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. */
44 #include "insn-config.h"
45 #include "insn-flags.h"
47 #include "hard-reg-set.h"
55 /* Information about the loop being processed used to compute
56 the number of loop iterations for loop unrolling and doloop
58 static struct loop_info this_loop_info;
60 /* Vector mapping INSN_UIDs to luids.
61 The luids are like uids but increase monotonically always.
62 We use them to see whether a jump comes from outside a given loop. */
66 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
67 number the insn is contained in. */
71 /* 1 + largest uid of any insn. */
75 /* 1 + luid of last insn. */
79 /* Number of loops detected in current function. Used as index to the
82 static int max_loop_num;
84 /* Indexed by loop number, contains the first and last insn of each loop. */
86 static rtx *loop_number_loop_starts, *loop_number_loop_ends;
88 /* Likewise for the continue insn */
89 static rtx *loop_number_loop_cont;
91 /* The first code_label that is reached in every loop iteration.
92 0 when not computed yet, initially const0_rtx if a jump couldn't be
94 Also set to 0 when there is no such label before the NOTE_INSN_LOOP_CONT
95 of this loop, or in verify_dominator, if a jump couldn't be followed. */
96 static rtx *loop_number_cont_dominator;
98 /* For each loop, gives the containing loop number, -1 if none. */
100 int *loop_outer_loop;
102 #ifdef HAVE_decrement_and_branch_on_count
103 /* Records whether resource in use by inner loop. */
105 int *loop_used_count_register;
106 #endif /* HAVE_decrement_and_branch_on_count */
108 /* Indexed by loop number, contains a nonzero value if the "loop" isn't
109 really a loop (an insn outside the loop branches into it). */
111 static char *loop_invalid;
113 /* Indexed by loop number, links together all LABEL_REFs which refer to
114 code labels outside the loop. Used by routines that need to know all
115 loop exits, such as final_biv_value and final_giv_value.
117 This does not include loop exits due to return instructions. This is
118 because all bivs and givs are pseudos, and hence must be dead after a
119 return, so the presense of a return does not affect any of the
120 optimizations that use this info. It is simpler to just not include return
121 instructions on this list. */
123 rtx *loop_number_exit_labels;
125 /* Indexed by loop number, counts the number of LABEL_REFs on
126 loop_number_exit_labels for this loop and all loops nested inside it. */
128 int *loop_number_exit_count;
130 /* Indexed by register number, contains the number of times the reg
131 is set during the loop being scanned.
132 During code motion, a negative value indicates a reg that has been
133 made a candidate; in particular -2 means that it is an candidate that
134 we know is equal to a constant and -1 means that it is an candidate
135 not known equal to a constant.
136 After code motion, regs moved have 0 (which is accurate now)
137 while the failed candidates have the original number of times set.
139 Therefore, at all times, == 0 indicates an invariant register;
140 < 0 a conditionally invariant one. */
142 static varray_type set_in_loop;
144 /* Original value of set_in_loop; same except that this value
145 is not set negative for a reg whose sets have been made candidates
146 and not set to 0 for a reg that is moved. */
148 static varray_type n_times_set;
150 /* Index by register number, 1 indicates that the register
151 cannot be moved or strength reduced. */
153 static varray_type may_not_optimize;
155 /* Contains the insn in which a register was used if it was used
156 exactly once; contains const0_rtx if it was used more than once. */
158 static varray_type reg_single_usage;
160 /* Nonzero means reg N has already been moved out of one loop.
161 This reduces the desire to move it out of another. */
163 static char *moved_once;
165 /* List of MEMs that are stored in this loop. */
167 static rtx loop_store_mems;
169 /* The insn where the first of these was found. */
170 static rtx first_loop_store_insn;
172 typedef struct loop_mem_info {
173 rtx mem; /* The MEM itself. */
174 rtx reg; /* Corresponding pseudo, if any. */
175 int optimize; /* Nonzero if we can optimize access to this MEM. */
178 /* Array of MEMs that are used (read or written) in this loop, but
179 cannot be aliased by anything in this loop, except perhaps
180 themselves. In other words, if loop_mems[i] is altered during the
181 loop, it is altered by an expression that is rtx_equal_p to it. */
183 static loop_mem_info *loop_mems;
185 /* The index of the next available slot in LOOP_MEMS. */
187 static int loop_mems_idx;
189 /* The number of elements allocated in LOOP_MEMs. */
191 static int loop_mems_allocated;
193 /* Nonzero if we don't know what MEMs were changed in the current
194 loop. This happens if the loop contains a call (in which case
195 `loop_info->has_call' will also be set) or if we store into more
196 than NUM_STORES MEMs. */
198 static int unknown_address_altered;
200 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
201 static int num_movables;
203 /* Count of memory write instructions discovered in the loop. */
204 static int num_mem_sets;
206 /* Bound on pseudo register number before loop optimization.
207 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
208 int max_reg_before_loop;
210 /* This obstack is used in product_cheap_p to allocate its rtl. It
211 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
212 If we used the same obstack that it did, we would be deallocating
215 static struct obstack temp_obstack;
217 /* This is where the pointer to the obstack being used for RTL is stored. */
219 extern struct obstack *rtl_obstack;
221 #define obstack_chunk_alloc xmalloc
222 #define obstack_chunk_free free
224 /* During the analysis of a loop, a chain of `struct movable's
225 is made to record all the movable insns found.
226 Then the entire chain can be scanned to decide which to move. */
230 rtx insn; /* A movable insn */
231 rtx set_src; /* The expression this reg is set from. */
232 rtx set_dest; /* The destination of this SET. */
233 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
234 of any registers used within the LIBCALL. */
235 int consec; /* Number of consecutive following insns
236 that must be moved with this one. */
237 int regno; /* The register it sets */
238 short lifetime; /* lifetime of that register;
239 may be adjusted when matching movables
240 that load the same value are found. */
241 short savings; /* Number of insns we can move for this reg,
242 including other movables that force this
243 or match this one. */
244 unsigned int cond : 1; /* 1 if only conditionally movable */
245 unsigned int force : 1; /* 1 means MUST move this insn */
246 unsigned int global : 1; /* 1 means reg is live outside this loop */
247 /* If PARTIAL is 1, GLOBAL means something different:
248 that the reg is live outside the range from where it is set
249 to the following label. */
250 unsigned int done : 1; /* 1 inhibits further processing of this */
252 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
253 In particular, moving it does not make it
255 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
256 load SRC, rather than copying INSN. */
257 unsigned int move_insn_first:1;/* Same as above, if this is necessary for the
258 first insn of a consecutive sets group. */
259 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
260 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
261 that we should avoid changing when clearing
262 the rest of the reg. */
263 struct movable *match; /* First entry for same value */
264 struct movable *forces; /* An insn that must be moved if this is */
265 struct movable *next;
268 static struct movable *the_movables;
270 FILE *loop_dump_stream;
272 /* For communicating return values from note_set_pseudo_multiple_uses. */
273 static int note_set_pseudo_multiple_uses_retval;
275 /* Forward declarations. */
277 static void verify_dominator PROTO((int));
278 static void find_and_verify_loops PROTO((rtx));
279 static void mark_loop_jump PROTO((rtx, int));
280 static void prescan_loop PROTO((rtx, rtx, struct loop_info *));
281 static int reg_in_basic_block_p PROTO((rtx, rtx));
282 static int consec_sets_invariant_p PROTO((rtx, int, rtx));
283 static int labels_in_range_p PROTO((rtx, int));
284 static void count_one_set PROTO((rtx, rtx, varray_type, rtx *));
286 static void count_loop_regs_set PROTO((rtx, rtx, varray_type, varray_type,
288 static void note_addr_stored PROTO((rtx, rtx));
289 static void note_set_pseudo_multiple_uses PROTO((rtx, rtx));
290 static int loop_reg_used_before_p PROTO((rtx, rtx, rtx, rtx, rtx));
291 static void scan_loop PROTO((rtx, rtx, rtx, int, int));
293 static void replace_call_address PROTO((rtx, rtx, rtx));
295 static rtx skip_consec_insns PROTO((rtx, int));
296 static int libcall_benefit PROTO((rtx));
297 static void ignore_some_movables PROTO((struct movable *));
298 static void force_movables PROTO((struct movable *));
299 static void combine_movables PROTO((struct movable *, int));
300 static int regs_match_p PROTO((rtx, rtx, struct movable *));
301 static int rtx_equal_for_loop_p PROTO((rtx, rtx, struct movable *));
302 static void add_label_notes PROTO((rtx, rtx));
303 static void move_movables PROTO((struct movable *, int, int, rtx, rtx, int));
304 static int count_nonfixed_reads PROTO((rtx));
305 static void strength_reduce PROTO((rtx, rtx, rtx, int, rtx, rtx,
306 struct loop_info *, rtx, int, int));
307 static void find_single_use_in_loop PROTO((rtx, rtx, varray_type));
308 static int valid_initial_value_p PROTO((rtx, rtx, int, rtx));
309 static void find_mem_givs PROTO((rtx, rtx, int, int, rtx, rtx));
310 static void record_biv PROTO((struct induction *, rtx, rtx, rtx, rtx, rtx *, int, int));
311 static void check_final_value PROTO((struct induction *, rtx, rtx,
312 unsigned HOST_WIDE_INT));
313 static void record_giv PROTO((struct induction *, rtx, rtx, rtx, rtx, rtx, int, enum g_types, int, int, rtx *, rtx, rtx));
314 static void update_giv_derive PROTO((rtx));
315 static int basic_induction_var PROTO((rtx, enum machine_mode, rtx, rtx, rtx *, rtx *, rtx **));
316 static rtx simplify_giv_expr PROTO((rtx, int *));
317 static int general_induction_var PROTO((rtx, rtx *, rtx *, rtx *, int, int *));
318 static int consec_sets_giv PROTO((int, rtx, rtx, rtx, rtx *, rtx *, rtx *));
319 static int check_dbra_loop PROTO((rtx, int, rtx, struct loop_info *));
320 static rtx express_from_1 PROTO((rtx, rtx, rtx));
321 static rtx combine_givs_p PROTO((struct induction *, struct induction *));
322 static void combine_givs PROTO((struct iv_class *));
323 struct recombine_givs_stats;
324 static int find_life_end PROTO((rtx, struct recombine_givs_stats *, rtx, rtx));
325 static void recombine_givs PROTO((struct iv_class *, rtx, rtx, int));
326 static int product_cheap_p PROTO((rtx, rtx));
327 static int maybe_eliminate_biv PROTO((struct iv_class *, rtx, rtx, int, int, int));
328 static int maybe_eliminate_biv_1 PROTO((rtx, rtx, struct iv_class *, int, rtx));
329 static int last_use_this_basic_block PROTO((rtx, rtx));
330 static void record_initial PROTO((rtx, rtx));
331 static void update_reg_last_use PROTO((rtx, rtx));
332 static rtx next_insn_in_loop PROTO((rtx, rtx, rtx, rtx));
333 static void load_mems_and_recount_loop_regs_set PROTO((rtx, rtx, rtx,
335 static void load_mems PROTO((rtx, rtx, rtx, rtx));
336 static int insert_loop_mem PROTO((rtx *, void *));
337 static int replace_loop_mem PROTO((rtx *, void *));
338 static int replace_label PROTO((rtx *, void *));
340 typedef struct rtx_and_int {
345 typedef struct rtx_pair {
350 /* Nonzero iff INSN is between START and END, inclusive. */
351 #define INSN_IN_RANGE_P(INSN, START, END) \
352 (INSN_UID (INSN) < max_uid_for_loop \
353 && INSN_LUID (INSN) >= INSN_LUID (START) \
354 && INSN_LUID (INSN) <= INSN_LUID (END))
356 #ifdef HAVE_decrement_and_branch_on_count
357 /* Test whether BCT applicable and safe. */
358 static void insert_bct PROTO((rtx, rtx, struct loop_info *));
360 /* Auxiliary function that inserts the BCT pattern into the loop. */
361 static void instrument_loop_bct PROTO((rtx, rtx, rtx));
362 #endif /* HAVE_decrement_and_branch_on_count */
364 /* Indirect_jump_in_function is computed once per function. */
365 int indirect_jump_in_function = 0;
366 static int indirect_jump_in_function_p PROTO((rtx));
368 static int compute_luids PROTO((rtx, rtx, int));
370 static int biv_elimination_giv_has_0_offset PROTO((struct induction *,
371 struct induction *, rtx));
373 /* Relative gain of eliminating various kinds of operations. */
376 static int shift_cost;
377 static int mult_cost;
380 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
381 copy the value of the strength reduced giv to its original register. */
382 static int copy_cost;
384 /* Cost of using a register, to normalize the benefits of a giv. */
385 static int reg_address_cost;
391 char *free_point = (char *) oballoc (1);
392 rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
394 add_cost = rtx_cost (gen_rtx_PLUS (word_mode, reg, reg), SET);
397 reg_address_cost = ADDRESS_COST (reg);
399 reg_address_cost = rtx_cost (reg, MEM);
402 /* We multiply by 2 to reconcile the difference in scale between
403 these two ways of computing costs. Otherwise the cost of a copy
404 will be far less than the cost of an add. */
408 /* Free the objects we just allocated. */
411 /* Initialize the obstack used for rtl in product_cheap_p. */
412 gcc_obstack_init (&temp_obstack);
415 /* Compute the mapping from uids to luids.
416 LUIDs are numbers assigned to insns, like uids,
417 except that luids increase monotonically through the code.
418 Start at insn START and stop just before END. Assign LUIDs
419 starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
421 compute_luids (start, end, prev_luid)
428 for (insn = start, i = prev_luid; insn != end; insn = NEXT_INSN (insn))
430 if (INSN_UID (insn) >= max_uid_for_loop)
432 /* Don't assign luids to line-number NOTEs, so that the distance in
433 luids between two insns is not affected by -g. */
434 if (GET_CODE (insn) != NOTE
435 || NOTE_LINE_NUMBER (insn) <= 0)
436 uid_luid[INSN_UID (insn)] = ++i;
438 /* Give a line number note the same luid as preceding insn. */
439 uid_luid[INSN_UID (insn)] = i;
444 /* Entry point of this file. Perform loop optimization
445 on the current function. F is the first insn of the function
446 and DUMPFILE is a stream for output of a trace of actions taken
447 (or 0 if none should be output). */
450 loop_optimize (f, dumpfile, unroll_p, bct_p)
451 /* f is the first instruction of a chain of insns for one function */
459 loop_dump_stream = dumpfile;
461 init_recog_no_volatile ();
463 max_reg_before_loop = max_reg_num ();
465 moved_once = (char *) alloca (max_reg_before_loop);
466 bzero (moved_once, max_reg_before_loop);
470 /* Count the number of loops. */
473 for (insn = f; insn; insn = NEXT_INSN (insn))
475 if (GET_CODE (insn) == NOTE
476 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
480 /* Don't waste time if no loops. */
481 if (max_loop_num == 0)
484 /* Get size to use for tables indexed by uids.
485 Leave some space for labels allocated by find_and_verify_loops. */
486 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
488 uid_luid = (int *) alloca (max_uid_for_loop * sizeof (int));
489 uid_loop_num = (int *) alloca (max_uid_for_loop * sizeof (int));
491 bzero ((char *) uid_luid, max_uid_for_loop * sizeof (int));
492 bzero ((char *) uid_loop_num, max_uid_for_loop * sizeof (int));
494 /* Allocate tables for recording each loop. We set each entry, so they need
496 loop_number_loop_starts = (rtx *) alloca (max_loop_num * sizeof (rtx));
497 loop_number_loop_ends = (rtx *) alloca (max_loop_num * sizeof (rtx));
498 loop_number_loop_cont = (rtx *) alloca (max_loop_num * sizeof (rtx));
499 loop_number_cont_dominator = (rtx *) alloca (max_loop_num * sizeof (rtx));
500 loop_outer_loop = (int *) alloca (max_loop_num * sizeof (int));
501 loop_invalid = (char *) alloca (max_loop_num * sizeof (char));
502 loop_number_exit_labels = (rtx *) alloca (max_loop_num * sizeof (rtx));
503 loop_number_exit_count = (int *) alloca (max_loop_num * sizeof (int));
505 #ifdef HAVE_decrement_and_branch_on_count
506 /* Allocate for BCT optimization */
507 loop_used_count_register = (int *) alloca (max_loop_num * sizeof (int));
508 bzero ((char *) loop_used_count_register, max_loop_num * sizeof (int));
509 #endif /* HAVE_decrement_and_branch_on_count */
511 /* Find and process each loop.
512 First, find them, and record them in order of their beginnings. */
513 find_and_verify_loops (f);
515 /* Now find all register lifetimes. This must be done after
516 find_and_verify_loops, because it might reorder the insns in the
518 reg_scan (f, max_reg_num (), 1);
520 /* This must occur after reg_scan so that registers created by gcse
521 will have entries in the register tables.
523 We could have added a call to reg_scan after gcse_main in toplev.c,
524 but moving this call to init_alias_analysis is more efficient. */
525 init_alias_analysis ();
527 /* See if we went too far. Note that get_max_uid already returns
528 one more that the maximum uid of all insn. */
529 if (get_max_uid () > max_uid_for_loop)
531 /* Now reset it to the actual size we need. See above. */
532 max_uid_for_loop = get_max_uid ();
534 /* find_and_verify_loops has already called compute_luids, but it might
535 have rearranged code afterwards, so we need to recompute the luids now. */
536 max_luid = compute_luids (f, NULL_RTX, 0);
538 /* Don't leave gaps in uid_luid for insns that have been
539 deleted. It is possible that the first or last insn
540 using some register has been deleted by cross-jumping.
541 Make sure that uid_luid for that former insn's uid
542 points to the general area where that insn used to be. */
543 for (i = 0; i < max_uid_for_loop; i++)
545 uid_luid[0] = uid_luid[i];
546 if (uid_luid[0] != 0)
549 for (i = 0; i < max_uid_for_loop; i++)
550 if (uid_luid[i] == 0)
551 uid_luid[i] = uid_luid[i - 1];
553 /* Create a mapping from loops to BLOCK tree nodes. */
554 if (unroll_p && write_symbols != NO_DEBUG)
555 find_loop_tree_blocks ();
557 /* Determine if the function has indirect jump. On some systems
558 this prevents low overhead loop instructions from being used. */
559 indirect_jump_in_function = indirect_jump_in_function_p (f);
561 /* Now scan the loops, last ones first, since this means inner ones are done
562 before outer ones. */
563 for (i = max_loop_num-1; i >= 0; i--)
564 if (! loop_invalid[i] && loop_number_loop_ends[i])
565 scan_loop (loop_number_loop_starts[i], loop_number_loop_ends[i],
566 loop_number_loop_cont[i], unroll_p, bct_p);
568 /* If debugging and unrolling loops, we must replicate the tree nodes
569 corresponding to the blocks inside the loop, so that the original one
570 to one mapping will remain. */
571 if (unroll_p && write_symbols != NO_DEBUG)
572 unroll_block_trees ();
574 end_alias_analysis ();
577 /* Returns the next insn, in execution order, after INSN. START and
578 END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
579 respectively. LOOP_TOP, if non-NULL, is the top of the loop in the
580 insn-stream; it is used with loops that are entered near the
584 next_insn_in_loop (insn, start, end, loop_top)
590 insn = NEXT_INSN (insn);
595 /* Go to the top of the loop, and continue there. */
609 /* Optimize one loop whose start is LOOP_START and end is END.
610 LOOP_START is the NOTE_INSN_LOOP_BEG and END is the matching
612 LOOP_CONT is the NOTE_INSN_LOOP_CONT. */
614 /* ??? Could also move memory writes out of loops if the destination address
615 is invariant, the source is invariant, the memory write is not volatile,
616 and if we can prove that no read inside the loop can read this address
617 before the write occurs. If there is a read of this address after the
618 write, then we can also mark the memory read as invariant. */
621 scan_loop (loop_start, end, loop_cont, unroll_p, bct_p)
622 rtx loop_start, end, loop_cont;
627 /* 1 if we are scanning insns that could be executed zero times. */
629 /* 1 if we are scanning insns that might never be executed
630 due to a subroutine call which might exit before they are reached. */
632 /* For a rotated loop that is entered near the bottom,
633 this is the label at the top. Otherwise it is zero. */
635 /* Jump insn that enters the loop, or 0 if control drops in. */
636 rtx loop_entry_jump = 0;
637 /* Place in the loop where control enters. */
639 /* Number of insns in the loop. */
644 /* The SET from an insn, if it is the only SET in the insn. */
646 /* Chain describing insns movable in current loop. */
647 struct movable *movables = 0;
648 /* Last element in `movables' -- so we can add elements at the end. */
649 struct movable *last_movable = 0;
650 /* Ratio of extra register life span we can justify
651 for saving an instruction. More if loop doesn't call subroutines
652 since in that case saving an insn makes more difference
653 and more registers are available. */
655 /* Nonzero if we are scanning instructions in a sub-loop. */
658 struct loop_info *loop_info = &this_loop_info;
660 /* Determine whether this loop starts with a jump down to a test at
661 the end. This will occur for a small number of loops with a test
662 that is too complex to duplicate in front of the loop.
664 We search for the first insn or label in the loop, skipping NOTEs.
665 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
666 (because we might have a loop executed only once that contains a
667 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
668 (in case we have a degenerate loop).
670 Note that if we mistakenly think that a loop is entered at the top
671 when, in fact, it is entered at the exit test, the only effect will be
672 slightly poorer optimization. Making the opposite error can generate
673 incorrect code. Since very few loops now start with a jump to the
674 exit test, the code here to detect that case is very conservative. */
676 for (p = NEXT_INSN (loop_start);
678 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
679 && (GET_CODE (p) != NOTE
680 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
681 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
687 /* Set up variables describing this loop. */
688 prescan_loop (loop_start, end, loop_info);
689 threshold = (loop_info->has_call ? 1 : 2) * (1 + n_non_fixed_regs);
691 /* If loop has a jump before the first label,
692 the true entry is the target of that jump.
693 Start scan from there.
694 But record in LOOP_TOP the place where the end-test jumps
695 back to so we can scan that after the end of the loop. */
696 if (GET_CODE (p) == JUMP_INSN)
700 /* Loop entry must be unconditional jump (and not a RETURN) */
702 && JUMP_LABEL (p) != 0
703 /* Check to see whether the jump actually
704 jumps out of the loop (meaning it's no loop).
705 This case can happen for things like
706 do {..} while (0). If this label was generated previously
707 by loop, we can't tell anything about it and have to reject
709 && INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, end))
711 loop_top = next_label (scan_start);
712 scan_start = JUMP_LABEL (p);
716 /* If SCAN_START was an insn created by loop, we don't know its luid
717 as required by loop_reg_used_before_p. So skip such loops. (This
718 test may never be true, but it's best to play it safe.)
720 Also, skip loops where we do not start scanning at a label. This
721 test also rejects loops starting with a JUMP_INSN that failed the
724 if (INSN_UID (scan_start) >= max_uid_for_loop
725 || GET_CODE (scan_start) != CODE_LABEL)
727 if (loop_dump_stream)
728 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
729 INSN_UID (loop_start), INSN_UID (end));
733 /* Count number of times each reg is set during this loop.
734 Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out
735 the setting of register I. Set VARRAY_RTX (reg_single_usage, I). */
737 /* Allocate extra space for REGS that might be created by
738 load_mems. We allocate a little extra slop as well, in the hopes
739 that even after the moving of movables creates some new registers
740 we won't have to reallocate these arrays. However, we do grow
741 the arrays, if necessary, in load_mems_recount_loop_regs_set. */
742 nregs = max_reg_num () + loop_mems_idx + 16;
743 VARRAY_INT_INIT (set_in_loop, nregs, "set_in_loop");
744 VARRAY_INT_INIT (n_times_set, nregs, "n_times_set");
745 VARRAY_CHAR_INIT (may_not_optimize, nregs, "may_not_optimize");
746 VARRAY_RTX_INIT (reg_single_usage, nregs, "reg_single_usage");
748 count_loop_regs_set (loop_top ? loop_top : loop_start, end,
749 may_not_optimize, reg_single_usage, &insn_count, nregs);
751 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
753 VARRAY_CHAR (may_not_optimize, i) = 1;
754 VARRAY_INT (set_in_loop, i) = 1;
757 #ifdef AVOID_CCMODE_COPIES
758 /* Don't try to move insns which set CC registers if we should not
759 create CCmode register copies. */
760 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
761 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
762 VARRAY_CHAR (may_not_optimize, i) = 1;
765 bcopy ((char *) &set_in_loop->data,
766 (char *) &n_times_set->data, nregs * sizeof (int));
768 if (loop_dump_stream)
770 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
771 INSN_UID (loop_start), INSN_UID (end), insn_count);
773 fprintf (loop_dump_stream, "Continue at insn %d.\n",
774 INSN_UID (loop_info->cont));
777 /* Scan through the loop finding insns that are safe to move.
778 Set set_in_loop negative for the reg being set, so that
779 this reg will be considered invariant for subsequent insns.
780 We consider whether subsequent insns use the reg
781 in deciding whether it is worth actually moving.
783 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
784 and therefore it is possible that the insns we are scanning
785 would never be executed. At such times, we must make sure
786 that it is safe to execute the insn once instead of zero times.
787 When MAYBE_NEVER is 0, all insns will be executed at least once
788 so that is not a problem. */
790 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
792 p = next_insn_in_loop (p, scan_start, end, loop_top))
794 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
795 && find_reg_note (p, REG_LIBCALL, NULL_RTX))
797 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
798 && find_reg_note (p, REG_RETVAL, NULL_RTX))
801 if (GET_CODE (p) == INSN
802 && (set = single_set (p))
803 && GET_CODE (SET_DEST (set)) == REG
804 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
809 rtx src = SET_SRC (set);
810 rtx dependencies = 0;
812 /* Figure out what to use as a source of this insn. If a REG_EQUIV
813 note is given or if a REG_EQUAL note with a constant operand is
814 specified, use it as the source and mark that we should move
815 this insn by calling emit_move_insn rather that duplicating the
818 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
820 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
822 src = XEXP (temp, 0), move_insn = 1;
825 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
826 if (temp && CONSTANT_P (XEXP (temp, 0)))
827 src = XEXP (temp, 0), move_insn = 1;
828 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
830 src = XEXP (temp, 0);
831 /* A libcall block can use regs that don't appear in
832 the equivalent expression. To move the libcall,
833 we must move those regs too. */
834 dependencies = libcall_other_reg (p, src);
838 /* Don't try to optimize a register that was made
839 by loop-optimization for an inner loop.
840 We don't know its life-span, so we can't compute the benefit. */
841 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
843 else if (/* The register is used in basic blocks other
844 than the one where it is set (meaning that
845 something after this point in the loop might
846 depend on its value before the set). */
847 ! reg_in_basic_block_p (p, SET_DEST (set))
848 /* And the set is not guaranteed to be executed one
849 the loop starts, or the value before the set is
850 needed before the set occurs...
852 ??? Note we have quadratic behaviour here, mitigated
853 by the fact that the previous test will often fail for
854 large loops. Rather than re-scanning the entire loop
855 each time for register usage, we should build tables
856 of the register usage and use them here instead. */
858 || loop_reg_used_before_p (set, p, loop_start,
860 /* It is unsafe to move the set.
862 This code used to consider it OK to move a set of a variable
863 which was not created by the user and not used in an exit test.
864 That behavior is incorrect and was removed. */
866 else if ((tem = invariant_p (src))
867 && (dependencies == 0
868 || (tem2 = invariant_p (dependencies)) != 0)
869 && (VARRAY_INT (set_in_loop,
870 REGNO (SET_DEST (set))) == 1
872 = consec_sets_invariant_p
874 VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))),
876 /* If the insn can cause a trap (such as divide by zero),
877 can't move it unless it's guaranteed to be executed
878 once loop is entered. Even a function call might
879 prevent the trap insn from being reached
880 (since it might exit!) */
881 && ! ((maybe_never || call_passed)
882 && may_trap_p (src)))
884 register struct movable *m;
885 register int regno = REGNO (SET_DEST (set));
887 /* A potential lossage is where we have a case where two insns
888 can be combined as long as they are both in the loop, but
889 we move one of them outside the loop. For large loops,
890 this can lose. The most common case of this is the address
891 of a function being called.
893 Therefore, if this register is marked as being used exactly
894 once if we are in a loop with calls (a "large loop"), see if
895 we can replace the usage of this register with the source
896 of this SET. If we can, delete this insn.
898 Don't do this if P has a REG_RETVAL note or if we have
899 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
901 if (loop_info->has_call
902 && VARRAY_RTX (reg_single_usage, regno) != 0
903 && VARRAY_RTX (reg_single_usage, regno) != const0_rtx
904 && REGNO_FIRST_UID (regno) == INSN_UID (p)
905 && (REGNO_LAST_UID (regno)
906 == INSN_UID (VARRAY_RTX (reg_single_usage, regno)))
907 && VARRAY_INT (set_in_loop, regno) == 1
908 && ! side_effects_p (SET_SRC (set))
909 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
910 && (! SMALL_REGISTER_CLASSES
911 || (! (GET_CODE (SET_SRC (set)) == REG
912 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)))
913 /* This test is not redundant; SET_SRC (set) might be
914 a call-clobbered register and the life of REGNO
915 might span a call. */
916 && ! modified_between_p (SET_SRC (set), p,
918 (reg_single_usage, regno))
919 && no_labels_between_p (p, VARRAY_RTX (reg_single_usage, regno))
920 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
922 (reg_single_usage, regno)))
924 /* Replace any usage in a REG_EQUAL note. Must copy the
925 new source, so that we don't get rtx sharing between the
926 SET_SOURCE and REG_NOTES of insn p. */
927 REG_NOTES (VARRAY_RTX (reg_single_usage, regno))
928 = replace_rtx (REG_NOTES (VARRAY_RTX
929 (reg_single_usage, regno)),
930 SET_DEST (set), copy_rtx (SET_SRC (set)));
933 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
934 NOTE_SOURCE_FILE (p) = 0;
935 VARRAY_INT (set_in_loop, regno) = 0;
939 m = (struct movable *) alloca (sizeof (struct movable));
943 m->dependencies = dependencies;
944 m->set_dest = SET_DEST (set);
946 m->consec = VARRAY_INT (set_in_loop,
947 REGNO (SET_DEST (set))) - 1;
951 m->move_insn = move_insn;
952 m->move_insn_first = 0;
953 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
954 m->savemode = VOIDmode;
956 /* Set M->cond if either invariant_p or consec_sets_invariant_p
957 returned 2 (only conditionally invariant). */
958 m->cond = ((tem | tem1 | tem2) > 1);
959 m->global = (uid_luid[REGNO_LAST_UID (regno)] > INSN_LUID (end)
960 || uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (loop_start));
962 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
963 - uid_luid[REGNO_FIRST_UID (regno)]);
964 m->savings = VARRAY_INT (n_times_set, regno);
965 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
966 m->savings += libcall_benefit (p);
967 VARRAY_INT (set_in_loop, regno) = move_insn ? -2 : -1;
968 /* Add M to the end of the chain MOVABLES. */
972 last_movable->next = m;
977 /* It is possible for the first instruction to have a
978 REG_EQUAL note but a non-invariant SET_SRC, so we must
979 remember the status of the first instruction in case
980 the last instruction doesn't have a REG_EQUAL note. */
981 m->move_insn_first = m->move_insn;
983 /* Skip this insn, not checking REG_LIBCALL notes. */
984 p = next_nonnote_insn (p);
985 /* Skip the consecutive insns, if there are any. */
986 p = skip_consec_insns (p, m->consec);
987 /* Back up to the last insn of the consecutive group. */
988 p = prev_nonnote_insn (p);
990 /* We must now reset m->move_insn, m->is_equiv, and possibly
991 m->set_src to correspond to the effects of all the
993 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
995 m->set_src = XEXP (temp, 0), m->move_insn = 1;
998 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
999 if (temp && CONSTANT_P (XEXP (temp, 0)))
1000 m->set_src = XEXP (temp, 0), m->move_insn = 1;
1005 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
1008 /* If this register is always set within a STRICT_LOW_PART
1009 or set to zero, then its high bytes are constant.
1010 So clear them outside the loop and within the loop
1011 just load the low bytes.
1012 We must check that the machine has an instruction to do so.
1013 Also, if the value loaded into the register
1014 depends on the same register, this cannot be done. */
1015 else if (SET_SRC (set) == const0_rtx
1016 && GET_CODE (NEXT_INSN (p)) == INSN
1017 && (set1 = single_set (NEXT_INSN (p)))
1018 && GET_CODE (set1) == SET
1019 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
1020 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
1021 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
1023 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
1025 register int regno = REGNO (SET_DEST (set));
1026 if (VARRAY_INT (set_in_loop, regno) == 2)
1028 register struct movable *m;
1029 m = (struct movable *) alloca (sizeof (struct movable));
1032 m->set_dest = SET_DEST (set);
1033 m->dependencies = 0;
1039 m->move_insn_first = 0;
1041 /* If the insn may not be executed on some cycles,
1042 we can't clear the whole reg; clear just high part.
1043 Not even if the reg is used only within this loop.
1050 Clearing x before the inner loop could clobber a value
1051 being saved from the last time around the outer loop.
1052 However, if the reg is not used outside this loop
1053 and all uses of the register are in the same
1054 basic block as the store, there is no problem.
1056 If this insn was made by loop, we don't know its
1057 INSN_LUID and hence must make a conservative
1059 m->global = (INSN_UID (p) >= max_uid_for_loop
1060 || (uid_luid[REGNO_LAST_UID (regno)]
1062 || (uid_luid[REGNO_FIRST_UID (regno)]
1064 || (labels_in_range_p
1065 (p, uid_luid[REGNO_FIRST_UID (regno)])));
1066 if (maybe_never && m->global)
1067 m->savemode = GET_MODE (SET_SRC (set1));
1069 m->savemode = VOIDmode;
1073 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
1074 - uid_luid[REGNO_FIRST_UID (regno)]);
1076 VARRAY_INT (set_in_loop, regno) = -1;
1077 /* Add M to the end of the chain MOVABLES. */
1081 last_movable->next = m;
1086 /* Past a call insn, we get to insns which might not be executed
1087 because the call might exit. This matters for insns that trap.
1088 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
1089 so they don't count. */
1090 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
1092 /* Past a label or a jump, we get to insns for which we
1093 can't count on whether or how many times they will be
1094 executed during each iteration. Therefore, we can
1095 only move out sets of trivial variables
1096 (those not used after the loop). */
1097 /* Similar code appears twice in strength_reduce. */
1098 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
1099 /* If we enter the loop in the middle, and scan around to the
1100 beginning, don't set maybe_never for that. This must be an
1101 unconditional jump, otherwise the code at the top of the
1102 loop might never be executed. Unconditional jumps are
1103 followed a by barrier then loop end. */
1104 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
1105 && NEXT_INSN (NEXT_INSN (p)) == end
1106 && simplejump_p (p)))
1108 else if (GET_CODE (p) == NOTE)
1110 /* At the virtual top of a converted loop, insns are again known to
1111 be executed: logically, the loop begins here even though the exit
1112 code has been duplicated. */
1113 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
1114 maybe_never = call_passed = 0;
1115 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
1117 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
1122 /* If one movable subsumes another, ignore that other. */
1124 ignore_some_movables (movables);
1126 /* For each movable insn, see if the reg that it loads
1127 leads when it dies right into another conditionally movable insn.
1128 If so, record that the second insn "forces" the first one,
1129 since the second can be moved only if the first is. */
1131 force_movables (movables);
1133 /* See if there are multiple movable insns that load the same value.
1134 If there are, make all but the first point at the first one
1135 through the `match' field, and add the priorities of them
1136 all together as the priority of the first. */
1138 combine_movables (movables, nregs);
1140 /* Now consider each movable insn to decide whether it is worth moving.
1141 Store 0 in set_in_loop for each reg that is moved.
1143 Generally this increases code size, so do not move moveables when
1144 optimizing for code size. */
1146 if (! optimize_size)
1147 move_movables (movables, threshold,
1148 insn_count, loop_start, end, nregs);
1150 /* Now candidates that still are negative are those not moved.
1151 Change set_in_loop to indicate that those are not actually invariant. */
1152 for (i = 0; i < nregs; i++)
1153 if (VARRAY_INT (set_in_loop, i) < 0)
1154 VARRAY_INT (set_in_loop, i) = VARRAY_INT (n_times_set, i);
1156 /* Now that we've moved some things out of the loop, we might be able to
1157 hoist even more memory references. */
1158 load_mems_and_recount_loop_regs_set (scan_start, end, loop_top,
1159 loop_start, &insn_count);
1161 if (flag_strength_reduce)
1163 the_movables = movables;
1164 strength_reduce (scan_start, end, loop_top,
1165 insn_count, loop_start, end,
1166 loop_info, loop_cont, unroll_p, bct_p);
1169 VARRAY_FREE (reg_single_usage);
1170 VARRAY_FREE (set_in_loop);
1171 VARRAY_FREE (n_times_set);
1172 VARRAY_FREE (may_not_optimize);
1175 /* Add elements to *OUTPUT to record all the pseudo-regs
1176 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1179 record_excess_regs (in_this, not_in_this, output)
1180 rtx in_this, not_in_this;
1187 code = GET_CODE (in_this);
1201 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1202 && ! reg_mentioned_p (in_this, not_in_this))
1203 *output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output);
1210 fmt = GET_RTX_FORMAT (code);
1211 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1218 for (j = 0; j < XVECLEN (in_this, i); j++)
1219 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1223 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1229 /* Check what regs are referred to in the libcall block ending with INSN,
1230 aside from those mentioned in the equivalent value.
1231 If there are none, return 0.
1232 If there are one or more, return an EXPR_LIST containing all of them. */
1235 libcall_other_reg (insn, equiv)
1238 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1239 rtx p = XEXP (note, 0);
1242 /* First, find all the regs used in the libcall block
1243 that are not mentioned as inputs to the result. */
1247 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1248 || GET_CODE (p) == CALL_INSN)
1249 record_excess_regs (PATTERN (p), equiv, &output);
1256 /* Return 1 if all uses of REG
1257 are between INSN and the end of the basic block. */
1260 reg_in_basic_block_p (insn, reg)
1263 int regno = REGNO (reg);
1266 if (REGNO_FIRST_UID (regno) != INSN_UID (insn))
1269 /* Search this basic block for the already recorded last use of the reg. */
1270 for (p = insn; p; p = NEXT_INSN (p))
1272 switch (GET_CODE (p))
1279 /* Ordinary insn: if this is the last use, we win. */
1280 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1285 /* Jump insn: if this is the last use, we win. */
1286 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1288 /* Otherwise, it's the end of the basic block, so we lose. */
1293 /* It's the end of the basic block, so we lose. */
1301 /* The "last use" doesn't follow the "first use"?? */
1305 /* Compute the benefit of eliminating the insns in the block whose
1306 last insn is LAST. This may be a group of insns used to compute a
1307 value directly or can contain a library call. */
1310 libcall_benefit (last)
1316 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1317 insn != last; insn = NEXT_INSN (insn))
1319 if (GET_CODE (insn) == CALL_INSN)
1320 benefit += 10; /* Assume at least this many insns in a library
1322 else if (GET_CODE (insn) == INSN
1323 && GET_CODE (PATTERN (insn)) != USE
1324 && GET_CODE (PATTERN (insn)) != CLOBBER)
1331 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1334 skip_consec_insns (insn, count)
1338 for (; count > 0; count--)
1342 /* If first insn of libcall sequence, skip to end. */
1343 /* Do this at start of loop, since INSN is guaranteed to
1345 if (GET_CODE (insn) != NOTE
1346 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1347 insn = XEXP (temp, 0);
1349 do insn = NEXT_INSN (insn);
1350 while (GET_CODE (insn) == NOTE);
1356 /* Ignore any movable whose insn falls within a libcall
1357 which is part of another movable.
1358 We make use of the fact that the movable for the libcall value
1359 was made later and so appears later on the chain. */
1362 ignore_some_movables (movables)
1363 struct movable *movables;
1365 register struct movable *m, *m1;
1367 for (m = movables; m; m = m->next)
1369 /* Is this a movable for the value of a libcall? */
1370 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1374 /* Check for earlier movables inside that range,
1375 and mark them invalid. We cannot use LUIDs here because
1376 insns created by loop.c for prior loops don't have LUIDs.
1377 Rather than reject all such insns from movables, we just
1378 explicitly check each insn in the libcall (since invariant
1379 libcalls aren't that common). */
1380 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1381 for (m1 = movables; m1 != m; m1 = m1->next)
1382 if (m1->insn == insn)
1388 /* For each movable insn, see if the reg that it loads
1389 leads when it dies right into another conditionally movable insn.
1390 If so, record that the second insn "forces" the first one,
1391 since the second can be moved only if the first is. */
1394 force_movables (movables)
1395 struct movable *movables;
1397 register struct movable *m, *m1;
1398 for (m1 = movables; m1; m1 = m1->next)
1399 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1400 if (!m1->partial && !m1->done)
1402 int regno = m1->regno;
1403 for (m = m1->next; m; m = m->next)
1404 /* ??? Could this be a bug? What if CSE caused the
1405 register of M1 to be used after this insn?
1406 Since CSE does not update regno_last_uid,
1407 this insn M->insn might not be where it dies.
1408 But very likely this doesn't matter; what matters is
1409 that M's reg is computed from M1's reg. */
1410 if (INSN_UID (m->insn) == REGNO_LAST_UID (regno)
1413 if (m != 0 && m->set_src == m1->set_dest
1414 /* If m->consec, m->set_src isn't valid. */
1418 /* Increase the priority of the moving the first insn
1419 since it permits the second to be moved as well. */
1423 m1->lifetime += m->lifetime;
1424 m1->savings += m->savings;
1429 /* Find invariant expressions that are equal and can be combined into
1433 combine_movables (movables, nregs)
1434 struct movable *movables;
1437 register struct movable *m;
1438 char *matched_regs = (char *) alloca (nregs);
1439 enum machine_mode mode;
1441 /* Regs that are set more than once are not allowed to match
1442 or be matched. I'm no longer sure why not. */
1443 /* Perhaps testing m->consec_sets would be more appropriate here? */
1445 for (m = movables; m; m = m->next)
1446 if (m->match == 0 && VARRAY_INT (n_times_set, m->regno) == 1 && !m->partial)
1448 register struct movable *m1;
1449 int regno = m->regno;
1451 bzero (matched_regs, nregs);
1452 matched_regs[regno] = 1;
1454 /* We want later insns to match the first one. Don't make the first
1455 one match any later ones. So start this loop at m->next. */
1456 for (m1 = m->next; m1; m1 = m1->next)
1457 if (m != m1 && m1->match == 0 && VARRAY_INT (n_times_set, m1->regno) == 1
1458 /* A reg used outside the loop mustn't be eliminated. */
1460 /* A reg used for zero-extending mustn't be eliminated. */
1462 && (matched_regs[m1->regno]
1465 /* Can combine regs with different modes loaded from the
1466 same constant only if the modes are the same or
1467 if both are integer modes with M wider or the same
1468 width as M1. The check for integer is redundant, but
1469 safe, since the only case of differing destination
1470 modes with equal sources is when both sources are
1471 VOIDmode, i.e., CONST_INT. */
1472 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1473 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1474 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1475 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1476 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1477 /* See if the source of M1 says it matches M. */
1478 && ((GET_CODE (m1->set_src) == REG
1479 && matched_regs[REGNO (m1->set_src)])
1480 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1482 && ((m->dependencies == m1->dependencies)
1483 || rtx_equal_p (m->dependencies, m1->dependencies)))
1485 m->lifetime += m1->lifetime;
1486 m->savings += m1->savings;
1489 matched_regs[m1->regno] = 1;
1493 /* Now combine the regs used for zero-extension.
1494 This can be done for those not marked `global'
1495 provided their lives don't overlap. */
1497 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1498 mode = GET_MODE_WIDER_MODE (mode))
1500 register struct movable *m0 = 0;
1502 /* Combine all the registers for extension from mode MODE.
1503 Don't combine any that are used outside this loop. */
1504 for (m = movables; m; m = m->next)
1505 if (m->partial && ! m->global
1506 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1508 register struct movable *m1;
1509 int first = uid_luid[REGNO_FIRST_UID (m->regno)];
1510 int last = uid_luid[REGNO_LAST_UID (m->regno)];
1514 /* First one: don't check for overlap, just record it. */
1519 /* Make sure they extend to the same mode.
1520 (Almost always true.) */
1521 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1524 /* We already have one: check for overlap with those
1525 already combined together. */
1526 for (m1 = movables; m1 != m; m1 = m1->next)
1527 if (m1 == m0 || (m1->partial && m1->match == m0))
1528 if (! (uid_luid[REGNO_FIRST_UID (m1->regno)] > last
1529 || uid_luid[REGNO_LAST_UID (m1->regno)] < first))
1532 /* No overlap: we can combine this with the others. */
1533 m0->lifetime += m->lifetime;
1534 m0->savings += m->savings;
1543 /* Return 1 if regs X and Y will become the same if moved. */
1546 regs_match_p (x, y, movables)
1548 struct movable *movables;
1552 struct movable *mx, *my;
1554 for (mx = movables; mx; mx = mx->next)
1555 if (mx->regno == xn)
1558 for (my = movables; my; my = my->next)
1559 if (my->regno == yn)
1563 && ((mx->match == my->match && mx->match != 0)
1565 || mx == my->match));
1568 /* Return 1 if X and Y are identical-looking rtx's.
1569 This is the Lisp function EQUAL for rtx arguments.
1571 If two registers are matching movables or a movable register and an
1572 equivalent constant, consider them equal. */
1575 rtx_equal_for_loop_p (x, y, movables)
1577 struct movable *movables;
1581 register struct movable *m;
1582 register enum rtx_code code;
1583 register const char *fmt;
1587 if (x == 0 || y == 0)
1590 code = GET_CODE (x);
1592 /* If we have a register and a constant, they may sometimes be
1594 if (GET_CODE (x) == REG && VARRAY_INT (set_in_loop, REGNO (x)) == -2
1597 for (m = movables; m; m = m->next)
1598 if (m->move_insn && m->regno == REGNO (x)
1599 && rtx_equal_p (m->set_src, y))
1602 else if (GET_CODE (y) == REG && VARRAY_INT (set_in_loop, REGNO (y)) == -2
1605 for (m = movables; m; m = m->next)
1606 if (m->move_insn && m->regno == REGNO (y)
1607 && rtx_equal_p (m->set_src, x))
1611 /* Otherwise, rtx's of different codes cannot be equal. */
1612 if (code != GET_CODE (y))
1615 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1616 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1618 if (GET_MODE (x) != GET_MODE (y))
1621 /* These three types of rtx's can be compared nonrecursively. */
1623 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1625 if (code == LABEL_REF)
1626 return XEXP (x, 0) == XEXP (y, 0);
1627 if (code == SYMBOL_REF)
1628 return XSTR (x, 0) == XSTR (y, 0);
1630 /* Compare the elements. If any pair of corresponding elements
1631 fail to match, return 0 for the whole things. */
1633 fmt = GET_RTX_FORMAT (code);
1634 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1639 if (XWINT (x, i) != XWINT (y, i))
1644 if (XINT (x, i) != XINT (y, i))
1649 /* Two vectors must have the same length. */
1650 if (XVECLEN (x, i) != XVECLEN (y, i))
1653 /* And the corresponding elements must match. */
1654 for (j = 0; j < XVECLEN (x, i); j++)
1655 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1660 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1665 if (strcmp (XSTR (x, i), XSTR (y, i)))
1670 /* These are just backpointers, so they don't matter. */
1676 /* It is believed that rtx's at this level will never
1677 contain anything but integers and other rtx's,
1678 except for within LABEL_REFs and SYMBOL_REFs. */
1686 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1687 insns in INSNS which use the reference. */
1690 add_label_notes (x, insns)
1694 enum rtx_code code = GET_CODE (x);
1699 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1701 /* This code used to ignore labels that referred to dispatch tables to
1702 avoid flow generating (slighly) worse code.
1704 We no longer ignore such label references (see LABEL_REF handling in
1705 mark_jump_label for additional information). */
1706 for (insn = insns; insn; insn = NEXT_INSN (insn))
1707 if (reg_mentioned_p (XEXP (x, 0), insn))
1708 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
1712 fmt = GET_RTX_FORMAT (code);
1713 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1716 add_label_notes (XEXP (x, i), insns);
1717 else if (fmt[i] == 'E')
1718 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1719 add_label_notes (XVECEXP (x, i, j), insns);
1723 /* Scan MOVABLES, and move the insns that deserve to be moved.
1724 If two matching movables are combined, replace one reg with the
1725 other throughout. */
1728 move_movables (movables, threshold, insn_count, loop_start, end, nregs)
1729 struct movable *movables;
1737 register struct movable *m;
1739 /* Map of pseudo-register replacements to handle combining
1740 when we move several insns that load the same value
1741 into different pseudo-registers. */
1742 rtx *reg_map = (rtx *) alloca (nregs * sizeof (rtx));
1743 char *already_moved = (char *) alloca (nregs);
1745 bzero (already_moved, nregs);
1746 bzero ((char *) reg_map, nregs * sizeof (rtx));
1750 for (m = movables; m; m = m->next)
1752 /* Describe this movable insn. */
1754 if (loop_dump_stream)
1756 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1757 INSN_UID (m->insn), m->regno, m->lifetime);
1759 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1761 fprintf (loop_dump_stream, "cond ");
1763 fprintf (loop_dump_stream, "force ");
1765 fprintf (loop_dump_stream, "global ");
1767 fprintf (loop_dump_stream, "done ");
1769 fprintf (loop_dump_stream, "move-insn ");
1771 fprintf (loop_dump_stream, "matches %d ",
1772 INSN_UID (m->match->insn));
1774 fprintf (loop_dump_stream, "forces %d ",
1775 INSN_UID (m->forces->insn));
1778 /* Count movables. Value used in heuristics in strength_reduce. */
1781 /* Ignore the insn if it's already done (it matched something else).
1782 Otherwise, see if it is now safe to move. */
1786 || (1 == invariant_p (m->set_src)
1787 && (m->dependencies == 0
1788 || 1 == invariant_p (m->dependencies))
1790 || 1 == consec_sets_invariant_p (m->set_dest,
1793 && (! m->forces || m->forces->done))
1797 int savings = m->savings;
1799 /* We have an insn that is safe to move.
1800 Compute its desirability. */
1805 if (loop_dump_stream)
1806 fprintf (loop_dump_stream, "savings %d ", savings);
1808 if (moved_once[regno] && loop_dump_stream)
1809 fprintf (loop_dump_stream, "halved since already moved ");
1811 /* An insn MUST be moved if we already moved something else
1812 which is safe only if this one is moved too: that is,
1813 if already_moved[REGNO] is nonzero. */
1815 /* An insn is desirable to move if the new lifetime of the
1816 register is no more than THRESHOLD times the old lifetime.
1817 If it's not desirable, it means the loop is so big
1818 that moving won't speed things up much,
1819 and it is liable to make register usage worse. */
1821 /* It is also desirable to move if it can be moved at no
1822 extra cost because something else was already moved. */
1824 if (already_moved[regno]
1825 || flag_move_all_movables
1826 || (threshold * savings * m->lifetime) >=
1827 (moved_once[regno] ? insn_count * 2 : insn_count)
1828 || (m->forces && m->forces->done
1829 && VARRAY_INT (n_times_set, m->forces->regno) == 1))
1832 register struct movable *m1;
1833 rtx first = NULL_RTX;
1835 /* Now move the insns that set the reg. */
1837 if (m->partial && m->match)
1841 /* Find the end of this chain of matching regs.
1842 Thus, we load each reg in the chain from that one reg.
1843 And that reg is loaded with 0 directly,
1844 since it has ->match == 0. */
1845 for (m1 = m; m1->match; m1 = m1->match);
1846 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1847 SET_DEST (PATTERN (m1->insn)));
1848 i1 = emit_insn_before (newpat, loop_start);
1850 /* Mark the moved, invariant reg as being allowed to
1851 share a hard reg with the other matching invariant. */
1852 REG_NOTES (i1) = REG_NOTES (m->insn);
1853 r1 = SET_DEST (PATTERN (m->insn));
1854 r2 = SET_DEST (PATTERN (m1->insn));
1856 = gen_rtx_EXPR_LIST (VOIDmode, r1,
1857 gen_rtx_EXPR_LIST (VOIDmode, r2,
1859 delete_insn (m->insn);
1864 if (loop_dump_stream)
1865 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1867 /* If we are to re-generate the item being moved with a
1868 new move insn, first delete what we have and then emit
1869 the move insn before the loop. */
1870 else if (m->move_insn)
1874 for (count = m->consec; count >= 0; count--)
1876 /* If this is the first insn of a library call sequence,
1878 if (GET_CODE (p) != NOTE
1879 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1882 /* If this is the last insn of a libcall sequence, then
1883 delete every insn in the sequence except the last.
1884 The last insn is handled in the normal manner. */
1885 if (GET_CODE (p) != NOTE
1886 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1888 temp = XEXP (temp, 0);
1890 temp = delete_insn (temp);
1894 p = delete_insn (p);
1896 /* simplify_giv_expr expects that it can walk the insns
1897 at m->insn forwards and see this old sequence we are
1898 tossing here. delete_insn does preserve the next
1899 pointers, but when we skip over a NOTE we must fix
1900 it up. Otherwise that code walks into the non-deleted
1902 while (p && GET_CODE (p) == NOTE)
1903 p = NEXT_INSN (temp) = NEXT_INSN (p);
1907 emit_move_insn (m->set_dest, m->set_src);
1908 temp = get_insns ();
1911 add_label_notes (m->set_src, temp);
1913 i1 = emit_insns_before (temp, loop_start);
1914 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1916 = gen_rtx_EXPR_LIST (m->is_equiv ? REG_EQUIV : REG_EQUAL,
1917 m->set_src, REG_NOTES (i1));
1919 if (loop_dump_stream)
1920 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1922 /* The more regs we move, the less we like moving them. */
1927 for (count = m->consec; count >= 0; count--)
1931 /* If first insn of libcall sequence, skip to end. */
1932 /* Do this at start of loop, since p is guaranteed to
1934 if (GET_CODE (p) != NOTE
1935 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1938 /* If last insn of libcall sequence, move all
1939 insns except the last before the loop. The last
1940 insn is handled in the normal manner. */
1941 if (GET_CODE (p) != NOTE
1942 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1946 rtx fn_address_insn = 0;
1949 for (temp = XEXP (temp, 0); temp != p;
1950 temp = NEXT_INSN (temp))
1956 if (GET_CODE (temp) == NOTE)
1959 body = PATTERN (temp);
1961 /* Find the next insn after TEMP,
1962 not counting USE or NOTE insns. */
1963 for (next = NEXT_INSN (temp); next != p;
1964 next = NEXT_INSN (next))
1965 if (! (GET_CODE (next) == INSN
1966 && GET_CODE (PATTERN (next)) == USE)
1967 && GET_CODE (next) != NOTE)
1970 /* If that is the call, this may be the insn
1971 that loads the function address.
1973 Extract the function address from the insn
1974 that loads it into a register.
1975 If this insn was cse'd, we get incorrect code.
1977 So emit a new move insn that copies the
1978 function address into the register that the
1979 call insn will use. flow.c will delete any
1980 redundant stores that we have created. */
1981 if (GET_CODE (next) == CALL_INSN
1982 && GET_CODE (body) == SET
1983 && GET_CODE (SET_DEST (body)) == REG
1984 && (n = find_reg_note (temp, REG_EQUAL,
1987 fn_reg = SET_SRC (body);
1988 if (GET_CODE (fn_reg) != REG)
1989 fn_reg = SET_DEST (body);
1990 fn_address = XEXP (n, 0);
1991 fn_address_insn = temp;
1993 /* We have the call insn.
1994 If it uses the register we suspect it might,
1995 load it with the correct address directly. */
1996 if (GET_CODE (temp) == CALL_INSN
1998 && reg_referenced_p (fn_reg, body))
1999 emit_insn_after (gen_move_insn (fn_reg,
2003 if (GET_CODE (temp) == CALL_INSN)
2005 i1 = emit_call_insn_before (body, loop_start);
2006 /* Because the USAGE information potentially
2007 contains objects other than hard registers
2008 we need to copy it. */
2009 if (CALL_INSN_FUNCTION_USAGE (temp))
2010 CALL_INSN_FUNCTION_USAGE (i1)
2011 = copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
2014 i1 = emit_insn_before (body, loop_start);
2017 if (temp == fn_address_insn)
2018 fn_address_insn = i1;
2019 REG_NOTES (i1) = REG_NOTES (temp);
2025 if (m->savemode != VOIDmode)
2027 /* P sets REG to zero; but we should clear only
2028 the bits that are not covered by the mode
2030 rtx reg = m->set_dest;
2036 (GET_MODE (reg), and_optab, reg,
2037 GEN_INT ((((HOST_WIDE_INT) 1
2038 << GET_MODE_BITSIZE (m->savemode)))
2040 reg, 1, OPTAB_LIB_WIDEN);
2044 emit_move_insn (reg, tem);
2045 sequence = gen_sequence ();
2047 i1 = emit_insn_before (sequence, loop_start);
2049 else if (GET_CODE (p) == CALL_INSN)
2051 i1 = emit_call_insn_before (PATTERN (p), loop_start);
2052 /* Because the USAGE information potentially
2053 contains objects other than hard registers
2054 we need to copy it. */
2055 if (CALL_INSN_FUNCTION_USAGE (p))
2056 CALL_INSN_FUNCTION_USAGE (i1)
2057 = copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
2059 else if (count == m->consec && m->move_insn_first)
2061 /* The SET_SRC might not be invariant, so we must
2062 use the REG_EQUAL note. */
2064 emit_move_insn (m->set_dest, m->set_src);
2065 temp = get_insns ();
2068 add_label_notes (m->set_src, temp);
2070 i1 = emit_insns_before (temp, loop_start);
2071 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
2073 = gen_rtx_EXPR_LIST ((m->is_equiv ? REG_EQUIV
2075 m->set_src, REG_NOTES (i1));
2078 i1 = emit_insn_before (PATTERN (p), loop_start);
2080 if (REG_NOTES (i1) == 0)
2082 REG_NOTES (i1) = REG_NOTES (p);
2084 /* If there is a REG_EQUAL note present whose value
2085 is not loop invariant, then delete it, since it
2086 may cause problems with later optimization passes.
2087 It is possible for cse to create such notes
2088 like this as a result of record_jump_cond. */
2090 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
2091 && ! invariant_p (XEXP (temp, 0)))
2092 remove_note (i1, temp);
2098 if (loop_dump_stream)
2099 fprintf (loop_dump_stream, " moved to %d",
2102 /* If library call, now fix the REG_NOTES that contain
2103 insn pointers, namely REG_LIBCALL on FIRST
2104 and REG_RETVAL on I1. */
2105 if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX)))
2107 XEXP (temp, 0) = first;
2108 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
2109 XEXP (temp, 0) = i1;
2116 /* simplify_giv_expr expects that it can walk the insns
2117 at m->insn forwards and see this old sequence we are
2118 tossing here. delete_insn does preserve the next
2119 pointers, but when we skip over a NOTE we must fix
2120 it up. Otherwise that code walks into the non-deleted
2122 while (p && GET_CODE (p) == NOTE)
2123 p = NEXT_INSN (temp) = NEXT_INSN (p);
2126 /* The more regs we move, the less we like moving them. */
2130 /* Any other movable that loads the same register
2132 already_moved[regno] = 1;
2134 /* This reg has been moved out of one loop. */
2135 moved_once[regno] = 1;
2137 /* The reg set here is now invariant. */
2139 VARRAY_INT (set_in_loop, regno) = 0;
2143 /* Change the length-of-life info for the register
2144 to say it lives at least the full length of this loop.
2145 This will help guide optimizations in outer loops. */
2147 if (uid_luid[REGNO_FIRST_UID (regno)] > INSN_LUID (loop_start))
2148 /* This is the old insn before all the moved insns.
2149 We can't use the moved insn because it is out of range
2150 in uid_luid. Only the old insns have luids. */
2151 REGNO_FIRST_UID (regno) = INSN_UID (loop_start);
2152 if (uid_luid[REGNO_LAST_UID (regno)] < INSN_LUID (end))
2153 REGNO_LAST_UID (regno) = INSN_UID (end);
2155 /* Combine with this moved insn any other matching movables. */
2158 for (m1 = movables; m1; m1 = m1->next)
2163 /* Schedule the reg loaded by M1
2164 for replacement so that shares the reg of M.
2165 If the modes differ (only possible in restricted
2166 circumstances, make a SUBREG.
2168 Note this assumes that the target dependent files
2169 treat REG and SUBREG equally, including within
2170 GO_IF_LEGITIMATE_ADDRESS and in all the
2171 predicates since we never verify that replacing the
2172 original register with a SUBREG results in a
2173 recognizable insn. */
2174 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
2175 reg_map[m1->regno] = m->set_dest;
2178 = gen_lowpart_common (GET_MODE (m1->set_dest),
2181 /* Get rid of the matching insn
2182 and prevent further processing of it. */
2185 /* if library call, delete all insn except last, which
2187 if ((temp = find_reg_note (m1->insn, REG_RETVAL,
2190 for (temp = XEXP (temp, 0); temp != m1->insn;
2191 temp = NEXT_INSN (temp))
2194 delete_insn (m1->insn);
2196 /* Any other movable that loads the same register
2198 already_moved[m1->regno] = 1;
2200 /* The reg merged here is now invariant,
2201 if the reg it matches is invariant. */
2203 VARRAY_INT (set_in_loop, m1->regno) = 0;
2206 else if (loop_dump_stream)
2207 fprintf (loop_dump_stream, "not desirable");
2209 else if (loop_dump_stream && !m->match)
2210 fprintf (loop_dump_stream, "not safe");
2212 if (loop_dump_stream)
2213 fprintf (loop_dump_stream, "\n");
2217 new_start = loop_start;
2219 /* Go through all the instructions in the loop, making
2220 all the register substitutions scheduled in REG_MAP. */
2221 for (p = new_start; p != end; p = NEXT_INSN (p))
2222 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
2223 || GET_CODE (p) == CALL_INSN)
2225 replace_regs (PATTERN (p), reg_map, nregs, 0);
2226 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
2232 /* Scan X and replace the address of any MEM in it with ADDR.
2233 REG is the address that MEM should have before the replacement. */
2236 replace_call_address (x, reg, addr)
2239 register enum rtx_code code;
2241 register const char *fmt;
2245 code = GET_CODE (x);
2259 /* Short cut for very common case. */
2260 replace_call_address (XEXP (x, 1), reg, addr);
2264 /* Short cut for very common case. */
2265 replace_call_address (XEXP (x, 0), reg, addr);
2269 /* If this MEM uses a reg other than the one we expected,
2270 something is wrong. */
2271 if (XEXP (x, 0) != reg)
2280 fmt = GET_RTX_FORMAT (code);
2281 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2284 replace_call_address (XEXP (x, i), reg, addr);
2288 for (j = 0; j < XVECLEN (x, i); j++)
2289 replace_call_address (XVECEXP (x, i, j), reg, addr);
2295 /* Return the number of memory refs to addresses that vary
2299 count_nonfixed_reads (x)
2302 register enum rtx_code code;
2304 register const char *fmt;
2310 code = GET_CODE (x);
2324 return ((invariant_p (XEXP (x, 0)) != 1)
2325 + count_nonfixed_reads (XEXP (x, 0)));
2332 fmt = GET_RTX_FORMAT (code);
2333 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2336 value += count_nonfixed_reads (XEXP (x, i));
2340 for (j = 0; j < XVECLEN (x, i); j++)
2341 value += count_nonfixed_reads (XVECEXP (x, i, j));
2349 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2350 Replace it with an instruction to load just the low bytes
2351 if the machine supports such an instruction,
2352 and insert above LOOP_START an instruction to clear the register. */
2355 constant_high_bytes (p, loop_start)
2359 register int insn_code_number;
2361 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2362 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2367 gen_rtx_STRICT_LOW_PART
2369 gen_rtx_SUBREG (GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2370 SET_DEST (PATTERN (p)), 0)),
2371 XEXP (SET_SRC (PATTERN (p)), 0));
2373 insn_code_number = recog (new, p);
2375 if (insn_code_number)
2379 /* Clear destination register before the loop. */
2380 emit_insn_before (gen_rtx_SET (VOIDmode,
2381 SET_DEST (PATTERN (p)), const0_rtx),
2384 /* Inside the loop, just load the low part. */
2390 /* Scan a loop setting the elements `cont', `vtop', `loops_enclosed',
2391 `has_call', `has_volatile', and `has_tablejump' within LOOP_INFO.
2392 Set the global variables `unknown_address_altered' and
2393 `num_mem_sets'. Also, fill in the array `loop_mems' and the list
2394 `loop_store_mems'. */
2397 prescan_loop (start, end, loop_info)
2399 struct loop_info *loop_info;
2401 register int level = 1;
2403 /* The label after END. Jumping here is just like falling off the
2404 end of the loop. We use next_nonnote_insn instead of next_label
2405 as a hedge against the (pathological) case where some actual insn
2406 might end up between the two. */
2407 rtx exit_target = next_nonnote_insn (end);
2409 loop_info->num = uid_loop_num [INSN_UID (start)];
2410 loop_info->has_indirect_jump = indirect_jump_in_function;
2411 loop_info->has_call = 0;
2412 loop_info->has_volatile = 0;
2413 loop_info->has_tablejump = 0;
2414 loop_info->loops_enclosed = 1;
2415 loop_info->has_multiple_exit_targets = 0;
2416 loop_info->cont = 0;
2417 loop_info->vtop = 0;
2419 unknown_address_altered = 0;
2420 loop_store_mems = NULL_RTX;
2421 first_loop_store_insn = NULL_RTX;
2425 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2426 insn = NEXT_INSN (insn))
2428 if (GET_CODE (insn) == NOTE)
2430 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2433 /* Count number of loops contained in this one. */
2434 loop_info->loops_enclosed++;
2436 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2445 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2448 loop_info->cont = insn;
2450 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_VTOP)
2452 /* If there is a NOTE_INSN_LOOP_VTOP, then this is a for
2453 or while style loop, with a loop exit test at the
2454 start. Thus, we can assume that the loop condition
2455 was true when the loop was entered. */
2457 loop_info->vtop = insn;
2460 else if (GET_CODE (insn) == CALL_INSN)
2462 if (! CONST_CALL_P (insn))
2463 unknown_address_altered = 1;
2464 loop_info->has_call = 1;
2466 else if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2468 rtx label1 = NULL_RTX;
2469 rtx label2 = NULL_RTX;
2471 if (volatile_refs_p (PATTERN (insn)))
2472 loop_info->has_volatile = 1;
2474 if (GET_CODE (insn) == JUMP_INSN
2475 && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
2476 || GET_CODE (PATTERN (insn)) == ADDR_VEC))
2477 loop_info->has_tablejump = 1;
2479 note_stores (PATTERN (insn), note_addr_stored);
2480 if (! first_loop_store_insn && loop_store_mems)
2481 first_loop_store_insn = insn;
2483 if (! loop_info->has_multiple_exit_targets
2484 && GET_CODE (insn) == JUMP_INSN
2485 && GET_CODE (PATTERN (insn)) == SET
2486 && SET_DEST (PATTERN (insn)) == pc_rtx)
2488 if (GET_CODE (SET_SRC (PATTERN (insn))) == IF_THEN_ELSE)
2490 label1 = XEXP (SET_SRC (PATTERN (insn)), 1);
2491 label2 = XEXP (SET_SRC (PATTERN (insn)), 2);
2495 label1 = SET_SRC (PATTERN (insn));
2499 if (label1 && label1 != pc_rtx)
2501 if (GET_CODE (label1) != LABEL_REF)
2503 /* Something tricky. */
2504 loop_info->has_multiple_exit_targets = 1;
2507 else if (XEXP (label1, 0) != exit_target
2508 && LABEL_OUTSIDE_LOOP_P (label1))
2510 /* A jump outside the current loop. */
2511 loop_info->has_multiple_exit_targets = 1;
2521 else if (GET_CODE (insn) == RETURN)
2522 loop_info->has_multiple_exit_targets = 1;
2525 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2526 if (/* We can't tell what MEMs are aliased by what. */
2527 !unknown_address_altered
2528 /* An exception thrown by a called function might land us
2530 && !loop_info->has_call
2531 /* We don't want loads for MEMs moved to a location before the
2532 one at which their stack memory becomes allocated. (Note
2533 that this is not a problem for malloc, etc., since those
2534 require actual function calls. */
2535 && !current_function_calls_alloca
2536 /* There are ways to leave the loop other than falling off the
2538 && !loop_info->has_multiple_exit_targets)
2539 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2540 insn = NEXT_INSN (insn))
2541 for_each_rtx (&insn, insert_loop_mem, 0);
2544 /* LOOP_NUMBER_CONT_DOMINATOR is now the last label between the loop start
2545 and the continue note that is a the destination of a (cond)jump after
2546 the continue note. If there is any (cond)jump between the loop start
2547 and what we have so far as LOOP_NUMBER_CONT_DOMINATOR that has a
2548 target between LOOP_DOMINATOR and the continue note, move
2549 LOOP_NUMBER_CONT_DOMINATOR forward to that label; if a jump's
2550 destination cannot be determined, clear LOOP_NUMBER_CONT_DOMINATOR. */
2553 verify_dominator (loop_number)
2558 if (! loop_number_cont_dominator[loop_number])
2559 /* This can happen for an empty loop, e.g. in
2560 gcc.c-torture/compile/920410-2.c */
2562 if (loop_number_cont_dominator[loop_number] == const0_rtx)
2564 loop_number_cont_dominator[loop_number] = 0;
2567 for (insn = loop_number_loop_starts[loop_number];
2568 insn != loop_number_cont_dominator[loop_number];
2569 insn = NEXT_INSN (insn))
2571 if (GET_CODE (insn) == JUMP_INSN
2572 && GET_CODE (PATTERN (insn)) != RETURN)
2574 rtx label = JUMP_LABEL (insn);
2577 /* If it is not a jump we can easily understand or for
2578 which we do not have jump target information in the JUMP_LABEL
2579 field (consider ADDR_VEC and ADDR_DIFF_VEC insns), then clear
2580 LOOP_NUMBER_CONT_DOMINATOR. */
2581 if ((! condjump_p (insn)
2582 && ! condjump_in_parallel_p (insn))
2583 || label == NULL_RTX)
2585 loop_number_cont_dominator[loop_number] = NULL_RTX;
2589 label_luid = INSN_LUID (label);
2590 if (label_luid < INSN_LUID (loop_number_loop_cont[loop_number])
2592 > INSN_LUID (loop_number_cont_dominator[loop_number])))
2593 loop_number_cont_dominator[loop_number] = label;
2598 /* Scan the function looking for loops. Record the start and end of each loop.
2599 Also mark as invalid loops any loops that contain a setjmp or are branched
2600 to from outside the loop. */
2603 find_and_verify_loops (f)
2607 int current_loop = -1;
2611 compute_luids (f, NULL_RTX, 0);
2613 /* If there are jumps to undefined labels,
2614 treat them as jumps out of any/all loops.
2615 This also avoids writing past end of tables when there are no loops. */
2616 uid_loop_num[0] = -1;
2618 /* Find boundaries of loops, mark which loops are contained within
2619 loops, and invalidate loops that have setjmp. */
2621 for (insn = f; insn; insn = NEXT_INSN (insn))
2623 if (GET_CODE (insn) == NOTE)
2624 switch (NOTE_LINE_NUMBER (insn))
2626 case NOTE_INSN_LOOP_BEG:
2627 loop_number_loop_starts[++next_loop] = insn;
2628 loop_number_loop_ends[next_loop] = 0;
2629 loop_number_loop_cont[next_loop] = 0;
2630 loop_number_cont_dominator[next_loop] = 0;
2631 loop_outer_loop[next_loop] = current_loop;
2632 loop_invalid[next_loop] = 0;
2633 loop_number_exit_labels[next_loop] = 0;
2634 loop_number_exit_count[next_loop] = 0;
2635 current_loop = next_loop;
2638 case NOTE_INSN_SETJMP:
2639 /* In this case, we must invalidate our current loop and any
2641 for (loop = current_loop; loop != -1; loop = loop_outer_loop[loop])
2643 loop_invalid[loop] = 1;
2644 if (loop_dump_stream)
2645 fprintf (loop_dump_stream,
2646 "\nLoop at %d ignored due to setjmp.\n",
2647 INSN_UID (loop_number_loop_starts[loop]));
2651 case NOTE_INSN_LOOP_CONT:
2652 loop_number_loop_cont[current_loop] = insn;
2654 case NOTE_INSN_LOOP_END:
2655 if (current_loop == -1)
2658 loop_number_loop_ends[current_loop] = insn;
2659 verify_dominator (current_loop);
2660 current_loop = loop_outer_loop[current_loop];
2666 /* If for any loop, this is a jump insn between the NOTE_INSN_LOOP_CONT
2667 and NOTE_INSN_LOOP_END notes, update loop_number_loop_dominator. */
2668 else if (GET_CODE (insn) == JUMP_INSN
2669 && GET_CODE (PATTERN (insn)) != RETURN
2670 && current_loop >= 0)
2673 rtx label = JUMP_LABEL (insn);
2675 if (! condjump_p (insn) && ! condjump_in_parallel_p (insn))
2678 this_loop_num = current_loop;
2681 /* First see if we care about this loop. */
2682 if (loop_number_loop_cont[this_loop_num]
2683 && loop_number_cont_dominator[this_loop_num] != const0_rtx)
2685 /* If the jump destination is not known, invalidate
2686 loop_number_const_dominator. */
2688 loop_number_cont_dominator[this_loop_num] = const0_rtx;
2690 /* Check if the destination is between loop start and
2692 if ((INSN_LUID (label)
2693 < INSN_LUID (loop_number_loop_cont[this_loop_num]))
2694 && (INSN_LUID (label)
2695 > INSN_LUID (loop_number_loop_starts[this_loop_num]))
2696 /* And if there is no later destination already
2698 && (! loop_number_cont_dominator[this_loop_num]
2699 || (INSN_LUID (label)
2700 > INSN_LUID (loop_number_cont_dominator
2702 loop_number_cont_dominator[this_loop_num] = label;
2704 this_loop_num = loop_outer_loop[this_loop_num];
2706 while (this_loop_num >= 0);
2709 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2710 enclosing loop, but this doesn't matter. */
2711 uid_loop_num[INSN_UID (insn)] = current_loop;
2714 /* Any loop containing a label used in an initializer must be invalidated,
2715 because it can be jumped into from anywhere. */
2717 for (label = forced_labels; label; label = XEXP (label, 1))
2721 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2723 loop_num = loop_outer_loop[loop_num])
2724 loop_invalid[loop_num] = 1;
2727 /* Any loop containing a label used for an exception handler must be
2728 invalidated, because it can be jumped into from anywhere. */
2730 for (label = exception_handler_labels; label; label = XEXP (label, 1))
2734 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2736 loop_num = loop_outer_loop[loop_num])
2737 loop_invalid[loop_num] = 1;
2740 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2741 loop that it is not contained within, that loop is marked invalid.
2742 If any INSN or CALL_INSN uses a label's address, then the loop containing
2743 that label is marked invalid, because it could be jumped into from
2746 Also look for blocks of code ending in an unconditional branch that
2747 exits the loop. If such a block is surrounded by a conditional
2748 branch around the block, move the block elsewhere (see below) and
2749 invert the jump to point to the code block. This may eliminate a
2750 label in our loop and will simplify processing by both us and a
2751 possible second cse pass. */
2753 for (insn = f; insn; insn = NEXT_INSN (insn))
2754 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2756 int this_loop_num = uid_loop_num[INSN_UID (insn)];
2758 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2760 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2765 for (loop_num = uid_loop_num[INSN_UID (XEXP (note, 0))];
2767 loop_num = loop_outer_loop[loop_num])
2768 loop_invalid[loop_num] = 1;
2772 if (GET_CODE (insn) != JUMP_INSN)
2775 mark_loop_jump (PATTERN (insn), this_loop_num);
2777 /* See if this is an unconditional branch outside the loop. */
2778 if (this_loop_num != -1
2779 && (GET_CODE (PATTERN (insn)) == RETURN
2780 || (simplejump_p (insn)
2781 && (uid_loop_num[INSN_UID (JUMP_LABEL (insn))]
2783 && get_max_uid () < max_uid_for_loop)
2786 rtx our_next = next_real_insn (insn);
2787 rtx last_insn_to_move = NEXT_INSN (insn);
2789 int outer_loop = -1;
2791 /* Go backwards until we reach the start of the loop, a label,
2793 for (p = PREV_INSN (insn);
2794 GET_CODE (p) != CODE_LABEL
2795 && ! (GET_CODE (p) == NOTE
2796 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2797 && GET_CODE (p) != JUMP_INSN;
2801 /* Check for the case where we have a jump to an inner nested
2802 loop, and do not perform the optimization in that case. */
2804 if (JUMP_LABEL (insn))
2806 dest_loop = uid_loop_num[INSN_UID (JUMP_LABEL (insn))];
2807 if (dest_loop != -1)
2809 for (outer_loop = dest_loop; outer_loop != -1;
2810 outer_loop = loop_outer_loop[outer_loop])
2811 if (outer_loop == this_loop_num)
2816 /* Make sure that the target of P is within the current loop. */
2818 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2819 && uid_loop_num[INSN_UID (JUMP_LABEL (p))] != this_loop_num)
2820 outer_loop = this_loop_num;
2822 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2823 we have a block of code to try to move.
2825 We look backward and then forward from the target of INSN
2826 to find a BARRIER at the same loop depth as the target.
2827 If we find such a BARRIER, we make a new label for the start
2828 of the block, invert the jump in P and point it to that label,
2829 and move the block of code to the spot we found. */
2831 if (outer_loop == -1
2832 && GET_CODE (p) == JUMP_INSN
2833 && JUMP_LABEL (p) != 0
2834 /* Just ignore jumps to labels that were never emitted.
2835 These always indicate compilation errors. */
2836 && INSN_UID (JUMP_LABEL (p)) != 0
2838 && ! simplejump_p (p)
2839 && next_real_insn (JUMP_LABEL (p)) == our_next
2840 /* If it's not safe to move the sequence, then we
2842 && insns_safe_to_move_p (p, NEXT_INSN (insn),
2843 &last_insn_to_move))
2846 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2847 int target_loop_num = uid_loop_num[INSN_UID (target)];
2850 for (loc = target; loc; loc = PREV_INSN (loc))
2851 if (GET_CODE (loc) == BARRIER
2852 /* Don't move things inside a tablejump. */
2853 && ((loc2 = next_nonnote_insn (loc)) == 0
2854 || GET_CODE (loc2) != CODE_LABEL
2855 || (loc2 = next_nonnote_insn (loc2)) == 0
2856 || GET_CODE (loc2) != JUMP_INSN
2857 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2858 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2859 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2863 for (loc = target; loc; loc = NEXT_INSN (loc))
2864 if (GET_CODE (loc) == BARRIER
2865 /* Don't move things inside a tablejump. */
2866 && ((loc2 = next_nonnote_insn (loc)) == 0
2867 || GET_CODE (loc2) != CODE_LABEL
2868 || (loc2 = next_nonnote_insn (loc2)) == 0
2869 || GET_CODE (loc2) != JUMP_INSN
2870 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2871 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2872 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2877 rtx cond_label = JUMP_LABEL (p);
2878 rtx new_label = get_label_after (p);
2880 /* Ensure our label doesn't go away. */
2881 LABEL_NUSES (cond_label)++;
2883 /* Verify that uid_loop_num is large enough and that
2885 if (invert_jump (p, new_label))
2889 /* If no suitable BARRIER was found, create a suitable
2890 one before TARGET. Since TARGET is a fall through
2891 path, we'll need to insert an jump around our block
2892 and a add a BARRIER before TARGET.
2894 This creates an extra unconditional jump outside
2895 the loop. However, the benefits of removing rarely
2896 executed instructions from inside the loop usually
2897 outweighs the cost of the extra unconditional jump
2898 outside the loop. */
2903 temp = gen_jump (JUMP_LABEL (insn));
2904 temp = emit_jump_insn_before (temp, target);
2905 JUMP_LABEL (temp) = JUMP_LABEL (insn);
2906 LABEL_NUSES (JUMP_LABEL (insn))++;
2907 loc = emit_barrier_before (target);
2910 /* Include the BARRIER after INSN and copy the
2912 new_label = squeeze_notes (new_label,
2914 reorder_insns (new_label, last_insn_to_move, loc);
2916 /* All those insns are now in TARGET_LOOP_NUM. */
2918 q != NEXT_INSN (last_insn_to_move);
2920 uid_loop_num[INSN_UID (q)] = target_loop_num;
2922 /* The label jumped to by INSN is no longer a loop exit.
2923 Unless INSN does not have a label (e.g., it is a
2924 RETURN insn), search loop_number_exit_labels to find
2925 its label_ref, and remove it. Also turn off
2926 LABEL_OUTSIDE_LOOP_P bit. */
2927 if (JUMP_LABEL (insn))
2932 r = loop_number_exit_labels[this_loop_num];
2933 r; q = r, r = LABEL_NEXTREF (r))
2934 if (XEXP (r, 0) == JUMP_LABEL (insn))
2936 LABEL_OUTSIDE_LOOP_P (r) = 0;
2938 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2940 loop_number_exit_labels[this_loop_num]
2941 = LABEL_NEXTREF (r);
2945 for (loop_num = this_loop_num;
2946 loop_num != -1 && loop_num != target_loop_num;
2947 loop_num = loop_outer_loop[loop_num])
2948 loop_number_exit_count[loop_num]--;
2950 /* If we didn't find it, then something is wrong. */
2955 /* P is now a jump outside the loop, so it must be put
2956 in loop_number_exit_labels, and marked as such.
2957 The easiest way to do this is to just call
2958 mark_loop_jump again for P. */
2959 mark_loop_jump (PATTERN (p), this_loop_num);
2961 /* If INSN now jumps to the insn after it,
2963 if (JUMP_LABEL (insn) != 0
2964 && (next_real_insn (JUMP_LABEL (insn))
2965 == next_real_insn (insn)))
2969 /* Continue the loop after where the conditional
2970 branch used to jump, since the only branch insn
2971 in the block (if it still remains) is an inter-loop
2972 branch and hence needs no processing. */
2973 insn = NEXT_INSN (cond_label);
2975 if (--LABEL_NUSES (cond_label) == 0)
2976 delete_insn (cond_label);
2978 /* This loop will be continued with NEXT_INSN (insn). */
2979 insn = PREV_INSN (insn);
2986 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2987 loops it is contained in, mark the target loop invalid.
2989 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2992 mark_loop_jump (x, loop_num)
3000 switch (GET_CODE (x))
3013 /* There could be a label reference in here. */
3014 mark_loop_jump (XEXP (x, 0), loop_num);
3020 mark_loop_jump (XEXP (x, 0), loop_num);
3021 mark_loop_jump (XEXP (x, 1), loop_num);
3025 /* This may refer to a LABEL_REF or SYMBOL_REF. */
3026 mark_loop_jump (XEXP (x, 1), loop_num);
3031 mark_loop_jump (XEXP (x, 0), loop_num);
3035 dest_loop = uid_loop_num[INSN_UID (XEXP (x, 0))];
3037 /* Link together all labels that branch outside the loop. This
3038 is used by final_[bg]iv_value and the loop unrolling code. Also
3039 mark this LABEL_REF so we know that this branch should predict
3042 /* A check to make sure the label is not in an inner nested loop,
3043 since this does not count as a loop exit. */
3044 if (dest_loop != -1)
3046 for (outer_loop = dest_loop; outer_loop != -1;
3047 outer_loop = loop_outer_loop[outer_loop])
3048 if (outer_loop == loop_num)
3054 if (loop_num != -1 && outer_loop == -1)
3056 LABEL_OUTSIDE_LOOP_P (x) = 1;
3057 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
3058 loop_number_exit_labels[loop_num] = x;
3060 for (outer_loop = loop_num;
3061 outer_loop != -1 && outer_loop != dest_loop;
3062 outer_loop = loop_outer_loop[outer_loop])
3063 loop_number_exit_count[outer_loop]++;
3066 /* If this is inside a loop, but not in the current loop or one enclosed
3067 by it, it invalidates at least one loop. */
3069 if (dest_loop == -1)
3072 /* We must invalidate every nested loop containing the target of this
3073 label, except those that also contain the jump insn. */
3075 for (; dest_loop != -1; dest_loop = loop_outer_loop[dest_loop])
3077 /* Stop when we reach a loop that also contains the jump insn. */
3078 for (outer_loop = loop_num; outer_loop != -1;
3079 outer_loop = loop_outer_loop[outer_loop])
3080 if (dest_loop == outer_loop)
3083 /* If we get here, we know we need to invalidate a loop. */
3084 if (loop_dump_stream && ! loop_invalid[dest_loop])
3085 fprintf (loop_dump_stream,
3086 "\nLoop at %d ignored due to multiple entry points.\n",
3087 INSN_UID (loop_number_loop_starts[dest_loop]));
3089 loop_invalid[dest_loop] = 1;
3094 /* If this is not setting pc, ignore. */
3095 if (SET_DEST (x) == pc_rtx)
3096 mark_loop_jump (SET_SRC (x), loop_num);
3100 mark_loop_jump (XEXP (x, 1), loop_num);
3101 mark_loop_jump (XEXP (x, 2), loop_num);
3106 for (i = 0; i < XVECLEN (x, 0); i++)
3107 mark_loop_jump (XVECEXP (x, 0, i), loop_num);
3111 for (i = 0; i < XVECLEN (x, 1); i++)
3112 mark_loop_jump (XVECEXP (x, 1, i), loop_num);
3116 /* Strictly speaking this is not a jump into the loop, only a possible
3117 jump out of the loop. However, we have no way to link the destination
3118 of this jump onto the list of exit labels. To be safe we mark this
3119 loop and any containing loops as invalid. */
3122 for (outer_loop = loop_num; outer_loop != -1;
3123 outer_loop = loop_outer_loop[outer_loop])
3125 if (loop_dump_stream && ! loop_invalid[outer_loop])
3126 fprintf (loop_dump_stream,
3127 "\nLoop at %d ignored due to unknown exit jump.\n",
3128 INSN_UID (loop_number_loop_starts[outer_loop]));
3129 loop_invalid[outer_loop] = 1;
3136 /* Return nonzero if there is a label in the range from
3137 insn INSN to and including the insn whose luid is END
3138 INSN must have an assigned luid (i.e., it must not have
3139 been previously created by loop.c). */
3142 labels_in_range_p (insn, end)
3146 while (insn && INSN_LUID (insn) <= end)
3148 if (GET_CODE (insn) == CODE_LABEL)
3150 insn = NEXT_INSN (insn);
3156 /* Record that a memory reference X is being set. */
3159 note_addr_stored (x, y)
3161 rtx y ATTRIBUTE_UNUSED;
3163 if (x == 0 || GET_CODE (x) != MEM)
3166 /* Count number of memory writes.
3167 This affects heuristics in strength_reduce. */
3170 /* BLKmode MEM means all memory is clobbered. */
3171 if (GET_MODE (x) == BLKmode)
3172 unknown_address_altered = 1;
3174 if (unknown_address_altered)
3177 loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, x, loop_store_mems);
3180 /* X is a value modified by an INSN that references a biv inside a loop
3181 exit test (ie, X is somehow related to the value of the biv). If X
3182 is a pseudo that is used more than once, then the biv is (effectively)
3183 used more than once. */
3186 note_set_pseudo_multiple_uses (x, y)
3188 rtx y ATTRIBUTE_UNUSED;
3193 while (GET_CODE (x) == STRICT_LOW_PART
3194 || GET_CODE (x) == SIGN_EXTRACT
3195 || GET_CODE (x) == ZERO_EXTRACT
3196 || GET_CODE (x) == SUBREG)
3199 if (GET_CODE (x) != REG || REGNO (x) < FIRST_PSEUDO_REGISTER)
3202 /* If we do not have usage information, or if we know the register
3203 is used more than once, note that fact for check_dbra_loop. */
3204 if (REGNO (x) >= max_reg_before_loop
3205 || ! VARRAY_RTX (reg_single_usage, REGNO (x))
3206 || VARRAY_RTX (reg_single_usage, REGNO (x)) == const0_rtx)
3207 note_set_pseudo_multiple_uses_retval = 1;
3210 /* Return nonzero if the rtx X is invariant over the current loop.
3212 The value is 2 if we refer to something only conditionally invariant.
3214 If `unknown_address_altered' is nonzero, no memory ref is invariant.
3215 Otherwise, a memory ref is invariant if it does not conflict with
3216 anything stored in `loop_store_mems'. */
3223 register enum rtx_code code;
3224 register const char *fmt;
3225 int conditional = 0;
3230 code = GET_CODE (x);
3240 /* A LABEL_REF is normally invariant, however, if we are unrolling
3241 loops, and this label is inside the loop, then it isn't invariant.
3242 This is because each unrolled copy of the loop body will have
3243 a copy of this label. If this was invariant, then an insn loading
3244 the address of this label into a register might get moved outside
3245 the loop, and then each loop body would end up using the same label.
3247 We don't know the loop bounds here though, so just fail for all
3249 if (flag_unroll_loops)
3256 case UNSPEC_VOLATILE:
3260 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
3261 since the reg might be set by initialization within the loop. */
3263 if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
3264 || x == arg_pointer_rtx)
3265 && ! current_function_has_nonlocal_goto)
3268 if (this_loop_info.has_call
3269 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
3272 if (VARRAY_INT (set_in_loop, REGNO (x)) < 0)
3275 return VARRAY_INT (set_in_loop, REGNO (x)) == 0;
3278 /* Volatile memory references must be rejected. Do this before
3279 checking for read-only items, so that volatile read-only items
3280 will be rejected also. */
3281 if (MEM_VOLATILE_P (x))
3284 /* Read-only items (such as constants in a constant pool) are
3285 invariant if their address is. */
3286 if (RTX_UNCHANGING_P (x))
3289 /* If we had a subroutine call, any location in memory could have been
3291 if (unknown_address_altered)
3294 /* See if there is any dependence between a store and this load. */
3295 mem_list_entry = loop_store_mems;
3296 while (mem_list_entry)
3298 if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
3301 mem_list_entry = XEXP (mem_list_entry, 1);
3304 /* It's not invalidated by a store in memory
3305 but we must still verify the address is invariant. */
3309 /* Don't mess with insns declared volatile. */
3310 if (MEM_VOLATILE_P (x))
3318 fmt = GET_RTX_FORMAT (code);
3319 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3323 int tem = invariant_p (XEXP (x, i));
3329 else if (fmt[i] == 'E')
3332 for (j = 0; j < XVECLEN (x, i); j++)
3334 int tem = invariant_p (XVECEXP (x, i, j));
3344 return 1 + conditional;
3348 /* Return nonzero if all the insns in the loop that set REG
3349 are INSN and the immediately following insns,
3350 and if each of those insns sets REG in an invariant way
3351 (not counting uses of REG in them).
3353 The value is 2 if some of these insns are only conditionally invariant.
3355 We assume that INSN itself is the first set of REG
3356 and that its source is invariant. */
3359 consec_sets_invariant_p (reg, n_sets, insn)
3363 register rtx p = insn;
3364 register int regno = REGNO (reg);
3366 /* Number of sets we have to insist on finding after INSN. */
3367 int count = n_sets - 1;
3368 int old = VARRAY_INT (set_in_loop, regno);
3372 /* If N_SETS hit the limit, we can't rely on its value. */
3376 VARRAY_INT (set_in_loop, regno) = 0;
3380 register enum rtx_code code;
3384 code = GET_CODE (p);
3386 /* If library call, skip to end of it. */
3387 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3392 && (set = single_set (p))
3393 && GET_CODE (SET_DEST (set)) == REG
3394 && REGNO (SET_DEST (set)) == regno)
3396 this = invariant_p (SET_SRC (set));
3399 else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)))
3401 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3402 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3404 this = (CONSTANT_P (XEXP (temp, 0))
3405 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
3406 && invariant_p (XEXP (temp, 0))));
3413 else if (code != NOTE)
3415 VARRAY_INT (set_in_loop, regno) = old;
3420 VARRAY_INT (set_in_loop, regno) = old;
3421 /* If invariant_p ever returned 2, we return 2. */
3422 return 1 + (value & 2);
3426 /* I don't think this condition is sufficient to allow INSN
3427 to be moved, so we no longer test it. */
3429 /* Return 1 if all insns in the basic block of INSN and following INSN
3430 that set REG are invariant according to TABLE. */
3433 all_sets_invariant_p (reg, insn, table)
3437 register rtx p = insn;
3438 register int regno = REGNO (reg);
3442 register enum rtx_code code;
3444 code = GET_CODE (p);
3445 if (code == CODE_LABEL || code == JUMP_INSN)
3447 if (code == INSN && GET_CODE (PATTERN (p)) == SET
3448 && GET_CODE (SET_DEST (PATTERN (p))) == REG
3449 && REGNO (SET_DEST (PATTERN (p))) == regno)
3451 if (!invariant_p (SET_SRC (PATTERN (p)), table))
3458 /* Look at all uses (not sets) of registers in X. For each, if it is
3459 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3460 a different insn, set USAGE[REGNO] to const0_rtx. */
3463 find_single_use_in_loop (insn, x, usage)
3468 enum rtx_code code = GET_CODE (x);
3469 const char *fmt = GET_RTX_FORMAT (code);
3473 VARRAY_RTX (usage, REGNO (x))
3474 = (VARRAY_RTX (usage, REGNO (x)) != 0
3475 && VARRAY_RTX (usage, REGNO (x)) != insn)
3476 ? const0_rtx : insn;
3478 else if (code == SET)
3480 /* Don't count SET_DEST if it is a REG; otherwise count things
3481 in SET_DEST because if a register is partially modified, it won't
3482 show up as a potential movable so we don't care how USAGE is set
3484 if (GET_CODE (SET_DEST (x)) != REG)
3485 find_single_use_in_loop (insn, SET_DEST (x), usage);
3486 find_single_use_in_loop (insn, SET_SRC (x), usage);
3489 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3491 if (fmt[i] == 'e' && XEXP (x, i) != 0)
3492 find_single_use_in_loop (insn, XEXP (x, i), usage);
3493 else if (fmt[i] == 'E')
3494 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3495 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
3499 /* Count and record any set in X which is contained in INSN. Update
3500 MAY_NOT_MOVE and LAST_SET for any register set in X. */
3503 count_one_set (insn, x, may_not_move, last_set)
3505 varray_type may_not_move;
3508 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3509 /* Don't move a reg that has an explicit clobber.
3510 It's not worth the pain to try to do it correctly. */
3511 VARRAY_CHAR (may_not_move, REGNO (XEXP (x, 0))) = 1;
3513 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3515 rtx dest = SET_DEST (x);
3516 while (GET_CODE (dest) == SUBREG
3517 || GET_CODE (dest) == ZERO_EXTRACT
3518 || GET_CODE (dest) == SIGN_EXTRACT
3519 || GET_CODE (dest) == STRICT_LOW_PART)
3520 dest = XEXP (dest, 0);
3521 if (GET_CODE (dest) == REG)
3523 register int regno = REGNO (dest);
3524 /* If this is the first setting of this reg
3525 in current basic block, and it was set before,
3526 it must be set in two basic blocks, so it cannot
3527 be moved out of the loop. */
3528 if (VARRAY_INT (set_in_loop, regno) > 0
3529 && last_set[regno] == 0)
3530 VARRAY_CHAR (may_not_move, regno) = 1;
3531 /* If this is not first setting in current basic block,
3532 see if reg was used in between previous one and this.
3533 If so, neither one can be moved. */
3534 if (last_set[regno] != 0
3535 && reg_used_between_p (dest, last_set[regno], insn))
3536 VARRAY_CHAR (may_not_move, regno) = 1;
3537 if (VARRAY_INT (set_in_loop, regno) < 127)
3538 ++VARRAY_INT (set_in_loop, regno);
3539 last_set[regno] = insn;
3544 /* Increment SET_IN_LOOP at the index of each register
3545 that is modified by an insn between FROM and TO.
3546 If the value of an element of SET_IN_LOOP becomes 127 or more,
3547 stop incrementing it, to avoid overflow.
3549 Store in SINGLE_USAGE[I] the single insn in which register I is
3550 used, if it is only used once. Otherwise, it is set to 0 (for no
3551 uses) or const0_rtx for more than one use. This parameter may be zero,
3552 in which case this processing is not done.
3554 Store in *COUNT_PTR the number of actual instruction
3555 in the loop. We use this to decide what is worth moving out. */
3557 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3558 In that case, it is the insn that last set reg n. */
3561 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
3562 register rtx from, to;
3563 varray_type may_not_move;
3564 varray_type single_usage;
3568 register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx));
3570 register int count = 0;
3572 bzero ((char *) last_set, nregs * sizeof (rtx));
3573 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3575 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3579 /* Record registers that have exactly one use. */
3580 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3582 /* Include uses in REG_EQUAL notes. */
3583 if (REG_NOTES (insn))
3584 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3586 if (GET_CODE (PATTERN (insn)) == SET
3587 || GET_CODE (PATTERN (insn)) == CLOBBER)
3588 count_one_set (insn, PATTERN (insn), may_not_move, last_set);
3589 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3592 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3593 count_one_set (insn, XVECEXP (PATTERN (insn), 0, i),
3594 may_not_move, last_set);
3598 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3599 bzero ((char *) last_set, nregs * sizeof (rtx));
3604 /* Given a loop that is bounded by LOOP_START and LOOP_END
3605 and that is entered at SCAN_START,
3606 return 1 if the register set in SET contained in insn INSN is used by
3607 any insn that precedes INSN in cyclic order starting
3608 from the loop entry point.
3610 We don't want to use INSN_LUID here because if we restrict INSN to those
3611 that have a valid INSN_LUID, it means we cannot move an invariant out
3612 from an inner loop past two loops. */
3615 loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end)
3616 rtx set, insn, loop_start, scan_start, loop_end;
3618 rtx reg = SET_DEST (set);
3621 /* Scan forward checking for register usage. If we hit INSN, we
3622 are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */
3623 for (p = scan_start; p != insn; p = NEXT_INSN (p))
3625 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
3626 && reg_overlap_mentioned_p (reg, PATTERN (p)))
3636 /* A "basic induction variable" or biv is a pseudo reg that is set
3637 (within this loop) only by incrementing or decrementing it. */
3638 /* A "general induction variable" or giv is a pseudo reg whose
3639 value is a linear function of a biv. */
3641 /* Bivs are recognized by `basic_induction_var';
3642 Givs by `general_induction_var'. */
3644 /* Indexed by register number, indicates whether or not register is an
3645 induction variable, and if so what type. */
3647 varray_type reg_iv_type;
3649 /* Indexed by register number, contains pointer to `struct induction'
3650 if register is an induction variable. This holds general info for
3651 all induction variables. */
3653 varray_type reg_iv_info;
3655 /* Indexed by register number, contains pointer to `struct iv_class'
3656 if register is a basic induction variable. This holds info describing
3657 the class (a related group) of induction variables that the biv belongs
3660 struct iv_class **reg_biv_class;
3662 /* The head of a list which links together (via the next field)
3663 every iv class for the current loop. */
3665 struct iv_class *loop_iv_list;
3667 /* Givs made from biv increments are always splittable for loop unrolling.
3668 Since there is no regscan info for them, we have to keep track of them
3670 int first_increment_giv, last_increment_giv;
3672 /* Communication with routines called via `note_stores'. */
3674 static rtx note_insn;
3676 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3678 static rtx addr_placeholder;
3680 /* ??? Unfinished optimizations, and possible future optimizations,
3681 for the strength reduction code. */
3683 /* ??? The interaction of biv elimination, and recognition of 'constant'
3684 bivs, may cause problems. */
3686 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3687 performance problems.
3689 Perhaps don't eliminate things that can be combined with an addressing
3690 mode. Find all givs that have the same biv, mult_val, and add_val;
3691 then for each giv, check to see if its only use dies in a following
3692 memory address. If so, generate a new memory address and check to see
3693 if it is valid. If it is valid, then store the modified memory address,
3694 otherwise, mark the giv as not done so that it will get its own iv. */
3696 /* ??? Could try to optimize branches when it is known that a biv is always
3699 /* ??? When replace a biv in a compare insn, we should replace with closest
3700 giv so that an optimized branch can still be recognized by the combiner,
3701 e.g. the VAX acb insn. */
3703 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3704 was rerun in loop_optimize whenever a register was added or moved.
3705 Also, some of the optimizations could be a little less conservative. */
3707 /* Perform strength reduction and induction variable elimination.
3709 Pseudo registers created during this function will be beyond the last
3710 valid index in several tables including n_times_set and regno_last_uid.
3711 This does not cause a problem here, because the added registers cannot be
3712 givs outside of their loop, and hence will never be reconsidered.
3713 But scan_loop must check regnos to make sure they are in bounds.
3715 SCAN_START is the first instruction in the loop, as the loop would
3716 actually be executed. END is the NOTE_INSN_LOOP_END. LOOP_TOP is
3717 the first instruction in the loop, as it is layed out in the
3718 instruction stream. LOOP_START is the NOTE_INSN_LOOP_BEG.
3719 LOOP_CONT is the NOTE_INSN_LOOP_CONT. */
3722 strength_reduce (scan_start, end, loop_top, insn_count,
3723 loop_start, loop_end, loop_info, loop_cont, unroll_p, bct_p)
3730 struct loop_info *loop_info;
3732 int unroll_p, bct_p ATTRIBUTE_UNUSED;
3740 /* This is 1 if current insn is not executed at least once for every loop
3742 int not_every_iteration = 0;
3743 /* This is 1 if current insn may be executed more than once for every
3745 int maybe_multiple = 0;
3746 /* This is 1 if we have past a branch back to the top of the loop
3747 (aka a loop latch). */
3748 int past_loop_latch = 0;
3749 /* Temporary list pointers for traversing loop_iv_list. */
3750 struct iv_class *bl, **backbl;
3751 /* Ratio of extra register life span we can justify
3752 for saving an instruction. More if loop doesn't call subroutines
3753 since in that case saving an insn makes more difference
3754 and more registers are available. */
3755 /* ??? could set this to last value of threshold in move_movables */
3756 int threshold = (loop_info->has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3757 /* Map of pseudo-register replacements. */
3762 rtx end_insert_before;
3764 int n_extra_increment;
3765 int unrolled_insn_copies = 0;
3767 /* If scan_start points to the loop exit test, we have to be wary of
3768 subversive use of gotos inside expression statements. */
3769 if (prev_nonnote_insn (scan_start) != prev_nonnote_insn (loop_start))
3770 maybe_multiple = back_branch_in_range_p (scan_start, loop_start, loop_end);
3772 VARRAY_INT_INIT (reg_iv_type, max_reg_before_loop, "reg_iv_type");
3773 VARRAY_GENERIC_PTR_INIT (reg_iv_info, max_reg_before_loop, "reg_iv_info");
3774 reg_biv_class = (struct iv_class **)
3775 alloca (max_reg_before_loop * sizeof (struct iv_class *));
3776 bzero ((char *) reg_biv_class, (max_reg_before_loop
3777 * sizeof (struct iv_class *)));
3780 addr_placeholder = gen_reg_rtx (Pmode);
3782 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3783 must be put before this insn, so that they will appear in the right
3784 order (i.e. loop order).
3786 If loop_end is the end of the current function, then emit a
3787 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3789 if (NEXT_INSN (loop_end) != 0)
3790 end_insert_before = NEXT_INSN (loop_end);
3792 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3794 /* Scan through loop to find all possible bivs. */
3796 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
3798 p = next_insn_in_loop (p, scan_start, end, loop_top))
3800 if (GET_CODE (p) == INSN
3801 && (set = single_set (p))
3802 && GET_CODE (SET_DEST (set)) == REG)
3804 dest_reg = SET_DEST (set);
3805 if (REGNO (dest_reg) < max_reg_before_loop
3806 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3807 && REG_IV_TYPE (REGNO (dest_reg)) != NOT_BASIC_INDUCT)
3809 if (basic_induction_var (SET_SRC (set), GET_MODE (SET_SRC (set)),
3810 dest_reg, p, &inc_val, &mult_val,
3813 /* It is a possible basic induction variable.
3814 Create and initialize an induction structure for it. */
3817 = (struct induction *) alloca (sizeof (struct induction));
3819 record_biv (v, p, dest_reg, inc_val, mult_val, location,
3820 not_every_iteration, maybe_multiple);
3821 REG_IV_TYPE (REGNO (dest_reg)) = BASIC_INDUCT;
3823 else if (REGNO (dest_reg) < max_reg_before_loop)
3824 REG_IV_TYPE (REGNO (dest_reg)) = NOT_BASIC_INDUCT;
3828 /* Past CODE_LABEL, we get to insns that may be executed multiple
3829 times. The only way we can be sure that they can't is if every
3830 jump insn between here and the end of the loop either
3831 returns, exits the loop, is a jump to a location that is still
3832 behind the label, or is a jump to the loop start. */
3834 if (GET_CODE (p) == CODE_LABEL)
3842 insn = NEXT_INSN (insn);
3843 if (insn == scan_start)
3851 if (insn == scan_start)
3855 if (GET_CODE (insn) == JUMP_INSN
3856 && GET_CODE (PATTERN (insn)) != RETURN
3857 && (! condjump_p (insn)
3858 || (JUMP_LABEL (insn) != 0
3859 && JUMP_LABEL (insn) != scan_start
3860 && ! loop_insn_first_p (p, JUMP_LABEL (insn)))))
3868 /* Past a jump, we get to insns for which we can't count
3869 on whether they will be executed during each iteration. */
3870 /* This code appears twice in strength_reduce. There is also similar
3871 code in scan_loop. */
3872 if (GET_CODE (p) == JUMP_INSN
3873 /* If we enter the loop in the middle, and scan around to the
3874 beginning, don't set not_every_iteration for that.
3875 This can be any kind of jump, since we want to know if insns
3876 will be executed if the loop is executed. */
3877 && ! (JUMP_LABEL (p) == loop_top
3878 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3879 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3883 /* If this is a jump outside the loop, then it also doesn't
3884 matter. Check to see if the target of this branch is on the
3885 loop_number_exits_labels list. */
3887 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3889 label = LABEL_NEXTREF (label))
3890 if (XEXP (label, 0) == JUMP_LABEL (p))
3894 not_every_iteration = 1;
3897 else if (GET_CODE (p) == NOTE)
3899 /* At the virtual top of a converted loop, insns are again known to
3900 be executed each iteration: logically, the loop begins here
3901 even though the exit code has been duplicated.
3903 Insns are also again known to be executed each iteration at
3904 the LOOP_CONT note. */
3905 if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
3906 || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
3908 not_every_iteration = 0;
3909 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3911 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3915 /* Note if we pass a loop latch. If we do, then we can not clear
3916 NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in
3917 a loop since a jump before the last CODE_LABEL may have started
3918 a new loop iteration.
3920 Note that LOOP_TOP is only set for rotated loops and we need
3921 this check for all loops, so compare against the CODE_LABEL
3922 which immediately follows LOOP_START. */
3923 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == NEXT_INSN (loop_start))
3924 past_loop_latch = 1;
3926 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3927 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3928 or not an insn is known to be executed each iteration of the
3929 loop, whether or not any iterations are known to occur.
3931 Therefore, if we have just passed a label and have no more labels
3932 between here and the test insn of the loop, and we have not passed
3933 a jump to the top of the loop, then we know these insns will be
3934 executed each iteration. */
3936 if (not_every_iteration
3937 && ! past_loop_latch
3938 && GET_CODE (p) == CODE_LABEL
3939 && no_labels_between_p (p, loop_end)
3940 && loop_insn_first_p (p, loop_cont))
3941 not_every_iteration = 0;
3944 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3945 Make a sanity check against n_times_set. */
3946 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3948 if (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3949 /* Above happens if register modified by subreg, etc. */
3950 /* Make sure it is not recognized as a basic induction var: */
3951 || VARRAY_INT (n_times_set, bl->regno) != bl->biv_count
3952 /* If never incremented, it is invariant that we decided not to
3953 move. So leave it alone. */
3954 || ! bl->incremented)
3956 if (loop_dump_stream)
3957 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3959 (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3960 ? "not induction variable"
3961 : (! bl->incremented ? "never incremented"
3964 REG_IV_TYPE (bl->regno) = NOT_BASIC_INDUCT;
3971 if (loop_dump_stream)
3972 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3976 /* Exit if there are no bivs. */
3979 /* Can still unroll the loop anyways, but indicate that there is no
3980 strength reduction info available. */
3982 unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
3988 /* Find initial value for each biv by searching backwards from loop_start,
3989 halting at first label. Also record any test condition. */
3992 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3996 if (GET_CODE (p) == CALL_INSN)
3999 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4000 || GET_CODE (p) == CALL_INSN)
4001 note_stores (PATTERN (p), record_initial);
4003 /* Record any test of a biv that branches around the loop if no store
4004 between it and the start of loop. We only care about tests with
4005 constants and registers and only certain of those. */
4006 if (GET_CODE (p) == JUMP_INSN
4007 && JUMP_LABEL (p) != 0
4008 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
4009 && (test = get_condition_for_loop (p)) != 0
4010 && GET_CODE (XEXP (test, 0)) == REG
4011 && REGNO (XEXP (test, 0)) < max_reg_before_loop
4012 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
4013 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
4014 && bl->init_insn == 0)
4016 /* If an NE test, we have an initial value! */
4017 if (GET_CODE (test) == NE)
4020 bl->init_set = gen_rtx_SET (VOIDmode,
4021 XEXP (test, 0), XEXP (test, 1));
4024 bl->initial_test = test;
4028 /* Look at the each biv and see if we can say anything better about its
4029 initial value from any initializing insns set up above. (This is done
4030 in two passes to avoid missing SETs in a PARALLEL.) */
4031 for (backbl = &loop_iv_list; (bl = *backbl); backbl = &bl->next)
4036 if (! bl->init_insn)
4039 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
4040 is a constant, use the value of that. */
4041 if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL
4042 && CONSTANT_P (XEXP (note, 0)))
4043 || ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL
4044 && CONSTANT_P (XEXP (note, 0))))
4045 src = XEXP (note, 0);
4047 src = SET_SRC (bl->init_set);
4049 if (loop_dump_stream)
4050 fprintf (loop_dump_stream,
4051 "Biv %d initialized at insn %d: initial value ",
4052 bl->regno, INSN_UID (bl->init_insn));
4054 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
4055 || GET_MODE (src) == VOIDmode)
4056 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
4058 bl->initial_value = src;
4060 if (loop_dump_stream)
4062 if (GET_CODE (src) == CONST_INT)
4064 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (src));
4065 fputc ('\n', loop_dump_stream);
4069 print_rtl (loop_dump_stream, src);
4070 fprintf (loop_dump_stream, "\n");
4076 struct iv_class *bl2 = 0;
4077 rtx increment = NULL_RTX;
4079 /* Biv initial value is not a simple move. If it is the sum of
4080 another biv and a constant, check if both bivs are incremented
4081 in lockstep. Then we are actually looking at a giv.
4082 For simplicity, we only handle the case where there is but a
4083 single increment, and the register is not used elsewhere. */
4084 if (bl->biv_count == 1
4085 && bl->regno < max_reg_before_loop
4086 && uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4087 && GET_CODE (src) == PLUS
4088 && GET_CODE (XEXP (src, 0)) == REG
4089 && CONSTANT_P (XEXP (src, 1))
4090 && ((increment = biv_total_increment (bl, loop_start, loop_end))
4093 int regno = REGNO (XEXP (src, 0));
4095 for (bl2 = loop_iv_list; bl2; bl2 = bl2->next)
4096 if (bl2->regno == regno)
4100 /* Now, can we transform this biv into a giv? */
4102 && bl2->biv_count == 1
4103 && rtx_equal_p (increment,
4104 biv_total_increment (bl2, loop_start, loop_end))
4105 /* init_insn is only set to insns that are before loop_start
4106 without any intervening labels. */
4107 && ! reg_set_between_p (bl2->biv->src_reg,
4108 PREV_INSN (bl->init_insn), loop_start)
4109 /* The register from BL2 must be set before the register from
4110 BL is set, or we must be able to move the latter set after
4111 the former set. Currently there can't be any labels
4112 in-between when biv_total_increment returns nonzero both times
4113 but we test it here in case some day some real cfg analysis
4114 gets used to set always_computable. */
4115 && (loop_insn_first_p (bl2->biv->insn, bl->biv->insn)
4116 ? no_labels_between_p (bl2->biv->insn, bl->biv->insn)
4117 : (! reg_used_between_p (bl->biv->src_reg, bl->biv->insn,
4119 && no_jumps_between_p (bl->biv->insn, bl2->biv->insn)))
4120 && validate_change (bl->biv->insn,
4121 &SET_SRC (single_set (bl->biv->insn)),
4124 int loop_num = uid_loop_num[INSN_UID (loop_start)];
4125 rtx dominator = loop_number_cont_dominator[loop_num];
4126 rtx giv = bl->biv->src_reg;
4127 rtx giv_insn = bl->biv->insn;
4128 rtx after_giv = NEXT_INSN (giv_insn);
4130 if (loop_dump_stream)
4131 fprintf (loop_dump_stream, "is giv of biv %d\n", bl2->regno);
4132 /* Let this giv be discovered by the generic code. */
4133 REG_IV_TYPE (bl->regno) = UNKNOWN_INDUCT;
4134 reg_biv_class[bl->regno] = NULL_PTR;
4135 /* We can get better optimization if we can move the giv setting
4136 before the first giv use. */
4138 && ! loop_insn_first_p (dominator, scan_start)
4139 && ! reg_set_between_p (bl2->biv->src_reg, loop_start,
4141 && ! reg_used_between_p (giv, loop_start, dominator)
4142 && ! reg_used_between_p (giv, giv_insn, loop_end))
4147 for (next = NEXT_INSN (dominator); ; next = NEXT_INSN (next))
4149 if ((GET_RTX_CLASS (GET_CODE (next)) == 'i'
4150 && (reg_mentioned_p (giv, PATTERN (next))
4151 || reg_set_p (bl2->biv->src_reg, next)))
4152 || GET_CODE (next) == JUMP_INSN)
4155 if (GET_RTX_CLASS (GET_CODE (next)) != 'i'
4156 || ! sets_cc0_p (PATTERN (next)))
4160 if (loop_dump_stream)
4161 fprintf (loop_dump_stream, "move after insn %d\n",
4162 INSN_UID (dominator));
4163 /* Avoid problems with luids by actually moving the insn
4164 and adjusting all luids in the range. */
4165 reorder_insns (giv_insn, giv_insn, dominator);
4166 for (p = dominator; INSN_UID (p) >= max_uid_for_loop; )
4168 compute_luids (giv_insn, after_giv, INSN_LUID (p));
4169 /* If the only purpose of the init insn is to initialize
4170 this giv, delete it. */
4171 if (single_set (bl->init_insn)
4172 && ! reg_used_between_p (giv, bl->init_insn, loop_start))
4173 delete_insn (bl->init_insn);
4175 else if (! loop_insn_first_p (bl2->biv->insn, bl->biv->insn))
4177 rtx p = PREV_INSN (giv_insn);
4178 while (INSN_UID (p) >= max_uid_for_loop)
4180 reorder_insns (giv_insn, giv_insn, bl2->biv->insn);
4181 compute_luids (after_giv, NEXT_INSN (giv_insn),
4184 /* Remove this biv from the chain. */
4187 /* We move the following giv from *bl->next into *bl.
4188 We have to update reg_biv_class for that moved biv
4189 to point to its new address. */
4191 reg_biv_class[bl->regno] = bl;
4200 /* If we can't make it a giv,
4201 let biv keep initial value of "itself". */
4202 else if (loop_dump_stream)
4203 fprintf (loop_dump_stream, "is complex\n");
4207 /* If a biv is unconditionally incremented several times in a row, convert
4208 all but the last increment into a giv. */
4210 /* Get an upper bound for the number of registers
4211 we might have after all bivs have been processed. */
4212 first_increment_giv = max_reg_num ();
4213 for (n_extra_increment = 0, bl = loop_iv_list; bl; bl = bl->next)
4214 n_extra_increment += bl->biv_count - 1;
4216 /* If the loop contains volatile memory references do not allow any
4217 replacements to take place, since this could loose the volatile markers. */
4218 if (n_extra_increment && ! loop_info->has_volatile)
4220 int nregs = first_increment_giv + n_extra_increment;
4222 /* Reallocate reg_iv_type and reg_iv_info. */
4223 VARRAY_GROW (reg_iv_type, nregs);
4224 VARRAY_GROW (reg_iv_info, nregs);
4226 for (bl = loop_iv_list; bl; bl = bl->next)
4228 struct induction **vp, *v, *next;
4229 int biv_dead_after_loop = 0;
4231 /* The biv increments lists are in reverse order. Fix this first. */
4232 for (v = bl->biv, bl->biv = 0; v; v = next)
4235 v->next_iv = bl->biv;
4239 /* We must guard against the case that an early exit between v->insn
4240 and next->insn leaves the biv live after the loop, since that
4241 would mean that we'd be missing an increment for the final
4242 value. The following test to set biv_dead_after_loop is like
4243 the first part of the test to set bl->eliminable.
4244 We don't check here if we can calculate the final value, since
4245 this can't succeed if we already know that there is a jump
4246 between v->insn and next->insn, yet next->always_executed is
4247 set and next->maybe_multiple is cleared. Such a combination
4248 implies that the jump destination is outside the loop.
4249 If we want to make this check more sophisticated, we should
4250 check each branch between v->insn and next->insn individually
4251 to see if the biv is dead at its destination. */
4253 if (uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4255 && INSN_UID (bl->init_insn) < max_uid_for_loop
4256 && (uid_luid[REGNO_FIRST_UID (bl->regno)]
4257 >= INSN_LUID (bl->init_insn))
4258 #ifdef HAVE_decrement_and_branch_until_zero
4261 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4262 biv_dead_after_loop = 1;
4264 for (vp = &bl->biv, next = *vp; v = next, next = v->next_iv;)
4266 HOST_WIDE_INT offset;
4267 rtx set, add_val, old_reg, dest_reg, last_use_insn, note;
4268 int old_regno, new_regno;
4270 if (! v->always_executed
4271 || v->maybe_multiple
4272 || GET_CODE (v->add_val) != CONST_INT
4273 || ! next->always_executed
4274 || next->maybe_multiple
4275 || ! CONSTANT_P (next->add_val)
4276 || v->mult_val != const1_rtx
4277 || next->mult_val != const1_rtx
4278 || ! (biv_dead_after_loop
4279 || no_jumps_between_p (v->insn, next->insn)))
4284 offset = INTVAL (v->add_val);
4285 set = single_set (v->insn);
4286 add_val = plus_constant (next->add_val, offset);
4287 old_reg = v->dest_reg;
4288 dest_reg = gen_reg_rtx (v->mode);
4290 /* Unlike reg_iv_type / reg_iv_info, the other three arrays
4291 have been allocated with some slop space, so we may not
4292 actually need to reallocate them. If we do, the following
4293 if statement will be executed just once in this loop. */
4294 if ((unsigned) max_reg_num () > n_times_set->num_elements)
4296 /* Grow all the remaining arrays. */
4297 VARRAY_GROW (set_in_loop, nregs);
4298 VARRAY_GROW (n_times_set, nregs);
4299 VARRAY_GROW (may_not_optimize, nregs);
4300 VARRAY_GROW (reg_single_usage, nregs);
4303 if (! validate_change (next->insn, next->location, add_val, 0))
4309 /* Here we can try to eliminate the increment by combining
4310 it into the uses. */
4312 /* Set last_use_insn so that we can check against it. */
4314 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4316 p = next_insn_in_loop (p, scan_start, end, loop_top))
4318 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
4320 if (reg_mentioned_p (old_reg, PATTERN (p)))
4326 /* If we can't get the LUIDs for the insns, we can't
4327 calculate the lifetime. This is likely from unrolling
4328 of an inner loop, so there is little point in making this
4329 a DEST_REG giv anyways. */
4330 if (INSN_UID (v->insn) >= max_uid_for_loop
4331 || INSN_UID (last_use_insn) >= max_uid_for_loop
4332 || ! validate_change (v->insn, &SET_DEST (set), dest_reg, 0))
4334 /* Change the increment at NEXT back to what it was. */
4335 if (! validate_change (next->insn, next->location,
4341 next->add_val = add_val;
4342 v->dest_reg = dest_reg;
4343 v->giv_type = DEST_REG;
4344 v->location = &SET_SRC (set);
4346 v->combined_with = 0;
4348 v->derive_adjustment = 0;
4354 v->auto_inc_opt = 0;
4357 v->derived_from = 0;
4358 v->always_computable = 1;
4359 v->always_executed = 1;
4361 v->no_const_addval = 0;
4363 old_regno = REGNO (old_reg);
4364 new_regno = REGNO (dest_reg);
4365 VARRAY_INT (set_in_loop, old_regno)--;
4366 VARRAY_INT (set_in_loop, new_regno) = 1;
4367 VARRAY_INT (n_times_set, old_regno)--;
4368 VARRAY_INT (n_times_set, new_regno) = 1;
4369 VARRAY_CHAR (may_not_optimize, new_regno) = 0;
4371 REG_IV_TYPE (new_regno) = GENERAL_INDUCT;
4372 REG_IV_INFO (new_regno) = v;
4374 /* If next_insn has a REG_EQUAL note that mentiones OLD_REG,
4375 it must be replaced. */
4376 note = find_reg_note (next->insn, REG_EQUAL, NULL_RTX);
4377 if (note && reg_mentioned_p (old_reg, XEXP (note, 0)))
4378 XEXP (note, 0) = copy_rtx (SET_SRC (single_set (next->insn)));
4380 /* Remove the increment from the list of biv increments,
4381 and record it as a giv. */
4384 v->next_iv = bl->giv;
4387 v->benefit = rtx_cost (SET_SRC (set), SET);
4388 bl->total_benefit += v->benefit;
4390 /* Now replace the biv with DEST_REG in all insns between
4391 the replaced increment and the next increment, and
4392 remember the last insn that needed a replacement. */
4393 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4395 p = next_insn_in_loop (p, scan_start, end, loop_top))
4399 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
4401 if (reg_mentioned_p (old_reg, PATTERN (p)))
4404 if (! validate_replace_rtx (old_reg, dest_reg, p))
4407 for (note = REG_NOTES (p); note; note = XEXP (note, 1))
4409 if (GET_CODE (note) == EXPR_LIST)
4411 = replace_rtx (XEXP (note, 0), old_reg, dest_reg);
4415 v->last_use = last_use_insn;
4416 v->lifetime = INSN_LUID (v->insn) - INSN_LUID (last_use_insn);
4417 /* If the lifetime is zero, it means that this register is really
4418 a dead store. So mark this as a giv that can be ignored.
4419 This will not prevent the biv from being eliminated. */
4420 if (v->lifetime == 0)
4423 if (loop_dump_stream)
4424 fprintf (loop_dump_stream,
4425 "Increment %d of biv %d converted to giv %d.\n\n",
4426 INSN_UID (v->insn), old_regno, new_regno);
4430 last_increment_giv = max_reg_num () - 1;
4432 /* Search the loop for general induction variables. */
4434 /* A register is a giv if: it is only set once, it is a function of a
4435 biv and a constant (or invariant), and it is not a biv. */
4437 not_every_iteration = 0;
4444 /* At end of a straight-in loop, we are done.
4445 At end of a loop entered at the bottom, scan the top. */
4446 if (p == scan_start)
4454 if (p == scan_start)
4458 /* Look for a general induction variable in a register. */
4459 if (GET_CODE (p) == INSN
4460 && (set = single_set (p))
4461 && GET_CODE (SET_DEST (set)) == REG
4462 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
4469 rtx last_consec_insn;
4471 dest_reg = SET_DEST (set);
4472 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
4475 if (/* SET_SRC is a giv. */
4476 (general_induction_var (SET_SRC (set), &src_reg, &add_val,
4477 &mult_val, 0, &benefit)
4478 /* Equivalent expression is a giv. */
4479 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
4480 && general_induction_var (XEXP (regnote, 0), &src_reg,
4481 &add_val, &mult_val, 0,
4483 /* Don't try to handle any regs made by loop optimization.
4484 We have nothing on them in regno_first_uid, etc. */
4485 && REGNO (dest_reg) < max_reg_before_loop
4486 /* Don't recognize a BASIC_INDUCT_VAR here. */
4487 && dest_reg != src_reg
4488 /* This must be the only place where the register is set. */
4489 && (VARRAY_INT (n_times_set, REGNO (dest_reg)) == 1
4490 /* or all sets must be consecutive and make a giv. */
4491 || (benefit = consec_sets_giv (benefit, p,
4493 &add_val, &mult_val,
4494 &last_consec_insn))))
4497 = (struct induction *) alloca (sizeof (struct induction));
4499 /* If this is a library call, increase benefit. */
4500 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
4501 benefit += libcall_benefit (p);
4503 /* Skip the consecutive insns, if there are any. */
4504 if (VARRAY_INT (n_times_set, REGNO (dest_reg)) != 1)
4505 p = last_consec_insn;
4507 record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit,
4508 DEST_REG, not_every_iteration, maybe_multiple,
4509 NULL_PTR, loop_start, loop_end);
4514 #ifndef DONT_REDUCE_ADDR
4515 /* Look for givs which are memory addresses. */
4516 /* This resulted in worse code on a VAX 8600. I wonder if it
4518 if (GET_CODE (p) == INSN)
4519 find_mem_givs (PATTERN (p), p, not_every_iteration, maybe_multiple,
4520 loop_start, loop_end);
4523 /* Update the status of whether giv can derive other givs. This can
4524 change when we pass a label or an insn that updates a biv. */
4525 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4526 || GET_CODE (p) == CODE_LABEL)
4527 update_giv_derive (p);
4529 /* Past CODE_LABEL, we get to insns that may be executed multiple
4530 times. The only way we can be sure that they can't is if every
4531 every jump insn between here and the end of the loop either
4532 returns, exits the loop, is a forward jump, or is a jump
4533 to the loop start. */
4535 if (GET_CODE (p) == CODE_LABEL)
4543 insn = NEXT_INSN (insn);
4544 if (insn == scan_start)
4552 if (insn == scan_start)
4556 if (GET_CODE (insn) == JUMP_INSN
4557 && GET_CODE (PATTERN (insn)) != RETURN
4558 && (! condjump_p (insn)
4559 || (JUMP_LABEL (insn) != 0
4560 && JUMP_LABEL (insn) != scan_start
4561 && (INSN_UID (JUMP_LABEL (insn)) >= max_uid_for_loop
4562 || INSN_UID (insn) >= max_uid_for_loop
4563 || (INSN_LUID (JUMP_LABEL (insn))
4564 < INSN_LUID (insn))))))
4572 /* Past a jump, we get to insns for which we can't count
4573 on whether they will be executed during each iteration. */
4574 /* This code appears twice in strength_reduce. There is also similar
4575 code in scan_loop. */
4576 if (GET_CODE (p) == JUMP_INSN
4577 /* If we enter the loop in the middle, and scan around to the
4578 beginning, don't set not_every_iteration for that.
4579 This can be any kind of jump, since we want to know if insns
4580 will be executed if the loop is executed. */
4581 && ! (JUMP_LABEL (p) == loop_top
4582 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
4583 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
4587 /* If this is a jump outside the loop, then it also doesn't
4588 matter. Check to see if the target of this branch is on the
4589 loop_number_exits_labels list. */
4591 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
4593 label = LABEL_NEXTREF (label))
4594 if (XEXP (label, 0) == JUMP_LABEL (p))
4598 not_every_iteration = 1;
4601 else if (GET_CODE (p) == NOTE)
4603 /* At the virtual top of a converted loop, insns are again known to
4604 be executed each iteration: logically, the loop begins here
4605 even though the exit code has been duplicated.
4607 Insns are also again known to be executed each iteration at
4608 the LOOP_CONT note. */
4609 if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
4610 || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
4612 not_every_iteration = 0;
4613 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
4615 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
4619 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
4620 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
4621 or not an insn is known to be executed each iteration of the
4622 loop, whether or not any iterations are known to occur.
4624 Therefore, if we have just passed a label and have no more labels
4625 between here and the test insn of the loop, we know these insns
4626 will be executed each iteration. */
4628 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
4629 && no_labels_between_p (p, loop_end)
4630 && loop_insn_first_p (p, loop_cont))
4631 not_every_iteration = 0;
4634 /* Try to calculate and save the number of loop iterations. This is
4635 set to zero if the actual number can not be calculated. This must
4636 be called after all giv's have been identified, since otherwise it may
4637 fail if the iteration variable is a giv. */
4639 loop_iterations (loop_start, loop_end, loop_info);
4641 /* Now for each giv for which we still don't know whether or not it is
4642 replaceable, check to see if it is replaceable because its final value
4643 can be calculated. This must be done after loop_iterations is called,
4644 so that final_giv_value will work correctly. */
4646 for (bl = loop_iv_list; bl; bl = bl->next)
4648 struct induction *v;
4650 for (v = bl->giv; v; v = v->next_iv)
4651 if (! v->replaceable && ! v->not_replaceable)
4652 check_final_value (v, loop_start, loop_end, loop_info->n_iterations);
4655 /* Try to prove that the loop counter variable (if any) is always
4656 nonnegative; if so, record that fact with a REG_NONNEG note
4657 so that "decrement and branch until zero" insn can be used. */
4658 check_dbra_loop (loop_end, insn_count, loop_start, loop_info);
4660 /* Create reg_map to hold substitutions for replaceable giv regs.
4661 Some givs might have been made from biv increments, so look at
4662 reg_iv_type for a suitable size. */
4663 reg_map_size = reg_iv_type->num_elements;
4664 reg_map = (rtx *) alloca (reg_map_size * sizeof (rtx));
4665 bzero ((char *) reg_map, reg_map_size * sizeof (rtx));
4667 /* Examine each iv class for feasibility of strength reduction/induction
4668 variable elimination. */
4670 for (bl = loop_iv_list; bl; bl = bl->next)
4672 struct induction *v;
4675 rtx final_value = 0;
4678 /* Test whether it will be possible to eliminate this biv
4679 provided all givs are reduced. This is possible if either
4680 the reg is not used outside the loop, or we can compute
4681 what its final value will be.
4683 For architectures with a decrement_and_branch_until_zero insn,
4684 don't do this if we put a REG_NONNEG note on the endtest for
4687 /* Compare against bl->init_insn rather than loop_start.
4688 We aren't concerned with any uses of the biv between
4689 init_insn and loop_start since these won't be affected
4690 by the value of the biv elsewhere in the function, so
4691 long as init_insn doesn't use the biv itself.
4692 March 14, 1989 -- self@bayes.arc.nasa.gov */
4694 if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4696 && INSN_UID (bl->init_insn) < max_uid_for_loop
4697 && uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)
4698 #ifdef HAVE_decrement_and_branch_until_zero
4701 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4702 || ((final_value = final_biv_value (bl, loop_start, loop_end,
4703 loop_info->n_iterations))
4704 #ifdef HAVE_decrement_and_branch_until_zero
4708 bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0,
4709 threshold, insn_count);
4712 if (loop_dump_stream)
4714 fprintf (loop_dump_stream,
4715 "Cannot eliminate biv %d.\n",
4717 fprintf (loop_dump_stream,
4718 "First use: insn %d, last use: insn %d.\n",
4719 REGNO_FIRST_UID (bl->regno),
4720 REGNO_LAST_UID (bl->regno));
4724 /* Combine all giv's for this iv_class. */
4727 /* This will be true at the end, if all givs which depend on this
4728 biv have been strength reduced.
4729 We can't (currently) eliminate the biv unless this is so. */
4732 /* Check each giv in this class to see if we will benefit by reducing
4733 it. Skip giv's combined with others. */
4734 for (v = bl->giv; v; v = v->next_iv)
4736 struct induction *tv;
4738 if (v->ignore || v->same)
4741 benefit = v->benefit;
4743 /* Reduce benefit if not replaceable, since we will insert
4744 a move-insn to replace the insn that calculates this giv.
4745 Don't do this unless the giv is a user variable, since it
4746 will often be marked non-replaceable because of the duplication
4747 of the exit code outside the loop. In such a case, the copies
4748 we insert are dead and will be deleted. So they don't have
4749 a cost. Similar situations exist. */
4750 /* ??? The new final_[bg]iv_value code does a much better job
4751 of finding replaceable giv's, and hence this code may no longer
4753 if (! v->replaceable && ! bl->eliminable
4754 && REG_USERVAR_P (v->dest_reg))
4755 benefit -= copy_cost;
4757 /* Decrease the benefit to count the add-insns that we will
4758 insert to increment the reduced reg for the giv. */
4759 benefit -= add_cost * bl->biv_count;
4761 /* Decide whether to strength-reduce this giv or to leave the code
4762 unchanged (recompute it from the biv each time it is used).
4763 This decision can be made independently for each giv. */
4766 /* Attempt to guess whether autoincrement will handle some of the
4767 new add insns; if so, increase BENEFIT (undo the subtraction of
4768 add_cost that was done above). */
4769 if (v->giv_type == DEST_ADDR
4770 && GET_CODE (v->mult_val) == CONST_INT)
4772 if (HAVE_POST_INCREMENT
4773 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4774 benefit += add_cost * bl->biv_count;
4775 else if (HAVE_PRE_INCREMENT
4776 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4777 benefit += add_cost * bl->biv_count;
4778 else if (HAVE_POST_DECREMENT
4779 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4780 benefit += add_cost * bl->biv_count;
4781 else if (HAVE_PRE_DECREMENT
4782 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4783 benefit += add_cost * bl->biv_count;
4787 /* If an insn is not to be strength reduced, then set its ignore
4788 flag, and clear all_reduced. */
4790 /* A giv that depends on a reversed biv must be reduced if it is
4791 used after the loop exit, otherwise, it would have the wrong
4792 value after the loop exit. To make it simple, just reduce all
4793 of such giv's whether or not we know they are used after the loop
4796 if ( ! flag_reduce_all_givs && v->lifetime * threshold * benefit < insn_count
4799 if (loop_dump_stream)
4800 fprintf (loop_dump_stream,
4801 "giv of insn %d not worth while, %d vs %d.\n",
4803 v->lifetime * threshold * benefit, insn_count);
4809 /* Check that we can increment the reduced giv without a
4810 multiply insn. If not, reject it. */
4812 for (tv = bl->biv; tv; tv = tv->next_iv)
4813 if (tv->mult_val == const1_rtx
4814 && ! product_cheap_p (tv->add_val, v->mult_val))
4816 if (loop_dump_stream)
4817 fprintf (loop_dump_stream,
4818 "giv of insn %d: would need a multiply.\n",
4819 INSN_UID (v->insn));
4827 /* Check for givs whose first use is their definition and whose
4828 last use is the definition of another giv. If so, it is likely
4829 dead and should not be used to derive another giv nor to
4831 for (v = bl->giv; v; v = v->next_iv)
4834 || (v->same && v->same->ignore))
4839 struct induction *v1;
4841 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4842 if (v->last_use == v1->insn)
4845 else if (v->giv_type == DEST_REG
4846 && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
4848 struct induction *v1;
4850 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4851 if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
4856 /* Now that we know which givs will be reduced, try to rearrange the
4857 combinations to reduce register pressure.
4858 recombine_givs calls find_life_end, which needs reg_iv_type and
4859 reg_iv_info to be valid for all pseudos. We do the necessary
4860 reallocation here since it allows to check if there are still
4861 more bivs to process. */
4862 nregs = max_reg_num ();
4863 if (nregs > reg_iv_type->num_elements)
4865 /* If there are still more bivs to process, allocate some slack
4866 space so that we're not constantly reallocating these arrays. */
4869 /* Reallocate reg_iv_type and reg_iv_info. */
4870 VARRAY_GROW (reg_iv_type, nregs);
4871 VARRAY_GROW (reg_iv_info, nregs);
4873 recombine_givs (bl, loop_start, loop_end, unroll_p);
4875 /* Reduce each giv that we decided to reduce. */
4877 for (v = bl->giv; v; v = v->next_iv)
4879 struct induction *tv;
4880 if (! v->ignore && v->same == 0)
4882 int auto_inc_opt = 0;
4884 /* If the code for derived givs immediately below has already
4885 allocated a new_reg, we must keep it. */
4887 v->new_reg = gen_reg_rtx (v->mode);
4889 if (v->derived_from)
4891 struct induction *d = v->derived_from;
4893 /* In case d->dest_reg is not replaceable, we have
4894 to replace it in v->insn now. */
4896 d->new_reg = gen_reg_rtx (d->mode);
4898 = replace_rtx (PATTERN (v->insn), d->dest_reg, d->new_reg);
4900 = replace_rtx (PATTERN (v->insn), v->dest_reg, v->new_reg);
4901 /* For each place where the biv is incremented, add an
4902 insn to set the new, reduced reg for the giv.
4903 We used to do this only for biv_count != 1, but
4904 this fails when there is a giv after a single biv
4905 increment, e.g. when the last giv was expressed as
4907 for (tv = bl->biv; tv; tv = tv->next_iv)
4909 /* We always emit reduced giv increments before the
4910 biv increment when bl->biv_count != 1. So by
4911 emitting the add insns for derived givs after the
4912 biv increment, they pick up the updated value of
4914 If the reduced giv is processed with
4915 auto_inc_opt == 1, then it is incremented earlier
4916 than the biv, hence we'll still pick up the right
4918 If it's processed with auto_inc_opt == -1,
4919 that implies that the biv increment is before the
4920 first reduced giv's use. The derived giv's lifetime
4921 is after the reduced giv's lifetime, hence in this
4922 case, the biv increment doesn't matter. */
4923 emit_insn_after (copy_rtx (PATTERN (v->insn)), tv->insn);
4929 /* If the target has auto-increment addressing modes, and
4930 this is an address giv, then try to put the increment
4931 immediately after its use, so that flow can create an
4932 auto-increment addressing mode. */
4933 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
4934 && bl->biv->always_executed && ! bl->biv->maybe_multiple
4935 /* We don't handle reversed biv's because bl->biv->insn
4936 does not have a valid INSN_LUID. */
4938 && v->always_executed && ! v->maybe_multiple
4939 && INSN_UID (v->insn) < max_uid_for_loop)
4941 /* If other giv's have been combined with this one, then
4942 this will work only if all uses of the other giv's occur
4943 before this giv's insn. This is difficult to check.
4945 We simplify this by looking for the common case where
4946 there is one DEST_REG giv, and this giv's insn is the
4947 last use of the dest_reg of that DEST_REG giv. If the
4948 increment occurs after the address giv, then we can
4949 perform the optimization. (Otherwise, the increment
4950 would have to go before other_giv, and we would not be
4951 able to combine it with the address giv to get an
4952 auto-inc address.) */
4953 if (v->combined_with)
4955 struct induction *other_giv = 0;
4957 for (tv = bl->giv; tv; tv = tv->next_iv)
4965 if (! tv && other_giv
4966 && REGNO (other_giv->dest_reg) < max_reg_before_loop
4967 && (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
4968 == INSN_UID (v->insn))
4969 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
4972 /* Check for case where increment is before the address
4973 giv. Do this test in "loop order". */
4974 else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn)
4975 && (INSN_LUID (v->insn) < INSN_LUID (scan_start)
4976 || (INSN_LUID (bl->biv->insn)
4977 > INSN_LUID (scan_start))))
4978 || (INSN_LUID (v->insn) < INSN_LUID (scan_start)
4979 && (INSN_LUID (scan_start)
4980 < INSN_LUID (bl->biv->insn))))
4989 /* We can't put an insn immediately after one setting
4990 cc0, or immediately before one using cc0. */
4991 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
4992 || (auto_inc_opt == -1
4993 && (prev = prev_nonnote_insn (v->insn)) != 0
4994 && GET_RTX_CLASS (GET_CODE (prev)) == 'i'
4995 && sets_cc0_p (PATTERN (prev))))
5001 v->auto_inc_opt = 1;
5005 /* For each place where the biv is incremented, add an insn
5006 to increment the new, reduced reg for the giv. */
5007 for (tv = bl->biv; tv; tv = tv->next_iv)
5012 insert_before = tv->insn;
5013 else if (auto_inc_opt == 1)
5014 insert_before = NEXT_INSN (v->insn);
5016 insert_before = v->insn;
5018 if (tv->mult_val == const1_rtx)
5019 emit_iv_add_mult (tv->add_val, v->mult_val,
5020 v->new_reg, v->new_reg, insert_before);
5021 else /* tv->mult_val == const0_rtx */
5022 /* A multiply is acceptable here
5023 since this is presumed to be seldom executed. */
5024 emit_iv_add_mult (tv->add_val, v->mult_val,
5025 v->add_val, v->new_reg, insert_before);
5028 /* Add code at loop start to initialize giv's reduced reg. */
5030 emit_iv_add_mult (bl->initial_value, v->mult_val,
5031 v->add_val, v->new_reg, loop_start);
5035 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
5038 For each giv register that can be reduced now: if replaceable,
5039 substitute reduced reg wherever the old giv occurs;
5040 else add new move insn "giv_reg = reduced_reg". */
5042 for (v = bl->giv; v; v = v->next_iv)
5044 if (v->same && v->same->ignore)
5050 /* Update expression if this was combined, in case other giv was
5053 v->new_reg = replace_rtx (v->new_reg,
5054 v->same->dest_reg, v->same->new_reg);
5056 if (v->giv_type == DEST_ADDR)
5057 /* Store reduced reg as the address in the memref where we found
5059 validate_change (v->insn, v->location, v->new_reg, 0);
5060 else if (v->replaceable)
5062 reg_map[REGNO (v->dest_reg)] = v->new_reg;
5065 /* I can no longer duplicate the original problem. Perhaps
5066 this is unnecessary now? */
5068 /* Replaceable; it isn't strictly necessary to delete the old
5069 insn and emit a new one, because v->dest_reg is now dead.
5071 However, especially when unrolling loops, the special
5072 handling for (set REG0 REG1) in the second cse pass may
5073 make v->dest_reg live again. To avoid this problem, emit
5074 an insn to set the original giv reg from the reduced giv.
5075 We can not delete the original insn, since it may be part
5076 of a LIBCALL, and the code in flow that eliminates dead
5077 libcalls will fail if it is deleted. */
5078 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
5084 /* Not replaceable; emit an insn to set the original giv reg from
5085 the reduced giv, same as above. */
5086 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
5090 /* When a loop is reversed, givs which depend on the reversed
5091 biv, and which are live outside the loop, must be set to their
5092 correct final value. This insn is only needed if the giv is
5093 not replaceable. The correct final value is the same as the
5094 value that the giv starts the reversed loop with. */
5095 if (bl->reversed && ! v->replaceable)
5096 emit_iv_add_mult (bl->initial_value, v->mult_val,
5097 v->add_val, v->dest_reg, end_insert_before);
5098 else if (v->final_value)
5102 /* If the loop has multiple exits, emit the insn before the
5103 loop to ensure that it will always be executed no matter
5104 how the loop exits. Otherwise, emit the insn after the loop,
5105 since this is slightly more efficient. */
5106 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
5107 insert_before = loop_start;
5109 insert_before = end_insert_before;
5110 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
5114 /* If the insn to set the final value of the giv was emitted
5115 before the loop, then we must delete the insn inside the loop
5116 that sets it. If this is a LIBCALL, then we must delete
5117 every insn in the libcall. Note, however, that
5118 final_giv_value will only succeed when there are multiple
5119 exits if the giv is dead at each exit, hence it does not
5120 matter that the original insn remains because it is dead
5122 /* Delete the insn inside the loop that sets the giv since
5123 the giv is now set before (or after) the loop. */
5124 delete_insn (v->insn);
5128 if (loop_dump_stream)
5130 fprintf (loop_dump_stream, "giv at %d reduced to ",
5131 INSN_UID (v->insn));
5132 print_rtl (loop_dump_stream, v->new_reg);
5133 fprintf (loop_dump_stream, "\n");
5137 /* All the givs based on the biv bl have been reduced if they
5140 /* For each giv not marked as maybe dead that has been combined with a
5141 second giv, clear any "maybe dead" mark on that second giv.
5142 v->new_reg will either be or refer to the register of the giv it
5145 Doing this clearing avoids problems in biv elimination where a
5146 giv's new_reg is a complex value that can't be put in the insn but
5147 the giv combined with (with a reg as new_reg) is marked maybe_dead.
5148 Since the register will be used in either case, we'd prefer it be
5149 used from the simpler giv. */
5151 for (v = bl->giv; v; v = v->next_iv)
5152 if (! v->maybe_dead && v->same)
5153 v->same->maybe_dead = 0;
5155 /* Try to eliminate the biv, if it is a candidate.
5156 This won't work if ! all_reduced,
5157 since the givs we planned to use might not have been reduced.
5159 We have to be careful that we didn't initially think we could eliminate
5160 this biv because of a giv that we now think may be dead and shouldn't
5161 be used as a biv replacement.
5163 Also, there is the possibility that we may have a giv that looks
5164 like it can be used to eliminate a biv, but the resulting insn
5165 isn't valid. This can happen, for example, on the 88k, where a
5166 JUMP_INSN can compare a register only with zero. Attempts to
5167 replace it with a compare with a constant will fail.
5169 Note that in cases where this call fails, we may have replaced some
5170 of the occurrences of the biv with a giv, but no harm was done in
5171 doing so in the rare cases where it can occur. */
5173 if (all_reduced == 1 && bl->eliminable
5174 && maybe_eliminate_biv (bl, loop_start, end, 1,
5175 threshold, insn_count))
5178 /* ?? If we created a new test to bypass the loop entirely,
5179 or otherwise drop straight in, based on this test, then
5180 we might want to rewrite it also. This way some later
5181 pass has more hope of removing the initialization of this
5184 /* If final_value != 0, then the biv may be used after loop end
5185 and we must emit an insn to set it just in case.
5187 Reversed bivs already have an insn after the loop setting their
5188 value, so we don't need another one. We can't calculate the
5189 proper final value for such a biv here anyways. */
5190 if (final_value != 0 && ! bl->reversed)
5194 /* If the loop has multiple exits, emit the insn before the
5195 loop to ensure that it will always be executed no matter
5196 how the loop exits. Otherwise, emit the insn after the
5197 loop, since this is slightly more efficient. */
5198 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
5199 insert_before = loop_start;
5201 insert_before = end_insert_before;
5203 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
5208 /* Delete all of the instructions inside the loop which set
5209 the biv, as they are all dead. If is safe to delete them,
5210 because an insn setting a biv will never be part of a libcall. */
5211 /* However, deleting them will invalidate the regno_last_uid info,
5212 so keeping them around is more convenient. Final_biv_value
5213 will only succeed when there are multiple exits if the biv
5214 is dead at each exit, hence it does not matter that the original
5215 insn remains, because it is dead anyways. */
5216 for (v = bl->biv; v; v = v->next_iv)
5217 delete_insn (v->insn);
5220 if (loop_dump_stream)
5221 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
5226 /* Go through all the instructions in the loop, making all the
5227 register substitutions scheduled in REG_MAP. */
5229 for (p = loop_start; p != end; p = NEXT_INSN (p))
5230 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5231 || GET_CODE (p) == CALL_INSN)
5233 replace_regs (PATTERN (p), reg_map, reg_map_size, 0);
5234 replace_regs (REG_NOTES (p), reg_map, reg_map_size, 0);
5238 if (loop_info->n_iterations > 0)
5240 /* When we completely unroll a loop we will likely not need the increment
5241 of the loop BIV and we will not need the conditional branch at the
5243 unrolled_insn_copies = insn_count - 2;
5246 /* When we completely unroll a loop on a HAVE_cc0 machine we will not
5247 need the comparison before the conditional branch at the end of the
5249 unrolled_insn_copies -= 1;
5252 /* We'll need one copy for each loop iteration. */
5253 unrolled_insn_copies *= loop_info->n_iterations;
5255 /* A little slop to account for the ability to remove initialization
5256 code, better CSE, and other secondary benefits of completely
5257 unrolling some loops. */
5258 unrolled_insn_copies -= 1;
5260 /* Clamp the value. */
5261 if (unrolled_insn_copies < 0)
5262 unrolled_insn_copies = 0;
5265 /* Unroll loops from within strength reduction so that we can use the
5266 induction variable information that strength_reduce has already
5267 collected. Always unroll loops that would be as small or smaller
5268 unrolled than when rolled. */
5270 || (loop_info->n_iterations > 0
5271 && unrolled_insn_copies <= insn_count))
5272 unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
5275 #ifdef HAVE_decrement_and_branch_on_count
5276 /* Instrument the loop with BCT insn. */
5277 if (HAVE_decrement_and_branch_on_count && bct_p
5278 && flag_branch_on_count_reg)
5279 insert_bct (loop_start, loop_end, loop_info);
5280 #endif /* HAVE_decrement_and_branch_on_count */
5282 if (loop_dump_stream)
5283 fprintf (loop_dump_stream, "\n");
5286 VARRAY_FREE (reg_iv_type);
5287 VARRAY_FREE (reg_iv_info);
5290 /* Return 1 if X is a valid source for an initial value (or as value being
5291 compared against in an initial test).
5293 X must be either a register or constant and must not be clobbered between
5294 the current insn and the start of the loop.
5296 INSN is the insn containing X. */
5299 valid_initial_value_p (x, insn, call_seen, loop_start)
5308 /* Only consider pseudos we know about initialized in insns whose luids
5310 if (GET_CODE (x) != REG
5311 || REGNO (x) >= max_reg_before_loop)
5314 /* Don't use call-clobbered registers across a call which clobbers it. On
5315 some machines, don't use any hard registers at all. */
5316 if (REGNO (x) < FIRST_PSEUDO_REGISTER
5317 && (SMALL_REGISTER_CLASSES
5318 || (call_used_regs[REGNO (x)] && call_seen)))
5321 /* Don't use registers that have been clobbered before the start of the
5323 if (reg_set_between_p (x, insn, loop_start))
5329 /* Scan X for memory refs and check each memory address
5330 as a possible giv. INSN is the insn whose pattern X comes from.
5331 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
5332 every loop iteration. MAYBE_MULTIPLE is 1 if the insn might be executed
5333 more thanonce in each loop iteration. */
5336 find_mem_givs (x, insn, not_every_iteration, maybe_multiple, loop_start,
5340 int not_every_iteration, maybe_multiple;
5341 rtx loop_start, loop_end;
5344 register enum rtx_code code;
5345 register const char *fmt;
5350 code = GET_CODE (x);
5374 /* This code used to disable creating GIVs with mult_val == 1 and
5375 add_val == 0. However, this leads to lost optimizations when
5376 it comes time to combine a set of related DEST_ADDR GIVs, since
5377 this one would not be seen. */
5379 if (general_induction_var (XEXP (x, 0), &src_reg, &add_val,
5380 &mult_val, 1, &benefit))
5382 /* Found one; record it. */
5384 = (struct induction *) oballoc (sizeof (struct induction));
5386 record_giv (v, insn, src_reg, addr_placeholder, mult_val,
5387 add_val, benefit, DEST_ADDR, not_every_iteration,
5388 maybe_multiple, &XEXP (x, 0), loop_start, loop_end);
5390 v->mem_mode = GET_MODE (x);
5399 /* Recursively scan the subexpressions for other mem refs. */
5401 fmt = GET_RTX_FORMAT (code);
5402 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5404 find_mem_givs (XEXP (x, i), insn, not_every_iteration, maybe_multiple,
5405 loop_start, loop_end);
5406 else if (fmt[i] == 'E')
5407 for (j = 0; j < XVECLEN (x, i); j++)
5408 find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration,
5409 maybe_multiple, loop_start, loop_end);
5412 /* Fill in the data about one biv update.
5413 V is the `struct induction' in which we record the biv. (It is
5414 allocated by the caller, with alloca.)
5415 INSN is the insn that sets it.
5416 DEST_REG is the biv's reg.
5418 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
5419 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
5420 being set to INC_VAL.
5422 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
5423 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
5424 can be executed more than once per iteration. If MAYBE_MULTIPLE
5425 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
5426 executed exactly once per iteration. */
5429 record_biv (v, insn, dest_reg, inc_val, mult_val, location,
5430 not_every_iteration, maybe_multiple)
5431 struct induction *v;
5437 int not_every_iteration;
5440 struct iv_class *bl;
5443 v->src_reg = dest_reg;
5444 v->dest_reg = dest_reg;
5445 v->mult_val = mult_val;
5446 v->add_val = inc_val;
5447 v->location = location;
5448 v->mode = GET_MODE (dest_reg);
5449 v->always_computable = ! not_every_iteration;
5450 v->always_executed = ! not_every_iteration;
5451 v->maybe_multiple = maybe_multiple;
5453 /* Add this to the reg's iv_class, creating a class
5454 if this is the first incrementation of the reg. */
5456 bl = reg_biv_class[REGNO (dest_reg)];
5459 /* Create and initialize new iv_class. */
5461 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
5463 bl->regno = REGNO (dest_reg);
5469 /* Set initial value to the reg itself. */
5470 bl->initial_value = dest_reg;
5471 /* We haven't seen the initializing insn yet */
5474 bl->initial_test = 0;
5475 bl->incremented = 0;
5479 bl->total_benefit = 0;
5481 /* Add this class to loop_iv_list. */
5482 bl->next = loop_iv_list;
5485 /* Put it in the array of biv register classes. */
5486 reg_biv_class[REGNO (dest_reg)] = bl;
5489 /* Update IV_CLASS entry for this biv. */
5490 v->next_iv = bl->biv;
5493 if (mult_val == const1_rtx)
5494 bl->incremented = 1;
5496 if (loop_dump_stream)
5498 fprintf (loop_dump_stream,
5499 "Insn %d: possible biv, reg %d,",
5500 INSN_UID (insn), REGNO (dest_reg));
5501 if (GET_CODE (inc_val) == CONST_INT)
5503 fprintf (loop_dump_stream, " const =");
5504 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (inc_val));
5505 fputc ('\n', loop_dump_stream);
5509 fprintf (loop_dump_stream, " const = ");
5510 print_rtl (loop_dump_stream, inc_val);
5511 fprintf (loop_dump_stream, "\n");
5516 /* Fill in the data about one giv.
5517 V is the `struct induction' in which we record the giv. (It is
5518 allocated by the caller, with alloca.)
5519 INSN is the insn that sets it.
5520 BENEFIT estimates the savings from deleting this insn.
5521 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
5522 into a register or is used as a memory address.
5524 SRC_REG is the biv reg which the giv is computed from.
5525 DEST_REG is the giv's reg (if the giv is stored in a reg).
5526 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
5527 LOCATION points to the place where this giv's value appears in INSN. */
5530 record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
5531 type, not_every_iteration, maybe_multiple, location, loop_start,
5533 struct induction *v;
5537 rtx mult_val, add_val;
5540 int not_every_iteration, maybe_multiple;
5542 rtx loop_start, loop_end;
5544 struct induction *b;
5545 struct iv_class *bl;
5546 rtx set = single_set (insn);
5549 v->src_reg = src_reg;
5551 v->dest_reg = dest_reg;
5552 v->mult_val = mult_val;
5553 v->add_val = add_val;
5554 v->benefit = benefit;
5555 v->location = location;
5557 v->combined_with = 0;
5558 v->maybe_multiple = maybe_multiple;
5560 v->derive_adjustment = 0;
5566 v->auto_inc_opt = 0;
5569 v->derived_from = 0;
5572 /* The v->always_computable field is used in update_giv_derive, to
5573 determine whether a giv can be used to derive another giv. For a
5574 DEST_REG giv, INSN computes a new value for the giv, so its value
5575 isn't computable if INSN insn't executed every iteration.
5576 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
5577 it does not compute a new value. Hence the value is always computable
5578 regardless of whether INSN is executed each iteration. */
5580 if (type == DEST_ADDR)
5581 v->always_computable = 1;
5583 v->always_computable = ! not_every_iteration;
5585 v->always_executed = ! not_every_iteration;
5587 if (type == DEST_ADDR)
5589 v->mode = GET_MODE (*location);
5592 else /* type == DEST_REG */
5594 v->mode = GET_MODE (SET_DEST (set));
5596 v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5597 - uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]);
5599 /* If the lifetime is zero, it means that this register is
5600 really a dead store. So mark this as a giv that can be
5601 ignored. This will not prevent the biv from being eliminated. */
5602 if (v->lifetime == 0)
5605 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
5606 REG_IV_INFO (REGNO (dest_reg)) = v;
5609 /* Add the giv to the class of givs computed from one biv. */
5611 bl = reg_biv_class[REGNO (src_reg)];
5614 v->next_iv = bl->giv;
5616 /* Don't count DEST_ADDR. This is supposed to count the number of
5617 insns that calculate givs. */
5618 if (type == DEST_REG)
5620 bl->total_benefit += benefit;
5623 /* Fatal error, biv missing for this giv? */
5626 if (type == DEST_ADDR)
5630 /* The giv can be replaced outright by the reduced register only if all
5631 of the following conditions are true:
5632 - the insn that sets the giv is always executed on any iteration
5633 on which the giv is used at all
5634 (there are two ways to deduce this:
5635 either the insn is executed on every iteration,
5636 or all uses follow that insn in the same basic block),
5637 - the giv is not used outside the loop
5638 - no assignments to the biv occur during the giv's lifetime. */
5640 if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
5641 /* Previous line always fails if INSN was moved by loop opt. */
5642 && uid_luid[REGNO_LAST_UID (REGNO (dest_reg))] < INSN_LUID (loop_end)
5643 && (! not_every_iteration
5644 || last_use_this_basic_block (dest_reg, insn)))
5646 /* Now check that there are no assignments to the biv within the
5647 giv's lifetime. This requires two separate checks. */
5649 /* Check each biv update, and fail if any are between the first
5650 and last use of the giv.
5652 If this loop contains an inner loop that was unrolled, then
5653 the insn modifying the biv may have been emitted by the loop
5654 unrolling code, and hence does not have a valid luid. Just
5655 mark the biv as not replaceable in this case. It is not very
5656 useful as a biv, because it is used in two different loops.
5657 It is very unlikely that we would be able to optimize the giv
5658 using this biv anyways. */
5661 for (b = bl->biv; b; b = b->next_iv)
5663 if (INSN_UID (b->insn) >= max_uid_for_loop
5664 || ((uid_luid[INSN_UID (b->insn)]
5665 >= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))])
5666 && (uid_luid[INSN_UID (b->insn)]
5667 <= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))])))
5670 v->not_replaceable = 1;
5675 /* If there are any backwards branches that go from after the
5676 biv update to before it, then this giv is not replaceable. */
5678 for (b = bl->biv; b; b = b->next_iv)
5679 if (back_branch_in_range_p (b->insn, loop_start, loop_end))
5682 v->not_replaceable = 1;
5688 /* May still be replaceable, we don't have enough info here to
5691 v->not_replaceable = 0;
5695 /* Record whether the add_val contains a const_int, for later use by
5700 v->no_const_addval = 1;
5701 if (tem == const0_rtx)
5703 else if (GET_CODE (tem) == CONST_INT)
5704 v->no_const_addval = 0;
5705 else if (GET_CODE (tem) == PLUS)
5709 if (GET_CODE (XEXP (tem, 0)) == PLUS)
5710 tem = XEXP (tem, 0);
5711 else if (GET_CODE (XEXP (tem, 1)) == PLUS)
5712 tem = XEXP (tem, 1);
5716 if (GET_CODE (XEXP (tem, 1)) == CONST_INT)
5717 v->no_const_addval = 0;
5721 if (loop_dump_stream)
5723 if (type == DEST_REG)
5724 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
5725 INSN_UID (insn), REGNO (dest_reg));
5727 fprintf (loop_dump_stream, "Insn %d: dest address",
5730 fprintf (loop_dump_stream, " src reg %d benefit %d",
5731 REGNO (src_reg), v->benefit);
5732 fprintf (loop_dump_stream, " lifetime %d",
5736 fprintf (loop_dump_stream, " replaceable");
5738 if (v->no_const_addval)
5739 fprintf (loop_dump_stream, " ncav");
5741 if (GET_CODE (mult_val) == CONST_INT)
5743 fprintf (loop_dump_stream, " mult ");
5744 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (mult_val));
5748 fprintf (loop_dump_stream, " mult ");
5749 print_rtl (loop_dump_stream, mult_val);
5752 if (GET_CODE (add_val) == CONST_INT)
5754 fprintf (loop_dump_stream, " add ");
5755 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (add_val));
5759 fprintf (loop_dump_stream, " add ");
5760 print_rtl (loop_dump_stream, add_val);
5764 if (loop_dump_stream)
5765 fprintf (loop_dump_stream, "\n");
5770 /* All this does is determine whether a giv can be made replaceable because
5771 its final value can be calculated. This code can not be part of record_giv
5772 above, because final_giv_value requires that the number of loop iterations
5773 be known, and that can not be accurately calculated until after all givs
5774 have been identified. */
5777 check_final_value (v, loop_start, loop_end, n_iterations)
5778 struct induction *v;
5779 rtx loop_start, loop_end;
5780 unsigned HOST_WIDE_INT n_iterations;
5782 struct iv_class *bl;
5783 rtx final_value = 0;
5785 bl = reg_biv_class[REGNO (v->src_reg)];
5787 /* DEST_ADDR givs will never reach here, because they are always marked
5788 replaceable above in record_giv. */
5790 /* The giv can be replaced outright by the reduced register only if all
5791 of the following conditions are true:
5792 - the insn that sets the giv is always executed on any iteration
5793 on which the giv is used at all
5794 (there are two ways to deduce this:
5795 either the insn is executed on every iteration,
5796 or all uses follow that insn in the same basic block),
5797 - its final value can be calculated (this condition is different
5798 than the one above in record_giv)
5799 - no assignments to the biv occur during the giv's lifetime. */
5802 /* This is only called now when replaceable is known to be false. */
5803 /* Clear replaceable, so that it won't confuse final_giv_value. */
5807 if ((final_value = final_giv_value (v, loop_start, loop_end, n_iterations))
5808 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
5810 int biv_increment_seen = 0;
5816 /* When trying to determine whether or not a biv increment occurs
5817 during the lifetime of the giv, we can ignore uses of the variable
5818 outside the loop because final_value is true. Hence we can not
5819 use regno_last_uid and regno_first_uid as above in record_giv. */
5821 /* Search the loop to determine whether any assignments to the
5822 biv occur during the giv's lifetime. Start with the insn
5823 that sets the giv, and search around the loop until we come
5824 back to that insn again.
5826 Also fail if there is a jump within the giv's lifetime that jumps
5827 to somewhere outside the lifetime but still within the loop. This
5828 catches spaghetti code where the execution order is not linear, and
5829 hence the above test fails. Here we assume that the giv lifetime
5830 does not extend from one iteration of the loop to the next, so as
5831 to make the test easier. Since the lifetime isn't known yet,
5832 this requires two loops. See also record_giv above. */
5834 last_giv_use = v->insn;
5840 p = NEXT_INSN (loop_start);
5844 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5845 || GET_CODE (p) == CALL_INSN)
5847 if (biv_increment_seen)
5849 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5852 v->not_replaceable = 1;
5856 else if (reg_set_p (v->src_reg, PATTERN (p)))
5857 biv_increment_seen = 1;
5858 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5863 /* Now that the lifetime of the giv is known, check for branches
5864 from within the lifetime to outside the lifetime if it is still
5874 p = NEXT_INSN (loop_start);
5875 if (p == last_giv_use)
5878 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
5879 && LABEL_NAME (JUMP_LABEL (p))
5880 && ((loop_insn_first_p (JUMP_LABEL (p), v->insn)
5881 && loop_insn_first_p (loop_start, JUMP_LABEL (p)))
5882 || (loop_insn_first_p (last_giv_use, JUMP_LABEL (p))
5883 && loop_insn_first_p (JUMP_LABEL (p), loop_end))))
5886 v->not_replaceable = 1;
5888 if (loop_dump_stream)
5889 fprintf (loop_dump_stream,
5890 "Found branch outside giv lifetime.\n");
5897 /* If it is replaceable, then save the final value. */
5899 v->final_value = final_value;
5902 if (loop_dump_stream && v->replaceable)
5903 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
5904 INSN_UID (v->insn), REGNO (v->dest_reg));
5907 /* Update the status of whether a giv can derive other givs.
5909 We need to do something special if there is or may be an update to the biv
5910 between the time the giv is defined and the time it is used to derive
5913 In addition, a giv that is only conditionally set is not allowed to
5914 derive another giv once a label has been passed.
5916 The cases we look at are when a label or an update to a biv is passed. */
5919 update_giv_derive (p)
5922 struct iv_class *bl;
5923 struct induction *biv, *giv;
5927 /* Search all IV classes, then all bivs, and finally all givs.
5929 There are three cases we are concerned with. First we have the situation
5930 of a giv that is only updated conditionally. In that case, it may not
5931 derive any givs after a label is passed.
5933 The second case is when a biv update occurs, or may occur, after the
5934 definition of a giv. For certain biv updates (see below) that are
5935 known to occur between the giv definition and use, we can adjust the
5936 giv definition. For others, or when the biv update is conditional,
5937 we must prevent the giv from deriving any other givs. There are two
5938 sub-cases within this case.
5940 If this is a label, we are concerned with any biv update that is done
5941 conditionally, since it may be done after the giv is defined followed by
5942 a branch here (actually, we need to pass both a jump and a label, but
5943 this extra tracking doesn't seem worth it).
5945 If this is a jump, we are concerned about any biv update that may be
5946 executed multiple times. We are actually only concerned about
5947 backward jumps, but it is probably not worth performing the test
5948 on the jump again here.
5950 If this is a biv update, we must adjust the giv status to show that a
5951 subsequent biv update was performed. If this adjustment cannot be done,
5952 the giv cannot derive further givs. */
5954 for (bl = loop_iv_list; bl; bl = bl->next)
5955 for (biv = bl->biv; biv; biv = biv->next_iv)
5956 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
5959 for (giv = bl->giv; giv; giv = giv->next_iv)
5961 /* If cant_derive is already true, there is no point in
5962 checking all of these conditions again. */
5963 if (giv->cant_derive)
5966 /* If this giv is conditionally set and we have passed a label,
5967 it cannot derive anything. */
5968 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
5969 giv->cant_derive = 1;
5971 /* Skip givs that have mult_val == 0, since
5972 they are really invariants. Also skip those that are
5973 replaceable, since we know their lifetime doesn't contain
5975 else if (giv->mult_val == const0_rtx || giv->replaceable)
5978 /* The only way we can allow this giv to derive another
5979 is if this is a biv increment and we can form the product
5980 of biv->add_val and giv->mult_val. In this case, we will
5981 be able to compute a compensation. */
5982 else if (biv->insn == p)
5986 if (biv->mult_val == const1_rtx)
5987 tem = simplify_giv_expr (gen_rtx_MULT (giv->mode,
5992 if (tem && giv->derive_adjustment)
5993 tem = simplify_giv_expr
5994 (gen_rtx_PLUS (giv->mode, tem, giv->derive_adjustment),
5998 giv->derive_adjustment = tem;
6000 giv->cant_derive = 1;
6002 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
6003 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
6004 giv->cant_derive = 1;
6009 /* Check whether an insn is an increment legitimate for a basic induction var.
6010 X is the source of insn P, or a part of it.
6011 MODE is the mode in which X should be interpreted.
6013 DEST_REG is the putative biv, also the destination of the insn.
6014 We accept patterns of these forms:
6015 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
6016 REG = INVARIANT + REG
6018 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
6019 store the additive term into *INC_VAL, and store the place where
6020 we found the additive term into *LOCATION.
6022 If X is an assignment of an invariant into DEST_REG, we set
6023 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
6025 We also want to detect a BIV when it corresponds to a variable
6026 whose mode was promoted via PROMOTED_MODE. In that case, an increment
6027 of the variable may be a PLUS that adds a SUBREG of that variable to
6028 an invariant and then sign- or zero-extends the result of the PLUS
6031 Most GIVs in such cases will be in the promoted mode, since that is the
6032 probably the natural computation mode (and almost certainly the mode
6033 used for addresses) on the machine. So we view the pseudo-reg containing
6034 the variable as the BIV, as if it were simply incremented.
6036 Note that treating the entire pseudo as a BIV will result in making
6037 simple increments to any GIVs based on it. However, if the variable
6038 overflows in its declared mode but not its promoted mode, the result will
6039 be incorrect. This is acceptable if the variable is signed, since
6040 overflows in such cases are undefined, but not if it is unsigned, since
6041 those overflows are defined. So we only check for SIGN_EXTEND and
6044 If we cannot find a biv, we return 0. */
6047 basic_induction_var (x, mode, dest_reg, p, inc_val, mult_val, location)
6049 enum machine_mode mode;
6056 register enum rtx_code code;
6060 code = GET_CODE (x);
6065 if (rtx_equal_p (XEXP (x, 0), dest_reg)
6066 || (GET_CODE (XEXP (x, 0)) == SUBREG
6067 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
6068 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
6070 argp = &XEXP (x, 1);
6072 else if (rtx_equal_p (XEXP (x, 1), dest_reg)
6073 || (GET_CODE (XEXP (x, 1)) == SUBREG
6074 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
6075 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
6077 argp = &XEXP (x, 0);
6083 if (invariant_p (arg) != 1)
6086 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
6087 *mult_val = const1_rtx;
6092 /* If this is a SUBREG for a promoted variable, check the inner
6094 if (SUBREG_PROMOTED_VAR_P (x))
6095 return basic_induction_var (SUBREG_REG (x), GET_MODE (SUBREG_REG (x)),
6096 dest_reg, p, inc_val, mult_val, location);
6100 /* If this register is assigned in a previous insn, look at its
6101 source, but don't go outside the loop or past a label. */
6107 insn = PREV_INSN (insn);
6108 } while (insn && GET_CODE (insn) == NOTE
6109 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6113 set = single_set (insn);
6117 if ((SET_DEST (set) == x
6118 || (GET_CODE (SET_DEST (set)) == SUBREG
6119 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
6121 && (GET_MODE_CLASS (GET_MODE (SET_DEST (set)))
6123 && SUBREG_REG (SET_DEST (set)) == x))
6124 && basic_induction_var (SET_SRC (set),
6125 (GET_MODE (SET_SRC (set)) == VOIDmode
6127 : GET_MODE (SET_SRC (set))),
6129 inc_val, mult_val, location))
6132 /* ... fall through ... */
6134 /* Can accept constant setting of biv only when inside inner most loop.
6135 Otherwise, a biv of an inner loop may be incorrectly recognized
6136 as a biv of the outer loop,
6137 causing code to be moved INTO the inner loop. */
6139 if (invariant_p (x) != 1)
6144 /* convert_modes aborts if we try to convert to or from CCmode, so just
6145 exclude that case. It is very unlikely that a condition code value
6146 would be a useful iterator anyways. */
6147 if (this_loop_info.loops_enclosed == 1
6148 && GET_MODE_CLASS (mode) != MODE_CC
6149 && GET_MODE_CLASS (GET_MODE (dest_reg)) != MODE_CC)
6151 /* Possible bug here? Perhaps we don't know the mode of X. */
6152 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
6153 *mult_val = const0_rtx;
6160 return basic_induction_var (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
6161 dest_reg, p, inc_val, mult_val, location);
6164 /* Similar, since this can be a sign extension. */
6165 for (insn = PREV_INSN (p);
6166 (insn && GET_CODE (insn) == NOTE
6167 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6168 insn = PREV_INSN (insn))
6172 set = single_set (insn);
6174 if (set && SET_DEST (set) == XEXP (x, 0)
6175 && GET_CODE (XEXP (x, 1)) == CONST_INT
6176 && INTVAL (XEXP (x, 1)) >= 0
6177 && GET_CODE (SET_SRC (set)) == ASHIFT
6178 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
6179 return basic_induction_var (XEXP (SET_SRC (set), 0),
6180 GET_MODE (XEXP (x, 0)),
6181 dest_reg, insn, inc_val, mult_val,
6190 /* A general induction variable (giv) is any quantity that is a linear
6191 function of a basic induction variable,
6192 i.e. giv = biv * mult_val + add_val.
6193 The coefficients can be any loop invariant quantity.
6194 A giv need not be computed directly from the biv;
6195 it can be computed by way of other givs. */
6197 /* Determine whether X computes a giv.
6198 If it does, return a nonzero value
6199 which is the benefit from eliminating the computation of X;
6200 set *SRC_REG to the register of the biv that it is computed from;
6201 set *ADD_VAL and *MULT_VAL to the coefficients,
6202 such that the value of X is biv * mult + add; */
6205 general_induction_var (x, src_reg, add_val, mult_val, is_addr, pbenefit)
6216 /* If this is an invariant, forget it, it isn't a giv. */
6217 if (invariant_p (x) == 1)
6220 /* See if the expression could be a giv and get its form.
6221 Mark our place on the obstack in case we don't find a giv. */
6222 storage = (char *) oballoc (0);
6224 x = simplify_giv_expr (x, pbenefit);
6231 switch (GET_CODE (x))
6235 /* Since this is now an invariant and wasn't before, it must be a giv
6236 with MULT_VAL == 0. It doesn't matter which BIV we associate this
6238 *src_reg = loop_iv_list->biv->dest_reg;
6239 *mult_val = const0_rtx;
6244 /* This is equivalent to a BIV. */
6246 *mult_val = const1_rtx;
6247 *add_val = const0_rtx;
6251 /* Either (plus (biv) (invar)) or
6252 (plus (mult (biv) (invar_1)) (invar_2)). */
6253 if (GET_CODE (XEXP (x, 0)) == MULT)
6255 *src_reg = XEXP (XEXP (x, 0), 0);
6256 *mult_val = XEXP (XEXP (x, 0), 1);
6260 *src_reg = XEXP (x, 0);
6261 *mult_val = const1_rtx;
6263 *add_val = XEXP (x, 1);
6267 /* ADD_VAL is zero. */
6268 *src_reg = XEXP (x, 0);
6269 *mult_val = XEXP (x, 1);
6270 *add_val = const0_rtx;
6277 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
6278 unless they are CONST_INT). */
6279 if (GET_CODE (*add_val) == USE)
6280 *add_val = XEXP (*add_val, 0);
6281 if (GET_CODE (*mult_val) == USE)
6282 *mult_val = XEXP (*mult_val, 0);
6287 *pbenefit += ADDRESS_COST (orig_x) - reg_address_cost;
6289 *pbenefit += rtx_cost (orig_x, MEM) - reg_address_cost;
6293 *pbenefit += rtx_cost (orig_x, SET);
6295 /* Always return true if this is a giv so it will be detected as such,
6296 even if the benefit is zero or negative. This allows elimination
6297 of bivs that might otherwise not be eliminated. */
6301 /* Given an expression, X, try to form it as a linear function of a biv.
6302 We will canonicalize it to be of the form
6303 (plus (mult (BIV) (invar_1))
6305 with possible degeneracies.
6307 The invariant expressions must each be of a form that can be used as a
6308 machine operand. We surround then with a USE rtx (a hack, but localized
6309 and certainly unambiguous!) if not a CONST_INT for simplicity in this
6310 routine; it is the caller's responsibility to strip them.
6312 If no such canonicalization is possible (i.e., two biv's are used or an
6313 expression that is neither invariant nor a biv or giv), this routine
6316 For a non-zero return, the result will have a code of CONST_INT, USE,
6317 REG (for a BIV), PLUS, or MULT. No other codes will occur.
6319 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
6321 static rtx sge_plus PARAMS ((enum machine_mode, rtx, rtx));
6322 static rtx sge_plus_constant PARAMS ((rtx, rtx));
6323 static int cmp_combine_givs_stats PARAMS ((const PTR, const PTR));
6324 static int cmp_recombine_givs_stats PARAMS ((const PTR, const PTR));
6327 simplify_giv_expr (x, benefit)
6331 enum machine_mode mode = GET_MODE (x);
6335 /* If this is not an integer mode, or if we cannot do arithmetic in this
6336 mode, this can't be a giv. */
6337 if (mode != VOIDmode
6338 && (GET_MODE_CLASS (mode) != MODE_INT
6339 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
6342 switch (GET_CODE (x))
6345 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
6346 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
6347 if (arg0 == 0 || arg1 == 0)
6350 /* Put constant last, CONST_INT last if both constant. */
6351 if ((GET_CODE (arg0) == USE
6352 || GET_CODE (arg0) == CONST_INT)
6353 && ! ((GET_CODE (arg0) == USE
6354 && GET_CODE (arg1) == USE)
6355 || GET_CODE (arg1) == CONST_INT))
6356 tem = arg0, arg0 = arg1, arg1 = tem;
6358 /* Handle addition of zero, then addition of an invariant. */
6359 if (arg1 == const0_rtx)
6361 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
6362 switch (GET_CODE (arg0))
6366 /* Adding two invariants must result in an invariant, so enclose
6367 addition operation inside a USE and return it. */
6368 if (GET_CODE (arg0) == USE)
6369 arg0 = XEXP (arg0, 0);
6370 if (GET_CODE (arg1) == USE)
6371 arg1 = XEXP (arg1, 0);
6373 if (GET_CODE (arg0) == CONST_INT)
6374 tem = arg0, arg0 = arg1, arg1 = tem;
6375 if (GET_CODE (arg1) == CONST_INT)
6376 tem = sge_plus_constant (arg0, arg1);
6378 tem = sge_plus (mode, arg0, arg1);
6380 if (GET_CODE (tem) != CONST_INT)
6381 tem = gen_rtx_USE (mode, tem);
6386 /* biv + invar or mult + invar. Return sum. */
6387 return gen_rtx_PLUS (mode, arg0, arg1);
6390 /* (a + invar_1) + invar_2. Associate. */
6392 simplify_giv_expr (gen_rtx_PLUS (mode,
6403 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
6404 MULT to reduce cases. */
6405 if (GET_CODE (arg0) == REG)
6406 arg0 = gen_rtx_MULT (mode, arg0, const1_rtx);
6407 if (GET_CODE (arg1) == REG)
6408 arg1 = gen_rtx_MULT (mode, arg1, const1_rtx);
6410 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
6411 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
6412 Recurse to associate the second PLUS. */
6413 if (GET_CODE (arg1) == MULT)
6414 tem = arg0, arg0 = arg1, arg1 = tem;
6416 if (GET_CODE (arg1) == PLUS)
6418 simplify_giv_expr (gen_rtx_PLUS (mode,
6419 gen_rtx_PLUS (mode, arg0,
6424 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
6425 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
6428 if (!rtx_equal_p (arg0, arg1))
6431 return simplify_giv_expr (gen_rtx_MULT (mode,
6439 /* Handle "a - b" as "a + b * (-1)". */
6440 return simplify_giv_expr (gen_rtx_PLUS (mode,
6448 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
6449 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
6450 if (arg0 == 0 || arg1 == 0)
6453 /* Put constant last, CONST_INT last if both constant. */
6454 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
6455 && GET_CODE (arg1) != CONST_INT)
6456 tem = arg0, arg0 = arg1, arg1 = tem;
6458 /* If second argument is not now constant, not giv. */
6459 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
6462 /* Handle multiply by 0 or 1. */
6463 if (arg1 == const0_rtx)
6466 else if (arg1 == const1_rtx)
6469 switch (GET_CODE (arg0))
6472 /* biv * invar. Done. */
6473 return gen_rtx_MULT (mode, arg0, arg1);
6476 /* Product of two constants. */
6477 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
6480 /* invar * invar. It is a giv, but very few of these will
6481 actually pay off, so limit to simple registers. */
6482 if (GET_CODE (arg1) != CONST_INT)
6485 arg0 = XEXP (arg0, 0);
6486 if (GET_CODE (arg0) == REG)
6487 tem = gen_rtx_MULT (mode, arg0, arg1);
6488 else if (GET_CODE (arg0) == MULT
6489 && GET_CODE (XEXP (arg0, 0)) == REG
6490 && GET_CODE (XEXP (arg0, 1)) == CONST_INT)
6492 tem = gen_rtx_MULT (mode, XEXP (arg0, 0),
6493 GEN_INT (INTVAL (XEXP (arg0, 1))
6498 return gen_rtx_USE (mode, tem);
6501 /* (a * invar_1) * invar_2. Associate. */
6502 return simplify_giv_expr (gen_rtx_MULT (mode,
6510 /* (a + invar_1) * invar_2. Distribute. */
6511 return simplify_giv_expr (gen_rtx_PLUS (mode,
6525 /* Shift by constant is multiply by power of two. */
6526 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6530 simplify_giv_expr (gen_rtx_MULT (mode,
6532 GEN_INT ((HOST_WIDE_INT) 1
6533 << INTVAL (XEXP (x, 1)))),
6537 /* "-a" is "a * (-1)" */
6538 return simplify_giv_expr (gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx),
6542 /* "~a" is "-a - 1". Silly, but easy. */
6543 return simplify_giv_expr (gen_rtx_MINUS (mode,
6544 gen_rtx_NEG (mode, XEXP (x, 0)),
6549 /* Already in proper form for invariant. */
6553 /* If this is a new register, we can't deal with it. */
6554 if (REGNO (x) >= max_reg_before_loop)
6557 /* Check for biv or giv. */
6558 switch (REG_IV_TYPE (REGNO (x)))
6562 case GENERAL_INDUCT:
6564 struct induction *v = REG_IV_INFO (REGNO (x));
6566 /* Form expression from giv and add benefit. Ensure this giv
6567 can derive another and subtract any needed adjustment if so. */
6568 *benefit += v->benefit;
6572 tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode,
6573 v->src_reg, v->mult_val),
6576 if (v->derive_adjustment)
6577 tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment);
6578 return simplify_giv_expr (tem, benefit);
6582 /* If it isn't an induction variable, and it is invariant, we
6583 may be able to simplify things further by looking through
6584 the bits we just moved outside the loop. */
6585 if (invariant_p (x) == 1)
6589 for (m = the_movables; m ; m = m->next)
6590 if (rtx_equal_p (x, m->set_dest))
6592 /* Ok, we found a match. Substitute and simplify. */
6594 /* If we match another movable, we must use that, as
6595 this one is going away. */
6597 return simplify_giv_expr (m->match->set_dest, benefit);
6599 /* If consec is non-zero, this is a member of a group of
6600 instructions that were moved together. We handle this
6601 case only to the point of seeking to the last insn and
6602 looking for a REG_EQUAL. Fail if we don't find one. */
6607 do { tem = NEXT_INSN (tem); } while (--i > 0);
6609 tem = find_reg_note (tem, REG_EQUAL, NULL_RTX);
6611 tem = XEXP (tem, 0);
6615 tem = single_set (m->insn);
6617 tem = SET_SRC (tem);
6622 /* What we are most interested in is pointer
6623 arithmetic on invariants -- only take
6624 patterns we may be able to do something with. */
6625 if (GET_CODE (tem) == PLUS
6626 || GET_CODE (tem) == MULT
6627 || GET_CODE (tem) == ASHIFT
6628 || GET_CODE (tem) == CONST_INT
6629 || GET_CODE (tem) == SYMBOL_REF)
6631 tem = simplify_giv_expr (tem, benefit);
6635 else if (GET_CODE (tem) == CONST
6636 && GET_CODE (XEXP (tem, 0)) == PLUS
6637 && GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF
6638 && GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT)
6640 tem = simplify_giv_expr (XEXP (tem, 0), benefit);
6651 /* Fall through to general case. */
6653 /* If invariant, return as USE (unless CONST_INT).
6654 Otherwise, not giv. */
6655 if (GET_CODE (x) == USE)
6658 if (invariant_p (x) == 1)
6660 if (GET_CODE (x) == CONST_INT)
6662 if (GET_CODE (x) == CONST
6663 && GET_CODE (XEXP (x, 0)) == PLUS
6664 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
6665 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
6667 return gen_rtx_USE (mode, x);
6674 /* This routine folds invariants such that there is only ever one
6675 CONST_INT in the summation. It is only used by simplify_giv_expr. */
6678 sge_plus_constant (x, c)
6681 if (GET_CODE (x) == CONST_INT)
6682 return GEN_INT (INTVAL (x) + INTVAL (c));
6683 else if (GET_CODE (x) != PLUS)
6684 return gen_rtx_PLUS (GET_MODE (x), x, c);
6685 else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6687 return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
6688 GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c)));
6690 else if (GET_CODE (XEXP (x, 0)) == PLUS
6691 || GET_CODE (XEXP (x, 1)) != PLUS)
6693 return gen_rtx_PLUS (GET_MODE (x),
6694 sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1));
6698 return gen_rtx_PLUS (GET_MODE (x),
6699 sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0));
6704 sge_plus (mode, x, y)
6705 enum machine_mode mode;
6708 while (GET_CODE (y) == PLUS)
6710 rtx a = XEXP (y, 0);
6711 if (GET_CODE (a) == CONST_INT)
6712 x = sge_plus_constant (x, a);
6714 x = gen_rtx_PLUS (mode, x, a);
6717 if (GET_CODE (y) == CONST_INT)
6718 x = sge_plus_constant (x, y);
6720 x = gen_rtx_PLUS (mode, x, y);
6724 /* Help detect a giv that is calculated by several consecutive insns;
6728 The caller has already identified the first insn P as having a giv as dest;
6729 we check that all other insns that set the same register follow
6730 immediately after P, that they alter nothing else,
6731 and that the result of the last is still a giv.
6733 The value is 0 if the reg set in P is not really a giv.
6734 Otherwise, the value is the amount gained by eliminating
6735 all the consecutive insns that compute the value.
6737 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
6738 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
6740 The coefficients of the ultimate giv value are stored in
6741 *MULT_VAL and *ADD_VAL. */
6744 consec_sets_giv (first_benefit, p, src_reg, dest_reg,
6745 add_val, mult_val, last_consec_insn)
6752 rtx *last_consec_insn;
6760 /* Indicate that this is a giv so that we can update the value produced in
6761 each insn of the multi-insn sequence.
6763 This induction structure will be used only by the call to
6764 general_induction_var below, so we can allocate it on our stack.
6765 If this is a giv, our caller will replace the induct var entry with
6766 a new induction structure. */
6768 = (struct induction *) alloca (sizeof (struct induction));
6769 v->src_reg = src_reg;
6770 v->mult_val = *mult_val;
6771 v->add_val = *add_val;
6772 v->benefit = first_benefit;
6774 v->derive_adjustment = 0;
6776 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
6777 REG_IV_INFO (REGNO (dest_reg)) = v;
6779 count = VARRAY_INT (n_times_set, REGNO (dest_reg)) - 1;
6784 code = GET_CODE (p);
6786 /* If libcall, skip to end of call sequence. */
6787 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
6791 && (set = single_set (p))
6792 && GET_CODE (SET_DEST (set)) == REG
6793 && SET_DEST (set) == dest_reg
6794 && (general_induction_var (SET_SRC (set), &src_reg,
6795 add_val, mult_val, 0, &benefit)
6796 /* Giv created by equivalent expression. */
6797 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
6798 && general_induction_var (XEXP (temp, 0), &src_reg,
6799 add_val, mult_val, 0, &benefit)))
6800 && src_reg == v->src_reg)
6802 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
6803 benefit += libcall_benefit (p);
6806 v->mult_val = *mult_val;
6807 v->add_val = *add_val;
6808 v->benefit = benefit;
6810 else if (code != NOTE)
6812 /* Allow insns that set something other than this giv to a
6813 constant. Such insns are needed on machines which cannot
6814 include long constants and should not disqualify a giv. */
6816 && (set = single_set (p))
6817 && SET_DEST (set) != dest_reg
6818 && CONSTANT_P (SET_SRC (set)))
6821 REG_IV_TYPE (REGNO (dest_reg)) = UNKNOWN_INDUCT;
6826 *last_consec_insn = p;
6830 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6831 represented by G1. If no such expression can be found, or it is clear that
6832 it cannot possibly be a valid address, 0 is returned.
6834 To perform the computation, we note that
6837 where `v' is the biv.
6839 So G2 = (y/b) * G1 + (b - a*y/x).
6841 Note that MULT = y/x.
6843 Update: A and B are now allowed to be additive expressions such that
6844 B contains all variables in A. That is, computing B-A will not require
6845 subtracting variables. */
6848 express_from_1 (a, b, mult)
6851 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
6853 if (mult == const0_rtx)
6856 /* If MULT is not 1, we cannot handle A with non-constants, since we
6857 would then be required to subtract multiples of the registers in A.
6858 This is theoretically possible, and may even apply to some Fortran
6859 constructs, but it is a lot of work and we do not attempt it here. */
6861 if (mult != const1_rtx && GET_CODE (a) != CONST_INT)
6864 /* In general these structures are sorted top to bottom (down the PLUS
6865 chain), but not left to right across the PLUS. If B is a higher
6866 order giv than A, we can strip one level and recurse. If A is higher
6867 order, we'll eventually bail out, but won't know that until the end.
6868 If they are the same, we'll strip one level around this loop. */
6870 while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS)
6872 rtx ra, rb, oa, ob, tmp;
6874 ra = XEXP (a, 0), oa = XEXP (a, 1);
6875 if (GET_CODE (ra) == PLUS)
6876 tmp = ra, ra = oa, oa = tmp;
6878 rb = XEXP (b, 0), ob = XEXP (b, 1);
6879 if (GET_CODE (rb) == PLUS)
6880 tmp = rb, rb = ob, ob = tmp;
6882 if (rtx_equal_p (ra, rb))
6883 /* We matched: remove one reg completely. */
6885 else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob))
6886 /* An alternate match. */
6888 else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb))
6889 /* An alternate match. */
6893 /* Indicates an extra register in B. Strip one level from B and
6894 recurse, hoping B was the higher order expression. */
6895 ob = express_from_1 (a, ob, mult);
6898 return gen_rtx_PLUS (GET_MODE (b), rb, ob);
6902 /* Here we are at the last level of A, go through the cases hoping to
6903 get rid of everything but a constant. */
6905 if (GET_CODE (a) == PLUS)
6909 ra = XEXP (a, 0), oa = XEXP (a, 1);
6910 if (rtx_equal_p (oa, b))
6912 else if (!rtx_equal_p (ra, b))
6915 if (GET_CODE (oa) != CONST_INT)
6918 return GEN_INT (-INTVAL (oa) * INTVAL (mult));
6920 else if (GET_CODE (a) == CONST_INT)
6922 return plus_constant (b, -INTVAL (a) * INTVAL (mult));
6924 else if (GET_CODE (b) == PLUS)
6926 if (rtx_equal_p (a, XEXP (b, 0)))
6928 else if (rtx_equal_p (a, XEXP (b, 1)))
6933 else if (rtx_equal_p (a, b))
6940 express_from (g1, g2)
6941 struct induction *g1, *g2;
6945 /* The value that G1 will be multiplied by must be a constant integer. Also,
6946 the only chance we have of getting a valid address is if b*c/a (see above
6947 for notation) is also an integer. */
6948 if (GET_CODE (g1->mult_val) == CONST_INT
6949 && GET_CODE (g2->mult_val) == CONST_INT)
6951 if (g1->mult_val == const0_rtx
6952 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
6954 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
6956 else if (rtx_equal_p (g1->mult_val, g2->mult_val))
6960 /* ??? Find out if the one is a multiple of the other? */
6964 add = express_from_1 (g1->add_val, g2->add_val, mult);
6965 if (add == NULL_RTX)
6967 /* Failed. If we've got a multiplication factor between G1 and G2,
6968 scale G1's addend and try again. */
6969 if (INTVAL (mult) > 1)
6971 rtx g1_add_val = g1->add_val;
6972 if (GET_CODE (g1_add_val) == MULT
6973 && GET_CODE (XEXP (g1_add_val, 1)) == CONST_INT)
6976 m = INTVAL (mult) * INTVAL (XEXP (g1_add_val, 1));
6977 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val),
6978 XEXP (g1_add_val, 0), GEN_INT (m));
6982 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val), g1_add_val,
6986 add = express_from_1 (g1_add_val, g2->add_val, const1_rtx);
6989 if (add == NULL_RTX)
6992 /* Form simplified final result. */
6993 if (mult == const0_rtx)
6995 else if (mult == const1_rtx)
6996 mult = g1->dest_reg;
6998 mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult);
7000 if (add == const0_rtx)
7004 if (GET_CODE (add) == PLUS
7005 && CONSTANT_P (XEXP (add, 1)))
7007 rtx tem = XEXP (add, 1);
7008 mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0));
7012 return gen_rtx_PLUS (g2->mode, mult, add);
7017 /* Return an rtx, if any, that expresses giv G2 as a function of the register
7018 represented by G1. This indicates that G2 should be combined with G1 and
7019 that G2 can use (either directly or via an address expression) a register
7020 used to represent G1. */
7023 combine_givs_p (g1, g2)
7024 struct induction *g1, *g2;
7026 rtx tem = express_from (g1, g2);
7028 /* If these givs are identical, they can be combined. We use the results
7029 of express_from because the addends are not in a canonical form, so
7030 rtx_equal_p is a weaker test. */
7031 /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
7032 combination to be the other way round. */
7033 if (tem == g1->dest_reg
7034 && (g1->giv_type == DEST_REG || g2->giv_type == DEST_ADDR))
7036 return g1->dest_reg;
7039 /* If G2 can be expressed as a function of G1 and that function is valid
7040 as an address and no more expensive than using a register for G2,
7041 the expression of G2 in terms of G1 can be used. */
7043 && g2->giv_type == DEST_ADDR
7044 && memory_address_p (g2->mem_mode, tem)
7045 /* ??? Looses, especially with -fforce-addr, where *g2->location
7046 will always be a register, and so anything more complicated
7050 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location)
7052 && rtx_cost (tem, MEM) <= rtx_cost (*g2->location, MEM)
7063 struct combine_givs_stats
7070 cmp_combine_givs_stats (xp, yp)
7074 const struct combine_givs_stats * const x =
7075 (const struct combine_givs_stats *) xp;
7076 const struct combine_givs_stats * const y =
7077 (const struct combine_givs_stats *) yp;
7079 d = y->total_benefit - x->total_benefit;
7080 /* Stabilize the sort. */
7082 d = x->giv_number - y->giv_number;
7086 /* Check all pairs of givs for iv_class BL and see if any can be combined with
7087 any other. If so, point SAME to the giv combined with and set NEW_REG to
7088 be an expression (in terms of the other giv's DEST_REG) equivalent to the
7089 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
7093 struct iv_class *bl;
7095 /* Additional benefit to add for being combined multiple times. */
7096 const int extra_benefit = 3;
7098 struct induction *g1, *g2, **giv_array;
7099 int i, j, k, giv_count;
7100 struct combine_givs_stats *stats;
7103 /* Count givs, because bl->giv_count is incorrect here. */
7105 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7110 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
7112 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7114 giv_array[i++] = g1;
7116 stats = (struct combine_givs_stats *) alloca (giv_count * sizeof (*stats));
7117 bzero ((char *) stats, giv_count * sizeof (*stats));
7119 can_combine = (rtx *) alloca (giv_count * giv_count * sizeof(rtx));
7120 bzero ((char *) can_combine, giv_count * giv_count * sizeof(rtx));
7122 for (i = 0; i < giv_count; i++)
7128 stats[i].giv_number = i;
7130 /* If a DEST_REG GIV is used only once, do not allow it to combine
7131 with anything, for in doing so we will gain nothing that cannot
7132 be had by simply letting the GIV with which we would have combined
7133 to be reduced on its own. The losage shows up in particular with
7134 DEST_ADDR targets on hosts with reg+reg addressing, though it can
7135 be seen elsewhere as well. */
7136 if (g1->giv_type == DEST_REG
7137 && (single_use = VARRAY_RTX (reg_single_usage, REGNO (g1->dest_reg)))
7138 && single_use != const0_rtx)
7141 this_benefit = g1->benefit;
7142 /* Add an additional weight for zero addends. */
7143 if (g1->no_const_addval)
7146 for (j = 0; j < giv_count; j++)
7152 && (this_combine = combine_givs_p (g1, g2)) != NULL_RTX)
7154 can_combine[i*giv_count + j] = this_combine;
7155 this_benefit += g2->benefit + extra_benefit;
7158 stats[i].total_benefit = this_benefit;
7161 /* Iterate, combining until we can't. */
7163 qsort (stats, giv_count, sizeof(*stats), cmp_combine_givs_stats);
7165 if (loop_dump_stream)
7167 fprintf (loop_dump_stream, "Sorted combine statistics:\n");
7168 for (k = 0; k < giv_count; k++)
7170 g1 = giv_array[stats[k].giv_number];
7171 if (!g1->combined_with && !g1->same)
7172 fprintf (loop_dump_stream, " {%d, %d}",
7173 INSN_UID (giv_array[stats[k].giv_number]->insn),
7174 stats[k].total_benefit);
7176 putc ('\n', loop_dump_stream);
7179 for (k = 0; k < giv_count; k++)
7181 int g1_add_benefit = 0;
7183 i = stats[k].giv_number;
7186 /* If it has already been combined, skip. */
7187 if (g1->combined_with || g1->same)
7190 for (j = 0; j < giv_count; j++)
7193 if (g1 != g2 && can_combine[i*giv_count + j]
7194 /* If it has already been combined, skip. */
7195 && ! g2->same && ! g2->combined_with)
7199 g2->new_reg = can_combine[i*giv_count + j];
7201 g1->combined_with++;
7202 g1->lifetime += g2->lifetime;
7204 g1_add_benefit += g2->benefit;
7206 /* ??? The new final_[bg]iv_value code does a much better job
7207 of finding replaceable giv's, and hence this code may no
7208 longer be necessary. */
7209 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
7210 g1_add_benefit -= copy_cost;
7212 /* To help optimize the next set of combinations, remove
7213 this giv from the benefits of other potential mates. */
7214 for (l = 0; l < giv_count; ++l)
7216 int m = stats[l].giv_number;
7217 if (can_combine[m*giv_count + j])
7218 stats[l].total_benefit -= g2->benefit + extra_benefit;
7221 if (loop_dump_stream)
7222 fprintf (loop_dump_stream,
7223 "giv at %d combined with giv at %d\n",
7224 INSN_UID (g2->insn), INSN_UID (g1->insn));
7228 /* To help optimize the next set of combinations, remove
7229 this giv from the benefits of other potential mates. */
7230 if (g1->combined_with)
7232 for (j = 0; j < giv_count; ++j)
7234 int m = stats[j].giv_number;
7235 if (can_combine[m*giv_count + i])
7236 stats[j].total_benefit -= g1->benefit + extra_benefit;
7239 g1->benefit += g1_add_benefit;
7241 /* We've finished with this giv, and everything it touched.
7242 Restart the combination so that proper weights for the
7243 rest of the givs are properly taken into account. */
7244 /* ??? Ideally we would compact the arrays at this point, so
7245 as to not cover old ground. But sanely compacting
7246 can_combine is tricky. */
7252 struct recombine_givs_stats
7255 int start_luid, end_luid;
7258 /* Used below as comparison function for qsort. We want a ascending luid
7259 when scanning the array starting at the end, thus the arguments are
7262 cmp_recombine_givs_stats (xp, yp)
7266 const struct recombine_givs_stats * const x =
7267 (const struct recombine_givs_stats *) xp;
7268 const struct recombine_givs_stats * const y =
7269 (const struct recombine_givs_stats *) yp;
7271 d = y->start_luid - x->start_luid;
7272 /* Stabilize the sort. */
7274 d = y->giv_number - x->giv_number;
7278 /* Scan X, which is a part of INSN, for the end of life of a giv. Also
7279 look for the start of life of a giv where the start has not been seen
7280 yet to unlock the search for the end of its life.
7281 Only consider givs that belong to BIV.
7282 Return the total number of lifetime ends that have been found. */
7284 find_life_end (x, stats, insn, biv)
7286 struct recombine_givs_stats *stats;
7293 code = GET_CODE (x);
7298 rtx reg = SET_DEST (x);
7299 if (GET_CODE (reg) == REG)
7301 int regno = REGNO (reg);
7302 struct induction *v = REG_IV_INFO (regno);
7304 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7306 && v->src_reg == biv
7307 && stats[v->ix].end_luid <= 0)
7309 /* If we see a 0 here for end_luid, it means that we have
7310 scanned the entire loop without finding any use at all.
7311 We must not predicate this code on a start_luid match
7312 since that would make the test fail for givs that have
7313 been hoisted out of inner loops. */
7314 if (stats[v->ix].end_luid == 0)
7316 stats[v->ix].end_luid = stats[v->ix].start_luid;
7317 return 1 + find_life_end (SET_SRC (x), stats, insn, biv);
7319 else if (stats[v->ix].start_luid == INSN_LUID (insn))
7320 stats[v->ix].end_luid = 0;
7322 return find_life_end (SET_SRC (x), stats, insn, biv);
7328 int regno = REGNO (x);
7329 struct induction *v = REG_IV_INFO (regno);
7331 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7333 && v->src_reg == biv
7334 && stats[v->ix].end_luid == 0)
7336 while (INSN_UID (insn) >= max_uid_for_loop)
7337 insn = NEXT_INSN (insn);
7338 stats[v->ix].end_luid = INSN_LUID (insn);
7351 fmt = GET_RTX_FORMAT (code);
7353 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7356 retval += find_life_end (XEXP (x, i), stats, insn, biv);
7358 else if (fmt[i] == 'E')
7359 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7360 retval += find_life_end (XVECEXP (x, i, j), stats, insn, biv);
7365 /* For each giv that has been combined with another, look if
7366 we can combine it with the most recently used one instead.
7367 This tends to shorten giv lifetimes, and helps the next step:
7368 try to derive givs from other givs. */
7370 recombine_givs (bl, loop_start, loop_end, unroll_p)
7371 struct iv_class *bl;
7372 rtx loop_start, loop_end;
7375 struct induction *v, **giv_array, *last_giv;
7376 struct recombine_givs_stats *stats;
7379 int ends_need_computing;
7381 for (giv_count = 0, v = bl->giv; v; v = v->next_iv)
7387 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
7388 stats = (struct recombine_givs_stats *) alloca (giv_count * sizeof *stats);
7390 /* Initialize stats and set up the ix field for each giv in stats to name
7391 the corresponding index into stats. */
7392 for (i = 0, v = bl->giv; v; v = v->next_iv)
7399 stats[i].giv_number = i;
7400 /* If this giv has been hoisted out of an inner loop, use the luid of
7401 the previous insn. */
7402 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7404 stats[i].start_luid = INSN_LUID (p);
7408 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7410 /* Set up the ix field for each giv in stats to name
7411 the corresponding index into stats, and
7412 do the actual most-recently-used recombination. */
7413 for (last_giv = 0, i = giv_count - 1; i >= 0; i--)
7415 v = giv_array[stats[i].giv_number];
7419 struct induction *old_same = v->same;
7422 /* combine_givs_p actually says if we can make this transformation.
7423 The other tests are here only to avoid keeping a giv alive
7424 that could otherwise be eliminated. */
7426 && ((old_same->maybe_dead && ! old_same->combined_with)
7427 || ! last_giv->maybe_dead
7428 || last_giv->combined_with)
7429 && (new_combine = combine_givs_p (last_giv, v)))
7431 old_same->combined_with--;
7432 v->new_reg = new_combine;
7434 last_giv->combined_with++;
7435 /* No need to update lifetimes / benefits here since we have
7436 already decided what to reduce. */
7438 if (loop_dump_stream)
7440 fprintf (loop_dump_stream,
7441 "giv at %d recombined with giv at %d as ",
7442 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7443 print_rtl (loop_dump_stream, v->new_reg);
7444 putc ('\n', loop_dump_stream);
7450 else if (v->giv_type != DEST_REG)
7453 || (last_giv->maybe_dead && ! last_giv->combined_with)
7455 || v->combined_with)
7459 ends_need_computing = 0;
7460 /* For each DEST_REG giv, compute lifetime starts, and try to compute
7461 lifetime ends from regscan info. */
7462 for (i = giv_count - 1; i >= 0; i--)
7464 v = giv_array[stats[i].giv_number];
7467 if (v->giv_type == DEST_ADDR)
7469 /* Loop unrolling of an inner loop can even create new DEST_REG
7472 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7474 stats[i].start_luid = stats[i].end_luid = INSN_LUID (p);
7476 stats[i].end_luid++;
7478 else /* v->giv_type == DEST_REG */
7482 stats[i].start_luid = INSN_LUID (v->insn);
7483 stats[i].end_luid = INSN_LUID (v->last_use);
7485 else if (INSN_UID (v->insn) >= max_uid_for_loop)
7488 /* This insn has been created by loop optimization on an inner
7489 loop. We don't have a proper start_luid that will match
7490 when we see the first set. But we do know that there will
7491 be no use before the set, so we can set end_luid to 0 so that
7492 we'll start looking for the last use right away. */
7493 for (p = PREV_INSN (v->insn); INSN_UID (p) >= max_uid_for_loop; )
7495 stats[i].start_luid = INSN_LUID (p);
7496 stats[i].end_luid = 0;
7497 ends_need_computing++;
7501 int regno = REGNO (v->dest_reg);
7502 int count = VARRAY_INT (n_times_set, regno) - 1;
7505 /* Find the first insn that sets the giv, so that we can verify
7506 if this giv's lifetime wraps around the loop. We also need
7507 the luid of the first setting insn in order to detect the
7508 last use properly. */
7511 p = prev_nonnote_insn (p);
7512 if (reg_set_p (v->dest_reg, p))
7516 stats[i].start_luid = INSN_LUID (p);
7517 if (stats[i].start_luid > uid_luid[REGNO_FIRST_UID (regno)])
7519 stats[i].end_luid = -1;
7520 ends_need_computing++;
7524 stats[i].end_luid = uid_luid[REGNO_LAST_UID (regno)];
7525 if (stats[i].end_luid > INSN_LUID (loop_end))
7527 stats[i].end_luid = -1;
7528 ends_need_computing++;
7535 /* If the regscan information was unconclusive for one or more DEST_REG
7536 givs, scan the all insn in the loop to find out lifetime ends. */
7537 if (ends_need_computing)
7539 rtx biv = bl->biv->src_reg;
7544 if (p == loop_start)
7547 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
7549 ends_need_computing -= find_life_end (PATTERN (p), stats, p, biv);
7551 while (ends_need_computing);
7554 /* Set start_luid back to the last insn that sets the giv. This allows
7555 more combinations. */
7556 for (i = giv_count - 1; i >= 0; i--)
7558 v = giv_array[stats[i].giv_number];
7561 if (INSN_UID (v->insn) < max_uid_for_loop)
7562 stats[i].start_luid = INSN_LUID (v->insn);
7565 /* Now adjust lifetime ends by taking combined givs into account. */
7566 for (i = giv_count - 1; i >= 0; i--)
7571 v = giv_array[stats[i].giv_number];
7574 if (v->same && ! v->same->ignore)
7577 luid = stats[i].start_luid;
7578 /* Use unsigned arithmetic to model loop wrap-around. */
7579 if (luid - stats[j].start_luid
7580 > (unsigned) stats[j].end_luid - stats[j].start_luid)
7581 stats[j].end_luid = luid;
7585 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7587 /* Try to derive DEST_REG givs from previous DEST_REG givs with the
7588 same mult_val and non-overlapping lifetime. This reduces register
7590 Once we find a DEST_REG giv that is suitable to derive others from,
7591 we set last_giv to this giv, and try to derive as many other DEST_REG
7592 givs from it without joining overlapping lifetimes. If we then
7593 encounter a DEST_REG giv that we can't derive, we set rescan to the
7594 index for this giv (unless rescan is already set).
7595 When we are finished with the current LAST_GIV (i.e. the inner loop
7596 terminates), we start again with rescan, which then becomes the new
7598 for (i = giv_count - 1; i >= 0; i = rescan)
7600 int life_start, life_end;
7602 for (last_giv = 0, rescan = -1; i >= 0; i--)
7606 v = giv_array[stats[i].giv_number];
7607 if (v->giv_type != DEST_REG || v->derived_from || v->same)
7611 /* Don't use a giv that's likely to be dead to derive
7612 others - that would be likely to keep that giv alive. */
7613 if (! v->maybe_dead || v->combined_with)
7616 life_start = stats[i].start_luid;
7617 life_end = stats[i].end_luid;
7621 /* Use unsigned arithmetic to model loop wrap around. */
7622 if (((unsigned) stats[i].start_luid - life_start
7623 >= (unsigned) life_end - life_start)
7624 && ((unsigned) stats[i].end_luid - life_start
7625 > (unsigned) life_end - life_start)
7626 /* Check that the giv insn we're about to use for deriving
7627 precedes all uses of that giv. Note that initializing the
7628 derived giv would defeat the purpose of reducing register
7630 ??? We could arrange to move the insn. */
7631 && ((unsigned) stats[i].end_luid - INSN_LUID (loop_start)
7632 > (unsigned) stats[i].start_luid - INSN_LUID (loop_start))
7633 && rtx_equal_p (last_giv->mult_val, v->mult_val)
7634 /* ??? Could handle libcalls, but would need more logic. */
7635 && ! find_reg_note (v->insn, REG_RETVAL, NULL_RTX)
7636 /* We would really like to know if for any giv that v
7637 is combined with, v->insn or any intervening biv increment
7638 dominates that combined giv. However, we
7639 don't have this detailed control flow information.
7640 N.B. since last_giv will be reduced, it is valid
7641 anywhere in the loop, so we don't need to check the
7642 validity of last_giv.
7643 We rely here on the fact that v->always_executed implies that
7644 there is no jump to someplace else in the loop before the
7645 giv insn, and hence any insn that is executed before the
7646 giv insn in the loop will have a lower luid. */
7647 && (v->always_executed || ! v->combined_with)
7648 && (sum = express_from (last_giv, v))
7649 /* Make sure we don't make the add more expensive. ADD_COST
7650 doesn't take different costs of registers and constants into
7651 account, so compare the cost of the actual SET_SRCs. */
7652 && (rtx_cost (sum, SET)
7653 <= rtx_cost (SET_SRC (single_set (v->insn)), SET))
7654 /* ??? unroll can't understand anything but reg + const_int
7655 sums. It would be cleaner to fix unroll. */
7656 && ((GET_CODE (sum) == PLUS
7657 && GET_CODE (XEXP (sum, 0)) == REG
7658 && GET_CODE (XEXP (sum, 1)) == CONST_INT)
7660 && validate_change (v->insn, &PATTERN (v->insn),
7661 gen_rtx_SET (VOIDmode, v->dest_reg, sum), 0))
7663 v->derived_from = last_giv;
7664 life_end = stats[i].end_luid;
7666 if (loop_dump_stream)
7668 fprintf (loop_dump_stream,
7669 "giv at %d derived from %d as ",
7670 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7671 print_rtl (loop_dump_stream, sum);
7672 putc ('\n', loop_dump_stream);
7675 else if (rescan < 0)
7681 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
7684 emit_iv_add_mult (b, m, a, reg, insert_before)
7685 rtx b; /* initial value of basic induction variable */
7686 rtx m; /* multiplicative constant */
7687 rtx a; /* additive constant */
7688 rtx reg; /* destination register */
7694 /* Prevent unexpected sharing of these rtx. */
7698 /* Increase the lifetime of any invariants moved further in code. */
7699 update_reg_last_use (a, insert_before);
7700 update_reg_last_use (b, insert_before);
7701 update_reg_last_use (m, insert_before);
7704 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
7706 emit_move_insn (reg, result);
7707 seq = gen_sequence ();
7710 emit_insn_before (seq, insert_before);
7712 /* It is entirely possible that the expansion created lots of new
7713 registers. Iterate over the sequence we just created and
7716 if (GET_CODE (seq) == SEQUENCE)
7719 for (i = 0; i < XVECLEN (seq, 0); ++i)
7721 rtx set = single_set (XVECEXP (seq, 0, i));
7722 if (set && GET_CODE (SET_DEST (set)) == REG)
7723 record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0);
7726 else if (GET_CODE (seq) == SET
7727 && GET_CODE (SET_DEST (seq)) == REG)
7728 record_base_value (REGNO (SET_DEST (seq)), SET_SRC (seq), 0);
7731 /* Test whether A * B can be computed without
7732 an actual multiply insn. Value is 1 if so. */
7735 product_cheap_p (a, b)
7741 struct obstack *old_rtl_obstack = rtl_obstack;
7742 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
7745 /* If only one is constant, make it B. */
7746 if (GET_CODE (a) == CONST_INT)
7747 tmp = a, a = b, b = tmp;
7749 /* If first constant, both constant, so don't need multiply. */
7750 if (GET_CODE (a) == CONST_INT)
7753 /* If second not constant, neither is constant, so would need multiply. */
7754 if (GET_CODE (b) != CONST_INT)
7757 /* One operand is constant, so might not need multiply insn. Generate the
7758 code for the multiply and see if a call or multiply, or long sequence
7759 of insns is generated. */
7761 rtl_obstack = &temp_obstack;
7763 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
7764 tmp = gen_sequence ();
7767 if (GET_CODE (tmp) == SEQUENCE)
7769 if (XVEC (tmp, 0) == 0)
7771 else if (XVECLEN (tmp, 0) > 3)
7774 for (i = 0; i < XVECLEN (tmp, 0); i++)
7776 rtx insn = XVECEXP (tmp, 0, i);
7778 if (GET_CODE (insn) != INSN
7779 || (GET_CODE (PATTERN (insn)) == SET
7780 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
7781 || (GET_CODE (PATTERN (insn)) == PARALLEL
7782 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
7783 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
7790 else if (GET_CODE (tmp) == SET
7791 && GET_CODE (SET_SRC (tmp)) == MULT)
7793 else if (GET_CODE (tmp) == PARALLEL
7794 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
7795 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
7798 /* Free any storage we obtained in generating this multiply and restore rtl
7799 allocation to its normal obstack. */
7800 obstack_free (&temp_obstack, storage);
7801 rtl_obstack = old_rtl_obstack;
7806 /* Check to see if loop can be terminated by a "decrement and branch until
7807 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
7808 Also try reversing an increment loop to a decrement loop
7809 to see if the optimization can be performed.
7810 Value is nonzero if optimization was performed. */
7812 /* This is useful even if the architecture doesn't have such an insn,
7813 because it might change a loops which increments from 0 to n to a loop
7814 which decrements from n to 0. A loop that decrements to zero is usually
7815 faster than one that increments from zero. */
7817 /* ??? This could be rewritten to use some of the loop unrolling procedures,
7818 such as approx_final_value, biv_total_increment, loop_iterations, and
7819 final_[bg]iv_value. */
7822 check_dbra_loop (loop_end, insn_count, loop_start, loop_info)
7826 struct loop_info *loop_info;
7828 struct iv_class *bl;
7835 rtx before_comparison;
7839 int compare_and_branch;
7841 /* If last insn is a conditional branch, and the insn before tests a
7842 register value, try to optimize it. Otherwise, we can't do anything. */
7844 jump = PREV_INSN (loop_end);
7845 comparison = get_condition_for_loop (jump);
7846 if (comparison == 0)
7849 /* Try to compute whether the compare/branch at the loop end is one or
7850 two instructions. */
7851 get_condition (jump, &first_compare);
7852 if (first_compare == jump)
7853 compare_and_branch = 1;
7854 else if (first_compare == prev_nonnote_insn (jump))
7855 compare_and_branch = 2;
7859 /* Check all of the bivs to see if the compare uses one of them.
7860 Skip biv's set more than once because we can't guarantee that
7861 it will be zero on the last iteration. Also skip if the biv is
7862 used between its update and the test insn. */
7864 for (bl = loop_iv_list; bl; bl = bl->next)
7866 if (bl->biv_count == 1
7867 && bl->biv->dest_reg == XEXP (comparison, 0)
7868 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
7876 /* Look for the case where the basic induction variable is always
7877 nonnegative, and equals zero on the last iteration.
7878 In this case, add a reg_note REG_NONNEG, which allows the
7879 m68k DBRA instruction to be used. */
7881 if (((GET_CODE (comparison) == GT
7882 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
7883 && INTVAL (XEXP (comparison, 1)) == -1)
7884 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
7885 && GET_CODE (bl->biv->add_val) == CONST_INT
7886 && INTVAL (bl->biv->add_val) < 0)
7888 /* Initial value must be greater than 0,
7889 init_val % -dec_value == 0 to ensure that it equals zero on
7890 the last iteration */
7892 if (GET_CODE (bl->initial_value) == CONST_INT
7893 && INTVAL (bl->initial_value) > 0
7894 && (INTVAL (bl->initial_value)
7895 % (-INTVAL (bl->biv->add_val))) == 0)
7897 /* register always nonnegative, add REG_NOTE to branch */
7898 REG_NOTES (PREV_INSN (loop_end))
7899 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7900 REG_NOTES (PREV_INSN (loop_end)));
7906 /* If the decrement is 1 and the value was tested as >= 0 before
7907 the loop, then we can safely optimize. */
7908 for (p = loop_start; p; p = PREV_INSN (p))
7910 if (GET_CODE (p) == CODE_LABEL)
7912 if (GET_CODE (p) != JUMP_INSN)
7915 before_comparison = get_condition_for_loop (p);
7916 if (before_comparison
7917 && XEXP (before_comparison, 0) == bl->biv->dest_reg
7918 && GET_CODE (before_comparison) == LT
7919 && XEXP (before_comparison, 1) == const0_rtx
7920 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
7921 && INTVAL (bl->biv->add_val) == -1)
7923 REG_NOTES (PREV_INSN (loop_end))
7924 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7925 REG_NOTES (PREV_INSN (loop_end)));
7932 else if (GET_CODE (bl->biv->add_val) == CONST_INT
7933 && INTVAL (bl->biv->add_val) > 0)
7935 /* Try to change inc to dec, so can apply above optimization. */
7937 all registers modified are induction variables or invariant,
7938 all memory references have non-overlapping addresses
7939 (obviously true if only one write)
7940 allow 2 insns for the compare/jump at the end of the loop. */
7941 /* Also, we must avoid any instructions which use both the reversed
7942 biv and another biv. Such instructions will fail if the loop is
7943 reversed. We meet this condition by requiring that either
7944 no_use_except_counting is true, or else that there is only
7946 int num_nonfixed_reads = 0;
7947 /* 1 if the iteration var is used only to count iterations. */
7948 int no_use_except_counting = 0;
7949 /* 1 if the loop has no memory store, or it has a single memory store
7950 which is reversible. */
7951 int reversible_mem_store = 1;
7953 if (bl->giv_count == 0
7954 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
7956 rtx bivreg = regno_reg_rtx[bl->regno];
7958 /* If there are no givs for this biv, and the only exit is the
7959 fall through at the end of the loop, then
7960 see if perhaps there are no uses except to count. */
7961 no_use_except_counting = 1;
7962 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7963 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
7965 rtx set = single_set (p);
7967 if (set && GET_CODE (SET_DEST (set)) == REG
7968 && REGNO (SET_DEST (set)) == bl->regno)
7969 /* An insn that sets the biv is okay. */
7971 else if ((p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
7972 || p == prev_nonnote_insn (loop_end))
7973 && reg_mentioned_p (bivreg, PATTERN (p)))
7975 /* If either of these insns uses the biv and sets a pseudo
7976 that has more than one usage, then the biv has uses
7977 other than counting since it's used to derive a value
7978 that is used more than one time. */
7979 note_set_pseudo_multiple_uses_retval = 0;
7980 note_stores (PATTERN (p), note_set_pseudo_multiple_uses);
7981 if (note_set_pseudo_multiple_uses_retval)
7983 no_use_except_counting = 0;
7987 else if (reg_mentioned_p (bivreg, PATTERN (p)))
7989 no_use_except_counting = 0;
7995 if (no_use_except_counting)
7996 ; /* no need to worry about MEMs. */
7997 else if (num_mem_sets <= 1)
7999 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8000 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
8001 num_nonfixed_reads += count_nonfixed_reads (PATTERN (p));
8003 /* If the loop has a single store, and the destination address is
8004 invariant, then we can't reverse the loop, because this address
8005 might then have the wrong value at loop exit.
8006 This would work if the source was invariant also, however, in that
8007 case, the insn should have been moved out of the loop. */
8009 if (num_mem_sets == 1)
8011 struct induction *v;
8013 reversible_mem_store
8014 = (! unknown_address_altered
8015 && ! invariant_p (XEXP (XEXP (loop_store_mems, 0), 0)));
8017 /* If the store depends on a register that is set after the
8018 store, it depends on the initial value, and is thus not
8020 for (v = bl->giv; reversible_mem_store && v; v = v->next_iv)
8022 if (v->giv_type == DEST_REG
8023 && reg_mentioned_p (v->dest_reg,
8024 XEXP (loop_store_mems, 0))
8025 && loop_insn_first_p (first_loop_store_insn, v->insn))
8026 reversible_mem_store = 0;
8033 /* This code only acts for innermost loops. Also it simplifies
8034 the memory address check by only reversing loops with
8035 zero or one memory access.
8036 Two memory accesses could involve parts of the same array,
8037 and that can't be reversed.
8038 If the biv is used only for counting, than we don't need to worry
8039 about all these things. */
8041 if ((num_nonfixed_reads <= 1
8042 && ! loop_info->has_call
8043 && ! loop_info->has_volatile
8044 && reversible_mem_store
8045 && (bl->giv_count + bl->biv_count + num_mem_sets
8046 + num_movables + compare_and_branch == insn_count)
8047 && (bl == loop_iv_list && bl->next == 0))
8048 || no_use_except_counting)
8052 /* Loop can be reversed. */
8053 if (loop_dump_stream)
8054 fprintf (loop_dump_stream, "Can reverse loop\n");
8056 /* Now check other conditions:
8058 The increment must be a constant, as must the initial value,
8059 and the comparison code must be LT.
8061 This test can probably be improved since +/- 1 in the constant
8062 can be obtained by changing LT to LE and vice versa; this is
8066 /* for constants, LE gets turned into LT */
8067 && (GET_CODE (comparison) == LT
8068 || (GET_CODE (comparison) == LE
8069 && no_use_except_counting)))
8071 HOST_WIDE_INT add_val, add_adjust, comparison_val = 0;
8072 rtx initial_value, comparison_value;
8074 enum rtx_code cmp_code;
8075 int comparison_const_width;
8076 unsigned HOST_WIDE_INT comparison_sign_mask;
8078 add_val = INTVAL (bl->biv->add_val);
8079 comparison_value = XEXP (comparison, 1);
8080 if (GET_MODE (comparison_value) == VOIDmode)
8081 comparison_const_width
8082 = GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0)));
8084 comparison_const_width
8085 = GET_MODE_BITSIZE (GET_MODE (comparison_value));
8086 if (comparison_const_width > HOST_BITS_PER_WIDE_INT)
8087 comparison_const_width = HOST_BITS_PER_WIDE_INT;
8088 comparison_sign_mask
8089 = (unsigned HOST_WIDE_INT)1 << (comparison_const_width - 1);
8091 /* If the comparison value is not a loop invariant, then we
8092 can not reverse this loop.
8094 ??? If the insns which initialize the comparison value as
8095 a whole compute an invariant result, then we could move
8096 them out of the loop and proceed with loop reversal. */
8097 if (!invariant_p (comparison_value))
8100 if (GET_CODE (comparison_value) == CONST_INT)
8101 comparison_val = INTVAL (comparison_value);
8102 initial_value = bl->initial_value;
8104 /* Normalize the initial value if it is an integer and
8105 has no other use except as a counter. This will allow
8106 a few more loops to be reversed. */
8107 if (no_use_except_counting
8108 && GET_CODE (comparison_value) == CONST_INT
8109 && GET_CODE (initial_value) == CONST_INT)
8111 comparison_val = comparison_val - INTVAL (bl->initial_value);
8112 /* The code below requires comparison_val to be a multiple
8113 of add_val in order to do the loop reversal, so
8114 round up comparison_val to a multiple of add_val.
8115 Since comparison_value is constant, we know that the
8116 current comparison code is LT. */
8117 comparison_val = comparison_val + add_val - 1;
8119 -= (unsigned HOST_WIDE_INT) comparison_val % add_val;
8120 /* We postpone overflow checks for COMPARISON_VAL here;
8121 even if there is an overflow, we might still be able to
8122 reverse the loop, if converting the loop exit test to
8124 initial_value = const0_rtx;
8127 /* First check if we can do a vanilla loop reversal. */
8128 if (initial_value == const0_rtx
8129 /* If we have a decrement_and_branch_on_count,
8130 prefer the NE test, since this will allow that
8131 instruction to be generated. Note that we must
8132 use a vanilla loop reversal if the biv is used to
8133 calculate a giv or has a non-counting use. */
8134 #if ! defined (HAVE_decrement_and_branch_until_zero) \
8135 && defined (HAVE_decrement_and_branch_on_count)
8136 && (! (add_val == 1 && loop_info->vtop
8137 && (bl->biv_count == 0
8138 || no_use_except_counting)))
8140 && GET_CODE (comparison_value) == CONST_INT
8141 /* Now do postponed overflow checks on COMPARISON_VAL. */
8142 && ! (((comparison_val - add_val) ^ INTVAL (comparison_value))
8143 & comparison_sign_mask))
8145 /* Register will always be nonnegative, with value
8146 0 on last iteration */
8147 add_adjust = add_val;
8151 else if (add_val == 1 && loop_info->vtop
8152 && (bl->biv_count == 0
8153 || no_use_except_counting))
8161 if (GET_CODE (comparison) == LE)
8162 add_adjust -= add_val;
8164 /* If the initial value is not zero, or if the comparison
8165 value is not an exact multiple of the increment, then we
8166 can not reverse this loop. */
8167 if (initial_value == const0_rtx
8168 && GET_CODE (comparison_value) == CONST_INT)
8170 if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0)
8175 if (! no_use_except_counting || add_val != 1)
8179 final_value = comparison_value;
8181 /* Reset these in case we normalized the initial value
8182 and comparison value above. */
8183 if (GET_CODE (comparison_value) == CONST_INT
8184 && GET_CODE (initial_value) == CONST_INT)
8186 comparison_value = GEN_INT (comparison_val);
8188 = GEN_INT (comparison_val + INTVAL (bl->initial_value));
8190 bl->initial_value = initial_value;
8192 /* Save some info needed to produce the new insns. */
8193 reg = bl->biv->dest_reg;
8194 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
8195 if (jump_label == pc_rtx)
8196 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
8197 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
8199 /* Set start_value; if this is not a CONST_INT, we need
8201 Initialize biv to start_value before loop start.
8202 The old initializing insn will be deleted as a
8203 dead store by flow.c. */
8204 if (initial_value == const0_rtx
8205 && GET_CODE (comparison_value) == CONST_INT)
8207 start_value = GEN_INT (comparison_val - add_adjust);
8208 emit_insn_before (gen_move_insn (reg, start_value),
8211 else if (GET_CODE (initial_value) == CONST_INT)
8213 rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust);
8214 enum machine_mode mode = GET_MODE (reg);
8215 enum insn_code icode
8216 = add_optab->handlers[(int) mode].insn_code;
8218 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8219 || ! ((*insn_data[icode].operand[1].predicate)
8220 (comparison_value, mode))
8221 || ! ((*insn_data[icode].operand[2].predicate)
8225 = gen_rtx_PLUS (mode, comparison_value, offset);
8226 emit_insn_before ((GEN_FCN (icode)
8227 (reg, comparison_value, offset)),
8229 if (GET_CODE (comparison) == LE)
8230 final_value = gen_rtx_PLUS (mode, comparison_value,
8233 else if (! add_adjust)
8235 enum machine_mode mode = GET_MODE (reg);
8236 enum insn_code icode
8237 = sub_optab->handlers[(int) mode].insn_code;
8238 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8239 || ! ((*insn_data[icode].operand[1].predicate)
8240 (comparison_value, mode))
8241 || ! ((*insn_data[icode].operand[2].predicate)
8242 (initial_value, mode)))
8245 = gen_rtx_MINUS (mode, comparison_value, initial_value);
8246 emit_insn_before ((GEN_FCN (icode)
8247 (reg, comparison_value, initial_value)),
8251 /* We could handle the other cases too, but it'll be
8252 better to have a testcase first. */
8255 /* We may not have a single insn which can increment a reg, so
8256 create a sequence to hold all the insns from expand_inc. */
8258 expand_inc (reg, new_add_val);
8259 tem = gen_sequence ();
8262 p = emit_insn_before (tem, bl->biv->insn);
8263 delete_insn (bl->biv->insn);
8265 /* Update biv info to reflect its new status. */
8267 bl->initial_value = start_value;
8268 bl->biv->add_val = new_add_val;
8270 /* Update loop info. */
8271 loop_info->initial_value = reg;
8272 loop_info->initial_equiv_value = reg;
8273 loop_info->final_value = const0_rtx;
8274 loop_info->final_equiv_value = const0_rtx;
8275 loop_info->comparison_value = const0_rtx;
8276 loop_info->comparison_code = cmp_code;
8277 loop_info->increment = new_add_val;
8279 /* Inc LABEL_NUSES so that delete_insn will
8280 not delete the label. */
8281 LABEL_NUSES (XEXP (jump_label, 0)) ++;
8283 /* Emit an insn after the end of the loop to set the biv's
8284 proper exit value if it is used anywhere outside the loop. */
8285 if ((REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare))
8287 || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
8288 emit_insn_after (gen_move_insn (reg, final_value),
8291 /* Delete compare/branch at end of loop. */
8292 delete_insn (PREV_INSN (loop_end));
8293 if (compare_and_branch == 2)
8294 delete_insn (first_compare);
8296 /* Add new compare/branch insn at end of loop. */
8298 emit_cmp_and_jump_insns (reg, const0_rtx, cmp_code, NULL_RTX,
8299 GET_MODE (reg), 0, 0,
8300 XEXP (jump_label, 0));
8301 tem = gen_sequence ();
8303 emit_jump_insn_before (tem, loop_end);
8305 for (tem = PREV_INSN (loop_end);
8306 tem && GET_CODE (tem) != JUMP_INSN;
8307 tem = PREV_INSN (tem))
8311 JUMP_LABEL (tem) = XEXP (jump_label, 0);
8317 /* Increment of LABEL_NUSES done above. */
8318 /* Register is now always nonnegative,
8319 so add REG_NONNEG note to the branch. */
8320 REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
8326 /* No insn may reference both the reversed and another biv or it
8327 will fail (see comment near the top of the loop reversal
8329 Earlier on, we have verified that the biv has no use except
8330 counting, or it is the only biv in this function.
8331 However, the code that computes no_use_except_counting does
8332 not verify reg notes. It's possible to have an insn that
8333 references another biv, and has a REG_EQUAL note with an
8334 expression based on the reversed biv. To avoid this case,
8335 remove all REG_EQUAL notes based on the reversed biv
8337 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8338 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
8341 rtx set = single_set (p);
8342 /* If this is a set of a GIV based on the reversed biv, any
8343 REG_EQUAL notes should still be correct. */
8345 || GET_CODE (SET_DEST (set)) != REG
8346 || (size_t) REGNO (SET_DEST (set)) >= reg_iv_type->num_elements
8347 || REG_IV_TYPE (REGNO (SET_DEST (set))) != GENERAL_INDUCT
8348 || REG_IV_INFO (REGNO (SET_DEST (set)))->src_reg != bl->biv->src_reg)
8349 for (pnote = ®_NOTES (p); *pnote;)
8351 if (REG_NOTE_KIND (*pnote) == REG_EQUAL
8352 && reg_mentioned_p (regno_reg_rtx[bl->regno],
8354 *pnote = XEXP (*pnote, 1);
8356 pnote = &XEXP (*pnote, 1);
8360 /* Mark that this biv has been reversed. Each giv which depends
8361 on this biv, and which is also live past the end of the loop
8362 will have to be fixed up. */
8366 if (loop_dump_stream)
8368 fprintf (loop_dump_stream, "Reversed loop");
8370 fprintf (loop_dump_stream, " and added reg_nonneg\n");
8372 fprintf (loop_dump_stream, "\n");
8383 /* Verify whether the biv BL appears to be eliminable,
8384 based on the insns in the loop that refer to it.
8385 LOOP_START is the first insn of the loop, and END is the end insn.
8387 If ELIMINATE_P is non-zero, actually do the elimination.
8389 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
8390 determine whether invariant insns should be placed inside or at the
8391 start of the loop. */
8394 maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count)
8395 struct iv_class *bl;
8399 int threshold, insn_count;
8401 rtx reg = bl->biv->dest_reg;
8404 /* Scan all insns in the loop, stopping if we find one that uses the
8405 biv in a way that we cannot eliminate. */
8407 for (p = loop_start; p != end; p = NEXT_INSN (p))
8409 enum rtx_code code = GET_CODE (p);
8410 rtx where = threshold >= insn_count ? loop_start : p;
8412 /* If this is a libcall that sets a giv, skip ahead to its end. */
8413 if (GET_RTX_CLASS (code) == 'i')
8415 rtx note = find_reg_note (p, REG_LIBCALL, NULL_RTX);
8419 rtx last = XEXP (note, 0);
8420 rtx set = single_set (last);
8422 if (set && GET_CODE (SET_DEST (set)) == REG)
8424 int regno = REGNO (SET_DEST (set));
8426 if (regno < max_reg_before_loop
8427 && REG_IV_TYPE (regno) == GENERAL_INDUCT
8428 && REG_IV_INFO (regno)->src_reg == bl->biv->src_reg)
8433 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
8434 && reg_mentioned_p (reg, PATTERN (p))
8435 && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where))
8437 if (loop_dump_stream)
8438 fprintf (loop_dump_stream,
8439 "Cannot eliminate biv %d: biv used in insn %d.\n",
8440 bl->regno, INSN_UID (p));
8447 if (loop_dump_stream)
8448 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
8449 bl->regno, eliminate_p ? "was" : "can be");
8456 /* INSN and REFERENCE are instructions in the same insn chain.
8457 Return non-zero if INSN is first. */
8460 loop_insn_first_p (insn, reference)
8461 rtx insn, reference;
8465 for (p = insn, q = reference; ;)
8467 /* Start with test for not first so that INSN == REFERENCE yields not
8469 if (q == insn || ! p)
8471 if (p == reference || ! q)
8474 /* Either of P or Q might be a NOTE. Notes have the same LUID as the
8475 previous insn, hence the <= comparison below does not work if
8477 if (INSN_UID (p) < max_uid_for_loop
8478 && INSN_UID (q) < max_uid_for_loop
8479 && GET_CODE (p) != NOTE)
8480 return INSN_LUID (p) <= INSN_LUID (q);
8482 if (INSN_UID (p) >= max_uid_for_loop
8483 || GET_CODE (p) == NOTE)
8485 if (INSN_UID (q) >= max_uid_for_loop)
8490 /* We are trying to eliminate BIV in INSN using GIV. Return non-zero if
8491 the offset that we have to take into account due to auto-increment /
8492 div derivation is zero. */
8494 biv_elimination_giv_has_0_offset (biv, giv, insn)
8495 struct induction *biv, *giv;
8498 /* If the giv V had the auto-inc address optimization applied
8499 to it, and INSN occurs between the giv insn and the biv
8500 insn, then we'd have to adjust the value used here.
8501 This is rare, so we don't bother to make this possible. */
8502 if (giv->auto_inc_opt
8503 && ((loop_insn_first_p (giv->insn, insn)
8504 && loop_insn_first_p (insn, biv->insn))
8505 || (loop_insn_first_p (biv->insn, insn)
8506 && loop_insn_first_p (insn, giv->insn))))
8509 /* If the giv V was derived from another giv, and INSN does
8510 not occur between the giv insn and the biv insn, then we'd
8511 have to adjust the value used here. This is rare, so we don't
8512 bother to make this possible. */
8513 if (giv->derived_from
8514 && ! (giv->always_executed
8515 && loop_insn_first_p (giv->insn, insn)
8516 && loop_insn_first_p (insn, biv->insn)))
8519 && giv->same->derived_from
8520 && ! (giv->same->always_executed
8521 && loop_insn_first_p (giv->same->insn, insn)
8522 && loop_insn_first_p (insn, biv->insn)))
8528 /* If BL appears in X (part of the pattern of INSN), see if we can
8529 eliminate its use. If so, return 1. If not, return 0.
8531 If BIV does not appear in X, return 1.
8533 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
8534 where extra insns should be added. Depending on how many items have been
8535 moved out of the loop, it will either be before INSN or at the start of
8539 maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where)
8541 struct iv_class *bl;
8545 enum rtx_code code = GET_CODE (x);
8546 rtx reg = bl->biv->dest_reg;
8547 enum machine_mode mode = GET_MODE (reg);
8548 struct induction *v;
8560 /* If we haven't already been able to do something with this BIV,
8561 we can't eliminate it. */
8567 /* If this sets the BIV, it is not a problem. */
8568 if (SET_DEST (x) == reg)
8571 /* If this is an insn that defines a giv, it is also ok because
8572 it will go away when the giv is reduced. */
8573 for (v = bl->giv; v; v = v->next_iv)
8574 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
8578 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
8580 /* Can replace with any giv that was reduced and
8581 that has (MULT_VAL != 0) and (ADD_VAL == 0).
8582 Require a constant for MULT_VAL, so we know it's nonzero.
8583 ??? We disable this optimization to avoid potential
8586 for (v = bl->giv; v; v = v->next_iv)
8587 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
8588 && v->add_val == const0_rtx
8589 && ! v->ignore && ! v->maybe_dead && v->always_computable
8593 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8599 /* If the giv has the opposite direction of change,
8600 then reverse the comparison. */
8601 if (INTVAL (v->mult_val) < 0)
8602 new = gen_rtx_COMPARE (GET_MODE (v->new_reg),
8603 const0_rtx, v->new_reg);
8607 /* We can probably test that giv's reduced reg. */
8608 if (validate_change (insn, &SET_SRC (x), new, 0))
8612 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
8613 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
8614 Require a constant for MULT_VAL, so we know it's nonzero.
8615 ??? Do this only if ADD_VAL is a pointer to avoid a potential
8616 overflow problem. */
8618 for (v = bl->giv; v; v = v->next_iv)
8619 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
8620 && ! v->ignore && ! v->maybe_dead && v->always_computable
8622 && (GET_CODE (v->add_val) == SYMBOL_REF
8623 || GET_CODE (v->add_val) == LABEL_REF
8624 || GET_CODE (v->add_val) == CONST
8625 || (GET_CODE (v->add_val) == REG
8626 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
8628 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8634 /* If the giv has the opposite direction of change,
8635 then reverse the comparison. */
8636 if (INTVAL (v->mult_val) < 0)
8637 new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val),
8640 new = gen_rtx_COMPARE (VOIDmode, v->new_reg,
8641 copy_rtx (v->add_val));
8643 /* Replace biv with the giv's reduced register. */
8644 update_reg_last_use (v->add_val, insn);
8645 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8648 /* Insn doesn't support that constant or invariant. Copy it
8649 into a register (it will be a loop invariant.) */
8650 tem = gen_reg_rtx (GET_MODE (v->new_reg));
8652 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
8655 /* Substitute the new register for its invariant value in
8656 the compare expression. */
8657 XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
8658 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8667 case GT: case GE: case GTU: case GEU:
8668 case LT: case LE: case LTU: case LEU:
8669 /* See if either argument is the biv. */
8670 if (XEXP (x, 0) == reg)
8671 arg = XEXP (x, 1), arg_operand = 1;
8672 else if (XEXP (x, 1) == reg)
8673 arg = XEXP (x, 0), arg_operand = 0;
8677 if (CONSTANT_P (arg))
8679 /* First try to replace with any giv that has constant positive
8680 mult_val and constant add_val. We might be able to support
8681 negative mult_val, but it seems complex to do it in general. */
8683 for (v = bl->giv; v; v = v->next_iv)
8684 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8685 && (GET_CODE (v->add_val) == SYMBOL_REF
8686 || GET_CODE (v->add_val) == LABEL_REF
8687 || GET_CODE (v->add_val) == CONST
8688 || (GET_CODE (v->add_val) == REG
8689 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
8690 && ! v->ignore && ! v->maybe_dead && v->always_computable
8693 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8699 /* Replace biv with the giv's reduced reg. */
8700 XEXP (x, 1-arg_operand) = v->new_reg;
8702 /* If all constants are actually constant integers and
8703 the derived constant can be directly placed in the COMPARE,
8705 if (GET_CODE (arg) == CONST_INT
8706 && GET_CODE (v->mult_val) == CONST_INT
8707 && GET_CODE (v->add_val) == CONST_INT
8708 && validate_change (insn, &XEXP (x, arg_operand),
8709 GEN_INT (INTVAL (arg)
8710 * INTVAL (v->mult_val)
8711 + INTVAL (v->add_val)), 0))
8714 /* Otherwise, load it into a register. */
8715 tem = gen_reg_rtx (mode);
8716 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8717 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
8720 /* If that failed, put back the change we made above. */
8721 XEXP (x, 1-arg_operand) = reg;
8724 /* Look for giv with positive constant mult_val and nonconst add_val.
8725 Insert insns to calculate new compare value.
8726 ??? Turn this off due to possible overflow. */
8728 for (v = bl->giv; v; v = v->next_iv)
8729 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8730 && ! v->ignore && ! v->maybe_dead && v->always_computable
8736 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8742 tem = gen_reg_rtx (mode);
8744 /* Replace biv with giv's reduced register. */
8745 validate_change (insn, &XEXP (x, 1 - arg_operand),
8748 /* Compute value to compare against. */
8749 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8750 /* Use it in this insn. */
8751 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8752 if (apply_change_group ())
8756 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
8758 if (invariant_p (arg) == 1)
8760 /* Look for giv with constant positive mult_val and nonconst
8761 add_val. Insert insns to compute new compare value.
8762 ??? Turn this off due to possible overflow. */
8764 for (v = bl->giv; v; v = v->next_iv)
8765 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8766 && ! v->ignore && ! v->maybe_dead && v->always_computable
8772 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8778 tem = gen_reg_rtx (mode);
8780 /* Replace biv with giv's reduced register. */
8781 validate_change (insn, &XEXP (x, 1 - arg_operand),
8784 /* Compute value to compare against. */
8785 emit_iv_add_mult (arg, v->mult_val, v->add_val,
8787 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8788 if (apply_change_group ())
8793 /* This code has problems. Basically, you can't know when
8794 seeing if we will eliminate BL, whether a particular giv
8795 of ARG will be reduced. If it isn't going to be reduced,
8796 we can't eliminate BL. We can try forcing it to be reduced,
8797 but that can generate poor code.
8799 The problem is that the benefit of reducing TV, below should
8800 be increased if BL can actually be eliminated, but this means
8801 we might have to do a topological sort of the order in which
8802 we try to process biv. It doesn't seem worthwhile to do
8803 this sort of thing now. */
8806 /* Otherwise the reg compared with had better be a biv. */
8807 if (GET_CODE (arg) != REG
8808 || REG_IV_TYPE (REGNO (arg)) != BASIC_INDUCT)
8811 /* Look for a pair of givs, one for each biv,
8812 with identical coefficients. */
8813 for (v = bl->giv; v; v = v->next_iv)
8815 struct induction *tv;
8817 if (v->ignore || v->maybe_dead || v->mode != mode)
8820 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
8821 if (! tv->ignore && ! tv->maybe_dead
8822 && rtx_equal_p (tv->mult_val, v->mult_val)
8823 && rtx_equal_p (tv->add_val, v->add_val)
8824 && tv->mode == mode)
8826 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8832 /* Replace biv with its giv's reduced reg. */
8833 XEXP (x, 1-arg_operand) = v->new_reg;
8834 /* Replace other operand with the other giv's
8836 XEXP (x, arg_operand) = tv->new_reg;
8843 /* If we get here, the biv can't be eliminated. */
8847 /* If this address is a DEST_ADDR giv, it doesn't matter if the
8848 biv is used in it, since it will be replaced. */
8849 for (v = bl->giv; v; v = v->next_iv)
8850 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
8858 /* See if any subexpression fails elimination. */
8859 fmt = GET_RTX_FORMAT (code);
8860 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
8865 if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl,
8866 eliminate_p, where))
8871 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8872 if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl,
8873 eliminate_p, where))
8882 /* Return nonzero if the last use of REG
8883 is in an insn following INSN in the same basic block. */
8886 last_use_this_basic_block (reg, insn)
8892 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
8895 if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
8901 /* Called via `note_stores' to record the initial value of a biv. Here we
8902 just record the location of the set and process it later. */
8905 record_initial (dest, set)
8909 struct iv_class *bl;
8911 if (GET_CODE (dest) != REG
8912 || REGNO (dest) >= max_reg_before_loop
8913 || REG_IV_TYPE (REGNO (dest)) != BASIC_INDUCT)
8916 bl = reg_biv_class[REGNO (dest)];
8918 /* If this is the first set found, record it. */
8919 if (bl->init_insn == 0)
8921 bl->init_insn = note_insn;
8926 /* If any of the registers in X are "old" and currently have a last use earlier
8927 than INSN, update them to have a last use of INSN. Their actual last use
8928 will be the previous insn but it will not have a valid uid_luid so we can't
8932 update_reg_last_use (x, insn)
8936 /* Check for the case where INSN does not have a valid luid. In this case,
8937 there is no need to modify the regno_last_uid, as this can only happen
8938 when code is inserted after the loop_end to set a pseudo's final value,
8939 and hence this insn will never be the last use of x. */
8940 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
8941 && INSN_UID (insn) < max_uid_for_loop
8942 && uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)])
8943 REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
8947 register const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
8948 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
8951 update_reg_last_use (XEXP (x, i), insn);
8952 else if (fmt[i] == 'E')
8953 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8954 update_reg_last_use (XVECEXP (x, i, j), insn);
8959 /* Given a jump insn JUMP, return the condition that will cause it to branch
8960 to its JUMP_LABEL. If the condition cannot be understood, or is an
8961 inequality floating-point comparison which needs to be reversed, 0 will
8964 If EARLIEST is non-zero, it is a pointer to a place where the earliest
8965 insn used in locating the condition was found. If a replacement test
8966 of the condition is desired, it should be placed in front of that
8967 insn and we will be sure that the inputs are still valid.
8969 The condition will be returned in a canonical form to simplify testing by
8970 callers. Specifically:
8972 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
8973 (2) Both operands will be machine operands; (cc0) will have been replaced.
8974 (3) If an operand is a constant, it will be the second operand.
8975 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
8976 for GE, GEU, and LEU. */
8979 get_condition (jump, earliest)
8988 int reverse_code = 0;
8989 int did_reverse_condition = 0;
8990 enum machine_mode mode;
8992 /* If this is not a standard conditional jump, we can't parse it. */
8993 if (GET_CODE (jump) != JUMP_INSN
8994 || ! condjump_p (jump) || simplejump_p (jump))
8997 code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0));
8998 mode = GET_MODE (XEXP (SET_SRC (PATTERN (jump)), 0));
8999 op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0);
9000 op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1);
9005 /* If this branches to JUMP_LABEL when the condition is false, reverse
9007 if (GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF
9008 && XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump))
9009 code = reverse_condition (code), did_reverse_condition ^= 1;
9011 /* If we are comparing a register with zero, see if the register is set
9012 in the previous insn to a COMPARE or a comparison operation. Perform
9013 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
9016 while (GET_RTX_CLASS (code) == '<' && op1 == CONST0_RTX (GET_MODE (op0)))
9018 /* Set non-zero when we find something of interest. */
9022 /* If comparison with cc0, import actual comparison from compare
9026 if ((prev = prev_nonnote_insn (prev)) == 0
9027 || GET_CODE (prev) != INSN
9028 || (set = single_set (prev)) == 0
9029 || SET_DEST (set) != cc0_rtx)
9032 op0 = SET_SRC (set);
9033 op1 = CONST0_RTX (GET_MODE (op0));
9039 /* If this is a COMPARE, pick up the two things being compared. */
9040 if (GET_CODE (op0) == COMPARE)
9042 op1 = XEXP (op0, 1);
9043 op0 = XEXP (op0, 0);
9046 else if (GET_CODE (op0) != REG)
9049 /* Go back to the previous insn. Stop if it is not an INSN. We also
9050 stop if it isn't a single set or if it has a REG_INC note because
9051 we don't want to bother dealing with it. */
9053 if ((prev = prev_nonnote_insn (prev)) == 0
9054 || GET_CODE (prev) != INSN
9055 || FIND_REG_INC_NOTE (prev, 0)
9056 || (set = single_set (prev)) == 0)
9059 /* If this is setting OP0, get what it sets it to if it looks
9061 if (rtx_equal_p (SET_DEST (set), op0))
9063 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
9065 /* ??? We may not combine comparisons done in a CCmode with
9066 comparisons not done in a CCmode. This is to aid targets
9067 like Alpha that have an IEEE compliant EQ instruction, and
9068 a non-IEEE compliant BEQ instruction. The use of CCmode is
9069 actually artificial, simply to prevent the combination, but
9070 should not affect other platforms.
9072 However, we must allow VOIDmode comparisons to match either
9073 CCmode or non-CCmode comparison, because some ports have
9074 modeless comparisons inside branch patterns.
9076 ??? This mode check should perhaps look more like the mode check
9077 in simplify_comparison in combine. */
9079 if ((GET_CODE (SET_SRC (set)) == COMPARE
9082 && GET_MODE_CLASS (inner_mode) == MODE_INT
9083 && (GET_MODE_BITSIZE (inner_mode)
9084 <= HOST_BITS_PER_WIDE_INT)
9085 && (STORE_FLAG_VALUE
9086 & ((HOST_WIDE_INT) 1
9087 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9088 #ifdef FLOAT_STORE_FLAG_VALUE
9090 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9091 && FLOAT_STORE_FLAG_VALUE < 0)
9094 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'))
9095 && (((GET_MODE_CLASS (mode) == MODE_CC)
9096 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9097 || mode == VOIDmode || inner_mode == VOIDmode))
9099 else if (((code == EQ
9101 && (GET_MODE_BITSIZE (inner_mode)
9102 <= HOST_BITS_PER_WIDE_INT)
9103 && GET_MODE_CLASS (inner_mode) == MODE_INT
9104 && (STORE_FLAG_VALUE
9105 & ((HOST_WIDE_INT) 1
9106 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9107 #ifdef FLOAT_STORE_FLAG_VALUE
9109 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9110 && FLOAT_STORE_FLAG_VALUE < 0)
9113 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'
9114 && (((GET_MODE_CLASS (mode) == MODE_CC)
9115 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9116 || mode == VOIDmode || inner_mode == VOIDmode))
9119 /* We might have reversed a LT to get a GE here. But this wasn't
9120 actually the comparison of data, so we don't flag that we
9121 have had to reverse the condition. */
9122 did_reverse_condition ^= 1;
9130 else if (reg_set_p (op0, prev))
9131 /* If this sets OP0, but not directly, we have to give up. */
9136 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
9137 code = GET_CODE (x);
9140 code = reverse_condition (code);
9141 did_reverse_condition ^= 1;
9145 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
9151 /* If constant is first, put it last. */
9152 if (CONSTANT_P (op0))
9153 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
9155 /* If OP0 is the result of a comparison, we weren't able to find what
9156 was really being compared, so fail. */
9157 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
9160 /* Canonicalize any ordered comparison with integers involving equality
9161 if we can do computations in the relevant mode and we do not
9164 if (GET_CODE (op1) == CONST_INT
9165 && GET_MODE (op0) != VOIDmode
9166 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
9168 HOST_WIDE_INT const_val = INTVAL (op1);
9169 unsigned HOST_WIDE_INT uconst_val = const_val;
9170 unsigned HOST_WIDE_INT max_val
9171 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
9176 if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
9177 code = LT, op1 = GEN_INT (const_val + 1);
9180 /* When cross-compiling, const_val might be sign-extended from
9181 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
9183 if ((HOST_WIDE_INT) (const_val & max_val)
9184 != (((HOST_WIDE_INT) 1
9185 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
9186 code = GT, op1 = GEN_INT (const_val - 1);
9190 if (uconst_val < max_val)
9191 code = LTU, op1 = GEN_INT (uconst_val + 1);
9195 if (uconst_val != 0)
9196 code = GTU, op1 = GEN_INT (uconst_val - 1);
9204 /* If this was floating-point and we reversed anything other than an
9205 EQ or NE, return zero. */
9206 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
9207 && did_reverse_condition && code != NE && code != EQ
9209 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
9213 /* Never return CC0; return zero instead. */
9218 return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);
9221 /* Similar to above routine, except that we also put an invariant last
9222 unless both operands are invariants. */
9225 get_condition_for_loop (x)
9228 rtx comparison = get_condition (x, NULL_PTR);
9231 || ! invariant_p (XEXP (comparison, 0))
9232 || invariant_p (XEXP (comparison, 1)))
9235 return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode,
9236 XEXP (comparison, 1), XEXP (comparison, 0));
9239 #ifdef HAVE_decrement_and_branch_on_count
9240 /* Instrument loop for insertion of bct instruction. We distinguish between
9241 loops with compile-time bounds and those with run-time bounds.
9242 Information from loop_iterations() is used to compute compile-time bounds.
9243 Run-time bounds should use loop preconditioning, but currently ignored.
9247 insert_bct (loop_start, loop_end, loop_info)
9248 rtx loop_start, loop_end;
9249 struct loop_info *loop_info;
9252 unsigned HOST_WIDE_INT n_iterations;
9254 int increment_direction, compare_direction;
9256 /* If the loop condition is <= or >=, the number of iteration
9257 is 1 more than the range of the bounds of the loop. */
9258 int add_iteration = 0;
9260 enum machine_mode loop_var_mode = word_mode;
9262 int loop_num = uid_loop_num [INSN_UID (loop_start)];
9264 /* It's impossible to instrument a competely unrolled loop. */
9265 if (loop_info->unroll_number == loop_info->n_iterations)
9268 /* Make sure that the count register is not in use. */
9269 if (loop_used_count_register [loop_num])
9271 if (loop_dump_stream)
9272 fprintf (loop_dump_stream,
9273 "insert_bct %d: BCT instrumentation failed: count register already in use\n",
9278 /* Make sure that the function has no indirect jumps. */
9279 if (indirect_jump_in_function)
9281 if (loop_dump_stream)
9282 fprintf (loop_dump_stream,
9283 "insert_bct %d: BCT instrumentation failed: indirect jump in function\n",
9288 /* Make sure that the last loop insn is a conditional jump. */
9289 if (GET_CODE (PREV_INSN (loop_end)) != JUMP_INSN
9290 || ! condjump_p (PREV_INSN (loop_end))
9291 || simplejump_p (PREV_INSN (loop_end)))
9293 if (loop_dump_stream)
9294 fprintf (loop_dump_stream,
9295 "insert_bct %d: BCT instrumentation failed: invalid jump at loop end\n",
9300 /* Make sure that the loop does not contain a function call
9301 (the count register might be altered by the called function). */
9302 if (loop_info->has_call)
9304 if (loop_dump_stream)
9305 fprintf (loop_dump_stream,
9306 "insert_bct %d: BCT instrumentation failed: function call in loop\n",
9311 /* Make sure that the loop does not jump via a table.
9312 (the count register might be used to perform the branch on table). */
9313 if (loop_info->has_tablejump)
9315 if (loop_dump_stream)
9316 fprintf (loop_dump_stream,
9317 "insert_bct %d: BCT instrumentation failed: computed branch in the loop\n",
9322 /* Account for loop unrolling in instrumented iteration count. */
9323 if (loop_info->unroll_number > 1)
9324 n_iterations = loop_info->n_iterations / loop_info->unroll_number;
9326 n_iterations = loop_info->n_iterations;
9328 if (n_iterations != 0 && n_iterations < 3)
9330 /* Allow an enclosing outer loop to benefit if possible. */
9331 if (loop_dump_stream)
9332 fprintf (loop_dump_stream,
9333 "insert_bct %d: Too few iterations to benefit from BCT optimization\n",
9338 /* Try to instrument the loop. */
9340 /* Handle the simpler case, where the bounds are known at compile time. */
9341 if (n_iterations > 0)
9343 /* Mark all enclosing loops that they cannot use count register. */
9344 for (i = loop_num; i != -1; i = loop_outer_loop[i])
9345 loop_used_count_register[i] = 1;
9346 instrument_loop_bct (loop_start, loop_end, GEN_INT (n_iterations));
9350 /* Handle the more complex case, that the bounds are NOT known
9351 at compile time. In this case we generate run_time calculation
9352 of the number of iterations. */
9354 if (loop_info->iteration_var == 0)
9356 if (loop_dump_stream)
9357 fprintf (loop_dump_stream,
9358 "insert_bct %d: BCT Runtime Instrumentation failed: no loop iteration variable found\n",
9363 if (GET_MODE_CLASS (GET_MODE (loop_info->iteration_var)) != MODE_INT
9364 || GET_MODE_SIZE (GET_MODE (loop_info->iteration_var)) != UNITS_PER_WORD)
9366 if (loop_dump_stream)
9367 fprintf (loop_dump_stream,
9368 "insert_bct %d: BCT Runtime Instrumentation failed: loop variable not integer\n",
9373 /* With runtime bounds, if the compare is of the form '!=' we give up */
9374 if (loop_info->comparison_code == NE)
9376 if (loop_dump_stream)
9377 fprintf (loop_dump_stream,
9378 "insert_bct %d: BCT Runtime Instrumentation failed: runtime bounds with != comparison\n",
9382 /* Use common loop preconditioning code instead. */
9386 /* We rely on the existence of run-time guard to ensure that the
9387 loop executes at least once. */
9389 rtx iterations_num_reg;
9391 unsigned HOST_WIDE_INT increment_value_abs
9392 = INTVAL (increment) * increment_direction;
9394 /* make sure that the increment is a power of two, otherwise (an
9395 expensive) divide is needed. */
9396 if (exact_log2 (increment_value_abs) == -1)
9398 if (loop_dump_stream)
9399 fprintf (loop_dump_stream,
9400 "insert_bct: not instrumenting BCT because the increment is not power of 2\n");
9404 /* compute the number of iterations */
9409 /* Again, the number of iterations is calculated by:
9411 ; compare-val - initial-val + (increment -1) + additional-iteration
9412 ; num_iterations = -----------------------------------------------------------------
9415 /* ??? Do we have to call copy_rtx here before passing rtx to
9417 if (compare_direction > 0)
9419 /* <, <= :the loop variable is increasing */
9420 temp_reg = expand_binop (loop_var_mode, sub_optab,
9421 comparison_value, initial_value,
9422 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9426 temp_reg = expand_binop (loop_var_mode, sub_optab,
9427 initial_value, comparison_value,
9428 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9431 if (increment_value_abs - 1 + add_iteration != 0)
9432 temp_reg = expand_binop (loop_var_mode, add_optab, temp_reg,
9433 GEN_INT (increment_value_abs - 1
9435 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9437 if (increment_value_abs != 1)
9438 iterations_num_reg = expand_binop (loop_var_mode, asr_optab,
9440 GEN_INT (exact_log2 (increment_value_abs)),
9441 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9443 iterations_num_reg = temp_reg;
9445 sequence = gen_sequence ();
9447 emit_insn_before (sequence, loop_start);
9448 instrument_loop_bct (loop_start, loop_end, iterations_num_reg);
9452 #endif /* Complex case */
9455 /* Instrument loop by inserting a bct in it as follows:
9456 1. A new counter register is created.
9457 2. In the head of the loop the new variable is initialized to the value
9458 passed in the loop_num_iterations parameter.
9459 3. At the end of the loop, comparison of the register with 0 is generated.
9460 The created comparison follows the pattern defined for the
9461 decrement_and_branch_on_count insn, so this insn will be generated.
9462 4. The branch on the old variable are deleted. The compare must remain
9463 because it might be used elsewhere. If the loop-variable or condition
9464 register are used elsewhere, they will be eliminated by flow. */
9467 instrument_loop_bct (loop_start, loop_end, loop_num_iterations)
9468 rtx loop_start, loop_end;
9469 rtx loop_num_iterations;
9475 if (HAVE_decrement_and_branch_on_count)
9477 if (loop_dump_stream)
9479 fputs ("instrument_bct: Inserting BCT (", loop_dump_stream);
9480 if (GET_CODE (loop_num_iterations) == CONST_INT)
9481 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
9482 INTVAL (loop_num_iterations));
9484 fputs ("runtime", loop_dump_stream);
9485 fputs (" iterations)", loop_dump_stream);
9488 /* Discard original jump to continue loop. Original compare result
9489 may still be live, so it cannot be discarded explicitly. */
9490 delete_insn (PREV_INSN (loop_end));
9492 /* Insert the label which will delimit the start of the loop. */
9493 start_label = gen_label_rtx ();
9494 emit_label_after (start_label, loop_start);
9496 /* Insert initialization of the count register into the loop header. */
9498 counter_reg = gen_reg_rtx (word_mode);
9499 emit_insn (gen_move_insn (counter_reg, loop_num_iterations));
9500 sequence = gen_sequence ();
9502 emit_insn_before (sequence, loop_start);
9504 /* Insert new comparison on the count register instead of the
9505 old one, generating the needed BCT pattern (that will be
9506 later recognized by assembly generation phase). */
9507 emit_jump_insn_before (gen_decrement_and_branch_on_count (counter_reg,
9510 LABEL_NUSES (start_label)++;
9514 #endif /* HAVE_decrement_and_branch_on_count */
9516 /* Scan the function and determine whether it has indirect (computed) jumps.
9518 This is taken mostly from flow.c; similar code exists elsewhere
9519 in the compiler. It may be useful to put this into rtlanal.c. */
9521 indirect_jump_in_function_p (start)
9526 for (insn = start; insn; insn = NEXT_INSN (insn))
9527 if (computed_jump_p (insn))
9533 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
9534 documentation for LOOP_MEMS for the definition of `appropriate'.
9535 This function is called from prescan_loop via for_each_rtx. */
9538 insert_loop_mem (mem, data)
9540 void *data ATTRIBUTE_UNUSED;
9548 switch (GET_CODE (m))
9554 /* We're not interested in the MEM associated with a
9555 CONST_DOUBLE, so there's no need to traverse into this. */
9559 /* This is not a MEM. */
9563 /* See if we've already seen this MEM. */
9564 for (i = 0; i < loop_mems_idx; ++i)
9565 if (rtx_equal_p (m, loop_mems[i].mem))
9567 if (GET_MODE (m) != GET_MODE (loop_mems[i].mem))
9568 /* The modes of the two memory accesses are different. If
9569 this happens, something tricky is going on, and we just
9570 don't optimize accesses to this MEM. */
9571 loop_mems[i].optimize = 0;
9576 /* Resize the array, if necessary. */
9577 if (loop_mems_idx == loop_mems_allocated)
9579 if (loop_mems_allocated != 0)
9580 loop_mems_allocated *= 2;
9582 loop_mems_allocated = 32;
9584 loop_mems = (loop_mem_info*)
9585 xrealloc (loop_mems,
9586 loop_mems_allocated * sizeof (loop_mem_info));
9589 /* Actually insert the MEM. */
9590 loop_mems[loop_mems_idx].mem = m;
9591 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
9592 because we can't put it in a register. We still store it in the
9593 table, though, so that if we see the same address later, but in a
9594 non-BLK mode, we'll not think we can optimize it at that point. */
9595 loop_mems[loop_mems_idx].optimize = (GET_MODE (m) != BLKmode);
9596 loop_mems[loop_mems_idx].reg = NULL_RTX;
9602 /* Like load_mems, but also ensures that SET_IN_LOOP,
9603 MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct
9604 values after load_mems. */
9607 load_mems_and_recount_loop_regs_set (scan_start, end, loop_top, start,
9615 int nregs = max_reg_num ();
9617 load_mems (scan_start, end, loop_top, start);
9619 /* Recalculate set_in_loop and friends since load_mems may have
9620 created new registers. */
9621 if (max_reg_num () > nregs)
9627 nregs = max_reg_num ();
9629 if ((unsigned) nregs > set_in_loop->num_elements)
9631 /* Grow all the arrays. */
9632 VARRAY_GROW (set_in_loop, nregs);
9633 VARRAY_GROW (n_times_set, nregs);
9634 VARRAY_GROW (may_not_optimize, nregs);
9635 VARRAY_GROW (reg_single_usage, nregs);
9637 /* Clear the arrays */
9638 bzero ((char *) &set_in_loop->data, nregs * sizeof (int));
9639 bzero ((char *) &may_not_optimize->data, nregs * sizeof (char));
9640 bzero ((char *) ®_single_usage->data, nregs * sizeof (rtx));
9642 count_loop_regs_set (loop_top ? loop_top : start, end,
9643 may_not_optimize, reg_single_usage,
9646 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
9648 VARRAY_CHAR (may_not_optimize, i) = 1;
9649 VARRAY_INT (set_in_loop, i) = 1;
9652 #ifdef AVOID_CCMODE_COPIES
9653 /* Don't try to move insns which set CC registers if we should not
9654 create CCmode register copies. */
9655 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
9656 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
9657 VARRAY_CHAR (may_not_optimize, i) = 1;
9660 /* Set n_times_set for the new registers. */
9661 bcopy ((char *) (&set_in_loop->data.i[0] + old_nregs),
9662 (char *) (&n_times_set->data.i[0] + old_nregs),
9663 (nregs - old_nregs) * sizeof (int));
9667 /* Move MEMs into registers for the duration of the loop. SCAN_START
9668 is the first instruction in the loop (as it is executed). The
9669 other parameters are as for next_insn_in_loop. */
9672 load_mems (scan_start, end, loop_top, start)
9678 int maybe_never = 0;
9681 rtx label = NULL_RTX;
9682 rtx end_label = NULL_RTX;
9684 if (loop_mems_idx > 0)
9686 /* Nonzero if the next instruction may never be executed. */
9687 int next_maybe_never = 0;
9689 /* Check to see if it's possible that some instructions in the
9690 loop are never executed. */
9691 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
9692 p != NULL_RTX && !maybe_never;
9693 p = next_insn_in_loop (p, scan_start, end, loop_top))
9695 if (GET_CODE (p) == CODE_LABEL)
9697 else if (GET_CODE (p) == JUMP_INSN
9698 /* If we enter the loop in the middle, and scan
9699 around to the beginning, don't set maybe_never
9700 for that. This must be an unconditional jump,
9701 otherwise the code at the top of the loop might
9702 never be executed. Unconditional jumps are
9703 followed a by barrier then loop end. */
9704 && ! (GET_CODE (p) == JUMP_INSN
9705 && JUMP_LABEL (p) == loop_top
9706 && NEXT_INSN (NEXT_INSN (p)) == end
9707 && simplejump_p (p)))
9709 if (!condjump_p (p))
9710 /* Something complicated. */
9713 /* If there are any more instructions in the loop, they
9714 might not be reached. */
9715 next_maybe_never = 1;
9717 else if (next_maybe_never)
9721 /* Actually move the MEMs. */
9722 for (i = 0; i < loop_mems_idx; ++i)
9726 rtx mem = loop_mems[i].mem;
9729 if (MEM_VOLATILE_P (mem)
9730 || invariant_p (XEXP (mem, 0)) != 1)
9731 /* There's no telling whether or not MEM is modified. */
9732 loop_mems[i].optimize = 0;
9734 /* Go through the MEMs written to in the loop to see if this
9735 one is aliased by one of them. */
9736 mem_list_entry = loop_store_mems;
9737 while (mem_list_entry)
9739 if (rtx_equal_p (mem, XEXP (mem_list_entry, 0)))
9741 else if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
9744 /* MEM is indeed aliased by this store. */
9745 loop_mems[i].optimize = 0;
9748 mem_list_entry = XEXP (mem_list_entry, 1);
9751 /* If this MEM is written to, we must be sure that there
9752 are no reads from another MEM that aliases this one. */
9753 if (loop_mems[i].optimize && written)
9757 for (j = 0; j < loop_mems_idx; ++j)
9761 else if (true_dependence (mem,
9766 /* It's not safe to hoist loop_mems[i] out of
9767 the loop because writes to it might not be
9768 seen by reads from loop_mems[j]. */
9769 loop_mems[i].optimize = 0;
9775 if (maybe_never && may_trap_p (mem))
9776 /* We can't access the MEM outside the loop; it might
9777 cause a trap that wouldn't have happened otherwise. */
9778 loop_mems[i].optimize = 0;
9780 if (!loop_mems[i].optimize)
9781 /* We thought we were going to lift this MEM out of the
9782 loop, but later discovered that we could not. */
9785 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
9786 order to keep scan_loop from moving stores to this MEM
9787 out of the loop just because this REG is neither a
9788 user-variable nor used in the loop test. */
9789 reg = gen_reg_rtx (GET_MODE (mem));
9790 REG_USERVAR_P (reg) = 1;
9791 loop_mems[i].reg = reg;
9793 /* Now, replace all references to the MEM with the
9794 corresponding pesudos. */
9795 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
9797 p = next_insn_in_loop (p, scan_start, end, loop_top))
9802 for_each_rtx (&p, replace_loop_mem, &ri);
9805 if (!apply_change_group ())
9806 /* We couldn't replace all occurrences of the MEM. */
9807 loop_mems[i].optimize = 0;
9812 /* Load the memory immediately before START, which is
9813 the NOTE_LOOP_BEG. */
9814 set = gen_move_insn (reg, mem);
9815 emit_insn_before (set, start);
9819 if (label == NULL_RTX)
9821 /* We must compute the former
9822 right-after-the-end label before we insert
9824 end_label = next_label (end);
9825 label = gen_label_rtx ();
9826 emit_label_after (label, end);
9829 /* Store the memory immediately after END, which is
9830 the NOTE_LOOP_END. */
9831 set = gen_move_insn (copy_rtx (mem), reg);
9832 emit_insn_after (set, label);
9835 if (loop_dump_stream)
9837 fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
9838 REGNO (reg), (written ? "r/w" : "r/o"));
9839 print_rtl (loop_dump_stream, mem);
9840 fputc ('\n', loop_dump_stream);
9846 if (label != NULL_RTX)
9848 /* Now, we need to replace all references to the previous exit
9849 label with the new one. */
9854 for (p = start; p != end; p = NEXT_INSN (p))
9856 for_each_rtx (&p, replace_label, &rr);
9858 /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
9859 field. This is not handled by for_each_rtx because it doesn't
9860 handle unprinted ('0') fields. We need to update JUMP_LABEL
9861 because the immediately following unroll pass will use it.
9862 replace_label would not work anyways, because that only handles
9864 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == end_label)
9865 JUMP_LABEL (p) = label;
9870 /* Replace MEM with its associated pseudo register. This function is
9871 called from load_mems via for_each_rtx. DATA is actually an
9872 rtx_and_int * describing the instruction currently being scanned
9873 and the MEM we are currently replacing. */
9876 replace_loop_mem (mem, data)
9888 switch (GET_CODE (m))
9894 /* We're not interested in the MEM associated with a
9895 CONST_DOUBLE, so there's no need to traverse into one. */
9899 /* This is not a MEM. */
9903 ri = (rtx_and_int*) data;
9906 if (!rtx_equal_p (loop_mems[i].mem, m))
9907 /* This is not the MEM we are currently replacing. */
9912 /* Actually replace the MEM. */
9913 validate_change (insn, mem, loop_mems[i].reg, 1);
9918 /* Replace occurrences of the old exit label for the loop with the new
9919 one. DATA is an rtx_pair containing the old and new labels,
9923 replace_label (x, data)
9928 rtx old_label = ((rtx_pair*) data)->r1;
9929 rtx new_label = ((rtx_pair*) data)->r2;
9934 if (GET_CODE (l) != LABEL_REF)
9937 if (XEXP (l, 0) != old_label)
9940 XEXP (l, 0) = new_label;
9941 ++LABEL_NUSES (new_label);
9942 --LABEL_NUSES (old_label);