1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 88, 89, 91-98, 1999 Free Software Foundation, Inc.
4 This file is part of GNU CC.
6 GNU CC is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2, or (at your option)
11 GNU CC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with GNU CC; see the file COPYING. If not, write to
18 the Free Software Foundation, 59 Temple Place - Suite 330,
19 Boston, MA 02111-1307, USA. */
22 /* This is the loop optimization pass of the compiler.
23 It finds invariant computations within loops and moves them
24 to the beginning of the loop. Then it identifies basic and
25 general induction variables. Strength reduction is applied to the general
26 induction variables, and induction variable elimination is applied to
27 the basic induction variables.
29 It also finds cases where
30 a register is set within the loop by zero-extending a narrower value
31 and changes these to zero the entire register once before the loop
32 and merely copy the low part within the loop.
34 Most of the complexity is in heuristics to decide when it is worth
35 while to do these things. */
43 #include "insn-config.h"
44 #include "insn-flags.h"
46 #include "hard-reg-set.h"
54 /* Vector mapping INSN_UIDs to luids.
55 The luids are like uids but increase monotonically always.
56 We use them to see whether a jump comes from outside a given loop. */
60 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
61 number the insn is contained in. */
65 /* 1 + largest uid of any insn. */
69 /* 1 + luid of last insn. */
73 /* Number of loops detected in current function. Used as index to the
76 static int max_loop_num;
78 /* Indexed by loop number, contains the first and last insn of each loop. */
80 static rtx *loop_number_loop_starts, *loop_number_loop_ends;
82 /* Likewise for the continue insn */
83 static rtx *loop_number_loop_cont;
85 /* The first code_label that is reached in every loop iteration.
86 0 when not computed yet, initially const0_rtx if a jump couldn't be
88 Also set to 0 when there is no such label before the NOTE_INSN_LOOP_CONT
89 of this loop, or in verify_dominator, if a jump couldn't be followed. */
90 static rtx *loop_number_cont_dominator;
92 /* For each loop, gives the containing loop number, -1 if none. */
96 #ifdef HAVE_decrement_and_branch_on_count
97 /* Records whether resource in use by inner loop. */
99 int *loop_used_count_register;
100 #endif /* HAVE_decrement_and_branch_on_count */
102 /* Indexed by loop number, contains a nonzero value if the "loop" isn't
103 really a loop (an insn outside the loop branches into it). */
105 static char *loop_invalid;
107 /* Indexed by loop number, links together all LABEL_REFs which refer to
108 code labels outside the loop. Used by routines that need to know all
109 loop exits, such as final_biv_value and final_giv_value.
111 This does not include loop exits due to return instructions. This is
112 because all bivs and givs are pseudos, and hence must be dead after a
113 return, so the presense of a return does not affect any of the
114 optimizations that use this info. It is simpler to just not include return
115 instructions on this list. */
117 rtx *loop_number_exit_labels;
119 /* Indexed by loop number, counts the number of LABEL_REFs on
120 loop_number_exit_labels for this loop and all loops nested inside it. */
122 int *loop_number_exit_count;
124 /* Nonzero if there is a subroutine call in the current loop. */
126 static int loop_has_call;
128 /* Nonzero if there is a volatile memory reference in the current
131 static int loop_has_volatile;
133 /* Nonzero if there is a tablejump in the current loop. */
135 static int loop_has_tablejump;
137 /* Added loop_continue which is the NOTE_INSN_LOOP_CONT of the
138 current loop. A continue statement will generate a branch to
139 NEXT_INSN (loop_continue). */
141 static rtx loop_continue;
143 /* Indexed by register number, contains the number of times the reg
144 is set during the loop being scanned.
145 During code motion, a negative value indicates a reg that has been
146 made a candidate; in particular -2 means that it is an candidate that
147 we know is equal to a constant and -1 means that it is an candidate
148 not known equal to a constant.
149 After code motion, regs moved have 0 (which is accurate now)
150 while the failed candidates have the original number of times set.
152 Therefore, at all times, == 0 indicates an invariant register;
153 < 0 a conditionally invariant one. */
155 static varray_type set_in_loop;
157 /* Original value of set_in_loop; same except that this value
158 is not set negative for a reg whose sets have been made candidates
159 and not set to 0 for a reg that is moved. */
161 static varray_type n_times_set;
163 /* Index by register number, 1 indicates that the register
164 cannot be moved or strength reduced. */
166 static varray_type may_not_optimize;
168 /* Contains the insn in which a register was used if it was used
169 exactly once; contains const0_rtx if it was used more than once. */
171 static varray_type reg_single_usage;
173 /* Nonzero means reg N has already been moved out of one loop.
174 This reduces the desire to move it out of another. */
176 static char *moved_once;
178 /* List of MEMs that are stored in this loop. */
180 static rtx loop_store_mems;
182 /* The insn where the first of these was found. */
183 static rtx first_loop_store_insn;
185 typedef struct loop_mem_info {
186 rtx mem; /* The MEM itself. */
187 rtx reg; /* Corresponding pseudo, if any. */
188 int optimize; /* Nonzero if we can optimize access to this MEM. */
191 /* Array of MEMs that are used (read or written) in this loop, but
192 cannot be aliased by anything in this loop, except perhaps
193 themselves. In other words, if loop_mems[i] is altered during the
194 loop, it is altered by an expression that is rtx_equal_p to it. */
196 static loop_mem_info *loop_mems;
198 /* The index of the next available slot in LOOP_MEMS. */
200 static int loop_mems_idx;
202 /* The number of elements allocated in LOOP_MEMs. */
204 static int loop_mems_allocated;
206 /* Nonzero if we don't know what MEMs were changed in the current loop.
207 This happens if the loop contains a call (in which case `loop_has_call'
208 will also be set) or if we store into more than NUM_STORES MEMs. */
210 static int unknown_address_altered;
212 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
213 static int num_movables;
215 /* Count of memory write instructions discovered in the loop. */
216 static int num_mem_sets;
218 /* Number of loops contained within the current one, including itself. */
219 static int loops_enclosed;
221 /* Bound on pseudo register number before loop optimization.
222 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
223 int max_reg_before_loop;
225 /* This obstack is used in product_cheap_p to allocate its rtl. It
226 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
227 If we used the same obstack that it did, we would be deallocating
230 static struct obstack temp_obstack;
232 /* This is where the pointer to the obstack being used for RTL is stored. */
234 extern struct obstack *rtl_obstack;
236 #define obstack_chunk_alloc xmalloc
237 #define obstack_chunk_free free
239 /* During the analysis of a loop, a chain of `struct movable's
240 is made to record all the movable insns found.
241 Then the entire chain can be scanned to decide which to move. */
245 rtx insn; /* A movable insn */
246 rtx set_src; /* The expression this reg is set from. */
247 rtx set_dest; /* The destination of this SET. */
248 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
249 of any registers used within the LIBCALL. */
250 int consec; /* Number of consecutive following insns
251 that must be moved with this one. */
252 int regno; /* The register it sets */
253 short lifetime; /* lifetime of that register;
254 may be adjusted when matching movables
255 that load the same value are found. */
256 short savings; /* Number of insns we can move for this reg,
257 including other movables that force this
258 or match this one. */
259 unsigned int cond : 1; /* 1 if only conditionally movable */
260 unsigned int force : 1; /* 1 means MUST move this insn */
261 unsigned int global : 1; /* 1 means reg is live outside this loop */
262 /* If PARTIAL is 1, GLOBAL means something different:
263 that the reg is live outside the range from where it is set
264 to the following label. */
265 unsigned int done : 1; /* 1 inhibits further processing of this */
267 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
268 In particular, moving it does not make it
270 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
271 load SRC, rather than copying INSN. */
272 unsigned int move_insn_first:1;/* Same as above, if this is necessary for the
273 first insn of a consecutive sets group. */
274 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
275 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
276 that we should avoid changing when clearing
277 the rest of the reg. */
278 struct movable *match; /* First entry for same value */
279 struct movable *forces; /* An insn that must be moved if this is */
280 struct movable *next;
283 static struct movable *the_movables;
285 FILE *loop_dump_stream;
287 /* Forward declarations. */
289 static void verify_dominator PROTO((int));
290 static void find_and_verify_loops PROTO((rtx));
291 static void mark_loop_jump PROTO((rtx, int));
292 static void prescan_loop PROTO((rtx, rtx));
293 static int reg_in_basic_block_p PROTO((rtx, rtx));
294 static int consec_sets_invariant_p PROTO((rtx, int, rtx));
295 static int labels_in_range_p PROTO((rtx, int));
296 static void count_one_set PROTO((rtx, rtx, varray_type, rtx *));
298 static void count_loop_regs_set PROTO((rtx, rtx, varray_type, varray_type,
300 static void note_addr_stored PROTO((rtx, rtx));
301 static int loop_reg_used_before_p PROTO((rtx, rtx, rtx, rtx, rtx));
302 static void scan_loop PROTO((rtx, rtx, rtx, int, int));
304 static void replace_call_address PROTO((rtx, rtx, rtx));
306 static rtx skip_consec_insns PROTO((rtx, int));
307 static int libcall_benefit PROTO((rtx));
308 static void ignore_some_movables PROTO((struct movable *));
309 static void force_movables PROTO((struct movable *));
310 static void combine_movables PROTO((struct movable *, int));
311 static int regs_match_p PROTO((rtx, rtx, struct movable *));
312 static int rtx_equal_for_loop_p PROTO((rtx, rtx, struct movable *));
313 static void add_label_notes PROTO((rtx, rtx));
314 static void move_movables PROTO((struct movable *, int, int, rtx, rtx, int));
315 static int count_nonfixed_reads PROTO((rtx));
316 static void strength_reduce PROTO((rtx, rtx, rtx, int, rtx, rtx, rtx, int, int));
317 static void find_single_use_in_loop PROTO((rtx, rtx, varray_type));
318 static int valid_initial_value_p PROTO((rtx, rtx, int, rtx));
319 static void find_mem_givs PROTO((rtx, rtx, int, rtx, rtx));
320 static void record_biv PROTO((struct induction *, rtx, rtx, rtx, rtx, rtx *, int, int));
321 static void check_final_value PROTO((struct induction *, rtx, rtx,
322 unsigned HOST_WIDE_INT));
323 static void record_giv PROTO((struct induction *, rtx, rtx, rtx, rtx, rtx, int, enum g_types, int, rtx *, rtx, rtx));
324 static void update_giv_derive PROTO((rtx));
325 static int basic_induction_var PROTO((rtx, enum machine_mode, rtx, rtx, rtx *, rtx *, rtx **));
326 static rtx simplify_giv_expr PROTO((rtx, int *));
327 static int general_induction_var PROTO((rtx, rtx *, rtx *, rtx *, int, int *));
328 static int consec_sets_giv PROTO((int, rtx, rtx, rtx, rtx *, rtx *, rtx *));
329 static int check_dbra_loop PROTO((rtx, int, rtx, struct loop_info *));
330 static rtx express_from_1 PROTO((rtx, rtx, rtx));
331 static rtx combine_givs_p PROTO((struct induction *, struct induction *));
332 static void combine_givs PROTO((struct iv_class *));
333 struct recombine_givs_stats;
334 static int find_life_end PROTO((rtx, struct recombine_givs_stats *, rtx, rtx));
335 static void recombine_givs PROTO((struct iv_class *, rtx, rtx, int));
336 static int product_cheap_p PROTO((rtx, rtx));
337 static int maybe_eliminate_biv PROTO((struct iv_class *, rtx, rtx, int, int, int));
338 static int maybe_eliminate_biv_1 PROTO((rtx, rtx, struct iv_class *, int, rtx));
339 static int last_use_this_basic_block PROTO((rtx, rtx));
340 static void record_initial PROTO((rtx, rtx));
341 static void update_reg_last_use PROTO((rtx, rtx));
342 static rtx next_insn_in_loop PROTO((rtx, rtx, rtx, rtx));
343 static void load_mems_and_recount_loop_regs_set PROTO((rtx, rtx, rtx,
345 static void load_mems PROTO((rtx, rtx, rtx, rtx));
346 static int insert_loop_mem PROTO((rtx *, void *));
347 static int replace_loop_mem PROTO((rtx *, void *));
348 static int replace_label PROTO((rtx *, void *));
350 typedef struct rtx_and_int {
355 typedef struct rtx_pair {
360 /* Nonzero iff INSN is between START and END, inclusive. */
361 #define INSN_IN_RANGE_P(INSN, START, END) \
362 (INSN_UID (INSN) < max_uid_for_loop \
363 && INSN_LUID (INSN) >= INSN_LUID (START) \
364 && INSN_LUID (INSN) <= INSN_LUID (END))
366 #ifdef HAVE_decrement_and_branch_on_count
367 /* Test whether BCT applicable and safe. */
368 static void insert_bct PROTO((rtx, rtx, struct loop_info *));
370 /* Auxiliary function that inserts the BCT pattern into the loop. */
371 static void instrument_loop_bct PROTO((rtx, rtx, rtx));
372 #endif /* HAVE_decrement_and_branch_on_count */
374 /* Indirect_jump_in_function is computed once per function. */
375 int indirect_jump_in_function = 0;
376 static int indirect_jump_in_function_p PROTO((rtx));
378 static int compute_luids PROTO((rtx, rtx, int));
380 static int biv_elimination_giv_has_0_offset PROTO((struct induction *,
381 struct induction *, rtx));
383 /* Relative gain of eliminating various kinds of operations. */
386 static int shift_cost;
387 static int mult_cost;
390 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
391 copy the value of the strength reduced giv to its original register. */
392 static int copy_cost;
394 /* Cost of using a register, to normalize the benefits of a giv. */
395 static int reg_address_cost;
401 char *free_point = (char *) oballoc (1);
402 rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
404 add_cost = rtx_cost (gen_rtx_PLUS (word_mode, reg, reg), SET);
407 reg_address_cost = ADDRESS_COST (reg);
409 reg_address_cost = rtx_cost (reg, MEM);
412 /* We multiply by 2 to reconcile the difference in scale between
413 these two ways of computing costs. Otherwise the cost of a copy
414 will be far less than the cost of an add. */
418 /* Free the objects we just allocated. */
421 /* Initialize the obstack used for rtl in product_cheap_p. */
422 gcc_obstack_init (&temp_obstack);
425 /* Compute the mapping from uids to luids.
426 LUIDs are numbers assigned to insns, like uids,
427 except that luids increase monotonically through the code.
428 Start at insn START and stop just before END. Assign LUIDs
429 starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
431 compute_luids (start, end, prev_luid)
438 for (insn = start, i = prev_luid; insn != end; insn = NEXT_INSN (insn))
440 if (INSN_UID (insn) >= max_uid_for_loop)
442 /* Don't assign luids to line-number NOTEs, so that the distance in
443 luids between two insns is not affected by -g. */
444 if (GET_CODE (insn) != NOTE
445 || NOTE_LINE_NUMBER (insn) <= 0)
446 uid_luid[INSN_UID (insn)] = ++i;
448 /* Give a line number note the same luid as preceding insn. */
449 uid_luid[INSN_UID (insn)] = i;
454 /* Entry point of this file. Perform loop optimization
455 on the current function. F is the first insn of the function
456 and DUMPFILE is a stream for output of a trace of actions taken
457 (or 0 if none should be output). */
460 loop_optimize (f, dumpfile, unroll_p, bct_p)
461 /* f is the first instruction of a chain of insns for one function */
469 loop_dump_stream = dumpfile;
471 init_recog_no_volatile ();
473 max_reg_before_loop = max_reg_num ();
475 moved_once = (char *) alloca (max_reg_before_loop);
476 bzero (moved_once, max_reg_before_loop);
480 /* Count the number of loops. */
483 for (insn = f; insn; insn = NEXT_INSN (insn))
485 if (GET_CODE (insn) == NOTE
486 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
490 /* Don't waste time if no loops. */
491 if (max_loop_num == 0)
494 /* Get size to use for tables indexed by uids.
495 Leave some space for labels allocated by find_and_verify_loops. */
496 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
498 uid_luid = (int *) alloca (max_uid_for_loop * sizeof (int));
499 uid_loop_num = (int *) alloca (max_uid_for_loop * sizeof (int));
501 bzero ((char *) uid_luid, max_uid_for_loop * sizeof (int));
502 bzero ((char *) uid_loop_num, max_uid_for_loop * sizeof (int));
504 /* Allocate tables for recording each loop. We set each entry, so they need
506 loop_number_loop_starts = (rtx *) alloca (max_loop_num * sizeof (rtx));
507 loop_number_loop_ends = (rtx *) alloca (max_loop_num * sizeof (rtx));
508 loop_number_loop_cont = (rtx *) alloca (max_loop_num * sizeof (rtx));
509 loop_number_cont_dominator = (rtx *) alloca (max_loop_num * sizeof (rtx));
510 loop_outer_loop = (int *) alloca (max_loop_num * sizeof (int));
511 loop_invalid = (char *) alloca (max_loop_num * sizeof (char));
512 loop_number_exit_labels = (rtx *) alloca (max_loop_num * sizeof (rtx));
513 loop_number_exit_count = (int *) alloca (max_loop_num * sizeof (int));
515 #ifdef HAVE_decrement_and_branch_on_count
516 /* Allocate for BCT optimization */
517 loop_used_count_register = (int *) alloca (max_loop_num * sizeof (int));
518 bzero ((char *) loop_used_count_register, max_loop_num * sizeof (int));
519 #endif /* HAVE_decrement_and_branch_on_count */
521 /* Find and process each loop.
522 First, find them, and record them in order of their beginnings. */
523 find_and_verify_loops (f);
525 /* Now find all register lifetimes. This must be done after
526 find_and_verify_loops, because it might reorder the insns in the
528 reg_scan (f, max_reg_num (), 1);
530 /* This must occur after reg_scan so that registers created by gcse
531 will have entries in the register tables.
533 We could have added a call to reg_scan after gcse_main in toplev.c,
534 but moving this call to init_alias_analysis is more efficient. */
535 init_alias_analysis ();
537 /* See if we went too far. Note that get_max_uid already returns
538 one more that the maximum uid of all insn. */
539 if (get_max_uid () > max_uid_for_loop)
541 /* Now reset it to the actual size we need. See above. */
542 max_uid_for_loop = get_max_uid ();
544 /* find_and_verify_loops has already called compute_luids, but it might
545 have rearranged code afterwards, so we need to recompute the luids now. */
546 max_luid = compute_luids (f, NULL_RTX, 0);
548 /* Don't leave gaps in uid_luid for insns that have been
549 deleted. It is possible that the first or last insn
550 using some register has been deleted by cross-jumping.
551 Make sure that uid_luid for that former insn's uid
552 points to the general area where that insn used to be. */
553 for (i = 0; i < max_uid_for_loop; i++)
555 uid_luid[0] = uid_luid[i];
556 if (uid_luid[0] != 0)
559 for (i = 0; i < max_uid_for_loop; i++)
560 if (uid_luid[i] == 0)
561 uid_luid[i] = uid_luid[i - 1];
563 /* Create a mapping from loops to BLOCK tree nodes. */
564 if (unroll_p && write_symbols != NO_DEBUG)
565 find_loop_tree_blocks ();
567 /* Determine if the function has indirect jump. On some systems
568 this prevents low overhead loop instructions from being used. */
569 indirect_jump_in_function = indirect_jump_in_function_p (f);
571 /* Now scan the loops, last ones first, since this means inner ones are done
572 before outer ones. */
573 for (i = max_loop_num-1; i >= 0; i--)
574 if (! loop_invalid[i] && loop_number_loop_ends[i])
575 scan_loop (loop_number_loop_starts[i], loop_number_loop_ends[i],
576 loop_number_loop_cont[i], unroll_p, bct_p);
578 /* If debugging and unrolling loops, we must replicate the tree nodes
579 corresponding to the blocks inside the loop, so that the original one
580 to one mapping will remain. */
581 if (unroll_p && write_symbols != NO_DEBUG)
582 unroll_block_trees ();
584 end_alias_analysis ();
587 /* Returns the next insn, in execution order, after INSN. START and
588 END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
589 respectively. LOOP_TOP, if non-NULL, is the top of the loop in the
590 insn-stream; it is used with loops that are entered near the
594 next_insn_in_loop (insn, start, end, loop_top)
600 insn = NEXT_INSN (insn);
605 /* Go to the top of the loop, and continue there. */
619 /* Optimize one loop whose start is LOOP_START and end is END.
620 LOOP_START is the NOTE_INSN_LOOP_BEG and END is the matching
622 LOOP_CONT is the NOTE_INSN_LOOP_CONT. */
624 /* ??? Could also move memory writes out of loops if the destination address
625 is invariant, the source is invariant, the memory write is not volatile,
626 and if we can prove that no read inside the loop can read this address
627 before the write occurs. If there is a read of this address after the
628 write, then we can also mark the memory read as invariant. */
631 scan_loop (loop_start, end, loop_cont, unroll_p, bct_p)
632 rtx loop_start, end, loop_cont;
637 /* 1 if we are scanning insns that could be executed zero times. */
639 /* 1 if we are scanning insns that might never be executed
640 due to a subroutine call which might exit before they are reached. */
642 /* For a rotated loop that is entered near the bottom,
643 this is the label at the top. Otherwise it is zero. */
645 /* Jump insn that enters the loop, or 0 if control drops in. */
646 rtx loop_entry_jump = 0;
647 /* Place in the loop where control enters. */
649 /* Number of insns in the loop. */
654 /* The SET from an insn, if it is the only SET in the insn. */
656 /* Chain describing insns movable in current loop. */
657 struct movable *movables = 0;
658 /* Last element in `movables' -- so we can add elements at the end. */
659 struct movable *last_movable = 0;
660 /* Ratio of extra register life span we can justify
661 for saving an instruction. More if loop doesn't call subroutines
662 since in that case saving an insn makes more difference
663 and more registers are available. */
665 /* Nonzero if we are scanning instructions in a sub-loop. */
669 /* Determine whether this loop starts with a jump down to a test at
670 the end. This will occur for a small number of loops with a test
671 that is too complex to duplicate in front of the loop.
673 We search for the first insn or label in the loop, skipping NOTEs.
674 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
675 (because we might have a loop executed only once that contains a
676 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
677 (in case we have a degenerate loop).
679 Note that if we mistakenly think that a loop is entered at the top
680 when, in fact, it is entered at the exit test, the only effect will be
681 slightly poorer optimization. Making the opposite error can generate
682 incorrect code. Since very few loops now start with a jump to the
683 exit test, the code here to detect that case is very conservative. */
685 for (p = NEXT_INSN (loop_start);
687 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
688 && (GET_CODE (p) != NOTE
689 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
690 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
696 /* Set up variables describing this loop. */
697 prescan_loop (loop_start, end);
698 threshold = (loop_has_call ? 1 : 2) * (1 + n_non_fixed_regs);
700 /* If loop has a jump before the first label,
701 the true entry is the target of that jump.
702 Start scan from there.
703 But record in LOOP_TOP the place where the end-test jumps
704 back to so we can scan that after the end of the loop. */
705 if (GET_CODE (p) == JUMP_INSN)
709 /* Loop entry must be unconditional jump (and not a RETURN) */
711 && JUMP_LABEL (p) != 0
712 /* Check to see whether the jump actually
713 jumps out of the loop (meaning it's no loop).
714 This case can happen for things like
715 do {..} while (0). If this label was generated previously
716 by loop, we can't tell anything about it and have to reject
718 && INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, end))
720 loop_top = next_label (scan_start);
721 scan_start = JUMP_LABEL (p);
725 /* If SCAN_START was an insn created by loop, we don't know its luid
726 as required by loop_reg_used_before_p. So skip such loops. (This
727 test may never be true, but it's best to play it safe.)
729 Also, skip loops where we do not start scanning at a label. This
730 test also rejects loops starting with a JUMP_INSN that failed the
733 if (INSN_UID (scan_start) >= max_uid_for_loop
734 || GET_CODE (scan_start) != CODE_LABEL)
736 if (loop_dump_stream)
737 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
738 INSN_UID (loop_start), INSN_UID (end));
742 /* Count number of times each reg is set during this loop.
743 Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out
744 the setting of register I. Set VARRAY_RTX (reg_single_usage, I). */
746 /* Allocate extra space for REGS that might be created by
747 load_mems. We allocate a little extra slop as well, in the hopes
748 that even after the moving of movables creates some new registers
749 we won't have to reallocate these arrays. However, we do grow
750 the arrays, if necessary, in load_mems_recount_loop_regs_set. */
751 nregs = max_reg_num () + loop_mems_idx + 16;
752 VARRAY_INT_INIT (set_in_loop, nregs, "set_in_loop");
753 VARRAY_INT_INIT (n_times_set, nregs, "n_times_set");
754 VARRAY_CHAR_INIT (may_not_optimize, nregs, "may_not_optimize");
755 VARRAY_RTX_INIT (reg_single_usage, nregs, "reg_single_usage");
757 count_loop_regs_set (loop_top ? loop_top : loop_start, end,
758 may_not_optimize, reg_single_usage, &insn_count, nregs);
760 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
762 VARRAY_CHAR (may_not_optimize, i) = 1;
763 VARRAY_INT (set_in_loop, i) = 1;
766 #ifdef AVOID_CCMODE_COPIES
767 /* Don't try to move insns which set CC registers if we should not
768 create CCmode register copies. */
769 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
770 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
771 VARRAY_CHAR (may_not_optimize, i) = 1;
774 bcopy ((char *) &set_in_loop->data,
775 (char *) &n_times_set->data, nregs * sizeof (int));
777 if (loop_dump_stream)
779 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
780 INSN_UID (loop_start), INSN_UID (end), insn_count);
782 fprintf (loop_dump_stream, "Continue at insn %d.\n",
783 INSN_UID (loop_continue));
786 /* Scan through the loop finding insns that are safe to move.
787 Set set_in_loop negative for the reg being set, so that
788 this reg will be considered invariant for subsequent insns.
789 We consider whether subsequent insns use the reg
790 in deciding whether it is worth actually moving.
792 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
793 and therefore it is possible that the insns we are scanning
794 would never be executed. At such times, we must make sure
795 that it is safe to execute the insn once instead of zero times.
796 When MAYBE_NEVER is 0, all insns will be executed at least once
797 so that is not a problem. */
799 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
801 p = next_insn_in_loop (p, scan_start, end, loop_top))
803 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
804 && find_reg_note (p, REG_LIBCALL, NULL_RTX))
806 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
807 && find_reg_note (p, REG_RETVAL, NULL_RTX))
810 if (GET_CODE (p) == INSN
811 && (set = single_set (p))
812 && GET_CODE (SET_DEST (set)) == REG
813 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
818 rtx src = SET_SRC (set);
819 rtx dependencies = 0;
821 /* Figure out what to use as a source of this insn. If a REG_EQUIV
822 note is given or if a REG_EQUAL note with a constant operand is
823 specified, use it as the source and mark that we should move
824 this insn by calling emit_move_insn rather that duplicating the
827 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
829 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
831 src = XEXP (temp, 0), move_insn = 1;
834 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
835 if (temp && CONSTANT_P (XEXP (temp, 0)))
836 src = XEXP (temp, 0), move_insn = 1;
837 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
839 src = XEXP (temp, 0);
840 /* A libcall block can use regs that don't appear in
841 the equivalent expression. To move the libcall,
842 we must move those regs too. */
843 dependencies = libcall_other_reg (p, src);
847 /* Don't try to optimize a register that was made
848 by loop-optimization for an inner loop.
849 We don't know its life-span, so we can't compute the benefit. */
850 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
852 else if (/* The register is used in basic blocks other
853 than the one where it is set (meaning that
854 something after this point in the loop might
855 depend on its value before the set). */
856 ! reg_in_basic_block_p (p, SET_DEST (set))
857 /* And the set is not guaranteed to be executed one
858 the loop starts, or the value before the set is
859 needed before the set occurs...
861 ??? Note we have quadratic behaviour here, mitigated
862 by the fact that the previous test will often fail for
863 large loops. Rather than re-scanning the entire loop
864 each time for register usage, we should build tables
865 of the register usage and use them here instead. */
867 || loop_reg_used_before_p (set, p, loop_start,
869 /* It is unsafe to move the set.
871 This code used to consider it OK to move a set of a variable
872 which was not created by the user and not used in an exit test.
873 That behavior is incorrect and was removed. */
875 else if ((tem = invariant_p (src))
876 && (dependencies == 0
877 || (tem2 = invariant_p (dependencies)) != 0)
878 && (VARRAY_INT (set_in_loop,
879 REGNO (SET_DEST (set))) == 1
881 = consec_sets_invariant_p
883 VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))),
885 /* If the insn can cause a trap (such as divide by zero),
886 can't move it unless it's guaranteed to be executed
887 once loop is entered. Even a function call might
888 prevent the trap insn from being reached
889 (since it might exit!) */
890 && ! ((maybe_never || call_passed)
891 && may_trap_p (src)))
893 register struct movable *m;
894 register int regno = REGNO (SET_DEST (set));
896 /* A potential lossage is where we have a case where two insns
897 can be combined as long as they are both in the loop, but
898 we move one of them outside the loop. For large loops,
899 this can lose. The most common case of this is the address
900 of a function being called.
902 Therefore, if this register is marked as being used exactly
903 once if we are in a loop with calls (a "large loop"), see if
904 we can replace the usage of this register with the source
905 of this SET. If we can, delete this insn.
907 Don't do this if P has a REG_RETVAL note or if we have
908 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
911 && VARRAY_RTX (reg_single_usage, regno) != 0
912 && VARRAY_RTX (reg_single_usage, regno) != const0_rtx
913 && REGNO_FIRST_UID (regno) == INSN_UID (p)
914 && (REGNO_LAST_UID (regno)
915 == INSN_UID (VARRAY_RTX (reg_single_usage, regno)))
916 && VARRAY_INT (set_in_loop, regno) == 1
917 && ! side_effects_p (SET_SRC (set))
918 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
919 && (! SMALL_REGISTER_CLASSES
920 || (! (GET_CODE (SET_SRC (set)) == REG
921 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)))
922 /* This test is not redundant; SET_SRC (set) might be
923 a call-clobbered register and the life of REGNO
924 might span a call. */
925 && ! modified_between_p (SET_SRC (set), p,
927 (reg_single_usage, regno))
928 && no_labels_between_p (p, VARRAY_RTX (reg_single_usage, regno))
929 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
931 (reg_single_usage, regno)))
933 /* Replace any usage in a REG_EQUAL note. Must copy the
934 new source, so that we don't get rtx sharing between the
935 SET_SOURCE and REG_NOTES of insn p. */
936 REG_NOTES (VARRAY_RTX (reg_single_usage, regno))
937 = replace_rtx (REG_NOTES (VARRAY_RTX
938 (reg_single_usage, regno)),
939 SET_DEST (set), copy_rtx (SET_SRC (set)));
942 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
943 NOTE_SOURCE_FILE (p) = 0;
944 VARRAY_INT (set_in_loop, regno) = 0;
948 m = (struct movable *) alloca (sizeof (struct movable));
952 m->dependencies = dependencies;
953 m->set_dest = SET_DEST (set);
955 m->consec = VARRAY_INT (set_in_loop,
956 REGNO (SET_DEST (set))) - 1;
960 m->move_insn = move_insn;
961 m->move_insn_first = 0;
962 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
963 m->savemode = VOIDmode;
965 /* Set M->cond if either invariant_p or consec_sets_invariant_p
966 returned 2 (only conditionally invariant). */
967 m->cond = ((tem | tem1 | tem2) > 1);
968 m->global = (uid_luid[REGNO_LAST_UID (regno)] > INSN_LUID (end)
969 || uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (loop_start));
971 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
972 - uid_luid[REGNO_FIRST_UID (regno)]);
973 m->savings = VARRAY_INT (n_times_set, regno);
974 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
975 m->savings += libcall_benefit (p);
976 VARRAY_INT (set_in_loop, regno) = move_insn ? -2 : -1;
977 /* Add M to the end of the chain MOVABLES. */
981 last_movable->next = m;
986 /* It is possible for the first instruction to have a
987 REG_EQUAL note but a non-invariant SET_SRC, so we must
988 remember the status of the first instruction in case
989 the last instruction doesn't have a REG_EQUAL note. */
990 m->move_insn_first = m->move_insn;
992 /* Skip this insn, not checking REG_LIBCALL notes. */
993 p = next_nonnote_insn (p);
994 /* Skip the consecutive insns, if there are any. */
995 p = skip_consec_insns (p, m->consec);
996 /* Back up to the last insn of the consecutive group. */
997 p = prev_nonnote_insn (p);
999 /* We must now reset m->move_insn, m->is_equiv, and possibly
1000 m->set_src to correspond to the effects of all the
1002 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
1004 m->set_src = XEXP (temp, 0), m->move_insn = 1;
1007 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
1008 if (temp && CONSTANT_P (XEXP (temp, 0)))
1009 m->set_src = XEXP (temp, 0), m->move_insn = 1;
1014 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
1017 /* If this register is always set within a STRICT_LOW_PART
1018 or set to zero, then its high bytes are constant.
1019 So clear them outside the loop and within the loop
1020 just load the low bytes.
1021 We must check that the machine has an instruction to do so.
1022 Also, if the value loaded into the register
1023 depends on the same register, this cannot be done. */
1024 else if (SET_SRC (set) == const0_rtx
1025 && GET_CODE (NEXT_INSN (p)) == INSN
1026 && (set1 = single_set (NEXT_INSN (p)))
1027 && GET_CODE (set1) == SET
1028 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
1029 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
1030 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
1032 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
1034 register int regno = REGNO (SET_DEST (set));
1035 if (VARRAY_INT (set_in_loop, regno) == 2)
1037 register struct movable *m;
1038 m = (struct movable *) alloca (sizeof (struct movable));
1041 m->set_dest = SET_DEST (set);
1042 m->dependencies = 0;
1048 m->move_insn_first = 0;
1050 /* If the insn may not be executed on some cycles,
1051 we can't clear the whole reg; clear just high part.
1052 Not even if the reg is used only within this loop.
1059 Clearing x before the inner loop could clobber a value
1060 being saved from the last time around the outer loop.
1061 However, if the reg is not used outside this loop
1062 and all uses of the register are in the same
1063 basic block as the store, there is no problem.
1065 If this insn was made by loop, we don't know its
1066 INSN_LUID and hence must make a conservative
1068 m->global = (INSN_UID (p) >= max_uid_for_loop
1069 || (uid_luid[REGNO_LAST_UID (regno)]
1071 || (uid_luid[REGNO_FIRST_UID (regno)]
1073 || (labels_in_range_p
1074 (p, uid_luid[REGNO_FIRST_UID (regno)])));
1075 if (maybe_never && m->global)
1076 m->savemode = GET_MODE (SET_SRC (set1));
1078 m->savemode = VOIDmode;
1082 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
1083 - uid_luid[REGNO_FIRST_UID (regno)]);
1085 VARRAY_INT (set_in_loop, regno) = -1;
1086 /* Add M to the end of the chain MOVABLES. */
1090 last_movable->next = m;
1095 /* Past a call insn, we get to insns which might not be executed
1096 because the call might exit. This matters for insns that trap.
1097 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
1098 so they don't count. */
1099 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
1101 /* Past a label or a jump, we get to insns for which we
1102 can't count on whether or how many times they will be
1103 executed during each iteration. Therefore, we can
1104 only move out sets of trivial variables
1105 (those not used after the loop). */
1106 /* Similar code appears twice in strength_reduce. */
1107 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
1108 /* If we enter the loop in the middle, and scan around to the
1109 beginning, don't set maybe_never for that. This must be an
1110 unconditional jump, otherwise the code at the top of the
1111 loop might never be executed. Unconditional jumps are
1112 followed a by barrier then loop end. */
1113 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
1114 && NEXT_INSN (NEXT_INSN (p)) == end
1115 && simplejump_p (p)))
1117 else if (GET_CODE (p) == NOTE)
1119 /* At the virtual top of a converted loop, insns are again known to
1120 be executed: logically, the loop begins here even though the exit
1121 code has been duplicated. */
1122 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
1123 maybe_never = call_passed = 0;
1124 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
1126 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
1131 /* If one movable subsumes another, ignore that other. */
1133 ignore_some_movables (movables);
1135 /* For each movable insn, see if the reg that it loads
1136 leads when it dies right into another conditionally movable insn.
1137 If so, record that the second insn "forces" the first one,
1138 since the second can be moved only if the first is. */
1140 force_movables (movables);
1142 /* See if there are multiple movable insns that load the same value.
1143 If there are, make all but the first point at the first one
1144 through the `match' field, and add the priorities of them
1145 all together as the priority of the first. */
1147 combine_movables (movables, nregs);
1149 /* Now consider each movable insn to decide whether it is worth moving.
1150 Store 0 in set_in_loop for each reg that is moved.
1152 Generally this increases code size, so do not move moveables when
1153 optimizing for code size. */
1155 if (! optimize_size)
1156 move_movables (movables, threshold,
1157 insn_count, loop_start, end, nregs);
1159 /* Now candidates that still are negative are those not moved.
1160 Change set_in_loop to indicate that those are not actually invariant. */
1161 for (i = 0; i < nregs; i++)
1162 if (VARRAY_INT (set_in_loop, i) < 0)
1163 VARRAY_INT (set_in_loop, i) = VARRAY_INT (n_times_set, i);
1165 /* Now that we've moved some things out of the loop, we might be able to
1166 hoist even more memory references. */
1167 load_mems_and_recount_loop_regs_set (scan_start, end, loop_top,
1168 loop_start, &insn_count);
1170 if (flag_strength_reduce)
1172 the_movables = movables;
1173 strength_reduce (scan_start, end, loop_top,
1174 insn_count, loop_start, end, loop_cont, unroll_p, bct_p);
1177 VARRAY_FREE (reg_single_usage);
1178 VARRAY_FREE (set_in_loop);
1179 VARRAY_FREE (n_times_set);
1180 VARRAY_FREE (may_not_optimize);
1183 /* Add elements to *OUTPUT to record all the pseudo-regs
1184 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1187 record_excess_regs (in_this, not_in_this, output)
1188 rtx in_this, not_in_this;
1195 code = GET_CODE (in_this);
1209 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1210 && ! reg_mentioned_p (in_this, not_in_this))
1211 *output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output);
1218 fmt = GET_RTX_FORMAT (code);
1219 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1226 for (j = 0; j < XVECLEN (in_this, i); j++)
1227 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1231 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1237 /* Check what regs are referred to in the libcall block ending with INSN,
1238 aside from those mentioned in the equivalent value.
1239 If there are none, return 0.
1240 If there are one or more, return an EXPR_LIST containing all of them. */
1243 libcall_other_reg (insn, equiv)
1246 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1247 rtx p = XEXP (note, 0);
1250 /* First, find all the regs used in the libcall block
1251 that are not mentioned as inputs to the result. */
1255 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1256 || GET_CODE (p) == CALL_INSN)
1257 record_excess_regs (PATTERN (p), equiv, &output);
1264 /* Return 1 if all uses of REG
1265 are between INSN and the end of the basic block. */
1268 reg_in_basic_block_p (insn, reg)
1271 int regno = REGNO (reg);
1274 if (REGNO_FIRST_UID (regno) != INSN_UID (insn))
1277 /* Search this basic block for the already recorded last use of the reg. */
1278 for (p = insn; p; p = NEXT_INSN (p))
1280 switch (GET_CODE (p))
1287 /* Ordinary insn: if this is the last use, we win. */
1288 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1293 /* Jump insn: if this is the last use, we win. */
1294 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1296 /* Otherwise, it's the end of the basic block, so we lose. */
1301 /* It's the end of the basic block, so we lose. */
1309 /* The "last use" doesn't follow the "first use"?? */
1313 /* Compute the benefit of eliminating the insns in the block whose
1314 last insn is LAST. This may be a group of insns used to compute a
1315 value directly or can contain a library call. */
1318 libcall_benefit (last)
1324 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1325 insn != last; insn = NEXT_INSN (insn))
1327 if (GET_CODE (insn) == CALL_INSN)
1328 benefit += 10; /* Assume at least this many insns in a library
1330 else if (GET_CODE (insn) == INSN
1331 && GET_CODE (PATTERN (insn)) != USE
1332 && GET_CODE (PATTERN (insn)) != CLOBBER)
1339 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1342 skip_consec_insns (insn, count)
1346 for (; count > 0; count--)
1350 /* If first insn of libcall sequence, skip to end. */
1351 /* Do this at start of loop, since INSN is guaranteed to
1353 if (GET_CODE (insn) != NOTE
1354 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1355 insn = XEXP (temp, 0);
1357 do insn = NEXT_INSN (insn);
1358 while (GET_CODE (insn) == NOTE);
1364 /* Ignore any movable whose insn falls within a libcall
1365 which is part of another movable.
1366 We make use of the fact that the movable for the libcall value
1367 was made later and so appears later on the chain. */
1370 ignore_some_movables (movables)
1371 struct movable *movables;
1373 register struct movable *m, *m1;
1375 for (m = movables; m; m = m->next)
1377 /* Is this a movable for the value of a libcall? */
1378 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1382 /* Check for earlier movables inside that range,
1383 and mark them invalid. We cannot use LUIDs here because
1384 insns created by loop.c for prior loops don't have LUIDs.
1385 Rather than reject all such insns from movables, we just
1386 explicitly check each insn in the libcall (since invariant
1387 libcalls aren't that common). */
1388 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1389 for (m1 = movables; m1 != m; m1 = m1->next)
1390 if (m1->insn == insn)
1396 /* For each movable insn, see if the reg that it loads
1397 leads when it dies right into another conditionally movable insn.
1398 If so, record that the second insn "forces" the first one,
1399 since the second can be moved only if the first is. */
1402 force_movables (movables)
1403 struct movable *movables;
1405 register struct movable *m, *m1;
1406 for (m1 = movables; m1; m1 = m1->next)
1407 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1408 if (!m1->partial && !m1->done)
1410 int regno = m1->regno;
1411 for (m = m1->next; m; m = m->next)
1412 /* ??? Could this be a bug? What if CSE caused the
1413 register of M1 to be used after this insn?
1414 Since CSE does not update regno_last_uid,
1415 this insn M->insn might not be where it dies.
1416 But very likely this doesn't matter; what matters is
1417 that M's reg is computed from M1's reg. */
1418 if (INSN_UID (m->insn) == REGNO_LAST_UID (regno)
1421 if (m != 0 && m->set_src == m1->set_dest
1422 /* If m->consec, m->set_src isn't valid. */
1426 /* Increase the priority of the moving the first insn
1427 since it permits the second to be moved as well. */
1431 m1->lifetime += m->lifetime;
1432 m1->savings += m->savings;
1437 /* Find invariant expressions that are equal and can be combined into
1441 combine_movables (movables, nregs)
1442 struct movable *movables;
1445 register struct movable *m;
1446 char *matched_regs = (char *) alloca (nregs);
1447 enum machine_mode mode;
1449 /* Regs that are set more than once are not allowed to match
1450 or be matched. I'm no longer sure why not. */
1451 /* Perhaps testing m->consec_sets would be more appropriate here? */
1453 for (m = movables; m; m = m->next)
1454 if (m->match == 0 && VARRAY_INT (n_times_set, m->regno) == 1 && !m->partial)
1456 register struct movable *m1;
1457 int regno = m->regno;
1459 bzero (matched_regs, nregs);
1460 matched_regs[regno] = 1;
1462 /* We want later insns to match the first one. Don't make the first
1463 one match any later ones. So start this loop at m->next. */
1464 for (m1 = m->next; m1; m1 = m1->next)
1465 if (m != m1 && m1->match == 0 && VARRAY_INT (n_times_set, m1->regno) == 1
1466 /* A reg used outside the loop mustn't be eliminated. */
1468 /* A reg used for zero-extending mustn't be eliminated. */
1470 && (matched_regs[m1->regno]
1473 /* Can combine regs with different modes loaded from the
1474 same constant only if the modes are the same or
1475 if both are integer modes with M wider or the same
1476 width as M1. The check for integer is redundant, but
1477 safe, since the only case of differing destination
1478 modes with equal sources is when both sources are
1479 VOIDmode, i.e., CONST_INT. */
1480 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1481 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1482 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1483 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1484 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1485 /* See if the source of M1 says it matches M. */
1486 && ((GET_CODE (m1->set_src) == REG
1487 && matched_regs[REGNO (m1->set_src)])
1488 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1490 && ((m->dependencies == m1->dependencies)
1491 || rtx_equal_p (m->dependencies, m1->dependencies)))
1493 m->lifetime += m1->lifetime;
1494 m->savings += m1->savings;
1497 matched_regs[m1->regno] = 1;
1501 /* Now combine the regs used for zero-extension.
1502 This can be done for those not marked `global'
1503 provided their lives don't overlap. */
1505 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1506 mode = GET_MODE_WIDER_MODE (mode))
1508 register struct movable *m0 = 0;
1510 /* Combine all the registers for extension from mode MODE.
1511 Don't combine any that are used outside this loop. */
1512 for (m = movables; m; m = m->next)
1513 if (m->partial && ! m->global
1514 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1516 register struct movable *m1;
1517 int first = uid_luid[REGNO_FIRST_UID (m->regno)];
1518 int last = uid_luid[REGNO_LAST_UID (m->regno)];
1522 /* First one: don't check for overlap, just record it. */
1527 /* Make sure they extend to the same mode.
1528 (Almost always true.) */
1529 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1532 /* We already have one: check for overlap with those
1533 already combined together. */
1534 for (m1 = movables; m1 != m; m1 = m1->next)
1535 if (m1 == m0 || (m1->partial && m1->match == m0))
1536 if (! (uid_luid[REGNO_FIRST_UID (m1->regno)] > last
1537 || uid_luid[REGNO_LAST_UID (m1->regno)] < first))
1540 /* No overlap: we can combine this with the others. */
1541 m0->lifetime += m->lifetime;
1542 m0->savings += m->savings;
1551 /* Return 1 if regs X and Y will become the same if moved. */
1554 regs_match_p (x, y, movables)
1556 struct movable *movables;
1560 struct movable *mx, *my;
1562 for (mx = movables; mx; mx = mx->next)
1563 if (mx->regno == xn)
1566 for (my = movables; my; my = my->next)
1567 if (my->regno == yn)
1571 && ((mx->match == my->match && mx->match != 0)
1573 || mx == my->match));
1576 /* Return 1 if X and Y are identical-looking rtx's.
1577 This is the Lisp function EQUAL for rtx arguments.
1579 If two registers are matching movables or a movable register and an
1580 equivalent constant, consider them equal. */
1583 rtx_equal_for_loop_p (x, y, movables)
1585 struct movable *movables;
1589 register struct movable *m;
1590 register enum rtx_code code;
1591 register const char *fmt;
1595 if (x == 0 || y == 0)
1598 code = GET_CODE (x);
1600 /* If we have a register and a constant, they may sometimes be
1602 if (GET_CODE (x) == REG && VARRAY_INT (set_in_loop, REGNO (x)) == -2
1605 for (m = movables; m; m = m->next)
1606 if (m->move_insn && m->regno == REGNO (x)
1607 && rtx_equal_p (m->set_src, y))
1610 else if (GET_CODE (y) == REG && VARRAY_INT (set_in_loop, REGNO (y)) == -2
1613 for (m = movables; m; m = m->next)
1614 if (m->move_insn && m->regno == REGNO (y)
1615 && rtx_equal_p (m->set_src, x))
1619 /* Otherwise, rtx's of different codes cannot be equal. */
1620 if (code != GET_CODE (y))
1623 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1624 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1626 if (GET_MODE (x) != GET_MODE (y))
1629 /* These three types of rtx's can be compared nonrecursively. */
1631 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1633 if (code == LABEL_REF)
1634 return XEXP (x, 0) == XEXP (y, 0);
1635 if (code == SYMBOL_REF)
1636 return XSTR (x, 0) == XSTR (y, 0);
1638 /* Compare the elements. If any pair of corresponding elements
1639 fail to match, return 0 for the whole things. */
1641 fmt = GET_RTX_FORMAT (code);
1642 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1647 if (XWINT (x, i) != XWINT (y, i))
1652 if (XINT (x, i) != XINT (y, i))
1657 /* Two vectors must have the same length. */
1658 if (XVECLEN (x, i) != XVECLEN (y, i))
1661 /* And the corresponding elements must match. */
1662 for (j = 0; j < XVECLEN (x, i); j++)
1663 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1668 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1673 if (strcmp (XSTR (x, i), XSTR (y, i)))
1678 /* These are just backpointers, so they don't matter. */
1684 /* It is believed that rtx's at this level will never
1685 contain anything but integers and other rtx's,
1686 except for within LABEL_REFs and SYMBOL_REFs. */
1694 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1695 insns in INSNS which use thet reference. */
1698 add_label_notes (x, insns)
1702 enum rtx_code code = GET_CODE (x);
1707 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1709 /* This code used to ignore labels that referred to dispatch tables to
1710 avoid flow generating (slighly) worse code.
1712 We no longer ignore such label references (see LABEL_REF handling in
1713 mark_jump_label for additional information). */
1714 for (insn = insns; insn; insn = NEXT_INSN (insn))
1715 if (reg_mentioned_p (XEXP (x, 0), insn))
1716 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
1720 fmt = GET_RTX_FORMAT (code);
1721 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1724 add_label_notes (XEXP (x, i), insns);
1725 else if (fmt[i] == 'E')
1726 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1727 add_label_notes (XVECEXP (x, i, j), insns);
1731 /* Scan MOVABLES, and move the insns that deserve to be moved.
1732 If two matching movables are combined, replace one reg with the
1733 other throughout. */
1736 move_movables (movables, threshold, insn_count, loop_start, end, nregs)
1737 struct movable *movables;
1745 register struct movable *m;
1747 /* Map of pseudo-register replacements to handle combining
1748 when we move several insns that load the same value
1749 into different pseudo-registers. */
1750 rtx *reg_map = (rtx *) alloca (nregs * sizeof (rtx));
1751 char *already_moved = (char *) alloca (nregs);
1753 bzero (already_moved, nregs);
1754 bzero ((char *) reg_map, nregs * sizeof (rtx));
1758 for (m = movables; m; m = m->next)
1760 /* Describe this movable insn. */
1762 if (loop_dump_stream)
1764 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1765 INSN_UID (m->insn), m->regno, m->lifetime);
1767 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1769 fprintf (loop_dump_stream, "cond ");
1771 fprintf (loop_dump_stream, "force ");
1773 fprintf (loop_dump_stream, "global ");
1775 fprintf (loop_dump_stream, "done ");
1777 fprintf (loop_dump_stream, "move-insn ");
1779 fprintf (loop_dump_stream, "matches %d ",
1780 INSN_UID (m->match->insn));
1782 fprintf (loop_dump_stream, "forces %d ",
1783 INSN_UID (m->forces->insn));
1786 /* Count movables. Value used in heuristics in strength_reduce. */
1789 /* Ignore the insn if it's already done (it matched something else).
1790 Otherwise, see if it is now safe to move. */
1794 || (1 == invariant_p (m->set_src)
1795 && (m->dependencies == 0
1796 || 1 == invariant_p (m->dependencies))
1798 || 1 == consec_sets_invariant_p (m->set_dest,
1801 && (! m->forces || m->forces->done))
1805 int savings = m->savings;
1807 /* We have an insn that is safe to move.
1808 Compute its desirability. */
1813 if (loop_dump_stream)
1814 fprintf (loop_dump_stream, "savings %d ", savings);
1816 if (moved_once[regno] && loop_dump_stream)
1817 fprintf (loop_dump_stream, "halved since already moved ");
1819 /* An insn MUST be moved if we already moved something else
1820 which is safe only if this one is moved too: that is,
1821 if already_moved[REGNO] is nonzero. */
1823 /* An insn is desirable to move if the new lifetime of the
1824 register is no more than THRESHOLD times the old lifetime.
1825 If it's not desirable, it means the loop is so big
1826 that moving won't speed things up much,
1827 and it is liable to make register usage worse. */
1829 /* It is also desirable to move if it can be moved at no
1830 extra cost because something else was already moved. */
1832 if (already_moved[regno]
1833 || flag_move_all_movables
1834 || (threshold * savings * m->lifetime) >=
1835 (moved_once[regno] ? insn_count * 2 : insn_count)
1836 || (m->forces && m->forces->done
1837 && VARRAY_INT (n_times_set, m->forces->regno) == 1))
1840 register struct movable *m1;
1841 rtx first = NULL_RTX;
1843 /* Now move the insns that set the reg. */
1845 if (m->partial && m->match)
1849 /* Find the end of this chain of matching regs.
1850 Thus, we load each reg in the chain from that one reg.
1851 And that reg is loaded with 0 directly,
1852 since it has ->match == 0. */
1853 for (m1 = m; m1->match; m1 = m1->match);
1854 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1855 SET_DEST (PATTERN (m1->insn)));
1856 i1 = emit_insn_before (newpat, loop_start);
1858 /* Mark the moved, invariant reg as being allowed to
1859 share a hard reg with the other matching invariant. */
1860 REG_NOTES (i1) = REG_NOTES (m->insn);
1861 r1 = SET_DEST (PATTERN (m->insn));
1862 r2 = SET_DEST (PATTERN (m1->insn));
1864 = gen_rtx_EXPR_LIST (VOIDmode, r1,
1865 gen_rtx_EXPR_LIST (VOIDmode, r2,
1867 delete_insn (m->insn);
1872 if (loop_dump_stream)
1873 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1875 /* If we are to re-generate the item being moved with a
1876 new move insn, first delete what we have and then emit
1877 the move insn before the loop. */
1878 else if (m->move_insn)
1882 for (count = m->consec; count >= 0; count--)
1884 /* If this is the first insn of a library call sequence,
1886 if (GET_CODE (p) != NOTE
1887 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1890 /* If this is the last insn of a libcall sequence, then
1891 delete every insn in the sequence except the last.
1892 The last insn is handled in the normal manner. */
1893 if (GET_CODE (p) != NOTE
1894 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1896 temp = XEXP (temp, 0);
1898 temp = delete_insn (temp);
1902 p = delete_insn (p);
1904 /* simplify_giv_expr expects that it can walk the insns
1905 at m->insn forwards and see this old sequence we are
1906 tossing here. delete_insn does preserve the next
1907 pointers, but when we skip over a NOTE we must fix
1908 it up. Otherwise that code walks into the non-deleted
1910 while (p && GET_CODE (p) == NOTE)
1911 p = NEXT_INSN (temp) = NEXT_INSN (p);
1915 emit_move_insn (m->set_dest, m->set_src);
1916 temp = get_insns ();
1919 add_label_notes (m->set_src, temp);
1921 i1 = emit_insns_before (temp, loop_start);
1922 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1924 = gen_rtx_EXPR_LIST (m->is_equiv ? REG_EQUIV : REG_EQUAL,
1925 m->set_src, REG_NOTES (i1));
1927 if (loop_dump_stream)
1928 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1930 /* The more regs we move, the less we like moving them. */
1935 for (count = m->consec; count >= 0; count--)
1939 /* If first insn of libcall sequence, skip to end. */
1940 /* Do this at start of loop, since p is guaranteed to
1942 if (GET_CODE (p) != NOTE
1943 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1946 /* If last insn of libcall sequence, move all
1947 insns except the last before the loop. The last
1948 insn is handled in the normal manner. */
1949 if (GET_CODE (p) != NOTE
1950 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1954 rtx fn_address_insn = 0;
1957 for (temp = XEXP (temp, 0); temp != p;
1958 temp = NEXT_INSN (temp))
1964 if (GET_CODE (temp) == NOTE)
1967 body = PATTERN (temp);
1969 /* Find the next insn after TEMP,
1970 not counting USE or NOTE insns. */
1971 for (next = NEXT_INSN (temp); next != p;
1972 next = NEXT_INSN (next))
1973 if (! (GET_CODE (next) == INSN
1974 && GET_CODE (PATTERN (next)) == USE)
1975 && GET_CODE (next) != NOTE)
1978 /* If that is the call, this may be the insn
1979 that loads the function address.
1981 Extract the function address from the insn
1982 that loads it into a register.
1983 If this insn was cse'd, we get incorrect code.
1985 So emit a new move insn that copies the
1986 function address into the register that the
1987 call insn will use. flow.c will delete any
1988 redundant stores that we have created. */
1989 if (GET_CODE (next) == CALL_INSN
1990 && GET_CODE (body) == SET
1991 && GET_CODE (SET_DEST (body)) == REG
1992 && (n = find_reg_note (temp, REG_EQUAL,
1995 fn_reg = SET_SRC (body);
1996 if (GET_CODE (fn_reg) != REG)
1997 fn_reg = SET_DEST (body);
1998 fn_address = XEXP (n, 0);
1999 fn_address_insn = temp;
2001 /* We have the call insn.
2002 If it uses the register we suspect it might,
2003 load it with the correct address directly. */
2004 if (GET_CODE (temp) == CALL_INSN
2006 && reg_referenced_p (fn_reg, body))
2007 emit_insn_after (gen_move_insn (fn_reg,
2011 if (GET_CODE (temp) == CALL_INSN)
2013 i1 = emit_call_insn_before (body, loop_start);
2014 /* Because the USAGE information potentially
2015 contains objects other than hard registers
2016 we need to copy it. */
2017 if (CALL_INSN_FUNCTION_USAGE (temp))
2018 CALL_INSN_FUNCTION_USAGE (i1)
2019 = copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
2022 i1 = emit_insn_before (body, loop_start);
2025 if (temp == fn_address_insn)
2026 fn_address_insn = i1;
2027 REG_NOTES (i1) = REG_NOTES (temp);
2033 if (m->savemode != VOIDmode)
2035 /* P sets REG to zero; but we should clear only
2036 the bits that are not covered by the mode
2038 rtx reg = m->set_dest;
2044 (GET_MODE (reg), and_optab, reg,
2045 GEN_INT ((((HOST_WIDE_INT) 1
2046 << GET_MODE_BITSIZE (m->savemode)))
2048 reg, 1, OPTAB_LIB_WIDEN);
2052 emit_move_insn (reg, tem);
2053 sequence = gen_sequence ();
2055 i1 = emit_insn_before (sequence, loop_start);
2057 else if (GET_CODE (p) == CALL_INSN)
2059 i1 = emit_call_insn_before (PATTERN (p), loop_start);
2060 /* Because the USAGE information potentially
2061 contains objects other than hard registers
2062 we need to copy it. */
2063 if (CALL_INSN_FUNCTION_USAGE (p))
2064 CALL_INSN_FUNCTION_USAGE (i1)
2065 = copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
2067 else if (count == m->consec && m->move_insn_first)
2069 /* The SET_SRC might not be invariant, so we must
2070 use the REG_EQUAL note. */
2072 emit_move_insn (m->set_dest, m->set_src);
2073 temp = get_insns ();
2076 add_label_notes (m->set_src, temp);
2078 i1 = emit_insns_before (temp, loop_start);
2079 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
2081 = gen_rtx_EXPR_LIST ((m->is_equiv ? REG_EQUIV
2083 m->set_src, REG_NOTES (i1));
2086 i1 = emit_insn_before (PATTERN (p), loop_start);
2088 if (REG_NOTES (i1) == 0)
2090 REG_NOTES (i1) = REG_NOTES (p);
2092 /* If there is a REG_EQUAL note present whose value
2093 is not loop invariant, then delete it, since it
2094 may cause problems with later optimization passes.
2095 It is possible for cse to create such notes
2096 like this as a result of record_jump_cond. */
2098 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
2099 && ! invariant_p (XEXP (temp, 0)))
2100 remove_note (i1, temp);
2106 if (loop_dump_stream)
2107 fprintf (loop_dump_stream, " moved to %d",
2110 /* If library call, now fix the REG_NOTES that contain
2111 insn pointers, namely REG_LIBCALL on FIRST
2112 and REG_RETVAL on I1. */
2113 if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX)))
2115 XEXP (temp, 0) = first;
2116 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
2117 XEXP (temp, 0) = i1;
2124 /* simplify_giv_expr expects that it can walk the insns
2125 at m->insn forwards and see this old sequence we are
2126 tossing here. delete_insn does preserve the next
2127 pointers, but when we skip over a NOTE we must fix
2128 it up. Otherwise that code walks into the non-deleted
2130 while (p && GET_CODE (p) == NOTE)
2131 p = NEXT_INSN (temp) = NEXT_INSN (p);
2134 /* The more regs we move, the less we like moving them. */
2138 /* Any other movable that loads the same register
2140 already_moved[regno] = 1;
2142 /* This reg has been moved out of one loop. */
2143 moved_once[regno] = 1;
2145 /* The reg set here is now invariant. */
2147 VARRAY_INT (set_in_loop, regno) = 0;
2151 /* Change the length-of-life info for the register
2152 to say it lives at least the full length of this loop.
2153 This will help guide optimizations in outer loops. */
2155 if (uid_luid[REGNO_FIRST_UID (regno)] > INSN_LUID (loop_start))
2156 /* This is the old insn before all the moved insns.
2157 We can't use the moved insn because it is out of range
2158 in uid_luid. Only the old insns have luids. */
2159 REGNO_FIRST_UID (regno) = INSN_UID (loop_start);
2160 if (uid_luid[REGNO_LAST_UID (regno)] < INSN_LUID (end))
2161 REGNO_LAST_UID (regno) = INSN_UID (end);
2163 /* Combine with this moved insn any other matching movables. */
2166 for (m1 = movables; m1; m1 = m1->next)
2171 /* Schedule the reg loaded by M1
2172 for replacement so that shares the reg of M.
2173 If the modes differ (only possible in restricted
2174 circumstances, make a SUBREG.
2176 Note this assumes that the target dependent files
2177 treat REG and SUBREG equally, including within
2178 GO_IF_LEGITIMATE_ADDRESS and in all the
2179 predicates since we never verify that replacing the
2180 original register with a SUBREG results in a
2181 recognizable insn. */
2182 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
2183 reg_map[m1->regno] = m->set_dest;
2186 = gen_lowpart_common (GET_MODE (m1->set_dest),
2189 /* Get rid of the matching insn
2190 and prevent further processing of it. */
2193 /* if library call, delete all insn except last, which
2195 if ((temp = find_reg_note (m1->insn, REG_RETVAL,
2198 for (temp = XEXP (temp, 0); temp != m1->insn;
2199 temp = NEXT_INSN (temp))
2202 delete_insn (m1->insn);
2204 /* Any other movable that loads the same register
2206 already_moved[m1->regno] = 1;
2208 /* The reg merged here is now invariant,
2209 if the reg it matches is invariant. */
2211 VARRAY_INT (set_in_loop, m1->regno) = 0;
2214 else if (loop_dump_stream)
2215 fprintf (loop_dump_stream, "not desirable");
2217 else if (loop_dump_stream && !m->match)
2218 fprintf (loop_dump_stream, "not safe");
2220 if (loop_dump_stream)
2221 fprintf (loop_dump_stream, "\n");
2225 new_start = loop_start;
2227 /* Go through all the instructions in the loop, making
2228 all the register substitutions scheduled in REG_MAP. */
2229 for (p = new_start; p != end; p = NEXT_INSN (p))
2230 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
2231 || GET_CODE (p) == CALL_INSN)
2233 replace_regs (PATTERN (p), reg_map, nregs, 0);
2234 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
2240 /* Scan X and replace the address of any MEM in it with ADDR.
2241 REG is the address that MEM should have before the replacement. */
2244 replace_call_address (x, reg, addr)
2247 register enum rtx_code code;
2249 register const char *fmt;
2253 code = GET_CODE (x);
2267 /* Short cut for very common case. */
2268 replace_call_address (XEXP (x, 1), reg, addr);
2272 /* Short cut for very common case. */
2273 replace_call_address (XEXP (x, 0), reg, addr);
2277 /* If this MEM uses a reg other than the one we expected,
2278 something is wrong. */
2279 if (XEXP (x, 0) != reg)
2288 fmt = GET_RTX_FORMAT (code);
2289 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2292 replace_call_address (XEXP (x, i), reg, addr);
2296 for (j = 0; j < XVECLEN (x, i); j++)
2297 replace_call_address (XVECEXP (x, i, j), reg, addr);
2303 /* Return the number of memory refs to addresses that vary
2307 count_nonfixed_reads (x)
2310 register enum rtx_code code;
2312 register const char *fmt;
2318 code = GET_CODE (x);
2332 return ((invariant_p (XEXP (x, 0)) != 1)
2333 + count_nonfixed_reads (XEXP (x, 0)));
2340 fmt = GET_RTX_FORMAT (code);
2341 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2344 value += count_nonfixed_reads (XEXP (x, i));
2348 for (j = 0; j < XVECLEN (x, i); j++)
2349 value += count_nonfixed_reads (XVECEXP (x, i, j));
2357 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2358 Replace it with an instruction to load just the low bytes
2359 if the machine supports such an instruction,
2360 and insert above LOOP_START an instruction to clear the register. */
2363 constant_high_bytes (p, loop_start)
2367 register int insn_code_number;
2369 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2370 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2372 new = gen_rtx_SET (VOIDmode,
2373 gen_rtx_STRICT_LOW_PART (VOIDmode,
2374 gen_rtx_SUBREG (GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2375 SET_DEST (PATTERN (p)),
2377 XEXP (SET_SRC (PATTERN (p)), 0));
2378 insn_code_number = recog (new, p);
2380 if (insn_code_number)
2384 /* Clear destination register before the loop. */
2385 emit_insn_before (gen_rtx_SET (VOIDmode, SET_DEST (PATTERN (p)),
2389 /* Inside the loop, just load the low part. */
2395 /* Scan a loop setting the variables `unknown_address_altered',
2396 `num_mem_sets', `loop_continue', `loops_enclosed', `loop_has_call',
2397 `loop_has_volatile', and `loop_has_tablejump'.
2398 Also, fill in the array `loop_mems' and the list `loop_store_mems'. */
2401 prescan_loop (start, end)
2404 register int level = 1;
2406 int loop_has_multiple_exit_targets = 0;
2407 /* The label after END. Jumping here is just like falling off the
2408 end of the loop. We use next_nonnote_insn instead of next_label
2409 as a hedge against the (pathological) case where some actual insn
2410 might end up between the two. */
2411 rtx exit_target = next_nonnote_insn (end);
2412 if (exit_target == NULL_RTX || GET_CODE (exit_target) != CODE_LABEL)
2413 loop_has_multiple_exit_targets = 1;
2415 unknown_address_altered = 0;
2417 loop_has_volatile = 0;
2418 loop_has_tablejump = 0;
2419 loop_store_mems = NULL_RTX;
2420 first_loop_store_insn = NULL_RTX;
2427 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2428 insn = NEXT_INSN (insn))
2430 if (GET_CODE (insn) == NOTE)
2432 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2435 /* Count number of loops contained in this one. */
2438 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2447 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2450 loop_continue = insn;
2453 else if (GET_CODE (insn) == CALL_INSN)
2455 if (! CONST_CALL_P (insn))
2456 unknown_address_altered = 1;
2459 else if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2461 rtx label1 = NULL_RTX;
2462 rtx label2 = NULL_RTX;
2464 if (volatile_refs_p (PATTERN (insn)))
2465 loop_has_volatile = 1;
2467 if (GET_CODE (insn) == JUMP_INSN
2468 && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
2469 || GET_CODE (PATTERN (insn)) == ADDR_VEC))
2470 loop_has_tablejump = 1;
2472 note_stores (PATTERN (insn), note_addr_stored);
2473 if (! first_loop_store_insn && loop_store_mems)
2474 first_loop_store_insn = insn;
2476 if (! loop_has_multiple_exit_targets
2477 && GET_CODE (insn) == JUMP_INSN
2478 && GET_CODE (PATTERN (insn)) == SET
2479 && SET_DEST (PATTERN (insn)) == pc_rtx)
2481 if (GET_CODE (SET_SRC (PATTERN (insn))) == IF_THEN_ELSE)
2483 label1 = XEXP (SET_SRC (PATTERN (insn)), 1);
2484 label2 = XEXP (SET_SRC (PATTERN (insn)), 2);
2488 label1 = SET_SRC (PATTERN (insn));
2492 if (label1 && label1 != pc_rtx)
2494 if (GET_CODE (label1) != LABEL_REF)
2496 /* Something tricky. */
2497 loop_has_multiple_exit_targets = 1;
2500 else if (XEXP (label1, 0) != exit_target
2501 && LABEL_OUTSIDE_LOOP_P (label1))
2503 /* A jump outside the current loop. */
2504 loop_has_multiple_exit_targets = 1;
2514 else if (GET_CODE (insn) == RETURN)
2515 loop_has_multiple_exit_targets = 1;
2518 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2519 if (/* We can't tell what MEMs are aliased by what. */
2520 !unknown_address_altered
2521 /* An exception thrown by a called function might land us
2524 /* We don't want loads for MEMs moved to a location before the
2525 one at which their stack memory becomes allocated. (Note
2526 that this is not a problem for malloc, etc., since those
2527 require actual function calls. */
2528 && !current_function_calls_alloca
2529 /* There are ways to leave the loop other than falling off the
2531 && !loop_has_multiple_exit_targets)
2532 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2533 insn = NEXT_INSN (insn))
2534 for_each_rtx (&insn, insert_loop_mem, 0);
2537 /* LOOP_NUMBER_CONT_DOMINATOR is now the last label between the loop start
2538 and the continue note that is a the destination of a (cond)jump after
2539 the continue note. If there is any (cond)jump between the loop start
2540 and what we have so far as LOOP_NUMBER_CONT_DOMINATOR that has a
2541 target between LOOP_DOMINATOR and the continue note, move
2542 LOOP_NUMBER_CONT_DOMINATOR forward to that label; if a jump's
2543 destination cannot be determined, clear LOOP_NUMBER_CONT_DOMINATOR. */
2546 verify_dominator (loop_number)
2551 if (! loop_number_cont_dominator[loop_number])
2552 /* This can happen for an empty loop, e.g. in
2553 gcc.c-torture/compile/920410-2.c */
2555 if (loop_number_cont_dominator[loop_number] == const0_rtx)
2557 loop_number_cont_dominator[loop_number] = 0;
2560 for (insn = loop_number_loop_starts[loop_number];
2561 insn != loop_number_cont_dominator[loop_number];
2562 insn = NEXT_INSN (insn))
2564 if (GET_CODE (insn) == JUMP_INSN
2565 && GET_CODE (PATTERN (insn)) != RETURN)
2567 rtx label = JUMP_LABEL (insn);
2570 /* If it is not a jump we can easily understand or for
2571 which we do not have jump target information in the JUMP_LABEL
2572 field (consider ADDR_VEC and ADDR_DIFF_VEC insns), then clear
2573 LOOP_NUMBER_CONT_DOMINATOR. */
2574 if ((! condjump_p (insn)
2575 && ! condjump_in_parallel_p (insn))
2576 || label == NULL_RTX)
2578 loop_number_cont_dominator[loop_number] = NULL_RTX;
2582 label_luid = INSN_LUID (label);
2583 if (label_luid < INSN_LUID (loop_number_loop_cont[loop_number])
2585 > INSN_LUID (loop_number_cont_dominator[loop_number])))
2586 loop_number_cont_dominator[loop_number] = label;
2591 /* Scan the function looking for loops. Record the start and end of each loop.
2592 Also mark as invalid loops any loops that contain a setjmp or are branched
2593 to from outside the loop. */
2596 find_and_verify_loops (f)
2600 int current_loop = -1;
2604 compute_luids (f, NULL_RTX, 0);
2606 /* If there are jumps to undefined labels,
2607 treat them as jumps out of any/all loops.
2608 This also avoids writing past end of tables when there are no loops. */
2609 uid_loop_num[0] = -1;
2611 /* Find boundaries of loops, mark which loops are contained within
2612 loops, and invalidate loops that have setjmp. */
2614 for (insn = f; insn; insn = NEXT_INSN (insn))
2616 if (GET_CODE (insn) == NOTE)
2617 switch (NOTE_LINE_NUMBER (insn))
2619 case NOTE_INSN_LOOP_BEG:
2620 loop_number_loop_starts[++next_loop] = insn;
2621 loop_number_loop_ends[next_loop] = 0;
2622 loop_number_loop_cont[next_loop] = 0;
2623 loop_number_cont_dominator[next_loop] = 0;
2624 loop_outer_loop[next_loop] = current_loop;
2625 loop_invalid[next_loop] = 0;
2626 loop_number_exit_labels[next_loop] = 0;
2627 loop_number_exit_count[next_loop] = 0;
2628 current_loop = next_loop;
2631 case NOTE_INSN_SETJMP:
2632 /* In this case, we must invalidate our current loop and any
2634 for (loop = current_loop; loop != -1; loop = loop_outer_loop[loop])
2636 loop_invalid[loop] = 1;
2637 if (loop_dump_stream)
2638 fprintf (loop_dump_stream,
2639 "\nLoop at %d ignored due to setjmp.\n",
2640 INSN_UID (loop_number_loop_starts[loop]));
2644 case NOTE_INSN_LOOP_CONT:
2645 loop_number_loop_cont[current_loop] = insn;
2647 case NOTE_INSN_LOOP_END:
2648 if (current_loop == -1)
2651 loop_number_loop_ends[current_loop] = insn;
2652 verify_dominator (current_loop);
2653 current_loop = loop_outer_loop[current_loop];
2659 /* If for any loop, this is a jump insn between the NOTE_INSN_LOOP_CONT
2660 and NOTE_INSN_LOOP_END notes, update loop_number_loop_dominator. */
2661 else if (GET_CODE (insn) == JUMP_INSN
2662 && GET_CODE (PATTERN (insn)) != RETURN
2663 && current_loop >= 0)
2666 rtx label = JUMP_LABEL (insn);
2668 if (! condjump_p (insn) && ! condjump_in_parallel_p (insn))
2671 this_loop = current_loop;
2674 /* First see if we care about this loop. */
2675 if (loop_number_loop_cont[this_loop]
2676 && loop_number_cont_dominator[this_loop] != const0_rtx)
2678 /* If the jump destination is not known, invalidate
2679 loop_number_const_dominator. */
2681 loop_number_cont_dominator[this_loop] = const0_rtx;
2683 /* Check if the destination is between loop start and
2685 if ((INSN_LUID (label)
2686 < INSN_LUID (loop_number_loop_cont[this_loop]))
2687 && (INSN_LUID (label)
2688 > INSN_LUID (loop_number_loop_starts[this_loop]))
2689 /* And if there is no later destination already
2691 && (! loop_number_cont_dominator[this_loop]
2692 || (INSN_LUID (label)
2693 > INSN_LUID (loop_number_cont_dominator
2695 loop_number_cont_dominator[this_loop] = label;
2697 this_loop = loop_outer_loop[this_loop];
2699 while (this_loop >= 0);
2702 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2703 enclosing loop, but this doesn't matter. */
2704 uid_loop_num[INSN_UID (insn)] = current_loop;
2707 /* Any loop containing a label used in an initializer must be invalidated,
2708 because it can be jumped into from anywhere. */
2710 for (label = forced_labels; label; label = XEXP (label, 1))
2714 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2716 loop_num = loop_outer_loop[loop_num])
2717 loop_invalid[loop_num] = 1;
2720 /* Any loop containing a label used for an exception handler must be
2721 invalidated, because it can be jumped into from anywhere. */
2723 for (label = exception_handler_labels; label; label = XEXP (label, 1))
2727 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2729 loop_num = loop_outer_loop[loop_num])
2730 loop_invalid[loop_num] = 1;
2733 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2734 loop that it is not contained within, that loop is marked invalid.
2735 If any INSN or CALL_INSN uses a label's address, then the loop containing
2736 that label is marked invalid, because it could be jumped into from
2739 Also look for blocks of code ending in an unconditional branch that
2740 exits the loop. If such a block is surrounded by a conditional
2741 branch around the block, move the block elsewhere (see below) and
2742 invert the jump to point to the code block. This may eliminate a
2743 label in our loop and will simplify processing by both us and a
2744 possible second cse pass. */
2746 for (insn = f; insn; insn = NEXT_INSN (insn))
2747 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2749 int this_loop_num = uid_loop_num[INSN_UID (insn)];
2751 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2753 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2758 for (loop_num = uid_loop_num[INSN_UID (XEXP (note, 0))];
2760 loop_num = loop_outer_loop[loop_num])
2761 loop_invalid[loop_num] = 1;
2765 if (GET_CODE (insn) != JUMP_INSN)
2768 mark_loop_jump (PATTERN (insn), this_loop_num);
2770 /* See if this is an unconditional branch outside the loop. */
2771 if (this_loop_num != -1
2772 && (GET_CODE (PATTERN (insn)) == RETURN
2773 || (simplejump_p (insn)
2774 && (uid_loop_num[INSN_UID (JUMP_LABEL (insn))]
2776 && get_max_uid () < max_uid_for_loop)
2779 rtx our_next = next_real_insn (insn);
2781 int outer_loop = -1;
2783 /* Go backwards until we reach the start of the loop, a label,
2785 for (p = PREV_INSN (insn);
2786 GET_CODE (p) != CODE_LABEL
2787 && ! (GET_CODE (p) == NOTE
2788 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2789 && GET_CODE (p) != JUMP_INSN;
2793 /* Check for the case where we have a jump to an inner nested
2794 loop, and do not perform the optimization in that case. */
2796 if (JUMP_LABEL (insn))
2798 dest_loop = uid_loop_num[INSN_UID (JUMP_LABEL (insn))];
2799 if (dest_loop != -1)
2801 for (outer_loop = dest_loop; outer_loop != -1;
2802 outer_loop = loop_outer_loop[outer_loop])
2803 if (outer_loop == this_loop_num)
2808 /* Make sure that the target of P is within the current loop. */
2810 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2811 && uid_loop_num[INSN_UID (JUMP_LABEL (p))] != this_loop_num)
2812 outer_loop = this_loop_num;
2814 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2815 we have a block of code to try to move.
2817 We look backward and then forward from the target of INSN
2818 to find a BARRIER at the same loop depth as the target.
2819 If we find such a BARRIER, we make a new label for the start
2820 of the block, invert the jump in P and point it to that label,
2821 and move the block of code to the spot we found. */
2823 if (outer_loop == -1
2824 && GET_CODE (p) == JUMP_INSN
2825 && JUMP_LABEL (p) != 0
2826 /* Just ignore jumps to labels that were never emitted.
2827 These always indicate compilation errors. */
2828 && INSN_UID (JUMP_LABEL (p)) != 0
2830 && ! simplejump_p (p)
2831 && next_real_insn (JUMP_LABEL (p)) == our_next)
2834 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2835 int target_loop_num = uid_loop_num[INSN_UID (target)];
2838 for (loc = target; loc; loc = PREV_INSN (loc))
2839 if (GET_CODE (loc) == BARRIER
2840 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2844 for (loc = target; loc; loc = NEXT_INSN (loc))
2845 if (GET_CODE (loc) == BARRIER
2846 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2851 rtx cond_label = JUMP_LABEL (p);
2852 rtx new_label = get_label_after (p);
2854 /* Ensure our label doesn't go away. */
2855 LABEL_NUSES (cond_label)++;
2857 /* Verify that uid_loop_num is large enough and that
2859 if (invert_jump (p, new_label))
2863 /* If no suitable BARRIER was found, create a suitable
2864 one before TARGET. Since TARGET is a fall through
2865 path, we'll need to insert an jump around our block
2866 and a add a BARRIER before TARGET.
2868 This creates an extra unconditional jump outside
2869 the loop. However, the benefits of removing rarely
2870 executed instructions from inside the loop usually
2871 outweighs the cost of the extra unconditional jump
2872 outside the loop. */
2877 temp = gen_jump (JUMP_LABEL (insn));
2878 temp = emit_jump_insn_before (temp, target);
2879 JUMP_LABEL (temp) = JUMP_LABEL (insn);
2880 LABEL_NUSES (JUMP_LABEL (insn))++;
2881 loc = emit_barrier_before (target);
2884 /* Include the BARRIER after INSN and copy the
2886 new_label = squeeze_notes (new_label, NEXT_INSN (insn));
2887 reorder_insns (new_label, NEXT_INSN (insn), loc);
2889 /* All those insns are now in TARGET_LOOP_NUM. */
2890 for (q = new_label; q != NEXT_INSN (NEXT_INSN (insn));
2892 uid_loop_num[INSN_UID (q)] = target_loop_num;
2894 /* The label jumped to by INSN is no longer a loop exit.
2895 Unless INSN does not have a label (e.g., it is a
2896 RETURN insn), search loop_number_exit_labels to find
2897 its label_ref, and remove it. Also turn off
2898 LABEL_OUTSIDE_LOOP_P bit. */
2899 if (JUMP_LABEL (insn))
2904 r = loop_number_exit_labels[this_loop_num];
2905 r; q = r, r = LABEL_NEXTREF (r))
2906 if (XEXP (r, 0) == JUMP_LABEL (insn))
2908 LABEL_OUTSIDE_LOOP_P (r) = 0;
2910 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2912 loop_number_exit_labels[this_loop_num]
2913 = LABEL_NEXTREF (r);
2917 for (loop_num = this_loop_num;
2918 loop_num != -1 && loop_num != target_loop_num;
2919 loop_num = loop_outer_loop[loop_num])
2920 loop_number_exit_count[loop_num]--;
2922 /* If we didn't find it, then something is wrong. */
2927 /* P is now a jump outside the loop, so it must be put
2928 in loop_number_exit_labels, and marked as such.
2929 The easiest way to do this is to just call
2930 mark_loop_jump again for P. */
2931 mark_loop_jump (PATTERN (p), this_loop_num);
2933 /* If INSN now jumps to the insn after it,
2935 if (JUMP_LABEL (insn) != 0
2936 && (next_real_insn (JUMP_LABEL (insn))
2937 == next_real_insn (insn)))
2941 /* Continue the loop after where the conditional
2942 branch used to jump, since the only branch insn
2943 in the block (if it still remains) is an inter-loop
2944 branch and hence needs no processing. */
2945 insn = NEXT_INSN (cond_label);
2947 if (--LABEL_NUSES (cond_label) == 0)
2948 delete_insn (cond_label);
2950 /* This loop will be continued with NEXT_INSN (insn). */
2951 insn = PREV_INSN (insn);
2958 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2959 loops it is contained in, mark the target loop invalid.
2961 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2964 mark_loop_jump (x, loop_num)
2972 switch (GET_CODE (x))
2985 /* There could be a label reference in here. */
2986 mark_loop_jump (XEXP (x, 0), loop_num);
2992 mark_loop_jump (XEXP (x, 0), loop_num);
2993 mark_loop_jump (XEXP (x, 1), loop_num);
2997 /* This may refer to a LABEL_REF or SYMBOL_REF. */
2998 mark_loop_jump (XEXP (x, 1), loop_num);
3003 mark_loop_jump (XEXP (x, 0), loop_num);
3007 dest_loop = uid_loop_num[INSN_UID (XEXP (x, 0))];
3009 /* Link together all labels that branch outside the loop. This
3010 is used by final_[bg]iv_value and the loop unrolling code. Also
3011 mark this LABEL_REF so we know that this branch should predict
3014 /* A check to make sure the label is not in an inner nested loop,
3015 since this does not count as a loop exit. */
3016 if (dest_loop != -1)
3018 for (outer_loop = dest_loop; outer_loop != -1;
3019 outer_loop = loop_outer_loop[outer_loop])
3020 if (outer_loop == loop_num)
3026 if (loop_num != -1 && outer_loop == -1)
3028 LABEL_OUTSIDE_LOOP_P (x) = 1;
3029 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
3030 loop_number_exit_labels[loop_num] = x;
3032 for (outer_loop = loop_num;
3033 outer_loop != -1 && outer_loop != dest_loop;
3034 outer_loop = loop_outer_loop[outer_loop])
3035 loop_number_exit_count[outer_loop]++;
3038 /* If this is inside a loop, but not in the current loop or one enclosed
3039 by it, it invalidates at least one loop. */
3041 if (dest_loop == -1)
3044 /* We must invalidate every nested loop containing the target of this
3045 label, except those that also contain the jump insn. */
3047 for (; dest_loop != -1; dest_loop = loop_outer_loop[dest_loop])
3049 /* Stop when we reach a loop that also contains the jump insn. */
3050 for (outer_loop = loop_num; outer_loop != -1;
3051 outer_loop = loop_outer_loop[outer_loop])
3052 if (dest_loop == outer_loop)
3055 /* If we get here, we know we need to invalidate a loop. */
3056 if (loop_dump_stream && ! loop_invalid[dest_loop])
3057 fprintf (loop_dump_stream,
3058 "\nLoop at %d ignored due to multiple entry points.\n",
3059 INSN_UID (loop_number_loop_starts[dest_loop]));
3061 loop_invalid[dest_loop] = 1;
3066 /* If this is not setting pc, ignore. */
3067 if (SET_DEST (x) == pc_rtx)
3068 mark_loop_jump (SET_SRC (x), loop_num);
3072 mark_loop_jump (XEXP (x, 1), loop_num);
3073 mark_loop_jump (XEXP (x, 2), loop_num);
3078 for (i = 0; i < XVECLEN (x, 0); i++)
3079 mark_loop_jump (XVECEXP (x, 0, i), loop_num);
3083 for (i = 0; i < XVECLEN (x, 1); i++)
3084 mark_loop_jump (XVECEXP (x, 1, i), loop_num);
3088 /* Strictly speaking this is not a jump into the loop, only a possible
3089 jump out of the loop. However, we have no way to link the destination
3090 of this jump onto the list of exit labels. To be safe we mark this
3091 loop and any containing loops as invalid. */
3094 for (outer_loop = loop_num; outer_loop != -1;
3095 outer_loop = loop_outer_loop[outer_loop])
3097 if (loop_dump_stream && ! loop_invalid[outer_loop])
3098 fprintf (loop_dump_stream,
3099 "\nLoop at %d ignored due to unknown exit jump.\n",
3100 INSN_UID (loop_number_loop_starts[outer_loop]));
3101 loop_invalid[outer_loop] = 1;
3108 /* Return nonzero if there is a label in the range from
3109 insn INSN to and including the insn whose luid is END
3110 INSN must have an assigned luid (i.e., it must not have
3111 been previously created by loop.c). */
3114 labels_in_range_p (insn, end)
3118 while (insn && INSN_LUID (insn) <= end)
3120 if (GET_CODE (insn) == CODE_LABEL)
3122 insn = NEXT_INSN (insn);
3128 /* Record that a memory reference X is being set. */
3131 note_addr_stored (x, y)
3133 rtx y ATTRIBUTE_UNUSED;
3135 if (x == 0 || GET_CODE (x) != MEM)
3138 /* Count number of memory writes.
3139 This affects heuristics in strength_reduce. */
3142 /* BLKmode MEM means all memory is clobbered. */
3143 if (GET_MODE (x) == BLKmode)
3144 unknown_address_altered = 1;
3146 if (unknown_address_altered)
3149 loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, x, loop_store_mems);
3152 /* Return nonzero if the rtx X is invariant over the current loop.
3154 The value is 2 if we refer to something only conditionally invariant.
3156 If `unknown_address_altered' is nonzero, no memory ref is invariant.
3157 Otherwise, a memory ref is invariant if it does not conflict with
3158 anything stored in `loop_store_mems'. */
3165 register enum rtx_code code;
3166 register const char *fmt;
3167 int conditional = 0;
3172 code = GET_CODE (x);
3182 /* A LABEL_REF is normally invariant, however, if we are unrolling
3183 loops, and this label is inside the loop, then it isn't invariant.
3184 This is because each unrolled copy of the loop body will have
3185 a copy of this label. If this was invariant, then an insn loading
3186 the address of this label into a register might get moved outside
3187 the loop, and then each loop body would end up using the same label.
3189 We don't know the loop bounds here though, so just fail for all
3191 if (flag_unroll_loops)
3198 case UNSPEC_VOLATILE:
3202 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
3203 since the reg might be set by initialization within the loop. */
3205 if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
3206 || x == arg_pointer_rtx)
3207 && ! current_function_has_nonlocal_goto)
3211 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
3214 if (VARRAY_INT (set_in_loop, REGNO (x)) < 0)
3217 return VARRAY_INT (set_in_loop, REGNO (x)) == 0;
3220 /* Volatile memory references must be rejected. Do this before
3221 checking for read-only items, so that volatile read-only items
3222 will be rejected also. */
3223 if (MEM_VOLATILE_P (x))
3226 /* Read-only items (such as constants in a constant pool) are
3227 invariant if their address is. */
3228 if (RTX_UNCHANGING_P (x))
3231 /* If we had a subroutine call, any location in memory could have been
3233 if (unknown_address_altered)
3236 /* See if there is any dependence between a store and this load. */
3237 mem_list_entry = loop_store_mems;
3238 while (mem_list_entry)
3240 if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
3243 mem_list_entry = XEXP (mem_list_entry, 1);
3246 /* It's not invalidated by a store in memory
3247 but we must still verify the address is invariant. */
3251 /* Don't mess with insns declared volatile. */
3252 if (MEM_VOLATILE_P (x))
3260 fmt = GET_RTX_FORMAT (code);
3261 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3265 int tem = invariant_p (XEXP (x, i));
3271 else if (fmt[i] == 'E')
3274 for (j = 0; j < XVECLEN (x, i); j++)
3276 int tem = invariant_p (XVECEXP (x, i, j));
3286 return 1 + conditional;
3290 /* Return nonzero if all the insns in the loop that set REG
3291 are INSN and the immediately following insns,
3292 and if each of those insns sets REG in an invariant way
3293 (not counting uses of REG in them).
3295 The value is 2 if some of these insns are only conditionally invariant.
3297 We assume that INSN itself is the first set of REG
3298 and that its source is invariant. */
3301 consec_sets_invariant_p (reg, n_sets, insn)
3305 register rtx p = insn;
3306 register int regno = REGNO (reg);
3308 /* Number of sets we have to insist on finding after INSN. */
3309 int count = n_sets - 1;
3310 int old = VARRAY_INT (set_in_loop, regno);
3314 /* If N_SETS hit the limit, we can't rely on its value. */
3318 VARRAY_INT (set_in_loop, regno) = 0;
3322 register enum rtx_code code;
3326 code = GET_CODE (p);
3328 /* If library call, skip to end of it. */
3329 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3334 && (set = single_set (p))
3335 && GET_CODE (SET_DEST (set)) == REG
3336 && REGNO (SET_DEST (set)) == regno)
3338 this = invariant_p (SET_SRC (set));
3341 else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)))
3343 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3344 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3346 this = (CONSTANT_P (XEXP (temp, 0))
3347 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
3348 && invariant_p (XEXP (temp, 0))));
3355 else if (code != NOTE)
3357 VARRAY_INT (set_in_loop, regno) = old;
3362 VARRAY_INT (set_in_loop, regno) = old;
3363 /* If invariant_p ever returned 2, we return 2. */
3364 return 1 + (value & 2);
3368 /* I don't think this condition is sufficient to allow INSN
3369 to be moved, so we no longer test it. */
3371 /* Return 1 if all insns in the basic block of INSN and following INSN
3372 that set REG are invariant according to TABLE. */
3375 all_sets_invariant_p (reg, insn, table)
3379 register rtx p = insn;
3380 register int regno = REGNO (reg);
3384 register enum rtx_code code;
3386 code = GET_CODE (p);
3387 if (code == CODE_LABEL || code == JUMP_INSN)
3389 if (code == INSN && GET_CODE (PATTERN (p)) == SET
3390 && GET_CODE (SET_DEST (PATTERN (p))) == REG
3391 && REGNO (SET_DEST (PATTERN (p))) == regno)
3393 if (!invariant_p (SET_SRC (PATTERN (p)), table))
3400 /* Look at all uses (not sets) of registers in X. For each, if it is
3401 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3402 a different insn, set USAGE[REGNO] to const0_rtx. */
3405 find_single_use_in_loop (insn, x, usage)
3410 enum rtx_code code = GET_CODE (x);
3411 const char *fmt = GET_RTX_FORMAT (code);
3415 VARRAY_RTX (usage, REGNO (x))
3416 = (VARRAY_RTX (usage, REGNO (x)) != 0
3417 && VARRAY_RTX (usage, REGNO (x)) != insn)
3418 ? const0_rtx : insn;
3420 else if (code == SET)
3422 /* Don't count SET_DEST if it is a REG; otherwise count things
3423 in SET_DEST because if a register is partially modified, it won't
3424 show up as a potential movable so we don't care how USAGE is set
3426 if (GET_CODE (SET_DEST (x)) != REG)
3427 find_single_use_in_loop (insn, SET_DEST (x), usage);
3428 find_single_use_in_loop (insn, SET_SRC (x), usage);
3431 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3433 if (fmt[i] == 'e' && XEXP (x, i) != 0)
3434 find_single_use_in_loop (insn, XEXP (x, i), usage);
3435 else if (fmt[i] == 'E')
3436 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3437 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
3441 /* Count and record any set in X which is contained in INSN. Update
3442 MAY_NOT_MOVE and LAST_SET for any register set in X. */
3445 count_one_set (insn, x, may_not_move, last_set)
3447 varray_type may_not_move;
3450 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3451 /* Don't move a reg that has an explicit clobber.
3452 It's not worth the pain to try to do it correctly. */
3453 VARRAY_CHAR (may_not_move, REGNO (XEXP (x, 0))) = 1;
3455 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3457 rtx dest = SET_DEST (x);
3458 while (GET_CODE (dest) == SUBREG
3459 || GET_CODE (dest) == ZERO_EXTRACT
3460 || GET_CODE (dest) == SIGN_EXTRACT
3461 || GET_CODE (dest) == STRICT_LOW_PART)
3462 dest = XEXP (dest, 0);
3463 if (GET_CODE (dest) == REG)
3465 register int regno = REGNO (dest);
3466 /* If this is the first setting of this reg
3467 in current basic block, and it was set before,
3468 it must be set in two basic blocks, so it cannot
3469 be moved out of the loop. */
3470 if (VARRAY_INT (set_in_loop, regno) > 0
3471 && last_set[regno] == 0)
3472 VARRAY_CHAR (may_not_move, regno) = 1;
3473 /* If this is not first setting in current basic block,
3474 see if reg was used in between previous one and this.
3475 If so, neither one can be moved. */
3476 if (last_set[regno] != 0
3477 && reg_used_between_p (dest, last_set[regno], insn))
3478 VARRAY_CHAR (may_not_move, regno) = 1;
3479 if (VARRAY_INT (set_in_loop, regno) < 127)
3480 ++VARRAY_INT (set_in_loop, regno);
3481 last_set[regno] = insn;
3486 /* Increment SET_IN_LOOP at the index of each register
3487 that is modified by an insn between FROM and TO.
3488 If the value of an element of SET_IN_LOOP becomes 127 or more,
3489 stop incrementing it, to avoid overflow.
3491 Store in SINGLE_USAGE[I] the single insn in which register I is
3492 used, if it is only used once. Otherwise, it is set to 0 (for no
3493 uses) or const0_rtx for more than one use. This parameter may be zero,
3494 in which case this processing is not done.
3496 Store in *COUNT_PTR the number of actual instruction
3497 in the loop. We use this to decide what is worth moving out. */
3499 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3500 In that case, it is the insn that last set reg n. */
3503 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
3504 register rtx from, to;
3505 varray_type may_not_move;
3506 varray_type single_usage;
3510 register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx));
3512 register int count = 0;
3514 bzero ((char *) last_set, nregs * sizeof (rtx));
3515 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3517 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3521 /* Record registers that have exactly one use. */
3522 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3524 /* Include uses in REG_EQUAL notes. */
3525 if (REG_NOTES (insn))
3526 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3528 if (GET_CODE (PATTERN (insn)) == SET
3529 || GET_CODE (PATTERN (insn)) == CLOBBER)
3530 count_one_set (insn, PATTERN (insn), may_not_move, last_set);
3531 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3534 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3535 count_one_set (insn, XVECEXP (PATTERN (insn), 0, i),
3536 may_not_move, last_set);
3540 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3541 bzero ((char *) last_set, nregs * sizeof (rtx));
3546 /* Given a loop that is bounded by LOOP_START and LOOP_END
3547 and that is entered at SCAN_START,
3548 return 1 if the register set in SET contained in insn INSN is used by
3549 any insn that precedes INSN in cyclic order starting
3550 from the loop entry point.
3552 We don't want to use INSN_LUID here because if we restrict INSN to those
3553 that have a valid INSN_LUID, it means we cannot move an invariant out
3554 from an inner loop past two loops. */
3557 loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end)
3558 rtx set, insn, loop_start, scan_start, loop_end;
3560 rtx reg = SET_DEST (set);
3563 /* Scan forward checking for register usage. If we hit INSN, we
3564 are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */
3565 for (p = scan_start; p != insn; p = NEXT_INSN (p))
3567 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
3568 && reg_overlap_mentioned_p (reg, PATTERN (p)))
3578 /* A "basic induction variable" or biv is a pseudo reg that is set
3579 (within this loop) only by incrementing or decrementing it. */
3580 /* A "general induction variable" or giv is a pseudo reg whose
3581 value is a linear function of a biv. */
3583 /* Bivs are recognized by `basic_induction_var';
3584 Givs by `general_induction_var'. */
3586 /* Indexed by register number, indicates whether or not register is an
3587 induction variable, and if so what type. */
3589 varray_type reg_iv_type;
3591 /* Indexed by register number, contains pointer to `struct induction'
3592 if register is an induction variable. This holds general info for
3593 all induction variables. */
3595 varray_type reg_iv_info;
3597 /* Indexed by register number, contains pointer to `struct iv_class'
3598 if register is a basic induction variable. This holds info describing
3599 the class (a related group) of induction variables that the biv belongs
3602 struct iv_class **reg_biv_class;
3604 /* The head of a list which links together (via the next field)
3605 every iv class for the current loop. */
3607 struct iv_class *loop_iv_list;
3609 /* Givs made from biv increments are always splittable for loop unrolling.
3610 Since there is no regscan info for them, we have to keep track of them
3612 int first_increment_giv, last_increment_giv;
3614 /* Communication with routines called via `note_stores'. */
3616 static rtx note_insn;
3618 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3620 static rtx addr_placeholder;
3622 /* ??? Unfinished optimizations, and possible future optimizations,
3623 for the strength reduction code. */
3625 /* ??? The interaction of biv elimination, and recognition of 'constant'
3626 bivs, may cause problems. */
3628 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3629 performance problems.
3631 Perhaps don't eliminate things that can be combined with an addressing
3632 mode. Find all givs that have the same biv, mult_val, and add_val;
3633 then for each giv, check to see if its only use dies in a following
3634 memory address. If so, generate a new memory address and check to see
3635 if it is valid. If it is valid, then store the modified memory address,
3636 otherwise, mark the giv as not done so that it will get its own iv. */
3638 /* ??? Could try to optimize branches when it is known that a biv is always
3641 /* ??? When replace a biv in a compare insn, we should replace with closest
3642 giv so that an optimized branch can still be recognized by the combiner,
3643 e.g. the VAX acb insn. */
3645 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3646 was rerun in loop_optimize whenever a register was added or moved.
3647 Also, some of the optimizations could be a little less conservative. */
3649 /* Perform strength reduction and induction variable elimination.
3651 Pseudo registers created during this function will be beyond the last
3652 valid index in several tables including n_times_set and regno_last_uid.
3653 This does not cause a problem here, because the added registers cannot be
3654 givs outside of their loop, and hence will never be reconsidered.
3655 But scan_loop must check regnos to make sure they are in bounds.
3657 SCAN_START is the first instruction in the loop, as the loop would
3658 actually be executed. END is the NOTE_INSN_LOOP_END. LOOP_TOP is
3659 the first instruction in the loop, as it is layed out in the
3660 instruction stream. LOOP_START is the NOTE_INSN_LOOP_BEG.
3661 LOOP_CONT is the NOTE_INSN_LOOP_CONT. */
3664 strength_reduce (scan_start, end, loop_top, insn_count,
3665 loop_start, loop_end, loop_cont, unroll_p, bct_p)
3673 int unroll_p, bct_p ATTRIBUTE_UNUSED;
3681 /* This is 1 if current insn is not executed at least once for every loop
3683 int not_every_iteration = 0;
3684 /* This is 1 if current insn may be executed more than once for every
3686 int maybe_multiple = 0;
3687 /* This is 1 if we have past a branch back to the top of the loop
3688 (aka a loop latch). */
3689 int past_loop_latch = 0;
3690 /* Temporary list pointers for traversing loop_iv_list. */
3691 struct iv_class *bl, **backbl;
3692 /* Ratio of extra register life span we can justify
3693 for saving an instruction. More if loop doesn't call subroutines
3694 since in that case saving an insn makes more difference
3695 and more registers are available. */
3696 /* ??? could set this to last value of threshold in move_movables */
3697 int threshold = (loop_has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3698 /* Map of pseudo-register replacements. */
3703 rtx end_insert_before;
3705 int n_extra_increment;
3706 struct loop_info loop_iteration_info;
3707 struct loop_info *loop_info = &loop_iteration_info;
3708 int unrolled_insn_copies;
3710 /* If scan_start points to the loop exit test, we have to be wary of
3711 subversive use of gotos inside expression statements. */
3712 if (prev_nonnote_insn (scan_start) != prev_nonnote_insn (loop_start))
3713 maybe_multiple = back_branch_in_range_p (scan_start, loop_start, loop_end);
3715 VARRAY_INT_INIT (reg_iv_type, max_reg_before_loop, "reg_iv_type");
3716 VARRAY_GENERIC_PTR_INIT (reg_iv_info, max_reg_before_loop, "reg_iv_info");
3717 reg_biv_class = (struct iv_class **)
3718 alloca (max_reg_before_loop * sizeof (struct iv_class *));
3719 bzero ((char *) reg_biv_class, (max_reg_before_loop
3720 * sizeof (struct iv_class *)));
3723 addr_placeholder = gen_reg_rtx (Pmode);
3725 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3726 must be put before this insn, so that they will appear in the right
3727 order (i.e. loop order).
3729 If loop_end is the end of the current function, then emit a
3730 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3732 if (NEXT_INSN (loop_end) != 0)
3733 end_insert_before = NEXT_INSN (loop_end);
3735 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3737 /* Scan through loop to find all possible bivs. */
3739 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
3741 p = next_insn_in_loop (p, scan_start, end, loop_top))
3743 if (GET_CODE (p) == INSN
3744 && (set = single_set (p))
3745 && GET_CODE (SET_DEST (set)) == REG)
3747 dest_reg = SET_DEST (set);
3748 if (REGNO (dest_reg) < max_reg_before_loop
3749 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3750 && REG_IV_TYPE (REGNO (dest_reg)) != NOT_BASIC_INDUCT)
3752 if (basic_induction_var (SET_SRC (set), GET_MODE (SET_SRC (set)),
3753 dest_reg, p, &inc_val, &mult_val,
3756 /* It is a possible basic induction variable.
3757 Create and initialize an induction structure for it. */
3760 = (struct induction *) alloca (sizeof (struct induction));
3762 record_biv (v, p, dest_reg, inc_val, mult_val, location,
3763 not_every_iteration, maybe_multiple);
3764 REG_IV_TYPE (REGNO (dest_reg)) = BASIC_INDUCT;
3766 else if (REGNO (dest_reg) < max_reg_before_loop)
3767 REG_IV_TYPE (REGNO (dest_reg)) = NOT_BASIC_INDUCT;
3771 /* Past CODE_LABEL, we get to insns that may be executed multiple
3772 times. The only way we can be sure that they can't is if every
3773 jump insn between here and the end of the loop either
3774 returns, exits the loop, is a jump to a location that is still
3775 behind the label, or is a jump to the loop start. */
3777 if (GET_CODE (p) == CODE_LABEL)
3785 insn = NEXT_INSN (insn);
3786 if (insn == scan_start)
3794 if (insn == scan_start)
3798 if (GET_CODE (insn) == JUMP_INSN
3799 && GET_CODE (PATTERN (insn)) != RETURN
3800 && (! condjump_p (insn)
3801 || (JUMP_LABEL (insn) != 0
3802 && JUMP_LABEL (insn) != scan_start
3803 && ! loop_insn_first_p (p, JUMP_LABEL (insn)))))
3811 /* Past a jump, we get to insns for which we can't count
3812 on whether they will be executed during each iteration. */
3813 /* This code appears twice in strength_reduce. There is also similar
3814 code in scan_loop. */
3815 if (GET_CODE (p) == JUMP_INSN
3816 /* If we enter the loop in the middle, and scan around to the
3817 beginning, don't set not_every_iteration for that.
3818 This can be any kind of jump, since we want to know if insns
3819 will be executed if the loop is executed. */
3820 && ! (JUMP_LABEL (p) == loop_top
3821 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3822 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3826 /* If this is a jump outside the loop, then it also doesn't
3827 matter. Check to see if the target of this branch is on the
3828 loop_number_exits_labels list. */
3830 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3832 label = LABEL_NEXTREF (label))
3833 if (XEXP (label, 0) == JUMP_LABEL (p))
3837 not_every_iteration = 1;
3840 else if (GET_CODE (p) == NOTE)
3842 /* At the virtual top of a converted loop, insns are again known to
3843 be executed each iteration: logically, the loop begins here
3844 even though the exit code has been duplicated.
3846 Insns are also again known to be executed each iteration at
3847 the LOOP_CONT note. */
3848 if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
3849 || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
3851 not_every_iteration = 0;
3852 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3854 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3858 /* Note if we pass a loop latch. If we do, then we can not clear
3859 NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in
3860 a loop since a jump before the last CODE_LABEL may have started
3861 a new loop iteration.
3863 Note that LOOP_TOP is only set for rotated loops and we need
3864 this check for all loops, so compare against the CODE_LABEL
3865 which immediately follows LOOP_START. */
3866 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == NEXT_INSN (loop_start))
3867 past_loop_latch = 1;
3869 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3870 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3871 or not an insn is known to be executed each iteration of the
3872 loop, whether or not any iterations are known to occur.
3874 Therefore, if we have just passed a label and have no more labels
3875 between here and the test insn of the loop, and we have not passed
3876 a jump to the top of the loop, then we know these insns will be
3877 executed each iteration. */
3879 if (not_every_iteration
3880 && ! past_loop_latch
3881 && GET_CODE (p) == CODE_LABEL
3882 && no_labels_between_p (p, loop_end)
3883 && loop_insn_first_p (p, loop_cont))
3884 not_every_iteration = 0;
3887 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3888 Make a sanity check against n_times_set. */
3889 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3891 if (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3892 /* Above happens if register modified by subreg, etc. */
3893 /* Make sure it is not recognized as a basic induction var: */
3894 || VARRAY_INT (n_times_set, bl->regno) != bl->biv_count
3895 /* If never incremented, it is invariant that we decided not to
3896 move. So leave it alone. */
3897 || ! bl->incremented)
3899 if (loop_dump_stream)
3900 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3902 (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3903 ? "not induction variable"
3904 : (! bl->incremented ? "never incremented"
3907 REG_IV_TYPE (bl->regno) = NOT_BASIC_INDUCT;
3914 if (loop_dump_stream)
3915 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3919 /* Exit if there are no bivs. */
3922 /* Can still unroll the loop anyways, but indicate that there is no
3923 strength reduction info available. */
3925 unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
3931 /* Find initial value for each biv by searching backwards from loop_start,
3932 halting at first label. Also record any test condition. */
3935 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3939 if (GET_CODE (p) == CALL_INSN)
3942 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3943 || GET_CODE (p) == CALL_INSN)
3944 note_stores (PATTERN (p), record_initial);
3946 /* Record any test of a biv that branches around the loop if no store
3947 between it and the start of loop. We only care about tests with
3948 constants and registers and only certain of those. */
3949 if (GET_CODE (p) == JUMP_INSN
3950 && JUMP_LABEL (p) != 0
3951 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3952 && (test = get_condition_for_loop (p)) != 0
3953 && GET_CODE (XEXP (test, 0)) == REG
3954 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3955 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3956 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3957 && bl->init_insn == 0)
3959 /* If an NE test, we have an initial value! */
3960 if (GET_CODE (test) == NE)
3963 bl->init_set = gen_rtx_SET (VOIDmode,
3964 XEXP (test, 0), XEXP (test, 1));
3967 bl->initial_test = test;
3971 /* Look at the each biv and see if we can say anything better about its
3972 initial value from any initializing insns set up above. (This is done
3973 in two passes to avoid missing SETs in a PARALLEL.) */
3974 for (backbl = &loop_iv_list; (bl = *backbl); backbl = &bl->next)
3979 if (! bl->init_insn)
3982 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
3983 is a constant, use the value of that. */
3984 if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL
3985 && CONSTANT_P (XEXP (note, 0)))
3986 || ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL
3987 && CONSTANT_P (XEXP (note, 0))))
3988 src = XEXP (note, 0);
3990 src = SET_SRC (bl->init_set);
3992 if (loop_dump_stream)
3993 fprintf (loop_dump_stream,
3994 "Biv %d initialized at insn %d: initial value ",
3995 bl->regno, INSN_UID (bl->init_insn));
3997 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
3998 || GET_MODE (src) == VOIDmode)
3999 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
4001 bl->initial_value = src;
4003 if (loop_dump_stream)
4005 if (GET_CODE (src) == CONST_INT)
4007 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (src));
4008 fputc ('\n', loop_dump_stream);
4012 print_rtl (loop_dump_stream, src);
4013 fprintf (loop_dump_stream, "\n");
4019 struct iv_class *bl2 = 0;
4020 rtx increment = NULL_RTX;
4022 /* Biv initial value is not a simple move. If it is the sum of
4023 another biv and a constant, check if both bivs are incremented
4024 in lockstep. Then we are actually looking at a giv.
4025 For simplicity, we only handle the case where there is but a
4026 single increment, and the register is not used elsewhere. */
4027 if (bl->biv_count == 1
4028 && bl->regno < max_reg_before_loop
4029 && uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4030 && GET_CODE (src) == PLUS
4031 && GET_CODE (XEXP (src, 0)) == REG
4032 && CONSTANT_P (XEXP (src, 1))
4033 && ((increment = biv_total_increment (bl, loop_start, loop_end))
4036 int regno = REGNO (XEXP (src, 0));
4038 for (bl2 = loop_iv_list; bl2; bl2 = bl2->next)
4039 if (bl2->regno == regno)
4043 /* Now, can we transform this biv into a giv? */
4045 && bl2->biv_count == 1
4046 && rtx_equal_p (increment,
4047 biv_total_increment (bl2, loop_start, loop_end))
4048 /* init_insn is only set to insns that are before loop_start
4049 without any intervening labels. */
4050 && ! reg_set_between_p (bl2->biv->src_reg,
4051 PREV_INSN (bl->init_insn), loop_start)
4052 /* The register from BL2 must be set before the register from
4053 BL is set, or we must be able to move the latter set after
4054 the former set. Currently there can't be any labels
4055 in-between when biv_toal_increment returns nonzero both times
4056 but we test it here in case some day some real cfg analysis
4057 gets used to set always_computable. */
4058 && ((loop_insn_first_p (bl2->biv->insn, bl->biv->insn)
4059 && no_labels_between_p (bl2->biv->insn, bl->biv->insn))
4060 || (! reg_used_between_p (bl->biv->src_reg, bl->biv->insn,
4062 && no_jumps_between_p (bl->biv->insn, bl2->biv->insn)))
4063 && validate_change (bl->biv->insn,
4064 &SET_SRC (single_set (bl->biv->insn)),
4067 int loop_num = uid_loop_num[INSN_UID (loop_start)];
4068 rtx dominator = loop_number_cont_dominator[loop_num];
4069 rtx giv = bl->biv->src_reg;
4070 rtx giv_insn = bl->biv->insn;
4071 rtx after_giv = NEXT_INSN (giv_insn);
4073 if (loop_dump_stream)
4074 fprintf (loop_dump_stream, "is giv of biv %d\n", bl2->regno);
4075 /* Let this giv be discovered by the generic code. */
4076 REG_IV_TYPE (bl->regno) = UNKNOWN_INDUCT;
4077 reg_biv_class[bl->regno] = NULL_PTR;
4078 /* We can get better optimization if we can move the giv setting
4079 before the first giv use. */
4081 && ! loop_insn_first_p (dominator, scan_start)
4082 && ! reg_set_between_p (bl2->biv->src_reg, loop_start,
4084 && ! reg_used_between_p (giv, loop_start, dominator)
4085 && ! reg_used_between_p (giv, giv_insn, loop_end))
4090 for (next = NEXT_INSN (dominator); ; next = NEXT_INSN (next))
4092 if ((GET_RTX_CLASS (GET_CODE (next)) == 'i'
4093 && (reg_mentioned_p (giv, PATTERN (next))
4094 || reg_set_p (bl2->biv->src_reg, next)))
4095 || GET_CODE (next) == JUMP_INSN)
4098 if (GET_RTX_CLASS (GET_CODE (next)) != 'i'
4099 || ! sets_cc0_p (PATTERN (next)))
4103 if (loop_dump_stream)
4104 fprintf (loop_dump_stream, "move after insn %d\n",
4105 INSN_UID (dominator));
4106 /* Avoid problems with luids by actually moving the insn
4107 and adjusting all luids in the range. */
4108 reorder_insns (giv_insn, giv_insn, dominator);
4109 for (p = dominator; INSN_UID (p) >= max_uid_for_loop; )
4111 compute_luids (giv_insn, after_giv, INSN_LUID (p));
4112 /* If the only purpose of the init insn is to initialize
4113 this giv, delete it. */
4114 if (single_set (bl->init_insn)
4115 && ! reg_used_between_p (giv, bl->init_insn, loop_start))
4116 delete_insn (bl->init_insn);
4118 else if (! loop_insn_first_p (bl2->biv->insn, bl->biv->insn))
4120 rtx p = PREV_INSN (giv_insn);
4121 while (INSN_UID (p) >= max_uid_for_loop)
4123 reorder_insns (giv_insn, giv_insn, bl2->biv->insn);
4124 compute_luids (after_giv, NEXT_INSN (giv_insn),
4127 /* Remove this biv from the chain. */
4130 /* We move the following giv from *bl->next into *bl.
4131 We have to update reg_biv_class for that moved biv
4132 to point to its new address. */
4134 reg_biv_class[bl->regno] = bl;
4143 /* If we can't make it a giv,
4144 let biv keep initial value of "itself". */
4145 else if (loop_dump_stream)
4146 fprintf (loop_dump_stream, "is complex\n");
4150 /* If a biv is unconditionally incremented several times in a row, convert
4151 all but the last increment into a giv. */
4153 /* Get an upper bound for the number of registers
4154 we might have after all bivs have been processed. */
4155 first_increment_giv = max_reg_num ();
4156 for (n_extra_increment = 0, bl = loop_iv_list; bl; bl = bl->next)
4157 n_extra_increment += bl->biv_count - 1;
4159 /* If the loop contains volatile memory references do not allow any
4160 replacements to take place, since this could loose the volatile markers. */
4161 if (n_extra_increment && ! loop_has_volatile)
4163 int nregs = first_increment_giv + n_extra_increment;
4165 /* Reallocate reg_iv_type and reg_iv_info. */
4166 VARRAY_GROW (reg_iv_type, nregs);
4167 VARRAY_GROW (reg_iv_info, nregs);
4169 for (bl = loop_iv_list; bl; bl = bl->next)
4171 struct induction **vp, *v, *next;
4172 int biv_dead_after_loop = 0;
4174 /* The biv increments lists are in reverse order. Fix this first. */
4175 for (v = bl->biv, bl->biv = 0; v; v = next)
4178 v->next_iv = bl->biv;
4182 /* We must guard against the case that an early exit between v->insn
4183 and next->insn leaves the biv live after the loop, since that
4184 would mean that we'd be missing an increment for the final
4185 value. The following test to set biv_dead_after_loop is like
4186 the first part of the test to set bl->eliminable.
4187 We don't check here if we can calculate the final value, since
4188 this can't succeed if we already know that there is a jump
4189 between v->insn and next->insn, yet next->always_executed is
4190 set and next->maybe_multiple is cleared. Such a combination
4191 implies that the jump destination is outside the loop.
4192 If we want to make this check more sophisticated, we should
4193 check each branch between v->insn and next->insn individually
4194 to see if the biv is dead at its destination. */
4196 if (uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4198 && INSN_UID (bl->init_insn) < max_uid_for_loop
4199 && (uid_luid[REGNO_FIRST_UID (bl->regno)]
4200 >= INSN_LUID (bl->init_insn))
4201 #ifdef HAVE_decrement_and_branch_until_zero
4204 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4205 biv_dead_after_loop = 1;
4207 for (vp = &bl->biv, next = *vp; v = next, next = v->next_iv;)
4209 HOST_WIDE_INT offset;
4210 rtx set, add_val, old_reg, dest_reg, last_use_insn, note;
4211 int old_regno, new_regno;
4213 if (! v->always_executed
4214 || v->maybe_multiple
4215 || GET_CODE (v->add_val) != CONST_INT
4216 || ! next->always_executed
4217 || next->maybe_multiple
4218 || ! CONSTANT_P (next->add_val)
4219 || v->mult_val != const1_rtx
4220 || next->mult_val != const1_rtx
4221 || ! (biv_dead_after_loop
4222 || no_jumps_between_p (v->insn, next->insn)))
4227 offset = INTVAL (v->add_val);
4228 set = single_set (v->insn);
4229 add_val = plus_constant (next->add_val, offset);
4230 old_reg = v->dest_reg;
4231 dest_reg = gen_reg_rtx (v->mode);
4233 /* Unlike reg_iv_type / reg_iv_info, the other three arrays
4234 have been allocated with some slop space, so we may not
4235 actually need to reallocate them. If we do, the following
4236 if statement will be executed just once in this loop. */
4237 if ((unsigned) max_reg_num () > n_times_set->num_elements)
4239 /* Grow all the remaining arrays. */
4240 VARRAY_GROW (set_in_loop, nregs);
4241 VARRAY_GROW (n_times_set, nregs);
4242 VARRAY_GROW (may_not_optimize, nregs);
4243 VARRAY_GROW (reg_single_usage, nregs);
4246 if (! validate_change (next->insn, next->location, add_val, 0))
4252 /* Here we can try to eliminate the increment by combining
4253 it into the uses. */
4255 /* Set last_use_insn so that we can check against it. */
4257 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4259 p = next_insn_in_loop (p, scan_start, end, loop_top))
4261 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
4263 if (reg_mentioned_p (old_reg, PATTERN (p)))
4269 /* If we can't get the LUIDs for the insns, we can't
4270 calculate the lifetime. This is likely from unrolling
4271 of an inner loop, so there is little point in making this
4272 a DEST_REG giv anyways. */
4273 if (INSN_UID (v->insn) >= max_uid_for_loop
4274 || INSN_UID (last_use_insn) >= max_uid_for_loop
4275 || ! validate_change (v->insn, &SET_DEST (set), dest_reg, 0))
4277 /* Change the increment at NEXT back to what it was. */
4278 if (! validate_change (next->insn, next->location,
4284 next->add_val = add_val;
4285 v->dest_reg = dest_reg;
4286 v->giv_type = DEST_REG;
4287 v->location = &SET_SRC (set);
4289 v->combined_with = 0;
4291 v->derive_adjustment = 0;
4297 v->auto_inc_opt = 0;
4300 v->derived_from = 0;
4301 v->always_computable = 1;
4302 v->always_executed = 1;
4304 v->no_const_addval = 0;
4306 old_regno = REGNO (old_reg);
4307 new_regno = REGNO (dest_reg);
4308 VARRAY_INT (set_in_loop, old_regno)--;
4309 VARRAY_INT (set_in_loop, new_regno) = 1;
4310 VARRAY_INT (n_times_set, old_regno)--;
4311 VARRAY_INT (n_times_set, new_regno) = 1;
4312 VARRAY_CHAR (may_not_optimize, new_regno) = 0;
4314 REG_IV_TYPE (new_regno) = GENERAL_INDUCT;
4315 REG_IV_INFO (new_regno) = v;
4317 /* If next_insn has a REG_EQUAL note that mentiones OLD_REG,
4318 it must be replaced. */
4319 note = find_reg_note (next->insn, REG_EQUAL, NULL_RTX);
4320 if (note && reg_mentioned_p (old_reg, XEXP (note, 0)))
4321 XEXP (note, 0) = copy_rtx (SET_SRC (single_set (next->insn)));
4323 /* Remove the increment from the list of biv increments,
4324 and record it as a giv. */
4327 v->next_iv = bl->giv;
4330 v->benefit = rtx_cost (SET_SRC (set), SET);
4331 bl->total_benefit += v->benefit;
4333 /* Now replace the biv with DEST_REG in all insns between
4334 the replaced increment and the next increment, and
4335 remember the last insn that needed a replacement. */
4336 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4338 p = next_insn_in_loop (p, scan_start, end, loop_top))
4342 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
4344 if (reg_mentioned_p (old_reg, PATTERN (p)))
4347 if (! validate_replace_rtx (old_reg, dest_reg, p))
4350 for (note = REG_NOTES (p); note; note = XEXP (note, 1))
4352 if (GET_CODE (note) == EXPR_LIST)
4354 = replace_rtx (XEXP (note, 0), old_reg, dest_reg);
4358 v->last_use = last_use_insn;
4359 v->lifetime = INSN_LUID (v->insn) - INSN_LUID (last_use_insn);
4360 /* If the lifetime is zero, it means that this register is really
4361 a dead store. So mark this as a giv that can be ignored.
4362 This will not prevent the biv from being eliminated. */
4363 if (v->lifetime == 0)
4366 if (loop_dump_stream)
4367 fprintf (loop_dump_stream,
4368 "Increment %d of biv %d converted to giv %d.\n\n",
4369 INSN_UID (v->insn), old_regno, new_regno);
4373 last_increment_giv = max_reg_num () - 1;
4375 /* Search the loop for general induction variables. */
4377 /* A register is a giv if: it is only set once, it is a function of a
4378 biv and a constant (or invariant), and it is not a biv. */
4380 not_every_iteration = 0;
4386 /* At end of a straight-in loop, we are done.
4387 At end of a loop entered at the bottom, scan the top. */
4388 if (p == scan_start)
4396 if (p == scan_start)
4400 /* Look for a general induction variable in a register. */
4401 if (GET_CODE (p) == INSN
4402 && (set = single_set (p))
4403 && GET_CODE (SET_DEST (set)) == REG
4404 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
4411 rtx last_consec_insn;
4413 dest_reg = SET_DEST (set);
4414 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
4417 if (/* SET_SRC is a giv. */
4418 (general_induction_var (SET_SRC (set), &src_reg, &add_val,
4419 &mult_val, 0, &benefit)
4420 /* Equivalent expression is a giv. */
4421 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
4422 && general_induction_var (XEXP (regnote, 0), &src_reg,
4423 &add_val, &mult_val, 0,
4425 /* Don't try to handle any regs made by loop optimization.
4426 We have nothing on them in regno_first_uid, etc. */
4427 && REGNO (dest_reg) < max_reg_before_loop
4428 /* Don't recognize a BASIC_INDUCT_VAR here. */
4429 && dest_reg != src_reg
4430 /* This must be the only place where the register is set. */
4431 && (VARRAY_INT (n_times_set, REGNO (dest_reg)) == 1
4432 /* or all sets must be consecutive and make a giv. */
4433 || (benefit = consec_sets_giv (benefit, p,
4435 &add_val, &mult_val,
4436 &last_consec_insn))))
4439 = (struct induction *) alloca (sizeof (struct induction));
4441 /* If this is a library call, increase benefit. */
4442 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
4443 benefit += libcall_benefit (p);
4445 /* Skip the consecutive insns, if there are any. */
4446 if (VARRAY_INT (n_times_set, REGNO (dest_reg)) != 1)
4447 p = last_consec_insn;
4449 record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit,
4450 DEST_REG, not_every_iteration, NULL_PTR, loop_start,
4456 #ifndef DONT_REDUCE_ADDR
4457 /* Look for givs which are memory addresses. */
4458 /* This resulted in worse code on a VAX 8600. I wonder if it
4460 if (GET_CODE (p) == INSN)
4461 find_mem_givs (PATTERN (p), p, not_every_iteration, loop_start,
4465 /* Update the status of whether giv can derive other givs. This can
4466 change when we pass a label or an insn that updates a biv. */
4467 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4468 || GET_CODE (p) == CODE_LABEL)
4469 update_giv_derive (p);
4471 /* Past a jump, we get to insns for which we can't count
4472 on whether they will be executed during each iteration. */
4473 /* This code appears twice in strength_reduce. There is also similar
4474 code in scan_loop. */
4475 if (GET_CODE (p) == JUMP_INSN
4476 /* If we enter the loop in the middle, and scan around to the
4477 beginning, don't set not_every_iteration for that.
4478 This can be any kind of jump, since we want to know if insns
4479 will be executed if the loop is executed. */
4480 && ! (JUMP_LABEL (p) == loop_top
4481 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
4482 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
4486 /* If this is a jump outside the loop, then it also doesn't
4487 matter. Check to see if the target of this branch is on the
4488 loop_number_exits_labels list. */
4490 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
4492 label = LABEL_NEXTREF (label))
4493 if (XEXP (label, 0) == JUMP_LABEL (p))
4497 not_every_iteration = 1;
4500 else if (GET_CODE (p) == NOTE)
4502 /* At the virtual top of a converted loop, insns are again known to
4503 be executed each iteration: logically, the loop begins here
4504 even though the exit code has been duplicated.
4506 Insns are also again known to be executed each iteration at
4507 the LOOP_CONT note. */
4508 if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
4509 || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
4511 not_every_iteration = 0;
4512 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
4514 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
4518 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
4519 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
4520 or not an insn is known to be executed each iteration of the
4521 loop, whether or not any iterations are known to occur.
4523 Therefore, if we have just passed a label and have no more labels
4524 between here and the test insn of the loop, we know these insns
4525 will be executed each iteration. */
4527 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
4528 && no_labels_between_p (p, loop_end)
4529 && loop_insn_first_p (p, loop_cont))
4530 not_every_iteration = 0;
4533 /* Try to calculate and save the number of loop iterations. This is
4534 set to zero if the actual number can not be calculated. This must
4535 be called after all giv's have been identified, since otherwise it may
4536 fail if the iteration variable is a giv. */
4538 loop_iterations (loop_start, loop_end, loop_info);
4540 /* Now for each giv for which we still don't know whether or not it is
4541 replaceable, check to see if it is replaceable because its final value
4542 can be calculated. This must be done after loop_iterations is called,
4543 so that final_giv_value will work correctly. */
4545 for (bl = loop_iv_list; bl; bl = bl->next)
4547 struct induction *v;
4549 for (v = bl->giv; v; v = v->next_iv)
4550 if (! v->replaceable && ! v->not_replaceable)
4551 check_final_value (v, loop_start, loop_end, loop_info->n_iterations);
4554 /* Try to prove that the loop counter variable (if any) is always
4555 nonnegative; if so, record that fact with a REG_NONNEG note
4556 so that "decrement and branch until zero" insn can be used. */
4557 check_dbra_loop (loop_end, insn_count, loop_start, loop_info);
4559 /* Create reg_map to hold substitutions for replaceable giv regs.
4560 Some givs might have been made from biv increments, so look at
4561 reg_iv_type for a suitable size. */
4562 reg_map_size = reg_iv_type->num_elements;
4563 reg_map = (rtx *) alloca (reg_map_size * sizeof (rtx));
4564 bzero ((char *) reg_map, reg_map_size * sizeof (rtx));
4566 /* Examine each iv class for feasibility of strength reduction/induction
4567 variable elimination. */
4569 for (bl = loop_iv_list; bl; bl = bl->next)
4571 struct induction *v;
4574 rtx final_value = 0;
4577 /* Test whether it will be possible to eliminate this biv
4578 provided all givs are reduced. This is possible if either
4579 the reg is not used outside the loop, or we can compute
4580 what its final value will be.
4582 For architectures with a decrement_and_branch_until_zero insn,
4583 don't do this if we put a REG_NONNEG note on the endtest for
4586 /* Compare against bl->init_insn rather than loop_start.
4587 We aren't concerned with any uses of the biv between
4588 init_insn and loop_start since these won't be affected
4589 by the value of the biv elsewhere in the function, so
4590 long as init_insn doesn't use the biv itself.
4591 March 14, 1989 -- self@bayes.arc.nasa.gov */
4593 if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4595 && INSN_UID (bl->init_insn) < max_uid_for_loop
4596 && uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)
4597 #ifdef HAVE_decrement_and_branch_until_zero
4600 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4601 || ((final_value = final_biv_value (bl, loop_start, loop_end,
4602 loop_info->n_iterations))
4603 #ifdef HAVE_decrement_and_branch_until_zero
4607 bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0,
4608 threshold, insn_count);
4611 if (loop_dump_stream)
4613 fprintf (loop_dump_stream,
4614 "Cannot eliminate biv %d.\n",
4616 fprintf (loop_dump_stream,
4617 "First use: insn %d, last use: insn %d.\n",
4618 REGNO_FIRST_UID (bl->regno),
4619 REGNO_LAST_UID (bl->regno));
4623 /* Combine all giv's for this iv_class. */
4626 /* This will be true at the end, if all givs which depend on this
4627 biv have been strength reduced.
4628 We can't (currently) eliminate the biv unless this is so. */
4631 /* Check each giv in this class to see if we will benefit by reducing
4632 it. Skip giv's combined with others. */
4633 for (v = bl->giv; v; v = v->next_iv)
4635 struct induction *tv;
4637 if (v->ignore || v->same)
4640 benefit = v->benefit;
4642 /* Reduce benefit if not replaceable, since we will insert
4643 a move-insn to replace the insn that calculates this giv.
4644 Don't do this unless the giv is a user variable, since it
4645 will often be marked non-replaceable because of the duplication
4646 of the exit code outside the loop. In such a case, the copies
4647 we insert are dead and will be deleted. So they don't have
4648 a cost. Similar situations exist. */
4649 /* ??? The new final_[bg]iv_value code does a much better job
4650 of finding replaceable giv's, and hence this code may no longer
4652 if (! v->replaceable && ! bl->eliminable
4653 && REG_USERVAR_P (v->dest_reg))
4654 benefit -= copy_cost;
4656 /* Decrease the benefit to count the add-insns that we will
4657 insert to increment the reduced reg for the giv. */
4658 benefit -= add_cost * bl->biv_count;
4660 /* Decide whether to strength-reduce this giv or to leave the code
4661 unchanged (recompute it from the biv each time it is used).
4662 This decision can be made independently for each giv. */
4665 /* Attempt to guess whether autoincrement will handle some of the
4666 new add insns; if so, increase BENEFIT (undo the subtraction of
4667 add_cost that was done above). */
4668 if (v->giv_type == DEST_ADDR
4669 && GET_CODE (v->mult_val) == CONST_INT)
4671 if (HAVE_POST_INCREMENT
4672 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4673 benefit += add_cost * bl->biv_count;
4674 else if (HAVE_PRE_INCREMENT
4675 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4676 benefit += add_cost * bl->biv_count;
4677 else if (HAVE_POST_DECREMENT
4678 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4679 benefit += add_cost * bl->biv_count;
4680 else if (HAVE_PRE_DECREMENT
4681 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4682 benefit += add_cost * bl->biv_count;
4686 /* If an insn is not to be strength reduced, then set its ignore
4687 flag, and clear all_reduced. */
4689 /* A giv that depends on a reversed biv must be reduced if it is
4690 used after the loop exit, otherwise, it would have the wrong
4691 value after the loop exit. To make it simple, just reduce all
4692 of such giv's whether or not we know they are used after the loop
4695 if ( ! flag_reduce_all_givs && v->lifetime * threshold * benefit < insn_count
4698 if (loop_dump_stream)
4699 fprintf (loop_dump_stream,
4700 "giv of insn %d not worth while, %d vs %d.\n",
4702 v->lifetime * threshold * benefit, insn_count);
4708 /* Check that we can increment the reduced giv without a
4709 multiply insn. If not, reject it. */
4711 for (tv = bl->biv; tv; tv = tv->next_iv)
4712 if (tv->mult_val == const1_rtx
4713 && ! product_cheap_p (tv->add_val, v->mult_val))
4715 if (loop_dump_stream)
4716 fprintf (loop_dump_stream,
4717 "giv of insn %d: would need a multiply.\n",
4718 INSN_UID (v->insn));
4726 /* Check for givs whose first use is their definition and whose
4727 last use is the definition of another giv. If so, it is likely
4728 dead and should not be used to derive another giv nor to
4730 for (v = bl->giv; v; v = v->next_iv)
4733 || (v->same && v->same->ignore))
4738 struct induction *v1;
4740 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4741 if (v->last_use == v1->insn)
4744 else if (v->giv_type == DEST_REG
4745 && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
4747 struct induction *v1;
4749 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4750 if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
4755 /* Now that we know which givs will be reduced, try to rearrange the
4756 combinations to reduce register pressure.
4757 recombine_givs calls find_life_end, which needs reg_iv_type and
4758 reg_iv_info to be valid for all pseudos. We do the necessary
4759 reallocation here since it allows to check if there are still
4760 more bivs to process. */
4761 nregs = max_reg_num ();
4762 if (nregs > reg_iv_type->num_elements)
4764 /* If there are still more bivs to process, allocate some slack
4765 space so that we're not constantly reallocating these arrays. */
4768 /* Reallocate reg_iv_type and reg_iv_info. */
4769 VARRAY_GROW (reg_iv_type, nregs);
4770 VARRAY_GROW (reg_iv_info, nregs);
4772 recombine_givs (bl, loop_start, loop_end, unroll_p);
4774 /* Reduce each giv that we decided to reduce. */
4776 for (v = bl->giv; v; v = v->next_iv)
4778 struct induction *tv;
4779 if (! v->ignore && v->same == 0)
4781 int auto_inc_opt = 0;
4783 /* If the code for derived givs immediately below has already
4784 allocated a new_reg, we must keep it. */
4786 v->new_reg = gen_reg_rtx (v->mode);
4788 if (v->derived_from)
4790 struct induction *d = v->derived_from;
4792 /* In case d->dest_reg is not replaceable, we have
4793 to replace it in v->insn now. */
4795 d->new_reg = gen_reg_rtx (d->mode);
4797 = replace_rtx (PATTERN (v->insn), d->dest_reg, d->new_reg);
4799 = replace_rtx (PATTERN (v->insn), v->dest_reg, v->new_reg);
4800 /* For each place where the biv is incremented, add an
4801 insn to set the new, reduced reg for the giv.
4802 We used to do this only for biv_count != 1, but
4803 this fails when there is a giv after a single biv
4804 increment, e.g. when the last giv was expressed as
4806 for (tv = bl->biv; tv; tv = tv->next_iv)
4808 /* We always emit reduced giv increments before the
4809 biv increment when bl->biv_count != 1. So by
4810 emitting the add insns for derived givs after the
4811 biv increment, they pick up the updated value of
4813 If the reduced giv is processed with
4814 auto_inc_opt == 1, then it is incremented earlier
4815 than the biv, hence we'll still pick up the right
4817 If it's processed with auto_inc_opt == -1,
4818 that implies that the biv increment is before the
4819 first reduced giv's use. The derived giv's lifetime
4820 is after the reduced giv's lifetime, hence in this
4821 case, the biv increment doesn't matter. */
4822 emit_insn_after (copy_rtx (PATTERN (v->insn)), tv->insn);
4828 /* If the target has auto-increment addressing modes, and
4829 this is an address giv, then try to put the increment
4830 immediately after its use, so that flow can create an
4831 auto-increment addressing mode. */
4832 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
4833 && bl->biv->always_executed && ! bl->biv->maybe_multiple
4834 /* We don't handle reversed biv's because bl->biv->insn
4835 does not have a valid INSN_LUID. */
4837 && v->always_executed && ! v->maybe_multiple
4838 && INSN_UID (v->insn) < max_uid_for_loop)
4840 /* If other giv's have been combined with this one, then
4841 this will work only if all uses of the other giv's occur
4842 before this giv's insn. This is difficult to check.
4844 We simplify this by looking for the common case where
4845 there is one DEST_REG giv, and this giv's insn is the
4846 last use of the dest_reg of that DEST_REG giv. If the
4847 increment occurs after the address giv, then we can
4848 perform the optimization. (Otherwise, the increment
4849 would have to go before other_giv, and we would not be
4850 able to combine it with the address giv to get an
4851 auto-inc address.) */
4852 if (v->combined_with)
4854 struct induction *other_giv = 0;
4856 for (tv = bl->giv; tv; tv = tv->next_iv)
4864 if (! tv && other_giv
4865 && REGNO (other_giv->dest_reg) < max_reg_before_loop
4866 && (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
4867 == INSN_UID (v->insn))
4868 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
4871 /* Check for case where increment is before the address
4872 giv. Do this test in "loop order". */
4873 else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn)
4874 && (INSN_LUID (v->insn) < INSN_LUID (scan_start)
4875 || (INSN_LUID (bl->biv->insn)
4876 > INSN_LUID (scan_start))))
4877 || (INSN_LUID (v->insn) < INSN_LUID (scan_start)
4878 && (INSN_LUID (scan_start)
4879 < INSN_LUID (bl->biv->insn))))
4888 /* We can't put an insn immediately after one setting
4889 cc0, or immediately before one using cc0. */
4890 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
4891 || (auto_inc_opt == -1
4892 && (prev = prev_nonnote_insn (v->insn)) != 0
4893 && GET_RTX_CLASS (GET_CODE (prev)) == 'i'
4894 && sets_cc0_p (PATTERN (prev))))
4900 v->auto_inc_opt = 1;
4904 /* For each place where the biv is incremented, add an insn
4905 to increment the new, reduced reg for the giv. */
4906 for (tv = bl->biv; tv; tv = tv->next_iv)
4911 insert_before = tv->insn;
4912 else if (auto_inc_opt == 1)
4913 insert_before = NEXT_INSN (v->insn);
4915 insert_before = v->insn;
4917 if (tv->mult_val == const1_rtx)
4918 emit_iv_add_mult (tv->add_val, v->mult_val,
4919 v->new_reg, v->new_reg, insert_before);
4920 else /* tv->mult_val == const0_rtx */
4921 /* A multiply is acceptable here
4922 since this is presumed to be seldom executed. */
4923 emit_iv_add_mult (tv->add_val, v->mult_val,
4924 v->add_val, v->new_reg, insert_before);
4927 /* Add code at loop start to initialize giv's reduced reg. */
4929 emit_iv_add_mult (bl->initial_value, v->mult_val,
4930 v->add_val, v->new_reg, loop_start);
4934 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4937 For each giv register that can be reduced now: if replaceable,
4938 substitute reduced reg wherever the old giv occurs;
4939 else add new move insn "giv_reg = reduced_reg". */
4941 for (v = bl->giv; v; v = v->next_iv)
4943 if (v->same && v->same->ignore)
4949 /* Update expression if this was combined, in case other giv was
4952 v->new_reg = replace_rtx (v->new_reg,
4953 v->same->dest_reg, v->same->new_reg);
4955 if (v->giv_type == DEST_ADDR)
4956 /* Store reduced reg as the address in the memref where we found
4958 validate_change (v->insn, v->location, v->new_reg, 0);
4959 else if (v->replaceable)
4961 reg_map[REGNO (v->dest_reg)] = v->new_reg;
4964 /* I can no longer duplicate the original problem. Perhaps
4965 this is unnecessary now? */
4967 /* Replaceable; it isn't strictly necessary to delete the old
4968 insn and emit a new one, because v->dest_reg is now dead.
4970 However, especially when unrolling loops, the special
4971 handling for (set REG0 REG1) in the second cse pass may
4972 make v->dest_reg live again. To avoid this problem, emit
4973 an insn to set the original giv reg from the reduced giv.
4974 We can not delete the original insn, since it may be part
4975 of a LIBCALL, and the code in flow that eliminates dead
4976 libcalls will fail if it is deleted. */
4977 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4983 /* Not replaceable; emit an insn to set the original giv reg from
4984 the reduced giv, same as above. */
4985 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4989 /* When a loop is reversed, givs which depend on the reversed
4990 biv, and which are live outside the loop, must be set to their
4991 correct final value. This insn is only needed if the giv is
4992 not replaceable. The correct final value is the same as the
4993 value that the giv starts the reversed loop with. */
4994 if (bl->reversed && ! v->replaceable)
4995 emit_iv_add_mult (bl->initial_value, v->mult_val,
4996 v->add_val, v->dest_reg, end_insert_before);
4997 else if (v->final_value)
5001 /* If the loop has multiple exits, emit the insn before the
5002 loop to ensure that it will always be executed no matter
5003 how the loop exits. Otherwise, emit the insn after the loop,
5004 since this is slightly more efficient. */
5005 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
5006 insert_before = loop_start;
5008 insert_before = end_insert_before;
5009 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
5013 /* If the insn to set the final value of the giv was emitted
5014 before the loop, then we must delete the insn inside the loop
5015 that sets it. If this is a LIBCALL, then we must delete
5016 every insn in the libcall. Note, however, that
5017 final_giv_value will only succeed when there are multiple
5018 exits if the giv is dead at each exit, hence it does not
5019 matter that the original insn remains because it is dead
5021 /* Delete the insn inside the loop that sets the giv since
5022 the giv is now set before (or after) the loop. */
5023 delete_insn (v->insn);
5027 if (loop_dump_stream)
5029 fprintf (loop_dump_stream, "giv at %d reduced to ",
5030 INSN_UID (v->insn));
5031 print_rtl (loop_dump_stream, v->new_reg);
5032 fprintf (loop_dump_stream, "\n");
5036 /* All the givs based on the biv bl have been reduced if they
5039 /* For each giv not marked as maybe dead that has been combined with a
5040 second giv, clear any "maybe dead" mark on that second giv.
5041 v->new_reg will either be or refer to the register of the giv it
5044 Doing this clearing avoids problems in biv elimination where a
5045 giv's new_reg is a complex value that can't be put in the insn but
5046 the giv combined with (with a reg as new_reg) is marked maybe_dead.
5047 Since the register will be used in either case, we'd prefer it be
5048 used from the simpler giv. */
5050 for (v = bl->giv; v; v = v->next_iv)
5051 if (! v->maybe_dead && v->same)
5052 v->same->maybe_dead = 0;
5054 /* Try to eliminate the biv, if it is a candidate.
5055 This won't work if ! all_reduced,
5056 since the givs we planned to use might not have been reduced.
5058 We have to be careful that we didn't initially think we could eliminate
5059 this biv because of a giv that we now think may be dead and shouldn't
5060 be used as a biv replacement.
5062 Also, there is the possibility that we may have a giv that looks
5063 like it can be used to eliminate a biv, but the resulting insn
5064 isn't valid. This can happen, for example, on the 88k, where a
5065 JUMP_INSN can compare a register only with zero. Attempts to
5066 replace it with a compare with a constant will fail.
5068 Note that in cases where this call fails, we may have replaced some
5069 of the occurrences of the biv with a giv, but no harm was done in
5070 doing so in the rare cases where it can occur. */
5072 if (all_reduced == 1 && bl->eliminable
5073 && maybe_eliminate_biv (bl, loop_start, end, 1,
5074 threshold, insn_count))
5077 /* ?? If we created a new test to bypass the loop entirely,
5078 or otherwise drop straight in, based on this test, then
5079 we might want to rewrite it also. This way some later
5080 pass has more hope of removing the initialization of this
5083 /* If final_value != 0, then the biv may be used after loop end
5084 and we must emit an insn to set it just in case.
5086 Reversed bivs already have an insn after the loop setting their
5087 value, so we don't need another one. We can't calculate the
5088 proper final value for such a biv here anyways. */
5089 if (final_value != 0 && ! bl->reversed)
5093 /* If the loop has multiple exits, emit the insn before the
5094 loop to ensure that it will always be executed no matter
5095 how the loop exits. Otherwise, emit the insn after the
5096 loop, since this is slightly more efficient. */
5097 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
5098 insert_before = loop_start;
5100 insert_before = end_insert_before;
5102 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
5107 /* Delete all of the instructions inside the loop which set
5108 the biv, as they are all dead. If is safe to delete them,
5109 because an insn setting a biv will never be part of a libcall. */
5110 /* However, deleting them will invalidate the regno_last_uid info,
5111 so keeping them around is more convenient. Final_biv_value
5112 will only succeed when there are multiple exits if the biv
5113 is dead at each exit, hence it does not matter that the original
5114 insn remains, because it is dead anyways. */
5115 for (v = bl->biv; v; v = v->next_iv)
5116 delete_insn (v->insn);
5119 if (loop_dump_stream)
5120 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
5125 /* Go through all the instructions in the loop, making all the
5126 register substitutions scheduled in REG_MAP. */
5128 for (p = loop_start; p != end; p = NEXT_INSN (p))
5129 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5130 || GET_CODE (p) == CALL_INSN)
5132 replace_regs (PATTERN (p), reg_map, reg_map_size, 0);
5133 replace_regs (REG_NOTES (p), reg_map, reg_map_size, 0);
5137 if (loop_info->n_iterations > 0)
5139 /* When we completely unroll a loop we will likely not need the increment
5140 of the loop BIV and we will not need the conditional branch at the
5142 unrolled_insn_copies = insn_count - 2;
5145 /* When we completely unroll a loop on a HAVE_cc0 machine we will not
5146 need the comparison before the conditional branch at the end of the
5148 unrolled_insn_copies -= 1;
5151 /* We'll need one copy for each loop iteration. */
5152 unrolled_insn_copies *= loop_info->n_iterations;
5154 /* A little slop to account for the ability to remove initialization
5155 code, better CSE, and other secondary benefits of completely
5156 unrolling some loops. */
5157 unrolled_insn_copies -= 1;
5159 /* Clamp the value. */
5160 if (unrolled_insn_copies < 0)
5161 unrolled_insn_copies = 0;
5164 /* Unroll loops from within strength reduction so that we can use the
5165 induction variable information that strength_reduce has already
5166 collected. Always unroll loops that would be as small or smaller
5167 unrolled than when rolled. */
5169 || (loop_info->n_iterations > 0
5170 && unrolled_insn_copies <= insn_count))
5171 unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
5174 #ifdef HAVE_decrement_and_branch_on_count
5175 /* Instrument the loop with BCT insn. */
5176 if (HAVE_decrement_and_branch_on_count && bct_p
5177 && flag_branch_on_count_reg)
5178 insert_bct (loop_start, loop_end, loop_info);
5179 #endif /* HAVE_decrement_and_branch_on_count */
5181 if (loop_dump_stream)
5182 fprintf (loop_dump_stream, "\n");
5183 VARRAY_FREE (reg_iv_type);
5184 VARRAY_FREE (reg_iv_info);
5187 /* Return 1 if X is a valid source for an initial value (or as value being
5188 compared against in an initial test).
5190 X must be either a register or constant and must not be clobbered between
5191 the current insn and the start of the loop.
5193 INSN is the insn containing X. */
5196 valid_initial_value_p (x, insn, call_seen, loop_start)
5205 /* Only consider pseudos we know about initialized in insns whose luids
5207 if (GET_CODE (x) != REG
5208 || REGNO (x) >= max_reg_before_loop)
5211 /* Don't use call-clobbered registers across a call which clobbers it. On
5212 some machines, don't use any hard registers at all. */
5213 if (REGNO (x) < FIRST_PSEUDO_REGISTER
5214 && (SMALL_REGISTER_CLASSES
5215 || (call_used_regs[REGNO (x)] && call_seen)))
5218 /* Don't use registers that have been clobbered before the start of the
5220 if (reg_set_between_p (x, insn, loop_start))
5226 /* Scan X for memory refs and check each memory address
5227 as a possible giv. INSN is the insn whose pattern X comes from.
5228 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
5229 every loop iteration. */
5232 find_mem_givs (x, insn, not_every_iteration, loop_start, loop_end)
5235 int not_every_iteration;
5236 rtx loop_start, loop_end;
5239 register enum rtx_code code;
5240 register const char *fmt;
5245 code = GET_CODE (x);
5269 /* This code used to disable creating GIVs with mult_val == 1 and
5270 add_val == 0. However, this leads to lost optimizations when
5271 it comes time to combine a set of related DEST_ADDR GIVs, since
5272 this one would not be seen. */
5274 if (general_induction_var (XEXP (x, 0), &src_reg, &add_val,
5275 &mult_val, 1, &benefit))
5277 /* Found one; record it. */
5279 = (struct induction *) oballoc (sizeof (struct induction));
5281 record_giv (v, insn, src_reg, addr_placeholder, mult_val,
5282 add_val, benefit, DEST_ADDR, not_every_iteration,
5283 &XEXP (x, 0), loop_start, loop_end);
5285 v->mem_mode = GET_MODE (x);
5294 /* Recursively scan the subexpressions for other mem refs. */
5296 fmt = GET_RTX_FORMAT (code);
5297 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5299 find_mem_givs (XEXP (x, i), insn, not_every_iteration, loop_start,
5301 else if (fmt[i] == 'E')
5302 for (j = 0; j < XVECLEN (x, i); j++)
5303 find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration,
5304 loop_start, loop_end);
5307 /* Fill in the data about one biv update.
5308 V is the `struct induction' in which we record the biv. (It is
5309 allocated by the caller, with alloca.)
5310 INSN is the insn that sets it.
5311 DEST_REG is the biv's reg.
5313 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
5314 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
5315 being set to INC_VAL.
5317 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
5318 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
5319 can be executed more than once per iteration. If MAYBE_MULTIPLE
5320 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
5321 executed exactly once per iteration. */
5324 record_biv (v, insn, dest_reg, inc_val, mult_val, location,
5325 not_every_iteration, maybe_multiple)
5326 struct induction *v;
5332 int not_every_iteration;
5335 struct iv_class *bl;
5338 v->src_reg = dest_reg;
5339 v->dest_reg = dest_reg;
5340 v->mult_val = mult_val;
5341 v->add_val = inc_val;
5342 v->location = location;
5343 v->mode = GET_MODE (dest_reg);
5344 v->always_computable = ! not_every_iteration;
5345 v->always_executed = ! not_every_iteration;
5346 v->maybe_multiple = maybe_multiple;
5348 /* Add this to the reg's iv_class, creating a class
5349 if this is the first incrementation of the reg. */
5351 bl = reg_biv_class[REGNO (dest_reg)];
5354 /* Create and initialize new iv_class. */
5356 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
5358 bl->regno = REGNO (dest_reg);
5364 /* Set initial value to the reg itself. */
5365 bl->initial_value = dest_reg;
5366 /* We haven't seen the initializing insn yet */
5369 bl->initial_test = 0;
5370 bl->incremented = 0;
5374 bl->total_benefit = 0;
5376 /* Add this class to loop_iv_list. */
5377 bl->next = loop_iv_list;
5380 /* Put it in the array of biv register classes. */
5381 reg_biv_class[REGNO (dest_reg)] = bl;
5384 /* Update IV_CLASS entry for this biv. */
5385 v->next_iv = bl->biv;
5388 if (mult_val == const1_rtx)
5389 bl->incremented = 1;
5391 if (loop_dump_stream)
5393 fprintf (loop_dump_stream,
5394 "Insn %d: possible biv, reg %d,",
5395 INSN_UID (insn), REGNO (dest_reg));
5396 if (GET_CODE (inc_val) == CONST_INT)
5398 fprintf (loop_dump_stream, " const =");
5399 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (inc_val));
5400 fputc ('\n', loop_dump_stream);
5404 fprintf (loop_dump_stream, " const = ");
5405 print_rtl (loop_dump_stream, inc_val);
5406 fprintf (loop_dump_stream, "\n");
5411 /* Fill in the data about one giv.
5412 V is the `struct induction' in which we record the giv. (It is
5413 allocated by the caller, with alloca.)
5414 INSN is the insn that sets it.
5415 BENEFIT estimates the savings from deleting this insn.
5416 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
5417 into a register or is used as a memory address.
5419 SRC_REG is the biv reg which the giv is computed from.
5420 DEST_REG is the giv's reg (if the giv is stored in a reg).
5421 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
5422 LOCATION points to the place where this giv's value appears in INSN. */
5425 record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
5426 type, not_every_iteration, location, loop_start, loop_end)
5427 struct induction *v;
5431 rtx mult_val, add_val;
5434 int not_every_iteration;
5436 rtx loop_start, loop_end;
5438 struct induction *b;
5439 struct iv_class *bl;
5440 rtx set = single_set (insn);
5443 v->src_reg = src_reg;
5445 v->dest_reg = dest_reg;
5446 v->mult_val = mult_val;
5447 v->add_val = add_val;
5448 v->benefit = benefit;
5449 v->location = location;
5451 v->combined_with = 0;
5452 v->maybe_multiple = 0;
5454 v->derive_adjustment = 0;
5460 v->auto_inc_opt = 0;
5463 v->derived_from = 0;
5466 /* The v->always_computable field is used in update_giv_derive, to
5467 determine whether a giv can be used to derive another giv. For a
5468 DEST_REG giv, INSN computes a new value for the giv, so its value
5469 isn't computable if INSN insn't executed every iteration.
5470 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
5471 it does not compute a new value. Hence the value is always computable
5472 regardless of whether INSN is executed each iteration. */
5474 if (type == DEST_ADDR)
5475 v->always_computable = 1;
5477 v->always_computable = ! not_every_iteration;
5479 v->always_executed = ! not_every_iteration;
5481 if (type == DEST_ADDR)
5483 v->mode = GET_MODE (*location);
5486 else /* type == DEST_REG */
5488 v->mode = GET_MODE (SET_DEST (set));
5490 v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5491 - uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]);
5493 /* If the lifetime is zero, it means that this register is
5494 really a dead store. So mark this as a giv that can be
5495 ignored. This will not prevent the biv from being eliminated. */
5496 if (v->lifetime == 0)
5499 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
5500 REG_IV_INFO (REGNO (dest_reg)) = v;
5503 /* Add the giv to the class of givs computed from one biv. */
5505 bl = reg_biv_class[REGNO (src_reg)];
5508 v->next_iv = bl->giv;
5510 /* Don't count DEST_ADDR. This is supposed to count the number of
5511 insns that calculate givs. */
5512 if (type == DEST_REG)
5514 bl->total_benefit += benefit;
5517 /* Fatal error, biv missing for this giv? */
5520 if (type == DEST_ADDR)
5524 /* The giv can be replaced outright by the reduced register only if all
5525 of the following conditions are true:
5526 - the insn that sets the giv is always executed on any iteration
5527 on which the giv is used at all
5528 (there are two ways to deduce this:
5529 either the insn is executed on every iteration,
5530 or all uses follow that insn in the same basic block),
5531 - the giv is not used outside the loop
5532 - no assignments to the biv occur during the giv's lifetime. */
5534 if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
5535 /* Previous line always fails if INSN was moved by loop opt. */
5536 && uid_luid[REGNO_LAST_UID (REGNO (dest_reg))] < INSN_LUID (loop_end)
5537 && (! not_every_iteration
5538 || last_use_this_basic_block (dest_reg, insn)))
5540 /* Now check that there are no assignments to the biv within the
5541 giv's lifetime. This requires two separate checks. */
5543 /* Check each biv update, and fail if any are between the first
5544 and last use of the giv.
5546 If this loop contains an inner loop that was unrolled, then
5547 the insn modifying the biv may have been emitted by the loop
5548 unrolling code, and hence does not have a valid luid. Just
5549 mark the biv as not replaceable in this case. It is not very
5550 useful as a biv, because it is used in two different loops.
5551 It is very unlikely that we would be able to optimize the giv
5552 using this biv anyways. */
5555 for (b = bl->biv; b; b = b->next_iv)
5557 if (INSN_UID (b->insn) >= max_uid_for_loop
5558 || ((uid_luid[INSN_UID (b->insn)]
5559 >= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))])
5560 && (uid_luid[INSN_UID (b->insn)]
5561 <= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))])))
5564 v->not_replaceable = 1;
5569 /* If there are any backwards branches that go from after the
5570 biv update to before it, then this giv is not replaceable. */
5572 for (b = bl->biv; b; b = b->next_iv)
5573 if (back_branch_in_range_p (b->insn, loop_start, loop_end))
5576 v->not_replaceable = 1;
5582 /* May still be replaceable, we don't have enough info here to
5585 v->not_replaceable = 0;
5589 /* Record whether the add_val contains a const_int, for later use by
5594 v->no_const_addval = 1;
5595 if (tem == const0_rtx)
5597 else if (GET_CODE (tem) == CONST_INT)
5598 v->no_const_addval = 0;
5599 else if (GET_CODE (tem) == PLUS)
5603 if (GET_CODE (XEXP (tem, 0)) == PLUS)
5604 tem = XEXP (tem, 0);
5605 else if (GET_CODE (XEXP (tem, 1)) == PLUS)
5606 tem = XEXP (tem, 1);
5610 if (GET_CODE (XEXP (tem, 1)) == CONST_INT)
5611 v->no_const_addval = 0;
5615 if (loop_dump_stream)
5617 if (type == DEST_REG)
5618 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
5619 INSN_UID (insn), REGNO (dest_reg));
5621 fprintf (loop_dump_stream, "Insn %d: dest address",
5624 fprintf (loop_dump_stream, " src reg %d benefit %d",
5625 REGNO (src_reg), v->benefit);
5626 fprintf (loop_dump_stream, " lifetime %d",
5630 fprintf (loop_dump_stream, " replaceable");
5632 if (v->no_const_addval)
5633 fprintf (loop_dump_stream, " ncav");
5635 if (GET_CODE (mult_val) == CONST_INT)
5637 fprintf (loop_dump_stream, " mult ");
5638 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (mult_val));
5642 fprintf (loop_dump_stream, " mult ");
5643 print_rtl (loop_dump_stream, mult_val);
5646 if (GET_CODE (add_val) == CONST_INT)
5648 fprintf (loop_dump_stream, " add ");
5649 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (add_val));
5653 fprintf (loop_dump_stream, " add ");
5654 print_rtl (loop_dump_stream, add_val);
5658 if (loop_dump_stream)
5659 fprintf (loop_dump_stream, "\n");
5664 /* All this does is determine whether a giv can be made replaceable because
5665 its final value can be calculated. This code can not be part of record_giv
5666 above, because final_giv_value requires that the number of loop iterations
5667 be known, and that can not be accurately calculated until after all givs
5668 have been identified. */
5671 check_final_value (v, loop_start, loop_end, n_iterations)
5672 struct induction *v;
5673 rtx loop_start, loop_end;
5674 unsigned HOST_WIDE_INT n_iterations;
5676 struct iv_class *bl;
5677 rtx final_value = 0;
5679 bl = reg_biv_class[REGNO (v->src_reg)];
5681 /* DEST_ADDR givs will never reach here, because they are always marked
5682 replaceable above in record_giv. */
5684 /* The giv can be replaced outright by the reduced register only if all
5685 of the following conditions are true:
5686 - the insn that sets the giv is always executed on any iteration
5687 on which the giv is used at all
5688 (there are two ways to deduce this:
5689 either the insn is executed on every iteration,
5690 or all uses follow that insn in the same basic block),
5691 - its final value can be calculated (this condition is different
5692 than the one above in record_giv)
5693 - no assignments to the biv occur during the giv's lifetime. */
5696 /* This is only called now when replaceable is known to be false. */
5697 /* Clear replaceable, so that it won't confuse final_giv_value. */
5701 if ((final_value = final_giv_value (v, loop_start, loop_end, n_iterations))
5702 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
5704 int biv_increment_seen = 0;
5710 /* When trying to determine whether or not a biv increment occurs
5711 during the lifetime of the giv, we can ignore uses of the variable
5712 outside the loop because final_value is true. Hence we can not
5713 use regno_last_uid and regno_first_uid as above in record_giv. */
5715 /* Search the loop to determine whether any assignments to the
5716 biv occur during the giv's lifetime. Start with the insn
5717 that sets the giv, and search around the loop until we come
5718 back to that insn again.
5720 Also fail if there is a jump within the giv's lifetime that jumps
5721 to somewhere outside the lifetime but still within the loop. This
5722 catches spaghetti code where the execution order is not linear, and
5723 hence the above test fails. Here we assume that the giv lifetime
5724 does not extend from one iteration of the loop to the next, so as
5725 to make the test easier. Since the lifetime isn't known yet,
5726 this requires two loops. See also record_giv above. */
5728 last_giv_use = v->insn;
5734 p = NEXT_INSN (loop_start);
5738 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5739 || GET_CODE (p) == CALL_INSN)
5741 if (biv_increment_seen)
5743 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5746 v->not_replaceable = 1;
5750 else if (reg_set_p (v->src_reg, PATTERN (p)))
5751 biv_increment_seen = 1;
5752 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5757 /* Now that the lifetime of the giv is known, check for branches
5758 from within the lifetime to outside the lifetime if it is still
5768 p = NEXT_INSN (loop_start);
5769 if (p == last_giv_use)
5772 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
5773 && LABEL_NAME (JUMP_LABEL (p))
5774 && ((loop_insn_first_p (JUMP_LABEL (p), v->insn)
5775 && loop_insn_first_p (loop_start, JUMP_LABEL (p)))
5776 || (loop_insn_first_p (last_giv_use, JUMP_LABEL (p))
5777 && loop_insn_first_p (JUMP_LABEL (p), loop_end))))
5780 v->not_replaceable = 1;
5782 if (loop_dump_stream)
5783 fprintf (loop_dump_stream,
5784 "Found branch outside giv lifetime.\n");
5791 /* If it is replaceable, then save the final value. */
5793 v->final_value = final_value;
5796 if (loop_dump_stream && v->replaceable)
5797 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
5798 INSN_UID (v->insn), REGNO (v->dest_reg));
5801 /* Update the status of whether a giv can derive other givs.
5803 We need to do something special if there is or may be an update to the biv
5804 between the time the giv is defined and the time it is used to derive
5807 In addition, a giv that is only conditionally set is not allowed to
5808 derive another giv once a label has been passed.
5810 The cases we look at are when a label or an update to a biv is passed. */
5813 update_giv_derive (p)
5816 struct iv_class *bl;
5817 struct induction *biv, *giv;
5821 /* Search all IV classes, then all bivs, and finally all givs.
5823 There are three cases we are concerned with. First we have the situation
5824 of a giv that is only updated conditionally. In that case, it may not
5825 derive any givs after a label is passed.
5827 The second case is when a biv update occurs, or may occur, after the
5828 definition of a giv. For certain biv updates (see below) that are
5829 known to occur between the giv definition and use, we can adjust the
5830 giv definition. For others, or when the biv update is conditional,
5831 we must prevent the giv from deriving any other givs. There are two
5832 sub-cases within this case.
5834 If this is a label, we are concerned with any biv update that is done
5835 conditionally, since it may be done after the giv is defined followed by
5836 a branch here (actually, we need to pass both a jump and a label, but
5837 this extra tracking doesn't seem worth it).
5839 If this is a jump, we are concerned about any biv update that may be
5840 executed multiple times. We are actually only concerned about
5841 backward jumps, but it is probably not worth performing the test
5842 on the jump again here.
5844 If this is a biv update, we must adjust the giv status to show that a
5845 subsequent biv update was performed. If this adjustment cannot be done,
5846 the giv cannot derive further givs. */
5848 for (bl = loop_iv_list; bl; bl = bl->next)
5849 for (biv = bl->biv; biv; biv = biv->next_iv)
5850 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
5853 for (giv = bl->giv; giv; giv = giv->next_iv)
5855 /* If cant_derive is already true, there is no point in
5856 checking all of these conditions again. */
5857 if (giv->cant_derive)
5860 /* If this giv is conditionally set and we have passed a label,
5861 it cannot derive anything. */
5862 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
5863 giv->cant_derive = 1;
5865 /* Skip givs that have mult_val == 0, since
5866 they are really invariants. Also skip those that are
5867 replaceable, since we know their lifetime doesn't contain
5869 else if (giv->mult_val == const0_rtx || giv->replaceable)
5872 /* The only way we can allow this giv to derive another
5873 is if this is a biv increment and we can form the product
5874 of biv->add_val and giv->mult_val. In this case, we will
5875 be able to compute a compensation. */
5876 else if (biv->insn == p)
5880 if (biv->mult_val == const1_rtx)
5881 tem = simplify_giv_expr (gen_rtx_MULT (giv->mode,
5886 if (tem && giv->derive_adjustment)
5887 tem = simplify_giv_expr (gen_rtx_PLUS (giv->mode, tem,
5888 giv->derive_adjustment),
5891 giv->derive_adjustment = tem;
5893 giv->cant_derive = 1;
5895 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
5896 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
5897 giv->cant_derive = 1;
5902 /* Check whether an insn is an increment legitimate for a basic induction var.
5903 X is the source of insn P, or a part of it.
5904 MODE is the mode in which X should be interpreted.
5906 DEST_REG is the putative biv, also the destination of the insn.
5907 We accept patterns of these forms:
5908 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
5909 REG = INVARIANT + REG
5911 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
5912 store the additive term into *INC_VAL, and store the place where
5913 we found the additive term into *LOCATION.
5915 If X is an assignment of an invariant into DEST_REG, we set
5916 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
5918 We also want to detect a BIV when it corresponds to a variable
5919 whose mode was promoted via PROMOTED_MODE. In that case, an increment
5920 of the variable may be a PLUS that adds a SUBREG of that variable to
5921 an invariant and then sign- or zero-extends the result of the PLUS
5924 Most GIVs in such cases will be in the promoted mode, since that is the
5925 probably the natural computation mode (and almost certainly the mode
5926 used for addresses) on the machine. So we view the pseudo-reg containing
5927 the variable as the BIV, as if it were simply incremented.
5929 Note that treating the entire pseudo as a BIV will result in making
5930 simple increments to any GIVs based on it. However, if the variable
5931 overflows in its declared mode but not its promoted mode, the result will
5932 be incorrect. This is acceptable if the variable is signed, since
5933 overflows in such cases are undefined, but not if it is unsigned, since
5934 those overflows are defined. So we only check for SIGN_EXTEND and
5937 If we cannot find a biv, we return 0. */
5940 basic_induction_var (x, mode, dest_reg, p, inc_val, mult_val, location)
5942 enum machine_mode mode;
5949 register enum rtx_code code;
5953 code = GET_CODE (x);
5957 if (rtx_equal_p (XEXP (x, 0), dest_reg)
5958 || (GET_CODE (XEXP (x, 0)) == SUBREG
5959 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
5960 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
5962 argp = &XEXP (x, 1);
5964 else if (rtx_equal_p (XEXP (x, 1), dest_reg)
5965 || (GET_CODE (XEXP (x, 1)) == SUBREG
5966 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
5967 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
5969 argp = &XEXP (x, 0);
5975 if (invariant_p (arg) != 1)
5978 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
5979 *mult_val = const1_rtx;
5984 /* If this is a SUBREG for a promoted variable, check the inner
5986 if (SUBREG_PROMOTED_VAR_P (x))
5987 return basic_induction_var (SUBREG_REG (x), GET_MODE (SUBREG_REG (x)),
5988 dest_reg, p, inc_val, mult_val, location);
5992 /* If this register is assigned in a previous insn, look at its
5993 source, but don't go outside the loop or past a label. */
5999 insn = PREV_INSN (insn);
6000 } while (insn && GET_CODE (insn) == NOTE
6001 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6005 set = single_set (insn);
6009 if ((SET_DEST (set) == x
6010 || (GET_CODE (SET_DEST (set)) == SUBREG
6011 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
6013 && SUBREG_REG (SET_DEST (set)) == x))
6014 && basic_induction_var (SET_SRC (set),
6015 (GET_MODE (SET_SRC (set)) == VOIDmode
6017 : GET_MODE (SET_SRC (set))),
6019 inc_val, mult_val, location))
6022 /* ... fall through ... */
6024 /* Can accept constant setting of biv only when inside inner most loop.
6025 Otherwise, a biv of an inner loop may be incorrectly recognized
6026 as a biv of the outer loop,
6027 causing code to be moved INTO the inner loop. */
6029 if (invariant_p (x) != 1)
6034 /* convert_modes aborts if we try to convert to or from CCmode, so just
6035 exclude that case. It is very unlikely that a condition code value
6036 would be a useful iterator anyways. */
6037 if (loops_enclosed == 1
6038 && GET_MODE_CLASS (mode) != MODE_CC
6039 && GET_MODE_CLASS (GET_MODE (dest_reg)) != MODE_CC)
6041 /* Possible bug here? Perhaps we don't know the mode of X. */
6042 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
6043 *mult_val = const0_rtx;
6050 return basic_induction_var (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
6051 dest_reg, p, inc_val, mult_val, location);
6054 /* Similar, since this can be a sign extension. */
6055 for (insn = PREV_INSN (p);
6056 (insn && GET_CODE (insn) == NOTE
6057 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6058 insn = PREV_INSN (insn))
6062 set = single_set (insn);
6064 if (set && SET_DEST (set) == XEXP (x, 0)
6065 && GET_CODE (XEXP (x, 1)) == CONST_INT
6066 && INTVAL (XEXP (x, 1)) >= 0
6067 && GET_CODE (SET_SRC (set)) == ASHIFT
6068 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
6069 return basic_induction_var (XEXP (SET_SRC (set), 0),
6070 GET_MODE (XEXP (x, 0)),
6071 dest_reg, insn, inc_val, mult_val,
6080 /* A general induction variable (giv) is any quantity that is a linear
6081 function of a basic induction variable,
6082 i.e. giv = biv * mult_val + add_val.
6083 The coefficients can be any loop invariant quantity.
6084 A giv need not be computed directly from the biv;
6085 it can be computed by way of other givs. */
6087 /* Determine whether X computes a giv.
6088 If it does, return a nonzero value
6089 which is the benefit from eliminating the computation of X;
6090 set *SRC_REG to the register of the biv that it is computed from;
6091 set *ADD_VAL and *MULT_VAL to the coefficients,
6092 such that the value of X is biv * mult + add; */
6095 general_induction_var (x, src_reg, add_val, mult_val, is_addr, pbenefit)
6106 /* If this is an invariant, forget it, it isn't a giv. */
6107 if (invariant_p (x) == 1)
6110 /* See if the expression could be a giv and get its form.
6111 Mark our place on the obstack in case we don't find a giv. */
6112 storage = (char *) oballoc (0);
6114 x = simplify_giv_expr (x, pbenefit);
6121 switch (GET_CODE (x))
6125 /* Since this is now an invariant and wasn't before, it must be a giv
6126 with MULT_VAL == 0. It doesn't matter which BIV we associate this
6128 *src_reg = loop_iv_list->biv->dest_reg;
6129 *mult_val = const0_rtx;
6134 /* This is equivalent to a BIV. */
6136 *mult_val = const1_rtx;
6137 *add_val = const0_rtx;
6141 /* Either (plus (biv) (invar)) or
6142 (plus (mult (biv) (invar_1)) (invar_2)). */
6143 if (GET_CODE (XEXP (x, 0)) == MULT)
6145 *src_reg = XEXP (XEXP (x, 0), 0);
6146 *mult_val = XEXP (XEXP (x, 0), 1);
6150 *src_reg = XEXP (x, 0);
6151 *mult_val = const1_rtx;
6153 *add_val = XEXP (x, 1);
6157 /* ADD_VAL is zero. */
6158 *src_reg = XEXP (x, 0);
6159 *mult_val = XEXP (x, 1);
6160 *add_val = const0_rtx;
6167 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
6168 unless they are CONST_INT). */
6169 if (GET_CODE (*add_val) == USE)
6170 *add_val = XEXP (*add_val, 0);
6171 if (GET_CODE (*mult_val) == USE)
6172 *mult_val = XEXP (*mult_val, 0);
6177 *pbenefit += ADDRESS_COST (orig_x) - reg_address_cost;
6179 *pbenefit += rtx_cost (orig_x, MEM) - reg_address_cost;
6183 *pbenefit += rtx_cost (orig_x, SET);
6185 /* Always return true if this is a giv so it will be detected as such,
6186 even if the benefit is zero or negative. This allows elimination
6187 of bivs that might otherwise not be eliminated. */
6191 /* Given an expression, X, try to form it as a linear function of a biv.
6192 We will canonicalize it to be of the form
6193 (plus (mult (BIV) (invar_1))
6195 with possible degeneracies.
6197 The invariant expressions must each be of a form that can be used as a
6198 machine operand. We surround then with a USE rtx (a hack, but localized
6199 and certainly unambiguous!) if not a CONST_INT for simplicity in this
6200 routine; it is the caller's responsibility to strip them.
6202 If no such canonicalization is possible (i.e., two biv's are used or an
6203 expression that is neither invariant nor a biv or giv), this routine
6206 For a non-zero return, the result will have a code of CONST_INT, USE,
6207 REG (for a BIV), PLUS, or MULT. No other codes will occur.
6209 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
6211 static rtx sge_plus PROTO ((enum machine_mode, rtx, rtx));
6212 static rtx sge_plus_constant PROTO ((rtx, rtx));
6215 simplify_giv_expr (x, benefit)
6219 enum machine_mode mode = GET_MODE (x);
6223 /* If this is not an integer mode, or if we cannot do arithmetic in this
6224 mode, this can't be a giv. */
6225 if (mode != VOIDmode
6226 && (GET_MODE_CLASS (mode) != MODE_INT
6227 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
6230 switch (GET_CODE (x))
6233 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
6234 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
6235 if (arg0 == 0 || arg1 == 0)
6238 /* Put constant last, CONST_INT last if both constant. */
6239 if ((GET_CODE (arg0) == USE
6240 || GET_CODE (arg0) == CONST_INT)
6241 && ! ((GET_CODE (arg0) == USE
6242 && GET_CODE (arg1) == USE)
6243 || GET_CODE (arg1) == CONST_INT))
6244 tem = arg0, arg0 = arg1, arg1 = tem;
6246 /* Handle addition of zero, then addition of an invariant. */
6247 if (arg1 == const0_rtx)
6249 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
6250 switch (GET_CODE (arg0))
6254 /* Adding two invariants must result in an invariant, so enclose
6255 addition operation inside a USE and return it. */
6256 if (GET_CODE (arg0) == USE)
6257 arg0 = XEXP (arg0, 0);
6258 if (GET_CODE (arg1) == USE)
6259 arg1 = XEXP (arg1, 0);
6261 if (GET_CODE (arg0) == CONST_INT)
6262 tem = arg0, arg0 = arg1, arg1 = tem;
6263 if (GET_CODE (arg1) == CONST_INT)
6264 tem = sge_plus_constant (arg0, arg1);
6266 tem = sge_plus (mode, arg0, arg1);
6268 if (GET_CODE (tem) != CONST_INT)
6269 tem = gen_rtx_USE (mode, tem);
6274 /* biv + invar or mult + invar. Return sum. */
6275 return gen_rtx_PLUS (mode, arg0, arg1);
6278 /* (a + invar_1) + invar_2. Associate. */
6279 return simplify_giv_expr (
6280 gen_rtx_PLUS (mode, XEXP (arg0, 0),
6281 gen_rtx_PLUS (mode, XEXP (arg0, 1), arg1)),
6288 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
6289 MULT to reduce cases. */
6290 if (GET_CODE (arg0) == REG)
6291 arg0 = gen_rtx_MULT (mode, arg0, const1_rtx);
6292 if (GET_CODE (arg1) == REG)
6293 arg1 = gen_rtx_MULT (mode, arg1, const1_rtx);
6295 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
6296 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
6297 Recurse to associate the second PLUS. */
6298 if (GET_CODE (arg1) == MULT)
6299 tem = arg0, arg0 = arg1, arg1 = tem;
6301 if (GET_CODE (arg1) == PLUS)
6302 return simplify_giv_expr (gen_rtx_PLUS (mode,
6303 gen_rtx_PLUS (mode, arg0,
6308 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
6309 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
6312 if (!rtx_equal_p (arg0, arg1))
6315 return simplify_giv_expr (gen_rtx_MULT (mode,
6323 /* Handle "a - b" as "a + b * (-1)". */
6324 return simplify_giv_expr (gen_rtx_PLUS (mode,
6326 gen_rtx_MULT (mode, XEXP (x, 1),
6331 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
6332 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
6333 if (arg0 == 0 || arg1 == 0)
6336 /* Put constant last, CONST_INT last if both constant. */
6337 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
6338 && GET_CODE (arg1) != CONST_INT)
6339 tem = arg0, arg0 = arg1, arg1 = tem;
6341 /* If second argument is not now constant, not giv. */
6342 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
6345 /* Handle multiply by 0 or 1. */
6346 if (arg1 == const0_rtx)
6349 else if (arg1 == const1_rtx)
6352 switch (GET_CODE (arg0))
6355 /* biv * invar. Done. */
6356 return gen_rtx_MULT (mode, arg0, arg1);
6359 /* Product of two constants. */
6360 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
6363 /* invar * invar. It is a giv, but very few of these will
6364 actually pay off, so limit to simple registers. */
6365 if (GET_CODE (arg1) != CONST_INT)
6368 arg0 = XEXP (arg0, 0);
6369 if (GET_CODE (arg0) == REG)
6370 tem = gen_rtx_MULT (mode, arg0, arg1);
6371 else if (GET_CODE (arg0) == MULT
6372 && GET_CODE (XEXP (arg0, 0)) == REG
6373 && GET_CODE (XEXP (arg0, 1)) == CONST_INT)
6375 tem = gen_rtx_MULT (mode, XEXP (arg0, 0),
6376 GEN_INT (INTVAL (XEXP (arg0, 1))
6381 return gen_rtx_USE (mode, tem);
6384 /* (a * invar_1) * invar_2. Associate. */
6385 return simplify_giv_expr (gen_rtx_MULT (mode, XEXP (arg0, 0),
6392 /* (a + invar_1) * invar_2. Distribute. */
6393 return simplify_giv_expr (gen_rtx_PLUS (mode,
6407 /* Shift by constant is multiply by power of two. */
6408 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6411 return simplify_giv_expr (gen_rtx_MULT (mode,
6413 GEN_INT ((HOST_WIDE_INT) 1
6414 << INTVAL (XEXP (x, 1)))),
6418 /* "-a" is "a * (-1)" */
6419 return simplify_giv_expr (gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx),
6423 /* "~a" is "-a - 1". Silly, but easy. */
6424 return simplify_giv_expr (gen_rtx_MINUS (mode,
6425 gen_rtx_NEG (mode, XEXP (x, 0)),
6430 /* Already in proper form for invariant. */
6434 /* If this is a new register, we can't deal with it. */
6435 if (REGNO (x) >= max_reg_before_loop)
6438 /* Check for biv or giv. */
6439 switch (REG_IV_TYPE (REGNO (x)))
6443 case GENERAL_INDUCT:
6445 struct induction *v = REG_IV_INFO (REGNO (x));
6447 /* Form expression from giv and add benefit. Ensure this giv
6448 can derive another and subtract any needed adjustment if so. */
6449 *benefit += v->benefit;
6453 tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode, v->src_reg,
6456 if (v->derive_adjustment)
6457 tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment);
6458 return simplify_giv_expr (tem, benefit);
6462 /* If it isn't an induction variable, and it is invariant, we
6463 may be able to simplify things further by looking through
6464 the bits we just moved outside the loop. */
6465 if (invariant_p (x) == 1)
6469 for (m = the_movables; m ; m = m->next)
6470 if (rtx_equal_p (x, m->set_dest))
6472 /* Ok, we found a match. Substitute and simplify. */
6474 /* If we match another movable, we must use that, as
6475 this one is going away. */
6477 return simplify_giv_expr (m->match->set_dest, benefit);
6479 /* If consec is non-zero, this is a member of a group of
6480 instructions that were moved together. We handle this
6481 case only to the point of seeking to the last insn and
6482 looking for a REG_EQUAL. Fail if we don't find one. */
6487 do { tem = NEXT_INSN (tem); } while (--i > 0);
6489 tem = find_reg_note (tem, REG_EQUAL, NULL_RTX);
6491 tem = XEXP (tem, 0);
6495 tem = single_set (m->insn);
6497 tem = SET_SRC (tem);
6502 /* What we are most interested in is pointer
6503 arithmetic on invariants -- only take
6504 patterns we may be able to do something with. */
6505 if (GET_CODE (tem) == PLUS
6506 || GET_CODE (tem) == MULT
6507 || GET_CODE (tem) == ASHIFT
6508 || GET_CODE (tem) == CONST_INT
6509 || GET_CODE (tem) == SYMBOL_REF)
6511 tem = simplify_giv_expr (tem, benefit);
6515 else if (GET_CODE (tem) == CONST
6516 && GET_CODE (XEXP (tem, 0)) == PLUS
6517 && GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF
6518 && GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT)
6520 tem = simplify_giv_expr (XEXP (tem, 0), benefit);
6531 /* Fall through to general case. */
6533 /* If invariant, return as USE (unless CONST_INT).
6534 Otherwise, not giv. */
6535 if (GET_CODE (x) == USE)
6538 if (invariant_p (x) == 1)
6540 if (GET_CODE (x) == CONST_INT)
6542 if (GET_CODE (x) == CONST
6543 && GET_CODE (XEXP (x, 0)) == PLUS
6544 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
6545 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
6547 return gen_rtx_USE (mode, x);
6554 /* This routine folds invariants such that there is only ever one
6555 CONST_INT in the summation. It is only used by simplify_giv_expr. */
6558 sge_plus_constant (x, c)
6561 if (GET_CODE (x) == CONST_INT)
6562 return GEN_INT (INTVAL (x) + INTVAL (c));
6563 else if (GET_CODE (x) != PLUS)
6564 return gen_rtx_PLUS (GET_MODE (x), x, c);
6565 else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6567 return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
6568 GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c)));
6570 else if (GET_CODE (XEXP (x, 0)) == PLUS
6571 || GET_CODE (XEXP (x, 1)) != PLUS)
6573 return gen_rtx_PLUS (GET_MODE (x),
6574 sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1));
6578 return gen_rtx_PLUS (GET_MODE (x),
6579 sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0));
6584 sge_plus (mode, x, y)
6585 enum machine_mode mode;
6588 while (GET_CODE (y) == PLUS)
6590 rtx a = XEXP (y, 0);
6591 if (GET_CODE (a) == CONST_INT)
6592 x = sge_plus_constant (x, a);
6594 x = gen_rtx_PLUS (mode, x, a);
6597 if (GET_CODE (y) == CONST_INT)
6598 x = sge_plus_constant (x, y);
6600 x = gen_rtx_PLUS (mode, x, y);
6604 /* Help detect a giv that is calculated by several consecutive insns;
6608 The caller has already identified the first insn P as having a giv as dest;
6609 we check that all other insns that set the same register follow
6610 immediately after P, that they alter nothing else,
6611 and that the result of the last is still a giv.
6613 The value is 0 if the reg set in P is not really a giv.
6614 Otherwise, the value is the amount gained by eliminating
6615 all the consecutive insns that compute the value.
6617 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
6618 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
6620 The coefficients of the ultimate giv value are stored in
6621 *MULT_VAL and *ADD_VAL. */
6624 consec_sets_giv (first_benefit, p, src_reg, dest_reg,
6625 add_val, mult_val, last_consec_insn)
6632 rtx *last_consec_insn;
6640 /* Indicate that this is a giv so that we can update the value produced in
6641 each insn of the multi-insn sequence.
6643 This induction structure will be used only by the call to
6644 general_induction_var below, so we can allocate it on our stack.
6645 If this is a giv, our caller will replace the induct var entry with
6646 a new induction structure. */
6648 = (struct induction *) alloca (sizeof (struct induction));
6649 v->src_reg = src_reg;
6650 v->mult_val = *mult_val;
6651 v->add_val = *add_val;
6652 v->benefit = first_benefit;
6654 v->derive_adjustment = 0;
6656 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
6657 REG_IV_INFO (REGNO (dest_reg)) = v;
6659 count = VARRAY_INT (n_times_set, REGNO (dest_reg)) - 1;
6664 code = GET_CODE (p);
6666 /* If libcall, skip to end of call sequence. */
6667 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
6671 && (set = single_set (p))
6672 && GET_CODE (SET_DEST (set)) == REG
6673 && SET_DEST (set) == dest_reg
6674 && (general_induction_var (SET_SRC (set), &src_reg,
6675 add_val, mult_val, 0, &benefit)
6676 /* Giv created by equivalent expression. */
6677 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
6678 && general_induction_var (XEXP (temp, 0), &src_reg,
6679 add_val, mult_val, 0, &benefit)))
6680 && src_reg == v->src_reg)
6682 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
6683 benefit += libcall_benefit (p);
6686 v->mult_val = *mult_val;
6687 v->add_val = *add_val;
6688 v->benefit = benefit;
6690 else if (code != NOTE)
6692 /* Allow insns that set something other than this giv to a
6693 constant. Such insns are needed on machines which cannot
6694 include long constants and should not disqualify a giv. */
6696 && (set = single_set (p))
6697 && SET_DEST (set) != dest_reg
6698 && CONSTANT_P (SET_SRC (set)))
6701 REG_IV_TYPE (REGNO (dest_reg)) = UNKNOWN_INDUCT;
6706 *last_consec_insn = p;
6710 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6711 represented by G1. If no such expression can be found, or it is clear that
6712 it cannot possibly be a valid address, 0 is returned.
6714 To perform the computation, we note that
6717 where `v' is the biv.
6719 So G2 = (y/b) * G1 + (b - a*y/x).
6721 Note that MULT = y/x.
6723 Update: A and B are now allowed to be additive expressions such that
6724 B contains all variables in A. That is, computing B-A will not require
6725 subtracting variables. */
6728 express_from_1 (a, b, mult)
6731 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
6733 if (mult == const0_rtx)
6736 /* If MULT is not 1, we cannot handle A with non-constants, since we
6737 would then be required to subtract multiples of the registers in A.
6738 This is theoretically possible, and may even apply to some Fortran
6739 constructs, but it is a lot of work and we do not attempt it here. */
6741 if (mult != const1_rtx && GET_CODE (a) != CONST_INT)
6744 /* In general these structures are sorted top to bottom (down the PLUS
6745 chain), but not left to right across the PLUS. If B is a higher
6746 order giv than A, we can strip one level and recurse. If A is higher
6747 order, we'll eventually bail out, but won't know that until the end.
6748 If they are the same, we'll strip one level around this loop. */
6750 while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS)
6752 rtx ra, rb, oa, ob, tmp;
6754 ra = XEXP (a, 0), oa = XEXP (a, 1);
6755 if (GET_CODE (ra) == PLUS)
6756 tmp = ra, ra = oa, oa = tmp;
6758 rb = XEXP (b, 0), ob = XEXP (b, 1);
6759 if (GET_CODE (rb) == PLUS)
6760 tmp = rb, rb = ob, ob = tmp;
6762 if (rtx_equal_p (ra, rb))
6763 /* We matched: remove one reg completely. */
6765 else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob))
6766 /* An alternate match. */
6768 else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb))
6769 /* An alternate match. */
6773 /* Indicates an extra register in B. Strip one level from B and
6774 recurse, hoping B was the higher order expression. */
6775 ob = express_from_1 (a, ob, mult);
6778 return gen_rtx_PLUS (GET_MODE (b), rb, ob);
6782 /* Here we are at the last level of A, go through the cases hoping to
6783 get rid of everything but a constant. */
6785 if (GET_CODE (a) == PLUS)
6789 ra = XEXP (a, 0), oa = XEXP (a, 1);
6790 if (rtx_equal_p (oa, b))
6792 else if (!rtx_equal_p (ra, b))
6795 if (GET_CODE (oa) != CONST_INT)
6798 return GEN_INT (-INTVAL (oa) * INTVAL (mult));
6800 else if (GET_CODE (a) == CONST_INT)
6802 return plus_constant (b, -INTVAL (a) * INTVAL (mult));
6804 else if (GET_CODE (b) == PLUS)
6806 if (rtx_equal_p (a, XEXP (b, 0)))
6808 else if (rtx_equal_p (a, XEXP (b, 1)))
6813 else if (rtx_equal_p (a, b))
6820 express_from (g1, g2)
6821 struct induction *g1, *g2;
6825 /* The value that G1 will be multiplied by must be a constant integer. Also,
6826 the only chance we have of getting a valid address is if b*c/a (see above
6827 for notation) is also an integer. */
6828 if (GET_CODE (g1->mult_val) == CONST_INT
6829 && GET_CODE (g2->mult_val) == CONST_INT)
6831 if (g1->mult_val == const0_rtx
6832 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
6834 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
6836 else if (rtx_equal_p (g1->mult_val, g2->mult_val))
6840 /* ??? Find out if the one is a multiple of the other? */
6844 add = express_from_1 (g1->add_val, g2->add_val, mult);
6845 if (add == NULL_RTX)
6847 /* Failed. If we've got a multiplication factor between G1 and G2,
6848 scale G1's addend and try again. */
6849 if (INTVAL (mult) > 1)
6851 rtx g1_add_val = g1->add_val;
6852 if (GET_CODE (g1_add_val) == MULT
6853 && GET_CODE (XEXP (g1_add_val, 1)) == CONST_INT)
6856 m = INTVAL (mult) * INTVAL (XEXP (g1_add_val, 1));
6857 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val),
6858 XEXP (g1_add_val, 0), GEN_INT (m));
6862 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val), g1_add_val,
6866 add = express_from_1 (g1_add_val, g2->add_val, const1_rtx);
6869 if (add == NULL_RTX)
6872 /* Form simplified final result. */
6873 if (mult == const0_rtx)
6875 else if (mult == const1_rtx)
6876 mult = g1->dest_reg;
6878 mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult);
6880 if (add == const0_rtx)
6884 if (GET_CODE (add) == PLUS
6885 && CONSTANT_P (XEXP (add, 1)))
6887 rtx tem = XEXP (add, 1);
6888 mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0));
6892 return gen_rtx_PLUS (g2->mode, mult, add);
6897 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6898 represented by G1. This indicates that G2 should be combined with G1 and
6899 that G2 can use (either directly or via an address expression) a register
6900 used to represent G1. */
6903 combine_givs_p (g1, g2)
6904 struct induction *g1, *g2;
6906 rtx tem = express_from (g1, g2);
6908 /* If these givs are identical, they can be combined. We use the results
6909 of express_from because the addends are not in a canonical form, so
6910 rtx_equal_p is a weaker test. */
6911 /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
6912 combination to be the other way round. */
6913 if (tem == g1->dest_reg
6914 && (g1->giv_type == DEST_REG || g2->giv_type == DEST_ADDR))
6916 return g1->dest_reg;
6919 /* If G2 can be expressed as a function of G1 and that function is valid
6920 as an address and no more expensive than using a register for G2,
6921 the expression of G2 in terms of G1 can be used. */
6923 && g2->giv_type == DEST_ADDR
6924 && memory_address_p (g2->mem_mode, tem)
6925 /* ??? Looses, especially with -fforce-addr, where *g2->location
6926 will always be a register, and so anything more complicated
6930 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location)
6932 && rtx_cost (tem, MEM) <= rtx_cost (*g2->location, MEM)
6943 struct combine_givs_stats
6950 cmp_combine_givs_stats (x, y)
6951 struct combine_givs_stats *x, *y;
6954 d = y->total_benefit - x->total_benefit;
6955 /* Stabilize the sort. */
6957 d = x->giv_number - y->giv_number;
6961 /* Check all pairs of givs for iv_class BL and see if any can be combined with
6962 any other. If so, point SAME to the giv combined with and set NEW_REG to
6963 be an expression (in terms of the other giv's DEST_REG) equivalent to the
6964 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
6968 struct iv_class *bl;
6970 /* Additional benefit to add for being combined multiple times. */
6971 const int extra_benefit = 3;
6973 struct induction *g1, *g2, **giv_array;
6974 int i, j, k, giv_count;
6975 struct combine_givs_stats *stats;
6978 /* Count givs, because bl->giv_count is incorrect here. */
6980 for (g1 = bl->giv; g1; g1 = g1->next_iv)
6985 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
6987 for (g1 = bl->giv; g1; g1 = g1->next_iv)
6989 giv_array[i++] = g1;
6991 stats = (struct combine_givs_stats *) alloca (giv_count * sizeof (*stats));
6992 bzero ((char *) stats, giv_count * sizeof (*stats));
6994 can_combine = (rtx *) alloca (giv_count * giv_count * sizeof(rtx));
6995 bzero ((char *) can_combine, giv_count * giv_count * sizeof(rtx));
6997 for (i = 0; i < giv_count; i++)
7003 stats[i].giv_number = i;
7005 /* If a DEST_REG GIV is used only once, do not allow it to combine
7006 with anything, for in doing so we will gain nothing that cannot
7007 be had by simply letting the GIV with which we would have combined
7008 to be reduced on its own. The losage shows up in particular with
7009 DEST_ADDR targets on hosts with reg+reg addressing, though it can
7010 be seen elsewhere as well. */
7011 if (g1->giv_type == DEST_REG
7012 && (single_use = VARRAY_RTX (reg_single_usage, REGNO (g1->dest_reg)))
7013 && single_use != const0_rtx)
7016 this_benefit = g1->benefit;
7017 /* Add an additional weight for zero addends. */
7018 if (g1->no_const_addval)
7021 for (j = 0; j < giv_count; j++)
7027 && (this_combine = combine_givs_p (g1, g2)) != NULL_RTX)
7029 can_combine[i*giv_count + j] = this_combine;
7030 this_benefit += g2->benefit + extra_benefit;
7033 stats[i].total_benefit = this_benefit;
7036 /* Iterate, combining until we can't. */
7038 qsort (stats, giv_count, sizeof(*stats), cmp_combine_givs_stats);
7040 if (loop_dump_stream)
7042 fprintf (loop_dump_stream, "Sorted combine statistics:\n");
7043 for (k = 0; k < giv_count; k++)
7045 g1 = giv_array[stats[k].giv_number];
7046 if (!g1->combined_with && !g1->same)
7047 fprintf (loop_dump_stream, " {%d, %d}",
7048 INSN_UID (giv_array[stats[k].giv_number]->insn),
7049 stats[k].total_benefit);
7051 putc ('\n', loop_dump_stream);
7054 for (k = 0; k < giv_count; k++)
7056 int g1_add_benefit = 0;
7058 i = stats[k].giv_number;
7061 /* If it has already been combined, skip. */
7062 if (g1->combined_with || g1->same)
7065 for (j = 0; j < giv_count; j++)
7068 if (g1 != g2 && can_combine[i*giv_count + j]
7069 /* If it has already been combined, skip. */
7070 && ! g2->same && ! g2->combined_with)
7074 g2->new_reg = can_combine[i*giv_count + j];
7076 g1->combined_with++;
7077 g1->lifetime += g2->lifetime;
7079 g1_add_benefit += g2->benefit;
7081 /* ??? The new final_[bg]iv_value code does a much better job
7082 of finding replaceable giv's, and hence this code may no
7083 longer be necessary. */
7084 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
7085 g1_add_benefit -= copy_cost;
7087 /* To help optimize the next set of combinations, remove
7088 this giv from the benefits of other potential mates. */
7089 for (l = 0; l < giv_count; ++l)
7091 int m = stats[l].giv_number;
7092 if (can_combine[m*giv_count + j])
7093 stats[l].total_benefit -= g2->benefit + extra_benefit;
7096 if (loop_dump_stream)
7097 fprintf (loop_dump_stream,
7098 "giv at %d combined with giv at %d\n",
7099 INSN_UID (g2->insn), INSN_UID (g1->insn));
7103 /* To help optimize the next set of combinations, remove
7104 this giv from the benefits of other potential mates. */
7105 if (g1->combined_with)
7107 for (j = 0; j < giv_count; ++j)
7109 int m = stats[j].giv_number;
7110 if (can_combine[m*giv_count + i])
7111 stats[j].total_benefit -= g1->benefit + extra_benefit;
7114 g1->benefit += g1_add_benefit;
7116 /* We've finished with this giv, and everything it touched.
7117 Restart the combination so that proper weights for the
7118 rest of the givs are properly taken into account. */
7119 /* ??? Ideally we would compact the arrays at this point, so
7120 as to not cover old ground. But sanely compacting
7121 can_combine is tricky. */
7127 struct recombine_givs_stats
7130 int start_luid, end_luid;
7133 /* Used below as comparison function for qsort. We want a ascending luid
7134 when scanning the array starting at the end, thus the arguments are
7137 cmp_recombine_givs_stats (x, y)
7138 struct recombine_givs_stats *x, *y;
7141 d = y->start_luid - x->start_luid;
7142 /* Stabilize the sort. */
7144 d = y->giv_number - x->giv_number;
7148 /* Scan X, which is a part of INSN, for the end of life of a giv. Also
7149 look for the start of life of a giv where the start has not been seen
7150 yet to unlock the search for the end of its life.
7151 Only consider givs that belong to BIV.
7152 Return the total number of lifetime ends that have been found. */
7154 find_life_end (x, stats, insn, biv)
7156 struct recombine_givs_stats *stats;
7163 code = GET_CODE (x);
7168 rtx reg = SET_DEST (x);
7169 if (GET_CODE (reg) == REG)
7171 int regno = REGNO (reg);
7172 struct induction *v = REG_IV_INFO (regno);
7174 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7176 && v->src_reg == biv
7177 && stats[v->ix].end_luid <= 0)
7179 /* If we see a 0 here for end_luid, it means that we have
7180 scanned the entire loop without finding any use at all.
7181 We must not predicate this code on a start_luid match
7182 since that would make the test fail for givs that have
7183 been hoisted out of inner loops. */
7184 if (stats[v->ix].end_luid == 0)
7186 stats[v->ix].end_luid = stats[v->ix].start_luid;
7187 return 1 + find_life_end (SET_SRC (x), stats, insn, biv);
7189 else if (stats[v->ix].start_luid == INSN_LUID (insn))
7190 stats[v->ix].end_luid = 0;
7192 return find_life_end (SET_SRC (x), stats, insn, biv);
7198 int regno = REGNO (x);
7199 struct induction *v = REG_IV_INFO (regno);
7201 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7203 && v->src_reg == biv
7204 && stats[v->ix].end_luid == 0)
7206 while (INSN_UID (insn) >= max_uid_for_loop)
7207 insn = NEXT_INSN (insn);
7208 stats[v->ix].end_luid = INSN_LUID (insn);
7221 fmt = GET_RTX_FORMAT (code);
7223 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7226 retval += find_life_end (XEXP (x, i), stats, insn, biv);
7228 else if (fmt[i] == 'E')
7229 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7230 retval += find_life_end (XVECEXP (x, i, j), stats, insn, biv);
7235 /* For each giv that has been combined with another, look if
7236 we can combine it with the most recently used one instead.
7237 This tends to shorten giv lifetimes, and helps the next step:
7238 try to derive givs from other givs. */
7240 recombine_givs (bl, loop_start, loop_end, unroll_p)
7241 struct iv_class *bl;
7242 rtx loop_start, loop_end;
7245 struct induction *v, **giv_array, *last_giv;
7246 struct recombine_givs_stats *stats;
7249 int ends_need_computing;
7251 for (giv_count = 0, v = bl->giv; v; v = v->next_iv)
7257 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
7258 stats = (struct recombine_givs_stats *) alloca (giv_count * sizeof *stats);
7260 /* Initialize stats and set up the ix field for each giv in stats to name
7261 the corresponding index into stats. */
7262 for (i = 0, v = bl->giv; v; v = v->next_iv)
7269 stats[i].giv_number = i;
7270 /* If this giv has been hoisted out of an inner loop, use the luid of
7271 the previous insn. */
7272 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7274 stats[i].start_luid = INSN_LUID (p);
7278 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7280 /* Set up the ix field for each giv in stats to name
7281 the corresponding index into stats, and
7282 do the actual most-recently-used recombination. */
7283 for (last_giv = 0, i = giv_count - 1; i >= 0; i--)
7285 v = giv_array[stats[i].giv_number];
7289 struct induction *old_same = v->same;
7292 /* combine_givs_p actually says if we can make this transformation.
7293 The other tests are here only to avoid keeping a giv alive
7294 that could otherwise be eliminated. */
7296 && ((old_same->maybe_dead && ! old_same->combined_with)
7297 || ! last_giv->maybe_dead
7298 || last_giv->combined_with)
7299 && (new_combine = combine_givs_p (last_giv, v)))
7301 old_same->combined_with--;
7302 v->new_reg = new_combine;
7304 last_giv->combined_with++;
7305 /* No need to update lifetimes / benefits here since we have
7306 already decided what to reduce. */
7308 if (loop_dump_stream)
7310 fprintf (loop_dump_stream,
7311 "giv at %d recombined with giv at %d as ",
7312 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7313 print_rtl (loop_dump_stream, v->new_reg);
7314 putc ('\n', loop_dump_stream);
7320 else if (v->giv_type != DEST_REG)
7323 || (last_giv->maybe_dead && ! last_giv->combined_with)
7325 || v->combined_with)
7329 ends_need_computing = 0;
7330 /* For each DEST_REG giv, compute lifetime starts, and try to compute
7331 lifetime ends from regscan info. */
7332 for (i = giv_count - 1; i >= 0; i--)
7334 v = giv_array[stats[i].giv_number];
7337 if (v->giv_type == DEST_ADDR)
7339 /* Loop unrolling of an inner loop can even create new DEST_REG
7342 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7344 stats[i].start_luid = stats[i].end_luid = INSN_LUID (p);
7346 stats[i].end_luid++;
7348 else /* v->giv_type == DEST_REG */
7352 stats[i].start_luid = INSN_LUID (v->insn);
7353 stats[i].end_luid = INSN_LUID (v->last_use);
7355 else if (INSN_UID (v->insn) >= max_uid_for_loop)
7358 /* This insn has been created by loop optimization on an inner
7359 loop. We don't have a proper start_luid that will match
7360 when we see the first set. But we do know that there will
7361 be no use before the set, so we can set end_luid to 0 so that
7362 we'll start looking for the last use right away. */
7363 for (p = PREV_INSN (v->insn); INSN_UID (p) >= max_uid_for_loop; )
7365 stats[i].start_luid = INSN_LUID (p);
7366 stats[i].end_luid = 0;
7367 ends_need_computing++;
7371 int regno = REGNO (v->dest_reg);
7372 int count = VARRAY_INT (n_times_set, regno) - 1;
7375 /* Find the first insn that sets the giv, so that we can verify
7376 if this giv's lifetime wraps around the loop. We also need
7377 the luid of the first setting insn in order to detect the
7378 last use properly. */
7381 p = prev_nonnote_insn (p);
7382 if (reg_set_p (v->dest_reg, p))
7386 stats[i].start_luid = INSN_LUID (p);
7387 if (stats[i].start_luid > uid_luid[REGNO_FIRST_UID (regno)])
7389 stats[i].end_luid = -1;
7390 ends_need_computing++;
7394 stats[i].end_luid = uid_luid[REGNO_LAST_UID (regno)];
7395 if (stats[i].end_luid > INSN_LUID (loop_end))
7397 stats[i].end_luid = -1;
7398 ends_need_computing++;
7405 /* If the regscan information was unconclusive for one or more DEST_REG
7406 givs, scan the all insn in the loop to find out lifetime ends. */
7407 if (ends_need_computing)
7409 rtx biv = bl->biv->src_reg;
7414 if (p == loop_start)
7417 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
7419 ends_need_computing -= find_life_end (PATTERN (p), stats, p, biv);
7421 while (ends_need_computing);
7424 /* Set start_luid back to the last insn that sets the giv. This allows
7425 more combinations. */
7426 for (i = giv_count - 1; i >= 0; i--)
7428 v = giv_array[stats[i].giv_number];
7431 if (INSN_UID (v->insn) < max_uid_for_loop)
7432 stats[i].start_luid = INSN_LUID (v->insn);
7435 /* Now adjust lifetime ends by taking combined givs into account. */
7436 for (i = giv_count - 1; i >= 0; i--)
7441 v = giv_array[stats[i].giv_number];
7444 if (v->same && ! v->same->ignore)
7447 luid = stats[i].start_luid;
7448 /* Use unsigned arithmetic to model loop wrap-around. */
7449 if (luid - stats[j].start_luid
7450 > (unsigned) stats[j].end_luid - stats[j].start_luid)
7451 stats[j].end_luid = luid;
7455 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7457 /* Try to derive DEST_REG givs from previous DEST_REG givs with the
7458 same mult_val and non-overlapping lifetime. This reduces register
7460 Once we find a DEST_REG giv that is suitable to derive others from,
7461 we set last_giv to this giv, and try to derive as many other DEST_REG
7462 givs from it without joining overlapping lifetimes. If we then
7463 encounter a DEST_REG giv that we can't derive, we set rescan to the
7464 index for this giv (unless rescan is already set).
7465 When we are finished with the current LAST_GIV (i.e. the inner loop
7466 terminates), we start again with rescan, which then becomes the new
7468 for (i = giv_count - 1; i >= 0; i = rescan)
7470 int life_start, life_end;
7472 for (last_giv = 0, rescan = -1; i >= 0; i--)
7476 v = giv_array[stats[i].giv_number];
7477 if (v->giv_type != DEST_REG || v->derived_from || v->same)
7481 /* Don't use a giv that's likely to be dead to derive
7482 others - that would be likely to keep that giv alive. */
7483 if (! v->maybe_dead || v->combined_with)
7486 life_start = stats[i].start_luid;
7487 life_end = stats[i].end_luid;
7491 /* Use unsigned arithmetic to model loop wrap around. */
7492 if (((unsigned) stats[i].start_luid - life_start
7493 >= (unsigned) life_end - life_start)
7494 && ((unsigned) stats[i].end_luid - life_start
7495 > (unsigned) life_end - life_start)
7496 /* Check that the giv insn we're about to use for deriving
7497 precedes all uses of that giv. Note that initializing the
7498 derived giv would defeat the purpose of reducing register
7500 ??? We could arrange to move the insn. */
7501 && ((unsigned) stats[i].end_luid - INSN_LUID (loop_start)
7502 > (unsigned) stats[i].start_luid - INSN_LUID (loop_start))
7503 && rtx_equal_p (last_giv->mult_val, v->mult_val)
7504 /* ??? Could handle libcalls, but would need more logic. */
7505 && ! find_reg_note (v->insn, REG_RETVAL, NULL_RTX)
7506 /* We would really like to know if for any giv that v
7507 is combined with, v->insn or any intervening biv increment
7508 dominates that combined giv. However, we
7509 don't have this detailed control flow information.
7510 N.B. since last_giv will be reduced, it is valid
7511 anywhere in the loop, so we don't need to check the
7512 validity of last_giv.
7513 We rely here on the fact that v->always_executed implies that
7514 there is no jump to someplace else in the loop before the
7515 giv insn, and hence any insn that is executed before the
7516 giv insn in the loop will have a lower luid. */
7517 && (v->always_executed || ! v->combined_with)
7518 && (sum = express_from (last_giv, v))
7519 /* Make sure we don't make the add more expensive. ADD_COST
7520 doesn't take different costs of registers and constants into
7521 account, so compare the cost of the actual SET_SRCs. */
7522 && (rtx_cost (sum, SET)
7523 <= rtx_cost (SET_SRC (single_set (v->insn)), SET))
7524 /* ??? unroll can't understand anything but reg + const_int
7525 sums. It would be cleaner to fix unroll. */
7526 && ((GET_CODE (sum) == PLUS
7527 && GET_CODE (XEXP (sum, 0)) == REG
7528 && GET_CODE (XEXP (sum, 1)) == CONST_INT)
7530 && validate_change (v->insn, &PATTERN (v->insn),
7531 gen_rtx_SET (VOIDmode, v->dest_reg, sum), 0))
7533 v->derived_from = last_giv;
7534 life_end = stats[i].end_luid;
7536 if (loop_dump_stream)
7538 fprintf (loop_dump_stream,
7539 "giv at %d derived from %d as ",
7540 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7541 print_rtl (loop_dump_stream, sum);
7542 putc ('\n', loop_dump_stream);
7545 else if (rescan < 0)
7551 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
7554 emit_iv_add_mult (b, m, a, reg, insert_before)
7555 rtx b; /* initial value of basic induction variable */
7556 rtx m; /* multiplicative constant */
7557 rtx a; /* additive constant */
7558 rtx reg; /* destination register */
7564 /* Prevent unexpected sharing of these rtx. */
7568 /* Increase the lifetime of any invariants moved further in code. */
7569 update_reg_last_use (a, insert_before);
7570 update_reg_last_use (b, insert_before);
7571 update_reg_last_use (m, insert_before);
7574 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
7576 emit_move_insn (reg, result);
7577 seq = gen_sequence ();
7580 emit_insn_before (seq, insert_before);
7582 /* It is entirely possible that the expansion created lots of new
7583 registers. Iterate over the sequence we just created and
7586 if (GET_CODE (seq) == SEQUENCE)
7589 for (i = 0; i < XVECLEN (seq, 0); ++i)
7591 rtx set = single_set (XVECEXP (seq, 0, i));
7592 if (set && GET_CODE (SET_DEST (set)) == REG)
7593 record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0);
7596 else if (GET_CODE (seq) == SET
7597 && GET_CODE (SET_DEST (seq)) == REG)
7598 record_base_value (REGNO (SET_DEST (seq)), SET_SRC (seq), 0);
7601 /* Test whether A * B can be computed without
7602 an actual multiply insn. Value is 1 if so. */
7605 product_cheap_p (a, b)
7611 struct obstack *old_rtl_obstack = rtl_obstack;
7612 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
7615 /* If only one is constant, make it B. */
7616 if (GET_CODE (a) == CONST_INT)
7617 tmp = a, a = b, b = tmp;
7619 /* If first constant, both constant, so don't need multiply. */
7620 if (GET_CODE (a) == CONST_INT)
7623 /* If second not constant, neither is constant, so would need multiply. */
7624 if (GET_CODE (b) != CONST_INT)
7627 /* One operand is constant, so might not need multiply insn. Generate the
7628 code for the multiply and see if a call or multiply, or long sequence
7629 of insns is generated. */
7631 rtl_obstack = &temp_obstack;
7633 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
7634 tmp = gen_sequence ();
7637 if (GET_CODE (tmp) == SEQUENCE)
7639 if (XVEC (tmp, 0) == 0)
7641 else if (XVECLEN (tmp, 0) > 3)
7644 for (i = 0; i < XVECLEN (tmp, 0); i++)
7646 rtx insn = XVECEXP (tmp, 0, i);
7648 if (GET_CODE (insn) != INSN
7649 || (GET_CODE (PATTERN (insn)) == SET
7650 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
7651 || (GET_CODE (PATTERN (insn)) == PARALLEL
7652 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
7653 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
7660 else if (GET_CODE (tmp) == SET
7661 && GET_CODE (SET_SRC (tmp)) == MULT)
7663 else if (GET_CODE (tmp) == PARALLEL
7664 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
7665 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
7668 /* Free any storage we obtained in generating this multiply and restore rtl
7669 allocation to its normal obstack. */
7670 obstack_free (&temp_obstack, storage);
7671 rtl_obstack = old_rtl_obstack;
7676 /* Check to see if loop can be terminated by a "decrement and branch until
7677 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
7678 Also try reversing an increment loop to a decrement loop
7679 to see if the optimization can be performed.
7680 Value is nonzero if optimization was performed. */
7682 /* This is useful even if the architecture doesn't have such an insn,
7683 because it might change a loops which increments from 0 to n to a loop
7684 which decrements from n to 0. A loop that decrements to zero is usually
7685 faster than one that increments from zero. */
7687 /* ??? This could be rewritten to use some of the loop unrolling procedures,
7688 such as approx_final_value, biv_total_increment, loop_iterations, and
7689 final_[bg]iv_value. */
7692 check_dbra_loop (loop_end, insn_count, loop_start, loop_info)
7696 struct loop_info *loop_info;
7698 struct iv_class *bl;
7705 rtx before_comparison;
7709 int compare_and_branch;
7711 /* If last insn is a conditional branch, and the insn before tests a
7712 register value, try to optimize it. Otherwise, we can't do anything. */
7714 jump = PREV_INSN (loop_end);
7715 comparison = get_condition_for_loop (jump);
7716 if (comparison == 0)
7719 /* Try to compute whether the compare/branch at the loop end is one or
7720 two instructions. */
7721 get_condition (jump, &first_compare);
7722 if (first_compare == jump)
7723 compare_and_branch = 1;
7724 else if (first_compare == prev_nonnote_insn (jump))
7725 compare_and_branch = 2;
7729 /* Check all of the bivs to see if the compare uses one of them.
7730 Skip biv's set more than once because we can't guarantee that
7731 it will be zero on the last iteration. Also skip if the biv is
7732 used between its update and the test insn. */
7734 for (bl = loop_iv_list; bl; bl = bl->next)
7736 if (bl->biv_count == 1
7737 && bl->biv->dest_reg == XEXP (comparison, 0)
7738 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
7746 /* Look for the case where the basic induction variable is always
7747 nonnegative, and equals zero on the last iteration.
7748 In this case, add a reg_note REG_NONNEG, which allows the
7749 m68k DBRA instruction to be used. */
7751 if (((GET_CODE (comparison) == GT
7752 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
7753 && INTVAL (XEXP (comparison, 1)) == -1)
7754 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
7755 && GET_CODE (bl->biv->add_val) == CONST_INT
7756 && INTVAL (bl->biv->add_val) < 0)
7758 /* Initial value must be greater than 0,
7759 init_val % -dec_value == 0 to ensure that it equals zero on
7760 the last iteration */
7762 if (GET_CODE (bl->initial_value) == CONST_INT
7763 && INTVAL (bl->initial_value) > 0
7764 && (INTVAL (bl->initial_value)
7765 % (-INTVAL (bl->biv->add_val))) == 0)
7767 /* register always nonnegative, add REG_NOTE to branch */
7768 REG_NOTES (PREV_INSN (loop_end))
7769 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7770 REG_NOTES (PREV_INSN (loop_end)));
7776 /* If the decrement is 1 and the value was tested as >= 0 before
7777 the loop, then we can safely optimize. */
7778 for (p = loop_start; p; p = PREV_INSN (p))
7780 if (GET_CODE (p) == CODE_LABEL)
7782 if (GET_CODE (p) != JUMP_INSN)
7785 before_comparison = get_condition_for_loop (p);
7786 if (before_comparison
7787 && XEXP (before_comparison, 0) == bl->biv->dest_reg
7788 && GET_CODE (before_comparison) == LT
7789 && XEXP (before_comparison, 1) == const0_rtx
7790 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
7791 && INTVAL (bl->biv->add_val) == -1)
7793 REG_NOTES (PREV_INSN (loop_end))
7794 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7795 REG_NOTES (PREV_INSN (loop_end)));
7802 else if (INTVAL (bl->biv->add_val) > 0)
7804 /* Try to change inc to dec, so can apply above optimization. */
7806 all registers modified are induction variables or invariant,
7807 all memory references have non-overlapping addresses
7808 (obviously true if only one write)
7809 allow 2 insns for the compare/jump at the end of the loop. */
7810 /* Also, we must avoid any instructions which use both the reversed
7811 biv and another biv. Such instructions will fail if the loop is
7812 reversed. We meet this condition by requiring that either
7813 no_use_except_counting is true, or else that there is only
7815 int num_nonfixed_reads = 0;
7816 /* 1 if the iteration var is used only to count iterations. */
7817 int no_use_except_counting = 0;
7818 /* 1 if the loop has no memory store, or it has a single memory store
7819 which is reversible. */
7820 int reversible_mem_store = 1;
7822 if (bl->giv_count == 0
7823 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
7825 rtx bivreg = regno_reg_rtx[bl->regno];
7827 /* If there are no givs for this biv, and the only exit is the
7828 fall through at the end of the loop, then
7829 see if perhaps there are no uses except to count. */
7830 no_use_except_counting = 1;
7831 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7832 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
7834 rtx set = single_set (p);
7836 if (set && GET_CODE (SET_DEST (set)) == REG
7837 && REGNO (SET_DEST (set)) == bl->regno)
7838 /* An insn that sets the biv is okay. */
7840 else if (p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
7841 || p == prev_nonnote_insn (loop_end))
7842 /* Don't bother about the end test. */
7844 else if (reg_mentioned_p (bivreg, PATTERN (p)))
7846 no_use_except_counting = 0;
7852 if (no_use_except_counting)
7853 ; /* no need to worry about MEMs. */
7854 else if (num_mem_sets <= 1)
7856 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7857 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
7858 num_nonfixed_reads += count_nonfixed_reads (PATTERN (p));
7860 /* If the loop has a single store, and the destination address is
7861 invariant, then we can't reverse the loop, because this address
7862 might then have the wrong value at loop exit.
7863 This would work if the source was invariant also, however, in that
7864 case, the insn should have been moved out of the loop. */
7866 if (num_mem_sets == 1)
7868 struct induction *v;
7870 reversible_mem_store
7871 = (! unknown_address_altered
7872 && ! invariant_p (XEXP (XEXP (loop_store_mems, 0), 0)));
7874 /* If the store depends on a register that is set after the
7875 store, it depends on the initial value, and is thus not
7877 for (v = bl->giv; reversible_mem_store && v; v = v->next_iv)
7879 if (v->giv_type == DEST_REG
7880 && reg_mentioned_p (v->dest_reg,
7881 XEXP (loop_store_mems, 0))
7882 && loop_insn_first_p (first_loop_store_insn, v->insn))
7883 reversible_mem_store = 0;
7890 /* This code only acts for innermost loops. Also it simplifies
7891 the memory address check by only reversing loops with
7892 zero or one memory access.
7893 Two memory accesses could involve parts of the same array,
7894 and that can't be reversed.
7895 If the biv is used only for counting, than we don't need to worry
7896 about all these things. */
7898 if ((num_nonfixed_reads <= 1
7900 && !loop_has_volatile
7901 && reversible_mem_store
7902 && (bl->giv_count + bl->biv_count + num_mem_sets
7903 + num_movables + compare_and_branch == insn_count)
7904 && (bl == loop_iv_list && bl->next == 0))
7905 || no_use_except_counting)
7909 /* Loop can be reversed. */
7910 if (loop_dump_stream)
7911 fprintf (loop_dump_stream, "Can reverse loop\n");
7913 /* Now check other conditions:
7915 The increment must be a constant, as must the initial value,
7916 and the comparison code must be LT.
7918 This test can probably be improved since +/- 1 in the constant
7919 can be obtained by changing LT to LE and vice versa; this is
7923 /* for constants, LE gets turned into LT */
7924 && (GET_CODE (comparison) == LT
7925 || (GET_CODE (comparison) == LE
7926 && no_use_except_counting)))
7928 HOST_WIDE_INT add_val, add_adjust, comparison_val = 0;
7929 rtx initial_value, comparison_value;
7931 enum rtx_code cmp_code;
7932 int comparison_const_width;
7933 unsigned HOST_WIDE_INT comparison_sign_mask;
7935 add_val = INTVAL (bl->biv->add_val);
7936 comparison_value = XEXP (comparison, 1);
7937 if (GET_MODE (comparison_value) == VOIDmode)
7938 comparison_const_width
7939 = GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0)));
7941 comparison_const_width
7942 = GET_MODE_BITSIZE (GET_MODE (comparison_value));
7943 if (comparison_const_width > HOST_BITS_PER_WIDE_INT)
7944 comparison_const_width = HOST_BITS_PER_WIDE_INT;
7945 comparison_sign_mask
7946 = (unsigned HOST_WIDE_INT)1 << (comparison_const_width - 1);
7948 /* If the comparison value is not a loop invariant, then we
7949 can not reverse this loop.
7951 ??? If the insns which initialize the comparison value as
7952 a whole compute an invariant result, then we could move
7953 them out of the loop and proceed with loop reversal. */
7954 if (!invariant_p (comparison_value))
7957 if (GET_CODE (comparison_value) == CONST_INT)
7958 comparison_val = INTVAL (comparison_value);
7959 initial_value = bl->initial_value;
7961 /* Normalize the initial value if it is an integer and
7962 has no other use except as a counter. This will allow
7963 a few more loops to be reversed. */
7964 if (no_use_except_counting
7965 && GET_CODE (comparison_value) == CONST_INT
7966 && GET_CODE (initial_value) == CONST_INT)
7968 comparison_val = comparison_val - INTVAL (bl->initial_value);
7969 /* The code below requires comparison_val to be a multiple
7970 of add_val in order to do the loop reversal, so
7971 round up comparison_val to a multiple of add_val.
7972 Since comparison_value is constant, we know that the
7973 current comparison code is LT. */
7974 comparison_val = comparison_val + add_val - 1;
7976 -= (unsigned HOST_WIDE_INT) comparison_val % add_val;
7977 /* We postpone overflow checks for COMPARISON_VAL here;
7978 even if there is an overflow, we might still be able to
7979 reverse the loop, if converting the loop exit test to
7981 initial_value = const0_rtx;
7984 /* First check if we can do a vanilla loop reversal. */
7985 if (initial_value == const0_rtx
7986 /* If we have a decrement_and_branch_on_count, prefer
7987 the NE test, since this will allow that instruction to
7988 be generated. Note that we must use a vanilla loop
7989 reversal if the biv is used to calculate a giv or has
7990 a non-counting use. */
7991 #if ! defined (HAVE_decrement_and_branch_until_zero) && defined (HAVE_decrement_and_branch_on_count)
7992 && (! (add_val == 1 && loop_info->vtop
7993 && (bl->biv_count == 0
7994 || no_use_except_counting)))
7996 && GET_CODE (comparison_value) == CONST_INT
7997 /* Now do postponed overflow checks on COMPARISON_VAL. */
7998 && ! (((comparison_val - add_val) ^ INTVAL (comparison_value))
7999 & comparison_sign_mask))
8001 /* Register will always be nonnegative, with value
8002 0 on last iteration */
8003 add_adjust = add_val;
8007 else if (add_val == 1 && loop_info->vtop
8008 && (bl->biv_count == 0
8009 || no_use_except_counting))
8017 if (GET_CODE (comparison) == LE)
8018 add_adjust -= add_val;
8020 /* If the initial value is not zero, or if the comparison
8021 value is not an exact multiple of the increment, then we
8022 can not reverse this loop. */
8023 if (initial_value == const0_rtx
8024 && GET_CODE (comparison_value) == CONST_INT)
8026 if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0)
8031 if (! no_use_except_counting || add_val != 1)
8035 final_value = comparison_value;
8037 /* Reset these in case we normalized the initial value
8038 and comparison value above. */
8039 if (GET_CODE (comparison_value) == CONST_INT
8040 && GET_CODE (initial_value) == CONST_INT)
8042 comparison_value = GEN_INT (comparison_val);
8044 = GEN_INT (comparison_val + INTVAL (bl->initial_value));
8046 bl->initial_value = initial_value;
8048 /* Save some info needed to produce the new insns. */
8049 reg = bl->biv->dest_reg;
8050 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
8051 if (jump_label == pc_rtx)
8052 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
8053 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
8055 /* Set start_value; if this is not a CONST_INT, we need
8057 Initialize biv to start_value before loop start.
8058 The old initializing insn will be deleted as a
8059 dead store by flow.c. */
8060 if (initial_value == const0_rtx
8061 && GET_CODE (comparison_value) == CONST_INT)
8063 start_value = GEN_INT (comparison_val - add_adjust);
8064 emit_insn_before (gen_move_insn (reg, start_value),
8067 else if (GET_CODE (initial_value) == CONST_INT)
8069 rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust);
8070 enum machine_mode mode = GET_MODE (reg);
8071 enum insn_code icode
8072 = add_optab->handlers[(int) mode].insn_code;
8073 if (! (*insn_operand_predicate[icode][0]) (reg, mode)
8074 || ! ((*insn_operand_predicate[icode][1])
8075 (comparison_value, mode))
8076 || ! (*insn_operand_predicate[icode][2]) (offset, mode))
8079 = gen_rtx_PLUS (mode, comparison_value, offset);
8080 emit_insn_before ((GEN_FCN (icode)
8081 (reg, comparison_value, offset)),
8083 if (GET_CODE (comparison) == LE)
8084 final_value = gen_rtx_PLUS (mode, comparison_value,
8087 else if (! add_adjust)
8089 enum machine_mode mode = GET_MODE (reg);
8090 enum insn_code icode
8091 = sub_optab->handlers[(int) mode].insn_code;
8092 if (! (*insn_operand_predicate[icode][0]) (reg, mode)
8093 || ! ((*insn_operand_predicate[icode][1])
8094 (comparison_value, mode))
8095 || ! ((*insn_operand_predicate[icode][2])
8096 (initial_value, mode)))
8099 = gen_rtx_MINUS (mode, comparison_value, initial_value);
8100 emit_insn_before ((GEN_FCN (icode)
8101 (reg, comparison_value, initial_value)),
8105 /* We could handle the other cases too, but it'll be
8106 better to have a testcase first. */
8109 /* We may not have a single insn which can increment a reg, so
8110 create a sequence to hold all the insns from expand_inc. */
8112 expand_inc (reg, new_add_val);
8113 tem = gen_sequence ();
8116 p = emit_insn_before (tem, bl->biv->insn);
8117 delete_insn (bl->biv->insn);
8119 /* Update biv info to reflect its new status. */
8121 bl->initial_value = start_value;
8122 bl->biv->add_val = new_add_val;
8124 /* Update loop info. */
8125 loop_info->initial_value = reg;
8126 loop_info->initial_equiv_value = reg;
8127 loop_info->final_value = const0_rtx;
8128 loop_info->final_equiv_value = const0_rtx;
8129 loop_info->comparison_value = const0_rtx;
8130 loop_info->comparison_code = cmp_code;
8131 loop_info->increment = new_add_val;
8133 /* Inc LABEL_NUSES so that delete_insn will
8134 not delete the label. */
8135 LABEL_NUSES (XEXP (jump_label, 0)) ++;
8137 /* Emit an insn after the end of the loop to set the biv's
8138 proper exit value if it is used anywhere outside the loop. */
8139 if ((REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare))
8141 || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
8142 emit_insn_after (gen_move_insn (reg, final_value),
8145 /* Delete compare/branch at end of loop. */
8146 delete_insn (PREV_INSN (loop_end));
8147 if (compare_and_branch == 2)
8148 delete_insn (first_compare);
8150 /* Add new compare/branch insn at end of loop. */
8152 emit_cmp_and_jump_insns (reg, const0_rtx, cmp_code, NULL_RTX,
8153 GET_MODE (reg), 0, 0,
8154 XEXP (jump_label, 0));
8155 tem = gen_sequence ();
8157 emit_jump_insn_before (tem, loop_end);
8159 for (tem = PREV_INSN (loop_end);
8160 tem && GET_CODE (tem) != JUMP_INSN;
8161 tem = PREV_INSN (tem))
8165 JUMP_LABEL (tem) = XEXP (jump_label, 0);
8171 /* Increment of LABEL_NUSES done above. */
8172 /* Register is now always nonnegative,
8173 so add REG_NONNEG note to the branch. */
8174 REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
8180 /* No insn may reference both the reversed and another biv or it
8181 will fail (see comment near the top of the loop reversal
8183 Earlier on, we have verified that the biv has no use except
8184 counting, or it is the only biv in this function.
8185 However, the code that computes no_use_except_counting does
8186 not verify reg notes. It's possible to have an insn that
8187 references another biv, and has a REG_EQUAL note with an
8188 expression based on the reversed biv. To avoid this case,
8189 remove all REG_EQUAL notes based on the reversed biv
8191 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8192 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
8195 rtx set = single_set (p);
8196 /* If this is a set of a GIV based on the reversed biv, any
8197 REG_EQUAL notes should still be correct. */
8199 || GET_CODE (SET_DEST (set)) != REG
8200 || (size_t) REGNO (SET_DEST (set)) >= reg_iv_type->num_elements
8201 || REG_IV_TYPE (REGNO (SET_DEST (set))) != GENERAL_INDUCT
8202 || REG_IV_INFO (REGNO (SET_DEST (set)))->src_reg != bl->biv->src_reg)
8203 for (pnote = ®_NOTES (p); *pnote;)
8205 if (REG_NOTE_KIND (*pnote) == REG_EQUAL
8206 && reg_mentioned_p (regno_reg_rtx[bl->regno],
8208 *pnote = XEXP (*pnote, 1);
8210 pnote = &XEXP (*pnote, 1);
8214 /* Mark that this biv has been reversed. Each giv which depends
8215 on this biv, and which is also live past the end of the loop
8216 will have to be fixed up. */
8220 if (loop_dump_stream)
8222 fprintf (loop_dump_stream, "Reversed loop");
8224 fprintf (loop_dump_stream, " and added reg_nonneg\n");
8226 fprintf (loop_dump_stream, "\n");
8237 /* Verify whether the biv BL appears to be eliminable,
8238 based on the insns in the loop that refer to it.
8239 LOOP_START is the first insn of the loop, and END is the end insn.
8241 If ELIMINATE_P is non-zero, actually do the elimination.
8243 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
8244 determine whether invariant insns should be placed inside or at the
8245 start of the loop. */
8248 maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count)
8249 struct iv_class *bl;
8253 int threshold, insn_count;
8255 rtx reg = bl->biv->dest_reg;
8258 /* Scan all insns in the loop, stopping if we find one that uses the
8259 biv in a way that we cannot eliminate. */
8261 for (p = loop_start; p != end; p = NEXT_INSN (p))
8263 enum rtx_code code = GET_CODE (p);
8264 rtx where = threshold >= insn_count ? loop_start : p;
8266 /* If this is a libcall that sets a giv, skip ahead to its end. */
8267 if (GET_RTX_CLASS (code) == 'i')
8269 rtx note = find_reg_note (p, REG_LIBCALL, NULL_RTX);
8273 rtx last = XEXP (note, 0);
8274 rtx set = single_set (last);
8276 if (set && GET_CODE (SET_DEST (set)) == REG)
8278 int regno = REGNO (SET_DEST (set));
8280 if (regno < max_reg_before_loop
8281 && REG_IV_TYPE (regno) == GENERAL_INDUCT
8282 && REG_IV_INFO (regno)->src_reg == bl->biv->src_reg)
8287 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
8288 && reg_mentioned_p (reg, PATTERN (p))
8289 && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where))
8291 if (loop_dump_stream)
8292 fprintf (loop_dump_stream,
8293 "Cannot eliminate biv %d: biv used in insn %d.\n",
8294 bl->regno, INSN_UID (p));
8301 if (loop_dump_stream)
8302 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
8303 bl->regno, eliminate_p ? "was" : "can be");
8310 /* INSN and REFERENCE are instructions in the same insn chain.
8311 Return non-zero if INSN is first. */
8314 loop_insn_first_p (insn, reference)
8315 rtx insn, reference;
8319 for (p = insn, q = reference; ;)
8321 /* Start with test for not first so that INSN == REFERENCE yields not
8323 if (q == insn || ! p)
8325 if (p == reference || ! q)
8328 /* Either of P or Q might be a NOTE. Notes have the same LUID as the
8329 previous insn, hence the <= comparison below does not work if
8331 if (INSN_UID (p) < max_uid_for_loop
8332 && INSN_UID (q) < max_uid_for_loop
8333 && GET_CODE (p) != NOTE)
8334 return INSN_LUID (p) <= INSN_LUID (q);
8336 if (INSN_UID (p) >= max_uid_for_loop
8337 || GET_CODE (p) == NOTE)
8339 if (INSN_UID (q) >= max_uid_for_loop)
8344 /* We are trying to eliminate BIV in INSN using GIV. Return non-zero if
8345 the offset that we have to take into account due to auto-increment /
8346 div derivation is zero. */
8348 biv_elimination_giv_has_0_offset (biv, giv, insn)
8349 struct induction *biv, *giv;
8352 /* If the giv V had the auto-inc address optimization applied
8353 to it, and INSN occurs between the giv insn and the biv
8354 insn, then we'd have to adjust the value used here.
8355 This is rare, so we don't bother to make this possible. */
8356 if (giv->auto_inc_opt
8357 && ((loop_insn_first_p (giv->insn, insn)
8358 && loop_insn_first_p (insn, biv->insn))
8359 || (loop_insn_first_p (biv->insn, insn)
8360 && loop_insn_first_p (insn, giv->insn))))
8363 /* If the giv V was derived from another giv, and INSN does
8364 not occur between the giv insn and the biv insn, then we'd
8365 have to adjust the value used here. This is rare, so we don't
8366 bother to make this possible. */
8367 if (giv->derived_from
8368 && ! (giv->always_executed
8369 && loop_insn_first_p (giv->insn, insn)
8370 && loop_insn_first_p (insn, biv->insn)))
8373 && giv->same->derived_from
8374 && ! (giv->same->always_executed
8375 && loop_insn_first_p (giv->same->insn, insn)
8376 && loop_insn_first_p (insn, biv->insn)))
8382 /* If BL appears in X (part of the pattern of INSN), see if we can
8383 eliminate its use. If so, return 1. If not, return 0.
8385 If BIV does not appear in X, return 1.
8387 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
8388 where extra insns should be added. Depending on how many items have been
8389 moved out of the loop, it will either be before INSN or at the start of
8393 maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where)
8395 struct iv_class *bl;
8399 enum rtx_code code = GET_CODE (x);
8400 rtx reg = bl->biv->dest_reg;
8401 enum machine_mode mode = GET_MODE (reg);
8402 struct induction *v;
8414 /* If we haven't already been able to do something with this BIV,
8415 we can't eliminate it. */
8421 /* If this sets the BIV, it is not a problem. */
8422 if (SET_DEST (x) == reg)
8425 /* If this is an insn that defines a giv, it is also ok because
8426 it will go away when the giv is reduced. */
8427 for (v = bl->giv; v; v = v->next_iv)
8428 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
8432 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
8434 /* Can replace with any giv that was reduced and
8435 that has (MULT_VAL != 0) and (ADD_VAL == 0).
8436 Require a constant for MULT_VAL, so we know it's nonzero.
8437 ??? We disable this optimization to avoid potential
8440 for (v = bl->giv; v; v = v->next_iv)
8441 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
8442 && v->add_val == const0_rtx
8443 && ! v->ignore && ! v->maybe_dead && v->always_computable
8447 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8453 /* If the giv has the opposite direction of change,
8454 then reverse the comparison. */
8455 if (INTVAL (v->mult_val) < 0)
8456 new = gen_rtx_COMPARE (GET_MODE (v->new_reg),
8457 const0_rtx, v->new_reg);
8461 /* We can probably test that giv's reduced reg. */
8462 if (validate_change (insn, &SET_SRC (x), new, 0))
8466 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
8467 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
8468 Require a constant for MULT_VAL, so we know it's nonzero.
8469 ??? Do this only if ADD_VAL is a pointer to avoid a potential
8470 overflow problem. */
8472 for (v = bl->giv; v; v = v->next_iv)
8473 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
8474 && ! v->ignore && ! v->maybe_dead && v->always_computable
8476 && (GET_CODE (v->add_val) == SYMBOL_REF
8477 || GET_CODE (v->add_val) == LABEL_REF
8478 || GET_CODE (v->add_val) == CONST
8479 || (GET_CODE (v->add_val) == REG
8480 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
8482 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8488 /* If the giv has the opposite direction of change,
8489 then reverse the comparison. */
8490 if (INTVAL (v->mult_val) < 0)
8491 new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val),
8494 new = gen_rtx_COMPARE (VOIDmode, v->new_reg,
8495 copy_rtx (v->add_val));
8497 /* Replace biv with the giv's reduced register. */
8498 update_reg_last_use (v->add_val, insn);
8499 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8502 /* Insn doesn't support that constant or invariant. Copy it
8503 into a register (it will be a loop invariant.) */
8504 tem = gen_reg_rtx (GET_MODE (v->new_reg));
8506 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
8509 /* Substitute the new register for its invariant value in
8510 the compare expression. */
8511 XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
8512 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8521 case GT: case GE: case GTU: case GEU:
8522 case LT: case LE: case LTU: case LEU:
8523 /* See if either argument is the biv. */
8524 if (XEXP (x, 0) == reg)
8525 arg = XEXP (x, 1), arg_operand = 1;
8526 else if (XEXP (x, 1) == reg)
8527 arg = XEXP (x, 0), arg_operand = 0;
8531 if (CONSTANT_P (arg))
8533 /* First try to replace with any giv that has constant positive
8534 mult_val and constant add_val. We might be able to support
8535 negative mult_val, but it seems complex to do it in general. */
8537 for (v = bl->giv; v; v = v->next_iv)
8538 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8539 && (GET_CODE (v->add_val) == SYMBOL_REF
8540 || GET_CODE (v->add_val) == LABEL_REF
8541 || GET_CODE (v->add_val) == CONST
8542 || (GET_CODE (v->add_val) == REG
8543 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
8544 && ! v->ignore && ! v->maybe_dead && v->always_computable
8547 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8553 /* Replace biv with the giv's reduced reg. */
8554 XEXP (x, 1-arg_operand) = v->new_reg;
8556 /* If all constants are actually constant integers and
8557 the derived constant can be directly placed in the COMPARE,
8559 if (GET_CODE (arg) == CONST_INT
8560 && GET_CODE (v->mult_val) == CONST_INT
8561 && GET_CODE (v->add_val) == CONST_INT
8562 && validate_change (insn, &XEXP (x, arg_operand),
8563 GEN_INT (INTVAL (arg)
8564 * INTVAL (v->mult_val)
8565 + INTVAL (v->add_val)), 0))
8568 /* Otherwise, load it into a register. */
8569 tem = gen_reg_rtx (mode);
8570 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8571 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
8574 /* If that failed, put back the change we made above. */
8575 XEXP (x, 1-arg_operand) = reg;
8578 /* Look for giv with positive constant mult_val and nonconst add_val.
8579 Insert insns to calculate new compare value.
8580 ??? Turn this off due to possible overflow. */
8582 for (v = bl->giv; v; v = v->next_iv)
8583 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8584 && ! v->ignore && ! v->maybe_dead && v->always_computable
8590 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8596 tem = gen_reg_rtx (mode);
8598 /* Replace biv with giv's reduced register. */
8599 validate_change (insn, &XEXP (x, 1 - arg_operand),
8602 /* Compute value to compare against. */
8603 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8604 /* Use it in this insn. */
8605 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8606 if (apply_change_group ())
8610 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
8612 if (invariant_p (arg) == 1)
8614 /* Look for giv with constant positive mult_val and nonconst
8615 add_val. Insert insns to compute new compare value.
8616 ??? Turn this off due to possible overflow. */
8618 for (v = bl->giv; v; v = v->next_iv)
8619 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8620 && ! v->ignore && ! v->maybe_dead && v->always_computable
8626 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8632 tem = gen_reg_rtx (mode);
8634 /* Replace biv with giv's reduced register. */
8635 validate_change (insn, &XEXP (x, 1 - arg_operand),
8638 /* Compute value to compare against. */
8639 emit_iv_add_mult (arg, v->mult_val, v->add_val,
8641 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8642 if (apply_change_group ())
8647 /* This code has problems. Basically, you can't know when
8648 seeing if we will eliminate BL, whether a particular giv
8649 of ARG will be reduced. If it isn't going to be reduced,
8650 we can't eliminate BL. We can try forcing it to be reduced,
8651 but that can generate poor code.
8653 The problem is that the benefit of reducing TV, below should
8654 be increased if BL can actually be eliminated, but this means
8655 we might have to do a topological sort of the order in which
8656 we try to process biv. It doesn't seem worthwhile to do
8657 this sort of thing now. */
8660 /* Otherwise the reg compared with had better be a biv. */
8661 if (GET_CODE (arg) != REG
8662 || REG_IV_TYPE (REGNO (arg)) != BASIC_INDUCT)
8665 /* Look for a pair of givs, one for each biv,
8666 with identical coefficients. */
8667 for (v = bl->giv; v; v = v->next_iv)
8669 struct induction *tv;
8671 if (v->ignore || v->maybe_dead || v->mode != mode)
8674 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
8675 if (! tv->ignore && ! tv->maybe_dead
8676 && rtx_equal_p (tv->mult_val, v->mult_val)
8677 && rtx_equal_p (tv->add_val, v->add_val)
8678 && tv->mode == mode)
8680 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8686 /* Replace biv with its giv's reduced reg. */
8687 XEXP (x, 1-arg_operand) = v->new_reg;
8688 /* Replace other operand with the other giv's
8690 XEXP (x, arg_operand) = tv->new_reg;
8697 /* If we get here, the biv can't be eliminated. */
8701 /* If this address is a DEST_ADDR giv, it doesn't matter if the
8702 biv is used in it, since it will be replaced. */
8703 for (v = bl->giv; v; v = v->next_iv)
8704 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
8712 /* See if any subexpression fails elimination. */
8713 fmt = GET_RTX_FORMAT (code);
8714 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
8719 if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl,
8720 eliminate_p, where))
8725 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8726 if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl,
8727 eliminate_p, where))
8736 /* Return nonzero if the last use of REG
8737 is in an insn following INSN in the same basic block. */
8740 last_use_this_basic_block (reg, insn)
8746 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
8749 if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
8755 /* Called via `note_stores' to record the initial value of a biv. Here we
8756 just record the location of the set and process it later. */
8759 record_initial (dest, set)
8763 struct iv_class *bl;
8765 if (GET_CODE (dest) != REG
8766 || REGNO (dest) >= max_reg_before_loop
8767 || REG_IV_TYPE (REGNO (dest)) != BASIC_INDUCT)
8770 bl = reg_biv_class[REGNO (dest)];
8772 /* If this is the first set found, record it. */
8773 if (bl->init_insn == 0)
8775 bl->init_insn = note_insn;
8780 /* If any of the registers in X are "old" and currently have a last use earlier
8781 than INSN, update them to have a last use of INSN. Their actual last use
8782 will be the previous insn but it will not have a valid uid_luid so we can't
8786 update_reg_last_use (x, insn)
8790 /* Check for the case where INSN does not have a valid luid. In this case,
8791 there is no need to modify the regno_last_uid, as this can only happen
8792 when code is inserted after the loop_end to set a pseudo's final value,
8793 and hence this insn will never be the last use of x. */
8794 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
8795 && INSN_UID (insn) < max_uid_for_loop
8796 && uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)])
8797 REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
8801 register const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
8802 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
8805 update_reg_last_use (XEXP (x, i), insn);
8806 else if (fmt[i] == 'E')
8807 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8808 update_reg_last_use (XVECEXP (x, i, j), insn);
8813 /* Given a jump insn JUMP, return the condition that will cause it to branch
8814 to its JUMP_LABEL. If the condition cannot be understood, or is an
8815 inequality floating-point comparison which needs to be reversed, 0 will
8818 If EARLIEST is non-zero, it is a pointer to a place where the earliest
8819 insn used in locating the condition was found. If a replacement test
8820 of the condition is desired, it should be placed in front of that
8821 insn and we will be sure that the inputs are still valid.
8823 The condition will be returned in a canonical form to simplify testing by
8824 callers. Specifically:
8826 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
8827 (2) Both operands will be machine operands; (cc0) will have been replaced.
8828 (3) If an operand is a constant, it will be the second operand.
8829 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
8830 for GE, GEU, and LEU. */
8833 get_condition (jump, earliest)
8842 int reverse_code = 0;
8843 int did_reverse_condition = 0;
8844 enum machine_mode mode;
8846 /* If this is not a standard conditional jump, we can't parse it. */
8847 if (GET_CODE (jump) != JUMP_INSN
8848 || ! condjump_p (jump) || simplejump_p (jump))
8851 code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0));
8852 mode = GET_MODE (XEXP (SET_SRC (PATTERN (jump)), 0));
8853 op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0);
8854 op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1);
8859 /* If this branches to JUMP_LABEL when the condition is false, reverse
8861 if (GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF
8862 && XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump))
8863 code = reverse_condition (code), did_reverse_condition ^= 1;
8865 /* If we are comparing a register with zero, see if the register is set
8866 in the previous insn to a COMPARE or a comparison operation. Perform
8867 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
8870 while (GET_RTX_CLASS (code) == '<' && op1 == CONST0_RTX (GET_MODE (op0)))
8872 /* Set non-zero when we find something of interest. */
8876 /* If comparison with cc0, import actual comparison from compare
8880 if ((prev = prev_nonnote_insn (prev)) == 0
8881 || GET_CODE (prev) != INSN
8882 || (set = single_set (prev)) == 0
8883 || SET_DEST (set) != cc0_rtx)
8886 op0 = SET_SRC (set);
8887 op1 = CONST0_RTX (GET_MODE (op0));
8893 /* If this is a COMPARE, pick up the two things being compared. */
8894 if (GET_CODE (op0) == COMPARE)
8896 op1 = XEXP (op0, 1);
8897 op0 = XEXP (op0, 0);
8900 else if (GET_CODE (op0) != REG)
8903 /* Go back to the previous insn. Stop if it is not an INSN. We also
8904 stop if it isn't a single set or if it has a REG_INC note because
8905 we don't want to bother dealing with it. */
8907 if ((prev = prev_nonnote_insn (prev)) == 0
8908 || GET_CODE (prev) != INSN
8909 || FIND_REG_INC_NOTE (prev, 0)
8910 || (set = single_set (prev)) == 0)
8913 /* If this is setting OP0, get what it sets it to if it looks
8915 if (rtx_equal_p (SET_DEST (set), op0))
8917 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
8919 /* ??? We may not combine comparisons done in a CCmode with
8920 comparisons not done in a CCmode. This is to aid targets
8921 like Alpha that have an IEEE compliant EQ instruction, and
8922 a non-IEEE compliant BEQ instruction. The use of CCmode is
8923 actually artificial, simply to prevent the combination, but
8924 should not affect other platforms.
8926 However, we must allow VOIDmode comparisons to match either
8927 CCmode or non-CCmode comparison, because some ports have
8928 modeless comparisons inside branch patterns.
8930 ??? This mode check should perhaps look more like the mode check
8931 in simplify_comparison in combine. */
8933 if ((GET_CODE (SET_SRC (set)) == COMPARE
8936 && GET_MODE_CLASS (inner_mode) == MODE_INT
8937 && (GET_MODE_BITSIZE (inner_mode)
8938 <= HOST_BITS_PER_WIDE_INT)
8939 && (STORE_FLAG_VALUE
8940 & ((HOST_WIDE_INT) 1
8941 << (GET_MODE_BITSIZE (inner_mode) - 1))))
8942 #ifdef FLOAT_STORE_FLAG_VALUE
8944 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
8945 && FLOAT_STORE_FLAG_VALUE < 0)
8948 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'))
8949 && (((GET_MODE_CLASS (mode) == MODE_CC)
8950 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
8951 || mode == VOIDmode || inner_mode == VOIDmode))
8953 else if (((code == EQ
8955 && (GET_MODE_BITSIZE (inner_mode)
8956 <= HOST_BITS_PER_WIDE_INT)
8957 && GET_MODE_CLASS (inner_mode) == MODE_INT
8958 && (STORE_FLAG_VALUE
8959 & ((HOST_WIDE_INT) 1
8960 << (GET_MODE_BITSIZE (inner_mode) - 1))))
8961 #ifdef FLOAT_STORE_FLAG_VALUE
8963 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
8964 && FLOAT_STORE_FLAG_VALUE < 0)
8967 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'
8968 && (((GET_MODE_CLASS (mode) == MODE_CC)
8969 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
8970 || mode == VOIDmode || inner_mode == VOIDmode))
8973 /* We might have reversed a LT to get a GE here. But this wasn't
8974 actually the comparison of data, so we don't flag that we
8975 have had to reverse the condition. */
8976 did_reverse_condition ^= 1;
8984 else if (reg_set_p (op0, prev))
8985 /* If this sets OP0, but not directly, we have to give up. */
8990 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
8991 code = GET_CODE (x);
8994 code = reverse_condition (code);
8995 did_reverse_condition ^= 1;
8999 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
9005 /* If constant is first, put it last. */
9006 if (CONSTANT_P (op0))
9007 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
9009 /* If OP0 is the result of a comparison, we weren't able to find what
9010 was really being compared, so fail. */
9011 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
9014 /* Canonicalize any ordered comparison with integers involving equality
9015 if we can do computations in the relevant mode and we do not
9018 if (GET_CODE (op1) == CONST_INT
9019 && GET_MODE (op0) != VOIDmode
9020 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
9022 HOST_WIDE_INT const_val = INTVAL (op1);
9023 unsigned HOST_WIDE_INT uconst_val = const_val;
9024 unsigned HOST_WIDE_INT max_val
9025 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
9030 if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
9031 code = LT, op1 = GEN_INT (const_val + 1);
9034 /* When cross-compiling, const_val might be sign-extended from
9035 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
9037 if ((HOST_WIDE_INT) (const_val & max_val)
9038 != (((HOST_WIDE_INT) 1
9039 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
9040 code = GT, op1 = GEN_INT (const_val - 1);
9044 if (uconst_val < max_val)
9045 code = LTU, op1 = GEN_INT (uconst_val + 1);
9049 if (uconst_val != 0)
9050 code = GTU, op1 = GEN_INT (uconst_val - 1);
9058 /* If this was floating-point and we reversed anything other than an
9059 EQ or NE, return zero. */
9060 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
9061 && did_reverse_condition && code != NE && code != EQ
9063 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
9067 /* Never return CC0; return zero instead. */
9072 return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);
9075 /* Similar to above routine, except that we also put an invariant last
9076 unless both operands are invariants. */
9079 get_condition_for_loop (x)
9082 rtx comparison = get_condition (x, NULL_PTR);
9085 || ! invariant_p (XEXP (comparison, 0))
9086 || invariant_p (XEXP (comparison, 1)))
9089 return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode,
9090 XEXP (comparison, 1), XEXP (comparison, 0));
9093 #ifdef HAVE_decrement_and_branch_on_count
9094 /* Instrument loop for insertion of bct instruction. We distinguish between
9095 loops with compile-time bounds and those with run-time bounds.
9096 Information from loop_iterations() is used to compute compile-time bounds.
9097 Run-time bounds should use loop preconditioning, but currently ignored.
9101 insert_bct (loop_start, loop_end, loop_info)
9102 rtx loop_start, loop_end;
9103 struct loop_info *loop_info;
9106 unsigned HOST_WIDE_INT n_iterations;
9108 int increment_direction, compare_direction;
9110 /* If the loop condition is <= or >=, the number of iteration
9111 is 1 more than the range of the bounds of the loop. */
9112 int add_iteration = 0;
9114 enum machine_mode loop_var_mode = word_mode;
9116 int loop_num = uid_loop_num [INSN_UID (loop_start)];
9118 /* It's impossible to instrument a competely unrolled loop. */
9119 if (loop_info->unroll_number == -1)
9122 /* Make sure that the count register is not in use. */
9123 if (loop_used_count_register [loop_num])
9125 if (loop_dump_stream)
9126 fprintf (loop_dump_stream,
9127 "insert_bct %d: BCT instrumentation failed: count register already in use\n",
9132 /* Make sure that the function has no indirect jumps. */
9133 if (indirect_jump_in_function)
9135 if (loop_dump_stream)
9136 fprintf (loop_dump_stream,
9137 "insert_bct %d: BCT instrumentation failed: indirect jump in function\n",
9142 /* Make sure that the last loop insn is a conditional jump. */
9143 if (GET_CODE (PREV_INSN (loop_end)) != JUMP_INSN
9144 || ! condjump_p (PREV_INSN (loop_end))
9145 || simplejump_p (PREV_INSN (loop_end)))
9147 if (loop_dump_stream)
9148 fprintf (loop_dump_stream,
9149 "insert_bct %d: BCT instrumentation failed: invalid jump at loop end\n",
9154 /* Make sure that the loop does not contain a function call
9155 (the count register might be altered by the called function). */
9158 if (loop_dump_stream)
9159 fprintf (loop_dump_stream,
9160 "insert_bct %d: BCT instrumentation failed: function call in loop\n",
9165 /* Make sure that the loop does not jump via a table.
9166 (the count register might be used to perform the branch on table). */
9167 if (loop_has_tablejump)
9169 if (loop_dump_stream)
9170 fprintf (loop_dump_stream,
9171 "insert_bct %d: BCT instrumentation failed: computed branch in the loop\n",
9176 /* Account for loop unrolling in instrumented iteration count. */
9177 if (loop_info->unroll_number > 1)
9178 n_iterations = loop_info->n_iterations / loop_info->unroll_number;
9180 n_iterations = loop_info->n_iterations;
9182 if (n_iterations != 0 && n_iterations < 3)
9184 /* Allow an enclosing outer loop to benefit if possible. */
9185 if (loop_dump_stream)
9186 fprintf (loop_dump_stream,
9187 "insert_bct %d: Too few iterations to benefit from BCT optimization\n",
9192 /* Try to instrument the loop. */
9194 /* Handle the simpler case, where the bounds are known at compile time. */
9195 if (n_iterations > 0)
9197 /* Mark all enclosing loops that they cannot use count register. */
9198 for (i = loop_num; i != -1; i = loop_outer_loop[i])
9199 loop_used_count_register[i] = 1;
9200 instrument_loop_bct (loop_start, loop_end, GEN_INT (n_iterations));
9204 /* Handle the more complex case, that the bounds are NOT known
9205 at compile time. In this case we generate run_time calculation
9206 of the number of iterations. */
9208 if (loop_info->iteration_var == 0)
9210 if (loop_dump_stream)
9211 fprintf (loop_dump_stream,
9212 "insert_bct %d: BCT Runtime Instrumentation failed: no loop iteration variable found\n",
9217 if (GET_MODE_CLASS (GET_MODE (loop_info->iteration_var)) != MODE_INT
9218 || GET_MODE_SIZE (GET_MODE (loop_info->iteration_var)) != UNITS_PER_WORD)
9220 if (loop_dump_stream)
9221 fprintf (loop_dump_stream,
9222 "insert_bct %d: BCT Runtime Instrumentation failed: loop variable not integer\n",
9227 /* With runtime bounds, if the compare is of the form '!=' we give up */
9228 if (loop_info->comparison_code == NE)
9230 if (loop_dump_stream)
9231 fprintf (loop_dump_stream,
9232 "insert_bct %d: BCT Runtime Instrumentation failed: runtime bounds with != comparison\n",
9236 /* Use common loop preconditioning code instead. */
9240 /* We rely on the existence of run-time guard to ensure that the
9241 loop executes at least once. */
9243 rtx iterations_num_reg;
9245 unsigned HOST_WIDE_INT increment_value_abs
9246 = INTVAL (increment) * increment_direction;
9248 /* make sure that the increment is a power of two, otherwise (an
9249 expensive) divide is needed. */
9250 if (exact_log2 (increment_value_abs) == -1)
9252 if (loop_dump_stream)
9253 fprintf (loop_dump_stream,
9254 "insert_bct: not instrumenting BCT because the increment is not power of 2\n");
9258 /* compute the number of iterations */
9263 /* Again, the number of iterations is calculated by:
9265 ; compare-val - initial-val + (increment -1) + additional-iteration
9266 ; num_iterations = -----------------------------------------------------------------
9269 /* ??? Do we have to call copy_rtx here before passing rtx to
9271 if (compare_direction > 0)
9273 /* <, <= :the loop variable is increasing */
9274 temp_reg = expand_binop (loop_var_mode, sub_optab,
9275 comparison_value, initial_value,
9276 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9280 temp_reg = expand_binop (loop_var_mode, sub_optab,
9281 initial_value, comparison_value,
9282 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9285 if (increment_value_abs - 1 + add_iteration != 0)
9286 temp_reg = expand_binop (loop_var_mode, add_optab, temp_reg,
9287 GEN_INT (increment_value_abs - 1
9289 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9291 if (increment_value_abs != 1)
9292 iterations_num_reg = expand_binop (loop_var_mode, asr_optab,
9294 GEN_INT (exact_log2 (increment_value_abs)),
9295 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9297 iterations_num_reg = temp_reg;
9299 sequence = gen_sequence ();
9301 emit_insn_before (sequence, loop_start);
9302 instrument_loop_bct (loop_start, loop_end, iterations_num_reg);
9306 #endif /* Complex case */
9309 /* Instrument loop by inserting a bct in it as follows:
9310 1. A new counter register is created.
9311 2. In the head of the loop the new variable is initialized to the value
9312 passed in the loop_num_iterations parameter.
9313 3. At the end of the loop, comparison of the register with 0 is generated.
9314 The created comparison follows the pattern defined for the
9315 decrement_and_branch_on_count insn, so this insn will be generated.
9316 4. The branch on the old variable are deleted. The compare must remain
9317 because it might be used elsewhere. If the loop-variable or condition
9318 register are used elsewhere, they will be eliminated by flow. */
9321 instrument_loop_bct (loop_start, loop_end, loop_num_iterations)
9322 rtx loop_start, loop_end;
9323 rtx loop_num_iterations;
9329 if (HAVE_decrement_and_branch_on_count)
9331 if (loop_dump_stream)
9333 fputs ("instrument_bct: Inserting BCT (", loop_dump_stream);
9334 if (GET_CODE (loop_num_iterations) == CONST_INT)
9335 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
9336 INTVAL (loop_num_iterations));
9338 fputs ("runtime", loop_dump_stream);
9339 fputs (" iterations)", loop_dump_stream);
9342 /* Discard original jump to continue loop. Original compare result
9343 may still be live, so it cannot be discarded explicitly. */
9344 delete_insn (PREV_INSN (loop_end));
9346 /* Insert the label which will delimit the start of the loop. */
9347 start_label = gen_label_rtx ();
9348 emit_label_after (start_label, loop_start);
9350 /* Insert initialization of the count register into the loop header. */
9352 counter_reg = gen_reg_rtx (word_mode);
9353 emit_insn (gen_move_insn (counter_reg, loop_num_iterations));
9354 sequence = gen_sequence ();
9356 emit_insn_before (sequence, loop_start);
9358 /* Insert new comparison on the count register instead of the
9359 old one, generating the needed BCT pattern (that will be
9360 later recognized by assembly generation phase). */
9361 emit_jump_insn_before (gen_decrement_and_branch_on_count (counter_reg,
9364 LABEL_NUSES (start_label)++;
9368 #endif /* HAVE_decrement_and_branch_on_count */
9370 /* Scan the function and determine whether it has indirect (computed) jumps.
9372 This is taken mostly from flow.c; similar code exists elsewhere
9373 in the compiler. It may be useful to put this into rtlanal.c. */
9375 indirect_jump_in_function_p (start)
9380 for (insn = start; insn; insn = NEXT_INSN (insn))
9381 if (computed_jump_p (insn))
9387 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
9388 documentation for LOOP_MEMS for the definition of `appropriate'.
9389 This function is called from prescan_loop via for_each_rtx. */
9392 insert_loop_mem (mem, data)
9394 void *data ATTRIBUTE_UNUSED;
9402 switch (GET_CODE (m))
9408 /* We're not interested in the MEM associated with a
9409 CONST_DOUBLE, so there's no need to traverse into this. */
9413 /* This is not a MEM. */
9417 /* See if we've already seen this MEM. */
9418 for (i = 0; i < loop_mems_idx; ++i)
9419 if (rtx_equal_p (m, loop_mems[i].mem))
9421 if (GET_MODE (m) != GET_MODE (loop_mems[i].mem))
9422 /* The modes of the two memory accesses are different. If
9423 this happens, something tricky is going on, and we just
9424 don't optimize accesses to this MEM. */
9425 loop_mems[i].optimize = 0;
9430 /* Resize the array, if necessary. */
9431 if (loop_mems_idx == loop_mems_allocated)
9433 if (loop_mems_allocated != 0)
9434 loop_mems_allocated *= 2;
9436 loop_mems_allocated = 32;
9438 loop_mems = (loop_mem_info*)
9439 xrealloc (loop_mems,
9440 loop_mems_allocated * sizeof (loop_mem_info));
9443 /* Actually insert the MEM. */
9444 loop_mems[loop_mems_idx].mem = m;
9445 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
9446 because we can't put it in a register. We still store it in the
9447 table, though, so that if we see the same address later, but in a
9448 non-BLK mode, we'll not think we can optimize it at that point. */
9449 loop_mems[loop_mems_idx].optimize = (GET_MODE (m) != BLKmode);
9450 loop_mems[loop_mems_idx].reg = NULL_RTX;
9456 /* Like load_mems, but also ensures that SET_IN_LOOP,
9457 MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct
9458 values after load_mems. */
9461 load_mems_and_recount_loop_regs_set (scan_start, end, loop_top, start,
9469 int nregs = max_reg_num ();
9471 load_mems (scan_start, end, loop_top, start);
9473 /* Recalculate set_in_loop and friends since load_mems may have
9474 created new registers. */
9475 if (max_reg_num () > nregs)
9481 nregs = max_reg_num ();
9483 if ((unsigned) nregs > set_in_loop->num_elements)
9485 /* Grow all the arrays. */
9486 VARRAY_GROW (set_in_loop, nregs);
9487 VARRAY_GROW (n_times_set, nregs);
9488 VARRAY_GROW (may_not_optimize, nregs);
9489 VARRAY_GROW (reg_single_usage, nregs);
9491 /* Clear the arrays */
9492 bzero ((char *) &set_in_loop->data, nregs * sizeof (int));
9493 bzero ((char *) &may_not_optimize->data, nregs * sizeof (char));
9494 bzero ((char *) ®_single_usage->data, nregs * sizeof (rtx));
9496 count_loop_regs_set (loop_top ? loop_top : start, end,
9497 may_not_optimize, reg_single_usage,
9500 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
9502 VARRAY_CHAR (may_not_optimize, i) = 1;
9503 VARRAY_INT (set_in_loop, i) = 1;
9506 #ifdef AVOID_CCMODE_COPIES
9507 /* Don't try to move insns which set CC registers if we should not
9508 create CCmode register copies. */
9509 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
9510 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
9511 VARRAY_CHAR (may_not_optimize, i) = 1;
9514 /* Set n_times_set for the new registers. */
9515 bcopy ((char *) (&set_in_loop->data.i[0] + old_nregs),
9516 (char *) (&n_times_set->data.i[0] + old_nregs),
9517 (nregs - old_nregs) * sizeof (int));
9521 /* Move MEMs into registers for the duration of the loop. SCAN_START
9522 is the first instruction in the loop (as it is executed). The
9523 other parameters are as for next_insn_in_loop. */
9526 load_mems (scan_start, end, loop_top, start)
9532 int maybe_never = 0;
9535 rtx label = NULL_RTX;
9536 rtx end_label = NULL_RTX;
9538 if (loop_mems_idx > 0)
9540 /* Nonzero if the next instruction may never be executed. */
9541 int next_maybe_never = 0;
9543 /* Check to see if it's possible that some instructions in the
9544 loop are never executed. */
9545 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
9546 p != NULL_RTX && !maybe_never;
9547 p = next_insn_in_loop (p, scan_start, end, loop_top))
9549 if (GET_CODE (p) == CODE_LABEL)
9551 else if (GET_CODE (p) == JUMP_INSN
9552 /* If we enter the loop in the middle, and scan
9553 around to the beginning, don't set maybe_never
9554 for that. This must be an unconditional jump,
9555 otherwise the code at the top of the loop might
9556 never be executed. Unconditional jumps are
9557 followed a by barrier then loop end. */
9558 && ! (GET_CODE (p) == JUMP_INSN
9559 && JUMP_LABEL (p) == loop_top
9560 && NEXT_INSN (NEXT_INSN (p)) == end
9561 && simplejump_p (p)))
9563 if (!condjump_p (p))
9564 /* Something complicated. */
9567 /* If there are any more instructions in the loop, they
9568 might not be reached. */
9569 next_maybe_never = 1;
9571 else if (next_maybe_never)
9575 /* Actually move the MEMs. */
9576 for (i = 0; i < loop_mems_idx; ++i)
9580 rtx mem = loop_mems[i].mem;
9583 if (MEM_VOLATILE_P (mem)
9584 || invariant_p (XEXP (mem, 0)) != 1)
9585 /* There's no telling whether or not MEM is modified. */
9586 loop_mems[i].optimize = 0;
9588 /* Go through the MEMs written to in the loop to see if this
9589 one is aliased by one of them. */
9590 mem_list_entry = loop_store_mems;
9591 while (mem_list_entry)
9593 if (rtx_equal_p (mem, XEXP (mem_list_entry, 0)))
9595 else if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
9598 /* MEM is indeed aliased by this store. */
9599 loop_mems[i].optimize = 0;
9602 mem_list_entry = XEXP (mem_list_entry, 1);
9605 /* If this MEM is written to, we must be sure that there
9606 are no reads from another MEM that aliases this one. */
9607 if (loop_mems[i].optimize && written)
9611 for (j = 0; j < loop_mems_idx; ++j)
9615 else if (true_dependence (mem,
9620 /* It's not safe to hoist loop_mems[i] out of
9621 the loop because writes to it might not be
9622 seen by reads from loop_mems[j]. */
9623 loop_mems[i].optimize = 0;
9629 if (maybe_never && may_trap_p (mem))
9630 /* We can't access the MEM outside the loop; it might
9631 cause a trap that wouldn't have happened otherwise. */
9632 loop_mems[i].optimize = 0;
9634 if (!loop_mems[i].optimize)
9635 /* We thought we were going to lift this MEM out of the
9636 loop, but later discovered that we could not. */
9639 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
9640 order to keep scan_loop from moving stores to this MEM
9641 out of the loop just because this REG is neither a
9642 user-variable nor used in the loop test. */
9643 reg = gen_reg_rtx (GET_MODE (mem));
9644 REG_USERVAR_P (reg) = 1;
9645 loop_mems[i].reg = reg;
9647 /* Now, replace all references to the MEM with the
9648 corresponding pesudos. */
9649 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
9651 p = next_insn_in_loop (p, scan_start, end, loop_top))
9656 for_each_rtx (&p, replace_loop_mem, &ri);
9659 if (!apply_change_group ())
9660 /* We couldn't replace all occurrences of the MEM. */
9661 loop_mems[i].optimize = 0;
9666 /* Load the memory immediately before START, which is
9667 the NOTE_LOOP_BEG. */
9668 set = gen_move_insn (reg, mem);
9669 emit_insn_before (set, start);
9673 if (label == NULL_RTX)
9675 /* We must compute the former
9676 right-after-the-end label before we insert
9678 end_label = next_label (end);
9679 label = gen_label_rtx ();
9680 emit_label_after (label, end);
9683 /* Store the memory immediately after END, which is
9684 the NOTE_LOOP_END. */
9685 set = gen_move_insn (copy_rtx (mem), reg);
9686 emit_insn_after (set, label);
9689 if (loop_dump_stream)
9691 fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
9692 REGNO (reg), (written ? "r/w" : "r/o"));
9693 print_rtl (loop_dump_stream, mem);
9694 fputc ('\n', loop_dump_stream);
9700 if (label != NULL_RTX)
9702 /* Now, we need to replace all references to the previous exit
9703 label with the new one. */
9708 for (p = start; p != end; p = NEXT_INSN (p))
9710 for_each_rtx (&p, replace_label, &rr);
9712 /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
9713 field. This is not handled by for_each_rtx because it doesn't
9714 handle unprinted ('0') fields. We need to update JUMP_LABEL
9715 because the immediately following unroll pass will use it.
9716 replace_label would not work anyways, because that only handles
9718 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == end_label)
9719 JUMP_LABEL (p) = label;
9724 /* Replace MEM with its associated pseudo register. This function is
9725 called from load_mems via for_each_rtx. DATA is actually an
9726 rtx_and_int * describing the instruction currently being scanned
9727 and the MEM we are currently replacing. */
9730 replace_loop_mem (mem, data)
9742 switch (GET_CODE (m))
9748 /* We're not interested in the MEM associated with a
9749 CONST_DOUBLE, so there's no need to traverse into one. */
9753 /* This is not a MEM. */
9757 ri = (rtx_and_int*) data;
9760 if (!rtx_equal_p (loop_mems[i].mem, m))
9761 /* This is not the MEM we are currently replacing. */
9766 /* Actually replace the MEM. */
9767 validate_change (insn, mem, loop_mems[i].reg, 1);
9772 /* Replace occurrences of the old exit label for the loop with the new
9773 one. DATA is an rtx_pair containing the old and new labels,
9777 replace_label (x, data)
9782 rtx old_label = ((rtx_pair*) data)->r1;
9783 rtx new_label = ((rtx_pair*) data)->r2;
9788 if (GET_CODE (l) != LABEL_REF)
9791 if (XEXP (l, 0) != old_label)
9794 XEXP (l, 0) = new_label;
9795 ++LABEL_NUSES (new_label);
9796 --LABEL_NUSES (old_label);