1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 1988, 1989, 1991, 1992, 1993, 1994, 1995, 1996, 1997,
3 1998, 1999, 2000 Free Software Foundation, Inc.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
23 /* This is the loop optimization pass of the compiler.
24 It finds invariant computations within loops and moves them
25 to the beginning of the loop. Then it identifies basic and
26 general induction variables. Strength reduction is applied to the general
27 induction variables, and induction variable elimination is applied to
28 the basic induction variables.
30 It also finds cases where
31 a register is set within the loop by zero-extending a narrower value
32 and changes these to zero the entire register once before the loop
33 and merely copy the low part within the loop.
35 Most of the complexity is in heuristics to decide when it is worth
36 while to do these things. */
45 #include "basic-block.h"
46 #include "insn-config.h"
47 #include "insn-flags.h"
49 #include "hard-reg-set.h"
57 /* Vector mapping INSN_UIDs to luids.
58 The luids are like uids but increase monotonically always.
59 We use them to see whether a jump comes from outside a given loop. */
63 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
64 number the insn is contained in. */
66 struct loop **uid_loop;
68 /* 1 + largest uid of any insn. */
72 /* 1 + luid of last insn. */
76 /* Number of loops detected in current function. Used as index to the
79 static int max_loop_num;
81 /* Indexed by register number, contains the number of times the reg
82 is set during the loop being scanned.
83 During code motion, a negative value indicates a reg that has been
84 made a candidate; in particular -2 means that it is an candidate that
85 we know is equal to a constant and -1 means that it is an candidate
86 not known equal to a constant.
87 After code motion, regs moved have 0 (which is accurate now)
88 while the failed candidates have the original number of times set.
90 Therefore, at all times, == 0 indicates an invariant register;
91 < 0 a conditionally invariant one. */
93 static varray_type set_in_loop;
95 /* Original value of set_in_loop; same except that this value
96 is not set negative for a reg whose sets have been made candidates
97 and not set to 0 for a reg that is moved. */
99 static varray_type n_times_set;
101 /* Index by register number, 1 indicates that the register
102 cannot be moved or strength reduced. */
104 static varray_type may_not_optimize;
106 /* Contains the insn in which a register was used if it was used
107 exactly once; contains const0_rtx if it was used more than once. */
109 static varray_type reg_single_usage;
111 /* Nonzero means reg N has already been moved out of one loop.
112 This reduces the desire to move it out of another. */
114 static char *moved_once;
116 /* List of MEMs that are stored in this loop. */
118 static rtx loop_store_mems;
120 /* The insn where the first of these was found. */
121 static rtx first_loop_store_insn;
123 typedef struct loop_mem_info {
124 rtx mem; /* The MEM itself. */
125 rtx reg; /* Corresponding pseudo, if any. */
126 int optimize; /* Nonzero if we can optimize access to this MEM. */
129 /* Array of MEMs that are used (read or written) in this loop, but
130 cannot be aliased by anything in this loop, except perhaps
131 themselves. In other words, if loop_mems[i] is altered during the
132 loop, it is altered by an expression that is rtx_equal_p to it. */
134 static loop_mem_info *loop_mems;
136 /* The index of the next available slot in LOOP_MEMS. */
138 static int loop_mems_idx;
140 /* The number of elements allocated in LOOP_MEMs. */
142 static int loop_mems_allocated;
144 /* Nonzero if we don't know what MEMs were changed in the current
145 loop. This happens if the loop contains a call (in which case
146 `loop_info->has_call' will also be set) or if we store into more
147 than NUM_STORES MEMs. */
149 static int unknown_address_altered;
151 /* The above doesn't count any readonly memory locations that are stored.
154 static int unknown_constant_address_altered;
156 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
157 static int num_movables;
159 /* Count of memory write instructions discovered in the loop. */
160 static int num_mem_sets;
162 /* Bound on pseudo register number before loop optimization.
163 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
164 int max_reg_before_loop;
166 /* The value to pass to the next call of reg_scan_update. */
167 static int loop_max_reg;
169 /* This obstack is used in product_cheap_p to allocate its rtl. It
170 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
171 If we used the same obstack that it did, we would be deallocating
174 static struct obstack temp_obstack;
176 /* This is where the pointer to the obstack being used for RTL is stored. */
178 extern struct obstack *rtl_obstack;
180 #define obstack_chunk_alloc xmalloc
181 #define obstack_chunk_free free
183 /* During the analysis of a loop, a chain of `struct movable's
184 is made to record all the movable insns found.
185 Then the entire chain can be scanned to decide which to move. */
189 rtx insn; /* A movable insn */
190 rtx set_src; /* The expression this reg is set from. */
191 rtx set_dest; /* The destination of this SET. */
192 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
193 of any registers used within the LIBCALL. */
194 int consec; /* Number of consecutive following insns
195 that must be moved with this one. */
196 int regno; /* The register it sets */
197 short lifetime; /* lifetime of that register;
198 may be adjusted when matching movables
199 that load the same value are found. */
200 short savings; /* Number of insns we can move for this reg,
201 including other movables that force this
202 or match this one. */
203 unsigned int cond : 1; /* 1 if only conditionally movable */
204 unsigned int force : 1; /* 1 means MUST move this insn */
205 unsigned int global : 1; /* 1 means reg is live outside this loop */
206 /* If PARTIAL is 1, GLOBAL means something different:
207 that the reg is live outside the range from where it is set
208 to the following label. */
209 unsigned int done : 1; /* 1 inhibits further processing of this */
211 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
212 In particular, moving it does not make it
214 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
215 load SRC, rather than copying INSN. */
216 unsigned int move_insn_first:1;/* Same as above, if this is necessary for the
217 first insn of a consecutive sets group. */
218 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
219 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
220 that we should avoid changing when clearing
221 the rest of the reg. */
222 struct movable *match; /* First entry for same value */
223 struct movable *forces; /* An insn that must be moved if this is */
224 struct movable *next;
227 static struct movable *the_movables;
229 FILE *loop_dump_stream;
231 /* Forward declarations. */
233 static void verify_dominator PARAMS ((struct loop *));
234 static void find_and_verify_loops PARAMS ((rtx, struct loops *));
235 static void mark_loop_jump PARAMS ((rtx, struct loop *));
236 static void prescan_loop PARAMS ((struct loop *));
237 static int reg_in_basic_block_p PARAMS ((rtx, rtx));
238 static int consec_sets_invariant_p PARAMS ((const struct loop *,
240 static int labels_in_range_p PARAMS ((rtx, int));
241 static void count_one_set PARAMS ((rtx, rtx, varray_type, rtx *));
243 static void count_loop_regs_set PARAMS ((rtx, rtx, varray_type, varray_type,
245 static void note_addr_stored PARAMS ((rtx, rtx, void *));
246 static void note_set_pseudo_multiple_uses PARAMS ((rtx, rtx, void *));
247 static int loop_reg_used_before_p PARAMS ((const struct loop *, rtx, rtx));
248 static void scan_loop PARAMS ((struct loop*, int, int));
250 static void replace_call_address PARAMS ((rtx, rtx, rtx));
252 static rtx skip_consec_insns PARAMS ((rtx, int));
253 static int libcall_benefit PARAMS ((rtx));
254 static void ignore_some_movables PARAMS ((struct movable *));
255 static void force_movables PARAMS ((struct movable *));
256 static void combine_movables PARAMS ((struct movable *, int));
257 static int regs_match_p PARAMS ((rtx, rtx, struct movable *));
258 static int rtx_equal_for_loop_p PARAMS ((rtx, rtx, struct movable *));
259 static void add_label_notes PARAMS ((rtx, rtx));
260 static void move_movables PARAMS ((struct loop *loop, struct movable *,
262 static int count_nonfixed_reads PARAMS ((const struct loop *, rtx));
263 static void strength_reduce PARAMS ((struct loop *, int, int, int));
264 static void find_single_use_in_loop PARAMS ((rtx, rtx, varray_type));
265 static int valid_initial_value_p PARAMS ((rtx, rtx, int, rtx));
266 static void find_mem_givs PARAMS ((const struct loop *, rtx, rtx, int, int));
267 static void record_biv PARAMS ((struct induction *, rtx, rtx, rtx, rtx, rtx *,
269 static void check_final_value PARAMS ((const struct loop *,
270 struct induction *));
271 static void record_giv PARAMS ((const struct loop *, struct induction *,
272 rtx, rtx, rtx, rtx, rtx, int, enum g_types,
274 static void update_giv_derive PARAMS ((const struct loop *, rtx));
275 static int basic_induction_var PARAMS ((const struct loop *, rtx,
276 enum machine_mode, rtx, rtx,
277 rtx *, rtx *, rtx **, int *));
278 static rtx simplify_giv_expr PARAMS ((const struct loop *, rtx, int *));
279 static int general_induction_var PARAMS ((const struct loop *loop, rtx, rtx *,
280 rtx *, rtx *, int, int *));
281 static int consec_sets_giv PARAMS ((const struct loop *, int, rtx,
282 rtx, rtx, rtx *, rtx *, rtx *));
283 static int check_dbra_loop PARAMS ((struct loop *, int));
284 static rtx express_from_1 PARAMS ((rtx, rtx, rtx));
285 static rtx combine_givs_p PARAMS ((struct induction *, struct induction *));
286 static void combine_givs PARAMS ((struct iv_class *));
287 struct recombine_givs_stats;
288 static int find_life_end PARAMS ((rtx, struct recombine_givs_stats *,
290 static void recombine_givs PARAMS ((const struct loop *, struct iv_class *,
292 static int product_cheap_p PARAMS ((rtx, rtx));
293 static int maybe_eliminate_biv PARAMS ((const struct loop *, struct iv_class *,
295 static int maybe_eliminate_biv_1 PARAMS ((const struct loop *, rtx, rtx,
296 struct iv_class *, int, rtx));
297 static int last_use_this_basic_block PARAMS ((rtx, rtx));
298 static void record_initial PARAMS ((rtx, rtx, void *));
299 static void update_reg_last_use PARAMS ((rtx, rtx));
300 static rtx next_insn_in_loop PARAMS ((const struct loop *, rtx));
301 static void load_mems_and_recount_loop_regs_set PARAMS ((const struct loop*,
303 static void load_mems PARAMS ((const struct loop *));
304 static int insert_loop_mem PARAMS ((rtx *, void *));
305 static int replace_loop_mem PARAMS ((rtx *, void *));
306 static int replace_loop_reg PARAMS ((rtx *, void *));
307 static void note_reg_stored PARAMS ((rtx, rtx, void *));
308 static void try_copy_prop PARAMS ((const struct loop *, rtx, int));
309 static int replace_label PARAMS ((rtx *, void *));
311 typedef struct rtx_and_int {
316 typedef struct rtx_pair {
321 /* Nonzero iff INSN is between START and END, inclusive. */
322 #define INSN_IN_RANGE_P(INSN, START, END) \
323 (INSN_UID (INSN) < max_uid_for_loop \
324 && INSN_LUID (INSN) >= INSN_LUID (START) \
325 && INSN_LUID (INSN) <= INSN_LUID (END))
327 #ifdef HAVE_decrement_and_branch_on_count
328 /* Test whether BCT applicable and safe. */
329 static void insert_bct PARAMS ((struct loop *));
331 /* Auxiliary function that inserts the BCT pattern into the loop. */
332 static void instrument_loop_bct PARAMS ((rtx, rtx, rtx));
333 #endif /* HAVE_decrement_and_branch_on_count */
335 /* Indirect_jump_in_function is computed once per function. */
336 int indirect_jump_in_function = 0;
337 static int indirect_jump_in_function_p PARAMS ((rtx));
339 static int compute_luids PARAMS ((rtx, rtx, int));
341 static int biv_elimination_giv_has_0_offset PARAMS ((struct induction *,
342 struct induction *, rtx));
344 /* Relative gain of eliminating various kinds of operations. */
347 static int shift_cost;
348 static int mult_cost;
351 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
352 copy the value of the strength reduced giv to its original register. */
353 static int copy_cost;
355 /* Cost of using a register, to normalize the benefits of a giv. */
356 static int reg_address_cost;
362 char *free_point = (char *) oballoc (1);
363 rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
365 add_cost = rtx_cost (gen_rtx_PLUS (word_mode, reg, reg), SET);
368 reg_address_cost = ADDRESS_COST (reg);
370 reg_address_cost = rtx_cost (reg, MEM);
373 /* We multiply by 2 to reconcile the difference in scale between
374 these two ways of computing costs. Otherwise the cost of a copy
375 will be far less than the cost of an add. */
379 /* Free the objects we just allocated. */
382 /* Initialize the obstack used for rtl in product_cheap_p. */
383 gcc_obstack_init (&temp_obstack);
386 /* Compute the mapping from uids to luids.
387 LUIDs are numbers assigned to insns, like uids,
388 except that luids increase monotonically through the code.
389 Start at insn START and stop just before END. Assign LUIDs
390 starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
392 compute_luids (start, end, prev_luid)
399 for (insn = start, i = prev_luid; insn != end; insn = NEXT_INSN (insn))
401 if (INSN_UID (insn) >= max_uid_for_loop)
403 /* Don't assign luids to line-number NOTEs, so that the distance in
404 luids between two insns is not affected by -g. */
405 if (GET_CODE (insn) != NOTE
406 || NOTE_LINE_NUMBER (insn) <= 0)
407 uid_luid[INSN_UID (insn)] = ++i;
409 /* Give a line number note the same luid as preceding insn. */
410 uid_luid[INSN_UID (insn)] = i;
415 /* Entry point of this file. Perform loop optimization
416 on the current function. F is the first insn of the function
417 and DUMPFILE is a stream for output of a trace of actions taken
418 (or 0 if none should be output). */
421 loop_optimize (f, dumpfile, unroll_p, bct_p)
422 /* f is the first instruction of a chain of insns for one function */
429 struct loops loops_data;
430 struct loops *loops = &loops_data;
431 struct loop_info *loops_info;
433 loop_dump_stream = dumpfile;
435 init_recog_no_volatile ();
437 max_reg_before_loop = max_reg_num ();
438 loop_max_reg = max_reg_before_loop;
442 /* Count the number of loops. */
445 for (insn = f; insn; insn = NEXT_INSN (insn))
447 if (GET_CODE (insn) == NOTE
448 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
452 /* Don't waste time if no loops. */
453 if (max_loop_num == 0)
456 loops->num = max_loop_num;
458 moved_once = (char *) xcalloc (max_reg_before_loop, sizeof (char));
460 /* Get size to use for tables indexed by uids.
461 Leave some space for labels allocated by find_and_verify_loops. */
462 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
464 uid_luid = (int *) xcalloc (max_uid_for_loop, sizeof (int));
465 uid_loop = (struct loop **) xcalloc (max_uid_for_loop,
466 sizeof (struct loop *));
468 /* Allocate storage for array of loops. */
469 loops->array = (struct loop *)
470 xcalloc (loops->num, sizeof (struct loop));
472 /* Find and process each loop.
473 First, find them, and record them in order of their beginnings. */
474 find_and_verify_loops (f, loops);
476 /* Allocate and initialize auxiliary loop information. */
477 loops_info = xcalloc (loops->num, sizeof (struct loop_info));
478 for (i = 0; i < loops->num; i++)
479 loops->array[i].aux = loops_info + i;
481 /* Now find all register lifetimes. This must be done after
482 find_and_verify_loops, because it might reorder the insns in the
484 reg_scan (f, max_reg_before_loop, 1);
486 /* This must occur after reg_scan so that registers created by gcse
487 will have entries in the register tables.
489 We could have added a call to reg_scan after gcse_main in toplev.c,
490 but moving this call to init_alias_analysis is more efficient. */
491 init_alias_analysis ();
493 /* See if we went too far. Note that get_max_uid already returns
494 one more that the maximum uid of all insn. */
495 if (get_max_uid () > max_uid_for_loop)
497 /* Now reset it to the actual size we need. See above. */
498 max_uid_for_loop = get_max_uid ();
500 /* find_and_verify_loops has already called compute_luids, but it
501 might have rearranged code afterwards, so we need to recompute
503 max_luid = compute_luids (f, NULL_RTX, 0);
505 /* Don't leave gaps in uid_luid for insns that have been
506 deleted. It is possible that the first or last insn
507 using some register has been deleted by cross-jumping.
508 Make sure that uid_luid for that former insn's uid
509 points to the general area where that insn used to be. */
510 for (i = 0; i < max_uid_for_loop; i++)
512 uid_luid[0] = uid_luid[i];
513 if (uid_luid[0] != 0)
516 for (i = 0; i < max_uid_for_loop; i++)
517 if (uid_luid[i] == 0)
518 uid_luid[i] = uid_luid[i - 1];
520 /* If debugging and unrolling loops, we must replicate the tree
521 nodes corresponding to the BLOCKs inside the loop, so that the
522 original one to one mapping will remain. We sometimes unroll
523 loops even when unroll_p is false, so we must always do this when
525 if (write_symbols != NO_DEBUG)
526 find_loop_tree_blocks ();
528 /* Determine if the function has indirect jump. On some systems
529 this prevents low overhead loop instructions from being used. */
530 indirect_jump_in_function = indirect_jump_in_function_p (f);
532 /* Now scan the loops, last ones first, since this means inner ones are done
533 before outer ones. */
534 for (i = max_loop_num - 1; i >= 0; i--)
536 struct loop *loop = &loops->array[i];
538 if (! loop->invalid && loop->end)
539 scan_loop (loop, unroll_p, bct_p);
542 /* Replicate the BLOCKs. */
543 if (write_symbols != NO_DEBUG)
544 unroll_block_trees ();
546 end_alias_analysis ();
556 /* Returns the next insn, in execution order, after INSN. START and
557 END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
558 respectively. LOOP->TOP, if non-NULL, is the top of the loop in the
559 insn-stream; it is used with loops that are entered near the
563 next_insn_in_loop (loop, insn)
564 const struct loop *loop;
567 insn = NEXT_INSN (insn);
569 if (insn == loop->end)
572 /* Go to the top of the loop, and continue there. */
579 if (insn == loop->scan_start)
586 /* Optimize one loop described by LOOP. */
588 /* ??? Could also move memory writes out of loops if the destination address
589 is invariant, the source is invariant, the memory write is not volatile,
590 and if we can prove that no read inside the loop can read this address
591 before the write occurs. If there is a read of this address after the
592 write, then we can also mark the memory read as invariant. */
595 scan_loop (loop, unroll_p, bct_p)
600 rtx loop_start = loop->start;
601 rtx loop_end = loop->end;
602 /* Additional information about the current loop being processed
603 that is used to compute the number of loop iterations for loop
604 unrolling and doloop optimization. */
605 struct loop_info *loop_info = LOOP_INFO (loop);
607 /* 1 if we are scanning insns that could be executed zero times. */
609 /* 1 if we are scanning insns that might never be executed
610 due to a subroutine call which might exit before they are reached. */
612 /* Jump insn that enters the loop, or 0 if control drops in. */
613 rtx loop_entry_jump = 0;
614 /* Number of insns in the loop. */
618 rtx temp, update_start, update_end;
619 /* The SET from an insn, if it is the only SET in the insn. */
621 /* Chain describing insns movable in current loop. */
622 struct movable *movables = 0;
623 /* Last element in `movables' -- so we can add elements at the end. */
624 struct movable *last_movable = 0;
625 /* Ratio of extra register life span we can justify
626 for saving an instruction. More if loop doesn't call subroutines
627 since in that case saving an insn makes more difference
628 and more registers are available. */
630 /* Nonzero if we are scanning instructions in a sub-loop. */
636 /* Determine whether this loop starts with a jump down to a test at
637 the end. This will occur for a small number of loops with a test
638 that is too complex to duplicate in front of the loop.
640 We search for the first insn or label in the loop, skipping NOTEs.
641 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
642 (because we might have a loop executed only once that contains a
643 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
644 (in case we have a degenerate loop).
646 Note that if we mistakenly think that a loop is entered at the top
647 when, in fact, it is entered at the exit test, the only effect will be
648 slightly poorer optimization. Making the opposite error can generate
649 incorrect code. Since very few loops now start with a jump to the
650 exit test, the code here to detect that case is very conservative. */
652 for (p = NEXT_INSN (loop_start);
654 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
655 && (GET_CODE (p) != NOTE
656 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
657 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
661 loop->scan_start = p;
663 /* Set up variables describing this loop. */
665 threshold = (loop_info->has_call ? 1 : 2) * (1 + n_non_fixed_regs);
667 /* If loop has a jump before the first label,
668 the true entry is the target of that jump.
669 Start scan from there.
670 But record in LOOP->TOP the place where the end-test jumps
671 back to so we can scan that after the end of the loop. */
672 if (GET_CODE (p) == JUMP_INSN)
676 /* Loop entry must be unconditional jump (and not a RETURN) */
678 && JUMP_LABEL (p) != 0
679 /* Check to see whether the jump actually
680 jumps out of the loop (meaning it's no loop).
681 This case can happen for things like
682 do {..} while (0). If this label was generated previously
683 by loop, we can't tell anything about it and have to reject
685 && INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, loop_end))
687 loop->top = next_label (loop->scan_start);
688 loop->scan_start = JUMP_LABEL (p);
692 /* If LOOP->SCAN_START was an insn created by loop, we don't know its luid
693 as required by loop_reg_used_before_p. So skip such loops. (This
694 test may never be true, but it's best to play it safe.)
696 Also, skip loops where we do not start scanning at a label. This
697 test also rejects loops starting with a JUMP_INSN that failed the
700 if (INSN_UID (loop->scan_start) >= max_uid_for_loop
701 || GET_CODE (loop->scan_start) != CODE_LABEL)
703 if (loop_dump_stream)
704 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
705 INSN_UID (loop_start), INSN_UID (loop_end));
709 /* Count number of times each reg is set during this loop.
710 Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out
711 the setting of register I. Set VARRAY_RTX (reg_single_usage, I). */
713 /* Allocate extra space for REGS that might be created by
714 load_mems. We allocate a little extra slop as well, in the hopes
715 that even after the moving of movables creates some new registers
716 we won't have to reallocate these arrays. However, we do grow
717 the arrays, if necessary, in load_mems_recount_loop_regs_set. */
718 nregs = max_reg_num () + loop_mems_idx + 16;
719 VARRAY_INT_INIT (set_in_loop, nregs, "set_in_loop");
720 VARRAY_INT_INIT (n_times_set, nregs, "n_times_set");
721 VARRAY_CHAR_INIT (may_not_optimize, nregs, "may_not_optimize");
722 VARRAY_RTX_INIT (reg_single_usage, nregs, "reg_single_usage");
724 count_loop_regs_set (loop->top ? loop->top : loop->start, loop->end,
725 may_not_optimize, reg_single_usage, &insn_count, nregs);
727 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
729 VARRAY_CHAR (may_not_optimize, i) = 1;
730 VARRAY_INT (set_in_loop, i) = 1;
733 #ifdef AVOID_CCMODE_COPIES
734 /* Don't try to move insns which set CC registers if we should not
735 create CCmode register copies. */
736 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
737 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
738 VARRAY_CHAR (may_not_optimize, i) = 1;
741 bcopy ((char *) &set_in_loop->data,
742 (char *) &n_times_set->data, nregs * sizeof (int));
744 if (loop_dump_stream)
746 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
747 INSN_UID (loop_start), INSN_UID (loop_end), insn_count);
749 fprintf (loop_dump_stream, "Continue at insn %d.\n",
750 INSN_UID (loop->cont));
753 /* Scan through the loop finding insns that are safe to move.
754 Set set_in_loop negative for the reg being set, so that
755 this reg will be considered invariant for subsequent insns.
756 We consider whether subsequent insns use the reg
757 in deciding whether it is worth actually moving.
759 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
760 and therefore it is possible that the insns we are scanning
761 would never be executed. At such times, we must make sure
762 that it is safe to execute the insn once instead of zero times.
763 When MAYBE_NEVER is 0, all insns will be executed at least once
764 so that is not a problem. */
766 for (p = next_insn_in_loop (loop, loop->scan_start);
768 p = next_insn_in_loop (loop, p))
770 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
771 && find_reg_note (p, REG_LIBCALL, NULL_RTX))
773 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
774 && find_reg_note (p, REG_RETVAL, NULL_RTX))
777 if (GET_CODE (p) == INSN
778 && (set = single_set (p))
779 && GET_CODE (SET_DEST (set)) == REG
780 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
785 rtx src = SET_SRC (set);
786 rtx dependencies = 0;
788 /* Figure out what to use as a source of this insn. If a REG_EQUIV
789 note is given or if a REG_EQUAL note with a constant operand is
790 specified, use it as the source and mark that we should move
791 this insn by calling emit_move_insn rather that duplicating the
794 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
796 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
798 src = XEXP (temp, 0), move_insn = 1;
801 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
802 if (temp && CONSTANT_P (XEXP (temp, 0)))
803 src = XEXP (temp, 0), move_insn = 1;
804 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
806 src = XEXP (temp, 0);
807 /* A libcall block can use regs that don't appear in
808 the equivalent expression. To move the libcall,
809 we must move those regs too. */
810 dependencies = libcall_other_reg (p, src);
814 /* Don't try to optimize a register that was made
815 by loop-optimization for an inner loop.
816 We don't know its life-span, so we can't compute the benefit. */
817 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
819 else if (/* The register is used in basic blocks other
820 than the one where it is set (meaning that
821 something after this point in the loop might
822 depend on its value before the set). */
823 ! reg_in_basic_block_p (p, SET_DEST (set))
824 /* And the set is not guaranteed to be executed one
825 the loop starts, or the value before the set is
826 needed before the set occurs...
828 ??? Note we have quadratic behaviour here, mitigated
829 by the fact that the previous test will often fail for
830 large loops. Rather than re-scanning the entire loop
831 each time for register usage, we should build tables
832 of the register usage and use them here instead. */
834 || loop_reg_used_before_p (loop, set, p)))
835 /* It is unsafe to move the set.
837 This code used to consider it OK to move a set of a variable
838 which was not created by the user and not used in an exit test.
839 That behavior is incorrect and was removed. */
841 else if ((tem = loop_invariant_p (loop, src))
842 && (dependencies == 0
843 || (tem2 = loop_invariant_p (loop, dependencies)) != 0)
844 && (VARRAY_INT (set_in_loop,
845 REGNO (SET_DEST (set))) == 1
847 = consec_sets_invariant_p
848 (loop, SET_DEST (set),
849 VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))),
851 /* If the insn can cause a trap (such as divide by zero),
852 can't move it unless it's guaranteed to be executed
853 once loop is entered. Even a function call might
854 prevent the trap insn from being reached
855 (since it might exit!) */
856 && ! ((maybe_never || call_passed)
857 && may_trap_p (src)))
859 register struct movable *m;
860 register int regno = REGNO (SET_DEST (set));
862 /* A potential lossage is where we have a case where two insns
863 can be combined as long as they are both in the loop, but
864 we move one of them outside the loop. For large loops,
865 this can lose. The most common case of this is the address
866 of a function being called.
868 Therefore, if this register is marked as being used exactly
869 once if we are in a loop with calls (a "large loop"), see if
870 we can replace the usage of this register with the source
871 of this SET. If we can, delete this insn.
873 Don't do this if P has a REG_RETVAL note or if we have
874 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
876 if (loop_info->has_call
877 && VARRAY_RTX (reg_single_usage, regno) != 0
878 && VARRAY_RTX (reg_single_usage, regno) != const0_rtx
879 && REGNO_FIRST_UID (regno) == INSN_UID (p)
880 && (REGNO_LAST_UID (regno)
881 == INSN_UID (VARRAY_RTX (reg_single_usage, regno)))
882 && VARRAY_INT (set_in_loop, regno) == 1
883 && ! side_effects_p (SET_SRC (set))
884 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
885 && (! SMALL_REGISTER_CLASSES
886 || (! (GET_CODE (SET_SRC (set)) == REG
887 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)))
888 /* This test is not redundant; SET_SRC (set) might be
889 a call-clobbered register and the life of REGNO
890 might span a call. */
891 && ! modified_between_p (SET_SRC (set), p,
893 (reg_single_usage, regno))
894 && no_labels_between_p (p, VARRAY_RTX (reg_single_usage, regno))
895 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
897 (reg_single_usage, regno)))
899 /* Replace any usage in a REG_EQUAL note. Must copy the
900 new source, so that we don't get rtx sharing between the
901 SET_SOURCE and REG_NOTES of insn p. */
902 REG_NOTES (VARRAY_RTX (reg_single_usage, regno))
903 = replace_rtx (REG_NOTES (VARRAY_RTX
904 (reg_single_usage, regno)),
905 SET_DEST (set), copy_rtx (SET_SRC (set)));
908 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
909 NOTE_SOURCE_FILE (p) = 0;
910 VARRAY_INT (set_in_loop, regno) = 0;
914 m = (struct movable *) alloca (sizeof (struct movable));
918 m->dependencies = dependencies;
919 m->set_dest = SET_DEST (set);
921 m->consec = VARRAY_INT (set_in_loop,
922 REGNO (SET_DEST (set))) - 1;
926 m->move_insn = move_insn;
927 m->move_insn_first = 0;
928 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
929 m->savemode = VOIDmode;
931 /* Set M->cond if either loop_invariant_p
932 or consec_sets_invariant_p returned 2
933 (only conditionally invariant). */
934 m->cond = ((tem | tem1 | tem2) > 1);
935 m->global = (uid_luid[REGNO_LAST_UID (regno)]
936 > INSN_LUID (loop_end)
937 || uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (loop_start));
939 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
940 - uid_luid[REGNO_FIRST_UID (regno)]);
941 m->savings = VARRAY_INT (n_times_set, regno);
942 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
943 m->savings += libcall_benefit (p);
944 VARRAY_INT (set_in_loop, regno) = move_insn ? -2 : -1;
945 /* Add M to the end of the chain MOVABLES. */
949 last_movable->next = m;
954 /* It is possible for the first instruction to have a
955 REG_EQUAL note but a non-invariant SET_SRC, so we must
956 remember the status of the first instruction in case
957 the last instruction doesn't have a REG_EQUAL note. */
958 m->move_insn_first = m->move_insn;
960 /* Skip this insn, not checking REG_LIBCALL notes. */
961 p = next_nonnote_insn (p);
962 /* Skip the consecutive insns, if there are any. */
963 p = skip_consec_insns (p, m->consec);
964 /* Back up to the last insn of the consecutive group. */
965 p = prev_nonnote_insn (p);
967 /* We must now reset m->move_insn, m->is_equiv, and possibly
968 m->set_src to correspond to the effects of all the
970 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
972 m->set_src = XEXP (temp, 0), m->move_insn = 1;
975 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
976 if (temp && CONSTANT_P (XEXP (temp, 0)))
977 m->set_src = XEXP (temp, 0), m->move_insn = 1;
982 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
985 /* If this register is always set within a STRICT_LOW_PART
986 or set to zero, then its high bytes are constant.
987 So clear them outside the loop and within the loop
988 just load the low bytes.
989 We must check that the machine has an instruction to do so.
990 Also, if the value loaded into the register
991 depends on the same register, this cannot be done. */
992 else if (SET_SRC (set) == const0_rtx
993 && GET_CODE (NEXT_INSN (p)) == INSN
994 && (set1 = single_set (NEXT_INSN (p)))
995 && GET_CODE (set1) == SET
996 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
997 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
998 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
1000 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
1002 register int regno = REGNO (SET_DEST (set));
1003 if (VARRAY_INT (set_in_loop, regno) == 2)
1005 register struct movable *m;
1006 m = (struct movable *) alloca (sizeof (struct movable));
1009 m->set_dest = SET_DEST (set);
1010 m->dependencies = 0;
1016 m->move_insn_first = 0;
1018 /* If the insn may not be executed on some cycles,
1019 we can't clear the whole reg; clear just high part.
1020 Not even if the reg is used only within this loop.
1027 Clearing x before the inner loop could clobber a value
1028 being saved from the last time around the outer loop.
1029 However, if the reg is not used outside this loop
1030 and all uses of the register are in the same
1031 basic block as the store, there is no problem.
1033 If this insn was made by loop, we don't know its
1034 INSN_LUID and hence must make a conservative
1036 m->global = (INSN_UID (p) >= max_uid_for_loop
1037 || (uid_luid[REGNO_LAST_UID (regno)]
1038 > INSN_LUID (loop_end))
1039 || (uid_luid[REGNO_FIRST_UID (regno)]
1041 || (labels_in_range_p
1042 (p, uid_luid[REGNO_FIRST_UID (regno)])));
1043 if (maybe_never && m->global)
1044 m->savemode = GET_MODE (SET_SRC (set1));
1046 m->savemode = VOIDmode;
1050 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
1051 - uid_luid[REGNO_FIRST_UID (regno)]);
1053 VARRAY_INT (set_in_loop, regno) = -1;
1054 /* Add M to the end of the chain MOVABLES. */
1058 last_movable->next = m;
1063 /* Past a call insn, we get to insns which might not be executed
1064 because the call might exit. This matters for insns that trap.
1065 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
1066 so they don't count. */
1067 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
1069 /* Past a label or a jump, we get to insns for which we
1070 can't count on whether or how many times they will be
1071 executed during each iteration. Therefore, we can
1072 only move out sets of trivial variables
1073 (those not used after the loop). */
1074 /* Similar code appears twice in strength_reduce. */
1075 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
1076 /* If we enter the loop in the middle, and scan around to the
1077 beginning, don't set maybe_never for that. This must be an
1078 unconditional jump, otherwise the code at the top of the
1079 loop might never be executed. Unconditional jumps are
1080 followed a by barrier then loop end. */
1081 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop->top
1082 && NEXT_INSN (NEXT_INSN (p)) == loop_end
1083 && simplejump_p (p)))
1085 else if (GET_CODE (p) == NOTE)
1087 /* At the virtual top of a converted loop, insns are again known to
1088 be executed: logically, the loop begins here even though the exit
1089 code has been duplicated. */
1090 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
1091 maybe_never = call_passed = 0;
1092 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
1094 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
1099 /* If one movable subsumes another, ignore that other. */
1101 ignore_some_movables (movables);
1103 /* For each movable insn, see if the reg that it loads
1104 leads when it dies right into another conditionally movable insn.
1105 If so, record that the second insn "forces" the first one,
1106 since the second can be moved only if the first is. */
1108 force_movables (movables);
1110 /* See if there are multiple movable insns that load the same value.
1111 If there are, make all but the first point at the first one
1112 through the `match' field, and add the priorities of them
1113 all together as the priority of the first. */
1115 combine_movables (movables, nregs);
1117 /* Now consider each movable insn to decide whether it is worth moving.
1118 Store 0 in set_in_loop for each reg that is moved.
1120 Generally this increases code size, so do not move moveables when
1121 optimizing for code size. */
1123 if (! optimize_size)
1124 move_movables (loop, movables, threshold, insn_count, nregs);
1126 /* Now candidates that still are negative are those not moved.
1127 Change set_in_loop to indicate that those are not actually invariant. */
1128 for (i = 0; i < nregs; i++)
1129 if (VARRAY_INT (set_in_loop, i) < 0)
1130 VARRAY_INT (set_in_loop, i) = VARRAY_INT (n_times_set, i);
1132 /* Now that we've moved some things out of the loop, we might be able to
1133 hoist even more memory references. */
1134 load_mems_and_recount_loop_regs_set (loop, &insn_count);
1136 for (update_start = loop_start;
1137 PREV_INSN (update_start)
1138 && GET_CODE (PREV_INSN (update_start)) != CODE_LABEL;
1139 update_start = PREV_INSN (update_start))
1141 update_end = NEXT_INSN (loop_end);
1143 reg_scan_update (update_start, update_end, loop_max_reg);
1144 loop_max_reg = max_reg_num ();
1146 if (flag_strength_reduce)
1148 the_movables = movables;
1149 strength_reduce (loop, insn_count, unroll_p, bct_p);
1151 reg_scan_update (update_start, update_end, loop_max_reg);
1152 loop_max_reg = max_reg_num ();
1155 VARRAY_FREE (reg_single_usage);
1156 VARRAY_FREE (set_in_loop);
1157 VARRAY_FREE (n_times_set);
1158 VARRAY_FREE (may_not_optimize);
1161 /* Add elements to *OUTPUT to record all the pseudo-regs
1162 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1165 record_excess_regs (in_this, not_in_this, output)
1166 rtx in_this, not_in_this;
1173 code = GET_CODE (in_this);
1187 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1188 && ! reg_mentioned_p (in_this, not_in_this))
1189 *output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output);
1196 fmt = GET_RTX_FORMAT (code);
1197 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1204 for (j = 0; j < XVECLEN (in_this, i); j++)
1205 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1209 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1215 /* Check what regs are referred to in the libcall block ending with INSN,
1216 aside from those mentioned in the equivalent value.
1217 If there are none, return 0.
1218 If there are one or more, return an EXPR_LIST containing all of them. */
1221 libcall_other_reg (insn, equiv)
1224 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1225 rtx p = XEXP (note, 0);
1228 /* First, find all the regs used in the libcall block
1229 that are not mentioned as inputs to the result. */
1233 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1234 || GET_CODE (p) == CALL_INSN)
1235 record_excess_regs (PATTERN (p), equiv, &output);
1242 /* Return 1 if all uses of REG
1243 are between INSN and the end of the basic block. */
1246 reg_in_basic_block_p (insn, reg)
1249 int regno = REGNO (reg);
1252 if (REGNO_FIRST_UID (regno) != INSN_UID (insn))
1255 /* Search this basic block for the already recorded last use of the reg. */
1256 for (p = insn; p; p = NEXT_INSN (p))
1258 switch (GET_CODE (p))
1265 /* Ordinary insn: if this is the last use, we win. */
1266 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1271 /* Jump insn: if this is the last use, we win. */
1272 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1274 /* Otherwise, it's the end of the basic block, so we lose. */
1279 /* It's the end of the basic block, so we lose. */
1287 /* The "last use" that was recorded can't be found after the first
1288 use. This can happen when the last use was deleted while
1289 processing an inner loop, this inner loop was then completely
1290 unrolled, and the outer loop is always exited after the inner loop,
1291 so that everything after the first use becomes a single basic block. */
1295 /* Compute the benefit of eliminating the insns in the block whose
1296 last insn is LAST. This may be a group of insns used to compute a
1297 value directly or can contain a library call. */
1300 libcall_benefit (last)
1306 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1307 insn != last; insn = NEXT_INSN (insn))
1309 if (GET_CODE (insn) == CALL_INSN)
1310 benefit += 10; /* Assume at least this many insns in a library
1312 else if (GET_CODE (insn) == INSN
1313 && GET_CODE (PATTERN (insn)) != USE
1314 && GET_CODE (PATTERN (insn)) != CLOBBER)
1321 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1324 skip_consec_insns (insn, count)
1328 for (; count > 0; count--)
1332 /* If first insn of libcall sequence, skip to end. */
1333 /* Do this at start of loop, since INSN is guaranteed to
1335 if (GET_CODE (insn) != NOTE
1336 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1337 insn = XEXP (temp, 0);
1339 do insn = NEXT_INSN (insn);
1340 while (GET_CODE (insn) == NOTE);
1346 /* Ignore any movable whose insn falls within a libcall
1347 which is part of another movable.
1348 We make use of the fact that the movable for the libcall value
1349 was made later and so appears later on the chain. */
1352 ignore_some_movables (movables)
1353 struct movable *movables;
1355 register struct movable *m, *m1;
1357 for (m = movables; m; m = m->next)
1359 /* Is this a movable for the value of a libcall? */
1360 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1364 /* Check for earlier movables inside that range,
1365 and mark them invalid. We cannot use LUIDs here because
1366 insns created by loop.c for prior loops don't have LUIDs.
1367 Rather than reject all such insns from movables, we just
1368 explicitly check each insn in the libcall (since invariant
1369 libcalls aren't that common). */
1370 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1371 for (m1 = movables; m1 != m; m1 = m1->next)
1372 if (m1->insn == insn)
1378 /* For each movable insn, see if the reg that it loads
1379 leads when it dies right into another conditionally movable insn.
1380 If so, record that the second insn "forces" the first one,
1381 since the second can be moved only if the first is. */
1384 force_movables (movables)
1385 struct movable *movables;
1387 register struct movable *m, *m1;
1388 for (m1 = movables; m1; m1 = m1->next)
1389 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1390 if (!m1->partial && !m1->done)
1392 int regno = m1->regno;
1393 for (m = m1->next; m; m = m->next)
1394 /* ??? Could this be a bug? What if CSE caused the
1395 register of M1 to be used after this insn?
1396 Since CSE does not update regno_last_uid,
1397 this insn M->insn might not be where it dies.
1398 But very likely this doesn't matter; what matters is
1399 that M's reg is computed from M1's reg. */
1400 if (INSN_UID (m->insn) == REGNO_LAST_UID (regno)
1403 if (m != 0 && m->set_src == m1->set_dest
1404 /* If m->consec, m->set_src isn't valid. */
1408 /* Increase the priority of the moving the first insn
1409 since it permits the second to be moved as well. */
1413 m1->lifetime += m->lifetime;
1414 m1->savings += m->savings;
1419 /* Find invariant expressions that are equal and can be combined into
1423 combine_movables (movables, nregs)
1424 struct movable *movables;
1427 register struct movable *m;
1428 char *matched_regs = (char *) xmalloc (nregs);
1429 enum machine_mode mode;
1431 /* Regs that are set more than once are not allowed to match
1432 or be matched. I'm no longer sure why not. */
1433 /* Perhaps testing m->consec_sets would be more appropriate here? */
1435 for (m = movables; m; m = m->next)
1436 if (m->match == 0 && VARRAY_INT (n_times_set, m->regno) == 1 && !m->partial)
1438 register struct movable *m1;
1439 int regno = m->regno;
1441 bzero (matched_regs, nregs);
1442 matched_regs[regno] = 1;
1444 /* We want later insns to match the first one. Don't make the first
1445 one match any later ones. So start this loop at m->next. */
1446 for (m1 = m->next; m1; m1 = m1->next)
1447 if (m != m1 && m1->match == 0 && VARRAY_INT (n_times_set, m1->regno) == 1
1448 /* A reg used outside the loop mustn't be eliminated. */
1450 /* A reg used for zero-extending mustn't be eliminated. */
1452 && (matched_regs[m1->regno]
1455 /* Can combine regs with different modes loaded from the
1456 same constant only if the modes are the same or
1457 if both are integer modes with M wider or the same
1458 width as M1. The check for integer is redundant, but
1459 safe, since the only case of differing destination
1460 modes with equal sources is when both sources are
1461 VOIDmode, i.e., CONST_INT. */
1462 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1463 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1464 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1465 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1466 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1467 /* See if the source of M1 says it matches M. */
1468 && ((GET_CODE (m1->set_src) == REG
1469 && matched_regs[REGNO (m1->set_src)])
1470 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1472 && ((m->dependencies == m1->dependencies)
1473 || rtx_equal_p (m->dependencies, m1->dependencies)))
1475 m->lifetime += m1->lifetime;
1476 m->savings += m1->savings;
1479 matched_regs[m1->regno] = 1;
1483 /* Now combine the regs used for zero-extension.
1484 This can be done for those not marked `global'
1485 provided their lives don't overlap. */
1487 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1488 mode = GET_MODE_WIDER_MODE (mode))
1490 register struct movable *m0 = 0;
1492 /* Combine all the registers for extension from mode MODE.
1493 Don't combine any that are used outside this loop. */
1494 for (m = movables; m; m = m->next)
1495 if (m->partial && ! m->global
1496 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1498 register struct movable *m1;
1499 int first = uid_luid[REGNO_FIRST_UID (m->regno)];
1500 int last = uid_luid[REGNO_LAST_UID (m->regno)];
1504 /* First one: don't check for overlap, just record it. */
1509 /* Make sure they extend to the same mode.
1510 (Almost always true.) */
1511 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1514 /* We already have one: check for overlap with those
1515 already combined together. */
1516 for (m1 = movables; m1 != m; m1 = m1->next)
1517 if (m1 == m0 || (m1->partial && m1->match == m0))
1518 if (! (uid_luid[REGNO_FIRST_UID (m1->regno)] > last
1519 || uid_luid[REGNO_LAST_UID (m1->regno)] < first))
1522 /* No overlap: we can combine this with the others. */
1523 m0->lifetime += m->lifetime;
1524 m0->savings += m->savings;
1533 free (matched_regs);
1536 /* Return 1 if regs X and Y will become the same if moved. */
1539 regs_match_p (x, y, movables)
1541 struct movable *movables;
1545 struct movable *mx, *my;
1547 for (mx = movables; mx; mx = mx->next)
1548 if (mx->regno == xn)
1551 for (my = movables; my; my = my->next)
1552 if (my->regno == yn)
1556 && ((mx->match == my->match && mx->match != 0)
1558 || mx == my->match));
1561 /* Return 1 if X and Y are identical-looking rtx's.
1562 This is the Lisp function EQUAL for rtx arguments.
1564 If two registers are matching movables or a movable register and an
1565 equivalent constant, consider them equal. */
1568 rtx_equal_for_loop_p (x, y, movables)
1570 struct movable *movables;
1574 register struct movable *m;
1575 register enum rtx_code code;
1576 register const char *fmt;
1580 if (x == 0 || y == 0)
1583 code = GET_CODE (x);
1585 /* If we have a register and a constant, they may sometimes be
1587 if (GET_CODE (x) == REG && VARRAY_INT (set_in_loop, REGNO (x)) == -2
1590 for (m = movables; m; m = m->next)
1591 if (m->move_insn && m->regno == REGNO (x)
1592 && rtx_equal_p (m->set_src, y))
1595 else if (GET_CODE (y) == REG && VARRAY_INT (set_in_loop, REGNO (y)) == -2
1598 for (m = movables; m; m = m->next)
1599 if (m->move_insn && m->regno == REGNO (y)
1600 && rtx_equal_p (m->set_src, x))
1604 /* Otherwise, rtx's of different codes cannot be equal. */
1605 if (code != GET_CODE (y))
1608 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1609 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1611 if (GET_MODE (x) != GET_MODE (y))
1614 /* These three types of rtx's can be compared nonrecursively. */
1616 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1618 if (code == LABEL_REF)
1619 return XEXP (x, 0) == XEXP (y, 0);
1620 if (code == SYMBOL_REF)
1621 return XSTR (x, 0) == XSTR (y, 0);
1623 /* Compare the elements. If any pair of corresponding elements
1624 fail to match, return 0 for the whole things. */
1626 fmt = GET_RTX_FORMAT (code);
1627 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1632 if (XWINT (x, i) != XWINT (y, i))
1637 if (XINT (x, i) != XINT (y, i))
1642 /* Two vectors must have the same length. */
1643 if (XVECLEN (x, i) != XVECLEN (y, i))
1646 /* And the corresponding elements must match. */
1647 for (j = 0; j < XVECLEN (x, i); j++)
1648 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1653 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1658 if (strcmp (XSTR (x, i), XSTR (y, i)))
1663 /* These are just backpointers, so they don't matter. */
1669 /* It is believed that rtx's at this level will never
1670 contain anything but integers and other rtx's,
1671 except for within LABEL_REFs and SYMBOL_REFs. */
1679 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1680 insns in INSNS which use the reference. */
1683 add_label_notes (x, insns)
1687 enum rtx_code code = GET_CODE (x);
1692 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1694 /* This code used to ignore labels that referred to dispatch tables to
1695 avoid flow generating (slighly) worse code.
1697 We no longer ignore such label references (see LABEL_REF handling in
1698 mark_jump_label for additional information). */
1699 for (insn = insns; insn; insn = NEXT_INSN (insn))
1700 if (reg_mentioned_p (XEXP (x, 0), insn))
1701 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
1705 fmt = GET_RTX_FORMAT (code);
1706 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1709 add_label_notes (XEXP (x, i), insns);
1710 else if (fmt[i] == 'E')
1711 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1712 add_label_notes (XVECEXP (x, i, j), insns);
1716 /* Scan MOVABLES, and move the insns that deserve to be moved.
1717 If two matching movables are combined, replace one reg with the
1718 other throughout. */
1721 move_movables (loop, movables, threshold, insn_count, nregs)
1723 struct movable *movables;
1729 register struct movable *m;
1731 rtx loop_start = loop->start;
1732 rtx loop_end = loop->end;
1733 /* Map of pseudo-register replacements to handle combining
1734 when we move several insns that load the same value
1735 into different pseudo-registers. */
1736 rtx *reg_map = (rtx *) xcalloc (nregs, sizeof (rtx));
1737 char *already_moved = (char *) xcalloc (nregs, sizeof (char));
1741 for (m = movables; m; m = m->next)
1743 /* Describe this movable insn. */
1745 if (loop_dump_stream)
1747 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1748 INSN_UID (m->insn), m->regno, m->lifetime);
1750 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1752 fprintf (loop_dump_stream, "cond ");
1754 fprintf (loop_dump_stream, "force ");
1756 fprintf (loop_dump_stream, "global ");
1758 fprintf (loop_dump_stream, "done ");
1760 fprintf (loop_dump_stream, "move-insn ");
1762 fprintf (loop_dump_stream, "matches %d ",
1763 INSN_UID (m->match->insn));
1765 fprintf (loop_dump_stream, "forces %d ",
1766 INSN_UID (m->forces->insn));
1769 /* Count movables. Value used in heuristics in strength_reduce. */
1772 /* Ignore the insn if it's already done (it matched something else).
1773 Otherwise, see if it is now safe to move. */
1777 || (1 == loop_invariant_p (loop, m->set_src)
1778 && (m->dependencies == 0
1779 || 1 == loop_invariant_p (loop, m->dependencies))
1781 || 1 == consec_sets_invariant_p (loop, m->set_dest,
1784 && (! m->forces || m->forces->done))
1788 int savings = m->savings;
1790 /* We have an insn that is safe to move.
1791 Compute its desirability. */
1796 if (loop_dump_stream)
1797 fprintf (loop_dump_stream, "savings %d ", savings);
1799 if (moved_once[regno] && loop_dump_stream)
1800 fprintf (loop_dump_stream, "halved since already moved ");
1802 /* An insn MUST be moved if we already moved something else
1803 which is safe only if this one is moved too: that is,
1804 if already_moved[REGNO] is nonzero. */
1806 /* An insn is desirable to move if the new lifetime of the
1807 register is no more than THRESHOLD times the old lifetime.
1808 If it's not desirable, it means the loop is so big
1809 that moving won't speed things up much,
1810 and it is liable to make register usage worse. */
1812 /* It is also desirable to move if it can be moved at no
1813 extra cost because something else was already moved. */
1815 if (already_moved[regno]
1816 || flag_move_all_movables
1817 || (threshold * savings * m->lifetime) >=
1818 (moved_once[regno] ? insn_count * 2 : insn_count)
1819 || (m->forces && m->forces->done
1820 && VARRAY_INT (n_times_set, m->forces->regno) == 1))
1823 register struct movable *m1;
1824 rtx first = NULL_RTX;
1826 /* Now move the insns that set the reg. */
1828 if (m->partial && m->match)
1832 /* Find the end of this chain of matching regs.
1833 Thus, we load each reg in the chain from that one reg.
1834 And that reg is loaded with 0 directly,
1835 since it has ->match == 0. */
1836 for (m1 = m; m1->match; m1 = m1->match);
1837 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1838 SET_DEST (PATTERN (m1->insn)));
1839 i1 = emit_insn_before (newpat, loop_start);
1841 /* Mark the moved, invariant reg as being allowed to
1842 share a hard reg with the other matching invariant. */
1843 REG_NOTES (i1) = REG_NOTES (m->insn);
1844 r1 = SET_DEST (PATTERN (m->insn));
1845 r2 = SET_DEST (PATTERN (m1->insn));
1847 = gen_rtx_EXPR_LIST (VOIDmode, r1,
1848 gen_rtx_EXPR_LIST (VOIDmode, r2,
1850 delete_insn (m->insn);
1855 if (loop_dump_stream)
1856 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1858 /* If we are to re-generate the item being moved with a
1859 new move insn, first delete what we have and then emit
1860 the move insn before the loop. */
1861 else if (m->move_insn)
1865 for (count = m->consec; count >= 0; count--)
1867 /* If this is the first insn of a library call sequence,
1869 if (GET_CODE (p) != NOTE
1870 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1873 /* If this is the last insn of a libcall sequence, then
1874 delete every insn in the sequence except the last.
1875 The last insn is handled in the normal manner. */
1876 if (GET_CODE (p) != NOTE
1877 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1879 temp = XEXP (temp, 0);
1881 temp = delete_insn (temp);
1885 p = delete_insn (p);
1887 /* simplify_giv_expr expects that it can walk the insns
1888 at m->insn forwards and see this old sequence we are
1889 tossing here. delete_insn does preserve the next
1890 pointers, but when we skip over a NOTE we must fix
1891 it up. Otherwise that code walks into the non-deleted
1893 while (p && GET_CODE (p) == NOTE)
1894 p = NEXT_INSN (temp) = NEXT_INSN (p);
1898 emit_move_insn (m->set_dest, m->set_src);
1899 temp = get_insns ();
1902 add_label_notes (m->set_src, temp);
1904 i1 = emit_insns_before (temp, loop_start);
1905 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1907 = gen_rtx_EXPR_LIST (m->is_equiv ? REG_EQUIV : REG_EQUAL,
1908 m->set_src, REG_NOTES (i1));
1910 if (loop_dump_stream)
1911 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1913 /* The more regs we move, the less we like moving them. */
1918 for (count = m->consec; count >= 0; count--)
1922 /* If first insn of libcall sequence, skip to end. */
1923 /* Do this at start of loop, since p is guaranteed to
1925 if (GET_CODE (p) != NOTE
1926 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1929 /* If last insn of libcall sequence, move all
1930 insns except the last before the loop. The last
1931 insn is handled in the normal manner. */
1932 if (GET_CODE (p) != NOTE
1933 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1937 rtx fn_address_insn = 0;
1940 for (temp = XEXP (temp, 0); temp != p;
1941 temp = NEXT_INSN (temp))
1947 if (GET_CODE (temp) == NOTE)
1950 body = PATTERN (temp);
1952 /* Find the next insn after TEMP,
1953 not counting USE or NOTE insns. */
1954 for (next = NEXT_INSN (temp); next != p;
1955 next = NEXT_INSN (next))
1956 if (! (GET_CODE (next) == INSN
1957 && GET_CODE (PATTERN (next)) == USE)
1958 && GET_CODE (next) != NOTE)
1961 /* If that is the call, this may be the insn
1962 that loads the function address.
1964 Extract the function address from the insn
1965 that loads it into a register.
1966 If this insn was cse'd, we get incorrect code.
1968 So emit a new move insn that copies the
1969 function address into the register that the
1970 call insn will use. flow.c will delete any
1971 redundant stores that we have created. */
1972 if (GET_CODE (next) == CALL_INSN
1973 && GET_CODE (body) == SET
1974 && GET_CODE (SET_DEST (body)) == REG
1975 && (n = find_reg_note (temp, REG_EQUAL,
1978 fn_reg = SET_SRC (body);
1979 if (GET_CODE (fn_reg) != REG)
1980 fn_reg = SET_DEST (body);
1981 fn_address = XEXP (n, 0);
1982 fn_address_insn = temp;
1984 /* We have the call insn.
1985 If it uses the register we suspect it might,
1986 load it with the correct address directly. */
1987 if (GET_CODE (temp) == CALL_INSN
1989 && reg_referenced_p (fn_reg, body))
1990 emit_insn_after (gen_move_insn (fn_reg,
1994 if (GET_CODE (temp) == CALL_INSN)
1996 i1 = emit_call_insn_before (body, loop_start);
1997 /* Because the USAGE information potentially
1998 contains objects other than hard registers
1999 we need to copy it. */
2000 if (CALL_INSN_FUNCTION_USAGE (temp))
2001 CALL_INSN_FUNCTION_USAGE (i1)
2002 = copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
2005 i1 = emit_insn_before (body, loop_start);
2008 if (temp == fn_address_insn)
2009 fn_address_insn = i1;
2010 REG_NOTES (i1) = REG_NOTES (temp);
2016 if (m->savemode != VOIDmode)
2018 /* P sets REG to zero; but we should clear only
2019 the bits that are not covered by the mode
2021 rtx reg = m->set_dest;
2027 (GET_MODE (reg), and_optab, reg,
2028 GEN_INT ((((HOST_WIDE_INT) 1
2029 << GET_MODE_BITSIZE (m->savemode)))
2031 reg, 1, OPTAB_LIB_WIDEN);
2035 emit_move_insn (reg, tem);
2036 sequence = gen_sequence ();
2038 i1 = emit_insn_before (sequence, loop_start);
2040 else if (GET_CODE (p) == CALL_INSN)
2042 i1 = emit_call_insn_before (PATTERN (p), loop_start);
2043 /* Because the USAGE information potentially
2044 contains objects other than hard registers
2045 we need to copy it. */
2046 if (CALL_INSN_FUNCTION_USAGE (p))
2047 CALL_INSN_FUNCTION_USAGE (i1)
2048 = copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
2050 else if (count == m->consec && m->move_insn_first)
2052 /* The SET_SRC might not be invariant, so we must
2053 use the REG_EQUAL note. */
2055 emit_move_insn (m->set_dest, m->set_src);
2056 temp = get_insns ();
2059 add_label_notes (m->set_src, temp);
2061 i1 = emit_insns_before (temp, loop_start);
2062 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
2064 = gen_rtx_EXPR_LIST ((m->is_equiv ? REG_EQUIV
2066 m->set_src, REG_NOTES (i1));
2069 i1 = emit_insn_before (PATTERN (p), loop_start);
2071 if (REG_NOTES (i1) == 0)
2073 REG_NOTES (i1) = REG_NOTES (p);
2075 /* If there is a REG_EQUAL note present whose value
2076 is not loop invariant, then delete it, since it
2077 may cause problems with later optimization passes.
2078 It is possible for cse to create such notes
2079 like this as a result of record_jump_cond. */
2081 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
2082 && ! loop_invariant_p (loop, XEXP (temp, 0)))
2083 remove_note (i1, temp);
2089 if (loop_dump_stream)
2090 fprintf (loop_dump_stream, " moved to %d",
2093 /* If library call, now fix the REG_NOTES that contain
2094 insn pointers, namely REG_LIBCALL on FIRST
2095 and REG_RETVAL on I1. */
2096 if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX)))
2098 XEXP (temp, 0) = first;
2099 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
2100 XEXP (temp, 0) = i1;
2107 /* simplify_giv_expr expects that it can walk the insns
2108 at m->insn forwards and see this old sequence we are
2109 tossing here. delete_insn does preserve the next
2110 pointers, but when we skip over a NOTE we must fix
2111 it up. Otherwise that code walks into the non-deleted
2113 while (p && GET_CODE (p) == NOTE)
2114 p = NEXT_INSN (temp) = NEXT_INSN (p);
2117 /* The more regs we move, the less we like moving them. */
2121 /* Any other movable that loads the same register
2123 already_moved[regno] = 1;
2125 /* This reg has been moved out of one loop. */
2126 moved_once[regno] = 1;
2128 /* The reg set here is now invariant. */
2130 VARRAY_INT (set_in_loop, regno) = 0;
2134 /* Change the length-of-life info for the register
2135 to say it lives at least the full length of this loop.
2136 This will help guide optimizations in outer loops. */
2138 if (uid_luid[REGNO_FIRST_UID (regno)] > INSN_LUID (loop_start))
2139 /* This is the old insn before all the moved insns.
2140 We can't use the moved insn because it is out of range
2141 in uid_luid. Only the old insns have luids. */
2142 REGNO_FIRST_UID (regno) = INSN_UID (loop_start);
2143 if (uid_luid[REGNO_LAST_UID (regno)] < INSN_LUID (loop_end))
2144 REGNO_LAST_UID (regno) = INSN_UID (loop_end);
2146 /* Combine with this moved insn any other matching movables. */
2149 for (m1 = movables; m1; m1 = m1->next)
2154 /* Schedule the reg loaded by M1
2155 for replacement so that shares the reg of M.
2156 If the modes differ (only possible in restricted
2157 circumstances, make a SUBREG.
2159 Note this assumes that the target dependent files
2160 treat REG and SUBREG equally, including within
2161 GO_IF_LEGITIMATE_ADDRESS and in all the
2162 predicates since we never verify that replacing the
2163 original register with a SUBREG results in a
2164 recognizable insn. */
2165 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
2166 reg_map[m1->regno] = m->set_dest;
2169 = gen_lowpart_common (GET_MODE (m1->set_dest),
2172 /* Get rid of the matching insn
2173 and prevent further processing of it. */
2176 /* if library call, delete all insn except last, which
2178 if ((temp = find_reg_note (m1->insn, REG_RETVAL,
2181 for (temp = XEXP (temp, 0); temp != m1->insn;
2182 temp = NEXT_INSN (temp))
2185 delete_insn (m1->insn);
2187 /* Any other movable that loads the same register
2189 already_moved[m1->regno] = 1;
2191 /* The reg merged here is now invariant,
2192 if the reg it matches is invariant. */
2194 VARRAY_INT (set_in_loop, m1->regno) = 0;
2197 else if (loop_dump_stream)
2198 fprintf (loop_dump_stream, "not desirable");
2200 else if (loop_dump_stream && !m->match)
2201 fprintf (loop_dump_stream, "not safe");
2203 if (loop_dump_stream)
2204 fprintf (loop_dump_stream, "\n");
2208 new_start = loop_start;
2210 /* Go through all the instructions in the loop, making
2211 all the register substitutions scheduled in REG_MAP. */
2212 for (p = new_start; p != loop_end; p = NEXT_INSN (p))
2213 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
2214 || GET_CODE (p) == CALL_INSN)
2216 replace_regs (PATTERN (p), reg_map, nregs, 0);
2217 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
2223 free (already_moved);
2227 /* Scan X and replace the address of any MEM in it with ADDR.
2228 REG is the address that MEM should have before the replacement. */
2231 replace_call_address (x, reg, addr)
2234 register enum rtx_code code;
2236 register const char *fmt;
2240 code = GET_CODE (x);
2254 /* Short cut for very common case. */
2255 replace_call_address (XEXP (x, 1), reg, addr);
2259 /* Short cut for very common case. */
2260 replace_call_address (XEXP (x, 0), reg, addr);
2264 /* If this MEM uses a reg other than the one we expected,
2265 something is wrong. */
2266 if (XEXP (x, 0) != reg)
2275 fmt = GET_RTX_FORMAT (code);
2276 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2279 replace_call_address (XEXP (x, i), reg, addr);
2280 else if (fmt[i] == 'E')
2283 for (j = 0; j < XVECLEN (x, i); j++)
2284 replace_call_address (XVECEXP (x, i, j), reg, addr);
2290 /* Return the number of memory refs to addresses that vary
2294 count_nonfixed_reads (loop, x)
2295 const struct loop *loop;
2298 register enum rtx_code code;
2300 register const char *fmt;
2306 code = GET_CODE (x);
2320 return ((loop_invariant_p (loop, XEXP (x, 0)) != 1)
2321 + count_nonfixed_reads (loop, XEXP (x, 0)));
2328 fmt = GET_RTX_FORMAT (code);
2329 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2332 value += count_nonfixed_reads (loop, XEXP (x, i));
2336 for (j = 0; j < XVECLEN (x, i); j++)
2337 value += count_nonfixed_reads (loop, XVECEXP (x, i, j));
2345 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2346 Replace it with an instruction to load just the low bytes
2347 if the machine supports such an instruction,
2348 and insert above LOOP_START an instruction to clear the register. */
2351 constant_high_bytes (p, loop_start)
2355 register int insn_code_number;
2357 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2358 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2363 gen_rtx_STRICT_LOW_PART
2365 gen_rtx_SUBREG (GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2366 SET_DEST (PATTERN (p)), 0)),
2367 XEXP (SET_SRC (PATTERN (p)), 0));
2369 insn_code_number = recog (new, p);
2371 if (insn_code_number)
2375 /* Clear destination register before the loop. */
2376 emit_insn_before (gen_rtx_SET (VOIDmode,
2377 SET_DEST (PATTERN (p)), const0_rtx),
2380 /* Inside the loop, just load the low part. */
2386 /* Scan a loop setting the elements `cont', `vtop', `loops_enclosed',
2387 `has_call', `has_volatile', and `has_tablejump' within LOOP.
2388 Set the global variables `unknown_address_altered',
2389 `unknown_constant_address_altered', and `num_mem_sets'. Also, fill
2390 in the array `loop_mems' and the list `loop_store_mems'. */
2396 register int level = 1;
2398 struct loop_info *loop_info = LOOP_INFO (loop);
2399 rtx start = loop->start;
2400 rtx end = loop->end;
2401 /* The label after END. Jumping here is just like falling off the
2402 end of the loop. We use next_nonnote_insn instead of next_label
2403 as a hedge against the (pathological) case where some actual insn
2404 might end up between the two. */
2405 rtx exit_target = next_nonnote_insn (end);
2407 loop_info->has_indirect_jump = indirect_jump_in_function;
2408 loop_info->has_call = 0;
2409 loop_info->has_volatile = 0;
2410 loop_info->has_tablejump = 0;
2411 loop_info->has_multiple_exit_targets = 0;
2416 unknown_address_altered = 0;
2417 unknown_constant_address_altered = 0;
2418 loop_store_mems = NULL_RTX;
2419 first_loop_store_insn = NULL_RTX;
2423 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2424 insn = NEXT_INSN (insn))
2426 if (GET_CODE (insn) == NOTE)
2428 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2431 /* Count number of loops contained in this one. */
2434 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2443 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2448 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_VTOP)
2450 /* If there is a NOTE_INSN_LOOP_VTOP, then this is a for
2451 or while style loop, with a loop exit test at the
2452 start. Thus, we can assume that the loop condition
2453 was true when the loop was entered. */
2458 else if (GET_CODE (insn) == CALL_INSN)
2460 if (! CONST_CALL_P (insn))
2461 unknown_address_altered = 1;
2462 loop_info->has_call = 1;
2464 else if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2466 rtx label1 = NULL_RTX;
2467 rtx label2 = NULL_RTX;
2469 if (volatile_refs_p (PATTERN (insn)))
2470 loop_info->has_volatile = 1;
2472 if (GET_CODE (insn) == JUMP_INSN
2473 && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
2474 || GET_CODE (PATTERN (insn)) == ADDR_VEC))
2475 loop_info->has_tablejump = 1;
2477 note_stores (PATTERN (insn), note_addr_stored, NULL);
2478 if (! first_loop_store_insn && loop_store_mems)
2479 first_loop_store_insn = insn;
2481 if (! loop_info->has_multiple_exit_targets
2482 && GET_CODE (insn) == JUMP_INSN
2483 && GET_CODE (PATTERN (insn)) == SET
2484 && SET_DEST (PATTERN (insn)) == pc_rtx)
2486 if (GET_CODE (SET_SRC (PATTERN (insn))) == IF_THEN_ELSE)
2488 label1 = XEXP (SET_SRC (PATTERN (insn)), 1);
2489 label2 = XEXP (SET_SRC (PATTERN (insn)), 2);
2493 label1 = SET_SRC (PATTERN (insn));
2497 if (label1 && label1 != pc_rtx)
2499 if (GET_CODE (label1) != LABEL_REF)
2501 /* Something tricky. */
2502 loop_info->has_multiple_exit_targets = 1;
2505 else if (XEXP (label1, 0) != exit_target
2506 && LABEL_OUTSIDE_LOOP_P (label1))
2508 /* A jump outside the current loop. */
2509 loop_info->has_multiple_exit_targets = 1;
2519 else if (GET_CODE (insn) == RETURN)
2520 loop_info->has_multiple_exit_targets = 1;
2523 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2524 if (/* We can't tell what MEMs are aliased by what. */
2525 ! unknown_address_altered
2526 /* An exception thrown by a called function might land us
2528 && ! loop_info->has_call
2529 /* We don't want loads for MEMs moved to a location before the
2530 one at which their stack memory becomes allocated. (Note
2531 that this is not a problem for malloc, etc., since those
2532 require actual function calls. */
2533 && ! current_function_calls_alloca
2534 /* There are ways to leave the loop other than falling off the
2536 && ! loop_info->has_multiple_exit_targets)
2537 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2538 insn = NEXT_INSN (insn))
2539 for_each_rtx (&insn, insert_loop_mem, 0);
2542 /* LOOP->CONT_DOMINATOR is now the last label between the loop start
2543 and the continue note that is a the destination of a (cond)jump after
2544 the continue note. If there is any (cond)jump between the loop start
2545 and what we have so far as LOOP->CONT_DOMINATOR that has a
2546 target between LOOP->DOMINATOR and the continue note, move
2547 LOOP->CONT_DOMINATOR forward to that label; if a jump's
2548 destination cannot be determined, clear LOOP->CONT_DOMINATOR. */
2551 verify_dominator (loop)
2556 if (! loop->cont_dominator)
2557 /* This can happen for an empty loop, e.g. in
2558 gcc.c-torture/compile/920410-2.c */
2560 if (loop->cont_dominator == const0_rtx)
2562 loop->cont_dominator = 0;
2565 for (insn = loop->start; insn != loop->cont_dominator;
2566 insn = NEXT_INSN (insn))
2568 if (GET_CODE (insn) == JUMP_INSN
2569 && GET_CODE (PATTERN (insn)) != RETURN)
2571 rtx label = JUMP_LABEL (insn);
2574 /* If it is not a jump we can easily understand or for
2575 which we do not have jump target information in the JUMP_LABEL
2576 field (consider ADDR_VEC and ADDR_DIFF_VEC insns), then clear
2577 LOOP->CONT_DOMINATOR. */
2578 if ((! condjump_p (insn)
2579 && ! condjump_in_parallel_p (insn))
2580 || label == NULL_RTX)
2582 loop->cont_dominator = NULL_RTX;
2586 label_luid = INSN_LUID (label);
2587 if (label_luid < INSN_LUID (loop->cont)
2589 > INSN_LUID (loop->cont)))
2590 loop->cont_dominator = label;
2595 /* Scan the function looking for loops. Record the start and end of each loop.
2596 Also mark as invalid loops any loops that contain a setjmp or are branched
2597 to from outside the loop. */
2600 find_and_verify_loops (f, loops)
2602 struct loops *loops;
2607 struct loop *current_loop;
2608 struct loop *next_loop;
2611 num_loops = loops->num;
2613 compute_luids (f, NULL_RTX, 0);
2615 /* If there are jumps to undefined labels,
2616 treat them as jumps out of any/all loops.
2617 This also avoids writing past end of tables when there are no loops. */
2620 /* Find boundaries of loops, mark which loops are contained within
2621 loops, and invalidate loops that have setjmp. */
2624 current_loop = NULL;
2625 for (insn = f; insn; insn = NEXT_INSN (insn))
2627 if (GET_CODE (insn) == NOTE)
2628 switch (NOTE_LINE_NUMBER (insn))
2630 case NOTE_INSN_LOOP_BEG:
2631 next_loop = loops->array + num_loops;
2632 next_loop->num = num_loops;
2634 next_loop->start = insn;
2635 next_loop->outer = current_loop;
2636 current_loop = next_loop;
2639 case NOTE_INSN_SETJMP:
2640 /* In this case, we must invalidate our current loop and any
2642 for (loop = current_loop; loop; loop = loop->outer)
2645 if (loop_dump_stream)
2646 fprintf (loop_dump_stream,
2647 "\nLoop at %d ignored due to setjmp.\n",
2648 INSN_UID (loop->start));
2652 case NOTE_INSN_LOOP_CONT:
2653 current_loop->cont = insn;
2655 case NOTE_INSN_LOOP_END:
2659 current_loop->end = insn;
2660 verify_dominator (current_loop);
2661 current_loop = current_loop->outer;
2667 /* If for any loop, this is a jump insn between the NOTE_INSN_LOOP_CONT
2668 and NOTE_INSN_LOOP_END notes, update loop->dominator. */
2669 else if (GET_CODE (insn) == JUMP_INSN
2670 && GET_CODE (PATTERN (insn)) != RETURN
2673 rtx label = JUMP_LABEL (insn);
2675 if (! condjump_p (insn) && ! condjump_in_parallel_p (insn))
2678 loop = current_loop;
2681 /* First see if we care about this loop. */
2682 if (loop->cont && loop->cont_dominator != const0_rtx)
2684 /* If the jump destination is not known, invalidate
2685 loop->const_dominator. */
2687 loop->cont_dominator = const0_rtx;
2689 /* Check if the destination is between loop start and
2691 if ((INSN_LUID (label)
2692 < INSN_LUID (loop->cont))
2693 && (INSN_LUID (label)
2694 > INSN_LUID (loop->start))
2695 /* And if there is no later destination already
2697 && (! loop->cont_dominator
2698 || (INSN_LUID (label)
2699 > INSN_LUID (loop->cont_dominator))))
2700 loop->cont_dominator = label;
2707 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2708 enclosing loop, but this doesn't matter. */
2709 uid_loop[INSN_UID (insn)] = current_loop;
2712 /* Any loop containing a label used in an initializer must be invalidated,
2713 because it can be jumped into from anywhere. */
2715 for (label = forced_labels; label; label = XEXP (label, 1))
2717 for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
2718 loop; loop = loop->outer)
2722 /* Any loop containing a label used for an exception handler must be
2723 invalidated, because it can be jumped into from anywhere. */
2725 for (label = exception_handler_labels; label; label = XEXP (label, 1))
2727 for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
2728 loop; loop = loop->outer)
2732 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2733 loop that it is not contained within, that loop is marked invalid.
2734 If any INSN or CALL_INSN uses a label's address, then the loop containing
2735 that label is marked invalid, because it could be jumped into from
2738 Also look for blocks of code ending in an unconditional branch that
2739 exits the loop. If such a block is surrounded by a conditional
2740 branch around the block, move the block elsewhere (see below) and
2741 invert the jump to point to the code block. This may eliminate a
2742 label in our loop and will simplify processing by both us and a
2743 possible second cse pass. */
2745 for (insn = f; insn; insn = NEXT_INSN (insn))
2746 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2748 struct loop *this_loop = uid_loop[INSN_UID (insn)];
2750 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2752 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2755 for (loop = uid_loop[INSN_UID (XEXP (note, 0))];
2756 loop; loop = loop->outer)
2761 if (GET_CODE (insn) != JUMP_INSN)
2764 mark_loop_jump (PATTERN (insn), this_loop);
2766 /* See if this is an unconditional branch outside the loop. */
2768 && (GET_CODE (PATTERN (insn)) == RETURN
2769 || (simplejump_p (insn)
2770 && (uid_loop[INSN_UID (JUMP_LABEL (insn))]
2772 && get_max_uid () < max_uid_for_loop)
2775 rtx our_next = next_real_insn (insn);
2776 rtx last_insn_to_move = NEXT_INSN (insn);
2777 struct loop *dest_loop;
2778 struct loop *outer_loop = NULL;
2780 /* Go backwards until we reach the start of the loop, a label,
2782 for (p = PREV_INSN (insn);
2783 GET_CODE (p) != CODE_LABEL
2784 && ! (GET_CODE (p) == NOTE
2785 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2786 && GET_CODE (p) != JUMP_INSN;
2790 /* Check for the case where we have a jump to an inner nested
2791 loop, and do not perform the optimization in that case. */
2793 if (JUMP_LABEL (insn))
2795 dest_loop = uid_loop[INSN_UID (JUMP_LABEL (insn))];
2798 for (outer_loop = dest_loop; outer_loop;
2799 outer_loop = outer_loop->outer)
2800 if (outer_loop == this_loop)
2805 /* Make sure that the target of P is within the current loop. */
2807 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2808 && uid_loop[INSN_UID (JUMP_LABEL (p))] != this_loop)
2809 outer_loop = this_loop;
2811 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2812 we have a block of code to try to move.
2814 We look backward and then forward from the target of INSN
2815 to find a BARRIER at the same loop depth as the target.
2816 If we find such a BARRIER, we make a new label for the start
2817 of the block, invert the jump in P and point it to that label,
2818 and move the block of code to the spot we found. */
2821 && GET_CODE (p) == JUMP_INSN
2822 && JUMP_LABEL (p) != 0
2823 /* Just ignore jumps to labels that were never emitted.
2824 These always indicate compilation errors. */
2825 && INSN_UID (JUMP_LABEL (p)) != 0
2827 && ! simplejump_p (p)
2828 && next_real_insn (JUMP_LABEL (p)) == our_next
2829 /* If it's not safe to move the sequence, then we
2831 && insns_safe_to_move_p (p, NEXT_INSN (insn),
2832 &last_insn_to_move))
2835 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2836 struct loop *target_loop = uid_loop[INSN_UID (target)];
2839 for (loc = target; loc; loc = PREV_INSN (loc))
2840 if (GET_CODE (loc) == BARRIER
2841 /* Don't move things inside a tablejump. */
2842 && ((loc2 = next_nonnote_insn (loc)) == 0
2843 || GET_CODE (loc2) != CODE_LABEL
2844 || (loc2 = next_nonnote_insn (loc2)) == 0
2845 || GET_CODE (loc2) != JUMP_INSN
2846 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2847 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2848 && uid_loop[INSN_UID (loc)] == target_loop)
2852 for (loc = target; loc; loc = NEXT_INSN (loc))
2853 if (GET_CODE (loc) == BARRIER
2854 /* Don't move things inside a tablejump. */
2855 && ((loc2 = next_nonnote_insn (loc)) == 0
2856 || GET_CODE (loc2) != CODE_LABEL
2857 || (loc2 = next_nonnote_insn (loc2)) == 0
2858 || GET_CODE (loc2) != JUMP_INSN
2859 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2860 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2861 && uid_loop[INSN_UID (loc)] == target_loop)
2866 rtx cond_label = JUMP_LABEL (p);
2867 rtx new_label = get_label_after (p);
2869 /* Ensure our label doesn't go away. */
2870 LABEL_NUSES (cond_label)++;
2872 /* Verify that uid_loop is large enough and that
2874 if (invert_jump (p, new_label))
2878 /* If no suitable BARRIER was found, create a suitable
2879 one before TARGET. Since TARGET is a fall through
2880 path, we'll need to insert an jump around our block
2881 and a add a BARRIER before TARGET.
2883 This creates an extra unconditional jump outside
2884 the loop. However, the benefits of removing rarely
2885 executed instructions from inside the loop usually
2886 outweighs the cost of the extra unconditional jump
2887 outside the loop. */
2892 temp = gen_jump (JUMP_LABEL (insn));
2893 temp = emit_jump_insn_before (temp, target);
2894 JUMP_LABEL (temp) = JUMP_LABEL (insn);
2895 LABEL_NUSES (JUMP_LABEL (insn))++;
2896 loc = emit_barrier_before (target);
2899 /* Include the BARRIER after INSN and copy the
2901 new_label = squeeze_notes (new_label,
2903 reorder_insns (new_label, last_insn_to_move, loc);
2905 /* All those insns are now in TARGET_LOOP. */
2907 q != NEXT_INSN (last_insn_to_move);
2909 uid_loop[INSN_UID (q)] = target_loop;
2911 /* The label jumped to by INSN is no longer a loop exit.
2912 Unless INSN does not have a label (e.g., it is a
2913 RETURN insn), search loop->exit_labels to find
2914 its label_ref, and remove it. Also turn off
2915 LABEL_OUTSIDE_LOOP_P bit. */
2916 if (JUMP_LABEL (insn))
2919 r = this_loop->exit_labels;
2920 r; q = r, r = LABEL_NEXTREF (r))
2921 if (XEXP (r, 0) == JUMP_LABEL (insn))
2923 LABEL_OUTSIDE_LOOP_P (r) = 0;
2925 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2927 this_loop->exit_labels = LABEL_NEXTREF (r);
2931 for (loop = this_loop; loop && loop != target_loop;
2935 /* If we didn't find it, then something is
2941 /* P is now a jump outside the loop, so it must be put
2942 in loop->exit_labels, and marked as such.
2943 The easiest way to do this is to just call
2944 mark_loop_jump again for P. */
2945 mark_loop_jump (PATTERN (p), this_loop);
2947 /* If INSN now jumps to the insn after it,
2949 if (JUMP_LABEL (insn) != 0
2950 && (next_real_insn (JUMP_LABEL (insn))
2951 == next_real_insn (insn)))
2955 /* Continue the loop after where the conditional
2956 branch used to jump, since the only branch insn
2957 in the block (if it still remains) is an inter-loop
2958 branch and hence needs no processing. */
2959 insn = NEXT_INSN (cond_label);
2961 if (--LABEL_NUSES (cond_label) == 0)
2962 delete_insn (cond_label);
2964 /* This loop will be continued with NEXT_INSN (insn). */
2965 insn = PREV_INSN (insn);
2972 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2973 loops it is contained in, mark the target loop invalid.
2975 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2978 mark_loop_jump (x, loop)
2982 struct loop *dest_loop;
2983 struct loop *outer_loop;
2986 switch (GET_CODE (x))
2999 /* There could be a label reference in here. */
3000 mark_loop_jump (XEXP (x, 0), loop);
3006 mark_loop_jump (XEXP (x, 0), loop);
3007 mark_loop_jump (XEXP (x, 1), loop);
3011 /* This may refer to a LABEL_REF or SYMBOL_REF. */
3012 mark_loop_jump (XEXP (x, 1), loop);
3017 mark_loop_jump (XEXP (x, 0), loop);
3021 dest_loop = uid_loop[INSN_UID (XEXP (x, 0))];
3023 /* Link together all labels that branch outside the loop. This
3024 is used by final_[bg]iv_value and the loop unrolling code. Also
3025 mark this LABEL_REF so we know that this branch should predict
3028 /* A check to make sure the label is not in an inner nested loop,
3029 since this does not count as a loop exit. */
3032 for (outer_loop = dest_loop; outer_loop;
3033 outer_loop = outer_loop->outer)
3034 if (outer_loop == loop)
3040 if (loop && ! outer_loop)
3042 LABEL_OUTSIDE_LOOP_P (x) = 1;
3043 LABEL_NEXTREF (x) = loop->exit_labels;
3044 loop->exit_labels = x;
3046 for (outer_loop = loop;
3047 outer_loop && outer_loop != dest_loop;
3048 outer_loop = outer_loop->outer)
3049 outer_loop->exit_count++;
3052 /* If this is inside a loop, but not in the current loop or one enclosed
3053 by it, it invalidates at least one loop. */
3058 /* We must invalidate every nested loop containing the target of this
3059 label, except those that also contain the jump insn. */
3061 for (; dest_loop; dest_loop = dest_loop->outer)
3063 /* Stop when we reach a loop that also contains the jump insn. */
3064 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3065 if (dest_loop == outer_loop)
3068 /* If we get here, we know we need to invalidate a loop. */
3069 if (loop_dump_stream && ! dest_loop->invalid)
3070 fprintf (loop_dump_stream,
3071 "\nLoop at %d ignored due to multiple entry points.\n",
3072 INSN_UID (dest_loop->start));
3074 dest_loop->invalid = 1;
3079 /* If this is not setting pc, ignore. */
3080 if (SET_DEST (x) == pc_rtx)
3081 mark_loop_jump (SET_SRC (x), loop);
3085 mark_loop_jump (XEXP (x, 1), loop);
3086 mark_loop_jump (XEXP (x, 2), loop);
3091 for (i = 0; i < XVECLEN (x, 0); i++)
3092 mark_loop_jump (XVECEXP (x, 0, i), loop);
3096 for (i = 0; i < XVECLEN (x, 1); i++)
3097 mark_loop_jump (XVECEXP (x, 1, i), loop);
3101 /* Strictly speaking this is not a jump into the loop, only a possible
3102 jump out of the loop. However, we have no way to link the destination
3103 of this jump onto the list of exit labels. To be safe we mark this
3104 loop and any containing loops as invalid. */
3107 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3109 if (loop_dump_stream && ! outer_loop->invalid)
3110 fprintf (loop_dump_stream,
3111 "\nLoop at %d ignored due to unknown exit jump.\n",
3112 INSN_UID (outer_loop->start));
3113 outer_loop->invalid = 1;
3120 /* Return nonzero if there is a label in the range from
3121 insn INSN to and including the insn whose luid is END
3122 INSN must have an assigned luid (i.e., it must not have
3123 been previously created by loop.c). */
3126 labels_in_range_p (insn, end)
3130 while (insn && INSN_LUID (insn) <= end)
3132 if (GET_CODE (insn) == CODE_LABEL)
3134 insn = NEXT_INSN (insn);
3140 /* Record that a memory reference X is being set. */
3143 note_addr_stored (x, y, data)
3145 rtx y ATTRIBUTE_UNUSED;
3146 void *data ATTRIBUTE_UNUSED;
3148 if (x == 0 || GET_CODE (x) != MEM)
3151 /* Count number of memory writes.
3152 This affects heuristics in strength_reduce. */
3155 /* BLKmode MEM means all memory is clobbered. */
3156 if (GET_MODE (x) == BLKmode)
3158 if (RTX_UNCHANGING_P (x))
3159 unknown_constant_address_altered = 1;
3161 unknown_address_altered = 1;
3166 loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, x, loop_store_mems);
3169 /* X is a value modified by an INSN that references a biv inside a loop
3170 exit test (ie, X is somehow related to the value of the biv). If X
3171 is a pseudo that is used more than once, then the biv is (effectively)
3172 used more than once. DATA is really an `int *', and is set if the
3173 biv is used more than once. */
3176 note_set_pseudo_multiple_uses (x, y, data)
3178 rtx y ATTRIBUTE_UNUSED;
3184 while (GET_CODE (x) == STRICT_LOW_PART
3185 || GET_CODE (x) == SIGN_EXTRACT
3186 || GET_CODE (x) == ZERO_EXTRACT
3187 || GET_CODE (x) == SUBREG)
3190 if (GET_CODE (x) != REG || REGNO (x) < FIRST_PSEUDO_REGISTER)
3193 /* If we do not have usage information, or if we know the register
3194 is used more than once, note that fact for check_dbra_loop. */
3195 if (REGNO (x) >= max_reg_before_loop
3196 || ! VARRAY_RTX (reg_single_usage, REGNO (x))
3197 || VARRAY_RTX (reg_single_usage, REGNO (x)) == const0_rtx)
3198 *((int *) data) = 1;
3201 /* Return nonzero if the rtx X is invariant over the current loop.
3203 The value is 2 if we refer to something only conditionally invariant.
3205 If `unknown_address_altered' is nonzero, no memory ref is invariant.
3206 Otherwise, a memory ref is invariant if it does not conflict with
3207 anything stored in `loop_store_mems'. */
3210 loop_invariant_p (loop, x)
3211 const struct loop *loop;
3215 register enum rtx_code code;
3216 register const char *fmt;
3217 int conditional = 0;
3222 code = GET_CODE (x);
3232 /* A LABEL_REF is normally invariant, however, if we are unrolling
3233 loops, and this label is inside the loop, then it isn't invariant.
3234 This is because each unrolled copy of the loop body will have
3235 a copy of this label. If this was invariant, then an insn loading
3236 the address of this label into a register might get moved outside
3237 the loop, and then each loop body would end up using the same label.
3239 We don't know the loop bounds here though, so just fail for all
3241 if (flag_unroll_loops)
3248 case UNSPEC_VOLATILE:
3252 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
3253 since the reg might be set by initialization within the loop. */
3255 if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
3256 || x == arg_pointer_rtx)
3257 && ! current_function_has_nonlocal_goto)
3260 if (LOOP_INFO (loop)->has_call
3261 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
3264 if (VARRAY_INT (set_in_loop, REGNO (x)) < 0)
3267 return VARRAY_INT (set_in_loop, REGNO (x)) == 0;
3270 /* Volatile memory references must be rejected. Do this before
3271 checking for read-only items, so that volatile read-only items
3272 will be rejected also. */
3273 if (MEM_VOLATILE_P (x))
3276 /* If we had a subroutine call, any location in memory could
3277 have been clobbered. We used to test here for volatile and
3278 readonly, but true_dependence knows how to do that better
3280 if (RTX_UNCHANGING_P (x)
3281 ? unknown_constant_address_altered : unknown_address_altered)
3284 /* See if there is any dependence between a store and this load. */
3285 mem_list_entry = loop_store_mems;
3286 while (mem_list_entry)
3288 if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
3292 mem_list_entry = XEXP (mem_list_entry, 1);
3295 /* It's not invalidated by a store in memory
3296 but we must still verify the address is invariant. */
3300 /* Don't mess with insns declared volatile. */
3301 if (MEM_VOLATILE_P (x))
3309 fmt = GET_RTX_FORMAT (code);
3310 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3314 int tem = loop_invariant_p (loop, XEXP (x, i));
3320 else if (fmt[i] == 'E')
3323 for (j = 0; j < XVECLEN (x, i); j++)
3325 int tem = loop_invariant_p (loop, XVECEXP (x, i, j));
3335 return 1 + conditional;
3339 /* Return nonzero if all the insns in the loop that set REG
3340 are INSN and the immediately following insns,
3341 and if each of those insns sets REG in an invariant way
3342 (not counting uses of REG in them).
3344 The value is 2 if some of these insns are only conditionally invariant.
3346 We assume that INSN itself is the first set of REG
3347 and that its source is invariant. */
3350 consec_sets_invariant_p (loop, reg, n_sets, insn)
3351 const struct loop *loop;
3355 register rtx p = insn;
3356 register int regno = REGNO (reg);
3358 /* Number of sets we have to insist on finding after INSN. */
3359 int count = n_sets - 1;
3360 int old = VARRAY_INT (set_in_loop, regno);
3364 /* If N_SETS hit the limit, we can't rely on its value. */
3368 VARRAY_INT (set_in_loop, regno) = 0;
3372 register enum rtx_code code;
3376 code = GET_CODE (p);
3378 /* If library call, skip to end of it. */
3379 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3384 && (set = single_set (p))
3385 && GET_CODE (SET_DEST (set)) == REG
3386 && REGNO (SET_DEST (set)) == regno)
3388 this = loop_invariant_p (loop, SET_SRC (set));
3391 else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)))
3393 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3394 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3396 this = (CONSTANT_P (XEXP (temp, 0))
3397 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
3398 && loop_invariant_p (loop, XEXP (temp, 0))));
3405 else if (code != NOTE)
3407 VARRAY_INT (set_in_loop, regno) = old;
3412 VARRAY_INT (set_in_loop, regno) = old;
3413 /* If loop_invariant_p ever returned 2, we return 2. */
3414 return 1 + (value & 2);
3418 /* I don't think this condition is sufficient to allow INSN
3419 to be moved, so we no longer test it. */
3421 /* Return 1 if all insns in the basic block of INSN and following INSN
3422 that set REG are invariant according to TABLE. */
3425 all_sets_invariant_p (reg, insn, table)
3429 register rtx p = insn;
3430 register int regno = REGNO (reg);
3434 register enum rtx_code code;
3436 code = GET_CODE (p);
3437 if (code == CODE_LABEL || code == JUMP_INSN)
3439 if (code == INSN && GET_CODE (PATTERN (p)) == SET
3440 && GET_CODE (SET_DEST (PATTERN (p))) == REG
3441 && REGNO (SET_DEST (PATTERN (p))) == regno)
3443 if (! loop_invariant_p (loop, SET_SRC (PATTERN (p)), table))
3450 /* Look at all uses (not sets) of registers in X. For each, if it is
3451 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3452 a different insn, set USAGE[REGNO] to const0_rtx. */
3455 find_single_use_in_loop (insn, x, usage)
3460 enum rtx_code code = GET_CODE (x);
3461 const char *fmt = GET_RTX_FORMAT (code);
3465 VARRAY_RTX (usage, REGNO (x))
3466 = (VARRAY_RTX (usage, REGNO (x)) != 0
3467 && VARRAY_RTX (usage, REGNO (x)) != insn)
3468 ? const0_rtx : insn;
3470 else if (code == SET)
3472 /* Don't count SET_DEST if it is a REG; otherwise count things
3473 in SET_DEST because if a register is partially modified, it won't
3474 show up as a potential movable so we don't care how USAGE is set
3476 if (GET_CODE (SET_DEST (x)) != REG)
3477 find_single_use_in_loop (insn, SET_DEST (x), usage);
3478 find_single_use_in_loop (insn, SET_SRC (x), usage);
3481 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3483 if (fmt[i] == 'e' && XEXP (x, i) != 0)
3484 find_single_use_in_loop (insn, XEXP (x, i), usage);
3485 else if (fmt[i] == 'E')
3486 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3487 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
3491 /* Count and record any set in X which is contained in INSN. Update
3492 MAY_NOT_MOVE and LAST_SET for any register set in X. */
3495 count_one_set (insn, x, may_not_move, last_set)
3497 varray_type may_not_move;
3500 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3501 /* Don't move a reg that has an explicit clobber.
3502 It's not worth the pain to try to do it correctly. */
3503 VARRAY_CHAR (may_not_move, REGNO (XEXP (x, 0))) = 1;
3505 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3507 rtx dest = SET_DEST (x);
3508 while (GET_CODE (dest) == SUBREG
3509 || GET_CODE (dest) == ZERO_EXTRACT
3510 || GET_CODE (dest) == SIGN_EXTRACT
3511 || GET_CODE (dest) == STRICT_LOW_PART)
3512 dest = XEXP (dest, 0);
3513 if (GET_CODE (dest) == REG)
3515 register int regno = REGNO (dest);
3516 /* If this is the first setting of this reg
3517 in current basic block, and it was set before,
3518 it must be set in two basic blocks, so it cannot
3519 be moved out of the loop. */
3520 if (VARRAY_INT (set_in_loop, regno) > 0
3521 && last_set[regno] == 0)
3522 VARRAY_CHAR (may_not_move, regno) = 1;
3523 /* If this is not first setting in current basic block,
3524 see if reg was used in between previous one and this.
3525 If so, neither one can be moved. */
3526 if (last_set[regno] != 0
3527 && reg_used_between_p (dest, last_set[regno], insn))
3528 VARRAY_CHAR (may_not_move, regno) = 1;
3529 if (VARRAY_INT (set_in_loop, regno) < 127)
3530 ++VARRAY_INT (set_in_loop, regno);
3531 last_set[regno] = insn;
3536 /* Increment SET_IN_LOOP at the index of each register
3537 that is modified by an insn between FROM and TO.
3538 If the value of an element of SET_IN_LOOP becomes 127 or more,
3539 stop incrementing it, to avoid overflow.
3541 Store in SINGLE_USAGE[I] the single insn in which register I is
3542 used, if it is only used once. Otherwise, it is set to 0 (for no
3543 uses) or const0_rtx for more than one use. This parameter may be zero,
3544 in which case this processing is not done.
3546 Store in *COUNT_PTR the number of actual instruction
3547 in the loop. We use this to decide what is worth moving out. */
3549 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3550 In that case, it is the insn that last set reg n. */
3553 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
3554 register rtx from, to;
3555 varray_type may_not_move;
3556 varray_type single_usage;
3560 register rtx *last_set = (rtx *) xcalloc (nregs, sizeof (rtx));
3562 register int count = 0;
3564 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3566 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3570 /* Record registers that have exactly one use. */
3571 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3573 /* Include uses in REG_EQUAL notes. */
3574 if (REG_NOTES (insn))
3575 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3577 if (GET_CODE (PATTERN (insn)) == SET
3578 || GET_CODE (PATTERN (insn)) == CLOBBER)
3579 count_one_set (insn, PATTERN (insn), may_not_move, last_set);
3580 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3583 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3584 count_one_set (insn, XVECEXP (PATTERN (insn), 0, i),
3585 may_not_move, last_set);
3589 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3590 bzero ((char *) last_set, nregs * sizeof (rtx));
3598 /* Given a loop that is bounded by LOOP->START and LOOP->END and that
3599 is entered at LOOP->SCAN_START, return 1 if the register set in SET
3600 contained in insn INSN is used by any insn that precedes INSN in
3601 cyclic order starting from the loop entry point.
3603 We don't want to use INSN_LUID here because if we restrict INSN to those
3604 that have a valid INSN_LUID, it means we cannot move an invariant out
3605 from an inner loop past two loops. */
3608 loop_reg_used_before_p (loop, set, insn)
3609 const struct loop *loop;
3612 rtx reg = SET_DEST (set);
3615 /* Scan forward checking for register usage. If we hit INSN, we
3616 are done. Otherwise, if we hit LOOP->END, wrap around to LOOP->START. */
3617 for (p = loop->scan_start; p != insn; p = NEXT_INSN (p))
3619 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
3620 && reg_overlap_mentioned_p (reg, PATTERN (p)))
3630 /* A "basic induction variable" or biv is a pseudo reg that is set
3631 (within this loop) only by incrementing or decrementing it. */
3632 /* A "general induction variable" or giv is a pseudo reg whose
3633 value is a linear function of a biv. */
3635 /* Bivs are recognized by `basic_induction_var';
3636 Givs by `general_induction_var'. */
3638 /* Indexed by register number, indicates whether or not register is an
3639 induction variable, and if so what type. */
3641 varray_type reg_iv_type;
3643 /* Indexed by register number, contains pointer to `struct induction'
3644 if register is an induction variable. This holds general info for
3645 all induction variables. */
3647 varray_type reg_iv_info;
3649 /* Indexed by register number, contains pointer to `struct iv_class'
3650 if register is a basic induction variable. This holds info describing
3651 the class (a related group) of induction variables that the biv belongs
3654 struct iv_class **reg_biv_class;
3656 /* The head of a list which links together (via the next field)
3657 every iv class for the current loop. */
3659 struct iv_class *loop_iv_list;
3661 /* Givs made from biv increments are always splittable for loop unrolling.
3662 Since there is no regscan info for them, we have to keep track of them
3664 int first_increment_giv, last_increment_giv;
3666 /* Communication with routines called via `note_stores'. */
3668 static rtx note_insn;
3670 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3672 static rtx addr_placeholder;
3674 /* ??? Unfinished optimizations, and possible future optimizations,
3675 for the strength reduction code. */
3677 /* ??? The interaction of biv elimination, and recognition of 'constant'
3678 bivs, may cause problems. */
3680 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3681 performance problems.
3683 Perhaps don't eliminate things that can be combined with an addressing
3684 mode. Find all givs that have the same biv, mult_val, and add_val;
3685 then for each giv, check to see if its only use dies in a following
3686 memory address. If so, generate a new memory address and check to see
3687 if it is valid. If it is valid, then store the modified memory address,
3688 otherwise, mark the giv as not done so that it will get its own iv. */
3690 /* ??? Could try to optimize branches when it is known that a biv is always
3693 /* ??? When replace a biv in a compare insn, we should replace with closest
3694 giv so that an optimized branch can still be recognized by the combiner,
3695 e.g. the VAX acb insn. */
3697 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3698 was rerun in loop_optimize whenever a register was added or moved.
3699 Also, some of the optimizations could be a little less conservative. */
3701 /* Perform strength reduction and induction variable elimination.
3703 Pseudo registers created during this function will be beyond the last
3704 valid index in several tables including n_times_set and regno_last_uid.
3705 This does not cause a problem here, because the added registers cannot be
3706 givs outside of their loop, and hence will never be reconsidered.
3707 But scan_loop must check regnos to make sure they are in bounds. */
3710 strength_reduce (loop, insn_count, unroll_p, bct_p)
3713 int unroll_p, bct_p ATTRIBUTE_UNUSED;
3721 /* This is 1 if current insn is not executed at least once for every loop
3723 int not_every_iteration = 0;
3724 /* This is 1 if current insn may be executed more than once for every
3726 int maybe_multiple = 0;
3727 /* This is 1 if we have past a branch back to the top of the loop
3728 (aka a loop latch). */
3729 int past_loop_latch = 0;
3730 /* Temporary list pointers for traversing loop_iv_list. */
3731 struct iv_class *bl, **backbl;
3732 struct loop_info *loop_info = LOOP_INFO (loop);
3733 /* Ratio of extra register life span we can justify
3734 for saving an instruction. More if loop doesn't call subroutines
3735 since in that case saving an insn makes more difference
3736 and more registers are available. */
3737 /* ??? could set this to last value of threshold in move_movables */
3738 int threshold = (loop_info->has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3739 /* Map of pseudo-register replacements. */
3740 rtx *reg_map = NULL;
3744 rtx end_insert_before;
3746 int n_extra_increment;
3747 int unrolled_insn_copies = 0;
3748 rtx loop_start = loop->start;
3749 rtx loop_end = loop->end;
3750 rtx loop_scan_start = loop->scan_start;
3751 rtx loop_top = loop->top;
3752 rtx loop_cont = loop->cont;
3754 /* If loop_scan_start points to the loop exit test, we have to be wary of
3755 subversive use of gotos inside expression statements. */
3756 if (prev_nonnote_insn (loop_scan_start) != prev_nonnote_insn (loop_start))
3757 maybe_multiple = back_branch_in_range_p (loop, loop_scan_start);
3759 VARRAY_INT_INIT (reg_iv_type, max_reg_before_loop, "reg_iv_type");
3760 VARRAY_GENERIC_PTR_INIT (reg_iv_info, max_reg_before_loop, "reg_iv_info");
3761 reg_biv_class = (struct iv_class **)
3762 xcalloc (max_reg_before_loop, sizeof (struct iv_class *));
3765 addr_placeholder = gen_reg_rtx (Pmode);
3767 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3768 must be put before this insn, so that they will appear in the right
3769 order (i.e. loop order).
3771 If loop_end is the end of the current function, then emit a
3772 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3774 if (NEXT_INSN (loop_end) != 0)
3775 end_insert_before = NEXT_INSN (loop_end);
3777 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3779 /* Scan through loop to find all possible bivs. */
3781 for (p = next_insn_in_loop (loop, loop_scan_start);
3783 p = next_insn_in_loop (loop, p))
3785 if (GET_CODE (p) == INSN
3786 && (set = single_set (p))
3787 && GET_CODE (SET_DEST (set)) == REG)
3789 dest_reg = SET_DEST (set);
3790 if (REGNO (dest_reg) < max_reg_before_loop
3791 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3792 && REG_IV_TYPE (REGNO (dest_reg)) != NOT_BASIC_INDUCT)
3794 int multi_insn_incr = 0;
3796 if (basic_induction_var (loop, SET_SRC (set),
3797 GET_MODE (SET_SRC (set)),
3798 dest_reg, p, &inc_val, &mult_val,
3799 &location, &multi_insn_incr))
3801 /* It is a possible basic induction variable.
3802 Create and initialize an induction structure for it. */
3805 = (struct induction *) alloca (sizeof (struct induction));
3807 record_biv (v, p, dest_reg, inc_val, mult_val, location,
3808 not_every_iteration, maybe_multiple,
3810 REG_IV_TYPE (REGNO (dest_reg)) = BASIC_INDUCT;
3812 else if (REGNO (dest_reg) < max_reg_before_loop)
3813 REG_IV_TYPE (REGNO (dest_reg)) = NOT_BASIC_INDUCT;
3817 /* Past CODE_LABEL, we get to insns that may be executed multiple
3818 times. The only way we can be sure that they can't is if every
3819 jump insn between here and the end of the loop either
3820 returns, exits the loop, is a jump to a location that is still
3821 behind the label, or is a jump to the loop start. */
3823 if (GET_CODE (p) == CODE_LABEL)
3831 insn = NEXT_INSN (insn);
3832 if (insn == loop_scan_start)
3834 if (insn == loop_end)
3840 if (insn == loop_scan_start)
3844 if (GET_CODE (insn) == JUMP_INSN
3845 && GET_CODE (PATTERN (insn)) != RETURN
3846 && (! condjump_p (insn)
3847 || (JUMP_LABEL (insn) != 0
3848 && JUMP_LABEL (insn) != loop_scan_start
3849 && ! loop_insn_first_p (p, JUMP_LABEL (insn)))))
3857 /* Past a jump, we get to insns for which we can't count
3858 on whether they will be executed during each iteration. */
3859 /* This code appears twice in strength_reduce. There is also similar
3860 code in scan_loop. */
3861 if (GET_CODE (p) == JUMP_INSN
3862 /* If we enter the loop in the middle, and scan around to the
3863 beginning, don't set not_every_iteration for that.
3864 This can be any kind of jump, since we want to know if insns
3865 will be executed if the loop is executed. */
3866 && ! (JUMP_LABEL (p) == loop_top
3867 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3868 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3872 /* If this is a jump outside the loop, then it also doesn't
3873 matter. Check to see if the target of this branch is on the
3874 loop->exits_labels list. */
3876 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
3877 if (XEXP (label, 0) == JUMP_LABEL (p))
3881 not_every_iteration = 1;
3884 else if (GET_CODE (p) == NOTE)
3886 /* At the virtual top of a converted loop, insns are again known to
3887 be executed each iteration: logically, the loop begins here
3888 even though the exit code has been duplicated.
3890 Insns are also again known to be executed each iteration at
3891 the LOOP_CONT note. */
3892 if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
3893 || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
3895 not_every_iteration = 0;
3896 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3898 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3902 /* Note if we pass a loop latch. If we do, then we can not clear
3903 NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in
3904 a loop since a jump before the last CODE_LABEL may have started
3905 a new loop iteration.
3907 Note that LOOP_TOP is only set for rotated loops and we need
3908 this check for all loops, so compare against the CODE_LABEL
3909 which immediately follows LOOP_START. */
3910 if (GET_CODE (p) == JUMP_INSN
3911 && JUMP_LABEL (p) == NEXT_INSN (loop_start))
3912 past_loop_latch = 1;
3914 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3915 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3916 or not an insn is known to be executed each iteration of the
3917 loop, whether or not any iterations are known to occur.
3919 Therefore, if we have just passed a label and have no more labels
3920 between here and the test insn of the loop, and we have not passed
3921 a jump to the top of the loop, then we know these insns will be
3922 executed each iteration. */
3924 if (not_every_iteration
3925 && ! past_loop_latch
3926 && GET_CODE (p) == CODE_LABEL
3927 && no_labels_between_p (p, loop_end)
3928 && loop_insn_first_p (p, loop_cont))
3929 not_every_iteration = 0;
3932 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3933 Make a sanity check against n_times_set. */
3934 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3938 if (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3939 /* Above happens if register modified by subreg, etc. */
3940 /* Make sure it is not recognized as a basic induction var: */
3941 || VARRAY_INT (n_times_set, bl->regno) != bl->biv_count
3942 /* If never incremented, it is invariant that we decided not to
3943 move. So leave it alone. */
3944 || ! bl->incremented)
3946 else if (bl->biv_count > 1)
3948 /* ??? If we have multiple increments for this BIV, and any of
3949 them take multiple insns to perform the increment, drop the
3950 BIV, since the bit below that converts the extra increments
3951 into GIVs can't handle the multiple insn increment. */
3953 struct induction *v;
3954 for (v = bl->biv; v ; v = v->next_iv)
3955 if (v->multi_insn_incr)
3961 if (loop_dump_stream)
3962 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3964 (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3965 ? "not induction variable"
3966 : (! bl->incremented ? "never incremented"
3969 REG_IV_TYPE (bl->regno) = NOT_BASIC_INDUCT;
3976 if (loop_dump_stream)
3977 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3981 /* Exit if there are no bivs. */
3984 /* Can still unroll the loop anyways, but indicate that there is no
3985 strength reduction info available. */
3987 unroll_loop (loop, insn_count, end_insert_before, 0);
3992 /* Find initial value for each biv by searching backwards from loop_start,
3993 halting at first label. Also record any test condition. */
3996 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
4000 if (GET_CODE (p) == CALL_INSN)
4003 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4004 || GET_CODE (p) == CALL_INSN)
4005 note_stores (PATTERN (p), record_initial, NULL);
4007 /* Record any test of a biv that branches around the loop if no store
4008 between it and the start of loop. We only care about tests with
4009 constants and registers and only certain of those. */
4010 if (GET_CODE (p) == JUMP_INSN
4011 && JUMP_LABEL (p) != 0
4012 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
4013 && (test = get_condition_for_loop (loop, p)) != 0
4014 && GET_CODE (XEXP (test, 0)) == REG
4015 && REGNO (XEXP (test, 0)) < max_reg_before_loop
4016 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
4017 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
4018 && bl->init_insn == 0)
4020 /* If an NE test, we have an initial value! */
4021 if (GET_CODE (test) == NE)
4024 bl->init_set = gen_rtx_SET (VOIDmode,
4025 XEXP (test, 0), XEXP (test, 1));
4028 bl->initial_test = test;
4032 /* Look at the each biv and see if we can say anything better about its
4033 initial value from any initializing insns set up above. (This is done
4034 in two passes to avoid missing SETs in a PARALLEL.) */
4035 for (backbl = &loop_iv_list; (bl = *backbl); backbl = &bl->next)
4040 if (! bl->init_insn)
4043 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
4044 is a constant, use the value of that. */
4045 if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL
4046 && CONSTANT_P (XEXP (note, 0)))
4047 || ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL
4048 && CONSTANT_P (XEXP (note, 0))))
4049 src = XEXP (note, 0);
4051 src = SET_SRC (bl->init_set);
4053 if (loop_dump_stream)
4054 fprintf (loop_dump_stream,
4055 "Biv %d initialized at insn %d: initial value ",
4056 bl->regno, INSN_UID (bl->init_insn));
4058 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
4059 || GET_MODE (src) == VOIDmode)
4060 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
4062 bl->initial_value = src;
4064 if (loop_dump_stream)
4066 if (GET_CODE (src) == CONST_INT)
4068 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (src));
4069 fputc ('\n', loop_dump_stream);
4073 print_rtl (loop_dump_stream, src);
4074 fprintf (loop_dump_stream, "\n");
4080 struct iv_class *bl2 = 0;
4081 rtx increment = NULL_RTX;
4083 /* Biv initial value is not a simple move. If it is the sum of
4084 another biv and a constant, check if both bivs are incremented
4085 in lockstep. Then we are actually looking at a giv.
4086 For simplicity, we only handle the case where there is but a
4087 single increment, and the register is not used elsewhere. */
4088 if (bl->biv_count == 1
4089 && bl->regno < max_reg_before_loop
4090 && uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4091 && GET_CODE (src) == PLUS
4092 && GET_CODE (XEXP (src, 0)) == REG
4093 && CONSTANT_P (XEXP (src, 1))
4094 && ((increment = biv_total_increment (bl)) != NULL_RTX))
4096 int regno = REGNO (XEXP (src, 0));
4098 for (bl2 = loop_iv_list; bl2; bl2 = bl2->next)
4099 if (bl2->regno == regno)
4103 /* Now, can we transform this biv into a giv? */
4105 && bl2->biv_count == 1
4106 && rtx_equal_p (increment, biv_total_increment (bl2))
4107 /* init_insn is only set to insns that are before loop_start
4108 without any intervening labels. */
4109 && ! reg_set_between_p (bl2->biv->src_reg,
4110 PREV_INSN (bl->init_insn), loop_start)
4111 /* The register from BL2 must be set before the register from
4112 BL is set, or we must be able to move the latter set after
4113 the former set. Currently there can't be any labels
4114 in-between when biv_total_increment returns nonzero both times
4115 but we test it here in case some day some real cfg analysis
4116 gets used to set always_computable. */
4117 && (loop_insn_first_p (bl2->biv->insn, bl->biv->insn)
4118 ? no_labels_between_p (bl2->biv->insn, bl->biv->insn)
4119 : (! reg_used_between_p (bl->biv->src_reg, bl->biv->insn,
4121 && no_jumps_between_p (bl->biv->insn, bl2->biv->insn)))
4122 && validate_change (bl->biv->insn,
4123 &SET_SRC (single_set (bl->biv->insn)),
4126 rtx dominator = loop->cont_dominator;
4127 rtx giv = bl->biv->src_reg;
4128 rtx giv_insn = bl->biv->insn;
4129 rtx after_giv = NEXT_INSN (giv_insn);
4131 if (loop_dump_stream)
4132 fprintf (loop_dump_stream, "is giv of biv %d\n", bl2->regno);
4133 /* Let this giv be discovered by the generic code. */
4134 REG_IV_TYPE (bl->regno) = UNKNOWN_INDUCT;
4135 reg_biv_class[bl->regno] = (struct iv_class *) NULL_PTR;
4136 /* We can get better optimization if we can move the giv setting
4137 before the first giv use. */
4139 && ! loop_insn_first_p (dominator, loop_scan_start)
4140 && ! reg_set_between_p (bl2->biv->src_reg, loop_start,
4142 && ! reg_used_between_p (giv, loop_start, dominator)
4143 && ! reg_used_between_p (giv, giv_insn, loop_end))
4148 for (next = NEXT_INSN (dominator); ; next = NEXT_INSN (next))
4150 if ((GET_RTX_CLASS (GET_CODE (next)) == 'i'
4151 && (reg_mentioned_p (giv, PATTERN (next))
4152 || reg_set_p (bl2->biv->src_reg, next)))
4153 || GET_CODE (next) == JUMP_INSN)
4156 if (GET_RTX_CLASS (GET_CODE (next)) != 'i'
4157 || ! sets_cc0_p (PATTERN (next)))
4161 if (loop_dump_stream)
4162 fprintf (loop_dump_stream, "move after insn %d\n",
4163 INSN_UID (dominator));
4164 /* Avoid problems with luids by actually moving the insn
4165 and adjusting all luids in the range. */
4166 reorder_insns (giv_insn, giv_insn, dominator);
4167 for (p = dominator; INSN_UID (p) >= max_uid_for_loop; )
4169 compute_luids (giv_insn, after_giv, INSN_LUID (p));
4170 /* If the only purpose of the init insn is to initialize
4171 this giv, delete it. */
4172 if (single_set (bl->init_insn)
4173 && ! reg_used_between_p (giv, bl->init_insn, loop_start))
4174 delete_insn (bl->init_insn);
4176 else if (! loop_insn_first_p (bl2->biv->insn, bl->biv->insn))
4178 rtx p = PREV_INSN (giv_insn);
4179 while (INSN_UID (p) >= max_uid_for_loop)
4181 reorder_insns (giv_insn, giv_insn, bl2->biv->insn);
4182 compute_luids (after_giv, NEXT_INSN (giv_insn),
4185 /* Remove this biv from the chain. */
4188 /* We move the following giv from *bl->next into *bl.
4189 We have to update reg_biv_class for that moved biv
4190 to point to its new address. */
4192 reg_biv_class[bl->regno] = bl;
4201 /* If we can't make it a giv,
4202 let biv keep initial value of "itself". */
4203 else if (loop_dump_stream)
4204 fprintf (loop_dump_stream, "is complex\n");
4208 /* If a biv is unconditionally incremented several times in a row, convert
4209 all but the last increment into a giv. */
4211 /* Get an upper bound for the number of registers
4212 we might have after all bivs have been processed. */
4213 first_increment_giv = max_reg_num ();
4214 for (n_extra_increment = 0, bl = loop_iv_list; bl; bl = bl->next)
4215 n_extra_increment += bl->biv_count - 1;
4217 /* If the loop contains volatile memory references do not allow any
4218 replacements to take place, since this could loose the volatile
4220 if (n_extra_increment && ! loop_info->has_volatile)
4222 int nregs = first_increment_giv + n_extra_increment;
4224 /* Reallocate reg_iv_type and reg_iv_info. */
4225 VARRAY_GROW (reg_iv_type, nregs);
4226 VARRAY_GROW (reg_iv_info, nregs);
4228 for (bl = loop_iv_list; bl; bl = bl->next)
4230 struct induction **vp, *v, *next;
4231 int biv_dead_after_loop = 0;
4233 /* The biv increments lists are in reverse order. Fix this
4235 for (v = bl->biv, bl->biv = 0; v; v = next)
4238 v->next_iv = bl->biv;
4242 /* We must guard against the case that an early exit between v->insn
4243 and next->insn leaves the biv live after the loop, since that
4244 would mean that we'd be missing an increment for the final
4245 value. The following test to set biv_dead_after_loop is like
4246 the first part of the test to set bl->eliminable.
4247 We don't check here if we can calculate the final value, since
4248 this can't succeed if we already know that there is a jump
4249 between v->insn and next->insn, yet next->always_executed is
4250 set and next->maybe_multiple is cleared. Such a combination
4251 implies that the jump destination is outside the loop.
4252 If we want to make this check more sophisticated, we should
4253 check each branch between v->insn and next->insn individually
4254 to see if the biv is dead at its destination. */
4256 if (uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4258 && INSN_UID (bl->init_insn) < max_uid_for_loop
4259 && (uid_luid[REGNO_FIRST_UID (bl->regno)]
4260 >= INSN_LUID (bl->init_insn))
4261 #ifdef HAVE_decrement_and_branch_until_zero
4264 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4265 biv_dead_after_loop = 1;
4267 for (vp = &bl->biv, next = *vp; v = next, next = v->next_iv;)
4269 HOST_WIDE_INT offset;
4270 rtx set, add_val, old_reg, dest_reg, last_use_insn, note;
4271 int old_regno, new_regno;
4273 if (! v->always_executed
4274 || v->maybe_multiple
4275 || GET_CODE (v->add_val) != CONST_INT
4276 || ! next->always_executed
4277 || next->maybe_multiple
4278 || ! CONSTANT_P (next->add_val)
4279 || v->mult_val != const1_rtx
4280 || next->mult_val != const1_rtx
4281 || ! (biv_dead_after_loop
4282 || no_jumps_between_p (v->insn, next->insn)))
4287 offset = INTVAL (v->add_val);
4288 set = single_set (v->insn);
4289 add_val = plus_constant (next->add_val, offset);
4290 old_reg = v->dest_reg;
4291 dest_reg = gen_reg_rtx (v->mode);
4293 /* Unlike reg_iv_type / reg_iv_info, the other three arrays
4294 have been allocated with some slop space, so we may not
4295 actually need to reallocate them. If we do, the following
4296 if statement will be executed just once in this loop. */
4297 if ((unsigned) max_reg_num () > n_times_set->num_elements)
4299 /* Grow all the remaining arrays. */
4300 VARRAY_GROW (set_in_loop, nregs);
4301 VARRAY_GROW (n_times_set, nregs);
4302 VARRAY_GROW (may_not_optimize, nregs);
4303 VARRAY_GROW (reg_single_usage, nregs);
4306 if (! validate_change (next->insn, next->location, add_val, 0))
4312 /* Here we can try to eliminate the increment by combining
4313 it into the uses. */
4315 /* Set last_use_insn so that we can check against it. */
4317 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4319 p = next_insn_in_loop (loop, p))
4321 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
4323 if (reg_mentioned_p (old_reg, PATTERN (p)))
4329 /* If we can't get the LUIDs for the insns, we can't
4330 calculate the lifetime. This is likely from unrolling
4331 of an inner loop, so there is little point in making this
4332 a DEST_REG giv anyways. */
4333 if (INSN_UID (v->insn) >= max_uid_for_loop
4334 || INSN_UID (last_use_insn) >= max_uid_for_loop
4335 || ! validate_change (v->insn, &SET_DEST (set), dest_reg, 0))
4337 /* Change the increment at NEXT back to what it was. */
4338 if (! validate_change (next->insn, next->location,
4344 next->add_val = add_val;
4345 v->dest_reg = dest_reg;
4346 v->giv_type = DEST_REG;
4347 v->location = &SET_SRC (set);
4349 v->combined_with = 0;
4351 v->derive_adjustment = 0;
4357 v->auto_inc_opt = 0;
4360 v->derived_from = 0;
4361 v->always_computable = 1;
4362 v->always_executed = 1;
4364 v->no_const_addval = 0;
4366 old_regno = REGNO (old_reg);
4367 new_regno = REGNO (dest_reg);
4368 VARRAY_INT (set_in_loop, old_regno)--;
4369 VARRAY_INT (set_in_loop, new_regno) = 1;
4370 VARRAY_INT (n_times_set, old_regno)--;
4371 VARRAY_INT (n_times_set, new_regno) = 1;
4372 VARRAY_CHAR (may_not_optimize, new_regno) = 0;
4374 REG_IV_TYPE (new_regno) = GENERAL_INDUCT;
4375 REG_IV_INFO (new_regno) = v;
4377 /* If next_insn has a REG_EQUAL note that mentiones OLD_REG,
4378 it must be replaced. */
4379 note = find_reg_note (next->insn, REG_EQUAL, NULL_RTX);
4380 if (note && reg_mentioned_p (old_reg, XEXP (note, 0)))
4381 XEXP (note, 0) = copy_rtx (SET_SRC (single_set (next->insn)));
4383 /* Remove the increment from the list of biv increments,
4384 and record it as a giv. */
4387 v->next_iv = bl->giv;
4390 v->benefit = rtx_cost (SET_SRC (set), SET);
4391 bl->total_benefit += v->benefit;
4393 /* Now replace the biv with DEST_REG in all insns between
4394 the replaced increment and the next increment, and
4395 remember the last insn that needed a replacement. */
4396 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4398 p = next_insn_in_loop (loop, p))
4402 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
4404 if (reg_mentioned_p (old_reg, PATTERN (p)))
4407 if (! validate_replace_rtx (old_reg, dest_reg, p))
4410 for (note = REG_NOTES (p); note; note = XEXP (note, 1))
4412 if (GET_CODE (note) == EXPR_LIST)
4414 = replace_rtx (XEXP (note, 0), old_reg, dest_reg);
4418 v->last_use = last_use_insn;
4419 v->lifetime = INSN_LUID (last_use_insn) - INSN_LUID (v->insn);
4420 /* If the lifetime is zero, it means that this register is really
4421 a dead store. So mark this as a giv that can be ignored.
4422 This will not prevent the biv from being eliminated. */
4423 if (v->lifetime == 0)
4426 if (loop_dump_stream)
4427 fprintf (loop_dump_stream,
4428 "Increment %d of biv %d converted to giv %d.\n",
4429 INSN_UID (v->insn), old_regno, new_regno);
4433 last_increment_giv = max_reg_num () - 1;
4435 /* Search the loop for general induction variables. */
4437 /* A register is a giv if: it is only set once, it is a function of a
4438 biv and a constant (or invariant), and it is not a biv. */
4440 not_every_iteration = 0;
4443 p = loop_scan_start;
4447 /* At end of a straight-in loop, we are done.
4448 At end of a loop entered at the bottom, scan the top. */
4449 if (p == loop_scan_start)
4457 if (p == loop_scan_start)
4461 /* Look for a general induction variable in a register. */
4462 if (GET_CODE (p) == INSN
4463 && (set = single_set (p))
4464 && GET_CODE (SET_DEST (set)) == REG
4465 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
4472 rtx last_consec_insn;
4474 dest_reg = SET_DEST (set);
4475 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
4478 if (/* SET_SRC is a giv. */
4479 (general_induction_var (loop, SET_SRC (set), &src_reg, &add_val,
4480 &mult_val, 0, &benefit)
4481 /* Equivalent expression is a giv. */
4482 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
4483 && general_induction_var (loop, XEXP (regnote, 0), &src_reg,
4484 &add_val, &mult_val, 0,
4486 /* Don't try to handle any regs made by loop optimization.
4487 We have nothing on them in regno_first_uid, etc. */
4488 && REGNO (dest_reg) < max_reg_before_loop
4489 /* Don't recognize a BASIC_INDUCT_VAR here. */
4490 && dest_reg != src_reg
4491 /* This must be the only place where the register is set. */
4492 && (VARRAY_INT (n_times_set, REGNO (dest_reg)) == 1
4493 /* or all sets must be consecutive and make a giv. */
4494 || (benefit = consec_sets_giv (loop, benefit, p,
4496 &add_val, &mult_val,
4497 &last_consec_insn))))
4500 = (struct induction *) alloca (sizeof (struct induction));
4502 /* If this is a library call, increase benefit. */
4503 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
4504 benefit += libcall_benefit (p);
4506 /* Skip the consecutive insns, if there are any. */
4507 if (VARRAY_INT (n_times_set, REGNO (dest_reg)) != 1)
4508 p = last_consec_insn;
4510 record_giv (loop, v, p, src_reg, dest_reg, mult_val, add_val,
4511 benefit, DEST_REG, not_every_iteration,
4512 maybe_multiple, NULL_PTR);
4517 #ifndef DONT_REDUCE_ADDR
4518 /* Look for givs which are memory addresses. */
4519 /* This resulted in worse code on a VAX 8600. I wonder if it
4521 if (GET_CODE (p) == INSN)
4522 find_mem_givs (loop, PATTERN (p), p, not_every_iteration,
4526 /* Update the status of whether giv can derive other givs. This can
4527 change when we pass a label or an insn that updates a biv. */
4528 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4529 || GET_CODE (p) == CODE_LABEL)
4530 update_giv_derive (loop, p);
4532 /* Past CODE_LABEL, we get to insns that may be executed multiple
4533 times. The only way we can be sure that they can't is if every
4534 every jump insn between here and the end of the loop either
4535 returns, exits the loop, is a forward jump, or is a jump
4536 to the loop start. */
4538 if (GET_CODE (p) == CODE_LABEL)
4546 insn = NEXT_INSN (insn);
4547 if (insn == loop_scan_start)
4549 if (insn == loop_end)
4555 if (insn == loop_scan_start)
4559 if (GET_CODE (insn) == JUMP_INSN
4560 && GET_CODE (PATTERN (insn)) != RETURN
4561 && (! condjump_p (insn)
4562 || (JUMP_LABEL (insn) != 0
4563 && JUMP_LABEL (insn) != loop_scan_start
4564 && (INSN_UID (JUMP_LABEL (insn)) >= max_uid_for_loop
4565 || INSN_UID (insn) >= max_uid_for_loop
4566 || (INSN_LUID (JUMP_LABEL (insn))
4567 < INSN_LUID (insn))))))
4575 /* Past a jump, we get to insns for which we can't count
4576 on whether they will be executed during each iteration. */
4577 /* This code appears twice in strength_reduce. There is also similar
4578 code in scan_loop. */
4579 if (GET_CODE (p) == JUMP_INSN
4580 /* If we enter the loop in the middle, and scan around to the
4581 beginning, don't set not_every_iteration for that.
4582 This can be any kind of jump, since we want to know if insns
4583 will be executed if the loop is executed. */
4584 && ! (JUMP_LABEL (p) == loop_top
4585 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
4586 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
4590 /* If this is a jump outside the loop, then it also doesn't
4591 matter. Check to see if the target of this branch is on the
4592 loop->exits_labels list. */
4594 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
4595 if (XEXP (label, 0) == JUMP_LABEL (p))
4599 not_every_iteration = 1;
4602 else if (GET_CODE (p) == NOTE)
4604 /* At the virtual top of a converted loop, insns are again known to
4605 be executed each iteration: logically, the loop begins here
4606 even though the exit code has been duplicated.
4608 Insns are also again known to be executed each iteration at
4609 the LOOP_CONT note. */
4610 if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
4611 || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
4613 not_every_iteration = 0;
4614 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
4616 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
4620 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
4621 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
4622 or not an insn is known to be executed each iteration of the
4623 loop, whether or not any iterations are known to occur.
4625 Therefore, if we have just passed a label and have no more labels
4626 between here and the test insn of the loop, we know these insns
4627 will be executed each iteration. */
4629 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
4630 && no_labels_between_p (p, loop_end)
4631 && loop_insn_first_p (p, loop_cont))
4632 not_every_iteration = 0;
4635 /* Try to calculate and save the number of loop iterations. This is
4636 set to zero if the actual number can not be calculated. This must
4637 be called after all giv's have been identified, since otherwise it may
4638 fail if the iteration variable is a giv. */
4640 loop_iterations (loop);
4642 /* Now for each giv for which we still don't know whether or not it is
4643 replaceable, check to see if it is replaceable because its final value
4644 can be calculated. This must be done after loop_iterations is called,
4645 so that final_giv_value will work correctly. */
4647 for (bl = loop_iv_list; bl; bl = bl->next)
4649 struct induction *v;
4651 for (v = bl->giv; v; v = v->next_iv)
4652 if (! v->replaceable && ! v->not_replaceable)
4653 check_final_value (loop, v);
4656 /* Try to prove that the loop counter variable (if any) is always
4657 nonnegative; if so, record that fact with a REG_NONNEG note
4658 so that "decrement and branch until zero" insn can be used. */
4659 check_dbra_loop (loop, insn_count);
4661 /* Create reg_map to hold substitutions for replaceable giv regs.
4662 Some givs might have been made from biv increments, so look at
4663 reg_iv_type for a suitable size. */
4664 reg_map_size = reg_iv_type->num_elements;
4665 reg_map = (rtx *) xcalloc (reg_map_size, sizeof (rtx));
4667 /* Examine each iv class for feasibility of strength reduction/induction
4668 variable elimination. */
4670 for (bl = loop_iv_list; bl; bl = bl->next)
4672 struct induction *v;
4675 rtx final_value = 0;
4678 /* Test whether it will be possible to eliminate this biv
4679 provided all givs are reduced. This is possible if either
4680 the reg is not used outside the loop, or we can compute
4681 what its final value will be.
4683 For architectures with a decrement_and_branch_until_zero insn,
4684 don't do this if we put a REG_NONNEG note on the endtest for
4687 /* Compare against bl->init_insn rather than loop_start.
4688 We aren't concerned with any uses of the biv between
4689 init_insn and loop_start since these won't be affected
4690 by the value of the biv elsewhere in the function, so
4691 long as init_insn doesn't use the biv itself.
4692 March 14, 1989 -- self@bayes.arc.nasa.gov */
4694 if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4696 && INSN_UID (bl->init_insn) < max_uid_for_loop
4697 && uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)
4698 #ifdef HAVE_decrement_and_branch_until_zero
4701 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4702 || ((final_value = final_biv_value (loop, bl))
4703 #ifdef HAVE_decrement_and_branch_until_zero
4707 bl->eliminable = maybe_eliminate_biv (loop, bl, 0, threshold,
4711 if (loop_dump_stream)
4713 fprintf (loop_dump_stream,
4714 "Cannot eliminate biv %d.\n",
4716 fprintf (loop_dump_stream,
4717 "First use: insn %d, last use: insn %d.\n",
4718 REGNO_FIRST_UID (bl->regno),
4719 REGNO_LAST_UID (bl->regno));
4723 /* Combine all giv's for this iv_class. */
4726 /* This will be true at the end, if all givs which depend on this
4727 biv have been strength reduced.
4728 We can't (currently) eliminate the biv unless this is so. */
4731 /* Check each giv in this class to see if we will benefit by reducing
4732 it. Skip giv's combined with others. */
4733 for (v = bl->giv; v; v = v->next_iv)
4735 struct induction *tv;
4737 if (v->ignore || v->same)
4740 benefit = v->benefit;
4742 /* Reduce benefit if not replaceable, since we will insert
4743 a move-insn to replace the insn that calculates this giv.
4744 Don't do this unless the giv is a user variable, since it
4745 will often be marked non-replaceable because of the duplication
4746 of the exit code outside the loop. In such a case, the copies
4747 we insert are dead and will be deleted. So they don't have
4748 a cost. Similar situations exist. */
4749 /* ??? The new final_[bg]iv_value code does a much better job
4750 of finding replaceable giv's, and hence this code may no longer
4752 if (! v->replaceable && ! bl->eliminable
4753 && REG_USERVAR_P (v->dest_reg))
4754 benefit -= copy_cost;
4756 /* Decrease the benefit to count the add-insns that we will
4757 insert to increment the reduced reg for the giv. */
4758 benefit -= add_cost * bl->biv_count;
4760 /* Decide whether to strength-reduce this giv or to leave the code
4761 unchanged (recompute it from the biv each time it is used).
4762 This decision can be made independently for each giv. */
4765 /* Attempt to guess whether autoincrement will handle some of the
4766 new add insns; if so, increase BENEFIT (undo the subtraction of
4767 add_cost that was done above). */
4768 if (v->giv_type == DEST_ADDR
4769 && GET_CODE (v->mult_val) == CONST_INT)
4771 if (HAVE_POST_INCREMENT
4772 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4773 benefit += add_cost * bl->biv_count;
4774 else if (HAVE_PRE_INCREMENT
4775 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4776 benefit += add_cost * bl->biv_count;
4777 else if (HAVE_POST_DECREMENT
4778 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4779 benefit += add_cost * bl->biv_count;
4780 else if (HAVE_PRE_DECREMENT
4781 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4782 benefit += add_cost * bl->biv_count;
4786 /* If an insn is not to be strength reduced, then set its ignore
4787 flag, and clear all_reduced. */
4789 /* A giv that depends on a reversed biv must be reduced if it is
4790 used after the loop exit, otherwise, it would have the wrong
4791 value after the loop exit. To make it simple, just reduce all
4792 of such giv's whether or not we know they are used after the loop
4795 if ( ! flag_reduce_all_givs && v->lifetime * threshold * benefit < insn_count
4798 if (loop_dump_stream)
4799 fprintf (loop_dump_stream,
4800 "giv of insn %d not worth while, %d vs %d.\n",
4802 v->lifetime * threshold * benefit, insn_count);
4808 /* Check that we can increment the reduced giv without a
4809 multiply insn. If not, reject it. */
4811 for (tv = bl->biv; tv; tv = tv->next_iv)
4812 if (tv->mult_val == const1_rtx
4813 && ! product_cheap_p (tv->add_val, v->mult_val))
4815 if (loop_dump_stream)
4816 fprintf (loop_dump_stream,
4817 "giv of insn %d: would need a multiply.\n",
4818 INSN_UID (v->insn));
4826 /* Check for givs whose first use is their definition and whose
4827 last use is the definition of another giv. If so, it is likely
4828 dead and should not be used to derive another giv nor to
4830 for (v = bl->giv; v; v = v->next_iv)
4833 || (v->same && v->same->ignore))
4838 struct induction *v1;
4840 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4841 if (v->last_use == v1->insn)
4844 else if (v->giv_type == DEST_REG
4845 && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
4847 struct induction *v1;
4849 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4850 if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
4855 /* Now that we know which givs will be reduced, try to rearrange the
4856 combinations to reduce register pressure.
4857 recombine_givs calls find_life_end, which needs reg_iv_type and
4858 reg_iv_info to be valid for all pseudos. We do the necessary
4859 reallocation here since it allows to check if there are still
4860 more bivs to process. */
4861 nregs = max_reg_num ();
4862 if (nregs > reg_iv_type->num_elements)
4864 /* If there are still more bivs to process, allocate some slack
4865 space so that we're not constantly reallocating these arrays. */
4868 /* Reallocate reg_iv_type and reg_iv_info. */
4869 VARRAY_GROW (reg_iv_type, nregs);
4870 VARRAY_GROW (reg_iv_info, nregs);
4872 recombine_givs (loop, bl, unroll_p);
4874 /* Reduce each giv that we decided to reduce. */
4876 for (v = bl->giv; v; v = v->next_iv)
4878 struct induction *tv;
4879 if (! v->ignore && v->same == 0)
4881 int auto_inc_opt = 0;
4883 /* If the code for derived givs immediately below has already
4884 allocated a new_reg, we must keep it. */
4886 v->new_reg = gen_reg_rtx (v->mode);
4888 if (v->derived_from)
4890 struct induction *d = v->derived_from;
4892 /* In case d->dest_reg is not replaceable, we have
4893 to replace it in v->insn now. */
4895 d->new_reg = gen_reg_rtx (d->mode);
4897 = replace_rtx (PATTERN (v->insn), d->dest_reg, d->new_reg);
4899 = replace_rtx (PATTERN (v->insn), v->dest_reg, v->new_reg);
4900 /* For each place where the biv is incremented, add an
4901 insn to set the new, reduced reg for the giv.
4902 We used to do this only for biv_count != 1, but
4903 this fails when there is a giv after a single biv
4904 increment, e.g. when the last giv was expressed as
4906 for (tv = bl->biv; tv; tv = tv->next_iv)
4908 /* We always emit reduced giv increments before the
4909 biv increment when bl->biv_count != 1. So by
4910 emitting the add insns for derived givs after the
4911 biv increment, they pick up the updated value of
4913 If the reduced giv is processed with
4914 auto_inc_opt == 1, then it is incremented earlier
4915 than the biv, hence we'll still pick up the right
4917 If it's processed with auto_inc_opt == -1,
4918 that implies that the biv increment is before the
4919 first reduced giv's use. The derived giv's lifetime
4920 is after the reduced giv's lifetime, hence in this
4921 case, the biv increment doesn't matter. */
4922 emit_insn_after (copy_rtx (PATTERN (v->insn)), tv->insn);
4928 /* If the target has auto-increment addressing modes, and
4929 this is an address giv, then try to put the increment
4930 immediately after its use, so that flow can create an
4931 auto-increment addressing mode. */
4932 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
4933 && bl->biv->always_executed && ! bl->biv->maybe_multiple
4934 /* We don't handle reversed biv's because bl->biv->insn
4935 does not have a valid INSN_LUID. */
4937 && v->always_executed && ! v->maybe_multiple
4938 && INSN_UID (v->insn) < max_uid_for_loop)
4940 /* If other giv's have been combined with this one, then
4941 this will work only if all uses of the other giv's occur
4942 before this giv's insn. This is difficult to check.
4944 We simplify this by looking for the common case where
4945 there is one DEST_REG giv, and this giv's insn is the
4946 last use of the dest_reg of that DEST_REG giv. If the
4947 increment occurs after the address giv, then we can
4948 perform the optimization. (Otherwise, the increment
4949 would have to go before other_giv, and we would not be
4950 able to combine it with the address giv to get an
4951 auto-inc address.) */
4952 if (v->combined_with)
4954 struct induction *other_giv = 0;
4956 for (tv = bl->giv; tv; tv = tv->next_iv)
4964 if (! tv && other_giv
4965 && REGNO (other_giv->dest_reg) < max_reg_before_loop
4966 && (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
4967 == INSN_UID (v->insn))
4968 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
4971 /* Check for case where increment is before the address
4972 giv. Do this test in "loop order". */
4973 else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn)
4974 && (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
4975 || (INSN_LUID (bl->biv->insn)
4976 > INSN_LUID (loop_scan_start))))
4977 || (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
4978 && (INSN_LUID (loop_scan_start)
4979 < INSN_LUID (bl->biv->insn))))
4988 /* We can't put an insn immediately after one setting
4989 cc0, or immediately before one using cc0. */
4990 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
4991 || (auto_inc_opt == -1
4992 && (prev = prev_nonnote_insn (v->insn)) != 0
4993 && GET_RTX_CLASS (GET_CODE (prev)) == 'i'
4994 && sets_cc0_p (PATTERN (prev))))
5000 v->auto_inc_opt = 1;
5004 /* For each place where the biv is incremented, add an insn
5005 to increment the new, reduced reg for the giv. */
5006 for (tv = bl->biv; tv; tv = tv->next_iv)
5011 insert_before = tv->insn;
5012 else if (auto_inc_opt == 1)
5013 insert_before = NEXT_INSN (v->insn);
5015 insert_before = v->insn;
5017 if (tv->mult_val == const1_rtx)
5018 emit_iv_add_mult (tv->add_val, v->mult_val,
5019 v->new_reg, v->new_reg, insert_before);
5020 else /* tv->mult_val == const0_rtx */
5021 /* A multiply is acceptable here
5022 since this is presumed to be seldom executed. */
5023 emit_iv_add_mult (tv->add_val, v->mult_val,
5024 v->add_val, v->new_reg, insert_before);
5027 /* Add code at loop start to initialize giv's reduced reg. */
5029 emit_iv_add_mult (bl->initial_value, v->mult_val,
5030 v->add_val, v->new_reg, loop_start);
5034 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
5037 For each giv register that can be reduced now: if replaceable,
5038 substitute reduced reg wherever the old giv occurs;
5039 else add new move insn "giv_reg = reduced_reg". */
5041 for (v = bl->giv; v; v = v->next_iv)
5043 if (v->same && v->same->ignore)
5049 /* Update expression if this was combined, in case other giv was
5052 v->new_reg = replace_rtx (v->new_reg,
5053 v->same->dest_reg, v->same->new_reg);
5055 if (v->giv_type == DEST_ADDR)
5056 /* Store reduced reg as the address in the memref where we found
5058 validate_change (v->insn, v->location, v->new_reg, 0);
5059 else if (v->replaceable)
5061 reg_map[REGNO (v->dest_reg)] = v->new_reg;
5064 /* I can no longer duplicate the original problem. Perhaps
5065 this is unnecessary now? */
5067 /* Replaceable; it isn't strictly necessary to delete the old
5068 insn and emit a new one, because v->dest_reg is now dead.
5070 However, especially when unrolling loops, the special
5071 handling for (set REG0 REG1) in the second cse pass may
5072 make v->dest_reg live again. To avoid this problem, emit
5073 an insn to set the original giv reg from the reduced giv.
5074 We can not delete the original insn, since it may be part
5075 of a LIBCALL, and the code in flow that eliminates dead
5076 libcalls will fail if it is deleted. */
5077 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
5083 /* Not replaceable; emit an insn to set the original giv reg from
5084 the reduced giv, same as above. */
5085 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
5089 /* When a loop is reversed, givs which depend on the reversed
5090 biv, and which are live outside the loop, must be set to their
5091 correct final value. This insn is only needed if the giv is
5092 not replaceable. The correct final value is the same as the
5093 value that the giv starts the reversed loop with. */
5094 if (bl->reversed && ! v->replaceable)
5095 emit_iv_add_mult (bl->initial_value, v->mult_val,
5096 v->add_val, v->dest_reg, end_insert_before);
5097 else if (v->final_value)
5101 /* If the loop has multiple exits, emit the insn before the
5102 loop to ensure that it will always be executed no matter
5103 how the loop exits. Otherwise, emit the insn after the loop,
5104 since this is slightly more efficient. */
5105 if (loop->exit_count)
5106 insert_before = loop_start;
5108 insert_before = end_insert_before;
5109 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
5113 /* If the insn to set the final value of the giv was emitted
5114 before the loop, then we must delete the insn inside the loop
5115 that sets it. If this is a LIBCALL, then we must delete
5116 every insn in the libcall. Note, however, that
5117 final_giv_value will only succeed when there are multiple
5118 exits if the giv is dead at each exit, hence it does not
5119 matter that the original insn remains because it is dead
5121 /* Delete the insn inside the loop that sets the giv since
5122 the giv is now set before (or after) the loop. */
5123 delete_insn (v->insn);
5127 if (loop_dump_stream)
5129 fprintf (loop_dump_stream, "giv at %d reduced to ",
5130 INSN_UID (v->insn));
5131 print_rtl (loop_dump_stream, v->new_reg);
5132 fprintf (loop_dump_stream, "\n");
5136 /* All the givs based on the biv bl have been reduced if they
5139 /* For each giv not marked as maybe dead that has been combined with a
5140 second giv, clear any "maybe dead" mark on that second giv.
5141 v->new_reg will either be or refer to the register of the giv it
5144 Doing this clearing avoids problems in biv elimination where a
5145 giv's new_reg is a complex value that can't be put in the insn but
5146 the giv combined with (with a reg as new_reg) is marked maybe_dead.
5147 Since the register will be used in either case, we'd prefer it be
5148 used from the simpler giv. */
5150 for (v = bl->giv; v; v = v->next_iv)
5151 if (! v->maybe_dead && v->same)
5152 v->same->maybe_dead = 0;
5154 /* Try to eliminate the biv, if it is a candidate.
5155 This won't work if ! all_reduced,
5156 since the givs we planned to use might not have been reduced.
5158 We have to be careful that we didn't initially think we could eliminate
5159 this biv because of a giv that we now think may be dead and shouldn't
5160 be used as a biv replacement.
5162 Also, there is the possibility that we may have a giv that looks
5163 like it can be used to eliminate a biv, but the resulting insn
5164 isn't valid. This can happen, for example, on the 88k, where a
5165 JUMP_INSN can compare a register only with zero. Attempts to
5166 replace it with a compare with a constant will fail.
5168 Note that in cases where this call fails, we may have replaced some
5169 of the occurrences of the biv with a giv, but no harm was done in
5170 doing so in the rare cases where it can occur. */
5172 if (all_reduced == 1 && bl->eliminable
5173 && maybe_eliminate_biv (loop, bl, 1, threshold, insn_count))
5175 /* ?? If we created a new test to bypass the loop entirely,
5176 or otherwise drop straight in, based on this test, then
5177 we might want to rewrite it also. This way some later
5178 pass has more hope of removing the initialization of this
5181 /* If final_value != 0, then the biv may be used after loop end
5182 and we must emit an insn to set it just in case.
5184 Reversed bivs already have an insn after the loop setting their
5185 value, so we don't need another one. We can't calculate the
5186 proper final value for such a biv here anyways. */
5187 if (final_value != 0 && ! bl->reversed)
5191 /* If the loop has multiple exits, emit the insn before the
5192 loop to ensure that it will always be executed no matter
5193 how the loop exits. Otherwise, emit the insn after the
5194 loop, since this is slightly more efficient. */
5195 if (loop->exit_count)
5196 insert_before = loop_start;
5198 insert_before = end_insert_before;
5200 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
5205 /* Delete all of the instructions inside the loop which set
5206 the biv, as they are all dead. If is safe to delete them,
5207 because an insn setting a biv will never be part of a libcall. */
5208 /* However, deleting them will invalidate the regno_last_uid info,
5209 so keeping them around is more convenient. Final_biv_value
5210 will only succeed when there are multiple exits if the biv
5211 is dead at each exit, hence it does not matter that the original
5212 insn remains, because it is dead anyways. */
5213 for (v = bl->biv; v; v = v->next_iv)
5214 delete_insn (v->insn);
5217 if (loop_dump_stream)
5218 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
5223 /* Go through all the instructions in the loop, making all the
5224 register substitutions scheduled in REG_MAP. */
5226 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5227 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5228 || GET_CODE (p) == CALL_INSN)
5230 replace_regs (PATTERN (p), reg_map, reg_map_size, 0);
5231 replace_regs (REG_NOTES (p), reg_map, reg_map_size, 0);
5235 if (loop_info->n_iterations > 0)
5237 /* When we completely unroll a loop we will likely not need the increment
5238 of the loop BIV and we will not need the conditional branch at the
5240 unrolled_insn_copies = insn_count - 2;
5243 /* When we completely unroll a loop on a HAVE_cc0 machine we will not
5244 need the comparison before the conditional branch at the end of the
5246 unrolled_insn_copies -= 1;
5249 /* We'll need one copy for each loop iteration. */
5250 unrolled_insn_copies *= loop_info->n_iterations;
5252 /* A little slop to account for the ability to remove initialization
5253 code, better CSE, and other secondary benefits of completely
5254 unrolling some loops. */
5255 unrolled_insn_copies -= 1;
5257 /* Clamp the value. */
5258 if (unrolled_insn_copies < 0)
5259 unrolled_insn_copies = 0;
5262 /* Unroll loops from within strength reduction so that we can use the
5263 induction variable information that strength_reduce has already
5264 collected. Always unroll loops that would be as small or smaller
5265 unrolled than when rolled. */
5267 || (loop_info->n_iterations > 0
5268 && unrolled_insn_copies <= insn_count))
5269 unroll_loop (loop, insn_count, end_insert_before, 1);
5271 #ifdef HAVE_decrement_and_branch_on_count
5272 /* Instrument the loop with BCT insn. */
5273 if (HAVE_decrement_and_branch_on_count && bct_p
5274 && flag_branch_on_count_reg)
5276 #endif /* HAVE_decrement_and_branch_on_count */
5278 if (loop_dump_stream)
5279 fprintf (loop_dump_stream, "\n");
5282 VARRAY_FREE (reg_iv_type);
5283 VARRAY_FREE (reg_iv_info);
5284 free (reg_biv_class);
5289 /* Return 1 if X is a valid source for an initial value (or as value being
5290 compared against in an initial test).
5292 X must be either a register or constant and must not be clobbered between
5293 the current insn and the start of the loop.
5295 INSN is the insn containing X. */
5298 valid_initial_value_p (x, insn, call_seen, loop_start)
5307 /* Only consider pseudos we know about initialized in insns whose luids
5309 if (GET_CODE (x) != REG
5310 || REGNO (x) >= max_reg_before_loop)
5313 /* Don't use call-clobbered registers across a call which clobbers it. On
5314 some machines, don't use any hard registers at all. */
5315 if (REGNO (x) < FIRST_PSEUDO_REGISTER
5316 && (SMALL_REGISTER_CLASSES
5317 || (call_used_regs[REGNO (x)] && call_seen)))
5320 /* Don't use registers that have been clobbered before the start of the
5322 if (reg_set_between_p (x, insn, loop_start))
5328 /* Scan X for memory refs and check each memory address
5329 as a possible giv. INSN is the insn whose pattern X comes from.
5330 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
5331 every loop iteration. MAYBE_MULTIPLE is 1 if the insn might be executed
5332 more thanonce in each loop iteration. */
5335 find_mem_givs (loop, x, insn, not_every_iteration, maybe_multiple)
5336 const struct loop *loop;
5339 int not_every_iteration, maybe_multiple;
5342 register enum rtx_code code;
5343 register const char *fmt;
5348 code = GET_CODE (x);
5372 /* This code used to disable creating GIVs with mult_val == 1 and
5373 add_val == 0. However, this leads to lost optimizations when
5374 it comes time to combine a set of related DEST_ADDR GIVs, since
5375 this one would not be seen. */
5377 if (general_induction_var (loop, XEXP (x, 0), &src_reg, &add_val,
5378 &mult_val, 1, &benefit))
5380 /* Found one; record it. */
5382 = (struct induction *) oballoc (sizeof (struct induction));
5384 record_giv (loop, v, insn, src_reg, addr_placeholder, mult_val,
5385 add_val, benefit, DEST_ADDR, not_every_iteration,
5386 maybe_multiple, &XEXP (x, 0));
5388 v->mem_mode = GET_MODE (x);
5397 /* Recursively scan the subexpressions for other mem refs. */
5399 fmt = GET_RTX_FORMAT (code);
5400 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5402 find_mem_givs (loop, XEXP (x, i), insn, not_every_iteration,
5404 else if (fmt[i] == 'E')
5405 for (j = 0; j < XVECLEN (x, i); j++)
5406 find_mem_givs (loop, XVECEXP (x, i, j), insn, not_every_iteration,
5410 /* Fill in the data about one biv update.
5411 V is the `struct induction' in which we record the biv. (It is
5412 allocated by the caller, with alloca.)
5413 INSN is the insn that sets it.
5414 DEST_REG is the biv's reg.
5416 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
5417 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
5418 being set to INC_VAL.
5420 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
5421 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
5422 can be executed more than once per iteration. If MAYBE_MULTIPLE
5423 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
5424 executed exactly once per iteration. */
5427 record_biv (v, insn, dest_reg, inc_val, mult_val, location,
5428 not_every_iteration, maybe_multiple, multi_insn_incr)
5429 struct induction *v;
5435 int not_every_iteration;
5437 int multi_insn_incr;
5439 struct iv_class *bl;
5442 v->src_reg = dest_reg;
5443 v->dest_reg = dest_reg;
5444 v->mult_val = mult_val;
5445 v->add_val = inc_val;
5446 v->location = location;
5447 v->mode = GET_MODE (dest_reg);
5448 v->always_computable = ! not_every_iteration;
5449 v->always_executed = ! not_every_iteration;
5450 v->maybe_multiple = maybe_multiple;
5451 v->multi_insn_incr = multi_insn_incr;
5453 /* Add this to the reg's iv_class, creating a class
5454 if this is the first incrementation of the reg. */
5456 bl = reg_biv_class[REGNO (dest_reg)];
5459 /* Create and initialize new iv_class. */
5461 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
5463 bl->regno = REGNO (dest_reg);
5469 /* Set initial value to the reg itself. */
5470 bl->initial_value = dest_reg;
5471 /* We haven't seen the initializing insn yet */
5474 bl->initial_test = 0;
5475 bl->incremented = 0;
5479 bl->total_benefit = 0;
5481 /* Add this class to loop_iv_list. */
5482 bl->next = loop_iv_list;
5485 /* Put it in the array of biv register classes. */
5486 reg_biv_class[REGNO (dest_reg)] = bl;
5489 /* Update IV_CLASS entry for this biv. */
5490 v->next_iv = bl->biv;
5493 if (mult_val == const1_rtx)
5494 bl->incremented = 1;
5496 if (loop_dump_stream)
5498 fprintf (loop_dump_stream,
5499 "Insn %d: possible biv, reg %d,",
5500 INSN_UID (insn), REGNO (dest_reg));
5501 if (GET_CODE (inc_val) == CONST_INT)
5503 fprintf (loop_dump_stream, " const =");
5504 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (inc_val));
5505 fputc ('\n', loop_dump_stream);
5509 fprintf (loop_dump_stream, " const = ");
5510 print_rtl (loop_dump_stream, inc_val);
5511 fprintf (loop_dump_stream, "\n");
5516 /* Fill in the data about one giv.
5517 V is the `struct induction' in which we record the giv. (It is
5518 allocated by the caller, with alloca.)
5519 INSN is the insn that sets it.
5520 BENEFIT estimates the savings from deleting this insn.
5521 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
5522 into a register or is used as a memory address.
5524 SRC_REG is the biv reg which the giv is computed from.
5525 DEST_REG is the giv's reg (if the giv is stored in a reg).
5526 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
5527 LOCATION points to the place where this giv's value appears in INSN. */
5530 record_giv (loop, v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
5531 type, not_every_iteration, maybe_multiple, location)
5532 const struct loop *loop;
5533 struct induction *v;
5537 rtx mult_val, add_val;
5540 int not_every_iteration, maybe_multiple;
5543 struct induction *b;
5544 struct iv_class *bl;
5545 rtx set = single_set (insn);
5548 v->src_reg = src_reg;
5550 v->dest_reg = dest_reg;
5551 v->mult_val = mult_val;
5552 v->add_val = add_val;
5553 v->benefit = benefit;
5554 v->location = location;
5556 v->combined_with = 0;
5557 v->maybe_multiple = maybe_multiple;
5558 v->multi_insn_incr = 0;
5560 v->derive_adjustment = 0;
5566 v->auto_inc_opt = 0;
5569 v->derived_from = 0;
5572 /* The v->always_computable field is used in update_giv_derive, to
5573 determine whether a giv can be used to derive another giv. For a
5574 DEST_REG giv, INSN computes a new value for the giv, so its value
5575 isn't computable if INSN insn't executed every iteration.
5576 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
5577 it does not compute a new value. Hence the value is always computable
5578 regardless of whether INSN is executed each iteration. */
5580 if (type == DEST_ADDR)
5581 v->always_computable = 1;
5583 v->always_computable = ! not_every_iteration;
5585 v->always_executed = ! not_every_iteration;
5587 if (type == DEST_ADDR)
5589 v->mode = GET_MODE (*location);
5592 else /* type == DEST_REG */
5594 v->mode = GET_MODE (SET_DEST (set));
5596 v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5597 - uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]);
5599 /* If the lifetime is zero, it means that this register is
5600 really a dead store. So mark this as a giv that can be
5601 ignored. This will not prevent the biv from being eliminated. */
5602 if (v->lifetime == 0)
5605 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
5606 REG_IV_INFO (REGNO (dest_reg)) = v;
5609 /* Add the giv to the class of givs computed from one biv. */
5611 bl = reg_biv_class[REGNO (src_reg)];
5614 v->next_iv = bl->giv;
5616 /* Don't count DEST_ADDR. This is supposed to count the number of
5617 insns that calculate givs. */
5618 if (type == DEST_REG)
5620 bl->total_benefit += benefit;
5623 /* Fatal error, biv missing for this giv? */
5626 if (type == DEST_ADDR)
5630 /* The giv can be replaced outright by the reduced register only if all
5631 of the following conditions are true:
5632 - the insn that sets the giv is always executed on any iteration
5633 on which the giv is used at all
5634 (there are two ways to deduce this:
5635 either the insn is executed on every iteration,
5636 or all uses follow that insn in the same basic block),
5637 - the giv is not used outside the loop
5638 - no assignments to the biv occur during the giv's lifetime. */
5640 if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
5641 /* Previous line always fails if INSN was moved by loop opt. */
5642 && uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5643 < INSN_LUID (loop->end)
5644 && (! not_every_iteration
5645 || last_use_this_basic_block (dest_reg, insn)))
5647 /* Now check that there are no assignments to the biv within the
5648 giv's lifetime. This requires two separate checks. */
5650 /* Check each biv update, and fail if any are between the first
5651 and last use of the giv.
5653 If this loop contains an inner loop that was unrolled, then
5654 the insn modifying the biv may have been emitted by the loop
5655 unrolling code, and hence does not have a valid luid. Just
5656 mark the biv as not replaceable in this case. It is not very
5657 useful as a biv, because it is used in two different loops.
5658 It is very unlikely that we would be able to optimize the giv
5659 using this biv anyways. */
5662 for (b = bl->biv; b; b = b->next_iv)
5664 if (INSN_UID (b->insn) >= max_uid_for_loop
5665 || ((uid_luid[INSN_UID (b->insn)]
5666 >= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))])
5667 && (uid_luid[INSN_UID (b->insn)]
5668 <= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))])))
5671 v->not_replaceable = 1;
5676 /* If there are any backwards branches that go from after the
5677 biv update to before it, then this giv is not replaceable. */
5679 for (b = bl->biv; b; b = b->next_iv)
5680 if (back_branch_in_range_p (loop, b->insn))
5683 v->not_replaceable = 1;
5689 /* May still be replaceable, we don't have enough info here to
5692 v->not_replaceable = 0;
5696 /* Record whether the add_val contains a const_int, for later use by
5701 v->no_const_addval = 1;
5702 if (tem == const0_rtx)
5704 else if (GET_CODE (tem) == CONST_INT)
5705 v->no_const_addval = 0;
5706 else if (GET_CODE (tem) == PLUS)
5710 if (GET_CODE (XEXP (tem, 0)) == PLUS)
5711 tem = XEXP (tem, 0);
5712 else if (GET_CODE (XEXP (tem, 1)) == PLUS)
5713 tem = XEXP (tem, 1);
5717 if (GET_CODE (XEXP (tem, 1)) == CONST_INT)
5718 v->no_const_addval = 0;
5722 if (loop_dump_stream)
5724 if (type == DEST_REG)
5725 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
5726 INSN_UID (insn), REGNO (dest_reg));
5728 fprintf (loop_dump_stream, "Insn %d: dest address",
5731 fprintf (loop_dump_stream, " src reg %d benefit %d",
5732 REGNO (src_reg), v->benefit);
5733 fprintf (loop_dump_stream, " lifetime %d",
5737 fprintf (loop_dump_stream, " replaceable");
5739 if (v->no_const_addval)
5740 fprintf (loop_dump_stream, " ncav");
5742 if (GET_CODE (mult_val) == CONST_INT)
5744 fprintf (loop_dump_stream, " mult ");
5745 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (mult_val));
5749 fprintf (loop_dump_stream, " mult ");
5750 print_rtl (loop_dump_stream, mult_val);
5753 if (GET_CODE (add_val) == CONST_INT)
5755 fprintf (loop_dump_stream, " add ");
5756 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (add_val));
5760 fprintf (loop_dump_stream, " add ");
5761 print_rtl (loop_dump_stream, add_val);
5765 if (loop_dump_stream)
5766 fprintf (loop_dump_stream, "\n");
5771 /* All this does is determine whether a giv can be made replaceable because
5772 its final value can be calculated. This code can not be part of record_giv
5773 above, because final_giv_value requires that the number of loop iterations
5774 be known, and that can not be accurately calculated until after all givs
5775 have been identified. */
5778 check_final_value (loop, v)
5779 const struct loop *loop;
5780 struct induction *v;
5782 struct iv_class *bl;
5783 rtx final_value = 0;
5785 bl = reg_biv_class[REGNO (v->src_reg)];
5787 /* DEST_ADDR givs will never reach here, because they are always marked
5788 replaceable above in record_giv. */
5790 /* The giv can be replaced outright by the reduced register only if all
5791 of the following conditions are true:
5792 - the insn that sets the giv is always executed on any iteration
5793 on which the giv is used at all
5794 (there are two ways to deduce this:
5795 either the insn is executed on every iteration,
5796 or all uses follow that insn in the same basic block),
5797 - its final value can be calculated (this condition is different
5798 than the one above in record_giv)
5799 - no assignments to the biv occur during the giv's lifetime. */
5802 /* This is only called now when replaceable is known to be false. */
5803 /* Clear replaceable, so that it won't confuse final_giv_value. */
5807 if ((final_value = final_giv_value (loop, v))
5808 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
5810 int biv_increment_seen = 0;
5816 /* When trying to determine whether or not a biv increment occurs
5817 during the lifetime of the giv, we can ignore uses of the variable
5818 outside the loop because final_value is true. Hence we can not
5819 use regno_last_uid and regno_first_uid as above in record_giv. */
5821 /* Search the loop to determine whether any assignments to the
5822 biv occur during the giv's lifetime. Start with the insn
5823 that sets the giv, and search around the loop until we come
5824 back to that insn again.
5826 Also fail if there is a jump within the giv's lifetime that jumps
5827 to somewhere outside the lifetime but still within the loop. This
5828 catches spaghetti code where the execution order is not linear, and
5829 hence the above test fails. Here we assume that the giv lifetime
5830 does not extend from one iteration of the loop to the next, so as
5831 to make the test easier. Since the lifetime isn't known yet,
5832 this requires two loops. See also record_giv above. */
5834 last_giv_use = v->insn;
5840 p = NEXT_INSN (loop->start);
5844 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5845 || GET_CODE (p) == CALL_INSN)
5847 if (biv_increment_seen)
5849 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5852 v->not_replaceable = 1;
5856 else if (reg_set_p (v->src_reg, PATTERN (p)))
5857 biv_increment_seen = 1;
5858 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5863 /* Now that the lifetime of the giv is known, check for branches
5864 from within the lifetime to outside the lifetime if it is still
5874 p = NEXT_INSN (loop->start);
5875 if (p == last_giv_use)
5878 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
5879 && LABEL_NAME (JUMP_LABEL (p))
5880 && ((loop_insn_first_p (JUMP_LABEL (p), v->insn)
5881 && loop_insn_first_p (loop->start, JUMP_LABEL (p)))
5882 || (loop_insn_first_p (last_giv_use, JUMP_LABEL (p))
5883 && loop_insn_first_p (JUMP_LABEL (p), loop->end))))
5886 v->not_replaceable = 1;
5888 if (loop_dump_stream)
5889 fprintf (loop_dump_stream,
5890 "Found branch outside giv lifetime.\n");
5897 /* If it is replaceable, then save the final value. */
5899 v->final_value = final_value;
5902 if (loop_dump_stream && v->replaceable)
5903 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
5904 INSN_UID (v->insn), REGNO (v->dest_reg));
5907 /* Update the status of whether a giv can derive other givs.
5909 We need to do something special if there is or may be an update to the biv
5910 between the time the giv is defined and the time it is used to derive
5913 In addition, a giv that is only conditionally set is not allowed to
5914 derive another giv once a label has been passed.
5916 The cases we look at are when a label or an update to a biv is passed. */
5919 update_giv_derive (loop, p)
5920 const struct loop *loop;
5923 struct iv_class *bl;
5924 struct induction *biv, *giv;
5928 /* Search all IV classes, then all bivs, and finally all givs.
5930 There are three cases we are concerned with. First we have the situation
5931 of a giv that is only updated conditionally. In that case, it may not
5932 derive any givs after a label is passed.
5934 The second case is when a biv update occurs, or may occur, after the
5935 definition of a giv. For certain biv updates (see below) that are
5936 known to occur between the giv definition and use, we can adjust the
5937 giv definition. For others, or when the biv update is conditional,
5938 we must prevent the giv from deriving any other givs. There are two
5939 sub-cases within this case.
5941 If this is a label, we are concerned with any biv update that is done
5942 conditionally, since it may be done after the giv is defined followed by
5943 a branch here (actually, we need to pass both a jump and a label, but
5944 this extra tracking doesn't seem worth it).
5946 If this is a jump, we are concerned about any biv update that may be
5947 executed multiple times. We are actually only concerned about
5948 backward jumps, but it is probably not worth performing the test
5949 on the jump again here.
5951 If this is a biv update, we must adjust the giv status to show that a
5952 subsequent biv update was performed. If this adjustment cannot be done,
5953 the giv cannot derive further givs. */
5955 for (bl = loop_iv_list; bl; bl = bl->next)
5956 for (biv = bl->biv; biv; biv = biv->next_iv)
5957 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
5960 for (giv = bl->giv; giv; giv = giv->next_iv)
5962 /* If cant_derive is already true, there is no point in
5963 checking all of these conditions again. */
5964 if (giv->cant_derive)
5967 /* If this giv is conditionally set and we have passed a label,
5968 it cannot derive anything. */
5969 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
5970 giv->cant_derive = 1;
5972 /* Skip givs that have mult_val == 0, since
5973 they are really invariants. Also skip those that are
5974 replaceable, since we know their lifetime doesn't contain
5976 else if (giv->mult_val == const0_rtx || giv->replaceable)
5979 /* The only way we can allow this giv to derive another
5980 is if this is a biv increment and we can form the product
5981 of biv->add_val and giv->mult_val. In this case, we will
5982 be able to compute a compensation. */
5983 else if (biv->insn == p)
5987 if (biv->mult_val == const1_rtx)
5988 tem = simplify_giv_expr (loop,
5989 gen_rtx_MULT (giv->mode,
5994 if (tem && giv->derive_adjustment)
5995 tem = simplify_giv_expr
5997 gen_rtx_PLUS (giv->mode, tem, giv->derive_adjustment),
6001 giv->derive_adjustment = tem;
6003 giv->cant_derive = 1;
6005 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
6006 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
6007 giv->cant_derive = 1;
6012 /* Check whether an insn is an increment legitimate for a basic induction var.
6013 X is the source of insn P, or a part of it.
6014 MODE is the mode in which X should be interpreted.
6016 DEST_REG is the putative biv, also the destination of the insn.
6017 We accept patterns of these forms:
6018 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
6019 REG = INVARIANT + REG
6021 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
6022 store the additive term into *INC_VAL, and store the place where
6023 we found the additive term into *LOCATION.
6025 If X is an assignment of an invariant into DEST_REG, we set
6026 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
6028 We also want to detect a BIV when it corresponds to a variable
6029 whose mode was promoted via PROMOTED_MODE. In that case, an increment
6030 of the variable may be a PLUS that adds a SUBREG of that variable to
6031 an invariant and then sign- or zero-extends the result of the PLUS
6034 Most GIVs in such cases will be in the promoted mode, since that is the
6035 probably the natural computation mode (and almost certainly the mode
6036 used for addresses) on the machine. So we view the pseudo-reg containing
6037 the variable as the BIV, as if it were simply incremented.
6039 Note that treating the entire pseudo as a BIV will result in making
6040 simple increments to any GIVs based on it. However, if the variable
6041 overflows in its declared mode but not its promoted mode, the result will
6042 be incorrect. This is acceptable if the variable is signed, since
6043 overflows in such cases are undefined, but not if it is unsigned, since
6044 those overflows are defined. So we only check for SIGN_EXTEND and
6047 If we cannot find a biv, we return 0. */
6050 basic_induction_var (loop, x, mode, dest_reg, p, inc_val, mult_val,
6051 location, multi_insn_incr)
6052 const struct loop *loop;
6054 enum machine_mode mode;
6060 int *multi_insn_incr;
6062 register enum rtx_code code;
6066 code = GET_CODE (x);
6071 if (rtx_equal_p (XEXP (x, 0), dest_reg)
6072 || (GET_CODE (XEXP (x, 0)) == SUBREG
6073 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
6074 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
6076 argp = &XEXP (x, 1);
6078 else if (rtx_equal_p (XEXP (x, 1), dest_reg)
6079 || (GET_CODE (XEXP (x, 1)) == SUBREG
6080 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
6081 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
6083 argp = &XEXP (x, 0);
6089 if (loop_invariant_p (loop, arg) != 1)
6092 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
6093 *mult_val = const1_rtx;
6098 /* If this is a SUBREG for a promoted variable, check the inner
6100 if (SUBREG_PROMOTED_VAR_P (x))
6101 return basic_induction_var (loop, SUBREG_REG (x),
6102 GET_MODE (SUBREG_REG (x)),
6103 dest_reg, p, inc_val, mult_val, location,
6108 /* If this register is assigned in a previous insn, look at its
6109 source, but don't go outside the loop or past a label. */
6115 insn = PREV_INSN (insn);
6116 } while (insn && GET_CODE (insn) == NOTE
6117 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6121 set = single_set (insn);
6125 if ((SET_DEST (set) == x
6126 || (GET_CODE (SET_DEST (set)) == SUBREG
6127 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
6129 && (GET_MODE_CLASS (GET_MODE (SET_DEST (set)))
6131 && SUBREG_REG (SET_DEST (set)) == x))
6132 && basic_induction_var (loop, SET_SRC (set),
6133 (GET_MODE (SET_SRC (set)) == VOIDmode
6135 : GET_MODE (SET_SRC (set))),
6137 inc_val, mult_val, location,
6140 *multi_insn_incr = 1;
6144 /* ... fall through ... */
6146 /* Can accept constant setting of biv only when inside inner most loop.
6147 Otherwise, a biv of an inner loop may be incorrectly recognized
6148 as a biv of the outer loop,
6149 causing code to be moved INTO the inner loop. */
6151 if (loop_invariant_p (loop, x) != 1)
6156 /* convert_modes aborts if we try to convert to or from CCmode, so just
6157 exclude that case. It is very unlikely that a condition code value
6158 would be a useful iterator anyways. */
6159 if (loop->level == 1
6160 && GET_MODE_CLASS (mode) != MODE_CC
6161 && GET_MODE_CLASS (GET_MODE (dest_reg)) != MODE_CC)
6163 /* Possible bug here? Perhaps we don't know the mode of X. */
6164 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
6165 *mult_val = const0_rtx;
6172 return basic_induction_var (loop, XEXP (x, 0), GET_MODE (XEXP (x, 0)),
6173 dest_reg, p, inc_val, mult_val, location,
6177 /* Similar, since this can be a sign extension. */
6178 for (insn = PREV_INSN (p);
6179 (insn && GET_CODE (insn) == NOTE
6180 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6181 insn = PREV_INSN (insn))
6185 set = single_set (insn);
6187 if (set && SET_DEST (set) == XEXP (x, 0)
6188 && GET_CODE (XEXP (x, 1)) == CONST_INT
6189 && INTVAL (XEXP (x, 1)) >= 0
6190 && GET_CODE (SET_SRC (set)) == ASHIFT
6191 && XEXP (x, 1) == XEXP (SET_SRC (set), 1)
6192 && basic_induction_var (loop, XEXP (SET_SRC (set), 0),
6193 GET_MODE (XEXP (x, 0)),
6194 dest_reg, insn, inc_val, mult_val,
6195 location, multi_insn_incr))
6197 *multi_insn_incr = 1;
6207 /* A general induction variable (giv) is any quantity that is a linear
6208 function of a basic induction variable,
6209 i.e. giv = biv * mult_val + add_val.
6210 The coefficients can be any loop invariant quantity.
6211 A giv need not be computed directly from the biv;
6212 it can be computed by way of other givs. */
6214 /* Determine whether X computes a giv.
6215 If it does, return a nonzero value
6216 which is the benefit from eliminating the computation of X;
6217 set *SRC_REG to the register of the biv that it is computed from;
6218 set *ADD_VAL and *MULT_VAL to the coefficients,
6219 such that the value of X is biv * mult + add; */
6222 general_induction_var (loop, x, src_reg, add_val, mult_val, is_addr, pbenefit)
6223 const struct loop *loop;
6234 /* If this is an invariant, forget it, it isn't a giv. */
6235 if (loop_invariant_p (loop, x) == 1)
6238 /* See if the expression could be a giv and get its form.
6239 Mark our place on the obstack in case we don't find a giv. */
6240 storage = (char *) oballoc (0);
6242 x = simplify_giv_expr (loop, x, pbenefit);
6249 switch (GET_CODE (x))
6253 /* Since this is now an invariant and wasn't before, it must be a giv
6254 with MULT_VAL == 0. It doesn't matter which BIV we associate this
6256 *src_reg = loop_iv_list->biv->dest_reg;
6257 *mult_val = const0_rtx;
6262 /* This is equivalent to a BIV. */
6264 *mult_val = const1_rtx;
6265 *add_val = const0_rtx;
6269 /* Either (plus (biv) (invar)) or
6270 (plus (mult (biv) (invar_1)) (invar_2)). */
6271 if (GET_CODE (XEXP (x, 0)) == MULT)
6273 *src_reg = XEXP (XEXP (x, 0), 0);
6274 *mult_val = XEXP (XEXP (x, 0), 1);
6278 *src_reg = XEXP (x, 0);
6279 *mult_val = const1_rtx;
6281 *add_val = XEXP (x, 1);
6285 /* ADD_VAL is zero. */
6286 *src_reg = XEXP (x, 0);
6287 *mult_val = XEXP (x, 1);
6288 *add_val = const0_rtx;
6295 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
6296 unless they are CONST_INT). */
6297 if (GET_CODE (*add_val) == USE)
6298 *add_val = XEXP (*add_val, 0);
6299 if (GET_CODE (*mult_val) == USE)
6300 *mult_val = XEXP (*mult_val, 0);
6305 *pbenefit += ADDRESS_COST (orig_x) - reg_address_cost;
6307 *pbenefit += rtx_cost (orig_x, MEM) - reg_address_cost;
6311 *pbenefit += rtx_cost (orig_x, SET);
6313 /* Always return true if this is a giv so it will be detected as such,
6314 even if the benefit is zero or negative. This allows elimination
6315 of bivs that might otherwise not be eliminated. */
6319 /* Given an expression, X, try to form it as a linear function of a biv.
6320 We will canonicalize it to be of the form
6321 (plus (mult (BIV) (invar_1))
6323 with possible degeneracies.
6325 The invariant expressions must each be of a form that can be used as a
6326 machine operand. We surround then with a USE rtx (a hack, but localized
6327 and certainly unambiguous!) if not a CONST_INT for simplicity in this
6328 routine; it is the caller's responsibility to strip them.
6330 If no such canonicalization is possible (i.e., two biv's are used or an
6331 expression that is neither invariant nor a biv or giv), this routine
6334 For a non-zero return, the result will have a code of CONST_INT, USE,
6335 REG (for a BIV), PLUS, or MULT. No other codes will occur.
6337 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
6339 static rtx sge_plus PARAMS ((enum machine_mode, rtx, rtx));
6340 static rtx sge_plus_constant PARAMS ((rtx, rtx));
6341 static int cmp_combine_givs_stats PARAMS ((const PTR, const PTR));
6342 static int cmp_recombine_givs_stats PARAMS ((const PTR, const PTR));
6345 simplify_giv_expr (loop, x, benefit)
6346 const struct loop *loop;
6350 enum machine_mode mode = GET_MODE (x);
6354 /* If this is not an integer mode, or if we cannot do arithmetic in this
6355 mode, this can't be a giv. */
6356 if (mode != VOIDmode
6357 && (GET_MODE_CLASS (mode) != MODE_INT
6358 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
6361 switch (GET_CODE (x))
6364 arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
6365 arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
6366 if (arg0 == 0 || arg1 == 0)
6369 /* Put constant last, CONST_INT last if both constant. */
6370 if ((GET_CODE (arg0) == USE
6371 || GET_CODE (arg0) == CONST_INT)
6372 && ! ((GET_CODE (arg0) == USE
6373 && GET_CODE (arg1) == USE)
6374 || GET_CODE (arg1) == CONST_INT))
6375 tem = arg0, arg0 = arg1, arg1 = tem;
6377 /* Handle addition of zero, then addition of an invariant. */
6378 if (arg1 == const0_rtx)
6380 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
6381 switch (GET_CODE (arg0))
6385 /* Adding two invariants must result in an invariant, so enclose
6386 addition operation inside a USE and return it. */
6387 if (GET_CODE (arg0) == USE)
6388 arg0 = XEXP (arg0, 0);
6389 if (GET_CODE (arg1) == USE)
6390 arg1 = XEXP (arg1, 0);
6392 if (GET_CODE (arg0) == CONST_INT)
6393 tem = arg0, arg0 = arg1, arg1 = tem;
6394 if (GET_CODE (arg1) == CONST_INT)
6395 tem = sge_plus_constant (arg0, arg1);
6397 tem = sge_plus (mode, arg0, arg1);
6399 if (GET_CODE (tem) != CONST_INT)
6400 tem = gen_rtx_USE (mode, tem);
6405 /* biv + invar or mult + invar. Return sum. */
6406 return gen_rtx_PLUS (mode, arg0, arg1);
6409 /* (a + invar_1) + invar_2. Associate. */
6411 simplify_giv_expr (loop,
6423 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
6424 MULT to reduce cases. */
6425 if (GET_CODE (arg0) == REG)
6426 arg0 = gen_rtx_MULT (mode, arg0, const1_rtx);
6427 if (GET_CODE (arg1) == REG)
6428 arg1 = gen_rtx_MULT (mode, arg1, const1_rtx);
6430 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
6431 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
6432 Recurse to associate the second PLUS. */
6433 if (GET_CODE (arg1) == MULT)
6434 tem = arg0, arg0 = arg1, arg1 = tem;
6436 if (GET_CODE (arg1) == PLUS)
6438 simplify_giv_expr (loop,
6440 gen_rtx_PLUS (mode, arg0,
6445 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
6446 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
6449 if (!rtx_equal_p (arg0, arg1))
6452 return simplify_giv_expr (loop,
6461 /* Handle "a - b" as "a + b * (-1)". */
6462 return simplify_giv_expr (loop,
6471 arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
6472 arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
6473 if (arg0 == 0 || arg1 == 0)
6476 /* Put constant last, CONST_INT last if both constant. */
6477 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
6478 && GET_CODE (arg1) != CONST_INT)
6479 tem = arg0, arg0 = arg1, arg1 = tem;
6481 /* If second argument is not now constant, not giv. */
6482 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
6485 /* Handle multiply by 0 or 1. */
6486 if (arg1 == const0_rtx)
6489 else if (arg1 == const1_rtx)
6492 switch (GET_CODE (arg0))
6495 /* biv * invar. Done. */
6496 return gen_rtx_MULT (mode, arg0, arg1);
6499 /* Product of two constants. */
6500 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
6503 /* invar * invar. It is a giv, but very few of these will
6504 actually pay off, so limit to simple registers. */
6505 if (GET_CODE (arg1) != CONST_INT)
6508 arg0 = XEXP (arg0, 0);
6509 if (GET_CODE (arg0) == REG)
6510 tem = gen_rtx_MULT (mode, arg0, arg1);
6511 else if (GET_CODE (arg0) == MULT
6512 && GET_CODE (XEXP (arg0, 0)) == REG
6513 && GET_CODE (XEXP (arg0, 1)) == CONST_INT)
6515 tem = gen_rtx_MULT (mode, XEXP (arg0, 0),
6516 GEN_INT (INTVAL (XEXP (arg0, 1))
6521 return gen_rtx_USE (mode, tem);
6524 /* (a * invar_1) * invar_2. Associate. */
6525 return simplify_giv_expr (loop,
6534 /* (a + invar_1) * invar_2. Distribute. */
6535 return simplify_giv_expr (loop,
6550 /* Shift by constant is multiply by power of two. */
6551 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6555 simplify_giv_expr (loop,
6558 GEN_INT ((HOST_WIDE_INT) 1
6559 << INTVAL (XEXP (x, 1)))),
6563 /* "-a" is "a * (-1)" */
6564 return simplify_giv_expr (loop,
6565 gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx),
6569 /* "~a" is "-a - 1". Silly, but easy. */
6570 return simplify_giv_expr (loop,
6571 gen_rtx_MINUS (mode,
6572 gen_rtx_NEG (mode, XEXP (x, 0)),
6577 /* Already in proper form for invariant. */
6581 /* If this is a new register, we can't deal with it. */
6582 if (REGNO (x) >= max_reg_before_loop)
6585 /* Check for biv or giv. */
6586 switch (REG_IV_TYPE (REGNO (x)))
6590 case GENERAL_INDUCT:
6592 struct induction *v = REG_IV_INFO (REGNO (x));
6594 /* Form expression from giv and add benefit. Ensure this giv
6595 can derive another and subtract any needed adjustment if so. */
6596 *benefit += v->benefit;
6600 tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode,
6601 v->src_reg, v->mult_val),
6604 if (v->derive_adjustment)
6605 tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment);
6606 return simplify_giv_expr (loop, tem, benefit);
6610 /* If it isn't an induction variable, and it is invariant, we
6611 may be able to simplify things further by looking through
6612 the bits we just moved outside the loop. */
6613 if (loop_invariant_p (loop, x) == 1)
6617 for (m = the_movables; m ; m = m->next)
6618 if (rtx_equal_p (x, m->set_dest))
6620 /* Ok, we found a match. Substitute and simplify. */
6622 /* If we match another movable, we must use that, as
6623 this one is going away. */
6625 return simplify_giv_expr (loop, m->match->set_dest,
6628 /* If consec is non-zero, this is a member of a group of
6629 instructions that were moved together. We handle this
6630 case only to the point of seeking to the last insn and
6631 looking for a REG_EQUAL. Fail if we don't find one. */
6636 do { tem = NEXT_INSN (tem); } while (--i > 0);
6638 tem = find_reg_note (tem, REG_EQUAL, NULL_RTX);
6640 tem = XEXP (tem, 0);
6644 tem = single_set (m->insn);
6646 tem = SET_SRC (tem);
6651 /* What we are most interested in is pointer
6652 arithmetic on invariants -- only take
6653 patterns we may be able to do something with. */
6654 if (GET_CODE (tem) == PLUS
6655 || GET_CODE (tem) == MULT
6656 || GET_CODE (tem) == ASHIFT
6657 || GET_CODE (tem) == CONST_INT
6658 || GET_CODE (tem) == SYMBOL_REF)
6660 tem = simplify_giv_expr (loop, tem, benefit);
6664 else if (GET_CODE (tem) == CONST
6665 && GET_CODE (XEXP (tem, 0)) == PLUS
6666 && GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF
6667 && GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT)
6669 tem = simplify_giv_expr (loop, XEXP (tem, 0),
6681 /* Fall through to general case. */
6683 /* If invariant, return as USE (unless CONST_INT).
6684 Otherwise, not giv. */
6685 if (GET_CODE (x) == USE)
6688 if (loop_invariant_p (loop, x) == 1)
6690 if (GET_CODE (x) == CONST_INT)
6692 if (GET_CODE (x) == CONST
6693 && GET_CODE (XEXP (x, 0)) == PLUS
6694 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
6695 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
6697 return gen_rtx_USE (mode, x);
6704 /* This routine folds invariants such that there is only ever one
6705 CONST_INT in the summation. It is only used by simplify_giv_expr. */
6708 sge_plus_constant (x, c)
6711 if (GET_CODE (x) == CONST_INT)
6712 return GEN_INT (INTVAL (x) + INTVAL (c));
6713 else if (GET_CODE (x) != PLUS)
6714 return gen_rtx_PLUS (GET_MODE (x), x, c);
6715 else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6717 return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
6718 GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c)));
6720 else if (GET_CODE (XEXP (x, 0)) == PLUS
6721 || GET_CODE (XEXP (x, 1)) != PLUS)
6723 return gen_rtx_PLUS (GET_MODE (x),
6724 sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1));
6728 return gen_rtx_PLUS (GET_MODE (x),
6729 sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0));
6734 sge_plus (mode, x, y)
6735 enum machine_mode mode;
6738 while (GET_CODE (y) == PLUS)
6740 rtx a = XEXP (y, 0);
6741 if (GET_CODE (a) == CONST_INT)
6742 x = sge_plus_constant (x, a);
6744 x = gen_rtx_PLUS (mode, x, a);
6747 if (GET_CODE (y) == CONST_INT)
6748 x = sge_plus_constant (x, y);
6750 x = gen_rtx_PLUS (mode, x, y);
6754 /* Help detect a giv that is calculated by several consecutive insns;
6758 The caller has already identified the first insn P as having a giv as dest;
6759 we check that all other insns that set the same register follow
6760 immediately after P, that they alter nothing else,
6761 and that the result of the last is still a giv.
6763 The value is 0 if the reg set in P is not really a giv.
6764 Otherwise, the value is the amount gained by eliminating
6765 all the consecutive insns that compute the value.
6767 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
6768 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
6770 The coefficients of the ultimate giv value are stored in
6771 *MULT_VAL and *ADD_VAL. */
6774 consec_sets_giv (loop, first_benefit, p, src_reg, dest_reg,
6775 add_val, mult_val, last_consec_insn)
6776 const struct loop *loop;
6783 rtx *last_consec_insn;
6791 /* Indicate that this is a giv so that we can update the value produced in
6792 each insn of the multi-insn sequence.
6794 This induction structure will be used only by the call to
6795 general_induction_var below, so we can allocate it on our stack.
6796 If this is a giv, our caller will replace the induct var entry with
6797 a new induction structure. */
6799 = (struct induction *) alloca (sizeof (struct induction));
6800 v->src_reg = src_reg;
6801 v->mult_val = *mult_val;
6802 v->add_val = *add_val;
6803 v->benefit = first_benefit;
6805 v->derive_adjustment = 0;
6807 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
6808 REG_IV_INFO (REGNO (dest_reg)) = v;
6810 count = VARRAY_INT (n_times_set, REGNO (dest_reg)) - 1;
6815 code = GET_CODE (p);
6817 /* If libcall, skip to end of call sequence. */
6818 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
6822 && (set = single_set (p))
6823 && GET_CODE (SET_DEST (set)) == REG
6824 && SET_DEST (set) == dest_reg
6825 && (general_induction_var (loop, SET_SRC (set), &src_reg,
6826 add_val, mult_val, 0, &benefit)
6827 /* Giv created by equivalent expression. */
6828 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
6829 && general_induction_var (loop, XEXP (temp, 0), &src_reg,
6830 add_val, mult_val, 0, &benefit)))
6831 && src_reg == v->src_reg)
6833 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
6834 benefit += libcall_benefit (p);
6837 v->mult_val = *mult_val;
6838 v->add_val = *add_val;
6839 v->benefit = benefit;
6841 else if (code != NOTE)
6843 /* Allow insns that set something other than this giv to a
6844 constant. Such insns are needed on machines which cannot
6845 include long constants and should not disqualify a giv. */
6847 && (set = single_set (p))
6848 && SET_DEST (set) != dest_reg
6849 && CONSTANT_P (SET_SRC (set)))
6852 REG_IV_TYPE (REGNO (dest_reg)) = UNKNOWN_INDUCT;
6857 *last_consec_insn = p;
6861 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6862 represented by G1. If no such expression can be found, or it is clear that
6863 it cannot possibly be a valid address, 0 is returned.
6865 To perform the computation, we note that
6868 where `v' is the biv.
6870 So G2 = (y/b) * G1 + (b - a*y/x).
6872 Note that MULT = y/x.
6874 Update: A and B are now allowed to be additive expressions such that
6875 B contains all variables in A. That is, computing B-A will not require
6876 subtracting variables. */
6879 express_from_1 (a, b, mult)
6882 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
6884 if (mult == const0_rtx)
6887 /* If MULT is not 1, we cannot handle A with non-constants, since we
6888 would then be required to subtract multiples of the registers in A.
6889 This is theoretically possible, and may even apply to some Fortran
6890 constructs, but it is a lot of work and we do not attempt it here. */
6892 if (mult != const1_rtx && GET_CODE (a) != CONST_INT)
6895 /* In general these structures are sorted top to bottom (down the PLUS
6896 chain), but not left to right across the PLUS. If B is a higher
6897 order giv than A, we can strip one level and recurse. If A is higher
6898 order, we'll eventually bail out, but won't know that until the end.
6899 If they are the same, we'll strip one level around this loop. */
6901 while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS)
6903 rtx ra, rb, oa, ob, tmp;
6905 ra = XEXP (a, 0), oa = XEXP (a, 1);
6906 if (GET_CODE (ra) == PLUS)
6907 tmp = ra, ra = oa, oa = tmp;
6909 rb = XEXP (b, 0), ob = XEXP (b, 1);
6910 if (GET_CODE (rb) == PLUS)
6911 tmp = rb, rb = ob, ob = tmp;
6913 if (rtx_equal_p (ra, rb))
6914 /* We matched: remove one reg completely. */
6916 else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob))
6917 /* An alternate match. */
6919 else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb))
6920 /* An alternate match. */
6924 /* Indicates an extra register in B. Strip one level from B and
6925 recurse, hoping B was the higher order expression. */
6926 ob = express_from_1 (a, ob, mult);
6929 return gen_rtx_PLUS (GET_MODE (b), rb, ob);
6933 /* Here we are at the last level of A, go through the cases hoping to
6934 get rid of everything but a constant. */
6936 if (GET_CODE (a) == PLUS)
6940 ra = XEXP (a, 0), oa = XEXP (a, 1);
6941 if (rtx_equal_p (oa, b))
6943 else if (!rtx_equal_p (ra, b))
6946 if (GET_CODE (oa) != CONST_INT)
6949 return GEN_INT (-INTVAL (oa) * INTVAL (mult));
6951 else if (GET_CODE (a) == CONST_INT)
6953 return plus_constant (b, -INTVAL (a) * INTVAL (mult));
6955 else if (GET_CODE (b) == PLUS)
6957 if (rtx_equal_p (a, XEXP (b, 0)))
6959 else if (rtx_equal_p (a, XEXP (b, 1)))
6964 else if (rtx_equal_p (a, b))
6971 express_from (g1, g2)
6972 struct induction *g1, *g2;
6976 /* The value that G1 will be multiplied by must be a constant integer. Also,
6977 the only chance we have of getting a valid address is if b*c/a (see above
6978 for notation) is also an integer. */
6979 if (GET_CODE (g1->mult_val) == CONST_INT
6980 && GET_CODE (g2->mult_val) == CONST_INT)
6982 if (g1->mult_val == const0_rtx
6983 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
6985 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
6987 else if (rtx_equal_p (g1->mult_val, g2->mult_val))
6991 /* ??? Find out if the one is a multiple of the other? */
6995 add = express_from_1 (g1->add_val, g2->add_val, mult);
6996 if (add == NULL_RTX)
6998 /* Failed. If we've got a multiplication factor between G1 and G2,
6999 scale G1's addend and try again. */
7000 if (INTVAL (mult) > 1)
7002 rtx g1_add_val = g1->add_val;
7003 if (GET_CODE (g1_add_val) == MULT
7004 && GET_CODE (XEXP (g1_add_val, 1)) == CONST_INT)
7007 m = INTVAL (mult) * INTVAL (XEXP (g1_add_val, 1));
7008 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val),
7009 XEXP (g1_add_val, 0), GEN_INT (m));
7013 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val), g1_add_val,
7017 add = express_from_1 (g1_add_val, g2->add_val, const1_rtx);
7020 if (add == NULL_RTX)
7023 /* Form simplified final result. */
7024 if (mult == const0_rtx)
7026 else if (mult == const1_rtx)
7027 mult = g1->dest_reg;
7029 mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult);
7031 if (add == const0_rtx)
7035 if (GET_CODE (add) == PLUS
7036 && CONSTANT_P (XEXP (add, 1)))
7038 rtx tem = XEXP (add, 1);
7039 mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0));
7043 return gen_rtx_PLUS (g2->mode, mult, add);
7048 /* Return an rtx, if any, that expresses giv G2 as a function of the register
7049 represented by G1. This indicates that G2 should be combined with G1 and
7050 that G2 can use (either directly or via an address expression) a register
7051 used to represent G1. */
7054 combine_givs_p (g1, g2)
7055 struct induction *g1, *g2;
7057 rtx tem = express_from (g1, g2);
7059 /* If these givs are identical, they can be combined. We use the results
7060 of express_from because the addends are not in a canonical form, so
7061 rtx_equal_p is a weaker test. */
7062 /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
7063 combination to be the other way round. */
7064 if (tem == g1->dest_reg
7065 && (g1->giv_type == DEST_REG || g2->giv_type == DEST_ADDR))
7067 return g1->dest_reg;
7070 /* If G2 can be expressed as a function of G1 and that function is valid
7071 as an address and no more expensive than using a register for G2,
7072 the expression of G2 in terms of G1 can be used. */
7074 && g2->giv_type == DEST_ADDR
7075 && memory_address_p (g2->mem_mode, tem)
7076 /* ??? Looses, especially with -fforce-addr, where *g2->location
7077 will always be a register, and so anything more complicated
7081 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location)
7083 && rtx_cost (tem, MEM) <= rtx_cost (*g2->location, MEM)
7094 struct combine_givs_stats
7101 cmp_combine_givs_stats (xp, yp)
7105 const struct combine_givs_stats * const x =
7106 (const struct combine_givs_stats *) xp;
7107 const struct combine_givs_stats * const y =
7108 (const struct combine_givs_stats *) yp;
7110 d = y->total_benefit - x->total_benefit;
7111 /* Stabilize the sort. */
7113 d = x->giv_number - y->giv_number;
7117 /* Check all pairs of givs for iv_class BL and see if any can be combined with
7118 any other. If so, point SAME to the giv combined with and set NEW_REG to
7119 be an expression (in terms of the other giv's DEST_REG) equivalent to the
7120 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
7124 struct iv_class *bl;
7126 /* Additional benefit to add for being combined multiple times. */
7127 const int extra_benefit = 3;
7129 struct induction *g1, *g2, **giv_array;
7130 int i, j, k, giv_count;
7131 struct combine_givs_stats *stats;
7134 /* Count givs, because bl->giv_count is incorrect here. */
7136 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7141 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
7143 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7145 giv_array[i++] = g1;
7147 stats = (struct combine_givs_stats *) xcalloc (giv_count, sizeof (*stats));
7148 can_combine = (rtx *) xcalloc (giv_count, giv_count * sizeof(rtx));
7150 for (i = 0; i < giv_count; i++)
7156 stats[i].giv_number = i;
7158 /* If a DEST_REG GIV is used only once, do not allow it to combine
7159 with anything, for in doing so we will gain nothing that cannot
7160 be had by simply letting the GIV with which we would have combined
7161 to be reduced on its own. The losage shows up in particular with
7162 DEST_ADDR targets on hosts with reg+reg addressing, though it can
7163 be seen elsewhere as well. */
7164 if (g1->giv_type == DEST_REG
7165 && (single_use = VARRAY_RTX (reg_single_usage, REGNO (g1->dest_reg)))
7166 && single_use != const0_rtx)
7169 this_benefit = g1->benefit;
7170 /* Add an additional weight for zero addends. */
7171 if (g1->no_const_addval)
7174 for (j = 0; j < giv_count; j++)
7180 && (this_combine = combine_givs_p (g1, g2)) != NULL_RTX)
7182 can_combine[i*giv_count + j] = this_combine;
7183 this_benefit += g2->benefit + extra_benefit;
7186 stats[i].total_benefit = this_benefit;
7189 /* Iterate, combining until we can't. */
7191 qsort (stats, giv_count, sizeof(*stats), cmp_combine_givs_stats);
7193 if (loop_dump_stream)
7195 fprintf (loop_dump_stream, "Sorted combine statistics:\n");
7196 for (k = 0; k < giv_count; k++)
7198 g1 = giv_array[stats[k].giv_number];
7199 if (!g1->combined_with && !g1->same)
7200 fprintf (loop_dump_stream, " {%d, %d}",
7201 INSN_UID (giv_array[stats[k].giv_number]->insn),
7202 stats[k].total_benefit);
7204 putc ('\n', loop_dump_stream);
7207 for (k = 0; k < giv_count; k++)
7209 int g1_add_benefit = 0;
7211 i = stats[k].giv_number;
7214 /* If it has already been combined, skip. */
7215 if (g1->combined_with || g1->same)
7218 for (j = 0; j < giv_count; j++)
7221 if (g1 != g2 && can_combine[i*giv_count + j]
7222 /* If it has already been combined, skip. */
7223 && ! g2->same && ! g2->combined_with)
7227 g2->new_reg = can_combine[i*giv_count + j];
7229 g1->combined_with++;
7230 g1->lifetime += g2->lifetime;
7232 g1_add_benefit += g2->benefit;
7234 /* ??? The new final_[bg]iv_value code does a much better job
7235 of finding replaceable giv's, and hence this code may no
7236 longer be necessary. */
7237 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
7238 g1_add_benefit -= copy_cost;
7240 /* To help optimize the next set of combinations, remove
7241 this giv from the benefits of other potential mates. */
7242 for (l = 0; l < giv_count; ++l)
7244 int m = stats[l].giv_number;
7245 if (can_combine[m*giv_count + j])
7246 stats[l].total_benefit -= g2->benefit + extra_benefit;
7249 if (loop_dump_stream)
7250 fprintf (loop_dump_stream,
7251 "giv at %d combined with giv at %d\n",
7252 INSN_UID (g2->insn), INSN_UID (g1->insn));
7256 /* To help optimize the next set of combinations, remove
7257 this giv from the benefits of other potential mates. */
7258 if (g1->combined_with)
7260 for (j = 0; j < giv_count; ++j)
7262 int m = stats[j].giv_number;
7263 if (can_combine[m*giv_count + i])
7264 stats[j].total_benefit -= g1->benefit + extra_benefit;
7267 g1->benefit += g1_add_benefit;
7269 /* We've finished with this giv, and everything it touched.
7270 Restart the combination so that proper weights for the
7271 rest of the givs are properly taken into account. */
7272 /* ??? Ideally we would compact the arrays at this point, so
7273 as to not cover old ground. But sanely compacting
7274 can_combine is tricky. */
7284 struct recombine_givs_stats
7287 int start_luid, end_luid;
7290 /* Used below as comparison function for qsort. We want a ascending luid
7291 when scanning the array starting at the end, thus the arguments are
7294 cmp_recombine_givs_stats (xp, yp)
7298 const struct recombine_givs_stats * const x =
7299 (const struct recombine_givs_stats *) xp;
7300 const struct recombine_givs_stats * const y =
7301 (const struct recombine_givs_stats *) yp;
7303 d = y->start_luid - x->start_luid;
7304 /* Stabilize the sort. */
7306 d = y->giv_number - x->giv_number;
7310 /* Scan X, which is a part of INSN, for the end of life of a giv. Also
7311 look for the start of life of a giv where the start has not been seen
7312 yet to unlock the search for the end of its life.
7313 Only consider givs that belong to BIV.
7314 Return the total number of lifetime ends that have been found. */
7316 find_life_end (x, stats, insn, biv)
7318 struct recombine_givs_stats *stats;
7325 code = GET_CODE (x);
7330 rtx reg = SET_DEST (x);
7331 if (GET_CODE (reg) == REG)
7333 int regno = REGNO (reg);
7334 struct induction *v = REG_IV_INFO (regno);
7336 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7338 && v->src_reg == biv
7339 && stats[v->ix].end_luid <= 0)
7341 /* If we see a 0 here for end_luid, it means that we have
7342 scanned the entire loop without finding any use at all.
7343 We must not predicate this code on a start_luid match
7344 since that would make the test fail for givs that have
7345 been hoisted out of inner loops. */
7346 if (stats[v->ix].end_luid == 0)
7348 stats[v->ix].end_luid = stats[v->ix].start_luid;
7349 return 1 + find_life_end (SET_SRC (x), stats, insn, biv);
7351 else if (stats[v->ix].start_luid == INSN_LUID (insn))
7352 stats[v->ix].end_luid = 0;
7354 return find_life_end (SET_SRC (x), stats, insn, biv);
7360 int regno = REGNO (x);
7361 struct induction *v = REG_IV_INFO (regno);
7363 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7365 && v->src_reg == biv
7366 && stats[v->ix].end_luid == 0)
7368 while (INSN_UID (insn) >= max_uid_for_loop)
7369 insn = NEXT_INSN (insn);
7370 stats[v->ix].end_luid = INSN_LUID (insn);
7383 fmt = GET_RTX_FORMAT (code);
7385 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7388 retval += find_life_end (XEXP (x, i), stats, insn, biv);
7390 else if (fmt[i] == 'E')
7391 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7392 retval += find_life_end (XVECEXP (x, i, j), stats, insn, biv);
7397 /* For each giv that has been combined with another, look if
7398 we can combine it with the most recently used one instead.
7399 This tends to shorten giv lifetimes, and helps the next step:
7400 try to derive givs from other givs. */
7402 recombine_givs (loop, bl, unroll_p)
7403 const struct loop *loop;
7404 struct iv_class *bl;
7407 struct induction *v, **giv_array, *last_giv;
7408 struct recombine_givs_stats *stats;
7411 int ends_need_computing;
7413 for (giv_count = 0, v = bl->giv; v; v = v->next_iv)
7419 = (struct induction **) xmalloc (giv_count * sizeof (struct induction *));
7420 stats = (struct recombine_givs_stats *) xmalloc (giv_count * sizeof *stats);
7422 /* Initialize stats and set up the ix field for each giv in stats to name
7423 the corresponding index into stats. */
7424 for (i = 0, v = bl->giv; v; v = v->next_iv)
7431 stats[i].giv_number = i;
7432 /* If this giv has been hoisted out of an inner loop, use the luid of
7433 the previous insn. */
7434 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7436 stats[i].start_luid = INSN_LUID (p);
7440 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7442 /* Set up the ix field for each giv in stats to name
7443 the corresponding index into stats, and
7444 do the actual most-recently-used recombination. */
7445 for (last_giv = 0, i = giv_count - 1; i >= 0; i--)
7447 v = giv_array[stats[i].giv_number];
7451 struct induction *old_same = v->same;
7454 /* combine_givs_p actually says if we can make this transformation.
7455 The other tests are here only to avoid keeping a giv alive
7456 that could otherwise be eliminated. */
7458 && ((old_same->maybe_dead && ! old_same->combined_with)
7459 || ! last_giv->maybe_dead
7460 || last_giv->combined_with)
7461 && (new_combine = combine_givs_p (last_giv, v)))
7463 old_same->combined_with--;
7464 v->new_reg = new_combine;
7466 last_giv->combined_with++;
7467 /* No need to update lifetimes / benefits here since we have
7468 already decided what to reduce. */
7470 if (loop_dump_stream)
7472 fprintf (loop_dump_stream,
7473 "giv at %d recombined with giv at %d as ",
7474 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7475 print_rtl (loop_dump_stream, v->new_reg);
7476 putc ('\n', loop_dump_stream);
7482 else if (v->giv_type != DEST_REG)
7485 || (last_giv->maybe_dead && ! last_giv->combined_with)
7487 || v->combined_with)
7491 ends_need_computing = 0;
7492 /* For each DEST_REG giv, compute lifetime starts, and try to compute
7493 lifetime ends from regscan info. */
7494 for (i = giv_count - 1; i >= 0; i--)
7496 v = giv_array[stats[i].giv_number];
7499 if (v->giv_type == DEST_ADDR)
7501 /* Loop unrolling of an inner loop can even create new DEST_REG
7504 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7506 stats[i].start_luid = stats[i].end_luid = INSN_LUID (p);
7508 stats[i].end_luid++;
7510 else /* v->giv_type == DEST_REG */
7514 stats[i].start_luid = INSN_LUID (v->insn);
7515 stats[i].end_luid = INSN_LUID (v->last_use);
7517 else if (INSN_UID (v->insn) >= max_uid_for_loop)
7520 /* This insn has been created by loop optimization on an inner
7521 loop. We don't have a proper start_luid that will match
7522 when we see the first set. But we do know that there will
7523 be no use before the set, so we can set end_luid to 0 so that
7524 we'll start looking for the last use right away. */
7525 for (p = PREV_INSN (v->insn); INSN_UID (p) >= max_uid_for_loop; )
7527 stats[i].start_luid = INSN_LUID (p);
7528 stats[i].end_luid = 0;
7529 ends_need_computing++;
7533 int regno = REGNO (v->dest_reg);
7534 int count = VARRAY_INT (n_times_set, regno) - 1;
7537 /* Find the first insn that sets the giv, so that we can verify
7538 if this giv's lifetime wraps around the loop. We also need
7539 the luid of the first setting insn in order to detect the
7540 last use properly. */
7543 p = prev_nonnote_insn (p);
7544 if (reg_set_p (v->dest_reg, p))
7548 stats[i].start_luid = INSN_LUID (p);
7549 if (stats[i].start_luid > uid_luid[REGNO_FIRST_UID (regno)])
7551 stats[i].end_luid = -1;
7552 ends_need_computing++;
7556 stats[i].end_luid = uid_luid[REGNO_LAST_UID (regno)];
7557 if (stats[i].end_luid > INSN_LUID (loop->end))
7559 stats[i].end_luid = -1;
7560 ends_need_computing++;
7567 /* If the regscan information was unconclusive for one or more DEST_REG
7568 givs, scan the all insn in the loop to find out lifetime ends. */
7569 if (ends_need_computing)
7571 rtx biv = bl->biv->src_reg;
7576 if (p == loop->start)
7579 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
7581 ends_need_computing -= find_life_end (PATTERN (p), stats, p, biv);
7583 while (ends_need_computing);
7586 /* Set start_luid back to the last insn that sets the giv. This allows
7587 more combinations. */
7588 for (i = giv_count - 1; i >= 0; i--)
7590 v = giv_array[stats[i].giv_number];
7593 if (INSN_UID (v->insn) < max_uid_for_loop)
7594 stats[i].start_luid = INSN_LUID (v->insn);
7597 /* Now adjust lifetime ends by taking combined givs into account. */
7598 for (i = giv_count - 1; i >= 0; i--)
7603 v = giv_array[stats[i].giv_number];
7606 if (v->same && ! v->same->ignore)
7609 luid = stats[i].start_luid;
7610 /* Use unsigned arithmetic to model loop wrap-around. */
7611 if (luid - stats[j].start_luid
7612 > (unsigned) stats[j].end_luid - stats[j].start_luid)
7613 stats[j].end_luid = luid;
7617 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7619 /* Try to derive DEST_REG givs from previous DEST_REG givs with the
7620 same mult_val and non-overlapping lifetime. This reduces register
7622 Once we find a DEST_REG giv that is suitable to derive others from,
7623 we set last_giv to this giv, and try to derive as many other DEST_REG
7624 givs from it without joining overlapping lifetimes. If we then
7625 encounter a DEST_REG giv that we can't derive, we set rescan to the
7626 index for this giv (unless rescan is already set).
7627 When we are finished with the current LAST_GIV (i.e. the inner loop
7628 terminates), we start again with rescan, which then becomes the new
7630 for (i = giv_count - 1; i >= 0; i = rescan)
7632 int life_start = 0, life_end = 0;
7634 for (last_giv = 0, rescan = -1; i >= 0; i--)
7638 v = giv_array[stats[i].giv_number];
7639 if (v->giv_type != DEST_REG || v->derived_from || v->same)
7643 /* Don't use a giv that's likely to be dead to derive
7644 others - that would be likely to keep that giv alive. */
7645 if (! v->maybe_dead || v->combined_with)
7648 life_start = stats[i].start_luid;
7649 life_end = stats[i].end_luid;
7653 /* Use unsigned arithmetic to model loop wrap around. */
7654 if (((unsigned) stats[i].start_luid - life_start
7655 >= (unsigned) life_end - life_start)
7656 && ((unsigned) stats[i].end_luid - life_start
7657 > (unsigned) life_end - life_start)
7658 /* Check that the giv insn we're about to use for deriving
7659 precedes all uses of that giv. Note that initializing the
7660 derived giv would defeat the purpose of reducing register
7662 ??? We could arrange to move the insn. */
7663 && ((unsigned) stats[i].end_luid - INSN_LUID (loop->start)
7664 > (unsigned) stats[i].start_luid - INSN_LUID (loop->start))
7665 && rtx_equal_p (last_giv->mult_val, v->mult_val)
7666 /* ??? Could handle libcalls, but would need more logic. */
7667 && ! find_reg_note (v->insn, REG_RETVAL, NULL_RTX)
7668 /* We would really like to know if for any giv that v
7669 is combined with, v->insn or any intervening biv increment
7670 dominates that combined giv. However, we
7671 don't have this detailed control flow information.
7672 N.B. since last_giv will be reduced, it is valid
7673 anywhere in the loop, so we don't need to check the
7674 validity of last_giv.
7675 We rely here on the fact that v->always_executed implies that
7676 there is no jump to someplace else in the loop before the
7677 giv insn, and hence any insn that is executed before the
7678 giv insn in the loop will have a lower luid. */
7679 && (v->always_executed || ! v->combined_with)
7680 && (sum = express_from (last_giv, v))
7681 /* Make sure we don't make the add more expensive. ADD_COST
7682 doesn't take different costs of registers and constants into
7683 account, so compare the cost of the actual SET_SRCs. */
7684 && (rtx_cost (sum, SET)
7685 <= rtx_cost (SET_SRC (single_set (v->insn)), SET))
7686 /* ??? unroll can't understand anything but reg + const_int
7687 sums. It would be cleaner to fix unroll. */
7688 && ((GET_CODE (sum) == PLUS
7689 && GET_CODE (XEXP (sum, 0)) == REG
7690 && GET_CODE (XEXP (sum, 1)) == CONST_INT)
7692 && validate_change (v->insn, &PATTERN (v->insn),
7693 gen_rtx_SET (VOIDmode, v->dest_reg, sum), 0))
7695 v->derived_from = last_giv;
7696 life_end = stats[i].end_luid;
7698 if (loop_dump_stream)
7700 fprintf (loop_dump_stream,
7701 "giv at %d derived from %d as ",
7702 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7703 print_rtl (loop_dump_stream, sum);
7704 putc ('\n', loop_dump_stream);
7707 else if (rescan < 0)
7717 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
7720 emit_iv_add_mult (b, m, a, reg, insert_before)
7721 rtx b; /* initial value of basic induction variable */
7722 rtx m; /* multiplicative constant */
7723 rtx a; /* additive constant */
7724 rtx reg; /* destination register */
7730 /* Prevent unexpected sharing of these rtx. */
7734 /* Increase the lifetime of any invariants moved further in code. */
7735 update_reg_last_use (a, insert_before);
7736 update_reg_last_use (b, insert_before);
7737 update_reg_last_use (m, insert_before);
7740 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
7742 emit_move_insn (reg, result);
7743 seq = gen_sequence ();
7746 emit_insn_before (seq, insert_before);
7748 /* It is entirely possible that the expansion created lots of new
7749 registers. Iterate over the sequence we just created and
7752 if (GET_CODE (seq) == SEQUENCE)
7755 for (i = 0; i < XVECLEN (seq, 0); ++i)
7757 rtx set = single_set (XVECEXP (seq, 0, i));
7758 if (set && GET_CODE (SET_DEST (set)) == REG)
7759 record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0);
7762 else if (GET_CODE (seq) == SET
7763 && GET_CODE (SET_DEST (seq)) == REG)
7764 record_base_value (REGNO (SET_DEST (seq)), SET_SRC (seq), 0);
7767 /* Test whether A * B can be computed without
7768 an actual multiply insn. Value is 1 if so. */
7771 product_cheap_p (a, b)
7777 struct obstack *old_rtl_obstack = rtl_obstack;
7778 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
7781 /* If only one is constant, make it B. */
7782 if (GET_CODE (a) == CONST_INT)
7783 tmp = a, a = b, b = tmp;
7785 /* If first constant, both constant, so don't need multiply. */
7786 if (GET_CODE (a) == CONST_INT)
7789 /* If second not constant, neither is constant, so would need multiply. */
7790 if (GET_CODE (b) != CONST_INT)
7793 /* One operand is constant, so might not need multiply insn. Generate the
7794 code for the multiply and see if a call or multiply, or long sequence
7795 of insns is generated. */
7797 rtl_obstack = &temp_obstack;
7799 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
7800 tmp = gen_sequence ();
7803 if (GET_CODE (tmp) == SEQUENCE)
7805 if (XVEC (tmp, 0) == 0)
7807 else if (XVECLEN (tmp, 0) > 3)
7810 for (i = 0; i < XVECLEN (tmp, 0); i++)
7812 rtx insn = XVECEXP (tmp, 0, i);
7814 if (GET_CODE (insn) != INSN
7815 || (GET_CODE (PATTERN (insn)) == SET
7816 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
7817 || (GET_CODE (PATTERN (insn)) == PARALLEL
7818 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
7819 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
7826 else if (GET_CODE (tmp) == SET
7827 && GET_CODE (SET_SRC (tmp)) == MULT)
7829 else if (GET_CODE (tmp) == PARALLEL
7830 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
7831 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
7834 /* Free any storage we obtained in generating this multiply and restore rtl
7835 allocation to its normal obstack. */
7836 obstack_free (&temp_obstack, storage);
7837 rtl_obstack = old_rtl_obstack;
7842 /* Check to see if loop can be terminated by a "decrement and branch until
7843 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
7844 Also try reversing an increment loop to a decrement loop
7845 to see if the optimization can be performed.
7846 Value is nonzero if optimization was performed. */
7848 /* This is useful even if the architecture doesn't have such an insn,
7849 because it might change a loops which increments from 0 to n to a loop
7850 which decrements from n to 0. A loop that decrements to zero is usually
7851 faster than one that increments from zero. */
7853 /* ??? This could be rewritten to use some of the loop unrolling procedures,
7854 such as approx_final_value, biv_total_increment, loop_iterations, and
7855 final_[bg]iv_value. */
7858 check_dbra_loop (loop, insn_count)
7862 struct iv_class *bl;
7869 rtx before_comparison;
7873 int compare_and_branch;
7874 rtx loop_start = loop->start;
7875 rtx loop_end = loop->end;
7876 struct loop_info *loop_info = LOOP_INFO (loop);
7878 /* If last insn is a conditional branch, and the insn before tests a
7879 register value, try to optimize it. Otherwise, we can't do anything. */
7881 jump = PREV_INSN (loop_end);
7882 comparison = get_condition_for_loop (loop, jump);
7883 if (comparison == 0)
7886 /* Try to compute whether the compare/branch at the loop end is one or
7887 two instructions. */
7888 get_condition (jump, &first_compare);
7889 if (first_compare == jump)
7890 compare_and_branch = 1;
7891 else if (first_compare == prev_nonnote_insn (jump))
7892 compare_and_branch = 2;
7896 /* Check all of the bivs to see if the compare uses one of them.
7897 Skip biv's set more than once because we can't guarantee that
7898 it will be zero on the last iteration. Also skip if the biv is
7899 used between its update and the test insn. */
7901 for (bl = loop_iv_list; bl; bl = bl->next)
7903 if (bl->biv_count == 1
7904 && ! bl->biv->maybe_multiple
7905 && bl->biv->dest_reg == XEXP (comparison, 0)
7906 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
7914 /* Look for the case where the basic induction variable is always
7915 nonnegative, and equals zero on the last iteration.
7916 In this case, add a reg_note REG_NONNEG, which allows the
7917 m68k DBRA instruction to be used. */
7919 if (((GET_CODE (comparison) == GT
7920 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
7921 && INTVAL (XEXP (comparison, 1)) == -1)
7922 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
7923 && GET_CODE (bl->biv->add_val) == CONST_INT
7924 && INTVAL (bl->biv->add_val) < 0)
7926 /* Initial value must be greater than 0,
7927 init_val % -dec_value == 0 to ensure that it equals zero on
7928 the last iteration */
7930 if (GET_CODE (bl->initial_value) == CONST_INT
7931 && INTVAL (bl->initial_value) > 0
7932 && (INTVAL (bl->initial_value)
7933 % (-INTVAL (bl->biv->add_val))) == 0)
7935 /* register always nonnegative, add REG_NOTE to branch */
7936 REG_NOTES (PREV_INSN (loop_end))
7937 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7938 REG_NOTES (PREV_INSN (loop_end)));
7944 /* If the decrement is 1 and the value was tested as >= 0 before
7945 the loop, then we can safely optimize. */
7946 for (p = loop_start; p; p = PREV_INSN (p))
7948 if (GET_CODE (p) == CODE_LABEL)
7950 if (GET_CODE (p) != JUMP_INSN)
7953 before_comparison = get_condition_for_loop (loop, p);
7954 if (before_comparison
7955 && XEXP (before_comparison, 0) == bl->biv->dest_reg
7956 && GET_CODE (before_comparison) == LT
7957 && XEXP (before_comparison, 1) == const0_rtx
7958 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
7959 && INTVAL (bl->biv->add_val) == -1)
7961 REG_NOTES (PREV_INSN (loop_end))
7962 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7963 REG_NOTES (PREV_INSN (loop_end)));
7970 else if (GET_CODE (bl->biv->add_val) == CONST_INT
7971 && INTVAL (bl->biv->add_val) > 0)
7973 /* Try to change inc to dec, so can apply above optimization. */
7975 all registers modified are induction variables or invariant,
7976 all memory references have non-overlapping addresses
7977 (obviously true if only one write)
7978 allow 2 insns for the compare/jump at the end of the loop. */
7979 /* Also, we must avoid any instructions which use both the reversed
7980 biv and another biv. Such instructions will fail if the loop is
7981 reversed. We meet this condition by requiring that either
7982 no_use_except_counting is true, or else that there is only
7984 int num_nonfixed_reads = 0;
7985 /* 1 if the iteration var is used only to count iterations. */
7986 int no_use_except_counting = 0;
7987 /* 1 if the loop has no memory store, or it has a single memory store
7988 which is reversible. */
7989 int reversible_mem_store = 1;
7991 if (bl->giv_count == 0 && ! loop->exit_count)
7993 rtx bivreg = regno_reg_rtx[bl->regno];
7995 /* If there are no givs for this biv, and the only exit is the
7996 fall through at the end of the loop, then
7997 see if perhaps there are no uses except to count. */
7998 no_use_except_counting = 1;
7999 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8000 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
8002 rtx set = single_set (p);
8004 if (set && GET_CODE (SET_DEST (set)) == REG
8005 && REGNO (SET_DEST (set)) == bl->regno)
8006 /* An insn that sets the biv is okay. */
8008 else if ((p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
8009 || p == prev_nonnote_insn (loop_end))
8010 && reg_mentioned_p (bivreg, PATTERN (p)))
8012 /* If either of these insns uses the biv and sets a pseudo
8013 that has more than one usage, then the biv has uses
8014 other than counting since it's used to derive a value
8015 that is used more than one time. */
8016 int note_set_pseudo_multiple_uses_retval = 0;
8017 note_stores (PATTERN (p), note_set_pseudo_multiple_uses,
8018 ¬e_set_pseudo_multiple_uses_retval);
8019 if (note_set_pseudo_multiple_uses_retval)
8021 no_use_except_counting = 0;
8025 else if (reg_mentioned_p (bivreg, PATTERN (p)))
8027 no_use_except_counting = 0;
8033 if (no_use_except_counting)
8034 ; /* no need to worry about MEMs. */
8035 else if (num_mem_sets <= 1)
8037 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8038 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
8039 num_nonfixed_reads += count_nonfixed_reads (loop, PATTERN (p));
8041 /* If the loop has a single store, and the destination address is
8042 invariant, then we can't reverse the loop, because this address
8043 might then have the wrong value at loop exit.
8044 This would work if the source was invariant also, however, in that
8045 case, the insn should have been moved out of the loop. */
8047 if (num_mem_sets == 1)
8049 struct induction *v;
8051 reversible_mem_store
8052 = (! unknown_address_altered
8053 && ! unknown_constant_address_altered
8054 && ! loop_invariant_p (loop,
8055 XEXP (XEXP (loop_store_mems, 0),
8058 /* If the store depends on a register that is set after the
8059 store, it depends on the initial value, and is thus not
8061 for (v = bl->giv; reversible_mem_store && v; v = v->next_iv)
8063 if (v->giv_type == DEST_REG
8064 && reg_mentioned_p (v->dest_reg,
8065 PATTERN (first_loop_store_insn))
8066 && loop_insn_first_p (first_loop_store_insn, v->insn))
8067 reversible_mem_store = 0;
8074 /* This code only acts for innermost loops. Also it simplifies
8075 the memory address check by only reversing loops with
8076 zero or one memory access.
8077 Two memory accesses could involve parts of the same array,
8078 and that can't be reversed.
8079 If the biv is used only for counting, than we don't need to worry
8080 about all these things. */
8082 if ((num_nonfixed_reads <= 1
8083 && ! loop_info->has_call
8084 && ! loop_info->has_volatile
8085 && reversible_mem_store
8086 && (bl->giv_count + bl->biv_count + num_mem_sets
8087 + num_movables + compare_and_branch == insn_count)
8088 && (bl == loop_iv_list && bl->next == 0))
8089 || no_use_except_counting)
8093 /* Loop can be reversed. */
8094 if (loop_dump_stream)
8095 fprintf (loop_dump_stream, "Can reverse loop\n");
8097 /* Now check other conditions:
8099 The increment must be a constant, as must the initial value,
8100 and the comparison code must be LT.
8102 This test can probably be improved since +/- 1 in the constant
8103 can be obtained by changing LT to LE and vice versa; this is
8107 /* for constants, LE gets turned into LT */
8108 && (GET_CODE (comparison) == LT
8109 || (GET_CODE (comparison) == LE
8110 && no_use_except_counting)))
8112 HOST_WIDE_INT add_val, add_adjust, comparison_val = 0;
8113 rtx initial_value, comparison_value;
8115 enum rtx_code cmp_code;
8116 int comparison_const_width;
8117 unsigned HOST_WIDE_INT comparison_sign_mask;
8119 add_val = INTVAL (bl->biv->add_val);
8120 comparison_value = XEXP (comparison, 1);
8121 if (GET_MODE (comparison_value) == VOIDmode)
8122 comparison_const_width
8123 = GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0)));
8125 comparison_const_width
8126 = GET_MODE_BITSIZE (GET_MODE (comparison_value));
8127 if (comparison_const_width > HOST_BITS_PER_WIDE_INT)
8128 comparison_const_width = HOST_BITS_PER_WIDE_INT;
8129 comparison_sign_mask
8130 = (unsigned HOST_WIDE_INT)1 << (comparison_const_width - 1);
8132 /* If the comparison value is not a loop invariant, then we
8133 can not reverse this loop.
8135 ??? If the insns which initialize the comparison value as
8136 a whole compute an invariant result, then we could move
8137 them out of the loop and proceed with loop reversal. */
8138 if (! loop_invariant_p (loop, comparison_value))
8141 if (GET_CODE (comparison_value) == CONST_INT)
8142 comparison_val = INTVAL (comparison_value);
8143 initial_value = bl->initial_value;
8145 /* Normalize the initial value if it is an integer and
8146 has no other use except as a counter. This will allow
8147 a few more loops to be reversed. */
8148 if (no_use_except_counting
8149 && GET_CODE (comparison_value) == CONST_INT
8150 && GET_CODE (initial_value) == CONST_INT)
8152 comparison_val = comparison_val - INTVAL (bl->initial_value);
8153 /* The code below requires comparison_val to be a multiple
8154 of add_val in order to do the loop reversal, so
8155 round up comparison_val to a multiple of add_val.
8156 Since comparison_value is constant, we know that the
8157 current comparison code is LT. */
8158 comparison_val = comparison_val + add_val - 1;
8160 -= (unsigned HOST_WIDE_INT) comparison_val % add_val;
8161 /* We postpone overflow checks for COMPARISON_VAL here;
8162 even if there is an overflow, we might still be able to
8163 reverse the loop, if converting the loop exit test to
8165 initial_value = const0_rtx;
8168 /* First check if we can do a vanilla loop reversal. */
8169 if (initial_value == const0_rtx
8170 /* If we have a decrement_and_branch_on_count,
8171 prefer the NE test, since this will allow that
8172 instruction to be generated. Note that we must
8173 use a vanilla loop reversal if the biv is used to
8174 calculate a giv or has a non-counting use. */
8175 #if ! defined (HAVE_decrement_and_branch_until_zero) \
8176 && defined (HAVE_decrement_and_branch_on_count)
8177 && (! (add_val == 1 && loop->vtop
8178 && (bl->biv_count == 0
8179 || no_use_except_counting)))
8181 && GET_CODE (comparison_value) == CONST_INT
8182 /* Now do postponed overflow checks on COMPARISON_VAL. */
8183 && ! (((comparison_val - add_val) ^ INTVAL (comparison_value))
8184 & comparison_sign_mask))
8186 /* Register will always be nonnegative, with value
8187 0 on last iteration */
8188 add_adjust = add_val;
8192 else if (add_val == 1 && loop->vtop
8193 && (bl->biv_count == 0
8194 || no_use_except_counting))
8202 if (GET_CODE (comparison) == LE)
8203 add_adjust -= add_val;
8205 /* If the initial value is not zero, or if the comparison
8206 value is not an exact multiple of the increment, then we
8207 can not reverse this loop. */
8208 if (initial_value == const0_rtx
8209 && GET_CODE (comparison_value) == CONST_INT)
8211 if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0)
8216 if (! no_use_except_counting || add_val != 1)
8220 final_value = comparison_value;
8222 /* Reset these in case we normalized the initial value
8223 and comparison value above. */
8224 if (GET_CODE (comparison_value) == CONST_INT
8225 && GET_CODE (initial_value) == CONST_INT)
8227 comparison_value = GEN_INT (comparison_val);
8229 = GEN_INT (comparison_val + INTVAL (bl->initial_value));
8231 bl->initial_value = initial_value;
8233 /* Save some info needed to produce the new insns. */
8234 reg = bl->biv->dest_reg;
8235 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
8236 if (jump_label == pc_rtx)
8237 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
8238 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
8240 /* Set start_value; if this is not a CONST_INT, we need
8242 Initialize biv to start_value before loop start.
8243 The old initializing insn will be deleted as a
8244 dead store by flow.c. */
8245 if (initial_value == const0_rtx
8246 && GET_CODE (comparison_value) == CONST_INT)
8248 start_value = GEN_INT (comparison_val - add_adjust);
8249 emit_insn_before (gen_move_insn (reg, start_value),
8252 else if (GET_CODE (initial_value) == CONST_INT)
8254 rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust);
8255 enum machine_mode mode = GET_MODE (reg);
8256 enum insn_code icode
8257 = add_optab->handlers[(int) mode].insn_code;
8259 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8260 || ! ((*insn_data[icode].operand[1].predicate)
8261 (comparison_value, mode))
8262 || ! ((*insn_data[icode].operand[2].predicate)
8266 = gen_rtx_PLUS (mode, comparison_value, offset);
8267 emit_insn_before ((GEN_FCN (icode)
8268 (reg, comparison_value, offset)),
8270 if (GET_CODE (comparison) == LE)
8271 final_value = gen_rtx_PLUS (mode, comparison_value,
8274 else if (! add_adjust)
8276 enum machine_mode mode = GET_MODE (reg);
8277 enum insn_code icode
8278 = sub_optab->handlers[(int) mode].insn_code;
8279 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8280 || ! ((*insn_data[icode].operand[1].predicate)
8281 (comparison_value, mode))
8282 || ! ((*insn_data[icode].operand[2].predicate)
8283 (initial_value, mode)))
8286 = gen_rtx_MINUS (mode, comparison_value, initial_value);
8287 emit_insn_before ((GEN_FCN (icode)
8288 (reg, comparison_value, initial_value)),
8292 /* We could handle the other cases too, but it'll be
8293 better to have a testcase first. */
8296 /* We may not have a single insn which can increment a reg, so
8297 create a sequence to hold all the insns from expand_inc. */
8299 expand_inc (reg, new_add_val);
8300 tem = gen_sequence ();
8303 p = emit_insn_before (tem, bl->biv->insn);
8304 delete_insn (bl->biv->insn);
8306 /* Update biv info to reflect its new status. */
8308 bl->initial_value = start_value;
8309 bl->biv->add_val = new_add_val;
8311 /* Update loop info. */
8312 loop_info->initial_value = reg;
8313 loop_info->initial_equiv_value = reg;
8314 loop_info->final_value = const0_rtx;
8315 loop_info->final_equiv_value = const0_rtx;
8316 loop_info->comparison_value = const0_rtx;
8317 loop_info->comparison_code = cmp_code;
8318 loop_info->increment = new_add_val;
8320 /* Inc LABEL_NUSES so that delete_insn will
8321 not delete the label. */
8322 LABEL_NUSES (XEXP (jump_label, 0)) ++;
8324 /* Emit an insn after the end of the loop to set the biv's
8325 proper exit value if it is used anywhere outside the loop. */
8326 if ((REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare))
8328 || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
8329 emit_insn_after (gen_move_insn (reg, final_value),
8332 /* Delete compare/branch at end of loop. */
8333 delete_insn (PREV_INSN (loop_end));
8334 if (compare_and_branch == 2)
8335 delete_insn (first_compare);
8337 /* Add new compare/branch insn at end of loop. */
8339 emit_cmp_and_jump_insns (reg, const0_rtx, cmp_code, NULL_RTX,
8340 GET_MODE (reg), 0, 0,
8341 XEXP (jump_label, 0));
8342 tem = gen_sequence ();
8344 emit_jump_insn_before (tem, loop_end);
8346 for (tem = PREV_INSN (loop_end);
8347 tem && GET_CODE (tem) != JUMP_INSN;
8348 tem = PREV_INSN (tem))
8352 JUMP_LABEL (tem) = XEXP (jump_label, 0);
8358 /* Increment of LABEL_NUSES done above. */
8359 /* Register is now always nonnegative,
8360 so add REG_NONNEG note to the branch. */
8361 REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
8367 /* No insn may reference both the reversed and another biv or it
8368 will fail (see comment near the top of the loop reversal
8370 Earlier on, we have verified that the biv has no use except
8371 counting, or it is the only biv in this function.
8372 However, the code that computes no_use_except_counting does
8373 not verify reg notes. It's possible to have an insn that
8374 references another biv, and has a REG_EQUAL note with an
8375 expression based on the reversed biv. To avoid this case,
8376 remove all REG_EQUAL notes based on the reversed biv
8378 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8379 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
8382 rtx set = single_set (p);
8383 /* If this is a set of a GIV based on the reversed biv, any
8384 REG_EQUAL notes should still be correct. */
8386 || GET_CODE (SET_DEST (set)) != REG
8387 || (size_t) REGNO (SET_DEST (set)) >= reg_iv_type->num_elements
8388 || REG_IV_TYPE (REGNO (SET_DEST (set))) != GENERAL_INDUCT
8389 || REG_IV_INFO (REGNO (SET_DEST (set)))->src_reg != bl->biv->src_reg)
8390 for (pnote = ®_NOTES (p); *pnote;)
8392 if (REG_NOTE_KIND (*pnote) == REG_EQUAL
8393 && reg_mentioned_p (regno_reg_rtx[bl->regno],
8395 *pnote = XEXP (*pnote, 1);
8397 pnote = &XEXP (*pnote, 1);
8401 /* Mark that this biv has been reversed. Each giv which depends
8402 on this biv, and which is also live past the end of the loop
8403 will have to be fixed up. */
8407 if (loop_dump_stream)
8409 fprintf (loop_dump_stream, "Reversed loop");
8411 fprintf (loop_dump_stream, " and added reg_nonneg\n");
8413 fprintf (loop_dump_stream, "\n");
8424 /* Verify whether the biv BL appears to be eliminable,
8425 based on the insns in the loop that refer to it.
8427 If ELIMINATE_P is non-zero, actually do the elimination.
8429 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
8430 determine whether invariant insns should be placed inside or at the
8431 start of the loop. */
8434 maybe_eliminate_biv (loop, bl, eliminate_p, threshold, insn_count)
8435 const struct loop *loop;
8436 struct iv_class *bl;
8438 int threshold, insn_count;
8440 rtx reg = bl->biv->dest_reg;
8441 rtx loop_start = loop->start;
8442 rtx loop_end = loop->end;
8445 /* Scan all insns in the loop, stopping if we find one that uses the
8446 biv in a way that we cannot eliminate. */
8448 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8450 enum rtx_code code = GET_CODE (p);
8451 rtx where = threshold >= insn_count ? loop_start : p;
8453 /* If this is a libcall that sets a giv, skip ahead to its end. */
8454 if (GET_RTX_CLASS (code) == 'i')
8456 rtx note = find_reg_note (p, REG_LIBCALL, NULL_RTX);
8460 rtx last = XEXP (note, 0);
8461 rtx set = single_set (last);
8463 if (set && GET_CODE (SET_DEST (set)) == REG)
8465 int regno = REGNO (SET_DEST (set));
8467 if (regno < max_reg_before_loop
8468 && REG_IV_TYPE (regno) == GENERAL_INDUCT
8469 && REG_IV_INFO (regno)->src_reg == bl->biv->src_reg)
8474 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
8475 && reg_mentioned_p (reg, PATTERN (p))
8476 && ! maybe_eliminate_biv_1 (loop, PATTERN (p), p, bl,
8477 eliminate_p, where))
8479 if (loop_dump_stream)
8480 fprintf (loop_dump_stream,
8481 "Cannot eliminate biv %d: biv used in insn %d.\n",
8482 bl->regno, INSN_UID (p));
8489 if (loop_dump_stream)
8490 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
8491 bl->regno, eliminate_p ? "was" : "can be");
8498 /* INSN and REFERENCE are instructions in the same insn chain.
8499 Return non-zero if INSN is first. */
8502 loop_insn_first_p (insn, reference)
8503 rtx insn, reference;
8507 for (p = insn, q = reference; ;)
8509 /* Start with test for not first so that INSN == REFERENCE yields not
8511 if (q == insn || ! p)
8513 if (p == reference || ! q)
8516 /* Either of P or Q might be a NOTE. Notes have the same LUID as the
8517 previous insn, hence the <= comparison below does not work if
8519 if (INSN_UID (p) < max_uid_for_loop
8520 && INSN_UID (q) < max_uid_for_loop
8521 && GET_CODE (p) != NOTE)
8522 return INSN_LUID (p) <= INSN_LUID (q);
8524 if (INSN_UID (p) >= max_uid_for_loop
8525 || GET_CODE (p) == NOTE)
8527 if (INSN_UID (q) >= max_uid_for_loop)
8532 /* We are trying to eliminate BIV in INSN using GIV. Return non-zero if
8533 the offset that we have to take into account due to auto-increment /
8534 div derivation is zero. */
8536 biv_elimination_giv_has_0_offset (biv, giv, insn)
8537 struct induction *biv, *giv;
8540 /* If the giv V had the auto-inc address optimization applied
8541 to it, and INSN occurs between the giv insn and the biv
8542 insn, then we'd have to adjust the value used here.
8543 This is rare, so we don't bother to make this possible. */
8544 if (giv->auto_inc_opt
8545 && ((loop_insn_first_p (giv->insn, insn)
8546 && loop_insn_first_p (insn, biv->insn))
8547 || (loop_insn_first_p (biv->insn, insn)
8548 && loop_insn_first_p (insn, giv->insn))))
8551 /* If the giv V was derived from another giv, and INSN does
8552 not occur between the giv insn and the biv insn, then we'd
8553 have to adjust the value used here. This is rare, so we don't
8554 bother to make this possible. */
8555 if (giv->derived_from
8556 && ! (giv->always_executed
8557 && loop_insn_first_p (giv->insn, insn)
8558 && loop_insn_first_p (insn, biv->insn)))
8561 && giv->same->derived_from
8562 && ! (giv->same->always_executed
8563 && loop_insn_first_p (giv->same->insn, insn)
8564 && loop_insn_first_p (insn, biv->insn)))
8570 /* If BL appears in X (part of the pattern of INSN), see if we can
8571 eliminate its use. If so, return 1. If not, return 0.
8573 If BIV does not appear in X, return 1.
8575 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
8576 where extra insns should be added. Depending on how many items have been
8577 moved out of the loop, it will either be before INSN or at the start of
8581 maybe_eliminate_biv_1 (loop, x, insn, bl, eliminate_p, where)
8582 const struct loop *loop;
8584 struct iv_class *bl;
8588 enum rtx_code code = GET_CODE (x);
8589 rtx reg = bl->biv->dest_reg;
8590 enum machine_mode mode = GET_MODE (reg);
8591 struct induction *v;
8603 /* If we haven't already been able to do something with this BIV,
8604 we can't eliminate it. */
8610 /* If this sets the BIV, it is not a problem. */
8611 if (SET_DEST (x) == reg)
8614 /* If this is an insn that defines a giv, it is also ok because
8615 it will go away when the giv is reduced. */
8616 for (v = bl->giv; v; v = v->next_iv)
8617 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
8621 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
8623 /* Can replace with any giv that was reduced and
8624 that has (MULT_VAL != 0) and (ADD_VAL == 0).
8625 Require a constant for MULT_VAL, so we know it's nonzero.
8626 ??? We disable this optimization to avoid potential
8629 for (v = bl->giv; v; v = v->next_iv)
8630 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
8631 && v->add_val == const0_rtx
8632 && ! v->ignore && ! v->maybe_dead && v->always_computable
8636 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8642 /* If the giv has the opposite direction of change,
8643 then reverse the comparison. */
8644 if (INTVAL (v->mult_val) < 0)
8645 new = gen_rtx_COMPARE (GET_MODE (v->new_reg),
8646 const0_rtx, v->new_reg);
8650 /* We can probably test that giv's reduced reg. */
8651 if (validate_change (insn, &SET_SRC (x), new, 0))
8655 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
8656 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
8657 Require a constant for MULT_VAL, so we know it's nonzero.
8658 ??? Do this only if ADD_VAL is a pointer to avoid a potential
8659 overflow problem. */
8661 for (v = bl->giv; v; v = v->next_iv)
8662 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
8663 && ! v->ignore && ! v->maybe_dead && v->always_computable
8665 && (GET_CODE (v->add_val) == SYMBOL_REF
8666 || GET_CODE (v->add_val) == LABEL_REF
8667 || GET_CODE (v->add_val) == CONST
8668 || (GET_CODE (v->add_val) == REG
8669 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
8671 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8677 /* If the giv has the opposite direction of change,
8678 then reverse the comparison. */
8679 if (INTVAL (v->mult_val) < 0)
8680 new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val),
8683 new = gen_rtx_COMPARE (VOIDmode, v->new_reg,
8684 copy_rtx (v->add_val));
8686 /* Replace biv with the giv's reduced register. */
8687 update_reg_last_use (v->add_val, insn);
8688 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8691 /* Insn doesn't support that constant or invariant. Copy it
8692 into a register (it will be a loop invariant.) */
8693 tem = gen_reg_rtx (GET_MODE (v->new_reg));
8695 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
8698 /* Substitute the new register for its invariant value in
8699 the compare expression. */
8700 XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
8701 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8710 case GT: case GE: case GTU: case GEU:
8711 case LT: case LE: case LTU: case LEU:
8712 /* See if either argument is the biv. */
8713 if (XEXP (x, 0) == reg)
8714 arg = XEXP (x, 1), arg_operand = 1;
8715 else if (XEXP (x, 1) == reg)
8716 arg = XEXP (x, 0), arg_operand = 0;
8720 if (CONSTANT_P (arg))
8722 /* First try to replace with any giv that has constant positive
8723 mult_val and constant add_val. We might be able to support
8724 negative mult_val, but it seems complex to do it in general. */
8726 for (v = bl->giv; v; v = v->next_iv)
8727 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8728 && (GET_CODE (v->add_val) == SYMBOL_REF
8729 || GET_CODE (v->add_val) == LABEL_REF
8730 || GET_CODE (v->add_val) == CONST
8731 || (GET_CODE (v->add_val) == REG
8732 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
8733 && ! v->ignore && ! v->maybe_dead && v->always_computable
8736 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8742 /* Replace biv with the giv's reduced reg. */
8743 XEXP (x, 1-arg_operand) = v->new_reg;
8745 /* If all constants are actually constant integers and
8746 the derived constant can be directly placed in the COMPARE,
8748 if (GET_CODE (arg) == CONST_INT
8749 && GET_CODE (v->mult_val) == CONST_INT
8750 && GET_CODE (v->add_val) == CONST_INT
8751 && validate_change (insn, &XEXP (x, arg_operand),
8752 GEN_INT (INTVAL (arg)
8753 * INTVAL (v->mult_val)
8754 + INTVAL (v->add_val)), 0))
8757 /* Otherwise, load it into a register. */
8758 tem = gen_reg_rtx (mode);
8759 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8760 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
8763 /* If that failed, put back the change we made above. */
8764 XEXP (x, 1-arg_operand) = reg;
8767 /* Look for giv with positive constant mult_val and nonconst add_val.
8768 Insert insns to calculate new compare value.
8769 ??? Turn this off due to possible overflow. */
8771 for (v = bl->giv; v; v = v->next_iv)
8772 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8773 && ! v->ignore && ! v->maybe_dead && v->always_computable
8779 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8785 tem = gen_reg_rtx (mode);
8787 /* Replace biv with giv's reduced register. */
8788 validate_change (insn, &XEXP (x, 1 - arg_operand),
8791 /* Compute value to compare against. */
8792 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8793 /* Use it in this insn. */
8794 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8795 if (apply_change_group ())
8799 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
8801 if (loop_invariant_p (loop, arg) == 1)
8803 /* Look for giv with constant positive mult_val and nonconst
8804 add_val. Insert insns to compute new compare value.
8805 ??? Turn this off due to possible overflow. */
8807 for (v = bl->giv; v; v = v->next_iv)
8808 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8809 && ! v->ignore && ! v->maybe_dead && v->always_computable
8815 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8821 tem = gen_reg_rtx (mode);
8823 /* Replace biv with giv's reduced register. */
8824 validate_change (insn, &XEXP (x, 1 - arg_operand),
8827 /* Compute value to compare against. */
8828 emit_iv_add_mult (arg, v->mult_val, v->add_val,
8830 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8831 if (apply_change_group ())
8836 /* This code has problems. Basically, you can't know when
8837 seeing if we will eliminate BL, whether a particular giv
8838 of ARG will be reduced. If it isn't going to be reduced,
8839 we can't eliminate BL. We can try forcing it to be reduced,
8840 but that can generate poor code.
8842 The problem is that the benefit of reducing TV, below should
8843 be increased if BL can actually be eliminated, but this means
8844 we might have to do a topological sort of the order in which
8845 we try to process biv. It doesn't seem worthwhile to do
8846 this sort of thing now. */
8849 /* Otherwise the reg compared with had better be a biv. */
8850 if (GET_CODE (arg) != REG
8851 || REG_IV_TYPE (REGNO (arg)) != BASIC_INDUCT)
8854 /* Look for a pair of givs, one for each biv,
8855 with identical coefficients. */
8856 for (v = bl->giv; v; v = v->next_iv)
8858 struct induction *tv;
8860 if (v->ignore || v->maybe_dead || v->mode != mode)
8863 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
8864 if (! tv->ignore && ! tv->maybe_dead
8865 && rtx_equal_p (tv->mult_val, v->mult_val)
8866 && rtx_equal_p (tv->add_val, v->add_val)
8867 && tv->mode == mode)
8869 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8875 /* Replace biv with its giv's reduced reg. */
8876 XEXP (x, 1-arg_operand) = v->new_reg;
8877 /* Replace other operand with the other giv's
8879 XEXP (x, arg_operand) = tv->new_reg;
8886 /* If we get here, the biv can't be eliminated. */
8890 /* If this address is a DEST_ADDR giv, it doesn't matter if the
8891 biv is used in it, since it will be replaced. */
8892 for (v = bl->giv; v; v = v->next_iv)
8893 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
8901 /* See if any subexpression fails elimination. */
8902 fmt = GET_RTX_FORMAT (code);
8903 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
8908 if (! maybe_eliminate_biv_1 (loop, XEXP (x, i), insn, bl,
8909 eliminate_p, where))
8914 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8915 if (! maybe_eliminate_biv_1 (loop, XVECEXP (x, i, j), insn, bl,
8916 eliminate_p, where))
8925 /* Return nonzero if the last use of REG
8926 is in an insn following INSN in the same basic block. */
8929 last_use_this_basic_block (reg, insn)
8935 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
8938 if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
8944 /* Called via `note_stores' to record the initial value of a biv. Here we
8945 just record the location of the set and process it later. */
8948 record_initial (dest, set, data)
8951 void *data ATTRIBUTE_UNUSED;
8953 struct iv_class *bl;
8955 if (GET_CODE (dest) != REG
8956 || REGNO (dest) >= max_reg_before_loop
8957 || REG_IV_TYPE (REGNO (dest)) != BASIC_INDUCT)
8960 bl = reg_biv_class[REGNO (dest)];
8962 /* If this is the first set found, record it. */
8963 if (bl->init_insn == 0)
8965 bl->init_insn = note_insn;
8970 /* If any of the registers in X are "old" and currently have a last use earlier
8971 than INSN, update them to have a last use of INSN. Their actual last use
8972 will be the previous insn but it will not have a valid uid_luid so we can't
8976 update_reg_last_use (x, insn)
8980 /* Check for the case where INSN does not have a valid luid. In this case,
8981 there is no need to modify the regno_last_uid, as this can only happen
8982 when code is inserted after the loop_end to set a pseudo's final value,
8983 and hence this insn will never be the last use of x. */
8984 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
8985 && INSN_UID (insn) < max_uid_for_loop
8986 && uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)])
8987 REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
8991 register const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
8992 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
8995 update_reg_last_use (XEXP (x, i), insn);
8996 else if (fmt[i] == 'E')
8997 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8998 update_reg_last_use (XVECEXP (x, i, j), insn);
9003 /* Given an insn INSN and condition COND, return the condition in a
9004 canonical form to simplify testing by callers. Specifically:
9006 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
9007 (2) Both operands will be machine operands; (cc0) will have been replaced.
9008 (3) If an operand is a constant, it will be the second operand.
9009 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
9010 for GE, GEU, and LEU.
9012 If the condition cannot be understood, or is an inequality floating-point
9013 comparison which needs to be reversed, 0 will be returned.
9015 If REVERSE is non-zero, then reverse the condition prior to canonizing it.
9017 If EARLIEST is non-zero, it is a pointer to a place where the earliest
9018 insn used in locating the condition was found. If a replacement test
9019 of the condition is desired, it should be placed in front of that
9020 insn and we will be sure that the inputs are still valid. */
9023 canonicalize_condition (insn, cond, reverse, earliest)
9034 int reverse_code = 0;
9035 int did_reverse_condition = 0;
9036 enum machine_mode mode;
9038 code = GET_CODE (cond);
9039 mode = GET_MODE (cond);
9040 op0 = XEXP (cond, 0);
9041 op1 = XEXP (cond, 1);
9045 code = reverse_condition (code);
9046 did_reverse_condition ^= 1;
9052 /* If we are comparing a register with zero, see if the register is set
9053 in the previous insn to a COMPARE or a comparison operation. Perform
9054 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
9057 while (GET_RTX_CLASS (code) == '<' && op1 == CONST0_RTX (GET_MODE (op0)))
9059 /* Set non-zero when we find something of interest. */
9063 /* If comparison with cc0, import actual comparison from compare
9067 if ((prev = prev_nonnote_insn (prev)) == 0
9068 || GET_CODE (prev) != INSN
9069 || (set = single_set (prev)) == 0
9070 || SET_DEST (set) != cc0_rtx)
9073 op0 = SET_SRC (set);
9074 op1 = CONST0_RTX (GET_MODE (op0));
9080 /* If this is a COMPARE, pick up the two things being compared. */
9081 if (GET_CODE (op0) == COMPARE)
9083 op1 = XEXP (op0, 1);
9084 op0 = XEXP (op0, 0);
9087 else if (GET_CODE (op0) != REG)
9090 /* Go back to the previous insn. Stop if it is not an INSN. We also
9091 stop if it isn't a single set or if it has a REG_INC note because
9092 we don't want to bother dealing with it. */
9094 if ((prev = prev_nonnote_insn (prev)) == 0
9095 || GET_CODE (prev) != INSN
9096 || FIND_REG_INC_NOTE (prev, 0)
9097 || (set = single_set (prev)) == 0)
9100 /* If this is setting OP0, get what it sets it to if it looks
9102 if (rtx_equal_p (SET_DEST (set), op0))
9104 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
9106 /* ??? We may not combine comparisons done in a CCmode with
9107 comparisons not done in a CCmode. This is to aid targets
9108 like Alpha that have an IEEE compliant EQ instruction, and
9109 a non-IEEE compliant BEQ instruction. The use of CCmode is
9110 actually artificial, simply to prevent the combination, but
9111 should not affect other platforms.
9113 However, we must allow VOIDmode comparisons to match either
9114 CCmode or non-CCmode comparison, because some ports have
9115 modeless comparisons inside branch patterns.
9117 ??? This mode check should perhaps look more like the mode check
9118 in simplify_comparison in combine. */
9120 if ((GET_CODE (SET_SRC (set)) == COMPARE
9123 && GET_MODE_CLASS (inner_mode) == MODE_INT
9124 && (GET_MODE_BITSIZE (inner_mode)
9125 <= HOST_BITS_PER_WIDE_INT)
9126 && (STORE_FLAG_VALUE
9127 & ((HOST_WIDE_INT) 1
9128 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9129 #ifdef FLOAT_STORE_FLAG_VALUE
9131 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9132 && (REAL_VALUE_NEGATIVE
9133 (FLOAT_STORE_FLAG_VALUE (inner_mode))))
9136 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'))
9137 && (((GET_MODE_CLASS (mode) == MODE_CC)
9138 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9139 || mode == VOIDmode || inner_mode == VOIDmode))
9141 else if (((code == EQ
9143 && (GET_MODE_BITSIZE (inner_mode)
9144 <= HOST_BITS_PER_WIDE_INT)
9145 && GET_MODE_CLASS (inner_mode) == MODE_INT
9146 && (STORE_FLAG_VALUE
9147 & ((HOST_WIDE_INT) 1
9148 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9149 #ifdef FLOAT_STORE_FLAG_VALUE
9151 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9152 && (REAL_VALUE_NEGATIVE
9153 (FLOAT_STORE_FLAG_VALUE (inner_mode))))
9156 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'
9157 && (((GET_MODE_CLASS (mode) == MODE_CC)
9158 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9159 || mode == VOIDmode || inner_mode == VOIDmode))
9162 /* We might have reversed a LT to get a GE here. But this wasn't
9163 actually the comparison of data, so we don't flag that we
9164 have had to reverse the condition. */
9165 did_reverse_condition ^= 1;
9173 else if (reg_set_p (op0, prev))
9174 /* If this sets OP0, but not directly, we have to give up. */
9179 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
9180 code = GET_CODE (x);
9183 code = reverse_condition (code);
9184 if (code == UNKNOWN)
9186 did_reverse_condition ^= 1;
9190 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
9196 /* If constant is first, put it last. */
9197 if (CONSTANT_P (op0))
9198 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
9200 /* If OP0 is the result of a comparison, we weren't able to find what
9201 was really being compared, so fail. */
9202 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
9205 /* Canonicalize any ordered comparison with integers involving equality
9206 if we can do computations in the relevant mode and we do not
9209 if (GET_CODE (op1) == CONST_INT
9210 && GET_MODE (op0) != VOIDmode
9211 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
9213 HOST_WIDE_INT const_val = INTVAL (op1);
9214 unsigned HOST_WIDE_INT uconst_val = const_val;
9215 unsigned HOST_WIDE_INT max_val
9216 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
9221 if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
9222 code = LT, op1 = GEN_INT (const_val + 1);
9225 /* When cross-compiling, const_val might be sign-extended from
9226 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
9228 if ((HOST_WIDE_INT) (const_val & max_val)
9229 != (((HOST_WIDE_INT) 1
9230 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
9231 code = GT, op1 = GEN_INT (const_val - 1);
9235 if (uconst_val < max_val)
9236 code = LTU, op1 = GEN_INT (uconst_val + 1);
9240 if (uconst_val != 0)
9241 code = GTU, op1 = GEN_INT (uconst_val - 1);
9249 /* If this was floating-point and we reversed anything other than an
9250 EQ or NE or (UN)ORDERED, return zero. */
9251 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
9252 && did_reverse_condition
9253 && code != NE && code != EQ && code != UNORDERED && code != ORDERED
9255 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
9259 /* Never return CC0; return zero instead. */
9264 return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);
9267 /* Given a jump insn JUMP, return the condition that will cause it to branch
9268 to its JUMP_LABEL. If the condition cannot be understood, or is an
9269 inequality floating-point comparison which needs to be reversed, 0 will
9272 If EARLIEST is non-zero, it is a pointer to a place where the earliest
9273 insn used in locating the condition was found. If a replacement test
9274 of the condition is desired, it should be placed in front of that
9275 insn and we will be sure that the inputs are still valid. */
9278 get_condition (jump, earliest)
9285 /* If this is not a standard conditional jump, we can't parse it. */
9286 if (GET_CODE (jump) != JUMP_INSN
9287 || ! condjump_p (jump) || simplejump_p (jump))
9290 cond = XEXP (SET_SRC (PATTERN (jump)), 0);
9292 /* If this branches to JUMP_LABEL when the condition is false, reverse
9295 = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF
9296 && XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump);
9298 return canonicalize_condition (jump, cond, reverse, earliest);
9301 /* Similar to above routine, except that we also put an invariant last
9302 unless both operands are invariants. */
9305 get_condition_for_loop (loop, x)
9306 const struct loop *loop;
9309 rtx comparison = get_condition (x, NULL_PTR);
9312 || ! loop_invariant_p (loop, XEXP (comparison, 0))
9313 || loop_invariant_p (loop, XEXP (comparison, 1)))
9316 return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode,
9317 XEXP (comparison, 1), XEXP (comparison, 0));
9320 #ifdef HAVE_decrement_and_branch_on_count
9321 /* Instrument loop for insertion of bct instruction. We distinguish between
9322 loops with compile-time bounds and those with run-time bounds.
9323 Information from loop_iterations() is used to compute compile-time bounds.
9324 Run-time bounds should use loop preconditioning, but currently ignored.
9331 unsigned HOST_WIDE_INT n_iterations;
9332 rtx loop_start = loop->start;
9333 rtx loop_end = loop->end;
9334 struct loop_info *loop_info = LOOP_INFO (loop);
9335 int loop_num = loop->num;
9338 int increment_direction, compare_direction;
9339 /* If the loop condition is <= or >=, the number of iteration
9340 is 1 more than the range of the bounds of the loop. */
9341 int add_iteration = 0;
9342 enum machine_mode loop_var_mode = word_mode;
9345 /* It's impossible to instrument a competely unrolled loop. */
9346 if (loop_info->unroll_number == loop_info->n_iterations)
9349 /* Make sure that the count register is not in use. */
9350 if (loop_info->used_count_register)
9352 if (loop_dump_stream)
9353 fprintf (loop_dump_stream,
9354 "insert_bct %d: BCT instrumentation failed: count register already in use\n",
9359 /* Make sure that the function has no indirect jumps. */
9360 if (indirect_jump_in_function)
9362 if (loop_dump_stream)
9363 fprintf (loop_dump_stream,
9364 "insert_bct %d: BCT instrumentation failed: indirect jump in function\n",
9369 /* Make sure that the last loop insn is a conditional jump. */
9370 if (GET_CODE (PREV_INSN (loop_end)) != JUMP_INSN
9371 || ! condjump_p (PREV_INSN (loop_end))
9372 || simplejump_p (PREV_INSN (loop_end)))
9374 if (loop_dump_stream)
9375 fprintf (loop_dump_stream,
9376 "insert_bct %d: BCT instrumentation failed: invalid jump at loop end\n",
9381 /* Make sure that the loop does not contain a function call
9382 (the count register might be altered by the called function). */
9383 if (loop_info->has_call)
9385 if (loop_dump_stream)
9386 fprintf (loop_dump_stream,
9387 "insert_bct %d: BCT instrumentation failed: function call in loop\n",
9392 /* Make sure that the loop does not jump via a table.
9393 (the count register might be used to perform the branch on table). */
9394 if (loop_info->has_tablejump)
9396 if (loop_dump_stream)
9397 fprintf (loop_dump_stream,
9398 "insert_bct %d: BCT instrumentation failed: computed branch in the loop\n",
9403 /* Account for loop unrolling in instrumented iteration count. */
9404 if (loop_info->unroll_number > 1)
9405 n_iterations = loop_info->n_iterations / loop_info->unroll_number;
9407 n_iterations = loop_info->n_iterations;
9409 if (n_iterations != 0 && n_iterations < 3)
9411 /* Allow an enclosing outer loop to benefit if possible. */
9412 if (loop_dump_stream)
9413 fprintf (loop_dump_stream,
9414 "insert_bct %d: Too few iterations to benefit from BCT optimization\n",
9419 /* Try to instrument the loop. */
9421 /* Handle the simpler case, where the bounds are known at compile time. */
9422 if (n_iterations > 0)
9424 struct loop *outer_loop;
9425 struct loop_info *outer_loop_info;
9427 /* Mark all enclosing loops that they cannot use count register. */
9428 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
9430 outer_loop_info = LOOP_INFO (outer_loop);
9431 outer_loop_info->used_count_register = 1;
9433 instrument_loop_bct (loop_start, loop_end, GEN_INT (n_iterations));
9437 /* Handle the more complex case, that the bounds are NOT known
9438 at compile time. In this case we generate run_time calculation
9439 of the number of iterations. */
9441 if (loop_info->iteration_var == 0)
9443 if (loop_dump_stream)
9444 fprintf (loop_dump_stream,
9445 "insert_bct %d: BCT Runtime Instrumentation failed: no loop iteration variable found\n",
9450 if (GET_MODE_CLASS (GET_MODE (loop_info->iteration_var)) != MODE_INT
9451 || GET_MODE_SIZE (GET_MODE (loop_info->iteration_var)) != UNITS_PER_WORD)
9453 if (loop_dump_stream)
9454 fprintf (loop_dump_stream,
9455 "insert_bct %d: BCT Runtime Instrumentation failed: loop variable not integer\n",
9460 /* With runtime bounds, if the compare is of the form '!=' we give up */
9461 if (loop_info->comparison_code == NE)
9463 if (loop_dump_stream)
9464 fprintf (loop_dump_stream,
9465 "insert_bct %d: BCT Runtime Instrumentation failed: runtime bounds with != comparison\n",
9469 /* Use common loop preconditioning code instead. */
9473 /* We rely on the existence of run-time guard to ensure that the
9474 loop executes at least once. */
9476 rtx iterations_num_reg;
9478 unsigned HOST_WIDE_INT increment_value_abs
9479 = INTVAL (increment) * increment_direction;
9481 /* make sure that the increment is a power of two, otherwise (an
9482 expensive) divide is needed. */
9483 if (exact_log2 (increment_value_abs) == -1)
9485 if (loop_dump_stream)
9486 fprintf (loop_dump_stream,
9487 "insert_bct: not instrumenting BCT because the increment is not power of 2\n");
9491 /* compute the number of iterations */
9496 /* Again, the number of iterations is calculated by:
9498 ; compare-val - initial-val + (increment -1) + additional-iteration
9499 ; num_iterations = -----------------------------------------------------------------
9502 /* ??? Do we have to call copy_rtx here before passing rtx to
9504 if (compare_direction > 0)
9506 /* <, <= :the loop variable is increasing */
9507 temp_reg = expand_binop (loop_var_mode, sub_optab,
9508 comparison_value, initial_value,
9509 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9513 temp_reg = expand_binop (loop_var_mode, sub_optab,
9514 initial_value, comparison_value,
9515 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9518 if (increment_value_abs - 1 + add_iteration != 0)
9519 temp_reg = expand_binop (loop_var_mode, add_optab, temp_reg,
9520 GEN_INT (increment_value_abs - 1
9522 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9524 if (increment_value_abs != 1)
9525 iterations_num_reg = expand_binop (loop_var_mode, asr_optab,
9527 GEN_INT (exact_log2 (increment_value_abs)),
9528 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9530 iterations_num_reg = temp_reg;
9532 sequence = gen_sequence ();
9534 emit_insn_before (sequence, loop_start);
9535 instrument_loop_bct (loop_start, loop_end, iterations_num_reg);
9539 #endif /* Complex case */
9542 /* Instrument loop by inserting a bct in it as follows:
9543 1. A new counter register is created.
9544 2. In the head of the loop the new variable is initialized to the value
9545 passed in the loop_num_iterations parameter.
9546 3. At the end of the loop, comparison of the register with 0 is generated.
9547 The created comparison follows the pattern defined for the
9548 decrement_and_branch_on_count insn, so this insn will be generated.
9549 4. The branch on the old variable are deleted. The compare must remain
9550 because it might be used elsewhere. If the loop-variable or condition
9551 register are used elsewhere, they will be eliminated by flow. */
9554 instrument_loop_bct (loop_start, loop_end, loop_num_iterations)
9555 rtx loop_start, loop_end;
9556 rtx loop_num_iterations;
9562 if (HAVE_decrement_and_branch_on_count)
9564 if (loop_dump_stream)
9566 fputs ("instrument_bct: Inserting BCT (", loop_dump_stream);
9567 if (GET_CODE (loop_num_iterations) == CONST_INT)
9568 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
9569 INTVAL (loop_num_iterations));
9571 fputs ("runtime", loop_dump_stream);
9572 fputs (" iterations)", loop_dump_stream);
9575 /* Discard original jump to continue loop. Original compare result
9576 may still be live, so it cannot be discarded explicitly. */
9577 delete_insn (PREV_INSN (loop_end));
9579 /* Insert the label which will delimit the start of the loop. */
9580 start_label = gen_label_rtx ();
9581 emit_label_after (start_label, loop_start);
9583 /* Insert initialization of the count register into the loop header. */
9585 counter_reg = gen_reg_rtx (word_mode);
9586 emit_insn (gen_move_insn (counter_reg, loop_num_iterations));
9587 sequence = gen_sequence ();
9589 emit_insn_before (sequence, loop_start);
9591 /* Insert new comparison on the count register instead of the
9592 old one, generating the needed BCT pattern (that will be
9593 later recognized by assembly generation phase). */
9594 emit_jump_insn_before (gen_decrement_and_branch_on_count (counter_reg,
9597 LABEL_NUSES (start_label)++;
9601 #endif /* HAVE_decrement_and_branch_on_count */
9603 /* Scan the function and determine whether it has indirect (computed) jumps.
9605 This is taken mostly from flow.c; similar code exists elsewhere
9606 in the compiler. It may be useful to put this into rtlanal.c. */
9608 indirect_jump_in_function_p (start)
9613 for (insn = start; insn; insn = NEXT_INSN (insn))
9614 if (computed_jump_p (insn))
9620 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
9621 documentation for LOOP_MEMS for the definition of `appropriate'.
9622 This function is called from prescan_loop via for_each_rtx. */
9625 insert_loop_mem (mem, data)
9627 void *data ATTRIBUTE_UNUSED;
9635 switch (GET_CODE (m))
9641 /* We're not interested in MEMs that are only clobbered. */
9645 /* We're not interested in the MEM associated with a
9646 CONST_DOUBLE, so there's no need to traverse into this. */
9650 /* We're not interested in any MEMs that only appear in notes. */
9654 /* This is not a MEM. */
9658 /* See if we've already seen this MEM. */
9659 for (i = 0; i < loop_mems_idx; ++i)
9660 if (rtx_equal_p (m, loop_mems[i].mem))
9662 if (GET_MODE (m) != GET_MODE (loop_mems[i].mem))
9663 /* The modes of the two memory accesses are different. If
9664 this happens, something tricky is going on, and we just
9665 don't optimize accesses to this MEM. */
9666 loop_mems[i].optimize = 0;
9671 /* Resize the array, if necessary. */
9672 if (loop_mems_idx == loop_mems_allocated)
9674 if (loop_mems_allocated != 0)
9675 loop_mems_allocated *= 2;
9677 loop_mems_allocated = 32;
9679 loop_mems = (loop_mem_info*)
9680 xrealloc (loop_mems,
9681 loop_mems_allocated * sizeof (loop_mem_info));
9684 /* Actually insert the MEM. */
9685 loop_mems[loop_mems_idx].mem = m;
9686 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
9687 because we can't put it in a register. We still store it in the
9688 table, though, so that if we see the same address later, but in a
9689 non-BLK mode, we'll not think we can optimize it at that point. */
9690 loop_mems[loop_mems_idx].optimize = (GET_MODE (m) != BLKmode);
9691 loop_mems[loop_mems_idx].reg = NULL_RTX;
9697 /* Like load_mems, but also ensures that SET_IN_LOOP,
9698 MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct
9699 values after load_mems. */
9702 load_mems_and_recount_loop_regs_set (loop, insn_count)
9703 const struct loop *loop;
9706 int nregs = max_reg_num ();
9710 /* Recalculate set_in_loop and friends since load_mems may have
9711 created new registers. */
9712 if (max_reg_num () > nregs)
9718 nregs = max_reg_num ();
9720 if ((unsigned) nregs > set_in_loop->num_elements)
9722 /* Grow all the arrays. */
9723 VARRAY_GROW (set_in_loop, nregs);
9724 VARRAY_GROW (n_times_set, nregs);
9725 VARRAY_GROW (may_not_optimize, nregs);
9726 VARRAY_GROW (reg_single_usage, nregs);
9728 /* Clear the arrays */
9729 bzero ((char *) &set_in_loop->data, nregs * sizeof (int));
9730 bzero ((char *) &may_not_optimize->data, nregs * sizeof (char));
9731 bzero ((char *) ®_single_usage->data, nregs * sizeof (rtx));
9733 count_loop_regs_set (loop->top ? loop->top : loop->start, loop->end,
9734 may_not_optimize, reg_single_usage,
9737 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
9739 VARRAY_CHAR (may_not_optimize, i) = 1;
9740 VARRAY_INT (set_in_loop, i) = 1;
9743 #ifdef AVOID_CCMODE_COPIES
9744 /* Don't try to move insns which set CC registers if we should not
9745 create CCmode register copies. */
9746 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
9747 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
9748 VARRAY_CHAR (may_not_optimize, i) = 1;
9751 /* Set n_times_set for the new registers. */
9752 bcopy ((char *) (&set_in_loop->data.i[0] + old_nregs),
9753 (char *) (&n_times_set->data.i[0] + old_nregs),
9754 (nregs - old_nregs) * sizeof (int));
9758 /* Move MEMs into registers for the duration of the loop. */
9762 const struct loop *loop;
9764 int maybe_never = 0;
9767 rtx label = NULL_RTX;
9768 rtx end_label = NULL_RTX;
9769 /* Nonzero if the next instruction may never be executed. */
9770 int next_maybe_never = 0;
9771 int last_max_reg = max_reg_num ();
9773 if (loop_mems_idx == 0)
9776 /* Check to see if it's possible that some instructions in the
9777 loop are never executed. */
9778 for (p = next_insn_in_loop (loop, loop->scan_start);
9779 p != NULL_RTX && ! maybe_never;
9780 p = next_insn_in_loop (loop, p))
9782 if (GET_CODE (p) == CODE_LABEL)
9784 else if (GET_CODE (p) == JUMP_INSN
9785 /* If we enter the loop in the middle, and scan
9786 around to the beginning, don't set maybe_never
9787 for that. This must be an unconditional jump,
9788 otherwise the code at the top of the loop might
9789 never be executed. Unconditional jumps are
9790 followed a by barrier then loop end. */
9791 && ! (GET_CODE (p) == JUMP_INSN
9792 && JUMP_LABEL (p) == loop->top
9793 && NEXT_INSN (NEXT_INSN (p)) == loop->end
9794 && simplejump_p (p)))
9796 if (!condjump_p (p))
9797 /* Something complicated. */
9800 /* If there are any more instructions in the loop, they
9801 might not be reached. */
9802 next_maybe_never = 1;
9804 else if (next_maybe_never)
9808 /* Actually move the MEMs. */
9809 for (i = 0; i < loop_mems_idx; ++i)
9814 rtx mem = loop_mems[i].mem;
9817 if (MEM_VOLATILE_P (mem)
9818 || loop_invariant_p (loop, XEXP (mem, 0)) != 1)
9819 /* There's no telling whether or not MEM is modified. */
9820 loop_mems[i].optimize = 0;
9822 /* Go through the MEMs written to in the loop to see if this
9823 one is aliased by one of them. */
9824 mem_list_entry = loop_store_mems;
9825 while (mem_list_entry)
9827 if (rtx_equal_p (mem, XEXP (mem_list_entry, 0)))
9829 else if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
9832 /* MEM is indeed aliased by this store. */
9833 loop_mems[i].optimize = 0;
9836 mem_list_entry = XEXP (mem_list_entry, 1);
9839 if (flag_float_store && written
9840 && GET_MODE_CLASS (GET_MODE (mem)) == MODE_FLOAT)
9841 loop_mems[i].optimize = 0;
9843 /* If this MEM is written to, we must be sure that there
9844 are no reads from another MEM that aliases this one. */
9845 if (loop_mems[i].optimize && written)
9849 for (j = 0; j < loop_mems_idx; ++j)
9853 else if (true_dependence (mem,
9858 /* It's not safe to hoist loop_mems[i] out of
9859 the loop because writes to it might not be
9860 seen by reads from loop_mems[j]. */
9861 loop_mems[i].optimize = 0;
9867 if (maybe_never && may_trap_p (mem))
9868 /* We can't access the MEM outside the loop; it might
9869 cause a trap that wouldn't have happened otherwise. */
9870 loop_mems[i].optimize = 0;
9872 if (!loop_mems[i].optimize)
9873 /* We thought we were going to lift this MEM out of the
9874 loop, but later discovered that we could not. */
9877 INIT_REG_SET (&copies);
9879 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
9880 order to keep scan_loop from moving stores to this MEM
9881 out of the loop just because this REG is neither a
9882 user-variable nor used in the loop test. */
9883 reg = gen_reg_rtx (GET_MODE (mem));
9884 REG_USERVAR_P (reg) = 1;
9885 loop_mems[i].reg = reg;
9887 /* Now, replace all references to the MEM with the
9888 corresponding pesudos. */
9890 for (p = next_insn_in_loop (loop, loop->scan_start);
9892 p = next_insn_in_loop (loop, p))
9897 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
9899 /* See if this copies the mem into a register that isn't
9900 modified afterwards. We'll try to do copy propagation
9901 a little further on. */
9902 set = single_set (p);
9904 /* @@@ This test is _way_ too conservative. */
9906 && GET_CODE (SET_DEST (set)) == REG
9907 && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER
9908 && REGNO (SET_DEST (set)) < last_max_reg
9909 && VARRAY_INT (n_times_set, REGNO (SET_DEST (set))) == 1
9910 && rtx_equal_p (SET_SRC (set), loop_mems[i].mem))
9911 SET_REGNO_REG_SET (&copies, REGNO (SET_DEST (set)));
9914 for_each_rtx (&p, replace_loop_mem, &ri);
9917 if (GET_CODE (p) == CODE_LABEL
9918 || GET_CODE (p) == JUMP_INSN)
9922 if (! apply_change_group ())
9923 /* We couldn't replace all occurrences of the MEM. */
9924 loop_mems[i].optimize = 0;
9930 /* Load the memory immediately before LOOP->START, which is
9931 the NOTE_LOOP_BEG. */
9932 set = gen_move_insn (reg, mem);
9933 emit_insn_before (set, loop->start);
9937 if (label == NULL_RTX)
9939 /* We must compute the former
9940 right-after-the-end label before we insert
9942 end_label = next_label (loop->end);
9943 label = gen_label_rtx ();
9944 emit_label_after (label, loop->end);
9947 /* Store the memory immediately after END, which is
9948 the NOTE_LOOP_END. */
9949 set = gen_move_insn (copy_rtx (mem), reg);
9950 emit_insn_after (set, label);
9953 if (loop_dump_stream)
9955 fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
9956 REGNO (reg), (written ? "r/w" : "r/o"));
9957 print_rtl (loop_dump_stream, mem);
9958 fputc ('\n', loop_dump_stream);
9961 /* Attempt a bit of copy propagation. This helps untangle the
9962 data flow, and enables {basic,general}_induction_var to find
9964 EXECUTE_IF_SET_IN_REG_SET
9965 (&copies, FIRST_PSEUDO_REGISTER, j,
9967 try_copy_prop (loop, loop_mems[i].reg, j);
9969 CLEAR_REG_SET (&copies);
9973 if (label != NULL_RTX)
9975 /* Now, we need to replace all references to the previous exit
9976 label with the new one. */
9981 for (p = loop->start; p != loop->end; p = NEXT_INSN (p))
9983 for_each_rtx (&p, replace_label, &rr);
9985 /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
9986 field. This is not handled by for_each_rtx because it doesn't
9987 handle unprinted ('0') fields. We need to update JUMP_LABEL
9988 because the immediately following unroll pass will use it.
9989 replace_label would not work anyways, because that only handles
9991 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == end_label)
9992 JUMP_LABEL (p) = label;
9997 /* For communication between note_reg_stored and its caller. */
9998 struct note_reg_stored_arg
10004 /* Called via note_stores, record in SET_SEEN whether X, which is written,
10005 is equal to ARG. */
10007 note_reg_stored (x, setter, arg)
10008 rtx x, setter ATTRIBUTE_UNUSED;
10011 struct note_reg_stored_arg *t = (struct note_reg_stored_arg *)arg;
10016 /* Try to replace every occurrence of pseudo REGNO with REPLACEMENT.
10017 There must be exactly one insn that sets this pseudo; it will be
10018 deleted if all replacements succeed and we can prove that the register
10019 is not used after the loop. */
10021 try_copy_prop (loop, replacement, regno)
10022 const struct loop *loop;
10026 /* This is the reg that we are copying from. */
10027 rtx reg_rtx = regno_reg_rtx[regno];
10030 /* These help keep track of whether we replaced all uses of the reg. */
10031 int replaced_last = 0;
10032 int store_is_first = 0;
10034 for (insn = next_insn_in_loop (loop, loop->scan_start);
10036 insn = next_insn_in_loop (loop, insn))
10040 /* Only substitute within one extended basic block from the initializing
10042 if (GET_CODE (insn) == CODE_LABEL && init_insn)
10045 if (GET_RTX_CLASS (GET_CODE (insn)) != 'i')
10048 /* Is this the initializing insn? */
10049 set = single_set (insn);
10051 && GET_CODE (SET_DEST (set)) == REG
10052 && REGNO (SET_DEST (set)) == regno)
10058 if (REGNO_FIRST_UID (regno) == INSN_UID (insn))
10059 store_is_first = 1;
10062 /* Only substitute after seeing the initializing insn. */
10063 if (init_insn && insn != init_insn)
10065 struct note_reg_stored_arg arg;
10067 array[0] = reg_rtx;
10068 array[1] = replacement;
10071 for_each_rtx (&insn, replace_loop_reg, array);
10072 if (REGNO_LAST_UID (regno) == INSN_UID (insn))
10075 /* Stop replacing when REPLACEMENT is modified. */
10076 arg.reg = replacement;
10078 note_stores (PATTERN (insn), note_reg_stored, &arg);
10085 if (apply_change_group ())
10087 if (loop_dump_stream)
10088 fprintf (loop_dump_stream, " Replaced reg %d", regno);
10089 if (store_is_first && replaced_last)
10091 PUT_CODE (init_insn, NOTE);
10092 NOTE_LINE_NUMBER (init_insn) = NOTE_INSN_DELETED;
10093 if (loop_dump_stream)
10094 fprintf (loop_dump_stream, ", deleting init_insn (%d)",
10095 INSN_UID (init_insn));
10097 if (loop_dump_stream)
10098 fprintf (loop_dump_stream, ".\n");
10102 /* Replace MEM with its associated pseudo register. This function is
10103 called from load_mems via for_each_rtx. DATA is actually an
10104 rtx_and_int * describing the instruction currently being scanned
10105 and the MEM we are currently replacing. */
10108 replace_loop_mem (mem, data)
10120 switch (GET_CODE (m))
10126 /* We're not interested in the MEM associated with a
10127 CONST_DOUBLE, so there's no need to traverse into one. */
10131 /* This is not a MEM. */
10135 ri = (rtx_and_int*) data;
10138 if (!rtx_equal_p (loop_mems[i].mem, m))
10139 /* This is not the MEM we are currently replacing. */
10144 /* Actually replace the MEM. */
10145 validate_change (insn, mem, loop_mems[i].reg, 1);
10150 /* Replace one register with another. Called through for_each_rtx; PX points
10151 to the rtx being scanned. DATA is actually an array of three rtx's; the
10152 first one is the one to be replaced, and the second one the replacement.
10153 The third one is the current insn. */
10156 replace_loop_reg (px, data)
10161 rtx *array = (rtx *)data;
10167 validate_change (array[2], px, array[1], 1);
10172 /* Replace occurrences of the old exit label for the loop with the new
10173 one. DATA is an rtx_pair containing the old and new labels,
10177 replace_label (x, data)
10182 rtx old_label = ((rtx_pair*) data)->r1;
10183 rtx new_label = ((rtx_pair*) data)->r2;
10188 if (GET_CODE (l) != LABEL_REF)
10191 if (XEXP (l, 0) != old_label)
10194 XEXP (l, 0) = new_label;
10195 ++LABEL_NUSES (new_label);
10196 --LABEL_NUSES (old_label);