2 Copyright (C) 1998-2013 Free Software Foundation, Inc.
4 This file is part of GCC.
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
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
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/>. */
22 #include "coretypes.h"
27 #include "hard-reg-set.h"
29 #include "insn-config.h"
31 #include "basic-block.h"
34 #include "tree-pass.h"
38 /* We want target macros for the mode switching code to be able to refer
39 to instruction attribute values. */
40 #include "insn-attr.h"
42 #ifdef OPTIMIZE_MODE_SWITCHING
44 /* The algorithm for setting the modes consists of scanning the insn list
45 and finding all the insns which require a specific mode. Each insn gets
46 a unique struct seginfo element. These structures are inserted into a list
47 for each basic block. For each entity, there is an array of bb_info over
48 the flow graph basic blocks (local var 'bb_info'), and contains a list
49 of all insns within that basic block, in the order they are encountered.
51 For each entity, any basic block WITHOUT any insns requiring a specific
52 mode are given a single entry, without a mode. (Each basic block
53 in the flow graph must have at least one entry in the segment table.)
55 The LCM algorithm is then run over the flow graph to determine where to
56 place the sets to the highest-priority value in respect of first the first
57 insn in any one block. Any adjustments required to the transparency
58 vectors are made, then the next iteration starts for the next-lower
59 priority mode, till for each entity all modes are exhausted.
61 More details are located in the code for optimize_mode_switching(). */
63 /* This structure contains the information for each insn which requires
64 either single or double mode to be set.
65 MODE is the mode this insn must be executed in.
66 INSN_PTR is the insn to be executed (may be the note that marks the
67 beginning of a basic block).
68 BBNUM is the flow graph basic block this insn occurs in.
69 NEXT is the next insn in the same basic block. */
76 HARD_REG_SET regs_live
;
81 struct seginfo
*seginfo
;
85 /* These bitmaps are used for the LCM algorithm. */
87 static sbitmap
*antic
;
88 static sbitmap
*transp
;
91 static struct seginfo
* new_seginfo (int, rtx
, int, HARD_REG_SET
);
92 static void add_seginfo (struct bb_info
*, struct seginfo
*);
93 static void reg_dies (rtx
, HARD_REG_SET
*);
94 static void reg_becomes_live (rtx
, const_rtx
, void *);
95 static void make_preds_opaque (basic_block
, int);
98 /* This function will allocate a new BBINFO structure, initialized
99 with the MODE, INSN, and basic block BB parameters. */
101 static struct seginfo
*
102 new_seginfo (int mode
, rtx insn
, int bb
, HARD_REG_SET regs_live
)
105 ptr
= XNEW (struct seginfo
);
107 ptr
->insn_ptr
= insn
;
110 COPY_HARD_REG_SET (ptr
->regs_live
, regs_live
);
114 /* Add a seginfo element to the end of a list.
115 HEAD is a pointer to the list beginning.
116 INFO is the structure to be linked in. */
119 add_seginfo (struct bb_info
*head
, struct seginfo
*info
)
123 if (head
->seginfo
== NULL
)
124 head
->seginfo
= info
;
128 while (ptr
->next
!= NULL
)
134 /* Make all predecessors of basic block B opaque, recursively, till we hit
135 some that are already non-transparent, or an edge where aux is set; that
136 denotes that a mode set is to be done on that edge.
137 J is the bit number in the bitmaps that corresponds to the entity that
138 we are currently handling mode-switching for. */
141 make_preds_opaque (basic_block b
, int j
)
146 FOR_EACH_EDGE (e
, ei
, b
->preds
)
148 basic_block pb
= e
->src
;
150 if (e
->aux
|| ! bitmap_bit_p (transp
[pb
->index
], j
))
153 bitmap_clear_bit (transp
[pb
->index
], j
);
154 make_preds_opaque (pb
, j
);
158 /* Record in LIVE that register REG died. */
161 reg_dies (rtx reg
, HARD_REG_SET
*live
)
169 if (regno
< FIRST_PSEUDO_REGISTER
)
170 remove_from_hard_reg_set (live
, GET_MODE (reg
), regno
);
173 /* Record in LIVE that register REG became live.
174 This is called via note_stores. */
177 reg_becomes_live (rtx reg
, const_rtx setter ATTRIBUTE_UNUSED
, void *live
)
181 if (GET_CODE (reg
) == SUBREG
)
182 reg
= SUBREG_REG (reg
);
188 if (regno
< FIRST_PSEUDO_REGISTER
)
189 add_to_hard_reg_set ((HARD_REG_SET
*) live
, GET_MODE (reg
), regno
);
192 /* Make sure if MODE_ENTRY is defined the MODE_EXIT is defined
194 #if defined (MODE_ENTRY) != defined (MODE_EXIT)
195 #error "Both MODE_ENTRY and MODE_EXIT must be defined"
198 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
199 /* Split the fallthrough edge to the exit block, so that we can note
200 that there NORMAL_MODE is required. Return the new block if it's
201 inserted before the exit block. Otherwise return null. */
204 create_pre_exit (int n_entities
, int *entity_map
, const int *num_modes
)
208 basic_block pre_exit
;
210 /* The only non-call predecessor at this stage is a block with a
211 fallthrough edge; there can be at most one, but there could be
212 none at all, e.g. when exit is called. */
214 FOR_EACH_EDGE (eg
, ei
, EXIT_BLOCK_PTR
->preds
)
215 if (eg
->flags
& EDGE_FALLTHRU
)
217 basic_block src_bb
= eg
->src
;
218 rtx last_insn
, ret_reg
;
220 gcc_assert (!pre_exit
);
221 /* If this function returns a value at the end, we have to
222 insert the final mode switch before the return value copy
223 to its hard register. */
224 if (EDGE_COUNT (EXIT_BLOCK_PTR
->preds
) == 1
225 && NONJUMP_INSN_P ((last_insn
= BB_END (src_bb
)))
226 && GET_CODE (PATTERN (last_insn
)) == USE
227 && GET_CODE ((ret_reg
= XEXP (PATTERN (last_insn
), 0))) == REG
)
229 int ret_start
= REGNO (ret_reg
);
230 int nregs
= hard_regno_nregs
[ret_start
][GET_MODE (ret_reg
)];
231 int ret_end
= ret_start
+ nregs
;
233 int maybe_builtin_apply
= 0;
234 int forced_late_switch
= 0;
235 rtx before_return_copy
;
239 rtx return_copy
= PREV_INSN (last_insn
);
240 rtx return_copy_pat
, copy_reg
;
241 int copy_start
, copy_num
;
244 if (NONDEBUG_INSN_P (return_copy
))
246 /* When using SJLJ exceptions, the call to the
247 unregister function is inserted between the
248 clobber of the return value and the copy.
249 We do not want to split the block before this
250 or any other call; if we have not found the
251 copy yet, the copy must have been deleted. */
252 if (CALL_P (return_copy
))
257 return_copy_pat
= PATTERN (return_copy
);
258 switch (GET_CODE (return_copy_pat
))
261 /* Skip __builtin_apply pattern. */
262 if (GET_CODE (XEXP (return_copy_pat
, 0)) == REG
263 && (targetm
.calls
.function_value_regno_p
264 (REGNO (XEXP (return_copy_pat
, 0)))))
266 maybe_builtin_apply
= 1;
267 last_insn
= return_copy
;
273 /* Skip barrier insns. */
274 if (!MEM_VOLATILE_P (return_copy_pat
))
280 case UNSPEC_VOLATILE
:
281 last_insn
= return_copy
;
288 /* If the return register is not (in its entirety)
289 likely spilled, the return copy might be
290 partially or completely optimized away. */
291 return_copy_pat
= single_set (return_copy
);
292 if (!return_copy_pat
)
294 return_copy_pat
= PATTERN (return_copy
);
295 if (GET_CODE (return_copy_pat
) != CLOBBER
)
299 /* This might be (clobber (reg [<result>]))
300 when not optimizing. Then check if
301 the previous insn is the clobber for
302 the return register. */
303 copy_reg
= SET_DEST (return_copy_pat
);
304 if (GET_CODE (copy_reg
) == REG
305 && !HARD_REGISTER_NUM_P (REGNO (copy_reg
)))
307 if (INSN_P (PREV_INSN (return_copy
)))
309 return_copy
= PREV_INSN (return_copy
);
310 return_copy_pat
= PATTERN (return_copy
);
311 if (GET_CODE (return_copy_pat
) != CLOBBER
)
317 copy_reg
= SET_DEST (return_copy_pat
);
318 if (GET_CODE (copy_reg
) == REG
)
319 copy_start
= REGNO (copy_reg
);
320 else if (GET_CODE (copy_reg
) == SUBREG
321 && GET_CODE (SUBREG_REG (copy_reg
)) == REG
)
322 copy_start
= REGNO (SUBREG_REG (copy_reg
));
325 /* When control reaches end of non-void function,
326 there are no return copy insns at all. This
327 avoids an ice on that invalid function. */
328 if (ret_start
+ nregs
== ret_end
)
332 if (!targetm
.calls
.function_value_regno_p (copy_start
))
334 last_insn
= return_copy
;
338 = hard_regno_nregs
[copy_start
][GET_MODE (copy_reg
)];
340 /* If the return register is not likely spilled, - as is
341 the case for floating point on SH4 - then it might
342 be set by an arithmetic operation that needs a
343 different mode than the exit block. */
344 for (j
= n_entities
- 1; j
>= 0; j
--)
346 int e
= entity_map
[j
];
347 int mode
= MODE_NEEDED (e
, return_copy
);
349 if (mode
!= num_modes
[e
] && mode
!= MODE_EXIT (e
))
354 /* __builtin_return emits a sequence of loads to all
355 return registers. One of them might require
356 another mode than MODE_EXIT, even if it is
357 unrelated to the return value, so we want to put
358 the final mode switch after it. */
359 if (maybe_builtin_apply
360 && targetm
.calls
.function_value_regno_p
362 forced_late_switch
= 1;
364 /* For the SH4, floating point loads depend on fpscr,
365 thus we might need to put the final mode switch
366 after the return value copy. That is still OK,
367 because a floating point return value does not
368 conflict with address reloads. */
369 if (copy_start
>= ret_start
370 && copy_start
+ copy_num
<= ret_end
371 && OBJECT_P (SET_SRC (return_copy_pat
)))
372 forced_late_switch
= 1;
376 if (copy_start
>= ret_start
377 && copy_start
+ copy_num
<= ret_end
)
379 else if (!maybe_builtin_apply
380 || !targetm
.calls
.function_value_regno_p
383 last_insn
= return_copy
;
385 /* ??? Exception handling can lead to the return value
386 copy being already separated from the return value use,
388 Similarly, conditionally returning without a value,
389 and conditionally using builtin_return can lead to an
391 if (return_copy
== BB_HEAD (src_bb
))
396 last_insn
= return_copy
;
400 /* If we didn't see a full return value copy, verify that there
401 is a plausible reason for this. If some, but not all of the
402 return register is likely spilled, we can expect that there
403 is a copy for the likely spilled part. */
405 || forced_late_switch
407 || !(targetm
.class_likely_spilled_p
408 (REGNO_REG_CLASS (ret_start
)))
410 != hard_regno_nregs
[ret_start
][GET_MODE (ret_reg
)])
411 /* For multi-hard-register floating point
412 values, sometimes the likely-spilled part
413 is ordinarily copied first, then the other
414 part is set with an arithmetic operation.
415 This doesn't actually cause reload
416 failures, so let it pass. */
417 || (GET_MODE_CLASS (GET_MODE (ret_reg
)) != MODE_INT
420 if (INSN_P (last_insn
))
423 = emit_note_before (NOTE_INSN_DELETED
, last_insn
);
424 /* Instructions preceding LAST_INSN in the same block might
425 require a different mode than MODE_EXIT, so if we might
426 have such instructions, keep them in a separate block
428 if (last_insn
!= BB_HEAD (src_bb
))
429 src_bb
= split_block (src_bb
,
430 PREV_INSN (before_return_copy
))->dest
;
433 before_return_copy
= last_insn
;
434 pre_exit
= split_block (src_bb
, before_return_copy
)->src
;
438 pre_exit
= split_edge (eg
);
446 /* Find all insns that need a particular mode setting, and insert the
447 necessary mode switches. Return true if we did work. */
450 optimize_mode_switching (void)
457 struct edge_list
*edge_list
;
458 static const int num_modes
[] = NUM_MODES_FOR_MODE_SWITCHING
;
459 #define N_ENTITIES ARRAY_SIZE (num_modes)
460 int entity_map
[N_ENTITIES
];
461 struct bb_info
*bb_info
[N_ENTITIES
];
464 int max_num_modes
= 0;
465 bool emitted ATTRIBUTE_UNUSED
= false;
466 basic_block post_entry ATTRIBUTE_UNUSED
, pre_exit ATTRIBUTE_UNUSED
;
468 for (e
= N_ENTITIES
- 1, n_entities
= 0; e
>= 0; e
--)
469 if (OPTIMIZE_MODE_SWITCHING (e
))
471 int entry_exit_extra
= 0;
473 /* Create the list of segments within each basic block.
474 If NORMAL_MODE is defined, allow for two extra
475 blocks split from the entry and exit block. */
476 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
477 entry_exit_extra
= 3;
480 = XCNEWVEC (struct bb_info
, last_basic_block
+ entry_exit_extra
);
481 entity_map
[n_entities
++] = e
;
482 if (num_modes
[e
] > max_num_modes
)
483 max_num_modes
= num_modes
[e
];
489 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
490 /* Split the edge from the entry block, so that we can note that
491 there NORMAL_MODE is supplied. */
492 post_entry
= split_edge (single_succ_edge (ENTRY_BLOCK_PTR
));
493 pre_exit
= create_pre_exit (n_entities
, entity_map
, num_modes
);
498 /* Create the bitmap vectors. */
500 antic
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
501 transp
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
502 comp
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
504 bitmap_vector_ones (transp
, last_basic_block
);
506 for (j
= n_entities
- 1; j
>= 0; j
--)
508 int e
= entity_map
[j
];
509 int no_mode
= num_modes
[e
];
510 struct bb_info
*info
= bb_info
[j
];
512 /* Determine what the first use (if any) need for a mode of entity E is.
513 This will be the mode that is anticipatable for this block.
514 Also compute the initial transparency settings. */
518 int last_mode
= no_mode
;
519 bool any_set_required
= false;
520 HARD_REG_SET live_now
;
522 REG_SET_TO_HARD_REG_SET (live_now
, df_get_live_in (bb
));
524 /* Pretend the mode is clobbered across abnormal edges. */
528 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
529 if (e
->flags
& EDGE_COMPLEX
)
533 ptr
= new_seginfo (no_mode
, BB_HEAD (bb
), bb
->index
, live_now
);
534 add_seginfo (info
+ bb
->index
, ptr
);
535 bitmap_clear_bit (transp
[bb
->index
], j
);
539 FOR_BB_INSNS (bb
, insn
)
543 int mode
= MODE_NEEDED (e
, insn
);
546 if (mode
!= no_mode
&& mode
!= last_mode
)
548 any_set_required
= true;
550 ptr
= new_seginfo (mode
, insn
, bb
->index
, live_now
);
551 add_seginfo (info
+ bb
->index
, ptr
);
552 bitmap_clear_bit (transp
[bb
->index
], j
);
555 last_mode
= MODE_AFTER (e
, last_mode
, insn
);
557 /* Update LIVE_NOW. */
558 for (link
= REG_NOTES (insn
); link
; link
= XEXP (link
, 1))
559 if (REG_NOTE_KIND (link
) == REG_DEAD
)
560 reg_dies (XEXP (link
, 0), &live_now
);
562 note_stores (PATTERN (insn
), reg_becomes_live
, &live_now
);
563 for (link
= REG_NOTES (insn
); link
; link
= XEXP (link
, 1))
564 if (REG_NOTE_KIND (link
) == REG_UNUSED
)
565 reg_dies (XEXP (link
, 0), &live_now
);
569 info
[bb
->index
].computing
= last_mode
;
570 /* Check for blocks without ANY mode requirements.
571 N.B. because of MODE_AFTER, last_mode might still be different
573 if (!any_set_required
)
575 ptr
= new_seginfo (no_mode
, BB_END (bb
), bb
->index
, live_now
);
576 add_seginfo (info
+ bb
->index
, ptr
);
579 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
581 int mode
= MODE_ENTRY (e
);
587 /* By always making this nontransparent, we save
588 an extra check in make_preds_opaque. We also
589 need this to avoid confusing pre_edge_lcm when
590 antic is cleared but transp and comp are set. */
591 bitmap_clear_bit (transp
[bb
->index
], j
);
593 /* Insert a fake computing definition of MODE into entry
594 blocks which compute no mode. This represents the mode on
596 info
[bb
->index
].computing
= mode
;
599 info
[pre_exit
->index
].seginfo
->mode
= MODE_EXIT (e
);
602 #endif /* NORMAL_MODE */
605 kill
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
606 for (i
= 0; i
< max_num_modes
; i
++)
608 int current_mode
[N_ENTITIES
];
612 /* Set the anticipatable and computing arrays. */
613 bitmap_vector_clear (antic
, last_basic_block
);
614 bitmap_vector_clear (comp
, last_basic_block
);
615 for (j
= n_entities
- 1; j
>= 0; j
--)
617 int m
= current_mode
[j
] = MODE_PRIORITY_TO_MODE (entity_map
[j
], i
);
618 struct bb_info
*info
= bb_info
[j
];
622 if (info
[bb
->index
].seginfo
->mode
== m
)
623 bitmap_set_bit (antic
[bb
->index
], j
);
625 if (info
[bb
->index
].computing
== m
)
626 bitmap_set_bit (comp
[bb
->index
], j
);
630 /* Calculate the optimal locations for the
631 placement mode switches to modes with priority I. */
634 bitmap_not (kill
[bb
->index
], transp
[bb
->index
]);
635 edge_list
= pre_edge_lcm (n_entities
, transp
, comp
, antic
,
636 kill
, &insert
, &del
);
638 for (j
= n_entities
- 1; j
>= 0; j
--)
640 /* Insert all mode sets that have been inserted by lcm. */
641 int no_mode
= num_modes
[entity_map
[j
]];
643 /* Wherever we have moved a mode setting upwards in the flow graph,
644 the blocks between the new setting site and the now redundant
645 computation ceases to be transparent for any lower-priority
646 mode of the same entity. First set the aux field of each
647 insertion site edge non-transparent, then propagate the new
648 non-transparency from the redundant computation upwards till
649 we hit an insertion site or an already non-transparent block. */
650 for (e
= NUM_EDGES (edge_list
) - 1; e
>= 0; e
--)
652 edge eg
= INDEX_EDGE (edge_list
, e
);
655 HARD_REG_SET live_at_edge
;
660 if (! bitmap_bit_p (insert
[e
], j
))
665 mode
= current_mode
[j
];
668 REG_SET_TO_HARD_REG_SET (live_at_edge
, df_get_live_out (src_bb
));
671 EMIT_MODE_SET (entity_map
[j
], mode
, live_at_edge
);
672 mode_set
= get_insns ();
675 /* Do not bother to insert empty sequence. */
676 if (mode_set
== NULL_RTX
)
679 /* We should not get an abnormal edge here. */
680 gcc_assert (! (eg
->flags
& EDGE_ABNORMAL
));
683 insert_insn_on_edge (mode_set
, eg
);
686 FOR_EACH_BB_REVERSE (bb
)
687 if (bitmap_bit_p (del
[bb
->index
], j
))
689 make_preds_opaque (bb
, j
);
690 /* Cancel the 'deleted' mode set. */
691 bb_info
[j
][bb
->index
].seginfo
->mode
= no_mode
;
695 sbitmap_vector_free (del
);
696 sbitmap_vector_free (insert
);
697 clear_aux_for_edges ();
698 free_edge_list (edge_list
);
701 /* Now output the remaining mode sets in all the segments. */
702 for (j
= n_entities
- 1; j
>= 0; j
--)
704 int no_mode
= num_modes
[entity_map
[j
]];
706 FOR_EACH_BB_REVERSE (bb
)
708 struct seginfo
*ptr
, *next
;
709 for (ptr
= bb_info
[j
][bb
->index
].seginfo
; ptr
; ptr
= next
)
712 if (ptr
->mode
!= no_mode
)
717 EMIT_MODE_SET (entity_map
[j
], ptr
->mode
, ptr
->regs_live
);
718 mode_set
= get_insns ();
721 /* Insert MODE_SET only if it is nonempty. */
722 if (mode_set
!= NULL_RTX
)
725 if (NOTE_INSN_BASIC_BLOCK_P (ptr
->insn_ptr
))
726 emit_insn_after (mode_set
, ptr
->insn_ptr
);
728 emit_insn_before (mode_set
, ptr
->insn_ptr
);
739 /* Finished. Free up all the things we've allocated. */
740 sbitmap_vector_free (kill
);
741 sbitmap_vector_free (antic
);
742 sbitmap_vector_free (transp
);
743 sbitmap_vector_free (comp
);
746 commit_edge_insertions ();
748 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
749 cleanup_cfg (CLEANUP_NO_INSN_DEL
);
751 if (!need_commit
&& !emitted
)
758 #endif /* OPTIMIZE_MODE_SWITCHING */
761 gate_mode_switching (void)
763 #ifdef OPTIMIZE_MODE_SWITCHING
771 rest_of_handle_mode_switching (void)
773 #ifdef OPTIMIZE_MODE_SWITCHING
774 optimize_mode_switching ();
775 #endif /* OPTIMIZE_MODE_SWITCHING */
780 struct rtl_opt_pass pass_mode_switching
=
784 "mode_sw", /* name */
785 OPTGROUP_NONE
, /* optinfo_flags */
786 gate_mode_switching
, /* gate */
787 rest_of_handle_mode_switching
, /* execute */
790 0, /* static_pass_number */
791 TV_MODE_SWITCH
, /* tv_id */
792 0, /* properties_required */
793 0, /* properties_provided */
794 0, /* properties_destroyed */
795 0, /* todo_flags_start */
796 TODO_df_finish
| TODO_verify_rtl_sharing
|
797 0 /* todo_flags_finish */