2 Copyright (C) 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
24 #include "coretypes.h"
28 #include "hard-reg-set.h"
31 #include "insn-config.h"
33 #include "basic-block.h"
37 #include "tree-pass.h"
40 /* We want target macros for the mode switching code to be able to refer
41 to instruction attribute values. */
42 #include "insn-attr.h"
44 #ifdef OPTIMIZE_MODE_SWITCHING
46 /* The algorithm for setting the modes consists of scanning the insn list
47 and finding all the insns which require a specific mode. Each insn gets
48 a unique struct seginfo element. These structures are inserted into a list
49 for each basic block. For each entity, there is an array of bb_info over
50 the flow graph basic blocks (local var 'bb_info'), and contains a list
51 of all insns within that basic block, in the order they are encountered.
53 For each entity, any basic block WITHOUT any insns requiring a specific
54 mode are given a single entry, without a mode. (Each basic block
55 in the flow graph must have at least one entry in the segment table.)
57 The LCM algorithm is then run over the flow graph to determine where to
58 place the sets to the highest-priority value in respect of first the first
59 insn in any one block. Any adjustments required to the transparency
60 vectors are made, then the next iteration starts for the next-lower
61 priority mode, till for each entity all modes are exhausted.
63 More details are located in the code for optimize_mode_switching(). */
65 /* This structure contains the information for each insn which requires
66 either single or double mode to be set.
67 MODE is the mode this insn must be executed in.
68 INSN_PTR is the insn to be executed (may be the note that marks the
69 beginning of a basic block).
70 BBNUM is the flow graph basic block this insn occurs in.
71 NEXT is the next insn in the same basic block. */
78 HARD_REG_SET regs_live
;
83 struct seginfo
*seginfo
;
87 /* These bitmaps are used for the LCM algorithm. */
89 static sbitmap
*antic
;
90 static sbitmap
*transp
;
93 static struct seginfo
* new_seginfo (int, rtx
, int, HARD_REG_SET
);
94 static void add_seginfo (struct bb_info
*, struct seginfo
*);
95 static void reg_dies (rtx
, HARD_REG_SET
);
96 static void reg_becomes_live (rtx
, rtx
, void *);
97 static void make_preds_opaque (basic_block
, int);
100 /* This function will allocate a new BBINFO structure, initialized
101 with the MODE, INSN, and basic block BB parameters. */
103 static struct seginfo
*
104 new_seginfo (int mode
, rtx insn
, int bb
, HARD_REG_SET regs_live
)
107 ptr
= XNEW (struct seginfo
);
109 ptr
->insn_ptr
= insn
;
112 COPY_HARD_REG_SET (ptr
->regs_live
, regs_live
);
116 /* Add a seginfo element to the end of a list.
117 HEAD is a pointer to the list beginning.
118 INFO is the structure to be linked in. */
121 add_seginfo (struct bb_info
*head
, struct seginfo
*info
)
125 if (head
->seginfo
== NULL
)
126 head
->seginfo
= info
;
130 while (ptr
->next
!= NULL
)
136 /* Make all predecessors of basic block B opaque, recursively, till we hit
137 some that are already non-transparent, or an edge where aux is set; that
138 denotes that a mode set is to be done on that edge.
139 J is the bit number in the bitmaps that corresponds to the entity that
140 we are currently handling mode-switching for. */
143 make_preds_opaque (basic_block b
, int j
)
148 FOR_EACH_EDGE (e
, ei
, b
->preds
)
150 basic_block pb
= e
->src
;
152 if (e
->aux
|| ! TEST_BIT (transp
[pb
->index
], j
))
155 RESET_BIT (transp
[pb
->index
], j
);
156 make_preds_opaque (pb
, j
);
160 /* Record in LIVE that register REG died. */
163 reg_dies (rtx reg
, HARD_REG_SET live
)
171 if (regno
< FIRST_PSEUDO_REGISTER
)
172 for (nregs
= hard_regno_nregs
[regno
][GET_MODE (reg
)] - 1; nregs
>= 0;
174 CLEAR_HARD_REG_BIT (live
, regno
+ nregs
);
177 /* Record in LIVE that register REG became live.
178 This is called via note_stores. */
181 reg_becomes_live (rtx reg
, rtx setter ATTRIBUTE_UNUSED
, void *live
)
185 if (GET_CODE (reg
) == SUBREG
)
186 reg
= SUBREG_REG (reg
);
192 if (regno
< FIRST_PSEUDO_REGISTER
)
193 for (nregs
= hard_regno_nregs
[regno
][GET_MODE (reg
)] - 1; nregs
>= 0;
195 SET_HARD_REG_BIT (* (HARD_REG_SET
*) live
, regno
+ nregs
);
198 /* Make sure if MODE_ENTRY is defined the MODE_EXIT is defined
200 #if defined (MODE_ENTRY) != defined (MODE_EXIT)
201 #error "Both MODE_ENTRY and MODE_EXIT must be defined"
204 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
205 /* Split the fallthrough edge to the exit block, so that we can note
206 that there NORMAL_MODE is required. Return the new block if it's
207 inserted before the exit block. Otherwise return null. */
210 create_pre_exit (int n_entities
, int *entity_map
, const int *num_modes
)
214 basic_block pre_exit
;
216 /* The only non-call predecessor at this stage is a block with a
217 fallthrough edge; there can be at most one, but there could be
218 none at all, e.g. when exit is called. */
220 FOR_EACH_EDGE (eg
, ei
, EXIT_BLOCK_PTR
->preds
)
221 if (eg
->flags
& EDGE_FALLTHRU
)
223 basic_block src_bb
= eg
->src
;
224 regset live_at_end
= src_bb
->il
.rtl
->global_live_at_end
;
225 rtx last_insn
, ret_reg
;
227 gcc_assert (!pre_exit
);
228 /* If this function returns a value at the end, we have to
229 insert the final mode switch before the return value copy
230 to its hard register. */
231 if (EDGE_COUNT (EXIT_BLOCK_PTR
->preds
) == 1
232 && NONJUMP_INSN_P ((last_insn
= BB_END (src_bb
)))
233 && GET_CODE (PATTERN (last_insn
)) == USE
234 && GET_CODE ((ret_reg
= XEXP (PATTERN (last_insn
), 0))) == REG
)
236 int ret_start
= REGNO (ret_reg
);
237 int nregs
= hard_regno_nregs
[ret_start
][GET_MODE (ret_reg
)];
238 int ret_end
= ret_start
+ nregs
;
240 int maybe_builtin_apply
= 0;
241 int forced_late_switch
= 0;
242 rtx before_return_copy
;
246 rtx return_copy
= PREV_INSN (last_insn
);
247 rtx return_copy_pat
, copy_reg
;
248 int copy_start
, copy_num
;
251 if (INSN_P (return_copy
))
253 if (GET_CODE (PATTERN (return_copy
)) == USE
254 && GET_CODE (XEXP (PATTERN (return_copy
), 0)) == REG
255 && (FUNCTION_VALUE_REGNO_P
256 (REGNO (XEXP (PATTERN (return_copy
), 0)))))
258 maybe_builtin_apply
= 1;
259 last_insn
= return_copy
;
262 /* If the return register is not (in its entirety)
263 likely spilled, the return copy might be
264 partially or completely optimized away. */
265 return_copy_pat
= single_set (return_copy
);
266 if (!return_copy_pat
)
268 return_copy_pat
= PATTERN (return_copy
);
269 if (GET_CODE (return_copy_pat
) != CLOBBER
)
272 copy_reg
= SET_DEST (return_copy_pat
);
273 if (GET_CODE (copy_reg
) == REG
)
274 copy_start
= REGNO (copy_reg
);
275 else if (GET_CODE (copy_reg
) == SUBREG
276 && GET_CODE (SUBREG_REG (copy_reg
)) == REG
)
277 copy_start
= REGNO (SUBREG_REG (copy_reg
));
280 if (copy_start
>= FIRST_PSEUDO_REGISTER
)
283 = hard_regno_nregs
[copy_start
][GET_MODE (copy_reg
)];
285 /* If the return register is not likely spilled, - as is
286 the case for floating point on SH4 - then it might
287 be set by an arithmetic operation that needs a
288 different mode than the exit block. */
289 for (j
= n_entities
- 1; j
>= 0; j
--)
291 int e
= entity_map
[j
];
292 int mode
= MODE_NEEDED (e
, return_copy
);
294 if (mode
!= num_modes
[e
] && mode
!= MODE_EXIT (e
))
299 /* For the SH4, floating point loads depend on fpscr,
300 thus we might need to put the final mode switch
301 after the return value copy. That is still OK,
302 because a floating point return value does not
303 conflict with address reloads. */
304 if (copy_start
>= ret_start
305 && copy_start
+ copy_num
<= ret_end
306 && OBJECT_P (SET_SRC (return_copy_pat
)))
307 forced_late_switch
= 1;
311 if (copy_start
>= ret_start
312 && copy_start
+ copy_num
<= ret_end
)
314 else if (!maybe_builtin_apply
315 || !FUNCTION_VALUE_REGNO_P (copy_start
))
317 last_insn
= return_copy
;
319 /* ??? Exception handling can lead to the return value
320 copy being already separated from the return value use,
322 Similarly, conditionally returning without a value,
323 and conditionally using builtin_return can lead to an
325 if (return_copy
== BB_HEAD (src_bb
))
330 last_insn
= return_copy
;
334 /* If we didn't see a full return value copy, verify that there
335 is a plausible reason for this. If some, but not all of the
336 return register is likely spilled, we can expect that there
337 is a copy for the likely spilled part. */
339 || forced_late_switch
341 || !(CLASS_LIKELY_SPILLED_P
342 (REGNO_REG_CLASS (ret_start
)))
344 != hard_regno_nregs
[ret_start
][GET_MODE (ret_reg
)])
345 /* For multi-hard-register floating point
346 values, sometimes the likely-spilled part
347 is ordinarily copied first, then the other
348 part is set with an arithmetic operation.
349 This doesn't actually cause reload
350 failures, so let it pass. */
351 || (GET_MODE_CLASS (GET_MODE (ret_reg
)) != MODE_INT
354 if (INSN_P (last_insn
))
357 = emit_note_before (NOTE_INSN_DELETED
, last_insn
);
358 /* Instructions preceding LAST_INSN in the same block might
359 require a different mode than MODE_EXIT, so if we might
360 have such instructions, keep them in a separate block
362 if (last_insn
!= BB_HEAD (src_bb
))
363 src_bb
= split_block (src_bb
,
364 PREV_INSN (before_return_copy
))->dest
;
367 before_return_copy
= last_insn
;
368 pre_exit
= split_block (src_bb
, before_return_copy
)->src
;
372 pre_exit
= split_edge (eg
);
373 COPY_REG_SET (pre_exit
->il
.rtl
->global_live_at_start
, live_at_end
);
374 COPY_REG_SET (pre_exit
->il
.rtl
->global_live_at_end
, live_at_end
);
382 /* Find all insns that need a particular mode setting, and insert the
383 necessary mode switches. Return true if we did work. */
386 optimize_mode_switching (void)
393 struct edge_list
*edge_list
;
394 static const int num_modes
[] = NUM_MODES_FOR_MODE_SWITCHING
;
395 #define N_ENTITIES ARRAY_SIZE (num_modes)
396 int entity_map
[N_ENTITIES
];
397 struct bb_info
*bb_info
[N_ENTITIES
];
400 int max_num_modes
= 0;
402 basic_block post_entry ATTRIBUTE_UNUSED
, pre_exit ATTRIBUTE_UNUSED
;
406 for (e
= N_ENTITIES
- 1, n_entities
= 0; e
>= 0; e
--)
407 if (OPTIMIZE_MODE_SWITCHING (e
))
409 int entry_exit_extra
= 0;
411 /* Create the list of segments within each basic block.
412 If NORMAL_MODE is defined, allow for two extra
413 blocks split from the entry and exit block. */
414 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
415 entry_exit_extra
= 3;
418 = XCNEWVEC (struct bb_info
, last_basic_block
+ entry_exit_extra
);
419 entity_map
[n_entities
++] = e
;
420 if (num_modes
[e
] > max_num_modes
)
421 max_num_modes
= num_modes
[e
];
427 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
428 /* Split the edge from the entry block, so that we can note that
429 there NORMAL_MODE is supplied. */
430 post_entry
= split_edge (single_succ_edge (ENTRY_BLOCK_PTR
));
431 pre_exit
= create_pre_exit (n_entities
, entity_map
, num_modes
);
434 /* Create the bitmap vectors. */
436 antic
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
437 transp
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
438 comp
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
440 sbitmap_vector_ones (transp
, last_basic_block
);
442 for (j
= n_entities
- 1; j
>= 0; j
--)
444 int e
= entity_map
[j
];
445 int no_mode
= num_modes
[e
];
446 struct bb_info
*info
= bb_info
[j
];
448 /* Determine what the first use (if any) need for a mode of entity E is.
449 This will be the mode that is anticipatable for this block.
450 Also compute the initial transparency settings. */
454 int last_mode
= no_mode
;
455 HARD_REG_SET live_now
;
457 REG_SET_TO_HARD_REG_SET (live_now
,
458 bb
->il
.rtl
->global_live_at_start
);
460 /* Pretend the mode is clobbered across abnormal edges. */
464 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
465 if (e
->flags
& EDGE_COMPLEX
)
468 RESET_BIT (transp
[bb
->index
], j
);
471 for (insn
= BB_HEAD (bb
);
472 insn
!= NULL
&& insn
!= NEXT_INSN (BB_END (bb
));
473 insn
= NEXT_INSN (insn
))
477 int mode
= MODE_NEEDED (e
, insn
);
480 if (mode
!= no_mode
&& mode
!= last_mode
)
483 ptr
= new_seginfo (mode
, insn
, bb
->index
, live_now
);
484 add_seginfo (info
+ bb
->index
, ptr
);
485 RESET_BIT (transp
[bb
->index
], j
);
488 last_mode
= MODE_AFTER (last_mode
, insn
);
490 /* Update LIVE_NOW. */
491 for (link
= REG_NOTES (insn
); link
; link
= XEXP (link
, 1))
492 if (REG_NOTE_KIND (link
) == REG_DEAD
)
493 reg_dies (XEXP (link
, 0), live_now
);
495 note_stores (PATTERN (insn
), reg_becomes_live
, &live_now
);
496 for (link
= REG_NOTES (insn
); link
; link
= XEXP (link
, 1))
497 if (REG_NOTE_KIND (link
) == REG_UNUSED
)
498 reg_dies (XEXP (link
, 0), live_now
);
502 info
[bb
->index
].computing
= last_mode
;
503 /* Check for blocks without ANY mode requirements. */
504 if (last_mode
== no_mode
)
506 ptr
= new_seginfo (no_mode
, BB_END (bb
), bb
->index
, live_now
);
507 add_seginfo (info
+ bb
->index
, ptr
);
510 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
512 int mode
= MODE_ENTRY (e
);
518 /* By always making this nontransparent, we save
519 an extra check in make_preds_opaque. We also
520 need this to avoid confusing pre_edge_lcm when
521 antic is cleared but transp and comp are set. */
522 RESET_BIT (transp
[bb
->index
], j
);
524 /* Insert a fake computing definition of MODE into entry
525 blocks which compute no mode. This represents the mode on
527 info
[bb
->index
].computing
= mode
;
530 info
[pre_exit
->index
].seginfo
->mode
= MODE_EXIT (e
);
533 #endif /* NORMAL_MODE */
536 kill
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
537 for (i
= 0; i
< max_num_modes
; i
++)
539 int current_mode
[N_ENTITIES
];
543 /* Set the anticipatable and computing arrays. */
544 sbitmap_vector_zero (antic
, last_basic_block
);
545 sbitmap_vector_zero (comp
, last_basic_block
);
546 for (j
= n_entities
- 1; j
>= 0; j
--)
548 int m
= current_mode
[j
] = MODE_PRIORITY_TO_MODE (entity_map
[j
], i
);
549 struct bb_info
*info
= bb_info
[j
];
553 if (info
[bb
->index
].seginfo
->mode
== m
)
554 SET_BIT (antic
[bb
->index
], j
);
556 if (info
[bb
->index
].computing
== m
)
557 SET_BIT (comp
[bb
->index
], j
);
561 /* Calculate the optimal locations for the
562 placement mode switches to modes with priority I. */
565 sbitmap_not (kill
[bb
->index
], transp
[bb
->index
]);
566 edge_list
= pre_edge_lcm (n_entities
, transp
, comp
, antic
,
567 kill
, &insert
, &delete);
569 for (j
= n_entities
- 1; j
>= 0; j
--)
571 /* Insert all mode sets that have been inserted by lcm. */
572 int no_mode
= num_modes
[entity_map
[j
]];
574 /* Wherever we have moved a mode setting upwards in the flow graph,
575 the blocks between the new setting site and the now redundant
576 computation ceases to be transparent for any lower-priority
577 mode of the same entity. First set the aux field of each
578 insertion site edge non-transparent, then propagate the new
579 non-transparency from the redundant computation upwards till
580 we hit an insertion site or an already non-transparent block. */
581 for (e
= NUM_EDGES (edge_list
) - 1; e
>= 0; e
--)
583 edge eg
= INDEX_EDGE (edge_list
, e
);
586 HARD_REG_SET live_at_edge
;
591 if (! TEST_BIT (insert
[e
], j
))
596 mode
= current_mode
[j
];
599 REG_SET_TO_HARD_REG_SET (live_at_edge
,
600 src_bb
->il
.rtl
->global_live_at_end
);
603 EMIT_MODE_SET (entity_map
[j
], mode
, live_at_edge
);
604 mode_set
= get_insns ();
607 /* Do not bother to insert empty sequence. */
608 if (mode_set
== NULL_RTX
)
611 /* If this is an abnormal edge, we'll insert at the end
612 of the previous block. */
613 if (eg
->flags
& EDGE_ABNORMAL
)
616 if (JUMP_P (BB_END (src_bb
)))
617 emit_insn_before (mode_set
, BB_END (src_bb
));
620 /* It doesn't make sense to switch to normal
621 mode after a CALL_INSN. The cases in which a
622 CALL_INSN may have an abnormal edge are
623 sibcalls and EH edges. In the case of
624 sibcalls, the dest basic-block is the
625 EXIT_BLOCK, that runs in normal mode; it is
626 assumed that a sibcall insn requires normal
627 mode itself, so no mode switch would be
628 required after the call (it wouldn't make
629 sense, anyway). In the case of EH edges, EH
630 entry points also start in normal mode, so a
631 similar reasoning applies. */
632 gcc_assert (NONJUMP_INSN_P (BB_END (src_bb
)));
633 emit_insn_after (mode_set
, BB_END (src_bb
));
635 bb_info
[j
][src_bb
->index
].computing
= mode
;
636 RESET_BIT (transp
[src_bb
->index
], j
);
641 insert_insn_on_edge (mode_set
, eg
);
645 FOR_EACH_BB_REVERSE (bb
)
646 if (TEST_BIT (delete[bb
->index
], j
))
648 make_preds_opaque (bb
, j
);
649 /* Cancel the 'deleted' mode set. */
650 bb_info
[j
][bb
->index
].seginfo
->mode
= no_mode
;
654 sbitmap_vector_free (delete);
655 sbitmap_vector_free (insert
);
656 clear_aux_for_edges ();
657 free_edge_list (edge_list
);
660 /* Now output the remaining mode sets in all the segments. */
661 for (j
= n_entities
- 1; j
>= 0; j
--)
663 int no_mode
= num_modes
[entity_map
[j
]];
665 FOR_EACH_BB_REVERSE (bb
)
667 struct seginfo
*ptr
, *next
;
668 for (ptr
= bb_info
[j
][bb
->index
].seginfo
; ptr
; ptr
= next
)
671 if (ptr
->mode
!= no_mode
)
676 EMIT_MODE_SET (entity_map
[j
], ptr
->mode
, ptr
->regs_live
);
677 mode_set
= get_insns ();
680 /* Insert MODE_SET only if it is nonempty. */
681 if (mode_set
!= NULL_RTX
)
684 if (NOTE_P (ptr
->insn_ptr
)
685 && (NOTE_LINE_NUMBER (ptr
->insn_ptr
)
686 == NOTE_INSN_BASIC_BLOCK
))
687 emit_insn_after (mode_set
, ptr
->insn_ptr
);
689 emit_insn_before (mode_set
, ptr
->insn_ptr
);
700 /* Finished. Free up all the things we've allocated. */
702 sbitmap_vector_free (kill
);
703 sbitmap_vector_free (antic
);
704 sbitmap_vector_free (transp
);
705 sbitmap_vector_free (comp
);
708 commit_edge_insertions ();
710 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
711 cleanup_cfg (CLEANUP_NO_INSN_DEL
);
713 if (!need_commit
&& !emited
)
717 max_regno
= max_reg_num ();
718 allocate_reg_info (max_regno
, FALSE
, FALSE
);
719 update_life_info_in_dirty_blocks (UPDATE_LIFE_GLOBAL_RM_NOTES
,
720 (PROP_DEATH_NOTES
| PROP_KILL_DEAD_CODE
721 | PROP_SCAN_DEAD_CODE
));
726 #endif /* OPTIMIZE_MODE_SWITCHING */
729 gate_mode_switching (void)
731 #ifdef OPTIMIZE_MODE_SWITCHING
739 rest_of_handle_mode_switching (void)
741 #ifdef OPTIMIZE_MODE_SWITCHING
743 optimize_mode_switching ();
745 #endif /* OPTIMIZE_MODE_SWITCHING */
750 struct tree_opt_pass pass_mode_switching
=
752 "mode-sw", /* name */
753 gate_mode_switching
, /* gate */
754 rest_of_handle_mode_switching
, /* execute */
757 0, /* static_pass_number */
758 TV_MODE_SWITCH
, /* tv_id */
759 0, /* properties_required */
760 0, /* properties_provided */
761 0, /* properties_destroyed */
762 0, /* todo_flags_start */
763 TODO_dump_func
, /* todo_flags_finish */