1 /* Perform instruction reorganizations for delay slot filling.
2 Copyright (C) 1992-2013 Free Software Foundation, Inc.
3 Contributed by Richard Kenner (kenner@vlsi1.ultra.nyu.edu).
4 Hacked by Michael Tiemann (tiemann@cygnus.com).
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /* Instruction reorganization pass.
24 This pass runs after register allocation and final jump
25 optimization. It should be the last pass to run before peephole.
26 It serves primarily to fill delay slots of insns, typically branch
27 and call insns. Other insns typically involve more complicated
28 interactions of data dependencies and resource constraints, and
29 are better handled by scheduling before register allocation (by the
30 function `schedule_insns').
32 The Branch Penalty is the number of extra cycles that are needed to
33 execute a branch insn. On an ideal machine, branches take a single
34 cycle, and the Branch Penalty is 0. Several RISC machines approach
35 branch delays differently:
37 The MIPS has a single branch delay slot. Most insns
38 (except other branches) can be used to fill this slot. When the
39 slot is filled, two insns execute in two cycles, reducing the
40 branch penalty to zero.
42 The SPARC always has a branch delay slot, but its effects can be
43 annulled when the branch is not taken. This means that failing to
44 find other sources of insns, we can hoist an insn from the branch
45 target that would only be safe to execute knowing that the branch
48 The HP-PA always has a branch delay slot. For unconditional branches
49 its effects can be annulled when the branch is taken. The effects
50 of the delay slot in a conditional branch can be nullified for forward
51 taken branches, or for untaken backward branches. This means
52 we can hoist insns from the fall-through path for forward branches or
53 steal insns from the target of backward branches.
55 The TMS320C3x and C4x have three branch delay slots. When the three
56 slots are filled, the branch penalty is zero. Most insns can fill the
57 delay slots except jump insns.
59 Three techniques for filling delay slots have been implemented so far:
61 (1) `fill_simple_delay_slots' is the simplest, most efficient way
62 to fill delay slots. This pass first looks for insns which come
63 from before the branch and which are safe to execute after the
64 branch. Then it searches after the insn requiring delay slots or,
65 in the case of a branch, for insns that are after the point at
66 which the branch merges into the fallthrough code, if such a point
67 exists. When such insns are found, the branch penalty decreases
68 and no code expansion takes place.
70 (2) `fill_eager_delay_slots' is more complicated: it is used for
71 scheduling conditional jumps, or for scheduling jumps which cannot
72 be filled using (1). A machine need not have annulled jumps to use
73 this strategy, but it helps (by keeping more options open).
74 `fill_eager_delay_slots' tries to guess the direction the branch
75 will go; if it guesses right 100% of the time, it can reduce the
76 branch penalty as much as `fill_simple_delay_slots' does. If it
77 guesses wrong 100% of the time, it might as well schedule nops. When
78 `fill_eager_delay_slots' takes insns from the fall-through path of
79 the jump, usually there is no code expansion; when it takes insns
80 from the branch target, there is code expansion if it is not the
81 only way to reach that target.
83 (3) `relax_delay_slots' uses a set of rules to simplify code that
84 has been reorganized by (1) and (2). It finds cases where
85 conditional test can be eliminated, jumps can be threaded, extra
86 insns can be eliminated, etc. It is the job of (1) and (2) to do a
87 good job of scheduling locally; `relax_delay_slots' takes care of
88 making the various individual schedules work well together. It is
89 especially tuned to handle the control flow interactions of branch
90 insns. It does nothing for insns with delay slots that do not
93 On machines that use CC0, we are very conservative. We will not make
94 a copy of an insn involving CC0 since we want to maintain a 1-1
95 correspondence between the insn that sets and uses CC0. The insns are
96 allowed to be separated by placing an insn that sets CC0 (but not an insn
97 that uses CC0; we could do this, but it doesn't seem worthwhile) in a
98 delay slot. In that case, we point each insn at the other with REG_CC_USER
99 and REG_CC_SETTER notes. Note that these restrictions affect very few
100 machines because most RISC machines with delay slots will not use CC0
101 (the RT is the only known exception at this point). */
105 #include "coretypes.h"
107 #include "diagnostic-core.h"
111 #include "function.h"
112 #include "insn-config.h"
113 #include "conditions.h"
114 #include "hard-reg-set.h"
115 #include "basic-block.h"
120 #include "insn-attr.h"
121 #include "resource.h"
125 #include "tree-pass.h"
126 #include "emit-rtl.h"
130 #ifndef ANNUL_IFTRUE_SLOTS
131 #define eligible_for_annul_true(INSN, SLOTS, TRIAL, FLAGS) 0
133 #ifndef ANNUL_IFFALSE_SLOTS
134 #define eligible_for_annul_false(INSN, SLOTS, TRIAL, FLAGS) 0
137 /* Insns which have delay slots that have not yet been filled. */
139 static struct obstack unfilled_slots_obstack
;
140 static rtx
*unfilled_firstobj
;
142 /* Define macros to refer to the first and last slot containing unfilled
143 insns. These are used because the list may move and its address
144 should be recomputed at each use. */
146 #define unfilled_slots_base \
147 ((rtx *) obstack_base (&unfilled_slots_obstack))
149 #define unfilled_slots_next \
150 ((rtx *) obstack_next_free (&unfilled_slots_obstack))
152 /* Points to the label before the end of the function, or before a
154 static rtx function_return_label
;
155 /* Likewise for a simple_return. */
156 static rtx function_simple_return_label
;
158 /* Mapping between INSN_UID's and position in the code since INSN_UID's do
159 not always monotonically increase. */
160 static int *uid_to_ruid
;
162 /* Highest valid index in `uid_to_ruid'. */
165 static int stop_search_p (rtx
, int);
166 static int resource_conflicts_p (struct resources
*, struct resources
*);
167 static int insn_references_resource_p (rtx
, struct resources
*, bool);
168 static int insn_sets_resource_p (rtx
, struct resources
*, bool);
169 static rtx
find_end_label (rtx
);
170 static rtx
emit_delay_sequence (rtx
, rtx
, int);
171 static rtx
add_to_delay_list (rtx
, rtx
);
172 static rtx
delete_from_delay_slot (rtx
);
173 static void delete_scheduled_jump (rtx
);
174 static void note_delay_statistics (int, int);
175 #if defined(ANNUL_IFFALSE_SLOTS) || defined(ANNUL_IFTRUE_SLOTS)
176 static rtx
optimize_skip (rtx
);
178 static int get_jump_flags (rtx
, rtx
);
179 static int mostly_true_jump (rtx
);
180 static rtx
get_branch_condition (rtx
, rtx
);
181 static int condition_dominates_p (rtx
, rtx
);
182 static int redirect_with_delay_slots_safe_p (rtx
, rtx
, rtx
);
183 static int redirect_with_delay_list_safe_p (rtx
, rtx
, rtx
);
184 static int check_annul_list_true_false (int, rtx
);
185 static rtx
steal_delay_list_from_target (rtx
, rtx
, rtx
, rtx
,
189 int, int *, int *, rtx
*);
190 static rtx
steal_delay_list_from_fallthrough (rtx
, rtx
, rtx
, rtx
,
195 static void try_merge_delay_insns (rtx
, rtx
);
196 static rtx
redundant_insn (rtx
, rtx
, rtx
);
197 static int own_thread_p (rtx
, rtx
, int);
198 static void update_block (rtx
, rtx
);
199 static int reorg_redirect_jump (rtx
, rtx
);
200 static void update_reg_dead_notes (rtx
, rtx
);
201 static void fix_reg_dead_note (rtx
, rtx
);
202 static void update_reg_unused_notes (rtx
, rtx
);
203 static void fill_simple_delay_slots (int);
204 static rtx
fill_slots_from_thread (rtx
, rtx
, rtx
, rtx
,
207 static void fill_eager_delay_slots (void);
208 static void relax_delay_slots (rtx
);
209 static void make_return_insns (rtx
);
211 /* A wrapper around next_active_insn which takes care to return ret_rtx
215 first_active_target_insn (rtx insn
)
217 if (ANY_RETURN_P (insn
))
219 return next_active_insn (insn
);
222 /* Return true iff INSN is a simplejump, or any kind of return insn. */
225 simplejump_or_return_p (rtx insn
)
227 return (JUMP_P (insn
)
228 && (simplejump_p (insn
) || ANY_RETURN_P (PATTERN (insn
))));
231 /* Return TRUE if this insn should stop the search for insn to fill delay
232 slots. LABELS_P indicates that labels should terminate the search.
233 In all cases, jumps terminate the search. */
236 stop_search_p (rtx insn
, int labels_p
)
241 /* If the insn can throw an exception that is caught within the function,
242 it may effectively perform a jump from the viewpoint of the function.
243 Therefore act like for a jump. */
244 if (can_throw_internal (insn
))
247 switch (GET_CODE (insn
))
261 /* OK unless it contains a delay slot or is an `asm' insn of some type.
262 We don't know anything about these. */
263 return (GET_CODE (PATTERN (insn
)) == SEQUENCE
264 || GET_CODE (PATTERN (insn
)) == ASM_INPUT
265 || asm_noperands (PATTERN (insn
)) >= 0);
272 /* Return TRUE if any resources are marked in both RES1 and RES2 or if either
273 resource set contains a volatile memory reference. Otherwise, return FALSE. */
276 resource_conflicts_p (struct resources
*res1
, struct resources
*res2
)
278 if ((res1
->cc
&& res2
->cc
) || (res1
->memory
&& res2
->memory
)
279 || (res1
->unch_memory
&& res2
->unch_memory
)
280 || res1
->volatil
|| res2
->volatil
)
283 return hard_reg_set_intersect_p (res1
->regs
, res2
->regs
);
286 /* Return TRUE if any resource marked in RES, a `struct resources', is
287 referenced by INSN. If INCLUDE_DELAYED_EFFECTS is set, return if the called
288 routine is using those resources.
290 We compute this by computing all the resources referenced by INSN and
291 seeing if this conflicts with RES. It might be faster to directly check
292 ourselves, and this is the way it used to work, but it means duplicating
293 a large block of complex code. */
296 insn_references_resource_p (rtx insn
, struct resources
*res
,
297 bool include_delayed_effects
)
299 struct resources insn_res
;
301 CLEAR_RESOURCE (&insn_res
);
302 mark_referenced_resources (insn
, &insn_res
, include_delayed_effects
);
303 return resource_conflicts_p (&insn_res
, res
);
306 /* Return TRUE if INSN modifies resources that are marked in RES.
307 INCLUDE_DELAYED_EFFECTS is set if the actions of that routine should be
308 included. CC0 is only modified if it is explicitly set; see comments
309 in front of mark_set_resources for details. */
312 insn_sets_resource_p (rtx insn
, struct resources
*res
,
313 bool include_delayed_effects
)
315 struct resources insn_sets
;
317 CLEAR_RESOURCE (&insn_sets
);
318 mark_set_resources (insn
, &insn_sets
, 0,
319 (include_delayed_effects
322 return resource_conflicts_p (&insn_sets
, res
);
325 /* Find a label at the end of the function or before a RETURN. If there
326 is none, try to make one. If that fails, returns 0.
328 The property of such a label is that it is placed just before the
329 epilogue or a bare RETURN insn, so that another bare RETURN can be
330 turned into a jump to the label unconditionally. In particular, the
331 label cannot be placed before a RETURN insn with a filled delay slot.
333 ??? There may be a problem with the current implementation. Suppose
334 we start with a bare RETURN insn and call find_end_label. It may set
335 function_return_label just before the RETURN. Suppose the machinery
336 is able to fill the delay slot of the RETURN insn afterwards. Then
337 function_return_label is no longer valid according to the property
338 described above and find_end_label will still return it unmodified.
339 Note that this is probably mitigated by the following observation:
340 once function_return_label is made, it is very likely the target of
341 a jump, so filling the delay slot of the RETURN will be much more
343 KIND is either simple_return_rtx or ret_rtx, indicating which type of
344 return we're looking for. */
347 find_end_label (rtx kind
)
353 plabel
= &function_return_label
;
356 gcc_assert (kind
== simple_return_rtx
);
357 plabel
= &function_simple_return_label
;
360 /* If we found one previously, return it. */
364 /* Otherwise, see if there is a label at the end of the function. If there
365 is, it must be that RETURN insns aren't needed, so that is our return
366 label and we don't have to do anything else. */
368 insn
= get_last_insn ();
370 || (NONJUMP_INSN_P (insn
)
371 && (GET_CODE (PATTERN (insn
)) == USE
372 || GET_CODE (PATTERN (insn
)) == CLOBBER
)))
373 insn
= PREV_INSN (insn
);
375 /* When a target threads its epilogue we might already have a
376 suitable return insn. If so put a label before it for the
377 function_return_label. */
379 && JUMP_P (PREV_INSN (insn
))
380 && PATTERN (PREV_INSN (insn
)) == kind
)
382 rtx temp
= PREV_INSN (PREV_INSN (insn
));
383 rtx label
= gen_label_rtx ();
384 LABEL_NUSES (label
) = 0;
386 /* Put the label before any USE insns that may precede the RETURN
388 while (GET_CODE (temp
) == USE
)
389 temp
= PREV_INSN (temp
);
391 emit_label_after (label
, temp
);
395 else if (LABEL_P (insn
))
399 rtx label
= gen_label_rtx ();
400 LABEL_NUSES (label
) = 0;
401 /* If the basic block reorder pass moves the return insn to
402 some other place try to locate it again and put our
403 function_return_label there. */
404 while (insn
&& ! (JUMP_P (insn
) && (PATTERN (insn
) == kind
)))
405 insn
= PREV_INSN (insn
);
408 insn
= PREV_INSN (insn
);
410 /* Put the label before any USE insns that may precede the
412 while (GET_CODE (insn
) == USE
)
413 insn
= PREV_INSN (insn
);
415 emit_label_after (label
, insn
);
425 /* The RETURN insn has its delay slot filled so we cannot
426 emit the label just before it. Since we already have
427 an epilogue and cannot emit a new RETURN, we cannot
428 emit the label at all. */
430 #endif /* HAVE_epilogue */
432 /* Otherwise, make a new label and emit a RETURN and BARRIER,
438 /* The return we make may have delay slots too. */
439 rtx insn
= gen_return ();
440 insn
= emit_jump_insn (insn
);
441 set_return_jump_label (insn
);
443 if (num_delay_slots (insn
) > 0)
444 obstack_ptr_grow (&unfilled_slots_obstack
, insn
);
451 /* Show one additional use for this label so it won't go away until
453 ++LABEL_NUSES (*plabel
);
458 /* Put INSN and LIST together in a SEQUENCE rtx of LENGTH, and replace
459 the pattern of INSN with the SEQUENCE.
461 Chain the insns so that NEXT_INSN of each insn in the sequence points to
462 the next and NEXT_INSN of the last insn in the sequence points to
463 the first insn after the sequence. Similarly for PREV_INSN. This makes
464 it easier to scan all insns.
466 Returns the SEQUENCE that replaces INSN. */
469 emit_delay_sequence (rtx insn
, rtx list
, int length
)
475 /* Allocate the rtvec to hold the insns and the SEQUENCE. */
476 rtvec seqv
= rtvec_alloc (length
+ 1);
477 rtx seq
= gen_rtx_SEQUENCE (VOIDmode
, seqv
);
478 rtx seq_insn
= make_insn_raw (seq
);
479 rtx first
= get_insns ();
480 rtx last
= get_last_insn ();
482 /* Make a copy of the insn having delay slots. */
483 rtx delay_insn
= copy_rtx (insn
);
485 /* If INSN is followed by a BARRIER, delete the BARRIER since it will only
486 confuse further processing. Update LAST in case it was the last insn.
487 We will put the BARRIER back in later. */
488 if (NEXT_INSN (insn
) && BARRIER_P (NEXT_INSN (insn
)))
490 delete_related_insns (NEXT_INSN (insn
));
491 last
= get_last_insn ();
495 /* Splice our SEQUENCE into the insn stream where INSN used to be. */
496 NEXT_INSN (seq_insn
) = NEXT_INSN (insn
);
497 PREV_INSN (seq_insn
) = PREV_INSN (insn
);
500 PREV_INSN (NEXT_INSN (seq_insn
)) = seq_insn
;
503 NEXT_INSN (PREV_INSN (seq_insn
)) = seq_insn
;
505 /* Note the calls to set_new_first_and_last_insn must occur after
506 SEQ_INSN has been completely spliced into the insn stream.
508 Otherwise CUR_INSN_UID will get set to an incorrect value because
509 set_new_first_and_last_insn will not find SEQ_INSN in the chain. */
511 set_new_first_and_last_insn (first
, seq_insn
);
514 set_new_first_and_last_insn (seq_insn
, last
);
516 /* Build our SEQUENCE and rebuild the insn chain. */
517 XVECEXP (seq
, 0, 0) = delay_insn
;
518 INSN_DELETED_P (delay_insn
) = 0;
519 PREV_INSN (delay_insn
) = PREV_INSN (seq_insn
);
521 INSN_LOCATION (seq_insn
) = INSN_LOCATION (delay_insn
);
523 for (li
= list
; li
; li
= XEXP (li
, 1), i
++)
525 rtx tem
= XEXP (li
, 0);
528 /* Show that this copy of the insn isn't deleted. */
529 INSN_DELETED_P (tem
) = 0;
531 XVECEXP (seq
, 0, i
) = tem
;
532 PREV_INSN (tem
) = XVECEXP (seq
, 0, i
- 1);
533 NEXT_INSN (XVECEXP (seq
, 0, i
- 1)) = tem
;
535 /* SPARC assembler, for instance, emit warning when debug info is output
536 into the delay slot. */
537 if (INSN_LOCATION (tem
) && !INSN_LOCATION (seq_insn
))
538 INSN_LOCATION (seq_insn
) = INSN_LOCATION (tem
);
539 INSN_LOCATION (tem
) = 0;
541 for (note
= REG_NOTES (tem
); note
; note
= next
)
543 next
= XEXP (note
, 1);
544 switch (REG_NOTE_KIND (note
))
547 /* Remove any REG_DEAD notes because we can't rely on them now
548 that the insn has been moved. */
549 remove_note (tem
, note
);
552 case REG_LABEL_OPERAND
:
553 case REG_LABEL_TARGET
:
554 /* Keep the label reference count up to date. */
555 if (LABEL_P (XEXP (note
, 0)))
556 LABEL_NUSES (XEXP (note
, 0)) ++;
565 NEXT_INSN (XVECEXP (seq
, 0, length
)) = NEXT_INSN (seq_insn
);
567 /* If the previous insn is a SEQUENCE, update the NEXT_INSN pointer on the
568 last insn in that SEQUENCE to point to us. Similarly for the first
569 insn in the following insn if it is a SEQUENCE. */
571 if (PREV_INSN (seq_insn
) && NONJUMP_INSN_P (PREV_INSN (seq_insn
))
572 && GET_CODE (PATTERN (PREV_INSN (seq_insn
))) == SEQUENCE
)
573 NEXT_INSN (XVECEXP (PATTERN (PREV_INSN (seq_insn
)), 0,
574 XVECLEN (PATTERN (PREV_INSN (seq_insn
)), 0) - 1))
577 if (NEXT_INSN (seq_insn
) && NONJUMP_INSN_P (NEXT_INSN (seq_insn
))
578 && GET_CODE (PATTERN (NEXT_INSN (seq_insn
))) == SEQUENCE
)
579 PREV_INSN (XVECEXP (PATTERN (NEXT_INSN (seq_insn
)), 0, 0)) = seq_insn
;
581 /* If there used to be a BARRIER, put it back. */
583 emit_barrier_after (seq_insn
);
585 gcc_assert (i
== length
+ 1);
590 /* Add INSN to DELAY_LIST and return the head of the new list. The list must
591 be in the order in which the insns are to be executed. */
594 add_to_delay_list (rtx insn
, rtx delay_list
)
596 /* If we have an empty list, just make a new list element. If
597 INSN has its block number recorded, clear it since we may
598 be moving the insn to a new block. */
602 clear_hashed_info_for_insn (insn
);
603 return gen_rtx_INSN_LIST (VOIDmode
, insn
, NULL_RTX
);
606 /* Otherwise this must be an INSN_LIST. Add INSN to the end of the
608 XEXP (delay_list
, 1) = add_to_delay_list (insn
, XEXP (delay_list
, 1));
613 /* Delete INSN from the delay slot of the insn that it is in, which may
614 produce an insn with no delay slots. Return the new insn. */
617 delete_from_delay_slot (rtx insn
)
619 rtx trial
, seq_insn
, seq
, prev
;
624 /* We first must find the insn containing the SEQUENCE with INSN in its
625 delay slot. Do this by finding an insn, TRIAL, where
626 PREV_INSN (NEXT_INSN (TRIAL)) != TRIAL. */
629 PREV_INSN (NEXT_INSN (trial
)) == trial
;
630 trial
= NEXT_INSN (trial
))
633 seq_insn
= PREV_INSN (NEXT_INSN (trial
));
634 seq
= PATTERN (seq_insn
);
636 if (NEXT_INSN (seq_insn
) && BARRIER_P (NEXT_INSN (seq_insn
)))
639 /* Create a delay list consisting of all the insns other than the one
640 we are deleting (unless we were the only one). */
641 if (XVECLEN (seq
, 0) > 2)
642 for (i
= 1; i
< XVECLEN (seq
, 0); i
++)
643 if (XVECEXP (seq
, 0, i
) != insn
)
644 delay_list
= add_to_delay_list (XVECEXP (seq
, 0, i
), delay_list
);
646 /* Delete the old SEQUENCE, re-emit the insn that used to have the delay
647 list, and rebuild the delay list if non-empty. */
648 prev
= PREV_INSN (seq_insn
);
649 trial
= XVECEXP (seq
, 0, 0);
650 delete_related_insns (seq_insn
);
651 add_insn_after (trial
, prev
, NULL
);
653 /* If there was a barrier after the old SEQUENCE, remit it. */
655 emit_barrier_after (trial
);
657 /* If there are any delay insns, remit them. Otherwise clear the
660 trial
= emit_delay_sequence (trial
, delay_list
, XVECLEN (seq
, 0) - 2);
661 else if (JUMP_P (trial
))
662 INSN_ANNULLED_BRANCH_P (trial
) = 0;
664 INSN_FROM_TARGET_P (insn
) = 0;
666 /* Show we need to fill this insn again. */
667 obstack_ptr_grow (&unfilled_slots_obstack
, trial
);
672 /* Delete INSN, a JUMP_INSN. If it is a conditional jump, we must track down
673 the insn that sets CC0 for it and delete it too. */
676 delete_scheduled_jump (rtx insn
)
678 /* Delete the insn that sets cc0 for us. On machines without cc0, we could
679 delete the insn that sets the condition code, but it is hard to find it.
680 Since this case is rare anyway, don't bother trying; there would likely
681 be other insns that became dead anyway, which we wouldn't know to
685 if (reg_mentioned_p (cc0_rtx
, insn
))
687 rtx note
= find_reg_note (insn
, REG_CC_SETTER
, NULL_RTX
);
689 /* If a reg-note was found, it points to an insn to set CC0. This
690 insn is in the delay list of some other insn. So delete it from
691 the delay list it was in. */
694 if (! FIND_REG_INC_NOTE (XEXP (note
, 0), NULL_RTX
)
695 && sets_cc0_p (PATTERN (XEXP (note
, 0))) == 1)
696 delete_from_delay_slot (XEXP (note
, 0));
700 /* The insn setting CC0 is our previous insn, but it may be in
701 a delay slot. It will be the last insn in the delay slot, if
703 rtx trial
= previous_insn (insn
);
705 trial
= prev_nonnote_insn (trial
);
706 if (sets_cc0_p (PATTERN (trial
)) != 1
707 || FIND_REG_INC_NOTE (trial
, NULL_RTX
))
709 if (PREV_INSN (NEXT_INSN (trial
)) == trial
)
710 delete_related_insns (trial
);
712 delete_from_delay_slot (trial
);
717 delete_related_insns (insn
);
720 /* Counters for delay-slot filling. */
722 #define NUM_REORG_FUNCTIONS 2
723 #define MAX_DELAY_HISTOGRAM 3
724 #define MAX_REORG_PASSES 2
726 static int num_insns_needing_delays
[NUM_REORG_FUNCTIONS
][MAX_REORG_PASSES
];
728 static int num_filled_delays
[NUM_REORG_FUNCTIONS
][MAX_DELAY_HISTOGRAM
+1][MAX_REORG_PASSES
];
730 static int reorg_pass_number
;
733 note_delay_statistics (int slots_filled
, int index
)
735 num_insns_needing_delays
[index
][reorg_pass_number
]++;
736 if (slots_filled
> MAX_DELAY_HISTOGRAM
)
737 slots_filled
= MAX_DELAY_HISTOGRAM
;
738 num_filled_delays
[index
][slots_filled
][reorg_pass_number
]++;
741 #if defined(ANNUL_IFFALSE_SLOTS) || defined(ANNUL_IFTRUE_SLOTS)
743 /* Optimize the following cases:
745 1. When a conditional branch skips over only one instruction,
746 use an annulling branch and put that insn in the delay slot.
747 Use either a branch that annuls when the condition if true or
748 invert the test with a branch that annuls when the condition is
749 false. This saves insns, since otherwise we must copy an insn
752 (orig) (skip) (otherwise)
753 Bcc.n L1 Bcc',a L1 Bcc,a L1'
760 2. When a conditional branch skips over only one instruction,
761 and after that, it unconditionally branches somewhere else,
762 perform the similar optimization. This saves executing the
763 second branch in the case where the inverted condition is true.
772 This should be expanded to skip over N insns, where N is the number
773 of delay slots required. */
776 optimize_skip (rtx insn
)
778 rtx trial
= next_nonnote_insn (insn
);
779 rtx next_trial
= next_active_insn (trial
);
783 flags
= get_jump_flags (insn
, JUMP_LABEL (insn
));
786 || !NONJUMP_INSN_P (trial
)
787 || GET_CODE (PATTERN (trial
)) == SEQUENCE
788 || recog_memoized (trial
) < 0
789 || (! eligible_for_annul_false (insn
, 0, trial
, flags
)
790 && ! eligible_for_annul_true (insn
, 0, trial
, flags
))
791 || can_throw_internal (trial
))
794 /* There are two cases where we are just executing one insn (we assume
795 here that a branch requires only one insn; this should be generalized
796 at some point): Where the branch goes around a single insn or where
797 we have one insn followed by a branch to the same label we branch to.
798 In both of these cases, inverting the jump and annulling the delay
799 slot give the same effect in fewer insns. */
800 if (next_trial
== next_active_insn (JUMP_LABEL (insn
))
802 && simplejump_or_return_p (next_trial
)
803 && JUMP_LABEL (insn
) == JUMP_LABEL (next_trial
)))
805 if (eligible_for_annul_false (insn
, 0, trial
, flags
))
807 if (invert_jump (insn
, JUMP_LABEL (insn
), 1))
808 INSN_FROM_TARGET_P (trial
) = 1;
809 else if (! eligible_for_annul_true (insn
, 0, trial
, flags
))
813 delay_list
= add_to_delay_list (trial
, NULL_RTX
);
814 next_trial
= next_active_insn (trial
);
815 update_block (trial
, trial
);
816 delete_related_insns (trial
);
818 /* Also, if we are targeting an unconditional
819 branch, thread our jump to the target of that branch. Don't
820 change this into a RETURN here, because it may not accept what
821 we have in the delay slot. We'll fix this up later. */
822 if (next_trial
&& simplejump_or_return_p (next_trial
))
824 rtx target_label
= JUMP_LABEL (next_trial
);
825 if (ANY_RETURN_P (target_label
))
826 target_label
= find_end_label (target_label
);
830 /* Recompute the flags based on TARGET_LABEL since threading
831 the jump to TARGET_LABEL may change the direction of the
832 jump (which may change the circumstances in which the
833 delay slot is nullified). */
834 flags
= get_jump_flags (insn
, target_label
);
835 if (eligible_for_annul_true (insn
, 0, trial
, flags
))
836 reorg_redirect_jump (insn
, target_label
);
840 INSN_ANNULLED_BRANCH_P (insn
) = 1;
847 /* Encode and return branch direction and prediction information for
848 INSN assuming it will jump to LABEL.
850 Non conditional branches return no direction information and
851 are predicted as very likely taken. */
854 get_jump_flags (rtx insn
, rtx label
)
858 /* get_jump_flags can be passed any insn with delay slots, these may
859 be INSNs, CALL_INSNs, or JUMP_INSNs. Only JUMP_INSNs have branch
860 direction information, and only if they are conditional jumps.
862 If LABEL is a return, then there is no way to determine the branch
865 && (condjump_p (insn
) || condjump_in_parallel_p (insn
))
866 && !ANY_RETURN_P (label
)
867 && INSN_UID (insn
) <= max_uid
868 && INSN_UID (label
) <= max_uid
)
870 = (uid_to_ruid
[INSN_UID (label
)] > uid_to_ruid
[INSN_UID (insn
)])
871 ? ATTR_FLAG_forward
: ATTR_FLAG_backward
;
872 /* No valid direction information. */
879 /* Return truth value of the statement that this branch
880 is mostly taken. If we think that the branch is extremely likely
881 to be taken, we return 2. If the branch is slightly more likely to be
882 taken, return 1. If the branch is slightly less likely to be taken,
883 return 0 and if the branch is highly unlikely to be taken, return -1. */
886 mostly_true_jump (rtx jump_insn
)
888 /* If branch probabilities are available, then use that number since it
889 always gives a correct answer. */
890 rtx note
= find_reg_note (jump_insn
, REG_BR_PROB
, 0);
893 int prob
= INTVAL (XEXP (note
, 0));
895 if (prob
>= REG_BR_PROB_BASE
* 9 / 10)
897 else if (prob
>= REG_BR_PROB_BASE
/ 2)
899 else if (prob
>= REG_BR_PROB_BASE
/ 10)
905 /* If there is no note, assume branches are not taken.
906 This should be rare. */
910 /* Return the condition under which INSN will branch to TARGET. If TARGET
911 is zero, return the condition under which INSN will return. If INSN is
912 an unconditional branch, return const_true_rtx. If INSN isn't a simple
913 type of jump, or it doesn't go to TARGET, return 0. */
916 get_branch_condition (rtx insn
, rtx target
)
918 rtx pat
= PATTERN (insn
);
921 if (condjump_in_parallel_p (insn
))
922 pat
= XVECEXP (pat
, 0, 0);
924 if (ANY_RETURN_P (pat
))
925 return pat
== target
? const_true_rtx
: 0;
927 if (GET_CODE (pat
) != SET
|| SET_DEST (pat
) != pc_rtx
)
931 if (GET_CODE (src
) == LABEL_REF
&& XEXP (src
, 0) == target
)
932 return const_true_rtx
;
934 else if (GET_CODE (src
) == IF_THEN_ELSE
935 && XEXP (src
, 2) == pc_rtx
936 && GET_CODE (XEXP (src
, 1)) == LABEL_REF
937 && XEXP (XEXP (src
, 1), 0) == target
)
938 return XEXP (src
, 0);
940 else if (GET_CODE (src
) == IF_THEN_ELSE
941 && XEXP (src
, 1) == pc_rtx
942 && GET_CODE (XEXP (src
, 2)) == LABEL_REF
943 && XEXP (XEXP (src
, 2), 0) == target
)
946 rev
= reversed_comparison_code (XEXP (src
, 0), insn
);
948 return gen_rtx_fmt_ee (rev
, GET_MODE (XEXP (src
, 0)),
949 XEXP (XEXP (src
, 0), 0),
950 XEXP (XEXP (src
, 0), 1));
956 /* Return nonzero if CONDITION is more strict than the condition of
957 INSN, i.e., if INSN will always branch if CONDITION is true. */
960 condition_dominates_p (rtx condition
, rtx insn
)
962 rtx other_condition
= get_branch_condition (insn
, JUMP_LABEL (insn
));
963 enum rtx_code code
= GET_CODE (condition
);
964 enum rtx_code other_code
;
966 if (rtx_equal_p (condition
, other_condition
)
967 || other_condition
== const_true_rtx
)
970 else if (condition
== const_true_rtx
|| other_condition
== 0)
973 other_code
= GET_CODE (other_condition
);
974 if (GET_RTX_LENGTH (code
) != 2 || GET_RTX_LENGTH (other_code
) != 2
975 || ! rtx_equal_p (XEXP (condition
, 0), XEXP (other_condition
, 0))
976 || ! rtx_equal_p (XEXP (condition
, 1), XEXP (other_condition
, 1)))
979 return comparison_dominates_p (code
, other_code
);
982 /* Return nonzero if redirecting JUMP to NEWLABEL does not invalidate
983 any insns already in the delay slot of JUMP. */
986 redirect_with_delay_slots_safe_p (rtx jump
, rtx newlabel
, rtx seq
)
989 rtx pat
= PATTERN (seq
);
991 /* Make sure all the delay slots of this jump would still
992 be valid after threading the jump. If they are still
993 valid, then return nonzero. */
995 flags
= get_jump_flags (jump
, newlabel
);
996 for (i
= 1; i
< XVECLEN (pat
, 0); i
++)
998 #ifdef ANNUL_IFFALSE_SLOTS
999 (INSN_ANNULLED_BRANCH_P (jump
)
1000 && INSN_FROM_TARGET_P (XVECEXP (pat
, 0, i
)))
1001 ? eligible_for_annul_false (jump
, i
- 1,
1002 XVECEXP (pat
, 0, i
), flags
) :
1004 #ifdef ANNUL_IFTRUE_SLOTS
1005 (INSN_ANNULLED_BRANCH_P (jump
)
1006 && ! INSN_FROM_TARGET_P (XVECEXP (pat
, 0, i
)))
1007 ? eligible_for_annul_true (jump
, i
- 1,
1008 XVECEXP (pat
, 0, i
), flags
) :
1010 eligible_for_delay (jump
, i
- 1, XVECEXP (pat
, 0, i
), flags
)))
1013 return (i
== XVECLEN (pat
, 0));
1016 /* Return nonzero if redirecting JUMP to NEWLABEL does not invalidate
1017 any insns we wish to place in the delay slot of JUMP. */
1020 redirect_with_delay_list_safe_p (rtx jump
, rtx newlabel
, rtx delay_list
)
1025 /* Make sure all the insns in DELAY_LIST would still be
1026 valid after threading the jump. If they are still
1027 valid, then return nonzero. */
1029 flags
= get_jump_flags (jump
, newlabel
);
1030 for (li
= delay_list
, i
= 0; li
; li
= XEXP (li
, 1), i
++)
1032 #ifdef ANNUL_IFFALSE_SLOTS
1033 (INSN_ANNULLED_BRANCH_P (jump
)
1034 && INSN_FROM_TARGET_P (XEXP (li
, 0)))
1035 ? eligible_for_annul_false (jump
, i
, XEXP (li
, 0), flags
) :
1037 #ifdef ANNUL_IFTRUE_SLOTS
1038 (INSN_ANNULLED_BRANCH_P (jump
)
1039 && ! INSN_FROM_TARGET_P (XEXP (li
, 0)))
1040 ? eligible_for_annul_true (jump
, i
, XEXP (li
, 0), flags
) :
1042 eligible_for_delay (jump
, i
, XEXP (li
, 0), flags
)))
1045 return (li
== NULL
);
1048 /* DELAY_LIST is a list of insns that have already been placed into delay
1049 slots. See if all of them have the same annulling status as ANNUL_TRUE_P.
1050 If not, return 0; otherwise return 1. */
1053 check_annul_list_true_false (int annul_true_p
, rtx delay_list
)
1059 for (temp
= delay_list
; temp
; temp
= XEXP (temp
, 1))
1061 rtx trial
= XEXP (temp
, 0);
1063 if ((annul_true_p
&& INSN_FROM_TARGET_P (trial
))
1064 || (!annul_true_p
&& !INSN_FROM_TARGET_P (trial
)))
1072 /* INSN branches to an insn whose pattern SEQ is a SEQUENCE. Given that
1073 the condition tested by INSN is CONDITION and the resources shown in
1074 OTHER_NEEDED are needed after INSN, see whether INSN can take all the insns
1075 from SEQ's delay list, in addition to whatever insns it may execute
1076 (in DELAY_LIST). SETS and NEEDED are denote resources already set and
1077 needed while searching for delay slot insns. Return the concatenated
1078 delay list if possible, otherwise, return 0.
1080 SLOTS_TO_FILL is the total number of slots required by INSN, and
1081 PSLOTS_FILLED points to the number filled so far (also the number of
1082 insns in DELAY_LIST). It is updated with the number that have been
1083 filled from the SEQUENCE, if any.
1085 PANNUL_P points to a nonzero value if we already know that we need
1086 to annul INSN. If this routine determines that annulling is needed,
1087 it may set that value nonzero.
1089 PNEW_THREAD points to a location that is to receive the place at which
1090 execution should continue. */
1093 steal_delay_list_from_target (rtx insn
, rtx condition
, rtx seq
,
1094 rtx delay_list
, struct resources
*sets
,
1095 struct resources
*needed
,
1096 struct resources
*other_needed
,
1097 int slots_to_fill
, int *pslots_filled
,
1098 int *pannul_p
, rtx
*pnew_thread
)
1101 int slots_remaining
= slots_to_fill
- *pslots_filled
;
1102 int total_slots_filled
= *pslots_filled
;
1103 rtx new_delay_list
= 0;
1104 int must_annul
= *pannul_p
;
1107 struct resources cc_set
;
1110 /* We can't do anything if there are more delay slots in SEQ than we
1111 can handle, or if we don't know that it will be a taken branch.
1112 We know that it will be a taken branch if it is either an unconditional
1113 branch or a conditional branch with a stricter branch condition.
1115 Also, exit if the branch has more than one set, since then it is computing
1116 other results that can't be ignored, e.g. the HPPA mov&branch instruction.
1117 ??? It may be possible to move other sets into INSN in addition to
1118 moving the instructions in the delay slots.
1120 We can not steal the delay list if one of the instructions in the
1121 current delay_list modifies the condition codes and the jump in the
1122 sequence is a conditional jump. We can not do this because we can
1123 not change the direction of the jump because the condition codes
1124 will effect the direction of the jump in the sequence. */
1126 CLEAR_RESOURCE (&cc_set
);
1127 for (temp
= delay_list
; temp
; temp
= XEXP (temp
, 1))
1129 rtx trial
= XEXP (temp
, 0);
1131 mark_set_resources (trial
, &cc_set
, 0, MARK_SRC_DEST_CALL
);
1132 if (insn_references_resource_p (XVECEXP (seq
, 0, 0), &cc_set
, false))
1136 if (XVECLEN (seq
, 0) - 1 > slots_remaining
1137 || ! condition_dominates_p (condition
, XVECEXP (seq
, 0, 0))
1138 || ! single_set (XVECEXP (seq
, 0, 0)))
1141 #ifdef MD_CAN_REDIRECT_BRANCH
1142 /* On some targets, branches with delay slots can have a limited
1143 displacement. Give the back end a chance to tell us we can't do
1145 if (! MD_CAN_REDIRECT_BRANCH (insn
, XVECEXP (seq
, 0, 0)))
1149 redundant
= XALLOCAVEC (bool, XVECLEN (seq
, 0));
1150 for (i
= 1; i
< XVECLEN (seq
, 0); i
++)
1152 rtx trial
= XVECEXP (seq
, 0, i
);
1155 if (insn_references_resource_p (trial
, sets
, false)
1156 || insn_sets_resource_p (trial
, needed
, false)
1157 || insn_sets_resource_p (trial
, sets
, false)
1159 /* If TRIAL sets CC0, we can't copy it, so we can't steal this
1161 || find_reg_note (trial
, REG_CC_USER
, NULL_RTX
)
1163 /* If TRIAL is from the fallthrough code of an annulled branch insn
1164 in SEQ, we cannot use it. */
1165 || (INSN_ANNULLED_BRANCH_P (XVECEXP (seq
, 0, 0))
1166 && ! INSN_FROM_TARGET_P (trial
)))
1169 /* If this insn was already done (usually in a previous delay slot),
1170 pretend we put it in our delay slot. */
1171 redundant
[i
] = redundant_insn (trial
, insn
, new_delay_list
);
1175 /* We will end up re-vectoring this branch, so compute flags
1176 based on jumping to the new label. */
1177 flags
= get_jump_flags (insn
, JUMP_LABEL (XVECEXP (seq
, 0, 0)));
1180 && ((condition
== const_true_rtx
1181 || (! insn_sets_resource_p (trial
, other_needed
, false)
1182 && ! may_trap_or_fault_p (PATTERN (trial
)))))
1183 ? eligible_for_delay (insn
, total_slots_filled
, trial
, flags
)
1184 : (must_annul
|| (delay_list
== NULL
&& new_delay_list
== NULL
))
1186 check_annul_list_true_false (0, delay_list
)
1187 && check_annul_list_true_false (0, new_delay_list
)
1188 && eligible_for_annul_false (insn
, total_slots_filled
,
1193 temp
= copy_delay_slot_insn (trial
);
1194 INSN_FROM_TARGET_P (temp
) = 1;
1195 new_delay_list
= add_to_delay_list (temp
, new_delay_list
);
1196 total_slots_filled
++;
1198 if (--slots_remaining
== 0)
1205 /* Record the effect of the instructions that were redundant and which
1206 we therefore decided not to copy. */
1207 for (i
= 1; i
< XVECLEN (seq
, 0); i
++)
1209 update_block (XVECEXP (seq
, 0, i
), insn
);
1211 /* Show the place to which we will be branching. */
1212 *pnew_thread
= first_active_target_insn (JUMP_LABEL (XVECEXP (seq
, 0, 0)));
1214 /* Add any new insns to the delay list and update the count of the
1215 number of slots filled. */
1216 *pslots_filled
= total_slots_filled
;
1220 if (delay_list
== 0)
1221 return new_delay_list
;
1223 for (temp
= new_delay_list
; temp
; temp
= XEXP (temp
, 1))
1224 delay_list
= add_to_delay_list (XEXP (temp
, 0), delay_list
);
1229 /* Similar to steal_delay_list_from_target except that SEQ is on the
1230 fallthrough path of INSN. Here we only do something if the delay insn
1231 of SEQ is an unconditional branch. In that case we steal its delay slot
1232 for INSN since unconditional branches are much easier to fill. */
1235 steal_delay_list_from_fallthrough (rtx insn
, rtx condition
, rtx seq
,
1236 rtx delay_list
, struct resources
*sets
,
1237 struct resources
*needed
,
1238 struct resources
*other_needed
,
1239 int slots_to_fill
, int *pslots_filled
,
1244 int must_annul
= *pannul_p
;
1247 flags
= get_jump_flags (insn
, JUMP_LABEL (insn
));
1249 /* We can't do anything if SEQ's delay insn isn't an
1250 unconditional branch. */
1252 if (! simplejump_or_return_p (XVECEXP (seq
, 0, 0)))
1255 for (i
= 1; i
< XVECLEN (seq
, 0); i
++)
1257 rtx trial
= XVECEXP (seq
, 0, i
);
1259 /* If TRIAL sets CC0, stealing it will move it too far from the use
1261 if (insn_references_resource_p (trial
, sets
, false)
1262 || insn_sets_resource_p (trial
, needed
, false)
1263 || insn_sets_resource_p (trial
, sets
, false)
1265 || sets_cc0_p (PATTERN (trial
))
1271 /* If this insn was already done, we don't need it. */
1272 if (redundant_insn (trial
, insn
, delay_list
))
1274 update_block (trial
, insn
);
1275 delete_from_delay_slot (trial
);
1280 && ((condition
== const_true_rtx
1281 || (! insn_sets_resource_p (trial
, other_needed
, false)
1282 && ! may_trap_or_fault_p (PATTERN (trial
)))))
1283 ? eligible_for_delay (insn
, *pslots_filled
, trial
, flags
)
1284 : (must_annul
|| delay_list
== NULL
) && (must_annul
= 1,
1285 check_annul_list_true_false (1, delay_list
)
1286 && eligible_for_annul_true (insn
, *pslots_filled
, trial
, flags
)))
1290 delete_from_delay_slot (trial
);
1291 delay_list
= add_to_delay_list (trial
, delay_list
);
1293 if (++(*pslots_filled
) == slots_to_fill
)
1305 /* Try merging insns starting at THREAD which match exactly the insns in
1308 If all insns were matched and the insn was previously annulling, the
1309 annul bit will be cleared.
1311 For each insn that is merged, if the branch is or will be non-annulling,
1312 we delete the merged insn. */
1315 try_merge_delay_insns (rtx insn
, rtx thread
)
1317 rtx trial
, next_trial
;
1318 rtx delay_insn
= XVECEXP (PATTERN (insn
), 0, 0);
1319 int annul_p
= JUMP_P (delay_insn
) && INSN_ANNULLED_BRANCH_P (delay_insn
);
1320 int slot_number
= 1;
1321 int num_slots
= XVECLEN (PATTERN (insn
), 0);
1322 rtx next_to_match
= XVECEXP (PATTERN (insn
), 0, slot_number
);
1323 struct resources set
, needed
;
1324 rtx merged_insns
= 0;
1328 flags
= get_jump_flags (delay_insn
, JUMP_LABEL (delay_insn
));
1330 CLEAR_RESOURCE (&needed
);
1331 CLEAR_RESOURCE (&set
);
1333 /* If this is not an annulling branch, take into account anything needed in
1334 INSN's delay slot. This prevents two increments from being incorrectly
1335 folded into one. If we are annulling, this would be the correct
1336 thing to do. (The alternative, looking at things set in NEXT_TO_MATCH
1337 will essentially disable this optimization. This method is somewhat of
1338 a kludge, but I don't see a better way.) */
1340 for (i
= 1 ; i
< num_slots
; i
++)
1341 if (XVECEXP (PATTERN (insn
), 0, i
))
1342 mark_referenced_resources (XVECEXP (PATTERN (insn
), 0, i
), &needed
,
1345 for (trial
= thread
; !stop_search_p (trial
, 1); trial
= next_trial
)
1347 rtx pat
= PATTERN (trial
);
1348 rtx oldtrial
= trial
;
1350 next_trial
= next_nonnote_insn (trial
);
1352 /* TRIAL must be a CALL_INSN or INSN. Skip USE and CLOBBER. */
1353 if (NONJUMP_INSN_P (trial
)
1354 && (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
))
1357 if (GET_CODE (next_to_match
) == GET_CODE (trial
)
1359 /* We can't share an insn that sets cc0. */
1360 && ! sets_cc0_p (pat
)
1362 && ! insn_references_resource_p (trial
, &set
, true)
1363 && ! insn_sets_resource_p (trial
, &set
, true)
1364 && ! insn_sets_resource_p (trial
, &needed
, true)
1365 && (trial
= try_split (pat
, trial
, 0)) != 0
1366 /* Update next_trial, in case try_split succeeded. */
1367 && (next_trial
= next_nonnote_insn (trial
))
1368 /* Likewise THREAD. */
1369 && (thread
= oldtrial
== thread
? trial
: thread
)
1370 && rtx_equal_p (PATTERN (next_to_match
), PATTERN (trial
))
1371 /* Have to test this condition if annul condition is different
1372 from (and less restrictive than) non-annulling one. */
1373 && eligible_for_delay (delay_insn
, slot_number
- 1, trial
, flags
))
1378 update_block (trial
, thread
);
1379 if (trial
== thread
)
1380 thread
= next_active_insn (thread
);
1382 delete_related_insns (trial
);
1383 INSN_FROM_TARGET_P (next_to_match
) = 0;
1386 merged_insns
= gen_rtx_INSN_LIST (VOIDmode
, trial
, merged_insns
);
1388 if (++slot_number
== num_slots
)
1391 next_to_match
= XVECEXP (PATTERN (insn
), 0, slot_number
);
1394 mark_set_resources (trial
, &set
, 0, MARK_SRC_DEST_CALL
);
1395 mark_referenced_resources (trial
, &needed
, true);
1398 /* See if we stopped on a filled insn. If we did, try to see if its
1399 delay slots match. */
1400 if (slot_number
!= num_slots
1401 && trial
&& NONJUMP_INSN_P (trial
)
1402 && GET_CODE (PATTERN (trial
)) == SEQUENCE
1403 && !(JUMP_P (XVECEXP (PATTERN (trial
), 0, 0))
1404 && INSN_ANNULLED_BRANCH_P (XVECEXP (PATTERN (trial
), 0, 0))))
1406 rtx pat
= PATTERN (trial
);
1407 rtx filled_insn
= XVECEXP (pat
, 0, 0);
1409 /* Account for resources set/needed by the filled insn. */
1410 mark_set_resources (filled_insn
, &set
, 0, MARK_SRC_DEST_CALL
);
1411 mark_referenced_resources (filled_insn
, &needed
, true);
1413 for (i
= 1; i
< XVECLEN (pat
, 0); i
++)
1415 rtx dtrial
= XVECEXP (pat
, 0, i
);
1417 if (! insn_references_resource_p (dtrial
, &set
, true)
1418 && ! insn_sets_resource_p (dtrial
, &set
, true)
1419 && ! insn_sets_resource_p (dtrial
, &needed
, true)
1421 && ! sets_cc0_p (PATTERN (dtrial
))
1423 && rtx_equal_p (PATTERN (next_to_match
), PATTERN (dtrial
))
1424 && eligible_for_delay (delay_insn
, slot_number
- 1, dtrial
, flags
))
1430 update_block (dtrial
, thread
);
1431 new_rtx
= delete_from_delay_slot (dtrial
);
1432 if (INSN_DELETED_P (thread
))
1434 INSN_FROM_TARGET_P (next_to_match
) = 0;
1437 merged_insns
= gen_rtx_INSN_LIST (SImode
, dtrial
,
1440 if (++slot_number
== num_slots
)
1443 next_to_match
= XVECEXP (PATTERN (insn
), 0, slot_number
);
1447 /* Keep track of the set/referenced resources for the delay
1448 slots of any trial insns we encounter. */
1449 mark_set_resources (dtrial
, &set
, 0, MARK_SRC_DEST_CALL
);
1450 mark_referenced_resources (dtrial
, &needed
, true);
1455 /* If all insns in the delay slot have been matched and we were previously
1456 annulling the branch, we need not any more. In that case delete all the
1457 merged insns. Also clear the INSN_FROM_TARGET_P bit of each insn in
1458 the delay list so that we know that it isn't only being used at the
1460 if (slot_number
== num_slots
&& annul_p
)
1462 for (; merged_insns
; merged_insns
= XEXP (merged_insns
, 1))
1464 if (GET_MODE (merged_insns
) == SImode
)
1468 update_block (XEXP (merged_insns
, 0), thread
);
1469 new_rtx
= delete_from_delay_slot (XEXP (merged_insns
, 0));
1470 if (INSN_DELETED_P (thread
))
1475 update_block (XEXP (merged_insns
, 0), thread
);
1476 delete_related_insns (XEXP (merged_insns
, 0));
1480 INSN_ANNULLED_BRANCH_P (delay_insn
) = 0;
1482 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
1483 INSN_FROM_TARGET_P (XVECEXP (PATTERN (insn
), 0, i
)) = 0;
1487 /* See if INSN is redundant with an insn in front of TARGET. Often this
1488 is called when INSN is a candidate for a delay slot of TARGET.
1489 DELAY_LIST are insns that will be placed in delay slots of TARGET in front
1490 of INSN. Often INSN will be redundant with an insn in a delay slot of
1491 some previous insn. This happens when we have a series of branches to the
1492 same label; in that case the first insn at the target might want to go
1493 into each of the delay slots.
1495 If we are not careful, this routine can take up a significant fraction
1496 of the total compilation time (4%), but only wins rarely. Hence we
1497 speed this routine up by making two passes. The first pass goes back
1498 until it hits a label and sees if it finds an insn with an identical
1499 pattern. Only in this (relatively rare) event does it check for
1502 We do not split insns we encounter. This could cause us not to find a
1503 redundant insn, but the cost of splitting seems greater than the possible
1504 gain in rare cases. */
1507 redundant_insn (rtx insn
, rtx target
, rtx delay_list
)
1509 rtx target_main
= target
;
1510 rtx ipat
= PATTERN (insn
);
1512 struct resources needed
, set
;
1514 unsigned insns_to_search
;
1516 /* If INSN has any REG_UNUSED notes, it can't match anything since we
1517 are allowed to not actually assign to such a register. */
1518 if (find_reg_note (insn
, REG_UNUSED
, NULL_RTX
) != 0)
1521 /* Scan backwards looking for a match. */
1522 for (trial
= PREV_INSN (target
),
1523 insns_to_search
= MAX_DELAY_SLOT_INSN_SEARCH
;
1524 trial
&& insns_to_search
> 0;
1525 trial
= PREV_INSN (trial
))
1527 if (LABEL_P (trial
))
1530 if (!INSN_P (trial
))
1534 pat
= PATTERN (trial
);
1535 if (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
)
1538 if (GET_CODE (pat
) == SEQUENCE
)
1540 /* Stop for a CALL and its delay slots because it is difficult to
1541 track its resource needs correctly. */
1542 if (CALL_P (XVECEXP (pat
, 0, 0)))
1545 /* Stop for an INSN or JUMP_INSN with delayed effects and its delay
1546 slots because it is difficult to track its resource needs
1549 #ifdef INSN_SETS_ARE_DELAYED
1550 if (INSN_SETS_ARE_DELAYED (XVECEXP (pat
, 0, 0)))
1554 #ifdef INSN_REFERENCES_ARE_DELAYED
1555 if (INSN_REFERENCES_ARE_DELAYED (XVECEXP (pat
, 0, 0)))
1559 /* See if any of the insns in the delay slot match, updating
1560 resource requirements as we go. */
1561 for (i
= XVECLEN (pat
, 0) - 1; i
> 0; i
--)
1562 if (GET_CODE (XVECEXP (pat
, 0, i
)) == GET_CODE (insn
)
1563 && rtx_equal_p (PATTERN (XVECEXP (pat
, 0, i
)), ipat
)
1564 && ! find_reg_note (XVECEXP (pat
, 0, i
), REG_UNUSED
, NULL_RTX
))
1567 /* If found a match, exit this loop early. */
1572 else if (GET_CODE (trial
) == GET_CODE (insn
) && rtx_equal_p (pat
, ipat
)
1573 && ! find_reg_note (trial
, REG_UNUSED
, NULL_RTX
))
1577 /* If we didn't find an insn that matches, return 0. */
1581 /* See what resources this insn sets and needs. If they overlap, or
1582 if this insn references CC0, it can't be redundant. */
1584 CLEAR_RESOURCE (&needed
);
1585 CLEAR_RESOURCE (&set
);
1586 mark_set_resources (insn
, &set
, 0, MARK_SRC_DEST_CALL
);
1587 mark_referenced_resources (insn
, &needed
, true);
1589 /* If TARGET is a SEQUENCE, get the main insn. */
1590 if (NONJUMP_INSN_P (target
) && GET_CODE (PATTERN (target
)) == SEQUENCE
)
1591 target_main
= XVECEXP (PATTERN (target
), 0, 0);
1593 if (resource_conflicts_p (&needed
, &set
)
1595 || reg_mentioned_p (cc0_rtx
, ipat
)
1597 /* The insn requiring the delay may not set anything needed or set by
1599 || insn_sets_resource_p (target_main
, &needed
, true)
1600 || insn_sets_resource_p (target_main
, &set
, true))
1603 /* Insns we pass may not set either NEEDED or SET, so merge them for
1605 needed
.memory
|= set
.memory
;
1606 needed
.unch_memory
|= set
.unch_memory
;
1607 IOR_HARD_REG_SET (needed
.regs
, set
.regs
);
1609 /* This insn isn't redundant if it conflicts with an insn that either is
1610 or will be in a delay slot of TARGET. */
1614 if (insn_sets_resource_p (XEXP (delay_list
, 0), &needed
, true))
1616 delay_list
= XEXP (delay_list
, 1);
1619 if (NONJUMP_INSN_P (target
) && GET_CODE (PATTERN (target
)) == SEQUENCE
)
1620 for (i
= 1; i
< XVECLEN (PATTERN (target
), 0); i
++)
1621 if (insn_sets_resource_p (XVECEXP (PATTERN (target
), 0, i
), &needed
,
1625 /* Scan backwards until we reach a label or an insn that uses something
1626 INSN sets or sets something insn uses or sets. */
1628 for (trial
= PREV_INSN (target
),
1629 insns_to_search
= MAX_DELAY_SLOT_INSN_SEARCH
;
1630 trial
&& !LABEL_P (trial
) && insns_to_search
> 0;
1631 trial
= PREV_INSN (trial
))
1633 if (!INSN_P (trial
))
1637 pat
= PATTERN (trial
);
1638 if (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
)
1641 if (GET_CODE (pat
) == SEQUENCE
)
1643 bool annul_p
= false;
1644 rtx control
= XVECEXP (pat
, 0, 0);
1646 /* If this is a CALL_INSN and its delay slots, it is hard to track
1647 the resource needs properly, so give up. */
1648 if (CALL_P (control
))
1651 /* If this is an INSN or JUMP_INSN with delayed effects, it
1652 is hard to track the resource needs properly, so give up. */
1654 #ifdef INSN_SETS_ARE_DELAYED
1655 if (INSN_SETS_ARE_DELAYED (control
))
1659 #ifdef INSN_REFERENCES_ARE_DELAYED
1660 if (INSN_REFERENCES_ARE_DELAYED (control
))
1664 if (JUMP_P (control
))
1665 annul_p
= INSN_ANNULLED_BRANCH_P (control
);
1667 /* See if any of the insns in the delay slot match, updating
1668 resource requirements as we go. */
1669 for (i
= XVECLEN (pat
, 0) - 1; i
> 0; i
--)
1671 rtx candidate
= XVECEXP (pat
, 0, i
);
1673 /* If an insn will be annulled if the branch is false, it isn't
1674 considered as a possible duplicate insn. */
1675 if (rtx_equal_p (PATTERN (candidate
), ipat
)
1676 && ! (annul_p
&& INSN_FROM_TARGET_P (candidate
)))
1678 /* Show that this insn will be used in the sequel. */
1679 INSN_FROM_TARGET_P (candidate
) = 0;
1683 /* Unless this is an annulled insn from the target of a branch,
1684 we must stop if it sets anything needed or set by INSN. */
1685 if ((!annul_p
|| !INSN_FROM_TARGET_P (candidate
))
1686 && insn_sets_resource_p (candidate
, &needed
, true))
1690 /* If the insn requiring the delay slot conflicts with INSN, we
1692 if (insn_sets_resource_p (control
, &needed
, true))
1697 /* See if TRIAL is the same as INSN. */
1698 pat
= PATTERN (trial
);
1699 if (rtx_equal_p (pat
, ipat
))
1702 /* Can't go any further if TRIAL conflicts with INSN. */
1703 if (insn_sets_resource_p (trial
, &needed
, true))
1711 /* Return 1 if THREAD can only be executed in one way. If LABEL is nonzero,
1712 it is the target of the branch insn being scanned. If ALLOW_FALLTHROUGH
1713 is nonzero, we are allowed to fall into this thread; otherwise, we are
1716 If LABEL is used more than one or we pass a label other than LABEL before
1717 finding an active insn, we do not own this thread. */
1720 own_thread_p (rtx thread
, rtx label
, int allow_fallthrough
)
1725 /* We don't own the function end. */
1726 if (thread
== 0 || ANY_RETURN_P (thread
))
1729 /* Get the first active insn, or THREAD, if it is an active insn. */
1730 active_insn
= next_active_insn (PREV_INSN (thread
));
1732 for (insn
= thread
; insn
!= active_insn
; insn
= NEXT_INSN (insn
))
1734 && (insn
!= label
|| LABEL_NUSES (insn
) != 1))
1737 if (allow_fallthrough
)
1740 /* Ensure that we reach a BARRIER before any insn or label. */
1741 for (insn
= prev_nonnote_insn (thread
);
1742 insn
== 0 || !BARRIER_P (insn
);
1743 insn
= prev_nonnote_insn (insn
))
1746 || (NONJUMP_INSN_P (insn
)
1747 && GET_CODE (PATTERN (insn
)) != USE
1748 && GET_CODE (PATTERN (insn
)) != CLOBBER
))
1754 /* Called when INSN is being moved from a location near the target of a jump.
1755 We leave a marker of the form (use (INSN)) immediately in front
1756 of WHERE for mark_target_live_regs. These markers will be deleted when
1759 We used to try to update the live status of registers if WHERE is at
1760 the start of a basic block, but that can't work since we may remove a
1761 BARRIER in relax_delay_slots. */
1764 update_block (rtx insn
, rtx where
)
1766 /* Ignore if this was in a delay slot and it came from the target of
1768 if (INSN_FROM_TARGET_P (insn
))
1771 emit_insn_before (gen_rtx_USE (VOIDmode
, insn
), where
);
1773 /* INSN might be making a value live in a block where it didn't use to
1774 be. So recompute liveness information for this block. */
1776 incr_ticks_for_insn (insn
);
1779 /* Similar to REDIRECT_JUMP except that we update the BB_TICKS entry for
1780 the basic block containing the jump. */
1783 reorg_redirect_jump (rtx jump
, rtx nlabel
)
1785 incr_ticks_for_insn (jump
);
1786 return redirect_jump (jump
, nlabel
, 1);
1789 /* Called when INSN is being moved forward into a delay slot of DELAYED_INSN.
1790 We check every instruction between INSN and DELAYED_INSN for REG_DEAD notes
1791 that reference values used in INSN. If we find one, then we move the
1792 REG_DEAD note to INSN.
1794 This is needed to handle the case where a later insn (after INSN) has a
1795 REG_DEAD note for a register used by INSN, and this later insn subsequently
1796 gets moved before a CODE_LABEL because it is a redundant insn. In this
1797 case, mark_target_live_regs may be confused into thinking the register
1798 is dead because it sees a REG_DEAD note immediately before a CODE_LABEL. */
1801 update_reg_dead_notes (rtx insn
, rtx delayed_insn
)
1805 for (p
= next_nonnote_insn (insn
); p
!= delayed_insn
;
1806 p
= next_nonnote_insn (p
))
1807 for (link
= REG_NOTES (p
); link
; link
= next
)
1809 next
= XEXP (link
, 1);
1811 if (REG_NOTE_KIND (link
) != REG_DEAD
1812 || !REG_P (XEXP (link
, 0)))
1815 if (reg_referenced_p (XEXP (link
, 0), PATTERN (insn
)))
1817 /* Move the REG_DEAD note from P to INSN. */
1818 remove_note (p
, link
);
1819 XEXP (link
, 1) = REG_NOTES (insn
);
1820 REG_NOTES (insn
) = link
;
1825 /* Called when an insn redundant with start_insn is deleted. If there
1826 is a REG_DEAD note for the target of start_insn between start_insn
1827 and stop_insn, then the REG_DEAD note needs to be deleted since the
1828 value no longer dies there.
1830 If the REG_DEAD note isn't deleted, then mark_target_live_regs may be
1831 confused into thinking the register is dead. */
1834 fix_reg_dead_note (rtx start_insn
, rtx stop_insn
)
1838 for (p
= next_nonnote_insn (start_insn
); p
!= stop_insn
;
1839 p
= next_nonnote_insn (p
))
1840 for (link
= REG_NOTES (p
); link
; link
= next
)
1842 next
= XEXP (link
, 1);
1844 if (REG_NOTE_KIND (link
) != REG_DEAD
1845 || !REG_P (XEXP (link
, 0)))
1848 if (reg_set_p (XEXP (link
, 0), PATTERN (start_insn
)))
1850 remove_note (p
, link
);
1856 /* Delete any REG_UNUSED notes that exist on INSN but not on REDUNDANT_INSN.
1858 This handles the case of udivmodXi4 instructions which optimize their
1859 output depending on whether any REG_UNUSED notes are present.
1860 we must make sure that INSN calculates as many results as REDUNDANT_INSN
1864 update_reg_unused_notes (rtx insn
, rtx redundant_insn
)
1868 for (link
= REG_NOTES (insn
); link
; link
= next
)
1870 next
= XEXP (link
, 1);
1872 if (REG_NOTE_KIND (link
) != REG_UNUSED
1873 || !REG_P (XEXP (link
, 0)))
1876 if (! find_regno_note (redundant_insn
, REG_UNUSED
,
1877 REGNO (XEXP (link
, 0))))
1878 remove_note (insn
, link
);
1882 /* Return the label before INSN, or put a new label there. */
1885 get_label_before (rtx insn
)
1889 /* Find an existing label at this point
1890 or make a new one if there is none. */
1891 label
= prev_nonnote_insn (insn
);
1893 if (label
== 0 || !LABEL_P (label
))
1895 rtx prev
= PREV_INSN (insn
);
1897 label
= gen_label_rtx ();
1898 emit_label_after (label
, prev
);
1899 LABEL_NUSES (label
) = 0;
1904 /* Scan a function looking for insns that need a delay slot and find insns to
1905 put into the delay slot.
1907 NON_JUMPS_P is nonzero if we are to only try to fill non-jump insns (such
1908 as calls). We do these first since we don't want jump insns (that are
1909 easier to fill) to get the only insns that could be used for non-jump insns.
1910 When it is zero, only try to fill JUMP_INSNs.
1912 When slots are filled in this manner, the insns (including the
1913 delay_insn) are put together in a SEQUENCE rtx. In this fashion,
1914 it is possible to tell whether a delay slot has really been filled
1915 or not. `final' knows how to deal with this, by communicating
1916 through FINAL_SEQUENCE. */
1919 fill_simple_delay_slots (int non_jumps_p
)
1921 rtx insn
, pat
, trial
, next_trial
;
1923 int num_unfilled_slots
= unfilled_slots_next
- unfilled_slots_base
;
1924 struct resources needed
, set
;
1925 int slots_to_fill
, slots_filled
;
1928 for (i
= 0; i
< num_unfilled_slots
; i
++)
1931 /* Get the next insn to fill. If it has already had any slots assigned,
1932 we can't do anything with it. Maybe we'll improve this later. */
1934 insn
= unfilled_slots_base
[i
];
1936 || INSN_DELETED_P (insn
)
1937 || (NONJUMP_INSN_P (insn
)
1938 && GET_CODE (PATTERN (insn
)) == SEQUENCE
)
1939 || (JUMP_P (insn
) && non_jumps_p
)
1940 || (!JUMP_P (insn
) && ! non_jumps_p
))
1943 /* It may have been that this insn used to need delay slots, but
1944 now doesn't; ignore in that case. This can happen, for example,
1945 on the HP PA RISC, where the number of delay slots depends on
1946 what insns are nearby. */
1947 slots_to_fill
= num_delay_slots (insn
);
1949 /* Some machine description have defined instructions to have
1950 delay slots only in certain circumstances which may depend on
1951 nearby insns (which change due to reorg's actions).
1953 For example, the PA port normally has delay slots for unconditional
1956 However, the PA port claims such jumps do not have a delay slot
1957 if they are immediate successors of certain CALL_INSNs. This
1958 allows the port to favor filling the delay slot of the call with
1959 the unconditional jump. */
1960 if (slots_to_fill
== 0)
1963 /* This insn needs, or can use, some delay slots. SLOTS_TO_FILL
1964 says how many. After initialization, first try optimizing
1967 nop add %o7,.-L1,%o7
1971 If this case applies, the delay slot of the call is filled with
1972 the unconditional jump. This is done first to avoid having the
1973 delay slot of the call filled in the backward scan. Also, since
1974 the unconditional jump is likely to also have a delay slot, that
1975 insn must exist when it is subsequently scanned.
1977 This is tried on each insn with delay slots as some machines
1978 have insns which perform calls, but are not represented as
1985 flags
= get_jump_flags (insn
, JUMP_LABEL (insn
));
1987 flags
= get_jump_flags (insn
, NULL_RTX
);
1989 if ((trial
= next_active_insn (insn
))
1991 && simplejump_p (trial
)
1992 && eligible_for_delay (insn
, slots_filled
, trial
, flags
)
1993 && no_labels_between_p (insn
, trial
)
1994 && ! can_throw_internal (trial
))
1998 delay_list
= add_to_delay_list (trial
, delay_list
);
2000 /* TRIAL may have had its delay slot filled, then unfilled. When
2001 the delay slot is unfilled, TRIAL is placed back on the unfilled
2002 slots obstack. Unfortunately, it is placed on the end of the
2003 obstack, not in its original location. Therefore, we must search
2004 from entry i + 1 to the end of the unfilled slots obstack to
2005 try and find TRIAL. */
2006 tmp
= &unfilled_slots_base
[i
+ 1];
2007 while (*tmp
!= trial
&& tmp
!= unfilled_slots_next
)
2010 /* Remove the unconditional jump from consideration for delay slot
2011 filling and unthread it. */
2015 rtx next
= NEXT_INSN (trial
);
2016 rtx prev
= PREV_INSN (trial
);
2018 NEXT_INSN (prev
) = next
;
2020 PREV_INSN (next
) = prev
;
2024 /* Now, scan backwards from the insn to search for a potential
2025 delay-slot candidate. Stop searching when a label or jump is hit.
2027 For each candidate, if it is to go into the delay slot (moved
2028 forward in execution sequence), it must not need or set any resources
2029 that were set by later insns and must not set any resources that
2030 are needed for those insns.
2032 The delay slot insn itself sets resources unless it is a call
2033 (in which case the called routine, not the insn itself, is doing
2036 if (slots_filled
< slots_to_fill
)
2038 CLEAR_RESOURCE (&needed
);
2039 CLEAR_RESOURCE (&set
);
2040 mark_set_resources (insn
, &set
, 0, MARK_SRC_DEST
);
2041 mark_referenced_resources (insn
, &needed
, false);
2043 for (trial
= prev_nonnote_insn (insn
); ! stop_search_p (trial
, 1);
2046 next_trial
= prev_nonnote_insn (trial
);
2048 /* This must be an INSN or CALL_INSN. */
2049 pat
= PATTERN (trial
);
2051 /* Stand-alone USE and CLOBBER are just for flow. */
2052 if (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
)
2055 /* Check for resource conflict first, to avoid unnecessary
2057 if (! insn_references_resource_p (trial
, &set
, true)
2058 && ! insn_sets_resource_p (trial
, &set
, true)
2059 && ! insn_sets_resource_p (trial
, &needed
, true)
2061 /* Can't separate set of cc0 from its use. */
2062 && ! (reg_mentioned_p (cc0_rtx
, pat
) && ! sets_cc0_p (pat
))
2064 && ! can_throw_internal (trial
))
2066 trial
= try_split (pat
, trial
, 1);
2067 next_trial
= prev_nonnote_insn (trial
);
2068 if (eligible_for_delay (insn
, slots_filled
, trial
, flags
))
2070 /* In this case, we are searching backward, so if we
2071 find insns to put on the delay list, we want
2072 to put them at the head, rather than the
2073 tail, of the list. */
2075 update_reg_dead_notes (trial
, insn
);
2076 delay_list
= gen_rtx_INSN_LIST (VOIDmode
,
2078 update_block (trial
, trial
);
2079 delete_related_insns (trial
);
2080 if (slots_to_fill
== ++slots_filled
)
2086 mark_set_resources (trial
, &set
, 0, MARK_SRC_DEST_CALL
);
2087 mark_referenced_resources (trial
, &needed
, true);
2091 /* If all needed slots haven't been filled, we come here. */
2093 /* Try to optimize case of jumping around a single insn. */
2094 #if defined(ANNUL_IFFALSE_SLOTS) || defined(ANNUL_IFTRUE_SLOTS)
2095 if (slots_filled
!= slots_to_fill
2098 && (condjump_p (insn
) || condjump_in_parallel_p (insn
))
2099 && !ANY_RETURN_P (JUMP_LABEL (insn
)))
2101 delay_list
= optimize_skip (insn
);
2107 /* Try to get insns from beyond the insn needing the delay slot.
2108 These insns can neither set or reference resources set in insns being
2109 skipped, cannot set resources in the insn being skipped, and, if this
2110 is a CALL_INSN (or a CALL_INSN is passed), cannot trap (because the
2111 call might not return).
2113 There used to be code which continued past the target label if
2114 we saw all uses of the target label. This code did not work,
2115 because it failed to account for some instructions which were
2116 both annulled and marked as from the target. This can happen as a
2117 result of optimize_skip. Since this code was redundant with
2118 fill_eager_delay_slots anyways, it was just deleted. */
2120 if (slots_filled
!= slots_to_fill
2121 /* If this instruction could throw an exception which is
2122 caught in the same function, then it's not safe to fill
2123 the delay slot with an instruction from beyond this
2124 point. For example, consider:
2135 Even though `i' is a local variable, we must be sure not
2136 to put `i = 3' in the delay slot if `f' might throw an
2139 Presumably, we should also check to see if we could get
2140 back to this function via `setjmp'. */
2141 && ! can_throw_internal (insn
)
2143 || ((condjump_p (insn
) || condjump_in_parallel_p (insn
))
2144 && ! simplejump_p (insn
)
2145 && !ANY_RETURN_P (JUMP_LABEL (insn
)))))
2147 /* Invariant: If insn is a JUMP_INSN, the insn's jump
2148 label. Otherwise, zero. */
2150 int maybe_never
= 0;
2151 rtx pat
, trial_delay
;
2153 CLEAR_RESOURCE (&needed
);
2154 CLEAR_RESOURCE (&set
);
2158 mark_set_resources (insn
, &set
, 0, MARK_SRC_DEST_CALL
);
2159 mark_referenced_resources (insn
, &needed
, true);
2164 mark_set_resources (insn
, &set
, 0, MARK_SRC_DEST_CALL
);
2165 mark_referenced_resources (insn
, &needed
, true);
2167 target
= JUMP_LABEL (insn
);
2170 if (target
== 0 || ANY_RETURN_P (target
))
2171 for (trial
= next_nonnote_insn (insn
); !stop_search_p (trial
, 1);
2174 next_trial
= next_nonnote_insn (trial
);
2176 /* This must be an INSN or CALL_INSN. */
2177 pat
= PATTERN (trial
);
2179 /* Stand-alone USE and CLOBBER are just for flow. */
2180 if (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
)
2183 /* If this already has filled delay slots, get the insn needing
2185 if (GET_CODE (pat
) == SEQUENCE
)
2186 trial_delay
= XVECEXP (pat
, 0, 0);
2188 trial_delay
= trial
;
2190 /* Stop our search when seeing a jump. */
2191 if (JUMP_P (trial_delay
))
2194 /* See if we have a resource problem before we try to
2196 if (GET_CODE (pat
) != SEQUENCE
2197 && ! insn_references_resource_p (trial
, &set
, true)
2198 && ! insn_sets_resource_p (trial
, &set
, true)
2199 && ! insn_sets_resource_p (trial
, &needed
, true)
2201 && ! (reg_mentioned_p (cc0_rtx
, pat
) && ! sets_cc0_p (pat
))
2203 && ! (maybe_never
&& may_trap_or_fault_p (pat
))
2204 && (trial
= try_split (pat
, trial
, 0))
2205 && eligible_for_delay (insn
, slots_filled
, trial
, flags
)
2206 && ! can_throw_internal(trial
))
2208 next_trial
= next_nonnote_insn (trial
);
2209 delay_list
= add_to_delay_list (trial
, delay_list
);
2212 if (reg_mentioned_p (cc0_rtx
, pat
))
2213 link_cc0_insns (trial
);
2216 delete_related_insns (trial
);
2217 if (slots_to_fill
== ++slots_filled
)
2222 mark_set_resources (trial
, &set
, 0, MARK_SRC_DEST_CALL
);
2223 mark_referenced_resources (trial
, &needed
, true);
2225 /* Ensure we don't put insns between the setting of cc and the
2226 comparison by moving a setting of cc into an earlier delay
2227 slot since these insns could clobber the condition code. */
2230 /* If this is a call or jump, we might not get here. */
2231 if (CALL_P (trial_delay
)
2232 || JUMP_P (trial_delay
))
2236 /* If there are slots left to fill and our search was stopped by an
2237 unconditional branch, try the insn at the branch target. We can
2238 redirect the branch if it works.
2240 Don't do this if the insn at the branch target is a branch. */
2241 if (slots_to_fill
!= slots_filled
2243 && jump_to_label_p (trial
)
2244 && simplejump_p (trial
)
2245 && (target
== 0 || JUMP_LABEL (trial
) == target
)
2246 && (next_trial
= next_active_insn (JUMP_LABEL (trial
))) != 0
2247 && ! (NONJUMP_INSN_P (next_trial
)
2248 && GET_CODE (PATTERN (next_trial
)) == SEQUENCE
)
2249 && !JUMP_P (next_trial
)
2250 && ! insn_references_resource_p (next_trial
, &set
, true)
2251 && ! insn_sets_resource_p (next_trial
, &set
, true)
2252 && ! insn_sets_resource_p (next_trial
, &needed
, true)
2254 && ! reg_mentioned_p (cc0_rtx
, PATTERN (next_trial
))
2256 && ! (maybe_never
&& may_trap_or_fault_p (PATTERN (next_trial
)))
2257 && (next_trial
= try_split (PATTERN (next_trial
), next_trial
, 0))
2258 && eligible_for_delay (insn
, slots_filled
, next_trial
, flags
)
2259 && ! can_throw_internal (trial
))
2261 /* See comment in relax_delay_slots about necessity of using
2262 next_real_insn here. */
2263 rtx new_label
= next_real_insn (next_trial
);
2266 new_label
= get_label_before (new_label
);
2268 new_label
= find_end_label (simple_return_rtx
);
2273 = add_to_delay_list (copy_delay_slot_insn (next_trial
),
2276 reorg_redirect_jump (trial
, new_label
);
2278 /* If we merged because we both jumped to the same place,
2279 redirect the original insn also. */
2281 reorg_redirect_jump (insn
, new_label
);
2286 /* If this is an unconditional jump, then try to get insns from the
2287 target of the jump. */
2289 && simplejump_p (insn
)
2290 && slots_filled
!= slots_to_fill
)
2292 = fill_slots_from_thread (insn
, const_true_rtx
,
2293 next_active_insn (JUMP_LABEL (insn
)),
2295 own_thread_p (JUMP_LABEL (insn
),
2296 JUMP_LABEL (insn
), 0),
2297 slots_to_fill
, &slots_filled
,
2301 unfilled_slots_base
[i
]
2302 = emit_delay_sequence (insn
, delay_list
, slots_filled
);
2304 if (slots_to_fill
== slots_filled
)
2305 unfilled_slots_base
[i
] = 0;
2307 note_delay_statistics (slots_filled
, 0);
2311 /* Follow any unconditional jump at LABEL, for the purpose of redirecting JUMP;
2312 return the ultimate label reached by any such chain of jumps.
2313 Return a suitable return rtx if the chain ultimately leads to a
2315 If LABEL is not followed by a jump, return LABEL.
2316 If the chain loops or we can't find end, return LABEL,
2317 since that tells caller to avoid changing the insn.
2318 If the returned label is obtained by following a REG_CROSSING_JUMP
2319 jump, set *CROSSING to true, otherwise set it to false. */
2322 follow_jumps (rtx label
, rtx jump
, bool *crossing
)
2330 if (ANY_RETURN_P (label
))
2334 && (insn
= next_active_insn (value
)) != 0
2336 && JUMP_LABEL (insn
) != NULL_RTX
2337 && ((any_uncondjump_p (insn
) && onlyjump_p (insn
))
2338 || ANY_RETURN_P (PATTERN (insn
)))
2339 && (next
= NEXT_INSN (insn
))
2340 && BARRIER_P (next
));
2343 rtx this_label
= JUMP_LABEL (insn
);
2346 /* If we have found a cycle, make the insn jump to itself. */
2347 if (this_label
== label
)
2349 if (ANY_RETURN_P (this_label
))
2351 tem
= next_active_insn (this_label
);
2353 && (GET_CODE (PATTERN (tem
)) == ADDR_VEC
2354 || GET_CODE (PATTERN (tem
)) == ADDR_DIFF_VEC
))
2357 if (!targetm
.can_follow_jump (jump
, insn
))
2361 = find_reg_note (insn
, REG_CROSSING_JUMP
, NULL_RTX
) != NULL_RTX
;
2369 /* Try to find insns to place in delay slots.
2371 INSN is the jump needing SLOTS_TO_FILL delay slots. It tests CONDITION
2372 or is an unconditional branch if CONDITION is const_true_rtx.
2373 *PSLOTS_FILLED is updated with the number of slots that we have filled.
2375 THREAD is a flow-of-control, either the insns to be executed if the
2376 branch is true or if the branch is false, THREAD_IF_TRUE says which.
2378 OPPOSITE_THREAD is the thread in the opposite direction. It is used
2379 to see if any potential delay slot insns set things needed there.
2381 LIKELY is nonzero if it is extremely likely that the branch will be
2382 taken and THREAD_IF_TRUE is set. This is used for the branch at the
2383 end of a loop back up to the top.
2385 OWN_THREAD and OWN_OPPOSITE_THREAD are true if we are the only user of the
2386 thread. I.e., it is the fallthrough code of our jump or the target of the
2387 jump when we are the only jump going there.
2389 If OWN_THREAD is false, it must be the "true" thread of a jump. In that
2390 case, we can only take insns from the head of the thread for our delay
2391 slot. We then adjust the jump to point after the insns we have taken. */
2394 fill_slots_from_thread (rtx insn
, rtx condition
, rtx thread
,
2395 rtx opposite_thread
, int likely
, int thread_if_true
,
2396 int own_thread
, int slots_to_fill
,
2397 int *pslots_filled
, rtx delay_list
)
2400 struct resources opposite_needed
, set
, needed
;
2406 /* Validate our arguments. */
2407 gcc_assert(condition
!= const_true_rtx
|| thread_if_true
);
2408 gcc_assert(own_thread
|| thread_if_true
);
2410 flags
= get_jump_flags (insn
, JUMP_LABEL (insn
));
2412 /* If our thread is the end of subroutine, we can't get any delay
2414 if (thread
== NULL_RTX
|| ANY_RETURN_P (thread
))
2417 /* If this is an unconditional branch, nothing is needed at the
2418 opposite thread. Otherwise, compute what is needed there. */
2419 if (condition
== const_true_rtx
)
2420 CLEAR_RESOURCE (&opposite_needed
);
2422 mark_target_live_regs (get_insns (), opposite_thread
, &opposite_needed
);
2424 /* If the insn at THREAD can be split, do it here to avoid having to
2425 update THREAD and NEW_THREAD if it is done in the loop below. Also
2426 initialize NEW_THREAD. */
2428 new_thread
= thread
= try_split (PATTERN (thread
), thread
, 0);
2430 /* Scan insns at THREAD. We are looking for an insn that can be removed
2431 from THREAD (it neither sets nor references resources that were set
2432 ahead of it and it doesn't set anything needs by the insns ahead of
2433 it) and that either can be placed in an annulling insn or aren't
2434 needed at OPPOSITE_THREAD. */
2436 CLEAR_RESOURCE (&needed
);
2437 CLEAR_RESOURCE (&set
);
2439 /* If we do not own this thread, we must stop as soon as we find
2440 something that we can't put in a delay slot, since all we can do
2441 is branch into THREAD at a later point. Therefore, labels stop
2442 the search if this is not the `true' thread. */
2444 for (trial
= thread
;
2445 ! stop_search_p (trial
, ! thread_if_true
) && (! lose
|| own_thread
);
2446 trial
= next_nonnote_insn (trial
))
2450 /* If we have passed a label, we no longer own this thread. */
2451 if (LABEL_P (trial
))
2457 pat
= PATTERN (trial
);
2458 if (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
)
2461 /* If TRIAL conflicts with the insns ahead of it, we lose. Also,
2462 don't separate or copy insns that set and use CC0. */
2463 if (! insn_references_resource_p (trial
, &set
, true)
2464 && ! insn_sets_resource_p (trial
, &set
, true)
2465 && ! insn_sets_resource_p (trial
, &needed
, true)
2467 && ! (reg_mentioned_p (cc0_rtx
, pat
)
2468 && (! own_thread
|| ! sets_cc0_p (pat
)))
2470 && ! can_throw_internal (trial
))
2474 /* If TRIAL is redundant with some insn before INSN, we don't
2475 actually need to add it to the delay list; we can merely pretend
2477 if ((prior_insn
= redundant_insn (trial
, insn
, delay_list
)))
2479 fix_reg_dead_note (prior_insn
, insn
);
2482 update_block (trial
, thread
);
2483 if (trial
== thread
)
2485 thread
= next_active_insn (thread
);
2486 if (new_thread
== trial
)
2487 new_thread
= thread
;
2490 delete_related_insns (trial
);
2494 update_reg_unused_notes (prior_insn
, trial
);
2495 new_thread
= next_active_insn (trial
);
2501 /* There are two ways we can win: If TRIAL doesn't set anything
2502 needed at the opposite thread and can't trap, or if it can
2503 go into an annulled delay slot. */
2505 && (condition
== const_true_rtx
2506 || (! insn_sets_resource_p (trial
, &opposite_needed
, true)
2507 && ! may_trap_or_fault_p (pat
)
2508 && ! RTX_FRAME_RELATED_P (trial
))))
2511 trial
= try_split (pat
, trial
, 0);
2512 if (new_thread
== old_trial
)
2514 if (thread
== old_trial
)
2516 pat
= PATTERN (trial
);
2517 if (eligible_for_delay (insn
, *pslots_filled
, trial
, flags
))
2521 #ifdef ANNUL_IFTRUE_SLOTS
2524 #ifdef ANNUL_IFFALSE_SLOTS
2530 trial
= try_split (pat
, trial
, 0);
2531 if (new_thread
== old_trial
)
2533 if (thread
== old_trial
)
2535 pat
= PATTERN (trial
);
2536 if ((must_annul
|| delay_list
== NULL
) && (thread_if_true
2537 ? check_annul_list_true_false (0, delay_list
)
2538 && eligible_for_annul_false (insn
, *pslots_filled
, trial
, flags
)
2539 : check_annul_list_true_false (1, delay_list
)
2540 && eligible_for_annul_true (insn
, *pslots_filled
, trial
, flags
)))
2548 if (reg_mentioned_p (cc0_rtx
, pat
))
2549 link_cc0_insns (trial
);
2552 /* If we own this thread, delete the insn. If this is the
2553 destination of a branch, show that a basic block status
2554 may have been updated. In any case, mark the new
2555 starting point of this thread. */
2560 update_block (trial
, thread
);
2561 if (trial
== thread
)
2563 thread
= next_active_insn (thread
);
2564 if (new_thread
== trial
)
2565 new_thread
= thread
;
2568 /* We are moving this insn, not deleting it. We must
2569 temporarily increment the use count on any referenced
2570 label lest it be deleted by delete_related_insns. */
2571 for (note
= REG_NOTES (trial
);
2573 note
= XEXP (note
, 1))
2574 if (REG_NOTE_KIND (note
) == REG_LABEL_OPERAND
2575 || REG_NOTE_KIND (note
) == REG_LABEL_TARGET
)
2577 /* REG_LABEL_OPERAND could be
2578 NOTE_INSN_DELETED_LABEL too. */
2579 if (LABEL_P (XEXP (note
, 0)))
2580 LABEL_NUSES (XEXP (note
, 0))++;
2582 gcc_assert (REG_NOTE_KIND (note
)
2583 == REG_LABEL_OPERAND
);
2585 if (jump_to_label_p (trial
))
2586 LABEL_NUSES (JUMP_LABEL (trial
))++;
2588 delete_related_insns (trial
);
2590 for (note
= REG_NOTES (trial
);
2592 note
= XEXP (note
, 1))
2593 if (REG_NOTE_KIND (note
) == REG_LABEL_OPERAND
2594 || REG_NOTE_KIND (note
) == REG_LABEL_TARGET
)
2596 /* REG_LABEL_OPERAND could be
2597 NOTE_INSN_DELETED_LABEL too. */
2598 if (LABEL_P (XEXP (note
, 0)))
2599 LABEL_NUSES (XEXP (note
, 0))--;
2601 gcc_assert (REG_NOTE_KIND (note
)
2602 == REG_LABEL_OPERAND
);
2604 if (jump_to_label_p (trial
))
2605 LABEL_NUSES (JUMP_LABEL (trial
))--;
2608 new_thread
= next_active_insn (trial
);
2610 temp
= own_thread
? trial
: copy_delay_slot_insn (trial
);
2612 INSN_FROM_TARGET_P (temp
) = 1;
2614 delay_list
= add_to_delay_list (temp
, delay_list
);
2616 if (slots_to_fill
== ++(*pslots_filled
))
2618 /* Even though we have filled all the slots, we
2619 may be branching to a location that has a
2620 redundant insn. Skip any if so. */
2621 while (new_thread
&& ! own_thread
2622 && ! insn_sets_resource_p (new_thread
, &set
, true)
2623 && ! insn_sets_resource_p (new_thread
, &needed
,
2625 && ! insn_references_resource_p (new_thread
,
2628 = redundant_insn (new_thread
, insn
,
2631 /* We know we do not own the thread, so no need
2632 to call update_block and delete_insn. */
2633 fix_reg_dead_note (prior_insn
, insn
);
2634 update_reg_unused_notes (prior_insn
, new_thread
);
2635 new_thread
= next_active_insn (new_thread
);
2645 /* This insn can't go into a delay slot. */
2647 mark_set_resources (trial
, &set
, 0, MARK_SRC_DEST_CALL
);
2648 mark_referenced_resources (trial
, &needed
, true);
2650 /* Ensure we don't put insns between the setting of cc and the comparison
2651 by moving a setting of cc into an earlier delay slot since these insns
2652 could clobber the condition code. */
2655 /* If this insn is a register-register copy and the next insn has
2656 a use of our destination, change it to use our source. That way,
2657 it will become a candidate for our delay slot the next time
2658 through this loop. This case occurs commonly in loops that
2661 We could check for more complex cases than those tested below,
2662 but it doesn't seem worth it. It might also be a good idea to try
2663 to swap the two insns. That might do better.
2665 We can't do this if the next insn modifies our destination, because
2666 that would make the replacement into the insn invalid. We also can't
2667 do this if it modifies our source, because it might be an earlyclobber
2668 operand. This latter test also prevents updating the contents of
2669 a PRE_INC. We also can't do this if there's overlap of source and
2670 destination. Overlap may happen for larger-than-register-size modes. */
2672 if (NONJUMP_INSN_P (trial
) && GET_CODE (pat
) == SET
2673 && REG_P (SET_SRC (pat
))
2674 && REG_P (SET_DEST (pat
))
2675 && !reg_overlap_mentioned_p (SET_DEST (pat
), SET_SRC (pat
)))
2677 rtx next
= next_nonnote_insn (trial
);
2679 if (next
&& NONJUMP_INSN_P (next
)
2680 && GET_CODE (PATTERN (next
)) != USE
2681 && ! reg_set_p (SET_DEST (pat
), next
)
2682 && ! reg_set_p (SET_SRC (pat
), next
)
2683 && reg_referenced_p (SET_DEST (pat
), PATTERN (next
))
2684 && ! modified_in_p (SET_DEST (pat
), next
))
2685 validate_replace_rtx (SET_DEST (pat
), SET_SRC (pat
), next
);
2689 /* If we stopped on a branch insn that has delay slots, see if we can
2690 steal some of the insns in those slots. */
2691 if (trial
&& NONJUMP_INSN_P (trial
)
2692 && GET_CODE (PATTERN (trial
)) == SEQUENCE
2693 && JUMP_P (XVECEXP (PATTERN (trial
), 0, 0)))
2695 /* If this is the `true' thread, we will want to follow the jump,
2696 so we can only do this if we have taken everything up to here. */
2697 if (thread_if_true
&& trial
== new_thread
)
2700 = steal_delay_list_from_target (insn
, condition
, PATTERN (trial
),
2701 delay_list
, &set
, &needed
,
2702 &opposite_needed
, slots_to_fill
,
2703 pslots_filled
, &must_annul
,
2705 /* If we owned the thread and are told that it branched
2706 elsewhere, make sure we own the thread at the new location. */
2707 if (own_thread
&& trial
!= new_thread
)
2708 own_thread
= own_thread_p (new_thread
, new_thread
, 0);
2710 else if (! thread_if_true
)
2712 = steal_delay_list_from_fallthrough (insn
, condition
,
2714 delay_list
, &set
, &needed
,
2715 &opposite_needed
, slots_to_fill
,
2716 pslots_filled
, &must_annul
);
2719 /* If we haven't found anything for this delay slot and it is very
2720 likely that the branch will be taken, see if the insn at our target
2721 increments or decrements a register with an increment that does not
2722 depend on the destination register. If so, try to place the opposite
2723 arithmetic insn after the jump insn and put the arithmetic insn in the
2724 delay slot. If we can't do this, return. */
2725 if (delay_list
== 0 && likely
2726 && new_thread
&& !ANY_RETURN_P (new_thread
)
2727 && NONJUMP_INSN_P (new_thread
)
2728 && !RTX_FRAME_RELATED_P (new_thread
)
2729 && GET_CODE (PATTERN (new_thread
)) != ASM_INPUT
2730 && asm_noperands (PATTERN (new_thread
)) < 0)
2732 rtx pat
= PATTERN (new_thread
);
2737 pat
= PATTERN (trial
);
2739 if (!NONJUMP_INSN_P (trial
)
2740 || GET_CODE (pat
) != SET
2741 || ! eligible_for_delay (insn
, 0, trial
, flags
)
2742 || can_throw_internal (trial
))
2745 dest
= SET_DEST (pat
), src
= SET_SRC (pat
);
2746 if ((GET_CODE (src
) == PLUS
|| GET_CODE (src
) == MINUS
)
2747 && rtx_equal_p (XEXP (src
, 0), dest
)
2748 && (!FLOAT_MODE_P (GET_MODE (src
))
2749 || flag_unsafe_math_optimizations
)
2750 && ! reg_overlap_mentioned_p (dest
, XEXP (src
, 1))
2751 && ! side_effects_p (pat
))
2753 rtx other
= XEXP (src
, 1);
2757 /* If this is a constant adjustment, use the same code with
2758 the negated constant. Otherwise, reverse the sense of the
2760 if (CONST_INT_P (other
))
2761 new_arith
= gen_rtx_fmt_ee (GET_CODE (src
), GET_MODE (src
), dest
,
2762 negate_rtx (GET_MODE (src
), other
));
2764 new_arith
= gen_rtx_fmt_ee (GET_CODE (src
) == PLUS
? MINUS
: PLUS
,
2765 GET_MODE (src
), dest
, other
);
2767 ninsn
= emit_insn_after (gen_rtx_SET (VOIDmode
, dest
, new_arith
),
2770 if (recog_memoized (ninsn
) < 0
2771 || (extract_insn (ninsn
), ! constrain_operands (1)))
2773 delete_related_insns (ninsn
);
2779 update_block (trial
, thread
);
2780 if (trial
== thread
)
2782 thread
= next_active_insn (thread
);
2783 if (new_thread
== trial
)
2784 new_thread
= thread
;
2786 delete_related_insns (trial
);
2789 new_thread
= next_active_insn (trial
);
2791 ninsn
= own_thread
? trial
: copy_delay_slot_insn (trial
);
2793 INSN_FROM_TARGET_P (ninsn
) = 1;
2795 delay_list
= add_to_delay_list (ninsn
, NULL_RTX
);
2800 if (delay_list
&& must_annul
)
2801 INSN_ANNULLED_BRANCH_P (insn
) = 1;
2803 /* If we are to branch into the middle of this thread, find an appropriate
2804 label or make a new one if none, and redirect INSN to it. If we hit the
2805 end of the function, use the end-of-function label. */
2806 if (new_thread
!= thread
)
2809 bool crossing
= false;
2811 gcc_assert (thread_if_true
);
2813 if (new_thread
&& simplejump_or_return_p (new_thread
)
2814 && redirect_with_delay_list_safe_p (insn
,
2815 JUMP_LABEL (new_thread
),
2817 new_thread
= follow_jumps (JUMP_LABEL (new_thread
), insn
, &crossing
);
2819 if (ANY_RETURN_P (new_thread
))
2820 label
= find_end_label (new_thread
);
2821 else if (LABEL_P (new_thread
))
2824 label
= get_label_before (new_thread
);
2828 reorg_redirect_jump (insn
, label
);
2830 set_unique_reg_note (insn
, REG_CROSSING_JUMP
, NULL_RTX
);
2837 /* Make another attempt to find insns to place in delay slots.
2839 We previously looked for insns located in front of the delay insn
2840 and, for non-jump delay insns, located behind the delay insn.
2842 Here only try to schedule jump insns and try to move insns from either
2843 the target or the following insns into the delay slot. If annulling is
2844 supported, we will be likely to do this. Otherwise, we can do this only
2848 fill_eager_delay_slots (void)
2852 int num_unfilled_slots
= unfilled_slots_next
- unfilled_slots_base
;
2854 for (i
= 0; i
< num_unfilled_slots
; i
++)
2857 rtx target_label
, insn_at_target
, fallthrough_insn
;
2860 int own_fallthrough
;
2861 int prediction
, slots_to_fill
, slots_filled
;
2863 insn
= unfilled_slots_base
[i
];
2865 || INSN_DELETED_P (insn
)
2867 || ! (condjump_p (insn
) || condjump_in_parallel_p (insn
)))
2870 slots_to_fill
= num_delay_slots (insn
);
2871 /* Some machine description have defined instructions to have
2872 delay slots only in certain circumstances which may depend on
2873 nearby insns (which change due to reorg's actions).
2875 For example, the PA port normally has delay slots for unconditional
2878 However, the PA port claims such jumps do not have a delay slot
2879 if they are immediate successors of certain CALL_INSNs. This
2880 allows the port to favor filling the delay slot of the call with
2881 the unconditional jump. */
2882 if (slots_to_fill
== 0)
2886 target_label
= JUMP_LABEL (insn
);
2887 condition
= get_branch_condition (insn
, target_label
);
2892 /* Get the next active fallthrough and target insns and see if we own
2893 them. Then see whether the branch is likely true. We don't need
2894 to do a lot of this for unconditional branches. */
2896 insn_at_target
= first_active_target_insn (target_label
);
2897 own_target
= own_thread_p (target_label
, target_label
, 0);
2899 if (condition
== const_true_rtx
)
2901 own_fallthrough
= 0;
2902 fallthrough_insn
= 0;
2907 fallthrough_insn
= next_active_insn (insn
);
2908 own_fallthrough
= own_thread_p (NEXT_INSN (insn
), NULL_RTX
, 1);
2909 prediction
= mostly_true_jump (insn
);
2912 /* If this insn is expected to branch, first try to get insns from our
2913 target, then our fallthrough insns. If it is not expected to branch,
2914 try the other order. */
2919 = fill_slots_from_thread (insn
, condition
, insn_at_target
,
2920 fallthrough_insn
, prediction
== 2, 1,
2922 slots_to_fill
, &slots_filled
, delay_list
);
2924 if (delay_list
== 0 && own_fallthrough
)
2926 /* Even though we didn't find anything for delay slots,
2927 we might have found a redundant insn which we deleted
2928 from the thread that was filled. So we have to recompute
2929 the next insn at the target. */
2930 target_label
= JUMP_LABEL (insn
);
2931 insn_at_target
= first_active_target_insn (target_label
);
2934 = fill_slots_from_thread (insn
, condition
, fallthrough_insn
,
2935 insn_at_target
, 0, 0,
2937 slots_to_fill
, &slots_filled
,
2943 if (own_fallthrough
)
2945 = fill_slots_from_thread (insn
, condition
, fallthrough_insn
,
2946 insn_at_target
, 0, 0,
2948 slots_to_fill
, &slots_filled
,
2951 if (delay_list
== 0)
2953 = fill_slots_from_thread (insn
, condition
, insn_at_target
,
2954 next_active_insn (insn
), 0, 1,
2956 slots_to_fill
, &slots_filled
,
2961 unfilled_slots_base
[i
]
2962 = emit_delay_sequence (insn
, delay_list
, slots_filled
);
2964 if (slots_to_fill
== slots_filled
)
2965 unfilled_slots_base
[i
] = 0;
2967 note_delay_statistics (slots_filled
, 1);
2971 static void delete_computation (rtx insn
);
2973 /* Recursively delete prior insns that compute the value (used only by INSN
2974 which the caller is deleting) stored in the register mentioned by NOTE
2975 which is a REG_DEAD note associated with INSN. */
2978 delete_prior_computation (rtx note
, rtx insn
)
2981 rtx reg
= XEXP (note
, 0);
2983 for (our_prev
= prev_nonnote_insn (insn
);
2984 our_prev
&& (NONJUMP_INSN_P (our_prev
)
2985 || CALL_P (our_prev
));
2986 our_prev
= prev_nonnote_insn (our_prev
))
2988 rtx pat
= PATTERN (our_prev
);
2990 /* If we reach a CALL which is not calling a const function
2991 or the callee pops the arguments, then give up. */
2992 if (CALL_P (our_prev
)
2993 && (! RTL_CONST_CALL_P (our_prev
)
2994 || GET_CODE (pat
) != SET
|| GET_CODE (SET_SRC (pat
)) != CALL
))
2997 /* If we reach a SEQUENCE, it is too complex to try to
2998 do anything with it, so give up. We can be run during
2999 and after reorg, so SEQUENCE rtl can legitimately show
3001 if (GET_CODE (pat
) == SEQUENCE
)
3004 if (GET_CODE (pat
) == USE
3005 && NONJUMP_INSN_P (XEXP (pat
, 0)))
3006 /* reorg creates USEs that look like this. We leave them
3007 alone because reorg needs them for its own purposes. */
3010 if (reg_set_p (reg
, pat
))
3012 if (side_effects_p (pat
) && !CALL_P (our_prev
))
3015 if (GET_CODE (pat
) == PARALLEL
)
3017 /* If we find a SET of something else, we can't
3022 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
3024 rtx part
= XVECEXP (pat
, 0, i
);
3026 if (GET_CODE (part
) == SET
3027 && SET_DEST (part
) != reg
)
3031 if (i
== XVECLEN (pat
, 0))
3032 delete_computation (our_prev
);
3034 else if (GET_CODE (pat
) == SET
3035 && REG_P (SET_DEST (pat
)))
3037 int dest_regno
= REGNO (SET_DEST (pat
));
3038 int dest_endregno
= END_REGNO (SET_DEST (pat
));
3039 int regno
= REGNO (reg
);
3040 int endregno
= END_REGNO (reg
);
3042 if (dest_regno
>= regno
3043 && dest_endregno
<= endregno
)
3044 delete_computation (our_prev
);
3046 /* We may have a multi-word hard register and some, but not
3047 all, of the words of the register are needed in subsequent
3048 insns. Write REG_UNUSED notes for those parts that were not
3050 else if (dest_regno
<= regno
3051 && dest_endregno
>= endregno
)
3055 add_reg_note (our_prev
, REG_UNUSED
, reg
);
3057 for (i
= dest_regno
; i
< dest_endregno
; i
++)
3058 if (! find_regno_note (our_prev
, REG_UNUSED
, i
))
3061 if (i
== dest_endregno
)
3062 delete_computation (our_prev
);
3069 /* If PAT references the register that dies here, it is an
3070 additional use. Hence any prior SET isn't dead. However, this
3071 insn becomes the new place for the REG_DEAD note. */
3072 if (reg_overlap_mentioned_p (reg
, pat
))
3074 XEXP (note
, 1) = REG_NOTES (our_prev
);
3075 REG_NOTES (our_prev
) = note
;
3081 /* Delete INSN and recursively delete insns that compute values used only
3082 by INSN. This uses the REG_DEAD notes computed during flow analysis.
3084 Look at all our REG_DEAD notes. If a previous insn does nothing other
3085 than set a register that dies in this insn, we can delete that insn
3088 On machines with CC0, if CC0 is used in this insn, we may be able to
3089 delete the insn that set it. */
3092 delete_computation (rtx insn
)
3097 if (reg_referenced_p (cc0_rtx
, PATTERN (insn
)))
3099 rtx prev
= prev_nonnote_insn (insn
);
3100 /* We assume that at this stage
3101 CC's are always set explicitly
3102 and always immediately before the jump that
3103 will use them. So if the previous insn
3104 exists to set the CC's, delete it
3105 (unless it performs auto-increments, etc.). */
3106 if (prev
&& NONJUMP_INSN_P (prev
)
3107 && sets_cc0_p (PATTERN (prev
)))
3109 if (sets_cc0_p (PATTERN (prev
)) > 0
3110 && ! side_effects_p (PATTERN (prev
)))
3111 delete_computation (prev
);
3113 /* Otherwise, show that cc0 won't be used. */
3114 add_reg_note (prev
, REG_UNUSED
, cc0_rtx
);
3119 for (note
= REG_NOTES (insn
); note
; note
= next
)
3121 next
= XEXP (note
, 1);
3123 if (REG_NOTE_KIND (note
) != REG_DEAD
3124 /* Verify that the REG_NOTE is legitimate. */
3125 || !REG_P (XEXP (note
, 0)))
3128 delete_prior_computation (note
, insn
);
3131 delete_related_insns (insn
);
3134 /* If all INSN does is set the pc, delete it,
3135 and delete the insn that set the condition codes for it
3136 if that's what the previous thing was. */
3139 delete_jump (rtx insn
)
3141 rtx set
= single_set (insn
);
3143 if (set
&& GET_CODE (SET_DEST (set
)) == PC
)
3144 delete_computation (insn
);
3148 label_before_next_insn (rtx x
, rtx scan_limit
)
3150 rtx insn
= next_active_insn (x
);
3153 insn
= PREV_INSN (insn
);
3154 if (insn
== scan_limit
|| insn
== NULL_RTX
)
3163 /* Once we have tried two ways to fill a delay slot, make a pass over the
3164 code to try to improve the results and to do such things as more jump
3168 relax_delay_slots (rtx first
)
3170 rtx insn
, next
, pat
;
3171 rtx trial
, delay_insn
, target_label
;
3173 /* Look at every JUMP_INSN and see if we can improve it. */
3174 for (insn
= first
; insn
; insn
= next
)
3179 next
= next_active_insn (insn
);
3181 /* If this is a jump insn, see if it now jumps to a jump, jumps to
3182 the next insn, or jumps to a label that is not the last of a
3183 group of consecutive labels. */
3185 && (condjump_p (insn
) || condjump_in_parallel_p (insn
))
3186 && !ANY_RETURN_P (target_label
= JUMP_LABEL (insn
)))
3189 = skip_consecutive_labels (follow_jumps (target_label
, insn
,
3191 if (ANY_RETURN_P (target_label
))
3192 target_label
= find_end_label (target_label
);
3194 if (target_label
&& next_active_insn (target_label
) == next
3195 && ! condjump_in_parallel_p (insn
))
3201 if (target_label
&& target_label
!= JUMP_LABEL (insn
))
3203 reorg_redirect_jump (insn
, target_label
);
3205 set_unique_reg_note (insn
, REG_CROSSING_JUMP
, NULL_RTX
);
3208 /* See if this jump conditionally branches around an unconditional
3209 jump. If so, invert this jump and point it to the target of the
3211 if (next
&& simplejump_or_return_p (next
)
3212 && any_condjump_p (insn
)
3214 && next_active_insn (target_label
) == next_active_insn (next
)
3215 && no_labels_between_p (insn
, next
))
3217 rtx label
= JUMP_LABEL (next
);
3219 /* Be careful how we do this to avoid deleting code or
3220 labels that are momentarily dead. See similar optimization
3223 We also need to ensure we properly handle the case when
3224 invert_jump fails. */
3226 ++LABEL_NUSES (target_label
);
3227 if (!ANY_RETURN_P (label
))
3228 ++LABEL_NUSES (label
);
3230 if (invert_jump (insn
, label
, 1))
3232 delete_related_insns (next
);
3236 if (!ANY_RETURN_P (label
))
3237 --LABEL_NUSES (label
);
3239 if (--LABEL_NUSES (target_label
) == 0)
3240 delete_related_insns (target_label
);
3246 /* If this is an unconditional jump and the previous insn is a
3247 conditional jump, try reversing the condition of the previous
3248 insn and swapping our targets. The next pass might be able to
3251 Don't do this if we expect the conditional branch to be true, because
3252 we would then be making the more common case longer. */
3254 if (simplejump_or_return_p (insn
)
3255 && (other
= prev_active_insn (insn
)) != 0
3256 && any_condjump_p (other
)
3257 && no_labels_between_p (other
, insn
)
3258 && 0 > mostly_true_jump (other
))
3260 rtx other_target
= JUMP_LABEL (other
);
3261 target_label
= JUMP_LABEL (insn
);
3263 if (invert_jump (other
, target_label
, 0))
3264 reorg_redirect_jump (insn
, other_target
);
3267 /* Now look only at cases where we have a filled delay slot. */
3268 if (!NONJUMP_INSN_P (insn
) || GET_CODE (PATTERN (insn
)) != SEQUENCE
)
3271 pat
= PATTERN (insn
);
3272 delay_insn
= XVECEXP (pat
, 0, 0);
3274 /* See if the first insn in the delay slot is redundant with some
3275 previous insn. Remove it from the delay slot if so; then set up
3276 to reprocess this insn. */
3277 if (redundant_insn (XVECEXP (pat
, 0, 1), delay_insn
, 0))
3279 update_block (XVECEXP (pat
, 0, 1), insn
);
3280 delete_from_delay_slot (XVECEXP (pat
, 0, 1));
3281 next
= prev_active_insn (next
);
3285 /* See if we have a RETURN insn with a filled delay slot followed
3286 by a RETURN insn with an unfilled a delay slot. If so, we can delete
3287 the first RETURN (but not its delay insn). This gives the same
3288 effect in fewer instructions.
3290 Only do so if optimizing for size since this results in slower, but
3292 if (optimize_function_for_size_p (cfun
)
3293 && ANY_RETURN_P (PATTERN (delay_insn
))
3296 && PATTERN (next
) == PATTERN (delay_insn
))
3301 /* Delete the RETURN and just execute the delay list insns.
3303 We do this by deleting the INSN containing the SEQUENCE, then
3304 re-emitting the insns separately, and then deleting the RETURN.
3305 This allows the count of the jump target to be properly
3308 Note that we need to change the INSN_UID of the re-emitted insns
3309 since it is used to hash the insns for mark_target_live_regs and
3310 the re-emitted insns will no longer be wrapped up in a SEQUENCE.
3312 Clear the from target bit, since these insns are no longer
3314 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
3315 INSN_FROM_TARGET_P (XVECEXP (pat
, 0, i
)) = 0;
3317 trial
= PREV_INSN (insn
);
3318 delete_related_insns (insn
);
3319 gcc_assert (GET_CODE (pat
) == SEQUENCE
);
3320 add_insn_after (delay_insn
, trial
, NULL
);
3322 for (i
= 1; i
< XVECLEN (pat
, 0); i
++)
3323 after
= emit_copy_of_insn_after (XVECEXP (pat
, 0, i
), after
);
3324 delete_scheduled_jump (delay_insn
);
3328 /* Now look only at the cases where we have a filled JUMP_INSN. */
3329 if (!JUMP_P (delay_insn
)
3330 || !(condjump_p (delay_insn
) || condjump_in_parallel_p (delay_insn
)))
3333 target_label
= JUMP_LABEL (delay_insn
);
3334 if (target_label
&& ANY_RETURN_P (target_label
))
3337 /* If this jump goes to another unconditional jump, thread it, but
3338 don't convert a jump into a RETURN here. */
3339 trial
= skip_consecutive_labels (follow_jumps (target_label
, delay_insn
,
3341 if (ANY_RETURN_P (trial
))
3342 trial
= find_end_label (trial
);
3344 if (trial
&& trial
!= target_label
3345 && redirect_with_delay_slots_safe_p (delay_insn
, trial
, insn
))
3347 reorg_redirect_jump (delay_insn
, trial
);
3348 target_label
= trial
;
3350 set_unique_reg_note (insn
, REG_CROSSING_JUMP
, NULL_RTX
);
3353 /* If the first insn at TARGET_LABEL is redundant with a previous
3354 insn, redirect the jump to the following insn and process again.
3355 We use next_real_insn instead of next_active_insn so we
3356 don't skip USE-markers, or we'll end up with incorrect
3358 trial
= next_real_insn (target_label
);
3359 if (trial
&& GET_CODE (PATTERN (trial
)) != SEQUENCE
3360 && redundant_insn (trial
, insn
, 0)
3361 && ! can_throw_internal (trial
))
3363 /* Figure out where to emit the special USE insn so we don't
3364 later incorrectly compute register live/death info. */
3365 rtx tmp
= next_active_insn (trial
);
3367 tmp
= find_end_label (simple_return_rtx
);
3371 /* Insert the special USE insn and update dataflow info. */
3372 update_block (trial
, tmp
);
3374 /* Now emit a label before the special USE insn, and
3375 redirect our jump to the new label. */
3376 target_label
= get_label_before (PREV_INSN (tmp
));
3377 reorg_redirect_jump (delay_insn
, target_label
);
3383 /* Similarly, if it is an unconditional jump with one insn in its
3384 delay list and that insn is redundant, thread the jump. */
3385 if (trial
&& GET_CODE (PATTERN (trial
)) == SEQUENCE
3386 && XVECLEN (PATTERN (trial
), 0) == 2
3387 && JUMP_P (XVECEXP (PATTERN (trial
), 0, 0))
3388 && simplejump_or_return_p (XVECEXP (PATTERN (trial
), 0, 0))
3389 && redundant_insn (XVECEXP (PATTERN (trial
), 0, 1), insn
, 0))
3391 target_label
= JUMP_LABEL (XVECEXP (PATTERN (trial
), 0, 0));
3392 if (ANY_RETURN_P (target_label
))
3393 target_label
= find_end_label (target_label
);
3396 && redirect_with_delay_slots_safe_p (delay_insn
, target_label
,
3399 update_block (XVECEXP (PATTERN (trial
), 0, 1), insn
);
3400 reorg_redirect_jump (delay_insn
, target_label
);
3406 /* See if we have a simple (conditional) jump that is useless. */
3407 if (! INSN_ANNULLED_BRANCH_P (delay_insn
)
3408 && ! condjump_in_parallel_p (delay_insn
)
3409 && prev_active_insn (target_label
) == insn
3410 && ! BARRIER_P (prev_nonnote_insn (target_label
))
3412 /* If the last insn in the delay slot sets CC0 for some insn,
3413 various code assumes that it is in a delay slot. We could
3414 put it back where it belonged and delete the register notes,
3415 but it doesn't seem worthwhile in this uncommon case. */
3416 && ! find_reg_note (XVECEXP (pat
, 0, XVECLEN (pat
, 0) - 1),
3417 REG_CC_USER
, NULL_RTX
)
3424 /* All this insn does is execute its delay list and jump to the
3425 following insn. So delete the jump and just execute the delay
3428 We do this by deleting the INSN containing the SEQUENCE, then
3429 re-emitting the insns separately, and then deleting the jump.
3430 This allows the count of the jump target to be properly
3433 Note that we need to change the INSN_UID of the re-emitted insns
3434 since it is used to hash the insns for mark_target_live_regs and
3435 the re-emitted insns will no longer be wrapped up in a SEQUENCE.
3437 Clear the from target bit, since these insns are no longer
3439 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
3440 INSN_FROM_TARGET_P (XVECEXP (pat
, 0, i
)) = 0;
3442 trial
= PREV_INSN (insn
);
3443 delete_related_insns (insn
);
3444 gcc_assert (GET_CODE (pat
) == SEQUENCE
);
3445 add_insn_after (delay_insn
, trial
, NULL
);
3447 for (i
= 1; i
< XVECLEN (pat
, 0); i
++)
3448 after
= emit_copy_of_insn_after (XVECEXP (pat
, 0, i
), after
);
3449 delete_scheduled_jump (delay_insn
);
3453 /* See if this is an unconditional jump around a single insn which is
3454 identical to the one in its delay slot. In this case, we can just
3455 delete the branch and the insn in its delay slot. */
3456 if (next
&& NONJUMP_INSN_P (next
)
3457 && label_before_next_insn (next
, insn
) == target_label
3458 && simplejump_p (insn
)
3459 && XVECLEN (pat
, 0) == 2
3460 && rtx_equal_p (PATTERN (next
), PATTERN (XVECEXP (pat
, 0, 1))))
3462 delete_related_insns (insn
);
3466 /* See if this jump (with its delay slots) conditionally branches
3467 around an unconditional jump (without delay slots). If so, invert
3468 this jump and point it to the target of the second jump. We cannot
3469 do this for annulled jumps, though. Again, don't convert a jump to
3471 if (! INSN_ANNULLED_BRANCH_P (delay_insn
)
3472 && any_condjump_p (delay_insn
)
3473 && next
&& simplejump_or_return_p (next
)
3474 && next_active_insn (target_label
) == next_active_insn (next
)
3475 && no_labels_between_p (insn
, next
))
3477 rtx label
= JUMP_LABEL (next
);
3478 rtx old_label
= JUMP_LABEL (delay_insn
);
3480 if (ANY_RETURN_P (label
))
3481 label
= find_end_label (label
);
3483 /* find_end_label can generate a new label. Check this first. */
3485 && no_labels_between_p (insn
, next
)
3486 && redirect_with_delay_slots_safe_p (delay_insn
, label
, insn
))
3488 /* Be careful how we do this to avoid deleting code or labels
3489 that are momentarily dead. See similar optimization in
3492 ++LABEL_NUSES (old_label
);
3494 if (invert_jump (delay_insn
, label
, 1))
3498 /* Must update the INSN_FROM_TARGET_P bits now that
3499 the branch is reversed, so that mark_target_live_regs
3500 will handle the delay slot insn correctly. */
3501 for (i
= 1; i
< XVECLEN (PATTERN (insn
), 0); i
++)
3503 rtx slot
= XVECEXP (PATTERN (insn
), 0, i
);
3504 INSN_FROM_TARGET_P (slot
) = ! INSN_FROM_TARGET_P (slot
);
3507 delete_related_insns (next
);
3511 if (old_label
&& --LABEL_NUSES (old_label
) == 0)
3512 delete_related_insns (old_label
);
3517 /* If we own the thread opposite the way this insn branches, see if we
3518 can merge its delay slots with following insns. */
3519 if (INSN_FROM_TARGET_P (XVECEXP (pat
, 0, 1))
3520 && own_thread_p (NEXT_INSN (insn
), 0, 1))
3521 try_merge_delay_insns (insn
, next
);
3522 else if (! INSN_FROM_TARGET_P (XVECEXP (pat
, 0, 1))
3523 && own_thread_p (target_label
, target_label
, 0))
3524 try_merge_delay_insns (insn
, next_active_insn (target_label
));
3526 /* If we get here, we haven't deleted INSN. But we may have deleted
3527 NEXT, so recompute it. */
3528 next
= next_active_insn (insn
);
3533 /* Look for filled jumps to the end of function label. We can try to convert
3534 them into RETURN insns if the insns in the delay slot are valid for the
3538 make_return_insns (rtx first
)
3540 rtx insn
, jump_insn
, pat
;
3541 rtx real_return_label
= function_return_label
;
3542 rtx real_simple_return_label
= function_simple_return_label
;
3545 /* See if there is a RETURN insn in the function other than the one we
3546 made for END_OF_FUNCTION_LABEL. If so, set up anything we can't change
3547 into a RETURN to jump to it. */
3548 for (insn
= first
; insn
; insn
= NEXT_INSN (insn
))
3549 if (JUMP_P (insn
) && ANY_RETURN_P (PATTERN (insn
)))
3551 rtx t
= get_label_before (insn
);
3552 if (PATTERN (insn
) == ret_rtx
)
3553 real_return_label
= t
;
3555 real_simple_return_label
= t
;
3559 /* Show an extra usage of REAL_RETURN_LABEL so it won't go away if it
3560 was equal to END_OF_FUNCTION_LABEL. */
3561 if (real_return_label
)
3562 LABEL_NUSES (real_return_label
)++;
3563 if (real_simple_return_label
)
3564 LABEL_NUSES (real_simple_return_label
)++;
3566 /* Clear the list of insns to fill so we can use it. */
3567 obstack_free (&unfilled_slots_obstack
, unfilled_firstobj
);
3569 for (insn
= first
; insn
; insn
= NEXT_INSN (insn
))
3572 rtx kind
, real_label
;
3574 /* Only look at filled JUMP_INSNs that go to the end of function
3576 if (!NONJUMP_INSN_P (insn
)
3577 || GET_CODE (PATTERN (insn
)) != SEQUENCE
3578 || !jump_to_label_p (XVECEXP (PATTERN (insn
), 0, 0)))
3581 if (JUMP_LABEL (XVECEXP (PATTERN (insn
), 0, 0)) == function_return_label
)
3584 real_label
= real_return_label
;
3586 else if (JUMP_LABEL (XVECEXP (PATTERN (insn
), 0, 0))
3587 == function_simple_return_label
)
3589 kind
= simple_return_rtx
;
3590 real_label
= real_simple_return_label
;
3595 pat
= PATTERN (insn
);
3596 jump_insn
= XVECEXP (pat
, 0, 0);
3598 /* If we can't make the jump into a RETURN, try to redirect it to the best
3599 RETURN and go on to the next insn. */
3600 if (!reorg_redirect_jump (jump_insn
, kind
))
3602 /* Make sure redirecting the jump will not invalidate the delay
3604 if (redirect_with_delay_slots_safe_p (jump_insn
, real_label
, insn
))
3605 reorg_redirect_jump (jump_insn
, real_label
);
3609 /* See if this RETURN can accept the insns current in its delay slot.
3610 It can if it has more or an equal number of slots and the contents
3611 of each is valid. */
3613 flags
= get_jump_flags (jump_insn
, JUMP_LABEL (jump_insn
));
3614 slots
= num_delay_slots (jump_insn
);
3615 if (slots
>= XVECLEN (pat
, 0) - 1)
3617 for (i
= 1; i
< XVECLEN (pat
, 0); i
++)
3619 #ifdef ANNUL_IFFALSE_SLOTS
3620 (INSN_ANNULLED_BRANCH_P (jump_insn
)
3621 && INSN_FROM_TARGET_P (XVECEXP (pat
, 0, i
)))
3622 ? eligible_for_annul_false (jump_insn
, i
- 1,
3623 XVECEXP (pat
, 0, i
), flags
) :
3625 #ifdef ANNUL_IFTRUE_SLOTS
3626 (INSN_ANNULLED_BRANCH_P (jump_insn
)
3627 && ! INSN_FROM_TARGET_P (XVECEXP (pat
, 0, i
)))
3628 ? eligible_for_annul_true (jump_insn
, i
- 1,
3629 XVECEXP (pat
, 0, i
), flags
) :
3631 eligible_for_delay (jump_insn
, i
- 1,
3632 XVECEXP (pat
, 0, i
), flags
)))
3638 if (i
== XVECLEN (pat
, 0))
3641 /* We have to do something with this insn. If it is an unconditional
3642 RETURN, delete the SEQUENCE and output the individual insns,
3643 followed by the RETURN. Then set things up so we try to find
3644 insns for its delay slots, if it needs some. */
3645 if (ANY_RETURN_P (PATTERN (jump_insn
)))
3647 rtx prev
= PREV_INSN (insn
);
3649 delete_related_insns (insn
);
3650 for (i
= 1; i
< XVECLEN (pat
, 0); i
++)
3651 prev
= emit_insn_after (PATTERN (XVECEXP (pat
, 0, i
)), prev
);
3653 insn
= emit_jump_insn_after (PATTERN (jump_insn
), prev
);
3654 emit_barrier_after (insn
);
3657 obstack_ptr_grow (&unfilled_slots_obstack
, insn
);
3660 /* It is probably more efficient to keep this with its current
3661 delay slot as a branch to a RETURN. */
3662 reorg_redirect_jump (jump_insn
, real_label
);
3665 /* Now delete REAL_RETURN_LABEL if we never used it. Then try to fill any
3666 new delay slots we have created. */
3667 if (real_return_label
!= NULL_RTX
&& --LABEL_NUSES (real_return_label
) == 0)
3668 delete_related_insns (real_return_label
);
3669 if (real_simple_return_label
!= NULL_RTX
3670 && --LABEL_NUSES (real_simple_return_label
) == 0)
3671 delete_related_insns (real_simple_return_label
);
3673 fill_simple_delay_slots (1);
3674 fill_simple_delay_slots (0);
3677 /* Try to find insns to place in delay slots. */
3680 dbr_schedule (rtx first
)
3682 rtx insn
, next
, epilogue_insn
= 0;
3684 bool need_return_insns
;
3686 /* If the current function has no insns other than the prologue and
3687 epilogue, then do not try to fill any delay slots. */
3688 if (n_basic_blocks
== NUM_FIXED_BLOCKS
)
3691 /* Find the highest INSN_UID and allocate and initialize our map from
3692 INSN_UID's to position in code. */
3693 for (max_uid
= 0, insn
= first
; insn
; insn
= NEXT_INSN (insn
))
3695 if (INSN_UID (insn
) > max_uid
)
3696 max_uid
= INSN_UID (insn
);
3698 && NOTE_KIND (insn
) == NOTE_INSN_EPILOGUE_BEG
)
3699 epilogue_insn
= insn
;
3702 uid_to_ruid
= XNEWVEC (int, max_uid
+ 1);
3703 for (i
= 0, insn
= first
; insn
; i
++, insn
= NEXT_INSN (insn
))
3704 uid_to_ruid
[INSN_UID (insn
)] = i
;
3706 /* Initialize the list of insns that need filling. */
3707 if (unfilled_firstobj
== 0)
3709 gcc_obstack_init (&unfilled_slots_obstack
);
3710 unfilled_firstobj
= XOBNEWVAR (&unfilled_slots_obstack
, rtx
, 0);
3713 for (insn
= next_active_insn (first
); insn
; insn
= next_active_insn (insn
))
3718 INSN_ANNULLED_BRANCH_P (insn
) = 0;
3719 INSN_FROM_TARGET_P (insn
) = 0;
3721 /* Skip vector tables. We can't get attributes for them. */
3722 if (JUMP_TABLE_DATA_P (insn
))
3725 if (num_delay_slots (insn
) > 0)
3726 obstack_ptr_grow (&unfilled_slots_obstack
, insn
);
3728 /* Ensure all jumps go to the last of a set of consecutive labels. */
3730 && (condjump_p (insn
) || condjump_in_parallel_p (insn
))
3731 && !ANY_RETURN_P (JUMP_LABEL (insn
))
3732 && ((target
= skip_consecutive_labels (JUMP_LABEL (insn
)))
3733 != JUMP_LABEL (insn
)))
3734 redirect_jump (insn
, target
, 1);
3737 init_resource_info (epilogue_insn
);
3739 /* Show we haven't computed an end-of-function label yet. */
3740 function_return_label
= function_simple_return_label
= NULL_RTX
;
3742 /* Initialize the statistics for this function. */
3743 memset (num_insns_needing_delays
, 0, sizeof num_insns_needing_delays
);
3744 memset (num_filled_delays
, 0, sizeof num_filled_delays
);
3746 /* Now do the delay slot filling. Try everything twice in case earlier
3747 changes make more slots fillable. */
3749 for (reorg_pass_number
= 0;
3750 reorg_pass_number
< MAX_REORG_PASSES
;
3751 reorg_pass_number
++)
3753 fill_simple_delay_slots (1);
3754 fill_simple_delay_slots (0);
3755 fill_eager_delay_slots ();
3756 relax_delay_slots (first
);
3759 /* If we made an end of function label, indicate that it is now
3760 safe to delete it by undoing our prior adjustment to LABEL_NUSES.
3761 If it is now unused, delete it. */
3762 if (function_return_label
&& --LABEL_NUSES (function_return_label
) == 0)
3763 delete_related_insns (function_return_label
);
3764 if (function_simple_return_label
3765 && --LABEL_NUSES (function_simple_return_label
) == 0)
3766 delete_related_insns (function_simple_return_label
);
3768 need_return_insns
= false;
3770 need_return_insns
|= HAVE_return
&& function_return_label
!= 0;
3772 #ifdef HAVE_simple_return
3773 need_return_insns
|= HAVE_simple_return
&& function_simple_return_label
!= 0;
3775 if (need_return_insns
)
3776 make_return_insns (first
);
3778 /* Delete any USE insns made by update_block; subsequent passes don't need
3779 them or know how to deal with them. */
3780 for (insn
= first
; insn
; insn
= next
)
3782 next
= NEXT_INSN (insn
);
3784 if (NONJUMP_INSN_P (insn
) && GET_CODE (PATTERN (insn
)) == USE
3785 && INSN_P (XEXP (PATTERN (insn
), 0)))
3786 next
= delete_related_insns (insn
);
3789 obstack_free (&unfilled_slots_obstack
, unfilled_firstobj
);
3791 /* It is not clear why the line below is needed, but it does seem to be. */
3792 unfilled_firstobj
= XOBNEWVAR (&unfilled_slots_obstack
, rtx
, 0);
3796 int i
, j
, need_comma
;
3797 int total_delay_slots
[MAX_DELAY_HISTOGRAM
+ 1];
3798 int total_annul_slots
[MAX_DELAY_HISTOGRAM
+ 1];
3800 for (reorg_pass_number
= 0;
3801 reorg_pass_number
< MAX_REORG_PASSES
;
3802 reorg_pass_number
++)
3804 fprintf (dump_file
, ";; Reorg pass #%d:\n", reorg_pass_number
+ 1);
3805 for (i
= 0; i
< NUM_REORG_FUNCTIONS
; i
++)
3808 fprintf (dump_file
, ";; Reorg function #%d\n", i
);
3810 fprintf (dump_file
, ";; %d insns needing delay slots\n;; ",
3811 num_insns_needing_delays
[i
][reorg_pass_number
]);
3813 for (j
= 0; j
< MAX_DELAY_HISTOGRAM
+ 1; j
++)
3814 if (num_filled_delays
[i
][j
][reorg_pass_number
])
3817 fprintf (dump_file
, ", ");
3819 fprintf (dump_file
, "%d got %d delays",
3820 num_filled_delays
[i
][j
][reorg_pass_number
], j
);
3822 fprintf (dump_file
, "\n");
3825 memset (total_delay_slots
, 0, sizeof total_delay_slots
);
3826 memset (total_annul_slots
, 0, sizeof total_annul_slots
);
3827 for (insn
= first
; insn
; insn
= NEXT_INSN (insn
))
3829 if (! INSN_DELETED_P (insn
)
3830 && NONJUMP_INSN_P (insn
)
3831 && GET_CODE (PATTERN (insn
)) != USE
3832 && GET_CODE (PATTERN (insn
)) != CLOBBER
)
3834 if (GET_CODE (PATTERN (insn
)) == SEQUENCE
)
3837 j
= XVECLEN (PATTERN (insn
), 0) - 1;
3838 if (j
> MAX_DELAY_HISTOGRAM
)
3839 j
= MAX_DELAY_HISTOGRAM
;
3840 control
= XVECEXP (PATTERN (insn
), 0, 0);
3841 if (JUMP_P (control
) && INSN_ANNULLED_BRANCH_P (control
))
3842 total_annul_slots
[j
]++;
3844 total_delay_slots
[j
]++;
3846 else if (num_delay_slots (insn
) > 0)
3847 total_delay_slots
[0]++;
3850 fprintf (dump_file
, ";; Reorg totals: ");
3852 for (j
= 0; j
< MAX_DELAY_HISTOGRAM
+ 1; j
++)
3854 if (total_delay_slots
[j
])
3857 fprintf (dump_file
, ", ");
3859 fprintf (dump_file
, "%d got %d delays", total_delay_slots
[j
], j
);
3862 fprintf (dump_file
, "\n");
3863 #if defined (ANNUL_IFTRUE_SLOTS) || defined (ANNUL_IFFALSE_SLOTS)
3864 fprintf (dump_file
, ";; Reorg annuls: ");
3866 for (j
= 0; j
< MAX_DELAY_HISTOGRAM
+ 1; j
++)
3868 if (total_annul_slots
[j
])
3871 fprintf (dump_file
, ", ");
3873 fprintf (dump_file
, "%d got %d delays", total_annul_slots
[j
], j
);
3876 fprintf (dump_file
, "\n");
3878 fprintf (dump_file
, "\n");
3881 free_resource_info ();
3883 crtl
->dbr_scheduled_p
= true;
3885 #endif /* DELAY_SLOTS */
3888 gate_handle_delay_slots (void)
3891 /* At -O0 dataflow info isn't updated after RA. */
3892 return optimize
> 0 && flag_delayed_branch
&& !crtl
->dbr_scheduled_p
;
3898 /* Run delay slot optimization. */
3900 rest_of_handle_delay_slots (void)
3903 dbr_schedule (get_insns ());
3908 struct rtl_opt_pass pass_delay_slots
=
3913 OPTGROUP_NONE
, /* optinfo_flags */
3914 gate_handle_delay_slots
, /* gate */
3915 rest_of_handle_delay_slots
, /* execute */
3918 0, /* static_pass_number */
3919 TV_DBR_SCHED
, /* tv_id */
3920 0, /* properties_required */
3921 0, /* properties_provided */
3922 0, /* properties_destroyed */
3923 0, /* todo_flags_start */
3924 TODO_ggc_collect
/* todo_flags_finish */
3928 /* Machine dependent reorg pass. */
3930 gate_handle_machine_reorg (void)
3932 return targetm
.machine_dependent_reorg
!= 0;
3937 rest_of_handle_machine_reorg (void)
3939 targetm
.machine_dependent_reorg ();
3943 struct rtl_opt_pass pass_machine_reorg
=
3948 OPTGROUP_NONE
, /* optinfo_flags */
3949 gate_handle_machine_reorg
, /* gate */
3950 rest_of_handle_machine_reorg
, /* execute */
3953 0, /* static_pass_number */
3954 TV_MACH_DEP
, /* tv_id */
3955 0, /* properties_required */
3956 0, /* properties_provided */
3957 0, /* properties_destroyed */
3958 0, /* todo_flags_start */
3959 TODO_ggc_collect
/* todo_flags_finish */