | |
4753 | return seqno; | |
4754 | } | |
4755 | ||
4756 | /* Insert bookkeeping copy of C_EXPS's insn after PLACE_TO_INSERT, assigning | |
4757 | NEW_SEQNO to it. Return created insn. */ | |
4758 | static insn_t | |
4759 | emit_bookkeeping_insn (insn_t place_to_insert, expr_t c_expr, int new_seqno) | |
4760 | { | |
4761 | rtx new_insn_rtx = create_copy_of_insn_rtx (EXPR_INSN_RTX (c_expr)); | |
4762 | ||
4763 | vinsn_t new_vinsn | |
4764 | = create_vinsn_from_insn_rtx (new_insn_rtx, | |
4765 | VINSN_UNIQUE_P (EXPR_VINSN (c_expr))); | |
4766 | ||
4767 | insn_t new_insn = emit_insn_from_expr_after (c_expr, new_vinsn, new_seqno, | |
4768 | place_to_insert); | |
4769 | ||
4770 | INSN_SCHED_TIMES (new_insn) = 0; | |
4771 | bitmap_set_bit (current_copies, INSN_UID (new_insn)); | |
4772 | ||
4773 | return new_insn; | |
4774 | } | |
4775 | ||
4776 | /* Generate a bookkeeping copy of C_EXPR's insn for path(s) incoming into to | |
4777 | E2->dest, except from E1->src (there may be a sequence of empty blocks | |
4778 | between E1->src and E2->dest). Return block containing the copy. | |
4779 | All scheduler data is initialized for the newly created insn. */ | |
4780 | static basic_block | |
4781 | generate_bookkeeping_insn (expr_t c_expr, edge e1, edge e2) | |
4782 | { | |
4783 | insn_t join_point, place_to_insert, new_insn; | |
4784 | int new_seqno; | |
4785 | bool need_to_exchange_data_sets; | |
4786 | ||
4787 | if (sched_verbose >= 4) | |
4788 | sel_print ("Generating bookkeeping insn (%d->%d)\n", e1->src->index, | |
4789 | e2->dest->index); | |
4790 | ||
4791 | join_point = sel_bb_head (e2->dest); | |
4792 | place_to_insert = find_place_for_bookkeeping (e1, e2); | |
b5b8b0ac AO |
4793 | if (!place_to_insert) |
4794 | return NULL; | |
e855c69d AB |
4795 | new_seqno = find_seqno_for_bookkeeping (place_to_insert, join_point); |
4796 | need_to_exchange_data_sets | |
4797 | = sel_bb_empty_p (BLOCK_FOR_INSN (place_to_insert)); | |
4798 | ||
4799 | new_insn = emit_bookkeeping_insn (place_to_insert, c_expr, new_seqno); | |
4800 | ||
4801 | /* When inserting bookkeeping insn in new block, av sets should be | |
4802 | following: old basic block (that now holds bookkeeping) data sets are | |
4803 | the same as was before generation of bookkeeping, and new basic block | |
4804 | (that now hold all other insns of old basic block) data sets are | |
4805 | invalid. So exchange data sets for these basic blocks as sel_split_block | |
4806 | mistakenly exchanges them in this case. Cannot do it earlier because | |
4807 | when single instruction is added to new basic block it should hold NULL | |
4808 | lv_set. */ | |
4809 | if (need_to_exchange_data_sets) | |
4810 | exchange_data_sets (BLOCK_FOR_INSN (new_insn), | |
4811 | BLOCK_FOR_INSN (join_point)); | |
4812 | ||
4813 | stat_bookkeeping_copies++; | |
4814 | return BLOCK_FOR_INSN (new_insn); | |
4815 | } | |
4816 | ||
4817 | /* Remove from AV_PTR all insns that may need bookkeeping when scheduling | |
4818 | on FENCE, but we are unable to copy them. */ | |
4819 | static void | |
4820 | remove_insns_that_need_bookkeeping (fence_t fence, av_set_t *av_ptr) | |
4821 | { | |
4822 | expr_t expr; | |
4823 | av_set_iterator i; | |
4824 | ||
4825 | /* An expression does not need bookkeeping if it is available on all paths | |
4826 | from current block to original block and current block dominates | |
4827 | original block. We check availability on all paths by examining | |
4828 | EXPR_SPEC; this is not equivalent, because it may be positive even | |
4829 | if expr is available on all paths (but if expr is not available on | |
4830 | any path, EXPR_SPEC will be positive). */ | |
4831 | ||
4832 | FOR_EACH_EXPR_1 (expr, i, av_ptr) | |
4833 | { | |
4834 | if (!control_flow_insn_p (EXPR_INSN_RTX (expr)) | |
4835 | && (!bookkeeping_p || VINSN_UNIQUE_P (EXPR_VINSN (expr))) | |
4836 | && (EXPR_SPEC (expr) | |
4837 | || !EXPR_ORIG_BB_INDEX (expr) | |
4838 | || !dominated_by_p (CDI_DOMINATORS, | |
4839 | BASIC_BLOCK (EXPR_ORIG_BB_INDEX (expr)), | |
4840 | BLOCK_FOR_INSN (FENCE_INSN (fence))))) | |
4841 | { | |
4842 | if (sched_verbose >= 4) | |
4843 | sel_print ("Expr %d removed because it would need bookkeeping, which " | |
4844 | "cannot be created\n", INSN_UID (EXPR_INSN_RTX (expr))); | |
4845 | av_set_iter_remove (&i); | |
4846 | } | |
4847 | } | |
4848 | } | |
4849 | ||
4850 | /* Moving conditional jump through some instructions. | |
4851 | ||
4852 | Consider example: | |
4853 | ||
4854 | ... <- current scheduling point | |
4855 | NOTE BASIC BLOCK: <- bb header | |
4856 | (p8) add r14=r14+0x9;; | |
4857 | (p8) mov [r14]=r23 | |
4858 | (!p8) jump L1;; | |
4859 | NOTE BASIC BLOCK: | |
4860 | ... | |
4861 | ||
4862 | We can schedule jump one cycle earlier, than mov, because they cannot be | |
4863 | executed together as their predicates are mutually exclusive. | |
4864 | ||
4865 | This is done in this way: first, new fallthrough basic block is created | |
4866 | after jump (it is always can be done, because there already should be a | |
4867 | fallthrough block, where control flow goes in case of predicate being true - | |
4868 | in our example; otherwise there should be a dependence between those | |
4869 | instructions and jump and we cannot schedule jump right now); | |
4870 | next, all instructions between jump and current scheduling point are moved | |
4871 | to this new block. And the result is this: | |
4872 | ||
4873 | NOTE BASIC BLOCK: | |
4874 | (!p8) jump L1 <- current scheduling point | |
4875 | NOTE BASIC BLOCK: <- bb header | |
4876 | (p8) add r14=r14+0x9;; | |
4877 | (p8) mov [r14]=r23 | |
4878 | NOTE BASIC BLOCK: | |
4879 | ... | |
4880 | */ | |
4881 | static void | |
4882 | move_cond_jump (rtx insn, bnd_t bnd) | |
4883 | { | |
4884 | edge ft_edge; | |
4885 | basic_block block_from, block_next, block_new; | |
4886 | rtx next, prev, link; | |
4887 | ||
4888 | /* BLOCK_FROM holds basic block of the jump. */ | |
4889 | block_from = BLOCK_FOR_INSN (insn); | |
4890 | ||
4891 | /* Moving of jump should not cross any other jumps or | |
4892 | beginnings of new basic blocks. */ | |
4893 | gcc_assert (block_from == BLOCK_FOR_INSN (BND_TO (bnd))); | |
4894 | ||
4895 | /* Jump is moved to the boundary. */ | |
4896 | prev = BND_TO (bnd); | |
4897 | next = PREV_INSN (insn); | |
4898 | BND_TO (bnd) = insn; | |
4899 | ||
4900 | ft_edge = find_fallthru_edge (block_from); | |
4901 | block_next = ft_edge->dest; | |
4902 | /* There must be a fallthrough block (or where should go | |
4903 | control flow in case of false jump predicate otherwise?). */ | |
4904 | gcc_assert (block_next); | |
4905 | ||
4906 | /* Create new empty basic block after source block. */ | |
4907 | block_new = sel_split_edge (ft_edge); | |
4908 | gcc_assert (block_new->next_bb == block_next | |
4909 | && block_from->next_bb == block_new); | |
4910 | ||
4911 | gcc_assert (BB_END (block_from) == insn); | |
4912 | ||
4913 | /* Move all instructions except INSN from BLOCK_FROM to | |
4914 | BLOCK_NEW. */ | |
4915 | for (link = prev; link != insn; link = NEXT_INSN (link)) | |
4916 | { | |
4917 | EXPR_ORIG_BB_INDEX (INSN_EXPR (link)) = block_new->index; | |
4918 | df_insn_change_bb (link, block_new); | |
4919 | } | |
4920 | ||
4921 | /* Set correct basic block and instructions properties. */ | |
4922 | BB_END (block_new) = PREV_INSN (insn); | |
4923 | ||
4924 | NEXT_INSN (PREV_INSN (prev)) = insn; | |
4925 | PREV_INSN (insn) = PREV_INSN (prev); | |
4926 | ||
4927 | /* Assert there is no jump to BLOCK_NEW, only fallthrough edge. */ | |
4928 | gcc_assert (NOTE_INSN_BASIC_BLOCK_P (BB_HEAD (block_new))); | |
4929 | PREV_INSN (prev) = BB_HEAD (block_new); | |
4930 | NEXT_INSN (next) = NEXT_INSN (BB_HEAD (block_new)); | |
4931 | NEXT_INSN (BB_HEAD (block_new)) = prev; | |
4932 | PREV_INSN (NEXT_INSN (next)) = next; | |
4933 | ||
4934 | gcc_assert (!sel_bb_empty_p (block_from) | |
4935 | && !sel_bb_empty_p (block_new)); | |
4936 | ||
4937 | /* Update data sets for BLOCK_NEW to represent that INSN and | |
4938 | instructions from the other branch of INSN is no longer | |
4939 | available at BLOCK_NEW. */ | |
4940 | BB_AV_LEVEL (block_new) = global_level; | |
4941 | gcc_assert (BB_LV_SET (block_new) == NULL); | |
4942 | BB_LV_SET (block_new) = get_clear_regset_from_pool (); | |
4943 | update_data_sets (sel_bb_head (block_new)); | |
4944 | ||
4945 | /* INSN is a new basic block header - so prepare its data | |
4946 | structures and update availability and liveness sets. */ | |
4947 | update_data_sets (insn); | |
4948 | ||
4949 | if (sched_verbose >= 4) | |
4950 | sel_print ("Moving jump %d\n", INSN_UID (insn)); | |
4951 | } | |
4952 | ||
4953 | /* Remove nops generated during move_op for preventing removal of empty | |
4954 | basic blocks. */ | |
4955 | static void | |
b5b8b0ac | 4956 | remove_temp_moveop_nops (bool full_tidying) |
e855c69d AB |
4957 | { |
4958 | int i; | |
4959 | insn_t insn; | |
4960 | ||
4961 | for (i = 0; VEC_iterate (insn_t, vec_temp_moveop_nops, i, insn); i++) | |
4962 | { | |
4963 | gcc_assert (INSN_NOP_P (insn)); | |
b5b8b0ac | 4964 | return_nop_to_pool (insn, full_tidying); |
e855c69d AB |
4965 | } |
4966 | ||
4967 | /* Empty the vector. */ | |
4968 | if (VEC_length (insn_t, vec_temp_moveop_nops) > 0) | |
4969 | VEC_block_remove (insn_t, vec_temp_moveop_nops, 0, | |
4970 | VEC_length (insn_t, vec_temp_moveop_nops)); | |
4971 | } | |
4972 | ||
4973 | /* Records the maximal UID before moving up an instruction. Used for | |
4974 | distinguishing between bookkeeping copies and original insns. */ | |
4975 | static int max_uid_before_move_op = 0; | |
4976 | ||
4977 | /* Remove from AV_VLIW_P all instructions but next when debug counter | |
4978 | tells us so. Next instruction is fetched from BNDS. */ | |
4979 | static void | |
4980 | remove_insns_for_debug (blist_t bnds, av_set_t *av_vliw_p) | |
4981 | { | |
4982 | if (! dbg_cnt (sel_sched_insn_cnt)) | |
4983 | /* Leave only the next insn in av_vliw. */ | |
4984 | { | |
4985 | av_set_iterator av_it; | |
4986 | expr_t expr; | |
4987 | bnd_t bnd = BLIST_BND (bnds); | |
4988 | insn_t next = BND_TO (bnd); | |
4989 | ||
4990 | gcc_assert (BLIST_NEXT (bnds) == NULL); | |
4991 | ||
4992 | FOR_EACH_EXPR_1 (expr, av_it, av_vliw_p) | |
4993 | if (EXPR_INSN_RTX (expr) != next) | |
4994 | av_set_iter_remove (&av_it); | |
4995 | } | |
4996 | } | |
4997 | ||
4998 | /* Compute available instructions on BNDS. FENCE is the current fence. Write | |
4999 | the computed set to *AV_VLIW_P. */ | |
5000 | static void | |
5001 | compute_av_set_on_boundaries (fence_t fence, blist_t bnds, av_set_t *av_vliw_p) | |
5002 | { | |
5003 | if (sched_verbose >= 2) | |
5004 | { | |
5005 | sel_print ("Boundaries: "); | |
5006 | dump_blist (bnds); | |
5007 | sel_print ("\n"); | |
5008 | } | |
5009 | ||
5010 | for (; bnds; bnds = BLIST_NEXT (bnds)) | |
5011 | { | |
5012 | bnd_t bnd = BLIST_BND (bnds); | |
5013 | av_set_t av1_copy; | |
5014 | insn_t bnd_to = BND_TO (bnd); | |
5015 | ||
5016 | /* Rewind BND->TO to the basic block header in case some bookkeeping | |
5017 | instructions were inserted before BND->TO and it needs to be | |
5018 | adjusted. */ | |
5019 | if (sel_bb_head_p (bnd_to)) | |
5020 | gcc_assert (INSN_SCHED_TIMES (bnd_to) == 0); | |
5021 | else | |
5022 | while (INSN_SCHED_TIMES (PREV_INSN (bnd_to)) == 0) | |
5023 | { | |
5024 | bnd_to = PREV_INSN (bnd_to); | |
5025 | if (sel_bb_head_p (bnd_to)) | |
5026 | break; | |
5027 | } | |
5028 | ||
5029 | if (BND_TO (bnd) != bnd_to) | |
5030 | { | |
5031 | gcc_assert (FENCE_INSN (fence) == BND_TO (bnd)); | |
5032 | FENCE_INSN (fence) = bnd_to; | |
5033 | BND_TO (bnd) = bnd_to; | |
5034 | } | |
5035 | ||
5036 | av_set_clear (&BND_AV (bnd)); | |
5037 | BND_AV (bnd) = compute_av_set (BND_TO (bnd), NULL, 0, true); | |
5038 | ||
5039 | av_set_clear (&BND_AV1 (bnd)); | |
5040 | BND_AV1 (bnd) = av_set_copy (BND_AV (bnd)); | |
5041 | ||
5042 | moveup_set_inside_insn_group (&BND_AV1 (bnd), NULL); | |
5043 | ||
5044 | av1_copy = av_set_copy (BND_AV1 (bnd)); | |
5045 | av_set_union_and_clear (av_vliw_p, &av1_copy, NULL); | |
5046 | } | |
5047 | ||
5048 | if (sched_verbose >= 2) | |
5049 | { | |
5050 | sel_print ("Available exprs (vliw form): "); | |
5051 | dump_av_set (*av_vliw_p); | |
5052 | sel_print ("\n"); | |
5053 | } | |
5054 | } | |
5055 | ||
5056 | /* Calculate the sequential av set on BND corresponding to the EXPR_VLIW | |
5057 | expression. When FOR_MOVEOP is true, also replace the register of | |
5058 | expressions found with the register from EXPR_VLIW. */ | |
5059 | static av_set_t | |
5060 | find_sequential_best_exprs (bnd_t bnd, expr_t expr_vliw, bool for_moveop) | |
5061 | { | |
5062 | av_set_t expr_seq = NULL; | |
5063 | expr_t expr; | |
5064 | av_set_iterator i; | |
5065 | ||
5066 | FOR_EACH_EXPR (expr, i, BND_AV (bnd)) | |
5067 | { | |
5068 | if (equal_after_moveup_path_p (expr, NULL, expr_vliw)) | |
5069 | { | |
5070 | if (for_moveop) | |
5071 | { | |
5072 | /* The sequential expression has the right form to pass | |
5073 | to move_op except when renaming happened. Put the | |
5074 | correct register in EXPR then. */ | |
5075 | if (EXPR_SEPARABLE_P (expr) && REG_P (EXPR_LHS (expr))) | |
5076 | { | |
5077 | if (expr_dest_regno (expr) != expr_dest_regno (expr_vliw)) | |
5078 | { | |
5079 | replace_dest_with_reg_in_expr (expr, EXPR_LHS (expr_vliw)); | |
5080 | stat_renamed_scheduled++; | |
5081 | } | |
5082 | /* Also put the correct TARGET_AVAILABLE bit on the expr. | |
5083 | This is needed when renaming came up with original | |
5084 | register. */ | |
5085 | else if (EXPR_TARGET_AVAILABLE (expr) | |
5086 | != EXPR_TARGET_AVAILABLE (expr_vliw)) | |
5087 | { | |
5088 | gcc_assert (EXPR_TARGET_AVAILABLE (expr_vliw) == 1); | |
5089 | EXPR_TARGET_AVAILABLE (expr) = 1; | |
5090 | } | |
5091 | } | |
5092 | if (EXPR_WAS_SUBSTITUTED (expr)) | |
5093 | stat_substitutions_total++; | |
5094 | } | |
5095 | ||
5096 | av_set_add (&expr_seq, expr); | |
5097 | ||
5098 | /* With substitution inside insn group, it is possible | |
5099 | that more than one expression in expr_seq will correspond | |
5100 | to expr_vliw. In this case, choose one as the attempt to | |
5101 | move both leads to miscompiles. */ | |
5102 | break; | |
5103 | } | |
5104 | } | |
5105 | ||
5106 | if (for_moveop && sched_verbose >= 2) | |
5107 | { | |
5108 | sel_print ("Best expression(s) (sequential form): "); | |
5109 | dump_av_set (expr_seq); | |
5110 | sel_print ("\n"); | |
5111 | } | |
5112 | ||
5113 | return expr_seq; | |
5114 | } | |
5115 | ||
5116 | ||
5117 | /* Move nop to previous block. */ | |
5118 | static void ATTRIBUTE_UNUSED | |
5119 | move_nop_to_previous_block (insn_t nop, basic_block prev_bb) | |
5120 | { | |
5121 | insn_t prev_insn, next_insn, note; | |
5122 | ||
5123 | gcc_assert (sel_bb_head_p (nop) | |
5124 | && prev_bb == BLOCK_FOR_INSN (nop)->prev_bb); | |
5125 | note = bb_note (BLOCK_FOR_INSN (nop)); | |
5126 | prev_insn = sel_bb_end (prev_bb); | |
5127 | next_insn = NEXT_INSN (nop); | |
5128 | gcc_assert (prev_insn != NULL_RTX | |
5129 | && PREV_INSN (note) == prev_insn); | |
5130 | ||
5131 | NEXT_INSN (prev_insn) = nop; | |
5132 | PREV_INSN (nop) = prev_insn; | |
5133 | ||
5134 | PREV_INSN (note) = nop; | |
5135 | NEXT_INSN (note) = next_insn; | |
5136 | ||
5137 | NEXT_INSN (nop) = note; | |
5138 | PREV_INSN (next_insn) = note; | |
5139 | ||
5140 | BB_END (prev_bb) = nop; | |
5141 | BLOCK_FOR_INSN (nop) = prev_bb; | |
5142 | } | |
5143 | ||
5144 | /* Prepare a place to insert the chosen expression on BND. */ | |
5145 | static insn_t | |
5146 | prepare_place_to_insert (bnd_t bnd) | |
5147 | { | |
5148 | insn_t place_to_insert; | |
5149 | ||
5150 | /* Init place_to_insert before calling move_op, as the later | |
5151 | can possibly remove BND_TO (bnd). */ | |
5152 | if (/* If this is not the first insn scheduled. */ | |
5153 | BND_PTR (bnd)) | |
5154 | { | |
5155 | /* Add it after last scheduled. */ | |
5156 | place_to_insert = ILIST_INSN (BND_PTR (bnd)); | |
b5b8b0ac AO |
5157 | if (DEBUG_INSN_P (place_to_insert)) |
5158 | { | |
5159 | ilist_t l = BND_PTR (bnd); | |
5160 | while ((l = ILIST_NEXT (l)) && | |
5161 | DEBUG_INSN_P (ILIST_INSN (l))) | |
5162 | ; | |
5163 | if (!l) | |
5164 | place_to_insert = NULL; | |
5165 | } | |
e855c69d AB |
5166 | } |
5167 | else | |
b5b8b0ac AO |
5168 | place_to_insert = NULL; |
5169 | ||
5170 | if (!place_to_insert) | |
e855c69d AB |
5171 | { |
5172 | /* Add it before BND_TO. The difference is in the | |
5173 | basic block, where INSN will be added. */ | |
5174 | place_to_insert = get_nop_from_pool (BND_TO (bnd)); | |
5175 | gcc_assert (BLOCK_FOR_INSN (place_to_insert) | |
5176 | == BLOCK_FOR_INSN (BND_TO (bnd))); | |
5177 | } | |
5178 | ||
5179 | return place_to_insert; | |
5180 | } | |
5181 | ||
5182 | /* Find original instructions for EXPR_SEQ and move it to BND boundary. | |
5183 | Return the expression to emit in C_EXPR. */ | |
72a54528 | 5184 | static bool |
e855c69d AB |
5185 | move_exprs_to_boundary (bnd_t bnd, expr_t expr_vliw, |
5186 | av_set_t expr_seq, expr_t c_expr) | |
5187 | { | |
72a54528 | 5188 | bool b, should_move; |
e855c69d AB |
5189 | unsigned book_uid; |
5190 | bitmap_iterator bi; | |
5191 | int n_bookkeeping_copies_before_moveop; | |
5192 | ||
5193 | /* Make a move. This call will remove the original operation, | |
5194 | insert all necessary bookkeeping instructions and update the | |
5195 | data sets. After that all we have to do is add the operation | |
5196 | at before BND_TO (BND). */ | |
5197 | n_bookkeeping_copies_before_moveop = stat_bookkeeping_copies; | |
5198 | max_uid_before_move_op = get_max_uid (); | |
5199 | bitmap_clear (current_copies); | |
5200 | bitmap_clear (current_originators); | |
5201 | ||
5202 | b = move_op (BND_TO (bnd), expr_seq, expr_vliw, | |
72a54528 | 5203 | get_dest_from_orig_ops (expr_seq), c_expr, &should_move); |
e855c69d AB |
5204 | |
5205 | /* We should be able to find the expression we've chosen for | |
5206 | scheduling. */ | |
72a54528 | 5207 | gcc_assert (b); |
e855c69d AB |
5208 | |
5209 | if (stat_bookkeeping_copies > n_bookkeeping_copies_before_moveop) | |
5210 | stat_insns_needed_bookkeeping++; | |
5211 | ||
5212 | EXECUTE_IF_SET_IN_BITMAP (current_copies, 0, book_uid, bi) | |
5213 | { | |
5214 | /* We allocate these bitmaps lazily. */ | |
5215 | if (! INSN_ORIGINATORS_BY_UID (book_uid)) | |
5216 | INSN_ORIGINATORS_BY_UID (book_uid) = BITMAP_ALLOC (NULL); | |
5217 | ||
5218 | bitmap_copy (INSN_ORIGINATORS_BY_UID (book_uid), | |
5219 | current_originators); | |
5220 | } | |
72a54528 AM |
5221 | |
5222 | return should_move; | |
e855c69d AB |
5223 | } |
5224 | ||
5225 | ||
5226 | /* Debug a DFA state as an array of bytes. */ | |
5227 | static void | |
5228 | debug_state (state_t state) | |
5229 | { | |
5230 | unsigned char *p; | |
5231 | unsigned int i, size = dfa_state_size; | |
5232 | ||
5233 | sel_print ("state (%u):", size); | |
5234 | for (i = 0, p = (unsigned char *) state; i < size; i++) | |
5235 | sel_print (" %d", p[i]); | |
5236 | sel_print ("\n"); | |
5237 | } | |
5238 | ||
5239 | /* Advance state on FENCE with INSN. Return true if INSN is | |
5240 | an ASM, and we should advance state once more. */ | |
5241 | static bool | |
5242 | advance_state_on_fence (fence_t fence, insn_t insn) | |
5243 | { | |
5244 | bool asm_p; | |
5245 | ||
5246 | if (recog_memoized (insn) >= 0) | |
5247 | { | |
5248 | int res; | |
5249 | state_t temp_state = alloca (dfa_state_size); | |
5250 | ||
5251 | gcc_assert (!INSN_ASM_P (insn)); | |
5252 | asm_p = false; | |
5253 | ||
5254 | memcpy (temp_state, FENCE_STATE (fence), dfa_state_size); | |
5255 | res = state_transition (FENCE_STATE (fence), insn); | |
5256 | gcc_assert (res < 0); | |
5257 | ||
5258 | if (memcmp (temp_state, FENCE_STATE (fence), dfa_state_size)) | |
5259 | { | |
5260 | FENCE_ISSUED_INSNS (fence)++; | |
5261 | ||
5262 | /* We should never issue more than issue_rate insns. */ | |
5263 | if (FENCE_ISSUED_INSNS (fence) > issue_rate) | |
5264 | gcc_unreachable (); | |
5265 | } | |
5266 | } | |
5267 | else | |
5268 | { | |
5269 | /* This could be an ASM insn which we'd like to schedule | |
5270 | on the next cycle. */ | |
5271 | asm_p = INSN_ASM_P (insn); | |
5272 | if (!FENCE_STARTS_CYCLE_P (fence) && asm_p) | |
5273 | advance_one_cycle (fence); | |
5274 | } | |
5275 | ||
5276 | if (sched_verbose >= 2) | |
5277 | debug_state (FENCE_STATE (fence)); | |
b5b8b0ac AO |
5278 | if (!DEBUG_INSN_P (insn)) |
5279 | FENCE_STARTS_CYCLE_P (fence) = 0; | |
e855c69d AB |
5280 | return asm_p; |
5281 | } | |
5282 | ||
5283 | /* Update FENCE on which INSN was scheduled and this INSN, too. NEED_STALL | |
5284 | is nonzero if we need to stall after issuing INSN. */ | |
5285 | static void | |
5286 | update_fence_and_insn (fence_t fence, insn_t insn, int need_stall) | |
5287 | { | |
5288 | bool asm_p; | |
5289 | ||
5290 | /* First, reflect that something is scheduled on this fence. */ | |
5291 | asm_p = advance_state_on_fence (fence, insn); | |
5292 | FENCE_LAST_SCHEDULED_INSN (fence) = insn; | |
5293 | VEC_safe_push (rtx, gc, FENCE_EXECUTING_INSNS (fence), insn); | |
5294 | if (SCHED_GROUP_P (insn)) | |
5295 | { | |
5296 | FENCE_SCHED_NEXT (fence) = INSN_SCHED_NEXT (insn); | |
5297 | SCHED_GROUP_P (insn) = 0; | |
5298 | } | |
5299 | else | |
5300 | FENCE_SCHED_NEXT (fence) = NULL_RTX; | |
5301 | if (INSN_UID (insn) < FENCE_READY_TICKS_SIZE (fence)) | |
5302 | FENCE_READY_TICKS (fence) [INSN_UID (insn)] = 0; | |
5303 | ||
5304 | /* Set instruction scheduling info. This will be used in bundling, | |
5305 | pipelining, tick computations etc. */ | |
5306 | ++INSN_SCHED_TIMES (insn); | |
5307 | EXPR_TARGET_AVAILABLE (INSN_EXPR (insn)) = true; | |
5308 | EXPR_ORIG_SCHED_CYCLE (INSN_EXPR (insn)) = FENCE_CYCLE (fence); | |
5309 | INSN_AFTER_STALL_P (insn) = FENCE_AFTER_STALL_P (fence); | |
5310 | INSN_SCHED_CYCLE (insn) = FENCE_CYCLE (fence); | |
5311 | ||
5312 | /* This does not account for adjust_cost hooks, just add the biggest | |
5313 | constant the hook may add to the latency. TODO: make this | |
5314 | a target dependent constant. */ | |
5315 | INSN_READY_CYCLE (insn) | |
5316 | = INSN_SCHED_CYCLE (insn) + (INSN_CODE (insn) < 0 | |
5317 | ? 1 | |
5318 | : maximal_insn_latency (insn) + 1); | |
5319 | ||
5320 | /* Change these fields last, as they're used above. */ | |
5321 | FENCE_AFTER_STALL_P (fence) = 0; | |
5322 | if (asm_p || need_stall) | |
5323 | advance_one_cycle (fence); | |
5324 | ||
5325 | /* Indicate that we've scheduled something on this fence. */ | |
5326 | FENCE_SCHEDULED_P (fence) = true; | |
5327 | scheduled_something_on_previous_fence = true; | |
5328 | ||
5329 | /* Print debug information when insn's fields are updated. */ | |
5330 | if (sched_verbose >= 2) | |
5331 | { | |
5332 | sel_print ("Scheduling insn: "); | |
5333 | dump_insn_1 (insn, 1); | |
5334 | sel_print ("\n"); | |
5335 | } | |
5336 | } | |
5337 | ||
b5b8b0ac AO |
5338 | /* Update boundary BND (and, if needed, FENCE) with INSN, remove the |
5339 | old boundary from BNDSP, add new boundaries to BNDS_TAIL_P and | |
5340 | return it. */ | |
e855c69d | 5341 | static blist_t * |
b5b8b0ac | 5342 | update_boundaries (fence_t fence, bnd_t bnd, insn_t insn, blist_t *bndsp, |
e855c69d AB |
5343 | blist_t *bnds_tailp) |
5344 | { | |
5345 | succ_iterator si; | |
5346 | insn_t succ; | |
5347 | ||
5348 | advance_deps_context (BND_DC (bnd), insn); | |
5349 | FOR_EACH_SUCC_1 (succ, si, insn, | |
5350 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
5351 | { | |
5352 | ilist_t ptr = ilist_copy (BND_PTR (bnd)); | |
5353 | ||
5354 | ilist_add (&ptr, insn); | |
b5b8b0ac AO |
5355 | |
5356 | if (DEBUG_INSN_P (insn) && sel_bb_end_p (insn) | |
5357 | && is_ineligible_successor (succ, ptr)) | |
5358 | { | |
5359 | ilist_clear (&ptr); | |
5360 | continue; | |
5361 | } | |
5362 | ||
5363 | if (FENCE_INSN (fence) == insn && !sel_bb_end_p (insn)) | |
5364 | { | |
5365 | if (sched_verbose >= 9) | |
5366 | sel_print ("Updating fence insn from %i to %i\n", | |
5367 | INSN_UID (insn), INSN_UID (succ)); | |
5368 | FENCE_INSN (fence) = succ; | |
5369 | } | |
e855c69d AB |
5370 | blist_add (bnds_tailp, succ, ptr, BND_DC (bnd)); |
5371 | bnds_tailp = &BLIST_NEXT (*bnds_tailp); | |
5372 | } | |
5373 | ||
5374 | blist_remove (bndsp); | |
5375 | return bnds_tailp; | |
5376 | } | |
5377 | ||
5378 | /* Schedule EXPR_VLIW on BND. Return the insn emitted. */ | |
5379 | static insn_t | |
5380 | schedule_expr_on_boundary (bnd_t bnd, expr_t expr_vliw, int seqno) | |
5381 | { | |
5382 | av_set_t expr_seq; | |
5383 | expr_t c_expr = XALLOCA (expr_def); | |
5384 | insn_t place_to_insert; | |
5385 | insn_t insn; | |
72a54528 | 5386 | bool should_move; |
e855c69d AB |
5387 | |
5388 | expr_seq = find_sequential_best_exprs (bnd, expr_vliw, true); | |
5389 | ||
5390 | /* In case of scheduling a jump skipping some other instructions, | |
5391 | prepare CFG. After this, jump is at the boundary and can be | |
5392 | scheduled as usual insn by MOVE_OP. */ | |
5393 | if (vinsn_cond_branch_p (EXPR_VINSN (expr_vliw))) | |
5394 | { | |
5395 | insn = EXPR_INSN_RTX (expr_vliw); | |
5396 | ||
5397 | /* Speculative jumps are not handled. */ | |
5398 | if (insn != BND_TO (bnd) | |
5399 | && !sel_insn_is_speculation_check (insn)) | |
5400 | move_cond_jump (insn, bnd); | |
5401 | } | |
5402 | ||
e855c69d AB |
5403 | /* Find a place for C_EXPR to schedule. */ |
5404 | place_to_insert = prepare_place_to_insert (bnd); | |
72a54528 | 5405 | should_move = move_exprs_to_boundary (bnd, expr_vliw, expr_seq, c_expr); |
e855c69d AB |
5406 | clear_expr (c_expr); |
5407 | ||
5408 | /* Add the instruction. The corner case to care about is when | |
5409 | the expr_seq set has more than one expr, and we chose the one that | |
5410 | is not equal to expr_vliw. Then expr_vliw may be insn in stream, and | |
5411 | we can't use it. Generate the new vinsn. */ | |
5412 | if (INSN_IN_STREAM_P (EXPR_INSN_RTX (expr_vliw))) | |
5413 | { | |
5414 | vinsn_t vinsn_new; | |
5415 | ||
5416 | vinsn_new = vinsn_copy (EXPR_VINSN (expr_vliw), false); | |
5417 | change_vinsn_in_expr (expr_vliw, vinsn_new); | |
72a54528 | 5418 | should_move = false; |
e855c69d | 5419 | } |
72a54528 AM |
5420 | if (should_move) |
5421 | insn = sel_move_insn (expr_vliw, seqno, place_to_insert); | |
5422 | else | |
e855c69d AB |
5423 | insn = emit_insn_from_expr_after (expr_vliw, NULL, seqno, |
5424 | place_to_insert); | |
e855c69d AB |
5425 | |
5426 | /* Return the nops generated for preserving of data sets back | |
5427 | into pool. */ | |
5428 | if (INSN_NOP_P (place_to_insert)) | |
b5b8b0ac AO |
5429 | return_nop_to_pool (place_to_insert, !DEBUG_INSN_P (insn)); |
5430 | remove_temp_moveop_nops (!DEBUG_INSN_P (insn)); | |
e855c69d AB |
5431 | |
5432 | av_set_clear (&expr_seq); | |
5433 | ||
5434 | /* Save the expression scheduled so to reset target availability if we'll | |
5435 | meet it later on the same fence. */ | |
5436 | if (EXPR_WAS_RENAMED (expr_vliw)) | |
5437 | vinsn_vec_add (&vec_target_unavailable_vinsns, INSN_EXPR (insn)); | |
5438 | ||
5439 | /* Check that the recent movement didn't destroyed loop | |
5440 | structure. */ | |
5441 | gcc_assert (!pipelining_p | |
5442 | || current_loop_nest == NULL | |
5443 | || loop_latch_edge (current_loop_nest)); | |
5444 | return insn; | |
5445 | } | |
5446 | ||
5447 | /* Stall for N cycles on FENCE. */ | |
5448 | static void | |
5449 | stall_for_cycles (fence_t fence, int n) | |
5450 | { | |
5451 | int could_more; | |
5452 | ||
5453 | could_more = n > 1 || FENCE_ISSUED_INSNS (fence) < issue_rate; | |
5454 | while (n--) | |
5455 | advance_one_cycle (fence); | |
5456 | if (could_more) | |
5457 | FENCE_AFTER_STALL_P (fence) = 1; | |
5458 | } | |
5459 | ||
5460 | /* Gather a parallel group of insns at FENCE and assign their seqno | |
5461 | to SEQNO. All scheduled insns are gathered in SCHEDULED_INSNS_TAILPP | |
5462 | list for later recalculation of seqnos. */ | |
5463 | static void | |
5464 | fill_insns (fence_t fence, int seqno, ilist_t **scheduled_insns_tailpp) | |
5465 | { | |
5466 | blist_t bnds = NULL, *bnds_tailp; | |
5467 | av_set_t av_vliw = NULL; | |
5468 | insn_t insn = FENCE_INSN (fence); | |
5469 | ||
5470 | if (sched_verbose >= 2) | |
5471 | sel_print ("Starting fill_insns for insn %d, cycle %d\n", | |
5472 | INSN_UID (insn), FENCE_CYCLE (fence)); | |
5473 | ||
5474 | blist_add (&bnds, insn, NULL, FENCE_DC (fence)); | |
5475 | bnds_tailp = &BLIST_NEXT (bnds); | |
5476 | set_target_context (FENCE_TC (fence)); | |
5477 | target_bb = INSN_BB (insn); | |
5478 | ||
5479 | /* Do while we can add any operation to the current group. */ | |
5480 | do | |
5481 | { | |
5482 | blist_t *bnds_tailp1, *bndsp; | |
5483 | expr_t expr_vliw; | |
5484 | int need_stall; | |
5485 | int was_stall = 0, scheduled_insns = 0, stall_iterations = 0; | |
5486 | int max_insns = pipelining_p ? issue_rate : 2 * issue_rate; | |
5487 | int max_stall = pipelining_p ? 1 : 3; | |
b5b8b0ac AO |
5488 | bool last_insn_was_debug = false; |
5489 | bool was_debug_bb_end_p = false; | |
5490 | ||
e855c69d AB |
5491 | compute_av_set_on_boundaries (fence, bnds, &av_vliw); |
5492 | remove_insns_that_need_bookkeeping (fence, &av_vliw); | |
5493 | remove_insns_for_debug (bnds, &av_vliw); | |
5494 | ||
5495 | /* Return early if we have nothing to schedule. */ | |
5496 | if (av_vliw == NULL) | |
5497 | break; | |
5498 | ||
5499 | /* Choose the best expression and, if needed, destination register | |
5500 | for it. */ | |
5501 | do | |
5502 | { | |
5503 | expr_vliw = find_best_expr (&av_vliw, bnds, fence, &need_stall); | |
5504 | if (!expr_vliw && need_stall) | |
5505 | { | |
5506 | /* All expressions required a stall. Do not recompute av sets | |
5507 | as we'll get the same answer (modulo the insns between | |
5508 | the fence and its boundary, which will not be available for | |
5509 | pipelining). */ | |
5510 | gcc_assert (! expr_vliw && stall_iterations < 2); | |
5511 | was_stall++; | |
5512 | /* If we are going to stall for too long, break to recompute av | |
5513 | sets and bring more insns for pipelining. */ | |
5514 | if (need_stall <= 3) | |
5515 | stall_for_cycles (fence, need_stall); | |
5516 | else | |
5517 | { | |
5518 | stall_for_cycles (fence, 1); | |
5519 | break; | |
5520 | } | |
5521 | } | |
5522 | } | |
5523 | while (! expr_vliw && need_stall); | |
5524 | ||
5525 | /* Now either we've selected expr_vliw or we have nothing to schedule. */ | |
5526 | if (!expr_vliw) | |
5527 | { | |
5528 | av_set_clear (&av_vliw); | |
5529 | break; | |
5530 | } | |
5531 | ||
5532 | bndsp = &bnds; | |
5533 | bnds_tailp1 = bnds_tailp; | |
5534 | ||
5535 | do | |
5536 | /* This code will be executed only once until we'd have several | |
5537 | boundaries per fence. */ | |
5538 | { | |
5539 | bnd_t bnd = BLIST_BND (*bndsp); | |
5540 | ||
5541 | if (!av_set_is_in_p (BND_AV1 (bnd), EXPR_VINSN (expr_vliw))) | |
5542 | { | |
5543 | bndsp = &BLIST_NEXT (*bndsp); | |
5544 | continue; | |
5545 | } | |
5546 | ||
5547 | insn = schedule_expr_on_boundary (bnd, expr_vliw, seqno); | |
b5b8b0ac AO |
5548 | last_insn_was_debug = DEBUG_INSN_P (insn); |
5549 | if (last_insn_was_debug) | |
5550 | was_debug_bb_end_p = (insn == BND_TO (bnd) && sel_bb_end_p (insn)); | |
e855c69d | 5551 | update_fence_and_insn (fence, insn, need_stall); |
b5b8b0ac | 5552 | bnds_tailp = update_boundaries (fence, bnd, insn, bndsp, bnds_tailp); |
e855c69d AB |
5553 | |
5554 | /* Add insn to the list of scheduled on this cycle instructions. */ | |
5555 | ilist_add (*scheduled_insns_tailpp, insn); | |
5556 | *scheduled_insns_tailpp = &ILIST_NEXT (**scheduled_insns_tailpp); | |
5557 | } | |
5558 | while (*bndsp != *bnds_tailp1); | |
5559 | ||
5560 | av_set_clear (&av_vliw); | |
b5b8b0ac AO |
5561 | if (!last_insn_was_debug) |
5562 | scheduled_insns++; | |
e855c69d AB |
5563 | |
5564 | /* We currently support information about candidate blocks only for | |
5565 | one 'target_bb' block. Hence we can't schedule after jump insn, | |
5566 | as this will bring two boundaries and, hence, necessity to handle | |
5567 | information for two or more blocks concurrently. */ | |
b5b8b0ac | 5568 | if ((last_insn_was_debug ? was_debug_bb_end_p : sel_bb_end_p (insn)) |
e855c69d AB |
5569 | || (was_stall |
5570 | && (was_stall >= max_stall | |
5571 | || scheduled_insns >= max_insns))) | |
5572 | break; | |
5573 | } | |
5574 | while (bnds); | |
5575 | ||
5576 | gcc_assert (!FENCE_BNDS (fence)); | |
5577 | ||
5578 | /* Update boundaries of the FENCE. */ | |
5579 | while (bnds) | |
5580 | { | |
5581 | ilist_t ptr = BND_PTR (BLIST_BND (bnds)); | |
5582 | ||
5583 | if (ptr) | |
5584 | { | |
5585 | insn = ILIST_INSN (ptr); | |
5586 | ||
5587 | if (!ilist_is_in_p (FENCE_BNDS (fence), insn)) | |
5588 | ilist_add (&FENCE_BNDS (fence), insn); | |
5589 | } | |
5590 | ||
5591 | blist_remove (&bnds); | |
5592 | } | |
5593 | ||
5594 | /* Update target context on the fence. */ | |
5595 | reset_target_context (FENCE_TC (fence), false); | |
5596 | } | |
5597 | ||
5598 | /* All exprs in ORIG_OPS must have the same destination register or memory. | |
5599 | Return that destination. */ | |
5600 | static rtx | |
5601 | get_dest_from_orig_ops (av_set_t orig_ops) | |
5602 | { | |
5603 | rtx dest = NULL_RTX; | |
5604 | av_set_iterator av_it; | |
5605 | expr_t expr; | |
5606 | bool first_p = true; | |
5607 | ||
5608 | FOR_EACH_EXPR (expr, av_it, orig_ops) | |
5609 | { | |
5610 | rtx x = EXPR_LHS (expr); | |
5611 | ||
5612 | if (first_p) | |
5613 | { | |
5614 | first_p = false; | |
5615 | dest = x; | |
5616 | } | |
5617 | else | |
5618 | gcc_assert (dest == x | |
5619 | || (dest != NULL_RTX && x != NULL_RTX | |
5620 | && rtx_equal_p (dest, x))); | |
5621 | } | |
5622 | ||
5623 | return dest; | |
5624 | } | |
5625 | ||
5626 | /* Update data sets for the bookkeeping block and record those expressions | |
5627 | which become no longer available after inserting this bookkeeping. */ | |
5628 | static void | |
5629 | update_and_record_unavailable_insns (basic_block book_block) | |
5630 | { | |
5631 | av_set_iterator i; | |
5632 | av_set_t old_av_set = NULL; | |
5633 | expr_t cur_expr; | |
5634 | rtx bb_end = sel_bb_end (book_block); | |
5635 | ||
5636 | /* First, get correct liveness in the bookkeeping block. The problem is | |
5637 | the range between the bookeeping insn and the end of block. */ | |
5638 | update_liveness_on_insn (bb_end); | |
5639 | if (control_flow_insn_p (bb_end)) | |
5640 | update_liveness_on_insn (PREV_INSN (bb_end)); | |
5641 | ||
5642 | /* If there's valid av_set on BOOK_BLOCK, then there might exist another | |
5643 | fence above, where we may choose to schedule an insn which is | |
5644 | actually blocked from moving up with the bookkeeping we create here. */ | |
5645 | if (AV_SET_VALID_P (sel_bb_head (book_block))) | |
5646 | { | |
5647 | old_av_set = av_set_copy (BB_AV_SET (book_block)); | |
5648 | update_data_sets (sel_bb_head (book_block)); | |
5649 | ||
5650 | /* Traverse all the expressions in the old av_set and check whether | |
5651 | CUR_EXPR is in new AV_SET. */ | |
5652 | FOR_EACH_EXPR (cur_expr, i, old_av_set) | |
5653 | { | |
5654 | expr_t new_expr = av_set_lookup (BB_AV_SET (book_block), | |
5655 | EXPR_VINSN (cur_expr)); | |
5656 | ||
5657 | if (! new_expr | |
5658 | /* In this case, we can just turn off the E_T_A bit, but we can't | |
5659 | represent this information with the current vector. */ | |
5660 | || EXPR_TARGET_AVAILABLE (new_expr) | |
5661 | != EXPR_TARGET_AVAILABLE (cur_expr)) | |
5662 | /* Unfortunately, the below code could be also fired up on | |
5663 | separable insns. | |
5664 | FIXME: add an example of how this could happen. */ | |
5665 | vinsn_vec_add (&vec_bookkeeping_blocked_vinsns, cur_expr); | |
5666 | } | |
5667 | ||
5668 | av_set_clear (&old_av_set); | |
5669 | } | |
5670 | } | |
5671 | ||
5672 | /* The main effect of this function is that sparams->c_expr is merged | |
5673 | with (or copied to) lparams->c_expr_merged. If there's only one successor, | |
5674 | we avoid merging anything by copying sparams->c_expr to lparams->c_expr_merged. | |
5675 | lparams->c_expr_merged is copied back to sparams->c_expr after all | |
5676 | successors has been traversed. lparams->c_expr_local is an expr allocated | |
5677 | on stack in the caller function, and is used if there is more than one | |
5678 | successor. | |
5679 | ||
5680 | SUCC is one of the SUCCS_NORMAL successors of INSN, | |
5681 | MOVEOP_DRV_CALL_RES is the result of call code_motion_path_driver on succ, | |
5682 | LPARAMS and STATIC_PARAMS contain the parameters described above. */ | |
5683 | static void | |
5684 | move_op_merge_succs (insn_t insn ATTRIBUTE_UNUSED, | |
5685 | insn_t succ ATTRIBUTE_UNUSED, | |
5686 | int moveop_drv_call_res, | |
5687 | cmpd_local_params_p lparams, void *static_params) | |
5688 | { | |
5689 | moveop_static_params_p sparams = (moveop_static_params_p) static_params; | |
5690 | ||
5691 | /* Nothing to do, if original expr wasn't found below. */ | |
5692 | if (moveop_drv_call_res != 1) | |
5693 | return; | |
5694 | ||
5695 | /* If this is a first successor. */ | |
5696 | if (!lparams->c_expr_merged) | |
5697 | { | |
5698 | lparams->c_expr_merged = sparams->c_expr; | |
5699 | sparams->c_expr = lparams->c_expr_local; | |
5700 | } | |
5701 | else | |
5702 | { | |
5703 | /* We must merge all found expressions to get reasonable | |
5704 | EXPR_SPEC_DONE_DS for the resulting insn. If we don't | |
5705 | do so then we can first find the expr with epsilon | |
5706 | speculation success probability and only then with the | |
5707 | good probability. As a result the insn will get epsilon | |
5708 | probability and will never be scheduled because of | |
5709 | weakness_cutoff in find_best_expr. | |
5710 | ||
5711 | We call merge_expr_data here instead of merge_expr | |
5712 | because due to speculation C_EXPR and X may have the | |
5713 | same insns with different speculation types. And as of | |
5714 | now such insns are considered non-equal. | |
5715 | ||
5716 | However, EXPR_SCHED_TIMES is different -- we must get | |
5717 | SCHED_TIMES from a real insn, not a bookkeeping copy. | |
5718 | We force this here. Instead, we may consider merging | |
5719 | SCHED_TIMES to the maximum instead of minimum in the | |
5720 | below function. */ | |
5721 | int old_times = EXPR_SCHED_TIMES (lparams->c_expr_merged); | |
5722 | ||
5723 | merge_expr_data (lparams->c_expr_merged, sparams->c_expr, NULL); | |
5724 | if (EXPR_SCHED_TIMES (sparams->c_expr) == 0) | |
5725 | EXPR_SCHED_TIMES (lparams->c_expr_merged) = old_times; | |
5726 | ||
5727 | clear_expr (sparams->c_expr); | |
5728 | } | |
5729 | } | |
5730 | ||
5731 | /* Add used regs for the successor SUCC into SPARAMS->USED_REGS. | |
5732 | ||
5733 | SUCC is one of the SUCCS_NORMAL successors of INSN, | |
5734 | MOVEOP_DRV_CALL_RES is the result of call code_motion_path_driver on succ or 0, | |
5735 | if SUCC is one of SUCCS_BACK or SUCCS_OUT. | |
5736 | STATIC_PARAMS contain USED_REGS set. */ | |
5737 | static void | |
5738 | fur_merge_succs (insn_t insn ATTRIBUTE_UNUSED, insn_t succ, | |
5739 | int moveop_drv_call_res, | |
5740 | cmpd_local_params_p lparams ATTRIBUTE_UNUSED, | |
5741 | void *static_params) | |
5742 | { | |
5743 | regset succ_live; | |
5744 | fur_static_params_p sparams = (fur_static_params_p) static_params; | |
5745 | ||
5746 | /* Here we compute live regsets only for branches that do not lie | |
5747 | on the code motion paths. These branches correspond to value | |
5748 | MOVEOP_DRV_CALL_RES==0 and include SUCCS_BACK and SUCCS_OUT, though | |
5749 | for such branches code_motion_path_driver is not called. */ | |
5750 | if (moveop_drv_call_res != 0) | |
5751 | return; | |
5752 | ||
5753 | /* Mark all registers that do not meet the following condition: | |
5754 | (3) not live on the other path of any conditional branch | |
5755 | that is passed by the operation, in case original | |
5756 | operations are not present on both paths of the | |
5757 | conditional branch. */ | |
5758 | succ_live = compute_live (succ); | |
5759 | IOR_REG_SET (sparams->used_regs, succ_live); | |
5760 | } | |
5761 | ||
5762 | /* This function is called after the last successor. Copies LP->C_EXPR_MERGED | |
5763 | into SP->CEXPR. */ | |
5764 | static void | |
5765 | move_op_after_merge_succs (cmpd_local_params_p lp, void *sparams) | |
5766 | { | |
5767 | moveop_static_params_p sp = (moveop_static_params_p) sparams; | |
5768 | ||
5769 | sp->c_expr = lp->c_expr_merged; | |
5770 | } | |
5771 | ||
5772 | /* Track bookkeeping copies created, insns scheduled, and blocks for | |
5773 | rescheduling when INSN is found by move_op. */ | |
5774 | static void | |
5775 | track_scheduled_insns_and_blocks (rtx insn) | |
5776 | { | |
5777 | /* Even if this insn can be a copy that will be removed during current move_op, | |
5778 | we still need to count it as an originator. */ | |
5779 | bitmap_set_bit (current_originators, INSN_UID (insn)); | |
5780 | ||
5781 | if (!bitmap_bit_p (current_copies, INSN_UID (insn))) | |
5782 | { | |
5783 | /* Note that original block needs to be rescheduled, as we pulled an | |
5784 | instruction out of it. */ | |
5785 | if (INSN_SCHED_TIMES (insn) > 0) | |
5786 | bitmap_set_bit (blocks_to_reschedule, BLOCK_FOR_INSN (insn)->index); | |
b5b8b0ac | 5787 | else if (INSN_UID (insn) < first_emitted_uid && !DEBUG_INSN_P (insn)) |
e855c69d AB |
5788 | num_insns_scheduled++; |
5789 | } | |
5790 | else | |
5791 | bitmap_clear_bit (current_copies, INSN_UID (insn)); | |
5792 | ||
5793 | /* For instructions we must immediately remove insn from the | |
5794 | stream, so subsequent update_data_sets () won't include this | |
5795 | insn into av_set. | |
5796 | For expr we must make insn look like "INSN_REG (insn) := c_expr". */ | |
5797 | if (INSN_UID (insn) > max_uid_before_move_op) | |
5798 | stat_bookkeeping_copies--; | |
5799 | } | |
5800 | ||
5801 | /* Emit a register-register copy for INSN if needed. Return true if | |
5802 | emitted one. PARAMS is the move_op static parameters. */ | |
5803 | static bool | |
5804 | maybe_emit_renaming_copy (rtx insn, | |
5805 | moveop_static_params_p params) | |
5806 | { | |
5807 | bool insn_emitted = false; | |
5808 | rtx cur_reg = expr_dest_reg (params->c_expr); | |
5809 | ||
5810 | gcc_assert (!cur_reg || (params->dest && REG_P (params->dest))); | |
5811 | ||
5812 | /* If original operation has expr and the register chosen for | |
5813 | that expr is not original operation's dest reg, substitute | |
5814 | operation's right hand side with the register chosen. */ | |
5815 | if (cur_reg != NULL_RTX && REGNO (params->dest) != REGNO (cur_reg)) | |
5816 | { | |
5817 | insn_t reg_move_insn, reg_move_insn_rtx; | |
5818 | ||
5819 | reg_move_insn_rtx = create_insn_rtx_with_rhs (INSN_VINSN (insn), | |
5820 | params->dest); | |
5821 | reg_move_insn = sel_gen_insn_from_rtx_after (reg_move_insn_rtx, | |
5822 | INSN_EXPR (insn), | |
5823 | INSN_SEQNO (insn), | |
5824 | insn); | |
5825 | EXPR_SPEC_DONE_DS (INSN_EXPR (reg_move_insn)) = 0; | |
5826 | replace_dest_with_reg_in_expr (params->c_expr, params->dest); | |
5827 | ||
5828 | insn_emitted = true; | |
5829 | params->was_renamed = true; | |
5830 | } | |
5831 | ||
5832 | return insn_emitted; | |
5833 | } | |
5834 | ||
5835 | /* Emit a speculative check for INSN speculated as EXPR if needed. | |
5836 | Return true if we've emitted one. PARAMS is the move_op static | |
5837 | parameters. */ | |
5838 | static bool | |
5839 | maybe_emit_speculative_check (rtx insn, expr_t expr, | |
5840 | moveop_static_params_p params) | |
5841 | { | |
5842 | bool insn_emitted = false; | |
5843 | insn_t x; | |
5844 | ds_t check_ds; | |
5845 | ||
5846 | check_ds = get_spec_check_type_for_insn (insn, expr); | |
5847 | if (check_ds != 0) | |
5848 | { | |
5849 | /* A speculation check should be inserted. */ | |
5850 | x = create_speculation_check (params->c_expr, check_ds, insn); | |
5851 | insn_emitted = true; | |
5852 | } | |
5853 | else | |
5854 | { | |
5855 | EXPR_SPEC_DONE_DS (INSN_EXPR (insn)) = 0; | |
5856 | x = insn; | |
5857 | } | |
5858 | ||
5859 | gcc_assert (EXPR_SPEC_DONE_DS (INSN_EXPR (x)) == 0 | |
5860 | && EXPR_SPEC_TO_CHECK_DS (INSN_EXPR (x)) == 0); | |
5861 | return insn_emitted; | |
5862 | } | |
5863 | ||
5864 | /* Handle transformations that leave an insn in place of original | |
5865 | insn such as renaming/speculation. Return true if one of such | |
5866 | transformations actually happened, and we have emitted this insn. */ | |
5867 | static bool | |
5868 | handle_emitting_transformations (rtx insn, expr_t expr, | |
5869 | moveop_static_params_p params) | |
5870 | { | |
5871 | bool insn_emitted = false; | |
5872 | ||
5873 | insn_emitted = maybe_emit_renaming_copy (insn, params); | |
5874 | insn_emitted |= maybe_emit_speculative_check (insn, expr, params); | |
5875 | ||
5876 | return insn_emitted; | |
5877 | } | |
5878 | ||
b5b8b0ac AO |
5879 | /* If INSN is the only insn in the basic block (not counting JUMP, |
5880 | which may be a jump to next insn, and DEBUG_INSNs), we want to | |
5881 | leave a NOP there till the return to fill_insns. */ | |
5882 | ||
5883 | static bool | |
5884 | need_nop_to_preserve_insn_bb (rtx insn) | |
e855c69d | 5885 | { |
b5b8b0ac | 5886 | insn_t bb_head, bb_end, bb_next, in_next; |
e855c69d AB |
5887 | basic_block bb = BLOCK_FOR_INSN (insn); |
5888 | ||
e855c69d AB |
5889 | bb_head = sel_bb_head (bb); |
5890 | bb_end = sel_bb_end (bb); | |
e855c69d | 5891 | |
b5b8b0ac AO |
5892 | if (bb_head == bb_end) |
5893 | return true; | |
5894 | ||
5895 | while (bb_head != bb_end && DEBUG_INSN_P (bb_head)) | |
5896 | bb_head = NEXT_INSN (bb_head); | |
5897 | ||
5898 | if (bb_head == bb_end) | |
5899 | return true; | |
5900 | ||
5901 | while (bb_head != bb_end && DEBUG_INSN_P (bb_end)) | |
5902 | bb_end = PREV_INSN (bb_end); | |
5903 | ||
5904 | if (bb_head == bb_end) | |
5905 | return true; | |
5906 | ||
5907 | bb_next = NEXT_INSN (bb_head); | |
5908 | while (bb_next != bb_end && DEBUG_INSN_P (bb_next)) | |
5909 | bb_next = NEXT_INSN (bb_next); | |
5910 | ||
5911 | if (bb_next == bb_end && JUMP_P (bb_end)) | |
5912 | return true; | |
5913 | ||
5914 | in_next = NEXT_INSN (insn); | |
5915 | while (DEBUG_INSN_P (in_next)) | |
5916 | in_next = NEXT_INSN (in_next); | |
5917 | ||
5918 | if (IN_CURRENT_FENCE_P (in_next)) | |
5919 | return true; | |
5920 | ||
5921 | return false; | |
5922 | } | |
5923 | ||
5924 | /* Remove INSN from stream. When ONLY_DISCONNECT is true, its data | |
5925 | is not removed but reused when INSN is re-emitted. */ | |
5926 | static void | |
5927 | remove_insn_from_stream (rtx insn, bool only_disconnect) | |
5928 | { | |
e855c69d AB |
5929 | /* If there's only one insn in the BB, make sure that a nop is |
5930 | inserted into it, so the basic block won't disappear when we'll | |
5931 | delete INSN below with sel_remove_insn. It should also survive | |
5932 | till the return to fill_insns. */ | |
b5b8b0ac | 5933 | if (need_nop_to_preserve_insn_bb (insn)) |
e855c69d | 5934 | { |
b5b8b0ac | 5935 | insn_t nop = get_nop_from_pool (insn); |
e855c69d AB |
5936 | gcc_assert (INSN_NOP_P (nop)); |
5937 | VEC_safe_push (insn_t, heap, vec_temp_moveop_nops, nop); | |
5938 | } | |
5939 | ||
5940 | sel_remove_insn (insn, only_disconnect, false); | |
5941 | } | |
5942 | ||
5943 | /* This function is called when original expr is found. | |
5944 | INSN - current insn traversed, EXPR - the corresponding expr found. | |
5945 | LPARAMS is the local parameters of code modion driver, STATIC_PARAMS | |
5946 | is static parameters of move_op. */ | |
5947 | static void | |
5948 | move_op_orig_expr_found (insn_t insn, expr_t expr, | |
5949 | cmpd_local_params_p lparams ATTRIBUTE_UNUSED, | |
5950 | void *static_params) | |
5951 | { | |
5952 | bool only_disconnect, insn_emitted; | |
5953 | moveop_static_params_p params = (moveop_static_params_p) static_params; | |
5954 | ||
5955 | copy_expr_onside (params->c_expr, INSN_EXPR (insn)); | |
5956 | track_scheduled_insns_and_blocks (insn); | |
5957 | insn_emitted = handle_emitting_transformations (insn, expr, params); | |
5958 | only_disconnect = (params->uid == INSN_UID (insn) | |
5959 | && ! insn_emitted && ! EXPR_WAS_CHANGED (expr)); | |
72a54528 AM |
5960 | |
5961 | /* Mark that we've disconnected an insn. */ | |
5962 | if (only_disconnect) | |
5963 | params->uid = -1; | |
e855c69d AB |
5964 | remove_insn_from_stream (insn, only_disconnect); |
5965 | } | |
5966 | ||
5967 | /* The function is called when original expr is found. | |
5968 | INSN - current insn traversed, EXPR - the corresponding expr found, | |
5969 | crosses_call and original_insns in STATIC_PARAMS are updated. */ | |
5970 | static void | |
5971 | fur_orig_expr_found (insn_t insn, expr_t expr ATTRIBUTE_UNUSED, | |
5972 | cmpd_local_params_p lparams ATTRIBUTE_UNUSED, | |
5973 | void *static_params) | |
5974 | { | |
5975 | fur_static_params_p params = (fur_static_params_p) static_params; | |
5976 | regset tmp; | |
5977 | ||
5978 | if (CALL_P (insn)) | |
5979 | params->crosses_call = true; | |
5980 | ||
5981 | def_list_add (params->original_insns, insn, params->crosses_call); | |
5982 | ||
5983 | /* Mark the registers that do not meet the following condition: | |
5984 | (2) not among the live registers of the point | |
5985 | immediately following the first original operation on | |
5986 | a given downward path, except for the original target | |
5987 | register of the operation. */ | |
5988 | tmp = get_clear_regset_from_pool (); | |
5989 | compute_live_below_insn (insn, tmp); | |
5990 | AND_COMPL_REG_SET (tmp, INSN_REG_SETS (insn)); | |
5991 | AND_COMPL_REG_SET (tmp, INSN_REG_CLOBBERS (insn)); | |
5992 | IOR_REG_SET (params->used_regs, tmp); | |
5993 | return_regset_to_pool (tmp); | |
5994 | ||
5995 | /* (*1) We need to add to USED_REGS registers that are read by | |
5996 | INSN's lhs. This may lead to choosing wrong src register. | |
5997 | E.g. (scheduling const expr enabled): | |
5998 | ||
5999 | 429: ax=0x0 <- Can't use AX for this expr (0x0) | |
6000 | 433: dx=[bp-0x18] | |
6001 | 427: [ax+dx+0x1]=ax | |
6002 | REG_DEAD: ax | |
6003 | 168: di=dx | |
6004 | REG_DEAD: dx | |
6005 | */ | |
6006 | /* FIXME: see comment above and enable MEM_P | |
6007 | in vinsn_separable_p. */ | |
6008 | gcc_assert (!VINSN_SEPARABLE_P (INSN_VINSN (insn)) | |
6009 | || !MEM_P (INSN_LHS (insn))); | |
6010 | } | |
6011 | ||
6012 | /* This function is called on the ascending pass, before returning from | |
6013 | current basic block. */ | |
6014 | static void | |
6015 | move_op_at_first_insn (insn_t insn, cmpd_local_params_p lparams, | |
6016 | void *static_params) | |
6017 | { | |
6018 | moveop_static_params_p sparams = (moveop_static_params_p) static_params; | |
6019 | basic_block book_block = NULL; | |
6020 | ||
6021 | /* When we have removed the boundary insn for scheduling, which also | |
6022 | happened to be the end insn in its bb, we don't need to update sets. */ | |
6023 | if (!lparams->removed_last_insn | |
6024 | && lparams->e1 | |
6025 | && sel_bb_head_p (insn)) | |
6026 | { | |
6027 | /* We should generate bookkeeping code only if we are not at the | |
6028 | top level of the move_op. */ | |
6029 | if (sel_num_cfg_preds_gt_1 (insn)) | |
6030 | book_block = generate_bookkeeping_insn (sparams->c_expr, | |
6031 | lparams->e1, lparams->e2); | |
6032 | /* Update data sets for the current insn. */ | |
6033 | update_data_sets (insn); | |
6034 | } | |
6035 | ||
6036 | /* If bookkeeping code was inserted, we need to update av sets of basic | |
6037 | block that received bookkeeping. After generation of bookkeeping insn, | |
6038 | bookkeeping block does not contain valid av set because we are not following | |
6039 | the original algorithm in every detail with regards to e.g. renaming | |
6040 | simple reg-reg copies. Consider example: | |
6041 | ||
6042 | bookkeeping block scheduling fence | |
6043 | \ / | |
6044 | \ join / | |
6045 | ---------- | |
6046 | | | | |
6047 | ---------- | |
6048 | / \ | |
6049 | / \ | |
6050 | r1 := r2 r1 := r3 | |
6051 | ||
6052 | We try to schedule insn "r1 := r3" on the current | |
6053 | scheduling fence. Also, note that av set of bookkeeping block | |
6054 | contain both insns "r1 := r2" and "r1 := r3". When the insn has | |
6055 | been scheduled, the CFG is as follows: | |
6056 | ||
6057 | r1 := r3 r1 := r3 | |
6058 | bookkeeping block scheduling fence | |
6059 | \ / | |
6060 | \ join / | |
6061 | ---------- | |
6062 | | | | |
6063 | ---------- | |
6064 | / \ | |
6065 | / \ | |
6066 | r1 := r2 | |
6067 | ||
6068 | Here, insn "r1 := r3" was scheduled at the current scheduling point | |
6069 | and bookkeeping code was generated at the bookeeping block. This | |
6070 | way insn "r1 := r2" is no longer available as a whole instruction | |
6071 | (but only as expr) ahead of insn "r1 := r3" in bookkeeping block. | |
6072 | This situation is handled by calling update_data_sets. | |
6073 | ||
6074 | Since update_data_sets is called only on the bookkeeping block, and | |
6075 | it also may have predecessors with av_sets, containing instructions that | |
6076 | are no longer available, we save all such expressions that become | |
6077 | unavailable during data sets update on the bookkeeping block in | |
6078 | VEC_BOOKKEEPING_BLOCKED_VINSNS. Later we avoid selecting such | |
6079 | expressions for scheduling. This allows us to avoid recomputation of | |
6080 | av_sets outside the code motion path. */ | |
6081 | ||
6082 | if (book_block) | |
6083 | update_and_record_unavailable_insns (book_block); | |
6084 | ||
6085 | /* If INSN was previously marked for deletion, it's time to do it. */ | |
6086 | if (lparams->removed_last_insn) | |
6087 | insn = PREV_INSN (insn); | |
6088 | ||
6089 | /* Do not tidy control flow at the topmost moveop, as we can erroneously | |
6090 | kill a block with a single nop in which the insn should be emitted. */ | |
6091 | if (lparams->e1) | |
6092 | tidy_control_flow (BLOCK_FOR_INSN (insn), true); | |
6093 | } | |
6094 | ||
6095 | /* This function is called on the ascending pass, before returning from the | |
6096 | current basic block. */ | |
6097 | static void | |
6098 | fur_at_first_insn (insn_t insn, | |
6099 | cmpd_local_params_p lparams ATTRIBUTE_UNUSED, | |
6100 | void *static_params ATTRIBUTE_UNUSED) | |
6101 | { | |
6102 | gcc_assert (!sel_bb_head_p (insn) || AV_SET_VALID_P (insn) | |
6103 | || AV_LEVEL (insn) == -1); | |
6104 | } | |
6105 | ||
6106 | /* Called on the backward stage of recursion to call moveup_expr for insn | |
6107 | and sparams->c_expr. */ | |
6108 | static void | |
6109 | move_op_ascend (insn_t insn, void *static_params) | |
6110 | { | |
6111 | enum MOVEUP_EXPR_CODE res; | |
6112 | moveop_static_params_p sparams = (moveop_static_params_p) static_params; | |
6113 | ||
6114 | if (! INSN_NOP_P (insn)) | |
6115 | { | |
6116 | res = moveup_expr_cached (sparams->c_expr, insn, false); | |
6117 | gcc_assert (res != MOVEUP_EXPR_NULL); | |
6118 | } | |
6119 | ||
6120 | /* Update liveness for this insn as it was invalidated. */ | |
6121 | update_liveness_on_insn (insn); | |
6122 | } | |
6123 | ||
6124 | /* This function is called on enter to the basic block. | |
6125 | Returns TRUE if this block already have been visited and | |
6126 | code_motion_path_driver should return 1, FALSE otherwise. */ | |
6127 | static int | |
6128 | fur_on_enter (insn_t insn ATTRIBUTE_UNUSED, cmpd_local_params_p local_params, | |
6129 | void *static_params, bool visited_p) | |
6130 | { | |
6131 | fur_static_params_p sparams = (fur_static_params_p) static_params; | |
6132 | ||
6133 | if (visited_p) | |
6134 | { | |
6135 | /* If we have found something below this block, there should be at | |
6136 | least one insn in ORIGINAL_INSNS. */ | |
6137 | gcc_assert (*sparams->original_insns); | |
6138 | ||
6139 | /* Adjust CROSSES_CALL, since we may have come to this block along | |
6140 | different path. */ | |
6141 | DEF_LIST_DEF (*sparams->original_insns)->crosses_call | |
6142 | |= sparams->crosses_call; | |
6143 | } | |
6144 | else | |
6145 | local_params->old_original_insns = *sparams->original_insns; | |
6146 | ||
6147 | return 1; | |
6148 | } | |
6149 | ||
6150 | /* Same as above but for move_op. */ | |
6151 | static int | |
6152 | move_op_on_enter (insn_t insn ATTRIBUTE_UNUSED, | |
6153 | cmpd_local_params_p local_params ATTRIBUTE_UNUSED, | |
6154 | void *static_params ATTRIBUTE_UNUSED, bool visited_p) | |
6155 | { | |
6156 | if (visited_p) | |
6157 | return -1; | |
6158 | return 1; | |
6159 | } | |
6160 | ||
6161 | /* This function is called while descending current basic block if current | |
6162 | insn is not the original EXPR we're searching for. | |
6163 | ||
6164 | Return value: FALSE, if code_motion_path_driver should perform a local | |
6165 | cleanup and return 0 itself; | |
6166 | TRUE, if code_motion_path_driver should continue. */ | |
6167 | static bool | |
6168 | move_op_orig_expr_not_found (insn_t insn, av_set_t orig_ops ATTRIBUTE_UNUSED, | |
6169 | void *static_params) | |
6170 | { | |
6171 | moveop_static_params_p sparams = (moveop_static_params_p) static_params; | |
6172 | ||
6173 | #ifdef ENABLE_CHECKING | |
6174 | sparams->failed_insn = insn; | |
6175 | #endif | |
6176 | ||
6177 | /* If we're scheduling separate expr, in order to generate correct code | |
6178 | we need to stop the search at bookkeeping code generated with the | |
6179 | same destination register or memory. */ | |
6180 | if (lhs_of_insn_equals_to_dest_p (insn, sparams->dest)) | |
6181 | return false; | |
6182 | return true; | |
6183 | } | |
6184 | ||
6185 | /* This function is called while descending current basic block if current | |
6186 | insn is not the original EXPR we're searching for. | |
6187 | ||
6188 | Return value: TRUE (code_motion_path_driver should continue). */ | |
6189 | static bool | |
6190 | fur_orig_expr_not_found (insn_t insn, av_set_t orig_ops, void *static_params) | |
6191 | { | |
6192 | bool mutexed; | |
6193 | expr_t r; | |
6194 | av_set_iterator avi; | |
6195 | fur_static_params_p sparams = (fur_static_params_p) static_params; | |
6196 | ||
6197 | if (CALL_P (insn)) | |
6198 | sparams->crosses_call = true; | |
b5b8b0ac AO |
6199 | else if (DEBUG_INSN_P (insn)) |
6200 | return true; | |
e855c69d AB |
6201 | |
6202 | /* If current insn we are looking at cannot be executed together | |
6203 | with original insn, then we can skip it safely. | |
6204 | ||
6205 | Example: ORIG_OPS = { (p6) r14 = sign_extend (r15); } | |
6206 | INSN = (!p6) r14 = r14 + 1; | |
6207 | ||
6208 | Here we can schedule ORIG_OP with lhs = r14, though only | |
6209 | looking at the set of used and set registers of INSN we must | |
6210 | forbid it. So, add set/used in INSN registers to the | |
6211 | untouchable set only if there is an insn in ORIG_OPS that can | |
6212 | affect INSN. */ | |
6213 | mutexed = true; | |
6214 | FOR_EACH_EXPR (r, avi, orig_ops) | |
6215 | if (!sched_insns_conditions_mutex_p (insn, EXPR_INSN_RTX (r))) | |
6216 | { | |
6217 | mutexed = false; | |
6218 | break; | |
6219 | } | |
6220 | ||
6221 | /* Mark all registers that do not meet the following condition: | |
6222 | (1) Not set or read on any path from xi to an instance of the | |
6223 | original operation. */ | |
6224 | if (!mutexed) | |
6225 | { | |
6226 | IOR_REG_SET (sparams->used_regs, INSN_REG_SETS (insn)); | |
6227 | IOR_REG_SET (sparams->used_regs, INSN_REG_USES (insn)); | |
6228 | IOR_REG_SET (sparams->used_regs, INSN_REG_CLOBBERS (insn)); | |
6229 | } | |
6230 | ||
6231 | return true; | |
6232 | } | |
6233 | ||
6234 | /* Hooks and data to perform move_op operations with code_motion_path_driver. */ | |
6235 | struct code_motion_path_driver_info_def move_op_hooks = { | |
6236 | move_op_on_enter, | |
6237 | move_op_orig_expr_found, | |
6238 | move_op_orig_expr_not_found, | |
6239 | move_op_merge_succs, | |
6240 | move_op_after_merge_succs, | |
6241 | move_op_ascend, | |
6242 | move_op_at_first_insn, | |
6243 | SUCCS_NORMAL, | |
6244 | "move_op" | |
6245 | }; | |
6246 | ||
6247 | /* Hooks and data to perform find_used_regs operations | |
6248 | with code_motion_path_driver. */ | |
6249 | struct code_motion_path_driver_info_def fur_hooks = { | |
6250 | fur_on_enter, | |
6251 | fur_orig_expr_found, | |
6252 | fur_orig_expr_not_found, | |
6253 | fur_merge_succs, | |
6254 | NULL, /* fur_after_merge_succs */ | |
6255 | NULL, /* fur_ascend */ | |
6256 | fur_at_first_insn, | |
6257 | SUCCS_ALL, | |
6258 | "find_used_regs" | |
6259 | }; | |
6260 | ||
6261 | /* Traverse all successors of INSN. For each successor that is SUCCS_NORMAL | |
6262 | code_motion_path_driver is called recursively. Original operation | |
6263 | was found at least on one path that is starting with one of INSN's | |
6264 | successors (this fact is asserted). ORIG_OPS is expressions we're looking | |
6265 | for, PATH is the path we've traversed, STATIC_PARAMS is the parameters | |
6266 | of either move_op or find_used_regs depending on the caller. | |
6267 | ||
6268 | Return 0 if we haven't found expression, 1 if we found it, -1 if we don't | |
6269 | know for sure at this point. */ | |
6270 | static int | |
6271 | code_motion_process_successors (insn_t insn, av_set_t orig_ops, | |
6272 | ilist_t path, void *static_params) | |
6273 | { | |
6274 | int res = 0; | |
6275 | succ_iterator succ_i; | |
6276 | rtx succ; | |
6277 | basic_block bb; | |
6278 | int old_index; | |
6279 | unsigned old_succs; | |
6280 | ||
6281 | struct cmpd_local_params lparams; | |
6282 | expr_def _x; | |
6283 | ||
6284 | lparams.c_expr_local = &_x; | |
6285 | lparams.c_expr_merged = NULL; | |
6286 | ||
6287 | /* We need to process only NORMAL succs for move_op, and collect live | |
6288 | registers from ALL branches (including those leading out of the | |
6289 | region) for find_used_regs. | |
6290 | ||
6291 | In move_op, there can be a case when insn's bb number has changed | |
6292 | due to created bookkeeping. This happens very rare, as we need to | |
6293 | move expression from the beginning to the end of the same block. | |
6294 | Rescan successors in this case. */ | |
6295 | ||
6296 | rescan: | |
6297 | bb = BLOCK_FOR_INSN (insn); | |
6298 | old_index = bb->index; | |
6299 | old_succs = EDGE_COUNT (bb->succs); | |
6300 | ||
6301 | FOR_EACH_SUCC_1 (succ, succ_i, insn, code_motion_path_driver_info->succ_flags) | |
6302 | { | |
6303 | int b; | |
6304 | ||
6305 | lparams.e1 = succ_i.e1; | |
6306 | lparams.e2 = succ_i.e2; | |
6307 | ||
6308 | /* Go deep into recursion only for NORMAL edges (non-backedges within the | |
6309 | current region). */ | |
6310 | if (succ_i.current_flags == SUCCS_NORMAL) | |
6311 | b = code_motion_path_driver (succ, orig_ops, path, &lparams, | |
6312 | static_params); | |
6313 | else | |
6314 | b = 0; | |
6315 | ||
6316 | /* Merge c_expres found or unify live register sets from different | |
6317 | successors. */ | |
6318 | code_motion_path_driver_info->merge_succs (insn, succ, b, &lparams, | |
6319 | static_params); | |
6320 | if (b == 1) | |
6321 | res = b; | |
6322 | else if (b == -1 && res != 1) | |
6323 | res = b; | |
6324 | ||
6325 | /* We have simplified the control flow below this point. In this case, | |
6326 | the iterator becomes invalid. We need to try again. */ | |
6327 | if (BLOCK_FOR_INSN (insn)->index != old_index | |
6328 | || EDGE_COUNT (bb->succs) != old_succs) | |
6329 | goto rescan; | |
6330 | } | |
6331 | ||
6332 | #ifdef ENABLE_CHECKING | |
6333 | /* Here, RES==1 if original expr was found at least for one of the | |
6334 | successors. After the loop, RES may happen to have zero value | |
6335 | only if at some point the expr searched is present in av_set, but is | |
6336 | not found below. In most cases, this situation is an error. | |
6337 | The exception is when the original operation is blocked by | |
6338 | bookkeeping generated for another fence or for another path in current | |
6339 | move_op. */ | |
6340 | gcc_assert (res == 1 | |
6341 | || (res == 0 | |
6342 | && av_set_could_be_blocked_by_bookkeeping_p (orig_ops, | |
6343 | static_params)) | |
6344 | || res == -1); | |
6345 | #endif | |
6346 | ||
6347 | /* Merge data, clean up, etc. */ | |
72a54528 | 6348 | if (res != -1 && code_motion_path_driver_info->after_merge_succs) |
e855c69d AB |
6349 | code_motion_path_driver_info->after_merge_succs (&lparams, static_params); |
6350 | ||
6351 | return res; | |
6352 | } | |
6353 | ||
6354 | ||
6355 | /* Perform a cleanup when the driver is about to terminate. ORIG_OPS_P | |
6356 | is the pointer to the av set with expressions we were looking for, | |
6357 | PATH_P is the pointer to the traversed path. */ | |
6358 | static inline void | |
6359 | code_motion_path_driver_cleanup (av_set_t *orig_ops_p, ilist_t *path_p) | |
6360 | { | |
6361 | ilist_remove (path_p); | |
6362 | av_set_clear (orig_ops_p); | |
6363 | } | |
6364 | ||
6365 | /* The driver function that implements move_op or find_used_regs | |
6366 | functionality dependent whether code_motion_path_driver_INFO is set to | |
6367 | &MOVE_OP_HOOKS or &FUR_HOOKS. This function implements the common parts | |
6368 | of code (CFG traversal etc) that are shared among both functions. INSN | |
6369 | is the insn we're starting the search from, ORIG_OPS are the expressions | |
6370 | we're searching for, PATH is traversed path, LOCAL_PARAMS_IN are local | |
6371 | parameters of the driver, and STATIC_PARAMS are static parameters of | |
6372 | the caller. | |
6373 | ||
6374 | Returns whether original instructions were found. Note that top-level | |
6375 | code_motion_path_driver always returns true. */ | |
72a54528 | 6376 | static int |
e855c69d AB |
6377 | code_motion_path_driver (insn_t insn, av_set_t orig_ops, ilist_t path, |
6378 | cmpd_local_params_p local_params_in, | |
6379 | void *static_params) | |
6380 | { | |
6381 | expr_t expr = NULL; | |
6382 | basic_block bb = BLOCK_FOR_INSN (insn); | |
6383 | insn_t first_insn, bb_tail, before_first; | |
6384 | bool removed_last_insn = false; | |
6385 | ||
6386 | if (sched_verbose >= 6) | |
6387 | { | |
6388 | sel_print ("%s (", code_motion_path_driver_info->routine_name); | |
6389 | dump_insn (insn); | |
6390 | sel_print (","); | |
6391 | dump_av_set (orig_ops); | |
6392 | sel_print (")\n"); | |
6393 | } | |
6394 | ||
6395 | gcc_assert (orig_ops); | |
6396 | ||
6397 | /* If no original operations exist below this insn, return immediately. */ | |
6398 | if (is_ineligible_successor (insn, path)) | |
6399 | { | |
6400 | if (sched_verbose >= 6) | |
6401 | sel_print ("Insn %d is ineligible successor\n", INSN_UID (insn)); | |
6402 | return false; | |
6403 | } | |
6404 | ||
6405 | /* The block can have invalid av set, in which case it was created earlier | |
6406 | during move_op. Return immediately. */ | |
6407 | if (sel_bb_head_p (insn)) | |
6408 | { | |
6409 | if (! AV_SET_VALID_P (insn)) | |
6410 | { | |
6411 | if (sched_verbose >= 6) | |
6412 | sel_print ("Returned from block %d as it had invalid av set\n", | |
6413 | bb->index); | |
6414 | return false; | |
6415 | } | |
6416 | ||
6417 | if (bitmap_bit_p (code_motion_visited_blocks, bb->index)) | |
6418 | { | |
6419 | /* We have already found an original operation on this branch, do not | |
6420 | go any further and just return TRUE here. If we don't stop here, | |
6421 | function can have exponential behaviour even on the small code | |
6422 | with many different paths (e.g. with data speculation and | |
6423 | recovery blocks). */ | |
6424 | if (sched_verbose >= 6) | |
6425 | sel_print ("Block %d already visited in this traversal\n", bb->index); | |
6426 | if (code_motion_path_driver_info->on_enter) | |
6427 | return code_motion_path_driver_info->on_enter (insn, | |
6428 | local_params_in, | |
6429 | static_params, | |
6430 | true); | |
6431 | } | |
6432 | } | |
6433 | ||
6434 | if (code_motion_path_driver_info->on_enter) | |
6435 | code_motion_path_driver_info->on_enter (insn, local_params_in, | |
6436 | static_params, false); | |
6437 | orig_ops = av_set_copy (orig_ops); | |
6438 | ||
6439 | /* Filter the orig_ops set. */ | |
6440 | if (AV_SET_VALID_P (insn)) | |
6441 | av_set_intersect (&orig_ops, AV_SET (insn)); | |
6442 | ||
6443 | /* If no more original ops, return immediately. */ | |
6444 | if (!orig_ops) | |
6445 | { | |
6446 | if (sched_verbose >= 6) | |
6447 | sel_print ("No intersection with av set of block %d\n", bb->index); | |
6448 | return false; | |
6449 | } | |
6450 | ||
6451 | /* For non-speculative insns we have to leave only one form of the | |
6452 | original operation, because if we don't, we may end up with | |
6453 | different C_EXPRes and, consequently, with bookkeepings for different | |
6454 | expression forms along the same code motion path. That may lead to | |
6455 | generation of incorrect code. So for each code motion we stick to | |
6456 | the single form of the instruction, except for speculative insns | |
6457 | which we need to keep in different forms with all speculation | |
6458 | types. */ | |
6459 | av_set_leave_one_nonspec (&orig_ops); | |
6460 | ||
6461 | /* It is not possible that all ORIG_OPS are filtered out. */ | |
6462 | gcc_assert (orig_ops); | |
6463 | ||
6464 | /* It is enough to place only heads and tails of visited basic blocks into | |
6465 | the PATH. */ | |
6466 | ilist_add (&path, insn); | |
6467 | first_insn = insn; | |
6468 | bb_tail = sel_bb_end (bb); | |
6469 | ||
6470 | /* Descend the basic block in search of the original expr; this part | |
6471 | corresponds to the part of the original move_op procedure executed | |
6472 | before the recursive call. */ | |
6473 | for (;;) | |
6474 | { | |
6475 | /* Look at the insn and decide if it could be an ancestor of currently | |
6476 | scheduling operation. If it is so, then the insn "dest = op" could | |
6477 | either be replaced with "dest = reg", because REG now holds the result | |
6478 | of OP, or just removed, if we've scheduled the insn as a whole. | |
6479 | ||
6480 | If this insn doesn't contain currently scheduling OP, then proceed | |
6481 | with searching and look at its successors. Operations we're searching | |
6482 | for could have changed when moving up through this insn via | |
6483 | substituting. In this case, perform unsubstitution on them first. | |
6484 | ||
6485 | When traversing the DAG below this insn is finished, insert | |
6486 | bookkeeping code, if the insn is a joint point, and remove | |
6487 | leftovers. */ | |
6488 | ||
6489 | expr = av_set_lookup (orig_ops, INSN_VINSN (insn)); | |
6490 | if (expr) | |
6491 | { | |
6492 | insn_t last_insn = PREV_INSN (insn); | |
6493 | ||
6494 | /* We have found the original operation. */ | |
6495 | if (sched_verbose >= 6) | |
6496 | sel_print ("Found original operation at insn %d\n", INSN_UID (insn)); | |
6497 | ||
6498 | code_motion_path_driver_info->orig_expr_found | |
6499 | (insn, expr, local_params_in, static_params); | |
6500 | ||
6501 | /* Step back, so on the way back we'll start traversing from the | |
6502 | previous insn (or we'll see that it's bb_note and skip that | |
6503 | loop). */ | |
6504 | if (insn == first_insn) | |
6505 | { | |
6506 | first_insn = NEXT_INSN (last_insn); | |
6507 | removed_last_insn = sel_bb_end_p (last_insn); | |
6508 | } | |
6509 | insn = last_insn; | |
6510 | break; | |
6511 | } | |
6512 | else | |
6513 | { | |
6514 | /* We haven't found the original expr, continue descending the basic | |
6515 | block. */ | |
6516 | if (code_motion_path_driver_info->orig_expr_not_found | |
6517 | (insn, orig_ops, static_params)) | |
6518 | { | |
6519 | /* Av set ops could have been changed when moving through this | |
6520 | insn. To find them below it, we have to un-substitute them. */ | |
6521 | undo_transformations (&orig_ops, insn); | |
6522 | } | |
6523 | else | |
6524 | { | |
6525 | /* Clean up and return, if the hook tells us to do so. It may | |
6526 | happen if we've encountered the previously created | |
6527 | bookkeeping. */ | |
6528 | code_motion_path_driver_cleanup (&orig_ops, &path); | |
6529 | return -1; | |
6530 | } | |
6531 | ||
6532 | gcc_assert (orig_ops); | |
6533 | } | |
6534 | ||
6535 | /* Stop at insn if we got to the end of BB. */ | |
6536 | if (insn == bb_tail) | |
6537 | break; | |
6538 | ||
6539 | insn = NEXT_INSN (insn); | |
6540 | } | |
6541 | ||
6542 | /* Here INSN either points to the insn before the original insn (may be | |
6543 | bb_note, if original insn was a bb_head) or to the bb_end. */ | |
6544 | if (!expr) | |
6545 | { | |
6546 | int res; | |
6547 | ||
6548 | gcc_assert (insn == sel_bb_end (bb)); | |
6549 | ||
6550 | /* Add bb tail to PATH (but it doesn't make any sense if it's a bb_head - | |
6551 | it's already in PATH then). */ | |
6552 | if (insn != first_insn) | |
6553 | ilist_add (&path, insn); | |
6554 | ||
6555 | /* Process_successors should be able to find at least one | |
6556 | successor for which code_motion_path_driver returns TRUE. */ | |
6557 | res = code_motion_process_successors (insn, orig_ops, | |
6558 | path, static_params); | |
6559 | ||
6560 | /* Remove bb tail from path. */ | |
6561 | if (insn != first_insn) | |
6562 | ilist_remove (&path); | |
6563 | ||
6564 | if (res != 1) | |
6565 | { | |
6566 | /* This is the case when one of the original expr is no longer available | |
6567 | due to bookkeeping created on this branch with the same register. | |
6568 | In the original algorithm, which doesn't have update_data_sets call | |
6569 | on a bookkeeping block, it would simply result in returning | |
6570 | FALSE when we've encountered a previously generated bookkeeping | |
6571 | insn in moveop_orig_expr_not_found. */ | |
6572 | code_motion_path_driver_cleanup (&orig_ops, &path); | |
6573 | return res; | |
6574 | } | |
6575 | } | |
6576 | ||
6577 | /* Don't need it any more. */ | |
6578 | av_set_clear (&orig_ops); | |
6579 | ||
6580 | /* Backward pass: now, when we have C_EXPR computed, we'll drag it to | |
6581 | the beginning of the basic block. */ | |
6582 | before_first = PREV_INSN (first_insn); | |
6583 | while (insn != before_first) | |
6584 | { | |
6585 | if (code_motion_path_driver_info->ascend) | |
6586 | code_motion_path_driver_info->ascend (insn, static_params); | |
6587 | ||
6588 | insn = PREV_INSN (insn); | |
6589 | } | |
6590 | ||
6591 | /* Now we're at the bb head. */ | |
6592 | insn = first_insn; | |
6593 | ilist_remove (&path); | |
6594 | local_params_in->removed_last_insn = removed_last_insn; | |
6595 | code_motion_path_driver_info->at_first_insn (insn, local_params_in, static_params); | |
6596 | ||
6597 | /* This should be the very last operation as at bb head we could change | |
6598 | the numbering by creating bookkeeping blocks. */ | |
6599 | if (removed_last_insn) | |
6600 | insn = PREV_INSN (insn); | |
6601 | bitmap_set_bit (code_motion_visited_blocks, BLOCK_FOR_INSN (insn)->index); | |
6602 | return true; | |
6603 | } | |
6604 | ||
6605 | /* Move up the operations from ORIG_OPS set traversing the dag starting | |
6606 | from INSN. PATH represents the edges traversed so far. | |
6607 | DEST is the register chosen for scheduling the current expr. Insert | |
6608 | bookkeeping code in the join points. EXPR_VLIW is the chosen expression, | |
6609 | C_EXPR is how it looks like at the given cfg point. | |
72a54528 AM |
6610 | Set *SHOULD_MOVE to indicate whether we have only disconnected |
6611 | one of the insns found. | |
e855c69d AB |
6612 | |
6613 | Returns whether original instructions were found, which is asserted | |
6614 | to be true in the caller. */ | |
6615 | static bool | |
6616 | move_op (insn_t insn, av_set_t orig_ops, expr_t expr_vliw, | |
72a54528 | 6617 | rtx dest, expr_t c_expr, bool *should_move) |
e855c69d AB |
6618 | { |
6619 | struct moveop_static_params sparams; | |
6620 | struct cmpd_local_params lparams; | |
6621 | bool res; | |
6622 | ||
6623 | /* Init params for code_motion_path_driver. */ | |
6624 | sparams.dest = dest; | |
6625 | sparams.c_expr = c_expr; | |
6626 | sparams.uid = INSN_UID (EXPR_INSN_RTX (expr_vliw)); | |
6627 | #ifdef ENABLE_CHECKING | |
6628 | sparams.failed_insn = NULL; | |
6629 | #endif | |
6630 | sparams.was_renamed = false; | |
6631 | lparams.e1 = NULL; | |
6632 | ||
6633 | /* We haven't visited any blocks yet. */ | |
6634 | bitmap_clear (code_motion_visited_blocks); | |
6635 | ||
6636 | /* Set appropriate hooks and data. */ | |
6637 | code_motion_path_driver_info = &move_op_hooks; | |
6638 | res = code_motion_path_driver (insn, orig_ops, NULL, &lparams, &sparams); | |
6639 | ||
6640 | if (sparams.was_renamed) | |
6641 | EXPR_WAS_RENAMED (expr_vliw) = true; | |
6642 | ||
72a54528 AM |
6643 | *should_move = (sparams.uid == -1); |
6644 | ||
e855c69d AB |
6645 | return res; |
6646 | } | |
6647 | \f | |
6648 | ||
6649 | /* Functions that work with regions. */ | |
6650 | ||
6651 | /* Current number of seqno used in init_seqno and init_seqno_1. */ | |
6652 | static int cur_seqno; | |
6653 | ||
6654 | /* A helper for init_seqno. Traverse the region starting from BB and | |
6655 | compute seqnos for visited insns, marking visited bbs in VISITED_BBS. | |
6656 | Clear visited blocks from BLOCKS_TO_RESCHEDULE. */ | |
6657 | static void | |
6658 | init_seqno_1 (basic_block bb, sbitmap visited_bbs, bitmap blocks_to_reschedule) | |
6659 | { | |
6660 | int bbi = BLOCK_TO_BB (bb->index); | |
6661 | insn_t insn, note = bb_note (bb); | |
6662 | insn_t succ_insn; | |
6663 | succ_iterator si; | |
6664 | ||
6665 | SET_BIT (visited_bbs, bbi); | |
6666 | if (blocks_to_reschedule) | |
6667 | bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
6668 | ||
6669 | FOR_EACH_SUCC_1 (succ_insn, si, BB_END (bb), | |
6670 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
6671 | { | |
6672 | basic_block succ = BLOCK_FOR_INSN (succ_insn); | |
6673 | int succ_bbi = BLOCK_TO_BB (succ->index); | |
6674 | ||
6675 | gcc_assert (in_current_region_p (succ)); | |
6676 | ||
6677 | if (!TEST_BIT (visited_bbs, succ_bbi)) | |
6678 | { | |
6679 | gcc_assert (succ_bbi > bbi); | |
6680 | ||
6681 | init_seqno_1 (succ, visited_bbs, blocks_to_reschedule); | |
6682 | } | |
6683 | } | |
6684 | ||
6685 | for (insn = BB_END (bb); insn != note; insn = PREV_INSN (insn)) | |
6686 | INSN_SEQNO (insn) = cur_seqno--; | |
6687 | } | |
6688 | ||
6689 | /* Initialize seqnos for the current region. NUMBER_OF_INSNS is the number | |
6690 | of instructions in the region, BLOCKS_TO_RESCHEDULE contains blocks on | |
6691 | which we're rescheduling when pipelining, FROM is the block where | |
6692 | traversing region begins (it may not be the head of the region when | |
6693 | pipelining, but the head of the loop instead). | |
6694 | ||
6695 | Returns the maximal seqno found. */ | |
6696 | static int | |
6697 | init_seqno (int number_of_insns, bitmap blocks_to_reschedule, basic_block from) | |
6698 | { | |
6699 | sbitmap visited_bbs; | |
6700 | bitmap_iterator bi; | |
6701 | unsigned bbi; | |
6702 | ||
6703 | visited_bbs = sbitmap_alloc (current_nr_blocks); | |
6704 | ||
6705 | if (blocks_to_reschedule) | |
6706 | { | |
6707 | sbitmap_ones (visited_bbs); | |
6708 | EXECUTE_IF_SET_IN_BITMAP (blocks_to_reschedule, 0, bbi, bi) | |
6709 | { | |
6710 | gcc_assert (BLOCK_TO_BB (bbi) < current_nr_blocks); | |
6711 | RESET_BIT (visited_bbs, BLOCK_TO_BB (bbi)); | |
6712 | } | |
6713 | } | |
6714 | else | |
6715 | { | |
6716 | sbitmap_zero (visited_bbs); | |
6717 | from = EBB_FIRST_BB (0); | |
6718 | } | |
6719 | ||
6720 | cur_seqno = number_of_insns > 0 ? number_of_insns : sched_max_luid - 1; | |
6721 | init_seqno_1 (from, visited_bbs, blocks_to_reschedule); | |
6722 | gcc_assert (cur_seqno == 0 || number_of_insns == 0); | |
6723 | ||
6724 | sbitmap_free (visited_bbs); | |
6725 | return sched_max_luid - 1; | |
6726 | } | |
6727 | ||
6728 | /* Initialize scheduling parameters for current region. */ | |
6729 | static void | |
6730 | sel_setup_region_sched_flags (void) | |
6731 | { | |
6732 | enable_schedule_as_rhs_p = 1; | |
6733 | bookkeeping_p = 1; | |
6734 | pipelining_p = (bookkeeping_p | |
6735 | && (flag_sel_sched_pipelining != 0) | |
6736 | && current_loop_nest != NULL); | |
6737 | max_insns_to_rename = PARAM_VALUE (PARAM_SELSCHED_INSNS_TO_RENAME); | |
6738 | max_ws = MAX_WS; | |
6739 | } | |
6740 | ||
6741 | /* Return true if all basic blocks of current region are empty. */ | |
6742 | static bool | |
6743 | current_region_empty_p (void) | |
6744 | { | |
6745 | int i; | |
6746 | for (i = 0; i < current_nr_blocks; i++) | |
6747 | if (! sel_bb_empty_p (BASIC_BLOCK (BB_TO_BLOCK (i)))) | |
6748 | return false; | |
6749 | ||
6750 | return true; | |
6751 | } | |
6752 | ||
6753 | /* Prepare and verify loop nest for pipelining. */ | |
6754 | static void | |
6755 | setup_current_loop_nest (int rgn) | |
6756 | { | |
6757 | current_loop_nest = get_loop_nest_for_rgn (rgn); | |
6758 | ||
6759 | if (!current_loop_nest) | |
6760 | return; | |
6761 | ||
6762 | /* If this loop has any saved loop preheaders from nested loops, | |
6763 | add these basic blocks to the current region. */ | |
6764 | sel_add_loop_preheaders (); | |
6765 | ||
6766 | /* Check that we're starting with a valid information. */ | |
6767 | gcc_assert (loop_latch_edge (current_loop_nest)); | |
6768 | gcc_assert (LOOP_MARKED_FOR_PIPELINING_P (current_loop_nest)); | |
6769 | } | |
6770 | ||
6771 | /* Purge meaningless empty blocks in the middle of a region. */ | |
6772 | static void | |
6773 | purge_empty_blocks (void) | |
6774 | { | |
f2c45f08 AM |
6775 | /* Do not attempt to delete preheader. */ |
6776 | int i = sel_is_loop_preheader_p (BASIC_BLOCK (BB_TO_BLOCK (0))) ? 1 : 0; | |
e855c69d | 6777 | |
f2c45f08 | 6778 | while (i < current_nr_blocks) |
e855c69d AB |
6779 | { |
6780 | basic_block b = BASIC_BLOCK (BB_TO_BLOCK (i)); | |
6781 | ||
6782 | if (maybe_tidy_empty_bb (b)) | |
6783 | continue; | |
6784 | ||
6785 | i++; | |
6786 | } | |
6787 | } | |
6788 | ||
6789 | /* Compute instruction priorities for current region. */ | |
6790 | static void | |
6791 | sel_compute_priorities (int rgn) | |
6792 | { | |
6793 | sched_rgn_compute_dependencies (rgn); | |
6794 | ||
6795 | /* Compute insn priorities in haifa style. Then free haifa style | |
6796 | dependencies that we've calculated for this. */ | |
6797 | compute_priorities (); | |
6798 | ||
6799 | if (sched_verbose >= 5) | |
6800 | debug_rgn_dependencies (0); | |
6801 | ||
6802 | free_rgn_deps (); | |
6803 | } | |
6804 | ||
6805 | /* Init scheduling data for RGN. Returns true when this region should not | |
6806 | be scheduled. */ | |
6807 | static bool | |
6808 | sel_region_init (int rgn) | |
6809 | { | |
6810 | int i; | |
6811 | bb_vec_t bbs; | |
6812 | ||
6813 | rgn_setup_region (rgn); | |
6814 | ||
6815 | /* Even if sched_is_disabled_for_current_region_p() is true, we still | |
6816 | do region initialization here so the region can be bundled correctly, | |
6817 | but we'll skip the scheduling in sel_sched_region (). */ | |
6818 | if (current_region_empty_p ()) | |
6819 | return true; | |
6820 | ||
6821 | if (flag_sel_sched_pipelining) | |
6822 | setup_current_loop_nest (rgn); | |
6823 | ||
6824 | sel_setup_region_sched_flags (); | |
6825 | ||
6826 | bbs = VEC_alloc (basic_block, heap, current_nr_blocks); | |
6827 | ||
6828 | for (i = 0; i < current_nr_blocks; i++) | |
6829 | VEC_quick_push (basic_block, bbs, BASIC_BLOCK (BB_TO_BLOCK (i))); | |
6830 | ||
6831 | sel_init_bbs (bbs, NULL); | |
6832 | ||
6833 | /* Initialize luids and dependence analysis which both sel-sched and haifa | |
6834 | need. */ | |
6835 | sched_init_luids (bbs, NULL, NULL, NULL); | |
6836 | sched_deps_init (false); | |
6837 | ||
6838 | /* Initialize haifa data. */ | |
6839 | rgn_setup_sched_infos (); | |
6840 | sel_set_sched_flags (); | |
6841 | haifa_init_h_i_d (bbs, NULL, NULL, NULL); | |
6842 | ||
6843 | sel_compute_priorities (rgn); | |
6844 | init_deps_global (); | |
6845 | ||
6846 | /* Main initialization. */ | |
6847 | sel_setup_sched_infos (); | |
6848 | sel_init_global_and_expr (bbs); | |
6849 | ||
6850 | VEC_free (basic_block, heap, bbs); | |
6851 | ||
6852 | blocks_to_reschedule = BITMAP_ALLOC (NULL); | |
6853 | ||
6854 | /* Init correct liveness sets on each instruction of a single-block loop. | |
6855 | This is the only situation when we can't update liveness when calling | |
6856 | compute_live for the first insn of the loop. */ | |
6857 | if (current_loop_nest) | |
6858 | { | |
6859 | int header = (sel_is_loop_preheader_p (BASIC_BLOCK (BB_TO_BLOCK (0))) | |
6860 | ? 1 | |
6861 | : 0); | |
6862 | ||
6863 | if (current_nr_blocks == header + 1) | |
6864 | update_liveness_on_insn | |
6865 | (sel_bb_head (BASIC_BLOCK (BB_TO_BLOCK (header)))); | |
6866 | } | |
6867 | ||
6868 | /* Set hooks so that no newly generated insn will go out unnoticed. */ | |
6869 | sel_register_cfg_hooks (); | |
6870 | ||
6871 | /* !!! We call target.sched.md_init () for the whole region, but we invoke | |
6872 | targetm.sched.md_finish () for every ebb. */ | |
6873 | if (targetm.sched.md_init) | |
6874 | /* None of the arguments are actually used in any target. */ | |
6875 | targetm.sched.md_init (sched_dump, sched_verbose, -1); | |
6876 | ||
6877 | first_emitted_uid = get_max_uid () + 1; | |
6878 | preheader_removed = false; | |
6879 | ||
6880 | /* Reset register allocation ticks array. */ | |
6881 | memset (reg_rename_tick, 0, sizeof reg_rename_tick); | |
6882 | reg_rename_this_tick = 0; | |
6883 | ||
6884 | bitmap_initialize (forced_ebb_heads, 0); | |
6885 | bitmap_clear (forced_ebb_heads); | |
6886 | ||
6887 | setup_nop_vinsn (); | |
6888 | current_copies = BITMAP_ALLOC (NULL); | |
6889 | current_originators = BITMAP_ALLOC (NULL); | |
6890 | code_motion_visited_blocks = BITMAP_ALLOC (NULL); | |
6891 | ||
6892 | return false; | |
6893 | } | |
6894 | ||
6895 | /* Simplify insns after the scheduling. */ | |
6896 | static void | |
6897 | simplify_changed_insns (void) | |
6898 | { | |
6899 | int i; | |
6900 | ||
6901 | for (i = 0; i < current_nr_blocks; i++) | |
6902 | { | |
6903 | basic_block bb = BASIC_BLOCK (BB_TO_BLOCK (i)); | |
6904 | rtx insn; | |
6905 | ||
6906 | FOR_BB_INSNS (bb, insn) | |
6907 | if (INSN_P (insn)) | |
6908 | { | |
6909 | expr_t expr = INSN_EXPR (insn); | |
6910 | ||
6911 | if (EXPR_WAS_SUBSTITUTED (expr)) | |
6912 | validate_simplify_insn (insn); | |
6913 | } | |
6914 | } | |
6915 | } | |
6916 | ||
6917 | /* Find boundaries of the EBB starting from basic block BB, marking blocks of | |
6918 | this EBB in SCHEDULED_BLOCKS and appropriately filling in HEAD, TAIL, | |
6919 | PREV_HEAD, and NEXT_TAIL fields of CURRENT_SCHED_INFO structure. */ | |
6920 | static void | |
6921 | find_ebb_boundaries (basic_block bb, bitmap scheduled_blocks) | |
6922 | { | |
6923 | insn_t head, tail; | |
6924 | basic_block bb1 = bb; | |
6925 | if (sched_verbose >= 2) | |
6926 | sel_print ("Finishing schedule in bbs: "); | |
6927 | ||
6928 | do | |
6929 | { | |
6930 | bitmap_set_bit (scheduled_blocks, BLOCK_TO_BB (bb1->index)); | |
6931 | ||
6932 | if (sched_verbose >= 2) | |
6933 | sel_print ("%d; ", bb1->index); | |
6934 | } | |
6935 | while (!bb_ends_ebb_p (bb1) && (bb1 = bb_next_bb (bb1))); | |
6936 | ||
6937 | if (sched_verbose >= 2) | |
6938 | sel_print ("\n"); | |
6939 | ||
6940 | get_ebb_head_tail (bb, bb1, &head, &tail); | |
6941 | ||
6942 | current_sched_info->head = head; | |
6943 | current_sched_info->tail = tail; | |
6944 | current_sched_info->prev_head = PREV_INSN (head); | |
6945 | current_sched_info->next_tail = NEXT_INSN (tail); | |
6946 | } | |
6947 | ||
6948 | /* Regenerate INSN_SCHED_CYCLEs for insns of current EBB. */ | |
6949 | static void | |
6950 | reset_sched_cycles_in_current_ebb (void) | |
6951 | { | |
6952 | int last_clock = 0; | |
6953 | int haifa_last_clock = -1; | |
6954 | int haifa_clock = 0; | |
6955 | insn_t insn; | |
6956 | ||
6957 | if (targetm.sched.md_init) | |
6958 | { | |
6959 | /* None of the arguments are actually used in any target. | |
6960 | NB: We should have md_reset () hook for cases like this. */ | |
6961 | targetm.sched.md_init (sched_dump, sched_verbose, -1); | |
6962 | } | |
6963 | ||
6964 | state_reset (curr_state); | |
6965 | advance_state (curr_state); | |
6966 | ||
6967 | for (insn = current_sched_info->head; | |
6968 | insn != current_sched_info->next_tail; | |
6969 | insn = NEXT_INSN (insn)) | |
6970 | { | |
6971 | int cost, haifa_cost; | |
6972 | int sort_p; | |
6973 | bool asm_p, real_insn, after_stall; | |
6974 | int clock; | |
6975 | ||
6976 | if (!INSN_P (insn)) | |
6977 | continue; | |
6978 | ||
6979 | asm_p = false; | |
6980 | real_insn = recog_memoized (insn) >= 0; | |
6981 | clock = INSN_SCHED_CYCLE (insn); | |
6982 | ||
6983 | cost = clock - last_clock; | |
6984 | ||
6985 | /* Initialize HAIFA_COST. */ | |
6986 | if (! real_insn) | |
6987 | { | |
6988 | asm_p = INSN_ASM_P (insn); | |
6989 | ||
6990 | if (asm_p) | |
6991 | /* This is asm insn which *had* to be scheduled first | |
6992 | on the cycle. */ | |
6993 | haifa_cost = 1; | |
6994 | else | |
6995 | /* This is a use/clobber insn. It should not change | |
6996 | cost. */ | |
6997 | haifa_cost = 0; | |
6998 | } | |
6999 | else | |
7000 | haifa_cost = estimate_insn_cost (insn, curr_state); | |
7001 | ||
7002 | /* Stall for whatever cycles we've stalled before. */ | |
7003 | after_stall = 0; | |
7004 | if (INSN_AFTER_STALL_P (insn) && cost > haifa_cost) | |
7005 | { | |
7006 | haifa_cost = cost; | |
7007 | after_stall = 1; | |
7008 | } | |
7009 | ||
7010 | if (haifa_cost > 0) | |
7011 | { | |
7012 | int i = 0; | |
7013 | ||
7014 | while (haifa_cost--) | |
7015 | { | |
7016 | advance_state (curr_state); | |
7017 | i++; | |
7018 | ||
7019 | if (sched_verbose >= 2) | |
7020 | { | |
7021 | sel_print ("advance_state (state_transition)\n"); | |
7022 | debug_state (curr_state); | |
7023 | } | |
7024 | ||
7025 | /* The DFA may report that e.g. insn requires 2 cycles to be | |
7026 | issued, but on the next cycle it says that insn is ready | |
7027 | to go. Check this here. */ | |
7028 | if (!after_stall | |
7029 | && real_insn | |
7030 | && haifa_cost > 0 | |
7031 | && estimate_insn_cost (insn, curr_state) == 0) | |
7032 | break; | |
7033 | } | |
7034 | ||
7035 | haifa_clock += i; | |
7036 | } | |
7037 | else | |
7038 | gcc_assert (haifa_cost == 0); | |
7039 | ||
7040 | if (sched_verbose >= 2) | |
7041 | sel_print ("Haifa cost for insn %d: %d\n", INSN_UID (insn), haifa_cost); | |
7042 | ||
7043 | if (targetm.sched.dfa_new_cycle) | |
7044 | while (targetm.sched.dfa_new_cycle (sched_dump, sched_verbose, insn, | |
7045 | haifa_last_clock, haifa_clock, | |
7046 | &sort_p)) | |
7047 | { | |
7048 | advance_state (curr_state); | |
7049 | haifa_clock++; | |
7050 | if (sched_verbose >= 2) | |
7051 | { | |
7052 | sel_print ("advance_state (dfa_new_cycle)\n"); | |
7053 | debug_state (curr_state); | |
7054 | } | |
7055 | } | |
7056 | ||
7057 | if (real_insn) | |
7058 | { | |
7059 | cost = state_transition (curr_state, insn); | |
7060 | ||
7061 | if (sched_verbose >= 2) | |
7062 | debug_state (curr_state); | |
7063 | ||
7064 | gcc_assert (cost < 0); | |
7065 | } | |
7066 | ||
7067 | if (targetm.sched.variable_issue) | |
7068 | targetm.sched.variable_issue (sched_dump, sched_verbose, insn, 0); | |
7069 | ||
7070 | INSN_SCHED_CYCLE (insn) = haifa_clock; | |
7071 | ||
7072 | last_clock = clock; | |
7073 | haifa_last_clock = haifa_clock; | |
7074 | } | |
7075 | } | |
7076 | ||
7077 | /* Put TImode markers on insns starting a new issue group. */ | |
7078 | static void | |
7079 | put_TImodes (void) | |
7080 | { | |
7081 | int last_clock = -1; | |
7082 | insn_t insn; | |
7083 | ||
7084 | for (insn = current_sched_info->head; insn != current_sched_info->next_tail; | |
7085 | insn = NEXT_INSN (insn)) | |
7086 | { | |
7087 | int cost, clock; | |
7088 | ||
7089 | if (!INSN_P (insn)) | |
7090 | continue; | |
7091 | ||
7092 | clock = INSN_SCHED_CYCLE (insn); | |
7093 | cost = (last_clock == -1) ? 1 : clock - last_clock; | |
7094 | ||
7095 | gcc_assert (cost >= 0); | |
7096 | ||
7097 | if (issue_rate > 1 | |
7098 | && GET_CODE (PATTERN (insn)) != USE | |
7099 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
7100 | { | |
7101 | if (reload_completed && cost > 0) | |
7102 | PUT_MODE (insn, TImode); | |
7103 | ||
7104 | last_clock = clock; | |
7105 | } | |
7106 | ||
7107 | if (sched_verbose >= 2) | |
7108 | sel_print ("Cost for insn %d is %d\n", INSN_UID (insn), cost); | |
7109 | } | |
7110 | } | |
7111 | ||
7112 | /* Perform MD_FINISH on EBBs comprising current region. When | |
7113 | RESET_SCHED_CYCLES_P is true, run a pass emulating the scheduler | |
7114 | to produce correct sched cycles on insns. */ | |
7115 | static void | |
7116 | sel_region_target_finish (bool reset_sched_cycles_p) | |
7117 | { | |
7118 | int i; | |
7119 | bitmap scheduled_blocks = BITMAP_ALLOC (NULL); | |
7120 | ||
7121 | for (i = 0; i < current_nr_blocks; i++) | |
7122 | { | |
7123 | if (bitmap_bit_p (scheduled_blocks, i)) | |
7124 | continue; | |
7125 | ||
7126 | /* While pipelining outer loops, skip bundling for loop | |
7127 | preheaders. Those will be rescheduled in the outer loop. */ | |
7128 | if (sel_is_loop_preheader_p (EBB_FIRST_BB (i))) | |
7129 | continue; | |
7130 | ||
7131 | find_ebb_boundaries (EBB_FIRST_BB (i), scheduled_blocks); | |
7132 | ||
7133 | if (no_real_insns_p (current_sched_info->head, current_sched_info->tail)) | |
7134 | continue; | |
7135 | ||
7136 | if (reset_sched_cycles_p) | |
7137 | reset_sched_cycles_in_current_ebb (); | |
7138 | ||
7139 | if (targetm.sched.md_init) | |
7140 | targetm.sched.md_init (sched_dump, sched_verbose, -1); | |
7141 | ||
7142 | put_TImodes (); | |
7143 | ||
7144 | if (targetm.sched.md_finish) | |
7145 | { | |
7146 | targetm.sched.md_finish (sched_dump, sched_verbose); | |
7147 | ||
7148 | /* Extend luids so that insns generated by the target will | |
7149 | get zero luid. */ | |
7150 | sched_init_luids (NULL, NULL, NULL, NULL); | |
7151 | } | |
7152 | } | |
7153 | ||
7154 | BITMAP_FREE (scheduled_blocks); | |
7155 | } | |
7156 | ||
7157 | /* Free the scheduling data for the current region. When RESET_SCHED_CYCLES_P | |
7158 | is true, make an additional pass emulating scheduler to get correct insn | |
7159 | cycles for md_finish calls. */ | |
7160 | static void | |
7161 | sel_region_finish (bool reset_sched_cycles_p) | |
7162 | { | |
7163 | simplify_changed_insns (); | |
7164 | sched_finish_ready_list (); | |
7165 | free_nop_pool (); | |
7166 | ||
7167 | /* Free the vectors. */ | |
7168 | if (vec_av_set) | |
7169 | VEC_free (expr_t, heap, vec_av_set); | |
7170 | BITMAP_FREE (current_copies); | |
7171 | BITMAP_FREE (current_originators); | |
7172 | BITMAP_FREE (code_motion_visited_blocks); | |
7173 | vinsn_vec_free (&vec_bookkeeping_blocked_vinsns); | |
7174 | vinsn_vec_free (&vec_target_unavailable_vinsns); | |
7175 | ||
7176 | /* If LV_SET of the region head should be updated, do it now because | |
7177 | there will be no other chance. */ | |
7178 | { | |
7179 | succ_iterator si; | |
7180 | insn_t insn; | |
7181 | ||
7182 | FOR_EACH_SUCC_1 (insn, si, bb_note (EBB_FIRST_BB (0)), | |
7183 | SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) | |
7184 | { | |
7185 | basic_block bb = BLOCK_FOR_INSN (insn); | |
7186 | ||
7187 | if (!BB_LV_SET_VALID_P (bb)) | |
7188 | compute_live (insn); | |
7189 | } | |
7190 | } | |
7191 | ||
7192 | /* Emulate the Haifa scheduler for bundling. */ | |
7193 | if (reload_completed) | |
7194 | sel_region_target_finish (reset_sched_cycles_p); | |
7195 | ||
7196 | sel_finish_global_and_expr (); | |
7197 | ||
7198 | bitmap_clear (forced_ebb_heads); | |
7199 | ||
7200 | free_nop_vinsn (); | |
7201 | ||
7202 | finish_deps_global (); | |
7203 | sched_finish_luids (); | |
7204 | ||
7205 | sel_finish_bbs (); | |
7206 | BITMAP_FREE (blocks_to_reschedule); | |
7207 | ||
7208 | sel_unregister_cfg_hooks (); | |
7209 | ||
7210 | max_issue_size = 0; | |
7211 | } | |
7212 | \f | |
7213 | ||
7214 | /* Functions that implement the scheduler driver. */ | |
7215 | ||
7216 | /* Schedule a parallel instruction group on each of FENCES. MAX_SEQNO | |
7217 | is the current maximum seqno. SCHEDULED_INSNS_TAILPP is the list | |
7218 | of insns scheduled -- these would be postprocessed later. */ | |
7219 | static void | |
7220 | schedule_on_fences (flist_t fences, int max_seqno, | |
7221 | ilist_t **scheduled_insns_tailpp) | |
7222 | { | |
7223 | flist_t old_fences = fences; | |
7224 | ||
7225 | if (sched_verbose >= 1) | |
7226 | { | |
7227 | sel_print ("\nScheduling on fences: "); | |
7228 | dump_flist (fences); | |
7229 | sel_print ("\n"); | |
7230 | } | |
7231 | ||
7232 | scheduled_something_on_previous_fence = false; | |
7233 | for (; fences; fences = FLIST_NEXT (fences)) | |
7234 | { | |
7235 | fence_t fence = NULL; | |
7236 | int seqno = 0; | |
7237 | flist_t fences2; | |
7238 | bool first_p = true; | |
7239 | ||
7240 | /* Choose the next fence group to schedule. | |
7241 | The fact that insn can be scheduled only once | |
7242 | on the cycle is guaranteed by two properties: | |
7243 | 1. seqnos of parallel groups decrease with each iteration. | |
7244 | 2. If is_ineligible_successor () sees the larger seqno, it | |
7245 | checks if candidate insn is_in_current_fence_p (). */ | |
7246 | for (fences2 = old_fences; fences2; fences2 = FLIST_NEXT (fences2)) | |
7247 | { | |
7248 | fence_t f = FLIST_FENCE (fences2); | |
7249 | ||
7250 | if (!FENCE_PROCESSED_P (f)) | |
7251 | { | |
7252 | int i = INSN_SEQNO (FENCE_INSN (f)); | |
7253 | ||
7254 | if (first_p || i > seqno) | |
7255 | { | |
7256 | seqno = i; | |
7257 | fence = f; | |
7258 | first_p = false; | |
7259 | } | |
7260 | else | |
7261 | /* ??? Seqnos of different groups should be different. */ | |
7262 | gcc_assert (1 || i != seqno); | |
7263 | } | |
7264 | } | |
7265 | ||
7266 | gcc_assert (fence); | |
7267 | ||
7268 | /* As FENCE is nonnull, SEQNO is initialized. */ | |
7269 | seqno -= max_seqno + 1; | |
7270 | fill_insns (fence, seqno, scheduled_insns_tailpp); | |
7271 | FENCE_PROCESSED_P (fence) = true; | |
7272 | } | |
7273 | ||
7274 | /* All av_sets are invalidated by GLOBAL_LEVEL increase, thus we | |
7275 | don't need to keep bookkeeping-invalidated and target-unavailable | |
7276 | vinsns any more. */ | |
7277 | vinsn_vec_clear (&vec_bookkeeping_blocked_vinsns); | |
7278 | vinsn_vec_clear (&vec_target_unavailable_vinsns); | |
7279 | } | |
7280 | ||
7281 | /* Calculate MIN_SEQNO and MAX_SEQNO. */ | |
7282 | static void | |
7283 | find_min_max_seqno (flist_t fences, int *min_seqno, int *max_seqno) | |
7284 | { | |
7285 | *min_seqno = *max_seqno = INSN_SEQNO (FENCE_INSN (FLIST_FENCE (fences))); | |
7286 | ||
7287 | /* The first element is already processed. */ | |
7288 | while ((fences = FLIST_NEXT (fences))) | |
7289 | { | |
7290 | int seqno = INSN_SEQNO (FENCE_INSN (FLIST_FENCE (fences))); | |
7291 | ||
7292 | if (*min_seqno > seqno) | |
7293 | *min_seqno = seqno; | |
7294 | else if (*max_seqno < seqno) | |
7295 | *max_seqno = seqno; | |
7296 | } | |
7297 | } | |
7298 | ||
7299 | /* Calculate new fences from FENCES. */ | |
7300 | static flist_t | |
7301 | calculate_new_fences (flist_t fences, int orig_max_seqno) | |
7302 | { | |
7303 | flist_t old_fences = fences; | |
7304 | struct flist_tail_def _new_fences, *new_fences = &_new_fences; | |
7305 | ||
7306 | flist_tail_init (new_fences); | |
7307 | for (; fences; fences = FLIST_NEXT (fences)) | |
7308 | { | |
7309 | fence_t fence = FLIST_FENCE (fences); | |
7310 | insn_t insn; | |
7311 | ||
7312 | if (!FENCE_BNDS (fence)) | |
7313 | { | |
7314 | /* This fence doesn't have any successors. */ | |
7315 | if (!FENCE_SCHEDULED_P (fence)) | |
7316 | { | |
7317 | /* Nothing was scheduled on this fence. */ | |
7318 | int seqno; | |
7319 | ||
7320 | insn = FENCE_INSN (fence); | |
7321 | seqno = INSN_SEQNO (insn); | |
7322 | gcc_assert (seqno > 0 && seqno <= orig_max_seqno); | |
7323 | ||
7324 | if (sched_verbose >= 1) | |
7325 | sel_print ("Fence %d[%d] has not changed\n", | |
7326 | INSN_UID (insn), | |
7327 | BLOCK_NUM (insn)); | |
7328 | move_fence_to_fences (fences, new_fences); | |
7329 | } | |
7330 | } | |
7331 | else | |
7332 | extract_new_fences_from (fences, new_fences, orig_max_seqno); | |
7333 | } | |
7334 | ||
7335 | flist_clear (&old_fences); | |
7336 | return FLIST_TAIL_HEAD (new_fences); | |
7337 | } | |
7338 | ||
7339 | /* Update seqnos of insns given by PSCHEDULED_INSNS. MIN_SEQNO and MAX_SEQNO | |
7340 | are the miminum and maximum seqnos of the group, HIGHEST_SEQNO_IN_USE is | |
7341 | the highest seqno used in a region. Return the updated highest seqno. */ | |
7342 | static int | |
7343 | update_seqnos_and_stage (int min_seqno, int max_seqno, | |
7344 | int highest_seqno_in_use, | |
7345 | ilist_t *pscheduled_insns) | |
7346 | { | |
7347 | int new_hs; | |
7348 | ilist_iterator ii; | |
7349 | insn_t insn; | |
7350 | ||
7351 | /* Actually, new_hs is the seqno of the instruction, that was | |
7352 | scheduled first (i.e. it is the first one in SCHEDULED_INSNS). */ | |
7353 | if (*pscheduled_insns) | |
7354 | { | |
7355 | new_hs = (INSN_SEQNO (ILIST_INSN (*pscheduled_insns)) | |
7356 | + highest_seqno_in_use + max_seqno - min_seqno + 2); | |
7357 | gcc_assert (new_hs > highest_seqno_in_use); | |
7358 | } | |
7359 | else | |
7360 | new_hs = highest_seqno_in_use; | |
7361 | ||
7362 | FOR_EACH_INSN (insn, ii, *pscheduled_insns) | |
7363 | { | |
7364 | gcc_assert (INSN_SEQNO (insn) < 0); | |
7365 | INSN_SEQNO (insn) += highest_seqno_in_use + max_seqno - min_seqno + 2; | |
7366 | gcc_assert (INSN_SEQNO (insn) <= new_hs); | |
bcf33775 AB |
7367 | |
7368 | /* When not pipelining, purge unneeded insn info on the scheduled insns. | |
7369 | For example, having reg_last array of INSN_DEPS_CONTEXT in memory may | |
7370 | require > 1GB of memory e.g. on limit-fnargs.c. */ | |
7371 | if (! pipelining_p) | |
7372 | free_data_for_scheduled_insn (insn); | |
e855c69d AB |
7373 | } |
7374 | ||
7375 | ilist_clear (pscheduled_insns); | |
7376 | global_level++; | |
7377 | ||
7378 | return new_hs; | |
7379 | } | |
7380 | ||
7381 | /* The main driver for scheduling a region. This function is responsible | |
7382 | for correct propagation of fences (i.e. scheduling points) and creating | |
7383 | a group of parallel insns at each of them. It also supports | |
7384 | pipelining. ORIG_MAX_SEQNO is the maximal seqno before this pass | |
7385 | of scheduling. */ | |
7386 | static void | |
7387 | sel_sched_region_2 (int orig_max_seqno) | |
7388 | { | |
7389 | int highest_seqno_in_use = orig_max_seqno; | |
7390 | ||
7391 | stat_bookkeeping_copies = 0; | |
7392 | stat_insns_needed_bookkeeping = 0; | |
7393 | stat_renamed_scheduled = 0; | |
7394 | stat_substitutions_total = 0; | |
7395 | num_insns_scheduled = 0; | |
7396 | ||
7397 | while (fences) | |
7398 | { | |
7399 | int min_seqno, max_seqno; | |
7400 | ilist_t scheduled_insns = NULL; | |
7401 | ilist_t *scheduled_insns_tailp = &scheduled_insns; | |
7402 | ||
7403 | find_min_max_seqno (fences, &min_seqno, &max_seqno); | |
7404 | schedule_on_fences (fences, max_seqno, &scheduled_insns_tailp); | |
7405 | fences = calculate_new_fences (fences, orig_max_seqno); | |
7406 | highest_seqno_in_use = update_seqnos_and_stage (min_seqno, max_seqno, | |
7407 | highest_seqno_in_use, | |
7408 | &scheduled_insns); | |
7409 | } | |
7410 | ||
7411 | if (sched_verbose >= 1) | |
7412 | sel_print ("Scheduled %d bookkeeping copies, %d insns needed " | |
7413 | "bookkeeping, %d insns renamed, %d insns substituted\n", | |
7414 | stat_bookkeeping_copies, | |
7415 | stat_insns_needed_bookkeeping, | |
7416 | stat_renamed_scheduled, | |
7417 | stat_substitutions_total); | |
7418 | } | |
7419 | ||
7420 | /* Schedule a region. When pipelining, search for possibly never scheduled | |
7421 | bookkeeping code and schedule it. Reschedule pipelined code without | |
7422 | pipelining after. */ | |
7423 | static void | |
7424 | sel_sched_region_1 (void) | |
7425 | { | |
7426 | int number_of_insns; | |
7427 | int orig_max_seqno; | |
7428 | ||
7429 | /* Remove empty blocks that might be in the region from the beginning. | |
7430 | We need to do save sched_max_luid before that, as it actually shows | |
7431 | the number of insns in the region, and purge_empty_blocks can | |
7432 | alter it. */ | |
7433 | number_of_insns = sched_max_luid - 1; | |
7434 | purge_empty_blocks (); | |
7435 | ||
7436 | orig_max_seqno = init_seqno (number_of_insns, NULL, NULL); | |
7437 | gcc_assert (orig_max_seqno >= 1); | |
7438 | ||
7439 | /* When pipelining outer loops, create fences on the loop header, | |
7440 | not preheader. */ | |
7441 | fences = NULL; | |
7442 | if (current_loop_nest) | |
7443 | init_fences (BB_END (EBB_FIRST_BB (0))); | |
7444 | else | |
7445 | init_fences (bb_note (EBB_FIRST_BB (0))); | |
7446 | global_level = 1; | |
7447 | ||
7448 | sel_sched_region_2 (orig_max_seqno); | |
7449 | ||
7450 | gcc_assert (fences == NULL); | |
7451 | ||
7452 | if (pipelining_p) | |
7453 | { | |
7454 | int i; | |
7455 | basic_block bb; | |
7456 | struct flist_tail_def _new_fences; | |
7457 | flist_tail_t new_fences = &_new_fences; | |
7458 | bool do_p = true; | |
7459 | ||
7460 | pipelining_p = false; | |
7461 | max_ws = MIN (max_ws, issue_rate * 3 / 2); | |
7462 | bookkeeping_p = false; | |
7463 | enable_schedule_as_rhs_p = false; | |
7464 | ||
7465 | /* Schedule newly created code, that has not been scheduled yet. */ | |
7466 | do_p = true; | |
7467 | ||
7468 | while (do_p) | |
7469 | { | |
7470 | do_p = false; | |
7471 | ||
7472 | for (i = 0; i < current_nr_blocks; i++) | |
7473 | { | |
7474 | basic_block bb = EBB_FIRST_BB (i); | |
7475 | ||
7476 | if (sel_bb_empty_p (bb)) | |
7477 | { | |
7478 | bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
7479 | continue; | |
7480 | } | |
7481 | ||
7482 | if (bitmap_bit_p (blocks_to_reschedule, bb->index)) | |
7483 | { | |
7484 | clear_outdated_rtx_info (bb); | |
7485 | if (sel_insn_is_speculation_check (BB_END (bb)) | |
7486 | && JUMP_P (BB_END (bb))) | |
7487 | bitmap_set_bit (blocks_to_reschedule, | |
7488 | BRANCH_EDGE (bb)->dest->index); | |
7489 | } | |
7490 | else if (INSN_SCHED_TIMES (sel_bb_head (bb)) <= 0) | |
7491 | bitmap_set_bit (blocks_to_reschedule, bb->index); | |
7492 | } | |
7493 | ||
7494 | for (i = 0; i < current_nr_blocks; i++) | |
7495 | { | |
7496 | bb = EBB_FIRST_BB (i); | |
7497 | ||
7498 | /* While pipelining outer loops, skip bundling for loop | |
7499 | preheaders. Those will be rescheduled in the outer | |
7500 | loop. */ | |
7501 | if (sel_is_loop_preheader_p (bb)) | |
7502 | { | |
7503 | clear_outdated_rtx_info (bb); | |
7504 | continue; | |
7505 | } | |
7506 | ||
7507 | if (bitmap_bit_p (blocks_to_reschedule, bb->index)) | |
7508 | { | |
7509 | flist_tail_init (new_fences); | |
7510 | ||
7511 | orig_max_seqno = init_seqno (0, blocks_to_reschedule, bb); | |
7512 | ||
7513 | /* Mark BB as head of the new ebb. */ | |
7514 | bitmap_set_bit (forced_ebb_heads, bb->index); | |
7515 | ||
7516 | bitmap_clear_bit (blocks_to_reschedule, bb->index); | |
7517 | ||
7518 | gcc_assert (fences == NULL); | |
7519 | ||
7520 | init_fences (bb_note (bb)); | |
7521 | ||
7522 | sel_sched_region_2 (orig_max_seqno); | |
7523 | ||
7524 | do_p = true; | |
7525 | break; | |
7526 | } | |
7527 | } | |
7528 | } | |
7529 | } | |
7530 | } | |
7531 | ||
7532 | /* Schedule the RGN region. */ | |
7533 | void | |
7534 | sel_sched_region (int rgn) | |
7535 | { | |
7536 | bool schedule_p; | |
7537 | bool reset_sched_cycles_p; | |
7538 | ||
7539 | if (sel_region_init (rgn)) | |
7540 | return; | |
7541 | ||
7542 | if (sched_verbose >= 1) | |
7543 | sel_print ("Scheduling region %d\n", rgn); | |
7544 | ||
7545 | schedule_p = (!sched_is_disabled_for_current_region_p () | |
7546 | && dbg_cnt (sel_sched_region_cnt)); | |
7547 | reset_sched_cycles_p = pipelining_p; | |
7548 | if (schedule_p) | |
7549 | sel_sched_region_1 (); | |
7550 | else | |
7551 | /* Force initialization of INSN_SCHED_CYCLEs for correct bundling. */ | |
7552 | reset_sched_cycles_p = true; | |
7553 | ||
7554 | sel_region_finish (reset_sched_cycles_p); | |
7555 | } | |
7556 | ||
7557 | /* Perform global init for the scheduler. */ | |
7558 | static void | |
7559 | sel_global_init (void) | |
7560 | { | |
7561 | calculate_dominance_info (CDI_DOMINATORS); | |
7562 | alloc_sched_pools (); | |
7563 | ||
7564 | /* Setup the infos for sched_init. */ | |
7565 | sel_setup_sched_infos (); | |
7566 | setup_sched_dump (); | |
7567 | ||
7568 | sched_rgn_init (false); | |
7569 | sched_init (); | |
7570 | ||
7571 | sched_init_bbs (); | |
7572 | /* Reset AFTER_RECOVERY if it has been set by the 1st scheduler pass. */ | |
7573 | after_recovery = 0; | |
7574 | can_issue_more = issue_rate; | |
7575 | ||
7576 | sched_extend_target (); | |
7577 | sched_deps_init (true); | |
7578 | setup_nop_and_exit_insns (); | |
7579 | sel_extend_global_bb_info (); | |
7580 | init_lv_sets (); | |
7581 | init_hard_regs_data (); | |
7582 | } | |
7583 | ||
7584 | /* Free the global data of the scheduler. */ | |
7585 | static void | |
7586 | sel_global_finish (void) | |
7587 | { | |
7588 | free_bb_note_pool (); | |
7589 | free_lv_sets (); | |
7590 | sel_finish_global_bb_info (); | |
7591 | ||
7592 | free_regset_pool (); | |
7593 | free_nop_and_exit_insns (); | |
7594 | ||
7595 | sched_rgn_finish (); | |
7596 | sched_deps_finish (); | |
7597 | sched_finish (); | |
7598 | ||
7599 | if (current_loops) | |
7600 | sel_finish_pipelining (); | |
7601 | ||
7602 | free_sched_pools (); | |
7603 | free_dominance_info (CDI_DOMINATORS); | |
7604 | } | |
7605 | ||
7606 | /* Return true when we need to skip selective scheduling. Used for debugging. */ | |
7607 | bool | |
7608 | maybe_skip_selective_scheduling (void) | |
7609 | { | |
7610 | return ! dbg_cnt (sel_sched_cnt); | |
7611 | } | |
7612 | ||
7613 | /* The entry point. */ | |
7614 | void | |
7615 | run_selective_scheduling (void) | |
7616 | { | |
7617 | int rgn; | |
7618 | ||
7619 | if (n_basic_blocks == NUM_FIXED_BLOCKS) | |
7620 | return; | |
7621 | ||
7622 | sel_global_init (); | |
7623 | ||
7624 | for (rgn = 0; rgn < nr_regions; rgn++) | |
7625 | sel_sched_region (rgn); | |
7626 | ||
7627 | sel_global_finish (); | |
7628 | } | |
7629 | ||
7630 | #endif |