1 /* Perform various loop optimizations, including strength reduction.
2 Copyright (C) 1987, 1988, 1989, 1991, 1992, 1993, 1994, 1995, 1996, 1997,
3 1998, 1999, 2000 Free Software Foundation, Inc.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
22 /* This is the loop optimization pass of the compiler.
23 It finds invariant computations within loops and moves them
24 to the beginning of the loop. Then it identifies basic and
25 general induction variables. Strength reduction is applied to the general
26 induction variables, and induction variable elimination is applied to
27 the basic induction variables.
29 It also finds cases where
30 a register is set within the loop by zero-extending a narrower value
31 and changes these to zero the entire register once before the loop
32 and merely copy the low part within the loop.
34 Most of the complexity is in heuristics to decide when it is worth
35 while to do these things. */
44 #include "hard-reg-set.h"
45 #include "basic-block.h"
46 #include "insn-config.h"
47 #include "insn-flags.h"
57 /* Vector mapping INSN_UIDs to luids.
58 The luids are like uids but increase monotonically always.
59 We use them to see whether a jump comes from outside a given loop. */
63 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
64 number the insn is contained in. */
66 struct loop **uid_loop;
68 /* 1 + largest uid of any insn. */
72 /* 1 + luid of last insn. */
76 /* Number of loops detected in current function. Used as index to the
79 static int max_loop_num;
81 /* Indexed by register number, contains the number of times the reg
82 is set during the loop being scanned.
83 During code motion, a negative value indicates a reg that has been
84 made a candidate; in particular -2 means that it is an candidate that
85 we know is equal to a constant and -1 means that it is an candidate
86 not known equal to a constant.
87 After code motion, regs moved have 0 (which is accurate now)
88 while the failed candidates have the original number of times set.
90 Therefore, at all times, == 0 indicates an invariant register;
91 < 0 a conditionally invariant one. */
93 static varray_type set_in_loop;
95 /* Original value of set_in_loop; same except that this value
96 is not set negative for a reg whose sets have been made candidates
97 and not set to 0 for a reg that is moved. */
99 static varray_type n_times_set;
101 /* Index by register number, 1 indicates that the register
102 cannot be moved or strength reduced. */
104 static varray_type may_not_optimize;
106 /* Contains the insn in which a register was used if it was used
107 exactly once; contains const0_rtx if it was used more than once. */
109 static varray_type reg_single_usage;
111 /* Nonzero means reg N has already been moved out of one loop.
112 This reduces the desire to move it out of another. */
114 static char *moved_once;
116 /* List of MEMs that are stored in this loop. */
118 static rtx loop_store_mems;
120 /* The insn where the first of these was found. */
121 static rtx first_loop_store_insn;
123 typedef struct loop_mem_info {
124 rtx mem; /* The MEM itself. */
125 rtx reg; /* Corresponding pseudo, if any. */
126 int optimize; /* Nonzero if we can optimize access to this MEM. */
129 /* Array of MEMs that are used (read or written) in this loop, but
130 cannot be aliased by anything in this loop, except perhaps
131 themselves. In other words, if loop_mems[i] is altered during the
132 loop, it is altered by an expression that is rtx_equal_p to it. */
134 static loop_mem_info *loop_mems;
136 /* The index of the next available slot in LOOP_MEMS. */
138 static int loop_mems_idx;
140 /* The number of elements allocated in LOOP_MEMs. */
142 static int loop_mems_allocated;
144 /* Nonzero if we don't know what MEMs were changed in the current
145 loop. This happens if the loop contains a call (in which case
146 `loop_info->has_call' will also be set) or if we store into more
147 than NUM_STORES MEMs. */
149 static int unknown_address_altered;
151 /* The above doesn't count any readonly memory locations that are stored.
154 static int unknown_constant_address_altered;
156 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
157 static int num_movables;
159 /* Count of memory write instructions discovered in the loop. */
160 static int num_mem_sets;
162 /* Bound on pseudo register number before loop optimization.
163 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
164 unsigned int max_reg_before_loop;
166 /* The value to pass to the next call of reg_scan_update. */
167 static int loop_max_reg;
169 /* This obstack is used in product_cheap_p to allocate its rtl. It
170 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
171 If we used the same obstack that it did, we would be deallocating
174 static struct obstack temp_obstack;
176 /* This is where the pointer to the obstack being used for RTL is stored. */
178 extern struct obstack *rtl_obstack;
180 #define obstack_chunk_alloc xmalloc
181 #define obstack_chunk_free free
183 /* During the analysis of a loop, a chain of `struct movable's
184 is made to record all the movable insns found.
185 Then the entire chain can be scanned to decide which to move. */
189 rtx insn; /* A movable insn */
190 rtx set_src; /* The expression this reg is set from. */
191 rtx set_dest; /* The destination of this SET. */
192 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
193 of any registers used within the LIBCALL. */
194 int consec; /* Number of consecutive following insns
195 that must be moved with this one. */
196 unsigned int regno; /* The register it sets */
197 short lifetime; /* lifetime of that register;
198 may be adjusted when matching movables
199 that load the same value are found. */
200 short savings; /* Number of insns we can move for this reg,
201 including other movables that force this
202 or match this one. */
203 unsigned int cond : 1; /* 1 if only conditionally movable */
204 unsigned int force : 1; /* 1 means MUST move this insn */
205 unsigned int global : 1; /* 1 means reg is live outside this loop */
206 /* If PARTIAL is 1, GLOBAL means something different:
207 that the reg is live outside the range from where it is set
208 to the following label. */
209 unsigned int done : 1; /* 1 inhibits further processing of this */
211 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
212 In particular, moving it does not make it
214 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
215 load SRC, rather than copying INSN. */
216 unsigned int move_insn_first:1;/* Same as above, if this is necessary for the
217 first insn of a consecutive sets group. */
218 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
219 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
220 that we should avoid changing when clearing
221 the rest of the reg. */
222 struct movable *match; /* First entry for same value */
223 struct movable *forces; /* An insn that must be moved if this is */
224 struct movable *next;
227 static struct movable *the_movables;
229 FILE *loop_dump_stream;
231 /* Forward declarations. */
233 static void verify_dominator PARAMS ((struct loop *));
234 static void find_and_verify_loops PARAMS ((rtx, struct loops *));
235 static void mark_loop_jump PARAMS ((rtx, struct loop *));
236 static void prescan_loop PARAMS ((struct loop *));
237 static int reg_in_basic_block_p PARAMS ((rtx, rtx));
238 static int consec_sets_invariant_p PARAMS ((const struct loop *,
240 static int labels_in_range_p PARAMS ((rtx, int));
241 static void count_one_set PARAMS ((rtx, rtx, varray_type, rtx *));
243 static void count_loop_regs_set PARAMS ((rtx, rtx, varray_type, varray_type,
245 static void note_addr_stored PARAMS ((rtx, rtx, void *));
246 static void note_set_pseudo_multiple_uses PARAMS ((rtx, rtx, void *));
247 static int loop_reg_used_before_p PARAMS ((const struct loop *, rtx, rtx));
248 static void scan_loop PARAMS ((struct loop*, int));
250 static void replace_call_address PARAMS ((rtx, rtx, rtx));
252 static rtx skip_consec_insns PARAMS ((rtx, int));
253 static int libcall_benefit PARAMS ((rtx));
254 static void ignore_some_movables PARAMS ((struct movable *));
255 static void force_movables PARAMS ((struct movable *));
256 static void combine_movables PARAMS ((struct movable *, int));
257 static int regs_match_p PARAMS ((rtx, rtx, struct movable *));
258 static int rtx_equal_for_loop_p PARAMS ((rtx, rtx, struct movable *));
259 static void add_label_notes PARAMS ((rtx, rtx));
260 static void move_movables PARAMS ((struct loop *loop, struct movable *,
262 static int count_nonfixed_reads PARAMS ((const struct loop *, rtx));
263 static void strength_reduce PARAMS ((struct loop *, int, int));
264 static void find_single_use_in_loop PARAMS ((rtx, rtx, varray_type));
265 static int valid_initial_value_p PARAMS ((rtx, rtx, int, rtx));
266 static void find_mem_givs PARAMS ((const struct loop *, rtx, rtx, int, int));
267 static void record_biv PARAMS ((struct induction *, rtx, rtx, rtx, rtx, rtx *,
269 static void check_final_value PARAMS ((const struct loop *,
270 struct induction *));
271 static void record_giv PARAMS ((const struct loop *, struct induction *,
272 rtx, rtx, rtx, rtx, rtx, int, enum g_types,
274 static void update_giv_derive PARAMS ((const struct loop *, rtx));
275 static int basic_induction_var PARAMS ((const struct loop *, rtx,
276 enum machine_mode, rtx, rtx,
277 rtx *, rtx *, rtx **));
278 static rtx simplify_giv_expr PARAMS ((const struct loop *, rtx, int *));
279 static int general_induction_var PARAMS ((const struct loop *loop, rtx, rtx *,
280 rtx *, rtx *, int, int *, enum machine_mode));
281 static int consec_sets_giv PARAMS ((const struct loop *, int, rtx,
282 rtx, rtx, rtx *, rtx *, rtx *));
283 static int check_dbra_loop PARAMS ((struct loop *, int));
284 static rtx express_from_1 PARAMS ((rtx, rtx, rtx));
285 static rtx combine_givs_p PARAMS ((struct induction *, struct induction *));
286 static void combine_givs PARAMS ((struct iv_class *));
287 struct recombine_givs_stats;
288 static int find_life_end PARAMS ((rtx, struct recombine_givs_stats *,
290 static void recombine_givs PARAMS ((const struct loop *, struct iv_class *,
292 static int product_cheap_p PARAMS ((rtx, rtx));
293 static int maybe_eliminate_biv PARAMS ((const struct loop *, struct iv_class *,
295 static int maybe_eliminate_biv_1 PARAMS ((const struct loop *, rtx, rtx,
296 struct iv_class *, int, rtx));
297 static int last_use_this_basic_block PARAMS ((rtx, rtx));
298 static void record_initial PARAMS ((rtx, rtx, void *));
299 static void update_reg_last_use PARAMS ((rtx, rtx));
300 static rtx next_insn_in_loop PARAMS ((const struct loop *, rtx));
301 static void load_mems_and_recount_loop_regs_set PARAMS ((const struct loop*,
303 static void load_mems PARAMS ((const struct loop *));
304 static int insert_loop_mem PARAMS ((rtx *, void *));
305 static int replace_loop_mem PARAMS ((rtx *, void *));
306 static int replace_loop_reg PARAMS ((rtx *, void *));
307 static void note_reg_stored PARAMS ((rtx, rtx, void *));
308 static void try_copy_prop PARAMS ((const struct loop *, rtx, unsigned int));
309 static void try_swap_copy_prop PARAMS ((const struct loop *, rtx,
311 static int replace_label PARAMS ((rtx *, void *));
312 static rtx check_insn_for_givs PARAMS((struct loop *, rtx, int, int));
313 static rtx check_insn_for_bivs PARAMS((struct loop *, rtx, int, int));
315 typedef struct rtx_and_int {
320 typedef struct rtx_pair {
325 /* Nonzero iff INSN is between START and END, inclusive. */
326 #define INSN_IN_RANGE_P(INSN, START, END) \
327 (INSN_UID (INSN) < max_uid_for_loop \
328 && INSN_LUID (INSN) >= INSN_LUID (START) \
329 && INSN_LUID (INSN) <= INSN_LUID (END))
331 /* Indirect_jump_in_function is computed once per function. */
332 static int indirect_jump_in_function;
333 static int indirect_jump_in_function_p PARAMS ((rtx));
335 static int compute_luids PARAMS ((rtx, rtx, int));
337 static int biv_elimination_giv_has_0_offset PARAMS ((struct induction *,
338 struct induction *, rtx));
340 /* Relative gain of eliminating various kinds of operations. */
343 static int shift_cost;
344 static int mult_cost;
347 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
348 copy the value of the strength reduced giv to its original register. */
349 static int copy_cost;
351 /* Cost of using a register, to normalize the benefits of a giv. */
352 static int reg_address_cost;
357 char *free_point = (char *) oballoc (1);
358 rtx reg = gen_rtx_REG (word_mode, LAST_VIRTUAL_REGISTER + 1);
360 add_cost = rtx_cost (gen_rtx_PLUS (word_mode, reg, reg), SET);
362 reg_address_cost = address_cost (reg, SImode);
364 /* We multiply by 2 to reconcile the difference in scale between
365 these two ways of computing costs. Otherwise the cost of a copy
366 will be far less than the cost of an add. */
370 /* Free the objects we just allocated. */
373 /* Initialize the obstack used for rtl in product_cheap_p. */
374 gcc_obstack_init (&temp_obstack);
377 /* Compute the mapping from uids to luids.
378 LUIDs are numbers assigned to insns, like uids,
379 except that luids increase monotonically through the code.
380 Start at insn START and stop just before END. Assign LUIDs
381 starting with PREV_LUID + 1. Return the last assigned LUID + 1. */
383 compute_luids (start, end, prev_luid)
390 for (insn = start, i = prev_luid; insn != end; insn = NEXT_INSN (insn))
392 if (INSN_UID (insn) >= max_uid_for_loop)
394 /* Don't assign luids to line-number NOTEs, so that the distance in
395 luids between two insns is not affected by -g. */
396 if (GET_CODE (insn) != NOTE
397 || NOTE_LINE_NUMBER (insn) <= 0)
398 uid_luid[INSN_UID (insn)] = ++i;
400 /* Give a line number note the same luid as preceding insn. */
401 uid_luid[INSN_UID (insn)] = i;
406 /* Entry point of this file. Perform loop optimization
407 on the current function. F is the first insn of the function
408 and DUMPFILE is a stream for output of a trace of actions taken
409 (or 0 if none should be output). */
412 loop_optimize (f, dumpfile, flags)
413 /* f is the first instruction of a chain of insns for one function */
420 struct loops loops_data;
421 struct loops *loops = &loops_data;
422 struct loop_info *loops_info;
424 loop_dump_stream = dumpfile;
426 init_recog_no_volatile ();
428 max_reg_before_loop = max_reg_num ();
429 loop_max_reg = max_reg_before_loop;
433 /* Count the number of loops. */
436 for (insn = f; insn; insn = NEXT_INSN (insn))
438 if (GET_CODE (insn) == NOTE
439 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
443 /* Don't waste time if no loops. */
444 if (max_loop_num == 0)
447 loops->num = max_loop_num;
449 moved_once = (char *) xcalloc (max_reg_before_loop, sizeof (char));
451 /* Get size to use for tables indexed by uids.
452 Leave some space for labels allocated by find_and_verify_loops. */
453 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 32;
455 uid_luid = (int *) xcalloc (max_uid_for_loop, sizeof (int));
456 uid_loop = (struct loop **) xcalloc (max_uid_for_loop,
457 sizeof (struct loop *));
459 /* Allocate storage for array of loops. */
460 loops->array = (struct loop *)
461 xcalloc (loops->num, sizeof (struct loop));
463 /* Find and process each loop.
464 First, find them, and record them in order of their beginnings. */
465 find_and_verify_loops (f, loops);
467 /* Allocate and initialize auxiliary loop information. */
468 loops_info = xcalloc (loops->num, sizeof (struct loop_info));
469 for (i = 0; i < loops->num; i++)
470 loops->array[i].aux = loops_info + i;
472 /* Now find all register lifetimes. This must be done after
473 find_and_verify_loops, because it might reorder the insns in the
475 reg_scan (f, max_reg_before_loop, 1);
477 /* This must occur after reg_scan so that registers created by gcse
478 will have entries in the register tables.
480 We could have added a call to reg_scan after gcse_main in toplev.c,
481 but moving this call to init_alias_analysis is more efficient. */
482 init_alias_analysis ();
484 /* See if we went too far. Note that get_max_uid already returns
485 one more that the maximum uid of all insn. */
486 if (get_max_uid () > max_uid_for_loop)
488 /* Now reset it to the actual size we need. See above. */
489 max_uid_for_loop = get_max_uid ();
491 /* find_and_verify_loops has already called compute_luids, but it
492 might have rearranged code afterwards, so we need to recompute
494 max_luid = compute_luids (f, NULL_RTX, 0);
496 /* Don't leave gaps in uid_luid for insns that have been
497 deleted. It is possible that the first or last insn
498 using some register has been deleted by cross-jumping.
499 Make sure that uid_luid for that former insn's uid
500 points to the general area where that insn used to be. */
501 for (i = 0; i < max_uid_for_loop; i++)
503 uid_luid[0] = uid_luid[i];
504 if (uid_luid[0] != 0)
507 for (i = 0; i < max_uid_for_loop; i++)
508 if (uid_luid[i] == 0)
509 uid_luid[i] = uid_luid[i - 1];
511 /* Determine if the function has indirect jump. On some systems
512 this prevents low overhead loop instructions from being used. */
513 indirect_jump_in_function = indirect_jump_in_function_p (f);
515 /* Now scan the loops, last ones first, since this means inner ones are done
516 before outer ones. */
517 for (i = max_loop_num - 1; i >= 0; i--)
519 struct loop *loop = &loops->array[i];
521 if (! loop->invalid && loop->end)
522 scan_loop (loop, flags);
525 /* If there were lexical blocks inside the loop, they have been
526 replicated. We will now have more than one NOTE_INSN_BLOCK_BEG
527 and NOTE_INSN_BLOCK_END for each such block. We must duplicate
528 the BLOCKs as well. */
529 if (write_symbols != NO_DEBUG)
532 end_alias_analysis ();
542 /* Returns the next insn, in execution order, after INSN. START and
543 END are the NOTE_INSN_LOOP_BEG and NOTE_INSN_LOOP_END for the loop,
544 respectively. LOOP->TOP, if non-NULL, is the top of the loop in the
545 insn-stream; it is used with loops that are entered near the
549 next_insn_in_loop (loop, insn)
550 const struct loop *loop;
553 insn = NEXT_INSN (insn);
555 if (insn == loop->end)
558 /* Go to the top of the loop, and continue there. */
565 if (insn == loop->scan_start)
572 /* Optimize one loop described by LOOP. */
574 /* ??? Could also move memory writes out of loops if the destination address
575 is invariant, the source is invariant, the memory write is not volatile,
576 and if we can prove that no read inside the loop can read this address
577 before the write occurs. If there is a read of this address after the
578 write, then we can also mark the memory read as invariant. */
581 scan_loop (loop, flags)
586 rtx loop_start = loop->start;
587 rtx loop_end = loop->end;
588 /* Additional information about the current loop being processed
589 that is used to compute the number of loop iterations for loop
590 unrolling and doloop optimization. */
591 struct loop_info *loop_info = LOOP_INFO (loop);
593 /* 1 if we are scanning insns that could be executed zero times. */
595 /* 1 if we are scanning insns that might never be executed
596 due to a subroutine call which might exit before they are reached. */
598 /* Jump insn that enters the loop, or 0 if control drops in. */
599 rtx loop_entry_jump = 0;
600 /* Number of insns in the loop. */
603 rtx temp, update_start, update_end;
604 /* The SET from an insn, if it is the only SET in the insn. */
606 /* Chain describing insns movable in current loop. */
607 struct movable *movables = 0;
608 /* Last element in `movables' -- so we can add elements at the end. */
609 struct movable *last_movable = 0;
610 /* Ratio of extra register life span we can justify
611 for saving an instruction. More if loop doesn't call subroutines
612 since in that case saving an insn makes more difference
613 and more registers are available. */
615 /* Nonzero if we are scanning instructions in a sub-loop. */
621 /* Determine whether this loop starts with a jump down to a test at
622 the end. This will occur for a small number of loops with a test
623 that is too complex to duplicate in front of the loop.
625 We search for the first insn or label in the loop, skipping NOTEs.
626 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
627 (because we might have a loop executed only once that contains a
628 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
629 (in case we have a degenerate loop).
631 Note that if we mistakenly think that a loop is entered at the top
632 when, in fact, it is entered at the exit test, the only effect will be
633 slightly poorer optimization. Making the opposite error can generate
634 incorrect code. Since very few loops now start with a jump to the
635 exit test, the code here to detect that case is very conservative. */
637 for (p = NEXT_INSN (loop_start);
639 && GET_CODE (p) != CODE_LABEL && ! INSN_P (p)
640 && (GET_CODE (p) != NOTE
641 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
642 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
646 loop->scan_start = p;
648 /* Set up variables describing this loop. */
650 threshold = (loop_info->has_call ? 1 : 2) * (1 + n_non_fixed_regs);
652 /* If loop has a jump before the first label,
653 the true entry is the target of that jump.
654 Start scan from there.
655 But record in LOOP->TOP the place where the end-test jumps
656 back to so we can scan that after the end of the loop. */
657 if (GET_CODE (p) == JUMP_INSN)
661 /* Loop entry must be unconditional jump (and not a RETURN) */
662 if (any_uncondjump_p (p)
663 && JUMP_LABEL (p) != 0
664 /* Check to see whether the jump actually
665 jumps out of the loop (meaning it's no loop).
666 This case can happen for things like
667 do {..} while (0). If this label was generated previously
668 by loop, we can't tell anything about it and have to reject
670 && INSN_IN_RANGE_P (JUMP_LABEL (p), loop_start, loop_end))
672 loop->top = next_label (loop->scan_start);
673 loop->scan_start = JUMP_LABEL (p);
677 /* If LOOP->SCAN_START was an insn created by loop, we don't know its luid
678 as required by loop_reg_used_before_p. So skip such loops. (This
679 test may never be true, but it's best to play it safe.)
681 Also, skip loops where we do not start scanning at a label. This
682 test also rejects loops starting with a JUMP_INSN that failed the
685 if (INSN_UID (loop->scan_start) >= max_uid_for_loop
686 || GET_CODE (loop->scan_start) != CODE_LABEL)
688 if (loop_dump_stream)
689 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
690 INSN_UID (loop_start), INSN_UID (loop_end));
694 /* Count number of times each reg is set during this loop.
695 Set VARRAY_CHAR (may_not_optimize, I) if it is not safe to move out
696 the setting of register I. Set VARRAY_RTX (reg_single_usage, I). */
698 /* Allocate extra space for REGS that might be created by
699 load_mems. We allocate a little extra slop as well, in the hopes
700 that even after the moving of movables creates some new registers
701 we won't have to reallocate these arrays. However, we do grow
702 the arrays, if necessary, in load_mems_recount_loop_regs_set. */
703 nregs = max_reg_num () + loop_mems_idx + 16;
704 VARRAY_INT_INIT (set_in_loop, nregs, "set_in_loop");
705 VARRAY_INT_INIT (n_times_set, nregs, "n_times_set");
706 VARRAY_CHAR_INIT (may_not_optimize, nregs, "may_not_optimize");
707 VARRAY_RTX_INIT (reg_single_usage, nregs, "reg_single_usage");
709 count_loop_regs_set (loop->top ? loop->top : loop->start, loop->end,
710 may_not_optimize, reg_single_usage, &insn_count, nregs);
712 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
714 VARRAY_CHAR (may_not_optimize, i) = 1;
715 VARRAY_INT (set_in_loop, i) = 1;
718 #ifdef AVOID_CCMODE_COPIES
719 /* Don't try to move insns which set CC registers if we should not
720 create CCmode register copies. */
721 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
722 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
723 VARRAY_CHAR (may_not_optimize, i) = 1;
726 bcopy ((char *) &set_in_loop->data,
727 (char *) &n_times_set->data, nregs * sizeof (int));
729 if (loop_dump_stream)
731 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
732 INSN_UID (loop_start), INSN_UID (loop_end), insn_count);
734 fprintf (loop_dump_stream, "Continue at insn %d.\n",
735 INSN_UID (loop->cont));
738 /* Scan through the loop finding insns that are safe to move.
739 Set set_in_loop negative for the reg being set, so that
740 this reg will be considered invariant for subsequent insns.
741 We consider whether subsequent insns use the reg
742 in deciding whether it is worth actually moving.
744 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
745 and therefore it is possible that the insns we are scanning
746 would never be executed. At such times, we must make sure
747 that it is safe to execute the insn once instead of zero times.
748 When MAYBE_NEVER is 0, all insns will be executed at least once
749 so that is not a problem. */
751 for (p = next_insn_in_loop (loop, loop->scan_start);
753 p = next_insn_in_loop (loop, p))
755 if (GET_CODE (p) == INSN
756 && (set = single_set (p))
757 && GET_CODE (SET_DEST (set)) == REG
758 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
763 rtx src = SET_SRC (set);
764 rtx dependencies = 0;
766 /* Figure out what to use as a source of this insn. If a REG_EQUIV
767 note is given or if a REG_EQUAL note with a constant operand is
768 specified, use it as the source and mark that we should move
769 this insn by calling emit_move_insn rather that duplicating the
772 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
774 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
776 src = XEXP (temp, 0), move_insn = 1;
779 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
780 if (temp && CONSTANT_P (XEXP (temp, 0)))
781 src = XEXP (temp, 0), move_insn = 1;
782 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
784 src = XEXP (temp, 0);
785 /* A libcall block can use regs that don't appear in
786 the equivalent expression. To move the libcall,
787 we must move those regs too. */
788 dependencies = libcall_other_reg (p, src);
792 /* Don't try to optimize a register that was made
793 by loop-optimization for an inner loop.
794 We don't know its life-span, so we can't compute the benefit. */
795 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
797 else if (/* The register is used in basic blocks other
798 than the one where it is set (meaning that
799 something after this point in the loop might
800 depend on its value before the set). */
801 ! reg_in_basic_block_p (p, SET_DEST (set))
802 /* And the set is not guaranteed to be executed one
803 the loop starts, or the value before the set is
804 needed before the set occurs...
806 ??? Note we have quadratic behaviour here, mitigated
807 by the fact that the previous test will often fail for
808 large loops. Rather than re-scanning the entire loop
809 each time for register usage, we should build tables
810 of the register usage and use them here instead. */
812 || loop_reg_used_before_p (loop, set, p)))
813 /* It is unsafe to move the set.
815 This code used to consider it OK to move a set of a variable
816 which was not created by the user and not used in an exit test.
817 That behavior is incorrect and was removed. */
819 else if ((tem = loop_invariant_p (loop, src))
820 && (dependencies == 0
821 || (tem2 = loop_invariant_p (loop, dependencies)) != 0)
822 && (VARRAY_INT (set_in_loop,
823 REGNO (SET_DEST (set))) == 1
825 = consec_sets_invariant_p
826 (loop, SET_DEST (set),
827 VARRAY_INT (set_in_loop, REGNO (SET_DEST (set))),
829 /* If the insn can cause a trap (such as divide by zero),
830 can't move it unless it's guaranteed to be executed
831 once loop is entered. Even a function call might
832 prevent the trap insn from being reached
833 (since it might exit!) */
834 && ! ((maybe_never || call_passed)
835 && may_trap_p (src)))
837 register struct movable *m;
838 register int regno = REGNO (SET_DEST (set));
840 /* A potential lossage is where we have a case where two insns
841 can be combined as long as they are both in the loop, but
842 we move one of them outside the loop. For large loops,
843 this can lose. The most common case of this is the address
844 of a function being called.
846 Therefore, if this register is marked as being used exactly
847 once if we are in a loop with calls (a "large loop"), see if
848 we can replace the usage of this register with the source
849 of this SET. If we can, delete this insn.
851 Don't do this if P has a REG_RETVAL note or if we have
852 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
854 if (loop_info->has_call
855 && VARRAY_RTX (reg_single_usage, regno) != 0
856 && VARRAY_RTX (reg_single_usage, regno) != const0_rtx
857 && REGNO_FIRST_UID (regno) == INSN_UID (p)
858 && (REGNO_LAST_UID (regno)
859 == INSN_UID (VARRAY_RTX (reg_single_usage, regno)))
860 && VARRAY_INT (set_in_loop, regno) == 1
861 && ! side_effects_p (SET_SRC (set))
862 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
863 && (! SMALL_REGISTER_CLASSES
864 || (! (GET_CODE (SET_SRC (set)) == REG
865 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)))
866 /* This test is not redundant; SET_SRC (set) might be
867 a call-clobbered register and the life of REGNO
868 might span a call. */
869 && ! modified_between_p (SET_SRC (set), p,
871 (reg_single_usage, regno))
872 && no_labels_between_p (p, VARRAY_RTX (reg_single_usage, regno))
873 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
875 (reg_single_usage, regno)))
877 /* Replace any usage in a REG_EQUAL note. Must copy the
878 new source, so that we don't get rtx sharing between the
879 SET_SOURCE and REG_NOTES of insn p. */
880 REG_NOTES (VARRAY_RTX (reg_single_usage, regno))
881 = replace_rtx (REG_NOTES (VARRAY_RTX
882 (reg_single_usage, regno)),
883 SET_DEST (set), copy_rtx (SET_SRC (set)));
886 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
887 NOTE_SOURCE_FILE (p) = 0;
888 VARRAY_INT (set_in_loop, regno) = 0;
892 m = (struct movable *) alloca (sizeof (struct movable));
896 m->dependencies = dependencies;
897 m->set_dest = SET_DEST (set);
899 m->consec = VARRAY_INT (set_in_loop,
900 REGNO (SET_DEST (set))) - 1;
904 m->move_insn = move_insn;
905 m->move_insn_first = 0;
906 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
907 m->savemode = VOIDmode;
909 /* Set M->cond if either loop_invariant_p
910 or consec_sets_invariant_p returned 2
911 (only conditionally invariant). */
912 m->cond = ((tem | tem1 | tem2) > 1);
913 m->global = (uid_luid[REGNO_LAST_UID (regno)]
914 > INSN_LUID (loop_end)
915 || uid_luid[REGNO_FIRST_UID (regno)] < INSN_LUID (loop_start));
917 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
918 - uid_luid[REGNO_FIRST_UID (regno)]);
919 m->savings = VARRAY_INT (n_times_set, regno);
920 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
921 m->savings += libcall_benefit (p);
922 VARRAY_INT (set_in_loop, regno) = move_insn ? -2 : -1;
923 /* Add M to the end of the chain MOVABLES. */
927 last_movable->next = m;
932 /* It is possible for the first instruction to have a
933 REG_EQUAL note but a non-invariant SET_SRC, so we must
934 remember the status of the first instruction in case
935 the last instruction doesn't have a REG_EQUAL note. */
936 m->move_insn_first = m->move_insn;
938 /* Skip this insn, not checking REG_LIBCALL notes. */
939 p = next_nonnote_insn (p);
940 /* Skip the consecutive insns, if there are any. */
941 p = skip_consec_insns (p, m->consec);
942 /* Back up to the last insn of the consecutive group. */
943 p = prev_nonnote_insn (p);
945 /* We must now reset m->move_insn, m->is_equiv, and possibly
946 m->set_src to correspond to the effects of all the
948 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
950 m->set_src = XEXP (temp, 0), m->move_insn = 1;
953 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
954 if (temp && CONSTANT_P (XEXP (temp, 0)))
955 m->set_src = XEXP (temp, 0), m->move_insn = 1;
960 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
963 /* If this register is always set within a STRICT_LOW_PART
964 or set to zero, then its high bytes are constant.
965 So clear them outside the loop and within the loop
966 just load the low bytes.
967 We must check that the machine has an instruction to do so.
968 Also, if the value loaded into the register
969 depends on the same register, this cannot be done. */
970 else if (SET_SRC (set) == const0_rtx
971 && GET_CODE (NEXT_INSN (p)) == INSN
972 && (set1 = single_set (NEXT_INSN (p)))
973 && GET_CODE (set1) == SET
974 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
975 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
976 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
978 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
980 register int regno = REGNO (SET_DEST (set));
981 if (VARRAY_INT (set_in_loop, regno) == 2)
983 register struct movable *m;
984 m = (struct movable *) alloca (sizeof (struct movable));
987 m->set_dest = SET_DEST (set);
994 m->move_insn_first = 0;
996 /* If the insn may not be executed on some cycles,
997 we can't clear the whole reg; clear just high part.
998 Not even if the reg is used only within this loop.
1005 Clearing x before the inner loop could clobber a value
1006 being saved from the last time around the outer loop.
1007 However, if the reg is not used outside this loop
1008 and all uses of the register are in the same
1009 basic block as the store, there is no problem.
1011 If this insn was made by loop, we don't know its
1012 INSN_LUID and hence must make a conservative
1014 m->global = (INSN_UID (p) >= max_uid_for_loop
1015 || (uid_luid[REGNO_LAST_UID (regno)]
1016 > INSN_LUID (loop_end))
1017 || (uid_luid[REGNO_FIRST_UID (regno)]
1019 || (labels_in_range_p
1020 (p, uid_luid[REGNO_FIRST_UID (regno)])));
1021 if (maybe_never && m->global)
1022 m->savemode = GET_MODE (SET_SRC (set1));
1024 m->savemode = VOIDmode;
1028 m->lifetime = (uid_luid[REGNO_LAST_UID (regno)]
1029 - uid_luid[REGNO_FIRST_UID (regno)]);
1031 VARRAY_INT (set_in_loop, regno) = -1;
1032 /* Add M to the end of the chain MOVABLES. */
1036 last_movable->next = m;
1041 /* Past a call insn, we get to insns which might not be executed
1042 because the call might exit. This matters for insns that trap.
1043 Constant and pure call insns always return, so they don't count. */
1044 else if (GET_CODE (p) == CALL_INSN && ! CONST_CALL_P (p))
1046 /* Past a label or a jump, we get to insns for which we
1047 can't count on whether or how many times they will be
1048 executed during each iteration. Therefore, we can
1049 only move out sets of trivial variables
1050 (those not used after the loop). */
1051 /* Similar code appears twice in strength_reduce. */
1052 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
1053 /* If we enter the loop in the middle, and scan around to the
1054 beginning, don't set maybe_never for that. This must be an
1055 unconditional jump, otherwise the code at the top of the
1056 loop might never be executed. Unconditional jumps are
1057 followed a by barrier then loop end. */
1058 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop->top
1059 && NEXT_INSN (NEXT_INSN (p)) == loop_end
1060 && any_uncondjump_p (p)))
1062 else if (GET_CODE (p) == NOTE)
1064 /* At the virtual top of a converted loop, insns are again known to
1065 be executed: logically, the loop begins here even though the exit
1066 code has been duplicated. */
1067 if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP && loop_depth == 0)
1068 maybe_never = call_passed = 0;
1069 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
1071 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
1076 /* If one movable subsumes another, ignore that other. */
1078 ignore_some_movables (movables);
1080 /* For each movable insn, see if the reg that it loads
1081 leads when it dies right into another conditionally movable insn.
1082 If so, record that the second insn "forces" the first one,
1083 since the second can be moved only if the first is. */
1085 force_movables (movables);
1087 /* See if there are multiple movable insns that load the same value.
1088 If there are, make all but the first point at the first one
1089 through the `match' field, and add the priorities of them
1090 all together as the priority of the first. */
1092 combine_movables (movables, nregs);
1094 /* Now consider each movable insn to decide whether it is worth moving.
1095 Store 0 in set_in_loop for each reg that is moved.
1097 Generally this increases code size, so do not move moveables when
1098 optimizing for code size. */
1100 if (! optimize_size)
1101 move_movables (loop, movables, threshold, insn_count, nregs);
1103 /* Now candidates that still are negative are those not moved.
1104 Change set_in_loop to indicate that those are not actually invariant. */
1105 for (i = 0; i < nregs; i++)
1106 if (VARRAY_INT (set_in_loop, i) < 0)
1107 VARRAY_INT (set_in_loop, i) = VARRAY_INT (n_times_set, i);
1109 /* Now that we've moved some things out of the loop, we might be able to
1110 hoist even more memory references. */
1111 load_mems_and_recount_loop_regs_set (loop, &insn_count);
1113 for (update_start = loop_start;
1114 PREV_INSN (update_start)
1115 && GET_CODE (PREV_INSN (update_start)) != CODE_LABEL;
1116 update_start = PREV_INSN (update_start))
1118 update_end = NEXT_INSN (loop_end);
1120 reg_scan_update (update_start, update_end, loop_max_reg);
1121 loop_max_reg = max_reg_num ();
1123 if (flag_strength_reduce)
1125 if (update_end && GET_CODE (update_end) == CODE_LABEL)
1126 /* Ensure our label doesn't go away. */
1127 LABEL_NUSES (update_end)++;
1129 the_movables = movables;
1130 strength_reduce (loop, insn_count, flags);
1132 reg_scan_update (update_start, update_end, loop_max_reg);
1133 loop_max_reg = max_reg_num ();
1135 if (update_end && GET_CODE (update_end) == CODE_LABEL
1136 && --LABEL_NUSES (update_end) == 0)
1137 delete_insn (update_end);
1140 VARRAY_FREE (reg_single_usage);
1141 VARRAY_FREE (set_in_loop);
1142 VARRAY_FREE (n_times_set);
1143 VARRAY_FREE (may_not_optimize);
1146 /* Add elements to *OUTPUT to record all the pseudo-regs
1147 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
1150 record_excess_regs (in_this, not_in_this, output)
1151 rtx in_this, not_in_this;
1158 code = GET_CODE (in_this);
1172 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
1173 && ! reg_mentioned_p (in_this, not_in_this))
1174 *output = gen_rtx_EXPR_LIST (VOIDmode, in_this, *output);
1181 fmt = GET_RTX_FORMAT (code);
1182 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1189 for (j = 0; j < XVECLEN (in_this, i); j++)
1190 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1194 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1200 /* Check what regs are referred to in the libcall block ending with INSN,
1201 aside from those mentioned in the equivalent value.
1202 If there are none, return 0.
1203 If there are one or more, return an EXPR_LIST containing all of them. */
1206 libcall_other_reg (insn, equiv)
1209 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1210 rtx p = XEXP (note, 0);
1213 /* First, find all the regs used in the libcall block
1214 that are not mentioned as inputs to the result. */
1218 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1219 || GET_CODE (p) == CALL_INSN)
1220 record_excess_regs (PATTERN (p), equiv, &output);
1227 /* Return 1 if all uses of REG
1228 are between INSN and the end of the basic block. */
1231 reg_in_basic_block_p (insn, reg)
1234 int regno = REGNO (reg);
1237 if (REGNO_FIRST_UID (regno) != INSN_UID (insn))
1240 /* Search this basic block for the already recorded last use of the reg. */
1241 for (p = insn; p; p = NEXT_INSN (p))
1243 switch (GET_CODE (p))
1250 /* Ordinary insn: if this is the last use, we win. */
1251 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1256 /* Jump insn: if this is the last use, we win. */
1257 if (REGNO_LAST_UID (regno) == INSN_UID (p))
1259 /* Otherwise, it's the end of the basic block, so we lose. */
1264 /* It's the end of the basic block, so we lose. */
1272 /* The "last use" that was recorded can't be found after the first
1273 use. This can happen when the last use was deleted while
1274 processing an inner loop, this inner loop was then completely
1275 unrolled, and the outer loop is always exited after the inner loop,
1276 so that everything after the first use becomes a single basic block. */
1280 /* Compute the benefit of eliminating the insns in the block whose
1281 last insn is LAST. This may be a group of insns used to compute a
1282 value directly or can contain a library call. */
1285 libcall_benefit (last)
1291 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1292 insn != last; insn = NEXT_INSN (insn))
1294 if (GET_CODE (insn) == CALL_INSN)
1295 benefit += 10; /* Assume at least this many insns in a library
1297 else if (GET_CODE (insn) == INSN
1298 && GET_CODE (PATTERN (insn)) != USE
1299 && GET_CODE (PATTERN (insn)) != CLOBBER)
1306 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1309 skip_consec_insns (insn, count)
1313 for (; count > 0; count--)
1317 /* If first insn of libcall sequence, skip to end. */
1318 /* Do this at start of loop, since INSN is guaranteed to
1320 if (GET_CODE (insn) != NOTE
1321 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1322 insn = XEXP (temp, 0);
1325 insn = NEXT_INSN (insn);
1326 while (GET_CODE (insn) == NOTE);
1332 /* Ignore any movable whose insn falls within a libcall
1333 which is part of another movable.
1334 We make use of the fact that the movable for the libcall value
1335 was made later and so appears later on the chain. */
1338 ignore_some_movables (movables)
1339 struct movable *movables;
1341 register struct movable *m, *m1;
1343 for (m = movables; m; m = m->next)
1345 /* Is this a movable for the value of a libcall? */
1346 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1350 /* Check for earlier movables inside that range,
1351 and mark them invalid. We cannot use LUIDs here because
1352 insns created by loop.c for prior loops don't have LUIDs.
1353 Rather than reject all such insns from movables, we just
1354 explicitly check each insn in the libcall (since invariant
1355 libcalls aren't that common). */
1356 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1357 for (m1 = movables; m1 != m; m1 = m1->next)
1358 if (m1->insn == insn)
1364 /* For each movable insn, see if the reg that it loads
1365 leads when it dies right into another conditionally movable insn.
1366 If so, record that the second insn "forces" the first one,
1367 since the second can be moved only if the first is. */
1370 force_movables (movables)
1371 struct movable *movables;
1373 register struct movable *m, *m1;
1374 for (m1 = movables; m1; m1 = m1->next)
1375 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1376 if (!m1->partial && !m1->done)
1378 int regno = m1->regno;
1379 for (m = m1->next; m; m = m->next)
1380 /* ??? Could this be a bug? What if CSE caused the
1381 register of M1 to be used after this insn?
1382 Since CSE does not update regno_last_uid,
1383 this insn M->insn might not be where it dies.
1384 But very likely this doesn't matter; what matters is
1385 that M's reg is computed from M1's reg. */
1386 if (INSN_UID (m->insn) == REGNO_LAST_UID (regno)
1389 if (m != 0 && m->set_src == m1->set_dest
1390 /* If m->consec, m->set_src isn't valid. */
1394 /* Increase the priority of the moving the first insn
1395 since it permits the second to be moved as well. */
1399 m1->lifetime += m->lifetime;
1400 m1->savings += m->savings;
1405 /* Find invariant expressions that are equal and can be combined into
1409 combine_movables (movables, nregs)
1410 struct movable *movables;
1413 register struct movable *m;
1414 char *matched_regs = (char *) xmalloc (nregs);
1415 enum machine_mode mode;
1417 /* Regs that are set more than once are not allowed to match
1418 or be matched. I'm no longer sure why not. */
1419 /* Perhaps testing m->consec_sets would be more appropriate here? */
1421 for (m = movables; m; m = m->next)
1422 if (m->match == 0 && VARRAY_INT (n_times_set, m->regno) == 1
1425 register struct movable *m1;
1426 int regno = m->regno;
1428 bzero (matched_regs, nregs);
1429 matched_regs[regno] = 1;
1431 /* We want later insns to match the first one. Don't make the first
1432 one match any later ones. So start this loop at m->next. */
1433 for (m1 = m->next; m1; m1 = m1->next)
1434 if (m != m1 && m1->match == 0 && VARRAY_INT (n_times_set, m1->regno) == 1
1435 /* A reg used outside the loop mustn't be eliminated. */
1437 /* A reg used for zero-extending mustn't be eliminated. */
1439 && (matched_regs[m1->regno]
1442 /* Can combine regs with different modes loaded from the
1443 same constant only if the modes are the same or
1444 if both are integer modes with M wider or the same
1445 width as M1. The check for integer is redundant, but
1446 safe, since the only case of differing destination
1447 modes with equal sources is when both sources are
1448 VOIDmode, i.e., CONST_INT. */
1449 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1450 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1451 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1452 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1453 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1454 /* See if the source of M1 says it matches M. */
1455 && ((GET_CODE (m1->set_src) == REG
1456 && matched_regs[REGNO (m1->set_src)])
1457 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1459 && ((m->dependencies == m1->dependencies)
1460 || rtx_equal_p (m->dependencies, m1->dependencies)))
1462 m->lifetime += m1->lifetime;
1463 m->savings += m1->savings;
1466 matched_regs[m1->regno] = 1;
1470 /* Now combine the regs used for zero-extension.
1471 This can be done for those not marked `global'
1472 provided their lives don't overlap. */
1474 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1475 mode = GET_MODE_WIDER_MODE (mode))
1477 register struct movable *m0 = 0;
1479 /* Combine all the registers for extension from mode MODE.
1480 Don't combine any that are used outside this loop. */
1481 for (m = movables; m; m = m->next)
1482 if (m->partial && ! m->global
1483 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1485 register struct movable *m1;
1486 int first = uid_luid[REGNO_FIRST_UID (m->regno)];
1487 int last = uid_luid[REGNO_LAST_UID (m->regno)];
1491 /* First one: don't check for overlap, just record it. */
1496 /* Make sure they extend to the same mode.
1497 (Almost always true.) */
1498 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1501 /* We already have one: check for overlap with those
1502 already combined together. */
1503 for (m1 = movables; m1 != m; m1 = m1->next)
1504 if (m1 == m0 || (m1->partial && m1->match == m0))
1505 if (! (uid_luid[REGNO_FIRST_UID (m1->regno)] > last
1506 || uid_luid[REGNO_LAST_UID (m1->regno)] < first))
1509 /* No overlap: we can combine this with the others. */
1510 m0->lifetime += m->lifetime;
1511 m0->savings += m->savings;
1521 free (matched_regs);
1524 /* Return 1 if regs X and Y will become the same if moved. */
1527 regs_match_p (x, y, movables)
1529 struct movable *movables;
1531 unsigned int xn = REGNO (x);
1532 unsigned int yn = REGNO (y);
1533 struct movable *mx, *my;
1535 for (mx = movables; mx; mx = mx->next)
1536 if (mx->regno == xn)
1539 for (my = movables; my; my = my->next)
1540 if (my->regno == yn)
1544 && ((mx->match == my->match && mx->match != 0)
1546 || mx == my->match));
1549 /* Return 1 if X and Y are identical-looking rtx's.
1550 This is the Lisp function EQUAL for rtx arguments.
1552 If two registers are matching movables or a movable register and an
1553 equivalent constant, consider them equal. */
1556 rtx_equal_for_loop_p (x, y, movables)
1558 struct movable *movables;
1562 register struct movable *m;
1563 register enum rtx_code code;
1564 register const char *fmt;
1568 if (x == 0 || y == 0)
1571 code = GET_CODE (x);
1573 /* If we have a register and a constant, they may sometimes be
1575 if (GET_CODE (x) == REG && VARRAY_INT (set_in_loop, REGNO (x)) == -2
1578 for (m = movables; m; m = m->next)
1579 if (m->move_insn && m->regno == REGNO (x)
1580 && rtx_equal_p (m->set_src, y))
1583 else if (GET_CODE (y) == REG && VARRAY_INT (set_in_loop, REGNO (y)) == -2
1586 for (m = movables; m; m = m->next)
1587 if (m->move_insn && m->regno == REGNO (y)
1588 && rtx_equal_p (m->set_src, x))
1592 /* Otherwise, rtx's of different codes cannot be equal. */
1593 if (code != GET_CODE (y))
1596 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1597 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1599 if (GET_MODE (x) != GET_MODE (y))
1602 /* These three types of rtx's can be compared nonrecursively. */
1604 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1606 if (code == LABEL_REF)
1607 return XEXP (x, 0) == XEXP (y, 0);
1608 if (code == SYMBOL_REF)
1609 return XSTR (x, 0) == XSTR (y, 0);
1611 /* Compare the elements. If any pair of corresponding elements
1612 fail to match, return 0 for the whole things. */
1614 fmt = GET_RTX_FORMAT (code);
1615 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1620 if (XWINT (x, i) != XWINT (y, i))
1625 if (XINT (x, i) != XINT (y, i))
1630 /* Two vectors must have the same length. */
1631 if (XVECLEN (x, i) != XVECLEN (y, i))
1634 /* And the corresponding elements must match. */
1635 for (j = 0; j < XVECLEN (x, i); j++)
1636 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1641 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1646 if (strcmp (XSTR (x, i), XSTR (y, i)))
1651 /* These are just backpointers, so they don't matter. */
1657 /* It is believed that rtx's at this level will never
1658 contain anything but integers and other rtx's,
1659 except for within LABEL_REFs and SYMBOL_REFs. */
1667 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1668 insns in INSNS which use the reference. */
1671 add_label_notes (x, insns)
1675 enum rtx_code code = GET_CODE (x);
1680 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
1682 /* This code used to ignore labels that referred to dispatch tables to
1683 avoid flow generating (slighly) worse code.
1685 We no longer ignore such label references (see LABEL_REF handling in
1686 mark_jump_label for additional information). */
1687 for (insn = insns; insn; insn = NEXT_INSN (insn))
1688 if (reg_mentioned_p (XEXP (x, 0), insn))
1689 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
1693 fmt = GET_RTX_FORMAT (code);
1694 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1697 add_label_notes (XEXP (x, i), insns);
1698 else if (fmt[i] == 'E')
1699 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1700 add_label_notes (XVECEXP (x, i, j), insns);
1704 /* Scan MOVABLES, and move the insns that deserve to be moved.
1705 If two matching movables are combined, replace one reg with the
1706 other throughout. */
1709 move_movables (loop, movables, threshold, insn_count, nregs)
1711 struct movable *movables;
1717 register struct movable *m;
1719 rtx loop_start = loop->start;
1720 rtx loop_end = loop->end;
1721 /* Map of pseudo-register replacements to handle combining
1722 when we move several insns that load the same value
1723 into different pseudo-registers. */
1724 rtx *reg_map = (rtx *) xcalloc (nregs, sizeof (rtx));
1725 char *already_moved = (char *) xcalloc (nregs, sizeof (char));
1729 for (m = movables; m; m = m->next)
1731 /* Describe this movable insn. */
1733 if (loop_dump_stream)
1735 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1736 INSN_UID (m->insn), m->regno, m->lifetime);
1738 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1740 fprintf (loop_dump_stream, "cond ");
1742 fprintf (loop_dump_stream, "force ");
1744 fprintf (loop_dump_stream, "global ");
1746 fprintf (loop_dump_stream, "done ");
1748 fprintf (loop_dump_stream, "move-insn ");
1750 fprintf (loop_dump_stream, "matches %d ",
1751 INSN_UID (m->match->insn));
1753 fprintf (loop_dump_stream, "forces %d ",
1754 INSN_UID (m->forces->insn));
1757 /* Count movables. Value used in heuristics in strength_reduce. */
1760 /* Ignore the insn if it's already done (it matched something else).
1761 Otherwise, see if it is now safe to move. */
1765 || (1 == loop_invariant_p (loop, m->set_src)
1766 && (m->dependencies == 0
1767 || 1 == loop_invariant_p (loop, m->dependencies))
1769 || 1 == consec_sets_invariant_p (loop, m->set_dest,
1772 && (! m->forces || m->forces->done))
1776 int savings = m->savings;
1778 /* We have an insn that is safe to move.
1779 Compute its desirability. */
1784 if (loop_dump_stream)
1785 fprintf (loop_dump_stream, "savings %d ", savings);
1787 if (moved_once[regno] && loop_dump_stream)
1788 fprintf (loop_dump_stream, "halved since already moved ");
1790 /* An insn MUST be moved if we already moved something else
1791 which is safe only if this one is moved too: that is,
1792 if already_moved[REGNO] is nonzero. */
1794 /* An insn is desirable to move if the new lifetime of the
1795 register is no more than THRESHOLD times the old lifetime.
1796 If it's not desirable, it means the loop is so big
1797 that moving won't speed things up much,
1798 and it is liable to make register usage worse. */
1800 /* It is also desirable to move if it can be moved at no
1801 extra cost because something else was already moved. */
1803 if (already_moved[regno]
1804 || flag_move_all_movables
1805 || (threshold * savings * m->lifetime) >=
1806 (moved_once[regno] ? insn_count * 2 : insn_count)
1807 || (m->forces && m->forces->done
1808 && VARRAY_INT (n_times_set, m->forces->regno) == 1))
1811 register struct movable *m1;
1812 rtx first = NULL_RTX;
1814 /* Now move the insns that set the reg. */
1816 if (m->partial && m->match)
1820 /* Find the end of this chain of matching regs.
1821 Thus, we load each reg in the chain from that one reg.
1822 And that reg is loaded with 0 directly,
1823 since it has ->match == 0. */
1824 for (m1 = m; m1->match; m1 = m1->match);
1825 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1826 SET_DEST (PATTERN (m1->insn)));
1827 i1 = emit_insn_before (newpat, loop_start);
1829 /* Mark the moved, invariant reg as being allowed to
1830 share a hard reg with the other matching invariant. */
1831 REG_NOTES (i1) = REG_NOTES (m->insn);
1832 r1 = SET_DEST (PATTERN (m->insn));
1833 r2 = SET_DEST (PATTERN (m1->insn));
1835 = gen_rtx_EXPR_LIST (VOIDmode, r1,
1836 gen_rtx_EXPR_LIST (VOIDmode, r2,
1838 delete_insn (m->insn);
1843 if (loop_dump_stream)
1844 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1846 /* If we are to re-generate the item being moved with a
1847 new move insn, first delete what we have and then emit
1848 the move insn before the loop. */
1849 else if (m->move_insn)
1853 for (count = m->consec; count >= 0; count--)
1855 /* If this is the first insn of a library call sequence,
1857 if (GET_CODE (p) != NOTE
1858 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1861 /* If this is the last insn of a libcall sequence, then
1862 delete every insn in the sequence except the last.
1863 The last insn is handled in the normal manner. */
1864 if (GET_CODE (p) != NOTE
1865 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1867 temp = XEXP (temp, 0);
1869 temp = delete_insn (temp);
1873 p = delete_insn (p);
1875 /* simplify_giv_expr expects that it can walk the insns
1876 at m->insn forwards and see this old sequence we are
1877 tossing here. delete_insn does preserve the next
1878 pointers, but when we skip over a NOTE we must fix
1879 it up. Otherwise that code walks into the non-deleted
1881 while (p && GET_CODE (p) == NOTE)
1882 p = NEXT_INSN (temp) = NEXT_INSN (p);
1886 emit_move_insn (m->set_dest, m->set_src);
1887 temp = get_insns ();
1890 add_label_notes (m->set_src, temp);
1892 i1 = emit_insns_before (temp, loop_start);
1893 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1895 = gen_rtx_EXPR_LIST (m->is_equiv ? REG_EQUIV : REG_EQUAL,
1896 m->set_src, REG_NOTES (i1));
1898 if (loop_dump_stream)
1899 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1901 /* The more regs we move, the less we like moving them. */
1906 for (count = m->consec; count >= 0; count--)
1910 /* If first insn of libcall sequence, skip to end. */
1911 /* Do this at start of loop, since p is guaranteed to
1913 if (GET_CODE (p) != NOTE
1914 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1917 /* If last insn of libcall sequence, move all
1918 insns except the last before the loop. The last
1919 insn is handled in the normal manner. */
1920 if (GET_CODE (p) != NOTE
1921 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1925 rtx fn_address_insn = 0;
1928 for (temp = XEXP (temp, 0); temp != p;
1929 temp = NEXT_INSN (temp))
1935 if (GET_CODE (temp) == NOTE)
1938 body = PATTERN (temp);
1940 /* Find the next insn after TEMP,
1941 not counting USE or NOTE insns. */
1942 for (next = NEXT_INSN (temp); next != p;
1943 next = NEXT_INSN (next))
1944 if (! (GET_CODE (next) == INSN
1945 && GET_CODE (PATTERN (next)) == USE)
1946 && GET_CODE (next) != NOTE)
1949 /* If that is the call, this may be the insn
1950 that loads the function address.
1952 Extract the function address from the insn
1953 that loads it into a register.
1954 If this insn was cse'd, we get incorrect code.
1956 So emit a new move insn that copies the
1957 function address into the register that the
1958 call insn will use. flow.c will delete any
1959 redundant stores that we have created. */
1960 if (GET_CODE (next) == CALL_INSN
1961 && GET_CODE (body) == SET
1962 && GET_CODE (SET_DEST (body)) == REG
1963 && (n = find_reg_note (temp, REG_EQUAL,
1966 fn_reg = SET_SRC (body);
1967 if (GET_CODE (fn_reg) != REG)
1968 fn_reg = SET_DEST (body);
1969 fn_address = XEXP (n, 0);
1970 fn_address_insn = temp;
1972 /* We have the call insn.
1973 If it uses the register we suspect it might,
1974 load it with the correct address directly. */
1975 if (GET_CODE (temp) == CALL_INSN
1977 && reg_referenced_p (fn_reg, body))
1978 emit_insn_after (gen_move_insn (fn_reg,
1982 if (GET_CODE (temp) == CALL_INSN)
1984 i1 = emit_call_insn_before (body, loop_start);
1985 /* Because the USAGE information potentially
1986 contains objects other than hard registers
1987 we need to copy it. */
1988 if (CALL_INSN_FUNCTION_USAGE (temp))
1989 CALL_INSN_FUNCTION_USAGE (i1)
1990 = copy_rtx (CALL_INSN_FUNCTION_USAGE (temp));
1993 i1 = emit_insn_before (body, loop_start);
1996 if (temp == fn_address_insn)
1997 fn_address_insn = i1;
1998 REG_NOTES (i1) = REG_NOTES (temp);
2004 if (m->savemode != VOIDmode)
2006 /* P sets REG to zero; but we should clear only
2007 the bits that are not covered by the mode
2009 rtx reg = m->set_dest;
2015 (GET_MODE (reg), and_optab, reg,
2016 GEN_INT ((((HOST_WIDE_INT) 1
2017 << GET_MODE_BITSIZE (m->savemode)))
2019 reg, 1, OPTAB_LIB_WIDEN);
2023 emit_move_insn (reg, tem);
2024 sequence = gen_sequence ();
2026 i1 = emit_insn_before (sequence, loop_start);
2028 else if (GET_CODE (p) == CALL_INSN)
2030 i1 = emit_call_insn_before (PATTERN (p), loop_start);
2031 /* Because the USAGE information potentially
2032 contains objects other than hard registers
2033 we need to copy it. */
2034 if (CALL_INSN_FUNCTION_USAGE (p))
2035 CALL_INSN_FUNCTION_USAGE (i1)
2036 = copy_rtx (CALL_INSN_FUNCTION_USAGE (p));
2038 else if (count == m->consec && m->move_insn_first)
2040 /* The SET_SRC might not be invariant, so we must
2041 use the REG_EQUAL note. */
2043 emit_move_insn (m->set_dest, m->set_src);
2044 temp = get_insns ();
2047 add_label_notes (m->set_src, temp);
2049 i1 = emit_insns_before (temp, loop_start);
2050 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
2052 = gen_rtx_EXPR_LIST ((m->is_equiv ? REG_EQUIV
2054 m->set_src, REG_NOTES (i1));
2057 i1 = emit_insn_before (PATTERN (p), loop_start);
2059 if (REG_NOTES (i1) == 0)
2061 REG_NOTES (i1) = REG_NOTES (p);
2063 /* If there is a REG_EQUAL note present whose value
2064 is not loop invariant, then delete it, since it
2065 may cause problems with later optimization passes.
2066 It is possible for cse to create such notes
2067 like this as a result of record_jump_cond. */
2069 if ((temp = find_reg_note (i1, REG_EQUAL, NULL_RTX))
2070 && ! loop_invariant_p (loop, XEXP (temp, 0)))
2071 remove_note (i1, temp);
2077 if (loop_dump_stream)
2078 fprintf (loop_dump_stream, " moved to %d",
2081 /* If library call, now fix the REG_NOTES that contain
2082 insn pointers, namely REG_LIBCALL on FIRST
2083 and REG_RETVAL on I1. */
2084 if ((temp = find_reg_note (i1, REG_RETVAL, NULL_RTX)))
2086 XEXP (temp, 0) = first;
2087 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
2088 XEXP (temp, 0) = i1;
2095 /* simplify_giv_expr expects that it can walk the insns
2096 at m->insn forwards and see this old sequence we are
2097 tossing here. delete_insn does preserve the next
2098 pointers, but when we skip over a NOTE we must fix
2099 it up. Otherwise that code walks into the non-deleted
2101 while (p && GET_CODE (p) == NOTE)
2102 p = NEXT_INSN (temp) = NEXT_INSN (p);
2105 /* The more regs we move, the less we like moving them. */
2109 /* Any other movable that loads the same register
2111 already_moved[regno] = 1;
2113 /* This reg has been moved out of one loop. */
2114 moved_once[regno] = 1;
2116 /* The reg set here is now invariant. */
2118 VARRAY_INT (set_in_loop, regno) = 0;
2122 /* Change the length-of-life info for the register
2123 to say it lives at least the full length of this loop.
2124 This will help guide optimizations in outer loops. */
2126 if (uid_luid[REGNO_FIRST_UID (regno)] > INSN_LUID (loop_start))
2127 /* This is the old insn before all the moved insns.
2128 We can't use the moved insn because it is out of range
2129 in uid_luid. Only the old insns have luids. */
2130 REGNO_FIRST_UID (regno) = INSN_UID (loop_start);
2131 if (uid_luid[REGNO_LAST_UID (regno)] < INSN_LUID (loop_end))
2132 REGNO_LAST_UID (regno) = INSN_UID (loop_end);
2134 /* Combine with this moved insn any other matching movables. */
2137 for (m1 = movables; m1; m1 = m1->next)
2142 /* Schedule the reg loaded by M1
2143 for replacement so that shares the reg of M.
2144 If the modes differ (only possible in restricted
2145 circumstances, make a SUBREG.
2147 Note this assumes that the target dependent files
2148 treat REG and SUBREG equally, including within
2149 GO_IF_LEGITIMATE_ADDRESS and in all the
2150 predicates since we never verify that replacing the
2151 original register with a SUBREG results in a
2152 recognizable insn. */
2153 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
2154 reg_map[m1->regno] = m->set_dest;
2157 = gen_lowpart_common (GET_MODE (m1->set_dest),
2160 /* Get rid of the matching insn
2161 and prevent further processing of it. */
2164 /* if library call, delete all insn except last, which
2166 if ((temp = find_reg_note (m1->insn, REG_RETVAL,
2169 for (temp = XEXP (temp, 0); temp != m1->insn;
2170 temp = NEXT_INSN (temp))
2173 delete_insn (m1->insn);
2175 /* Any other movable that loads the same register
2177 already_moved[m1->regno] = 1;
2179 /* The reg merged here is now invariant,
2180 if the reg it matches is invariant. */
2182 VARRAY_INT (set_in_loop, m1->regno) = 0;
2185 else if (loop_dump_stream)
2186 fprintf (loop_dump_stream, "not desirable");
2188 else if (loop_dump_stream && !m->match)
2189 fprintf (loop_dump_stream, "not safe");
2191 if (loop_dump_stream)
2192 fprintf (loop_dump_stream, "\n");
2196 new_start = loop_start;
2198 /* Go through all the instructions in the loop, making
2199 all the register substitutions scheduled in REG_MAP. */
2200 for (p = new_start; p != loop_end; p = NEXT_INSN (p))
2201 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
2202 || GET_CODE (p) == CALL_INSN)
2204 replace_regs (PATTERN (p), reg_map, nregs, 0);
2205 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
2211 free (already_moved);
2215 /* Scan X and replace the address of any MEM in it with ADDR.
2216 REG is the address that MEM should have before the replacement. */
2219 replace_call_address (x, reg, addr)
2222 register enum rtx_code code;
2224 register const char *fmt;
2228 code = GET_CODE (x);
2242 /* Short cut for very common case. */
2243 replace_call_address (XEXP (x, 1), reg, addr);
2247 /* Short cut for very common case. */
2248 replace_call_address (XEXP (x, 0), reg, addr);
2252 /* If this MEM uses a reg other than the one we expected,
2253 something is wrong. */
2254 if (XEXP (x, 0) != reg)
2263 fmt = GET_RTX_FORMAT (code);
2264 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2267 replace_call_address (XEXP (x, i), reg, addr);
2268 else if (fmt[i] == 'E')
2271 for (j = 0; j < XVECLEN (x, i); j++)
2272 replace_call_address (XVECEXP (x, i, j), reg, addr);
2278 /* Return the number of memory refs to addresses that vary
2282 count_nonfixed_reads (loop, x)
2283 const struct loop *loop;
2286 register enum rtx_code code;
2288 register const char *fmt;
2294 code = GET_CODE (x);
2308 return ((loop_invariant_p (loop, XEXP (x, 0)) != 1)
2309 + count_nonfixed_reads (loop, XEXP (x, 0)));
2316 fmt = GET_RTX_FORMAT (code);
2317 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2320 value += count_nonfixed_reads (loop, XEXP (x, i));
2324 for (j = 0; j < XVECLEN (x, i); j++)
2325 value += count_nonfixed_reads (loop, XVECEXP (x, i, j));
2332 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2333 Replace it with an instruction to load just the low bytes
2334 if the machine supports such an instruction,
2335 and insert above LOOP_START an instruction to clear the register. */
2338 constant_high_bytes (p, loop_start)
2342 register int insn_code_number;
2344 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2345 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2350 gen_rtx_STRICT_LOW_PART
2352 gen_rtx_SUBREG (GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2353 SET_DEST (PATTERN (p)), 0)),
2354 XEXP (SET_SRC (PATTERN (p)), 0));
2356 insn_code_number = recog (new, p);
2358 if (insn_code_number)
2362 /* Clear destination register before the loop. */
2363 emit_insn_before (gen_rtx_SET (VOIDmode,
2364 SET_DEST (PATTERN (p)), const0_rtx),
2367 /* Inside the loop, just load the low part. */
2373 /* Scan a loop setting the elements `cont', `vtop', `loops_enclosed',
2374 `has_call', `has_volatile', and `has_tablejump' within LOOP.
2375 Set the global variables `unknown_address_altered',
2376 `unknown_constant_address_altered', and `num_mem_sets'. Also, fill
2377 in the array `loop_mems' and the list `loop_store_mems'. */
2383 register int level = 1;
2385 struct loop_info *loop_info = LOOP_INFO (loop);
2386 rtx start = loop->start;
2387 rtx end = loop->end;
2388 /* The label after END. Jumping here is just like falling off the
2389 end of the loop. We use next_nonnote_insn instead of next_label
2390 as a hedge against the (pathological) case where some actual insn
2391 might end up between the two. */
2392 rtx exit_target = next_nonnote_insn (end);
2394 loop_info->has_indirect_jump = indirect_jump_in_function;
2395 loop_info->has_call = 0;
2396 loop_info->has_volatile = 0;
2397 loop_info->has_tablejump = 0;
2398 loop_info->has_multiple_exit_targets = 0;
2401 unknown_address_altered = 0;
2402 unknown_constant_address_altered = 0;
2403 loop_store_mems = NULL_RTX;
2404 first_loop_store_insn = NULL_RTX;
2408 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2409 insn = NEXT_INSN (insn))
2411 if (GET_CODE (insn) == NOTE)
2413 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2416 /* Count number of loops contained in this one. */
2419 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2424 else if (GET_CODE (insn) == CALL_INSN)
2426 if (! CONST_CALL_P (insn))
2427 unknown_address_altered = 1;
2428 loop_info->has_call = 1;
2430 else if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2432 rtx label1 = NULL_RTX;
2433 rtx label2 = NULL_RTX;
2435 if (volatile_refs_p (PATTERN (insn)))
2436 loop_info->has_volatile = 1;
2438 if (GET_CODE (insn) == JUMP_INSN
2439 && (GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
2440 || GET_CODE (PATTERN (insn)) == ADDR_VEC))
2441 loop_info->has_tablejump = 1;
2443 note_stores (PATTERN (insn), note_addr_stored, NULL);
2444 if (! first_loop_store_insn && loop_store_mems)
2445 first_loop_store_insn = insn;
2447 if (! loop_info->has_multiple_exit_targets
2448 && GET_CODE (insn) == JUMP_INSN
2449 && GET_CODE (PATTERN (insn)) == SET
2450 && SET_DEST (PATTERN (insn)) == pc_rtx)
2452 if (GET_CODE (SET_SRC (PATTERN (insn))) == IF_THEN_ELSE)
2454 label1 = XEXP (SET_SRC (PATTERN (insn)), 1);
2455 label2 = XEXP (SET_SRC (PATTERN (insn)), 2);
2459 label1 = SET_SRC (PATTERN (insn));
2464 if (label1 && label1 != pc_rtx)
2466 if (GET_CODE (label1) != LABEL_REF)
2468 /* Something tricky. */
2469 loop_info->has_multiple_exit_targets = 1;
2472 else if (XEXP (label1, 0) != exit_target
2473 && LABEL_OUTSIDE_LOOP_P (label1))
2475 /* A jump outside the current loop. */
2476 loop_info->has_multiple_exit_targets = 1;
2487 else if (GET_CODE (insn) == RETURN)
2488 loop_info->has_multiple_exit_targets = 1;
2491 /* Now, rescan the loop, setting up the LOOP_MEMS array. */
2492 if (/* We can't tell what MEMs are aliased by what. */
2493 ! unknown_address_altered
2494 /* An exception thrown by a called function might land us
2496 && ! loop_info->has_call
2497 /* We don't want loads for MEMs moved to a location before the
2498 one at which their stack memory becomes allocated. (Note
2499 that this is not a problem for malloc, etc., since those
2500 require actual function calls. */
2501 && ! current_function_calls_alloca
2502 /* There are ways to leave the loop other than falling off the
2504 && ! loop_info->has_multiple_exit_targets)
2505 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2506 insn = NEXT_INSN (insn))
2507 for_each_rtx (&insn, insert_loop_mem, 0);
2510 /* LOOP->CONT_DOMINATOR is now the last label between the loop start
2511 and the continue note that is a the destination of a (cond)jump after
2512 the continue note. If there is any (cond)jump between the loop start
2513 and what we have so far as LOOP->CONT_DOMINATOR that has a
2514 target between LOOP->CONT_DOMINATOR and the continue note, move
2515 LOOP->CONT_DOMINATOR forward to that label; if a jump's
2516 destination cannot be determined, clear LOOP->CONT_DOMINATOR. */
2519 verify_dominator (loop)
2524 if (! loop->cont_dominator)
2525 /* This can happen for an empty loop, e.g. in
2526 gcc.c-torture/compile/920410-2.c */
2528 if (loop->cont_dominator == const0_rtx)
2530 loop->cont_dominator = 0;
2533 for (insn = loop->start; insn != loop->cont_dominator;
2534 insn = NEXT_INSN (insn))
2536 if (GET_CODE (insn) == JUMP_INSN
2537 && GET_CODE (PATTERN (insn)) != RETURN)
2539 rtx label = JUMP_LABEL (insn);
2542 /* If it is not a jump we can easily understand or for
2543 which we do not have jump target information in the JUMP_LABEL
2544 field (consider ADDR_VEC and ADDR_DIFF_VEC insns), then clear
2545 LOOP->CONT_DOMINATOR. */
2546 if (! any_condjump_p (insn)
2547 || label == NULL_RTX)
2549 loop->cont_dominator = NULL_RTX;
2553 label_luid = INSN_LUID (label);
2554 if (label_luid < INSN_LUID (loop->cont)
2556 > INSN_LUID (loop->cont)))
2557 loop->cont_dominator = label;
2562 /* Scan the function looking for loops. Record the start and end of each loop.
2563 Also mark as invalid loops any loops that contain a setjmp or are branched
2564 to from outside the loop. */
2567 find_and_verify_loops (f, loops)
2569 struct loops *loops;
2574 struct loop *current_loop;
2575 struct loop *next_loop;
2578 num_loops = loops->num;
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. */
2587 /* Find boundaries of loops, mark which loops are contained within
2588 loops, and invalidate loops that have setjmp. */
2591 current_loop = NULL;
2592 for (insn = f; insn; insn = NEXT_INSN (insn))
2594 if (GET_CODE (insn) == NOTE)
2595 switch (NOTE_LINE_NUMBER (insn))
2597 case NOTE_INSN_LOOP_BEG:
2598 next_loop = loops->array + num_loops;
2599 next_loop->num = num_loops;
2601 next_loop->start = insn;
2602 next_loop->outer = current_loop;
2603 current_loop = next_loop;
2606 case NOTE_INSN_SETJMP:
2607 /* In this case, we must invalidate our current loop and any
2609 for (loop = current_loop; loop; loop = loop->outer)
2612 if (loop_dump_stream)
2613 fprintf (loop_dump_stream,
2614 "\nLoop at %d ignored due to setjmp.\n",
2615 INSN_UID (loop->start));
2619 case NOTE_INSN_LOOP_CONT:
2620 current_loop->cont = insn;
2623 case NOTE_INSN_LOOP_VTOP:
2624 current_loop->vtop = insn;
2627 case NOTE_INSN_LOOP_END:
2631 current_loop->end = insn;
2632 verify_dominator (current_loop);
2633 current_loop = current_loop->outer;
2639 /* If for any loop, this is a jump insn between the NOTE_INSN_LOOP_CONT
2640 and NOTE_INSN_LOOP_END notes, update loop->cont_dominator. */
2641 else if (GET_CODE (insn) == JUMP_INSN
2642 && GET_CODE (PATTERN (insn)) != RETURN
2645 rtx label = JUMP_LABEL (insn);
2647 if (! any_condjump_p (insn))
2650 loop = current_loop;
2653 /* First see if we care about this loop. */
2654 if (loop->cont && loop->cont_dominator != const0_rtx)
2656 /* If the jump destination is not known, invalidate
2657 loop->cont_dominator. */
2659 loop->cont_dominator = const0_rtx;
2661 /* Check if the destination is between loop start and
2663 if ((INSN_LUID (label)
2664 < INSN_LUID (loop->cont))
2665 && (INSN_LUID (label)
2666 > INSN_LUID (loop->start))
2667 /* And if there is no later destination already
2669 && (! loop->cont_dominator
2670 || (INSN_LUID (label)
2671 > INSN_LUID (loop->cont_dominator))))
2672 loop->cont_dominator = label;
2679 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2680 enclosing loop, but this doesn't matter. */
2681 uid_loop[INSN_UID (insn)] = current_loop;
2684 /* Any loop containing a label used in an initializer must be invalidated,
2685 because it can be jumped into from anywhere. */
2687 for (label = forced_labels; label; label = XEXP (label, 1))
2689 for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
2690 loop; loop = loop->outer)
2694 /* Any loop containing a label used for an exception handler must be
2695 invalidated, because it can be jumped into from anywhere. */
2697 for (label = exception_handler_labels; label; label = XEXP (label, 1))
2699 for (loop = uid_loop[INSN_UID (XEXP (label, 0))];
2700 loop; loop = loop->outer)
2704 /* Now scan all insn's in the function. If any JUMP_INSN branches into a
2705 loop that it is not contained within, that loop is marked invalid.
2706 If any INSN or CALL_INSN uses a label's address, then the loop containing
2707 that label is marked invalid, because it could be jumped into from
2710 Also look for blocks of code ending in an unconditional branch that
2711 exits the loop. If such a block is surrounded by a conditional
2712 branch around the block, move the block elsewhere (see below) and
2713 invert the jump to point to the code block. This may eliminate a
2714 label in our loop and will simplify processing by both us and a
2715 possible second cse pass. */
2717 for (insn = f; insn; insn = NEXT_INSN (insn))
2720 struct loop *this_loop = uid_loop[INSN_UID (insn)];
2722 if (GET_CODE (insn) == INSN || GET_CODE (insn) == CALL_INSN)
2724 rtx note = find_reg_note (insn, REG_LABEL, NULL_RTX);
2727 for (loop = uid_loop[INSN_UID (XEXP (note, 0))];
2728 loop; loop = loop->outer)
2733 if (GET_CODE (insn) != JUMP_INSN)
2736 mark_loop_jump (PATTERN (insn), this_loop);
2738 /* See if this is an unconditional branch outside the loop. */
2740 && (GET_CODE (PATTERN (insn)) == RETURN
2741 || (any_uncondjump_p (insn)
2742 && onlyjump_p (insn)
2743 && (uid_loop[INSN_UID (JUMP_LABEL (insn))]
2745 && get_max_uid () < max_uid_for_loop)
2748 rtx our_next = next_real_insn (insn);
2749 rtx last_insn_to_move = NEXT_INSN (insn);
2750 struct loop *dest_loop;
2751 struct loop *outer_loop = NULL;
2753 /* Go backwards until we reach the start of the loop, a label,
2755 for (p = PREV_INSN (insn);
2756 GET_CODE (p) != CODE_LABEL
2757 && ! (GET_CODE (p) == NOTE
2758 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2759 && GET_CODE (p) != JUMP_INSN;
2763 /* Check for the case where we have a jump to an inner nested
2764 loop, and do not perform the optimization in that case. */
2766 if (JUMP_LABEL (insn))
2768 dest_loop = uid_loop[INSN_UID (JUMP_LABEL (insn))];
2771 for (outer_loop = dest_loop; outer_loop;
2772 outer_loop = outer_loop->outer)
2773 if (outer_loop == this_loop)
2778 /* Make sure that the target of P is within the current loop. */
2780 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
2781 && uid_loop[INSN_UID (JUMP_LABEL (p))] != this_loop)
2782 outer_loop = this_loop;
2784 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2785 we have a block of code to try to move.
2787 We look backward and then forward from the target of INSN
2788 to find a BARRIER at the same loop depth as the target.
2789 If we find such a BARRIER, we make a new label for the start
2790 of the block, invert the jump in P and point it to that label,
2791 and move the block of code to the spot we found. */
2794 && GET_CODE (p) == JUMP_INSN
2795 && JUMP_LABEL (p) != 0
2796 /* Just ignore jumps to labels that were never emitted.
2797 These always indicate compilation errors. */
2798 && INSN_UID (JUMP_LABEL (p)) != 0
2799 && any_condjump_p (p) && onlyjump_p (p)
2800 && next_real_insn (JUMP_LABEL (p)) == our_next
2801 /* If it's not safe to move the sequence, then we
2803 && insns_safe_to_move_p (p, NEXT_INSN (insn),
2804 &last_insn_to_move))
2807 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2808 struct loop *target_loop = uid_loop[INSN_UID (target)];
2811 for (loc = target; loc; loc = PREV_INSN (loc))
2812 if (GET_CODE (loc) == BARRIER
2813 /* Don't move things inside a tablejump. */
2814 && ((loc2 = next_nonnote_insn (loc)) == 0
2815 || GET_CODE (loc2) != CODE_LABEL
2816 || (loc2 = next_nonnote_insn (loc2)) == 0
2817 || GET_CODE (loc2) != JUMP_INSN
2818 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2819 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2820 && uid_loop[INSN_UID (loc)] == target_loop)
2824 for (loc = target; loc; loc = NEXT_INSN (loc))
2825 if (GET_CODE (loc) == BARRIER
2826 /* Don't move things inside a tablejump. */
2827 && ((loc2 = next_nonnote_insn (loc)) == 0
2828 || GET_CODE (loc2) != CODE_LABEL
2829 || (loc2 = next_nonnote_insn (loc2)) == 0
2830 || GET_CODE (loc2) != JUMP_INSN
2831 || (GET_CODE (PATTERN (loc2)) != ADDR_VEC
2832 && GET_CODE (PATTERN (loc2)) != ADDR_DIFF_VEC))
2833 && uid_loop[INSN_UID (loc)] == target_loop)
2838 rtx cond_label = JUMP_LABEL (p);
2839 rtx new_label = get_label_after (p);
2841 /* Ensure our label doesn't go away. */
2842 LABEL_NUSES (cond_label)++;
2844 /* Verify that uid_loop is large enough and that
2846 if (invert_jump (p, new_label, 1))
2850 /* If no suitable BARRIER was found, create a suitable
2851 one before TARGET. Since TARGET is a fall through
2852 path, we'll need to insert an jump around our block
2853 and a add a BARRIER before TARGET.
2855 This creates an extra unconditional jump outside
2856 the loop. However, the benefits of removing rarely
2857 executed instructions from inside the loop usually
2858 outweighs the cost of the extra unconditional jump
2859 outside the loop. */
2864 temp = gen_jump (JUMP_LABEL (insn));
2865 temp = emit_jump_insn_before (temp, target);
2866 JUMP_LABEL (temp) = JUMP_LABEL (insn);
2867 LABEL_NUSES (JUMP_LABEL (insn))++;
2868 loc = emit_barrier_before (target);
2871 /* Include the BARRIER after INSN and copy the
2873 new_label = squeeze_notes (new_label,
2875 reorder_insns (new_label, last_insn_to_move, loc);
2877 /* All those insns are now in TARGET_LOOP. */
2879 q != NEXT_INSN (last_insn_to_move);
2881 uid_loop[INSN_UID (q)] = target_loop;
2883 /* The label jumped to by INSN is no longer a loop
2884 exit. Unless INSN does not have a label (e.g.,
2885 it is a RETURN insn), search loop->exit_labels
2886 to find its label_ref, and remove it. Also turn
2887 off LABEL_OUTSIDE_LOOP_P bit. */
2888 if (JUMP_LABEL (insn))
2891 r = this_loop->exit_labels;
2892 r; q = r, r = LABEL_NEXTREF (r))
2893 if (XEXP (r, 0) == JUMP_LABEL (insn))
2895 LABEL_OUTSIDE_LOOP_P (r) = 0;
2897 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2899 this_loop->exit_labels = LABEL_NEXTREF (r);
2903 for (loop = this_loop; loop && loop != target_loop;
2907 /* If we didn't find it, then something is
2913 /* P is now a jump outside the loop, so it must be put
2914 in loop->exit_labels, and marked as such.
2915 The easiest way to do this is to just call
2916 mark_loop_jump again for P. */
2917 mark_loop_jump (PATTERN (p), this_loop);
2919 /* If INSN now jumps to the insn after it,
2921 if (JUMP_LABEL (insn) != 0
2922 && (next_real_insn (JUMP_LABEL (insn))
2923 == next_real_insn (insn)))
2927 /* Continue the loop after where the conditional
2928 branch used to jump, since the only branch insn
2929 in the block (if it still remains) is an inter-loop
2930 branch and hence needs no processing. */
2931 insn = NEXT_INSN (cond_label);
2933 if (--LABEL_NUSES (cond_label) == 0)
2934 delete_insn (cond_label);
2936 /* This loop will be continued with NEXT_INSN (insn). */
2937 insn = PREV_INSN (insn);
2944 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2945 loops it is contained in, mark the target loop invalid.
2947 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2950 mark_loop_jump (x, loop)
2954 struct loop *dest_loop;
2955 struct loop *outer_loop;
2958 switch (GET_CODE (x))
2971 /* There could be a label reference in here. */
2972 mark_loop_jump (XEXP (x, 0), loop);
2978 mark_loop_jump (XEXP (x, 0), loop);
2979 mark_loop_jump (XEXP (x, 1), loop);
2983 /* This may refer to a LABEL_REF or SYMBOL_REF. */
2984 mark_loop_jump (XEXP (x, 1), loop);
2989 mark_loop_jump (XEXP (x, 0), loop);
2993 dest_loop = uid_loop[INSN_UID (XEXP (x, 0))];
2995 /* Link together all labels that branch outside the loop. This
2996 is used by final_[bg]iv_value and the loop unrolling code. Also
2997 mark this LABEL_REF so we know that this branch should predict
3000 /* A check to make sure the label is not in an inner nested loop,
3001 since this does not count as a loop exit. */
3004 for (outer_loop = dest_loop; outer_loop;
3005 outer_loop = outer_loop->outer)
3006 if (outer_loop == loop)
3012 if (loop && ! outer_loop)
3014 LABEL_OUTSIDE_LOOP_P (x) = 1;
3015 LABEL_NEXTREF (x) = loop->exit_labels;
3016 loop->exit_labels = x;
3018 for (outer_loop = loop;
3019 outer_loop && outer_loop != dest_loop;
3020 outer_loop = outer_loop->outer)
3021 outer_loop->exit_count++;
3024 /* If this is inside a loop, but not in the current loop or one enclosed
3025 by it, it invalidates at least one loop. */
3030 /* We must invalidate every nested loop containing the target of this
3031 label, except those that also contain the jump insn. */
3033 for (; dest_loop; dest_loop = dest_loop->outer)
3035 /* Stop when we reach a loop that also contains the jump insn. */
3036 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3037 if (dest_loop == outer_loop)
3040 /* If we get here, we know we need to invalidate a loop. */
3041 if (loop_dump_stream && ! dest_loop->invalid)
3042 fprintf (loop_dump_stream,
3043 "\nLoop at %d ignored due to multiple entry points.\n",
3044 INSN_UID (dest_loop->start));
3046 dest_loop->invalid = 1;
3051 /* If this is not setting pc, ignore. */
3052 if (SET_DEST (x) == pc_rtx)
3053 mark_loop_jump (SET_SRC (x), loop);
3057 mark_loop_jump (XEXP (x, 1), loop);
3058 mark_loop_jump (XEXP (x, 2), loop);
3063 for (i = 0; i < XVECLEN (x, 0); i++)
3064 mark_loop_jump (XVECEXP (x, 0, i), loop);
3068 for (i = 0; i < XVECLEN (x, 1); i++)
3069 mark_loop_jump (XVECEXP (x, 1, i), loop);
3073 /* Strictly speaking this is not a jump into the loop, only a possible
3074 jump out of the loop. However, we have no way to link the destination
3075 of this jump onto the list of exit labels. To be safe we mark this
3076 loop and any containing loops as invalid. */
3079 for (outer_loop = loop; outer_loop; outer_loop = outer_loop->outer)
3081 if (loop_dump_stream && ! outer_loop->invalid)
3082 fprintf (loop_dump_stream,
3083 "\nLoop at %d ignored due to unknown exit jump.\n",
3084 INSN_UID (outer_loop->start));
3085 outer_loop->invalid = 1;
3092 /* Return nonzero if there is a label in the range from
3093 insn INSN to and including the insn whose luid is END
3094 INSN must have an assigned luid (i.e., it must not have
3095 been previously created by loop.c). */
3098 labels_in_range_p (insn, end)
3102 while (insn && INSN_LUID (insn) <= end)
3104 if (GET_CODE (insn) == CODE_LABEL)
3106 insn = NEXT_INSN (insn);
3112 /* Record that a memory reference X is being set. */
3115 note_addr_stored (x, y, data)
3117 rtx y ATTRIBUTE_UNUSED;
3118 void *data ATTRIBUTE_UNUSED;
3120 if (x == 0 || GET_CODE (x) != MEM)
3123 /* Count number of memory writes.
3124 This affects heuristics in strength_reduce. */
3127 /* BLKmode MEM means all memory is clobbered. */
3128 if (GET_MODE (x) == BLKmode)
3130 if (RTX_UNCHANGING_P (x))
3131 unknown_constant_address_altered = 1;
3133 unknown_address_altered = 1;
3138 loop_store_mems = gen_rtx_EXPR_LIST (VOIDmode, x, loop_store_mems);
3141 /* X is a value modified by an INSN that references a biv inside a loop
3142 exit test (ie, X is somehow related to the value of the biv). If X
3143 is a pseudo that is used more than once, then the biv is (effectively)
3144 used more than once. DATA is really an `int *', and is set if the
3145 biv is used more than once. */
3148 note_set_pseudo_multiple_uses (x, y, data)
3150 rtx y ATTRIBUTE_UNUSED;
3156 while (GET_CODE (x) == STRICT_LOW_PART
3157 || GET_CODE (x) == SIGN_EXTRACT
3158 || GET_CODE (x) == ZERO_EXTRACT
3159 || GET_CODE (x) == SUBREG)
3162 if (GET_CODE (x) != REG || REGNO (x) < FIRST_PSEUDO_REGISTER)
3165 /* If we do not have usage information, or if we know the register
3166 is used more than once, note that fact for check_dbra_loop. */
3167 if (REGNO (x) >= max_reg_before_loop
3168 || ! VARRAY_RTX (reg_single_usage, REGNO (x))
3169 || VARRAY_RTX (reg_single_usage, REGNO (x)) == const0_rtx)
3170 *((int *) data) = 1;
3173 /* Return nonzero if the rtx X is invariant over the current loop.
3175 The value is 2 if we refer to something only conditionally invariant.
3177 If `unknown_address_altered' is nonzero, no memory ref is invariant.
3178 Otherwise, a memory ref is invariant if it does not conflict with
3179 anything stored in `loop_store_mems'. */
3182 loop_invariant_p (loop, x)
3183 const struct loop *loop;
3187 register enum rtx_code code;
3188 register const char *fmt;
3189 int conditional = 0;
3194 code = GET_CODE (x);
3204 /* A LABEL_REF is normally invariant, however, if we are unrolling
3205 loops, and this label is inside the loop, then it isn't invariant.
3206 This is because each unrolled copy of the loop body will have
3207 a copy of this label. If this was invariant, then an insn loading
3208 the address of this label into a register might get moved outside
3209 the loop, and then each loop body would end up using the same label.
3211 We don't know the loop bounds here though, so just fail for all
3213 if (flag_unroll_loops)
3220 case UNSPEC_VOLATILE:
3224 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
3225 since the reg might be set by initialization within the loop. */
3227 if ((x == frame_pointer_rtx || x == hard_frame_pointer_rtx
3228 || x == arg_pointer_rtx)
3229 && ! current_function_has_nonlocal_goto)
3232 if (LOOP_INFO (loop)->has_call
3233 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
3236 if (VARRAY_INT (set_in_loop, REGNO (x)) < 0)
3239 return VARRAY_INT (set_in_loop, REGNO (x)) == 0;
3242 /* Volatile memory references must be rejected. Do this before
3243 checking for read-only items, so that volatile read-only items
3244 will be rejected also. */
3245 if (MEM_VOLATILE_P (x))
3248 /* If we had a subroutine call, any location in memory could
3249 have been clobbered. We used to test here for volatile and
3250 readonly, but true_dependence knows how to do that better
3252 if (RTX_UNCHANGING_P (x)
3253 ? unknown_constant_address_altered : unknown_address_altered)
3256 /* See if there is any dependence between a store and this load. */
3257 mem_list_entry = loop_store_mems;
3258 while (mem_list_entry)
3260 if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
3264 mem_list_entry = XEXP (mem_list_entry, 1);
3267 /* It's not invalidated by a store in memory
3268 but we must still verify the address is invariant. */
3272 /* Don't mess with insns declared volatile. */
3273 if (MEM_VOLATILE_P (x))
3281 fmt = GET_RTX_FORMAT (code);
3282 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3286 int tem = loop_invariant_p (loop, XEXP (x, i));
3292 else if (fmt[i] == 'E')
3295 for (j = 0; j < XVECLEN (x, i); j++)
3297 int tem = loop_invariant_p (loop, XVECEXP (x, i, j));
3307 return 1 + conditional;
3310 /* Return nonzero if all the insns in the loop that set REG
3311 are INSN and the immediately following insns,
3312 and if each of those insns sets REG in an invariant way
3313 (not counting uses of REG in them).
3315 The value is 2 if some of these insns are only conditionally invariant.
3317 We assume that INSN itself is the first set of REG
3318 and that its source is invariant. */
3321 consec_sets_invariant_p (loop, reg, n_sets, insn)
3322 const struct loop *loop;
3327 unsigned int regno = REGNO (reg);
3329 /* Number of sets we have to insist on finding after INSN. */
3330 int count = n_sets - 1;
3331 int old = VARRAY_INT (set_in_loop, regno);
3335 /* If N_SETS hit the limit, we can't rely on its value. */
3339 VARRAY_INT (set_in_loop, regno) = 0;
3343 register enum rtx_code code;
3347 code = GET_CODE (p);
3349 /* If library call, skip to end of it. */
3350 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3355 && (set = single_set (p))
3356 && GET_CODE (SET_DEST (set)) == REG
3357 && REGNO (SET_DEST (set)) == regno)
3359 this = loop_invariant_p (loop, SET_SRC (set));
3362 else if ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX)))
3364 /* If this is a libcall, then any invariant REG_EQUAL note is OK.
3365 If this is an ordinary insn, then only CONSTANT_P REG_EQUAL
3367 this = (CONSTANT_P (XEXP (temp, 0))
3368 || (find_reg_note (p, REG_RETVAL, NULL_RTX)
3369 && loop_invariant_p (loop, XEXP (temp, 0))));
3376 else if (code != NOTE)
3378 VARRAY_INT (set_in_loop, regno) = old;
3383 VARRAY_INT (set_in_loop, regno) = old;
3384 /* If loop_invariant_p ever returned 2, we return 2. */
3385 return 1 + (value & 2);
3389 /* I don't think this condition is sufficient to allow INSN
3390 to be moved, so we no longer test it. */
3392 /* Return 1 if all insns in the basic block of INSN and following INSN
3393 that set REG are invariant according to TABLE. */
3396 all_sets_invariant_p (reg, insn, table)
3400 register rtx p = insn;
3401 register int regno = REGNO (reg);
3405 register enum rtx_code code;
3407 code = GET_CODE (p);
3408 if (code == CODE_LABEL || code == JUMP_INSN)
3410 if (code == INSN && GET_CODE (PATTERN (p)) == SET
3411 && GET_CODE (SET_DEST (PATTERN (p))) == REG
3412 && REGNO (SET_DEST (PATTERN (p))) == regno)
3414 if (! loop_invariant_p (loop, SET_SRC (PATTERN (p)), table))
3421 /* Look at all uses (not sets) of registers in X. For each, if it is
3422 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
3423 a different insn, set USAGE[REGNO] to const0_rtx. */
3426 find_single_use_in_loop (insn, x, usage)
3431 enum rtx_code code = GET_CODE (x);
3432 const char *fmt = GET_RTX_FORMAT (code);
3436 VARRAY_RTX (usage, REGNO (x))
3437 = (VARRAY_RTX (usage, REGNO (x)) != 0
3438 && VARRAY_RTX (usage, REGNO (x)) != insn)
3439 ? const0_rtx : insn;
3441 else if (code == SET)
3443 /* Don't count SET_DEST if it is a REG; otherwise count things
3444 in SET_DEST because if a register is partially modified, it won't
3445 show up as a potential movable so we don't care how USAGE is set
3447 if (GET_CODE (SET_DEST (x)) != REG)
3448 find_single_use_in_loop (insn, SET_DEST (x), usage);
3449 find_single_use_in_loop (insn, SET_SRC (x), usage);
3452 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3454 if (fmt[i] == 'e' && XEXP (x, i) != 0)
3455 find_single_use_in_loop (insn, XEXP (x, i), usage);
3456 else if (fmt[i] == 'E')
3457 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3458 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
3462 /* Count and record any set in X which is contained in INSN. Update
3463 MAY_NOT_MOVE and LAST_SET for any register set in X. */
3466 count_one_set (insn, x, may_not_move, last_set)
3468 varray_type may_not_move;
3471 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
3472 /* Don't move a reg that has an explicit clobber.
3473 It's not worth the pain to try to do it correctly. */
3474 VARRAY_CHAR (may_not_move, REGNO (XEXP (x, 0))) = 1;
3476 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
3478 rtx dest = SET_DEST (x);
3479 while (GET_CODE (dest) == SUBREG
3480 || GET_CODE (dest) == ZERO_EXTRACT
3481 || GET_CODE (dest) == SIGN_EXTRACT
3482 || GET_CODE (dest) == STRICT_LOW_PART)
3483 dest = XEXP (dest, 0);
3484 if (GET_CODE (dest) == REG)
3486 register int regno = REGNO (dest);
3487 /* If this is the first setting of this reg
3488 in current basic block, and it was set before,
3489 it must be set in two basic blocks, so it cannot
3490 be moved out of the loop. */
3491 if (VARRAY_INT (set_in_loop, regno) > 0
3492 && last_set[regno] == 0)
3493 VARRAY_CHAR (may_not_move, regno) = 1;
3494 /* If this is not first setting in current basic block,
3495 see if reg was used in between previous one and this.
3496 If so, neither one can be moved. */
3497 if (last_set[regno] != 0
3498 && reg_used_between_p (dest, last_set[regno], insn))
3499 VARRAY_CHAR (may_not_move, regno) = 1;
3500 if (VARRAY_INT (set_in_loop, regno) < 127)
3501 ++VARRAY_INT (set_in_loop, regno);
3502 last_set[regno] = insn;
3507 /* Increment SET_IN_LOOP at the index of each register
3508 that is modified by an insn between FROM and TO.
3509 If the value of an element of SET_IN_LOOP becomes 127 or more,
3510 stop incrementing it, to avoid overflow.
3512 Store in SINGLE_USAGE[I] the single insn in which register I is
3513 used, if it is only used once. Otherwise, it is set to 0 (for no
3514 uses) or const0_rtx for more than one use. This parameter may be zero,
3515 in which case this processing is not done.
3517 Store in *COUNT_PTR the number of actual instruction
3518 in the loop. We use this to decide what is worth moving out. */
3520 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
3521 In that case, it is the insn that last set reg n. */
3524 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
3525 register rtx from, to;
3526 varray_type may_not_move;
3527 varray_type single_usage;
3531 register rtx *last_set = (rtx *) xcalloc (nregs, sizeof (rtx));
3533 register int count = 0;
3535 for (insn = from; insn != to; insn = NEXT_INSN (insn))
3541 /* Record registers that have exactly one use. */
3542 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
3544 /* Include uses in REG_EQUAL notes. */
3545 if (REG_NOTES (insn))
3546 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
3548 if (GET_CODE (PATTERN (insn)) == SET
3549 || GET_CODE (PATTERN (insn)) == CLOBBER)
3550 count_one_set (insn, PATTERN (insn), may_not_move, last_set);
3551 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
3554 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
3555 count_one_set (insn, XVECEXP (PATTERN (insn), 0, i),
3556 may_not_move, last_set);
3560 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
3561 bzero ((char *) last_set, nregs * sizeof (rtx));
3569 /* Given a loop that is bounded by LOOP->START and LOOP->END and that
3570 is entered at LOOP->SCAN_START, return 1 if the register set in SET
3571 contained in insn INSN is used by any insn that precedes INSN in
3572 cyclic order starting from the loop entry point.
3574 We don't want to use INSN_LUID here because if we restrict INSN to those
3575 that have a valid INSN_LUID, it means we cannot move an invariant out
3576 from an inner loop past two loops. */
3579 loop_reg_used_before_p (loop, set, insn)
3580 const struct loop *loop;
3583 rtx reg = SET_DEST (set);
3586 /* Scan forward checking for register usage. If we hit INSN, we
3587 are done. Otherwise, if we hit LOOP->END, wrap around to LOOP->START. */
3588 for (p = loop->scan_start; p != insn; p = NEXT_INSN (p))
3590 if (INSN_P (p) && reg_overlap_mentioned_p (reg, PATTERN (p)))
3600 /* A "basic induction variable" or biv is a pseudo reg that is set
3601 (within this loop) only by incrementing or decrementing it. */
3602 /* A "general induction variable" or giv is a pseudo reg whose
3603 value is a linear function of a biv. */
3605 /* Bivs are recognized by `basic_induction_var';
3606 Givs by `general_induction_var'. */
3608 /* Indexed by register number, indicates whether or not register is an
3609 induction variable, and if so what type. */
3611 varray_type reg_iv_type;
3613 /* Indexed by register number, contains pointer to `struct induction'
3614 if register is an induction variable. This holds general info for
3615 all induction variables. */
3617 varray_type reg_iv_info;
3619 /* Indexed by register number, contains pointer to `struct iv_class'
3620 if register is a basic induction variable. This holds info describing
3621 the class (a related group) of induction variables that the biv belongs
3624 struct iv_class **reg_biv_class;
3626 /* The head of a list which links together (via the next field)
3627 every iv class for the current loop. */
3629 struct iv_class *loop_iv_list;
3631 /* Givs made from biv increments are always splittable for loop unrolling.
3632 Since there is no regscan info for them, we have to keep track of them
3634 unsigned int first_increment_giv, last_increment_giv;
3636 /* Communication with routines called via `note_stores'. */
3638 static rtx note_insn;
3640 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3642 static rtx addr_placeholder;
3644 /* ??? Unfinished optimizations, and possible future optimizations,
3645 for the strength reduction code. */
3647 /* ??? The interaction of biv elimination, and recognition of 'constant'
3648 bivs, may cause problems. */
3650 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3651 performance problems.
3653 Perhaps don't eliminate things that can be combined with an addressing
3654 mode. Find all givs that have the same biv, mult_val, and add_val;
3655 then for each giv, check to see if its only use dies in a following
3656 memory address. If so, generate a new memory address and check to see
3657 if it is valid. If it is valid, then store the modified memory address,
3658 otherwise, mark the giv as not done so that it will get its own iv. */
3660 /* ??? Could try to optimize branches when it is known that a biv is always
3663 /* ??? When replace a biv in a compare insn, we should replace with closest
3664 giv so that an optimized branch can still be recognized by the combiner,
3665 e.g. the VAX acb insn. */
3667 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3668 was rerun in loop_optimize whenever a register was added or moved.
3669 Also, some of the optimizations could be a little less conservative. */
3671 /* Scan the loop body and call FNCALL for each insn. In the addition to the
3672 LOOP and INSN parameters pass MAYBE_MULTIPLE and NOT_EVERY_ITERATION to the
3675 NOT_EVERY_ITERATION if current insn is not executed at least once for every
3676 loop iteration except for the last one.
3678 MAYBE_MULTIPLE is 1 if current insn may be executed more than once for every
3682 for_each_insn_in_loop (loop, fncall)
3684 loop_insn_callback fncall;
3686 /* This is 1 if current insn is not executed at least once for every loop
3688 int not_every_iteration = 0;
3689 int maybe_multiple = 0;
3690 int past_loop_latch = 0;
3694 /* If loop_scan_start points to the loop exit test, we have to be wary of
3695 subversive use of gotos inside expression statements. */
3696 if (prev_nonnote_insn (loop->scan_start) != prev_nonnote_insn (loop->start))
3697 maybe_multiple = back_branch_in_range_p (loop, loop->scan_start);
3699 /* Scan through loop to find all possible bivs. */
3701 for (p = next_insn_in_loop (loop, loop->scan_start);
3703 p = next_insn_in_loop (loop, p))
3705 p = fncall (loop, p, not_every_iteration, maybe_multiple);
3707 /* Past CODE_LABEL, we get to insns that may be executed multiple
3708 times. The only way we can be sure that they can't is if every
3709 jump insn between here and the end of the loop either
3710 returns, exits the loop, is a jump to a location that is still
3711 behind the label, or is a jump to the loop start. */
3713 if (GET_CODE (p) == CODE_LABEL)
3721 insn = NEXT_INSN (insn);
3722 if (insn == loop->scan_start)
3724 if (insn == loop->end)
3730 if (insn == loop->scan_start)
3734 if (GET_CODE (insn) == JUMP_INSN
3735 && GET_CODE (PATTERN (insn)) != RETURN
3736 && (!any_condjump_p (insn)
3737 || (JUMP_LABEL (insn) != 0
3738 && JUMP_LABEL (insn) != loop->scan_start
3739 && !loop_insn_first_p (p, JUMP_LABEL (insn)))))
3747 /* Past a jump, we get to insns for which we can't count
3748 on whether they will be executed during each iteration. */
3749 /* This code appears twice in strength_reduce. There is also similar
3750 code in scan_loop. */
3751 if (GET_CODE (p) == JUMP_INSN
3752 /* If we enter the loop in the middle, and scan around to the
3753 beginning, don't set not_every_iteration for that.
3754 This can be any kind of jump, since we want to know if insns
3755 will be executed if the loop is executed. */
3756 && !(JUMP_LABEL (p) == loop->top
3757 && ((NEXT_INSN (NEXT_INSN (p)) == loop->end
3758 && any_uncondjump_p (p))
3759 || (NEXT_INSN (p) == loop->end && any_condjump_p (p)))))
3763 /* If this is a jump outside the loop, then it also doesn't
3764 matter. Check to see if the target of this branch is on the
3765 loop->exits_labels list. */
3767 for (label = loop->exit_labels; label; label = LABEL_NEXTREF (label))
3768 if (XEXP (label, 0) == JUMP_LABEL (p))
3772 not_every_iteration = 1;
3775 else if (GET_CODE (p) == NOTE)
3777 /* At the virtual top of a converted loop, insns are again known to
3778 be executed each iteration: logically, the loop begins here
3779 even though the exit code has been duplicated.
3781 Insns are also again known to be executed each iteration at
3782 the LOOP_CONT note. */
3783 if ((NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP
3784 || NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_CONT)
3786 not_every_iteration = 0;
3787 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
3789 else if (NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_END)
3793 /* Note if we pass a loop latch. If we do, then we can not clear
3794 NOT_EVERY_ITERATION below when we pass the last CODE_LABEL in
3795 a loop since a jump before the last CODE_LABEL may have started
3796 a new loop iteration.
3798 Note that LOOP_TOP is only set for rotated loops and we need
3799 this check for all loops, so compare against the CODE_LABEL
3800 which immediately follows LOOP_START. */
3801 if (GET_CODE (p) == JUMP_INSN
3802 && JUMP_LABEL (p) == NEXT_INSN (loop->start))
3803 past_loop_latch = 1;
3805 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3806 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3807 or not an insn is known to be executed each iteration of the
3808 loop, whether or not any iterations are known to occur.
3810 Therefore, if we have just passed a label and have no more labels
3811 between here and the test insn of the loop, and we have not passed
3812 a jump to the top of the loop, then we know these insns will be
3813 executed each iteration. */
3815 if (not_every_iteration
3817 && GET_CODE (p) == CODE_LABEL
3818 && no_labels_between_p (p, loop->end)
3819 && loop_insn_first_p (p, loop->cont))
3820 not_every_iteration = 0;
3824 /* Perform strength reduction and induction variable elimination.
3826 Pseudo registers created during this function will be beyond the last
3827 valid index in several tables including n_times_set and regno_last_uid.
3828 This does not cause a problem here, because the added registers cannot be
3829 givs outside of their loop, and hence will never be reconsidered.
3830 But scan_loop must check regnos to make sure they are in bounds. */
3833 strength_reduce (loop, insn_count, flags)
3839 /* Temporary list pointers for traversing loop_iv_list. */
3840 struct iv_class *bl, **backbl;
3841 struct loop_info *loop_info = LOOP_INFO (loop);
3842 /* Ratio of extra register life span we can justify
3843 for saving an instruction. More if loop doesn't call subroutines
3844 since in that case saving an insn makes more difference
3845 and more registers are available. */
3846 /* ??? could set this to last value of threshold in move_movables */
3847 int threshold = (loop_info->has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3848 /* Map of pseudo-register replacements. */
3849 rtx *reg_map = NULL;
3853 rtx end_insert_before;
3854 int n_extra_increment;
3855 int unrolled_insn_copies = 0;
3856 rtx loop_start = loop->start;
3857 rtx loop_end = loop->end;
3858 rtx loop_scan_start = loop->scan_start;
3860 VARRAY_INT_INIT (reg_iv_type, max_reg_before_loop, "reg_iv_type");
3861 VARRAY_GENERIC_PTR_INIT (reg_iv_info, max_reg_before_loop, "reg_iv_info");
3862 reg_biv_class = (struct iv_class **)
3863 xcalloc (max_reg_before_loop, sizeof (struct iv_class *));
3866 addr_placeholder = gen_reg_rtx (Pmode);
3868 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3869 must be put before this insn, so that they will appear in the right
3870 order (i.e. loop order).
3872 If loop_end is the end of the current function, then emit a
3873 NOTE_INSN_DELETED after loop_end and set end_insert_before to the
3875 if (NEXT_INSN (loop_end) != 0)
3876 end_insert_before = NEXT_INSN (loop_end);
3878 end_insert_before = emit_note_after (NOTE_INSN_DELETED, loop_end);
3880 for_each_insn_in_loop (loop, check_insn_for_bivs);
3882 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3883 Make a sanity check against n_times_set. */
3884 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3886 if (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3887 /* Above happens if register modified by subreg, etc. */
3888 /* Make sure it is not recognized as a basic induction var: */
3889 || VARRAY_INT (n_times_set, bl->regno) != bl->biv_count
3890 /* If never incremented, it is invariant that we decided not to
3891 move. So leave it alone. */
3892 || ! bl->incremented)
3894 if (loop_dump_stream)
3895 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3897 (REG_IV_TYPE (bl->regno) != BASIC_INDUCT
3898 ? "not induction variable"
3899 : (! bl->incremented ? "never incremented"
3902 REG_IV_TYPE (bl->regno) = NOT_BASIC_INDUCT;
3909 if (loop_dump_stream)
3910 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3914 /* Exit if there are no bivs. */
3917 /* Can still unroll the loop anyways, but indicate that there is no
3918 strength reduction info available. */
3919 if (flags & LOOP_UNROLL)
3920 unroll_loop (loop, insn_count, end_insert_before, 0);
3925 /* Find initial value for each biv by searching backwards from loop_start,
3926 halting at first label. Also record any test condition. */
3929 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3933 if (GET_CODE (p) == CALL_INSN)
3937 note_stores (PATTERN (p), record_initial, NULL);
3939 /* Record any test of a biv that branches around the loop if no store
3940 between it and the start of loop. We only care about tests with
3941 constants and registers and only certain of those. */
3942 if (GET_CODE (p) == JUMP_INSN
3943 && JUMP_LABEL (p) != 0
3944 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3945 && (test = get_condition_for_loop (loop, p)) != 0
3946 && GET_CODE (XEXP (test, 0)) == REG
3947 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3948 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3949 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3950 && bl->init_insn == 0)
3952 /* If an NE test, we have an initial value! */
3953 if (GET_CODE (test) == NE)
3956 bl->init_set = gen_rtx_SET (VOIDmode,
3957 XEXP (test, 0), XEXP (test, 1));
3960 bl->initial_test = test;
3964 /* Look at the each biv and see if we can say anything better about its
3965 initial value from any initializing insns set up above. (This is done
3966 in two passes to avoid missing SETs in a PARALLEL.) */
3967 for (backbl = &loop_iv_list; (bl = *backbl); backbl = &bl->next)
3972 if (! bl->init_insn)
3975 /* IF INIT_INSN has a REG_EQUAL or REG_EQUIV note and the value
3976 is a constant, use the value of that. */
3977 if (((note = find_reg_note (bl->init_insn, REG_EQUAL, 0)) != NULL
3978 && CONSTANT_P (XEXP (note, 0)))
3979 || ((note = find_reg_note (bl->init_insn, REG_EQUIV, 0)) != NULL
3980 && CONSTANT_P (XEXP (note, 0))))
3981 src = XEXP (note, 0);
3983 src = SET_SRC (bl->init_set);
3985 if (loop_dump_stream)
3986 fprintf (loop_dump_stream,
3987 "Biv %d initialized at insn %d: initial value ",
3988 bl->regno, INSN_UID (bl->init_insn));
3990 if ((GET_MODE (src) == GET_MODE (regno_reg_rtx[bl->regno])
3991 || GET_MODE (src) == VOIDmode)
3992 && valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
3994 bl->initial_value = src;
3996 if (loop_dump_stream)
3998 if (GET_CODE (src) == CONST_INT)
4000 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (src));
4001 fputc ('\n', loop_dump_stream);
4005 print_rtl (loop_dump_stream, src);
4006 fprintf (loop_dump_stream, "\n");
4012 struct iv_class *bl2 = 0;
4013 rtx increment = NULL_RTX;
4015 /* Biv initial value is not a simple move. If it is the sum of
4016 another biv and a constant, check if both bivs are incremented
4017 in lockstep. Then we are actually looking at a giv.
4018 For simplicity, we only handle the case where there is but a
4019 single increment, and the register is not used elsewhere. */
4020 if (bl->biv_count == 1
4021 && bl->regno < max_reg_before_loop
4022 && uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4023 && GET_CODE (src) == PLUS
4024 && GET_CODE (XEXP (src, 0)) == REG
4025 && CONSTANT_P (XEXP (src, 1))
4026 && ((increment = biv_total_increment (bl)) != NULL_RTX))
4028 unsigned int regno = REGNO (XEXP (src, 0));
4030 for (bl2 = loop_iv_list; bl2; bl2 = bl2->next)
4031 if (bl2->regno == regno)
4035 /* Now, can we transform this biv into a giv? */
4037 && bl2->biv_count == 1
4038 && rtx_equal_p (increment, biv_total_increment (bl2))
4039 /* init_insn is only set to insns that are before loop_start
4040 without any intervening labels. */
4041 && ! reg_set_between_p (bl2->biv->src_reg,
4042 PREV_INSN (bl->init_insn), loop_start)
4043 /* The register from BL2 must be set before the register from
4044 BL is set, or we must be able to move the latter set after
4045 the former set. Currently there can't be any labels
4046 in-between when biv_total_increment returns nonzero both times
4047 but we test it here in case some day some real cfg analysis
4048 gets used to set always_computable. */
4049 && (loop_insn_first_p (bl2->biv->insn, bl->biv->insn)
4050 ? no_labels_between_p (bl2->biv->insn, bl->biv->insn)
4051 : (! reg_used_between_p (bl->biv->src_reg, bl->biv->insn,
4053 && no_jumps_between_p (bl->biv->insn, bl2->biv->insn)))
4054 && validate_change (bl->biv->insn,
4055 &SET_SRC (single_set (bl->biv->insn)),
4058 rtx dominator = loop->cont_dominator;
4059 rtx giv = bl->biv->src_reg;
4060 rtx giv_insn = bl->biv->insn;
4061 rtx after_giv = NEXT_INSN (giv_insn);
4063 if (loop_dump_stream)
4064 fprintf (loop_dump_stream, "is giv of biv %d\n", bl2->regno);
4065 /* Let this giv be discovered by the generic code. */
4066 REG_IV_TYPE (bl->regno) = UNKNOWN_INDUCT;
4067 reg_biv_class[bl->regno] = (struct iv_class *) NULL_PTR;
4068 /* We can get better optimization if we can move the giv setting
4069 before the first giv use. */
4071 && ! loop_insn_first_p (dominator, loop_scan_start)
4072 && ! reg_set_between_p (bl2->biv->src_reg, loop_start,
4074 && ! reg_used_between_p (giv, loop_start, dominator)
4075 && ! reg_used_between_p (giv, giv_insn, loop_end))
4080 for (next = NEXT_INSN (dominator);; next = NEXT_INSN (next))
4082 if (GET_CODE (next) == JUMP_INSN
4084 && insn_dependent_p (giv_insn, next)))
4087 if (! INSN_P (next) || ! sets_cc0_p (PATTERN (next)))
4091 if (loop_dump_stream)
4092 fprintf (loop_dump_stream, "move after insn %d\n",
4093 INSN_UID (dominator));
4094 /* Avoid problems with luids by actually moving the insn
4095 and adjusting all luids in the range. */
4096 reorder_insns (giv_insn, giv_insn, dominator);
4097 for (p = dominator; INSN_UID (p) >= max_uid_for_loop;)
4099 compute_luids (giv_insn, after_giv, INSN_LUID (p));
4100 /* If the only purpose of the init insn is to initialize
4101 this giv, delete it. */
4102 if (single_set (bl->init_insn)
4103 && ! reg_used_between_p (giv, bl->init_insn, loop_start))
4104 delete_insn (bl->init_insn);
4106 else if (! loop_insn_first_p (bl2->biv->insn, bl->biv->insn))
4108 rtx p = PREV_INSN (giv_insn);
4109 while (INSN_UID (p) >= max_uid_for_loop)
4111 reorder_insns (giv_insn, giv_insn, bl2->biv->insn);
4112 compute_luids (after_giv, NEXT_INSN (giv_insn),
4115 /* Remove this biv from the chain. */
4119 /* If we can't make it a giv,
4120 let biv keep initial value of "itself". */
4121 else if (loop_dump_stream)
4122 fprintf (loop_dump_stream, "is complex\n");
4126 /* If a biv is unconditionally incremented several times in a row, convert
4127 all but the last increment into a giv. */
4129 /* Get an upper bound for the number of registers
4130 we might have after all bivs have been processed. */
4131 first_increment_giv = max_reg_num ();
4132 for (n_extra_increment = 0, bl = loop_iv_list; bl; bl = bl->next)
4133 n_extra_increment += bl->biv_count - 1;
4135 /* If the loop contains volatile memory references do not allow any
4136 replacements to take place, since this could loose the volatile
4138 if (n_extra_increment && ! loop_info->has_volatile)
4140 unsigned int nregs = first_increment_giv + n_extra_increment;
4142 /* Reallocate reg_iv_type and reg_iv_info. */
4143 VARRAY_GROW (reg_iv_type, nregs);
4144 VARRAY_GROW (reg_iv_info, nregs);
4146 for (bl = loop_iv_list; bl; bl = bl->next)
4148 struct induction **vp, *v, *next;
4149 int biv_dead_after_loop = 0;
4151 /* The biv increments lists are in reverse order. Fix this
4153 for (v = bl->biv, bl->biv = 0; v; v = next)
4156 v->next_iv = bl->biv;
4160 /* We must guard against the case that an early exit between v->insn
4161 and next->insn leaves the biv live after the loop, since that
4162 would mean that we'd be missing an increment for the final
4163 value. The following test to set biv_dead_after_loop is like
4164 the first part of the test to set bl->eliminable.
4165 We don't check here if we can calculate the final value, since
4166 this can't succeed if we already know that there is a jump
4167 between v->insn and next->insn, yet next->always_executed is
4168 set and next->maybe_multiple is cleared. Such a combination
4169 implies that the jump destination is outside the loop.
4170 If we want to make this check more sophisticated, we should
4171 check each branch between v->insn and next->insn individually
4172 to see if the biv is dead at its destination. */
4174 if (uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4176 && INSN_UID (bl->init_insn) < max_uid_for_loop
4177 && (uid_luid[REGNO_FIRST_UID (bl->regno)]
4178 >= INSN_LUID (bl->init_insn))
4179 #ifdef HAVE_decrement_and_branch_until_zero
4182 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4183 biv_dead_after_loop = 1;
4185 for (vp = &bl->biv, next = *vp; v = next, next = v->next_iv;)
4187 HOST_WIDE_INT offset;
4188 rtx set, add_val, old_reg, dest_reg, last_use_insn, note;
4189 int old_regno, new_regno;
4192 if (! v->always_executed
4193 || v->maybe_multiple
4194 || GET_CODE (v->add_val) != CONST_INT
4195 || ! next->always_executed
4196 || next->maybe_multiple
4197 || ! CONSTANT_P (next->add_val)
4198 || v->mult_val != const1_rtx
4199 || next->mult_val != const1_rtx
4200 || ! (biv_dead_after_loop
4201 || no_jumps_between_p (v->insn, next->insn)))
4206 offset = INTVAL (v->add_val);
4207 set = single_set (v->insn);
4208 add_val = plus_constant (next->add_val, offset);
4209 old_reg = v->dest_reg;
4210 dest_reg = gen_reg_rtx (v->mode);
4212 /* Unlike reg_iv_type / reg_iv_info, the other three arrays
4213 have been allocated with some slop space, so we may not
4214 actually need to reallocate them. If we do, the following
4215 if statement will be executed just once in this loop. */
4216 if ((unsigned) max_reg_num () > n_times_set->num_elements)
4218 /* Grow all the remaining arrays. */
4219 VARRAY_GROW (set_in_loop, nregs);
4220 VARRAY_GROW (n_times_set, nregs);
4221 VARRAY_GROW (may_not_optimize, nregs);
4222 VARRAY_GROW (reg_single_usage, nregs);
4225 /* Some bivs are incremented with a multi-insn sequence.
4226 The first insn contains the add. */
4227 next_loc_insn = next->insn;
4228 while (! loc_mentioned_in_p (next->location,
4229 PATTERN (next_loc_insn)))
4230 next_loc_insn = PREV_INSN (next_loc_insn);
4232 if (next_loc_insn == v->insn)
4235 if (! validate_change (next_loc_insn, next->location, add_val, 0))
4241 /* Here we can try to eliminate the increment by combining
4242 it into the uses. */
4244 /* Set last_use_insn so that we can check against it. */
4246 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4248 p = next_insn_in_loop (loop, p))
4252 if (reg_mentioned_p (old_reg, PATTERN (p)))
4258 /* If we can't get the LUIDs for the insns, we can't
4259 calculate the lifetime. This is likely from unrolling
4260 of an inner loop, so there is little point in making this
4261 a DEST_REG giv anyways. */
4262 if (INSN_UID (v->insn) >= max_uid_for_loop
4263 || INSN_UID (last_use_insn) >= max_uid_for_loop
4264 || ! validate_change (v->insn, &SET_DEST (set), dest_reg, 0))
4266 /* Change the increment at NEXT back to what it was. */
4267 if (! validate_change (next_loc_insn, next->location,
4273 next->add_val = add_val;
4274 v->dest_reg = dest_reg;
4275 v->giv_type = DEST_REG;
4276 v->location = &SET_SRC (set);
4278 v->combined_with = 0;
4280 v->derive_adjustment = 0;
4286 v->auto_inc_opt = 0;
4289 v->derived_from = 0;
4290 v->always_computable = 1;
4291 v->always_executed = 1;
4293 v->no_const_addval = 0;
4295 old_regno = REGNO (old_reg);
4296 new_regno = REGNO (dest_reg);
4297 VARRAY_INT (set_in_loop, old_regno)--;
4298 VARRAY_INT (set_in_loop, new_regno) = 1;
4299 VARRAY_INT (n_times_set, old_regno)--;
4300 VARRAY_INT (n_times_set, new_regno) = 1;
4301 VARRAY_CHAR (may_not_optimize, new_regno) = 0;
4303 REG_IV_TYPE (new_regno) = GENERAL_INDUCT;
4304 REG_IV_INFO (new_regno) = v;
4306 /* If next_insn has a REG_EQUAL note that mentiones OLD_REG,
4307 it must be replaced. */
4308 note = find_reg_note (next->insn, REG_EQUAL, NULL_RTX);
4309 if (note && reg_mentioned_p (old_reg, XEXP (note, 0)))
4310 XEXP (note, 0) = copy_rtx (SET_SRC (single_set (next->insn)));
4312 /* Remove the increment from the list of biv increments,
4313 and record it as a giv. */
4316 v->next_iv = bl->giv;
4319 v->benefit = rtx_cost (SET_SRC (set), SET);
4320 bl->total_benefit += v->benefit;
4322 /* Now replace the biv with DEST_REG in all insns between
4323 the replaced increment and the next increment, and
4324 remember the last insn that needed a replacement. */
4325 for (last_use_insn = v->insn, p = NEXT_INSN (v->insn);
4327 p = next_insn_in_loop (loop, p))
4333 if (reg_mentioned_p (old_reg, PATTERN (p)))
4336 if (! validate_replace_rtx (old_reg, dest_reg, p))
4339 for (note = REG_NOTES (p); note; note = XEXP (note, 1))
4341 if (GET_CODE (note) == EXPR_LIST)
4343 = replace_rtx (XEXP (note, 0), old_reg, dest_reg);
4347 v->last_use = last_use_insn;
4348 v->lifetime = INSN_LUID (last_use_insn) - INSN_LUID (v->insn);
4349 /* If the lifetime is zero, it means that this register is really
4350 a dead store. So mark this as a giv that can be ignored.
4351 This will not prevent the biv from being eliminated. */
4352 if (v->lifetime == 0)
4355 if (loop_dump_stream)
4356 fprintf (loop_dump_stream,
4357 "Increment %d of biv %d converted to giv %d.\n\n",
4358 INSN_UID (v->insn), old_regno, new_regno);
4362 last_increment_giv = max_reg_num () - 1;
4364 /* Search the loop for general induction variables. */
4366 for_each_insn_in_loop (loop, check_insn_for_givs);
4368 /* Try to calculate and save the number of loop iterations. This is
4369 set to zero if the actual number can not be calculated. This must
4370 be called after all giv's have been identified, since otherwise it may
4371 fail if the iteration variable is a giv. */
4373 loop_iterations (loop);
4375 /* Now for each giv for which we still don't know whether or not it is
4376 replaceable, check to see if it is replaceable because its final value
4377 can be calculated. This must be done after loop_iterations is called,
4378 so that final_giv_value will work correctly. */
4380 for (bl = loop_iv_list; bl; bl = bl->next)
4382 struct induction *v;
4384 for (v = bl->giv; v; v = v->next_iv)
4385 if (! v->replaceable && ! v->not_replaceable)
4386 check_final_value (loop, v);
4389 /* Try to prove that the loop counter variable (if any) is always
4390 nonnegative; if so, record that fact with a REG_NONNEG note
4391 so that "decrement and branch until zero" insn can be used. */
4392 check_dbra_loop (loop, insn_count);
4394 /* Create reg_map to hold substitutions for replaceable giv regs.
4395 Some givs might have been made from biv increments, so look at
4396 reg_iv_type for a suitable size. */
4397 reg_map_size = reg_iv_type->num_elements;
4398 reg_map = (rtx *) xcalloc (reg_map_size, sizeof (rtx));
4400 /* Examine each iv class for feasibility of strength reduction/induction
4401 variable elimination. */
4403 for (bl = loop_iv_list; bl; bl = bl->next)
4405 struct induction *v;
4408 rtx final_value = 0;
4411 /* Test whether it will be possible to eliminate this biv
4412 provided all givs are reduced. This is possible if either
4413 the reg is not used outside the loop, or we can compute
4414 what its final value will be.
4416 For architectures with a decrement_and_branch_until_zero insn,
4417 don't do this if we put a REG_NONNEG note on the endtest for
4420 /* Compare against bl->init_insn rather than loop_start.
4421 We aren't concerned with any uses of the biv between
4422 init_insn and loop_start since these won't be affected
4423 by the value of the biv elsewhere in the function, so
4424 long as init_insn doesn't use the biv itself.
4425 March 14, 1989 -- self@bayes.arc.nasa.gov */
4427 if ((uid_luid[REGNO_LAST_UID (bl->regno)] < INSN_LUID (loop_end)
4429 && INSN_UID (bl->init_insn) < max_uid_for_loop
4430 && uid_luid[REGNO_FIRST_UID (bl->regno)] >= INSN_LUID (bl->init_insn)
4431 #ifdef HAVE_decrement_and_branch_until_zero
4434 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
4435 || ((final_value = final_biv_value (loop, bl))
4436 #ifdef HAVE_decrement_and_branch_until_zero
4440 bl->eliminable = maybe_eliminate_biv (loop, bl, 0, threshold,
4444 if (loop_dump_stream)
4446 fprintf (loop_dump_stream,
4447 "Cannot eliminate biv %d.\n",
4449 fprintf (loop_dump_stream,
4450 "First use: insn %d, last use: insn %d.\n",
4451 REGNO_FIRST_UID (bl->regno),
4452 REGNO_LAST_UID (bl->regno));
4456 /* Combine all giv's for this iv_class. */
4459 /* This will be true at the end, if all givs which depend on this
4460 biv have been strength reduced.
4461 We can't (currently) eliminate the biv unless this is so. */
4464 /* Check each giv in this class to see if we will benefit by reducing
4465 it. Skip giv's combined with others. */
4466 for (v = bl->giv; v; v = v->next_iv)
4468 struct induction *tv;
4470 if (v->ignore || v->same)
4473 benefit = v->benefit;
4475 /* Reduce benefit if not replaceable, since we will insert
4476 a move-insn to replace the insn that calculates this giv.
4477 Don't do this unless the giv is a user variable, since it
4478 will often be marked non-replaceable because of the duplication
4479 of the exit code outside the loop. In such a case, the copies
4480 we insert are dead and will be deleted. So they don't have
4481 a cost. Similar situations exist. */
4482 /* ??? The new final_[bg]iv_value code does a much better job
4483 of finding replaceable giv's, and hence this code may no longer
4485 if (! v->replaceable && ! bl->eliminable
4486 && REG_USERVAR_P (v->dest_reg))
4487 benefit -= copy_cost;
4489 /* Decrease the benefit to count the add-insns that we will
4490 insert to increment the reduced reg for the giv. */
4491 benefit -= add_cost * bl->biv_count;
4493 /* Decide whether to strength-reduce this giv or to leave the code
4494 unchanged (recompute it from the biv each time it is used).
4495 This decision can be made independently for each giv. */
4498 /* Attempt to guess whether autoincrement will handle some of the
4499 new add insns; if so, increase BENEFIT (undo the subtraction of
4500 add_cost that was done above). */
4501 if (v->giv_type == DEST_ADDR
4502 && GET_CODE (v->mult_val) == CONST_INT)
4504 if (HAVE_POST_INCREMENT
4505 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4506 benefit += add_cost * bl->biv_count;
4507 else if (HAVE_PRE_INCREMENT
4508 && INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4509 benefit += add_cost * bl->biv_count;
4510 else if (HAVE_POST_DECREMENT
4511 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4512 benefit += add_cost * bl->biv_count;
4513 else if (HAVE_PRE_DECREMENT
4514 && -INTVAL (v->mult_val) == GET_MODE_SIZE (v->mem_mode))
4515 benefit += add_cost * bl->biv_count;
4519 /* If an insn is not to be strength reduced, then set its ignore
4520 flag, and clear all_reduced. */
4522 /* A giv that depends on a reversed biv must be reduced if it is
4523 used after the loop exit, otherwise, it would have the wrong
4524 value after the loop exit. To make it simple, just reduce all
4525 of such giv's whether or not we know they are used after the loop
4528 if ( ! flag_reduce_all_givs && v->lifetime * threshold * benefit < insn_count
4531 if (loop_dump_stream)
4532 fprintf (loop_dump_stream,
4533 "giv of insn %d not worth while, %d vs %d.\n",
4535 v->lifetime * threshold * benefit, insn_count);
4541 /* Check that we can increment the reduced giv without a
4542 multiply insn. If not, reject it. */
4544 for (tv = bl->biv; tv; tv = tv->next_iv)
4545 if (tv->mult_val == const1_rtx
4546 && ! product_cheap_p (tv->add_val, v->mult_val))
4548 if (loop_dump_stream)
4549 fprintf (loop_dump_stream,
4550 "giv of insn %d: would need a multiply.\n",
4551 INSN_UID (v->insn));
4559 /* Check for givs whose first use is their definition and whose
4560 last use is the definition of another giv. If so, it is likely
4561 dead and should not be used to derive another giv nor to
4563 for (v = bl->giv; v; v = v->next_iv)
4566 || (v->same && v->same->ignore))
4571 struct induction *v1;
4573 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4574 if (v->last_use == v1->insn)
4577 else if (v->giv_type == DEST_REG
4578 && REGNO_FIRST_UID (REGNO (v->dest_reg)) == INSN_UID (v->insn))
4580 struct induction *v1;
4582 for (v1 = bl->giv; v1; v1 = v1->next_iv)
4583 if (REGNO_LAST_UID (REGNO (v->dest_reg)) == INSN_UID (v1->insn))
4588 /* Now that we know which givs will be reduced, try to rearrange the
4589 combinations to reduce register pressure.
4590 recombine_givs calls find_life_end, which needs reg_iv_type and
4591 reg_iv_info to be valid for all pseudos. We do the necessary
4592 reallocation here since it allows to check if there are still
4593 more bivs to process. */
4594 nregs = max_reg_num ();
4595 if (nregs > reg_iv_type->num_elements)
4597 /* If there are still more bivs to process, allocate some slack
4598 space so that we're not constantly reallocating these arrays. */
4601 /* Reallocate reg_iv_type and reg_iv_info. */
4602 VARRAY_GROW (reg_iv_type, nregs);
4603 VARRAY_GROW (reg_iv_info, nregs);
4605 recombine_givs (loop, bl, flags & LOOP_UNROLL);
4607 /* Reduce each giv that we decided to reduce. */
4609 for (v = bl->giv; v; v = v->next_iv)
4611 struct induction *tv;
4612 if (! v->ignore && v->same == 0)
4614 int auto_inc_opt = 0;
4616 /* If the code for derived givs immediately below has already
4617 allocated a new_reg, we must keep it. */
4619 v->new_reg = gen_reg_rtx (v->mode);
4621 if (v->derived_from)
4623 struct induction *d = v->derived_from;
4625 /* In case d->dest_reg is not replaceable, we have
4626 to replace it in v->insn now. */
4628 d->new_reg = gen_reg_rtx (d->mode);
4630 = replace_rtx (PATTERN (v->insn), d->dest_reg, d->new_reg);
4632 = replace_rtx (PATTERN (v->insn), v->dest_reg, v->new_reg);
4633 /* For each place where the biv is incremented, add an
4634 insn to set the new, reduced reg for the giv.
4635 We used to do this only for biv_count != 1, but
4636 this fails when there is a giv after a single biv
4637 increment, e.g. when the last giv was expressed as
4639 for (tv = bl->biv; tv; tv = tv->next_iv)
4641 /* We always emit reduced giv increments before the
4642 biv increment when bl->biv_count != 1. So by
4643 emitting the add insns for derived givs after the
4644 biv increment, they pick up the updated value of
4646 If the reduced giv is processed with
4647 auto_inc_opt == 1, then it is incremented earlier
4648 than the biv, hence we'll still pick up the right
4650 If it's processed with auto_inc_opt == -1,
4651 that implies that the biv increment is before the
4652 first reduced giv's use. The derived giv's lifetime
4653 is after the reduced giv's lifetime, hence in this
4654 case, the biv increment doesn't matter. */
4655 emit_insn_after (copy_rtx (PATTERN (v->insn)), tv->insn);
4661 /* If the target has auto-increment addressing modes, and
4662 this is an address giv, then try to put the increment
4663 immediately after its use, so that flow can create an
4664 auto-increment addressing mode. */
4665 if (v->giv_type == DEST_ADDR && bl->biv_count == 1
4666 && bl->biv->always_executed && ! bl->biv->maybe_multiple
4667 /* We don't handle reversed biv's because bl->biv->insn
4668 does not have a valid INSN_LUID. */
4670 && v->always_executed && ! v->maybe_multiple
4671 && INSN_UID (v->insn) < max_uid_for_loop)
4673 /* If other giv's have been combined with this one, then
4674 this will work only if all uses of the other giv's occur
4675 before this giv's insn. This is difficult to check.
4677 We simplify this by looking for the common case where
4678 there is one DEST_REG giv, and this giv's insn is the
4679 last use of the dest_reg of that DEST_REG giv. If the
4680 increment occurs after the address giv, then we can
4681 perform the optimization. (Otherwise, the increment
4682 would have to go before other_giv, and we would not be
4683 able to combine it with the address giv to get an
4684 auto-inc address.) */
4685 if (v->combined_with)
4687 struct induction *other_giv = 0;
4689 for (tv = bl->giv; tv; tv = tv->next_iv)
4697 if (! tv && other_giv
4698 && REGNO (other_giv->dest_reg) < max_reg_before_loop
4699 && (REGNO_LAST_UID (REGNO (other_giv->dest_reg))
4700 == INSN_UID (v->insn))
4701 && INSN_LUID (v->insn) < INSN_LUID (bl->biv->insn))
4704 /* Check for case where increment is before the address
4705 giv. Do this test in "loop order". */
4706 else if ((INSN_LUID (v->insn) > INSN_LUID (bl->biv->insn)
4707 && (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
4708 || (INSN_LUID (bl->biv->insn)
4709 > INSN_LUID (loop_scan_start))))
4710 || (INSN_LUID (v->insn) < INSN_LUID (loop_scan_start)
4711 && (INSN_LUID (loop_scan_start)
4712 < INSN_LUID (bl->biv->insn))))
4721 /* We can't put an insn immediately after one setting
4722 cc0, or immediately before one using cc0. */
4723 if ((auto_inc_opt == 1 && sets_cc0_p (PATTERN (v->insn)))
4724 || (auto_inc_opt == -1
4725 && (prev = prev_nonnote_insn (v->insn)) != 0
4727 && sets_cc0_p (PATTERN (prev))))
4733 v->auto_inc_opt = 1;
4737 /* For each place where the biv is incremented, add an insn
4738 to increment the new, reduced reg for the giv. */
4739 for (tv = bl->biv; tv; tv = tv->next_iv)
4744 insert_before = tv->insn;
4745 else if (auto_inc_opt == 1)
4746 insert_before = NEXT_INSN (v->insn);
4748 insert_before = v->insn;
4750 if (tv->mult_val == const1_rtx)
4751 emit_iv_add_mult (tv->add_val, v->mult_val,
4752 v->new_reg, v->new_reg, insert_before);
4753 else /* tv->mult_val == const0_rtx */
4754 /* A multiply is acceptable here
4755 since this is presumed to be seldom executed. */
4756 emit_iv_add_mult (tv->add_val, v->mult_val,
4757 v->add_val, v->new_reg, insert_before);
4760 /* Add code at loop start to initialize giv's reduced reg. */
4762 emit_iv_add_mult (bl->initial_value, v->mult_val,
4763 v->add_val, v->new_reg, loop_start);
4767 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
4770 For each giv register that can be reduced now: if replaceable,
4771 substitute reduced reg wherever the old giv occurs;
4772 else add new move insn "giv_reg = reduced_reg". */
4774 for (v = bl->giv; v; v = v->next_iv)
4776 if (v->same && v->same->ignore)
4782 /* Update expression if this was combined, in case other giv was
4785 v->new_reg = replace_rtx (v->new_reg,
4786 v->same->dest_reg, v->same->new_reg);
4788 if (v->giv_type == DEST_ADDR)
4789 /* Store reduced reg as the address in the memref where we found
4791 validate_change (v->insn, v->location, v->new_reg, 0);
4792 else if (v->replaceable)
4794 reg_map[REGNO (v->dest_reg)] = v->new_reg;
4797 /* I can no longer duplicate the original problem. Perhaps
4798 this is unnecessary now? */
4800 /* Replaceable; it isn't strictly necessary to delete the old
4801 insn and emit a new one, because v->dest_reg is now dead.
4803 However, especially when unrolling loops, the special
4804 handling for (set REG0 REG1) in the second cse pass may
4805 make v->dest_reg live again. To avoid this problem, emit
4806 an insn to set the original giv reg from the reduced giv.
4807 We can not delete the original insn, since it may be part
4808 of a LIBCALL, and the code in flow that eliminates dead
4809 libcalls will fail if it is deleted. */
4810 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4816 /* Not replaceable; emit an insn to set the original giv reg from
4817 the reduced giv, same as above. */
4818 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
4822 /* When a loop is reversed, givs which depend on the reversed
4823 biv, and which are live outside the loop, must be set to their
4824 correct final value. This insn is only needed if the giv is
4825 not replaceable. The correct final value is the same as the
4826 value that the giv starts the reversed loop with. */
4827 if (bl->reversed && ! v->replaceable)
4828 emit_iv_add_mult (bl->initial_value, v->mult_val,
4829 v->add_val, v->dest_reg, end_insert_before);
4830 else if (v->final_value)
4834 /* If the loop has multiple exits, emit the insn before the
4835 loop to ensure that it will always be executed no matter
4836 how the loop exits. Otherwise, emit the insn after the loop,
4837 since this is slightly more efficient. */
4838 if (loop->exit_count)
4839 insert_before = loop_start;
4841 insert_before = end_insert_before;
4842 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
4846 /* If the insn to set the final value of the giv was emitted
4847 before the loop, then we must delete the insn inside the loop
4848 that sets it. If this is a LIBCALL, then we must delete
4849 every insn in the libcall. Note, however, that
4850 final_giv_value will only succeed when there are multiple
4851 exits if the giv is dead at each exit, hence it does not
4852 matter that the original insn remains because it is dead
4854 /* Delete the insn inside the loop that sets the giv since
4855 the giv is now set before (or after) the loop. */
4856 delete_insn (v->insn);
4860 if (loop_dump_stream)
4862 fprintf (loop_dump_stream, "giv at %d reduced to ",
4863 INSN_UID (v->insn));
4864 print_rtl (loop_dump_stream, v->new_reg);
4865 fprintf (loop_dump_stream, "\n");
4869 /* All the givs based on the biv bl have been reduced if they
4872 /* For each giv not marked as maybe dead that has been combined with a
4873 second giv, clear any "maybe dead" mark on that second giv.
4874 v->new_reg will either be or refer to the register of the giv it
4877 Doing this clearing avoids problems in biv elimination where a
4878 giv's new_reg is a complex value that can't be put in the insn but
4879 the giv combined with (with a reg as new_reg) is marked maybe_dead.
4880 Since the register will be used in either case, we'd prefer it be
4881 used from the simpler giv. */
4883 for (v = bl->giv; v; v = v->next_iv)
4884 if (! v->maybe_dead && v->same)
4885 v->same->maybe_dead = 0;
4887 /* Try to eliminate the biv, if it is a candidate.
4888 This won't work if ! all_reduced,
4889 since the givs we planned to use might not have been reduced.
4891 We have to be careful that we didn't initially think we could eliminate
4892 this biv because of a giv that we now think may be dead and shouldn't
4893 be used as a biv replacement.
4895 Also, there is the possibility that we may have a giv that looks
4896 like it can be used to eliminate a biv, but the resulting insn
4897 isn't valid. This can happen, for example, on the 88k, where a
4898 JUMP_INSN can compare a register only with zero. Attempts to
4899 replace it with a compare with a constant will fail.
4901 Note that in cases where this call fails, we may have replaced some
4902 of the occurrences of the biv with a giv, but no harm was done in
4903 doing so in the rare cases where it can occur. */
4905 if (all_reduced == 1 && bl->eliminable
4906 && maybe_eliminate_biv (loop, bl, 1, threshold, insn_count))
4908 /* ?? If we created a new test to bypass the loop entirely,
4909 or otherwise drop straight in, based on this test, then
4910 we might want to rewrite it also. This way some later
4911 pass has more hope of removing the initialization of this
4914 /* If final_value != 0, then the biv may be used after loop end
4915 and we must emit an insn to set it just in case.
4917 Reversed bivs already have an insn after the loop setting their
4918 value, so we don't need another one. We can't calculate the
4919 proper final value for such a biv here anyways. */
4920 if (final_value != 0 && ! bl->reversed)
4924 /* If the loop has multiple exits, emit the insn before the
4925 loop to ensure that it will always be executed no matter
4926 how the loop exits. Otherwise, emit the insn after the
4927 loop, since this is slightly more efficient. */
4928 if (loop->exit_count)
4929 insert_before = loop_start;
4931 insert_before = end_insert_before;
4933 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
4938 /* Delete all of the instructions inside the loop which set
4939 the biv, as they are all dead. If is safe to delete them,
4940 because an insn setting a biv will never be part of a libcall. */
4941 /* However, deleting them will invalidate the regno_last_uid info,
4942 so keeping them around is more convenient. Final_biv_value
4943 will only succeed when there are multiple exits if the biv
4944 is dead at each exit, hence it does not matter that the original
4945 insn remains, because it is dead anyways. */
4946 for (v = bl->biv; v; v = v->next_iv)
4947 delete_insn (v->insn);
4950 if (loop_dump_stream)
4951 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
4956 /* Go through all the instructions in the loop, making all the
4957 register substitutions scheduled in REG_MAP. */
4959 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
4960 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4961 || GET_CODE (p) == CALL_INSN)
4963 replace_regs (PATTERN (p), reg_map, reg_map_size, 0);
4964 replace_regs (REG_NOTES (p), reg_map, reg_map_size, 0);
4968 if (loop_info->n_iterations > 0)
4970 /* When we completely unroll a loop we will likely not need the increment
4971 of the loop BIV and we will not need the conditional branch at the
4973 unrolled_insn_copies = insn_count - 2;
4976 /* When we completely unroll a loop on a HAVE_cc0 machine we will not
4977 need the comparison before the conditional branch at the end of the
4979 unrolled_insn_copies -= 1;
4982 /* We'll need one copy for each loop iteration. */
4983 unrolled_insn_copies *= loop_info->n_iterations;
4985 /* A little slop to account for the ability to remove initialization
4986 code, better CSE, and other secondary benefits of completely
4987 unrolling some loops. */
4988 unrolled_insn_copies -= 1;
4990 /* Clamp the value. */
4991 if (unrolled_insn_copies < 0)
4992 unrolled_insn_copies = 0;
4995 /* Unroll loops from within strength reduction so that we can use the
4996 induction variable information that strength_reduce has already
4997 collected. Always unroll loops that would be as small or smaller
4998 unrolled than when rolled. */
4999 if ((flags & LOOP_UNROLL)
5000 || (loop_info->n_iterations > 0
5001 && unrolled_insn_copies <= insn_count))
5002 unroll_loop (loop, insn_count, end_insert_before, 1);
5004 #ifdef HAVE_doloop_end
5005 if (HAVE_doloop_end && (flags & LOOP_BCT) && flag_branch_on_count_reg)
5006 doloop_optimize (loop);
5007 #endif /* HAVE_doloop_end */
5009 if (loop_dump_stream)
5010 fprintf (loop_dump_stream, "\n");
5013 VARRAY_FREE (reg_iv_type);
5014 VARRAY_FREE (reg_iv_info);
5015 free (reg_biv_class);
5020 /*Record all basic induction variables calculated in the insn. */
5022 check_insn_for_bivs (loop, p, not_every_iteration, maybe_multiple)
5025 int not_every_iteration;
5034 if (GET_CODE (p) == INSN
5035 && (set = single_set (p))
5036 && GET_CODE (SET_DEST (set)) == REG)
5038 dest_reg = SET_DEST (set);
5039 if (REGNO (dest_reg) < max_reg_before_loop
5040 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
5041 && REG_IV_TYPE (REGNO (dest_reg)) != NOT_BASIC_INDUCT)
5043 if (basic_induction_var (loop, SET_SRC (set),
5044 GET_MODE (SET_SRC (set)),
5045 dest_reg, p, &inc_val, &mult_val,
5048 /* It is a possible basic induction variable.
5049 Create and initialize an induction structure for it. */
5052 = (struct induction *) oballoc (sizeof (struct induction));
5054 record_biv (v, p, dest_reg, inc_val, mult_val, location,
5055 not_every_iteration, maybe_multiple);
5056 REG_IV_TYPE (REGNO (dest_reg)) = BASIC_INDUCT;
5058 else if (REGNO (dest_reg) < max_reg_before_loop)
5059 REG_IV_TYPE (REGNO (dest_reg)) = NOT_BASIC_INDUCT;
5065 /* Record all givs calculated in the insn.
5066 A register is a giv if: it is only set once, it is a function of a
5067 biv and a constant (or invariant), and it is not a biv. */
5069 check_insn_for_givs (loop, p, not_every_iteration, maybe_multiple)
5072 int not_every_iteration;
5076 /* Look for a general induction variable in a register. */
5077 if (GET_CODE (p) == INSN
5078 && (set = single_set (p))
5079 && GET_CODE (SET_DEST (set)) == REG
5080 && ! VARRAY_CHAR (may_not_optimize, REGNO (SET_DEST (set))))
5088 rtx last_consec_insn;
5090 dest_reg = SET_DEST (set);
5091 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
5094 if (/* SET_SRC is a giv. */
5095 (general_induction_var (loop, SET_SRC (set), &src_reg, &add_val,
5096 &mult_val, 0, &benefit, VOIDmode)
5097 /* Equivalent expression is a giv. */
5098 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
5099 && general_induction_var (loop, XEXP (regnote, 0), &src_reg,
5100 &add_val, &mult_val, 0,
5101 &benefit, VOIDmode)))
5102 /* Don't try to handle any regs made by loop optimization.
5103 We have nothing on them in regno_first_uid, etc. */
5104 && REGNO (dest_reg) < max_reg_before_loop
5105 /* Don't recognize a BASIC_INDUCT_VAR here. */
5106 && dest_reg != src_reg
5107 /* This must be the only place where the register is set. */
5108 && (VARRAY_INT (n_times_set, REGNO (dest_reg)) == 1
5109 /* or all sets must be consecutive and make a giv. */
5110 || (benefit = consec_sets_giv (loop, benefit, p,
5112 &add_val, &mult_val,
5113 &last_consec_insn))))
5116 = (struct induction *) oballoc (sizeof (struct induction));
5118 /* If this is a library call, increase benefit. */
5119 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
5120 benefit += libcall_benefit (p);
5122 /* Skip the consecutive insns, if there are any. */
5123 if (VARRAY_INT (n_times_set, REGNO (dest_reg)) != 1)
5124 p = last_consec_insn;
5126 record_giv (loop, v, p, src_reg, dest_reg, mult_val, add_val,
5127 benefit, DEST_REG, not_every_iteration,
5128 maybe_multiple, NULL_PTR);
5133 #ifndef DONT_REDUCE_ADDR
5134 /* Look for givs which are memory addresses. */
5135 /* This resulted in worse code on a VAX 8600. I wonder if it
5137 if (GET_CODE (p) == INSN)
5138 find_mem_givs (loop, PATTERN (p), p, not_every_iteration,
5142 /* Update the status of whether giv can derive other givs. This can
5143 change when we pass a label or an insn that updates a biv. */
5144 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5145 || GET_CODE (p) == CODE_LABEL)
5146 update_giv_derive (loop, p);
5150 /* Return 1 if X is a valid source for an initial value (or as value being
5151 compared against in an initial test).
5153 X must be either a register or constant and must not be clobbered between
5154 the current insn and the start of the loop.
5156 INSN is the insn containing X. */
5159 valid_initial_value_p (x, insn, call_seen, loop_start)
5168 /* Only consider pseudos we know about initialized in insns whose luids
5170 if (GET_CODE (x) != REG
5171 || REGNO (x) >= max_reg_before_loop)
5174 /* Don't use call-clobbered registers across a call which clobbers it. On
5175 some machines, don't use any hard registers at all. */
5176 if (REGNO (x) < FIRST_PSEUDO_REGISTER
5177 && (SMALL_REGISTER_CLASSES
5178 || (call_used_regs[REGNO (x)] && call_seen)))
5181 /* Don't use registers that have been clobbered before the start of the
5183 if (reg_set_between_p (x, insn, loop_start))
5189 /* Scan X for memory refs and check each memory address
5190 as a possible giv. INSN is the insn whose pattern X comes from.
5191 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
5192 every loop iteration. MAYBE_MULTIPLE is 1 if the insn might be executed
5193 more thanonce in each loop iteration. */
5196 find_mem_givs (loop, x, insn, not_every_iteration, maybe_multiple)
5197 const struct loop *loop;
5200 int not_every_iteration, maybe_multiple;
5203 register enum rtx_code code;
5204 register const char *fmt;
5209 code = GET_CODE (x);
5233 /* This code used to disable creating GIVs with mult_val == 1 and
5234 add_val == 0. However, this leads to lost optimizations when
5235 it comes time to combine a set of related DEST_ADDR GIVs, since
5236 this one would not be seen. */
5238 if (general_induction_var (loop, XEXP (x, 0), &src_reg, &add_val,
5239 &mult_val, 1, &benefit, GET_MODE (x)))
5241 /* Found one; record it. */
5243 = (struct induction *) oballoc (sizeof (struct induction));
5245 record_giv (loop, v, insn, src_reg, addr_placeholder, mult_val,
5246 add_val, benefit, DEST_ADDR, not_every_iteration,
5247 maybe_multiple, &XEXP (x, 0));
5249 v->mem_mode = GET_MODE (x);
5258 /* Recursively scan the subexpressions for other mem refs. */
5260 fmt = GET_RTX_FORMAT (code);
5261 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5263 find_mem_givs (loop, XEXP (x, i), insn, not_every_iteration,
5265 else if (fmt[i] == 'E')
5266 for (j = 0; j < XVECLEN (x, i); j++)
5267 find_mem_givs (loop, XVECEXP (x, i, j), insn, not_every_iteration,
5271 /* Fill in the data about one biv update.
5272 V is the `struct induction' in which we record the biv. (It is
5273 allocated by the caller, with alloca.)
5274 INSN is the insn that sets it.
5275 DEST_REG is the biv's reg.
5277 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
5278 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
5279 being set to INC_VAL.
5281 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
5282 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
5283 can be executed more than once per iteration. If MAYBE_MULTIPLE
5284 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
5285 executed exactly once per iteration. */
5288 record_biv (v, insn, dest_reg, inc_val, mult_val, location,
5289 not_every_iteration, maybe_multiple)
5290 struct induction *v;
5296 int not_every_iteration;
5299 struct iv_class *bl;
5302 v->src_reg = dest_reg;
5303 v->dest_reg = dest_reg;
5304 v->mult_val = mult_val;
5305 v->add_val = inc_val;
5306 v->location = location;
5307 v->mode = GET_MODE (dest_reg);
5308 v->always_computable = ! not_every_iteration;
5309 v->always_executed = ! not_every_iteration;
5310 v->maybe_multiple = maybe_multiple;
5312 /* Add this to the reg's iv_class, creating a class
5313 if this is the first incrementation of the reg. */
5315 bl = reg_biv_class[REGNO (dest_reg)];
5318 /* Create and initialize new iv_class. */
5320 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
5322 bl->regno = REGNO (dest_reg);
5328 /* Set initial value to the reg itself. */
5329 bl->initial_value = dest_reg;
5330 /* We haven't seen the initializing insn yet */
5333 bl->initial_test = 0;
5334 bl->incremented = 0;
5338 bl->total_benefit = 0;
5340 /* Add this class to loop_iv_list. */
5341 bl->next = loop_iv_list;
5344 /* Put it in the array of biv register classes. */
5345 reg_biv_class[REGNO (dest_reg)] = bl;
5348 /* Update IV_CLASS entry for this biv. */
5349 v->next_iv = bl->biv;
5352 if (mult_val == const1_rtx)
5353 bl->incremented = 1;
5355 if (loop_dump_stream)
5357 fprintf (loop_dump_stream,
5358 "Insn %d: possible biv, reg %d,",
5359 INSN_UID (insn), REGNO (dest_reg));
5360 if (GET_CODE (inc_val) == CONST_INT)
5362 fprintf (loop_dump_stream, " const =");
5363 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (inc_val));
5364 fputc ('\n', loop_dump_stream);
5368 fprintf (loop_dump_stream, " const = ");
5369 print_rtl (loop_dump_stream, inc_val);
5370 fprintf (loop_dump_stream, "\n");
5375 /* Fill in the data about one giv.
5376 V is the `struct induction' in which we record the giv. (It is
5377 allocated by the caller, with alloca.)
5378 INSN is the insn that sets it.
5379 BENEFIT estimates the savings from deleting this insn.
5380 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
5381 into a register or is used as a memory address.
5383 SRC_REG is the biv reg which the giv is computed from.
5384 DEST_REG is the giv's reg (if the giv is stored in a reg).
5385 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
5386 LOCATION points to the place where this giv's value appears in INSN. */
5389 record_giv (loop, v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
5390 type, not_every_iteration, maybe_multiple, location)
5391 const struct loop *loop;
5392 struct induction *v;
5396 rtx mult_val, add_val;
5399 int not_every_iteration, maybe_multiple;
5402 struct induction *b;
5403 struct iv_class *bl;
5404 rtx set = single_set (insn);
5407 /* Attempt to prove constantness of the values. */
5408 temp = simplify_rtx (add_val);
5413 v->src_reg = src_reg;
5415 v->dest_reg = dest_reg;
5416 v->mult_val = mult_val;
5417 v->add_val = add_val;
5418 v->benefit = benefit;
5419 v->location = location;
5421 v->combined_with = 0;
5422 v->maybe_multiple = maybe_multiple;
5424 v->derive_adjustment = 0;
5430 v->auto_inc_opt = 0;
5433 v->derived_from = 0;
5436 /* The v->always_computable field is used in update_giv_derive, to
5437 determine whether a giv can be used to derive another giv. For a
5438 DEST_REG giv, INSN computes a new value for the giv, so its value
5439 isn't computable if INSN insn't executed every iteration.
5440 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
5441 it does not compute a new value. Hence the value is always computable
5442 regardless of whether INSN is executed each iteration. */
5444 if (type == DEST_ADDR)
5445 v->always_computable = 1;
5447 v->always_computable = ! not_every_iteration;
5449 v->always_executed = ! not_every_iteration;
5451 if (type == DEST_ADDR)
5453 v->mode = GET_MODE (*location);
5456 else /* type == DEST_REG */
5458 v->mode = GET_MODE (SET_DEST (set));
5460 v->lifetime = (uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5461 - uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))]);
5463 /* If the lifetime is zero, it means that this register is
5464 really a dead store. So mark this as a giv that can be
5465 ignored. This will not prevent the biv from being eliminated. */
5466 if (v->lifetime == 0)
5469 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
5470 REG_IV_INFO (REGNO (dest_reg)) = v;
5473 /* Add the giv to the class of givs computed from one biv. */
5475 bl = reg_biv_class[REGNO (src_reg)];
5478 v->next_iv = bl->giv;
5480 /* Don't count DEST_ADDR. This is supposed to count the number of
5481 insns that calculate givs. */
5482 if (type == DEST_REG)
5484 bl->total_benefit += benefit;
5487 /* Fatal error, biv missing for this giv? */
5490 if (type == DEST_ADDR)
5494 /* The giv can be replaced outright by the reduced register only if all
5495 of the following conditions are true:
5496 - the insn that sets the giv is always executed on any iteration
5497 on which the giv is used at all
5498 (there are two ways to deduce this:
5499 either the insn is executed on every iteration,
5500 or all uses follow that insn in the same basic block),
5501 - the giv is not used outside the loop
5502 - no assignments to the biv occur during the giv's lifetime. */
5504 if (REGNO_FIRST_UID (REGNO (dest_reg)) == INSN_UID (insn)
5505 /* Previous line always fails if INSN was moved by loop opt. */
5506 && uid_luid[REGNO_LAST_UID (REGNO (dest_reg))]
5507 < INSN_LUID (loop->end)
5508 && (! not_every_iteration
5509 || last_use_this_basic_block (dest_reg, insn)))
5511 /* Now check that there are no assignments to the biv within the
5512 giv's lifetime. This requires two separate checks. */
5514 /* Check each biv update, and fail if any are between the first
5515 and last use of the giv.
5517 If this loop contains an inner loop that was unrolled, then
5518 the insn modifying the biv may have been emitted by the loop
5519 unrolling code, and hence does not have a valid luid. Just
5520 mark the biv as not replaceable in this case. It is not very
5521 useful as a biv, because it is used in two different loops.
5522 It is very unlikely that we would be able to optimize the giv
5523 using this biv anyways. */
5526 for (b = bl->biv; b; b = b->next_iv)
5528 if (INSN_UID (b->insn) >= max_uid_for_loop
5529 || ((uid_luid[INSN_UID (b->insn)]
5530 >= uid_luid[REGNO_FIRST_UID (REGNO (dest_reg))])
5531 && (uid_luid[INSN_UID (b->insn)]
5532 <= uid_luid[REGNO_LAST_UID (REGNO (dest_reg))])))
5535 v->not_replaceable = 1;
5540 /* If there are any backwards branches that go from after the
5541 biv update to before it, then this giv is not replaceable. */
5543 for (b = bl->biv; b; b = b->next_iv)
5544 if (back_branch_in_range_p (loop, b->insn))
5547 v->not_replaceable = 1;
5553 /* May still be replaceable, we don't have enough info here to
5556 v->not_replaceable = 0;
5560 /* Record whether the add_val contains a const_int, for later use by
5565 v->no_const_addval = 1;
5566 if (tem == const0_rtx)
5568 else if (CONSTANT_P (add_val))
5569 v->no_const_addval = 0;
5570 if (GET_CODE (tem) == PLUS)
5574 if (GET_CODE (XEXP (tem, 0)) == PLUS)
5575 tem = XEXP (tem, 0);
5576 else if (GET_CODE (XEXP (tem, 1)) == PLUS)
5577 tem = XEXP (tem, 1);
5581 if (CONSTANT_P (XEXP (tem, 1)))
5582 v->no_const_addval = 0;
5586 if (loop_dump_stream)
5588 if (type == DEST_REG)
5589 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
5590 INSN_UID (insn), REGNO (dest_reg));
5592 fprintf (loop_dump_stream, "Insn %d: dest address",
5595 fprintf (loop_dump_stream, " src reg %d benefit %d",
5596 REGNO (src_reg), v->benefit);
5597 fprintf (loop_dump_stream, " lifetime %d",
5601 fprintf (loop_dump_stream, " replaceable");
5603 if (v->no_const_addval)
5604 fprintf (loop_dump_stream, " ncav");
5606 if (GET_CODE (mult_val) == CONST_INT)
5608 fprintf (loop_dump_stream, " mult ");
5609 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (mult_val));
5613 fprintf (loop_dump_stream, " mult ");
5614 print_rtl (loop_dump_stream, mult_val);
5617 if (GET_CODE (add_val) == CONST_INT)
5619 fprintf (loop_dump_stream, " add ");
5620 fprintf (loop_dump_stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (add_val));
5624 fprintf (loop_dump_stream, " add ");
5625 print_rtl (loop_dump_stream, add_val);
5629 if (loop_dump_stream)
5630 fprintf (loop_dump_stream, "\n");
5634 /* All this does is determine whether a giv can be made replaceable because
5635 its final value can be calculated. This code can not be part of record_giv
5636 above, because final_giv_value requires that the number of loop iterations
5637 be known, and that can not be accurately calculated until after all givs
5638 have been identified. */
5641 check_final_value (loop, v)
5642 const struct loop *loop;
5643 struct induction *v;
5645 struct iv_class *bl;
5646 rtx final_value = 0;
5648 bl = reg_biv_class[REGNO (v->src_reg)];
5650 /* DEST_ADDR givs will never reach here, because they are always marked
5651 replaceable above in record_giv. */
5653 /* The giv can be replaced outright by the reduced register only if all
5654 of the following conditions are true:
5655 - the insn that sets the giv is always executed on any iteration
5656 on which the giv is used at all
5657 (there are two ways to deduce this:
5658 either the insn is executed on every iteration,
5659 or all uses follow that insn in the same basic block),
5660 - its final value can be calculated (this condition is different
5661 than the one above in record_giv)
5662 - no assignments to the biv occur during the giv's lifetime. */
5665 /* This is only called now when replaceable is known to be false. */
5666 /* Clear replaceable, so that it won't confuse final_giv_value. */
5670 if ((final_value = final_giv_value (loop, v))
5671 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
5673 int biv_increment_seen = 0;
5679 /* When trying to determine whether or not a biv increment occurs
5680 during the lifetime of the giv, we can ignore uses of the variable
5681 outside the loop because final_value is true. Hence we can not
5682 use regno_last_uid and regno_first_uid as above in record_giv. */
5684 /* Search the loop to determine whether any assignments to the
5685 biv occur during the giv's lifetime. Start with the insn
5686 that sets the giv, and search around the loop until we come
5687 back to that insn again.
5689 Also fail if there is a jump within the giv's lifetime that jumps
5690 to somewhere outside the lifetime but still within the loop. This
5691 catches spaghetti code where the execution order is not linear, and
5692 hence the above test fails. Here we assume that the giv lifetime
5693 does not extend from one iteration of the loop to the next, so as
5694 to make the test easier. Since the lifetime isn't known yet,
5695 this requires two loops. See also record_giv above. */
5697 last_giv_use = v->insn;
5703 p = NEXT_INSN (loop->start);
5707 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
5708 || GET_CODE (p) == CALL_INSN)
5710 if (biv_increment_seen)
5712 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5715 v->not_replaceable = 1;
5719 else if (reg_set_p (v->src_reg, PATTERN (p)))
5720 biv_increment_seen = 1;
5721 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
5726 /* Now that the lifetime of the giv is known, check for branches
5727 from within the lifetime to outside the lifetime if it is still
5737 p = NEXT_INSN (loop->start);
5738 if (p == last_giv_use)
5741 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
5742 && LABEL_NAME (JUMP_LABEL (p))
5743 && ((loop_insn_first_p (JUMP_LABEL (p), v->insn)
5744 && loop_insn_first_p (loop->start, JUMP_LABEL (p)))
5745 || (loop_insn_first_p (last_giv_use, JUMP_LABEL (p))
5746 && loop_insn_first_p (JUMP_LABEL (p), loop->end))))
5749 v->not_replaceable = 1;
5751 if (loop_dump_stream)
5752 fprintf (loop_dump_stream,
5753 "Found branch outside giv lifetime.\n");
5760 /* If it is replaceable, then save the final value. */
5762 v->final_value = final_value;
5765 if (loop_dump_stream && v->replaceable)
5766 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
5767 INSN_UID (v->insn), REGNO (v->dest_reg));
5770 /* Update the status of whether a giv can derive other givs.
5772 We need to do something special if there is or may be an update to the biv
5773 between the time the giv is defined and the time it is used to derive
5776 In addition, a giv that is only conditionally set is not allowed to
5777 derive another giv once a label has been passed.
5779 The cases we look at are when a label or an update to a biv is passed. */
5782 update_giv_derive (loop, p)
5783 const struct loop *loop;
5786 struct iv_class *bl;
5787 struct induction *biv, *giv;
5791 /* Search all IV classes, then all bivs, and finally all givs.
5793 There are three cases we are concerned with. First we have the situation
5794 of a giv that is only updated conditionally. In that case, it may not
5795 derive any givs after a label is passed.
5797 The second case is when a biv update occurs, or may occur, after the
5798 definition of a giv. For certain biv updates (see below) that are
5799 known to occur between the giv definition and use, we can adjust the
5800 giv definition. For others, or when the biv update is conditional,
5801 we must prevent the giv from deriving any other givs. There are two
5802 sub-cases within this case.
5804 If this is a label, we are concerned with any biv update that is done
5805 conditionally, since it may be done after the giv is defined followed by
5806 a branch here (actually, we need to pass both a jump and a label, but
5807 this extra tracking doesn't seem worth it).
5809 If this is a jump, we are concerned about any biv update that may be
5810 executed multiple times. We are actually only concerned about
5811 backward jumps, but it is probably not worth performing the test
5812 on the jump again here.
5814 If this is a biv update, we must adjust the giv status to show that a
5815 subsequent biv update was performed. If this adjustment cannot be done,
5816 the giv cannot derive further givs. */
5818 for (bl = loop_iv_list; bl; bl = bl->next)
5819 for (biv = bl->biv; biv; biv = biv->next_iv)
5820 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
5823 for (giv = bl->giv; giv; giv = giv->next_iv)
5825 /* If cant_derive is already true, there is no point in
5826 checking all of these conditions again. */
5827 if (giv->cant_derive)
5830 /* If this giv is conditionally set and we have passed a label,
5831 it cannot derive anything. */
5832 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
5833 giv->cant_derive = 1;
5835 /* Skip givs that have mult_val == 0, since
5836 they are really invariants. Also skip those that are
5837 replaceable, since we know their lifetime doesn't contain
5839 else if (giv->mult_val == const0_rtx || giv->replaceable)
5842 /* The only way we can allow this giv to derive another
5843 is if this is a biv increment and we can form the product
5844 of biv->add_val and giv->mult_val. In this case, we will
5845 be able to compute a compensation. */
5846 else if (biv->insn == p)
5850 if (biv->mult_val == const1_rtx)
5851 tem = simplify_giv_expr (loop,
5852 gen_rtx_MULT (giv->mode,
5857 if (tem && giv->derive_adjustment)
5858 tem = simplify_giv_expr
5860 gen_rtx_PLUS (giv->mode, tem, giv->derive_adjustment),
5864 giv->derive_adjustment = tem;
5866 giv->cant_derive = 1;
5868 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
5869 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
5870 giv->cant_derive = 1;
5875 /* Check whether an insn is an increment legitimate for a basic induction var.
5876 X is the source of insn P, or a part of it.
5877 MODE is the mode in which X should be interpreted.
5879 DEST_REG is the putative biv, also the destination of the insn.
5880 We accept patterns of these forms:
5881 REG = REG + INVARIANT (includes REG = REG - CONSTANT)
5882 REG = INVARIANT + REG
5884 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
5885 store the additive term into *INC_VAL, and store the place where
5886 we found the additive term into *LOCATION.
5888 If X is an assignment of an invariant into DEST_REG, we set
5889 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
5891 We also want to detect a BIV when it corresponds to a variable
5892 whose mode was promoted via PROMOTED_MODE. In that case, an increment
5893 of the variable may be a PLUS that adds a SUBREG of that variable to
5894 an invariant and then sign- or zero-extends the result of the PLUS
5897 Most GIVs in such cases will be in the promoted mode, since that is the
5898 probably the natural computation mode (and almost certainly the mode
5899 used for addresses) on the machine. So we view the pseudo-reg containing
5900 the variable as the BIV, as if it were simply incremented.
5902 Note that treating the entire pseudo as a BIV will result in making
5903 simple increments to any GIVs based on it. However, if the variable
5904 overflows in its declared mode but not its promoted mode, the result will
5905 be incorrect. This is acceptable if the variable is signed, since
5906 overflows in such cases are undefined, but not if it is unsigned, since
5907 those overflows are defined. So we only check for SIGN_EXTEND and
5910 If we cannot find a biv, we return 0. */
5913 basic_induction_var (loop, x, mode, dest_reg, p, inc_val, mult_val, location)
5914 const struct loop *loop;
5916 enum machine_mode mode;
5923 register enum rtx_code code;
5927 code = GET_CODE (x);
5932 if (rtx_equal_p (XEXP (x, 0), dest_reg)
5933 || (GET_CODE (XEXP (x, 0)) == SUBREG
5934 && SUBREG_PROMOTED_VAR_P (XEXP (x, 0))
5935 && SUBREG_REG (XEXP (x, 0)) == dest_reg))
5937 argp = &XEXP (x, 1);
5939 else if (rtx_equal_p (XEXP (x, 1), dest_reg)
5940 || (GET_CODE (XEXP (x, 1)) == SUBREG
5941 && SUBREG_PROMOTED_VAR_P (XEXP (x, 1))
5942 && SUBREG_REG (XEXP (x, 1)) == dest_reg))
5944 argp = &XEXP (x, 0);
5950 if (loop_invariant_p (loop, arg) != 1)
5953 *inc_val = convert_modes (GET_MODE (dest_reg), GET_MODE (x), arg, 0);
5954 *mult_val = const1_rtx;
5959 /* If this is a SUBREG for a promoted variable, check the inner
5961 if (SUBREG_PROMOTED_VAR_P (x))
5962 return basic_induction_var (loop, SUBREG_REG (x),
5963 GET_MODE (SUBREG_REG (x)),
5964 dest_reg, p, inc_val, mult_val, location);
5968 /* If this register is assigned in a previous insn, look at its
5969 source, but don't go outside the loop or past a label. */
5971 /* If this sets a register to itself, we would repeat any previous
5972 biv increment if we applied this strategy blindly. */
5973 if (rtx_equal_p (dest_reg, x))
5981 insn = PREV_INSN (insn);
5983 while (insn && GET_CODE (insn) == NOTE
5984 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
5988 set = single_set (insn);
5992 if ((SET_DEST (set) == x
5993 || (GET_CODE (SET_DEST (set)) == SUBREG
5994 && (GET_MODE_SIZE (GET_MODE (SET_DEST (set)))
5996 && (GET_MODE_CLASS (GET_MODE (SET_DEST (set)))
5998 && SUBREG_REG (SET_DEST (set)) == x))
5999 && basic_induction_var (loop, SET_SRC (set),
6000 (GET_MODE (SET_SRC (set)) == VOIDmode
6002 : GET_MODE (SET_SRC (set))),
6004 inc_val, mult_val, location))
6007 /* ... fall through ... */
6009 /* Can accept constant setting of biv only when inside inner most loop.
6010 Otherwise, a biv of an inner loop may be incorrectly recognized
6011 as a biv of the outer loop,
6012 causing code to be moved INTO the inner loop. */
6014 if (loop_invariant_p (loop, x) != 1)
6019 /* convert_modes aborts if we try to convert to or from CCmode, so just
6020 exclude that case. It is very unlikely that a condition code value
6021 would be a useful iterator anyways. */
6022 if (loop->level == 1
6023 && GET_MODE_CLASS (mode) != MODE_CC
6024 && GET_MODE_CLASS (GET_MODE (dest_reg)) != MODE_CC)
6026 /* Possible bug here? Perhaps we don't know the mode of X. */
6027 *inc_val = convert_modes (GET_MODE (dest_reg), mode, x, 0);
6028 *mult_val = const0_rtx;
6035 return basic_induction_var (loop, XEXP (x, 0), GET_MODE (XEXP (x, 0)),
6036 dest_reg, p, inc_val, mult_val, location);
6039 /* Similar, since this can be a sign extension. */
6040 for (insn = PREV_INSN (p);
6041 (insn && GET_CODE (insn) == NOTE
6042 && NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_BEG);
6043 insn = PREV_INSN (insn))
6047 set = single_set (insn);
6049 if (! rtx_equal_p (dest_reg, XEXP (x, 0))
6050 && set && SET_DEST (set) == XEXP (x, 0)
6051 && GET_CODE (XEXP (x, 1)) == CONST_INT
6052 && INTVAL (XEXP (x, 1)) >= 0
6053 && GET_CODE (SET_SRC (set)) == ASHIFT
6054 && XEXP (x, 1) == XEXP (SET_SRC (set), 1))
6055 return basic_induction_var (loop, XEXP (SET_SRC (set), 0),
6056 GET_MODE (XEXP (x, 0)),
6057 dest_reg, insn, inc_val, mult_val,
6066 /* A general induction variable (giv) is any quantity that is a linear
6067 function of a basic induction variable,
6068 i.e. giv = biv * mult_val + add_val.
6069 The coefficients can be any loop invariant quantity.
6070 A giv need not be computed directly from the biv;
6071 it can be computed by way of other givs. */
6073 /* Determine whether X computes a giv.
6074 If it does, return a nonzero value
6075 which is the benefit from eliminating the computation of X;
6076 set *SRC_REG to the register of the biv that it is computed from;
6077 set *ADD_VAL and *MULT_VAL to the coefficients,
6078 such that the value of X is biv * mult + add; */
6081 general_induction_var (loop, x, src_reg, add_val, mult_val, is_addr,
6082 pbenefit, addr_mode)
6083 const struct loop *loop;
6090 enum machine_mode addr_mode;
6095 /* If this is an invariant, forget it, it isn't a giv. */
6096 if (loop_invariant_p (loop, x) == 1)
6099 /* See if the expression could be a giv and get its form.
6100 Mark our place on the obstack in case we don't find a giv. */
6101 storage = (char *) oballoc (0);
6103 x = simplify_giv_expr (loop, x, pbenefit);
6110 switch (GET_CODE (x))
6114 /* Since this is now an invariant and wasn't before, it must be a giv
6115 with MULT_VAL == 0. It doesn't matter which BIV we associate this
6117 *src_reg = loop_iv_list->biv->dest_reg;
6118 *mult_val = const0_rtx;
6123 /* This is equivalent to a BIV. */
6125 *mult_val = const1_rtx;
6126 *add_val = const0_rtx;
6130 /* Either (plus (biv) (invar)) or
6131 (plus (mult (biv) (invar_1)) (invar_2)). */
6132 if (GET_CODE (XEXP (x, 0)) == MULT)
6134 *src_reg = XEXP (XEXP (x, 0), 0);
6135 *mult_val = XEXP (XEXP (x, 0), 1);
6139 *src_reg = XEXP (x, 0);
6140 *mult_val = const1_rtx;
6142 *add_val = XEXP (x, 1);
6146 /* ADD_VAL is zero. */
6147 *src_reg = XEXP (x, 0);
6148 *mult_val = XEXP (x, 1);
6149 *add_val = const0_rtx;
6156 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
6157 unless they are CONST_INT). */
6158 if (GET_CODE (*add_val) == USE)
6159 *add_val = XEXP (*add_val, 0);
6160 if (GET_CODE (*mult_val) == USE)
6161 *mult_val = XEXP (*mult_val, 0);
6164 *pbenefit += address_cost (orig_x, addr_mode) - reg_address_cost;
6166 *pbenefit += rtx_cost (orig_x, SET);
6168 /* Always return true if this is a giv so it will be detected as such,
6169 even if the benefit is zero or negative. This allows elimination
6170 of bivs that might otherwise not be eliminated. */
6174 /* Given an expression, X, try to form it as a linear function of a biv.
6175 We will canonicalize it to be of the form
6176 (plus (mult (BIV) (invar_1))
6178 with possible degeneracies.
6180 The invariant expressions must each be of a form that can be used as a
6181 machine operand. We surround then with a USE rtx (a hack, but localized
6182 and certainly unambiguous!) if not a CONST_INT for simplicity in this
6183 routine; it is the caller's responsibility to strip them.
6185 If no such canonicalization is possible (i.e., two biv's are used or an
6186 expression that is neither invariant nor a biv or giv), this routine
6189 For a non-zero return, the result will have a code of CONST_INT, USE,
6190 REG (for a BIV), PLUS, or MULT. No other codes will occur.
6192 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
6194 static rtx sge_plus PARAMS ((enum machine_mode, rtx, rtx));
6195 static rtx sge_plus_constant PARAMS ((rtx, rtx));
6196 static int cmp_combine_givs_stats PARAMS ((const PTR, const PTR));
6197 static int cmp_recombine_givs_stats PARAMS ((const PTR, const PTR));
6200 simplify_giv_expr (loop, x, benefit)
6201 const struct loop *loop;
6205 enum machine_mode mode = GET_MODE (x);
6209 /* If this is not an integer mode, or if we cannot do arithmetic in this
6210 mode, this can't be a giv. */
6211 if (mode != VOIDmode
6212 && (GET_MODE_CLASS (mode) != MODE_INT
6213 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
6216 switch (GET_CODE (x))
6219 arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
6220 arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
6221 if (arg0 == 0 || arg1 == 0)
6224 /* Put constant last, CONST_INT last if both constant. */
6225 if ((GET_CODE (arg0) == USE
6226 || GET_CODE (arg0) == CONST_INT)
6227 && ! ((GET_CODE (arg0) == USE
6228 && GET_CODE (arg1) == USE)
6229 || GET_CODE (arg1) == CONST_INT))
6230 tem = arg0, arg0 = arg1, arg1 = tem;
6232 /* Handle addition of zero, then addition of an invariant. */
6233 if (arg1 == const0_rtx)
6235 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
6236 switch (GET_CODE (arg0))
6240 /* Adding two invariants must result in an invariant, so enclose
6241 addition operation inside a USE and return it. */
6242 if (GET_CODE (arg0) == USE)
6243 arg0 = XEXP (arg0, 0);
6244 if (GET_CODE (arg1) == USE)
6245 arg1 = XEXP (arg1, 0);
6247 if (GET_CODE (arg0) == CONST_INT)
6248 tem = arg0, arg0 = arg1, arg1 = tem;
6249 if (GET_CODE (arg1) == CONST_INT)
6250 tem = sge_plus_constant (arg0, arg1);
6252 tem = sge_plus (mode, arg0, arg1);
6254 if (GET_CODE (tem) != CONST_INT)
6255 tem = gen_rtx_USE (mode, tem);
6260 /* biv + invar or mult + invar. Return sum. */
6261 return gen_rtx_PLUS (mode, arg0, arg1);
6264 /* (a + invar_1) + invar_2. Associate. */
6266 simplify_giv_expr (loop,
6278 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
6279 MULT to reduce cases. */
6280 if (GET_CODE (arg0) == REG)
6281 arg0 = gen_rtx_MULT (mode, arg0, const1_rtx);
6282 if (GET_CODE (arg1) == REG)
6283 arg1 = gen_rtx_MULT (mode, arg1, const1_rtx);
6285 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
6286 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
6287 Recurse to associate the second PLUS. */
6288 if (GET_CODE (arg1) == MULT)
6289 tem = arg0, arg0 = arg1, arg1 = tem;
6291 if (GET_CODE (arg1) == PLUS)
6293 simplify_giv_expr (loop,
6295 gen_rtx_PLUS (mode, arg0,
6300 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
6301 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
6304 if (!rtx_equal_p (arg0, arg1))
6307 return simplify_giv_expr (loop,
6316 /* Handle "a - b" as "a + b * (-1)". */
6317 return simplify_giv_expr (loop,
6326 arg0 = simplify_giv_expr (loop, XEXP (x, 0), benefit);
6327 arg1 = simplify_giv_expr (loop, XEXP (x, 1), benefit);
6328 if (arg0 == 0 || arg1 == 0)
6331 /* Put constant last, CONST_INT last if both constant. */
6332 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
6333 && GET_CODE (arg1) != CONST_INT)
6334 tem = arg0, arg0 = arg1, arg1 = tem;
6336 /* If second argument is not now constant, not giv. */
6337 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
6340 /* Handle multiply by 0 or 1. */
6341 if (arg1 == const0_rtx)
6344 else if (arg1 == const1_rtx)
6347 switch (GET_CODE (arg0))
6350 /* biv * invar. Done. */
6351 return gen_rtx_MULT (mode, arg0, arg1);
6354 /* Product of two constants. */
6355 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
6358 /* invar * invar is a giv, but attempt to simplify it somehow. */
6359 if (GET_CODE (arg1) != CONST_INT)
6362 arg0 = XEXP (arg0, 0);
6363 if (GET_CODE (arg0) == MULT)
6365 /* (invar_0 * invar_1) * invar_2. Associate. */
6366 return simplify_giv_expr (loop,
6375 /* Porpagate the MULT expressions to the intermost nodes. */
6376 else if (GET_CODE (arg0) == PLUS)
6378 /* (invar_0 + invar_1) * invar_2. Distribute. */
6379 return simplify_giv_expr (loop,
6391 return gen_rtx_USE (mode, gen_rtx_MULT (mode, arg0, arg1));
6394 /* (a * invar_1) * invar_2. Associate. */
6395 return simplify_giv_expr (loop,
6404 /* (a + invar_1) * invar_2. Distribute. */
6405 return simplify_giv_expr (loop,
6420 /* Shift by constant is multiply by power of two. */
6421 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
6425 simplify_giv_expr (loop,
6428 GEN_INT ((HOST_WIDE_INT) 1
6429 << INTVAL (XEXP (x, 1)))),
6433 /* "-a" is "a * (-1)" */
6434 return simplify_giv_expr (loop,
6435 gen_rtx_MULT (mode, XEXP (x, 0), constm1_rtx),
6439 /* "~a" is "-a - 1". Silly, but easy. */
6440 return simplify_giv_expr (loop,
6441 gen_rtx_MINUS (mode,
6442 gen_rtx_NEG (mode, XEXP (x, 0)),
6447 /* Already in proper form for invariant. */
6451 /* If this is a new register, we can't deal with it. */
6452 if (REGNO (x) >= max_reg_before_loop)
6455 /* Check for biv or giv. */
6456 switch (REG_IV_TYPE (REGNO (x)))
6460 case GENERAL_INDUCT:
6462 struct induction *v = REG_IV_INFO (REGNO (x));
6464 /* Form expression from giv and add benefit. Ensure this giv
6465 can derive another and subtract any needed adjustment if so. */
6466 *benefit += v->benefit;
6470 tem = gen_rtx_PLUS (mode, gen_rtx_MULT (mode,
6471 v->src_reg, v->mult_val),
6474 if (v->derive_adjustment)
6475 tem = gen_rtx_MINUS (mode, tem, v->derive_adjustment);
6476 return simplify_giv_expr (loop, tem, benefit);
6480 /* If it isn't an induction variable, and it is invariant, we
6481 may be able to simplify things further by looking through
6482 the bits we just moved outside the loop. */
6483 if (loop_invariant_p (loop, x) == 1)
6487 for (m = the_movables; m; m = m->next)
6488 if (rtx_equal_p (x, m->set_dest))
6490 /* Ok, we found a match. Substitute and simplify. */
6492 /* If we match another movable, we must use that, as
6493 this one is going away. */
6495 return simplify_giv_expr (loop, m->match->set_dest,
6498 /* If consec is non-zero, this is a member of a group of
6499 instructions that were moved together. We handle this
6500 case only to the point of seeking to the last insn and
6501 looking for a REG_EQUAL. Fail if we don't find one. */
6506 do { tem = NEXT_INSN (tem); } while (--i > 0);
6508 tem = find_reg_note (tem, REG_EQUAL, NULL_RTX);
6510 tem = XEXP (tem, 0);
6514 tem = single_set (m->insn);
6516 tem = SET_SRC (tem);
6521 /* What we are most interested in is pointer
6522 arithmetic on invariants -- only take
6523 patterns we may be able to do something with. */
6524 if (GET_CODE (tem) == PLUS
6525 || GET_CODE (tem) == MULT
6526 || GET_CODE (tem) == ASHIFT
6527 || GET_CODE (tem) == CONST_INT
6528 || GET_CODE (tem) == SYMBOL_REF)
6530 tem = simplify_giv_expr (loop, tem, benefit);
6534 else if (GET_CODE (tem) == CONST
6535 && GET_CODE (XEXP (tem, 0)) == PLUS
6536 && GET_CODE (XEXP (XEXP (tem, 0), 0)) == SYMBOL_REF
6537 && GET_CODE (XEXP (XEXP (tem, 0), 1)) == CONST_INT)
6539 tem = simplify_giv_expr (loop, XEXP (tem, 0),
6551 /* Fall through to general case. */
6553 /* If invariant, return as USE (unless CONST_INT).
6554 Otherwise, not giv. */
6555 if (GET_CODE (x) == USE)
6558 if (loop_invariant_p (loop, x) == 1)
6560 if (GET_CODE (x) == CONST_INT)
6562 if (GET_CODE (x) == CONST
6563 && GET_CODE (XEXP (x, 0)) == PLUS
6564 && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF
6565 && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
6567 return gen_rtx_USE (mode, x);
6574 /* This routine folds invariants such that there is only ever one
6575 CONST_INT in the summation. It is only used by simplify_giv_expr. */
6578 sge_plus_constant (x, c)
6581 if (GET_CODE (x) == CONST_INT)
6582 return GEN_INT (INTVAL (x) + INTVAL (c));
6583 else if (GET_CODE (x) != PLUS)
6584 return gen_rtx_PLUS (GET_MODE (x), x, c);
6585 else if (GET_CODE (XEXP (x, 1)) == CONST_INT)
6587 return gen_rtx_PLUS (GET_MODE (x), XEXP (x, 0),
6588 GEN_INT (INTVAL (XEXP (x, 1)) + INTVAL (c)));
6590 else if (GET_CODE (XEXP (x, 0)) == PLUS
6591 || GET_CODE (XEXP (x, 1)) != PLUS)
6593 return gen_rtx_PLUS (GET_MODE (x),
6594 sge_plus_constant (XEXP (x, 0), c), XEXP (x, 1));
6598 return gen_rtx_PLUS (GET_MODE (x),
6599 sge_plus_constant (XEXP (x, 1), c), XEXP (x, 0));
6604 sge_plus (mode, x, y)
6605 enum machine_mode mode;
6608 while (GET_CODE (y) == PLUS)
6610 rtx a = XEXP (y, 0);
6611 if (GET_CODE (a) == CONST_INT)
6612 x = sge_plus_constant (x, a);
6614 x = gen_rtx_PLUS (mode, x, a);
6617 if (GET_CODE (y) == CONST_INT)
6618 x = sge_plus_constant (x, y);
6620 x = gen_rtx_PLUS (mode, x, y);
6624 /* Help detect a giv that is calculated by several consecutive insns;
6628 The caller has already identified the first insn P as having a giv as dest;
6629 we check that all other insns that set the same register follow
6630 immediately after P, that they alter nothing else,
6631 and that the result of the last is still a giv.
6633 The value is 0 if the reg set in P is not really a giv.
6634 Otherwise, the value is the amount gained by eliminating
6635 all the consecutive insns that compute the value.
6637 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
6638 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
6640 The coefficients of the ultimate giv value are stored in
6641 *MULT_VAL and *ADD_VAL. */
6644 consec_sets_giv (loop, first_benefit, p, src_reg, dest_reg,
6645 add_val, mult_val, last_consec_insn)
6646 const struct loop *loop;
6653 rtx *last_consec_insn;
6661 /* Indicate that this is a giv so that we can update the value produced in
6662 each insn of the multi-insn sequence.
6664 This induction structure will be used only by the call to
6665 general_induction_var below, so we can allocate it on our stack.
6666 If this is a giv, our caller will replace the induct var entry with
6667 a new induction structure. */
6669 = (struct induction *) alloca (sizeof (struct induction));
6670 v->src_reg = src_reg;
6671 v->mult_val = *mult_val;
6672 v->add_val = *add_val;
6673 v->benefit = first_benefit;
6675 v->derive_adjustment = 0;
6677 REG_IV_TYPE (REGNO (dest_reg)) = GENERAL_INDUCT;
6678 REG_IV_INFO (REGNO (dest_reg)) = v;
6680 count = VARRAY_INT (n_times_set, REGNO (dest_reg)) - 1;
6685 code = GET_CODE (p);
6687 /* If libcall, skip to end of call sequence. */
6688 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
6692 && (set = single_set (p))
6693 && GET_CODE (SET_DEST (set)) == REG
6694 && SET_DEST (set) == dest_reg
6695 && (general_induction_var (loop, SET_SRC (set), &src_reg,
6696 add_val, mult_val, 0, &benefit, VOIDmode)
6697 /* Giv created by equivalent expression. */
6698 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
6699 && general_induction_var (loop, XEXP (temp, 0), &src_reg,
6700 add_val, mult_val, 0, &benefit,
6702 && src_reg == v->src_reg)
6704 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
6705 benefit += libcall_benefit (p);
6708 v->mult_val = *mult_val;
6709 v->add_val = *add_val;
6710 v->benefit = benefit;
6712 else if (code != NOTE)
6714 /* Allow insns that set something other than this giv to a
6715 constant. Such insns are needed on machines which cannot
6716 include long constants and should not disqualify a giv. */
6718 && (set = single_set (p))
6719 && SET_DEST (set) != dest_reg
6720 && CONSTANT_P (SET_SRC (set)))
6723 REG_IV_TYPE (REGNO (dest_reg)) = UNKNOWN_INDUCT;
6728 *last_consec_insn = p;
6732 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6733 represented by G1. If no such expression can be found, or it is clear that
6734 it cannot possibly be a valid address, 0 is returned.
6736 To perform the computation, we note that
6739 where `v' is the biv.
6741 So G2 = (y/b) * G1 + (b - a*y/x).
6743 Note that MULT = y/x.
6745 Update: A and B are now allowed to be additive expressions such that
6746 B contains all variables in A. That is, computing B-A will not require
6747 subtracting variables. */
6750 express_from_1 (a, b, mult)
6753 /* If MULT is zero, then A*MULT is zero, and our expression is B. */
6755 if (mult == const0_rtx)
6758 /* If MULT is not 1, we cannot handle A with non-constants, since we
6759 would then be required to subtract multiples of the registers in A.
6760 This is theoretically possible, and may even apply to some Fortran
6761 constructs, but it is a lot of work and we do not attempt it here. */
6763 if (mult != const1_rtx && GET_CODE (a) != CONST_INT)
6766 /* In general these structures are sorted top to bottom (down the PLUS
6767 chain), but not left to right across the PLUS. If B is a higher
6768 order giv than A, we can strip one level and recurse. If A is higher
6769 order, we'll eventually bail out, but won't know that until the end.
6770 If they are the same, we'll strip one level around this loop. */
6772 while (GET_CODE (a) == PLUS && GET_CODE (b) == PLUS)
6774 rtx ra, rb, oa, ob, tmp;
6776 ra = XEXP (a, 0), oa = XEXP (a, 1);
6777 if (GET_CODE (ra) == PLUS)
6778 tmp = ra, ra = oa, oa = tmp;
6780 rb = XEXP (b, 0), ob = XEXP (b, 1);
6781 if (GET_CODE (rb) == PLUS)
6782 tmp = rb, rb = ob, ob = tmp;
6784 if (rtx_equal_p (ra, rb))
6785 /* We matched: remove one reg completely. */
6787 else if (GET_CODE (ob) != PLUS && rtx_equal_p (ra, ob))
6788 /* An alternate match. */
6790 else if (GET_CODE (oa) != PLUS && rtx_equal_p (oa, rb))
6791 /* An alternate match. */
6795 /* Indicates an extra register in B. Strip one level from B and
6796 recurse, hoping B was the higher order expression. */
6797 ob = express_from_1 (a, ob, mult);
6800 return gen_rtx_PLUS (GET_MODE (b), rb, ob);
6804 /* Here we are at the last level of A, go through the cases hoping to
6805 get rid of everything but a constant. */
6807 if (GET_CODE (a) == PLUS)
6811 ra = XEXP (a, 0), oa = XEXP (a, 1);
6812 if (rtx_equal_p (oa, b))
6814 else if (!rtx_equal_p (ra, b))
6817 if (GET_CODE (oa) != CONST_INT)
6820 return GEN_INT (-INTVAL (oa) * INTVAL (mult));
6822 else if (GET_CODE (a) == CONST_INT)
6824 return plus_constant (b, -INTVAL (a) * INTVAL (mult));
6826 else if (CONSTANT_P (a))
6828 return simplify_gen_binary (MINUS, GET_MODE (b) != VOIDmode ? GET_MODE (b) : GET_MODE (a), const0_rtx, a);
6830 else if (GET_CODE (b) == PLUS)
6832 if (rtx_equal_p (a, XEXP (b, 0)))
6834 else if (rtx_equal_p (a, XEXP (b, 1)))
6839 else if (rtx_equal_p (a, b))
6846 express_from (g1, g2)
6847 struct induction *g1, *g2;
6851 /* The value that G1 will be multiplied by must be a constant integer. Also,
6852 the only chance we have of getting a valid address is if b*c/a (see above
6853 for notation) is also an integer. */
6854 if (GET_CODE (g1->mult_val) == CONST_INT
6855 && GET_CODE (g2->mult_val) == CONST_INT)
6857 if (g1->mult_val == const0_rtx
6858 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
6860 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
6862 else if (rtx_equal_p (g1->mult_val, g2->mult_val))
6866 /* ??? Find out if the one is a multiple of the other? */
6870 add = express_from_1 (g1->add_val, g2->add_val, mult);
6871 if (add == NULL_RTX)
6873 /* Failed. If we've got a multiplication factor between G1 and G2,
6874 scale G1's addend and try again. */
6875 if (INTVAL (mult) > 1)
6877 rtx g1_add_val = g1->add_val;
6878 if (GET_CODE (g1_add_val) == MULT
6879 && GET_CODE (XEXP (g1_add_val, 1)) == CONST_INT)
6882 m = INTVAL (mult) * INTVAL (XEXP (g1_add_val, 1));
6883 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val),
6884 XEXP (g1_add_val, 0), GEN_INT (m));
6888 g1_add_val = gen_rtx_MULT (GET_MODE (g1_add_val), g1_add_val,
6892 add = express_from_1 (g1_add_val, g2->add_val, const1_rtx);
6895 if (add == NULL_RTX)
6898 /* Form simplified final result. */
6899 if (mult == const0_rtx)
6901 else if (mult == const1_rtx)
6902 mult = g1->dest_reg;
6904 mult = gen_rtx_MULT (g2->mode, g1->dest_reg, mult);
6906 if (add == const0_rtx)
6910 if (GET_CODE (add) == PLUS
6911 && CONSTANT_P (XEXP (add, 1)))
6913 rtx tem = XEXP (add, 1);
6914 mult = gen_rtx_PLUS (g2->mode, mult, XEXP (add, 0));
6918 return gen_rtx_PLUS (g2->mode, mult, add);
6922 /* Return an rtx, if any, that expresses giv G2 as a function of the register
6923 represented by G1. This indicates that G2 should be combined with G1 and
6924 that G2 can use (either directly or via an address expression) a register
6925 used to represent G1. */
6928 combine_givs_p (g1, g2)
6929 struct induction *g1, *g2;
6931 rtx tem = express_from (g1, g2);
6933 /* If these givs are identical, they can be combined. We use the results
6934 of express_from because the addends are not in a canonical form, so
6935 rtx_equal_p is a weaker test. */
6936 /* But don't combine a DEST_REG giv with a DEST_ADDR giv; we want the
6937 combination to be the other way round. */
6938 if (tem == g1->dest_reg
6939 && (g1->giv_type == DEST_REG || g2->giv_type == DEST_ADDR))
6941 return g1->dest_reg;
6944 /* If G2 can be expressed as a function of G1 and that function is valid
6945 as an address and no more expensive than using a register for G2,
6946 the expression of G2 in terms of G1 can be used. */
6948 && g2->giv_type == DEST_ADDR
6949 && memory_address_p (g2->mem_mode, tem)
6950 /* ??? Looses, especially with -fforce-addr, where *g2->location
6951 will always be a register, and so anything more complicated
6955 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location)
6957 && rtx_cost (tem, MEM) <= rtx_cost (*g2->location, MEM)
6968 struct combine_givs_stats
6975 cmp_combine_givs_stats (xp, yp)
6979 const struct combine_givs_stats * const x =
6980 (const struct combine_givs_stats *) xp;
6981 const struct combine_givs_stats * const y =
6982 (const struct combine_givs_stats *) yp;
6984 d = y->total_benefit - x->total_benefit;
6985 /* Stabilize the sort. */
6987 d = x->giv_number - y->giv_number;
6991 /* Check all pairs of givs for iv_class BL and see if any can be combined with
6992 any other. If so, point SAME to the giv combined with and set NEW_REG to
6993 be an expression (in terms of the other giv's DEST_REG) equivalent to the
6994 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
6998 struct iv_class *bl;
7000 /* Additional benefit to add for being combined multiple times. */
7001 const int extra_benefit = 3;
7003 struct induction *g1, *g2, **giv_array;
7004 int i, j, k, giv_count;
7005 struct combine_givs_stats *stats;
7008 /* Count givs, because bl->giv_count is incorrect here. */
7010 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7015 = (struct induction **) alloca (giv_count * sizeof (struct induction *));
7017 for (g1 = bl->giv; g1; g1 = g1->next_iv)
7019 giv_array[i++] = g1;
7021 stats = (struct combine_givs_stats *) xcalloc (giv_count, sizeof (*stats));
7022 can_combine = (rtx *) xcalloc (giv_count, giv_count * sizeof (rtx));
7024 for (i = 0; i < giv_count; i++)
7030 stats[i].giv_number = i;
7032 /* If a DEST_REG GIV is used only once, do not allow it to combine
7033 with anything, for in doing so we will gain nothing that cannot
7034 be had by simply letting the GIV with which we would have combined
7035 to be reduced on its own. The losage shows up in particular with
7036 DEST_ADDR targets on hosts with reg+reg addressing, though it can
7037 be seen elsewhere as well. */
7038 if (g1->giv_type == DEST_REG
7039 && (single_use = VARRAY_RTX (reg_single_usage, REGNO (g1->dest_reg)))
7040 && single_use != const0_rtx)
7043 this_benefit = g1->benefit;
7044 /* Add an additional weight for zero addends. */
7045 if (g1->no_const_addval)
7048 for (j = 0; j < giv_count; j++)
7054 && (this_combine = combine_givs_p (g1, g2)) != NULL_RTX)
7056 can_combine[i * giv_count + j] = this_combine;
7057 this_benefit += g2->benefit + extra_benefit;
7060 stats[i].total_benefit = this_benefit;
7063 /* Iterate, combining until we can't. */
7065 qsort (stats, giv_count, sizeof (*stats), cmp_combine_givs_stats);
7067 if (loop_dump_stream)
7069 fprintf (loop_dump_stream, "Sorted combine statistics:\n");
7070 for (k = 0; k < giv_count; k++)
7072 g1 = giv_array[stats[k].giv_number];
7073 if (!g1->combined_with && !g1->same)
7074 fprintf (loop_dump_stream, " {%d, %d}",
7075 INSN_UID (giv_array[stats[k].giv_number]->insn),
7076 stats[k].total_benefit);
7078 putc ('\n', loop_dump_stream);
7081 for (k = 0; k < giv_count; k++)
7083 int g1_add_benefit = 0;
7085 i = stats[k].giv_number;
7088 /* If it has already been combined, skip. */
7089 if (g1->combined_with || g1->same)
7092 for (j = 0; j < giv_count; j++)
7095 if (g1 != g2 && can_combine[i * giv_count + j]
7096 /* If it has already been combined, skip. */
7097 && ! g2->same && ! g2->combined_with)
7101 g2->new_reg = can_combine[i * giv_count + j];
7103 g1->combined_with++;
7104 g1->lifetime += g2->lifetime;
7106 g1_add_benefit += g2->benefit;
7108 /* ??? The new final_[bg]iv_value code does a much better job
7109 of finding replaceable giv's, and hence this code may no
7110 longer be necessary. */
7111 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
7112 g1_add_benefit -= copy_cost;
7114 /* To help optimize the next set of combinations, remove
7115 this giv from the benefits of other potential mates. */
7116 for (l = 0; l < giv_count; ++l)
7118 int m = stats[l].giv_number;
7119 if (can_combine[m * giv_count + j])
7120 stats[l].total_benefit -= g2->benefit + extra_benefit;
7123 if (loop_dump_stream)
7124 fprintf (loop_dump_stream,
7125 "giv at %d combined with giv at %d\n",
7126 INSN_UID (g2->insn), INSN_UID (g1->insn));
7130 /* To help optimize the next set of combinations, remove
7131 this giv from the benefits of other potential mates. */
7132 if (g1->combined_with)
7134 for (j = 0; j < giv_count; ++j)
7136 int m = stats[j].giv_number;
7137 if (can_combine[m * giv_count + i])
7138 stats[j].total_benefit -= g1->benefit + extra_benefit;
7141 g1->benefit += g1_add_benefit;
7143 /* We've finished with this giv, and everything it touched.
7144 Restart the combination so that proper weights for the
7145 rest of the givs are properly taken into account. */
7146 /* ??? Ideally we would compact the arrays at this point, so
7147 as to not cover old ground. But sanely compacting
7148 can_combine is tricky. */
7158 struct recombine_givs_stats
7161 int start_luid, end_luid;
7164 /* Used below as comparison function for qsort. We want a ascending luid
7165 when scanning the array starting at the end, thus the arguments are
7168 cmp_recombine_givs_stats (xp, yp)
7172 const struct recombine_givs_stats * const x =
7173 (const struct recombine_givs_stats *) xp;
7174 const struct recombine_givs_stats * const y =
7175 (const struct recombine_givs_stats *) yp;
7177 d = y->start_luid - x->start_luid;
7178 /* Stabilize the sort. */
7180 d = y->giv_number - x->giv_number;
7184 /* Scan X, which is a part of INSN, for the end of life of a giv. Also
7185 look for the start of life of a giv where the start has not been seen
7186 yet to unlock the search for the end of its life.
7187 Only consider givs that belong to BIV.
7188 Return the total number of lifetime ends that have been found. */
7190 find_life_end (x, stats, insn, biv)
7192 struct recombine_givs_stats *stats;
7199 code = GET_CODE (x);
7204 rtx reg = SET_DEST (x);
7205 if (GET_CODE (reg) == REG)
7207 int regno = REGNO (reg);
7208 struct induction *v = REG_IV_INFO (regno);
7210 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7212 && v->src_reg == biv
7213 && stats[v->ix].end_luid <= 0)
7215 /* If we see a 0 here for end_luid, it means that we have
7216 scanned the entire loop without finding any use at all.
7217 We must not predicate this code on a start_luid match
7218 since that would make the test fail for givs that have
7219 been hoisted out of inner loops. */
7220 if (stats[v->ix].end_luid == 0)
7222 stats[v->ix].end_luid = stats[v->ix].start_luid;
7223 return 1 + find_life_end (SET_SRC (x), stats, insn, biv);
7225 else if (stats[v->ix].start_luid == INSN_LUID (insn))
7226 stats[v->ix].end_luid = 0;
7228 return find_life_end (SET_SRC (x), stats, insn, biv);
7234 int regno = REGNO (x);
7235 struct induction *v = REG_IV_INFO (regno);
7237 if (REG_IV_TYPE (regno) == GENERAL_INDUCT
7239 && v->src_reg == biv
7240 && stats[v->ix].end_luid == 0)
7242 while (INSN_UID (insn) >= max_uid_for_loop)
7243 insn = NEXT_INSN (insn);
7244 stats[v->ix].end_luid = INSN_LUID (insn);
7257 fmt = GET_RTX_FORMAT (code);
7259 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
7262 retval += find_life_end (XEXP (x, i), stats, insn, biv);
7264 else if (fmt[i] == 'E')
7265 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
7266 retval += find_life_end (XVECEXP (x, i, j), stats, insn, biv);
7271 /* For each giv that has been combined with another, look if
7272 we can combine it with the most recently used one instead.
7273 This tends to shorten giv lifetimes, and helps the next step:
7274 try to derive givs from other givs. */
7276 recombine_givs (loop, bl, unroll_p)
7277 const struct loop *loop;
7278 struct iv_class *bl;
7281 struct induction *v, **giv_array, *last_giv;
7282 struct recombine_givs_stats *stats;
7285 int ends_need_computing;
7287 for (giv_count = 0, v = bl->giv; v; v = v->next_iv)
7293 = (struct induction **) xmalloc (giv_count * sizeof (struct induction *));
7294 stats = (struct recombine_givs_stats *) xmalloc (giv_count * sizeof *stats);
7296 /* Initialize stats and set up the ix field for each giv in stats to name
7297 the corresponding index into stats. */
7298 for (i = 0, v = bl->giv; v; v = v->next_iv)
7305 stats[i].giv_number = i;
7306 /* If this giv has been hoisted out of an inner loop, use the luid of
7307 the previous insn. */
7308 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop; )
7310 stats[i].start_luid = INSN_LUID (p);
7314 qsort (stats, giv_count, sizeof (*stats), cmp_recombine_givs_stats);
7316 /* Set up the ix field for each giv in stats to name
7317 the corresponding index into stats, and
7318 do the actual most-recently-used recombination. */
7319 for (last_giv = 0, i = giv_count - 1; i >= 0; i--)
7321 v = giv_array[stats[i].giv_number];
7325 struct induction *old_same = v->same;
7328 /* combine_givs_p actually says if we can make this transformation.
7329 The other tests are here only to avoid keeping a giv alive
7330 that could otherwise be eliminated. */
7332 && ((old_same->maybe_dead && ! old_same->combined_with)
7333 || ! last_giv->maybe_dead
7334 || last_giv->combined_with)
7335 && (new_combine = combine_givs_p (last_giv, v)))
7337 old_same->combined_with--;
7338 v->new_reg = new_combine;
7340 last_giv->combined_with++;
7341 /* No need to update lifetimes / benefits here since we have
7342 already decided what to reduce. */
7344 if (loop_dump_stream)
7346 fprintf (loop_dump_stream,
7347 "giv at %d recombined with giv at %d as ",
7348 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7349 print_rtl (loop_dump_stream, v->new_reg);
7350 putc ('\n', loop_dump_stream);
7356 else if (v->giv_type != DEST_REG)
7359 || (last_giv->maybe_dead && ! last_giv->combined_with)
7361 || v->combined_with)
7365 ends_need_computing = 0;
7366 /* For each DEST_REG giv, compute lifetime starts, and try to compute
7367 lifetime ends from regscan info. */
7368 for (i = giv_count - 1; i >= 0; i--)
7370 v = giv_array[stats[i].giv_number];
7373 if (v->giv_type == DEST_ADDR)
7375 /* Loop unrolling of an inner loop can even create new DEST_REG
7378 for (p = v->insn; INSN_UID (p) >= max_uid_for_loop;)
7380 stats[i].start_luid = stats[i].end_luid = INSN_LUID (p);
7382 stats[i].end_luid++;
7384 else /* v->giv_type == DEST_REG */
7388 stats[i].start_luid = INSN_LUID (v->insn);
7389 stats[i].end_luid = INSN_LUID (v->last_use);
7391 else if (INSN_UID (v->insn) >= max_uid_for_loop)
7394 /* This insn has been created by loop optimization on an inner
7395 loop. We don't have a proper start_luid that will match
7396 when we see the first set. But we do know that there will
7397 be no use before the set, so we can set end_luid to 0 so that
7398 we'll start looking for the last use right away. */
7399 for (p = PREV_INSN (v->insn); INSN_UID (p) >= max_uid_for_loop; )
7401 stats[i].start_luid = INSN_LUID (p);
7402 stats[i].end_luid = 0;
7403 ends_need_computing++;
7407 int regno = REGNO (v->dest_reg);
7408 int count = VARRAY_INT (n_times_set, regno) - 1;
7411 /* Find the first insn that sets the giv, so that we can verify
7412 if this giv's lifetime wraps around the loop. We also need
7413 the luid of the first setting insn in order to detect the
7414 last use properly. */
7417 p = prev_nonnote_insn (p);
7418 if (reg_set_p (v->dest_reg, p))
7422 stats[i].start_luid = INSN_LUID (p);
7423 if (stats[i].start_luid > uid_luid[REGNO_FIRST_UID (regno)])
7425 stats[i].end_luid = -1;
7426 ends_need_computing++;
7430 stats[i].end_luid = uid_luid[REGNO_LAST_UID (regno)];
7431 if (stats[i].end_luid > INSN_LUID (loop->end))
7433 stats[i].end_luid = -1;
7434 ends_need_computing++;
7441 /* If the regscan information was unconclusive for one or more DEST_REG
7442 givs, scan the all insn in the loop to find out lifetime ends. */
7443 if (ends_need_computing)
7445 rtx biv = bl->biv->src_reg;
7450 if (p == loop->start)
7455 ends_need_computing -= find_life_end (PATTERN (p), stats, p, biv);
7457 while (ends_need_computing);
7460 /* Set start_luid back to the last insn that sets the giv. This allows
7461 more combinations. */
7462 for (i = giv_count - 1; i >= 0; i--)
7464 v = giv_array[stats[i].giv_number];
7467 if (INSN_UID (v->insn) < max_uid_for_loop)
7468 stats[i].start_luid = INSN_LUID (v->insn);
7471 /* Now adjust lifetime ends by taking combined givs into account. */
7472 for (i = giv_count - 1; i >= 0; i--)
7477 v = giv_array[stats[i].giv_number];
7480 if (v->same && ! v->same->ignore)
7483 luid = stats[i].start_luid;
7484 /* Use unsigned arithmetic to model loop wrap-around. */
7485 if (luid - stats[j].start_luid
7486 > (unsigned) stats[j].end_luid - stats[j].start_luid)
7487 stats[j].end_luid = luid;
7491 qsort (stats, giv_count, sizeof (*stats), cmp_recombine_givs_stats);
7493 /* Try to derive DEST_REG givs from previous DEST_REG givs with the
7494 same mult_val and non-overlapping lifetime. This reduces register
7496 Once we find a DEST_REG giv that is suitable to derive others from,
7497 we set last_giv to this giv, and try to derive as many other DEST_REG
7498 givs from it without joining overlapping lifetimes. If we then
7499 encounter a DEST_REG giv that we can't derive, we set rescan to the
7500 index for this giv (unless rescan is already set).
7501 When we are finished with the current LAST_GIV (i.e. the inner loop
7502 terminates), we start again with rescan, which then becomes the new
7504 for (i = giv_count - 1; i >= 0; i = rescan)
7506 int life_start = 0, life_end = 0;
7508 for (last_giv = 0, rescan = -1; i >= 0; i--)
7512 v = giv_array[stats[i].giv_number];
7513 if (v->giv_type != DEST_REG || v->derived_from || v->same)
7517 /* Don't use a giv that's likely to be dead to derive
7518 others - that would be likely to keep that giv alive. */
7519 if (! v->maybe_dead || v->combined_with)
7522 life_start = stats[i].start_luid;
7523 life_end = stats[i].end_luid;
7527 /* Use unsigned arithmetic to model loop wrap around. */
7528 if (((unsigned) stats[i].start_luid - life_start
7529 >= (unsigned) life_end - life_start)
7530 && ((unsigned) stats[i].end_luid - life_start
7531 > (unsigned) life_end - life_start)
7532 /* Check that the giv insn we're about to use for deriving
7533 precedes all uses of that giv. Note that initializing the
7534 derived giv would defeat the purpose of reducing register
7536 ??? We could arrange to move the insn. */
7537 && ((unsigned) stats[i].end_luid - INSN_LUID (loop->start)
7538 > (unsigned) stats[i].start_luid - INSN_LUID (loop->start))
7539 && rtx_equal_p (last_giv->mult_val, v->mult_val)
7540 /* ??? Could handle libcalls, but would need more logic. */
7541 && ! find_reg_note (v->insn, REG_RETVAL, NULL_RTX)
7542 /* We would really like to know if for any giv that v
7543 is combined with, v->insn or any intervening biv increment
7544 dominates that combined giv. However, we
7545 don't have this detailed control flow information.
7546 N.B. since last_giv will be reduced, it is valid
7547 anywhere in the loop, so we don't need to check the
7548 validity of last_giv.
7549 We rely here on the fact that v->always_executed implies that
7550 there is no jump to someplace else in the loop before the
7551 giv insn, and hence any insn that is executed before the
7552 giv insn in the loop will have a lower luid. */
7553 && (v->always_executed || ! v->combined_with)
7554 && (sum = express_from (last_giv, v))
7555 /* Make sure we don't make the add more expensive. ADD_COST
7556 doesn't take different costs of registers and constants into
7557 account, so compare the cost of the actual SET_SRCs. */
7558 && (rtx_cost (sum, SET)
7559 <= rtx_cost (SET_SRC (single_set (v->insn)), SET))
7560 /* ??? unroll can't understand anything but reg + const_int
7561 sums. It would be cleaner to fix unroll. */
7562 && ((GET_CODE (sum) == PLUS
7563 && GET_CODE (XEXP (sum, 0)) == REG
7564 && GET_CODE (XEXP (sum, 1)) == CONST_INT)
7566 && validate_change (v->insn, &PATTERN (v->insn),
7567 gen_rtx_SET (VOIDmode, v->dest_reg, sum), 0))
7569 v->derived_from = last_giv;
7570 life_end = stats[i].end_luid;
7572 if (loop_dump_stream)
7574 fprintf (loop_dump_stream,
7575 "giv at %d derived from %d as ",
7576 INSN_UID (v->insn), INSN_UID (last_giv->insn));
7577 print_rtl (loop_dump_stream, sum);
7578 putc ('\n', loop_dump_stream);
7581 else if (rescan < 0)
7591 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
7594 emit_iv_add_mult (b, m, a, reg, insert_before)
7595 rtx b; /* initial value of basic induction variable */
7596 rtx m; /* multiplicative constant */
7597 rtx a; /* additive constant */
7598 rtx reg; /* destination register */
7604 /* Prevent unexpected sharing of these rtx. */
7608 /* Increase the lifetime of any invariants moved further in code. */
7609 update_reg_last_use (a, insert_before);
7610 update_reg_last_use (b, insert_before);
7611 update_reg_last_use (m, insert_before);
7614 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
7616 emit_move_insn (reg, result);
7617 seq = gen_sequence ();
7620 emit_insn_before (seq, insert_before);
7622 /* It is entirely possible that the expansion created lots of new
7623 registers. Iterate over the sequence we just created and
7626 if (GET_CODE (seq) == SEQUENCE)
7629 for (i = 0; i < XVECLEN (seq, 0); ++i)
7631 rtx set = single_set (XVECEXP (seq, 0, i));
7632 if (set && GET_CODE (SET_DEST (set)) == REG)
7633 record_base_value (REGNO (SET_DEST (set)), SET_SRC (set), 0);
7636 else if (GET_CODE (seq) == SET
7637 && GET_CODE (SET_DEST (seq)) == REG)
7638 record_base_value (REGNO (SET_DEST (seq)), SET_SRC (seq), 0);
7641 /* Test whether A * B can be computed without
7642 an actual multiply insn. Value is 1 if so. */
7645 product_cheap_p (a, b)
7651 struct obstack *old_rtl_obstack = rtl_obstack;
7652 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
7655 /* If only one is constant, make it B. */
7656 if (GET_CODE (a) == CONST_INT)
7657 tmp = a, a = b, b = tmp;
7659 /* If first constant, both constant, so don't need multiply. */
7660 if (GET_CODE (a) == CONST_INT)
7663 /* If second not constant, neither is constant, so would need multiply. */
7664 if (GET_CODE (b) != CONST_INT)
7667 /* One operand is constant, so might not need multiply insn. Generate the
7668 code for the multiply and see if a call or multiply, or long sequence
7669 of insns is generated. */
7671 rtl_obstack = &temp_obstack;
7673 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
7674 tmp = gen_sequence ();
7677 if (GET_CODE (tmp) == SEQUENCE)
7679 if (XVEC (tmp, 0) == 0)
7681 else if (XVECLEN (tmp, 0) > 3)
7684 for (i = 0; i < XVECLEN (tmp, 0); i++)
7686 rtx insn = XVECEXP (tmp, 0, i);
7688 if (GET_CODE (insn) != INSN
7689 || (GET_CODE (PATTERN (insn)) == SET
7690 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
7691 || (GET_CODE (PATTERN (insn)) == PARALLEL
7692 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
7693 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
7700 else if (GET_CODE (tmp) == SET
7701 && GET_CODE (SET_SRC (tmp)) == MULT)
7703 else if (GET_CODE (tmp) == PARALLEL
7704 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
7705 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
7708 /* Free any storage we obtained in generating this multiply and restore rtl
7709 allocation to its normal obstack. */
7710 obstack_free (&temp_obstack, storage);
7711 rtl_obstack = old_rtl_obstack;
7716 /* Check to see if loop can be terminated by a "decrement and branch until
7717 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
7718 Also try reversing an increment loop to a decrement loop
7719 to see if the optimization can be performed.
7720 Value is nonzero if optimization was performed. */
7722 /* This is useful even if the architecture doesn't have such an insn,
7723 because it might change a loops which increments from 0 to n to a loop
7724 which decrements from n to 0. A loop that decrements to zero is usually
7725 faster than one that increments from zero. */
7727 /* ??? This could be rewritten to use some of the loop unrolling procedures,
7728 such as approx_final_value, biv_total_increment, loop_iterations, and
7729 final_[bg]iv_value. */
7732 check_dbra_loop (loop, insn_count)
7736 struct iv_class *bl;
7743 rtx before_comparison;
7747 int compare_and_branch;
7748 rtx loop_start = loop->start;
7749 rtx loop_end = loop->end;
7750 struct loop_info *loop_info = LOOP_INFO (loop);
7752 /* If last insn is a conditional branch, and the insn before tests a
7753 register value, try to optimize it. Otherwise, we can't do anything. */
7755 jump = PREV_INSN (loop_end);
7756 comparison = get_condition_for_loop (loop, jump);
7757 if (comparison == 0)
7759 if (!onlyjump_p (jump))
7762 /* Try to compute whether the compare/branch at the loop end is one or
7763 two instructions. */
7764 get_condition (jump, &first_compare);
7765 if (first_compare == jump)
7766 compare_and_branch = 1;
7767 else if (first_compare == prev_nonnote_insn (jump))
7768 compare_and_branch = 2;
7773 /* If more than one condition is present to control the loop, then
7774 do not proceed, as this function does not know how to rewrite
7775 loop tests with more than one condition.
7777 Look backwards from the first insn in the last comparison
7778 sequence and see if we've got another comparison sequence. */
7781 if ((jump1 = prev_nonnote_insn (first_compare)) != loop->cont)
7782 if (GET_CODE (jump1) == JUMP_INSN)
7786 /* Check all of the bivs to see if the compare uses one of them.
7787 Skip biv's set more than once because we can't guarantee that
7788 it will be zero on the last iteration. Also skip if the biv is
7789 used between its update and the test insn. */
7791 for (bl = loop_iv_list; bl; bl = bl->next)
7793 if (bl->biv_count == 1
7794 && ! bl->biv->maybe_multiple
7795 && bl->biv->dest_reg == XEXP (comparison, 0)
7796 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
7804 /* Look for the case where the basic induction variable is always
7805 nonnegative, and equals zero on the last iteration.
7806 In this case, add a reg_note REG_NONNEG, which allows the
7807 m68k DBRA instruction to be used. */
7809 if (((GET_CODE (comparison) == GT
7810 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
7811 && INTVAL (XEXP (comparison, 1)) == -1)
7812 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
7813 && GET_CODE (bl->biv->add_val) == CONST_INT
7814 && INTVAL (bl->biv->add_val) < 0)
7816 /* Initial value must be greater than 0,
7817 init_val % -dec_value == 0 to ensure that it equals zero on
7818 the last iteration */
7820 if (GET_CODE (bl->initial_value) == CONST_INT
7821 && INTVAL (bl->initial_value) > 0
7822 && (INTVAL (bl->initial_value)
7823 % (-INTVAL (bl->biv->add_val))) == 0)
7825 /* register always nonnegative, add REG_NOTE to branch */
7826 if (! find_reg_note (jump, REG_NONNEG, NULL_RTX))
7828 = gen_rtx_EXPR_LIST (REG_NONNEG, bl->biv->dest_reg,
7835 /* If the decrement is 1 and the value was tested as >= 0 before
7836 the loop, then we can safely optimize. */
7837 for (p = loop_start; p; p = PREV_INSN (p))
7839 if (GET_CODE (p) == CODE_LABEL)
7841 if (GET_CODE (p) != JUMP_INSN)
7844 before_comparison = get_condition_for_loop (loop, p);
7845 if (before_comparison
7846 && XEXP (before_comparison, 0) == bl->biv->dest_reg
7847 && GET_CODE (before_comparison) == LT
7848 && XEXP (before_comparison, 1) == const0_rtx
7849 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
7850 && INTVAL (bl->biv->add_val) == -1)
7852 if (! find_reg_note (jump, REG_NONNEG, NULL_RTX))
7854 = gen_rtx_EXPR_LIST (REG_NONNEG, bl->biv->dest_reg,
7862 else if (GET_CODE (bl->biv->add_val) == CONST_INT
7863 && INTVAL (bl->biv->add_val) > 0)
7865 /* Try to change inc to dec, so can apply above optimization. */
7867 all registers modified are induction variables or invariant,
7868 all memory references have non-overlapping addresses
7869 (obviously true if only one write)
7870 allow 2 insns for the compare/jump at the end of the loop. */
7871 /* Also, we must avoid any instructions which use both the reversed
7872 biv and another biv. Such instructions will fail if the loop is
7873 reversed. We meet this condition by requiring that either
7874 no_use_except_counting is true, or else that there is only
7876 int num_nonfixed_reads = 0;
7877 /* 1 if the iteration var is used only to count iterations. */
7878 int no_use_except_counting = 0;
7879 /* 1 if the loop has no memory store, or it has a single memory store
7880 which is reversible. */
7881 int reversible_mem_store = 1;
7883 if (bl->giv_count == 0 && ! loop->exit_count)
7885 rtx bivreg = regno_reg_rtx[bl->regno];
7887 /* If there are no givs for this biv, and the only exit is the
7888 fall through at the end of the loop, then
7889 see if perhaps there are no uses except to count. */
7890 no_use_except_counting = 1;
7891 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7894 rtx set = single_set (p);
7896 if (set && GET_CODE (SET_DEST (set)) == REG
7897 && REGNO (SET_DEST (set)) == bl->regno)
7898 /* An insn that sets the biv is okay. */
7900 else if ((p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
7901 || p == prev_nonnote_insn (loop_end))
7902 && reg_mentioned_p (bivreg, PATTERN (p)))
7904 /* If either of these insns uses the biv and sets a pseudo
7905 that has more than one usage, then the biv has uses
7906 other than counting since it's used to derive a value
7907 that is used more than one time. */
7908 int note_set_pseudo_multiple_uses_retval = 0;
7909 note_stores (PATTERN (p), note_set_pseudo_multiple_uses,
7910 ¬e_set_pseudo_multiple_uses_retval);
7911 if (note_set_pseudo_multiple_uses_retval)
7913 no_use_except_counting = 0;
7917 else if (reg_mentioned_p (bivreg, PATTERN (p)))
7919 no_use_except_counting = 0;
7925 if (no_use_except_counting)
7926 /* No need to worry about MEMs. */
7928 else if (num_mem_sets <= 1)
7930 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
7932 num_nonfixed_reads += count_nonfixed_reads (loop, PATTERN (p));
7934 /* If the loop has a single store, and the destination address is
7935 invariant, then we can't reverse the loop, because this address
7936 might then have the wrong value at loop exit.
7937 This would work if the source was invariant also, however, in that
7938 case, the insn should have been moved out of the loop. */
7940 if (num_mem_sets == 1)
7942 struct induction *v;
7944 reversible_mem_store
7945 = (! unknown_address_altered
7946 && ! unknown_constant_address_altered
7947 && ! loop_invariant_p (loop,
7948 XEXP (XEXP (loop_store_mems, 0),
7951 /* If the store depends on a register that is set after the
7952 store, it depends on the initial value, and is thus not
7954 for (v = bl->giv; reversible_mem_store && v; v = v->next_iv)
7956 if (v->giv_type == DEST_REG
7957 && reg_mentioned_p (v->dest_reg,
7958 PATTERN (first_loop_store_insn))
7959 && loop_insn_first_p (first_loop_store_insn, v->insn))
7960 reversible_mem_store = 0;
7967 /* This code only acts for innermost loops. Also it simplifies
7968 the memory address check by only reversing loops with
7969 zero or one memory access.
7970 Two memory accesses could involve parts of the same array,
7971 and that can't be reversed.
7972 If the biv is used only for counting, than we don't need to worry
7973 about all these things. */
7975 if ((num_nonfixed_reads <= 1
7976 && ! loop_info->has_call
7977 && ! loop_info->has_volatile
7978 && reversible_mem_store
7979 && (bl->giv_count + bl->biv_count + num_mem_sets
7980 + num_movables + compare_and_branch == insn_count)
7981 && (bl == loop_iv_list && bl->next == 0))
7982 || no_use_except_counting)
7986 /* Loop can be reversed. */
7987 if (loop_dump_stream)
7988 fprintf (loop_dump_stream, "Can reverse loop\n");
7990 /* Now check other conditions:
7992 The increment must be a constant, as must the initial value,
7993 and the comparison code must be LT.
7995 This test can probably be improved since +/- 1 in the constant
7996 can be obtained by changing LT to LE and vice versa; this is
8000 /* for constants, LE gets turned into LT */
8001 && (GET_CODE (comparison) == LT
8002 || (GET_CODE (comparison) == LE
8003 && no_use_except_counting)))
8005 HOST_WIDE_INT add_val, add_adjust, comparison_val = 0;
8006 rtx initial_value, comparison_value;
8008 enum rtx_code cmp_code;
8009 int comparison_const_width;
8010 unsigned HOST_WIDE_INT comparison_sign_mask;
8012 add_val = INTVAL (bl->biv->add_val);
8013 comparison_value = XEXP (comparison, 1);
8014 if (GET_MODE (comparison_value) == VOIDmode)
8015 comparison_const_width
8016 = GET_MODE_BITSIZE (GET_MODE (XEXP (comparison, 0)));
8018 comparison_const_width
8019 = GET_MODE_BITSIZE (GET_MODE (comparison_value));
8020 if (comparison_const_width > HOST_BITS_PER_WIDE_INT)
8021 comparison_const_width = HOST_BITS_PER_WIDE_INT;
8022 comparison_sign_mask
8023 = (unsigned HOST_WIDE_INT) 1 << (comparison_const_width - 1);
8025 /* If the comparison value is not a loop invariant, then we
8026 can not reverse this loop.
8028 ??? If the insns which initialize the comparison value as
8029 a whole compute an invariant result, then we could move
8030 them out of the loop and proceed with loop reversal. */
8031 if (! loop_invariant_p (loop, comparison_value))
8034 if (GET_CODE (comparison_value) == CONST_INT)
8035 comparison_val = INTVAL (comparison_value);
8036 initial_value = bl->initial_value;
8038 /* Normalize the initial value if it is an integer and
8039 has no other use except as a counter. This will allow
8040 a few more loops to be reversed. */
8041 if (no_use_except_counting
8042 && GET_CODE (comparison_value) == CONST_INT
8043 && GET_CODE (initial_value) == CONST_INT)
8045 comparison_val = comparison_val - INTVAL (bl->initial_value);
8046 /* The code below requires comparison_val to be a multiple
8047 of add_val in order to do the loop reversal, so
8048 round up comparison_val to a multiple of add_val.
8049 Since comparison_value is constant, we know that the
8050 current comparison code is LT. */
8051 comparison_val = comparison_val + add_val - 1;
8053 -= (unsigned HOST_WIDE_INT) comparison_val % add_val;
8054 /* We postpone overflow checks for COMPARISON_VAL here;
8055 even if there is an overflow, we might still be able to
8056 reverse the loop, if converting the loop exit test to
8058 initial_value = const0_rtx;
8061 /* First check if we can do a vanilla loop reversal. */
8062 if (initial_value == const0_rtx
8063 /* If we have a decrement_and_branch_on_count,
8064 prefer the NE test, since this will allow that
8065 instruction to be generated. Note that we must
8066 use a vanilla loop reversal if the biv is used to
8067 calculate a giv or has a non-counting use. */
8068 #if ! defined (HAVE_decrement_and_branch_until_zero) \
8069 && defined (HAVE_decrement_and_branch_on_count)
8070 && (! (add_val == 1 && loop->vtop
8071 && (bl->biv_count == 0
8072 || no_use_except_counting)))
8074 && GET_CODE (comparison_value) == CONST_INT
8075 /* Now do postponed overflow checks on COMPARISON_VAL. */
8076 && ! (((comparison_val - add_val) ^ INTVAL (comparison_value))
8077 & comparison_sign_mask))
8079 /* Register will always be nonnegative, with value
8080 0 on last iteration */
8081 add_adjust = add_val;
8085 else if (add_val == 1 && loop->vtop
8086 && (bl->biv_count == 0
8087 || no_use_except_counting))
8095 if (GET_CODE (comparison) == LE)
8096 add_adjust -= add_val;
8098 /* If the initial value is not zero, or if the comparison
8099 value is not an exact multiple of the increment, then we
8100 can not reverse this loop. */
8101 if (initial_value == const0_rtx
8102 && GET_CODE (comparison_value) == CONST_INT)
8104 if (((unsigned HOST_WIDE_INT) comparison_val % add_val) != 0)
8109 if (! no_use_except_counting || add_val != 1)
8113 final_value = comparison_value;
8115 /* Reset these in case we normalized the initial value
8116 and comparison value above. */
8117 if (GET_CODE (comparison_value) == CONST_INT
8118 && GET_CODE (initial_value) == CONST_INT)
8120 comparison_value = GEN_INT (comparison_val);
8122 = GEN_INT (comparison_val + INTVAL (bl->initial_value));
8124 bl->initial_value = initial_value;
8126 /* Save some info needed to produce the new insns. */
8127 reg = bl->biv->dest_reg;
8128 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
8129 if (jump_label == pc_rtx)
8130 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 2);
8131 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
8133 /* Set start_value; if this is not a CONST_INT, we need
8135 Initialize biv to start_value before loop start.
8136 The old initializing insn will be deleted as a
8137 dead store by flow.c. */
8138 if (initial_value == const0_rtx
8139 && GET_CODE (comparison_value) == CONST_INT)
8141 start_value = GEN_INT (comparison_val - add_adjust);
8142 emit_insn_before (gen_move_insn (reg, start_value),
8145 else if (GET_CODE (initial_value) == CONST_INT)
8147 rtx offset = GEN_INT (-INTVAL (initial_value) - add_adjust);
8148 enum machine_mode mode = GET_MODE (reg);
8149 enum insn_code icode
8150 = add_optab->handlers[(int) mode].insn_code;
8152 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8153 || ! ((*insn_data[icode].operand[1].predicate)
8154 (comparison_value, mode))
8155 || ! ((*insn_data[icode].operand[2].predicate)
8159 = gen_rtx_PLUS (mode, comparison_value, offset);
8160 emit_insn_before ((GEN_FCN (icode)
8161 (reg, comparison_value, offset)),
8163 if (GET_CODE (comparison) == LE)
8164 final_value = gen_rtx_PLUS (mode, comparison_value,
8167 else if (! add_adjust)
8169 enum machine_mode mode = GET_MODE (reg);
8170 enum insn_code icode
8171 = sub_optab->handlers[(int) mode].insn_code;
8172 if (! (*insn_data[icode].operand[0].predicate) (reg, mode)
8173 || ! ((*insn_data[icode].operand[1].predicate)
8174 (comparison_value, mode))
8175 || ! ((*insn_data[icode].operand[2].predicate)
8176 (initial_value, mode)))
8179 = gen_rtx_MINUS (mode, comparison_value, initial_value);
8180 emit_insn_before ((GEN_FCN (icode)
8181 (reg, comparison_value, initial_value)),
8185 /* We could handle the other cases too, but it'll be
8186 better to have a testcase first. */
8189 /* We may not have a single insn which can increment a reg, so
8190 create a sequence to hold all the insns from expand_inc. */
8192 expand_inc (reg, new_add_val);
8193 tem = gen_sequence ();
8196 p = emit_insn_before (tem, bl->biv->insn);
8197 delete_insn (bl->biv->insn);
8199 /* Update biv info to reflect its new status. */
8201 bl->initial_value = start_value;
8202 bl->biv->add_val = new_add_val;
8204 /* Update loop info. */
8205 loop_info->initial_value = reg;
8206 loop_info->initial_equiv_value = reg;
8207 loop_info->final_value = const0_rtx;
8208 loop_info->final_equiv_value = const0_rtx;
8209 loop_info->comparison_value = const0_rtx;
8210 loop_info->comparison_code = cmp_code;
8211 loop_info->increment = new_add_val;
8213 /* Inc LABEL_NUSES so that delete_insn will
8214 not delete the label. */
8215 LABEL_NUSES (XEXP (jump_label, 0)) ++;
8217 /* Emit an insn after the end of the loop to set the biv's
8218 proper exit value if it is used anywhere outside the loop. */
8219 if ((REGNO_LAST_UID (bl->regno) != INSN_UID (first_compare))
8221 || REGNO_FIRST_UID (bl->regno) != INSN_UID (bl->init_insn))
8222 emit_insn_after (gen_move_insn (reg, final_value),
8225 /* Delete compare/branch at end of loop. */
8226 delete_insn (PREV_INSN (loop_end));
8227 if (compare_and_branch == 2)
8228 delete_insn (first_compare);
8230 /* Add new compare/branch insn at end of loop. */
8232 emit_cmp_and_jump_insns (reg, const0_rtx, cmp_code, NULL_RTX,
8233 GET_MODE (reg), 0, 0,
8234 XEXP (jump_label, 0));
8235 tem = gen_sequence ();
8237 emit_jump_insn_before (tem, loop_end);
8239 for (tem = PREV_INSN (loop_end);
8240 tem && GET_CODE (tem) != JUMP_INSN;
8241 tem = PREV_INSN (tem))
8245 JUMP_LABEL (tem) = XEXP (jump_label, 0);
8251 /* Increment of LABEL_NUSES done above. */
8252 /* Register is now always nonnegative,
8253 so add REG_NONNEG note to the branch. */
8254 REG_NOTES (tem) = gen_rtx_EXPR_LIST (REG_NONNEG, reg,
8260 /* No insn may reference both the reversed and another biv or it
8261 will fail (see comment near the top of the loop reversal
8263 Earlier on, we have verified that the biv has no use except
8264 counting, or it is the only biv in this function.
8265 However, the code that computes no_use_except_counting does
8266 not verify reg notes. It's possible to have an insn that
8267 references another biv, and has a REG_EQUAL note with an
8268 expression based on the reversed biv. To avoid this case,
8269 remove all REG_EQUAL notes based on the reversed biv
8271 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8275 rtx set = single_set (p);
8276 /* If this is a set of a GIV based on the reversed biv, any
8277 REG_EQUAL notes should still be correct. */
8279 || GET_CODE (SET_DEST (set)) != REG
8280 || (size_t) REGNO (SET_DEST (set)) >= reg_iv_type->num_elements
8281 || REG_IV_TYPE (REGNO (SET_DEST (set))) != GENERAL_INDUCT
8282 || REG_IV_INFO (REGNO (SET_DEST (set)))->src_reg != bl->biv->src_reg)
8283 for (pnote = ®_NOTES (p); *pnote;)
8285 if (REG_NOTE_KIND (*pnote) == REG_EQUAL
8286 && reg_mentioned_p (regno_reg_rtx[bl->regno],
8288 *pnote = XEXP (*pnote, 1);
8290 pnote = &XEXP (*pnote, 1);
8294 /* Mark that this biv has been reversed. Each giv which depends
8295 on this biv, and which is also live past the end of the loop
8296 will have to be fixed up. */
8300 if (loop_dump_stream)
8302 fprintf (loop_dump_stream, "Reversed loop");
8304 fprintf (loop_dump_stream, " and added reg_nonneg\n");
8306 fprintf (loop_dump_stream, "\n");
8317 /* Verify whether the biv BL appears to be eliminable,
8318 based on the insns in the loop that refer to it.
8320 If ELIMINATE_P is non-zero, actually do the elimination.
8322 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
8323 determine whether invariant insns should be placed inside or at the
8324 start of the loop. */
8327 maybe_eliminate_biv (loop, bl, eliminate_p, threshold, insn_count)
8328 const struct loop *loop;
8329 struct iv_class *bl;
8331 int threshold, insn_count;
8333 rtx reg = bl->biv->dest_reg;
8334 rtx loop_start = loop->start;
8335 rtx loop_end = loop->end;
8338 /* Scan all insns in the loop, stopping if we find one that uses the
8339 biv in a way that we cannot eliminate. */
8341 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
8343 enum rtx_code code = GET_CODE (p);
8344 rtx where = threshold >= insn_count ? loop_start : p;
8346 /* If this is a libcall that sets a giv, skip ahead to its end. */
8347 if (GET_RTX_CLASS (code) == 'i')
8349 rtx note = find_reg_note (p, REG_LIBCALL, NULL_RTX);
8353 rtx last = XEXP (note, 0);
8354 rtx set = single_set (last);
8356 if (set && GET_CODE (SET_DEST (set)) == REG)
8358 unsigned int regno = REGNO (SET_DEST (set));
8360 if (regno < max_reg_before_loop
8361 && REG_IV_TYPE (regno) == GENERAL_INDUCT
8362 && REG_IV_INFO (regno)->src_reg == bl->biv->src_reg)
8367 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
8368 && reg_mentioned_p (reg, PATTERN (p))
8369 && ! maybe_eliminate_biv_1 (loop, PATTERN (p), p, bl,
8370 eliminate_p, where))
8372 if (loop_dump_stream)
8373 fprintf (loop_dump_stream,
8374 "Cannot eliminate biv %d: biv used in insn %d.\n",
8375 bl->regno, INSN_UID (p));
8382 if (loop_dump_stream)
8383 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
8384 bl->regno, eliminate_p ? "was" : "can be");
8391 /* INSN and REFERENCE are instructions in the same insn chain.
8392 Return non-zero if INSN is first. */
8395 loop_insn_first_p (insn, reference)
8396 rtx insn, reference;
8400 for (p = insn, q = reference;;)
8402 /* Start with test for not first so that INSN == REFERENCE yields not
8404 if (q == insn || ! p)
8406 if (p == reference || ! q)
8409 /* Either of P or Q might be a NOTE. Notes have the same LUID as the
8410 previous insn, hence the <= comparison below does not work if
8412 if (INSN_UID (p) < max_uid_for_loop
8413 && INSN_UID (q) < max_uid_for_loop
8414 && GET_CODE (p) != NOTE)
8415 return INSN_LUID (p) <= INSN_LUID (q);
8417 if (INSN_UID (p) >= max_uid_for_loop
8418 || GET_CODE (p) == NOTE)
8420 if (INSN_UID (q) >= max_uid_for_loop)
8425 /* We are trying to eliminate BIV in INSN using GIV. Return non-zero if
8426 the offset that we have to take into account due to auto-increment /
8427 div derivation is zero. */
8429 biv_elimination_giv_has_0_offset (biv, giv, insn)
8430 struct induction *biv, *giv;
8433 /* If the giv V had the auto-inc address optimization applied
8434 to it, and INSN occurs between the giv insn and the biv
8435 insn, then we'd have to adjust the value used here.
8436 This is rare, so we don't bother to make this possible. */
8437 if (giv->auto_inc_opt
8438 && ((loop_insn_first_p (giv->insn, insn)
8439 && loop_insn_first_p (insn, biv->insn))
8440 || (loop_insn_first_p (biv->insn, insn)
8441 && loop_insn_first_p (insn, giv->insn))))
8444 /* If the giv V was derived from another giv, and INSN does
8445 not occur between the giv insn and the biv insn, then we'd
8446 have to adjust the value used here. This is rare, so we don't
8447 bother to make this possible. */
8448 if (giv->derived_from
8449 && ! (giv->always_executed
8450 && loop_insn_first_p (giv->insn, insn)
8451 && loop_insn_first_p (insn, biv->insn)))
8454 && giv->same->derived_from
8455 && ! (giv->same->always_executed
8456 && loop_insn_first_p (giv->same->insn, insn)
8457 && loop_insn_first_p (insn, biv->insn)))
8463 /* If BL appears in X (part of the pattern of INSN), see if we can
8464 eliminate its use. If so, return 1. If not, return 0.
8466 If BIV does not appear in X, return 1.
8468 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
8469 where extra insns should be added. Depending on how many items have been
8470 moved out of the loop, it will either be before INSN or at the start of
8474 maybe_eliminate_biv_1 (loop, x, insn, bl, eliminate_p, where)
8475 const struct loop *loop;
8477 struct iv_class *bl;
8481 enum rtx_code code = GET_CODE (x);
8482 rtx reg = bl->biv->dest_reg;
8483 enum machine_mode mode = GET_MODE (reg);
8484 struct induction *v;
8496 /* If we haven't already been able to do something with this BIV,
8497 we can't eliminate it. */
8503 /* If this sets the BIV, it is not a problem. */
8504 if (SET_DEST (x) == reg)
8507 /* If this is an insn that defines a giv, it is also ok because
8508 it will go away when the giv is reduced. */
8509 for (v = bl->giv; v; v = v->next_iv)
8510 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
8514 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
8516 /* Can replace with any giv that was reduced and
8517 that has (MULT_VAL != 0) and (ADD_VAL == 0).
8518 Require a constant for MULT_VAL, so we know it's nonzero.
8519 ??? We disable this optimization to avoid potential
8522 for (v = bl->giv; v; v = v->next_iv)
8523 if (GET_CODE (v->mult_val) == CONST_INT && v->mult_val != const0_rtx
8524 && v->add_val == const0_rtx
8525 && ! v->ignore && ! v->maybe_dead && v->always_computable
8529 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8535 /* If the giv has the opposite direction of change,
8536 then reverse the comparison. */
8537 if (INTVAL (v->mult_val) < 0)
8538 new = gen_rtx_COMPARE (GET_MODE (v->new_reg),
8539 const0_rtx, v->new_reg);
8543 /* We can probably test that giv's reduced reg. */
8544 if (validate_change (insn, &SET_SRC (x), new, 0))
8548 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
8549 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
8550 Require a constant for MULT_VAL, so we know it's nonzero.
8551 ??? Do this only if ADD_VAL is a pointer to avoid a potential
8552 overflow problem. */
8554 for (v = bl->giv; v; v = v->next_iv)
8555 if (GET_CODE (v->mult_val) == CONST_INT
8556 && v->mult_val != const0_rtx
8557 && ! v->ignore && ! v->maybe_dead && v->always_computable
8559 && (GET_CODE (v->add_val) == SYMBOL_REF
8560 || GET_CODE (v->add_val) == LABEL_REF
8561 || GET_CODE (v->add_val) == CONST
8562 || (GET_CODE (v->add_val) == REG
8563 && REGNO_POINTER_FLAG (REGNO (v->add_val)))))
8565 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8571 /* If the giv has the opposite direction of change,
8572 then reverse the comparison. */
8573 if (INTVAL (v->mult_val) < 0)
8574 new = gen_rtx_COMPARE (VOIDmode, copy_rtx (v->add_val),
8577 new = gen_rtx_COMPARE (VOIDmode, v->new_reg,
8578 copy_rtx (v->add_val));
8580 /* Replace biv with the giv's reduced register. */
8581 update_reg_last_use (v->add_val, insn);
8582 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8585 /* Insn doesn't support that constant or invariant. Copy it
8586 into a register (it will be a loop invariant.) */
8587 tem = gen_reg_rtx (GET_MODE (v->new_reg));
8589 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
8592 /* Substitute the new register for its invariant value in
8593 the compare expression. */
8594 XEXP (new, (INTVAL (v->mult_val) < 0) ? 0 : 1) = tem;
8595 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
8604 case GT: case GE: case GTU: case GEU:
8605 case LT: case LE: case LTU: case LEU:
8606 /* See if either argument is the biv. */
8607 if (XEXP (x, 0) == reg)
8608 arg = XEXP (x, 1), arg_operand = 1;
8609 else if (XEXP (x, 1) == reg)
8610 arg = XEXP (x, 0), arg_operand = 0;
8614 if (CONSTANT_P (arg))
8616 /* First try to replace with any giv that has constant positive
8617 mult_val and constant add_val. We might be able to support
8618 negative mult_val, but it seems complex to do it in general. */
8620 for (v = bl->giv; v; v = v->next_iv)
8621 if (GET_CODE (v->mult_val) == CONST_INT
8622 && INTVAL (v->mult_val) > 0
8623 && (GET_CODE (v->add_val) == SYMBOL_REF
8624 || GET_CODE (v->add_val) == LABEL_REF
8625 || GET_CODE (v->add_val) == CONST
8626 || (GET_CODE (v->add_val) == REG
8627 && REGNO_POINTER_FLAG (REGNO (v->add_val))))
8628 && ! v->ignore && ! v->maybe_dead && v->always_computable
8631 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8637 /* Replace biv with the giv's reduced reg. */
8638 validate_change (insn, &XEXP (x, 1 - arg_operand), v->new_reg, 1);
8640 /* If all constants are actually constant integers and
8641 the derived constant can be directly placed in the COMPARE,
8643 if (GET_CODE (arg) == CONST_INT
8644 && GET_CODE (v->mult_val) == CONST_INT
8645 && GET_CODE (v->add_val) == CONST_INT)
8647 validate_change (insn, &XEXP (x, arg_operand),
8648 GEN_INT (INTVAL (arg)
8649 * INTVAL (v->mult_val)
8650 + INTVAL (v->add_val)), 1);
8654 /* Otherwise, load it into a register. */
8655 tem = gen_reg_rtx (mode);
8656 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8657 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8659 if (apply_change_group ())
8663 /* Look for giv with positive constant mult_val and nonconst add_val.
8664 Insert insns to calculate new compare value.
8665 ??? Turn this off due to possible overflow. */
8667 for (v = bl->giv; v; v = v->next_iv)
8668 if (GET_CODE (v->mult_val) == CONST_INT
8669 && INTVAL (v->mult_val) > 0
8670 && ! v->ignore && ! v->maybe_dead && v->always_computable
8676 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8682 tem = gen_reg_rtx (mode);
8684 /* Replace biv with giv's reduced register. */
8685 validate_change (insn, &XEXP (x, 1 - arg_operand),
8688 /* Compute value to compare against. */
8689 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
8690 /* Use it in this insn. */
8691 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8692 if (apply_change_group ())
8696 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
8698 if (loop_invariant_p (loop, arg) == 1)
8700 /* Look for giv with constant positive mult_val and nonconst
8701 add_val. Insert insns to compute new compare value.
8702 ??? Turn this off due to possible overflow. */
8704 for (v = bl->giv; v; v = v->next_iv)
8705 if (GET_CODE (v->mult_val) == CONST_INT && INTVAL (v->mult_val) > 0
8706 && ! v->ignore && ! v->maybe_dead && v->always_computable
8712 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8718 tem = gen_reg_rtx (mode);
8720 /* Replace biv with giv's reduced register. */
8721 validate_change (insn, &XEXP (x, 1 - arg_operand),
8724 /* Compute value to compare against. */
8725 emit_iv_add_mult (arg, v->mult_val, v->add_val,
8727 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
8728 if (apply_change_group ())
8733 /* This code has problems. Basically, you can't know when
8734 seeing if we will eliminate BL, whether a particular giv
8735 of ARG will be reduced. If it isn't going to be reduced,
8736 we can't eliminate BL. We can try forcing it to be reduced,
8737 but that can generate poor code.
8739 The problem is that the benefit of reducing TV, below should
8740 be increased if BL can actually be eliminated, but this means
8741 we might have to do a topological sort of the order in which
8742 we try to process biv. It doesn't seem worthwhile to do
8743 this sort of thing now. */
8746 /* Otherwise the reg compared with had better be a biv. */
8747 if (GET_CODE (arg) != REG
8748 || REG_IV_TYPE (REGNO (arg)) != BASIC_INDUCT)
8751 /* Look for a pair of givs, one for each biv,
8752 with identical coefficients. */
8753 for (v = bl->giv; v; v = v->next_iv)
8755 struct induction *tv;
8757 if (v->ignore || v->maybe_dead || v->mode != mode)
8760 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
8761 if (! tv->ignore && ! tv->maybe_dead
8762 && rtx_equal_p (tv->mult_val, v->mult_val)
8763 && rtx_equal_p (tv->add_val, v->add_val)
8764 && tv->mode == mode)
8766 if (! biv_elimination_giv_has_0_offset (bl->biv, v, insn))
8772 /* Replace biv with its giv's reduced reg. */
8773 XEXP (x, 1 - arg_operand) = v->new_reg;
8774 /* Replace other operand with the other giv's
8776 XEXP (x, arg_operand) = tv->new_reg;
8783 /* If we get here, the biv can't be eliminated. */
8787 /* If this address is a DEST_ADDR giv, it doesn't matter if the
8788 biv is used in it, since it will be replaced. */
8789 for (v = bl->giv; v; v = v->next_iv)
8790 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
8798 /* See if any subexpression fails elimination. */
8799 fmt = GET_RTX_FORMAT (code);
8800 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
8805 if (! maybe_eliminate_biv_1 (loop, XEXP (x, i), insn, bl,
8806 eliminate_p, where))
8811 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8812 if (! maybe_eliminate_biv_1 (loop, XVECEXP (x, i, j), insn, bl,
8813 eliminate_p, where))
8822 /* Return nonzero if the last use of REG
8823 is in an insn following INSN in the same basic block. */
8826 last_use_this_basic_block (reg, insn)
8832 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
8835 if (REGNO_LAST_UID (REGNO (reg)) == INSN_UID (n))
8841 /* Called via `note_stores' to record the initial value of a biv. Here we
8842 just record the location of the set and process it later. */
8845 record_initial (dest, set, data)
8848 void *data ATTRIBUTE_UNUSED;
8850 struct iv_class *bl;
8852 if (GET_CODE (dest) != REG
8853 || REGNO (dest) >= max_reg_before_loop
8854 || REG_IV_TYPE (REGNO (dest)) != BASIC_INDUCT)
8857 bl = reg_biv_class[REGNO (dest)];
8859 /* If this is the first set found, record it. */
8860 if (bl->init_insn == 0)
8862 bl->init_insn = note_insn;
8867 /* If any of the registers in X are "old" and currently have a last use earlier
8868 than INSN, update them to have a last use of INSN. Their actual last use
8869 will be the previous insn but it will not have a valid uid_luid so we can't
8873 update_reg_last_use (x, insn)
8877 /* Check for the case where INSN does not have a valid luid. In this case,
8878 there is no need to modify the regno_last_uid, as this can only happen
8879 when code is inserted after the loop_end to set a pseudo's final value,
8880 and hence this insn will never be the last use of x. */
8881 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
8882 && INSN_UID (insn) < max_uid_for_loop
8883 && uid_luid[REGNO_LAST_UID (REGNO (x))] < uid_luid[INSN_UID (insn)])
8884 REGNO_LAST_UID (REGNO (x)) = INSN_UID (insn);
8888 register const char *fmt = GET_RTX_FORMAT (GET_CODE (x));
8889 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
8892 update_reg_last_use (XEXP (x, i), insn);
8893 else if (fmt[i] == 'E')
8894 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
8895 update_reg_last_use (XVECEXP (x, i, j), insn);
8900 /* Given an insn INSN and condition COND, return the condition in a
8901 canonical form to simplify testing by callers. Specifically:
8903 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
8904 (2) Both operands will be machine operands; (cc0) will have been replaced.
8905 (3) If an operand is a constant, it will be the second operand.
8906 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
8907 for GE, GEU, and LEU.
8909 If the condition cannot be understood, or is an inequality floating-point
8910 comparison which needs to be reversed, 0 will be returned.
8912 If REVERSE is non-zero, then reverse the condition prior to canonizing it.
8914 If EARLIEST is non-zero, it is a pointer to a place where the earliest
8915 insn used in locating the condition was found. If a replacement test
8916 of the condition is desired, it should be placed in front of that
8917 insn and we will be sure that the inputs are still valid.
8919 If WANT_REG is non-zero, we wish the condition to be relative to that
8920 register, if possible. Therefore, do not canonicalize the condition
8924 canonicalize_condition (insn, cond, reverse, earliest, want_reg)
8936 int reverse_code = 0;
8937 int did_reverse_condition = 0;
8938 enum machine_mode mode;
8940 code = GET_CODE (cond);
8941 mode = GET_MODE (cond);
8942 op0 = XEXP (cond, 0);
8943 op1 = XEXP (cond, 1);
8947 code = reverse_condition (code);
8948 did_reverse_condition ^= 1;
8954 /* If we are comparing a register with zero, see if the register is set
8955 in the previous insn to a COMPARE or a comparison operation. Perform
8956 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
8959 while (GET_RTX_CLASS (code) == '<'
8960 && op1 == CONST0_RTX (GET_MODE (op0))
8963 /* Set non-zero when we find something of interest. */
8967 /* If comparison with cc0, import actual comparison from compare
8971 if ((prev = prev_nonnote_insn (prev)) == 0
8972 || GET_CODE (prev) != INSN
8973 || (set = single_set (prev)) == 0
8974 || SET_DEST (set) != cc0_rtx)
8977 op0 = SET_SRC (set);
8978 op1 = CONST0_RTX (GET_MODE (op0));
8984 /* If this is a COMPARE, pick up the two things being compared. */
8985 if (GET_CODE (op0) == COMPARE)
8987 op1 = XEXP (op0, 1);
8988 op0 = XEXP (op0, 0);
8991 else if (GET_CODE (op0) != REG)
8994 /* Go back to the previous insn. Stop if it is not an INSN. We also
8995 stop if it isn't a single set or if it has a REG_INC note because
8996 we don't want to bother dealing with it. */
8998 if ((prev = prev_nonnote_insn (prev)) == 0
8999 || GET_CODE (prev) != INSN
9000 || FIND_REG_INC_NOTE (prev, 0)
9001 || (set = single_set (prev)) == 0)
9004 /* If this is setting OP0, get what it sets it to if it looks
9006 if (rtx_equal_p (SET_DEST (set), op0))
9008 enum machine_mode inner_mode = GET_MODE (SET_DEST (set));
9010 /* ??? We may not combine comparisons done in a CCmode with
9011 comparisons not done in a CCmode. This is to aid targets
9012 like Alpha that have an IEEE compliant EQ instruction, and
9013 a non-IEEE compliant BEQ instruction. The use of CCmode is
9014 actually artificial, simply to prevent the combination, but
9015 should not affect other platforms.
9017 However, we must allow VOIDmode comparisons to match either
9018 CCmode or non-CCmode comparison, because some ports have
9019 modeless comparisons inside branch patterns.
9021 ??? This mode check should perhaps look more like the mode check
9022 in simplify_comparison in combine. */
9024 if ((GET_CODE (SET_SRC (set)) == COMPARE
9027 && GET_MODE_CLASS (inner_mode) == MODE_INT
9028 && (GET_MODE_BITSIZE (inner_mode)
9029 <= HOST_BITS_PER_WIDE_INT)
9030 && (STORE_FLAG_VALUE
9031 & ((HOST_WIDE_INT) 1
9032 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9033 #ifdef FLOAT_STORE_FLAG_VALUE
9035 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9036 && (REAL_VALUE_NEGATIVE
9037 (FLOAT_STORE_FLAG_VALUE (inner_mode))))
9040 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'))
9041 && (((GET_MODE_CLASS (mode) == MODE_CC)
9042 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9043 || mode == VOIDmode || inner_mode == VOIDmode))
9045 else if (((code == EQ
9047 && (GET_MODE_BITSIZE (inner_mode)
9048 <= HOST_BITS_PER_WIDE_INT)
9049 && GET_MODE_CLASS (inner_mode) == MODE_INT
9050 && (STORE_FLAG_VALUE
9051 & ((HOST_WIDE_INT) 1
9052 << (GET_MODE_BITSIZE (inner_mode) - 1))))
9053 #ifdef FLOAT_STORE_FLAG_VALUE
9055 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
9056 && (REAL_VALUE_NEGATIVE
9057 (FLOAT_STORE_FLAG_VALUE (inner_mode))))
9060 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<'
9061 && (((GET_MODE_CLASS (mode) == MODE_CC)
9062 == (GET_MODE_CLASS (inner_mode) == MODE_CC))
9063 || mode == VOIDmode || inner_mode == VOIDmode))
9066 /* We might have reversed a LT to get a GE here. But this wasn't
9067 actually the comparison of data, so we don't flag that we
9068 have had to reverse the condition. */
9069 did_reverse_condition ^= 1;
9077 else if (reg_set_p (op0, prev))
9078 /* If this sets OP0, but not directly, we have to give up. */
9083 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
9084 code = GET_CODE (x);
9087 code = reverse_condition (code);
9088 if (code == UNKNOWN)
9090 did_reverse_condition ^= 1;
9094 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
9100 /* If constant is first, put it last. */
9101 if (CONSTANT_P (op0))
9102 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
9104 /* If OP0 is the result of a comparison, we weren't able to find what
9105 was really being compared, so fail. */
9106 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
9109 /* Canonicalize any ordered comparison with integers involving equality
9110 if we can do computations in the relevant mode and we do not
9113 if (GET_CODE (op1) == CONST_INT
9114 && GET_MODE (op0) != VOIDmode
9115 && GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
9117 HOST_WIDE_INT const_val = INTVAL (op1);
9118 unsigned HOST_WIDE_INT uconst_val = const_val;
9119 unsigned HOST_WIDE_INT max_val
9120 = (unsigned HOST_WIDE_INT) GET_MODE_MASK (GET_MODE (op0));
9125 if ((unsigned HOST_WIDE_INT) const_val != max_val >> 1)
9126 code = LT, op1 = GEN_INT (const_val + 1);
9129 /* When cross-compiling, const_val might be sign-extended from
9130 BITS_PER_WORD to HOST_BITS_PER_WIDE_INT */
9132 if ((HOST_WIDE_INT) (const_val & max_val)
9133 != (((HOST_WIDE_INT) 1
9134 << (GET_MODE_BITSIZE (GET_MODE (op0)) - 1))))
9135 code = GT, op1 = GEN_INT (const_val - 1);
9139 if (uconst_val < max_val)
9140 code = LTU, op1 = GEN_INT (uconst_val + 1);
9144 if (uconst_val != 0)
9145 code = GTU, op1 = GEN_INT (uconst_val - 1);
9153 /* If this was floating-point and we reversed anything other than an
9154 EQ or NE or (UN)ORDERED, return zero. */
9155 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
9156 && did_reverse_condition
9157 && code != NE && code != EQ && code != UNORDERED && code != ORDERED
9159 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
9163 /* Never return CC0; return zero instead. */
9168 return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);
9171 /* Given a jump insn JUMP, return the condition that will cause it to branch
9172 to its JUMP_LABEL. If the condition cannot be understood, or is an
9173 inequality floating-point comparison which needs to be reversed, 0 will
9176 If EARLIEST is non-zero, it is a pointer to a place where the earliest
9177 insn used in locating the condition was found. If a replacement test
9178 of the condition is desired, it should be placed in front of that
9179 insn and we will be sure that the inputs are still valid. */
9182 get_condition (jump, earliest)
9190 /* If this is not a standard conditional jump, we can't parse it. */
9191 if (GET_CODE (jump) != JUMP_INSN
9192 || ! any_condjump_p (jump))
9194 set = pc_set (jump);
9196 cond = XEXP (SET_SRC (set), 0);
9198 /* If this branches to JUMP_LABEL when the condition is false, reverse
9201 = GET_CODE (XEXP (SET_SRC (set), 2)) == LABEL_REF
9202 && XEXP (XEXP (SET_SRC (set), 2), 0) == JUMP_LABEL (jump);
9204 return canonicalize_condition (jump, cond, reverse, earliest, NULL_RTX);
9207 /* Similar to above routine, except that we also put an invariant last
9208 unless both operands are invariants. */
9211 get_condition_for_loop (loop, x)
9212 const struct loop *loop;
9215 rtx comparison = get_condition (x, NULL_PTR);
9218 || ! loop_invariant_p (loop, XEXP (comparison, 0))
9219 || loop_invariant_p (loop, XEXP (comparison, 1)))
9222 return gen_rtx_fmt_ee (swap_condition (GET_CODE (comparison)), VOIDmode,
9223 XEXP (comparison, 1), XEXP (comparison, 0));
9226 /* Scan the function and determine whether it has indirect (computed) jumps.
9228 This is taken mostly from flow.c; similar code exists elsewhere
9229 in the compiler. It may be useful to put this into rtlanal.c. */
9231 indirect_jump_in_function_p (start)
9236 for (insn = start; insn; insn = NEXT_INSN (insn))
9237 if (computed_jump_p (insn))
9243 /* Add MEM to the LOOP_MEMS array, if appropriate. See the
9244 documentation for LOOP_MEMS for the definition of `appropriate'.
9245 This function is called from prescan_loop via for_each_rtx. */
9248 insert_loop_mem (mem, data)
9250 void *data ATTRIBUTE_UNUSED;
9258 switch (GET_CODE (m))
9264 /* We're not interested in MEMs that are only clobbered. */
9268 /* We're not interested in the MEM associated with a
9269 CONST_DOUBLE, so there's no need to traverse into this. */
9273 /* We're not interested in any MEMs that only appear in notes. */
9277 /* This is not a MEM. */
9281 /* See if we've already seen this MEM. */
9282 for (i = 0; i < loop_mems_idx; ++i)
9283 if (rtx_equal_p (m, loop_mems[i].mem))
9285 if (GET_MODE (m) != GET_MODE (loop_mems[i].mem))
9286 /* The modes of the two memory accesses are different. If
9287 this happens, something tricky is going on, and we just
9288 don't optimize accesses to this MEM. */
9289 loop_mems[i].optimize = 0;
9294 /* Resize the array, if necessary. */
9295 if (loop_mems_idx == loop_mems_allocated)
9297 if (loop_mems_allocated != 0)
9298 loop_mems_allocated *= 2;
9300 loop_mems_allocated = 32;
9302 loop_mems = (loop_mem_info*)
9303 xrealloc (loop_mems,
9304 loop_mems_allocated * sizeof (loop_mem_info));
9307 /* Actually insert the MEM. */
9308 loop_mems[loop_mems_idx].mem = m;
9309 /* We can't hoist this MEM out of the loop if it's a BLKmode MEM
9310 because we can't put it in a register. We still store it in the
9311 table, though, so that if we see the same address later, but in a
9312 non-BLK mode, we'll not think we can optimize it at that point. */
9313 loop_mems[loop_mems_idx].optimize = (GET_MODE (m) != BLKmode);
9314 loop_mems[loop_mems_idx].reg = NULL_RTX;
9320 /* Like load_mems, but also ensures that SET_IN_LOOP,
9321 MAY_NOT_OPTIMIZE, REG_SINGLE_USAGE, and INSN_COUNT have the correct
9322 values after load_mems. */
9325 load_mems_and_recount_loop_regs_set (loop, insn_count)
9326 const struct loop *loop;
9329 int nregs = max_reg_num ();
9333 /* Recalculate set_in_loop and friends since load_mems may have
9334 created new registers. */
9335 if (max_reg_num () > nregs)
9341 nregs = max_reg_num ();
9343 if ((unsigned) nregs > set_in_loop->num_elements)
9345 /* Grow all the arrays. */
9346 VARRAY_GROW (set_in_loop, nregs);
9347 VARRAY_GROW (n_times_set, nregs);
9348 VARRAY_GROW (may_not_optimize, nregs);
9349 VARRAY_GROW (reg_single_usage, nregs);
9351 /* Clear the arrays */
9352 bzero ((char *) &set_in_loop->data, nregs * sizeof (int));
9353 bzero ((char *) &may_not_optimize->data, nregs * sizeof (char));
9354 bzero ((char *) ®_single_usage->data, nregs * sizeof (rtx));
9356 count_loop_regs_set (loop->top ? loop->top : loop->start, loop->end,
9357 may_not_optimize, reg_single_usage,
9360 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
9362 VARRAY_CHAR (may_not_optimize, i) = 1;
9363 VARRAY_INT (set_in_loop, i) = 1;
9366 #ifdef AVOID_CCMODE_COPIES
9367 /* Don't try to move insns which set CC registers if we should not
9368 create CCmode register copies. */
9369 for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--)
9370 if (GET_MODE_CLASS (GET_MODE (regno_reg_rtx[i])) == MODE_CC)
9371 VARRAY_CHAR (may_not_optimize, i) = 1;
9374 /* Set n_times_set for the new registers. */
9375 bcopy ((char *) (&set_in_loop->data.i[0] + old_nregs),
9376 (char *) (&n_times_set->data.i[0] + old_nregs),
9377 (nregs - old_nregs) * sizeof (int));
9381 /* Move MEMs into registers for the duration of the loop. */
9385 const struct loop *loop;
9387 int maybe_never = 0;
9390 rtx label = NULL_RTX;
9391 rtx end_label = NULL_RTX;
9392 /* Nonzero if the next instruction may never be executed. */
9393 int next_maybe_never = 0;
9394 int last_max_reg = max_reg_num ();
9396 if (loop_mems_idx == 0)
9399 /* Find start of the extended basic block that enters the loop. */
9400 for (p = loop->start;
9401 PREV_INSN (p) && GET_CODE (p) != CODE_LABEL;
9407 /* Build table of mems that get set to constant values before the
9409 for (; p != loop->start; p = NEXT_INSN (p))
9410 cselib_process_insn (p);
9412 /* Check to see if it's possible that some instructions in the
9413 loop are never executed. */
9414 for (p = next_insn_in_loop (loop, loop->scan_start);
9415 p != NULL_RTX && ! maybe_never;
9416 p = next_insn_in_loop (loop, p))
9418 if (GET_CODE (p) == CODE_LABEL)
9420 else if (GET_CODE (p) == JUMP_INSN
9421 /* If we enter the loop in the middle, and scan
9422 around to the beginning, don't set maybe_never
9423 for that. This must be an unconditional jump,
9424 otherwise the code at the top of the loop might
9425 never be executed. Unconditional jumps are
9426 followed a by barrier then loop end. */
9427 && ! (GET_CODE (p) == JUMP_INSN
9428 && JUMP_LABEL (p) == loop->top
9429 && NEXT_INSN (NEXT_INSN (p)) == loop->end
9430 && any_uncondjump_p (p)))
9432 if (!any_condjump_p (p))
9433 /* Something complicated. */
9436 /* If there are any more instructions in the loop, they
9437 might not be reached. */
9438 next_maybe_never = 1;
9440 else if (next_maybe_never)
9444 /* Actually move the MEMs. */
9445 for (i = 0; i < loop_mems_idx; ++i)
9447 regset_head load_copies;
9448 regset_head store_copies;
9451 rtx mem = loop_mems[i].mem;
9454 if (MEM_VOLATILE_P (mem)
9455 || loop_invariant_p (loop, XEXP (mem, 0)) != 1)
9456 /* There's no telling whether or not MEM is modified. */
9457 loop_mems[i].optimize = 0;
9459 /* Go through the MEMs written to in the loop to see if this
9460 one is aliased by one of them. */
9461 mem_list_entry = loop_store_mems;
9462 while (mem_list_entry)
9464 if (rtx_equal_p (mem, XEXP (mem_list_entry, 0)))
9466 else if (true_dependence (XEXP (mem_list_entry, 0), VOIDmode,
9469 /* MEM is indeed aliased by this store. */
9470 loop_mems[i].optimize = 0;
9473 mem_list_entry = XEXP (mem_list_entry, 1);
9476 if (flag_float_store && written
9477 && GET_MODE_CLASS (GET_MODE (mem)) == MODE_FLOAT)
9478 loop_mems[i].optimize = 0;
9480 /* If this MEM is written to, we must be sure that there
9481 are no reads from another MEM that aliases this one. */
9482 if (loop_mems[i].optimize && written)
9486 for (j = 0; j < loop_mems_idx; ++j)
9490 else if (true_dependence (mem,
9495 /* It's not safe to hoist loop_mems[i] out of
9496 the loop because writes to it might not be
9497 seen by reads from loop_mems[j]. */
9498 loop_mems[i].optimize = 0;
9504 if (maybe_never && may_trap_p (mem))
9505 /* We can't access the MEM outside the loop; it might
9506 cause a trap that wouldn't have happened otherwise. */
9507 loop_mems[i].optimize = 0;
9509 if (!loop_mems[i].optimize)
9510 /* We thought we were going to lift this MEM out of the
9511 loop, but later discovered that we could not. */
9514 INIT_REG_SET (&load_copies);
9515 INIT_REG_SET (&store_copies);
9517 /* Allocate a pseudo for this MEM. We set REG_USERVAR_P in
9518 order to keep scan_loop from moving stores to this MEM
9519 out of the loop just because this REG is neither a
9520 user-variable nor used in the loop test. */
9521 reg = gen_reg_rtx (GET_MODE (mem));
9522 REG_USERVAR_P (reg) = 1;
9523 loop_mems[i].reg = reg;
9525 /* Now, replace all references to the MEM with the
9526 corresponding pesudos. */
9528 for (p = next_insn_in_loop (loop, loop->scan_start);
9530 p = next_insn_in_loop (loop, p))
9538 set = single_set (p);
9540 /* See if this copies the mem into a register that isn't
9541 modified afterwards. We'll try to do copy propagation
9542 a little further on. */
9544 /* @@@ This test is _way_ too conservative. */
9546 && GET_CODE (SET_DEST (set)) == REG
9547 && REGNO (SET_DEST (set)) >= FIRST_PSEUDO_REGISTER
9548 && REGNO (SET_DEST (set)) < last_max_reg
9549 && VARRAY_INT (n_times_set, REGNO (SET_DEST (set))) == 1
9550 && rtx_equal_p (SET_SRC (set), mem))
9551 SET_REGNO_REG_SET (&load_copies, REGNO (SET_DEST (set)));
9553 /* See if this copies the mem from a register that isn't
9554 modified afterwards. We'll try to remove the
9555 redundant copy later on by doing a little register
9556 renaming and copy propagation. This will help
9557 to untangle things for the BIV detection code. */
9560 && GET_CODE (SET_SRC (set)) == REG
9561 && REGNO (SET_SRC (set)) >= FIRST_PSEUDO_REGISTER
9562 && REGNO (SET_SRC (set)) < last_max_reg
9563 && VARRAY_INT (n_times_set, REGNO (SET_SRC (set))) == 1
9564 && rtx_equal_p (SET_DEST (set), mem))
9565 SET_REGNO_REG_SET (&store_copies, REGNO (SET_SRC (set)));
9567 /* Replace the memory reference with the shadow register. */
9570 for_each_rtx (&p, replace_loop_mem, &ri);
9573 if (GET_CODE (p) == CODE_LABEL
9574 || GET_CODE (p) == JUMP_INSN)
9578 if (! apply_change_group ())
9579 /* We couldn't replace all occurrences of the MEM. */
9580 loop_mems[i].optimize = 0;
9583 /* Load the memory immediately before LOOP->START, which is
9584 the NOTE_LOOP_BEG. */
9585 cselib_val *e = cselib_lookup (mem, VOIDmode, 0);
9589 struct elt_loc_list *const_equiv = 0;
9593 struct elt_loc_list *equiv;
9594 struct elt_loc_list *best_equiv = 0;
9595 for (equiv = e->locs; equiv; equiv = equiv->next)
9597 if (CONSTANT_P (equiv->loc))
9598 const_equiv = equiv;
9599 else if (GET_CODE (equiv->loc) == REG
9600 /* Extending hard register lifetimes cuases crash
9601 on SRC targets. Doing so on non-SRC is
9602 probably also not good idea, since we most
9603 probably have pseudoregister equivalence as
9605 && REGNO (equiv->loc) >= FIRST_PSEUDO_REGISTER)
9608 /* Use the constant equivalence if that is cheap enough. */
9610 best_equiv = const_equiv;
9611 else if (const_equiv
9612 && (rtx_cost (const_equiv->loc, SET)
9613 <= rtx_cost (best_equiv->loc, SET)))
9615 best_equiv = const_equiv;
9619 /* If best_equiv is nonzero, we know that MEM is set to a
9620 constant or register before the loop. We will use this
9621 knowledge to initialize the shadow register with that
9622 constant or reg rather than by loading from MEM. */
9624 best = copy_rtx (best_equiv->loc);
9626 set = gen_move_insn (reg, best);
9627 set = emit_insn_before (set, loop->start);
9629 REG_NOTES (set) = gen_rtx_EXPR_LIST (REG_EQUAL,
9630 copy_rtx (const_equiv->loc),
9635 if (label == NULL_RTX)
9637 /* We must compute the former
9638 right-after-the-end label before we insert
9640 end_label = next_label (loop->end);
9641 label = gen_label_rtx ();
9642 emit_label_after (label, loop->end);
9645 /* Store the memory immediately after END, which is
9646 the NOTE_LOOP_END. */
9647 set = gen_move_insn (copy_rtx (mem), reg);
9648 emit_insn_after (set, label);
9651 if (loop_dump_stream)
9653 fprintf (loop_dump_stream, "Hoisted regno %d %s from ",
9654 REGNO (reg), (written ? "r/w" : "r/o"));
9655 print_rtl (loop_dump_stream, mem);
9656 fputc ('\n', loop_dump_stream);
9659 /* Attempt a bit of copy propagation. This helps untangle the
9660 data flow, and enables {basic,general}_induction_var to find
9662 EXECUTE_IF_SET_IN_REG_SET
9663 (&load_copies, FIRST_PSEUDO_REGISTER, j,
9665 try_copy_prop (loop, reg, j);
9667 CLEAR_REG_SET (&load_copies);
9669 EXECUTE_IF_SET_IN_REG_SET
9670 (&store_copies, FIRST_PSEUDO_REGISTER, j,
9672 try_swap_copy_prop (loop, reg, j);
9674 CLEAR_REG_SET (&store_copies);
9678 if (label != NULL_RTX)
9680 /* Now, we need to replace all references to the previous exit
9681 label with the new one. */
9686 for (p = loop->start; p != loop->end; p = NEXT_INSN (p))
9688 for_each_rtx (&p, replace_label, &rr);
9690 /* If this is a JUMP_INSN, then we also need to fix the JUMP_LABEL
9691 field. This is not handled by for_each_rtx because it doesn't
9692 handle unprinted ('0') fields. We need to update JUMP_LABEL
9693 because the immediately following unroll pass will use it.
9694 replace_label would not work anyways, because that only handles
9696 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == end_label)
9697 JUMP_LABEL (p) = label;
9704 /* For communication between note_reg_stored and its caller. */
9705 struct note_reg_stored_arg
9711 /* Called via note_stores, record in SET_SEEN whether X, which is written,
9714 note_reg_stored (x, setter, arg)
9715 rtx x, setter ATTRIBUTE_UNUSED;
9718 struct note_reg_stored_arg *t = (struct note_reg_stored_arg *) arg;
9723 /* Try to replace every occurrence of pseudo REGNO with REPLACEMENT.
9724 There must be exactly one insn that sets this pseudo; it will be
9725 deleted if all replacements succeed and we can prove that the register
9726 is not used after the loop. */
9729 try_copy_prop (loop, replacement, regno)
9730 const struct loop *loop;
9734 /* This is the reg that we are copying from. */
9735 rtx reg_rtx = regno_reg_rtx[regno];
9738 /* These help keep track of whether we replaced all uses of the reg. */
9739 int replaced_last = 0;
9740 int store_is_first = 0;
9742 for (insn = next_insn_in_loop (loop, loop->scan_start);
9744 insn = next_insn_in_loop (loop, insn))
9748 /* Only substitute within one extended basic block from the initializing
9750 if (GET_CODE (insn) == CODE_LABEL && init_insn)
9753 if (! INSN_P (insn))
9756 /* Is this the initializing insn? */
9757 set = single_set (insn);
9759 && GET_CODE (SET_DEST (set)) == REG
9760 && REGNO (SET_DEST (set)) == regno)
9766 if (REGNO_FIRST_UID (regno) == INSN_UID (insn))
9770 /* Only substitute after seeing the initializing insn. */
9771 if (init_insn && insn != init_insn)
9773 struct note_reg_stored_arg arg;
9776 array[1] = replacement;
9779 for_each_rtx (&insn, replace_loop_reg, array);
9780 if (REGNO_LAST_UID (regno) == INSN_UID (insn))
9783 /* Stop replacing when REPLACEMENT is modified. */
9784 arg.reg = replacement;
9786 note_stores (PATTERN (insn), note_reg_stored, &arg);
9793 if (apply_change_group ())
9795 if (loop_dump_stream)
9796 fprintf (loop_dump_stream, " Replaced reg %d", regno);
9797 if (store_is_first && replaced_last)
9799 PUT_CODE (init_insn, NOTE);
9800 NOTE_LINE_NUMBER (init_insn) = NOTE_INSN_DELETED;
9801 if (loop_dump_stream)
9802 fprintf (loop_dump_stream, ", deleting init_insn (%d)",
9803 INSN_UID (init_insn));
9805 if (loop_dump_stream)
9806 fprintf (loop_dump_stream, ".\n");
9811 /* Try to replace occurrences of pseudo REGNO with REPLACEMENT within
9812 loop LOOP if the order of the sets of these registers can be
9813 swapped. There must be exactly one insn within the loop that sets
9814 this pseudo followed immediately by a move insn that sets
9815 REPLACEMENT with REGNO. */
9817 try_swap_copy_prop (loop, replacement, regno)
9818 const struct loop *loop;
9824 unsigned int new_regno;
9826 new_regno = REGNO (replacement);
9828 for (insn = next_insn_in_loop (loop, loop->scan_start);
9830 insn = next_insn_in_loop (loop, insn))
9832 /* Search for the insn that copies REGNO to NEW_REGNO? */
9833 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i'
9834 && (set = single_set (insn))
9835 && GET_CODE (SET_DEST (set)) == REG
9836 && REGNO (SET_DEST (set)) == new_regno
9837 && GET_CODE (SET_SRC (set)) == REG
9838 && REGNO (SET_SRC (set)) == regno)
9842 if (insn != NULL_RTX)
9847 /* Some DEF-USE info would come in handy here to make this
9848 function more general. For now, just check the previous insn
9849 which is the most likely candidate for setting REGNO. */
9851 prev_insn = PREV_INSN (insn);
9853 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i'
9854 && (prev_set = single_set (prev_insn))
9855 && GET_CODE (SET_DEST (prev_set)) == REG
9856 && REGNO (SET_DEST (prev_set)) == regno)
9859 (set (reg regno) (expr))
9860 (set (reg new_regno) (reg regno))
9862 so try converting this to:
9863 (set (reg new_regno) (expr))
9864 (set (reg regno) (reg new_regno))
9866 The former construct is often generated when a global
9867 variable used for an induction variable is shadowed by a
9868 register (NEW_REGNO). The latter construct improves the
9869 chances of GIV replacement and BIV elimination. */
9871 validate_change (prev_insn, &SET_DEST (prev_set),
9873 validate_change (insn, &SET_DEST (set),
9875 validate_change (insn, &SET_SRC (set),
9878 if (apply_change_group ())
9880 if (loop_dump_stream)
9881 fprintf (loop_dump_stream,
9882 " Swapped set of reg %d at %d with reg %d at %d.\n",
9883 regno, INSN_UID (insn),
9884 new_regno, INSN_UID (prev_insn));
9886 /* Update first use of REGNO. */
9887 if (REGNO_FIRST_UID (regno) == INSN_UID (prev_insn))
9888 REGNO_FIRST_UID (regno) = INSN_UID (insn);
9890 /* Now perform copy propagation to hopefully
9891 remove all uses of REGNO within the loop. */
9892 try_copy_prop (loop, replacement, regno);
9899 /* Replace MEM with its associated pseudo register. This function is
9900 called from load_mems via for_each_rtx. DATA is actually an
9901 rtx_and_int * describing the instruction currently being scanned
9902 and the MEM we are currently replacing. */
9905 replace_loop_mem (mem, data)
9917 switch (GET_CODE (m))
9923 /* We're not interested in the MEM associated with a
9924 CONST_DOUBLE, so there's no need to traverse into one. */
9928 /* This is not a MEM. */
9932 ri = (rtx_and_int *) data;
9935 if (!rtx_equal_p (loop_mems[i].mem, m))
9936 /* This is not the MEM we are currently replacing. */
9941 /* Actually replace the MEM. */
9942 validate_change (insn, mem, loop_mems[i].reg, 1);
9947 /* Replace one register with another. Called through for_each_rtx; PX points
9948 to the rtx being scanned. DATA is actually an array of three rtx's; the
9949 first one is the one to be replaced, and the second one the replacement.
9950 The third one is the current insn. */
9953 replace_loop_reg (px, data)
9958 rtx *array = (rtx *) data;
9964 validate_change (array[2], px, array[1], 1);
9969 /* Replace occurrences of the old exit label for the loop with the new
9970 one. DATA is an rtx_pair containing the old and new labels,
9974 replace_label (x, data)
9979 rtx old_label = ((rtx_pair *) data)->r1;
9980 rtx new_label = ((rtx_pair *) data)->r2;
9985 if (GET_CODE (l) != LABEL_REF)
9988 if (XEXP (l, 0) != old_label)
9991 XEXP (l, 0) = new_label;
9992 ++LABEL_NUSES (new_label);
9993 --LABEL_NUSES (old_label);