2 Copyright (C) 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2007, 2008,
3 2009, 2010 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 3, 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 COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
28 #include "hard-reg-set.h"
30 #include "insn-config.h"
32 #include "basic-block.h"
36 #include "tree-pass.h"
41 /* We want target macros for the mode switching code to be able to refer
42 to instruction attribute values. */
43 #include "insn-attr.h"
45 #ifdef OPTIMIZE_MODE_SWITCHING
47 /* The algorithm for setting the modes consists of scanning the insn list
48 and finding all the insns which require a specific mode. Each insn gets
49 a unique struct seginfo element. These structures are inserted into a list
50 for each basic block. For each entity, there is an array of bb_info over
51 the flow graph basic blocks (local var 'bb_info'), and contains a list
52 of all insns within that basic block, in the order they are encountered.
54 For each entity, any basic block WITHOUT any insns requiring a specific
55 mode are given a single entry, without a mode. (Each basic block
56 in the flow graph must have at least one entry in the segment table.)
58 The LCM algorithm is then run over the flow graph to determine where to
59 place the sets to the highest-priority value in respect of first the first
60 insn in any one block. Any adjustments required to the transparency
61 vectors are made, then the next iteration starts for the next-lower
62 priority mode, till for each entity all modes are exhausted.
64 More details are located in the code for optimize_mode_switching(). */
66 /* This structure contains the information for each insn which requires
67 either single or double mode to be set.
68 MODE is the mode this insn must be executed in.
69 INSN_PTR is the insn to be executed (may be the note that marks the
70 beginning of a basic block).
71 BBNUM is the flow graph basic block this insn occurs in.
72 NEXT is the next insn in the same basic block. */
79 HARD_REG_SET regs_live
;
84 struct seginfo
*seginfo
;
88 /* These bitmaps are used for the LCM algorithm. */
90 static sbitmap
*antic
;
91 static sbitmap
*transp
;
94 static struct seginfo
* new_seginfo (int, rtx
, int, HARD_REG_SET
);
95 static void add_seginfo (struct bb_info
*, struct seginfo
*);
96 static void reg_dies (rtx
, HARD_REG_SET
*);
97 static void reg_becomes_live (rtx
, const_rtx
, void *);
98 static void make_preds_opaque (basic_block
, int);
101 /* This function will allocate a new BBINFO structure, initialized
102 with the MODE, INSN, and basic block BB parameters. */
104 static struct seginfo
*
105 new_seginfo (int mode
, rtx insn
, int bb
, HARD_REG_SET regs_live
)
108 ptr
= XNEW (struct seginfo
);
110 ptr
->insn_ptr
= insn
;
113 COPY_HARD_REG_SET (ptr
->regs_live
, regs_live
);
117 /* Add a seginfo element to the end of a list.
118 HEAD is a pointer to the list beginning.
119 INFO is the structure to be linked in. */
122 add_seginfo (struct bb_info
*head
, struct seginfo
*info
)
126 if (head
->seginfo
== NULL
)
127 head
->seginfo
= info
;
131 while (ptr
->next
!= NULL
)
137 /* Make all predecessors of basic block B opaque, recursively, till we hit
138 some that are already non-transparent, or an edge where aux is set; that
139 denotes that a mode set is to be done on that edge.
140 J is the bit number in the bitmaps that corresponds to the entity that
141 we are currently handling mode-switching for. */
144 make_preds_opaque (basic_block b
, int j
)
149 FOR_EACH_EDGE (e
, ei
, b
->preds
)
151 basic_block pb
= e
->src
;
153 if (e
->aux
|| ! TEST_BIT (transp
[pb
->index
], j
))
156 RESET_BIT (transp
[pb
->index
], j
);
157 make_preds_opaque (pb
, j
);
161 /* Record in LIVE that register REG died. */
164 reg_dies (rtx reg
, HARD_REG_SET
*live
)
172 if (regno
< FIRST_PSEUDO_REGISTER
)
173 remove_from_hard_reg_set (live
, GET_MODE (reg
), regno
);
176 /* Record in LIVE that register REG became live.
177 This is called via note_stores. */
180 reg_becomes_live (rtx reg
, const_rtx setter ATTRIBUTE_UNUSED
, void *live
)
184 if (GET_CODE (reg
) == SUBREG
)
185 reg
= SUBREG_REG (reg
);
191 if (regno
< FIRST_PSEUDO_REGISTER
)
192 add_to_hard_reg_set ((HARD_REG_SET
*) live
, GET_MODE (reg
), regno
);
195 /* Make sure if MODE_ENTRY is defined the MODE_EXIT is defined
197 #if defined (MODE_ENTRY) != defined (MODE_EXIT)
198 #error "Both MODE_ENTRY and MODE_EXIT must be defined"
201 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
202 /* Split the fallthrough edge to the exit block, so that we can note
203 that there NORMAL_MODE is required. Return the new block if it's
204 inserted before the exit block. Otherwise return null. */
207 create_pre_exit (int n_entities
, int *entity_map
, const int *num_modes
)
211 basic_block pre_exit
;
213 /* The only non-call predecessor at this stage is a block with a
214 fallthrough edge; there can be at most one, but there could be
215 none at all, e.g. when exit is called. */
217 FOR_EACH_EDGE (eg
, ei
, EXIT_BLOCK_PTR
->preds
)
218 if (eg
->flags
& EDGE_FALLTHRU
)
220 basic_block src_bb
= eg
->src
;
221 rtx last_insn
, ret_reg
;
223 gcc_assert (!pre_exit
);
224 /* If this function returns a value at the end, we have to
225 insert the final mode switch before the return value copy
226 to its hard register. */
227 if (EDGE_COUNT (EXIT_BLOCK_PTR
->preds
) == 1
228 && NONJUMP_INSN_P ((last_insn
= BB_END (src_bb
)))
229 && GET_CODE (PATTERN (last_insn
)) == USE
230 && GET_CODE ((ret_reg
= XEXP (PATTERN (last_insn
), 0))) == REG
)
232 int ret_start
= REGNO (ret_reg
);
233 int nregs
= hard_regno_nregs
[ret_start
][GET_MODE (ret_reg
)];
234 int ret_end
= ret_start
+ nregs
;
236 int maybe_builtin_apply
= 0;
237 int forced_late_switch
= 0;
238 rtx before_return_copy
;
242 rtx return_copy
= PREV_INSN (last_insn
);
243 rtx return_copy_pat
, copy_reg
;
244 int copy_start
, copy_num
;
247 if (INSN_P (return_copy
))
249 /* When using SJLJ exceptions, the call to the
250 unregister function is inserted between the
251 clobber of the return value and the copy.
252 We do not want to split the block before this
253 or any other call; if we have not found the
254 copy yet, the copy must have been deleted. */
255 if (CALL_P (return_copy
))
260 return_copy_pat
= PATTERN (return_copy
);
261 switch (GET_CODE (return_copy_pat
))
264 /* Skip __builtin_apply pattern. */
265 if (GET_CODE (XEXP (return_copy_pat
, 0)) == REG
266 && (targetm
.calls
.function_value_regno_p
267 (REGNO (XEXP (return_copy_pat
, 0)))))
269 maybe_builtin_apply
= 1;
270 last_insn
= return_copy
;
276 /* Skip barrier insns. */
277 if (!MEM_VOLATILE_P (return_copy_pat
))
283 case UNSPEC_VOLATILE
:
284 last_insn
= return_copy
;
291 /* If the return register is not (in its entirety)
292 likely spilled, the return copy might be
293 partially or completely optimized away. */
294 return_copy_pat
= single_set (return_copy
);
295 if (!return_copy_pat
)
297 return_copy_pat
= PATTERN (return_copy
);
298 if (GET_CODE (return_copy_pat
) != CLOBBER
)
302 /* This might be (clobber (reg [<result>]))
303 when not optimizing. Then check if
304 the previous insn is the clobber for
305 the return register. */
306 copy_reg
= SET_DEST (return_copy_pat
);
307 if (GET_CODE (copy_reg
) == REG
308 && !HARD_REGISTER_NUM_P (REGNO (copy_reg
)))
310 if (INSN_P (PREV_INSN (return_copy
)))
312 return_copy
= PREV_INSN (return_copy
);
313 return_copy_pat
= PATTERN (return_copy
);
314 if (GET_CODE (return_copy_pat
) != CLOBBER
)
320 copy_reg
= SET_DEST (return_copy_pat
);
321 if (GET_CODE (copy_reg
) == REG
)
322 copy_start
= REGNO (copy_reg
);
323 else if (GET_CODE (copy_reg
) == SUBREG
324 && GET_CODE (SUBREG_REG (copy_reg
)) == REG
)
325 copy_start
= REGNO (SUBREG_REG (copy_reg
));
328 if (copy_start
>= FIRST_PSEUDO_REGISTER
)
331 = hard_regno_nregs
[copy_start
][GET_MODE (copy_reg
)];
333 /* If the return register is not likely spilled, - as is
334 the case for floating point on SH4 - then it might
335 be set by an arithmetic operation that needs a
336 different mode than the exit block. */
337 for (j
= n_entities
- 1; j
>= 0; j
--)
339 int e
= entity_map
[j
];
340 int mode
= MODE_NEEDED (e
, return_copy
);
342 if (mode
!= num_modes
[e
] && mode
!= MODE_EXIT (e
))
347 /* For the SH4, floating point loads depend on fpscr,
348 thus we might need to put the final mode switch
349 after the return value copy. That is still OK,
350 because a floating point return value does not
351 conflict with address reloads. */
352 if (copy_start
>= ret_start
353 && copy_start
+ copy_num
<= ret_end
354 && OBJECT_P (SET_SRC (return_copy_pat
)))
355 forced_late_switch
= 1;
359 if (copy_start
>= ret_start
360 && copy_start
+ copy_num
<= ret_end
)
362 else if (!maybe_builtin_apply
363 || !targetm
.calls
.function_value_regno_p
366 last_insn
= return_copy
;
368 /* ??? Exception handling can lead to the return value
369 copy being already separated from the return value use,
371 Similarly, conditionally returning without a value,
372 and conditionally using builtin_return can lead to an
374 if (return_copy
== BB_HEAD (src_bb
))
379 last_insn
= return_copy
;
383 /* If we didn't see a full return value copy, verify that there
384 is a plausible reason for this. If some, but not all of the
385 return register is likely spilled, we can expect that there
386 is a copy for the likely spilled part. */
388 || forced_late_switch
390 || !(targetm
.class_likely_spilled_p
391 (REGNO_REG_CLASS (ret_start
)))
393 != hard_regno_nregs
[ret_start
][GET_MODE (ret_reg
)])
394 /* For multi-hard-register floating point
395 values, sometimes the likely-spilled part
396 is ordinarily copied first, then the other
397 part is set with an arithmetic operation.
398 This doesn't actually cause reload
399 failures, so let it pass. */
400 || (GET_MODE_CLASS (GET_MODE (ret_reg
)) != MODE_INT
403 if (INSN_P (last_insn
))
406 = emit_note_before (NOTE_INSN_DELETED
, last_insn
);
407 /* Instructions preceding LAST_INSN in the same block might
408 require a different mode than MODE_EXIT, so if we might
409 have such instructions, keep them in a separate block
411 if (last_insn
!= BB_HEAD (src_bb
))
412 src_bb
= split_block (src_bb
,
413 PREV_INSN (before_return_copy
))->dest
;
416 before_return_copy
= last_insn
;
417 pre_exit
= split_block (src_bb
, before_return_copy
)->src
;
421 pre_exit
= split_edge (eg
);
429 /* Find all insns that need a particular mode setting, and insert the
430 necessary mode switches. Return true if we did work. */
433 optimize_mode_switching (void)
440 struct edge_list
*edge_list
;
441 static const int num_modes
[] = NUM_MODES_FOR_MODE_SWITCHING
;
442 #define N_ENTITIES ARRAY_SIZE (num_modes)
443 int entity_map
[N_ENTITIES
];
444 struct bb_info
*bb_info
[N_ENTITIES
];
447 int max_num_modes
= 0;
448 bool emited ATTRIBUTE_UNUSED
= false;
449 basic_block post_entry ATTRIBUTE_UNUSED
, pre_exit ATTRIBUTE_UNUSED
;
451 for (e
= N_ENTITIES
- 1, n_entities
= 0; e
>= 0; e
--)
452 if (OPTIMIZE_MODE_SWITCHING (e
))
454 int entry_exit_extra
= 0;
456 /* Create the list of segments within each basic block.
457 If NORMAL_MODE is defined, allow for two extra
458 blocks split from the entry and exit block. */
459 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
460 entry_exit_extra
= 3;
463 = XCNEWVEC (struct bb_info
, last_basic_block
+ entry_exit_extra
);
464 entity_map
[n_entities
++] = e
;
465 if (num_modes
[e
] > max_num_modes
)
466 max_num_modes
= num_modes
[e
];
472 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
473 /* Split the edge from the entry block, so that we can note that
474 there NORMAL_MODE is supplied. */
475 post_entry
= split_edge (single_succ_edge (ENTRY_BLOCK_PTR
));
476 pre_exit
= create_pre_exit (n_entities
, entity_map
, num_modes
);
481 /* Create the bitmap vectors. */
483 antic
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
484 transp
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
485 comp
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
487 sbitmap_vector_ones (transp
, last_basic_block
);
489 for (j
= n_entities
- 1; j
>= 0; j
--)
491 int e
= entity_map
[j
];
492 int no_mode
= num_modes
[e
];
493 struct bb_info
*info
= bb_info
[j
];
495 /* Determine what the first use (if any) need for a mode of entity E is.
496 This will be the mode that is anticipatable for this block.
497 Also compute the initial transparency settings. */
501 int last_mode
= no_mode
;
502 HARD_REG_SET live_now
;
504 REG_SET_TO_HARD_REG_SET (live_now
, df_get_live_in (bb
));
506 /* Pretend the mode is clobbered across abnormal edges. */
510 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
511 if (e
->flags
& EDGE_COMPLEX
)
515 ptr
= new_seginfo (no_mode
, BB_HEAD (bb
), bb
->index
, live_now
);
516 add_seginfo (info
+ bb
->index
, ptr
);
517 RESET_BIT (transp
[bb
->index
], j
);
521 for (insn
= BB_HEAD (bb
);
522 insn
!= NULL
&& insn
!= NEXT_INSN (BB_END (bb
));
523 insn
= NEXT_INSN (insn
))
527 int mode
= MODE_NEEDED (e
, insn
);
530 if (mode
!= no_mode
&& mode
!= last_mode
)
533 ptr
= new_seginfo (mode
, insn
, bb
->index
, live_now
);
534 add_seginfo (info
+ bb
->index
, ptr
);
535 RESET_BIT (transp
[bb
->index
], j
);
538 last_mode
= MODE_AFTER (last_mode
, insn
);
540 /* Update LIVE_NOW. */
541 for (link
= REG_NOTES (insn
); link
; link
= XEXP (link
, 1))
542 if (REG_NOTE_KIND (link
) == REG_DEAD
)
543 reg_dies (XEXP (link
, 0), &live_now
);
545 note_stores (PATTERN (insn
), reg_becomes_live
, &live_now
);
546 for (link
= REG_NOTES (insn
); link
; link
= XEXP (link
, 1))
547 if (REG_NOTE_KIND (link
) == REG_UNUSED
)
548 reg_dies (XEXP (link
, 0), &live_now
);
552 info
[bb
->index
].computing
= last_mode
;
553 /* Check for blocks without ANY mode requirements. */
554 if (last_mode
== no_mode
)
556 ptr
= new_seginfo (no_mode
, BB_END (bb
), bb
->index
, live_now
);
557 add_seginfo (info
+ bb
->index
, ptr
);
560 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
562 int mode
= MODE_ENTRY (e
);
568 /* By always making this nontransparent, we save
569 an extra check in make_preds_opaque. We also
570 need this to avoid confusing pre_edge_lcm when
571 antic is cleared but transp and comp are set. */
572 RESET_BIT (transp
[bb
->index
], j
);
574 /* Insert a fake computing definition of MODE into entry
575 blocks which compute no mode. This represents the mode on
577 info
[bb
->index
].computing
= mode
;
580 info
[pre_exit
->index
].seginfo
->mode
= MODE_EXIT (e
);
583 #endif /* NORMAL_MODE */
586 kill
= sbitmap_vector_alloc (last_basic_block
, n_entities
);
587 for (i
= 0; i
< max_num_modes
; i
++)
589 int current_mode
[N_ENTITIES
];
593 /* Set the anticipatable and computing arrays. */
594 sbitmap_vector_zero (antic
, last_basic_block
);
595 sbitmap_vector_zero (comp
, last_basic_block
);
596 for (j
= n_entities
- 1; j
>= 0; j
--)
598 int m
= current_mode
[j
] = MODE_PRIORITY_TO_MODE (entity_map
[j
], i
);
599 struct bb_info
*info
= bb_info
[j
];
603 if (info
[bb
->index
].seginfo
->mode
== m
)
604 SET_BIT (antic
[bb
->index
], j
);
606 if (info
[bb
->index
].computing
== m
)
607 SET_BIT (comp
[bb
->index
], j
);
611 /* Calculate the optimal locations for the
612 placement mode switches to modes with priority I. */
615 sbitmap_not (kill
[bb
->index
], transp
[bb
->index
]);
616 edge_list
= pre_edge_lcm (n_entities
, transp
, comp
, antic
,
617 kill
, &insert
, &del
);
619 for (j
= n_entities
- 1; j
>= 0; j
--)
621 /* Insert all mode sets that have been inserted by lcm. */
622 int no_mode
= num_modes
[entity_map
[j
]];
624 /* Wherever we have moved a mode setting upwards in the flow graph,
625 the blocks between the new setting site and the now redundant
626 computation ceases to be transparent for any lower-priority
627 mode of the same entity. First set the aux field of each
628 insertion site edge non-transparent, then propagate the new
629 non-transparency from the redundant computation upwards till
630 we hit an insertion site or an already non-transparent block. */
631 for (e
= NUM_EDGES (edge_list
) - 1; e
>= 0; e
--)
633 edge eg
= INDEX_EDGE (edge_list
, e
);
636 HARD_REG_SET live_at_edge
;
641 if (! TEST_BIT (insert
[e
], j
))
646 mode
= current_mode
[j
];
649 REG_SET_TO_HARD_REG_SET (live_at_edge
, df_get_live_out (src_bb
));
652 EMIT_MODE_SET (entity_map
[j
], mode
, live_at_edge
);
653 mode_set
= get_insns ();
656 /* Do not bother to insert empty sequence. */
657 if (mode_set
== NULL_RTX
)
660 /* We should not get an abnormal edge here. */
661 gcc_assert (! (eg
->flags
& EDGE_ABNORMAL
));
664 insert_insn_on_edge (mode_set
, eg
);
667 FOR_EACH_BB_REVERSE (bb
)
668 if (TEST_BIT (del
[bb
->index
], j
))
670 make_preds_opaque (bb
, j
);
671 /* Cancel the 'deleted' mode set. */
672 bb_info
[j
][bb
->index
].seginfo
->mode
= no_mode
;
676 sbitmap_vector_free (del
);
677 sbitmap_vector_free (insert
);
678 clear_aux_for_edges ();
679 free_edge_list (edge_list
);
682 /* Now output the remaining mode sets in all the segments. */
683 for (j
= n_entities
- 1; j
>= 0; j
--)
685 int no_mode
= num_modes
[entity_map
[j
]];
687 FOR_EACH_BB_REVERSE (bb
)
689 struct seginfo
*ptr
, *next
;
690 for (ptr
= bb_info
[j
][bb
->index
].seginfo
; ptr
; ptr
= next
)
693 if (ptr
->mode
!= no_mode
)
698 EMIT_MODE_SET (entity_map
[j
], ptr
->mode
, ptr
->regs_live
);
699 mode_set
= get_insns ();
702 /* Insert MODE_SET only if it is nonempty. */
703 if (mode_set
!= NULL_RTX
)
706 if (NOTE_INSN_BASIC_BLOCK_P (ptr
->insn_ptr
))
707 emit_insn_after (mode_set
, ptr
->insn_ptr
);
709 emit_insn_before (mode_set
, ptr
->insn_ptr
);
720 /* Finished. Free up all the things we've allocated. */
721 sbitmap_vector_free (kill
);
722 sbitmap_vector_free (antic
);
723 sbitmap_vector_free (transp
);
724 sbitmap_vector_free (comp
);
727 commit_edge_insertions ();
729 #if defined (MODE_ENTRY) && defined (MODE_EXIT)
730 cleanup_cfg (CLEANUP_NO_INSN_DEL
);
732 if (!need_commit
&& !emited
)
739 #endif /* OPTIMIZE_MODE_SWITCHING */
742 gate_mode_switching (void)
744 #ifdef OPTIMIZE_MODE_SWITCHING
752 rest_of_handle_mode_switching (void)
754 #ifdef OPTIMIZE_MODE_SWITCHING
755 optimize_mode_switching ();
756 #endif /* OPTIMIZE_MODE_SWITCHING */
761 struct rtl_opt_pass pass_mode_switching
=
765 "mode_sw", /* name */
766 gate_mode_switching
, /* gate */
767 rest_of_handle_mode_switching
, /* execute */
770 0, /* static_pass_number */
771 TV_MODE_SWITCH
, /* tv_id */
772 0, /* properties_required */
773 0, /* properties_provided */
774 0, /* properties_destroyed */
775 0, /* todo_flags_start */
776 TODO_df_finish
| TODO_verify_rtl_sharing
|
777 TODO_dump_func
/* todo_flags_finish */