1 /* Perform instruction reorganizations for delay slot filling.
2 Copyright (C) 1992-2015 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"
112 #include "hash-set.h"
114 #include "machmode.h"
115 #include "hard-reg-set.h"
117 #include "function.h"
119 #include "statistics.h"
120 #include "double-int.h"
122 #include "fixed-value.h"
124 #include "wide-int.h"
127 #include "insn-config.h"
132 #include "emit-rtl.h"
136 #include "conditions.h"
138 #include "dominance.h"
140 #include "basic-block.h"
144 #include "insn-attr.h"
145 #include "resource.h"
149 #include "tree-pass.h"
153 #ifndef ANNUL_IFTRUE_SLOTS
154 #define eligible_for_annul_true(INSN, SLOTS, TRIAL, FLAGS) 0
156 #ifndef ANNUL_IFFALSE_SLOTS
157 #define eligible_for_annul_false(INSN, SLOTS, TRIAL, FLAGS) 0
161 /* First, some functions that were used before GCC got a control flow graph.
162 These functions are now only used here in reorg.c, and have therefore
163 been moved here to avoid inadvertent misuse elsewhere in the compiler. */
165 /* Return the last label to mark the same position as LABEL. Return LABEL
166 itself if it is null or any return rtx. */
169 skip_consecutive_labels (rtx label_or_return
)
173 if (label_or_return
&& ANY_RETURN_P (label_or_return
))
174 return label_or_return
;
176 rtx_insn
*label
= as_a
<rtx_insn
*> (label_or_return
);
178 for (insn
= label
; insn
!= 0 && !INSN_P (insn
); insn
= NEXT_INSN (insn
))
186 /* INSN uses CC0 and is being moved into a delay slot. Set up REG_CC_SETTER
187 and REG_CC_USER notes so we can find it. */
190 link_cc0_insns (rtx insn
)
192 rtx user
= next_nonnote_insn (insn
);
194 if (NONJUMP_INSN_P (user
) && GET_CODE (PATTERN (user
)) == SEQUENCE
)
195 user
= XVECEXP (PATTERN (user
), 0, 0);
197 add_reg_note (user
, REG_CC_SETTER
, insn
);
198 add_reg_note (insn
, REG_CC_USER
, user
);
202 /* Insns which have delay slots that have not yet been filled. */
204 static struct obstack unfilled_slots_obstack
;
205 static rtx
*unfilled_firstobj
;
207 /* Define macros to refer to the first and last slot containing unfilled
208 insns. These are used because the list may move and its address
209 should be recomputed at each use. */
211 #define unfilled_slots_base \
212 ((rtx_insn **) obstack_base (&unfilled_slots_obstack))
214 #define unfilled_slots_next \
215 ((rtx_insn **) obstack_next_free (&unfilled_slots_obstack))
217 /* Points to the label before the end of the function, or before a
219 static rtx_code_label
*function_return_label
;
220 /* Likewise for a simple_return. */
221 static rtx_code_label
*function_simple_return_label
;
223 /* Mapping between INSN_UID's and position in the code since INSN_UID's do
224 not always monotonically increase. */
225 static int *uid_to_ruid
;
227 /* Highest valid index in `uid_to_ruid'. */
230 static int stop_search_p (rtx
, int);
231 static int resource_conflicts_p (struct resources
*, struct resources
*);
232 static int insn_references_resource_p (rtx
, struct resources
*, bool);
233 static int insn_sets_resource_p (rtx
, struct resources
*, bool);
234 static rtx_code_label
*find_end_label (rtx
);
235 static rtx_insn
*emit_delay_sequence (rtx_insn
*, rtx_insn_list
*, int);
236 static rtx_insn_list
*add_to_delay_list (rtx_insn
*, rtx_insn_list
*);
237 static rtx_insn
*delete_from_delay_slot (rtx_insn
*);
238 static void delete_scheduled_jump (rtx_insn
*);
239 static void note_delay_statistics (int, int);
240 #if defined(ANNUL_IFFALSE_SLOTS) || defined(ANNUL_IFTRUE_SLOTS)
241 static rtx_insn_list
*optimize_skip (rtx_insn
*);
243 static int get_jump_flags (const rtx_insn
*, rtx
);
244 static int mostly_true_jump (rtx
);
245 static rtx
get_branch_condition (const rtx_insn
*, rtx
);
246 static int condition_dominates_p (rtx
, const rtx_insn
*);
247 static int redirect_with_delay_slots_safe_p (rtx_insn
*, rtx
, rtx
);
248 static int redirect_with_delay_list_safe_p (rtx_insn
*, rtx
, rtx_insn_list
*);
249 static int check_annul_list_true_false (int, rtx
);
250 static rtx_insn_list
*steal_delay_list_from_target (rtx_insn
*, rtx
,
258 static rtx_insn_list
*steal_delay_list_from_fallthrough (rtx_insn
*, rtx
,
265 static void try_merge_delay_insns (rtx
, rtx_insn
*);
266 static rtx
redundant_insn (rtx
, rtx_insn
*, rtx
);
267 static int own_thread_p (rtx
, rtx
, int);
268 static void update_block (rtx_insn
*, rtx
);
269 static int reorg_redirect_jump (rtx_insn
*, rtx
);
270 static void update_reg_dead_notes (rtx
, rtx
);
271 static void fix_reg_dead_note (rtx
, rtx
);
272 static void update_reg_unused_notes (rtx
, rtx
);
273 static void fill_simple_delay_slots (int);
274 static rtx_insn_list
*fill_slots_from_thread (rtx_insn
*, rtx
, rtx
, rtx
,
276 int *, rtx_insn_list
*);
277 static void fill_eager_delay_slots (void);
278 static void relax_delay_slots (rtx_insn
*);
279 static void make_return_insns (rtx_insn
*);
281 /* A wrapper around next_active_insn which takes care to return ret_rtx
285 first_active_target_insn (rtx insn
)
287 if (ANY_RETURN_P (insn
))
289 return next_active_insn (as_a
<rtx_insn
*> (insn
));
292 /* Return true iff INSN is a simplejump, or any kind of return insn. */
295 simplejump_or_return_p (rtx insn
)
297 return (JUMP_P (insn
)
298 && (simplejump_p (as_a
<rtx_insn
*> (insn
))
299 || ANY_RETURN_P (PATTERN (insn
))));
302 /* Return TRUE if this insn should stop the search for insn to fill delay
303 slots. LABELS_P indicates that labels should terminate the search.
304 In all cases, jumps terminate the search. */
307 stop_search_p (rtx insn
, int labels_p
)
312 /* If the insn can throw an exception that is caught within the function,
313 it may effectively perform a jump from the viewpoint of the function.
314 Therefore act like for a jump. */
315 if (can_throw_internal (insn
))
318 switch (GET_CODE (insn
))
332 /* OK unless it contains a delay slot or is an `asm' insn of some type.
333 We don't know anything about these. */
334 return (GET_CODE (PATTERN (insn
)) == SEQUENCE
335 || GET_CODE (PATTERN (insn
)) == ASM_INPUT
336 || asm_noperands (PATTERN (insn
)) >= 0);
343 /* Return TRUE if any resources are marked in both RES1 and RES2 or if either
344 resource set contains a volatile memory reference. Otherwise, return FALSE. */
347 resource_conflicts_p (struct resources
*res1
, struct resources
*res2
)
349 if ((res1
->cc
&& res2
->cc
) || (res1
->memory
&& res2
->memory
)
350 || res1
->volatil
|| res2
->volatil
)
353 return hard_reg_set_intersect_p (res1
->regs
, res2
->regs
);
356 /* Return TRUE if any resource marked in RES, a `struct resources', is
357 referenced by INSN. If INCLUDE_DELAYED_EFFECTS is set, return if the called
358 routine is using those resources.
360 We compute this by computing all the resources referenced by INSN and
361 seeing if this conflicts with RES. It might be faster to directly check
362 ourselves, and this is the way it used to work, but it means duplicating
363 a large block of complex code. */
366 insn_references_resource_p (rtx insn
, struct resources
*res
,
367 bool include_delayed_effects
)
369 struct resources insn_res
;
371 CLEAR_RESOURCE (&insn_res
);
372 mark_referenced_resources (insn
, &insn_res
, include_delayed_effects
);
373 return resource_conflicts_p (&insn_res
, res
);
376 /* Return TRUE if INSN modifies resources that are marked in RES.
377 INCLUDE_DELAYED_EFFECTS is set if the actions of that routine should be
378 included. CC0 is only modified if it is explicitly set; see comments
379 in front of mark_set_resources for details. */
382 insn_sets_resource_p (rtx insn
, struct resources
*res
,
383 bool include_delayed_effects
)
385 struct resources insn_sets
;
387 CLEAR_RESOURCE (&insn_sets
);
388 mark_set_resources (insn
, &insn_sets
, 0,
389 (include_delayed_effects
392 return resource_conflicts_p (&insn_sets
, res
);
395 /* Find a label at the end of the function or before a RETURN. If there
396 is none, try to make one. If that fails, returns 0.
398 The property of such a label is that it is placed just before the
399 epilogue or a bare RETURN insn, so that another bare RETURN can be
400 turned into a jump to the label unconditionally. In particular, the
401 label cannot be placed before a RETURN insn with a filled delay slot.
403 ??? There may be a problem with the current implementation. Suppose
404 we start with a bare RETURN insn and call find_end_label. It may set
405 function_return_label just before the RETURN. Suppose the machinery
406 is able to fill the delay slot of the RETURN insn afterwards. Then
407 function_return_label is no longer valid according to the property
408 described above and find_end_label will still return it unmodified.
409 Note that this is probably mitigated by the following observation:
410 once function_return_label is made, it is very likely the target of
411 a jump, so filling the delay slot of the RETURN will be much more
413 KIND is either simple_return_rtx or ret_rtx, indicating which type of
414 return we're looking for. */
416 static rtx_code_label
*
417 find_end_label (rtx kind
)
420 rtx_code_label
**plabel
;
423 plabel
= &function_return_label
;
426 gcc_assert (kind
== simple_return_rtx
);
427 plabel
= &function_simple_return_label
;
430 /* If we found one previously, return it. */
434 /* Otherwise, see if there is a label at the end of the function. If there
435 is, it must be that RETURN insns aren't needed, so that is our return
436 label and we don't have to do anything else. */
438 insn
= get_last_insn ();
440 || (NONJUMP_INSN_P (insn
)
441 && (GET_CODE (PATTERN (insn
)) == USE
442 || GET_CODE (PATTERN (insn
)) == CLOBBER
)))
443 insn
= PREV_INSN (insn
);
445 /* When a target threads its epilogue we might already have a
446 suitable return insn. If so put a label before it for the
447 function_return_label. */
449 && JUMP_P (PREV_INSN (insn
))
450 && PATTERN (PREV_INSN (insn
)) == kind
)
452 rtx_insn
*temp
= PREV_INSN (PREV_INSN (insn
));
453 rtx_code_label
*label
= gen_label_rtx ();
454 LABEL_NUSES (label
) = 0;
456 /* Put the label before any USE insns that may precede the RETURN
458 while (GET_CODE (temp
) == USE
)
459 temp
= PREV_INSN (temp
);
461 emit_label_after (label
, temp
);
465 else if (LABEL_P (insn
))
466 *plabel
= as_a
<rtx_code_label
*> (insn
);
469 rtx_code_label
*label
= gen_label_rtx ();
470 LABEL_NUSES (label
) = 0;
471 /* If the basic block reorder pass moves the return insn to
472 some other place try to locate it again and put our
473 function_return_label there. */
474 while (insn
&& ! (JUMP_P (insn
) && (PATTERN (insn
) == kind
)))
475 insn
= PREV_INSN (insn
);
478 insn
= PREV_INSN (insn
);
480 /* Put the label before any USE insns that may precede the
482 while (GET_CODE (insn
) == USE
)
483 insn
= PREV_INSN (insn
);
485 emit_label_after (label
, insn
);
495 /* The RETURN insn has its delay slot filled so we cannot
496 emit the label just before it. Since we already have
497 an epilogue and cannot emit a new RETURN, we cannot
498 emit the label at all. */
500 #endif /* HAVE_epilogue */
502 /* Otherwise, make a new label and emit a RETURN and BARRIER,
508 /* The return we make may have delay slots too. */
509 rtx pat
= gen_return ();
510 rtx_insn
*insn
= emit_jump_insn (pat
);
511 set_return_jump_label (insn
);
513 if (num_delay_slots (insn
) > 0)
514 obstack_ptr_grow (&unfilled_slots_obstack
, insn
);
521 /* Show one additional use for this label so it won't go away until
523 ++LABEL_NUSES (*plabel
);
528 /* Put INSN and LIST together in a SEQUENCE rtx of LENGTH, and replace
529 the pattern of INSN with the SEQUENCE.
531 Returns the insn containing the SEQUENCE that replaces INSN. */
534 emit_delay_sequence (rtx_insn
*insn
, rtx_insn_list
*list
, int length
)
536 /* Allocate the rtvec to hold the insns and the SEQUENCE. */
537 rtvec seqv
= rtvec_alloc (length
+ 1);
538 rtx seq
= gen_rtx_SEQUENCE (VOIDmode
, seqv
);
539 rtx_insn
*seq_insn
= make_insn_raw (seq
);
541 /* If DELAY_INSN has a location, use it for SEQ_INSN. If DELAY_INSN does
542 not have a location, but one of the delayed insns does, we pick up a
543 location from there later. */
544 INSN_LOCATION (seq_insn
) = INSN_LOCATION (insn
);
546 /* Unlink INSN from the insn chain, so that we can put it into
547 the SEQUENCE. Remember where we want to emit SEQUENCE in AFTER. */
548 rtx after
= PREV_INSN (insn
);
550 SET_NEXT_INSN (insn
) = SET_PREV_INSN (insn
) = NULL
;
552 /* Build our SEQUENCE and rebuild the insn chain. */
555 XVECEXP (seq
, 0, 0) = emit_insn (insn
);
556 for (rtx_insn_list
*li
= list
; li
; li
= li
->next (), i
++)
558 rtx_insn
*tem
= li
->insn ();
561 /* Show that this copy of the insn isn't deleted. */
562 tem
->set_undeleted ();
564 /* Unlink insn from its original place, and re-emit it into
566 SET_NEXT_INSN (tem
) = SET_PREV_INSN (tem
) = NULL
;
567 XVECEXP (seq
, 0, i
) = emit_insn (tem
);
569 /* SPARC assembler, for instance, emit warning when debug info is output
570 into the delay slot. */
571 if (INSN_LOCATION (tem
) && !INSN_LOCATION (seq_insn
))
572 INSN_LOCATION (seq_insn
) = INSN_LOCATION (tem
);
573 INSN_LOCATION (tem
) = 0;
575 for (note
= REG_NOTES (tem
); note
; note
= next
)
577 next
= XEXP (note
, 1);
578 switch (REG_NOTE_KIND (note
))
581 /* Remove any REG_DEAD notes because we can't rely on them now
582 that the insn has been moved. */
583 remove_note (tem
, note
);
586 case REG_LABEL_OPERAND
:
587 case REG_LABEL_TARGET
:
588 /* Keep the label reference count up to date. */
589 if (LABEL_P (XEXP (note
, 0)))
590 LABEL_NUSES (XEXP (note
, 0)) ++;
599 gcc_assert (i
== length
+ 1);
601 /* Splice our SEQUENCE into the insn stream where INSN used to be. */
602 add_insn_after (seq_insn
, after
, NULL
);
607 /* Add INSN to DELAY_LIST and return the head of the new list. The list must
608 be in the order in which the insns are to be executed. */
610 static rtx_insn_list
*
611 add_to_delay_list (rtx_insn
*insn
, rtx_insn_list
*delay_list
)
613 /* If we have an empty list, just make a new list element. If
614 INSN has its block number recorded, clear it since we may
615 be moving the insn to a new block. */
619 clear_hashed_info_for_insn (insn
);
620 return gen_rtx_INSN_LIST (VOIDmode
, insn
, NULL_RTX
);
623 /* Otherwise this must be an INSN_LIST. Add INSN to the end of the
625 XEXP (delay_list
, 1) = add_to_delay_list (insn
, delay_list
->next ());
630 /* Delete INSN from the delay slot of the insn that it is in, which may
631 produce an insn with no delay slots. Return the new insn. */
634 delete_from_delay_slot (rtx_insn
*insn
)
636 rtx_insn
*trial
, *seq_insn
, *prev
;
638 rtx_insn_list
*delay_list
= 0;
642 /* We first must find the insn containing the SEQUENCE with INSN in its
643 delay slot. Do this by finding an insn, TRIAL, where
644 PREV_INSN (NEXT_INSN (TRIAL)) != TRIAL. */
647 PREV_INSN (NEXT_INSN (trial
)) == trial
;
648 trial
= NEXT_INSN (trial
))
651 seq_insn
= PREV_INSN (NEXT_INSN (trial
));
652 seq
= as_a
<rtx_sequence
*> (PATTERN (seq_insn
));
654 if (NEXT_INSN (seq_insn
) && BARRIER_P (NEXT_INSN (seq_insn
)))
657 /* Create a delay list consisting of all the insns other than the one
658 we are deleting (unless we were the only one). */
660 for (i
= 1; i
< seq
->len (); i
++)
661 if (seq
->insn (i
) != insn
)
662 delay_list
= add_to_delay_list (seq
->insn (i
), delay_list
);
664 /* Delete the old SEQUENCE, re-emit the insn that used to have the delay
665 list, and rebuild the delay list if non-empty. */
666 prev
= PREV_INSN (seq_insn
);
667 trial
= seq
->insn (0);
668 delete_related_insns (seq_insn
);
669 add_insn_after (trial
, prev
, NULL
);
671 /* If there was a barrier after the old SEQUENCE, remit it. */
673 emit_barrier_after (trial
);
675 /* If there are any delay insns, remit them. Otherwise clear the
678 trial
= emit_delay_sequence (trial
, delay_list
, XVECLEN (seq
, 0) - 2);
679 else if (JUMP_P (trial
))
680 INSN_ANNULLED_BRANCH_P (trial
) = 0;
682 INSN_FROM_TARGET_P (insn
) = 0;
684 /* Show we need to fill this insn again. */
685 obstack_ptr_grow (&unfilled_slots_obstack
, trial
);
690 /* Delete INSN, a JUMP_INSN. If it is a conditional jump, we must track down
691 the insn that sets CC0 for it and delete it too. */
694 delete_scheduled_jump (rtx_insn
*insn
)
696 /* Delete the insn that sets cc0 for us. On machines without cc0, we could
697 delete the insn that sets the condition code, but it is hard to find it.
698 Since this case is rare anyway, don't bother trying; there would likely
699 be other insns that became dead anyway, which we wouldn't know to
703 if (reg_mentioned_p (cc0_rtx
, insn
))
705 rtx note
= find_reg_note (insn
, REG_CC_SETTER
, NULL_RTX
);
707 /* If a reg-note was found, it points to an insn to set CC0. This
708 insn is in the delay list of some other insn. So delete it from
709 the delay list it was in. */
712 if (! FIND_REG_INC_NOTE (XEXP (note
, 0), NULL_RTX
)
713 && sets_cc0_p (PATTERN (XEXP (note
, 0))) == 1)
714 delete_from_delay_slot (as_a
<rtx_insn
*> (XEXP (note
, 0)));
718 /* The insn setting CC0 is our previous insn, but it may be in
719 a delay slot. It will be the last insn in the delay slot, if
721 rtx_insn
*trial
= previous_insn (insn
);
723 trial
= prev_nonnote_insn (trial
);
724 if (sets_cc0_p (PATTERN (trial
)) != 1
725 || FIND_REG_INC_NOTE (trial
, NULL_RTX
))
727 if (PREV_INSN (NEXT_INSN (trial
)) == trial
)
728 delete_related_insns (trial
);
730 delete_from_delay_slot (trial
);
735 delete_related_insns (insn
);
738 /* Counters for delay-slot filling. */
740 #define NUM_REORG_FUNCTIONS 2
741 #define MAX_DELAY_HISTOGRAM 3
742 #define MAX_REORG_PASSES 2
744 static int num_insns_needing_delays
[NUM_REORG_FUNCTIONS
][MAX_REORG_PASSES
];
746 static int num_filled_delays
[NUM_REORG_FUNCTIONS
][MAX_DELAY_HISTOGRAM
+1][MAX_REORG_PASSES
];
748 static int reorg_pass_number
;
751 note_delay_statistics (int slots_filled
, int index
)
753 num_insns_needing_delays
[index
][reorg_pass_number
]++;
754 if (slots_filled
> MAX_DELAY_HISTOGRAM
)
755 slots_filled
= MAX_DELAY_HISTOGRAM
;
756 num_filled_delays
[index
][slots_filled
][reorg_pass_number
]++;
759 #if defined(ANNUL_IFFALSE_SLOTS) || defined(ANNUL_IFTRUE_SLOTS)
761 /* Optimize the following cases:
763 1. When a conditional branch skips over only one instruction,
764 use an annulling branch and put that insn in the delay slot.
765 Use either a branch that annuls when the condition if true or
766 invert the test with a branch that annuls when the condition is
767 false. This saves insns, since otherwise we must copy an insn
770 (orig) (skip) (otherwise)
771 Bcc.n L1 Bcc',a L1 Bcc,a L1'
778 2. When a conditional branch skips over only one instruction,
779 and after that, it unconditionally branches somewhere else,
780 perform the similar optimization. This saves executing the
781 second branch in the case where the inverted condition is true.
790 This should be expanded to skip over N insns, where N is the number
791 of delay slots required. */
793 static rtx_insn_list
*
794 optimize_skip (rtx_insn
*insn
)
796 rtx_insn
*trial
= next_nonnote_insn (insn
);
797 rtx_insn
*next_trial
= next_active_insn (trial
);
798 rtx_insn_list
*delay_list
= 0;
801 flags
= get_jump_flags (insn
, JUMP_LABEL (insn
));
804 || !NONJUMP_INSN_P (trial
)
805 || GET_CODE (PATTERN (trial
)) == SEQUENCE
806 || recog_memoized (trial
) < 0
807 || (! eligible_for_annul_false (insn
, 0, trial
, flags
)
808 && ! eligible_for_annul_true (insn
, 0, trial
, flags
))
809 || can_throw_internal (trial
))
812 /* There are two cases where we are just executing one insn (we assume
813 here that a branch requires only one insn; this should be generalized
814 at some point): Where the branch goes around a single insn or where
815 we have one insn followed by a branch to the same label we branch to.
816 In both of these cases, inverting the jump and annulling the delay
817 slot give the same effect in fewer insns. */
818 if (next_trial
== next_active_insn (JUMP_LABEL (insn
))
820 && simplejump_or_return_p (next_trial
)
821 && JUMP_LABEL (insn
) == JUMP_LABEL (next_trial
)))
823 if (eligible_for_annul_false (insn
, 0, trial
, flags
))
825 if (invert_jump (insn
, JUMP_LABEL (insn
), 1))
826 INSN_FROM_TARGET_P (trial
) = 1;
827 else if (! eligible_for_annul_true (insn
, 0, trial
, flags
))
831 delay_list
= add_to_delay_list (trial
, NULL
);
832 next_trial
= next_active_insn (trial
);
833 update_block (trial
, trial
);
834 delete_related_insns (trial
);
836 /* Also, if we are targeting an unconditional
837 branch, thread our jump to the target of that branch. Don't
838 change this into a RETURN here, because it may not accept what
839 we have in the delay slot. We'll fix this up later. */
840 if (next_trial
&& simplejump_or_return_p (next_trial
))
842 rtx target_label
= JUMP_LABEL (next_trial
);
843 if (ANY_RETURN_P (target_label
))
844 target_label
= find_end_label (target_label
);
848 /* Recompute the flags based on TARGET_LABEL since threading
849 the jump to TARGET_LABEL may change the direction of the
850 jump (which may change the circumstances in which the
851 delay slot is nullified). */
852 flags
= get_jump_flags (insn
, target_label
);
853 if (eligible_for_annul_true (insn
, 0, trial
, flags
))
854 reorg_redirect_jump (insn
, target_label
);
858 INSN_ANNULLED_BRANCH_P (insn
) = 1;
865 /* Encode and return branch direction and prediction information for
866 INSN assuming it will jump to LABEL.
868 Non conditional branches return no direction information and
869 are predicted as very likely taken. */
872 get_jump_flags (const rtx_insn
*insn
, rtx label
)
876 /* get_jump_flags can be passed any insn with delay slots, these may
877 be INSNs, CALL_INSNs, or JUMP_INSNs. Only JUMP_INSNs have branch
878 direction information, and only if they are conditional jumps.
880 If LABEL is a return, then there is no way to determine the branch
883 && (condjump_p (insn
) || condjump_in_parallel_p (insn
))
884 && !ANY_RETURN_P (label
)
885 && INSN_UID (insn
) <= max_uid
886 && INSN_UID (label
) <= max_uid
)
888 = (uid_to_ruid
[INSN_UID (label
)] > uid_to_ruid
[INSN_UID (insn
)])
889 ? ATTR_FLAG_forward
: ATTR_FLAG_backward
;
890 /* No valid direction information. */
897 /* Return truth value of the statement that this branch
898 is mostly taken. If we think that the branch is extremely likely
899 to be taken, we return 2. If the branch is slightly more likely to be
900 taken, return 1. If the branch is slightly less likely to be taken,
901 return 0 and if the branch is highly unlikely to be taken, return -1. */
904 mostly_true_jump (rtx jump_insn
)
906 /* If branch probabilities are available, then use that number since it
907 always gives a correct answer. */
908 rtx note
= find_reg_note (jump_insn
, REG_BR_PROB
, 0);
911 int prob
= XINT (note
, 0);
913 if (prob
>= REG_BR_PROB_BASE
* 9 / 10)
915 else if (prob
>= REG_BR_PROB_BASE
/ 2)
917 else if (prob
>= REG_BR_PROB_BASE
/ 10)
923 /* If there is no note, assume branches are not taken.
924 This should be rare. */
928 /* Return the condition under which INSN will branch to TARGET. If TARGET
929 is zero, return the condition under which INSN will return. If INSN is
930 an unconditional branch, return const_true_rtx. If INSN isn't a simple
931 type of jump, or it doesn't go to TARGET, return 0. */
934 get_branch_condition (const rtx_insn
*insn
, rtx target
)
936 rtx pat
= PATTERN (insn
);
939 if (condjump_in_parallel_p (insn
))
940 pat
= XVECEXP (pat
, 0, 0);
942 if (ANY_RETURN_P (pat
) && pat
== target
)
943 return const_true_rtx
;
945 if (GET_CODE (pat
) != SET
|| SET_DEST (pat
) != pc_rtx
)
949 if (GET_CODE (src
) == LABEL_REF
&& LABEL_REF_LABEL (src
) == target
)
950 return const_true_rtx
;
952 else if (GET_CODE (src
) == IF_THEN_ELSE
953 && XEXP (src
, 2) == pc_rtx
954 && ((GET_CODE (XEXP (src
, 1)) == LABEL_REF
955 && LABEL_REF_LABEL (XEXP (src
, 1)) == target
)
956 || (ANY_RETURN_P (XEXP (src
, 1)) && XEXP (src
, 1) == target
)))
957 return XEXP (src
, 0);
959 else if (GET_CODE (src
) == IF_THEN_ELSE
960 && XEXP (src
, 1) == pc_rtx
961 && ((GET_CODE (XEXP (src
, 2)) == LABEL_REF
962 && LABEL_REF_LABEL (XEXP (src
, 2)) == target
)
963 || (ANY_RETURN_P (XEXP (src
, 2)) && XEXP (src
, 2) == target
)))
966 rev
= reversed_comparison_code (XEXP (src
, 0), insn
);
968 return gen_rtx_fmt_ee (rev
, GET_MODE (XEXP (src
, 0)),
969 XEXP (XEXP (src
, 0), 0),
970 XEXP (XEXP (src
, 0), 1));
976 /* Return nonzero if CONDITION is more strict than the condition of
977 INSN, i.e., if INSN will always branch if CONDITION is true. */
980 condition_dominates_p (rtx condition
, const rtx_insn
*insn
)
982 rtx other_condition
= get_branch_condition (insn
, JUMP_LABEL (insn
));
983 enum rtx_code code
= GET_CODE (condition
);
984 enum rtx_code other_code
;
986 if (rtx_equal_p (condition
, other_condition
)
987 || other_condition
== const_true_rtx
)
990 else if (condition
== const_true_rtx
|| other_condition
== 0)
993 other_code
= GET_CODE (other_condition
);
994 if (GET_RTX_LENGTH (code
) != 2 || GET_RTX_LENGTH (other_code
) != 2
995 || ! rtx_equal_p (XEXP (condition
, 0), XEXP (other_condition
, 0))
996 || ! rtx_equal_p (XEXP (condition
, 1), XEXP (other_condition
, 1)))
999 return comparison_dominates_p (code
, other_code
);
1002 /* Return nonzero if redirecting JUMP to NEWLABEL does not invalidate
1003 any insns already in the delay slot of JUMP. */
1006 redirect_with_delay_slots_safe_p (rtx_insn
*jump
, rtx newlabel
, rtx seq
)
1009 rtx_sequence
*pat
= as_a
<rtx_sequence
*> (PATTERN (seq
));
1011 /* Make sure all the delay slots of this jump would still
1012 be valid after threading the jump. If they are still
1013 valid, then return nonzero. */
1015 flags
= get_jump_flags (jump
, newlabel
);
1016 for (i
= 1; i
< pat
->len (); i
++)
1018 #ifdef ANNUL_IFFALSE_SLOTS
1019 (INSN_ANNULLED_BRANCH_P (jump
)
1020 && INSN_FROM_TARGET_P (pat
->insn (i
)))
1021 ? eligible_for_annul_false (jump
, i
- 1, pat
->insn (i
), flags
) :
1023 #ifdef ANNUL_IFTRUE_SLOTS
1024 (INSN_ANNULLED_BRANCH_P (jump
)
1025 && ! INSN_FROM_TARGET_P (XVECEXP (pat
, 0, i
)))
1026 ? eligible_for_annul_true (jump
, i
- 1, pat
->insn (i
), flags
) :
1028 eligible_for_delay (jump
, i
- 1, pat
->insn (i
), flags
)))
1031 return (i
== pat
->len ());
1034 /* Return nonzero if redirecting JUMP to NEWLABEL does not invalidate
1035 any insns we wish to place in the delay slot of JUMP. */
1038 redirect_with_delay_list_safe_p (rtx_insn
*jump
, rtx newlabel
,
1039 rtx_insn_list
*delay_list
)
1044 /* Make sure all the insns in DELAY_LIST would still be
1045 valid after threading the jump. If they are still
1046 valid, then return nonzero. */
1048 flags
= get_jump_flags (jump
, newlabel
);
1049 for (li
= delay_list
, i
= 0; li
; li
= li
->next (), i
++)
1051 #ifdef ANNUL_IFFALSE_SLOTS
1052 (INSN_ANNULLED_BRANCH_P (jump
)
1053 && INSN_FROM_TARGET_P (li
->insn ()))
1054 ? eligible_for_annul_false (jump
, i
, li
->insn (), flags
) :
1056 #ifdef ANNUL_IFTRUE_SLOTS
1057 (INSN_ANNULLED_BRANCH_P (jump
)
1058 && ! INSN_FROM_TARGET_P (XEXP (li
, 0)))
1059 ? eligible_for_annul_true (jump
, i
, li
->insn (), flags
) :
1061 eligible_for_delay (jump
, i
, li
->insn (), flags
)))
1064 return (li
== NULL
);
1067 /* DELAY_LIST is a list of insns that have already been placed into delay
1068 slots. See if all of them have the same annulling status as ANNUL_TRUE_P.
1069 If not, return 0; otherwise return 1. */
1072 check_annul_list_true_false (int annul_true_p
, rtx delay_list
)
1078 for (temp
= delay_list
; temp
; temp
= XEXP (temp
, 1))
1080 rtx trial
= XEXP (temp
, 0);
1082 if ((annul_true_p
&& INSN_FROM_TARGET_P (trial
))
1083 || (!annul_true_p
&& !INSN_FROM_TARGET_P (trial
)))
1091 /* INSN branches to an insn whose pattern SEQ is a SEQUENCE. Given that
1092 the condition tested by INSN is CONDITION and the resources shown in
1093 OTHER_NEEDED are needed after INSN, see whether INSN can take all the insns
1094 from SEQ's delay list, in addition to whatever insns it may execute
1095 (in DELAY_LIST). SETS and NEEDED are denote resources already set and
1096 needed while searching for delay slot insns. Return the concatenated
1097 delay list if possible, otherwise, return 0.
1099 SLOTS_TO_FILL is the total number of slots required by INSN, and
1100 PSLOTS_FILLED points to the number filled so far (also the number of
1101 insns in DELAY_LIST). It is updated with the number that have been
1102 filled from the SEQUENCE, if any.
1104 PANNUL_P points to a nonzero value if we already know that we need
1105 to annul INSN. If this routine determines that annulling is needed,
1106 it may set that value nonzero.
1108 PNEW_THREAD points to a location that is to receive the place at which
1109 execution should continue. */
1111 static rtx_insn_list
*
1112 steal_delay_list_from_target (rtx_insn
*insn
, rtx condition
, rtx_sequence
*seq
,
1113 rtx_insn_list
*delay_list
, struct resources
*sets
,
1114 struct resources
*needed
,
1115 struct resources
*other_needed
,
1116 int slots_to_fill
, int *pslots_filled
,
1117 int *pannul_p
, rtx
*pnew_thread
)
1119 int slots_remaining
= slots_to_fill
- *pslots_filled
;
1120 int total_slots_filled
= *pslots_filled
;
1121 rtx_insn_list
*new_delay_list
= 0;
1122 int must_annul
= *pannul_p
;
1125 struct resources cc_set
;
1128 /* We can't do anything if there are more delay slots in SEQ than we
1129 can handle, or if we don't know that it will be a taken branch.
1130 We know that it will be a taken branch if it is either an unconditional
1131 branch or a conditional branch with a stricter branch condition.
1133 Also, exit if the branch has more than one set, since then it is computing
1134 other results that can't be ignored, e.g. the HPPA mov&branch instruction.
1135 ??? It may be possible to move other sets into INSN in addition to
1136 moving the instructions in the delay slots.
1138 We can not steal the delay list if one of the instructions in the
1139 current delay_list modifies the condition codes and the jump in the
1140 sequence is a conditional jump. We can not do this because we can
1141 not change the direction of the jump because the condition codes
1142 will effect the direction of the jump in the sequence. */
1144 CLEAR_RESOURCE (&cc_set
);
1145 for (rtx_insn_list
*temp
= delay_list
; temp
; temp
= temp
->next ())
1147 rtx_insn
*trial
= temp
->insn ();
1149 mark_set_resources (trial
, &cc_set
, 0, MARK_SRC_DEST_CALL
);
1150 if (insn_references_resource_p (seq
->insn (0), &cc_set
, false))
1154 if (XVECLEN (seq
, 0) - 1 > slots_remaining
1155 || ! condition_dominates_p (condition
, seq
->insn (0))
1156 || ! single_set (seq
->insn (0)))
1159 #ifdef MD_CAN_REDIRECT_BRANCH
1160 /* On some targets, branches with delay slots can have a limited
1161 displacement. Give the back end a chance to tell us we can't do
1163 if (! MD_CAN_REDIRECT_BRANCH (insn
, seq
->insn (0)))
1167 redundant
= XALLOCAVEC (bool, XVECLEN (seq
, 0));
1168 for (i
= 1; i
< seq
->len (); i
++)
1170 rtx_insn
*trial
= seq
->insn (i
);
1173 if (insn_references_resource_p (trial
, sets
, false)
1174 || insn_sets_resource_p (trial
, needed
, false)
1175 || insn_sets_resource_p (trial
, sets
, false)
1177 /* If TRIAL sets CC0, we can't copy it, so we can't steal this
1179 || find_reg_note (trial
, REG_CC_USER
, NULL_RTX
)
1181 /* If TRIAL is from the fallthrough code of an annulled branch insn
1182 in SEQ, we cannot use it. */
1183 || (INSN_ANNULLED_BRANCH_P (seq
->insn (0))
1184 && ! INSN_FROM_TARGET_P (trial
)))
1187 /* If this insn was already done (usually in a previous delay slot),
1188 pretend we put it in our delay slot. */
1189 redundant
[i
] = redundant_insn (trial
, insn
, new_delay_list
);
1193 /* We will end up re-vectoring this branch, so compute flags
1194 based on jumping to the new label. */
1195 flags
= get_jump_flags (insn
, JUMP_LABEL (seq
->insn (0)));
1198 && ((condition
== const_true_rtx
1199 || (! insn_sets_resource_p (trial
, other_needed
, false)
1200 && ! may_trap_or_fault_p (PATTERN (trial
)))))
1201 ? eligible_for_delay (insn
, total_slots_filled
, trial
, flags
)
1202 : (must_annul
|| (delay_list
== NULL
&& new_delay_list
== NULL
))
1204 check_annul_list_true_false (0, delay_list
)
1205 && check_annul_list_true_false (0, new_delay_list
)
1206 && eligible_for_annul_false (insn
, total_slots_filled
,
1211 rtx_insn
*temp
= copy_delay_slot_insn (trial
);
1212 INSN_FROM_TARGET_P (temp
) = 1;
1213 new_delay_list
= add_to_delay_list (temp
, new_delay_list
);
1214 total_slots_filled
++;
1216 if (--slots_remaining
== 0)
1223 /* Record the effect of the instructions that were redundant and which
1224 we therefore decided not to copy. */
1225 for (i
= 1; i
< seq
->len (); i
++)
1227 update_block (seq
->insn (i
), insn
);
1229 /* Show the place to which we will be branching. */
1230 *pnew_thread
= first_active_target_insn (JUMP_LABEL (seq
->insn (0)));
1232 /* Add any new insns to the delay list and update the count of the
1233 number of slots filled. */
1234 *pslots_filled
= total_slots_filled
;
1238 if (delay_list
== 0)
1239 return new_delay_list
;
1241 for (rtx_insn_list
*temp
= new_delay_list
; temp
; temp
= temp
->next ())
1242 delay_list
= add_to_delay_list (temp
->insn (), delay_list
);
1247 /* Similar to steal_delay_list_from_target except that SEQ is on the
1248 fallthrough path of INSN. Here we only do something if the delay insn
1249 of SEQ is an unconditional branch. In that case we steal its delay slot
1250 for INSN since unconditional branches are much easier to fill. */
1252 static rtx_insn_list
*
1253 steal_delay_list_from_fallthrough (rtx_insn
*insn
, rtx condition
,
1255 rtx_insn_list
*delay_list
,
1256 struct resources
*sets
,
1257 struct resources
*needed
,
1258 struct resources
*other_needed
,
1259 int slots_to_fill
, int *pslots_filled
,
1264 int must_annul
= *pannul_p
;
1267 flags
= get_jump_flags (insn
, JUMP_LABEL (insn
));
1269 /* We can't do anything if SEQ's delay insn isn't an
1270 unconditional branch. */
1272 if (! simplejump_or_return_p (seq
->insn (0)))
1275 for (i
= 1; i
< seq
->len (); i
++)
1277 rtx_insn
*trial
= seq
->insn (i
);
1279 /* If TRIAL sets CC0, stealing it will move it too far from the use
1281 if (insn_references_resource_p (trial
, sets
, false)
1282 || insn_sets_resource_p (trial
, needed
, false)
1283 || insn_sets_resource_p (trial
, sets
, false)
1285 || sets_cc0_p (PATTERN (trial
))
1291 /* If this insn was already done, we don't need it. */
1292 if (redundant_insn (trial
, insn
, delay_list
))
1294 update_block (trial
, insn
);
1295 delete_from_delay_slot (trial
);
1300 && ((condition
== const_true_rtx
1301 || (! insn_sets_resource_p (trial
, other_needed
, false)
1302 && ! may_trap_or_fault_p (PATTERN (trial
)))))
1303 ? eligible_for_delay (insn
, *pslots_filled
, trial
, flags
)
1304 : (must_annul
|| delay_list
== NULL
) && (must_annul
= 1,
1305 check_annul_list_true_false (1, delay_list
)
1306 && eligible_for_annul_true (insn
, *pslots_filled
, trial
, flags
)))
1310 delete_from_delay_slot (trial
);
1311 delay_list
= add_to_delay_list (trial
, delay_list
);
1313 if (++(*pslots_filled
) == slots_to_fill
)
1325 /* Try merging insns starting at THREAD which match exactly the insns in
1328 If all insns were matched and the insn was previously annulling, the
1329 annul bit will be cleared.
1331 For each insn that is merged, if the branch is or will be non-annulling,
1332 we delete the merged insn. */
1335 try_merge_delay_insns (rtx insn
, rtx_insn
*thread
)
1337 rtx_insn
*trial
, *next_trial
;
1338 rtx_insn
*delay_insn
= as_a
<rtx_insn
*> (XVECEXP (PATTERN (insn
), 0, 0));
1339 int annul_p
= JUMP_P (delay_insn
) && INSN_ANNULLED_BRANCH_P (delay_insn
);
1340 int slot_number
= 1;
1341 int num_slots
= XVECLEN (PATTERN (insn
), 0);
1342 rtx next_to_match
= XVECEXP (PATTERN (insn
), 0, slot_number
);
1343 struct resources set
, needed
;
1344 rtx_insn_list
*merged_insns
= 0;
1348 flags
= get_jump_flags (delay_insn
, JUMP_LABEL (delay_insn
));
1350 CLEAR_RESOURCE (&needed
);
1351 CLEAR_RESOURCE (&set
);
1353 /* If this is not an annulling branch, take into account anything needed in
1354 INSN's delay slot. This prevents two increments from being incorrectly
1355 folded into one. If we are annulling, this would be the correct
1356 thing to do. (The alternative, looking at things set in NEXT_TO_MATCH
1357 will essentially disable this optimization. This method is somewhat of
1358 a kludge, but I don't see a better way.) */
1360 for (i
= 1 ; i
< num_slots
; i
++)
1361 if (XVECEXP (PATTERN (insn
), 0, i
))
1362 mark_referenced_resources (XVECEXP (PATTERN (insn
), 0, i
), &needed
,
1365 for (trial
= thread
; !stop_search_p (trial
, 1); trial
= next_trial
)
1367 rtx pat
= PATTERN (trial
);
1368 rtx oldtrial
= trial
;
1370 next_trial
= next_nonnote_insn (trial
);
1372 /* TRIAL must be a CALL_INSN or INSN. Skip USE and CLOBBER. */
1373 if (NONJUMP_INSN_P (trial
)
1374 && (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
))
1377 if (GET_CODE (next_to_match
) == GET_CODE (trial
)
1379 /* We can't share an insn that sets cc0. */
1380 && ! sets_cc0_p (pat
)
1382 && ! insn_references_resource_p (trial
, &set
, true)
1383 && ! insn_sets_resource_p (trial
, &set
, true)
1384 && ! insn_sets_resource_p (trial
, &needed
, true)
1385 && (trial
= try_split (pat
, trial
, 0)) != 0
1386 /* Update next_trial, in case try_split succeeded. */
1387 && (next_trial
= next_nonnote_insn (trial
))
1388 /* Likewise THREAD. */
1389 && (thread
= oldtrial
== thread
? trial
: thread
)
1390 && rtx_equal_p (PATTERN (next_to_match
), PATTERN (trial
))
1391 /* Have to test this condition if annul condition is different
1392 from (and less restrictive than) non-annulling one. */
1393 && eligible_for_delay (delay_insn
, slot_number
- 1, trial
, flags
))
1398 update_block (trial
, thread
);
1399 if (trial
== thread
)
1400 thread
= next_active_insn (thread
);
1402 delete_related_insns (trial
);
1403 INSN_FROM_TARGET_P (next_to_match
) = 0;
1406 merged_insns
= gen_rtx_INSN_LIST (VOIDmode
, trial
, merged_insns
);
1408 if (++slot_number
== num_slots
)
1411 next_to_match
= XVECEXP (PATTERN (insn
), 0, slot_number
);
1414 mark_set_resources (trial
, &set
, 0, MARK_SRC_DEST_CALL
);
1415 mark_referenced_resources (trial
, &needed
, true);
1418 /* See if we stopped on a filled insn. If we did, try to see if its
1419 delay slots match. */
1420 if (slot_number
!= num_slots
1421 && trial
&& NONJUMP_INSN_P (trial
)
1422 && GET_CODE (PATTERN (trial
)) == SEQUENCE
1423 && !(JUMP_P (XVECEXP (PATTERN (trial
), 0, 0))
1424 && INSN_ANNULLED_BRANCH_P (XVECEXP (PATTERN (trial
), 0, 0))))
1426 rtx_sequence
*pat
= as_a
<rtx_sequence
*> (PATTERN (trial
));
1427 rtx filled_insn
= XVECEXP (pat
, 0, 0);
1429 /* Account for resources set/needed by the filled insn. */
1430 mark_set_resources (filled_insn
, &set
, 0, MARK_SRC_DEST_CALL
);
1431 mark_referenced_resources (filled_insn
, &needed
, true);
1433 for (i
= 1; i
< pat
->len (); i
++)
1435 rtx_insn
*dtrial
= pat
->insn (i
);
1437 if (! insn_references_resource_p (dtrial
, &set
, true)
1438 && ! insn_sets_resource_p (dtrial
, &set
, true)
1439 && ! insn_sets_resource_p (dtrial
, &needed
, true)
1441 && ! sets_cc0_p (PATTERN (dtrial
))
1443 && rtx_equal_p (PATTERN (next_to_match
), PATTERN (dtrial
))
1444 && eligible_for_delay (delay_insn
, slot_number
- 1, dtrial
, flags
))
1450 update_block (dtrial
, thread
);
1451 new_rtx
= delete_from_delay_slot (dtrial
);
1452 if (thread
->deleted ())
1454 INSN_FROM_TARGET_P (next_to_match
) = 0;
1457 merged_insns
= gen_rtx_INSN_LIST (SImode
, dtrial
,
1460 if (++slot_number
== num_slots
)
1463 next_to_match
= XVECEXP (PATTERN (insn
), 0, slot_number
);
1467 /* Keep track of the set/referenced resources for the delay
1468 slots of any trial insns we encounter. */
1469 mark_set_resources (dtrial
, &set
, 0, MARK_SRC_DEST_CALL
);
1470 mark_referenced_resources (dtrial
, &needed
, true);
1475 /* If all insns in the delay slot have been matched and we were previously
1476 annulling the branch, we need not any more. In that case delete all the
1477 merged insns. Also clear the INSN_FROM_TARGET_P bit of each insn in
1478 the delay list so that we know that it isn't only being used at the
1480 if (slot_number
== num_slots
&& annul_p
)
1482 for (; merged_insns
; merged_insns
= merged_insns
->next ())
1484 if (GET_MODE (merged_insns
) == SImode
)
1488 update_block (merged_insns
->insn (), thread
);
1489 new_rtx
= delete_from_delay_slot (merged_insns
->insn ());
1490 if (thread
->deleted ())
1495 update_block (merged_insns
->insn (), thread
);
1496 delete_related_insns (merged_insns
->insn ());
1500 INSN_ANNULLED_BRANCH_P (delay_insn
) = 0;
1502 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
1503 INSN_FROM_TARGET_P (XVECEXP (PATTERN (insn
), 0, i
)) = 0;
1507 /* See if INSN is redundant with an insn in front of TARGET. Often this
1508 is called when INSN is a candidate for a delay slot of TARGET.
1509 DELAY_LIST are insns that will be placed in delay slots of TARGET in front
1510 of INSN. Often INSN will be redundant with an insn in a delay slot of
1511 some previous insn. This happens when we have a series of branches to the
1512 same label; in that case the first insn at the target might want to go
1513 into each of the delay slots.
1515 If we are not careful, this routine can take up a significant fraction
1516 of the total compilation time (4%), but only wins rarely. Hence we
1517 speed this routine up by making two passes. The first pass goes back
1518 until it hits a label and sees if it finds an insn with an identical
1519 pattern. Only in this (relatively rare) event does it check for
1522 We do not split insns we encounter. This could cause us not to find a
1523 redundant insn, but the cost of splitting seems greater than the possible
1524 gain in rare cases. */
1527 redundant_insn (rtx insn
, rtx_insn
*target
, rtx delay_list
)
1529 rtx target_main
= target
;
1530 rtx ipat
= PATTERN (insn
);
1533 struct resources needed
, set
;
1535 unsigned insns_to_search
;
1537 /* If INSN has any REG_UNUSED notes, it can't match anything since we
1538 are allowed to not actually assign to such a register. */
1539 if (find_reg_note (insn
, REG_UNUSED
, NULL_RTX
) != 0)
1542 /* Scan backwards looking for a match. */
1543 for (trial
= PREV_INSN (target
),
1544 insns_to_search
= MAX_DELAY_SLOT_INSN_SEARCH
;
1545 trial
&& insns_to_search
> 0;
1546 trial
= PREV_INSN (trial
))
1548 /* (use (insn))s can come immediately after a barrier if the
1549 label that used to precede them has been deleted as dead.
1550 See delete_related_insns. */
1551 if (LABEL_P (trial
) || BARRIER_P (trial
))
1554 if (!INSN_P (trial
))
1558 pat
= PATTERN (trial
);
1559 if (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
)
1562 if (rtx_sequence
*seq
= dyn_cast
<rtx_sequence
*> (pat
))
1564 /* Stop for a CALL and its delay slots because it is difficult to
1565 track its resource needs correctly. */
1566 if (CALL_P (seq
->element (0)))
1569 /* Stop for an INSN or JUMP_INSN with delayed effects and its delay
1570 slots because it is difficult to track its resource needs
1573 #ifdef INSN_SETS_ARE_DELAYED
1574 if (INSN_SETS_ARE_DELAYED (seq
->insn (0)))
1578 #ifdef INSN_REFERENCES_ARE_DELAYED
1579 if (INSN_REFERENCES_ARE_DELAYED (seq
->insn (0)))
1583 /* See if any of the insns in the delay slot match, updating
1584 resource requirements as we go. */
1585 for (i
= seq
->len () - 1; i
> 0; i
--)
1586 if (GET_CODE (seq
->element (i
)) == GET_CODE (insn
)
1587 && rtx_equal_p (PATTERN (seq
->element (i
)), ipat
)
1588 && ! find_reg_note (seq
->element (i
), REG_UNUSED
, NULL_RTX
))
1591 /* If found a match, exit this loop early. */
1596 else if (GET_CODE (trial
) == GET_CODE (insn
) && rtx_equal_p (pat
, ipat
)
1597 && ! find_reg_note (trial
, REG_UNUSED
, NULL_RTX
))
1601 /* If we didn't find an insn that matches, return 0. */
1605 /* See what resources this insn sets and needs. If they overlap, or
1606 if this insn references CC0, it can't be redundant. */
1608 CLEAR_RESOURCE (&needed
);
1609 CLEAR_RESOURCE (&set
);
1610 mark_set_resources (insn
, &set
, 0, MARK_SRC_DEST_CALL
);
1611 mark_referenced_resources (insn
, &needed
, true);
1613 /* If TARGET is a SEQUENCE, get the main insn. */
1614 if (NONJUMP_INSN_P (target
) && GET_CODE (PATTERN (target
)) == SEQUENCE
)
1615 target_main
= XVECEXP (PATTERN (target
), 0, 0);
1617 if (resource_conflicts_p (&needed
, &set
)
1619 || reg_mentioned_p (cc0_rtx
, ipat
)
1621 /* The insn requiring the delay may not set anything needed or set by
1623 || insn_sets_resource_p (target_main
, &needed
, true)
1624 || insn_sets_resource_p (target_main
, &set
, true))
1627 /* Insns we pass may not set either NEEDED or SET, so merge them for
1629 needed
.memory
|= set
.memory
;
1630 IOR_HARD_REG_SET (needed
.regs
, set
.regs
);
1632 /* This insn isn't redundant if it conflicts with an insn that either is
1633 or will be in a delay slot of TARGET. */
1637 if (insn_sets_resource_p (XEXP (delay_list
, 0), &needed
, true))
1639 delay_list
= XEXP (delay_list
, 1);
1642 if (NONJUMP_INSN_P (target
) && GET_CODE (PATTERN (target
)) == SEQUENCE
)
1643 for (i
= 1; i
< XVECLEN (PATTERN (target
), 0); i
++)
1644 if (insn_sets_resource_p (XVECEXP (PATTERN (target
), 0, i
), &needed
,
1648 /* Scan backwards until we reach a label or an insn that uses something
1649 INSN sets or sets something insn uses or sets. */
1651 for (trial
= PREV_INSN (target
),
1652 insns_to_search
= MAX_DELAY_SLOT_INSN_SEARCH
;
1653 trial
&& !LABEL_P (trial
) && insns_to_search
> 0;
1654 trial
= PREV_INSN (trial
))
1656 if (!INSN_P (trial
))
1660 pat
= PATTERN (trial
);
1661 if (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
)
1664 if (rtx_sequence
*seq
= dyn_cast
<rtx_sequence
*> (pat
))
1666 bool annul_p
= false;
1667 rtx_insn
*control
= seq
->insn (0);
1669 /* If this is a CALL_INSN and its delay slots, it is hard to track
1670 the resource needs properly, so give up. */
1671 if (CALL_P (control
))
1674 /* If this is an INSN or JUMP_INSN with delayed effects, it
1675 is hard to track the resource needs properly, so give up. */
1677 #ifdef INSN_SETS_ARE_DELAYED
1678 if (INSN_SETS_ARE_DELAYED (control
))
1682 #ifdef INSN_REFERENCES_ARE_DELAYED
1683 if (INSN_REFERENCES_ARE_DELAYED (control
))
1687 if (JUMP_P (control
))
1688 annul_p
= INSN_ANNULLED_BRANCH_P (control
);
1690 /* See if any of the insns in the delay slot match, updating
1691 resource requirements as we go. */
1692 for (i
= seq
->len () - 1; i
> 0; i
--)
1694 rtx candidate
= seq
->element (i
);
1696 /* If an insn will be annulled if the branch is false, it isn't
1697 considered as a possible duplicate insn. */
1698 if (rtx_equal_p (PATTERN (candidate
), ipat
)
1699 && ! (annul_p
&& INSN_FROM_TARGET_P (candidate
)))
1701 /* Show that this insn will be used in the sequel. */
1702 INSN_FROM_TARGET_P (candidate
) = 0;
1706 /* Unless this is an annulled insn from the target of a branch,
1707 we must stop if it sets anything needed or set by INSN. */
1708 if ((!annul_p
|| !INSN_FROM_TARGET_P (candidate
))
1709 && insn_sets_resource_p (candidate
, &needed
, true))
1713 /* If the insn requiring the delay slot conflicts with INSN, we
1715 if (insn_sets_resource_p (control
, &needed
, true))
1720 /* See if TRIAL is the same as INSN. */
1721 pat
= PATTERN (trial
);
1722 if (rtx_equal_p (pat
, ipat
))
1725 /* Can't go any further if TRIAL conflicts with INSN. */
1726 if (insn_sets_resource_p (trial
, &needed
, true))
1734 /* Return 1 if THREAD can only be executed in one way. If LABEL is nonzero,
1735 it is the target of the branch insn being scanned. If ALLOW_FALLTHROUGH
1736 is nonzero, we are allowed to fall into this thread; otherwise, we are
1739 If LABEL is used more than one or we pass a label other than LABEL before
1740 finding an active insn, we do not own this thread. */
1743 own_thread_p (rtx thread
, rtx label
, int allow_fallthrough
)
1745 rtx_insn
*active_insn
;
1748 /* We don't own the function end. */
1749 if (thread
== 0 || ANY_RETURN_P (thread
))
1752 /* We have a non-NULL insn. */
1753 rtx_insn
*thread_insn
= as_a
<rtx_insn
*> (thread
);
1755 /* Get the first active insn, or THREAD_INSN, if it is an active insn. */
1756 active_insn
= next_active_insn (PREV_INSN (thread_insn
));
1758 for (insn
= thread_insn
; insn
!= active_insn
; insn
= NEXT_INSN (insn
))
1760 && (insn
!= label
|| LABEL_NUSES (insn
) != 1))
1763 if (allow_fallthrough
)
1766 /* Ensure that we reach a BARRIER before any insn or label. */
1767 for (insn
= prev_nonnote_insn (thread_insn
);
1768 insn
== 0 || !BARRIER_P (insn
);
1769 insn
= prev_nonnote_insn (insn
))
1772 || (NONJUMP_INSN_P (insn
)
1773 && GET_CODE (PATTERN (insn
)) != USE
1774 && GET_CODE (PATTERN (insn
)) != CLOBBER
))
1780 /* Called when INSN is being moved from a location near the target of a jump.
1781 We leave a marker of the form (use (INSN)) immediately in front
1782 of WHERE for mark_target_live_regs. These markers will be deleted when
1785 We used to try to update the live status of registers if WHERE is at
1786 the start of a basic block, but that can't work since we may remove a
1787 BARRIER in relax_delay_slots. */
1790 update_block (rtx_insn
*insn
, rtx where
)
1792 /* Ignore if this was in a delay slot and it came from the target of
1794 if (INSN_FROM_TARGET_P (insn
))
1797 emit_insn_before (gen_rtx_USE (VOIDmode
, insn
), where
);
1799 /* INSN might be making a value live in a block where it didn't use to
1800 be. So recompute liveness information for this block. */
1802 incr_ticks_for_insn (insn
);
1805 /* Similar to REDIRECT_JUMP except that we update the BB_TICKS entry for
1806 the basic block containing the jump. */
1809 reorg_redirect_jump (rtx_insn
*jump
, rtx nlabel
)
1811 incr_ticks_for_insn (jump
);
1812 return redirect_jump (jump
, nlabel
, 1);
1815 /* Called when INSN is being moved forward into a delay slot of DELAYED_INSN.
1816 We check every instruction between INSN and DELAYED_INSN for REG_DEAD notes
1817 that reference values used in INSN. If we find one, then we move the
1818 REG_DEAD note to INSN.
1820 This is needed to handle the case where a later insn (after INSN) has a
1821 REG_DEAD note for a register used by INSN, and this later insn subsequently
1822 gets moved before a CODE_LABEL because it is a redundant insn. In this
1823 case, mark_target_live_regs may be confused into thinking the register
1824 is dead because it sees a REG_DEAD note immediately before a CODE_LABEL. */
1827 update_reg_dead_notes (rtx insn
, rtx delayed_insn
)
1831 for (p
= next_nonnote_insn (insn
); p
!= delayed_insn
;
1832 p
= next_nonnote_insn (p
))
1833 for (link
= REG_NOTES (p
); link
; link
= next
)
1835 next
= XEXP (link
, 1);
1837 if (REG_NOTE_KIND (link
) != REG_DEAD
1838 || !REG_P (XEXP (link
, 0)))
1841 if (reg_referenced_p (XEXP (link
, 0), PATTERN (insn
)))
1843 /* Move the REG_DEAD note from P to INSN. */
1844 remove_note (p
, link
);
1845 XEXP (link
, 1) = REG_NOTES (insn
);
1846 REG_NOTES (insn
) = link
;
1851 /* Called when an insn redundant with start_insn is deleted. If there
1852 is a REG_DEAD note for the target of start_insn between start_insn
1853 and stop_insn, then the REG_DEAD note needs to be deleted since the
1854 value no longer dies there.
1856 If the REG_DEAD note isn't deleted, then mark_target_live_regs may be
1857 confused into thinking the register is dead. */
1860 fix_reg_dead_note (rtx start_insn
, rtx stop_insn
)
1864 for (p
= next_nonnote_insn (start_insn
); p
!= stop_insn
;
1865 p
= next_nonnote_insn (p
))
1866 for (link
= REG_NOTES (p
); link
; link
= next
)
1868 next
= XEXP (link
, 1);
1870 if (REG_NOTE_KIND (link
) != REG_DEAD
1871 || !REG_P (XEXP (link
, 0)))
1874 if (reg_set_p (XEXP (link
, 0), PATTERN (start_insn
)))
1876 remove_note (p
, link
);
1882 /* Delete any REG_UNUSED notes that exist on INSN but not on REDUNDANT_INSN.
1884 This handles the case of udivmodXi4 instructions which optimize their
1885 output depending on whether any REG_UNUSED notes are present.
1886 we must make sure that INSN calculates as many results as REDUNDANT_INSN
1890 update_reg_unused_notes (rtx insn
, rtx redundant_insn
)
1894 for (link
= REG_NOTES (insn
); link
; link
= next
)
1896 next
= XEXP (link
, 1);
1898 if (REG_NOTE_KIND (link
) != REG_UNUSED
1899 || !REG_P (XEXP (link
, 0)))
1902 if (! find_regno_note (redundant_insn
, REG_UNUSED
,
1903 REGNO (XEXP (link
, 0))))
1904 remove_note (insn
, link
);
1908 static vec
<rtx
> sibling_labels
;
1910 /* Return the label before INSN, or put a new label there. If SIBLING is
1911 non-zero, it is another label associated with the new label (if any),
1912 typically the former target of the jump that will be redirected to
1916 get_label_before (rtx_insn
*insn
, rtx sibling
)
1920 /* Find an existing label at this point
1921 or make a new one if there is none. */
1922 label
= prev_nonnote_insn (insn
);
1924 if (label
== 0 || !LABEL_P (label
))
1926 rtx_insn
*prev
= PREV_INSN (insn
);
1928 label
= gen_label_rtx ();
1929 emit_label_after (label
, prev
);
1930 LABEL_NUSES (label
) = 0;
1933 sibling_labels
.safe_push (label
);
1934 sibling_labels
.safe_push (sibling
);
1940 /* Scan a function looking for insns that need a delay slot and find insns to
1941 put into the delay slot.
1943 NON_JUMPS_P is nonzero if we are to only try to fill non-jump insns (such
1944 as calls). We do these first since we don't want jump insns (that are
1945 easier to fill) to get the only insns that could be used for non-jump insns.
1946 When it is zero, only try to fill JUMP_INSNs.
1948 When slots are filled in this manner, the insns (including the
1949 delay_insn) are put together in a SEQUENCE rtx. In this fashion,
1950 it is possible to tell whether a delay slot has really been filled
1951 or not. `final' knows how to deal with this, by communicating
1952 through FINAL_SEQUENCE. */
1955 fill_simple_delay_slots (int non_jumps_p
)
1957 rtx_insn
*insn
, *trial
, *next_trial
;
1960 int num_unfilled_slots
= unfilled_slots_next
- unfilled_slots_base
;
1961 struct resources needed
, set
;
1962 int slots_to_fill
, slots_filled
;
1963 rtx_insn_list
*delay_list
;
1965 for (i
= 0; i
< num_unfilled_slots
; i
++)
1968 /* Get the next insn to fill. If it has already had any slots assigned,
1969 we can't do anything with it. Maybe we'll improve this later. */
1971 insn
= unfilled_slots_base
[i
];
1974 || (NONJUMP_INSN_P (insn
)
1975 && GET_CODE (PATTERN (insn
)) == SEQUENCE
)
1976 || (JUMP_P (insn
) && non_jumps_p
)
1977 || (!JUMP_P (insn
) && ! non_jumps_p
))
1980 /* It may have been that this insn used to need delay slots, but
1981 now doesn't; ignore in that case. This can happen, for example,
1982 on the HP PA RISC, where the number of delay slots depends on
1983 what insns are nearby. */
1984 slots_to_fill
= num_delay_slots (insn
);
1986 /* Some machine description have defined instructions to have
1987 delay slots only in certain circumstances which may depend on
1988 nearby insns (which change due to reorg's actions).
1990 For example, the PA port normally has delay slots for unconditional
1993 However, the PA port claims such jumps do not have a delay slot
1994 if they are immediate successors of certain CALL_INSNs. This
1995 allows the port to favor filling the delay slot of the call with
1996 the unconditional jump. */
1997 if (slots_to_fill
== 0)
2000 /* This insn needs, or can use, some delay slots. SLOTS_TO_FILL
2001 says how many. After initialization, first try optimizing
2004 nop add %o7,.-L1,%o7
2008 If this case applies, the delay slot of the call is filled with
2009 the unconditional jump. This is done first to avoid having the
2010 delay slot of the call filled in the backward scan. Also, since
2011 the unconditional jump is likely to also have a delay slot, that
2012 insn must exist when it is subsequently scanned.
2014 This is tried on each insn with delay slots as some machines
2015 have insns which perform calls, but are not represented as
2022 flags
= get_jump_flags (insn
, JUMP_LABEL (insn
));
2024 flags
= get_jump_flags (insn
, NULL_RTX
);
2026 if ((trial
= next_active_insn (insn
))
2028 && simplejump_p (trial
)
2029 && eligible_for_delay (insn
, slots_filled
, trial
, flags
)
2030 && no_labels_between_p (insn
, trial
)
2031 && ! can_throw_internal (trial
))
2035 delay_list
= add_to_delay_list (trial
, delay_list
);
2037 /* TRIAL may have had its delay slot filled, then unfilled. When
2038 the delay slot is unfilled, TRIAL is placed back on the unfilled
2039 slots obstack. Unfortunately, it is placed on the end of the
2040 obstack, not in its original location. Therefore, we must search
2041 from entry i + 1 to the end of the unfilled slots obstack to
2042 try and find TRIAL. */
2043 tmp
= &unfilled_slots_base
[i
+ 1];
2044 while (*tmp
!= trial
&& tmp
!= unfilled_slots_next
)
2047 /* Remove the unconditional jump from consideration for delay slot
2048 filling and unthread it. */
2052 rtx_insn
*next
= NEXT_INSN (trial
);
2053 rtx_insn
*prev
= PREV_INSN (trial
);
2055 SET_NEXT_INSN (prev
) = next
;
2057 SET_PREV_INSN (next
) = prev
;
2061 /* Now, scan backwards from the insn to search for a potential
2062 delay-slot candidate. Stop searching when a label or jump is hit.
2064 For each candidate, if it is to go into the delay slot (moved
2065 forward in execution sequence), it must not need or set any resources
2066 that were set by later insns and must not set any resources that
2067 are needed for those insns.
2069 The delay slot insn itself sets resources unless it is a call
2070 (in which case the called routine, not the insn itself, is doing
2073 if (slots_filled
< slots_to_fill
)
2075 /* If the flags register is dead after the insn, then we want to be
2076 able to accept a candidate that clobbers it. For this purpose,
2077 we need to filter the flags register during life analysis, so
2078 that it doesn't create RAW and WAW dependencies, while still
2079 creating the necessary WAR dependencies. */
2081 = (slots_to_fill
== 1
2082 && targetm
.flags_regnum
!= INVALID_REGNUM
2083 && find_regno_note (insn
, REG_DEAD
, targetm
.flags_regnum
));
2084 struct resources fset
;
2085 CLEAR_RESOURCE (&needed
);
2086 CLEAR_RESOURCE (&set
);
2087 mark_set_resources (insn
, &set
, 0, MARK_SRC_DEST
);
2090 CLEAR_RESOURCE (&fset
);
2091 mark_set_resources (insn
, &fset
, 0, MARK_SRC_DEST
);
2093 mark_referenced_resources (insn
, &needed
, false);
2095 for (trial
= prev_nonnote_insn (insn
); ! stop_search_p (trial
, 1);
2098 next_trial
= prev_nonnote_insn (trial
);
2100 /* This must be an INSN or CALL_INSN. */
2101 pat
= PATTERN (trial
);
2103 /* Stand-alone USE and CLOBBER are just for flow. */
2104 if (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
)
2107 /* Check for resource conflict first, to avoid unnecessary
2109 if (! insn_references_resource_p (trial
, &set
, true)
2110 && ! insn_sets_resource_p (trial
,
2111 filter_flags
? &fset
: &set
,
2113 && ! insn_sets_resource_p (trial
, &needed
, true)
2115 /* Can't separate set of cc0 from its use. */
2116 && ! (reg_mentioned_p (cc0_rtx
, pat
) && ! sets_cc0_p (pat
))
2118 && ! can_throw_internal (trial
))
2120 trial
= try_split (pat
, trial
, 1);
2121 next_trial
= prev_nonnote_insn (trial
);
2122 if (eligible_for_delay (insn
, slots_filled
, trial
, flags
))
2124 /* In this case, we are searching backward, so if we
2125 find insns to put on the delay list, we want
2126 to put them at the head, rather than the
2127 tail, of the list. */
2129 update_reg_dead_notes (trial
, insn
);
2130 delay_list
= gen_rtx_INSN_LIST (VOIDmode
,
2132 update_block (trial
, trial
);
2133 delete_related_insns (trial
);
2134 if (slots_to_fill
== ++slots_filled
)
2140 mark_set_resources (trial
, &set
, 0, MARK_SRC_DEST_CALL
);
2143 mark_set_resources (trial
, &fset
, 0, MARK_SRC_DEST_CALL
);
2144 /* If the flags register is set, then it doesn't create RAW
2145 dependencies any longer and it also doesn't create WAW
2146 dependencies since it's dead after the original insn. */
2147 if (TEST_HARD_REG_BIT (fset
.regs
, targetm
.flags_regnum
))
2149 CLEAR_HARD_REG_BIT (needed
.regs
, targetm
.flags_regnum
);
2150 CLEAR_HARD_REG_BIT (fset
.regs
, targetm
.flags_regnum
);
2153 mark_referenced_resources (trial
, &needed
, true);
2157 /* If all needed slots haven't been filled, we come here. */
2159 /* Try to optimize case of jumping around a single insn. */
2160 #if defined(ANNUL_IFFALSE_SLOTS) || defined(ANNUL_IFTRUE_SLOTS)
2161 if (slots_filled
!= slots_to_fill
2164 && (condjump_p (insn
) || condjump_in_parallel_p (insn
))
2165 && !ANY_RETURN_P (JUMP_LABEL (insn
)))
2167 delay_list
= optimize_skip (insn
);
2173 /* Try to get insns from beyond the insn needing the delay slot.
2174 These insns can neither set or reference resources set in insns being
2175 skipped, cannot set resources in the insn being skipped, and, if this
2176 is a CALL_INSN (or a CALL_INSN is passed), cannot trap (because the
2177 call might not return).
2179 There used to be code which continued past the target label if
2180 we saw all uses of the target label. This code did not work,
2181 because it failed to account for some instructions which were
2182 both annulled and marked as from the target. This can happen as a
2183 result of optimize_skip. Since this code was redundant with
2184 fill_eager_delay_slots anyways, it was just deleted. */
2186 if (slots_filled
!= slots_to_fill
2187 /* If this instruction could throw an exception which is
2188 caught in the same function, then it's not safe to fill
2189 the delay slot with an instruction from beyond this
2190 point. For example, consider:
2201 Even though `i' is a local variable, we must be sure not
2202 to put `i = 3' in the delay slot if `f' might throw an
2205 Presumably, we should also check to see if we could get
2206 back to this function via `setjmp'. */
2207 && ! can_throw_internal (insn
)
2210 int maybe_never
= 0;
2211 rtx pat
, trial_delay
;
2213 CLEAR_RESOURCE (&needed
);
2214 CLEAR_RESOURCE (&set
);
2215 mark_set_resources (insn
, &set
, 0, MARK_SRC_DEST_CALL
);
2216 mark_referenced_resources (insn
, &needed
, true);
2221 for (trial
= next_nonnote_insn (insn
); !stop_search_p (trial
, 1);
2224 next_trial
= next_nonnote_insn (trial
);
2226 /* This must be an INSN or CALL_INSN. */
2227 pat
= PATTERN (trial
);
2229 /* Stand-alone USE and CLOBBER are just for flow. */
2230 if (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
)
2233 /* If this already has filled delay slots, get the insn needing
2235 if (GET_CODE (pat
) == SEQUENCE
)
2236 trial_delay
= XVECEXP (pat
, 0, 0);
2238 trial_delay
= trial
;
2240 /* Stop our search when seeing a jump. */
2241 if (JUMP_P (trial_delay
))
2244 /* See if we have a resource problem before we try to split. */
2245 if (GET_CODE (pat
) != SEQUENCE
2246 && ! insn_references_resource_p (trial
, &set
, true)
2247 && ! insn_sets_resource_p (trial
, &set
, true)
2248 && ! insn_sets_resource_p (trial
, &needed
, true)
2250 && ! (reg_mentioned_p (cc0_rtx
, pat
) && ! sets_cc0_p (pat
))
2252 && ! (maybe_never
&& may_trap_or_fault_p (pat
))
2253 && (trial
= try_split (pat
, trial
, 0))
2254 && eligible_for_delay (insn
, slots_filled
, trial
, flags
)
2255 && ! can_throw_internal (trial
))
2257 next_trial
= next_nonnote_insn (trial
);
2258 delay_list
= add_to_delay_list (trial
, delay_list
);
2260 if (reg_mentioned_p (cc0_rtx
, pat
))
2261 link_cc0_insns (trial
);
2263 delete_related_insns (trial
);
2264 if (slots_to_fill
== ++slots_filled
)
2269 mark_set_resources (trial
, &set
, 0, MARK_SRC_DEST_CALL
);
2270 mark_referenced_resources (trial
, &needed
, true);
2272 /* Ensure we don't put insns between the setting of cc and the
2273 comparison by moving a setting of cc into an earlier delay
2274 slot since these insns could clobber the condition code. */
2277 /* If this is a call, we might not get here. */
2278 if (CALL_P (trial_delay
))
2282 /* If there are slots left to fill and our search was stopped by an
2283 unconditional branch, try the insn at the branch target. We can
2284 redirect the branch if it works.
2286 Don't do this if the insn at the branch target is a branch. */
2287 if (slots_to_fill
!= slots_filled
2289 && jump_to_label_p (trial
)
2290 && simplejump_p (trial
)
2291 && (next_trial
= next_active_insn (JUMP_LABEL (trial
))) != 0
2292 && ! (NONJUMP_INSN_P (next_trial
)
2293 && GET_CODE (PATTERN (next_trial
)) == SEQUENCE
)
2294 && !JUMP_P (next_trial
)
2295 && ! insn_references_resource_p (next_trial
, &set
, true)
2296 && ! insn_sets_resource_p (next_trial
, &set
, true)
2297 && ! insn_sets_resource_p (next_trial
, &needed
, true)
2299 && ! reg_mentioned_p (cc0_rtx
, PATTERN (next_trial
))
2301 && ! (maybe_never
&& may_trap_or_fault_p (PATTERN (next_trial
)))
2302 && (next_trial
= try_split (PATTERN (next_trial
), next_trial
, 0))
2303 && eligible_for_delay (insn
, slots_filled
, next_trial
, flags
)
2304 && ! can_throw_internal (trial
))
2306 /* See comment in relax_delay_slots about necessity of using
2307 next_real_insn here. */
2308 rtx_insn
*new_label
= next_real_insn (next_trial
);
2311 new_label
= get_label_before (new_label
, JUMP_LABEL (trial
));
2313 new_label
= find_end_label (simple_return_rtx
);
2318 = add_to_delay_list (copy_delay_slot_insn (next_trial
),
2321 reorg_redirect_jump (trial
, new_label
);
2326 /* If this is an unconditional jump, then try to get insns from the
2327 target of the jump. */
2329 && simplejump_p (insn
)
2330 && slots_filled
!= slots_to_fill
)
2332 = fill_slots_from_thread (insn
, const_true_rtx
,
2333 next_active_insn (JUMP_LABEL (insn
)),
2335 own_thread_p (JUMP_LABEL (insn
),
2336 JUMP_LABEL (insn
), 0),
2337 slots_to_fill
, &slots_filled
,
2341 unfilled_slots_base
[i
]
2342 = emit_delay_sequence (insn
, delay_list
, slots_filled
);
2344 if (slots_to_fill
== slots_filled
)
2345 unfilled_slots_base
[i
] = 0;
2347 note_delay_statistics (slots_filled
, 0);
2351 /* Follow any unconditional jump at LABEL, for the purpose of redirecting JUMP;
2352 return the ultimate label reached by any such chain of jumps.
2353 Return a suitable return rtx if the chain ultimately leads to a
2355 If LABEL is not followed by a jump, return LABEL.
2356 If the chain loops or we can't find end, return LABEL,
2357 since that tells caller to avoid changing the insn.
2358 If the returned label is obtained by following a crossing jump,
2359 set *CROSSING to true, otherwise set it to false. */
2362 follow_jumps (rtx label
, rtx_insn
*jump
, bool *crossing
)
2369 if (ANY_RETURN_P (label
))
2372 rtx_insn
*value
= as_a
<rtx_insn
*> (label
);
2376 && (insn
= next_active_insn (value
)) != 0
2378 && JUMP_LABEL (insn
) != NULL_RTX
2379 && ((any_uncondjump_p (insn
) && onlyjump_p (insn
))
2380 || ANY_RETURN_P (PATTERN (insn
)))
2381 && (next
= NEXT_INSN (insn
))
2382 && BARRIER_P (next
));
2385 rtx this_label_or_return
= JUMP_LABEL (insn
);
2387 /* If we have found a cycle, make the insn jump to itself. */
2388 if (this_label_or_return
== label
)
2391 /* Cannot follow returns and cannot look through tablejumps. */
2392 if (ANY_RETURN_P (this_label_or_return
))
2393 return this_label_or_return
;
2395 rtx_insn
*this_label
= as_a
<rtx_insn
*> (this_label_or_return
);
2396 if (NEXT_INSN (this_label
)
2397 && JUMP_TABLE_DATA_P (NEXT_INSN (this_label
)))
2400 if (!targetm
.can_follow_jump (jump
, insn
))
2403 *crossing
= CROSSING_JUMP_P (jump
);
2411 /* Try to find insns to place in delay slots.
2413 INSN is the jump needing SLOTS_TO_FILL delay slots. It tests CONDITION
2414 or is an unconditional branch if CONDITION is const_true_rtx.
2415 *PSLOTS_FILLED is updated with the number of slots that we have filled.
2417 THREAD is a flow-of-control, either the insns to be executed if the
2418 branch is true or if the branch is false, THREAD_IF_TRUE says which.
2420 OPPOSITE_THREAD is the thread in the opposite direction. It is used
2421 to see if any potential delay slot insns set things needed there.
2423 LIKELY is nonzero if it is extremely likely that the branch will be
2424 taken and THREAD_IF_TRUE is set. This is used for the branch at the
2425 end of a loop back up to the top.
2427 OWN_THREAD and OWN_OPPOSITE_THREAD are true if we are the only user of the
2428 thread. I.e., it is the fallthrough code of our jump or the target of the
2429 jump when we are the only jump going there.
2431 If OWN_THREAD is false, it must be the "true" thread of a jump. In that
2432 case, we can only take insns from the head of the thread for our delay
2433 slot. We then adjust the jump to point after the insns we have taken. */
2435 static rtx_insn_list
*
2436 fill_slots_from_thread (rtx_insn
*insn
, rtx condition
, rtx thread_or_return
,
2437 rtx opposite_thread
, int likely
,
2439 int own_thread
, int slots_to_fill
,
2440 int *pslots_filled
, rtx_insn_list
*delay_list
)
2443 struct resources opposite_needed
, set
, needed
;
2449 /* Validate our arguments. */
2450 gcc_assert (condition
!= const_true_rtx
|| thread_if_true
);
2451 gcc_assert (own_thread
|| thread_if_true
);
2453 flags
= get_jump_flags (insn
, JUMP_LABEL (insn
));
2455 /* If our thread is the end of subroutine, we can't get any delay
2457 if (thread_or_return
== NULL_RTX
|| ANY_RETURN_P (thread_or_return
))
2460 rtx_insn
*thread
= as_a
<rtx_insn
*> (thread_or_return
);
2462 /* If this is an unconditional branch, nothing is needed at the
2463 opposite thread. Otherwise, compute what is needed there. */
2464 if (condition
== const_true_rtx
)
2465 CLEAR_RESOURCE (&opposite_needed
);
2467 mark_target_live_regs (get_insns (), opposite_thread
, &opposite_needed
);
2469 /* If the insn at THREAD can be split, do it here to avoid having to
2470 update THREAD and NEW_THREAD if it is done in the loop below. Also
2471 initialize NEW_THREAD. */
2473 new_thread
= thread
= try_split (PATTERN (thread
), thread
, 0);
2475 /* Scan insns at THREAD. We are looking for an insn that can be removed
2476 from THREAD (it neither sets nor references resources that were set
2477 ahead of it and it doesn't set anything needs by the insns ahead of
2478 it) and that either can be placed in an annulling insn or aren't
2479 needed at OPPOSITE_THREAD. */
2481 CLEAR_RESOURCE (&needed
);
2482 CLEAR_RESOURCE (&set
);
2484 /* If we do not own this thread, we must stop as soon as we find
2485 something that we can't put in a delay slot, since all we can do
2486 is branch into THREAD at a later point. Therefore, labels stop
2487 the search if this is not the `true' thread. */
2489 for (trial
= thread
;
2490 ! stop_search_p (trial
, ! thread_if_true
) && (! lose
|| own_thread
);
2491 trial
= next_nonnote_insn (trial
))
2495 /* If we have passed a label, we no longer own this thread. */
2496 if (LABEL_P (trial
))
2502 pat
= PATTERN (trial
);
2503 if (GET_CODE (pat
) == USE
|| GET_CODE (pat
) == CLOBBER
)
2506 /* If TRIAL conflicts with the insns ahead of it, we lose. Also,
2507 don't separate or copy insns that set and use CC0. */
2508 if (! insn_references_resource_p (trial
, &set
, true)
2509 && ! insn_sets_resource_p (trial
, &set
, true)
2510 && ! insn_sets_resource_p (trial
, &needed
, true)
2512 && ! (reg_mentioned_p (cc0_rtx
, pat
)
2513 && (! own_thread
|| ! sets_cc0_p (pat
)))
2515 && ! can_throw_internal (trial
))
2519 /* If TRIAL is redundant with some insn before INSN, we don't
2520 actually need to add it to the delay list; we can merely pretend
2522 if ((prior_insn
= redundant_insn (trial
, insn
, delay_list
)))
2524 fix_reg_dead_note (prior_insn
, insn
);
2527 update_block (trial
, thread
);
2528 if (trial
== thread
)
2530 thread
= next_active_insn (thread
);
2531 if (new_thread
== trial
)
2532 new_thread
= thread
;
2535 delete_related_insns (trial
);
2539 update_reg_unused_notes (prior_insn
, trial
);
2540 new_thread
= next_active_insn (trial
);
2546 /* There are two ways we can win: If TRIAL doesn't set anything
2547 needed at the opposite thread and can't trap, or if it can
2548 go into an annulled delay slot. But we want neither to copy
2549 nor to speculate frame-related insns. */
2551 && ((condition
== const_true_rtx
2552 && (own_thread
|| !RTX_FRAME_RELATED_P (trial
)))
2553 || (! insn_sets_resource_p (trial
, &opposite_needed
, true)
2554 && ! may_trap_or_fault_p (pat
)
2555 && ! RTX_FRAME_RELATED_P (trial
))))
2558 trial
= try_split (pat
, trial
, 0);
2559 if (new_thread
== old_trial
)
2561 if (thread
== old_trial
)
2563 pat
= PATTERN (trial
);
2564 if (eligible_for_delay (insn
, *pslots_filled
, trial
, flags
))
2568 #ifdef ANNUL_IFTRUE_SLOTS
2571 #ifdef ANNUL_IFFALSE_SLOTS
2577 trial
= try_split (pat
, trial
, 0);
2578 if (new_thread
== old_trial
)
2580 if (thread
== old_trial
)
2582 pat
= PATTERN (trial
);
2583 if ((must_annul
|| delay_list
== NULL
) && (thread_if_true
2584 ? check_annul_list_true_false (0, delay_list
)
2585 && eligible_for_annul_false (insn
, *pslots_filled
, trial
, flags
)
2586 : check_annul_list_true_false (1, delay_list
)
2587 && eligible_for_annul_true (insn
, *pslots_filled
, trial
, flags
)))
2595 if (reg_mentioned_p (cc0_rtx
, pat
))
2596 link_cc0_insns (trial
);
2599 /* If we own this thread, delete the insn. If this is the
2600 destination of a branch, show that a basic block status
2601 may have been updated. In any case, mark the new
2602 starting point of this thread. */
2607 update_block (trial
, thread
);
2608 if (trial
== thread
)
2610 thread
= next_active_insn (thread
);
2611 if (new_thread
== trial
)
2612 new_thread
= thread
;
2615 /* We are moving this insn, not deleting it. We must
2616 temporarily increment the use count on any referenced
2617 label lest it be deleted by delete_related_insns. */
2618 for (note
= REG_NOTES (trial
);
2620 note
= XEXP (note
, 1))
2621 if (REG_NOTE_KIND (note
) == REG_LABEL_OPERAND
2622 || REG_NOTE_KIND (note
) == REG_LABEL_TARGET
)
2624 /* REG_LABEL_OPERAND could be
2625 NOTE_INSN_DELETED_LABEL too. */
2626 if (LABEL_P (XEXP (note
, 0)))
2627 LABEL_NUSES (XEXP (note
, 0))++;
2629 gcc_assert (REG_NOTE_KIND (note
)
2630 == REG_LABEL_OPERAND
);
2632 if (jump_to_label_p (trial
))
2633 LABEL_NUSES (JUMP_LABEL (trial
))++;
2635 delete_related_insns (trial
);
2637 for (note
= REG_NOTES (trial
);
2639 note
= XEXP (note
, 1))
2640 if (REG_NOTE_KIND (note
) == REG_LABEL_OPERAND
2641 || REG_NOTE_KIND (note
) == REG_LABEL_TARGET
)
2643 /* REG_LABEL_OPERAND could be
2644 NOTE_INSN_DELETED_LABEL too. */
2645 if (LABEL_P (XEXP (note
, 0)))
2646 LABEL_NUSES (XEXP (note
, 0))--;
2648 gcc_assert (REG_NOTE_KIND (note
)
2649 == REG_LABEL_OPERAND
);
2651 if (jump_to_label_p (trial
))
2652 LABEL_NUSES (JUMP_LABEL (trial
))--;
2655 new_thread
= next_active_insn (trial
);
2657 temp
= own_thread
? trial
: copy_delay_slot_insn (trial
);
2659 INSN_FROM_TARGET_P (temp
) = 1;
2661 delay_list
= add_to_delay_list (temp
, delay_list
);
2663 if (slots_to_fill
== ++(*pslots_filled
))
2665 /* Even though we have filled all the slots, we
2666 may be branching to a location that has a
2667 redundant insn. Skip any if so. */
2668 while (new_thread
&& ! own_thread
2669 && ! insn_sets_resource_p (new_thread
, &set
, true)
2670 && ! insn_sets_resource_p (new_thread
, &needed
,
2672 && ! insn_references_resource_p (new_thread
,
2675 = redundant_insn (new_thread
, insn
,
2678 /* We know we do not own the thread, so no need
2679 to call update_block and delete_insn. */
2680 fix_reg_dead_note (prior_insn
, insn
);
2681 update_reg_unused_notes (prior_insn
, new_thread
);
2682 new_thread
= next_active_insn (new_thread
);
2692 /* This insn can't go into a delay slot. */
2694 mark_set_resources (trial
, &set
, 0, MARK_SRC_DEST_CALL
);
2695 mark_referenced_resources (trial
, &needed
, true);
2697 /* Ensure we don't put insns between the setting of cc and the comparison
2698 by moving a setting of cc into an earlier delay slot since these insns
2699 could clobber the condition code. */
2702 /* If this insn is a register-register copy and the next insn has
2703 a use of our destination, change it to use our source. That way,
2704 it will become a candidate for our delay slot the next time
2705 through this loop. This case occurs commonly in loops that
2708 We could check for more complex cases than those tested below,
2709 but it doesn't seem worth it. It might also be a good idea to try
2710 to swap the two insns. That might do better.
2712 We can't do this if the next insn modifies our destination, because
2713 that would make the replacement into the insn invalid. We also can't
2714 do this if it modifies our source, because it might be an earlyclobber
2715 operand. This latter test also prevents updating the contents of
2716 a PRE_INC. We also can't do this if there's overlap of source and
2717 destination. Overlap may happen for larger-than-register-size modes. */
2719 if (NONJUMP_INSN_P (trial
) && GET_CODE (pat
) == SET
2720 && REG_P (SET_SRC (pat
))
2721 && REG_P (SET_DEST (pat
))
2722 && !reg_overlap_mentioned_p (SET_DEST (pat
), SET_SRC (pat
)))
2724 rtx next
= next_nonnote_insn (trial
);
2726 if (next
&& NONJUMP_INSN_P (next
)
2727 && GET_CODE (PATTERN (next
)) != USE
2728 && ! reg_set_p (SET_DEST (pat
), next
)
2729 && ! reg_set_p (SET_SRC (pat
), next
)
2730 && reg_referenced_p (SET_DEST (pat
), PATTERN (next
))
2731 && ! modified_in_p (SET_DEST (pat
), next
))
2732 validate_replace_rtx (SET_DEST (pat
), SET_SRC (pat
), next
);
2736 /* If we stopped on a branch insn that has delay slots, see if we can
2737 steal some of the insns in those slots. */
2738 if (trial
&& NONJUMP_INSN_P (trial
)
2739 && GET_CODE (PATTERN (trial
)) == SEQUENCE
2740 && JUMP_P (XVECEXP (PATTERN (trial
), 0, 0)))
2742 rtx_sequence
*sequence
= as_a
<rtx_sequence
*> (PATTERN (trial
));
2743 /* If this is the `true' thread, we will want to follow the jump,
2744 so we can only do this if we have taken everything up to here. */
2745 if (thread_if_true
&& trial
== new_thread
)
2748 = steal_delay_list_from_target (insn
, condition
, sequence
,
2749 delay_list
, &set
, &needed
,
2750 &opposite_needed
, slots_to_fill
,
2751 pslots_filled
, &must_annul
,
2753 /* If we owned the thread and are told that it branched
2754 elsewhere, make sure we own the thread at the new location. */
2755 if (own_thread
&& trial
!= new_thread
)
2756 own_thread
= own_thread_p (new_thread
, new_thread
, 0);
2758 else if (! thread_if_true
)
2760 = steal_delay_list_from_fallthrough (insn
, condition
,
2762 delay_list
, &set
, &needed
,
2763 &opposite_needed
, slots_to_fill
,
2764 pslots_filled
, &must_annul
);
2767 /* If we haven't found anything for this delay slot and it is very
2768 likely that the branch will be taken, see if the insn at our target
2769 increments or decrements a register with an increment that does not
2770 depend on the destination register. If so, try to place the opposite
2771 arithmetic insn after the jump insn and put the arithmetic insn in the
2772 delay slot. If we can't do this, return. */
2773 if (delay_list
== 0 && likely
2774 && new_thread
&& !ANY_RETURN_P (new_thread
)
2775 && NONJUMP_INSN_P (new_thread
)
2776 && !RTX_FRAME_RELATED_P (new_thread
)
2777 && GET_CODE (PATTERN (new_thread
)) != ASM_INPUT
2778 && asm_noperands (PATTERN (new_thread
)) < 0)
2780 rtx pat
= PATTERN (new_thread
);
2784 /* We know "new_thread" is an insn due to NONJUMP_INSN_P (new_thread)
2786 trial
= as_a
<rtx_insn
*> (new_thread
);
2787 pat
= PATTERN (trial
);
2789 if (!NONJUMP_INSN_P (trial
)
2790 || GET_CODE (pat
) != SET
2791 || ! eligible_for_delay (insn
, 0, trial
, flags
)
2792 || can_throw_internal (trial
))
2795 dest
= SET_DEST (pat
), src
= SET_SRC (pat
);
2796 if ((GET_CODE (src
) == PLUS
|| GET_CODE (src
) == MINUS
)
2797 && rtx_equal_p (XEXP (src
, 0), dest
)
2798 && (!FLOAT_MODE_P (GET_MODE (src
))
2799 || flag_unsafe_math_optimizations
)
2800 && ! reg_overlap_mentioned_p (dest
, XEXP (src
, 1))
2801 && ! side_effects_p (pat
))
2803 rtx other
= XEXP (src
, 1);
2807 /* If this is a constant adjustment, use the same code with
2808 the negated constant. Otherwise, reverse the sense of the
2810 if (CONST_INT_P (other
))
2811 new_arith
= gen_rtx_fmt_ee (GET_CODE (src
), GET_MODE (src
), dest
,
2812 negate_rtx (GET_MODE (src
), other
));
2814 new_arith
= gen_rtx_fmt_ee (GET_CODE (src
) == PLUS
? MINUS
: PLUS
,
2815 GET_MODE (src
), dest
, other
);
2817 ninsn
= emit_insn_after (gen_rtx_SET (VOIDmode
, dest
, new_arith
),
2820 if (recog_memoized (ninsn
) < 0
2821 || (extract_insn (ninsn
),
2822 !constrain_operands (1, get_preferred_alternatives (ninsn
))))
2824 delete_related_insns (ninsn
);
2830 update_block (trial
, thread
);
2831 if (trial
== thread
)
2833 thread
= next_active_insn (thread
);
2834 if (new_thread
== trial
)
2835 new_thread
= thread
;
2837 delete_related_insns (trial
);
2840 new_thread
= next_active_insn (trial
);
2842 ninsn
= own_thread
? trial
: copy_delay_slot_insn (trial
);
2844 INSN_FROM_TARGET_P (ninsn
) = 1;
2846 delay_list
= add_to_delay_list (ninsn
, NULL
);
2851 if (delay_list
&& must_annul
)
2852 INSN_ANNULLED_BRANCH_P (insn
) = 1;
2854 /* If we are to branch into the middle of this thread, find an appropriate
2855 label or make a new one if none, and redirect INSN to it. If we hit the
2856 end of the function, use the end-of-function label. */
2857 if (new_thread
!= thread
)
2860 bool crossing
= false;
2862 gcc_assert (thread_if_true
);
2864 if (new_thread
&& simplejump_or_return_p (new_thread
)
2865 && redirect_with_delay_list_safe_p (insn
,
2866 JUMP_LABEL (new_thread
),
2868 new_thread
= follow_jumps (JUMP_LABEL (new_thread
), insn
,
2871 if (ANY_RETURN_P (new_thread
))
2872 label
= find_end_label (new_thread
);
2873 else if (LABEL_P (new_thread
))
2876 label
= get_label_before (as_a
<rtx_insn
*> (new_thread
),
2881 reorg_redirect_jump (insn
, label
);
2883 CROSSING_JUMP_P (insn
) = 1;
2890 /* Make another attempt to find insns to place in delay slots.
2892 We previously looked for insns located in front of the delay insn
2893 and, for non-jump delay insns, located behind the delay insn.
2895 Here only try to schedule jump insns and try to move insns from either
2896 the target or the following insns into the delay slot. If annulling is
2897 supported, we will be likely to do this. Otherwise, we can do this only
2901 fill_eager_delay_slots (void)
2905 int num_unfilled_slots
= unfilled_slots_next
- unfilled_slots_base
;
2907 for (i
= 0; i
< num_unfilled_slots
; i
++)
2910 rtx target_label
, insn_at_target
;
2911 rtx_insn
*fallthrough_insn
;
2912 rtx_insn_list
*delay_list
= 0;
2914 int own_fallthrough
;
2915 int prediction
, slots_to_fill
, slots_filled
;
2917 insn
= unfilled_slots_base
[i
];
2921 || ! (condjump_p (insn
) || condjump_in_parallel_p (insn
)))
2924 slots_to_fill
= num_delay_slots (insn
);
2925 /* Some machine description have defined instructions to have
2926 delay slots only in certain circumstances which may depend on
2927 nearby insns (which change due to reorg's actions).
2929 For example, the PA port normally has delay slots for unconditional
2932 However, the PA port claims such jumps do not have a delay slot
2933 if they are immediate successors of certain CALL_INSNs. This
2934 allows the port to favor filling the delay slot of the call with
2935 the unconditional jump. */
2936 if (slots_to_fill
== 0)
2940 target_label
= JUMP_LABEL (insn
);
2941 condition
= get_branch_condition (insn
, target_label
);
2946 /* Get the next active fallthrough and target insns and see if we own
2947 them. Then see whether the branch is likely true. We don't need
2948 to do a lot of this for unconditional branches. */
2950 insn_at_target
= first_active_target_insn (target_label
);
2951 own_target
= own_thread_p (target_label
, target_label
, 0);
2953 if (condition
== const_true_rtx
)
2955 own_fallthrough
= 0;
2956 fallthrough_insn
= 0;
2961 fallthrough_insn
= next_active_insn (insn
);
2962 own_fallthrough
= own_thread_p (NEXT_INSN (insn
), NULL_RTX
, 1);
2963 prediction
= mostly_true_jump (insn
);
2966 /* If this insn is expected to branch, first try to get insns from our
2967 target, then our fallthrough insns. If it is not expected to branch,
2968 try the other order. */
2973 = fill_slots_from_thread (insn
, condition
, insn_at_target
,
2974 fallthrough_insn
, prediction
== 2, 1,
2976 slots_to_fill
, &slots_filled
, delay_list
);
2978 if (delay_list
== 0 && own_fallthrough
)
2980 /* Even though we didn't find anything for delay slots,
2981 we might have found a redundant insn which we deleted
2982 from the thread that was filled. So we have to recompute
2983 the next insn at the target. */
2984 target_label
= JUMP_LABEL (insn
);
2985 insn_at_target
= first_active_target_insn (target_label
);
2988 = fill_slots_from_thread (insn
, condition
, fallthrough_insn
,
2989 insn_at_target
, 0, 0,
2991 slots_to_fill
, &slots_filled
,
2997 if (own_fallthrough
)
2999 = fill_slots_from_thread (insn
, condition
, fallthrough_insn
,
3000 insn_at_target
, 0, 0,
3002 slots_to_fill
, &slots_filled
,
3005 if (delay_list
== 0)
3007 = fill_slots_from_thread (insn
, condition
, insn_at_target
,
3008 next_active_insn (insn
), 0, 1,
3010 slots_to_fill
, &slots_filled
,
3015 unfilled_slots_base
[i
]
3016 = emit_delay_sequence (insn
, delay_list
, slots_filled
);
3018 if (slots_to_fill
== slots_filled
)
3019 unfilled_slots_base
[i
] = 0;
3021 note_delay_statistics (slots_filled
, 1);
3025 static void delete_computation (rtx insn
);
3027 /* Recursively delete prior insns that compute the value (used only by INSN
3028 which the caller is deleting) stored in the register mentioned by NOTE
3029 which is a REG_DEAD note associated with INSN. */
3032 delete_prior_computation (rtx note
, rtx insn
)
3035 rtx reg
= XEXP (note
, 0);
3037 for (our_prev
= prev_nonnote_insn (insn
);
3038 our_prev
&& (NONJUMP_INSN_P (our_prev
)
3039 || CALL_P (our_prev
));
3040 our_prev
= prev_nonnote_insn (our_prev
))
3042 rtx pat
= PATTERN (our_prev
);
3044 /* If we reach a CALL which is not calling a const function
3045 or the callee pops the arguments, then give up. */
3046 if (CALL_P (our_prev
)
3047 && (! RTL_CONST_CALL_P (our_prev
)
3048 || GET_CODE (pat
) != SET
|| GET_CODE (SET_SRC (pat
)) != CALL
))
3051 /* If we reach a SEQUENCE, it is too complex to try to
3052 do anything with it, so give up. We can be run during
3053 and after reorg, so SEQUENCE rtl can legitimately show
3055 if (GET_CODE (pat
) == SEQUENCE
)
3058 if (GET_CODE (pat
) == USE
3059 && NONJUMP_INSN_P (XEXP (pat
, 0)))
3060 /* reorg creates USEs that look like this. We leave them
3061 alone because reorg needs them for its own purposes. */
3064 if (reg_set_p (reg
, pat
))
3066 if (side_effects_p (pat
) && !CALL_P (our_prev
))
3069 if (GET_CODE (pat
) == PARALLEL
)
3071 /* If we find a SET of something else, we can't
3076 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
3078 rtx part
= XVECEXP (pat
, 0, i
);
3080 if (GET_CODE (part
) == SET
3081 && SET_DEST (part
) != reg
)
3085 if (i
== XVECLEN (pat
, 0))
3086 delete_computation (our_prev
);
3088 else if (GET_CODE (pat
) == SET
3089 && REG_P (SET_DEST (pat
)))
3091 int dest_regno
= REGNO (SET_DEST (pat
));
3092 int dest_endregno
= END_REGNO (SET_DEST (pat
));
3093 int regno
= REGNO (reg
);
3094 int endregno
= END_REGNO (reg
);
3096 if (dest_regno
>= regno
3097 && dest_endregno
<= endregno
)
3098 delete_computation (our_prev
);
3100 /* We may have a multi-word hard register and some, but not
3101 all, of the words of the register are needed in subsequent
3102 insns. Write REG_UNUSED notes for those parts that were not
3104 else if (dest_regno
<= regno
3105 && dest_endregno
>= endregno
)
3109 add_reg_note (our_prev
, REG_UNUSED
, reg
);
3111 for (i
= dest_regno
; i
< dest_endregno
; i
++)
3112 if (! find_regno_note (our_prev
, REG_UNUSED
, i
))
3115 if (i
== dest_endregno
)
3116 delete_computation (our_prev
);
3123 /* If PAT references the register that dies here, it is an
3124 additional use. Hence any prior SET isn't dead. However, this
3125 insn becomes the new place for the REG_DEAD note. */
3126 if (reg_overlap_mentioned_p (reg
, pat
))
3128 XEXP (note
, 1) = REG_NOTES (our_prev
);
3129 REG_NOTES (our_prev
) = note
;
3135 /* Delete INSN and recursively delete insns that compute values used only
3136 by INSN. This uses the REG_DEAD notes computed during flow analysis.
3138 Look at all our REG_DEAD notes. If a previous insn does nothing other
3139 than set a register that dies in this insn, we can delete that insn
3142 On machines with CC0, if CC0 is used in this insn, we may be able to
3143 delete the insn that set it. */
3146 delete_computation (rtx insn
)
3151 if (reg_referenced_p (cc0_rtx
, PATTERN (insn
)))
3153 rtx prev
= prev_nonnote_insn (insn
);
3154 /* We assume that at this stage
3155 CC's are always set explicitly
3156 and always immediately before the jump that
3157 will use them. So if the previous insn
3158 exists to set the CC's, delete it
3159 (unless it performs auto-increments, etc.). */
3160 if (prev
&& NONJUMP_INSN_P (prev
)
3161 && sets_cc0_p (PATTERN (prev
)))
3163 if (sets_cc0_p (PATTERN (prev
)) > 0
3164 && ! side_effects_p (PATTERN (prev
)))
3165 delete_computation (prev
);
3167 /* Otherwise, show that cc0 won't be used. */
3168 add_reg_note (prev
, REG_UNUSED
, cc0_rtx
);
3173 for (note
= REG_NOTES (insn
); note
; note
= next
)
3175 next
= XEXP (note
, 1);
3177 if (REG_NOTE_KIND (note
) != REG_DEAD
3178 /* Verify that the REG_NOTE is legitimate. */
3179 || !REG_P (XEXP (note
, 0)))
3182 delete_prior_computation (note
, insn
);
3185 delete_related_insns (insn
);
3188 /* If all INSN does is set the pc, delete it,
3189 and delete the insn that set the condition codes for it
3190 if that's what the previous thing was. */
3193 delete_jump (rtx_insn
*insn
)
3195 rtx set
= single_set (insn
);
3197 if (set
&& GET_CODE (SET_DEST (set
)) == PC
)
3198 delete_computation (insn
);
3202 label_before_next_insn (rtx x
, rtx scan_limit
)
3204 rtx_insn
*insn
= next_active_insn (x
);
3207 insn
= PREV_INSN (insn
);
3208 if (insn
== scan_limit
|| insn
== NULL_RTX
)
3217 /* Once we have tried two ways to fill a delay slot, make a pass over the
3218 code to try to improve the results and to do such things as more jump
3222 relax_delay_slots (rtx_insn
*first
)
3224 rtx_insn
*insn
, *next
;
3227 rtx_insn
*delay_insn
;
3230 /* Look at every JUMP_INSN and see if we can improve it. */
3231 for (insn
= first
; insn
; insn
= next
)
3236 next
= next_active_insn (insn
);
3238 /* If this is a jump insn, see if it now jumps to a jump, jumps to
3239 the next insn, or jumps to a label that is not the last of a
3240 group of consecutive labels. */
3242 && (condjump_p (insn
) || condjump_in_parallel_p (insn
))
3243 && !ANY_RETURN_P (target_label
= JUMP_LABEL (insn
)))
3246 = skip_consecutive_labels (follow_jumps (target_label
, insn
,
3248 if (ANY_RETURN_P (target_label
))
3249 target_label
= find_end_label (target_label
);
3251 if (target_label
&& next_active_insn (target_label
) == next
3252 && ! condjump_in_parallel_p (insn
))
3258 if (target_label
&& target_label
!= JUMP_LABEL (insn
))
3260 reorg_redirect_jump (insn
, target_label
);
3262 CROSSING_JUMP_P (insn
) = 1;
3265 /* See if this jump conditionally branches around an unconditional
3266 jump. If so, invert this jump and point it to the target of the
3268 if (next
&& simplejump_or_return_p (next
)
3269 && any_condjump_p (insn
)
3271 && next_active_insn (target_label
) == next_active_insn (next
)
3272 && no_labels_between_p (insn
, next
))
3274 rtx label
= JUMP_LABEL (next
);
3276 /* Be careful how we do this to avoid deleting code or
3277 labels that are momentarily dead. See similar optimization
3280 We also need to ensure we properly handle the case when
3281 invert_jump fails. */
3283 ++LABEL_NUSES (target_label
);
3284 if (!ANY_RETURN_P (label
))
3285 ++LABEL_NUSES (label
);
3287 if (invert_jump (insn
, label
, 1))
3289 delete_related_insns (next
);
3293 if (!ANY_RETURN_P (label
))
3294 --LABEL_NUSES (label
);
3296 if (--LABEL_NUSES (target_label
) == 0)
3297 delete_related_insns (target_label
);
3303 /* If this is an unconditional jump and the previous insn is a
3304 conditional jump, try reversing the condition of the previous
3305 insn and swapping our targets. The next pass might be able to
3308 Don't do this if we expect the conditional branch to be true, because
3309 we would then be making the more common case longer. */
3311 if (simplejump_or_return_p (insn
)
3312 && (other
= prev_active_insn (insn
)) != 0
3313 && any_condjump_p (other
)
3314 && no_labels_between_p (other
, insn
)
3315 && 0 > mostly_true_jump (other
))
3317 rtx other_target
= JUMP_LABEL (other
);
3318 target_label
= JUMP_LABEL (insn
);
3320 if (invert_jump (other
, target_label
, 0))
3321 reorg_redirect_jump (insn
, other_target
);
3324 /* Now look only at cases where we have a filled delay slot. */
3325 if (!NONJUMP_INSN_P (insn
) || GET_CODE (PATTERN (insn
)) != SEQUENCE
)
3328 pat
= as_a
<rtx_sequence
*> (PATTERN (insn
));
3329 delay_insn
= pat
->insn (0);
3331 /* See if the first insn in the delay slot is redundant with some
3332 previous insn. Remove it from the delay slot if so; then set up
3333 to reprocess this insn. */
3334 if (redundant_insn (pat
->insn (1), delay_insn
, 0))
3336 update_block (pat
->insn (1), insn
);
3337 delete_from_delay_slot (pat
->insn (1));
3338 next
= prev_active_insn (next
);
3342 /* See if we have a RETURN insn with a filled delay slot followed
3343 by a RETURN insn with an unfilled a delay slot. If so, we can delete
3344 the first RETURN (but not its delay insn). This gives the same
3345 effect in fewer instructions.
3347 Only do so if optimizing for size since this results in slower, but
3349 if (optimize_function_for_size_p (cfun
)
3350 && ANY_RETURN_P (PATTERN (delay_insn
))
3353 && PATTERN (next
) == PATTERN (delay_insn
))
3358 /* Delete the RETURN and just execute the delay list insns.
3360 We do this by deleting the INSN containing the SEQUENCE, then
3361 re-emitting the insns separately, and then deleting the RETURN.
3362 This allows the count of the jump target to be properly
3365 Note that we need to change the INSN_UID of the re-emitted insns
3366 since it is used to hash the insns for mark_target_live_regs and
3367 the re-emitted insns will no longer be wrapped up in a SEQUENCE.
3369 Clear the from target bit, since these insns are no longer
3371 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
3372 INSN_FROM_TARGET_P (XVECEXP (pat
, 0, i
)) = 0;
3374 trial
= PREV_INSN (insn
);
3375 delete_related_insns (insn
);
3376 gcc_assert (GET_CODE (pat
) == SEQUENCE
);
3377 add_insn_after (delay_insn
, trial
, NULL
);
3379 for (i
= 1; i
< pat
->len (); i
++)
3380 after
= emit_copy_of_insn_after (pat
->insn (i
), after
);
3381 delete_scheduled_jump (delay_insn
);
3385 /* Now look only at the cases where we have a filled JUMP_INSN. */
3386 if (!JUMP_P (delay_insn
)
3387 || !(condjump_p (delay_insn
) || condjump_in_parallel_p (delay_insn
)))
3390 target_label
= JUMP_LABEL (delay_insn
);
3391 if (target_label
&& ANY_RETURN_P (target_label
))
3394 /* If this jump goes to another unconditional jump, thread it, but
3395 don't convert a jump into a RETURN here. */
3396 trial
= skip_consecutive_labels (follow_jumps (target_label
, delay_insn
,
3398 if (ANY_RETURN_P (trial
))
3399 trial
= find_end_label (trial
);
3401 if (trial
&& trial
!= target_label
3402 && redirect_with_delay_slots_safe_p (delay_insn
, trial
, insn
))
3404 reorg_redirect_jump (delay_insn
, trial
);
3405 target_label
= trial
;
3407 CROSSING_JUMP_P (insn
) = 1;
3410 /* If the first insn at TARGET_LABEL is redundant with a previous
3411 insn, redirect the jump to the following insn and process again.
3412 We use next_real_insn instead of next_active_insn so we
3413 don't skip USE-markers, or we'll end up with incorrect
3415 trial
= next_real_insn (target_label
);
3416 if (trial
&& GET_CODE (PATTERN (trial
)) != SEQUENCE
3417 && redundant_insn (trial
, insn
, 0)
3418 && ! can_throw_internal (trial
))
3420 /* Figure out where to emit the special USE insn so we don't
3421 later incorrectly compute register live/death info. */
3422 rtx_insn
*tmp
= next_active_insn (trial
);
3424 tmp
= find_end_label (simple_return_rtx
);
3428 /* Insert the special USE insn and update dataflow info.
3429 We know "trial" is an insn here as it is the output of
3430 next_real_insn () above. */
3431 update_block (as_a
<rtx_insn
*> (trial
), tmp
);
3433 /* Now emit a label before the special USE insn, and
3434 redirect our jump to the new label. */
3435 target_label
= get_label_before (PREV_INSN (tmp
), target_label
);
3436 reorg_redirect_jump (delay_insn
, target_label
);
3442 /* Similarly, if it is an unconditional jump with one insn in its
3443 delay list and that insn is redundant, thread the jump. */
3444 rtx_sequence
*trial_seq
=
3445 trial
? dyn_cast
<rtx_sequence
*> (PATTERN (trial
)) : NULL
;
3447 && trial_seq
->len () == 2
3448 && JUMP_P (trial_seq
->insn (0))
3449 && simplejump_or_return_p (trial_seq
->insn (0))
3450 && redundant_insn (trial_seq
->insn (1), insn
, 0))
3452 target_label
= JUMP_LABEL (trial_seq
->insn (0));
3453 if (ANY_RETURN_P (target_label
))
3454 target_label
= find_end_label (target_label
);
3457 && redirect_with_delay_slots_safe_p (delay_insn
, target_label
,
3460 update_block (trial_seq
->insn (1), insn
);
3461 reorg_redirect_jump (delay_insn
, target_label
);
3467 /* See if we have a simple (conditional) jump that is useless. */
3468 if (! INSN_ANNULLED_BRANCH_P (delay_insn
)
3469 && ! condjump_in_parallel_p (delay_insn
)
3470 && prev_active_insn (target_label
) == insn
3471 && ! BARRIER_P (prev_nonnote_insn (target_label
))
3473 /* If the last insn in the delay slot sets CC0 for some insn,
3474 various code assumes that it is in a delay slot. We could
3475 put it back where it belonged and delete the register notes,
3476 but it doesn't seem worthwhile in this uncommon case. */
3477 && ! find_reg_note (XVECEXP (pat
, 0, XVECLEN (pat
, 0) - 1),
3478 REG_CC_USER
, NULL_RTX
)
3485 /* All this insn does is execute its delay list and jump to the
3486 following insn. So delete the jump and just execute the delay
3489 We do this by deleting the INSN containing the SEQUENCE, then
3490 re-emitting the insns separately, and then deleting the jump.
3491 This allows the count of the jump target to be properly
3494 Note that we need to change the INSN_UID of the re-emitted insns
3495 since it is used to hash the insns for mark_target_live_regs and
3496 the re-emitted insns will no longer be wrapped up in a SEQUENCE.
3498 Clear the from target bit, since these insns are no longer
3500 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
3501 INSN_FROM_TARGET_P (XVECEXP (pat
, 0, i
)) = 0;
3503 trial
= PREV_INSN (insn
);
3504 delete_related_insns (insn
);
3505 gcc_assert (GET_CODE (pat
) == SEQUENCE
);
3506 add_insn_after (delay_insn
, trial
, NULL
);
3508 for (i
= 1; i
< pat
->len (); i
++)
3509 after
= emit_copy_of_insn_after (pat
->insn (i
), after
);
3510 delete_scheduled_jump (delay_insn
);
3514 /* See if this is an unconditional jump around a single insn which is
3515 identical to the one in its delay slot. In this case, we can just
3516 delete the branch and the insn in its delay slot. */
3517 if (next
&& NONJUMP_INSN_P (next
)
3518 && label_before_next_insn (next
, insn
) == target_label
3519 && simplejump_p (insn
)
3520 && XVECLEN (pat
, 0) == 2
3521 && rtx_equal_p (PATTERN (next
), PATTERN (pat
->insn (1))))
3523 delete_related_insns (insn
);
3527 /* See if this jump (with its delay slots) conditionally branches
3528 around an unconditional jump (without delay slots). If so, invert
3529 this jump and point it to the target of the second jump. We cannot
3530 do this for annulled jumps, though. Again, don't convert a jump to
3532 if (! INSN_ANNULLED_BRANCH_P (delay_insn
)
3533 && any_condjump_p (delay_insn
)
3534 && next
&& simplejump_or_return_p (next
)
3535 && next_active_insn (target_label
) == next_active_insn (next
)
3536 && no_labels_between_p (insn
, next
))
3538 rtx label
= JUMP_LABEL (next
);
3539 rtx old_label
= JUMP_LABEL (delay_insn
);
3541 if (ANY_RETURN_P (label
))
3542 label
= find_end_label (label
);
3544 /* find_end_label can generate a new label. Check this first. */
3546 && no_labels_between_p (insn
, next
)
3547 && redirect_with_delay_slots_safe_p (delay_insn
, label
, insn
))
3549 /* Be careful how we do this to avoid deleting code or labels
3550 that are momentarily dead. See similar optimization in
3553 ++LABEL_NUSES (old_label
);
3555 if (invert_jump (delay_insn
, label
, 1))
3559 /* Must update the INSN_FROM_TARGET_P bits now that
3560 the branch is reversed, so that mark_target_live_regs
3561 will handle the delay slot insn correctly. */
3562 for (i
= 1; i
< XVECLEN (PATTERN (insn
), 0); i
++)
3564 rtx slot
= XVECEXP (PATTERN (insn
), 0, i
);
3565 INSN_FROM_TARGET_P (slot
) = ! INSN_FROM_TARGET_P (slot
);
3568 delete_related_insns (next
);
3572 if (old_label
&& --LABEL_NUSES (old_label
) == 0)
3573 delete_related_insns (old_label
);
3578 /* If we own the thread opposite the way this insn branches, see if we
3579 can merge its delay slots with following insns. */
3580 if (INSN_FROM_TARGET_P (pat
->insn (1))
3581 && own_thread_p (NEXT_INSN (insn
), 0, 1))
3582 try_merge_delay_insns (insn
, next
);
3583 else if (! INSN_FROM_TARGET_P (pat
->insn (1))
3584 && own_thread_p (target_label
, target_label
, 0))
3585 try_merge_delay_insns (insn
, next_active_insn (target_label
));
3587 /* If we get here, we haven't deleted INSN. But we may have deleted
3588 NEXT, so recompute it. */
3589 next
= next_active_insn (insn
);
3594 /* Look for filled jumps to the end of function label. We can try to convert
3595 them into RETURN insns if the insns in the delay slot are valid for the
3599 make_return_insns (rtx_insn
*first
)
3602 rtx_insn
*jump_insn
;
3603 rtx real_return_label
= function_return_label
;
3604 rtx real_simple_return_label
= function_simple_return_label
;
3607 /* See if there is a RETURN insn in the function other than the one we
3608 made for END_OF_FUNCTION_LABEL. If so, set up anything we can't change
3609 into a RETURN to jump to it. */
3610 for (insn
= first
; insn
; insn
= NEXT_INSN (insn
))
3611 if (JUMP_P (insn
) && ANY_RETURN_P (PATTERN (insn
)))
3613 rtx t
= get_label_before (insn
, NULL_RTX
);
3614 if (PATTERN (insn
) == ret_rtx
)
3615 real_return_label
= t
;
3617 real_simple_return_label
= t
;
3621 /* Show an extra usage of REAL_RETURN_LABEL so it won't go away if it
3622 was equal to END_OF_FUNCTION_LABEL. */
3623 if (real_return_label
)
3624 LABEL_NUSES (real_return_label
)++;
3625 if (real_simple_return_label
)
3626 LABEL_NUSES (real_simple_return_label
)++;
3628 /* Clear the list of insns to fill so we can use it. */
3629 obstack_free (&unfilled_slots_obstack
, unfilled_firstobj
);
3631 for (insn
= first
; insn
; insn
= NEXT_INSN (insn
))
3634 rtx kind
, real_label
;
3636 /* Only look at filled JUMP_INSNs that go to the end of function
3638 if (!NONJUMP_INSN_P (insn
))
3641 if (GET_CODE (PATTERN (insn
)) != SEQUENCE
)
3644 rtx_sequence
*pat
= as_a
<rtx_sequence
*> (PATTERN (insn
));
3646 if (!jump_to_label_p (pat
->insn (0)))
3649 if (JUMP_LABEL (pat
->insn (0)) == function_return_label
)
3652 real_label
= real_return_label
;
3654 else if (JUMP_LABEL (pat
->insn (0)) == function_simple_return_label
)
3656 kind
= simple_return_rtx
;
3657 real_label
= real_simple_return_label
;
3662 jump_insn
= pat
->insn (0);
3664 /* If we can't make the jump into a RETURN, try to redirect it to the best
3665 RETURN and go on to the next insn. */
3666 if (!reorg_redirect_jump (jump_insn
, kind
))
3668 /* Make sure redirecting the jump will not invalidate the delay
3670 if (redirect_with_delay_slots_safe_p (jump_insn
, real_label
, insn
))
3671 reorg_redirect_jump (jump_insn
, real_label
);
3675 /* See if this RETURN can accept the insns current in its delay slot.
3676 It can if it has more or an equal number of slots and the contents
3677 of each is valid. */
3679 flags
= get_jump_flags (jump_insn
, JUMP_LABEL (jump_insn
));
3680 slots
= num_delay_slots (jump_insn
);
3681 if (slots
>= XVECLEN (pat
, 0) - 1)
3683 for (i
= 1; i
< XVECLEN (pat
, 0); i
++)
3685 #ifdef ANNUL_IFFALSE_SLOTS
3686 (INSN_ANNULLED_BRANCH_P (jump_insn
)
3687 && INSN_FROM_TARGET_P (pat
->insn (i
)))
3688 ? eligible_for_annul_false (jump_insn
, i
- 1,
3689 pat
->insn (i
), flags
) :
3691 #ifdef ANNUL_IFTRUE_SLOTS
3692 (INSN_ANNULLED_BRANCH_P (jump_insn
)
3693 && ! INSN_FROM_TARGET_P (pat
->insn (i
)))
3694 ? eligible_for_annul_true (jump_insn
, i
- 1,
3695 pat
->insn (i
), flags
) :
3697 eligible_for_delay (jump_insn
, i
- 1,
3698 pat
->insn (i
), flags
)))
3704 if (i
== XVECLEN (pat
, 0))
3707 /* We have to do something with this insn. If it is an unconditional
3708 RETURN, delete the SEQUENCE and output the individual insns,
3709 followed by the RETURN. Then set things up so we try to find
3710 insns for its delay slots, if it needs some. */
3711 if (ANY_RETURN_P (PATTERN (jump_insn
)))
3713 rtx_insn
*prev
= PREV_INSN (insn
);
3715 delete_related_insns (insn
);
3716 for (i
= 1; i
< XVECLEN (pat
, 0); i
++)
3717 prev
= emit_insn_after (PATTERN (XVECEXP (pat
, 0, i
)), prev
);
3719 insn
= emit_jump_insn_after (PATTERN (jump_insn
), prev
);
3720 emit_barrier_after (insn
);
3723 obstack_ptr_grow (&unfilled_slots_obstack
, insn
);
3726 /* It is probably more efficient to keep this with its current
3727 delay slot as a branch to a RETURN. */
3728 reorg_redirect_jump (jump_insn
, real_label
);
3731 /* Now delete REAL_RETURN_LABEL if we never used it. Then try to fill any
3732 new delay slots we have created. */
3733 if (real_return_label
!= NULL_RTX
&& --LABEL_NUSES (real_return_label
) == 0)
3734 delete_related_insns (real_return_label
);
3735 if (real_simple_return_label
!= NULL_RTX
3736 && --LABEL_NUSES (real_simple_return_label
) == 0)
3737 delete_related_insns (real_simple_return_label
);
3739 fill_simple_delay_slots (1);
3740 fill_simple_delay_slots (0);
3743 /* Try to find insns to place in delay slots. */
3746 dbr_schedule (rtx_insn
*first
)
3748 rtx_insn
*insn
, *next
, *epilogue_insn
= 0;
3750 bool need_return_insns
;
3752 /* If the current function has no insns other than the prologue and
3753 epilogue, then do not try to fill any delay slots. */
3754 if (n_basic_blocks_for_fn (cfun
) == NUM_FIXED_BLOCKS
)
3757 /* Find the highest INSN_UID and allocate and initialize our map from
3758 INSN_UID's to position in code. */
3759 for (max_uid
= 0, insn
= first
; insn
; insn
= NEXT_INSN (insn
))
3761 if (INSN_UID (insn
) > max_uid
)
3762 max_uid
= INSN_UID (insn
);
3764 && NOTE_KIND (insn
) == NOTE_INSN_EPILOGUE_BEG
)
3765 epilogue_insn
= insn
;
3768 uid_to_ruid
= XNEWVEC (int, max_uid
+ 1);
3769 for (i
= 0, insn
= first
; insn
; i
++, insn
= NEXT_INSN (insn
))
3770 uid_to_ruid
[INSN_UID (insn
)] = i
;
3772 /* Initialize the list of insns that need filling. */
3773 if (unfilled_firstobj
== 0)
3775 gcc_obstack_init (&unfilled_slots_obstack
);
3776 unfilled_firstobj
= XOBNEWVAR (&unfilled_slots_obstack
, rtx
, 0);
3779 for (insn
= next_active_insn (first
); insn
; insn
= next_active_insn (insn
))
3783 /* Skip vector tables. We can't get attributes for them. */
3784 if (JUMP_TABLE_DATA_P (insn
))
3788 INSN_ANNULLED_BRANCH_P (insn
) = 0;
3789 INSN_FROM_TARGET_P (insn
) = 0;
3791 if (num_delay_slots (insn
) > 0)
3792 obstack_ptr_grow (&unfilled_slots_obstack
, insn
);
3794 /* Ensure all jumps go to the last of a set of consecutive labels. */
3796 && (condjump_p (insn
) || condjump_in_parallel_p (insn
))
3797 && !ANY_RETURN_P (JUMP_LABEL (insn
))
3798 && ((target
= skip_consecutive_labels (JUMP_LABEL (insn
)))
3799 != JUMP_LABEL (insn
)))
3800 redirect_jump (insn
, target
, 1);
3803 init_resource_info (epilogue_insn
);
3805 /* Show we haven't computed an end-of-function label yet. */
3806 function_return_label
= function_simple_return_label
= NULL
;
3808 /* Initialize the statistics for this function. */
3809 memset (num_insns_needing_delays
, 0, sizeof num_insns_needing_delays
);
3810 memset (num_filled_delays
, 0, sizeof num_filled_delays
);
3812 /* Now do the delay slot filling. Try everything twice in case earlier
3813 changes make more slots fillable. */
3815 for (reorg_pass_number
= 0;
3816 reorg_pass_number
< MAX_REORG_PASSES
;
3817 reorg_pass_number
++)
3819 fill_simple_delay_slots (1);
3820 fill_simple_delay_slots (0);
3821 fill_eager_delay_slots ();
3822 relax_delay_slots (first
);
3825 /* If we made an end of function label, indicate that it is now
3826 safe to delete it by undoing our prior adjustment to LABEL_NUSES.
3827 If it is now unused, delete it. */
3828 if (function_return_label
&& --LABEL_NUSES (function_return_label
) == 0)
3829 delete_related_insns (function_return_label
);
3830 if (function_simple_return_label
3831 && --LABEL_NUSES (function_simple_return_label
) == 0)
3832 delete_related_insns (function_simple_return_label
);
3834 need_return_insns
= false;
3836 need_return_insns
|= HAVE_return
&& function_return_label
!= 0;
3838 #ifdef HAVE_simple_return
3839 need_return_insns
|= HAVE_simple_return
&& function_simple_return_label
!= 0;
3841 if (need_return_insns
)
3842 make_return_insns (first
);
3844 /* Delete any USE insns made by update_block; subsequent passes don't need
3845 them or know how to deal with them. */
3846 for (insn
= first
; insn
; insn
= next
)
3848 next
= NEXT_INSN (insn
);
3850 if (NONJUMP_INSN_P (insn
) && GET_CODE (PATTERN (insn
)) == USE
3851 && INSN_P (XEXP (PATTERN (insn
), 0)))
3852 next
= delete_related_insns (insn
);
3855 obstack_free (&unfilled_slots_obstack
, unfilled_firstobj
);
3857 /* It is not clear why the line below is needed, but it does seem to be. */
3858 unfilled_firstobj
= XOBNEWVAR (&unfilled_slots_obstack
, rtx
, 0);
3862 int i
, j
, need_comma
;
3863 int total_delay_slots
[MAX_DELAY_HISTOGRAM
+ 1];
3864 int total_annul_slots
[MAX_DELAY_HISTOGRAM
+ 1];
3866 for (reorg_pass_number
= 0;
3867 reorg_pass_number
< MAX_REORG_PASSES
;
3868 reorg_pass_number
++)
3870 fprintf (dump_file
, ";; Reorg pass #%d:\n", reorg_pass_number
+ 1);
3871 for (i
= 0; i
< NUM_REORG_FUNCTIONS
; i
++)
3874 fprintf (dump_file
, ";; Reorg function #%d\n", i
);
3876 fprintf (dump_file
, ";; %d insns needing delay slots\n;; ",
3877 num_insns_needing_delays
[i
][reorg_pass_number
]);
3879 for (j
= 0; j
< MAX_DELAY_HISTOGRAM
+ 1; j
++)
3880 if (num_filled_delays
[i
][j
][reorg_pass_number
])
3883 fprintf (dump_file
, ", ");
3885 fprintf (dump_file
, "%d got %d delays",
3886 num_filled_delays
[i
][j
][reorg_pass_number
], j
);
3888 fprintf (dump_file
, "\n");
3891 memset (total_delay_slots
, 0, sizeof total_delay_slots
);
3892 memset (total_annul_slots
, 0, sizeof total_annul_slots
);
3893 for (insn
= first
; insn
; insn
= NEXT_INSN (insn
))
3895 if (! insn
->deleted ()
3896 && NONJUMP_INSN_P (insn
)
3897 && GET_CODE (PATTERN (insn
)) != USE
3898 && GET_CODE (PATTERN (insn
)) != CLOBBER
)
3900 if (GET_CODE (PATTERN (insn
)) == SEQUENCE
)
3903 j
= XVECLEN (PATTERN (insn
), 0) - 1;
3904 if (j
> MAX_DELAY_HISTOGRAM
)
3905 j
= MAX_DELAY_HISTOGRAM
;
3906 control
= XVECEXP (PATTERN (insn
), 0, 0);
3907 if (JUMP_P (control
) && INSN_ANNULLED_BRANCH_P (control
))
3908 total_annul_slots
[j
]++;
3910 total_delay_slots
[j
]++;
3912 else if (num_delay_slots (insn
) > 0)
3913 total_delay_slots
[0]++;
3916 fprintf (dump_file
, ";; Reorg totals: ");
3918 for (j
= 0; j
< MAX_DELAY_HISTOGRAM
+ 1; j
++)
3920 if (total_delay_slots
[j
])
3923 fprintf (dump_file
, ", ");
3925 fprintf (dump_file
, "%d got %d delays", total_delay_slots
[j
], j
);
3928 fprintf (dump_file
, "\n");
3929 #if defined (ANNUL_IFTRUE_SLOTS) || defined (ANNUL_IFFALSE_SLOTS)
3930 fprintf (dump_file
, ";; Reorg annuls: ");
3932 for (j
= 0; j
< MAX_DELAY_HISTOGRAM
+ 1; j
++)
3934 if (total_annul_slots
[j
])
3937 fprintf (dump_file
, ", ");
3939 fprintf (dump_file
, "%d got %d delays", total_annul_slots
[j
], j
);
3942 fprintf (dump_file
, "\n");
3944 fprintf (dump_file
, "\n");
3947 if (!sibling_labels
.is_empty ())
3949 update_alignments (sibling_labels
);
3950 sibling_labels
.release ();
3953 free_resource_info ();
3955 crtl
->dbr_scheduled_p
= true;
3957 #endif /* DELAY_SLOTS */
3959 /* Run delay slot optimization. */
3961 rest_of_handle_delay_slots (void)
3964 dbr_schedule (get_insns ());
3971 const pass_data pass_data_delay_slots
=
3973 RTL_PASS
, /* type */
3975 OPTGROUP_NONE
, /* optinfo_flags */
3976 TV_DBR_SCHED
, /* tv_id */
3977 0, /* properties_required */
3978 0, /* properties_provided */
3979 0, /* properties_destroyed */
3980 0, /* todo_flags_start */
3981 0, /* todo_flags_finish */
3984 class pass_delay_slots
: public rtl_opt_pass
3987 pass_delay_slots (gcc::context
*ctxt
)
3988 : rtl_opt_pass (pass_data_delay_slots
, ctxt
)
3991 /* opt_pass methods: */
3992 virtual bool gate (function
*);
3993 virtual unsigned int execute (function
*)
3995 return rest_of_handle_delay_slots ();
3998 }; // class pass_delay_slots
4001 pass_delay_slots::gate (function
*)
4004 /* At -O0 dataflow info isn't updated after RA. */
4005 return optimize
> 0 && flag_delayed_branch
&& !crtl
->dbr_scheduled_p
;
4014 make_pass_delay_slots (gcc::context
*ctxt
)
4016 return new pass_delay_slots (ctxt
);
4019 /* Machine dependent reorg pass. */
4023 const pass_data pass_data_machine_reorg
=
4025 RTL_PASS
, /* type */
4027 OPTGROUP_NONE
, /* optinfo_flags */
4028 TV_MACH_DEP
, /* tv_id */
4029 0, /* properties_required */
4030 0, /* properties_provided */
4031 0, /* properties_destroyed */
4032 0, /* todo_flags_start */
4033 0, /* todo_flags_finish */
4036 class pass_machine_reorg
: public rtl_opt_pass
4039 pass_machine_reorg (gcc::context
*ctxt
)
4040 : rtl_opt_pass (pass_data_machine_reorg
, ctxt
)
4043 /* opt_pass methods: */
4044 virtual bool gate (function
*)
4046 return targetm
.machine_dependent_reorg
!= 0;
4049 virtual unsigned int execute (function
*)
4051 targetm
.machine_dependent_reorg ();
4055 }; // class pass_machine_reorg
4060 make_pass_machine_reorg (gcc::context
*ctxt
)
4062 return new pass_machine_reorg (ctxt
);