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. */
42 #include "insn-config.h"
43 #include "insn-flags.h"
45 #include "hard-reg-set.h"
53 /* Vector mapping INSN_UIDs to luids.
54 The luids are like uids but increase monotonically always.
55 We use them to see whether a jump comes from outside a given loop. */
59 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
60 number the insn is contained in. */
64 /* 1 + largest uid of any insn. */
68 /* 1 + luid of last insn. */
72 /* Number of loops detected in current function. Used as index to the
75 static int max_loop_num;
77 /* Indexed by loop number, contains the first and last insn of each loop. */
79 static rtx *loop_number_loop_starts, *loop_number_loop_ends;
81 /* Likewise for the continue insn */
82 static rtx *loop_number_loop_cont;
84 /* The first code_label that is reached in every loop iteration.
85 0 when not computed yet, initially const0_rtx if a jump couldn't be
87 Also set to 0 when there is no such label before the NOTE_INSN_LOOP_CONT
88 of this loop, or in verify_dominator, if a jump couldn't be followed. */
89 static rtx *loop_number_cont_dominator;
91 /* For each loop, gives the containing loop number, -1 if none. */
95 #ifdef HAVE_decrement_and_branch_on_count
96 /* Records whether resource in use by inner loop. */
98 int *loop_used_count_register;
99 #endif /* HAVE_decrement_and_branch_on_count */
101 /* Indexed by loop number, contains a nonzero value if the "loop" isn't
102 really a loop (an insn outside the loop branches into it). */
104 static char *loop_invalid;
106 /* Indexed by loop number, links together all LABEL_REFs which refer to
107 code labels outside the loop. Used by routines that need to know all
108 loop exits, such as final_biv_value and final_giv_value.
110 This does not include loop exits due to return instructions. This is
111 because all bivs and givs are pseudos, and hence must be dead after a
112 return, so the presense of a return does not affect any of the
113 optimizations that use this info. It is simpler to just not include return
114 instructions on this list. */
116 rtx *loop_number_exit_labels;
118 /* Indexed by loop number, counts the number of LABEL_REFs on
119 loop_number_exit_labels for this loop and all loops nested inside it. */
121 int *loop_number_exit_count;
123 /* Nonzero if there is a subroutine call in the current loop. */
125 static int loop_has_call;
127 /* Nonzero if there is a volatile memory reference in the current
130 static int loop_has_volatile;
132 /* Nonzero if there is a tablejump in the current loop. */
134 static int loop_has_tablejump;
136 /* Added loop_continue which is the NOTE_INSN_LOOP_CONT of the
137 current loop. A continue statement will generate a branch to
138 NEXT_INSN (loop_continue). */
140 static rtx loop_continue;
142 /* Indexed by register number, contains the number of times the reg
143 is set during the loop being scanned.
144 During code motion, a negative value indicates a reg that has been
145 made a candidate; in particular -2 means that it is an candidate that
146 we know is equal to a constant and -1 means that it is an candidate
147 not known equal to a constant.
148 After code motion, regs moved have 0 (which is accurate now)
149 while the failed candidates have the original number of times set.
151 Therefore, at all times, == 0 indicates an invariant register;
152 < 0 a conditionally invariant one. */
154 static varray_type set_in_loop;
156 /* Original value of set_in_loop; same except that this value
157 is not set negative for a reg whose sets have been made candidates
158 and not set to 0 for a reg that is moved. */
160 static varray_type n_times_set;
162 /* Index by register number, 1 indicates that the register
163 cannot be moved or strength reduced. */
165 static varray_type may_not_optimize;
167 /* Nonzero means reg N has already been moved out of one loop.
168 This reduces the desire to move it out of another. */
170 static char *moved_once;
172 /* List of MEMs that are stored in this loop. */
174 static rtx loop_store_mems;
176 typedef struct loop_mem_info {
177 rtx mem; /* The MEM itself. */
178 rtx reg; /* Corresponding pseudo, if any. */
179 int optimize; /* Nonzero if we can optimize access to this MEM. */
182 /* Array of MEMs that are used (read or written) in this loop, but
183 cannot be aliased by anything in this loop, except perhaps
184 themselves. In other words, if loop_mems[i] is altered during the
185 loop, it is altered by an expression that is rtx_equal_p to it. */
187 static loop_mem_info *loop_mems;
189 /* The index of the next available slot in LOOP_MEMS. */
191 static int loop_mems_idx;
193 /* The number of elements allocated in LOOP_MEMs. */
195 static int loop_mems_allocated;
197 /* Nonzero if we don't know what MEMs were changed in the current loop.
198 This happens if the loop contains a call (in which case `loop_has_call'
199 will also be set) or if we store into more than NUM_STORES MEMs. */
201 static int unknown_address_altered;
203 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
204 static int num_movables;
206 /* Count of memory write instructions discovered in the loop. */
207 static int num_mem_sets;
209 /* Number of loops contained within the current one, including itself. */
210 static int loops_enclosed;
212 /* Bound on pseudo register number before loop optimization.
213 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
214 int max_reg_before_loop;
216 /* This obstack is used in product_cheap_p to allocate its rtl. It
217 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
218 If we used the same obstack that it did, we would be deallocating
221 static struct obstack temp_obstack;
223 /* This is where the pointer to the obstack being used for RTL is stored. */
225 extern struct obstack *rtl_obstack;
227 #define obstack_chunk_alloc xmalloc
228 #define obstack_chunk_free free
230 /* During the analysis of a loop, a chain of `struct movable's
231 is made to record all the movable insns found.
232 Then the entire chain can be scanned to decide which to move. */
236 rtx insn; /* A movable insn */
237 rtx set_src; /* The expression this reg is set from. */
238 rtx set_dest; /* The destination of this SET. */
239 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
240 of any registers used within the LIBCALL. */
241 int consec; /* Number of consecutive following insns
242 that must be moved with this one. */
243 int regno; /* The register it sets */
244 short lifetime; /* lifetime of that register;
245 may be adjusted when matching movables
246 that load the same value are found. */
247 short savings; /* Number of insns we can move for this reg,
248 including other movables that force this
249 or match this one. */
250 unsigned int cond : 1; /* 1 if only conditionally movable */
251 unsigned int force : 1; /* 1 means MUST move this insn */
252 unsigned int global : 1; /* 1 means reg is live outside this loop */
253 /* If PARTIAL is 1, GLOBAL means something different:
254 that the reg is live outside the range from where it is set
255 to the following label. */
256 unsigned int done : 1; /* 1 inhibits further processing of this */
258 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
259 In particular, moving it does not make it
261 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
262 load SRC, rather than copying INSN. */
263 unsigned int move_insn_first:1;/* Same as above, if this is necessary for the
264 first insn of a consecutive sets group. */
265 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
266 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
267 that we should avoid changing when clearing
268 the rest of the reg. */
269 struct movable *match; /* First entry for same value */
270 struct movable *forces; /* An insn that must be moved if this is */
271 struct movable *next;
274 static struct movable *the_movables;
276 FILE *loop_dump_stream;
278 /* Forward declarations. */
280 static void verify_dominator PROTO((int));
281 static void find_and_verify_loops PROTO((rtx));
282 static void mark_loop_jump PROTO((rtx, int));
283 static void prescan_loop PROTO((rtx, rtx));
284 static int reg_in_basic_block_p PROTO((rtx, rtx));
285 static int consec_sets_invariant_p PROTO((rtx, int, rtx));
286 static rtx libcall_other_reg PROTO((rtx, rtx));
287 static int labels_in_range_p PROTO((rtx, int));
288 static void count_one_set PROTO((rtx, rtx, varray_type, rtx *));
290 static void count_loop_regs_set PROTO((rtx, rtx, varray_type, varray_type,
292 static void note_addr_stored PROTO((rtx, rtx));
293 static int loop_reg_used_before_p PROTO((rtx, rtx, rtx, rtx, rtx));
294 static void scan_loop PROTO((rtx, rtx, int, int));
296 static void replace_call_address PROTO((rtx, rtx, rtx));
298 static rtx skip_consec_insns PROTO((rtx, int));
299 static int libcall_benefit PROTO((rtx));
300 static void ignore_some_movables PROTO((struct movable *));
301 static void force_movables PROTO((struct movable *));
302 static void combine_movables PROTO((struct movable *, int));
303 static int regs_match_p PROTO((rtx, rtx, struct movable *));
304 static int rtx_equal_for_loop_p PROTO((rtx, rtx, struct movable *));
305 static void add_label_notes PROTO((rtx, rtx));
306 static void move_movables PROTO((struct movable *, int, int, rtx, rtx, int));
307 static int count_nonfixed_reads PROTO((rtx));
308 static void strength_reduce PROTO((rtx, rtx, rtx, int, rtx, rtx, int, int));
309 static void find_single_use_in_loop PROTO((rtx, rtx, varray_type));
310 static int valid_initial_value_p PROTO((rtx, rtx, int, rtx));
311 static void find_mem_givs PROTO((rtx, rtx, int, rtx, rtx));
312 static void record_biv PROTO((struct induction *, rtx, rtx, rtx, rtx, rtx *, int, int));
313 static void check_final_value PROTO((struct induction *, rtx, rtx,
314 unsigned HOST_WIDE_INT));
315 static void record_giv PROTO((struct induction *, rtx, rtx, rtx, rtx, rtx, int, enum g_types, int, rtx *, rtx, rtx));
316 static void update_giv_derive PROTO((rtx));
317 static int basic_induction_var PROTO((rtx, enum machine_mode, rtx, rtx, rtx *, rtx *, rtx **));
318 static rtx simplify_giv_expr PROTO((rtx, int *));
319 static int general_induction_var PROTO((rtx, rtx *, rtx *, rtx *, int, int *));
320 static int consec_sets_giv PROTO((int, rtx, rtx, rtx, rtx *, rtx *, rtx *));
321 static int check_dbra_loop PROTO((rtx, int, rtx, struct loop_info *));
322 static rtx express_from_1 PROTO((rtx, rtx, rtx));
323 static rtx express_from PROTO((struct induction *, struct induction *));
324 static rtx combine_givs_p PROTO((struct induction *, struct induction *));
325 static void combine_givs PROTO((struct iv_class *));
326 struct recombine_givs_stats;
327 static int find_life_end PROTO((rtx, struct recombine_givs_stats *, rtx, rtx));
328 static void recombine_givs PROTO((struct iv_class *, rtx, rtx, int));
329 static int product_cheap_p PROTO((rtx, rtx));
330 static int maybe_eliminate_biv PROTO((struct iv_class *, rtx, rtx, int, int, int));
331 static int maybe_eliminate_biv_1 PROTO((rtx, rtx, struct iv_class *, int, rtx));
332 static int last_use_this_basic_block PROTO((rtx, rtx));
333 static void record_initial PROTO((rtx, rtx));
334 static void update_reg_last_use PROTO((rtx, rtx));
335 static rtx next_insn_in_loop PROTO((rtx, rtx, rtx, rtx));
336 static void load_mems_and_recount_loop_regs_set PROTO((rtx, rtx, rtx,
339 static void load_mems PROTO((rtx, rtx, rtx, rtx));
340 static int insert_loop_mem PROTO((rtx *, void *));
341 static int replace_loop_mem PROTO((rtx *, void *));
342 static int replace_label PROTO((rtx *, void *));
344 typedef struct rtx_and_int {
349 typedef struct rtx_pair {
354 /* Nonzero iff INSN is between START and END, inclusive. */
355 #define INSN_IN_RANGE_P(INSN, START, END) \
356 (INSN_UID (INSN) < max_uid_for_loop \
357 && INSN_LUID (INSN) >= INSN_LUID (START) \
358 && INSN_LUID (INSN) <= INSN_LUID (END))
360 #ifdef HAVE_decrement_and_branch_on_count
361 /* Test whether BCT applicable and safe. */
362 static void insert_bct PROTO((rtx, rtx, struct loop_info *));
364 /* Auxiliary function that inserts the BCT pattern into the loop. */
365 static void instrument_loop_bct PROTO((rtx, rtx, rtx));
366 #endif /* HAVE_decrement_and_branch_on_count */
368 /* Indirect_jump_in_function is computed once per function. */
369 int indirect_jump_in_function = 0;
370 static int indirect_jump_in_function_p PROTO((rtx));
372 static int compute_luids PROTO ((rtx, rtx, int));
374 /* Relative gain of eliminating various kinds of operations. */
377 static int shift_cost;
378 static int mult_cost;
381 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
382 copy the value of the strength reduced giv to its original register. */
383 static int copy_cost;
385 /* Cost of using a register, to normalize the benefits of a giv. */
386 static int reg_address_cost;
392 char *free_point = (char *) oballoc (1);
393 rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
395 add_cost = rtx_cost (gen_rtx_PLUS (word_mode, reg, reg), SET);
398 reg_address_cost = ADDRESS_COST (reg);
400 reg_address_cost = rtx_cost (reg, MEM);
403 /* We multiply by 2 to reconcile the difference in scale between
404 these two ways of computing costs. Otherwise the cost of a copy
405 will be far less than the cost of an add. */
409 /* Free the objects we just allocated. */
412 /* Initialize the obstack used for rtl in product_cheap_p. */
413 gcc_obstack_init (&temp_obstack);
416 /* Compute the mapping from uids to luids.
417 LUIDs are numbers assigned to insns, like uids,
418 except that luids increase monotonically through the code.
419 Start at insn START and stop just before END. Assign LUIDs
420 starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
422 compute_luids (start, end, prev_luid)
429 for (insn = start, i = prev_luid; insn != end; insn = NEXT_INSN (insn))
431 if (INSN_UID (insn) >= max_uid_for_loop)
433 /* Don't assign luids to line-number NOTEs, so that the distance in
434 luids between two insns is not affected by -g. */
435 if (GET_CODE (insn) != NOTE
436 || NOTE_LINE_NUMBER (insn) <= 0)
437 uid_luid[INSN_UID (insn)] = ++i;
439 /* Give a line number note the same luid as preceding insn. */
440 uid_luid[INSN_UID (insn)] = i;
445 /* Entry point of this file. Perform loop optimization
446 on the current function. F is the first insn of the function
447 and DUMPFILE is a stream for output of a trace of actions taken
448 (or 0 if none should be output). */
451 loop_optimize (f, dumpfile, unroll_p, bct_p)
452 /* f is the first instruction of a chain of insns for one function */
460 loop_dump_stream = dumpfile;
462 init_recog_no_volatile ();
464 max_reg_before_loop = max_reg_num ();
466 moved_once = (char *) alloca (max_reg_before_loop);
467 bzero (moved_once, max_reg_before_loop);
471 /* Count the number of loops. */
474 for (insn = f; insn; insn = NEXT_INSN (insn))
476 if (GET_CODE (insn) == NOTE
477 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
481 /* Don't waste time if no loops. */
482 if (max_loop_num == 0)
485 /* Get size to use for tables indexed by uids.
486 Leave some space for labels allocated by find_and_verify_loops. */
487 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
489 uid_luid = (int *) alloca (max_uid_for_loop * sizeof (int));
490 uid_loop_num = (int *) alloca (max_uid_for_loop * sizeof (int));
492 bzero ((char *) uid_luid, max_uid_for_loop * sizeof (int));
493 bzero ((char *) uid_loop_num, max_uid_for_loop * sizeof (int));
495 /* Allocate tables for recording each loop. We set each entry, so they need
497 loop_number_loop_starts = (rtx *) alloca (max_loop_num * sizeof (rtx));
498 loop_number_loop_ends = (rtx *) alloca (max_loop_num * sizeof (rtx));
499 loop_number_loop_cont = (rtx *) alloca (max_loop_num * sizeof (rtx));
500 loop_number_cont_dominator = (rtx *) alloca (max_loop_num * sizeof (rtx));
501 loop_outer_loop = (int *) alloca (max_loop_num * sizeof (int));
502 loop_invalid = (char *) alloca (max_loop_num * sizeof (char));
503 loop_number_exit_labels = (rtx *) alloca (max_loop_num * sizeof (rtx));
504 loop_number_exit_count = (int *) alloca (max_loop_num * sizeof (int));
506 #ifdef HAVE_decrement_and_branch_on_count
507 /* Allocate for BCT optimization */
508 loop_used_count_register = (int *) alloca (max_loop_num * sizeof (int));
509 bzero ((char *) loop_used_count_register, max_loop_num * sizeof (int));
510 #endif /* HAVE_decrement_and_branch_on_count */
512 /* Find and process each loop.
513 First, find them, and record them in order of their beginnings. */
514 find_and_verify_loops (f);
516 /* Now find all register lifetimes. This must be done after
517 find_and_verify_loops, because it might reorder the insns in the
519 reg_scan (f, max_reg_num (), 1);
521 /* This must occur after reg_scan so that registers created by gcse
522 will have entries in the register tables.
524 We could have added a call to reg_scan after gcse_main in toplev.c,
525 but moving this call to init_alias_analysis is more efficient. */
526 init_alias_analysis ();
528 /* See if we went too far. */
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 () + 1;
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],
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
611 NOTE_INSN_LOOP_END. */
613 /* ??? Could also move memory writes out of loops if the destination address
614 is invariant, the source is invariant, the memory write is not volatile,
615 and if we can prove that no read inside the loop can read this address
616 before the write occurs. If there is a read of this address after the
617 write, then we can also mark the memory read as invariant. */
620 scan_loop (loop_start, end, unroll_p, bct_p)
626 /* 1 if we are scanning insns that could be executed zero times. */
628 /* 1 if we are scanning insns that might never be executed
629 due to a subroutine call which might exit before they are reached. */
631 /* For a rotated loop that is entered near the bottom,
632 this is the label at the top. Otherwise it is zero. */
634 /* Jump insn that enters the loop, or 0 if control drops in. */
635 rtx loop_entry_jump = 0;
636 /* Place in the loop where control enters. */
638 /* Number of insns in the loop. */
643 /* The SET from an insn, if it is the only SET in the insn. */
645 /* Chain describing insns movable in current loop. */
646 struct movable *movables = 0;
647 /* Last element in `movables' -- so we can add elements at the end. */
648 struct movable *last_movable = 0;
649 /* Ratio of extra register life span we can justify
650 for saving an instruction. More if loop doesn't call subroutines
651 since in that case saving an insn makes more difference
652 and more registers are available. */
654 /* If we have calls, contains the insn in which a register was used
655 if it was used exactly once; contains const0_rtx if it was used more
657 varray_type reg_single_usage = 0;
658 /* Nonzero if we are scanning instructions in a sub-loop. */
662 /* Determine whether this loop starts with a jump down to a test at
663 the end. This will occur for a small number of loops with a test
664 that is too complex to duplicate in front of the loop.
666 We search for the first insn or label in the loop, skipping NOTEs.
667 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
668 (because we might have a loop executed only once that contains a
669 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
670 (in case we have a degenerate loop).
672 Note that if we mistakenly think that a loop is entered at the top
673 when, in fact, it is entered at the exit test, the only effect will be
674 slightly poorer optimization. Making the opposite error can generate
675 incorrect code. Since very few loops now start with a jump to the
676 exit test, the code here to detect that case is very conservative. */
678 for (p = NEXT_INSN (loop_start);
680 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
681 && (GET_CODE (p) != NOTE
682 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
683 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
689 /* Set up variables describing this loop. */
690 prescan_loop (loop_start, end);
691 threshold = (loop_has_call ? 1 : 2) * (1 + n_non_fixed_regs);
693 /* If loop has a jump before the first label,
694 the true entry is the target of that jump.
695 Start scan from there.
696 But record in LOOP_TOP the place where the end-test jumps
697 back to so we can scan that after the end of the loop. */
698 if (GET_CODE (p) == JUMP_INSN)
702 /* Loop entry must be unconditional jump (and not a RETURN) */
704 && JUMP_LABEL (p) != 0
705 /* Check to see whether the jump actually
706 jumps out of the loop (meaning it's no loop).
707 This case can happen for things like
708 do {..} while (0). If this label was generated previously
709 by loop, we can't tell anything about it and have to reject
711 && INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, end))
713 loop_top = next_label (scan_start);
714 scan_start = JUMP_LABEL (p);
718 /* If SCAN_START was an insn created by loop, we don't know its luid
719 as required by loop_reg_used_before_p. So skip such loops. (This
720 test may never be true, but it's best to play it safe.)
722 Also, skip loops where we do not start scanning at a label. This
723 test also rejects loops starting with a JUMP_INSN that failed the
726 if (INSN_UID (scan_start) >= max_uid_for_loop
727 || GET_CODE (scan_start) != CODE_LABEL)
729 if (loop_dump_stream)
730 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
731 INSN_UID (loop_start), INSN_UID (end));
735 /* Count number of times each reg is set during this loop.
736 Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out
737 the setting of register I. If this loop has calls, set
738 VARRAY_RTX (reg_single_usage, I). */
740 /* Allocate extra space for REGS that might be created by
741 load_mems. We allocate a little extra slop as well, in the hopes
742 that even after the moving of movables creates some new registers
743 we won't have to reallocate these arrays. However, we do grow
744 the arrays, if necessary, in load_mems_recount_loop_regs_set. */
745 nregs = max_reg_num () + loop_mems_idx + 16;
746 VARRAY_INT_INIT (set_in_loop, nregs, "set_in_loop");
747 VARRAY_INT_INIT (n_times_set, nregs, "n_times_set");
748 VARRAY_CHAR_INIT (may_not_optimize, nregs, "may_not_optimize");
751 VARRAY_RTX_INIT (reg_single_usage, nregs, "reg_single_usage");
753 count_loop_regs_set (loop_top ? loop_top : loop_start, end,
754 may_not_optimize, reg_single_usage, &insn_count, nregs);
756 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
758 VARRAY_CHAR (may_not_optimize, i) = 1;
759 VARRAY_INT (set_in_loop, i) = 1;
762 #ifdef AVOID_CCMODE_COPIES
763 /* Don't try to move insns which set CC registers if we should not
764 create CCmode register copies. */
765 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
766 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
767 VARRAY_CHAR (may_not_optimize, i) = 1;
770 bcopy ((char *) &set_in_loop->data,
771 (char *) &n_times_set->data, nregs * sizeof (int));
773 if (loop_dump_stream)
775 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
776 INSN_UID (loop_start), INSN_UID (end), insn_count);
778 fprintf (loop_dump_stream, "Continue at insn %d.\n",
779 INSN_UID (loop_continue));
782 /* Scan through the loop finding insns that are safe to move.
783 Set set_in_loop negative for the reg being set, so that
784 this reg will be considered invariant for subsequent insns.
785 We consider whether subsequent insns use the reg
786 in deciding whether it is worth actually moving.
788 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
789 and therefore it is possible that the insns we are scanning
790 would never be executed. At such times, we must make sure
791 that it is safe to execute the insn once instead of zero times.
792 When MAYBE_NEVER is 0, all insns will be executed at least once
793 so that is not a problem. */
795 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
797 p = next_insn_in_loop (p, scan_start, end, loop_top))
799 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
800 && find_reg_note (p, REG_LIBCALL, NULL_RTX))
802 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
803 && find_reg_note (p, REG_RETVAL, NULL_RTX))
806 if (GET_CODE (p) == INSN
807 && (set = single_set (p))
808 && GET_CODE (SET_DEST (set)) == REG
809 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
814 rtx src = SET_SRC (set);
815 rtx dependencies = 0;
817 /* Figure out what to use as a source of this insn. If a REG_EQUIV
818 note is given or if a REG_EQUAL note with a constant operand is
819 specified, use it as the source and mark that we should move
820 this insn by calling emit_move_insn rather that duplicating the
823 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
825 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
827 src = XEXP (temp, 0), move_insn = 1;
830 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
831 if (temp && CONSTANT_P (XEXP (temp, 0)))
832 src = XEXP (temp, 0), move_insn = 1;
833 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
835 src = XEXP (temp, 0);
836 /* A libcall block can use regs that don't appear in
837 the equivalent expression. To move the libcall,
838 we must move those regs too. */
839 dependencies = libcall_other_reg (p, src);
843 /* Don't try to optimize a register that was made
844 by loop-optimization for an inner loop.
845 We don't know its life-span, so we can't compute the benefit. */
846 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
848 else if (/* 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 || loop_reg_used_before_p (set, p, loop_start,
854 /* And the register is used in basic blocks other
855 than the one where it is set (meaning that
856 something after this point in the loop might
857 depend on its value before the set). */
858 && !reg_in_basic_block_p (p, SET_DEST (set)))
859 /* It is unsafe to move the set.
861 This code used to consider it OK to move a set of a variable
862 which was not created by the user and not used in an exit test.
863 That behavior is incorrect and was removed. */
865 else if ((tem = invariant_p (src))
866 && (dependencies == 0
867 || (tem2 = invariant_p (dependencies)) != 0)
868 && (VARRAY_INT (set_in_loop,
869 REGNO (SET_DEST (set))) == 1
871 = consec_sets_invariant_p
873 VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))),
875 /* If the insn can cause a trap (such as divide by zero),
876 can't move it unless it's guaranteed to be executed
877 once loop is entered. Even a function call might
878 prevent the trap insn from being reached
879 (since it might exit!) */
880 && ! ((maybe_never || call_passed)
881 && may_trap_p (src)))
883 register struct movable *m;
884 register int regno = REGNO (SET_DEST (set));
886 /* A potential lossage is where we have a case where two insns
887 can be combined as long as they are both in the loop, but
888 we move one of them outside the loop. For large loops,
889 this can lose. The most common case of this is the address
890 of a function being called.
892 Therefore, if this register is marked as being used exactly
893 once if we are in a loop with calls (a "large loop"), see if
894 we can replace the usage of this register with the source
895 of this SET. If we can, delete this insn.
897 Don't do this if P has a REG_RETVAL note or if we have
898 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
900 if (reg_single_usage && VARRAY_RTX (reg_single_usage, regno) != 0
901 && VARRAY_RTX (reg_single_usage, regno) != const0_rtx
902 && REGNO_FIRST_UID (regno) == INSN_UID (p)
903 && (REGNO_LAST_UID (regno)
904 == INSN_UID (VARRAY_RTX (reg_single_usage, regno)))
905 && VARRAY_INT (set_in_loop, regno) == 1
906 && ! side_effects_p (SET_SRC (set))
907 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
908 && (! SMALL_REGISTER_CLASSES
909 || (! (GET_CODE (SET_SRC (set)) == REG
910 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)))
911 /* This test is not redundant; SET_SRC (set) might be
912 a call-clobbered register and the life of REGNO
913 might span a call. */
914 && ! modified_between_p (SET_SRC (set), p,
916 (reg_single_usage, regno))
917 && no_labels_between_p (p, VARRAY_RTX (reg_single_usage, regno))
918 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
920 (reg_single_usage, regno)))
922 /* Replace any usage in a REG_EQUAL note. Must copy the
923 new source, so that we don't get rtx sharing between the
924 SET_SOURCE and REG_NOTES of insn p. */
925 REG_NOTES (VARRAY_RTX (reg_single_usage, regno))
926 = replace_rtx (REG_NOTES (VARRAY_RTX
927 (reg_single_usage, regno)),
928 SET_DEST (set), copy_rtx (SET_SRC (set)));
931 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
932 NOTE_SOURCE_FILE (p) = 0;
933 VARRAY_INT (set_in_loop, regno) = 0;
937 m = (struct movable *) alloca (sizeof (struct movable));
941 m->dependencies = dependencies;
942 m->set_dest = SET_DEST (set);
944 m->consec = VARRAY_INT (set_in_loop,
945 REGNO (SET_DEST (set))) - 1;
949 m->move_insn = move_insn;
950 m->move_insn_first = 0;
951 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
952 m->savemode = VOIDmode;
954 /* Set M->cond if either invariant_p or consec_sets_invariant_p
955 returned 2 (only conditionally invariant). */
956 m->cond = ((tem | tem1 | tem2) > 1);
957 m->global = (uid_luid[REGNO_LAST_UID (regno)] > INSN_LUID (end)
958 || uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (loop_start));
960 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
961 - uid_luid[REGNO_FIRST_UID (regno)]);
962 m->savings = VARRAY_INT (n_times_set, regno);
963 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
964 m->savings += libcall_benefit (p);
965 VARRAY_INT (set_in_loop, regno) = move_insn ? -2 : -1;
966 /* Add M to the end of the chain MOVABLES. */
970 last_movable->next = m;
975 /* It is possible for the first instruction to have a
976 REG_EQUAL note but a non-invariant SET_SRC, so we must
977 remember the status of the first instruction in case
978 the last instruction doesn't have a REG_EQUAL note. */
979 m->move_insn_first = m->move_insn;
981 /* Skip this insn, not checking REG_LIBCALL notes. */
982 p = next_nonnote_insn (p);
983 /* Skip the consecutive insns, if there are any. */
984 p = skip_consec_insns (p, m->consec);
985 /* Back up to the last insn of the consecutive group. */
986 p = prev_nonnote_insn (p);
988 /* We must now reset m->move_insn, m->is_equiv, and possibly
989 m->set_src to correspond to the effects of all the
991 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
993 m->set_src = XEXP (temp, 0), m->move_insn = 1;
996 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
997 if (temp && CONSTANT_P (XEXP (temp, 0)))
998 m->set_src = XEXP (temp, 0), m->move_insn = 1;
1003 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
1006 /* If this register is always set within a STRICT_LOW_PART
1007 or set to zero, then its high bytes are constant.
1008 So clear them outside the loop and within the loop
1009 just load the low bytes.
1010 We must check that the machine has an instruction to do so.
1011 Also, if the value loaded into the register
1012 depends on the same register, this cannot be done. */
1013 else if (SET_SRC (set) == const0_rtx
1014 && GET_CODE (NEXT_INSN (p)) == INSN
1015 && (set1 = single_set (NEXT_INSN (p)))
1016 && GET_CODE (set1) == SET
1017 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
1018 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
1019 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
1021 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
1023 register int regno = REGNO (SET_DEST (set));
1024 if (VARRAY_INT (set_in_loop, regno) == 2)
1026 register struct movable *m;
1027 m = (struct movable *) alloca (sizeof (struct movable));
1030 m->set_dest = SET_DEST (set);
1031 m->dependencies = 0;
1037 m->move_insn_first = 0;
1039 /* If the insn may not be executed on some cycles,
1040 we can't clear the whole reg; clear just high part.
1041 Not even if the reg is used only within this loop.
1048 Clearing x before the inner loop could clobber a value
1049 being saved from the last time around the outer loop.
1050 However, if the reg is not used outside this loop
1051 and all uses of the register are in the same
1052 basic block as the store, there is no problem.
1054 If this insn was made by loop, we don't know its
1055 INSN_LUID and hence must make a conservative
1057 m->global = (INSN_UID (p) >= max_uid_for_loop
1058 || (uid_luid[REGNO_LAST_UID (regno)]
1060 || (uid_luid[REGNO_FIRST_UID (regno)]
1062 || (labels_in_range_p
1063 (p, uid_luid[REGNO_FIRST_UID (regno)])));
1064 if (maybe_never && m->global)
1065 m->savemode = GET_MODE (SET_SRC (set1));
1067 m->savemode = VOIDmode;
1071 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
1072 - uid_luid[REGNO_FIRST_UID (regno)]);
1074 VARRAY_INT (set_in_loop, regno) = -1;
1075 /* Add M to the end of the chain MOVABLES. */
1079 last_movable->next = m;
1084 /* Past a call insn, we get to insns which might not be executed
1085 because the call might exit. This matters for insns that trap.
1086 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
1087 so they don't count. */
1088 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
1090 /* Past a label or a jump, we get to insns for which we
1091 can't count on whether or how many times they will be
1092 executed during each iteration. Therefore, we can
1093 only move out sets of trivial variables
1094 (those not used after the loop). */
1095 /* Similar code appears twice in strength_reduce. */
1096 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
1097 /* If we enter the loop in the middle, and scan around to the
1098 beginning, don't set maybe_never for that. This must be an
1099 unconditional jump, otherwise the code at the top of the
1100 loop might never be executed. Unconditional jumps are
1101 followed a by barrier then loop end. */
1102 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
1103 && NEXT_INSN (NEXT_INSN (p)) == end
1104 && simplejump_p (p)))
1106 else if (GET_CODE (p) == NOTE)
1108 /* At the virtual top of a converted loop, insns are again known to
1109 be executed: logically, the loop begins here even though the exit
1110 code has been duplicated. */
1111 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
1112 maybe_never = call_passed = 0;
1113 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
1115 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
1120 /* If one movable subsumes another, ignore that other. */
1122 ignore_some_movables (movables);
1124 /* For each movable insn, see if the reg that it loads
1125 leads when it dies right into another conditionally movable insn.
1126 If so, record that the second insn "forces" the first one,
1127 since the second can be moved only if the first is. */
1129 force_movables (movables);
1131 /* See if there are multiple movable insns that load the same value.
1132 If there are, make all but the first point at the first one
1133 through the `match' field, and add the priorities of them
1134 all together as the priority of the first. */
1136 combine_movables (movables, nregs);
1138 /* Now consider each movable insn to decide whether it is worth moving.
1139 Store 0 in set_in_loop for each reg that is moved.
1141 Generally this increases code size, so do not move moveables when
1142 optimizing for code size. */
1144 if (! optimize_size)
1145 move_movables (movables, threshold,
1146 insn_count, loop_start, end, nregs);
1148 /* Now candidates that still are negative are those not moved.
1149 Change set_in_loop to indicate that those are not actually invariant. */
1150 for (i = 0; i < nregs; i++)
1151 if (VARRAY_INT (set_in_loop, i) < 0)
1152 VARRAY_INT (set_in_loop, i) = VARRAY_INT (n_times_set, i);
1154 /* Now that we've moved some things out of the loop, we might be able to
1155 hoist even more memory references. There's no need to pass
1156 reg_single_usage this time, since we're done with it. */
1157 load_mems_and_recount_loop_regs_set (scan_start, end, loop_top,
1161 /* set_in_loop is still used by invariant_p, so we can't free it now. */
1162 VARRAY_FREE (reg_single_usage);
1164 if (flag_strength_reduce)
1166 the_movables = movables;
1167 strength_reduce (scan_start, end, loop_top,
1168 insn_count, loop_start, end, unroll_p, bct_p);
1171 VARRAY_FREE (set_in_loop);
1172 VARRAY_FREE (n_times_set);
1173 VARRAY_FREE (may_not_optimize);
1176 /* Add elements to *OUTPUT to record all the pseudo-regs
1177 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1180 record_excess_regs (in_this, not_in_this, output)
1181 rtx in_this, not_in_this;
1188 code = GET_CODE (in_this);
1202 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1203 && ! reg_mentioned_p (in_this, not_in_this))
1204 *output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output);
1211 fmt = GET_RTX_FORMAT (code);
1212 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1219 for (j = 0; j < XVECLEN (in_this, i); j++)
1220 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1224 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1230 /* Check what regs are referred to in the libcall block ending with INSN,
1231 aside from those mentioned in the equivalent value.
1232 If there are none, return 0.
1233 If there are one or more, return an EXPR_LIST containing all of them. */
1236 libcall_other_reg (insn, equiv)
1239 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1240 rtx p = XEXP (note, 0);
1243 /* First, find all the regs used in the libcall block
1244 that are not mentioned as inputs to the result. */
1248 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1249 || GET_CODE (p) == CALL_INSN)
1250 record_excess_regs (PATTERN (p), equiv, &output);
1257 /* Return 1 if all uses of REG
1258 are between INSN and the end of the basic block. */
1261 reg_in_basic_block_p (insn, reg)
1264 int regno = REGNO (reg);
1267 if (REGNO_FIRST_UID (regno) != INSN_UID (insn))
1270 /* Search this basic block for the already recorded last use of the reg. */
1271 for (p = insn; p; p = NEXT_INSN (p))
1273 switch (GET_CODE (p))
1280 /* Ordinary insn: if this is the last use, we win. */
1281 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1286 /* Jump insn: if this is the last use, we win. */
1287 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1289 /* Otherwise, it's the end of the basic block, so we lose. */
1294 /* It's the end of the basic block, so we lose. */
1302 /* The "last use" doesn't follow the "first use"?? */
1306 /* Compute the benefit of eliminating the insns in the block whose
1307 last insn is LAST. This may be a group of insns used to compute a
1308 value directly or can contain a library call. */
1311 libcall_benefit (last)
1317 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1318 insn != last; insn = NEXT_INSN (insn))
1320 if (GET_CODE (insn) == CALL_INSN)
1321 benefit += 10; /* Assume at least this many insns in a library
1323 else if (GET_CODE (insn) == INSN
1324 && GET_CODE (PATTERN (insn)) != USE
1325 && GET_CODE (PATTERN (insn)) != CLOBBER)
1332 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1335 skip_consec_insns (insn, count)
1339 for (; count > 0; count--)
1343 /* If first insn of libcall sequence, skip to end. */
1344 /* Do this at start of loop, since INSN is guaranteed to
1346 if (GET_CODE (insn) != NOTE
1347 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1348 insn = XEXP (temp, 0);
1350 do insn = NEXT_INSN (insn);
1351 while (GET_CODE (insn) == NOTE);
1357 /* Ignore any movable whose insn falls within a libcall
1358 which is part of another movable.
1359 We make use of the fact that the movable for the libcall value
1360 was made later and so appears later on the chain. */
1363 ignore_some_movables (movables)
1364 struct movable *movables;
1366 register struct movable *m, *m1;
1368 for (m = movables; m; m = m->next)
1370 /* Is this a movable for the value of a libcall? */
1371 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1375 /* Check for earlier movables inside that range,
1376 and mark them invalid. We cannot use LUIDs here because
1377 insns created by loop.c for prior loops don't have LUIDs.
1378 Rather than reject all such insns from movables, we just
1379 explicitly check each insn in the libcall (since invariant
1380 libcalls aren't that common). */
1381 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1382 for (m1 = movables; m1 != m; m1 = m1->next)
1383 if (m1->insn == insn)
1389 /* For each movable insn, see if the reg that it loads
1390 leads when it dies right into another conditionally movable insn.
1391 If so, record that the second insn "forces" the first one,
1392 since the second can be moved only if the first is. */
1395 force_movables (movables)
1396 struct movable *movables;
1398 register struct movable *m, *m1;
1399 for (m1 = movables; m1; m1 = m1->next)
1400 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1401 if (!m1->partial && !m1->done)
1403 int regno = m1->regno;
1404 for (m = m1->next; m; m = m->next)
1405 /* ??? Could this be a bug? What if CSE caused the
1406 register of M1 to be used after this insn?
1407 Since CSE does not update regno_last_uid,
1408 this insn M->insn might not be where it dies.
1409 But very likely this doesn't matter; what matters is
1410 that M's reg is computed from M1's reg. */
1411 if (INSN_UID (m->insn) == REGNO_LAST_UID (regno)
1414 if (m != 0 && m->set_src == m1->set_dest
1415 /* If m->consec, m->set_src isn't valid. */
1419 /* Increase the priority of the moving the first insn
1420 since it permits the second to be moved as well. */
1424 m1->lifetime += m->lifetime;
1425 m1->savings += m->savings;
1430 /* Find invariant expressions that are equal and can be combined into
1434 combine_movables (movables, nregs)
1435 struct movable *movables;
1438 register struct movable *m;
1439 char *matched_regs = (char *) alloca (nregs);
1440 enum machine_mode mode;
1442 /* Regs that are set more than once are not allowed to match
1443 or be matched. I'm no longer sure why not. */
1444 /* Perhaps testing m->consec_sets would be more appropriate here? */
1446 for (m = movables; m; m = m->next)
1447 if (m->match == 0 && VARRAY_INT (n_times_set, m->regno) == 1 && !m->partial)
1449 register struct movable *m1;
1450 int regno = m->regno;
1452 bzero (matched_regs, nregs);
1453 matched_regs[regno] = 1;
1455 /* We want later insns to match the first one. Don't make the first
1456 one match any later ones. So start this loop at m->next. */
1457 for (m1 = m->next; m1; m1 = m1->next)
1458 if (m != m1 && m1->match == 0 && VARRAY_INT (n_times_set, m1->regno) == 1
1459 /* A reg used outside the loop mustn't be eliminated. */
1461 /* A reg used for zero-extending mustn't be eliminated. */
1463 && (matched_regs[m1->regno]
1466 /* Can combine regs with different modes loaded from the
1467 same constant only if the modes are the same or
1468 if both are integer modes with M wider or the same
1469 width as M1. The check for integer is redundant, but
1470 safe, since the only case of differing destination
1471 modes with equal sources is when both sources are
1472 VOIDmode, i.e., CONST_INT. */
1473 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1474 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1475 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1476 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1477 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1478 /* See if the source of M1 says it matches M. */
1479 && ((GET_CODE (m1->set_src) == REG
1480 && matched_regs[REGNO (m1->set_src)])
1481 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1483 && ((m->dependencies == m1->dependencies)
1484 || rtx_equal_p (m->dependencies, m1->dependencies)))
1486 m->lifetime += m1->lifetime;
1487 m->savings += m1->savings;
1490 matched_regs[m1->regno] = 1;
1494 /* Now combine the regs used for zero-extension.
1495 This can be done for those not marked `global'
1496 provided their lives don't overlap. */
1498 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1499 mode = GET_MODE_WIDER_MODE (mode))
1501 register struct movable *m0 = 0;
1503 /* Combine all the registers for extension from mode MODE.
1504 Don't combine any that are used outside this loop. */
1505 for (m = movables; m; m = m->next)
1506 if (m->partial && ! m->global
1507 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1509 register struct movable *m1;
1510 int first = uid_luid[REGNO_FIRST_UID (m->regno)];
1511 int last = uid_luid[REGNO_LAST_UID (m->regno)];
1515 /* First one: don't check for overlap, just record it. */
1520 /* Make sure they extend to the same mode.
1521 (Almost always true.) */
1522 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1525 /* We already have one: check for overlap with those
1526 already combined together. */
1527 for (m1 = movables; m1 != m; m1 = m1->next)
1528 if (m1 == m0 || (m1->partial && m1->match == m0))
1529 if (! (uid_luid[REGNO_FIRST_UID (m1->regno)] > last
1530 || uid_luid[REGNO_LAST_UID (m1->regno)] < first))
1533 /* No overlap: we can combine this with the others. */
1534 m0->lifetime += m->lifetime;
1535 m0->savings += m->savings;
1544 /* Return 1 if regs X and Y will become the same if moved. */
1547 regs_match_p (x, y, movables)
1549 struct movable *movables;
1553 struct movable *mx, *my;
1555 for (mx = movables; mx; mx = mx->next)
1556 if (mx->regno == xn)
1559 for (my = movables; my; my = my->next)
1560 if (my->regno == yn)
1564 && ((mx->match == my->match && mx->match != 0)
1566 || mx == my->match));
1569 /* Return 1 if X and Y are identical-looking rtx's.
1570 This is the Lisp function EQUAL for rtx arguments.
1572 If two registers are matching movables or a movable register and an
1573 equivalent constant, consider them equal. */
1576 rtx_equal_for_loop_p (x, y, movables)
1578 struct movable *movables;
1582 register struct movable *m;
1583 register enum rtx_code code;
1588 if (x == 0 || y == 0)
1591 code = GET_CODE (x);
1593 /* If we have a register and a constant, they may sometimes be
1595 if (GET_CODE (x) == REG && VARRAY_INT (set_in_loop, REGNO (x)) == -2
1598 for (m = movables; m; m = m->next)
1599 if (m->move_insn && m->regno == REGNO (x)
1600 && rtx_equal_p (m->set_src, y))
1603 else if (GET_CODE (y) == REG && VARRAY_INT (set_in_loop, REGNO (y)) == -2
1606 for (m = movables; m; m = m->next)
1607 if (m->move_insn && m->regno == REGNO (y)
1608 && rtx_equal_p (m->set_src, x))
1612 /* Otherwise, rtx's of different codes cannot be equal. */
1613 if (code != GET_CODE (y))
1616 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1617 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1619 if (GET_MODE (x) != GET_MODE (y))
1622 /* These three types of rtx's can be compared nonrecursively. */
1624 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1626 if (code == LABEL_REF)
1627 return XEXP (x, 0) == XEXP (y, 0);
1628 if (code == SYMBOL_REF)
1629 return XSTR (x, 0) == XSTR (y, 0);
1631 /* Compare the elements. If any pair of corresponding elements
1632 fail to match, return 0 for the whole things. */
1634 fmt = GET_RTX_FORMAT (code);
1635 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1640 if (XWINT (x, i) != XWINT (y, i))
1645 if (XINT (x, i) != XINT (y, i))
1650 /* Two vectors must have the same length. */
1651 if (XVECLEN (x, i) != XVECLEN (y, i))
1654 /* And the corresponding elements must match. */
1655 for (j = 0; j < XVECLEN (x, i); j++)
1656 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1661 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1666 if (strcmp (XSTR (x, i), XSTR (y, i)))
1671 /* These are just backpointers, so they don't matter. */
1677 /* It is believed that rtx's at this level will never
1678 contain anything but integers and other rtx's,
1679 except for within LABEL_REFs and SYMBOL_REFs. */
1687 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1688 insns in INSNS which use thet reference. */
1691 add_label_notes (x, insns)
1695 enum rtx_code code = GET_CODE (x);
1700 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1702 /* This code used to ignore labels that referred to dispatch tables to
1703 avoid flow generating (slighly) worse code.
1705 We no longer ignore such label references (see LABEL_REF handling in
1706 mark_jump_label for additional information). */
1707 for (insn = insns; insn; insn = NEXT_INSN (insn))
1708 if (reg_mentioned_p (XEXP (x, 0), insn))
1709 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
1713 fmt = GET_RTX_FORMAT (code);
1714 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1717 add_label_notes (XEXP (x, i), insns);
1718 else if (fmt[i] == 'E')
1719 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1720 add_label_notes (XVECEXP (x, i, j), insns);
1724 /* Scan MOVABLES, and move the insns that deserve to be moved.
1725 If two matching movables are combined, replace one reg with the
1726 other throughout. */
1729 move_movables (movables, threshold, insn_count, loop_start, end, nregs)
1730 struct movable *movables;
1738 register struct movable *m;
1740 /* Map of pseudo-register replacements to handle combining
1741 when we move several insns that load the same value
1742 into different pseudo-registers. */
1743 rtx *reg_map = (rtx *) alloca (nregs * sizeof (rtx));
1744 char *already_moved = (char *) alloca (nregs);
1746 bzero (already_moved, nregs);
1747 bzero ((char *) reg_map, nregs * sizeof (rtx));
1751 for (m = movables; m; m = m->next)
1753 /* Describe this movable insn. */
1755 if (loop_dump_stream)
1757 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1758 INSN_UID (m->insn), m->regno, m->lifetime);
1760 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1762 fprintf (loop_dump_stream, "cond ");
1764 fprintf (loop_dump_stream, "force ");
1766 fprintf (loop_dump_stream, "global ");
1768 fprintf (loop_dump_stream, "done ");
1770 fprintf (loop_dump_stream, "move-insn ");
1772 fprintf (loop_dump_stream, "matches %d ",
1773 INSN_UID (m->match->insn));
1775 fprintf (loop_dump_stream, "forces %d ",
1776 INSN_UID (m->forces->insn));
1779 /* Count movables. Value used in heuristics in strength_reduce. */
1782 /* Ignore the insn if it's already done (it matched something else).
1783 Otherwise, see if it is now safe to move. */
1787 || (1 == invariant_p (m->set_src)
1788 && (m->dependencies == 0
1789 || 1 == invariant_p (m->dependencies))
1791 || 1 == consec_sets_invariant_p (m->set_dest,
1794 && (! m->forces || m->forces->done))
1798 int savings = m->savings;
1800 /* We have an insn that is safe to move.
1801 Compute its desirability. */
1806 if (loop_dump_stream)
1807 fprintf (loop_dump_stream, "savings %d ", savings);
1809 if (moved_once[regno] && loop_dump_stream)
1810 fprintf (loop_dump_stream, "halved since already moved ");
1812 /* An insn MUST be moved if we already moved something else
1813 which is safe only if this one is moved too: that is,
1814 if already_moved[REGNO] is nonzero. */
1816 /* An insn is desirable to move if the new lifetime of the
1817 register is no more than THRESHOLD times the old lifetime.
1818 If it's not desirable, it means the loop is so big
1819 that moving won't speed things up much,
1820 and it is liable to make register usage worse. */
1822 /* It is also desirable to move if it can be moved at no
1823 extra cost because something else was already moved. */
1825 if (already_moved[regno]
1826 || flag_move_all_movables
1827 || (threshold * savings * m->lifetime) >=
1828 (moved_once[regno] ? insn_count * 2 : insn_count)
1829 || (m->forces && m->forces->done
1830 && VARRAY_INT (n_times_set, m->forces->regno) == 1))
1833 register struct movable *m1;
1836 /* Now move the insns that set the reg. */
1838 if (m->partial && m->match)
1842 /* Find the end of this chain of matching regs.
1843 Thus, we load each reg in the chain from that one reg.
1844 And that reg is loaded with 0 directly,
1845 since it has ->match == 0. */
1846 for (m1 = m; m1->match; m1 = m1->match);
1847 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1848 SET_DEST (PATTERN (m1->insn)));
1849 i1 = emit_insn_before (newpat, loop_start);
1851 /* Mark the moved, invariant reg as being allowed to
1852 share a hard reg with the other matching invariant. */
1853 REG_NOTES (i1) = REG_NOTES (m->insn);
1854 r1 = SET_DEST (PATTERN (m->insn));
1855 r2 = SET_DEST (PATTERN (m1->insn));
1857 = gen_rtx_EXPR_LIST (VOIDmode, r1,
1858 gen_rtx_EXPR_LIST (VOIDmode, r2,
1860 delete_insn (m->insn);
1865 if (loop_dump_stream)
1866 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1868 /* If we are to re-generate the item being moved with a
1869 new move insn, first delete what we have and then emit
1870 the move insn before the loop. */
1871 else if (m->move_insn)
1875 for (count = m->consec; count >= 0; count--)
1877 /* If this is the first insn of a library call sequence,
1879 if (GET_CODE (p) != NOTE
1880 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1883 /* If this is the last insn of a libcall sequence, then
1884 delete every insn in the sequence except the last.
1885 The last insn is handled in the normal manner. */
1886 if (GET_CODE (p) != NOTE
1887 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1889 temp = XEXP (temp, 0);
1891 temp = delete_insn (temp);
1895 p = delete_insn (p);
1897 /* simplify_giv_expr expects that it can walk the insns
1898 at m->insn forwards and see this old sequence we are
1899 tossing here. delete_insn does preserve the next
1900 pointers, but when we skip over a NOTE we must fix
1901 it up. Otherwise that code walks into the non-deleted
1903 while (p && GET_CODE (p) == NOTE)
1904 p = NEXT_INSN (temp) = NEXT_INSN (p);
1908 emit_move_insn (m->set_dest, m->set_src);
1909 temp = get_insns ();
1912 add_label_notes (m->set_src, temp);
1914 i1 = emit_insns_before (temp, loop_start);
1915 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1917 = gen_rtx_EXPR_LIST (m->is_equiv ? REG_EQUIV : REG_EQUAL,
1918 m->set_src, REG_NOTES (i1));
1920 if (loop_dump_stream)
1921 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1923 /* The more regs we move, the less we like moving them. */
1928 for (count = m->consec; count >= 0; count--)
1932 /* If first insn of libcall sequence, skip to end. */
1933 /* Do this at start of loop, since p is guaranteed to
1935 if (GET_CODE (p) != NOTE
1936 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1939 /* If last insn of libcall sequence, move all
1940 insns except the last before the loop. The last
1941 insn is handled in the normal manner. */
1942 if (GET_CODE (p) != NOTE
1943 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1947 rtx fn_address_insn = 0;
1950 for (temp = XEXP (temp, 0); temp != p;
1951 temp = NEXT_INSN (temp))
1957 if (GET_CODE (temp) == NOTE)
1960 body = PATTERN (temp);
1962 /* Find the next insn after TEMP,
1963 not counting USE or NOTE insns. */
1964 for (next = NEXT_INSN (temp); next != p;
1965 next = NEXT_INSN (next))
1966 if (! (GET_CODE (next) == INSN
1967 && GET_CODE (PATTERN (next)) == USE)
1968 && GET_CODE (next) != NOTE)
1971 /* If that is the call, this may be the insn
1972 that loads the function address.
1974 Extract the function address from the insn
1975 that loads it into a register.
1976 If this insn was cse'd, we get incorrect code.
1978 So emit a new move insn that copies the
1979 function address into the register that the
1980 call insn will use. flow.c will delete any
1981 redundant stores that we have created. */
1982 if (GET_CODE (next) == CALL_INSN
1983 && GET_CODE (body) == SET
1984 && GET_CODE (SET_DEST (body)) == REG
1985 && (n = find_reg_note (temp, REG_EQUAL,
1988 fn_reg = SET_SRC (body);
1989 if (GET_CODE (fn_reg) != REG)
1990 fn_reg = SET_DEST (body);
1991 fn_address = XEXP (n, 0);
1992 fn_address_insn = temp;
1994 /* We have the call insn.
1995 If it uses the register we suspect it might,
1996 load it with the correct address directly. */
1997 if (GET_CODE (temp) == CALL_INSN
1999 && reg_referenced_p (fn_reg, body))
2000 emit_insn_after (gen_move_insn (fn_reg,
2004 if (GET_CODE (temp) == CALL_INSN)
2006 i1 = emit_call_insn_before (body, loop_start);
2007 /* Because the USAGE information potentially
2008 contains objects other than hard registers
2009 we need to copy it. */
2010 if (CALL_INSN_FUNCTION_USAGE (temp))
2011 CALL_INSN_FUNCTION_USAGE (i1)
2012 = copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
2015 i1 = emit_insn_before (body, loop_start);
2018 if (temp == fn_address_insn)
2019 fn_address_insn = i1;
2020 REG_NOTES (i1) = REG_NOTES (temp);
2026 if (m->savemode != VOIDmode)
2028 /* P sets REG to zero; but we should clear only
2029 the bits that are not covered by the mode
2031 rtx reg = m->set_dest;
2037 (GET_MODE (reg), and_optab, reg,
2038 GEN_INT ((((HOST_WIDE_INT) 1
2039 << GET_MODE_BITSIZE (m->savemode)))
2041 reg, 1, OPTAB_LIB_WIDEN);
2045 emit_move_insn (reg, tem);
2046 sequence = gen_sequence ();
2048 i1 = emit_insn_before (sequence, loop_start);
2050 else if (GET_CODE (p) == CALL_INSN)
2052 i1 = emit_call_insn_before (PATTERN (p), loop_start);
2053 /* Because the USAGE information potentially
2054 contains objects other than hard registers
2055 we need to copy it. */
2056 if (CALL_INSN_FUNCTION_USAGE (p))
2057 CALL_INSN_FUNCTION_USAGE (i1)
2058 = copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
2060 else if (count == m->consec && m->move_insn_first)
2062 /* The SET_SRC might not be invariant, so we must
2063 use the REG_EQUAL note. */
2065 emit_move_insn (m->set_dest, m->set_src);
2066 temp = get_insns ();
2069 add_label_notes (m->set_src, temp);
2071 i1 = emit_insns_before (temp, loop_start);
2072 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
2074 = gen_rtx_EXPR_LIST ((m->is_equiv ? REG_EQUIV
2076 m->set_src, REG_NOTES (i1));
2079 i1 = emit_insn_before (PATTERN (p), loop_start);
2081 if (REG_NOTES (i1) == 0)
2083 REG_NOTES (i1) = REG_NOTES (p);
2085 /* If there is a REG_EQUAL note present whose value
2086 is not loop invariant, then delete it, since it
2087 may cause problems with later optimization passes.
2088 It is possible for cse to create such notes
2089 like this as a result of record_jump_cond. */
2091 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
2092 && ! invariant_p (XEXP (temp, 0)))
2093 remove_note (i1, temp);
2099 if (loop_dump_stream)
2100 fprintf (loop_dump_stream, " moved to %d",
2103 /* If library call, now fix the REG_NOTES that contain
2104 insn pointers, namely REG_LIBCALL on FIRST
2105 and REG_RETVAL on I1. */
2106 if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX)))
2108 XEXP (temp, 0) = first;
2109 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
2110 XEXP (temp, 0) = i1;
2117 /* simplify_giv_expr expects that it can walk the insns
2118 at m->insn forwards and see this old sequence we are
2119 tossing here. delete_insn does preserve the next
2120 pointers, but when we skip over a NOTE we must fix
2121 it up. Otherwise that code walks into the non-deleted
2123 while (p && GET_CODE (p) == NOTE)
2124 p = NEXT_INSN (temp) = NEXT_INSN (p);
2127 /* The more regs we move, the less we like moving them. */
2131 /* Any other movable that loads the same register
2133 already_moved[regno] = 1;
2135 /* This reg has been moved out of one loop. */
2136 moved_once[regno] = 1;
2138 /* The reg set here is now invariant. */
2140 VARRAY_INT (set_in_loop, regno) = 0;
2144 /* Change the length-of-life info for the register
2145 to say it lives at least the full length of this loop.
2146 This will help guide optimizations in outer loops. */
2148 if (uid_luid[REGNO_FIRST_UID (regno)] > INSN_LUID (loop_start))
2149 /* This is the old insn before all the moved insns.
2150 We can't use the moved insn because it is out of range
2151 in uid_luid. Only the old insns have luids. */
2152 REGNO_FIRST_UID (regno) = INSN_UID (loop_start);
2153 if (uid_luid[REGNO_LAST_UID (regno)] < INSN_LUID (end))
2154 REGNO_LAST_UID (regno) = INSN_UID (end);
2156 /* Combine with this moved insn any other matching movables. */
2159 for (m1 = movables; m1; m1 = m1->next)
2164 /* Schedule the reg loaded by M1
2165 for replacement so that shares the reg of M.
2166 If the modes differ (only possible in restricted
2167 circumstances, make a SUBREG. */
2168 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
2169 reg_map[m1->regno] = m->set_dest;
2172 = gen_lowpart_common (GET_MODE (m1->set_dest),
2175 /* Get rid of the matching insn
2176 and prevent further processing of it. */
2179 /* if library call, delete all insn except last, which
2181 if ((temp = find_reg_note (m1->insn, REG_RETVAL,
2184 for (temp = XEXP (temp, 0); temp != m1->insn;
2185 temp = NEXT_INSN (temp))
2188 delete_insn (m1->insn);
2190 /* Any other movable that loads the same register
2192 already_moved[m1->regno] = 1;
2194 /* The reg merged here is now invariant,
2195 if the reg it matches is invariant. */
2197 VARRAY_INT (set_in_loop, m1->regno) = 0;
2200 else if (loop_dump_stream)
2201 fprintf (loop_dump_stream, "not desirable");
2203 else if (loop_dump_stream && !m->match)
2204 fprintf (loop_dump_stream, "not safe");
2206 if (loop_dump_stream)
2207 fprintf (loop_dump_stream, "\n");
2211 new_start = loop_start;
2213 /* Go through all the instructions in the loop, making
2214 all the register substitutions scheduled in REG_MAP. */
2215 for (p = new_start; p != end; p = NEXT_INSN (p))
2216 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
2217 || GET_CODE (p) == CALL_INSN)
2219 replace_regs (PATTERN (p), reg_map, nregs, 0);
2220 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
2226 /* Scan X and replace the address of any MEM in it with ADDR.
2227 REG is the address that MEM should have before the replacement. */
2230 replace_call_address (x, reg, addr)
2233 register enum rtx_code code;
2239 code = GET_CODE (x);
2253 /* Short cut for very common case. */
2254 replace_call_address (XEXP (x, 1), reg, addr);
2258 /* Short cut for very common case. */
2259 replace_call_address (XEXP (x, 0), reg, addr);
2263 /* If this MEM uses a reg other than the one we expected,
2264 something is wrong. */
2265 if (XEXP (x, 0) != reg)
2274 fmt = GET_RTX_FORMAT (code);
2275 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2278 replace_call_address (XEXP (x, i), reg, addr);
2282 for (j = 0; j < XVECLEN (x, i); j++)
2283 replace_call_address (XVECEXP (x, i, j), reg, addr);
2289 /* Return the number of memory refs to addresses that vary
2293 count_nonfixed_reads (x)
2296 register enum rtx_code code;
2304 code = GET_CODE (x);
2318 return ((invariant_p (XEXP (x, 0)) != 1)
2319 + count_nonfixed_reads (XEXP (x, 0)));
2326 fmt = GET_RTX_FORMAT (code);
2327 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2330 value += count_nonfixed_reads (XEXP (x, i));
2334 for (j = 0; j < XVECLEN (x, i); j++)
2335 value += count_nonfixed_reads (XVECEXP (x, i, j));
2343 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2344 Replace it with an instruction to load just the low bytes
2345 if the machine supports such an instruction,
2346 and insert above LOOP_START an instruction to clear the register. */
2349 constant_high_bytes (p, loop_start)
2353 register int insn_code_number;
2355 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2356 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2358 new = gen_rtx_SET (VOIDmode,
2359 gen_rtx_STRICT_LOW_PART (VOIDmode,
2360 gen_rtx_SUBREG (GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2361 SET_DEST (PATTERN (p)),
2363 XEXP (SET_SRC (PATTERN (p)), 0));
2364 insn_code_number = recog (new, p);
2366 if (insn_code_number)
2370 /* Clear destination register before the loop. */
2371 emit_insn_before (gen_rtx_SET (VOIDmode, SET_DEST (PATTERN (p)),
2375 /* Inside the loop, just load the low part. */
2381 /* Scan a loop setting the variables `unknown_address_altered',
2382 `num_mem_sets', `loop_continue', `loops_enclosed', `loop_has_call',
2383 `loop_has_volatile', and `loop_has_tablejump'.
2384 Also, fill in the array `loop_mems' and the list `loop_store_mems'. */
2387 prescan_loop (start, end)
2390 register int level = 1;
2392 int loop_has_multiple_exit_targets = 0;
2393 /* The label after END. Jumping here is just like falling off the
2394 end of the loop. We use next_nonnote_insn instead of next_label
2395 as a hedge against the (pathological) case where some actual insn
2396 might end up between the two. */
2397 rtx exit_target = next_nonnote_insn (end);
2398 if (exit_target == NULL_RTX || GET_CODE (exit_target) != CODE_LABEL)
2399 loop_has_multiple_exit_targets = 1;
2401 unknown_address_altered = 0;
2403 loop_has_volatile = 0;
2404 loop_has_tablejump = 0;
2405 loop_store_mems = NULL_RTX;
2412 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2413 insn = NEXT_INSN (insn))
2415 if (GET_CODE (insn) == NOTE)
2417 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2420 /* Count number of loops contained in this one. */
2423 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2432 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2435 loop_continue = insn;
2438 else if (GET_CODE (insn) == CALL_INSN)
2440 if (! CONST_CALL_P (insn))
2441 unknown_address_altered = 1;
2444 else if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2446 rtx label1 = NULL_RTX;
2447 rtx label2 = NULL_RTX;
2449 if (volatile_refs_p (PATTERN (insn)))
2450 loop_has_volatile = 1;
2452 if (GET_CODE (insn) == JUMP_INSN
2453 && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
2454 || GET_CODE (PATTERN (insn)) == ADDR_VEC))
2455 loop_has_tablejump = 1;
2457 note_stores (PATTERN (insn), note_addr_stored);
2459 if (! loop_has_multiple_exit_targets
2460 && GET_CODE (insn) == JUMP_INSN
2461 && GET_CODE (PATTERN (insn)) == SET
2462 && SET_DEST (PATTERN (insn)) == pc_rtx)
2464 if (GET_CODE (SET_SRC (PATTERN (insn))) == IF_THEN_ELSE)
2466 label1 = XEXP (SET_SRC (PATTERN (insn)), 1);
2467 label2 = XEXP (SET_SRC (PATTERN (insn)), 2);
2471 label1 = SET_SRC (PATTERN (insn));
2475 if (label1 && label1 != pc_rtx)
2477 if (GET_CODE (label1) != LABEL_REF)
2479 /* Something tricky. */
2480 loop_has_multiple_exit_targets = 1;
2483 else if (XEXP (label1, 0) != exit_target
2484 && LABEL_OUTSIDE_LOOP_P (label1))
2486 /* A jump outside the current loop. */
2487 loop_has_multiple_exit_targets = 1;
2497 else if (GET_CODE (insn) == RETURN)
2498 loop_has_multiple_exit_targets = 1;
2501 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2502 if (/* We can't tell what MEMs are aliased by what. */
2503 !unknown_address_altered
2504 /* An exception thrown by a called function might land us
2507 /* We don't want loads for MEMs moved to a location before the
2508 one at which their stack memory becomes allocated. (Note
2509 that this is not a problem for malloc, etc., since those
2510 require actual function calls. */
2511 && !current_function_calls_alloca
2512 /* There are ways to leave the loop other than falling off the
2514 && !loop_has_multiple_exit_targets)
2515 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2516 insn = NEXT_INSN (insn))
2517 for_each_rtx (&insn, insert_loop_mem, 0);
2520 /* LOOP_NUMBER_CONT_DOMINATOR is now the last label between the loop start
2521 and the continue note that is a the destination of a (cond)jump after
2522 the continue note. If there is any (cond)jump between the loop start
2523 and what we have so far as LOOP_NUMBER_CONT_DOMINATOR that has a
2524 target between LOOP_DOMINATOR and the continue note, move
2525 LOOP_NUMBER_CONT_DOMINATOR forward to that label; if a jump's
2526 destination cannot be determined, clear LOOP_NUMBER_CONT_DOMINATOR. */
2529 verify_dominator (loop_number)
2534 if (! loop_number_cont_dominator[loop_number])
2535 /* This can happen for an empty loop, e.g. in
2536 gcc.c-torture/compile/920410-2.c */
2538 if (loop_number_cont_dominator[loop_number] == const0_rtx)
2540 loop_number_cont_dominator[loop_number] = 0;
2543 for (insn = loop_number_loop_starts[loop_number];
2544 insn != loop_number_cont_dominator[loop_number];
2545 insn = NEXT_INSN (insn))
2547 if (GET_CODE (insn) == JUMP_INSN
2548 && GET_CODE (PATTERN (insn)) != RETURN)
2550 rtx label = JUMP_LABEL (insn);
2551 int label_luid = INSN_LUID (label);
2553 if (! condjump_p (insn)
2554 && ! condjump_in_parallel_p (insn))
2556 loop_number_cont_dominator[loop_number] = NULL_RTX;
2559 if (label_luid < INSN_LUID (loop_number_loop_cont[loop_number])
2561 > INSN_LUID (loop_number_cont_dominator[loop_number])))
2562 loop_number_cont_dominator[loop_number] = label;
2567 /* Scan the function looking for loops. Record the start and end of each loop.
2568 Also mark as invalid loops any loops that contain a setjmp or are branched
2569 to from outside the loop. */
2572 find_and_verify_loops (f)
2576 int current_loop = -1;
2580 compute_luids (f, NULL_RTX, 0);
2582 /* If there are jumps to undefined labels,
2583 treat them as jumps out of any/all loops.
2584 This also avoids writing past end of tables when there are no loops. */
2585 uid_loop_num[0] = -1;
2587 /* Find boundaries of loops, mark which loops are contained within
2588 loops, and invalidate loops that have setjmp. */
2590 for (insn = f; insn; insn = NEXT_INSN (insn))
2592 if (GET_CODE (insn) == NOTE)
2593 switch (NOTE_LINE_NUMBER (insn))
2595 case NOTE_INSN_LOOP_BEG:
2596 loop_number_loop_starts[++next_loop] = insn;
2597 loop_number_loop_ends[next_loop] = 0;
2598 loop_number_loop_cont[next_loop] = 0;
2599 loop_number_cont_dominator[next_loop] = 0;
2600 loop_outer_loop[next_loop] = current_loop;
2601 loop_invalid[next_loop] = 0;
2602 loop_number_exit_labels[next_loop] = 0;
2603 loop_number_exit_count[next_loop] = 0;
2604 current_loop = next_loop;
2607 case NOTE_INSN_SETJMP:
2608 /* In this case, we must invalidate our current loop and any
2610 for (loop = current_loop; loop != -1; loop = loop_outer_loop[loop])
2612 loop_invalid[loop] = 1;
2613 if (loop_dump_stream)
2614 fprintf (loop_dump_stream,
2615 "\nLoop at %d ignored due to setjmp.\n",
2616 INSN_UID (loop_number_loop_starts[loop]));
2620 case NOTE_INSN_LOOP_CONT:
2621 loop_number_loop_cont[current_loop] = insn;
2623 case NOTE_INSN_LOOP_END:
2624 if (current_loop == -1)
2627 loop_number_loop_ends[current_loop] = insn;
2628 verify_dominator (current_loop);
2629 current_loop = loop_outer_loop[current_loop];
2635 /* If for any loop, this is a jump insn between the NOTE_INSN_LOOP_CONT
2636 and NOTE_INSN_LOOP_END notes, update loop_number_loop_dominator. */
2637 else if (GET_CODE (insn) == JUMP_INSN
2638 && GET_CODE (PATTERN (insn)) != RETURN
2639 && current_loop >= 0)
2642 rtx label = JUMP_LABEL (insn);
2644 if (! condjump_p (insn) && ! condjump_in_parallel_p (insn))
2647 this_loop = current_loop;
2650 /* First see if we care about this loop. */
2651 if (loop_number_loop_cont[this_loop]
2652 && loop_number_cont_dominator[this_loop] != const0_rtx)
2654 /* If the jump destination is not known, invalidate
2655 loop_number_const_dominator. */
2657 loop_number_cont_dominator[this_loop] = const0_rtx;
2659 /* Check if the destination is between loop start and
2661 if ((INSN_LUID (label)
2662 < INSN_LUID (loop_number_loop_cont[this_loop]))
2663 && (INSN_LUID (label)
2664 > INSN_LUID (loop_number_loop_starts[this_loop]))
2665 /* And if there is no later destination already
2667 && (! loop_number_cont_dominator[this_loop]
2668 || (INSN_LUID (label)
2669 > INSN_LUID (loop_number_cont_dominator
2671 loop_number_cont_dominator[this_loop] = label;
2673 this_loop = loop_outer_loop[this_loop];
2675 while (this_loop >= 0);
2678 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2679 enclosing loop, but this doesn't matter. */
2680 uid_loop_num[INSN_UID (insn)] = current_loop;
2683 /* Any loop containing a label used in an initializer must be invalidated,
2684 because it can be jumped into from anywhere. */
2686 for (label = forced_labels; label; label = XEXP (label, 1))
2690 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2692 loop_num = loop_outer_loop[loop_num])
2693 loop_invalid[loop_num] = 1;
2696 /* Any loop containing a label used for an exception handler must be
2697 invalidated, because it can be jumped into from anywhere. */
2699 for (label = exception_handler_labels; label; label = XEXP (label, 1))
2703 for (loop_num = uid_loop_num[INSN_UID (XEXP (label, 0))];
2705 loop_num = loop_outer_loop[loop_num])
2706 loop_invalid[loop_num] = 1;
2709 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2710 loop that it is not contained within, that loop is marked invalid.
2711 If any INSN or CALL_INSN uses a label's address, then the loop containing
2712 that label is marked invalid, because it could be jumped into from
2715 Also look for blocks of code ending in an unconditional branch that
2716 exits the loop. If such a block is surrounded by a conditional
2717 branch around the block, move the block elsewhere (see below) and
2718 invert the jump to point to the code block. This may eliminate a
2719 label in our loop and will simplify processing by both us and a
2720 possible second cse pass. */
2722 for (insn = f; insn; insn = NEXT_INSN (insn))
2723 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2725 int this_loop_num = uid_loop_num[INSN_UID (insn)];
2727 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2729 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2734 for (loop_num = uid_loop_num[INSN_UID (XEXP (note, 0))];
2736 loop_num = loop_outer_loop[loop_num])
2737 loop_invalid[loop_num] = 1;
2741 if (GET_CODE (insn) != JUMP_INSN)
2744 mark_loop_jump (PATTERN (insn), this_loop_num);
2746 /* See if this is an unconditional branch outside the loop. */
2747 if (this_loop_num != -1
2748 && (GET_CODE (PATTERN (insn)) == RETURN
2749 || (simplejump_p (insn)
2750 && (uid_loop_num[INSN_UID (JUMP_LABEL (insn))]
2752 && get_max_uid () < max_uid_for_loop)
2755 rtx our_next = next_real_insn (insn);
2757 int outer_loop = -1;
2759 /* Go backwards until we reach the start of the loop, a label,
2761 for (p = PREV_INSN (insn);
2762 GET_CODE (p) != CODE_LABEL
2763 && ! (GET_CODE (p) == NOTE
2764 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2765 && GET_CODE (p) != JUMP_INSN;
2769 /* Check for the case where we have a jump to an inner nested
2770 loop, and do not perform the optimization in that case. */
2772 if (JUMP_LABEL (insn))
2774 dest_loop = uid_loop_num[INSN_UID (JUMP_LABEL (insn))];
2775 if (dest_loop != -1)
2777 for (outer_loop = dest_loop; outer_loop != -1;
2778 outer_loop = loop_outer_loop[outer_loop])
2779 if (outer_loop == this_loop_num)
2784 /* Make sure that the target of P is within the current loop. */
2786 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2787 && uid_loop_num[INSN_UID (JUMP_LABEL (p))] != this_loop_num)
2788 outer_loop = this_loop_num;
2790 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2791 we have a block of code to try to move.
2793 We look backward and then forward from the target of INSN
2794 to find a BARRIER at the same loop depth as the target.
2795 If we find such a BARRIER, we make a new label for the start
2796 of the block, invert the jump in P and point it to that label,
2797 and move the block of code to the spot we found. */
2799 if (outer_loop == -1
2800 && GET_CODE (p) == JUMP_INSN
2801 && JUMP_LABEL (p) != 0
2802 /* Just ignore jumps to labels that were never emitted.
2803 These always indicate compilation errors. */
2804 && INSN_UID (JUMP_LABEL (p)) != 0
2806 && ! simplejump_p (p)
2807 && next_real_insn (JUMP_LABEL (p)) == our_next)
2810 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2811 int target_loop_num = uid_loop_num[INSN_UID (target)];
2814 for (loc = target; loc; loc = PREV_INSN (loc))
2815 if (GET_CODE (loc) == BARRIER
2816 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2820 for (loc = target; loc; loc = NEXT_INSN (loc))
2821 if (GET_CODE (loc) == BARRIER
2822 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2827 rtx cond_label = JUMP_LABEL (p);
2828 rtx new_label = get_label_after (p);
2830 /* Ensure our label doesn't go away. */
2831 LABEL_NUSES (cond_label)++;
2833 /* Verify that uid_loop_num is large enough and that
2835 if (invert_jump (p, new_label))
2839 /* If no suitable BARRIER was found, create a suitable
2840 one before TARGET. Since TARGET is a fall through
2841 path, we'll need to insert an jump around our block
2842 and a add a BARRIER before TARGET.
2844 This creates an extra unconditional jump outside
2845 the loop. However, the benefits of removing rarely
2846 executed instructions from inside the loop usually
2847 outweighs the cost of the extra unconditional jump
2848 outside the loop. */
2853 temp = gen_jump (JUMP_LABEL (insn));
2854 temp = emit_jump_insn_before (temp, target);
2855 JUMP_LABEL (temp) = JUMP_LABEL (insn);
2856 LABEL_NUSES (JUMP_LABEL (insn))++;
2857 loc = emit_barrier_before (target);
2860 /* Include the BARRIER after INSN and copy the
2862 new_label = squeeze_notes (new_label, NEXT_INSN (insn));
2863 reorder_insns (new_label, NEXT_INSN (insn), loc);
2865 /* All those insns are now in TARGET_LOOP_NUM. */
2866 for (q = new_label; q != NEXT_INSN (NEXT_INSN (insn));
2868 uid_loop_num[INSN_UID (q)] = target_loop_num;
2870 /* The label jumped to by INSN is no longer a loop exit.
2871 Unless INSN does not have a label (e.g., it is a
2872 RETURN insn), search loop_number_exit_labels to find
2873 its label_ref, and remove it. Also turn off
2874 LABEL_OUTSIDE_LOOP_P bit. */
2875 if (JUMP_LABEL (insn))
2880 r = loop_number_exit_labels[this_loop_num];
2881 r; q = r, r = LABEL_NEXTREF (r))
2882 if (XEXP (r, 0) == JUMP_LABEL (insn))
2884 LABEL_OUTSIDE_LOOP_P (r) = 0;
2886 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2888 loop_number_exit_labels[this_loop_num]
2889 = LABEL_NEXTREF (r);
2893 for (loop_num = this_loop_num;
2894 loop_num != -1 && loop_num != target_loop_num;
2895 loop_num = loop_outer_loop[loop_num])
2896 loop_number_exit_count[loop_num]--;
2898 /* If we didn't find it, then something is wrong. */
2903 /* P is now a jump outside the loop, so it must be put
2904 in loop_number_exit_labels, and marked as such.
2905 The easiest way to do this is to just call
2906 mark_loop_jump again for P. */
2907 mark_loop_jump (PATTERN (p), this_loop_num);
2909 /* If INSN now jumps to the insn after it,
2911 if (JUMP_LABEL (insn) != 0
2912 && (next_real_insn (JUMP_LABEL (insn))
2913 == next_real_insn (insn)))
2917 /* Continue the loop after where the conditional
2918 branch used to jump, since the only branch insn
2919 in the block (if it still remains) is an inter-loop
2920 branch and hence needs no processing. */
2921 insn = NEXT_INSN (cond_label);
2923 if (--LABEL_NUSES (cond_label) == 0)
2924 delete_insn (cond_label);
2926 /* This loop will be continued with NEXT_INSN (insn). */
2927 insn = PREV_INSN (insn);
2934 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2935 loops it is contained in, mark the target loop invalid.
2937 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2940 mark_loop_jump (x, loop_num)
2948 switch (GET_CODE (x))
2961 /* There could be a label reference in here. */
2962 mark_loop_jump (XEXP (x, 0), loop_num);
2968 mark_loop_jump (XEXP (x, 0), loop_num);
2969 mark_loop_jump (XEXP (x, 1), loop_num);
2974 mark_loop_jump (XEXP (x, 0), loop_num);
2978 dest_loop = uid_loop_num[INSN_UID (XEXP (x, 0))];
2980 /* Link together all labels that branch outside the loop. This
2981 is used by final_[bg]iv_value and the loop unrolling code. Also
2982 mark this LABEL_REF so we know that this branch should predict
2985 /* A check to make sure the label is not in an inner nested loop,
2986 since this does not count as a loop exit. */
2987 if (dest_loop != -1)
2989 for (outer_loop = dest_loop; outer_loop != -1;
2990 outer_loop = loop_outer_loop[outer_loop])
2991 if (outer_loop == loop_num)
2997 if (loop_num != -1 && outer_loop == -1)
2999 LABEL_OUTSIDE_LOOP_P (x) = 1;
3000 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
3001 loop_number_exit_labels[loop_num] = x;
3003 for (outer_loop = loop_num;
3004 outer_loop != -1 && outer_loop != dest_loop;
3005 outer_loop = loop_outer_loop[outer_loop])
3006 loop_number_exit_count[outer_loop]++;
3009 /* If this is inside a loop, but not in the current loop or one enclosed
3010 by it, it invalidates at least one loop. */
3012 if (dest_loop == -1)
3015 /* We must invalidate every nested loop containing the target of this
3016 label, except those that also contain the jump insn. */
3018 for (; dest_loop != -1; dest_loop = loop_outer_loop[dest_loop])
3020 /* Stop when we reach a loop that also contains the jump insn. */
3021 for (outer_loop = loop_num; outer_loop != -1;
3022 outer_loop = loop_outer_loop[outer_loop])
3023 if (dest_loop == outer_loop)
3026 /* If we get here, we know we need to invalidate a loop. */
3027 if (loop_dump_stream && ! loop_invalid[dest_loop])
3028 fprintf (loop_dump_stream,
3029 "\nLoop at %d ignored due to multiple entry points.\n",
3030 INSN_UID (loop_number_loop_starts[dest_loop]));
3032 loop_invalid[dest_loop] = 1;
3037 /* If this is not setting pc, ignore. */
3038 if (SET_DEST (x) == pc_rtx)
3039 mark_loop_jump (SET_SRC (x), loop_num);
3043 mark_loop_jump (XEXP (x, 1), loop_num);
3044 mark_loop_jump (XEXP (x, 2), loop_num);
3049 for (i = 0; i < XVECLEN (x, 0); i++)
3050 mark_loop_jump (XVECEXP (x, 0, i), loop_num);
3054 for (i = 0; i < XVECLEN (x, 1); i++)
3055 mark_loop_jump (XVECEXP (x, 1, i), loop_num);
3059 /* Treat anything else (such as a symbol_ref)
3060 as a branch out of this loop, but not into any loop. */
3064 #ifdef HAVE_decrement_and_branch_on_count
3065 LABEL_OUTSIDE_LOOP_P (x) = 1;
3066 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
3067 #endif /* HAVE_decrement_and_branch_on_count */
3069 loop_number_exit_labels[loop_num] = x;
3071 for (outer_loop = loop_num; outer_loop != -1;
3072 outer_loop = loop_outer_loop[outer_loop])
3073 loop_number_exit_count[outer_loop]++;
3079 /* Return nonzero if there is a label in the range from
3080 insn INSN to and including the insn whose luid is END
3081 INSN must have an assigned luid (i.e., it must not have
3082 been previously created by loop.c). */
3085 labels_in_range_p (insn, end)
3089 while (insn && INSN_LUID (insn) <= end)
3091 if (GET_CODE (insn) == CODE_LABEL)
3093 insn = NEXT_INSN (insn);
3099 /* Record that a memory reference X is being set. */
3102 note_addr_stored (x, y)
3104 rtx y ATTRIBUTE_UNUSED;
3106 if (x == 0 || GET_CODE (x) != MEM)
3109 /* Count number of memory writes.
3110 This affects heuristics in strength_reduce. */
3113 /* BLKmode MEM means all memory is clobbered. */
3114 if (GET_MODE (x) == BLKmode)
3115 unknown_address_altered = 1;
3117 if (unknown_address_altered)
3120 loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, x, loop_store_mems);
3123 /* Return nonzero if the rtx X is invariant over the current loop.
3125 The value is 2 if we refer to something only conditionally invariant.
3127 If `unknown_address_altered' is nonzero, no memory ref is invariant.
3128 Otherwise, a memory ref is invariant if it does not conflict with
3129 anything stored in `loop_store_mems'. */
3136 register enum rtx_code code;
3138 int conditional = 0;
3143 code = GET_CODE (x);
3153 /* A LABEL_REF is normally invariant, however, if we are unrolling
3154 loops, and this label is inside the loop, then it isn't invariant.
3155 This is because each unrolled copy of the loop body will have
3156 a copy of this label. If this was invariant, then an insn loading
3157 the address of this label into a register might get moved outside
3158 the loop, and then each loop body would end up using the same label.
3160 We don't know the loop bounds here though, so just fail for all
3162 if (flag_unroll_loops)
3169 case UNSPEC_VOLATILE:
3173 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
3174 since the reg might be set by initialization within the loop. */
3176 if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
3177 || x == arg_pointer_rtx)
3178 && ! current_function_has_nonlocal_goto)
3182 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
3185 if (VARRAY_INT (set_in_loop, REGNO (x)) < 0)
3188 return VARRAY_INT (set_in_loop, REGNO (x)) == 0;
3191 /* Volatile memory references must be rejected. Do this before
3192 checking for read-only items, so that volatile read-only items
3193 will be rejected also. */
3194 if (MEM_VOLATILE_P (x))
3197 /* Read-only items (such as constants in a constant pool) are
3198 invariant if their address is. */
3199 if (RTX_UNCHANGING_P (x))
3202 /* If we had a subroutine call, any location in memory could have been
3204 if (unknown_address_altered)
3207 /* See if there is any dependence between a store and this load. */
3208 mem_list_entry = loop_store_mems;
3209 while (mem_list_entry)
3211 if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
3214 mem_list_entry = XEXP (mem_list_entry, 1);
3217 /* It's not invalidated by a store in memory
3218 but we must still verify the address is invariant. */
3222 /* Don't mess with insns declared volatile. */
3223 if (MEM_VOLATILE_P (x))
3231 fmt = GET_RTX_FORMAT (code);
3232 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3236 int tem = invariant_p (XEXP (x, i));
3242 else if (fmt[i] == 'E')
3245 for (j = 0; j < XVECLEN (x, i); j++)
3247 int tem = invariant_p (XVECEXP (x, i, j));
3257 return 1 + conditional;
3261 /* Return nonzero if all the insns in the loop that set REG
3262 are INSN and the immediately following insns,
3263 and if each of those insns sets REG in an invariant way
3264 (not counting uses of REG in them).
3266 The value is 2 if some of these insns are only conditionally invariant.
3268 We assume that INSN itself is the first set of REG
3269 and that its source is invariant. */
3272 consec_sets_invariant_p (reg, n_sets, insn)
3276 register rtx p = insn;
3277 register int regno = REGNO (reg);
3279 /* Number of sets we have to insist on finding after INSN. */
3280 int count = n_sets - 1;
3281 int old = VARRAY_INT (set_in_loop, regno);
3285 /* If N_SETS hit the limit, we can't rely on its value. */
3289 VARRAY_INT (set_in_loop, regno) = 0;
3293 register enum rtx_code code;
3297 code = GET_CODE (p);
3299 /* If library call, skip to end of it. */
3300 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3305 && (set = single_set (p))
3306 && GET_CODE (SET_DEST (set)) == REG
3307 && REGNO (SET_DEST (set)) == regno)
3309 this = invariant_p (SET_SRC (set));
3312 else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)))
3314 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3315 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3317 this = (CONSTANT_P (XEXP (temp, 0))
3318 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
3319 && invariant_p (XEXP (temp, 0))));
3326 else if (code != NOTE)
3328 VARRAY_INT (set_in_loop, regno) = old;
3333 VARRAY_INT (set_in_loop, regno) = old;
3334 /* If invariant_p ever returned 2, we return 2. */
3335 return 1 + (value & 2);
3339 /* I don't think this condition is sufficient to allow INSN
3340 to be moved, so we no longer test it. */
3342 /* Return 1 if all insns in the basic block of INSN and following INSN
3343 that set REG are invariant according to TABLE. */
3346 all_sets_invariant_p (reg, insn, table)
3350 register rtx p = insn;
3351 register int regno = REGNO (reg);
3355 register enum rtx_code code;
3357 code = GET_CODE (p);
3358 if (code == CODE_LABEL || code == JUMP_INSN)
3360 if (code == INSN && GET_CODE (PATTERN (p)) == SET
3361 && GET_CODE (SET_DEST (PATTERN (p))) == REG
3362 && REGNO (SET_DEST (PATTERN (p))) == regno)
3364 if (!invariant_p (SET_SRC (PATTERN (p)), table))
3371 /* Look at all uses (not sets) of registers in X. For each, if it is
3372 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3373 a different insn, set USAGE[REGNO] to const0_rtx. */
3376 find_single_use_in_loop (insn, x, usage)
3381 enum rtx_code code = GET_CODE (x);
3382 char *fmt = GET_RTX_FORMAT (code);
3386 VARRAY_RTX (usage, REGNO (x))
3387 = (VARRAY_RTX (usage, REGNO (x)) != 0
3388 && VARRAY_RTX (usage, REGNO (x)) != insn)
3389 ? const0_rtx : insn;
3391 else if (code == SET)
3393 /* Don't count SET_DEST if it is a REG; otherwise count things
3394 in SET_DEST because if a register is partially modified, it won't
3395 show up as a potential movable so we don't care how USAGE is set
3397 if (GET_CODE (SET_DEST (x)) != REG)
3398 find_single_use_in_loop (insn, SET_DEST (x), usage);
3399 find_single_use_in_loop (insn, SET_SRC (x), usage);
3402 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3404 if (fmt[i] == 'e' && XEXP (x, i) != 0)
3405 find_single_use_in_loop (insn, XEXP (x, i), usage);
3406 else if (fmt[i] == 'E')
3407 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3408 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
3412 /* Count and record any set in X which is contained in INSN. Update
3413 MAY_NOT_MOVE and LAST_SET for any register set in X. */
3416 count_one_set (insn, x, may_not_move, last_set)
3418 varray_type may_not_move;
3421 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3422 /* Don't move a reg that has an explicit clobber.
3423 It's not worth the pain to try to do it correctly. */
3424 VARRAY_CHAR (may_not_move, REGNO (XEXP (x, 0))) = 1;
3426 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3428 rtx dest = SET_DEST (x);
3429 while (GET_CODE (dest) == SUBREG
3430 || GET_CODE (dest) == ZERO_EXTRACT
3431 || GET_CODE (dest) == SIGN_EXTRACT
3432 || GET_CODE (dest) == STRICT_LOW_PART)
3433 dest = XEXP (dest, 0);
3434 if (GET_CODE (dest) == REG)
3436 register int regno = REGNO (dest);
3437 /* If this is the first setting of this reg
3438 in current basic block, and it was set before,
3439 it must be set in two basic blocks, so it cannot
3440 be moved out of the loop. */
3441 if (VARRAY_INT (set_in_loop, regno) > 0
3442 && last_set[regno] == 0)
3443 VARRAY_CHAR (may_not_move, regno) = 1;
3444 /* If this is not first setting in current basic block,
3445 see if reg was used in between previous one and this.
3446 If so, neither one can be moved. */
3447 if (last_set[regno] != 0
3448 && reg_used_between_p (dest, last_set[regno], insn))
3449 VARRAY_CHAR (may_not_move, regno) = 1;
3450 if (VARRAY_INT (set_in_loop, regno) < 127)
3451 ++VARRAY_INT (set_in_loop, regno);
3452 last_set[regno] = insn;
3457 /* Increment SET_IN_LOOP at the index of each register
3458 that is modified by an insn between FROM and TO.
3459 If the value of an element of SET_IN_LOOP becomes 127 or more,
3460 stop incrementing it, to avoid overflow.
3462 Store in SINGLE_USAGE[I] the single insn in which register I is
3463 used, if it is only used once. Otherwise, it is set to 0 (for no
3464 uses) or const0_rtx for more than one use. This parameter may be zero,
3465 in which case this processing is not done.
3467 Store in *COUNT_PTR the number of actual instruction
3468 in the loop. We use this to decide what is worth moving out. */
3470 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3471 In that case, it is the insn that last set reg n. */
3474 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
3475 register rtx from, to;
3476 varray_type may_not_move;
3477 varray_type single_usage;
3481 register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx));
3483 register int count = 0;
3485 bzero ((char *) last_set, nregs * sizeof (rtx));
3486 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3488 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3492 /* If requested, record registers that have exactly one use. */
3495 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3497 /* Include uses in REG_EQUAL notes. */
3498 if (REG_NOTES (insn))
3499 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3502 if (GET_CODE (PATTERN (insn)) == SET
3503 || GET_CODE (PATTERN (insn)) == CLOBBER)
3504 count_one_set (insn, PATTERN (insn), may_not_move, last_set);
3505 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3508 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3509 count_one_set (insn, XVECEXP (PATTERN (insn), 0, i),
3510 may_not_move, last_set);
3514 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3515 bzero ((char *) last_set, nregs * sizeof (rtx));
3520 /* Given a loop that is bounded by LOOP_START and LOOP_END
3521 and that is entered at SCAN_START,
3522 return 1 if the register set in SET contained in insn INSN is used by
3523 any insn that precedes INSN in cyclic order starting
3524 from the loop entry point.
3526 We don't want to use INSN_LUID here because if we restrict INSN to those
3527 that have a valid INSN_LUID, it means we cannot move an invariant out
3528 from an inner loop past two loops. */
3531 loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end)
3532 rtx set, insn, loop_start, scan_start, loop_end;
3534 rtx reg = SET_DEST (set);
3537 /* Scan forward checking for register usage. If we hit INSN, we
3538 are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */
3539 for (p = scan_start; p != insn; p = NEXT_INSN (p))
3541 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
3542 && reg_overlap_mentioned_p (reg, PATTERN (p)))
3552 /* A "basic induction variable" or biv is a pseudo reg that is set
3553 (within this loop) only by incrementing or decrementing it. */
3554 /* A "general induction variable" or giv is a pseudo reg whose
3555 value is a linear function of a biv. */
3557 /* Bivs are recognized by `basic_induction_var';
3558 Givs by `general_induction_var'. */
3560 /* Indexed by register number, indicates whether or not register is an
3561 induction variable, and if so what type. */
3563 varray_type reg_iv_type;
3565 /* Indexed by register number, contains pointer to `struct induction'
3566 if register is an induction variable. This holds general info for
3567 all induction variables. */
3569 varray_type reg_iv_info;
3571 /* Indexed by register number, contains pointer to `struct iv_class'
3572 if register is a basic induction variable. This holds info describing
3573 the class (a related group) of induction variables that the biv belongs
3576 struct iv_class **reg_biv_class;
3578 /* The head of a list which links together (via the next field)
3579 every iv class for the current loop. */
3581 struct iv_class *loop_iv_list;
3583 /* Givs made from biv increments are always splittable for loop unrolling.
3584 Since there is no regscan info for them, we have to keep track of them
3586 int first_increment_giv, last_increment_giv;
3588 /* Communication with routines called via `note_stores'. */
3590 static rtx note_insn;
3592 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3594 static rtx addr_placeholder;
3596 /* ??? Unfinished optimizations, and possible future optimizations,
3597 for the strength reduction code. */
3599 /* ??? The interaction of biv elimination, and recognition of 'constant'
3600 bivs, may cause problems. */
3602 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3603 performance problems.
3605 Perhaps don't eliminate things that can be combined with an addressing
3606 mode. Find all givs that have the same biv, mult_val, and add_val;
3607 then for each giv, check to see if its only use dies in a following
3608 memory address. If so, generate a new memory address and check to see
3609 if it is valid. If it is valid, then store the modified memory address,
3610 otherwise, mark the giv as not done so that it will get its own iv. */
3612 /* ??? Could try to optimize branches when it is known that a biv is always
3615 /* ??? When replace a biv in a compare insn, we should replace with closest
3616 giv so that an optimized branch can still be recognized by the combiner,
3617 e.g. the VAX acb insn. */
3619 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3620 was rerun in loop_optimize whenever a register was added or moved.
3621 Also, some of the optimizations could be a little less conservative. */
3623 /* Perform strength reduction and induction variable elimination.
3625 Pseudo registers created during this function will be beyond the last
3626 valid index in several tables including n_times_set and regno_last_uid.
3627 This does not cause a problem here, because the added registers cannot be
3628 givs outside of their loop, and hence will never be reconsidered.
3629 But scan_loop must check regnos to make sure they are in bounds.
3631 SCAN_START is the first instruction in the loop, as the loop would
3632 actually be executed. END is the NOTE_INSN_LOOP_END. LOOP_TOP is
3633 the first instruction in the loop, as it is layed out in the
3634 instruction stream. LOOP_START is the NOTE_INSN_LOOP_BEG. */
3637 strength_reduce (scan_start, end, loop_top, insn_count,
3638 loop_start, loop_end, unroll_p, bct_p)
3645 int unroll_p, bct_p ATTRIBUTE_UNUSED;
3653 /* This is 1 if current insn is not executed at least once for every loop
3655 int not_every_iteration = 0;
3656 /* This is 1 if current insn may be executed more than once for every
3658 int maybe_multiple = 0;
3659 /* Temporary list pointers for traversing loop_iv_list. */
3660 struct iv_class *bl, **backbl;
3661 /* Ratio of extra register life span we can justify
3662 for saving an instruction. More if loop doesn't call subroutines
3663 since in that case saving an insn makes more difference
3664 and more registers are available. */
3665 /* ??? could set this to last value of threshold in move_movables */
3666 int threshold = (loop_has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3667 /* Map of pseudo-register replacements. */
3672 rtx end_insert_before;
3674 int n_extra_increment;
3675 struct loop_info loop_iteration_info;
3676 struct loop_info *loop_info = &loop_iteration_info;
3678 /* If scan_start points to the loop exit test, we have to be wary of
3679 subversive use of gotos inside expression statements. */
3680 if (prev_nonnote_insn (scan_start) != prev_nonnote_insn (loop_start))
3681 maybe_multiple = back_branch_in_range_p (scan_start, loop_start, loop_end);
3683 VARRAY_INT_INIT (reg_iv_type, max_reg_before_loop, "reg_iv_type");
3684 VARRAY_GENERIC_PTR_INIT (reg_iv_info, max_reg_before_loop, "reg_iv_info");
3685 reg_biv_class = (struct iv_class **)
3686 alloca (max_reg_before_loop * sizeof (struct iv_class *));
3687 bzero ((char *) reg_biv_class, (max_reg_before_loop
3688 * sizeof (struct iv_class *)));
3691 addr_placeholder = gen_reg_rtx (Pmode);
3693 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3694 must be put before this insn, so that they will appear in the right
3695 order (i.e. loop order).
3697 If loop_end is the end of the current function, then emit a
3698 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3700 if (NEXT_INSN (loop_end) != 0)
3701 end_insert_before = NEXT_INSN (loop_end);
3703 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3705 /* Scan through loop to find all possible bivs. */
3707 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
3709 p = next_insn_in_loop (p, scan_start, end, loop_top))
3711 if (GET_CODE (p) == INSN
3712 && (set = single_set (p))
3713 && GET_CODE (SET_DEST (set)) == REG)
3715 dest_reg = SET_DEST (set);
3716 if (REGNO (dest_reg) < max_reg_before_loop
3717 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3718 && REG_IV_TYPE (REGNO (dest_reg)) != NOT_BASIC_INDUCT)
3720 if (basic_induction_var (SET_SRC (set), GET_MODE (SET_SRC (set)),
3721 dest_reg, p, &inc_val, &mult_val,
3724 /* It is a possible basic induction variable.
3725 Create and initialize an induction structure for it. */
3728 = (struct induction *) alloca (sizeof (struct induction));
3730 record_biv (v, p, dest_reg, inc_val, mult_val, location,
3731 not_every_iteration, maybe_multiple);
3732 REG_IV_TYPE (REGNO (dest_reg)) = BASIC_INDUCT;
3734 else if (REGNO (dest_reg) < max_reg_before_loop)
3735 REG_IV_TYPE (REGNO (dest_reg)) = NOT_BASIC_INDUCT;
3739 /* Past CODE_LABEL, we get to insns that may be executed multiple
3740 times. The only way we can be sure that they can't is if every
3741 jump insn between here and the end of the loop either
3742 returns, exits the loop, is a jump to a location that is still
3743 behind the label, or is a jump to the loop start. */
3745 if (GET_CODE (p) == CODE_LABEL)
3753 insn = NEXT_INSN (insn);
3754 if (insn == scan_start)
3762 if (insn == scan_start)
3766 if (GET_CODE (insn) == JUMP_INSN
3767 && GET_CODE (PATTERN (insn)) != RETURN
3768 && (! condjump_p (insn)
3769 || (JUMP_LABEL (insn) != 0
3770 && JUMP_LABEL (insn) != scan_start
3771 && (INSN_UID (JUMP_LABEL (insn)) >= max_uid_for_loop
3772 || (INSN_UID (p) < max_uid_for_loop
3773 ? (INSN_LUID (JUMP_LABEL (insn))
3775 : (INSN_UID (insn) >= max_uid_for_loop
3776 || (INSN_LUID (JUMP_LABEL (insn))
3777 < INSN_LUID (insn))))))))
3785 /* Past a jump, we get to insns for which we can't count
3786 on whether they will be executed during each iteration. */
3787 /* This code appears twice in strength_reduce. There is also similar
3788 code in scan_loop. */
3789 if (GET_CODE (p) == JUMP_INSN
3790 /* If we enter the loop in the middle, and scan around to the
3791 beginning, don't set not_every_iteration for that.
3792 This can be any kind of jump, since we want to know if insns
3793 will be executed if the loop is executed. */
3794 && ! (JUMP_LABEL (p) == loop_top
3795 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3796 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3800 /* If this is a jump outside the loop, then it also doesn't
3801 matter. Check to see if the target of this branch is on the
3802 loop_number_exits_labels list. */
3804 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
3806 label = LABEL_NEXTREF (label))
3807 if (XEXP (label, 0) == JUMP_LABEL (p))
3811 not_every_iteration = 1;
3814 else if (GET_CODE (p) == NOTE)
3816 /* At the virtual top of a converted loop, insns are again known to
3817 be executed each iteration: logically, the loop begins here
3818 even though the exit code has been duplicated. */
3819 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
3820 not_every_iteration = 0;
3821 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3823 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3827 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3828 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3829 or not an insn is known to be executed each iteration of the
3830 loop, whether or not any iterations are known to occur.
3832 Therefore, if we have just passed a label and have no more labels
3833 between here and the test insn of the loop, we know these insns
3834 will be executed each iteration. */
3836 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3837 && no_labels_between_p (p, loop_end))
3838 not_every_iteration = 0;
3841 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3842 Make a sanity check against n_times_set. */
3843 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3845 if (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3846 /* Above happens if register modified by subreg, etc. */
3847 /* Make sure it is not recognized as a basic induction var: */
3848 || VARRAY_INT (n_times_set, bl->regno) != bl->biv_count
3849 /* If never incremented, it is invariant that we decided not to
3850 move. So leave it alone. */
3851 || ! bl->incremented)
3853 if (loop_dump_stream)
3854 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3856 (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3857 ? "not induction variable"
3858 : (! bl->incremented ? "never incremented"
3861 REG_IV_TYPE (bl->regno) = NOT_BASIC_INDUCT;
3868 if (loop_dump_stream)
3869 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3873 /* Exit if there are no bivs. */
3876 /* Can still unroll the loop anyways, but indicate that there is no
3877 strength reduction info available. */
3879 unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
3885 /* Find initial value for each biv by searching backwards from loop_start,
3886 halting at first label. Also record any test condition. */
3889 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3893 if (GET_CODE (p) == CALL_INSN)
3896 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3897 || GET_CODE (p) == CALL_INSN)
3898 note_stores (PATTERN (p), record_initial);
3900 /* Record any test of a biv that branches around the loop if no store
3901 between it and the start of loop. We only care about tests with
3902 constants and registers and only certain of those. */
3903 if (GET_CODE (p) == JUMP_INSN
3904 && JUMP_LABEL (p) != 0
3905 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3906 && (test = get_condition_for_loop (p)) != 0
3907 && GET_CODE (XEXP (test, 0)) == REG
3908 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3909 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3910 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3911 && bl->init_insn == 0)
3913 /* If an NE test, we have an initial value! */
3914 if (GET_CODE (test) == NE)
3917 bl->init_set = gen_rtx_SET (VOIDmode,
3918 XEXP (test, 0), XEXP (test, 1));
3921 bl->initial_test = test;
3925 /* Look at the each biv and see if we can say anything better about its
3926 initial value from any initializing insns set up above. (This is done
3927 in two passes to avoid missing SETs in a PARALLEL.) */
3928 for (backbl = &loop_iv_list; (bl = *backbl); backbl = &bl->next)
3933 if (! bl->init_insn)
3936 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
3937 is a constant, use the value of that. */
3938 if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL
3939 && CONSTANT_P (XEXP (note, 0)))
3940 || ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL
3941 && CONSTANT_P (XEXP (note, 0))))
3942 src = XEXP (note, 0);
3944 src = SET_SRC (bl->init_set);
3946 if (loop_dump_stream)
3947 fprintf (loop_dump_stream,
3948 "Biv %d initialized at insn %d: initial value ",
3949 bl->regno, INSN_UID (bl->init_insn));
3951 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
3952 || GET_MODE (src) == VOIDmode)
3953 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
3955 bl->initial_value = src;
3957 if (loop_dump_stream)
3959 if (GET_CODE (src) == CONST_INT)
3961 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (src));
3962 fputc ('\n', loop_dump_stream);
3966 print_rtl (loop_dump_stream, src);
3967 fprintf (loop_dump_stream, "\n");
3973 struct iv_class *bl2 = 0;
3976 /* Biv initial value is not a simple move. If it is the sum of
3977 another biv and a constant, check if both bivs are incremented
3978 in lockstep. Then we are actually looking at a giv.
3979 For simplicity, we only handle the case where there is but a
3980 single increment, and the register is not used elsewhere. */
3981 if (bl->biv_count == 1
3982 && bl->regno < max_reg_before_loop
3983 && uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
3984 && GET_CODE (src) == PLUS
3985 && GET_CODE (XEXP (src, 0)) == REG
3986 && CONSTANT_P (XEXP (src, 1))
3987 && ((increment = biv_total_increment (bl, loop_start, loop_end))
3990 int regno = REGNO (XEXP (src, 0));
3992 for (bl2 = loop_iv_list; bl2; bl2 = bl2->next)
3993 if (bl2->regno == regno)
3997 /* Now, can we transform this biv into a giv? */
3999 && bl2->biv_count == 1
4000 && rtx_equal_p (increment,
4001 biv_total_increment (bl2, loop_start, loop_end))
4002 /* init_insn is only set to insns that are before loop_start
4003 without any intervening labels. */
4004 && ! reg_set_between_p (bl2->biv->src_reg,
4005 PREV_INSN (bl->init_insn), loop_start)
4006 /* The register from BL2 must be set before the register from
4007 BL is set, or we must be able to move the latter set after
4008 the former set. Currently there can't be any labels
4009 in-between when biv_toal_increment returns nonzero both times
4010 but we test it here in case some day some real cfg analysis
4011 gets used to set always_computable. */
4012 && ((insn_first_p (bl2->biv->insn, bl->biv->insn)
4013 && no_labels_between_p (bl2->biv->insn, bl->biv->insn))
4014 || (! reg_used_between_p (bl->biv->src_reg, bl->biv->insn,
4016 && no_jumps_between_p (bl->biv->insn, bl2->biv->insn)))
4017 && validate_change (bl->biv->insn,
4018 &SET_SRC (single_set (bl->biv->insn)),
4021 int loop_num = uid_loop_num[INSN_UID (loop_start)];
4022 rtx dominator = loop_number_cont_dominator[loop_num];
4023 rtx giv = bl->biv->src_reg;
4024 rtx giv_insn = bl->biv->insn;
4025 rtx after_giv = NEXT_INSN (giv_insn);
4027 if (loop_dump_stream)
4028 fprintf (loop_dump_stream, "is giv of biv %d\n", bl2->regno);
4029 /* Let this giv be discovered by the generic code. */
4030 REG_IV_TYPE (bl->regno) = UNKNOWN_INDUCT;
4031 /* We can get better optimization if we can move the giv setting
4032 before the first giv use. */
4034 && ! reg_set_between_p (bl2->biv->src_reg, loop_start,
4036 && ! reg_used_between_p (giv, loop_start, dominator)
4037 && ! reg_used_between_p (giv, giv_insn, loop_end))
4042 for (next = NEXT_INSN (dominator); ; next = NEXT_INSN (next))
4044 if ((GET_RTX_CLASS (GET_CODE (next)) == 'i'
4045 && (reg_mentioned_p (giv, PATTERN (next))
4046 || reg_set_p (bl2->biv->src_reg, next)))
4047 || GET_CODE (next) == JUMP_INSN)
4050 if (GET_RTX_CLASS (GET_CODE (next)) != 'i'
4051 || ! sets_cc0_p (PATTERN (next)))
4055 if (loop_dump_stream)
4056 fprintf (loop_dump_stream, "move after insn %d\n",
4057 INSN_UID (dominator));
4058 /* Avoid problems with luids by actually moving the insn
4059 and adjusting all luids in the range. */
4060 reorder_insns (giv_insn, giv_insn, dominator);
4061 for (p = dominator; INSN_UID (p) >= max_uid_for_loop; )
4063 compute_luids (giv_insn, after_giv, INSN_LUID (p));
4064 /* If the only purpose of the init insn is to initialize
4065 this giv, delete it. */
4066 if (single_set (bl->init_insn)
4067 && ! reg_used_between_p (giv, bl->init_insn, loop_start))
4068 delete_insn (bl->init_insn);
4070 else if (! insn_first_p (bl2->biv->insn, bl->biv->insn))
4072 rtx p = PREV_INSN (giv_insn);
4073 while (INSN_UID (p) >= max_uid_for_loop)
4075 reorder_insns (giv_insn, giv_insn, bl2->biv->insn);
4076 compute_luids (after_giv, NEXT_INSN (giv_insn),
4079 /* Remove this biv from the chain. */
4089 /* If we can't make it a giv,
4090 let biv keep initial value of "itself". */
4091 else if (loop_dump_stream)
4092 fprintf (loop_dump_stream, "is complex\n");
4096 /* If a biv is unconditionally incremented several times in a row, convert
4097 all but the last increment into a giv. */
4099 /* Get an upper bound for the number of registers
4100 we might have after all bivs have been processed. */
4101 first_increment_giv = max_reg_num ();
4102 for (n_extra_increment = 0, bl = loop_iv_list; bl; bl = bl->next)
4103 n_extra_increment += bl->biv_count - 1;
4104 /* XXX Temporary. */
4105 if (0 && n_extra_increment)
4107 int nregs = first_increment_giv + n_extra_increment;
4109 /* Reallocate reg_iv_type and reg_iv_info. */
4110 VARRAY_GROW (reg_iv_type, nregs);
4111 VARRAY_GROW (reg_iv_info, nregs);
4113 for (bl = loop_iv_list; bl; bl = bl->next)
4115 struct induction **vp, *v, *next;
4117 /* The biv increments lists are in reverse order. Fix this first. */
4118 for (v = bl->biv, bl->biv = 0; v; v = next)
4121 v->next_iv = bl->biv;
4125 for (vp = &bl->biv, next = *vp; v = next, next = v->next_iv;)
4127 HOST_WIDE_INT offset;
4128 rtx set, add_val, old_reg, dest_reg, last_use_insn;
4129 int old_regno, new_regno;
4131 if (! v->always_executed
4132 || v->maybe_multiple
4133 || GET_CODE (v->add_val) != CONST_INT
4134 || ! next->always_executed
4135 || next->maybe_multiple
4136 || ! CONSTANT_P (next->add_val))
4141 offset = INTVAL (v->add_val);
4142 set = single_set (v->insn);
4143 add_val = plus_constant (next->add_val, offset);
4144 old_reg = v->dest_reg;
4145 dest_reg = gen_reg_rtx (v->mode);
4147 /* Unlike reg_iv_type / reg_iv_info, the other three arrays
4148 have been allocated with some slop space, so we may not
4149 actually need to reallocate them. If we do, the following
4150 if statement will be executed just once in this loop. */
4151 if ((unsigned) max_reg_num () > n_times_set->num_elements)
4153 /* Grow all the remaining arrays. */
4154 VARRAY_GROW (set_in_loop, nregs);
4155 VARRAY_GROW (n_times_set, nregs);
4156 VARRAY_GROW (may_not_optimize, nregs);
4159 validate_change (v->insn, &SET_DEST (set), dest_reg, 1);
4160 validate_change (next->insn, next->location, add_val, 1);
4161 if (! apply_change_group ())
4166 next->add_val = add_val;
4167 v->dest_reg = dest_reg;
4168 v->giv_type = DEST_REG;
4169 v->location = &SET_SRC (set);
4171 v->combined_with = 0;
4173 v->derive_adjustment = 0;
4179 v->auto_inc_opt = 0;
4183 v->always_computable = 1;
4184 v->always_executed = 1;
4186 v->no_const_addval = 0;
4188 old_regno = REGNO (old_reg);
4189 new_regno = REGNO (dest_reg);
4190 VARRAY_INT (set_in_loop, old_regno)--;
4191 VARRAY_INT (set_in_loop, new_regno) = 1;
4192 VARRAY_INT (n_times_set, old_regno)--;
4193 VARRAY_INT (n_times_set, new_regno) = 1;
4194 VARRAY_CHAR (may_not_optimize, new_regno) = 0;
4196 REG_IV_TYPE (new_regno) = GENERAL_INDUCT;
4197 REG_IV_INFO (new_regno) = v;
4199 /* Remove the increment from the list of biv increments,
4200 and record it as a giv. */
4203 v->next_iv = bl->giv;
4206 v->benefit = rtx_cost (SET_SRC (set), SET);
4207 bl->total_benefit += v->benefit;
4209 /* Now replace the biv with DEST_REG in all insns between
4210 the replaced increment and the next increment, and
4211 remember the last insn that needed a replacement. */
4212 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4214 p = next_insn_in_loop (p, scan_start, end, loop_top))
4218 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
4220 if (reg_mentioned_p (old_reg, PATTERN (p)))
4223 if (! validate_replace_rtx (old_reg, dest_reg, p))
4226 for (note = REG_NOTES (p); note; note = XEXP (note, 1))
4228 if (GET_CODE (note) == EXPR_LIST)
4230 = replace_rtx (XEXP (note, 0), old_reg, dest_reg);
4234 v->last_use = last_use_insn;
4235 v->lifetime = INSN_LUID (v->insn) - INSN_LUID (last_use_insn);
4236 /* If the lifetime is zero, it means that this register is really
4237 a dead store. So mark this as a giv that can be ignored.
4238 This will not prevent the biv from being eliminated. */
4239 if (v->lifetime == 0)
4244 last_increment_giv = max_reg_num () - 1;
4246 /* Search the loop for general induction variables. */
4248 /* A register is a giv if: it is only set once, it is a function of a
4249 biv and a constant (or invariant), and it is not a biv. */
4251 not_every_iteration = 0;
4257 /* At end of a straight-in loop, we are done.
4258 At end of a loop entered at the bottom, scan the top. */
4259 if (p == scan_start)
4267 if (p == scan_start)
4271 /* Look for a general induction variable in a register. */
4272 if (GET_CODE (p) == INSN
4273 && (set = single_set (p))
4274 && GET_CODE (SET_DEST (set)) == REG
4275 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
4282 rtx last_consec_insn;
4284 dest_reg = SET_DEST (set);
4285 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
4288 if (/* SET_SRC is a giv. */
4289 (general_induction_var (SET_SRC (set), &src_reg, &add_val,
4290 &mult_val, 0, &benefit)
4291 /* Equivalent expression is a giv. */
4292 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
4293 && general_induction_var (XEXP (regnote, 0), &src_reg,
4294 &add_val, &mult_val, 0,
4296 /* Don't try to handle any regs made by loop optimization.
4297 We have nothing on them in regno_first_uid, etc. */
4298 && REGNO (dest_reg) < max_reg_before_loop
4299 /* Don't recognize a BASIC_INDUCT_VAR here. */
4300 && dest_reg != src_reg
4301 /* This must be the only place where the register is set. */
4302 && (VARRAY_INT (n_times_set, REGNO (dest_reg)) == 1
4303 /* or all sets must be consecutive and make a giv. */
4304 || (benefit = consec_sets_giv (benefit, p,
4306 &add_val, &mult_val,
4307 &last_consec_insn))))
4310 = (struct induction *) alloca (sizeof (struct induction));
4312 /* If this is a library call, increase benefit. */
4313 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
4314 benefit += libcall_benefit (p);
4316 /* Skip the consecutive insns, if there are any. */
4317 if (VARRAY_INT (n_times_set, REGNO (dest_reg)) != 1)
4318 p = last_consec_insn;
4320 record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit,
4321 DEST_REG, not_every_iteration, NULL_PTR, loop_start,
4327 #ifndef DONT_REDUCE_ADDR
4328 /* Look for givs which are memory addresses. */
4329 /* This resulted in worse code on a VAX 8600. I wonder if it
4331 if (GET_CODE (p) == INSN)
4332 find_mem_givs (PATTERN (p), p, not_every_iteration, loop_start,
4336 /* Update the status of whether giv can derive other givs. This can
4337 change when we pass a label or an insn that updates a biv. */
4338 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4339 || GET_CODE (p) == CODE_LABEL)
4340 update_giv_derive (p);
4342 /* Past a jump, we get to insns for which we can't count
4343 on whether they will be executed during each iteration. */
4344 /* This code appears twice in strength_reduce. There is also similar
4345 code in scan_loop. */
4346 if (GET_CODE (p) == JUMP_INSN
4347 /* If we enter the loop in the middle, and scan around to the
4348 beginning, don't set not_every_iteration for that.
4349 This can be any kind of jump, since we want to know if insns
4350 will be executed if the loop is executed. */
4351 && ! (JUMP_LABEL (p) == loop_top
4352 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
4353 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
4357 /* If this is a jump outside the loop, then it also doesn't
4358 matter. Check to see if the target of this branch is on the
4359 loop_number_exits_labels list. */
4361 for (label = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
4363 label = LABEL_NEXTREF (label))
4364 if (XEXP (label, 0) == JUMP_LABEL (p))
4368 not_every_iteration = 1;
4371 else if (GET_CODE (p) == NOTE)
4373 /* At the virtual top of a converted loop, insns are again known to
4374 be executed each iteration: logically, the loop begins here
4375 even though the exit code has been duplicated. */
4376 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
4377 not_every_iteration = 0;
4378 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
4380 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
4384 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
4385 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
4386 or not an insn is known to be executed each iteration of the
4387 loop, whether or not any iterations are known to occur.
4389 Therefore, if we have just passed a label and have no more labels
4390 between here and the test insn of the loop, we know these insns
4391 will be executed each iteration. */
4393 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
4394 && no_labels_between_p (p, loop_end))
4395 not_every_iteration = 0;
4398 /* Try to calculate and save the number of loop iterations. This is
4399 set to zero if the actual number can not be calculated. This must
4400 be called after all giv's have been identified, since otherwise it may
4401 fail if the iteration variable is a giv. */
4403 loop_iterations (loop_start, loop_end, loop_info);
4405 /* Now for each giv for which we still don't know whether or not it is
4406 replaceable, check to see if it is replaceable because its final value
4407 can be calculated. This must be done after loop_iterations is called,
4408 so that final_giv_value will work correctly. */
4410 for (bl = loop_iv_list; bl; bl = bl->next)
4412 struct induction *v;
4414 for (v = bl->giv; v; v = v->next_iv)
4415 if (! v->replaceable && ! v->not_replaceable)
4416 check_final_value (v, loop_start, loop_end, loop_info->n_iterations);
4419 /* Try to prove that the loop counter variable (if any) is always
4420 nonnegative; if so, record that fact with a REG_NONNEG note
4421 so that "decrement and branch until zero" insn can be used. */
4422 check_dbra_loop (loop_end, insn_count, loop_start, loop_info);
4424 /* Create reg_map to hold substitutions for replaceable giv regs.
4425 Some givs might have been made from biv increments, so look at
4426 reg_iv_type for a suitable size. */
4427 reg_map_size = reg_iv_type->num_elements;
4428 reg_map = (rtx *) alloca (reg_map_size * sizeof (rtx));
4429 bzero ((char *) reg_map, reg_map_size * sizeof (rtx));
4431 /* Examine each iv class for feasibility of strength reduction/induction
4432 variable elimination. */
4434 for (bl = loop_iv_list; bl; bl = bl->next)
4436 struct induction *v;
4439 rtx final_value = 0;
4442 /* Test whether it will be possible to eliminate this biv
4443 provided all givs are reduced. This is possible if either
4444 the reg is not used outside the loop, or we can compute
4445 what its final value will be.
4447 For architectures with a decrement_and_branch_until_zero insn,
4448 don't do this if we put a REG_NONNEG note on the endtest for
4451 /* Compare against bl->init_insn rather than loop_start.
4452 We aren't concerned with any uses of the biv between
4453 init_insn and loop_start since these won't be affected
4454 by the value of the biv elsewhere in the function, so
4455 long as init_insn doesn't use the biv itself.
4456 March 14, 1989 -- self@bayes.arc.nasa.gov */
4458 if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4460 && INSN_UID (bl->init_insn) < max_uid_for_loop
4461 && uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)
4462 #ifdef HAVE_decrement_and_branch_until_zero
4465 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4466 || ((final_value = final_biv_value (bl, loop_start, loop_end,
4467 loop_info->n_iterations))
4468 #ifdef HAVE_decrement_and_branch_until_zero
4472 bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0,
4473 threshold, insn_count);
4476 if (loop_dump_stream)
4478 fprintf (loop_dump_stream,
4479 "Cannot eliminate biv %d.\n",
4481 fprintf (loop_dump_stream,
4482 "First use: insn %d, last use: insn %d.\n",
4483 REGNO_FIRST_UID (bl->regno),
4484 REGNO_LAST_UID (bl->regno));
4488 /* Combine all giv's for this iv_class. */
4491 /* This will be true at the end, if all givs which depend on this
4492 biv have been strength reduced.
4493 We can't (currently) eliminate the biv unless this is so. */
4496 /* Check each giv in this class to see if we will benefit by reducing
4497 it. Skip giv's combined with others. */
4498 for (v = bl->giv; v; v = v->next_iv)
4500 struct induction *tv;
4502 if (v->ignore || v->same)
4505 benefit = v->benefit;
4507 /* Reduce benefit if not replaceable, since we will insert
4508 a move-insn to replace the insn that calculates this giv.
4509 Don't do this unless the giv is a user variable, since it
4510 will often be marked non-replaceable because of the duplication
4511 of the exit code outside the loop. In such a case, the copies
4512 we insert are dead and will be deleted. So they don't have
4513 a cost. Similar situations exist. */
4514 /* ??? The new final_[bg]iv_value code does a much better job
4515 of finding replaceable giv's, and hence this code may no longer
4517 if (! v->replaceable && ! bl->eliminable
4518 && REG_USERVAR_P (v->dest_reg))
4519 benefit -= copy_cost;
4521 /* Decrease the benefit to count the add-insns that we will
4522 insert to increment the reduced reg for the giv. */
4523 benefit -= add_cost * bl->biv_count;
4525 /* Decide whether to strength-reduce this giv or to leave the code
4526 unchanged (recompute it from the biv each time it is used).
4527 This decision can be made independently for each giv. */
4530 /* Attempt to guess whether autoincrement will handle some of the
4531 new add insns; if so, increase BENEFIT (undo the subtraction of
4532 add_cost that was done above). */
4533 if (v->giv_type == DEST_ADDR
4534 && GET_CODE (v->mult_val) == CONST_INT)
4536 if (HAVE_POST_INCREMENT
4537 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4538 benefit += add_cost * bl->biv_count;
4539 else if (HAVE_PRE_INCREMENT
4540 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4541 benefit += add_cost * bl->biv_count;
4542 else if (HAVE_POST_DECREMENT
4543 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4544 benefit += add_cost * bl->biv_count;
4545 else if (HAVE_PRE_DECREMENT
4546 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4547 benefit += add_cost * bl->biv_count;
4551 /* If an insn is not to be strength reduced, then set its ignore
4552 flag, and clear all_reduced. */
4554 /* A giv that depends on a reversed biv must be reduced if it is
4555 used after the loop exit, otherwise, it would have the wrong
4556 value after the loop exit. To make it simple, just reduce all
4557 of such giv's whether or not we know they are used after the loop
4560 if ( ! flag_reduce_all_givs && v->lifetime * threshold * benefit < insn_count
4563 if (loop_dump_stream)
4564 fprintf (loop_dump_stream,
4565 "giv of insn %d not worth while, %d vs %d.\n",
4567 v->lifetime * threshold * benefit, insn_count);
4573 /* Check that we can increment the reduced giv without a
4574 multiply insn. If not, reject it. */
4576 for (tv = bl->biv; tv; tv = tv->next_iv)
4577 if (tv->mult_val == const1_rtx
4578 && ! product_cheap_p (tv->add_val, v->mult_val))
4580 if (loop_dump_stream)
4581 fprintf (loop_dump_stream,
4582 "giv of insn %d: would need a multiply.\n",
4583 INSN_UID (v->insn));
4592 /* XXX Temporary. */
4593 /* Now that we know which givs will be reduced, try to rearrange the
4594 combinations to reduce register pressure.
4595 recombine_givs calls find_life_end, which needs reg_iv_type and
4596 reg_iv_info to be valid for all pseudos. We do the necessary
4597 reallocation here since it allows to check if there are still
4598 more bivs to process. */
4599 nregs = max_reg_num ();
4600 if (nregs > reg_iv_type->num_elements)
4602 /* If there are still more bivs to process, allocate some slack
4603 space so that we're not constantly reallocating these arrays. */
4606 /* Reallocate reg_iv_type and reg_iv_info. */
4607 VARRAY_GROW (reg_iv_type, nregs);
4608 VARRAY_GROW (reg_iv_info, nregs);
4610 recombine_givs (bl, loop_start, loop_end, unroll_p);
4613 /* Reduce each giv that we decided to reduce. */
4615 for (v = bl->giv; v; v = v->next_iv)
4617 struct induction *tv;
4618 if (! v->ignore && v->same == 0)
4620 int auto_inc_opt = 0;
4622 v->new_reg = gen_reg_rtx (v->mode);
4627 = replace_rtx (PATTERN (v->insn), v->dest_reg, v->new_reg);
4628 if (bl->biv_count != 1)
4630 /* For each place where the biv is incremented, add an
4631 insn to set the new, reduced reg for the giv. */
4632 for (tv = bl->biv; tv; tv = tv->next_iv)
4634 /* We always emit reduced giv increments before the
4635 biv increment when bl->biv_count != 1. So by
4636 emitting the add insns for derived givs after the
4637 biv increment, they pick up the updated value of
4639 emit_insn_after (copy_rtx (PATTERN (v->insn)),
4648 /* If the target has auto-increment addressing modes, and
4649 this is an address giv, then try to put the increment
4650 immediately after its use, so that flow can create an
4651 auto-increment addressing mode. */
4652 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
4653 && bl->biv->always_executed && ! bl->biv->maybe_multiple
4654 /* We don't handle reversed biv's because bl->biv->insn
4655 does not have a valid INSN_LUID. */
4657 && v->always_executed && ! v->maybe_multiple
4658 && INSN_UID (v->insn) < max_uid_for_loop)
4660 /* If other giv's have been combined with this one, then
4661 this will work only if all uses of the other giv's occur
4662 before this giv's insn. This is difficult to check.
4664 We simplify this by looking for the common case where
4665 there is one DEST_REG giv, and this giv's insn is the
4666 last use of the dest_reg of that DEST_REG giv. If the
4667 increment occurs after the address giv, then we can
4668 perform the optimization. (Otherwise, the increment
4669 would have to go before other_giv, and we would not be
4670 able to combine it with the address giv to get an
4671 auto-inc address.) */
4672 if (v->combined_with)
4674 struct induction *other_giv = 0;
4676 for (tv = bl->giv; tv; tv = tv->next_iv)
4684 if (! tv && other_giv
4685 && REGNO (other_giv->dest_reg) < max_reg_before_loop
4686 && (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
4687 == INSN_UID (v->insn))
4688 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
4691 /* Check for case where increment is before the address
4692 giv. Do this test in "loop order". */
4693 else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn)
4694 && (INSN_LUID (v->insn) < INSN_LUID (scan_start)
4695 || (INSN_LUID (bl->biv->insn)
4696 > INSN_LUID (scan_start))))
4697 || (INSN_LUID (v->insn) < INSN_LUID (scan_start)
4698 && (INSN_LUID (scan_start)
4699 < INSN_LUID (bl->biv->insn))))
4708 /* We can't put an insn immediately after one setting
4709 cc0, or immediately before one using cc0. */
4710 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
4711 || (auto_inc_opt == -1
4712 && (prev = prev_nonnote_insn (v->insn)) != 0
4713 && GET_RTX_CLASS (GET_CODE (prev)) == 'i'
4714 && sets_cc0_p (PATTERN (prev))))
4720 v->auto_inc_opt = 1;
4724 /* For each place where the biv is incremented, add an insn
4725 to increment the new, reduced reg for the giv. */
4726 for (tv = bl->biv; tv; tv = tv->next_iv)
4731 insert_before = tv->insn;
4732 else if (auto_inc_opt == 1)
4733 insert_before = NEXT_INSN (v->insn);
4735 insert_before = v->insn;
4737 if (tv->mult_val == const1_rtx)
4738 emit_iv_add_mult (tv->add_val, v->mult_val,
4739 v->new_reg, v->new_reg, insert_before);
4740 else /* tv->mult_val == const0_rtx */
4741 /* A multiply is acceptable here
4742 since this is presumed to be seldom executed. */
4743 emit_iv_add_mult (tv->add_val, v->mult_val,
4744 v->add_val, v->new_reg, insert_before);
4747 /* Add code at loop start to initialize giv's reduced reg. */
4749 emit_iv_add_mult (bl->initial_value, v->mult_val,
4750 v->add_val, v->new_reg, loop_start);
4754 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4757 For each giv register that can be reduced now: if replaceable,
4758 substitute reduced reg wherever the old giv occurs;
4759 else add new move insn "giv_reg = reduced_reg".
4761 Also check for givs whose first use is their definition and whose
4762 last use is the definition of another giv. If so, it is likely
4763 dead and should not be used to eliminate a biv. */
4764 for (v = bl->giv; v; v = v->next_iv)
4766 if (v->same && v->same->ignore)
4774 struct induction *v1;
4776 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4777 if (v->last_use == v1->insn)
4780 else if (v->giv_type == DEST_REG
4781 && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
4783 struct induction *v1;
4785 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4786 if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
4790 /* Update expression if this was combined, in case other giv was
4793 v->new_reg = replace_rtx (v->new_reg,
4794 v->same->dest_reg, v->same->new_reg);
4796 if (v->giv_type == DEST_ADDR)
4797 /* Store reduced reg as the address in the memref where we found
4799 validate_change (v->insn, v->location, v->new_reg, 0);
4800 else if (v->replaceable)
4802 reg_map[REGNO (v->dest_reg)] = v->new_reg;
4805 /* I can no longer duplicate the original problem. Perhaps
4806 this is unnecessary now? */
4808 /* Replaceable; it isn't strictly necessary to delete the old
4809 insn and emit a new one, because v->dest_reg is now dead.
4811 However, especially when unrolling loops, the special
4812 handling for (set REG0 REG1) in the second cse pass may
4813 make v->dest_reg live again. To avoid this problem, emit
4814 an insn to set the original giv reg from the reduced giv.
4815 We can not delete the original insn, since it may be part
4816 of a LIBCALL, and the code in flow that eliminates dead
4817 libcalls will fail if it is deleted. */
4818 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4824 /* Not replaceable; emit an insn to set the original giv reg from
4825 the reduced giv, same as above. */
4826 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4830 /* When a loop is reversed, givs which depend on the reversed
4831 biv, and which are live outside the loop, must be set to their
4832 correct final value. This insn is only needed if the giv is
4833 not replaceable. The correct final value is the same as the
4834 value that the giv starts the reversed loop with. */
4835 if (bl->reversed && ! v->replaceable)
4836 emit_iv_add_mult (bl->initial_value, v->mult_val,
4837 v->add_val, v->dest_reg, end_insert_before);
4838 else if (v->final_value)
4842 /* If the loop has multiple exits, emit the insn before the
4843 loop to ensure that it will always be executed no matter
4844 how the loop exits. Otherwise, emit the insn after the loop,
4845 since this is slightly more efficient. */
4846 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4847 insert_before = loop_start;
4849 insert_before = end_insert_before;
4850 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
4854 /* If the insn to set the final value of the giv was emitted
4855 before the loop, then we must delete the insn inside the loop
4856 that sets it. If this is a LIBCALL, then we must delete
4857 every insn in the libcall. Note, however, that
4858 final_giv_value will only succeed when there are multiple
4859 exits if the giv is dead at each exit, hence it does not
4860 matter that the original insn remains because it is dead
4862 /* Delete the insn inside the loop that sets the giv since
4863 the giv is now set before (or after) the loop. */
4864 delete_insn (v->insn);
4868 if (loop_dump_stream)
4870 fprintf (loop_dump_stream, "giv at %d reduced to ",
4871 INSN_UID (v->insn));
4872 print_rtl (loop_dump_stream, v->new_reg);
4873 fprintf (loop_dump_stream, "\n");
4877 /* All the givs based on the biv bl have been reduced if they
4880 /* For each giv not marked as maybe dead that has been combined with a
4881 second giv, clear any "maybe dead" mark on that second giv.
4882 v->new_reg will either be or refer to the register of the giv it
4885 Doing this clearing avoids problems in biv elimination where a
4886 giv's new_reg is a complex value that can't be put in the insn but
4887 the giv combined with (with a reg as new_reg) is marked maybe_dead.
4888 Since the register will be used in either case, we'd prefer it be
4889 used from the simpler giv. */
4891 for (v = bl->giv; v; v = v->next_iv)
4892 if (! v->maybe_dead && v->same)
4893 v->same->maybe_dead = 0;
4895 /* Try to eliminate the biv, if it is a candidate.
4896 This won't work if ! all_reduced,
4897 since the givs we planned to use might not have been reduced.
4899 We have to be careful that we didn't initially think we could eliminate
4900 this biv because of a giv that we now think may be dead and shouldn't
4901 be used as a biv replacement.
4903 Also, there is the possibility that we may have a giv that looks
4904 like it can be used to eliminate a biv, but the resulting insn
4905 isn't valid. This can happen, for example, on the 88k, where a
4906 JUMP_INSN can compare a register only with zero. Attempts to
4907 replace it with a compare with a constant will fail.
4909 Note that in cases where this call fails, we may have replaced some
4910 of the occurrences of the biv with a giv, but no harm was done in
4911 doing so in the rare cases where it can occur. */
4913 if (all_reduced == 1 && bl->eliminable
4914 && maybe_eliminate_biv (bl, loop_start, end, 1,
4915 threshold, insn_count))
4918 /* ?? If we created a new test to bypass the loop entirely,
4919 or otherwise drop straight in, based on this test, then
4920 we might want to rewrite it also. This way some later
4921 pass has more hope of removing the initialization of this
4924 /* If final_value != 0, then the biv may be used after loop end
4925 and we must emit an insn to set it just in case.
4927 Reversed bivs already have an insn after the loop setting their
4928 value, so we don't need another one. We can't calculate the
4929 proper final value for such a biv here anyways. */
4930 if (final_value != 0 && ! bl->reversed)
4934 /* If the loop has multiple exits, emit the insn before the
4935 loop to ensure that it will always be executed no matter
4936 how the loop exits. Otherwise, emit the insn after the
4937 loop, since this is slightly more efficient. */
4938 if (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
4939 insert_before = loop_start;
4941 insert_before = end_insert_before;
4943 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
4948 /* Delete all of the instructions inside the loop which set
4949 the biv, as they are all dead. If is safe to delete them,
4950 because an insn setting a biv will never be part of a libcall. */
4951 /* However, deleting them will invalidate the regno_last_uid info,
4952 so keeping them around is more convenient. Final_biv_value
4953 will only succeed when there are multiple exits if the biv
4954 is dead at each exit, hence it does not matter that the original
4955 insn remains, because it is dead anyways. */
4956 for (v = bl->biv; v; v = v->next_iv)
4957 delete_insn (v->insn);
4960 if (loop_dump_stream)
4961 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
4966 /* Go through all the instructions in the loop, making all the
4967 register substitutions scheduled in REG_MAP. */
4969 for (p = loop_start; p != end; p = NEXT_INSN (p))
4970 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4971 || GET_CODE (p) == CALL_INSN)
4973 replace_regs (PATTERN (p), reg_map, reg_map_size, 0);
4974 replace_regs (REG_NOTES (p), reg_map, reg_map_size, 0);
4978 /* Unroll loops from within strength reduction so that we can use the
4979 induction variable information that strength_reduce has already
4983 unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
4986 #ifdef HAVE_decrement_and_branch_on_count
4987 /* Instrument the loop with BCT insn. */
4988 if (HAVE_decrement_and_branch_on_count && bct_p
4989 && flag_branch_on_count_reg)
4990 insert_bct (loop_start, loop_end, loop_info);
4991 #endif /* HAVE_decrement_and_branch_on_count */
4993 if (loop_dump_stream)
4994 fprintf (loop_dump_stream, "\n");
4995 VARRAY_FREE (reg_iv_type);
4996 VARRAY_FREE (reg_iv_info);
4999 /* Return 1 if X is a valid source for an initial value (or as value being
5000 compared against in an initial test).
5002 X must be either a register or constant and must not be clobbered between
5003 the current insn and the start of the loop.
5005 INSN is the insn containing X. */
5008 valid_initial_value_p (x, insn, call_seen, loop_start)
5017 /* Only consider pseudos we know about initialized in insns whose luids
5019 if (GET_CODE (x) != REG
5020 || REGNO (x) >= max_reg_before_loop)
5023 /* Don't use call-clobbered registers across a call which clobbers it. On
5024 some machines, don't use any hard registers at all. */
5025 if (REGNO (x) < FIRST_PSEUDO_REGISTER
5026 && (SMALL_REGISTER_CLASSES
5027 || (call_used_regs[REGNO (x)] && call_seen)))
5030 /* Don't use registers that have been clobbered before the start of the
5032 if (reg_set_between_p (x, insn, loop_start))
5038 /* Scan X for memory refs and check each memory address
5039 as a possible giv. INSN is the insn whose pattern X comes from.
5040 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
5041 every loop iteration. */
5044 find_mem_givs (x, insn, not_every_iteration, loop_start, loop_end)
5047 int not_every_iteration;
5048 rtx loop_start, loop_end;
5051 register enum rtx_code code;
5057 code = GET_CODE (x);
5081 /* This code used to disable creating GIVs with mult_val == 1 and
5082 add_val == 0. However, this leads to lost optimizations when
5083 it comes time to combine a set of related DEST_ADDR GIVs, since
5084 this one would not be seen. */
5086 if (general_induction_var (XEXP (x, 0), &src_reg, &add_val,
5087 &mult_val, 1, &benefit))
5089 /* Found one; record it. */
5091 = (struct induction *) oballoc (sizeof (struct induction));
5093 record_giv (v, insn, src_reg, addr_placeholder, mult_val,
5094 add_val, benefit, DEST_ADDR, not_every_iteration,
5095 &XEXP (x, 0), loop_start, loop_end);
5097 v->mem_mode = GET_MODE (x);
5106 /* Recursively scan the subexpressions for other mem refs. */
5108 fmt = GET_RTX_FORMAT (code);
5109 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5111 find_mem_givs (XEXP (x, i), insn, not_every_iteration, loop_start,
5113 else if (fmt[i] == 'E')
5114 for (j = 0; j < XVECLEN (x, i); j++)
5115 find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration,
5116 loop_start, loop_end);
5119 /* Fill in the data about one biv update.
5120 V is the `struct induction' in which we record the biv. (It is
5121 allocated by the caller, with alloca.)
5122 INSN is the insn that sets it.
5123 DEST_REG is the biv's reg.
5125 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
5126 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
5127 being set to INC_VAL.
5129 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
5130 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
5131 can be executed more than once per iteration. If MAYBE_MULTIPLE
5132 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
5133 executed exactly once per iteration. */
5136 record_biv (v, insn, dest_reg, inc_val, mult_val, location,
5137 not_every_iteration, maybe_multiple)
5138 struct induction *v;
5144 int not_every_iteration;
5147 struct iv_class *bl;
5150 v->src_reg = dest_reg;
5151 v->dest_reg = dest_reg;
5152 v->mult_val = mult_val;
5153 v->add_val = inc_val;
5154 v->location = location;
5155 v->mode = GET_MODE (dest_reg);
5156 v->always_computable = ! not_every_iteration;
5157 v->always_executed = ! not_every_iteration;
5158 v->maybe_multiple = maybe_multiple;
5160 /* Add this to the reg's iv_class, creating a class
5161 if this is the first incrementation of the reg. */
5163 bl = reg_biv_class[REGNO (dest_reg)];
5166 /* Create and initialize new iv_class. */
5168 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
5170 bl->regno = REGNO (dest_reg);
5176 /* Set initial value to the reg itself. */
5177 bl->initial_value = dest_reg;
5178 /* We haven't seen the initializing insn yet */
5181 bl->initial_test = 0;
5182 bl->incremented = 0;
5186 bl->total_benefit = 0;
5188 /* Add this class to loop_iv_list. */
5189 bl->next = loop_iv_list;
5192 /* Put it in the array of biv register classes. */
5193 reg_biv_class[REGNO (dest_reg)] = bl;
5196 /* Update IV_CLASS entry for this biv. */
5197 v->next_iv = bl->biv;
5200 if (mult_val == const1_rtx)
5201 bl->incremented = 1;
5203 if (loop_dump_stream)
5205 fprintf (loop_dump_stream,
5206 "Insn %d: possible biv, reg %d,",
5207 INSN_UID (insn), REGNO (dest_reg));
5208 if (GET_CODE (inc_val) == CONST_INT)
5210 fprintf (loop_dump_stream, " const =");
5211 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (inc_val));
5212 fputc ('\n', loop_dump_stream);
5216 fprintf (loop_dump_stream, " const = ");
5217 print_rtl (loop_dump_stream, inc_val);
5218 fprintf (loop_dump_stream, "\n");
5223 /* Fill in the data about one giv.
5224 V is the `struct induction' in which we record the giv. (It is
5225 allocated by the caller, with alloca.)
5226 INSN is the insn that sets it.
5227 BENEFIT estimates the savings from deleting this insn.
5228 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
5229 into a register or is used as a memory address.
5231 SRC_REG is the biv reg which the giv is computed from.
5232 DEST_REG is the giv's reg (if the giv is stored in a reg).
5233 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
5234 LOCATION points to the place where this giv's value appears in INSN. */
5237 record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
5238 type, not_every_iteration, location, loop_start, loop_end)
5239 struct induction *v;
5243 rtx mult_val, add_val;
5246 int not_every_iteration;
5248 rtx loop_start, loop_end;
5250 struct induction *b;
5251 struct iv_class *bl;
5252 rtx set = single_set (insn);
5255 v->src_reg = src_reg;
5257 v->dest_reg = dest_reg;
5258 v->mult_val = mult_val;
5259 v->add_val = add_val;
5260 v->benefit = benefit;
5261 v->location = location;
5263 v->combined_with = 0;
5264 v->maybe_multiple = 0;
5266 v->derive_adjustment = 0;
5272 v->auto_inc_opt = 0;
5278 /* The v->always_computable field is used in update_giv_derive, to
5279 determine whether a giv can be used to derive another giv. For a
5280 DEST_REG giv, INSN computes a new value for the giv, so its value
5281 isn't computable if INSN insn't executed every iteration.
5282 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
5283 it does not compute a new value. Hence the value is always computable
5284 regardless of whether INSN is executed each iteration. */
5286 if (type == DEST_ADDR)
5287 v->always_computable = 1;
5289 v->always_computable = ! not_every_iteration;
5291 v->always_executed = ! not_every_iteration;
5293 if (type == DEST_ADDR)
5295 v->mode = GET_MODE (*location);
5298 else /* type == DEST_REG */
5300 v->mode = GET_MODE (SET_DEST (set));
5302 v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5303 - uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]);
5305 /* If the lifetime is zero, it means that this register is
5306 really a dead store. So mark this as a giv that can be
5307 ignored. This will not prevent the biv from being eliminated. */
5308 if (v->lifetime == 0)
5311 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
5312 REG_IV_INFO (REGNO (dest_reg)) = v;
5315 /* Add the giv to the class of givs computed from one biv. */
5317 bl = reg_biv_class[REGNO (src_reg)];
5320 v->next_iv = bl->giv;
5322 /* Don't count DEST_ADDR. This is supposed to count the number of
5323 insns that calculate givs. */
5324 if (type == DEST_REG)
5326 bl->total_benefit += benefit;
5329 /* Fatal error, biv missing for this giv? */
5332 if (type == DEST_ADDR)
5336 /* The giv can be replaced outright by the reduced register only if all
5337 of the following conditions are true:
5338 - the insn that sets the giv is always executed on any iteration
5339 on which the giv is used at all
5340 (there are two ways to deduce this:
5341 either the insn is executed on every iteration,
5342 or all uses follow that insn in the same basic block),
5343 - the giv is not used outside the loop
5344 - no assignments to the biv occur during the giv's lifetime. */
5346 if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
5347 /* Previous line always fails if INSN was moved by loop opt. */
5348 && uid_luid[REGNO_LAST_UID (REGNO (dest_reg))] < INSN_LUID (loop_end)
5349 && (! not_every_iteration
5350 || last_use_this_basic_block (dest_reg, insn)))
5352 /* Now check that there are no assignments to the biv within the
5353 giv's lifetime. This requires two separate checks. */
5355 /* Check each biv update, and fail if any are between the first
5356 and last use of the giv.
5358 If this loop contains an inner loop that was unrolled, then
5359 the insn modifying the biv may have been emitted by the loop
5360 unrolling code, and hence does not have a valid luid. Just
5361 mark the biv as not replaceable in this case. It is not very
5362 useful as a biv, because it is used in two different loops.
5363 It is very unlikely that we would be able to optimize the giv
5364 using this biv anyways. */
5367 for (b = bl->biv; b; b = b->next_iv)
5369 if (INSN_UID (b->insn) >= max_uid_for_loop
5370 || ((uid_luid[INSN_UID (b->insn)]
5371 >= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))])
5372 && (uid_luid[INSN_UID (b->insn)]
5373 <= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))])))
5376 v->not_replaceable = 1;
5381 /* If there are any backwards branches that go from after the
5382 biv update to before it, then this giv is not replaceable. */
5384 for (b = bl->biv; b; b = b->next_iv)
5385 if (back_branch_in_range_p (b->insn, loop_start, loop_end))
5388 v->not_replaceable = 1;
5394 /* May still be replaceable, we don't have enough info here to
5397 v->not_replaceable = 0;
5401 /* Record whether the add_val contains a const_int, for later use by
5406 v->no_const_addval = 1;
5407 if (tem == const0_rtx)
5409 else if (GET_CODE (tem) == CONST_INT)
5410 v->no_const_addval = 0;
5411 else if (GET_CODE (tem) == PLUS)
5415 if (GET_CODE (XEXP (tem, 0)) == PLUS)
5416 tem = XEXP (tem, 0);
5417 else if (GET_CODE (XEXP (tem, 1)) == PLUS)
5418 tem = XEXP (tem, 1);
5422 if (GET_CODE (XEXP (tem, 1)) == CONST_INT)
5423 v->no_const_addval = 0;
5427 if (loop_dump_stream)
5429 if (type == DEST_REG)
5430 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
5431 INSN_UID (insn), REGNO (dest_reg));
5433 fprintf (loop_dump_stream, "Insn %d: dest address",
5436 fprintf (loop_dump_stream, " src reg %d benefit %d",
5437 REGNO (src_reg), v->benefit);
5438 fprintf (loop_dump_stream, " lifetime %d",
5442 fprintf (loop_dump_stream, " replaceable");
5444 if (v->no_const_addval)
5445 fprintf (loop_dump_stream, " ncav");
5447 if (GET_CODE (mult_val) == CONST_INT)
5449 fprintf (loop_dump_stream, " mult ");
5450 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (mult_val));
5454 fprintf (loop_dump_stream, " mult ");
5455 print_rtl (loop_dump_stream, mult_val);
5458 if (GET_CODE (add_val) == CONST_INT)
5460 fprintf (loop_dump_stream, " add ");
5461 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (add_val));
5465 fprintf (loop_dump_stream, " add ");
5466 print_rtl (loop_dump_stream, add_val);
5470 if (loop_dump_stream)
5471 fprintf (loop_dump_stream, "\n");
5476 /* All this does is determine whether a giv can be made replaceable because
5477 its final value can be calculated. This code can not be part of record_giv
5478 above, because final_giv_value requires that the number of loop iterations
5479 be known, and that can not be accurately calculated until after all givs
5480 have been identified. */
5483 check_final_value (v, loop_start, loop_end, n_iterations)
5484 struct induction *v;
5485 rtx loop_start, loop_end;
5486 unsigned HOST_WIDE_INT n_iterations;
5488 struct iv_class *bl;
5489 rtx final_value = 0;
5491 bl = reg_biv_class[REGNO (v->src_reg)];
5493 /* DEST_ADDR givs will never reach here, because they are always marked
5494 replaceable above in record_giv. */
5496 /* The giv can be replaced outright by the reduced register only if all
5497 of the following conditions are true:
5498 - the insn that sets the giv is always executed on any iteration
5499 on which the giv is used at all
5500 (there are two ways to deduce this:
5501 either the insn is executed on every iteration,
5502 or all uses follow that insn in the same basic block),
5503 - its final value can be calculated (this condition is different
5504 than the one above in record_giv)
5505 - no assignments to the biv occur during the giv's lifetime. */
5508 /* This is only called now when replaceable is known to be false. */
5509 /* Clear replaceable, so that it won't confuse final_giv_value. */
5513 if ((final_value = final_giv_value (v, loop_start, loop_end, n_iterations))
5514 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
5516 int biv_increment_seen = 0;
5522 /* When trying to determine whether or not a biv increment occurs
5523 during the lifetime of the giv, we can ignore uses of the variable
5524 outside the loop because final_value is true. Hence we can not
5525 use regno_last_uid and regno_first_uid as above in record_giv. */
5527 /* Search the loop to determine whether any assignments to the
5528 biv occur during the giv's lifetime. Start with the insn
5529 that sets the giv, and search around the loop until we come
5530 back to that insn again.
5532 Also fail if there is a jump within the giv's lifetime that jumps
5533 to somewhere outside the lifetime but still within the loop. This
5534 catches spaghetti code where the execution order is not linear, and
5535 hence the above test fails. Here we assume that the giv lifetime
5536 does not extend from one iteration of the loop to the next, so as
5537 to make the test easier. Since the lifetime isn't known yet,
5538 this requires two loops. See also record_giv above. */
5540 last_giv_use = v->insn;
5546 p = NEXT_INSN (loop_start);
5550 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5551 || GET_CODE (p) == CALL_INSN)
5553 if (biv_increment_seen)
5555 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5558 v->not_replaceable = 1;
5562 else if (reg_set_p (v->src_reg, PATTERN (p)))
5563 biv_increment_seen = 1;
5564 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5569 /* Now that the lifetime of the giv is known, check for branches
5570 from within the lifetime to outside the lifetime if it is still
5580 p = NEXT_INSN (loop_start);
5581 if (p == last_giv_use)
5584 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
5585 && LABEL_NAME (JUMP_LABEL (p))
5586 && ((INSN_UID (JUMP_LABEL (p)) >= max_uid_for_loop)
5587 || (INSN_UID (v->insn) >= max_uid_for_loop)
5588 || (INSN_UID (last_giv_use) >= max_uid_for_loop)
5589 || (INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (v->insn)
5590 && INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start))
5591 || (INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (last_giv_use)
5592 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end))))
5595 v->not_replaceable = 1;
5597 if (loop_dump_stream)
5598 fprintf (loop_dump_stream,
5599 "Found branch outside giv lifetime.\n");
5606 /* If it is replaceable, then save the final value. */
5608 v->final_value = final_value;
5611 if (loop_dump_stream && v->replaceable)
5612 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
5613 INSN_UID (v->insn), REGNO (v->dest_reg));
5616 /* Update the status of whether a giv can derive other givs.
5618 We need to do something special if there is or may be an update to the biv
5619 between the time the giv is defined and the time it is used to derive
5622 In addition, a giv that is only conditionally set is not allowed to
5623 derive another giv once a label has been passed.
5625 The cases we look at are when a label or an update to a biv is passed. */
5628 update_giv_derive (p)
5631 struct iv_class *bl;
5632 struct induction *biv, *giv;
5636 /* Search all IV classes, then all bivs, and finally all givs.
5638 There are three cases we are concerned with. First we have the situation
5639 of a giv that is only updated conditionally. In that case, it may not
5640 derive any givs after a label is passed.
5642 The second case is when a biv update occurs, or may occur, after the
5643 definition of a giv. For certain biv updates (see below) that are
5644 known to occur between the giv definition and use, we can adjust the
5645 giv definition. For others, or when the biv update is conditional,
5646 we must prevent the giv from deriving any other givs. There are two
5647 sub-cases within this case.
5649 If this is a label, we are concerned with any biv update that is done
5650 conditionally, since it may be done after the giv is defined followed by
5651 a branch here (actually, we need to pass both a jump and a label, but
5652 this extra tracking doesn't seem worth it).
5654 If this is a jump, we are concerned about any biv update that may be
5655 executed multiple times. We are actually only concerned about
5656 backward jumps, but it is probably not worth performing the test
5657 on the jump again here.
5659 If this is a biv update, we must adjust the giv status to show that a
5660 subsequent biv update was performed. If this adjustment cannot be done,
5661 the giv cannot derive further givs. */
5663 for (bl = loop_iv_list; bl; bl = bl->next)
5664 for (biv = bl->biv; biv; biv = biv->next_iv)
5665 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
5668 for (giv = bl->giv; giv; giv = giv->next_iv)
5670 /* If cant_derive is already true, there is no point in
5671 checking all of these conditions again. */
5672 if (giv->cant_derive)
5675 /* If this giv is conditionally set and we have passed a label,
5676 it cannot derive anything. */
5677 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
5678 giv->cant_derive = 1;
5680 /* Skip givs that have mult_val == 0, since
5681 they are really invariants. Also skip those that are
5682 replaceable, since we know their lifetime doesn't contain
5684 else if (giv->mult_val == const0_rtx || giv->replaceable)
5687 /* The only way we can allow this giv to derive another
5688 is if this is a biv increment and we can form the product
5689 of biv->add_val and giv->mult_val. In this case, we will
5690 be able to compute a compensation. */
5691 else if (biv->insn == p)
5695 if (biv->mult_val == const1_rtx)
5696 tem = simplify_giv_expr (gen_rtx_MULT (giv->mode,
5701 if (tem && giv->derive_adjustment)
5702 tem = simplify_giv_expr (gen_rtx_PLUS (giv->mode, tem,
5703 giv->derive_adjustment),
5706 giv->derive_adjustment = tem;
5708 giv->cant_derive = 1;
5710 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
5711 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
5712 giv->cant_derive = 1;
5717 /* Check whether an insn is an increment legitimate for a basic induction var.
5718 X is the source of insn P, or a part of it.
5719 MODE is the mode in which X should be interpreted.
5721 DEST_REG is the putative biv, also the destination of the insn.
5722 We accept patterns of these forms:
5723 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
5724 REG = INVARIANT + REG
5726 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
5727 store the additive term into *INC_VAL, and store the place where
5728 we found the additive term into *LOCATION.
5730 If X is an assignment of an invariant into DEST_REG, we set
5731 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
5733 We also want to detect a BIV when it corresponds to a variable
5734 whose mode was promoted via PROMOTED_MODE. In that case, an increment
5735 of the variable may be a PLUS that adds a SUBREG of that variable to
5736 an invariant and then sign- or zero-extends the result of the PLUS
5739 Most GIVs in such cases will be in the promoted mode, since that is the
5740 probably the natural computation mode (and almost certainly the mode
5741 used for addresses) on the machine. So we view the pseudo-reg containing
5742 the variable as the BIV, as if it were simply incremented.
5744 Note that treating the entire pseudo as a BIV will result in making
5745 simple increments to any GIVs based on it. However, if the variable
5746 overflows in its declared mode but not its promoted mode, the result will
5747 be incorrect. This is acceptable if the variable is signed, since
5748 overflows in such cases are undefined, but not if it is unsigned, since
5749 those overflows are defined. So we only check for SIGN_EXTEND and
5752 If we cannot find a biv, we return 0. */
5755 basic_induction_var (x, mode, dest_reg, p, inc_val, mult_val, location)
5757 enum machine_mode mode;
5764 register enum rtx_code code;
5768 code = GET_CODE (x);
5772 if (rtx_equal_p (XEXP (x, 0), dest_reg)
5773 || (GET_CODE (XEXP (x, 0)) == SUBREG
5774 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
5775 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
5777 argp = &XEXP (x, 1);
5779 else if (rtx_equal_p (XEXP (x, 1), dest_reg)
5780 || (GET_CODE (XEXP (x, 1)) == SUBREG
5781 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
5782 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
5784 argp = &XEXP (x, 0);
5790 if (invariant_p (arg) != 1)
5793 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
5794 *mult_val = const1_rtx;
5799 /* If this is a SUBREG for a promoted variable, check the inner
5801 if (SUBREG_PROMOTED_VAR_P (x))
5802 return basic_induction_var (SUBREG_REG (x), GET_MODE (SUBREG_REG (x)),
5803 dest_reg, p, inc_val, mult_val, location);
5807 /* If this register is assigned in a previous insn, look at its
5808 source, but don't go outside the loop or past a label. */
5814 insn = PREV_INSN (insn);
5815 } while (insn && GET_CODE (insn) == NOTE
5816 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5820 set = single_set (insn);
5824 if ((SET_DEST (set) == x
5825 || (GET_CODE (SET_DEST (set)) == SUBREG
5826 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
5828 && SUBREG_REG (SET_DEST (set)) == x))
5829 && basic_induction_var (SET_SRC (set),
5830 (GET_MODE (SET_SRC (set)) == VOIDmode
5832 : GET_MODE (SET_SRC (set))),
5834 inc_val, mult_val, location))
5837 /* ... fall through ... */
5839 /* Can accept constant setting of biv only when inside inner most loop.
5840 Otherwise, a biv of an inner loop may be incorrectly recognized
5841 as a biv of the outer loop,
5842 causing code to be moved INTO the inner loop. */
5844 if (invariant_p (x) != 1)
5849 /* convert_modes aborts if we try to convert to or from CCmode, so just
5850 exclude that case. It is very unlikely that a condition code value
5851 would be a useful iterator anyways. */
5852 if (loops_enclosed == 1
5853 && GET_MODE_CLASS (mode) != MODE_CC
5854 && GET_MODE_CLASS (GET_MODE (dest_reg)) != MODE_CC)
5856 /* Possible bug here? Perhaps we don't know the mode of X. */
5857 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
5858 *mult_val = const0_rtx;
5865 return basic_induction_var (XEXP (x, 0), GET_MODE (XEXP (x, 0)),
5866 dest_reg, p, inc_val, mult_val, location);
5869 /* Similar, since this can be a sign extension. */
5870 for (insn = PREV_INSN (p);
5871 (insn && GET_CODE (insn) == NOTE
5872 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5873 insn = PREV_INSN (insn))
5877 set = single_set (insn);
5879 if (set && SET_DEST (set) == XEXP (x, 0)
5880 && GET_CODE (XEXP (x, 1)) == CONST_INT
5881 && INTVAL (XEXP (x, 1)) >= 0
5882 && GET_CODE (SET_SRC (set)) == ASHIFT
5883 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
5884 return basic_induction_var (XEXP (SET_SRC (set), 0),
5885 GET_MODE (XEXP (x, 0)),
5886 dest_reg, insn, inc_val, mult_val,
5895 /* A general induction variable (giv) is any quantity that is a linear
5896 function of a basic induction variable,
5897 i.e. giv = biv * mult_val + add_val.
5898 The coefficients can be any loop invariant quantity.
5899 A giv need not be computed directly from the biv;
5900 it can be computed by way of other givs. */
5902 /* Determine whether X computes a giv.
5903 If it does, return a nonzero value
5904 which is the benefit from eliminating the computation of X;
5905 set *SRC_REG to the register of the biv that it is computed from;
5906 set *ADD_VAL and *MULT_VAL to the coefficients,
5907 such that the value of X is biv * mult + add; */
5910 general_induction_var (x, src_reg, add_val, mult_val, is_addr, pbenefit)
5921 /* If this is an invariant, forget it, it isn't a giv. */
5922 if (invariant_p (x) == 1)
5925 /* See if the expression could be a giv and get its form.
5926 Mark our place on the obstack in case we don't find a giv. */
5927 storage = (char *) oballoc (0);
5929 x = simplify_giv_expr (x, pbenefit);
5936 switch (GET_CODE (x))
5940 /* Since this is now an invariant and wasn't before, it must be a giv
5941 with MULT_VAL == 0. It doesn't matter which BIV we associate this
5943 *src_reg = loop_iv_list->biv->dest_reg;
5944 *mult_val = const0_rtx;
5949 /* This is equivalent to a BIV. */
5951 *mult_val = const1_rtx;
5952 *add_val = const0_rtx;
5956 /* Either (plus (biv) (invar)) or
5957 (plus (mult (biv) (invar_1)) (invar_2)). */
5958 if (GET_CODE (XEXP (x, 0)) == MULT)
5960 *src_reg = XEXP (XEXP (x, 0), 0);
5961 *mult_val = XEXP (XEXP (x, 0), 1);
5965 *src_reg = XEXP (x, 0);
5966 *mult_val = const1_rtx;
5968 *add_val = XEXP (x, 1);
5972 /* ADD_VAL is zero. */
5973 *src_reg = XEXP (x, 0);
5974 *mult_val = XEXP (x, 1);
5975 *add_val = const0_rtx;
5982 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
5983 unless they are CONST_INT). */
5984 if (GET_CODE (*add_val) == USE)
5985 *add_val = XEXP (*add_val, 0);
5986 if (GET_CODE (*mult_val) == USE)
5987 *mult_val = XEXP (*mult_val, 0);
5992 *pbenefit += ADDRESS_COST (orig_x) - reg_address_cost;
5994 *pbenefit += rtx_cost (orig_x, MEM) - reg_address_cost;
5998 *pbenefit += rtx_cost (orig_x, SET);
6000 /* Always return true if this is a giv so it will be detected as such,
6001 even if the benefit is zero or negative. This allows elimination
6002 of bivs that might otherwise not be eliminated. */
6006 /* Given an expression, X, try to form it as a linear function of a biv.
6007 We will canonicalize it to be of the form
6008 (plus (mult (BIV) (invar_1))
6010 with possible degeneracies.
6012 The invariant expressions must each be of a form that can be used as a
6013 machine operand. We surround then with a USE rtx (a hack, but localized
6014 and certainly unambiguous!) if not a CONST_INT for simplicity in this
6015 routine; it is the caller's responsibility to strip them.
6017 If no such canonicalization is possible (i.e., two biv's are used or an
6018 expression that is neither invariant nor a biv or giv), this routine
6021 For a non-zero return, the result will have a code of CONST_INT, USE,
6022 REG (for a BIV), PLUS, or MULT. No other codes will occur.
6024 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
6026 static rtx sge_plus PROTO ((enum machine_mode, rtx, rtx));
6027 static rtx sge_plus_constant PROTO ((rtx, rtx));
6030 simplify_giv_expr (x, benefit)
6034 enum machine_mode mode = GET_MODE (x);
6038 /* If this is not an integer mode, or if we cannot do arithmetic in this
6039 mode, this can't be a giv. */
6040 if (mode != VOIDmode
6041 && (GET_MODE_CLASS (mode) != MODE_INT
6042 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
6045 switch (GET_CODE (x))
6048 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
6049 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
6050 if (arg0 == 0 || arg1 == 0)
6053 /* Put constant last, CONST_INT last if both constant. */
6054 if ((GET_CODE (arg0) == USE
6055 || GET_CODE (arg0) == CONST_INT)
6056 && ! ((GET_CODE (arg0) == USE
6057 && GET_CODE (arg1) == USE)
6058 || GET_CODE (arg1) == CONST_INT))
6059 tem = arg0, arg0 = arg1, arg1 = tem;
6061 /* Handle addition of zero, then addition of an invariant. */
6062 if (arg1 == const0_rtx)
6064 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
6065 switch (GET_CODE (arg0))
6069 /* Adding two invariants must result in an invariant, so enclose
6070 addition operation inside a USE and return it. */
6071 if (GET_CODE (arg0) == USE)
6072 arg0 = XEXP (arg0, 0);
6073 if (GET_CODE (arg1) == USE)
6074 arg1 = XEXP (arg1, 0);
6076 if (GET_CODE (arg0) == CONST_INT)
6077 tem = arg0, arg0 = arg1, arg1 = tem;
6078 if (GET_CODE (arg1) == CONST_INT)
6079 tem = sge_plus_constant (arg0, arg1);
6081 tem = sge_plus (mode, arg0, arg1);
6083 if (GET_CODE (tem) != CONST_INT)
6084 tem = gen_rtx_USE (mode, tem);
6089 /* biv + invar or mult + invar. Return sum. */
6090 return gen_rtx_PLUS (mode, arg0, arg1);
6093 /* (a + invar_1) + invar_2. Associate. */
6094 return simplify_giv_expr (
6095 gen_rtx_PLUS (mode, XEXP (arg0, 0),
6096 gen_rtx_PLUS (mode, XEXP (arg0, 1), arg1)),
6103 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
6104 MULT to reduce cases. */
6105 if (GET_CODE (arg0) == REG)
6106 arg0 = gen_rtx_MULT (mode, arg0, const1_rtx);
6107 if (GET_CODE (arg1) == REG)
6108 arg1 = gen_rtx_MULT (mode, arg1, const1_rtx);
6110 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
6111 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
6112 Recurse to associate the second PLUS. */
6113 if (GET_CODE (arg1) == MULT)
6114 tem = arg0, arg0 = arg1, arg1 = tem;
6116 if (GET_CODE (arg1) == PLUS)
6117 return simplify_giv_expr (gen_rtx_PLUS (mode,
6118 gen_rtx_PLUS (mode, arg0,
6123 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
6124 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
6127 if (!rtx_equal_p (arg0, arg1))
6130 return simplify_giv_expr (gen_rtx_MULT (mode,
6138 /* Handle "a - b" as "a + b * (-1)". */
6139 return simplify_giv_expr (gen_rtx_PLUS (mode,
6141 gen_rtx_MULT (mode, XEXP (x, 1),
6146 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
6147 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
6148 if (arg0 == 0 || arg1 == 0)
6151 /* Put constant last, CONST_INT last if both constant. */
6152 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
6153 && GET_CODE (arg1) != CONST_INT)
6154 tem = arg0, arg0 = arg1, arg1 = tem;
6156 /* If second argument is not now constant, not giv. */
6157 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
6160 /* Handle multiply by 0 or 1. */
6161 if (arg1 == const0_rtx)
6164 else if (arg1 == const1_rtx)
6167 switch (GET_CODE (arg0))
6170 /* biv * invar. Done. */
6171 return gen_rtx_MULT (mode, arg0, arg1);
6174 /* Product of two constants. */
6175 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
6178 /* invar * invar. It is a giv, but very few of these will
6179 actually pay off, so limit to simple registers. */
6180 if (GET_CODE (arg1) != CONST_INT)
6183 arg0 = XEXP (arg0, 0);
6184 if (GET_CODE (arg0) == REG)
6185 tem = gen_rtx_MULT (mode, arg0, arg1);
6186 else if (GET_CODE (arg0) == MULT
6187 && GET_CODE (XEXP (arg0, 0)) == REG
6188 && GET_CODE (XEXP (arg0, 1)) == CONST_INT)
6190 tem = gen_rtx_MULT (mode, XEXP (arg0, 0),
6191 GEN_INT (INTVAL (XEXP (arg0, 1))
6196 return gen_rtx_USE (mode, tem);
6199 /* (a * invar_1) * invar_2. Associate. */
6200 return simplify_giv_expr (gen_rtx_MULT (mode, XEXP (arg0, 0),
6207 /* (a + invar_1) * invar_2. Distribute. */
6208 return simplify_giv_expr (gen_rtx_PLUS (mode,
6222 /* Shift by constant is multiply by power of two. */
6223 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6226 return simplify_giv_expr (gen_rtx_MULT (mode,
6228 GEN_INT ((HOST_WIDE_INT) 1
6229 << INTVAL (XEXP (x, 1)))),
6233 /* "-a" is "a * (-1)" */
6234 return simplify_giv_expr (gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx),
6238 /* "~a" is "-a - 1". Silly, but easy. */
6239 return simplify_giv_expr (gen_rtx_MINUS (mode,
6240 gen_rtx_NEG (mode, XEXP (x, 0)),
6245 /* Already in proper form for invariant. */
6249 /* If this is a new register, we can't deal with it. */
6250 if (REGNO (x) >= max_reg_before_loop)
6253 /* Check for biv or giv. */
6254 switch (REG_IV_TYPE (REGNO (x)))
6258 case GENERAL_INDUCT:
6260 struct induction *v = REG_IV_INFO (REGNO (x));
6262 /* Form expression from giv and add benefit. Ensure this giv
6263 can derive another and subtract any needed adjustment if so. */
6264 *benefit += v->benefit;
6268 tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode, v->src_reg,
6271 if (v->derive_adjustment)
6272 tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment);
6273 return simplify_giv_expr (tem, benefit);
6277 /* If it isn't an induction variable, and it is invariant, we
6278 may be able to simplify things further by looking through
6279 the bits we just moved outside the loop. */
6280 if (invariant_p (x) == 1)
6284 for (m = the_movables; m ; m = m->next)
6285 if (rtx_equal_p (x, m->set_dest))
6287 /* Ok, we found a match. Substitute and simplify. */
6289 /* If we match another movable, we must use that, as
6290 this one is going away. */
6292 return simplify_giv_expr (m->match->set_dest, benefit);
6294 /* If consec is non-zero, this is a member of a group of
6295 instructions that were moved together. We handle this
6296 case only to the point of seeking to the last insn and
6297 looking for a REG_EQUAL. Fail if we don't find one. */
6302 do { tem = NEXT_INSN (tem); } while (--i > 0);
6304 tem = find_reg_note (tem, REG_EQUAL, NULL_RTX);
6306 tem = XEXP (tem, 0);
6310 tem = single_set (m->insn);
6312 tem = SET_SRC (tem);
6317 /* What we are most interested in is pointer
6318 arithmetic on invariants -- only take
6319 patterns we may be able to do something with. */
6320 if (GET_CODE (tem) == PLUS
6321 || GET_CODE (tem) == MULT
6322 || GET_CODE (tem) == ASHIFT
6323 || GET_CODE (tem) == CONST_INT
6324 || GET_CODE (tem) == SYMBOL_REF)
6326 tem = simplify_giv_expr (tem, benefit);
6330 else if (GET_CODE (tem) == CONST
6331 && GET_CODE (XEXP (tem, 0)) == PLUS
6332 && GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF
6333 && GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT)
6335 tem = simplify_giv_expr (XEXP (tem, 0), benefit);
6346 /* Fall through to general case. */
6348 /* If invariant, return as USE (unless CONST_INT).
6349 Otherwise, not giv. */
6350 if (GET_CODE (x) == USE)
6353 if (invariant_p (x) == 1)
6355 if (GET_CODE (x) == CONST_INT)
6357 if (GET_CODE (x) == CONST
6358 && GET_CODE (XEXP (x, 0)) == PLUS
6359 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
6360 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
6362 return gen_rtx_USE (mode, x);
6369 /* This routine folds invariants such that there is only ever one
6370 CONST_INT in the summation. It is only used by simplify_giv_expr. */
6373 sge_plus_constant (x, c)
6376 if (GET_CODE (x) == CONST_INT)
6377 return GEN_INT (INTVAL (x) + INTVAL (c));
6378 else if (GET_CODE (x) != PLUS)
6379 return gen_rtx_PLUS (GET_MODE (x), x, c);
6380 else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6382 return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
6383 GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c)));
6385 else if (GET_CODE (XEXP (x, 0)) == PLUS
6386 || GET_CODE (XEXP (x, 1)) != PLUS)
6388 return gen_rtx_PLUS (GET_MODE (x),
6389 sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1));
6393 return gen_rtx_PLUS (GET_MODE (x),
6394 sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0));
6399 sge_plus (mode, x, y)
6400 enum machine_mode mode;
6403 while (GET_CODE (y) == PLUS)
6405 rtx a = XEXP (y, 0);
6406 if (GET_CODE (a) == CONST_INT)
6407 x = sge_plus_constant (x, a);
6409 x = gen_rtx_PLUS (mode, x, a);
6412 if (GET_CODE (y) == CONST_INT)
6413 x = sge_plus_constant (x, y);
6415 x = gen_rtx_PLUS (mode, x, y);
6419 /* Help detect a giv that is calculated by several consecutive insns;
6423 The caller has already identified the first insn P as having a giv as dest;
6424 we check that all other insns that set the same register follow
6425 immediately after P, that they alter nothing else,
6426 and that the result of the last is still a giv.
6428 The value is 0 if the reg set in P is not really a giv.
6429 Otherwise, the value is the amount gained by eliminating
6430 all the consecutive insns that compute the value.
6432 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
6433 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
6435 The coefficients of the ultimate giv value are stored in
6436 *MULT_VAL and *ADD_VAL. */
6439 consec_sets_giv (first_benefit, p, src_reg, dest_reg,
6440 add_val, mult_val, last_consec_insn)
6447 rtx *last_consec_insn;
6455 /* Indicate that this is a giv so that we can update the value produced in
6456 each insn of the multi-insn sequence.
6458 This induction structure will be used only by the call to
6459 general_induction_var below, so we can allocate it on our stack.
6460 If this is a giv, our caller will replace the induct var entry with
6461 a new induction structure. */
6463 = (struct induction *) alloca (sizeof (struct induction));
6464 v->src_reg = src_reg;
6465 v->mult_val = *mult_val;
6466 v->add_val = *add_val;
6467 v->benefit = first_benefit;
6469 v->derive_adjustment = 0;
6471 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
6472 REG_IV_INFO (REGNO (dest_reg)) = v;
6474 count = VARRAY_INT (n_times_set, REGNO (dest_reg)) - 1;
6479 code = GET_CODE (p);
6481 /* If libcall, skip to end of call sequence. */
6482 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
6486 && (set = single_set (p))
6487 && GET_CODE (SET_DEST (set)) == REG
6488 && SET_DEST (set) == dest_reg
6489 && (general_induction_var (SET_SRC (set), &src_reg,
6490 add_val, mult_val, 0, &benefit)
6491 /* Giv created by equivalent expression. */
6492 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
6493 && general_induction_var (XEXP (temp, 0), &src_reg,
6494 add_val, mult_val, 0, &benefit)))
6495 && src_reg == v->src_reg)
6497 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
6498 benefit += libcall_benefit (p);
6501 v->mult_val = *mult_val;
6502 v->add_val = *add_val;
6503 v->benefit = benefit;
6505 else if (code != NOTE)
6507 /* Allow insns that set something other than this giv to a
6508 constant. Such insns are needed on machines which cannot
6509 include long constants and should not disqualify a giv. */
6511 && (set = single_set (p))
6512 && SET_DEST (set) != dest_reg
6513 && CONSTANT_P (SET_SRC (set)))
6516 REG_IV_TYPE (REGNO (dest_reg)) = UNKNOWN_INDUCT;
6521 *last_consec_insn = p;
6525 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6526 represented by G1. If no such expression can be found, or it is clear that
6527 it cannot possibly be a valid address, 0 is returned.
6529 To perform the computation, we note that
6532 where `v' is the biv.
6534 So G2 = (y/b) * G1 + (b - a*y/x).
6536 Note that MULT = y/x.
6538 Update: A and B are now allowed to be additive expressions such that
6539 B contains all variables in A. That is, computing B-A will not require
6540 subtracting variables. */
6543 express_from_1 (a, b, mult)
6546 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
6548 if (mult == const0_rtx)
6551 /* If MULT is not 1, we cannot handle A with non-constants, since we
6552 would then be required to subtract multiples of the registers in A.
6553 This is theoretically possible, and may even apply to some Fortran
6554 constructs, but it is a lot of work and we do not attempt it here. */
6556 if (mult != const1_rtx && GET_CODE (a) != CONST_INT)
6559 /* In general these structures are sorted top to bottom (down the PLUS
6560 chain), but not left to right across the PLUS. If B is a higher
6561 order giv than A, we can strip one level and recurse. If A is higher
6562 order, we'll eventually bail out, but won't know that until the end.
6563 If they are the same, we'll strip one level around this loop. */
6565 while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS)
6567 rtx ra, rb, oa, ob, tmp;
6569 ra = XEXP (a, 0), oa = XEXP (a, 1);
6570 if (GET_CODE (ra) == PLUS)
6571 tmp = ra, ra = oa, oa = tmp;
6573 rb = XEXP (b, 0), ob = XEXP (b, 1);
6574 if (GET_CODE (rb) == PLUS)
6575 tmp = rb, rb = ob, ob = tmp;
6577 if (rtx_equal_p (ra, rb))
6578 /* We matched: remove one reg completely. */
6580 else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob))
6581 /* An alternate match. */
6583 else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb))
6584 /* An alternate match. */
6588 /* Indicates an extra register in B. Strip one level from B and
6589 recurse, hoping B was the higher order expression. */
6590 ob = express_from_1 (a, ob, mult);
6593 return gen_rtx_PLUS (GET_MODE (b), rb, ob);
6597 /* Here we are at the last level of A, go through the cases hoping to
6598 get rid of everything but a constant. */
6600 if (GET_CODE (a) == PLUS)
6604 ra = XEXP (a, 0), oa = XEXP (a, 1);
6605 if (rtx_equal_p (oa, b))
6607 else if (!rtx_equal_p (ra, b))
6610 if (GET_CODE (oa) != CONST_INT)
6613 return GEN_INT (-INTVAL (oa) * INTVAL (mult));
6615 else if (GET_CODE (a) == CONST_INT)
6617 return plus_constant (b, -INTVAL (a) * INTVAL (mult));
6619 else if (GET_CODE (b) == PLUS)
6621 if (rtx_equal_p (a, XEXP (b, 0)))
6623 else if (rtx_equal_p (a, XEXP (b, 1)))
6628 else if (rtx_equal_p (a, b))
6635 express_from (g1, g2)
6636 struct induction *g1, *g2;
6640 /* The value that G1 will be multiplied by must be a constant integer. Also,
6641 the only chance we have of getting a valid address is if b*c/a (see above
6642 for notation) is also an integer. */
6643 if (GET_CODE (g1->mult_val) == CONST_INT
6644 && GET_CODE (g2->mult_val) == CONST_INT)
6646 if (g1->mult_val == const0_rtx
6647 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
6649 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
6651 else if (rtx_equal_p (g1->mult_val, g2->mult_val))
6655 /* ??? Find out if the one is a multiple of the other? */
6659 add = express_from_1 (g1->add_val, g2->add_val, mult);
6660 if (add == NULL_RTX)
6663 /* Form simplified final result. */
6664 if (mult == const0_rtx)
6666 else if (mult == const1_rtx)
6667 mult = g1->dest_reg;
6669 mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult);
6671 if (add == const0_rtx)
6675 if (GET_CODE (add) == PLUS
6676 && CONSTANT_P (XEXP (add, 1)))
6678 rtx tem = XEXP (add, 1);
6679 mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0));
6683 return gen_rtx_PLUS (g2->mode, mult, add);
6688 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6689 represented by G1. This indicates that G2 should be combined with G1 and
6690 that G2 can use (either directly or via an address expression) a register
6691 used to represent G1. */
6694 combine_givs_p (g1, g2)
6695 struct induction *g1, *g2;
6697 rtx tem = express_from (g1, g2);
6699 /* If these givs are identical, they can be combined. We use the results
6700 of express_from because the addends are not in a canonical form, so
6701 rtx_equal_p is a weaker test. */
6702 /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
6703 combination to be the other way round. */
6704 if (tem == g1->dest_reg
6705 && (g1->giv_type == DEST_REG || g2->giv_type == DEST_ADDR))
6707 return g1->dest_reg;
6710 /* If G2 can be expressed as a function of G1 and that function is valid
6711 as an address and no more expensive than using a register for G2,
6712 the expression of G2 in terms of G1 can be used. */
6714 && g2->giv_type == DEST_ADDR
6715 && memory_address_p (g2->mem_mode, tem)
6716 /* ??? Looses, especially with -fforce-addr, where *g2->location
6717 will always be a register, and so anything more complicated
6721 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location)
6723 && rtx_cost (tem, MEM) <= rtx_cost (*g2->location, MEM)
6734 struct combine_givs_stats
6741 cmp_combine_givs_stats (x, y)
6742 struct combine_givs_stats *x, *y;
6745 d = y->total_benefit - x->total_benefit;
6746 /* Stabilize the sort. */
6748 d = x->giv_number - y->giv_number;
6752 /* If one of these givs is a DEST_REG that was used by the other giv,
6753 this is actually a single use. Return 0 if this is not
6754 the case, -1 if g1 is the DEST_REG involved, and 1 if it was g2. */
6757 combine_givs_used_by_other (g1, g2)
6758 struct induction *g1, *g2;
6760 if (g1->giv_type == DEST_REG
6761 && reg_mentioned_p (g1->dest_reg, PATTERN (g2->insn)))
6764 if (g2->giv_type == DEST_REG
6765 && reg_mentioned_p (g2->dest_reg, PATTERN (g1->insn)))
6772 combine_givs_benefit_from (g1, g2)
6773 struct induction *g1, *g2;
6775 int tmp = combine_givs_used_by_other (g1, g2);
6779 return g2->benefit - g1->benefit;
6784 /* Check all pairs of givs for iv_class BL and see if any can be combined with
6785 any other. If so, point SAME to the giv combined with and set NEW_REG to
6786 be an expression (in terms of the other giv's DEST_REG) equivalent to the
6787 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
6791 struct iv_class *bl;
6793 struct induction *g1, *g2, **giv_array;
6794 int i, j, k, giv_count;
6795 struct combine_givs_stats *stats;
6798 /* Count givs, because bl->giv_count is incorrect here. */
6800 for (g1 = bl->giv; g1; g1 = g1->next_iv)
6805 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
6807 for (g1 = bl->giv; g1; g1 = g1->next_iv)
6809 giv_array[i++] = g1;
6811 stats = (struct combine_givs_stats *) alloca (giv_count * sizeof (*stats));
6812 bzero ((char *) stats, giv_count * sizeof (*stats));
6814 can_combine = (rtx *) alloca (giv_count * giv_count * sizeof(rtx));
6815 bzero ((char *) can_combine, giv_count * giv_count * sizeof(rtx));
6817 for (i = 0; i < giv_count; i++)
6823 this_benefit = g1->benefit;
6824 /* Add an additional weight for zero addends. */
6825 if (g1->no_const_addval)
6827 for (j = 0; j < giv_count; j++)
6833 && (this_combine = combine_givs_p (g1, g2)) != NULL_RTX)
6835 can_combine[i*giv_count + j] = this_combine;
6836 this_benefit += combine_givs_benefit_from (g1, g2);
6837 /* Add an additional weight for being reused more times. */
6841 stats[i].giv_number = i;
6842 stats[i].total_benefit = this_benefit;
6845 /* Iterate, combining until we can't. */
6847 qsort (stats, giv_count, sizeof(*stats), cmp_combine_givs_stats);
6849 if (loop_dump_stream)
6851 fprintf (loop_dump_stream, "Sorted combine statistics:\n");
6852 for (k = 0; k < giv_count; k++)
6854 g1 = giv_array[stats[k].giv_number];
6855 if (!g1->combined_with && !g1->same)
6856 fprintf (loop_dump_stream, " {%d, %d}",
6857 INSN_UID (giv_array[stats[k].giv_number]->insn),
6858 stats[k].total_benefit);
6860 putc ('\n', loop_dump_stream);
6863 for (k = 0; k < giv_count; k++)
6865 int g1_add_benefit = 0;
6867 i = stats[k].giv_number;
6870 /* If it has already been combined, skip. */
6871 if (g1->combined_with || g1->same)
6874 for (j = 0; j < giv_count; j++)
6877 if (g1 != g2 && can_combine[i*giv_count + j]
6878 /* If it has already been combined, skip. */
6879 && ! g2->same && ! g2->combined_with)
6883 g2->new_reg = can_combine[i*giv_count + j];
6885 g1->combined_with++;
6886 g1->lifetime += g2->lifetime;
6888 g1_add_benefit += combine_givs_benefit_from (g1, g2);
6890 /* ??? The new final_[bg]iv_value code does a much better job
6891 of finding replaceable giv's, and hence this code may no
6892 longer be necessary. */
6893 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
6894 g1_add_benefit -= copy_cost;
6896 /* To help optimize the next set of combinations, remove
6897 this giv from the benefits of other potential mates. */
6898 for (l = 0; l < giv_count; ++l)
6900 int m = stats[l].giv_number;
6901 if (can_combine[m*giv_count + j])
6903 /* Remove additional weight for being reused. */
6904 stats[l].total_benefit -= 3 +
6905 combine_givs_benefit_from (giv_array[m], g2);
6909 if (loop_dump_stream)
6910 fprintf (loop_dump_stream,
6911 "giv at %d combined with giv at %d\n",
6912 INSN_UID (g2->insn), INSN_UID (g1->insn));
6916 /* To help optimize the next set of combinations, remove
6917 this giv from the benefits of other potential mates. */
6918 if (g1->combined_with)
6920 for (j = 0; j < giv_count; ++j)
6922 int m = stats[j].giv_number;
6923 if (can_combine[m*giv_count + j])
6925 /* Remove additional weight for being reused. */
6926 stats[j].total_benefit -= 3 +
6927 combine_givs_benefit_from (giv_array[m], g1);
6931 g1->benefit += g1_add_benefit;
6933 /* We've finished with this giv, and everything it touched.
6934 Restart the combination so that proper weights for the
6935 rest of the givs are properly taken into account. */
6936 /* ??? Ideally we would compact the arrays at this point, so
6937 as to not cover old ground. But sanely compacting
6938 can_combine is tricky. */
6944 struct recombine_givs_stats
6947 int start_luid, end_luid;
6950 /* Used below as comparison function for qsort. We want a ascending luid
6951 when scanning the array starting at the end, thus the arguments are
6954 cmp_recombine_givs_stats (x, y)
6955 struct recombine_givs_stats *x, *y;
6958 d = y->start_luid - x->start_luid;
6959 /* Stabilize the sort. */
6961 d = y->giv_number - x->giv_number;
6965 /* Scan X, which is a part of INSN, for the end of life of a giv. Also
6966 look for the start of life of a giv where the start has not been seen
6967 yet to unlock the search for the end of its life.
6968 Only consider givs that belong to BIV.
6969 Return the total number of lifetime ends that have been found. */
6971 find_life_end (x, stats, insn, biv)
6973 struct recombine_givs_stats *stats;
6980 code = GET_CODE (x);
6985 rtx reg = SET_DEST (x);
6986 if (GET_CODE (reg) == REG)
6988 int regno = REGNO (reg);
6989 struct induction *v = REG_IV_INFO (regno);
6991 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
6993 && v->src_reg == biv
6994 && stats[v->ix].end_luid <= 0)
6996 /* If we see a 0 here for end_luid, it means that we have
6997 scanned the entire loop without finding any use at all.
6998 We must not predicate this code on a start_luid match
6999 since that would make the test fail for givs that have
7000 been hoisted out of inner loops. */
7001 if (stats[v->ix].end_luid == 0)
7003 stats[v->ix].end_luid = stats[v->ix].start_luid;
7004 return 1 + find_life_end (SET_SRC (x), stats, insn, biv);
7006 else if (stats[v->ix].start_luid == INSN_LUID (insn))
7007 stats[v->ix].end_luid = 0;
7009 return find_life_end (SET_SRC (x), stats, insn, biv);
7015 int regno = REGNO (x);
7016 struct induction *v = REG_IV_INFO (regno);
7018 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7020 && v->src_reg == biv
7021 && stats[v->ix].end_luid == 0)
7023 while (INSN_UID (insn) >= max_uid_for_loop)
7024 insn = NEXT_INSN (insn);
7025 stats[v->ix].end_luid = INSN_LUID (insn);
7038 fmt = GET_RTX_FORMAT (code);
7040 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7043 retval += find_life_end (XEXP (x, i), stats, insn, biv);
7045 else if (fmt[i] == 'E')
7046 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7047 retval += find_life_end (XVECEXP (x, i, j), stats, insn, biv);
7052 /* For each giv that has been combined with another, look if
7053 we can combine it with the most recently used one instead.
7054 This tends to shorten giv lifetimes, and helps the next step:
7055 try to derive givs from other givs. */
7057 recombine_givs (bl, loop_start, loop_end, unroll_p)
7058 struct iv_class *bl;
7059 rtx loop_start, loop_end;
7062 struct induction *v, **giv_array, *last_giv;
7063 struct recombine_givs_stats *stats;
7066 int ends_need_computing;
7068 for (giv_count = 0, v = bl->giv; v; v = v->next_iv)
7074 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
7075 stats = (struct recombine_givs_stats *) alloca (giv_count * sizeof *stats);
7077 /* Initialize stats and set up the ix field for each giv in stats to name
7078 the corresponding index into stats. */
7079 for (i = 0, v = bl->giv; v; v = v->next_iv)
7086 stats[i].giv_number = i;
7087 /* If this giv has been hoisted out of an inner loop, use the luid of
7088 the previous insn. */
7089 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7091 stats[i].start_luid = INSN_LUID (p);
7096 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7098 /* Do the actual most-recently-used recombination. */
7099 for (last_giv = 0, i = giv_count - 1; i >= 0; i--)
7101 v = giv_array[stats[i].giv_number];
7104 struct induction *old_same = v->same;
7107 /* combine_givs_p actually says if we can make this transformation.
7108 The other tests are here only to avoid keeping a giv alive
7109 that could otherwise be eliminated. */
7111 && ((old_same->maybe_dead && ! old_same->combined_with)
7112 || ! last_giv->maybe_dead
7113 || last_giv->combined_with)
7114 && (new_combine = combine_givs_p (last_giv, v)))
7116 old_same->combined_with--;
7117 v->new_reg = new_combine;
7119 last_giv->combined_with++;
7120 /* No need to update lifetimes / benefits here since we have
7121 already decided what to reduce. */
7123 if (loop_dump_stream)
7125 fprintf (loop_dump_stream,
7126 "giv at %d recombined with giv at %d as ",
7127 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7128 print_rtl (loop_dump_stream, v->new_reg);
7129 putc ('\n', loop_dump_stream);
7135 else if (v->giv_type != DEST_REG)
7138 || (last_giv->maybe_dead && ! last_giv->combined_with)
7140 || v->combined_with)
7144 ends_need_computing = 0;
7145 /* For each DEST_REG giv, compute lifetime starts, and try to compute
7146 lifetime ends from regscan info. */
7147 for (i = 0, v = bl->giv; v; v = v->next_iv)
7151 if (v->giv_type == DEST_ADDR)
7153 /* Loop unrolling of an inner loop can even create new DEST_REG
7156 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7158 stats[i].start_luid = stats[i].end_luid = INSN_LUID (p);
7160 stats[i].end_luid++;
7162 else /* v->giv_type == DEST_REG */
7166 stats[i].start_luid = INSN_LUID (v->insn);
7167 stats[i].end_luid = INSN_LUID (v->last_use);
7169 else if (INSN_UID (v->insn) >= max_uid_for_loop)
7172 /* This insn has been created by loop optimization on an inner
7173 loop. We don't have a proper start_luid that will match
7174 when we see the first set. But we do know that there will
7175 be no use before the set, so we can set end_luid to 0 so that
7176 we'll start looking for the last use right away. */
7177 for (p = PREV_INSN (v->insn); INSN_UID (p) >= max_uid_for_loop; )
7179 stats[i].start_luid = INSN_LUID (p);
7180 stats[i].end_luid = 0;
7181 ends_need_computing++;
7185 int regno = REGNO (v->dest_reg);
7186 int count = VARRAY_INT (n_times_set, regno) - 1;
7189 /* Find the first insn that sets the giv, so that we can verify
7190 if this giv's lifetime wraps around the loop. We also need
7191 the luid of the first setting insn in order to detect the
7192 last use properly. */
7195 p = prev_nonnote_insn (p);
7196 if (reg_set_p (v->dest_reg, p))
7200 stats[i].start_luid = INSN_LUID (p);
7201 if (stats[i].start_luid > uid_luid[REGNO_FIRST_UID (regno)])
7203 stats[i].end_luid = -1;
7204 ends_need_computing++;
7208 stats[i].end_luid = uid_luid[REGNO_LAST_UID (regno)];
7209 if (stats[i].end_luid > INSN_LUID (loop_end))
7211 stats[i].end_luid = -1;
7212 ends_need_computing++;
7220 /* If the regscan information was unconclusive for one or more DEST_REG
7221 givs, scan the all insn in the loop to find out lifetime ends. */
7222 if (ends_need_computing)
7224 rtx biv = bl->biv->src_reg;
7229 if (p == loop_start)
7232 if (GET_RTX_CLASS (GET_CODE (p)) != 'i')
7234 ends_need_computing -= find_life_end (PATTERN (p), stats, p, biv);
7236 while (ends_need_computing);
7239 /* Set start_luid back to the last insn that sets the giv. This allows
7240 more combinations. */
7241 for (i = 0, v = bl->giv; v; v = v->next_iv)
7245 if (INSN_UID (v->insn) < max_uid_for_loop)
7246 stats[i].start_luid = INSN_LUID (v->insn);
7250 /* Now adjust lifetime ends by taking combined givs into account. */
7251 for (i = 0, v = bl->giv; v; v = v->next_iv)
7258 if (v->same && ! v->same->ignore)
7261 luid = stats[i].start_luid;
7262 /* Use unsigned arithmetic to model loop wrap-around. */
7263 if (luid - stats[j].start_luid
7264 > (unsigned) stats[j].end_luid - stats[j].start_luid)
7265 stats[j].end_luid = luid;
7270 qsort (stats, giv_count, sizeof(*stats), cmp_recombine_givs_stats);
7272 /* Try to derive DEST_REG givs from previous DEST_REG givs with the
7273 same mult_val and non-overlapping lifetime. This reduces register
7275 Once we find a DEST_REG giv that is suitable to derive others from,
7276 we set last_giv to this giv, and try to derive as many other DEST_REG
7277 givs from it without joining overlapping lifetimes. If we then
7278 encounter a DEST_REG giv that we can't derive, we set rescan to the
7279 index for this giv (unless rescan is already set).
7280 When we are finished with the current LAST_GIV (i.e. the inner loop
7281 terminates), we start again with rescan, which then becomes the new
7283 for (i = giv_count - 1; i >= 0; i = rescan)
7285 int life_start, life_end;
7287 for (last_giv = 0, rescan = -1; i >= 0; i--)
7291 v = giv_array[stats[i].giv_number];
7292 if (v->giv_type != DEST_REG || v->derived || v->same)
7297 life_start = stats[i].start_luid;
7298 life_end = stats[i].end_luid;
7301 /* Use unsigned arithmetic to model loop wrap around. */
7302 if (((unsigned) stats[i].start_luid - life_start
7303 >= (unsigned) life_end - life_start)
7304 && ((unsigned) stats[i].end_luid - life_start
7305 >= (unsigned) life_end - life_start)
7306 && rtx_equal_p (last_giv->mult_val, v->mult_val)
7307 /* ??? Could handle libcalls, but would need more logic. */
7308 && ! find_reg_note (v->insn, REG_RETVAL, NULL_RTX)
7309 /* We would really like to know if for any giv that v
7310 is combined with, v->insn or any intervening biv increment
7311 dominates that combined giv. However, we
7312 don't have this detailed control flow information.
7313 N.B. since last_giv will be reduced, it is valid
7314 anywhere in the loop, so we don't need to check the
7315 validity of last_giv. */
7316 && (v->always_executed || ! v->combined_with)
7317 && (sum = express_from (last_giv, v))
7318 /* Make sure we don't make the add more expensive. ADD_COST
7319 doesn't take different costs of registers and constants into
7320 account, so compare the cost of the actual SET_SRCs. */
7321 && (rtx_cost (sum, SET)
7322 <= rtx_cost (SET_SRC (single_set (v->insn)), SET))
7323 /* ??? unroll can't understand anything but reg + const_int
7324 sums. It would be cleaner to fix unroll. */
7325 && ((GET_CODE (sum) == PLUS
7326 && GET_CODE (XEXP (sum, 0)) == REG
7327 && GET_CODE (XEXP (sum, 1)) == CONST_INT)
7329 && validate_change (v->insn, &PATTERN (v->insn),
7330 gen_rtx_SET (GET_MODE (v->dest_reg),
7331 v->dest_reg, sum), 0))
7334 v->new_reg = v->dest_reg;
7335 life_end = stats[i].end_luid;
7337 if (loop_dump_stream)
7339 fprintf (loop_dump_stream,
7340 "giv at %d derived from %d as ",
7341 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7342 print_rtl (loop_dump_stream, v->new_reg);
7343 putc ('\n', loop_dump_stream);
7346 else if (rescan < 0)
7352 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
7355 emit_iv_add_mult (b, m, a, reg, insert_before)
7356 rtx b; /* initial value of basic induction variable */
7357 rtx m; /* multiplicative constant */
7358 rtx a; /* additive constant */
7359 rtx reg; /* destination register */
7365 /* Prevent unexpected sharing of these rtx. */
7369 /* Increase the lifetime of any invariants moved further in code. */
7370 update_reg_last_use (a, insert_before);
7371 update_reg_last_use (b, insert_before);
7372 update_reg_last_use (m, insert_before);
7375 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
7377 emit_move_insn (reg, result);
7378 seq = gen_sequence ();
7381 emit_insn_before (seq, insert_before);
7383 /* It is entirely possible that the expansion created lots of new
7384 registers. Iterate over the sequence we just created and
7387 if (GET_CODE (seq) == SEQUENCE)
7390 for (i = 0; i < XVECLEN (seq, 0); ++i)
7392 rtx set = single_set (XVECEXP (seq, 0, i));
7393 if (set && GET_CODE (SET_DEST (set)) == REG)
7394 record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0);
7397 else if (GET_CODE (seq) == SET
7398 && GET_CODE (SET_DEST (seq)) == REG)
7399 record_base_value (REGNO (SET_DEST (seq)), SET_SRC (seq), 0);
7402 /* Test whether A * B can be computed without
7403 an actual multiply insn. Value is 1 if so. */
7406 product_cheap_p (a, b)
7412 struct obstack *old_rtl_obstack = rtl_obstack;
7413 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
7416 /* If only one is constant, make it B. */
7417 if (GET_CODE (a) == CONST_INT)
7418 tmp = a, a = b, b = tmp;
7420 /* If first constant, both constant, so don't need multiply. */
7421 if (GET_CODE (a) == CONST_INT)
7424 /* If second not constant, neither is constant, so would need multiply. */
7425 if (GET_CODE (b) != CONST_INT)
7428 /* One operand is constant, so might not need multiply insn. Generate the
7429 code for the multiply and see if a call or multiply, or long sequence
7430 of insns is generated. */
7432 rtl_obstack = &temp_obstack;
7434 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
7435 tmp = gen_sequence ();
7438 if (GET_CODE (tmp) == SEQUENCE)
7440 if (XVEC (tmp, 0) == 0)
7442 else if (XVECLEN (tmp, 0) > 3)
7445 for (i = 0; i < XVECLEN (tmp, 0); i++)
7447 rtx insn = XVECEXP (tmp, 0, i);
7449 if (GET_CODE (insn) != INSN
7450 || (GET_CODE (PATTERN (insn)) == SET
7451 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
7452 || (GET_CODE (PATTERN (insn)) == PARALLEL
7453 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
7454 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
7461 else if (GET_CODE (tmp) == SET
7462 && GET_CODE (SET_SRC (tmp)) == MULT)
7464 else if (GET_CODE (tmp) == PARALLEL
7465 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
7466 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
7469 /* Free any storage we obtained in generating this multiply and restore rtl
7470 allocation to its normal obstack. */
7471 obstack_free (&temp_obstack, storage);
7472 rtl_obstack = old_rtl_obstack;
7477 /* Check to see if loop can be terminated by a "decrement and branch until
7478 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
7479 Also try reversing an increment loop to a decrement loop
7480 to see if the optimization can be performed.
7481 Value is nonzero if optimization was performed. */
7483 /* This is useful even if the architecture doesn't have such an insn,
7484 because it might change a loops which increments from 0 to n to a loop
7485 which decrements from n to 0. A loop that decrements to zero is usually
7486 faster than one that increments from zero. */
7488 /* ??? This could be rewritten to use some of the loop unrolling procedures,
7489 such as approx_final_value, biv_total_increment, loop_iterations, and
7490 final_[bg]iv_value. */
7493 check_dbra_loop (loop_end, insn_count, loop_start, loop_info)
7497 struct loop_info *loop_info;
7499 struct iv_class *bl;
7506 rtx before_comparison;
7510 int compare_and_branch;
7512 /* If last insn is a conditional branch, and the insn before tests a
7513 register value, try to optimize it. Otherwise, we can't do anything. */
7515 jump = PREV_INSN (loop_end);
7516 comparison = get_condition_for_loop (jump);
7517 if (comparison == 0)
7520 /* Try to compute whether the compare/branch at the loop end is one or
7521 two instructions. */
7522 get_condition (jump, &first_compare);
7523 if (first_compare == jump)
7524 compare_and_branch = 1;
7525 else if (first_compare == prev_nonnote_insn (jump))
7526 compare_and_branch = 2;
7530 /* Check all of the bivs to see if the compare uses one of them.
7531 Skip biv's set more than once because we can't guarantee that
7532 it will be zero on the last iteration. Also skip if the biv is
7533 used between its update and the test insn. */
7535 for (bl = loop_iv_list; bl; bl = bl->next)
7537 if (bl->biv_count == 1
7538 && bl->biv->dest_reg == XEXP (comparison, 0)
7539 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
7547 /* Look for the case where the basic induction variable is always
7548 nonnegative, and equals zero on the last iteration.
7549 In this case, add a reg_note REG_NONNEG, which allows the
7550 m68k DBRA instruction to be used. */
7552 if (((GET_CODE (comparison) == GT
7553 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
7554 && INTVAL (XEXP (comparison, 1)) == -1)
7555 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
7556 && GET_CODE (bl->biv->add_val) == CONST_INT
7557 && INTVAL (bl->biv->add_val) < 0)
7559 /* Initial value must be greater than 0,
7560 init_val % -dec_value == 0 to ensure that it equals zero on
7561 the last iteration */
7563 if (GET_CODE (bl->initial_value) == CONST_INT
7564 && INTVAL (bl->initial_value) > 0
7565 && (INTVAL (bl->initial_value)
7566 % (-INTVAL (bl->biv->add_val))) == 0)
7568 /* register always nonnegative, add REG_NOTE to branch */
7569 REG_NOTES (PREV_INSN (loop_end))
7570 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7571 REG_NOTES (PREV_INSN (loop_end)));
7577 /* If the decrement is 1 and the value was tested as >= 0 before
7578 the loop, then we can safely optimize. */
7579 for (p = loop_start; p; p = PREV_INSN (p))
7581 if (GET_CODE (p) == CODE_LABEL)
7583 if (GET_CODE (p) != JUMP_INSN)
7586 before_comparison = get_condition_for_loop (p);
7587 if (before_comparison
7588 && XEXP (before_comparison, 0) == bl->biv->dest_reg
7589 && GET_CODE (before_comparison) == LT
7590 && XEXP (before_comparison, 1) == const0_rtx
7591 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
7592 && INTVAL (bl->biv->add_val) == -1)
7594 REG_NOTES (PREV_INSN (loop_end))
7595 = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7596 REG_NOTES (PREV_INSN (loop_end)));
7603 else if (INTVAL (bl->biv->add_val) > 0)
7605 /* Try to change inc to dec, so can apply above optimization. */
7607 all registers modified are induction variables or invariant,
7608 all memory references have non-overlapping addresses
7609 (obviously true if only one write)
7610 allow 2 insns for the compare/jump at the end of the loop. */
7611 /* Also, we must avoid any instructions which use both the reversed
7612 biv and another biv. Such instructions will fail if the loop is
7613 reversed. We meet this condition by requiring that either
7614 no_use_except_counting is true, or else that there is only
7616 int num_nonfixed_reads = 0;
7617 /* 1 if the iteration var is used only to count iterations. */
7618 int no_use_except_counting = 0;
7619 /* 1 if the loop has no memory store, or it has a single memory store
7620 which is reversible. */
7621 int reversible_mem_store = 1;
7623 if (bl->giv_count == 0
7624 && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]])
7626 rtx bivreg = regno_reg_rtx[bl->regno];
7628 /* If there are no givs for this biv, and the only exit is the
7629 fall through at the end of the loop, then
7630 see if perhaps there are no uses except to count. */
7631 no_use_except_counting = 1;
7632 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7633 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
7635 rtx set = single_set (p);
7637 if (set && GET_CODE (SET_DEST (set)) == REG
7638 && REGNO (SET_DEST (set)) == bl->regno)
7639 /* An insn that sets the biv is okay. */
7641 else if (p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
7642 || p == prev_nonnote_insn (loop_end))
7643 /* Don't bother about the end test. */
7645 else if (reg_mentioned_p (bivreg, PATTERN (p)))
7647 no_use_except_counting = 0;
7653 if (no_use_except_counting)
7654 ; /* no need to worry about MEMs. */
7655 else if (num_mem_sets <= 1)
7657 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7658 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
7659 num_nonfixed_reads += count_nonfixed_reads (PATTERN (p));
7661 /* If the loop has a single store, and the destination address is
7662 invariant, then we can't reverse the loop, because this address
7663 might then have the wrong value at loop exit.
7664 This would work if the source was invariant also, however, in that
7665 case, the insn should have been moved out of the loop. */
7667 if (num_mem_sets == 1)
7668 reversible_mem_store
7669 = (! unknown_address_altered
7670 && ! invariant_p (XEXP (loop_store_mems, 0)));
7675 /* This code only acts for innermost loops. Also it simplifies
7676 the memory address check by only reversing loops with
7677 zero or one memory access.
7678 Two memory accesses could involve parts of the same array,
7679 and that can't be reversed.
7680 If the biv is used only for counting, than we don't need to worry
7681 about all these things. */
7683 if ((num_nonfixed_reads <= 1
7685 && !loop_has_volatile
7686 && reversible_mem_store
7687 && (bl->giv_count + bl->biv_count + num_mem_sets
7688 + num_movables + compare_and_branch == insn_count)
7689 && (bl == loop_iv_list && bl->next == 0))
7690 || no_use_except_counting)
7694 /* Loop can be reversed. */
7695 if (loop_dump_stream)
7696 fprintf (loop_dump_stream, "Can reverse loop\n");
7698 /* Now check other conditions:
7700 The increment must be a constant, as must the initial value,
7701 and the comparison code must be LT.
7703 This test can probably be improved since +/- 1 in the constant
7704 can be obtained by changing LT to LE and vice versa; this is
7708 /* for constants, LE gets turned into LT */
7709 && (GET_CODE (comparison) == LT
7710 || (GET_CODE (comparison) == LE
7711 && no_use_except_counting)))
7713 HOST_WIDE_INT add_val, add_adjust, comparison_val;
7714 rtx initial_value, comparison_value;
7716 enum rtx_code cmp_code;
7717 int comparison_const_width;
7718 unsigned HOST_WIDE_INT comparison_sign_mask;
7720 add_val = INTVAL (bl->biv->add_val);
7721 comparison_value = XEXP (comparison, 1);
7722 if (GET_MODE (comparison_value) == VOIDmode)
7723 comparison_const_width
7724 = GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0)));
7726 comparison_const_width
7727 = GET_MODE_BITSIZE (GET_MODE (comparison_value));
7728 if (comparison_const_width > HOST_BITS_PER_WIDE_INT)
7729 comparison_const_width = HOST_BITS_PER_WIDE_INT;
7730 comparison_sign_mask
7731 = (unsigned HOST_WIDE_INT)1 << (comparison_const_width - 1);
7733 /* If the comparison value is not a loop invariant, then we
7734 can not reverse this loop.
7736 ??? If the insns which initialize the comparison value as
7737 a whole compute an invariant result, then we could move
7738 them out of the loop and proceed with loop reversal. */
7739 if (!invariant_p (comparison_value))
7742 if (GET_CODE (comparison_value) == CONST_INT)
7743 comparison_val = INTVAL (comparison_value);
7744 initial_value = bl->initial_value;
7746 /* Normalize the initial value if it is an integer and
7747 has no other use except as a counter. This will allow
7748 a few more loops to be reversed. */
7749 if (no_use_except_counting
7750 && GET_CODE (comparison_value) == CONST_INT
7751 && GET_CODE (initial_value) == CONST_INT)
7753 comparison_val = comparison_val - INTVAL (bl->initial_value);
7754 /* The code below requires comparison_val to be a multiple
7755 of add_val in order to do the loop reversal, so
7756 round up comparison_val to a multiple of add_val.
7757 Since comparison_value is constant, we know that the
7758 current comparison code is LT. */
7759 comparison_val = comparison_val + add_val - 1;
7761 -= (unsigned HOST_WIDE_INT) comparison_val % add_val;
7762 /* We postpone overflow checks for COMPARISON_VAL here;
7763 even if there is an overflow, we might still be able to
7764 reverse the loop, if converting the loop exit test to
7766 initial_value = const0_rtx;
7769 /* First check if we can do a vanilla loop reversal. */
7770 if (initial_value == const0_rtx
7771 /* If we have a decrement_and_branch_on_count, prefer
7772 the NE test, since this will allow that instruction to
7773 be generated. Note that we must use a vanilla loop
7774 reversal if the biv is used to calculate a giv or has
7775 a non-counting use. */
7776 #if ! defined (HAVE_decrement_and_branch_until_zero) && defined (HAVE_decrement_and_branch_on_count)
7777 && (! (add_val == 1 && loop_info->vtop
7778 && (bl->biv_count == 0
7779 || no_use_except_counting)))
7781 && GET_CODE (comparison_value) == CONST_INT
7782 /* Now do postponed overflow checks on COMPARISON_VAL. */
7783 && ! (((comparison_val - add_val) ^ INTVAL (comparison_value))
7784 & comparison_sign_mask))
7786 /* Register will always be nonnegative, with value
7787 0 on last iteration */
7788 add_adjust = add_val;
7792 else if (add_val == 1 && loop_info->vtop
7793 && (bl->biv_count == 0
7794 || no_use_except_counting))
7802 if (GET_CODE (comparison) == LE)
7803 add_adjust -= add_val;
7805 /* If the initial value is not zero, or if the comparison
7806 value is not an exact multiple of the increment, then we
7807 can not reverse this loop. */
7808 if (initial_value == const0_rtx
7809 && GET_CODE (comparison_value) == CONST_INT)
7811 if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0)
7816 if (! no_use_except_counting || add_val != 1)
7820 final_value = comparison_value;
7822 /* Reset these in case we normalized the initial value
7823 and comparison value above. */
7824 if (GET_CODE (comparison_value) == CONST_INT
7825 && GET_CODE (initial_value) == CONST_INT)
7827 comparison_value = GEN_INT (comparison_val);
7829 = GEN_INT (comparison_val + INTVAL (bl->initial_value));
7831 bl->initial_value = initial_value;
7833 /* Save some info needed to produce the new insns. */
7834 reg = bl->biv->dest_reg;
7835 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
7836 if (jump_label == pc_rtx)
7837 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
7838 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
7840 /* Set start_value; if this is not a CONST_INT, we need
7842 Initialize biv to start_value before loop start.
7843 The old initializing insn will be deleted as a
7844 dead store by flow.c. */
7845 if (initial_value == const0_rtx
7846 && GET_CODE (comparison_value) == CONST_INT)
7848 start_value = GEN_INT (comparison_val - add_adjust);
7849 emit_insn_before (gen_move_insn (reg, start_value),
7852 else if (GET_CODE (initial_value) == CONST_INT)
7854 rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust);
7855 enum machine_mode mode = GET_MODE (reg);
7856 enum insn_code icode
7857 = add_optab->handlers[(int) mode].insn_code;
7858 if (! (*insn_operand_predicate[icode][0]) (reg, mode)
7859 || ! ((*insn_operand_predicate[icode][1])
7860 (comparison_value, mode))
7861 || ! (*insn_operand_predicate[icode][2]) (offset, mode))
7864 = gen_rtx_PLUS (mode, comparison_value, offset);
7865 emit_insn_before ((GEN_FCN (icode)
7866 (reg, comparison_value, offset)),
7868 if (GET_CODE (comparison) == LE)
7869 final_value = gen_rtx_PLUS (mode, comparison_value,
7872 else if (! add_adjust)
7874 enum machine_mode mode = GET_MODE (reg);
7875 enum insn_code icode
7876 = sub_optab->handlers[(int) mode].insn_code;
7877 if (! (*insn_operand_predicate[icode][0]) (reg, mode)
7878 || ! ((*insn_operand_predicate[icode][1])
7879 (comparison_value, mode))
7880 || ! ((*insn_operand_predicate[icode][2])
7881 (initial_value, mode)))
7884 = gen_rtx_MINUS (mode, comparison_value, initial_value);
7885 emit_insn_before ((GEN_FCN (icode)
7886 (reg, comparison_value, initial_value)),
7890 /* We could handle the other cases too, but it'll be
7891 better to have a testcase first. */
7894 /* We may not have a single insn which can increment a reg, so
7895 create a sequence to hold all the insns from expand_inc. */
7897 expand_inc (reg, new_add_val);
7898 tem = gen_sequence ();
7901 p = emit_insn_before (tem, bl->biv->insn);
7902 delete_insn (bl->biv->insn);
7904 /* Update biv info to reflect its new status. */
7906 bl->initial_value = start_value;
7907 bl->biv->add_val = new_add_val;
7909 /* Update loop info. */
7910 loop_info->initial_value = reg;
7911 loop_info->initial_equiv_value = reg;
7912 loop_info->final_value = const0_rtx;
7913 loop_info->final_equiv_value = const0_rtx;
7914 loop_info->comparison_value = const0_rtx;
7915 loop_info->comparison_code = cmp_code;
7916 loop_info->increment = new_add_val;
7918 /* Inc LABEL_NUSES so that delete_insn will
7919 not delete the label. */
7920 LABEL_NUSES (XEXP (jump_label, 0)) ++;
7922 /* Emit an insn after the end of the loop to set the biv's
7923 proper exit value if it is used anywhere outside the loop. */
7924 if ((REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare))
7926 || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
7927 emit_insn_after (gen_move_insn (reg, final_value),
7930 /* Delete compare/branch at end of loop. */
7931 delete_insn (PREV_INSN (loop_end));
7932 if (compare_and_branch == 2)
7933 delete_insn (first_compare);
7935 /* Add new compare/branch insn at end of loop. */
7937 emit_cmp_and_jump_insns (reg, const0_rtx, cmp_code, NULL_RTX,
7938 GET_MODE (reg), 0, 0,
7939 XEXP (jump_label, 0));
7940 tem = gen_sequence ();
7942 emit_jump_insn_before (tem, loop_end);
7944 for (tem = PREV_INSN (loop_end);
7945 tem && GET_CODE (tem) != JUMP_INSN;
7946 tem = PREV_INSN (tem))
7950 JUMP_LABEL (tem) = XEXP (jump_label, 0);
7956 /* Increment of LABEL_NUSES done above. */
7957 /* Register is now always nonnegative,
7958 so add REG_NONNEG note to the branch. */
7959 REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX,
7965 /* Mark that this biv has been reversed. Each giv which depends
7966 on this biv, and which is also live past the end of the loop
7967 will have to be fixed up. */
7971 if (loop_dump_stream)
7972 fprintf (loop_dump_stream,
7973 "Reversed loop and added reg_nonneg\n");
7983 /* Verify whether the biv BL appears to be eliminable,
7984 based on the insns in the loop that refer to it.
7985 LOOP_START is the first insn of the loop, and END is the end insn.
7987 If ELIMINATE_P is non-zero, actually do the elimination.
7989 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
7990 determine whether invariant insns should be placed inside or at the
7991 start of the loop. */
7994 maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count)
7995 struct iv_class *bl;
7999 int threshold, insn_count;
8001 rtx reg = bl->biv->dest_reg;
8004 /* Scan all insns in the loop, stopping if we find one that uses the
8005 biv in a way that we cannot eliminate. */
8007 for (p = loop_start; p != end; p = NEXT_INSN (p))
8009 enum rtx_code code = GET_CODE (p);
8010 rtx where = threshold >= insn_count ? loop_start : p;
8012 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
8013 && reg_mentioned_p (reg, PATTERN (p))
8014 && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where))
8016 if (loop_dump_stream)
8017 fprintf (loop_dump_stream,
8018 "Cannot eliminate biv %d: biv used in insn %d.\n",
8019 bl->regno, INSN_UID (p));
8026 if (loop_dump_stream)
8027 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
8028 bl->regno, eliminate_p ? "was" : "can be");
8035 /* If BL appears in X (part of the pattern of INSN), see if we can
8036 eliminate its use. If so, return 1. If not, return 0.
8038 If BIV does not appear in X, return 1.
8040 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
8041 where extra insns should be added. Depending on how many items have been
8042 moved out of the loop, it will either be before INSN or at the start of
8046 maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where)
8048 struct iv_class *bl;
8052 enum rtx_code code = GET_CODE (x);
8053 rtx reg = bl->biv->dest_reg;
8054 enum machine_mode mode = GET_MODE (reg);
8055 struct induction *v;
8067 /* If we haven't already been able to do something with this BIV,
8068 we can't eliminate it. */
8074 /* If this sets the BIV, it is not a problem. */
8075 if (SET_DEST (x) == reg)
8078 /* If this is an insn that defines a giv, it is also ok because
8079 it will go away when the giv is reduced. */
8080 for (v = bl->giv; v; v = v->next_iv)
8081 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
8085 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
8087 /* Can replace with any giv that was reduced and
8088 that has (MULT_VAL != 0) and (ADD_VAL == 0).
8089 Require a constant for MULT_VAL, so we know it's nonzero.
8090 ??? We disable this optimization to avoid potential
8093 for (v = bl->giv; v; v = v->next_iv)
8094 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
8095 && v->add_val == const0_rtx
8096 && ! v->ignore && ! v->maybe_dead && v->always_computable
8100 /* If the giv V had the auto-inc address optimization applied
8101 to it, and INSN occurs between the giv insn and the biv
8102 insn, then we must adjust the value used here.
8103 This is rare, so we don't bother to do so. */
8105 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
8106 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
8107 || (INSN_LUID (v->insn) > INSN_LUID (insn)
8108 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
8114 /* If the giv has the opposite direction of change,
8115 then reverse the comparison. */
8116 if (INTVAL (v->mult_val) < 0)
8117 new = gen_rtx_COMPARE (GET_MODE (v->new_reg),
8118 const0_rtx, v->new_reg);
8122 /* We can probably test that giv's reduced reg. */
8123 if (validate_change (insn, &SET_SRC (x), new, 0))
8127 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
8128 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
8129 Require a constant for MULT_VAL, so we know it's nonzero.
8130 ??? Do this only if ADD_VAL is a pointer to avoid a potential
8131 overflow problem. */
8133 for (v = bl->giv; v; v = v->next_iv)
8134 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
8135 && ! v->ignore && ! v->maybe_dead && v->always_computable
8137 && (GET_CODE (v->add_val) == SYMBOL_REF
8138 || GET_CODE (v->add_val) == LABEL_REF
8139 || GET_CODE (v->add_val) == CONST
8140 || (GET_CODE (v->add_val) == REG
8141 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
8143 /* If the giv V had the auto-inc address optimization applied
8144 to it, and INSN occurs between the giv insn and the biv
8145 insn, then we must adjust the value used here.
8146 This is rare, so we don't bother to do so. */
8148 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
8149 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
8150 || (INSN_LUID (v->insn) > INSN_LUID (insn)
8151 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
8157 /* If the giv has the opposite direction of change,
8158 then reverse the comparison. */
8159 if (INTVAL (v->mult_val) < 0)
8160 new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val),
8163 new = gen_rtx_COMPARE (VOIDmode, v->new_reg,
8164 copy_rtx (v->add_val));
8166 /* Replace biv with the giv's reduced register. */
8167 update_reg_last_use (v->add_val, insn);
8168 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8171 /* Insn doesn't support that constant or invariant. Copy it
8172 into a register (it will be a loop invariant.) */
8173 tem = gen_reg_rtx (GET_MODE (v->new_reg));
8175 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
8178 /* Substitute the new register for its invariant value in
8179 the compare expression. */
8180 XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
8181 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8190 case GT: case GE: case GTU: case GEU:
8191 case LT: case LE: case LTU: case LEU:
8192 /* See if either argument is the biv. */
8193 if (XEXP (x, 0) == reg)
8194 arg = XEXP (x, 1), arg_operand = 1;
8195 else if (XEXP (x, 1) == reg)
8196 arg = XEXP (x, 0), arg_operand = 0;
8200 if (CONSTANT_P (arg))
8202 /* First try to replace with any giv that has constant positive
8203 mult_val and constant add_val. We might be able to support
8204 negative mult_val, but it seems complex to do it in general. */
8206 for (v = bl->giv; v; v = v->next_iv)
8207 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8208 && (GET_CODE (v->add_val) == SYMBOL_REF
8209 || GET_CODE (v->add_val) == LABEL_REF
8210 || GET_CODE (v->add_val) == CONST
8211 || (GET_CODE (v->add_val) == REG
8212 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
8213 && ! v->ignore && ! v->maybe_dead && v->always_computable
8216 /* If the giv V had the auto-inc address optimization applied
8217 to it, and INSN occurs between the giv insn and the biv
8218 insn, then we must adjust the value used here.
8219 This is rare, so we don't bother to do so. */
8221 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
8222 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
8223 || (INSN_LUID (v->insn) > INSN_LUID (insn)
8224 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
8230 /* Replace biv with the giv's reduced reg. */
8231 XEXP (x, 1-arg_operand) = v->new_reg;
8233 /* If all constants are actually constant integers and
8234 the derived constant can be directly placed in the COMPARE,
8236 if (GET_CODE (arg) == CONST_INT
8237 && GET_CODE (v->mult_val) == CONST_INT
8238 && GET_CODE (v->add_val) == CONST_INT
8239 && validate_change (insn, &XEXP (x, arg_operand),
8240 GEN_INT (INTVAL (arg)
8241 * INTVAL (v->mult_val)
8242 + INTVAL (v->add_val)), 0))
8245 /* Otherwise, load it into a register. */
8246 tem = gen_reg_rtx (mode);
8247 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8248 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
8251 /* If that failed, put back the change we made above. */
8252 XEXP (x, 1-arg_operand) = reg;
8255 /* Look for giv with positive constant mult_val and nonconst add_val.
8256 Insert insns to calculate new compare value.
8257 ??? Turn this off due to possible overflow. */
8259 for (v = bl->giv; v; v = v->next_iv)
8260 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8261 && ! v->ignore && ! v->maybe_dead && v->always_computable
8267 /* If the giv V had the auto-inc address optimization applied
8268 to it, and INSN occurs between the giv insn and the biv
8269 insn, then we must adjust the value used here.
8270 This is rare, so we don't bother to do so. */
8272 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
8273 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
8274 || (INSN_LUID (v->insn) > INSN_LUID (insn)
8275 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
8281 tem = gen_reg_rtx (mode);
8283 /* Replace biv with giv's reduced register. */
8284 validate_change (insn, &XEXP (x, 1 - arg_operand),
8287 /* Compute value to compare against. */
8288 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8289 /* Use it in this insn. */
8290 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8291 if (apply_change_group ())
8295 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
8297 if (invariant_p (arg) == 1)
8299 /* Look for giv with constant positive mult_val and nonconst
8300 add_val. Insert insns to compute new compare value.
8301 ??? Turn this off due to possible overflow. */
8303 for (v = bl->giv; v; v = v->next_iv)
8304 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
8305 && ! v->ignore && ! v->maybe_dead && v->always_computable
8311 /* If the giv V had the auto-inc address optimization applied
8312 to it, and INSN occurs between the giv insn and the biv
8313 insn, then we must adjust the value used here.
8314 This is rare, so we don't bother to do so. */
8316 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
8317 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
8318 || (INSN_LUID (v->insn) > INSN_LUID (insn)
8319 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
8325 tem = gen_reg_rtx (mode);
8327 /* Replace biv with giv's reduced register. */
8328 validate_change (insn, &XEXP (x, 1 - arg_operand),
8331 /* Compute value to compare against. */
8332 emit_iv_add_mult (arg, v->mult_val, v->add_val,
8334 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8335 if (apply_change_group ())
8340 /* This code has problems. Basically, you can't know when
8341 seeing if we will eliminate BL, whether a particular giv
8342 of ARG will be reduced. If it isn't going to be reduced,
8343 we can't eliminate BL. We can try forcing it to be reduced,
8344 but that can generate poor code.
8346 The problem is that the benefit of reducing TV, below should
8347 be increased if BL can actually be eliminated, but this means
8348 we might have to do a topological sort of the order in which
8349 we try to process biv. It doesn't seem worthwhile to do
8350 this sort of thing now. */
8353 /* Otherwise the reg compared with had better be a biv. */
8354 if (GET_CODE (arg) != REG
8355 || REG_IV_TYPE (REGNO (arg)) != BASIC_INDUCT)
8358 /* Look for a pair of givs, one for each biv,
8359 with identical coefficients. */
8360 for (v = bl->giv; v; v = v->next_iv)
8362 struct induction *tv;
8364 if (v->ignore || v->maybe_dead || v->mode != mode)
8367 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
8368 if (! tv->ignore && ! tv->maybe_dead
8369 && rtx_equal_p (tv->mult_val, v->mult_val)
8370 && rtx_equal_p (tv->add_val, v->add_val)
8371 && tv->mode == mode)
8373 /* If the giv V had the auto-inc address optimization applied
8374 to it, and INSN occurs between the giv insn and the biv
8375 insn, then we must adjust the value used here.
8376 This is rare, so we don't bother to do so. */
8378 && ((INSN_LUID (v->insn) < INSN_LUID (insn)
8379 && INSN_LUID (insn) < INSN_LUID (bl->biv->insn))
8380 || (INSN_LUID (v->insn) > INSN_LUID (insn)
8381 && INSN_LUID (insn) > INSN_LUID (bl->biv->insn))))
8387 /* Replace biv with its giv's reduced reg. */
8388 XEXP (x, 1-arg_operand) = v->new_reg;
8389 /* Replace other operand with the other giv's
8391 XEXP (x, arg_operand) = tv->new_reg;
8398 /* If we get here, the biv can't be eliminated. */
8402 /* If this address is a DEST_ADDR giv, it doesn't matter if the
8403 biv is used in it, since it will be replaced. */
8404 for (v = bl->giv; v; v = v->next_iv)
8405 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
8413 /* See if any subexpression fails elimination. */
8414 fmt = GET_RTX_FORMAT (code);
8415 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
8420 if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl,
8421 eliminate_p, where))
8426 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8427 if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl,
8428 eliminate_p, where))
8437 /* Return nonzero if the last use of REG
8438 is in an insn following INSN in the same basic block. */
8441 last_use_this_basic_block (reg, insn)
8447 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
8450 if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
8456 /* Called via `note_stores' to record the initial value of a biv. Here we
8457 just record the location of the set and process it later. */
8460 record_initial (dest, set)
8464 struct iv_class *bl;
8466 if (GET_CODE (dest) != REG
8467 || REGNO (dest) >= max_reg_before_loop
8468 || REG_IV_TYPE (REGNO (dest)) != BASIC_INDUCT)
8471 bl = reg_biv_class[REGNO (dest)];
8473 /* If this is the first set found, record it. */
8474 if (bl->init_insn == 0)
8476 bl->init_insn = note_insn;
8481 /* If any of the registers in X are "old" and currently have a last use earlier
8482 than INSN, update them to have a last use of INSN. Their actual last use
8483 will be the previous insn but it will not have a valid uid_luid so we can't
8487 update_reg_last_use (x, insn)
8491 /* Check for the case where INSN does not have a valid luid. In this case,
8492 there is no need to modify the regno_last_uid, as this can only happen
8493 when code is inserted after the loop_end to set a pseudo's final value,
8494 and hence this insn will never be the last use of x. */
8495 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
8496 && INSN_UID (insn) < max_uid_for_loop
8497 && uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)])
8498 REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
8502 register char *fmt = GET_RTX_FORMAT (GET_CODE (x));
8503 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
8506 update_reg_last_use (XEXP (x, i), insn);
8507 else if (fmt[i] == 'E')
8508 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8509 update_reg_last_use (XVECEXP (x, i, j), insn);
8514 /* Given a jump insn JUMP, return the condition that will cause it to branch
8515 to its JUMP_LABEL. If the condition cannot be understood, or is an
8516 inequality floating-point comparison which needs to be reversed, 0 will
8519 If EARLIEST is non-zero, it is a pointer to a place where the earliest
8520 insn used in locating the condition was found. If a replacement test
8521 of the condition is desired, it should be placed in front of that
8522 insn and we will be sure that the inputs are still valid.
8524 The condition will be returned in a canonical form to simplify testing by
8525 callers. Specifically:
8527 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
8528 (2) Both operands will be machine operands; (cc0) will have been replaced.
8529 (3) If an operand is a constant, it will be the second operand.
8530 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
8531 for GE, GEU, and LEU. */
8534 get_condition (jump, earliest)
8543 int reverse_code = 0;
8544 int did_reverse_condition = 0;
8545 enum machine_mode mode;
8547 /* If this is not a standard conditional jump, we can't parse it. */
8548 if (GET_CODE (jump) != JUMP_INSN
8549 || ! condjump_p (jump) || simplejump_p (jump))
8552 code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0));
8553 mode = GET_MODE (XEXP (SET_SRC (PATTERN (jump)), 0));
8554 op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0);
8555 op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1);
8560 /* If this branches to JUMP_LABEL when the condition is false, reverse
8562 if (GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 2)) == LABEL_REF
8563 && XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump))
8564 code = reverse_condition (code), did_reverse_condition ^= 1;
8566 /* If we are comparing a register with zero, see if the register is set
8567 in the previous insn to a COMPARE or a comparison operation. Perform
8568 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
8571 while (GET_RTX_CLASS (code) == '<' && op1 == CONST0_RTX (GET_MODE (op0)))
8573 /* Set non-zero when we find something of interest. */
8577 /* If comparison with cc0, import actual comparison from compare
8581 if ((prev = prev_nonnote_insn (prev)) == 0
8582 || GET_CODE (prev) != INSN
8583 || (set = single_set (prev)) == 0
8584 || SET_DEST (set) != cc0_rtx)
8587 op0 = SET_SRC (set);
8588 op1 = CONST0_RTX (GET_MODE (op0));
8594 /* If this is a COMPARE, pick up the two things being compared. */
8595 if (GET_CODE (op0) == COMPARE)
8597 op1 = XEXP (op0, 1);
8598 op0 = XEXP (op0, 0);
8601 else if (GET_CODE (op0) != REG)
8604 /* Go back to the previous insn. Stop if it is not an INSN. We also
8605 stop if it isn't a single set or if it has a REG_INC note because
8606 we don't want to bother dealing with it. */
8608 if ((prev = prev_nonnote_insn (prev)) == 0
8609 || GET_CODE (prev) != INSN
8610 || FIND_REG_INC_NOTE (prev, 0)
8611 || (set = single_set (prev)) == 0)
8614 /* If this is setting OP0, get what it sets it to if it looks
8616 if (rtx_equal_p (SET_DEST (set), op0))
8618 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
8620 /* ??? We may not combine comparisons done in a CCmode with
8621 comparisons not done in a CCmode. This is to aid targets
8622 like Alpha that have an IEEE compliant EQ instruction, and
8623 a non-IEEE compliant BEQ instruction. The use of CCmode is
8624 actually artificial, simply to prevent the combination, but
8625 should not affect other platforms.
8627 However, we must allow VOIDmode comparisons to match either
8628 CCmode or non-CCmode comparison, because some ports have
8629 modeless comparisons inside branch patterns.
8631 ??? This mode check should perhaps look more like the mode check
8632 in simplify_comparison in combine. */
8634 if ((GET_CODE (SET_SRC (set)) == COMPARE
8637 && GET_MODE_CLASS (inner_mode) == MODE_INT
8638 && (GET_MODE_BITSIZE (inner_mode)
8639 <= HOST_BITS_PER_WIDE_INT)
8640 && (STORE_FLAG_VALUE
8641 & ((HOST_WIDE_INT) 1
8642 << (GET_MODE_BITSIZE (inner_mode) - 1))))
8643 #ifdef FLOAT_STORE_FLAG_VALUE
8645 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
8646 && FLOAT_STORE_FLAG_VALUE < 0)
8649 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'))
8650 && (((GET_MODE_CLASS (mode) == MODE_CC)
8651 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
8652 || mode == VOIDmode || inner_mode == VOIDmode))
8654 else if (((code == EQ
8656 && (GET_MODE_BITSIZE (inner_mode)
8657 <= HOST_BITS_PER_WIDE_INT)
8658 && GET_MODE_CLASS (inner_mode) == MODE_INT
8659 && (STORE_FLAG_VALUE
8660 & ((HOST_WIDE_INT) 1
8661 << (GET_MODE_BITSIZE (inner_mode) - 1))))
8662 #ifdef FLOAT_STORE_FLAG_VALUE
8664 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
8665 && FLOAT_STORE_FLAG_VALUE < 0)
8668 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'
8669 && (((GET_MODE_CLASS (mode) == MODE_CC)
8670 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
8671 || mode == VOIDmode || inner_mode == VOIDmode))
8674 /* We might have reversed a LT to get a GE here. But this wasn't
8675 actually the comparison of data, so we don't flag that we
8676 have had to reverse the condition. */
8677 did_reverse_condition ^= 1;
8685 else if (reg_set_p (op0, prev))
8686 /* If this sets OP0, but not directly, we have to give up. */
8691 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
8692 code = GET_CODE (x);
8695 code = reverse_condition (code);
8696 did_reverse_condition ^= 1;
8700 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
8706 /* If constant is first, put it last. */
8707 if (CONSTANT_P (op0))
8708 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
8710 /* If OP0 is the result of a comparison, we weren't able to find what
8711 was really being compared, so fail. */
8712 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
8715 /* Canonicalize any ordered comparison with integers involving equality
8716 if we can do computations in the relevant mode and we do not
8719 if (GET_CODE (op1) == CONST_INT
8720 && GET_MODE (op0) != VOIDmode
8721 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
8723 HOST_WIDE_INT const_val = INTVAL (op1);
8724 unsigned HOST_WIDE_INT uconst_val = const_val;
8725 unsigned HOST_WIDE_INT max_val
8726 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
8731 if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
8732 code = LT, op1 = GEN_INT (const_val + 1);
8735 /* When cross-compiling, const_val might be sign-extended from
8736 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
8738 if ((HOST_WIDE_INT) (const_val & max_val)
8739 != (((HOST_WIDE_INT) 1
8740 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
8741 code = GT, op1 = GEN_INT (const_val - 1);
8745 if (uconst_val < max_val)
8746 code = LTU, op1 = GEN_INT (uconst_val + 1);
8750 if (uconst_val != 0)
8751 code = GTU, op1 = GEN_INT (uconst_val - 1);
8759 /* If this was floating-point and we reversed anything other than an
8760 EQ or NE, return zero. */
8761 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
8762 && did_reverse_condition && code != NE && code != EQ
8764 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
8768 /* Never return CC0; return zero instead. */
8773 return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);
8776 /* Similar to above routine, except that we also put an invariant last
8777 unless both operands are invariants. */
8780 get_condition_for_loop (x)
8783 rtx comparison = get_condition (x, NULL_PTR);
8786 || ! invariant_p (XEXP (comparison, 0))
8787 || invariant_p (XEXP (comparison, 1)))
8790 return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode,
8791 XEXP (comparison, 1), XEXP (comparison, 0));
8794 #ifdef HAVE_decrement_and_branch_on_count
8795 /* Instrument loop for insertion of bct instruction. We distinguish between
8796 loops with compile-time bounds and those with run-time bounds.
8797 Information from loop_iterations() is used to compute compile-time bounds.
8798 Run-time bounds should use loop preconditioning, but currently ignored.
8802 insert_bct (loop_start, loop_end, loop_info)
8803 rtx loop_start, loop_end;
8804 struct loop_info *loop_info;
8807 unsigned HOST_WIDE_INT n_iterations;
8810 int increment_direction, compare_direction;
8812 /* If the loop condition is <= or >=, the number of iteration
8813 is 1 more than the range of the bounds of the loop. */
8814 int add_iteration = 0;
8816 enum machine_mode loop_var_mode = word_mode;
8819 int loop_num = uid_loop_num [INSN_UID (loop_start)];
8821 /* It's impossible to instrument a competely unrolled loop. */
8822 if (loop_info->unroll_number == -1)
8825 /* Make sure that the count register is not in use. */
8826 if (loop_used_count_register [loop_num])
8828 if (loop_dump_stream)
8829 fprintf (loop_dump_stream,
8830 "insert_bct %d: BCT instrumentation failed: count register already in use\n",
8835 /* Make sure that the function has no indirect jumps. */
8836 if (indirect_jump_in_function)
8838 if (loop_dump_stream)
8839 fprintf (loop_dump_stream,
8840 "insert_bct %d: BCT instrumentation failed: indirect jump in function\n",
8845 /* Make sure that the last loop insn is a conditional jump. */
8846 if (GET_CODE (PREV_INSN (loop_end)) != JUMP_INSN
8847 || ! condjump_p (PREV_INSN (loop_end))
8848 || simplejump_p (PREV_INSN (loop_end)))
8850 if (loop_dump_stream)
8851 fprintf (loop_dump_stream,
8852 "insert_bct %d: BCT instrumentation failed: invalid jump at loop end\n",
8857 /* Make sure that the loop does not contain a function call
8858 (the count register might be altered by the called function). */
8861 if (loop_dump_stream)
8862 fprintf (loop_dump_stream,
8863 "insert_bct %d: BCT instrumentation failed: function call in loop\n",
8868 /* Make sure that the loop does not jump via a table.
8869 (the count register might be used to perform the branch on table). */
8870 if (loop_has_tablejump)
8872 if (loop_dump_stream)
8873 fprintf (loop_dump_stream,
8874 "insert_bct %d: BCT instrumentation failed: computed branch in the loop\n",
8879 /* Account for loop unrolling in instrumented iteration count. */
8880 if (loop_info->unroll_number > 1)
8881 n_iterations = loop_info->n_iterations / loop_info->unroll_number;
8883 n_iterations = loop_info->n_iterations;
8885 if (n_iterations != 0 && n_iterations < 3)
8887 /* Allow an enclosing outer loop to benefit if possible. */
8888 if (loop_dump_stream)
8889 fprintf (loop_dump_stream,
8890 "insert_bct %d: Too few iterations to benefit from BCT optimization\n",
8895 /* Try to instrument the loop. */
8897 /* Handle the simpler case, where the bounds are known at compile time. */
8898 if (n_iterations > 0)
8900 /* Mark all enclosing loops that they cannot use count register. */
8901 for (i = loop_num; i != -1; i = loop_outer_loop[i])
8902 loop_used_count_register[i] = 1;
8903 instrument_loop_bct (loop_start, loop_end, GEN_INT (n_iterations));
8907 /* Handle the more complex case, that the bounds are NOT known
8908 at compile time. In this case we generate run_time calculation
8909 of the number of iterations. */
8911 if (loop_info->iteration_var == 0)
8913 if (loop_dump_stream)
8914 fprintf (loop_dump_stream,
8915 "insert_bct %d: BCT Runtime Instrumentation failed: no loop iteration variable found\n",
8920 if (GET_MODE_CLASS (GET_MODE (loop_info->iteration_var)) != MODE_INT
8921 || GET_MODE_SIZE (GET_MODE (loop_info->iteration_var)) != UNITS_PER_WORD)
8923 if (loop_dump_stream)
8924 fprintf (loop_dump_stream,
8925 "insert_bct %d: BCT Runtime Instrumentation failed: loop variable not integer\n",
8930 /* With runtime bounds, if the compare is of the form '!=' we give up */
8931 if (loop_info->comparison_code == NE)
8933 if (loop_dump_stream)
8934 fprintf (loop_dump_stream,
8935 "insert_bct %d: BCT Runtime Instrumentation failed: runtime bounds with != comparison\n",
8939 /* Use common loop preconditioning code instead. */
8943 /* We rely on the existence of run-time guard to ensure that the
8944 loop executes at least once. */
8946 rtx iterations_num_reg;
8948 unsigned HOST_WIDE_INT increment_value_abs
8949 = INTVAL (increment) * increment_direction;
8951 /* make sure that the increment is a power of two, otherwise (an
8952 expensive) divide is needed. */
8953 if (exact_log2 (increment_value_abs) == -1)
8955 if (loop_dump_stream)
8956 fprintf (loop_dump_stream,
8957 "insert_bct: not instrumenting BCT because the increment is not power of 2\n");
8961 /* compute the number of iterations */
8966 /* Again, the number of iterations is calculated by:
8968 ; compare-val - initial-val + (increment -1) + additional-iteration
8969 ; num_iterations = -----------------------------------------------------------------
8972 /* ??? Do we have to call copy_rtx here before passing rtx to
8974 if (compare_direction > 0)
8976 /* <, <= :the loop variable is increasing */
8977 temp_reg = expand_binop (loop_var_mode, sub_optab,
8978 comparison_value, initial_value,
8979 NULL_RTX, 0, OPTAB_LIB_WIDEN);
8983 temp_reg = expand_binop (loop_var_mode, sub_optab,
8984 initial_value, comparison_value,
8985 NULL_RTX, 0, OPTAB_LIB_WIDEN);
8988 if (increment_value_abs - 1 + add_iteration != 0)
8989 temp_reg = expand_binop (loop_var_mode, add_optab, temp_reg,
8990 GEN_INT (increment_value_abs - 1
8992 NULL_RTX, 0, OPTAB_LIB_WIDEN);
8994 if (increment_value_abs != 1)
8996 /* ??? This will generate an expensive divide instruction for
8997 most targets. The original authors apparently expected this
8998 to be a shift, since they test for power-of-2 divisors above,
8999 but just naively generating a divide instruction will not give
9000 a shift. It happens to work for the PowerPC target because
9001 the rs6000.md file has a divide pattern that emits shifts.
9002 It will probably not work for any other target. */
9003 iterations_num_reg = expand_binop (loop_var_mode, sdiv_optab,
9005 GEN_INT (increment_value_abs),
9006 NULL_RTX, 0, OPTAB_LIB_WIDEN);
9009 iterations_num_reg = temp_reg;
9011 sequence = gen_sequence ();
9013 emit_insn_before (sequence, loop_start);
9014 instrument_loop_bct (loop_start, loop_end, iterations_num_reg);
9018 #endif /* Complex case */
9021 /* Instrument loop by inserting a bct in it as follows:
9022 1. A new counter register is created.
9023 2. In the head of the loop the new variable is initialized to the value
9024 passed in the loop_num_iterations parameter.
9025 3. At the end of the loop, comparison of the register with 0 is generated.
9026 The created comparison follows the pattern defined for the
9027 decrement_and_branch_on_count insn, so this insn will be generated.
9028 4. The branch on the old variable are deleted. The compare must remain
9029 because it might be used elsewhere. If the loop-variable or condition
9030 register are used elsewhere, they will be eliminated by flow. */
9033 instrument_loop_bct (loop_start, loop_end, loop_num_iterations)
9034 rtx loop_start, loop_end;
9035 rtx loop_num_iterations;
9041 if (HAVE_decrement_and_branch_on_count)
9043 if (loop_dump_stream)
9045 fputs ("instrument_bct: Inserting BCT (", loop_dump_stream);
9046 if (GET_CODE (loop_num_iterations) == CONST_INT)
9047 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC,
9048 INTVAL (loop_num_iterations));
9050 fputs ("runtime", loop_dump_stream);
9051 fputs (" iterations)", loop_dump_stream);
9054 /* Discard original jump to continue loop. Original compare result
9055 may still be live, so it cannot be discarded explicitly. */
9056 delete_insn (PREV_INSN (loop_end));
9058 /* Insert the label which will delimit the start of the loop. */
9059 start_label = gen_label_rtx ();
9060 emit_label_after (start_label, loop_start);
9062 /* Insert initialization of the count register into the loop header. */
9064 counter_reg = gen_reg_rtx (word_mode);
9065 emit_insn (gen_move_insn (counter_reg, loop_num_iterations));
9066 sequence = gen_sequence ();
9068 emit_insn_before (sequence, loop_start);
9070 /* Insert new comparison on the count register instead of the
9071 old one, generating the needed BCT pattern (that will be
9072 later recognized by assembly generation phase). */
9073 emit_jump_insn_before (gen_decrement_and_branch_on_count (counter_reg,
9076 LABEL_NUSES (start_label)++;
9080 #endif /* HAVE_decrement_and_branch_on_count */
9082 /* Scan the function and determine whether it has indirect (computed) jumps.
9084 This is taken mostly from flow.c; similar code exists elsewhere
9085 in the compiler. It may be useful to put this into rtlanal.c. */
9087 indirect_jump_in_function_p (start)
9092 for (insn = start; insn; insn = NEXT_INSN (insn))
9093 if (computed_jump_p (insn))
9099 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
9100 documentation for LOOP_MEMS for the definition of `appropriate'.
9101 This function is called from prescan_loop via for_each_rtx. */
9104 insert_loop_mem (mem, data)
9106 void *data ATTRIBUTE_UNUSED;
9114 switch (GET_CODE (m))
9120 /* We're not interested in the MEM associated with a
9121 CONST_DOUBLE, so there's no need to traverse into this. */
9125 /* This is not a MEM. */
9129 /* See if we've already seen this MEM. */
9130 for (i = 0; i < loop_mems_idx; ++i)
9131 if (rtx_equal_p (m, loop_mems[i].mem))
9133 if (GET_MODE (m) != GET_MODE (loop_mems[i].mem))
9134 /* The modes of the two memory accesses are different. If
9135 this happens, something tricky is going on, and we just
9136 don't optimize accesses to this MEM. */
9137 loop_mems[i].optimize = 0;
9142 /* Resize the array, if necessary. */
9143 if (loop_mems_idx == loop_mems_allocated)
9145 if (loop_mems_allocated != 0)
9146 loop_mems_allocated *= 2;
9148 loop_mems_allocated = 32;
9150 loop_mems = (loop_mem_info*)
9151 xrealloc (loop_mems,
9152 loop_mems_allocated * sizeof (loop_mem_info));
9155 /* Actually insert the MEM. */
9156 loop_mems[loop_mems_idx].mem = m;
9157 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
9158 because we can't put it in a register. We still store it in the
9159 table, though, so that if we see the same address later, but in a
9160 non-BLK mode, we'll not think we can optimize it at that point. */
9161 loop_mems[loop_mems_idx].optimize = (GET_MODE (m) != BLKmode);
9162 loop_mems[loop_mems_idx].reg = NULL_RTX;
9168 /* Like load_mems, but also ensures that SET_IN_LOOP,
9169 MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct
9170 values after load_mems. */
9173 load_mems_and_recount_loop_regs_set (scan_start, end, loop_top, start,
9174 reg_single_usage, insn_count)
9179 varray_type reg_single_usage;
9182 int nregs = max_reg_num ();
9184 load_mems (scan_start, end, loop_top, start);
9186 /* Recalculate set_in_loop and friends since load_mems may have
9187 created new registers. */
9188 if (max_reg_num () > nregs)
9194 nregs = max_reg_num ();
9196 if ((unsigned) nregs > set_in_loop->num_elements)
9198 /* Grow all the arrays. */
9199 VARRAY_GROW (set_in_loop, nregs);
9200 VARRAY_GROW (n_times_set, nregs);
9201 VARRAY_GROW (may_not_optimize, nregs);
9202 if (reg_single_usage)
9203 VARRAY_GROW (reg_single_usage, nregs);
9205 /* Clear the arrays */
9206 bzero ((char *) &set_in_loop->data, nregs * sizeof (int));
9207 bzero ((char *) &may_not_optimize->data, nregs * sizeof (char));
9208 if (reg_single_usage)
9209 bzero ((char *) ®_single_usage->data, nregs * sizeof (rtx));
9211 count_loop_regs_set (loop_top ? loop_top : start, end,
9212 may_not_optimize, reg_single_usage,
9215 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
9217 VARRAY_CHAR (may_not_optimize, i) = 1;
9218 VARRAY_INT (set_in_loop, i) = 1;
9221 #ifdef AVOID_CCMODE_COPIES
9222 /* Don't try to move insns which set CC registers if we should not
9223 create CCmode register copies. */
9224 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
9225 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
9226 VARRAY_CHAR (may_not_optimize, i) = 1;
9229 /* Set n_times_set for the new registers. */
9230 bcopy ((char *) (&set_in_loop->data.i[0] + old_nregs),
9231 (char *) (&n_times_set->data.i[0] + old_nregs),
9232 (nregs - old_nregs) * sizeof (int));
9236 /* Move MEMs into registers for the duration of the loop. SCAN_START
9237 is the first instruction in the loop (as it is executed). The
9238 other parameters are as for next_insn_in_loop. */
9241 load_mems (scan_start, end, loop_top, start)
9247 int maybe_never = 0;
9250 rtx label = NULL_RTX;
9253 if (loop_mems_idx > 0)
9255 /* Nonzero if the next instruction may never be executed. */
9256 int next_maybe_never = 0;
9258 /* Check to see if it's possible that some instructions in the
9259 loop are never executed. */
9260 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
9261 p != NULL_RTX && !maybe_never;
9262 p = next_insn_in_loop (p, scan_start, end, loop_top))
9264 if (GET_CODE (p) == CODE_LABEL)
9266 else if (GET_CODE (p) == JUMP_INSN
9267 /* If we enter the loop in the middle, and scan
9268 around to the beginning, don't set maybe_never
9269 for that. This must be an unconditional jump,
9270 otherwise the code at the top of the loop might
9271 never be executed. Unconditional jumps are
9272 followed a by barrier then loop end. */
9273 && ! (GET_CODE (p) == JUMP_INSN
9274 && JUMP_LABEL (p) == loop_top
9275 && NEXT_INSN (NEXT_INSN (p)) == end
9276 && simplejump_p (p)))
9278 if (!condjump_p (p))
9279 /* Something complicated. */
9282 /* If there are any more instructions in the loop, they
9283 might not be reached. */
9284 next_maybe_never = 1;
9286 else if (next_maybe_never)
9290 /* Actually move the MEMs. */
9291 for (i = 0; i < loop_mems_idx; ++i)
9295 rtx mem = loop_mems[i].mem;
9298 if (MEM_VOLATILE_P (mem)
9299 || invariant_p (XEXP (mem, 0)) != 1)
9300 /* There's no telling whether or not MEM is modified. */
9301 loop_mems[i].optimize = 0;
9303 /* Go through the MEMs written to in the loop to see if this
9304 one is aliased by one of them. */
9305 mem_list_entry = loop_store_mems;
9306 while (mem_list_entry)
9308 if (rtx_equal_p (mem, XEXP (mem_list_entry, 0)))
9310 else if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
9313 /* MEM is indeed aliased by this store. */
9314 loop_mems[i].optimize = 0;
9317 mem_list_entry = XEXP (mem_list_entry, 1);
9320 /* If this MEM is written to, we must be sure that there
9321 are no reads from another MEM that aliases this one. */
9322 if (loop_mems[i].optimize && written)
9326 for (j = 0; j < loop_mems_idx; ++j)
9330 else if (true_dependence (mem,
9335 /* It's not safe to hoist loop_mems[i] out of
9336 the loop because writes to it might not be
9337 seen by reads from loop_mems[j]. */
9338 loop_mems[i].optimize = 0;
9344 if (maybe_never && may_trap_p (mem))
9345 /* We can't access the MEM outside the loop; it might
9346 cause a trap that wouldn't have happened otherwise. */
9347 loop_mems[i].optimize = 0;
9349 if (!loop_mems[i].optimize)
9350 /* We thought we were going to lift this MEM out of the
9351 loop, but later discovered that we could not. */
9354 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
9355 order to keep scan_loop from moving stores to this MEM
9356 out of the loop just because this REG is neither a
9357 user-variable nor used in the loop test. */
9358 reg = gen_reg_rtx (GET_MODE (mem));
9359 REG_USERVAR_P (reg) = 1;
9360 loop_mems[i].reg = reg;
9362 /* Now, replace all references to the MEM with the
9363 corresponding pesudos. */
9364 for (p = next_insn_in_loop (scan_start, scan_start, end, loop_top);
9366 p = next_insn_in_loop (p, scan_start, end, loop_top))
9371 for_each_rtx (&p, replace_loop_mem, &ri);
9374 if (!apply_change_group ())
9375 /* We couldn't replace all occurrences of the MEM. */
9376 loop_mems[i].optimize = 0;
9381 /* Load the memory immediately before START, which is
9382 the NOTE_LOOP_BEG. */
9383 set = gen_rtx_SET (GET_MODE (reg), reg, mem);
9384 emit_insn_before (set, start);
9388 if (label == NULL_RTX)
9390 /* We must compute the former
9391 right-after-the-end label before we insert
9393 end_label = next_label (end);
9394 label = gen_label_rtx ();
9395 emit_label_after (label, end);
9398 /* Store the memory immediately after END, which is
9399 the NOTE_LOOP_END. */
9400 set = gen_rtx_SET (GET_MODE (reg), copy_rtx (mem), reg);
9401 emit_insn_after (set, label);
9404 if (loop_dump_stream)
9406 fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
9407 REGNO (reg), (written ? "r/w" : "r/o"));
9408 print_rtl (loop_dump_stream, mem);
9409 fputc ('\n', loop_dump_stream);
9415 if (label != NULL_RTX)
9417 /* Now, we need to replace all references to the previous exit
9418 label with the new one. */
9423 for (p = start; p != end; p = NEXT_INSN (p))
9425 for_each_rtx (&p, replace_label, &rr);
9427 /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
9428 field. This is not handled by for_each_rtx because it doesn't
9429 handle unprinted ('0') fields. We need to update JUMP_LABEL
9430 because the immediately following unroll pass will use it.
9431 replace_label would not work anyways, because that only handles
9433 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == end_label)
9434 JUMP_LABEL (p) = label;
9439 /* Replace MEM with its associated pseudo register. This function is
9440 called from load_mems via for_each_rtx. DATA is actually an
9441 rtx_and_int * describing the instruction currently being scanned
9442 and the MEM we are currently replacing. */
9445 replace_loop_mem (mem, data)
9457 switch (GET_CODE (m))
9463 /* We're not interested in the MEM associated with a
9464 CONST_DOUBLE, so there's no need to traverse into one. */
9468 /* This is not a MEM. */
9472 ri = (rtx_and_int*) data;
9475 if (!rtx_equal_p (loop_mems[i].mem, m))
9476 /* This is not the MEM we are currently replacing. */
9481 /* Actually replace the MEM. */
9482 validate_change (insn, mem, loop_mems[i].reg, 1);
9487 /* Replace occurrences of the old exit label for the loop with the new
9488 one. DATA is an rtx_pair containing the old and new labels,
9492 replace_label (x, data)
9497 rtx old_label = ((rtx_pair*) data)->r1;
9498 rtx new_label = ((rtx_pair*) data)->r2;
9503 if (GET_CODE (l) != LABEL_REF)
9506 if (XEXP (l, 0) != old_label)
9509 XEXP (l, 0) = new_label;
9510 ++LABEL_NUSES (new_label);
9511 --LABEL_NUSES (old_label);