1 /* Move constant computations out of loops.
2 Copyright (C) 1987, 1988, 1989, 1991, 1992 Free Software Foundation, Inc.
4 This file is part of GNU CC.
6 GNU CC is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2, or (at your option)
11 GNU CC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with GNU CC; see the file COPYING. If not, write to
18 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. */
21 /* This is the loop optimization pass of the compiler.
22 It finds invariant computations within loops and moves them
23 to the beginning of the loop. Then it identifies basic and
24 general induction variables. Strength reduction is applied to the general
25 induction variables, and induction variable elimination is applied to
26 the basic induction variables.
28 It also finds cases where
29 a register is set within the loop by zero-extending a narrower value
30 and changes these to zero the entire register once before the loop
31 and merely copy the low part within the loop.
33 Most of the complexity is in heuristics to decide when it is worth
34 while to do these things. */
41 #include "insn-config.h"
42 #include "insn-flags.h"
44 #include "hard-reg-set.h"
50 /* Vector mapping INSN_UIDs to luids.
51 The luids are like uids but increase monotonically always.
52 We use them to see whether a jump comes from outside a given loop. */
56 /* Indexed by INSN_UID, contains the ordinal giving the (innermost) loop
57 number the insn is contained in. */
61 /* 1 + largest uid of any insn. */
65 /* 1 + luid of last insn. */
69 /* Number of loops detected in current function. Used as index to the
72 static int max_loop_num;
74 /* Indexed by loop number, contains the first and last insn of each loop. */
76 static rtx *loop_number_loop_starts, *loop_number_loop_ends;
78 /* For each loop, gives the containing loop number, -1 if none. */
82 /* Indexed by loop number, contains a nonzero value if the "loop" isn't
83 really a loop (an insn outside the loop branches into it). */
85 static char *loop_invalid;
87 /* Indexed by loop number, links together all LABEL_REFs which refer to
88 code labels outside the loop. Used by routines that need to know all
89 loop exits, such as final_biv_value and final_giv_value.
91 This does not include loop exits due to return instructions. This is
92 because all bivs and givs are pseudos, and hence must be dead after a
93 return, so the presense of a return does not affect any of the
94 optimizations that use this info. It is simpler to just not include return
95 instructions on this list. */
97 rtx *loop_number_exit_labels;
99 /* Holds the number of loop iterations. It is zero if the number could not be
100 calculated. Must be unsigned since the number of iterations can
101 be as high as 2^wordsize-1. For loops with a wider iterator, this number
102 will will be zero if the number of loop iterations is too large for an
103 unsigned integer to hold. */
105 unsigned HOST_WIDE_INT loop_n_iterations;
107 /* Nonzero if there is a subroutine call in the current loop.
108 (unknown_address_altered is also nonzero in this case.) */
110 static int loop_has_call;
112 /* Nonzero if there is a volatile memory reference in the current
115 static int loop_has_volatile;
117 /* Added loop_continue which is the NOTE_INSN_LOOP_CONT of the
118 current loop. A continue statement will generate a branch to
119 NEXT_INSN (loop_continue). */
121 static rtx loop_continue;
123 /* Indexed by register number, contains the number of times the reg
124 is set during the loop being scanned.
125 During code motion, a negative value indicates a reg that has been
126 made a candidate; in particular -2 means that it is an candidate that
127 we know is equal to a constant and -1 means that it is an candidate
128 not known equal to a constant.
129 After code motion, regs moved have 0 (which is accurate now)
130 while the failed candidates have the original number of times set.
132 Therefore, at all times, == 0 indicates an invariant register;
133 < 0 a conditionally invariant one. */
135 static short *n_times_set;
137 /* Original value of n_times_set; same except that this value
138 is not set negative for a reg whose sets have been made candidates
139 and not set to 0 for a reg that is moved. */
141 static short *n_times_used;
143 /* Index by register number, 1 indicates that the register
144 cannot be moved or strength reduced. */
146 static char *may_not_optimize;
148 /* Nonzero means reg N has already been moved out of one loop.
149 This reduces the desire to move it out of another. */
151 static char *moved_once;
153 /* Array of MEMs that are stored in this loop. If there are too many to fit
154 here, we just turn on unknown_address_altered. */
156 #define NUM_STORES 20
157 static rtx loop_store_mems[NUM_STORES];
159 /* Index of first available slot in above array. */
160 static int loop_store_mems_idx;
162 /* Nonzero if we don't know what MEMs were changed in the current loop.
163 This happens if the loop contains a call (in which case `loop_has_call'
164 will also be set) or if we store into more than NUM_STORES MEMs. */
166 static int unknown_address_altered;
168 /* Count of movable (i.e. invariant) instructions discovered in the loop. */
169 static int num_movables;
171 /* Count of memory write instructions discovered in the loop. */
172 static int num_mem_sets;
174 /* Number of loops contained within the current one, including itself. */
175 static int loops_enclosed;
177 /* Bound on pseudo register number before loop optimization.
178 A pseudo has valid regscan info if its number is < max_reg_before_loop. */
179 int max_reg_before_loop;
181 /* This obstack is used in product_cheap_p to allocate its rtl. It
182 may call gen_reg_rtx which, in turn, may reallocate regno_reg_rtx.
183 If we used the same obstack that it did, we would be deallocating
186 static struct obstack temp_obstack;
188 /* This is where the pointer to the obstack being used for RTL is stored. */
190 extern struct obstack *rtl_obstack;
192 #define obstack_chunk_alloc xmalloc
193 #define obstack_chunk_free free
195 extern char *oballoc ();
197 /* During the analysis of a loop, a chain of `struct movable's
198 is made to record all the movable insns found.
199 Then the entire chain can be scanned to decide which to move. */
203 rtx insn; /* A movable insn */
204 rtx set_src; /* The expression this reg is set from. */
205 rtx set_dest; /* The destination of this SET. */
206 rtx dependencies; /* When INSN is libcall, this is an EXPR_LIST
207 of any registers used within the LIBCALL. */
208 int consec; /* Number of consecutive following insns
209 that must be moved with this one. */
210 int regno; /* The register it sets */
211 short lifetime; /* lifetime of that register;
212 may be adjusted when matching movables
213 that load the same value are found. */
214 short savings; /* Number of insns we can move for this reg,
215 including other movables that force this
216 or match this one. */
217 unsigned int cond : 1; /* 1 if only conditionally movable */
218 unsigned int force : 1; /* 1 means MUST move this insn */
219 unsigned int global : 1; /* 1 means reg is live outside this loop */
220 /* If PARTIAL is 1, GLOBAL means something different:
221 that the reg is live outside the range from where it is set
222 to the following label. */
223 unsigned int done : 1; /* 1 inhibits further processing of this */
225 unsigned int partial : 1; /* 1 means this reg is used for zero-extending.
226 In particular, moving it does not make it
228 unsigned int move_insn : 1; /* 1 means that we call emit_move_insn to
229 load SRC, rather than copying INSN. */
230 unsigned int is_equiv : 1; /* 1 means a REG_EQUIV is present on INSN. */
231 enum machine_mode savemode; /* Nonzero means it is a mode for a low part
232 that we should avoid changing when clearing
233 the rest of the reg. */
234 struct movable *match; /* First entry for same value */
235 struct movable *forces; /* An insn that must be moved if this is */
236 struct movable *next;
239 FILE *loop_dump_stream;
241 /* Forward declarations. */
243 static void find_and_verify_loops ();
244 static void mark_loop_jump ();
245 static void prescan_loop ();
246 static int reg_in_basic_block_p ();
247 static int consec_sets_invariant_p ();
248 static rtx libcall_other_reg ();
249 static int labels_in_range_p ();
250 static void count_loop_regs_set ();
251 static void note_addr_stored ();
252 static int loop_reg_used_before_p ();
253 static void scan_loop ();
254 static void replace_call_address ();
255 static rtx skip_consec_insns ();
256 static int libcall_benefit ();
257 static void ignore_some_movables ();
258 static void force_movables ();
259 static void combine_movables ();
260 static int rtx_equal_for_loop_p ();
261 static void move_movables ();
262 static void strength_reduce ();
263 static int valid_initial_value_p ();
264 static void find_mem_givs ();
265 static void record_biv ();
266 static void check_final_value ();
267 static void record_giv ();
268 static void update_giv_derive ();
269 static void delete_insn_forces ();
270 static int basic_induction_var ();
271 static rtx simplify_giv_expr ();
272 static int general_induction_var ();
273 static int consec_sets_giv ();
274 static int check_dbra_loop ();
275 static rtx express_from ();
276 static int combine_givs_p ();
277 static void combine_givs ();
278 static int product_cheap_p ();
279 static int maybe_eliminate_biv ();
280 static int maybe_eliminate_biv_1 ();
281 static int last_use_this_basic_block ();
282 static void record_initial ();
283 static void update_reg_last_use ();
285 /* Relative gain of eliminating various kinds of operations. */
292 /* Benefit penalty, if a giv is not replaceable, i.e. must emit an insn to
293 copy the value of the strength reduced giv to its original register. */
299 char *free_point = (char *) oballoc (1);
300 rtx reg = gen_rtx (REG, word_mode, 0);
301 rtx pow2 = GEN_INT (32);
305 add_cost = rtx_cost (gen_rtx (PLUS, word_mode, reg, reg), SET);
307 /* We multiply by 2 to reconcile the difference in scale between
308 these two ways of computing costs. Otherwise the cost of a copy
309 will be far less than the cost of an add. */
313 /* Free the objects we just allocated. */
316 /* Initialize the obstack used for rtl in product_cheap_p. */
317 gcc_obstack_init (&temp_obstack);
320 /* Entry point of this file. Perform loop optimization
321 on the current function. F is the first insn of the function
322 and DUMPFILE is a stream for output of a trace of actions taken
323 (or 0 if none should be output). */
326 loop_optimize (f, dumpfile)
327 /* f is the first instruction of a chain of insns for one function */
336 loop_dump_stream = dumpfile;
338 init_recog_no_volatile ();
339 init_alias_analysis ();
341 max_reg_before_loop = max_reg_num ();
343 moved_once = (char *) alloca (max_reg_before_loop);
344 bzero (moved_once, max_reg_before_loop);
348 /* Count the number of loops. */
351 for (insn = f; insn; insn = NEXT_INSN (insn))
353 if (GET_CODE (insn) == NOTE
354 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
358 /* Don't waste time if no loops. */
359 if (max_loop_num == 0)
362 /* Get size to use for tables indexed by uids.
363 Leave some space for labels allocated by find_and_verify_loops. */
364 max_uid_for_loop = get_max_uid () + 1 + max_loop_num * 4;
366 uid_luid = (int *) alloca (max_uid_for_loop * sizeof (int));
367 uid_loop_num = (int *) alloca (max_uid_for_loop * sizeof (int));
369 bzero (uid_luid, max_uid_for_loop * sizeof (int));
370 bzero (uid_loop_num, max_uid_for_loop * sizeof (int));
372 /* Allocate tables for recording each loop. We set each entry, so they need
374 loop_number_loop_starts = (rtx *) alloca (max_loop_num * sizeof (rtx));
375 loop_number_loop_ends = (rtx *) alloca (max_loop_num * sizeof (rtx));
376 loop_outer_loop = (int *) alloca (max_loop_num * sizeof (int));
377 loop_invalid = (char *) alloca (max_loop_num * sizeof (char));
378 loop_number_exit_labels = (rtx *) alloca (max_loop_num * sizeof (rtx));
380 if (flag_unroll_loops && write_symbols != NO_DEBUG)
382 loop_number_first_block
383 = (union tree_node **) alloca (max_loop_num
384 * sizeof (union tree_node *));
385 loop_number_last_block
386 = (union tree_node **) alloca (max_loop_num
387 * sizeof (union tree_node *));
388 loop_number_block_level = (int *) alloca (max_loop_num * sizeof (int));
391 /* Find and process each loop.
392 First, find them, and record them in order of their beginnings. */
393 find_and_verify_loops (f);
395 /* Now find all register lifetimes. This must be done after
396 find_and_verify_loops, because it might reorder the insns in the
398 reg_scan (f, max_reg_num (), 1);
400 /* Compute the mapping from uids to luids.
401 LUIDs are numbers assigned to insns, like uids,
402 except that luids increase monotonically through the code.
403 Don't assign luids to line-number NOTEs, so that the distance in luids
404 between two insns is not affected by -g. */
406 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
409 if (GET_CODE (insn) != NOTE
410 || NOTE_LINE_NUMBER (insn) <= 0)
411 uid_luid[INSN_UID (insn)] = ++i;
413 /* Give a line number note the same luid as preceding insn. */
414 uid_luid[INSN_UID (insn)] = i;
419 /* Don't leave gaps in uid_luid for insns that have been
420 deleted. It is possible that the first or last insn
421 using some register has been deleted by cross-jumping.
422 Make sure that uid_luid for that former insn's uid
423 points to the general area where that insn used to be. */
424 for (i = 0; i < max_uid_for_loop; i++)
426 uid_luid[0] = uid_luid[i];
427 if (uid_luid[0] != 0)
430 for (i = 0; i < max_uid_for_loop; i++)
431 if (uid_luid[i] == 0)
432 uid_luid[i] = uid_luid[i - 1];
434 /* Create a mapping from loops to BLOCK tree nodes. */
435 if (flag_unroll_loops && write_symbols != NO_DEBUG)
436 find_loop_tree_blocks (f);
438 /* Now scan the loops, last ones first, since this means inner ones are done
439 before outer ones. */
440 for (i = max_loop_num-1; i >= 0; i--)
441 if (! loop_invalid[i] && loop_number_loop_ends[i])
442 scan_loop (loop_number_loop_starts[i], loop_number_loop_ends[i],
446 /* Optimize one loop whose start is LOOP_START and end is END.
447 LOOP_START is the NOTE_INSN_LOOP_BEG and END is the matching
448 NOTE_INSN_LOOP_END. */
450 /* ??? Could also move memory writes out of loops if the destination address
451 is invariant, the source is invariant, the memory write is not volatile,
452 and if we can prove that no read inside the loop can read this address
453 before the write occurs. If there is a read of this address after the
454 write, then we can also mark the memory read as invariant. */
457 scan_loop (loop_start, end, nregs)
463 /* 1 if we are scanning insns that could be executed zero times. */
465 /* 1 if we are scanning insns that might never be executed
466 due to a subroutine call which might exit before they are reached. */
468 /* For a rotated loop that is entered near the bottom,
469 this is the label at the top. Otherwise it is zero. */
471 /* Jump insn that enters the loop, or 0 if control drops in. */
472 rtx loop_entry_jump = 0;
473 /* Place in the loop where control enters. */
475 /* Number of insns in the loop. */
480 /* The SET from an insn, if it is the only SET in the insn. */
482 /* Chain describing insns movable in current loop. */
483 struct movable *movables = 0;
484 /* Last element in `movables' -- so we can add elements at the end. */
485 struct movable *last_movable = 0;
486 /* Ratio of extra register life span we can justify
487 for saving an instruction. More if loop doesn't call subroutines
488 since in that case saving an insn makes more difference
489 and more registers are available. */
491 /* If we have calls, contains the insn in which a register was used
492 if it was used exactly once; contains const0_rtx if it was used more
494 rtx *reg_single_usage = 0;
496 n_times_set = (short *) alloca (nregs * sizeof (short));
497 n_times_used = (short *) alloca (nregs * sizeof (short));
498 may_not_optimize = (char *) alloca (nregs);
500 /* Determine whether this loop starts with a jump down to a test at
501 the end. This will occur for a small number of loops with a test
502 that is too complex to duplicate in front of the loop.
504 We search for the first insn or label in the loop, skipping NOTEs.
505 However, we must be careful not to skip past a NOTE_INSN_LOOP_BEG
506 (because we might have a loop executed only once that contains a
507 loop which starts with a jump to its exit test) or a NOTE_INSN_LOOP_END
508 (in case we have a degenerate loop).
510 Note that if we mistakenly think that a loop is entered at the top
511 when, in fact, it is entered at the exit test, the only effect will be
512 slightly poorer optimization. Making the opposite error can generate
513 incorrect code. Since very few loops now start with a jump to the
514 exit test, the code here to detect that case is very conservative. */
516 for (p = NEXT_INSN (loop_start);
518 && GET_CODE (p) != CODE_LABEL && GET_RTX_CLASS (GET_CODE (p)) != 'i'
519 && (GET_CODE (p) != NOTE
520 || (NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_BEG
521 && NOTE_LINE_NUMBER (p) != NOTE_INSN_LOOP_END));
527 /* Set up variables describing this loop. */
528 prescan_loop (loop_start, end);
529 threshold = (loop_has_call ? 1 : 2) * (1 + n_non_fixed_regs);
531 /* If loop has a jump before the first label,
532 the true entry is the target of that jump.
533 Start scan from there.
534 But record in LOOP_TOP the place where the end-test jumps
535 back to so we can scan that after the end of the loop. */
536 if (GET_CODE (p) == JUMP_INSN)
540 /* Loop entry must be unconditional jump (and not a RETURN) */
542 && JUMP_LABEL (p) != 0
543 /* Check to see whether the jump actually
544 jumps out of the loop (meaning it's no loop).
545 This case can happen for things like
546 do {..} while (0). If this label was generated previously
547 by loop, we can't tell anything about it and have to reject
549 && INSN_UID (JUMP_LABEL (p)) < max_uid_for_loop
550 && INSN_LUID (JUMP_LABEL (p)) >= INSN_LUID (loop_start)
551 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (end))
553 loop_top = next_label (scan_start);
554 scan_start = JUMP_LABEL (p);
558 /* If SCAN_START was an insn created by loop, we don't know its luid
559 as required by loop_reg_used_before_p. So skip such loops. (This
560 test may never be true, but it's best to play it safe.)
562 Also, skip loops where we do not start scanning at a label. This
563 test also rejects loops starting with a JUMP_INSN that failed the
566 if (INSN_UID (scan_start) >= max_uid_for_loop
567 || GET_CODE (scan_start) != CODE_LABEL)
569 if (loop_dump_stream)
570 fprintf (loop_dump_stream, "\nLoop from %d to %d is phony.\n\n",
571 INSN_UID (loop_start), INSN_UID (end));
575 /* Count number of times each reg is set during this loop.
576 Set may_not_optimize[I] if it is not safe to move out
577 the setting of register I. If this loop has calls, set
578 reg_single_usage[I]. */
580 bzero (n_times_set, nregs * sizeof (short));
581 bzero (may_not_optimize, nregs);
585 reg_single_usage = (rtx *) alloca (nregs * sizeof (rtx));
586 bzero (reg_single_usage, nregs * sizeof (rtx));
589 count_loop_regs_set (loop_top ? loop_top : loop_start, end,
590 may_not_optimize, reg_single_usage, &insn_count, nregs);
592 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
593 may_not_optimize[i] = 1, n_times_set[i] = 1;
594 bcopy (n_times_set, n_times_used, nregs * sizeof (short));
596 if (loop_dump_stream)
598 fprintf (loop_dump_stream, "\nLoop from %d to %d: %d real insns.\n",
599 INSN_UID (loop_start), INSN_UID (end), insn_count);
601 fprintf (loop_dump_stream, "Continue at insn %d.\n",
602 INSN_UID (loop_continue));
605 /* Scan through the loop finding insns that are safe to move.
606 Set n_times_set negative for the reg being set, so that
607 this reg will be considered invariant for subsequent insns.
608 We consider whether subsequent insns use the reg
609 in deciding whether it is worth actually moving.
611 MAYBE_NEVER is nonzero if we have passed a conditional jump insn
612 and therefore it is possible that the insns we are scanning
613 would never be executed. At such times, we must make sure
614 that it is safe to execute the insn once instead of zero times.
615 When MAYBE_NEVER is 0, all insns will be executed at least once
616 so that is not a problem. */
622 /* At end of a straight-in loop, we are done.
623 At end of a loop entered at the bottom, scan the top. */
629 p = NEXT_INSN (loop_top);
636 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
637 && find_reg_note (p, REG_LIBCALL, NULL_RTX))
639 else if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
640 && find_reg_note (p, REG_RETVAL, NULL_RTX))
643 if (GET_CODE (p) == INSN
644 && (set = single_set (p))
645 && GET_CODE (SET_DEST (set)) == REG
646 && ! may_not_optimize[REGNO (SET_DEST (set))])
651 rtx src = SET_SRC (set);
652 rtx dependencies = 0;
654 /* Figure out what to use as a source of this insn. If a REG_EQUIV
655 note is given or if a REG_EQUAL note with a constant operand is
656 specified, use it as the source and mark that we should move
657 this insn by calling emit_move_insn rather that duplicating the
660 Otherwise, only use the REG_EQUAL contents if a REG_RETVAL note
662 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
664 src = XEXP (temp, 0), move_insn = 1;
667 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
668 if (temp && CONSTANT_P (XEXP (temp, 0)))
669 src = XEXP (temp, 0), move_insn = 1;
670 if (temp && find_reg_note (p, REG_RETVAL, NULL_RTX))
672 src = XEXP (temp, 0);
673 /* A libcall block can use regs that don't appear in
674 the equivalent expression. To move the libcall,
675 we must move those regs too. */
676 dependencies = libcall_other_reg (p, src);
680 /* Don't try to optimize a register that was made
681 by loop-optimization for an inner loop.
682 We don't know its life-span, so we can't compute the benefit. */
683 if (REGNO (SET_DEST (set)) >= max_reg_before_loop)
685 /* In order to move a register, we need to have one of three cases:
686 (1) it is used only in the same basic block as the set
687 (2) it is not a user variable.
688 (3) the set is guaranteed to be executed once the loop starts,
689 and the reg is not used until after that. */
690 else if (! ((! maybe_never
691 && ! loop_reg_used_before_p (set, p, loop_start,
693 || ! REG_USERVAR_P (SET_DEST (PATTERN (p)))
694 || reg_in_basic_block_p (p, SET_DEST (PATTERN (p)))))
696 else if ((tem = invariant_p (src))
697 && (dependencies == 0
698 || (tem2 = invariant_p (dependencies)) != 0)
699 && (n_times_set[REGNO (SET_DEST (set))] == 1
701 = consec_sets_invariant_p (SET_DEST (set),
702 n_times_set[REGNO (SET_DEST (set))],
704 /* If the insn can cause a trap (such as divide by zero),
705 can't move it unless it's guaranteed to be executed
706 once loop is entered. Even a function call might
707 prevent the trap insn from being reached
708 (since it might exit!) */
709 && ! ((maybe_never || call_passed)
710 && may_trap_p (src)))
712 register struct movable *m;
713 register int regno = REGNO (SET_DEST (set));
715 /* A potential lossage is where we have a case where two insns
716 can be combined as long as they are both in the loop, but
717 we move one of them outside the loop. For large loops,
718 this can lose. The most common case of this is the address
719 of a function being called.
721 Therefore, if this register is marked as being used exactly
722 once if we are in a loop with calls (a "large loop"), see if
723 we can replace the usage of this register with the source
724 of this SET. If we can, delete this insn.
726 Don't do this if P has a REG_RETVAL note or if we have
727 SMALL_REGISTER_CLASSES and SET_SRC is a hard register. */
729 if (reg_single_usage && reg_single_usage[regno] != 0
730 && reg_single_usage[regno] != const0_rtx
731 && regno_first_uid[regno] == INSN_UID (p)
732 && (regno_last_uid[regno]
733 == INSN_UID (reg_single_usage[regno]))
734 && n_times_set[REGNO (SET_DEST (set))] == 1
735 && ! side_effects_p (SET_SRC (set))
736 && ! find_reg_note (p, REG_RETVAL, NULL_RTX)
737 #ifdef SMALL_REGISTER_CLASSES
738 && ! (GET_CODE (SET_SRC (set)) == REG
739 && REGNO (SET_SRC (set)) < FIRST_PSEUDO_REGISTER)
741 /* This test is not redundant; SET_SRC (set) might be
742 a call-clobbered register and the life of REGNO
743 might span a call. */
744 && ! modified_between_p (SET_SRC (set), p,
745 reg_single_usage[regno])
746 && validate_replace_rtx (SET_DEST (set), SET_SRC (set),
747 reg_single_usage[regno]))
749 /* Replace any usage in a REG_EQUAL note. */
750 REG_NOTES (reg_single_usage[regno])
751 = replace_rtx (REG_NOTES (reg_single_usage[regno]),
752 SET_DEST (set), SET_SRC (set));
755 NOTE_LINE_NUMBER (p) = NOTE_INSN_DELETED;
756 NOTE_SOURCE_FILE (p) = 0;
757 n_times_set[regno] = 0;
761 m = (struct movable *) alloca (sizeof (struct movable));
765 m->dependencies = dependencies;
766 m->set_dest = SET_DEST (set);
768 m->consec = n_times_set[REGNO (SET_DEST (set))] - 1;
772 m->move_insn = move_insn;
773 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
774 m->savemode = VOIDmode;
776 /* Set M->cond if either invariant_p or consec_sets_invariant_p
777 returned 2 (only conditionally invariant). */
778 m->cond = ((tem | tem1 | tem2) > 1);
779 m->global = (uid_luid[regno_last_uid[regno]] > INSN_LUID (end)
780 || uid_luid[regno_first_uid[regno]] < INSN_LUID (loop_start));
782 m->lifetime = (uid_luid[regno_last_uid[regno]]
783 - uid_luid[regno_first_uid[regno]]);
784 m->savings = n_times_used[regno];
785 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
786 m->savings += libcall_benefit (p);
787 n_times_set[regno] = move_insn ? -2 : -1;
788 /* Add M to the end of the chain MOVABLES. */
792 last_movable->next = m;
797 /* Skip this insn, not checking REG_LIBCALL notes. */
799 /* Skip the consecutive insns, if there are any. */
800 p = skip_consec_insns (p, m->consec);
801 /* Back up to the last insn of the consecutive group. */
802 p = prev_nonnote_insn (p);
804 /* We must now reset m->move_insn, m->is_equiv, and possibly
805 m->set_src to correspond to the effects of all the
807 temp = find_reg_note (p, REG_EQUIV, NULL_RTX);
809 m->set_src = XEXP (temp, 0), m->move_insn = 1;
812 temp = find_reg_note (p, REG_EQUAL, NULL_RTX);
813 if (temp && CONSTANT_P (XEXP (temp, 0)))
814 m->set_src = XEXP (temp, 0), m->move_insn = 1;
819 m->is_equiv = (find_reg_note (p, REG_EQUIV, NULL_RTX) != 0);
822 /* If this register is always set within a STRICT_LOW_PART
823 or set to zero, then its high bytes are constant.
824 So clear them outside the loop and within the loop
825 just load the low bytes.
826 We must check that the machine has an instruction to do so.
827 Also, if the value loaded into the register
828 depends on the same register, this cannot be done. */
829 else if (SET_SRC (set) == const0_rtx
830 && GET_CODE (NEXT_INSN (p)) == INSN
831 && (set1 = single_set (NEXT_INSN (p)))
832 && GET_CODE (set1) == SET
833 && (GET_CODE (SET_DEST (set1)) == STRICT_LOW_PART)
834 && (GET_CODE (XEXP (SET_DEST (set1), 0)) == SUBREG)
835 && (SUBREG_REG (XEXP (SET_DEST (set1), 0))
837 && !reg_mentioned_p (SET_DEST (set), SET_SRC (set1)))
839 register int regno = REGNO (SET_DEST (set));
840 if (n_times_set[regno] == 2)
842 register struct movable *m;
843 m = (struct movable *) alloca (sizeof (struct movable));
846 m->set_dest = SET_DEST (set);
854 /* If the insn may not be executed on some cycles,
855 we can't clear the whole reg; clear just high part.
856 Not even if the reg is used only within this loop.
863 Clearing x before the inner loop could clobber a value
864 being saved from the last time around the outer loop.
865 However, if the reg is not used outside this loop
866 and all uses of the register are in the same
867 basic block as the store, there is no problem.
869 If this insn was made by loop, we don't know its
870 INSN_LUID and hence must make a conservative
872 m->global = (INSN_UID (p) >= max_uid_for_loop
873 || (uid_luid[regno_last_uid[regno]]
875 || (uid_luid[regno_first_uid[regno]]
877 || (labels_in_range_p
878 (p, uid_luid[regno_first_uid[regno]])));
879 if (maybe_never && m->global)
880 m->savemode = GET_MODE (SET_SRC (set1));
882 m->savemode = VOIDmode;
886 m->lifetime = (uid_luid[regno_last_uid[regno]]
887 - uid_luid[regno_first_uid[regno]]);
889 n_times_set[regno] = -1;
890 /* Add M to the end of the chain MOVABLES. */
894 last_movable->next = m;
899 /* Past a call insn, we get to insns which might not be executed
900 because the call might exit. This matters for insns that trap.
901 Call insns inside a REG_LIBCALL/REG_RETVAL block always return,
902 so they don't count. */
903 else if (GET_CODE (p) == CALL_INSN && ! in_libcall)
905 /* Past a label or a jump, we get to insns for which we
906 can't count on whether or how many times they will be
907 executed during each iteration. Therefore, we can
908 only move out sets of trivial variables
909 (those not used after the loop). */
910 /* This code appears in three places, once in scan_loop, and twice
911 in strength_reduce. */
912 else if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
913 /* If we enter the loop in the middle, and scan around to the
914 beginning, don't set maybe_never for that. This must be an
915 unconditional jump, otherwise the code at the top of the
916 loop might never be executed. Unconditional jumps are
917 followed a by barrier then loop end. */
918 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
919 && NEXT_INSN (NEXT_INSN (p)) == end
920 && simplejump_p (p)))
922 /* At the virtual top of a converted loop, insns are again known to
923 be executed: logically, the loop begins here even though the exit
924 code has been duplicated. */
925 else if (GET_CODE (p) == NOTE
926 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP)
927 maybe_never = call_passed = 0;
930 /* If one movable subsumes another, ignore that other. */
932 ignore_some_movables (movables);
934 /* For each movable insn, see if the reg that it loads
935 leads when it dies right into another conditionally movable insn.
936 If so, record that the second insn "forces" the first one,
937 since the second can be moved only if the first is. */
939 force_movables (movables);
941 /* See if there are multiple movable insns that load the same value.
942 If there are, make all but the first point at the first one
943 through the `match' field, and add the priorities of them
944 all together as the priority of the first. */
946 combine_movables (movables, nregs);
948 /* Now consider each movable insn to decide whether it is worth moving.
949 Store 0 in n_times_set for each reg that is moved. */
951 move_movables (movables, threshold,
952 insn_count, loop_start, end, nregs);
954 /* Now candidates that still are negative are those not moved.
955 Change n_times_set to indicate that those are not actually invariant. */
956 for (i = 0; i < nregs; i++)
957 if (n_times_set[i] < 0)
958 n_times_set[i] = n_times_used[i];
960 if (flag_strength_reduce)
961 strength_reduce (scan_start, end, loop_top,
962 insn_count, loop_start, end);
965 /* Add elements to *OUTPUT to record all the pseudo-regs
966 mentioned in IN_THIS but not mentioned in NOT_IN_THIS. */
969 record_excess_regs (in_this, not_in_this, output)
970 rtx in_this, not_in_this;
977 code = GET_CODE (in_this);
991 if (REGNO (in_this) >= FIRST_PSEUDO_REGISTER
992 && ! reg_mentioned_p (in_this, not_in_this))
993 *output = gen_rtx (EXPR_LIST, VOIDmode, in_this, *output);
997 fmt = GET_RTX_FORMAT (code);
998 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1005 for (j = 0; j < XVECLEN (in_this, i); j++)
1006 record_excess_regs (XVECEXP (in_this, i, j), not_in_this, output);
1010 record_excess_regs (XEXP (in_this, i), not_in_this, output);
1016 /* Check what regs are referred to in the libcall block ending with INSN,
1017 aside from those mentioned in the equivalent value.
1018 If there are none, return 0.
1019 If there are one or more, return an EXPR_LIST containing all of them. */
1022 libcall_other_reg (insn, equiv)
1025 rtx note = find_reg_note (insn, REG_RETVAL, NULL_RTX);
1026 rtx p = XEXP (note, 0);
1029 /* First, find all the regs used in the libcall block
1030 that are not mentioned as inputs to the result. */
1034 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1035 || GET_CODE (p) == CALL_INSN)
1036 record_excess_regs (PATTERN (p), equiv, &output);
1043 /* Return 1 if all uses of REG
1044 are between INSN and the end of the basic block. */
1047 reg_in_basic_block_p (insn, reg)
1050 int regno = REGNO (reg);
1053 if (regno_first_uid[regno] != INSN_UID (insn))
1056 /* Search this basic block for the already recorded last use of the reg. */
1057 for (p = insn; p; p = NEXT_INSN (p))
1059 switch (GET_CODE (p))
1066 /* Ordinary insn: if this is the last use, we win. */
1067 if (regno_last_uid[regno] == INSN_UID (p))
1072 /* Jump insn: if this is the last use, we win. */
1073 if (regno_last_uid[regno] == INSN_UID (p))
1075 /* Otherwise, it's the end of the basic block, so we lose. */
1080 /* It's the end of the basic block, so we lose. */
1085 /* The "last use" doesn't follow the "first use"?? */
1089 /* Compute the benefit of eliminating the insns in the block whose
1090 last insn is LAST. This may be a group of insns used to compute a
1091 value directly or can contain a library call. */
1094 libcall_benefit (last)
1100 for (insn = XEXP (find_reg_note (last, REG_RETVAL, NULL_RTX), 0);
1101 insn != last; insn = NEXT_INSN (insn))
1103 if (GET_CODE (insn) == CALL_INSN)
1104 benefit += 10; /* Assume at least this many insns in a library
1106 else if (GET_CODE (insn) == INSN
1107 && GET_CODE (PATTERN (insn)) != USE
1108 && GET_CODE (PATTERN (insn)) != CLOBBER)
1115 /* Skip COUNT insns from INSN, counting library calls as 1 insn. */
1118 skip_consec_insns (insn, count)
1122 for (; count > 0; count--)
1126 /* If first insn of libcall sequence, skip to end. */
1127 /* Do this at start of loop, since INSN is guaranteed to
1129 if (GET_CODE (insn) != NOTE
1130 && (temp = find_reg_note (insn, REG_LIBCALL, NULL_RTX)))
1131 insn = XEXP (temp, 0);
1133 do insn = NEXT_INSN (insn);
1134 while (GET_CODE (insn) == NOTE);
1140 /* Ignore any movable whose insn falls within a libcall
1141 which is part of another movable.
1142 We make use of the fact that the movable for the libcall value
1143 was made later and so appears later on the chain. */
1146 ignore_some_movables (movables)
1147 struct movable *movables;
1149 register struct movable *m, *m1;
1151 for (m = movables; m; m = m->next)
1153 /* Is this a movable for the value of a libcall? */
1154 rtx note = find_reg_note (m->insn, REG_RETVAL, NULL_RTX);
1158 /* Check for earlier movables inside that range,
1159 and mark them invalid. We cannot use LUIDs here because
1160 insns created by loop.c for prior loops don't have LUIDs.
1161 Rather than reject all such insns from movables, we just
1162 explicitly check each insn in the libcall (since invariant
1163 libcalls aren't that common). */
1164 for (insn = XEXP (note, 0); insn != m->insn; insn = NEXT_INSN (insn))
1165 for (m1 = movables; m1 != m; m1 = m1->next)
1166 if (m1->insn == insn)
1172 /* For each movable insn, see if the reg that it loads
1173 leads when it dies right into another conditionally movable insn.
1174 If so, record that the second insn "forces" the first one,
1175 since the second can be moved only if the first is. */
1178 force_movables (movables)
1179 struct movable *movables;
1181 register struct movable *m, *m1;
1182 for (m1 = movables; m1; m1 = m1->next)
1183 /* Omit this if moving just the (SET (REG) 0) of a zero-extend. */
1184 if (!m1->partial && !m1->done)
1186 int regno = m1->regno;
1187 for (m = m1->next; m; m = m->next)
1188 /* ??? Could this be a bug? What if CSE caused the
1189 register of M1 to be used after this insn?
1190 Since CSE does not update regno_last_uid,
1191 this insn M->insn might not be where it dies.
1192 But very likely this doesn't matter; what matters is
1193 that M's reg is computed from M1's reg. */
1194 if (INSN_UID (m->insn) == regno_last_uid[regno]
1197 if (m != 0 && m->set_src == m1->set_dest
1198 /* If m->consec, m->set_src isn't valid. */
1202 /* Increase the priority of the moving the first insn
1203 since it permits the second to be moved as well. */
1207 m1->lifetime += m->lifetime;
1208 m1->savings += m1->savings;
1213 /* Find invariant expressions that are equal and can be combined into
1217 combine_movables (movables, nregs)
1218 struct movable *movables;
1221 register struct movable *m;
1222 char *matched_regs = (char *) alloca (nregs);
1223 enum machine_mode mode;
1225 /* Regs that are set more than once are not allowed to match
1226 or be matched. I'm no longer sure why not. */
1227 /* Perhaps testing m->consec_sets would be more appropriate here? */
1229 for (m = movables; m; m = m->next)
1230 if (m->match == 0 && n_times_used[m->regno] == 1 && !m->partial)
1232 register struct movable *m1;
1233 int regno = m->regno;
1234 rtx reg_note, reg_note1;
1236 bzero (matched_regs, nregs);
1237 matched_regs[regno] = 1;
1239 for (m1 = movables; m1; m1 = m1->next)
1240 if (m != m1 && m1->match == 0 && n_times_used[m1->regno] == 1
1241 /* A reg used outside the loop mustn't be eliminated. */
1243 /* A reg used for zero-extending mustn't be eliminated. */
1245 && (matched_regs[m1->regno]
1248 /* Can combine regs with different modes loaded from the
1249 same constant only if the modes are the same or
1250 if both are integer modes with M wider or the same
1251 width as M1. The check for integer is redundant, but
1252 safe, since the only case of differing destination
1253 modes with equal sources is when both sources are
1254 VOIDmode, i.e., CONST_INT. */
1255 (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest)
1256 || (GET_MODE_CLASS (GET_MODE (m->set_dest)) == MODE_INT
1257 && GET_MODE_CLASS (GET_MODE (m1->set_dest)) == MODE_INT
1258 && (GET_MODE_BITSIZE (GET_MODE (m->set_dest))
1259 >= GET_MODE_BITSIZE (GET_MODE (m1->set_dest)))))
1260 /* See if the source of M1 says it matches M. */
1261 && ((GET_CODE (m1->set_src) == REG
1262 && matched_regs[REGNO (m1->set_src)])
1263 || rtx_equal_for_loop_p (m->set_src, m1->set_src,
1265 && ((m->dependencies == m1->dependencies)
1266 || rtx_equal_p (m->dependencies, m1->dependencies)))
1268 m->lifetime += m1->lifetime;
1269 m->savings += m1->savings;
1272 matched_regs[m1->regno] = 1;
1276 /* Now combine the regs used for zero-extension.
1277 This can be done for those not marked `global'
1278 provided their lives don't overlap. */
1280 for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
1281 mode = GET_MODE_WIDER_MODE (mode))
1283 register struct movable *m0 = 0;
1285 /* Combine all the registers for extension from mode MODE.
1286 Don't combine any that are used outside this loop. */
1287 for (m = movables; m; m = m->next)
1288 if (m->partial && ! m->global
1289 && mode == GET_MODE (SET_SRC (PATTERN (NEXT_INSN (m->insn)))))
1291 register struct movable *m1;
1292 int first = uid_luid[regno_first_uid[m->regno]];
1293 int last = uid_luid[regno_last_uid[m->regno]];
1297 /* First one: don't check for overlap, just record it. */
1302 /* Make sure they extend to the same mode.
1303 (Almost always true.) */
1304 if (GET_MODE (m->set_dest) != GET_MODE (m0->set_dest))
1307 /* We already have one: check for overlap with those
1308 already combined together. */
1309 for (m1 = movables; m1 != m; m1 = m1->next)
1310 if (m1 == m0 || (m1->partial && m1->match == m0))
1311 if (! (uid_luid[regno_first_uid[m1->regno]] > last
1312 || uid_luid[regno_last_uid[m1->regno]] < first))
1315 /* No overlap: we can combine this with the others. */
1316 m0->lifetime += m->lifetime;
1317 m0->savings += m->savings;
1326 /* Return 1 if regs X and Y will become the same if moved. */
1329 regs_match_p (x, y, movables)
1331 struct movable *movables;
1335 struct movable *mx, *my;
1337 for (mx = movables; mx; mx = mx->next)
1338 if (mx->regno == xn)
1341 for (my = movables; my; my = my->next)
1342 if (my->regno == yn)
1346 && ((mx->match == my->match && mx->match != 0)
1348 || mx == my->match));
1351 /* Return 1 if X and Y are identical-looking rtx's.
1352 This is the Lisp function EQUAL for rtx arguments.
1354 If two registers are matching movables or a movable register and an
1355 equivalent constant, consider them equal. */
1358 rtx_equal_for_loop_p (x, y, movables)
1360 struct movable *movables;
1364 register struct movable *m;
1365 register enum rtx_code code;
1370 if (x == 0 || y == 0)
1373 code = GET_CODE (x);
1375 /* If we have a register and a constant, they may sometimes be
1377 if (GET_CODE (x) == REG && n_times_set[REGNO (x)] == -2
1379 for (m = movables; m; m = m->next)
1380 if (m->move_insn && m->regno == REGNO (x)
1381 && rtx_equal_p (m->set_src, y))
1384 else if (GET_CODE (y) == REG && n_times_set[REGNO (y)] == -2
1386 for (m = movables; m; m = m->next)
1387 if (m->move_insn && m->regno == REGNO (y)
1388 && rtx_equal_p (m->set_src, x))
1391 /* Otherwise, rtx's of different codes cannot be equal. */
1392 if (code != GET_CODE (y))
1395 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent.
1396 (REG:SI x) and (REG:HI x) are NOT equivalent. */
1398 if (GET_MODE (x) != GET_MODE (y))
1401 /* These three types of rtx's can be compared nonrecursively. */
1403 return (REGNO (x) == REGNO (y) || regs_match_p (x, y, movables));
1405 if (code == LABEL_REF)
1406 return XEXP (x, 0) == XEXP (y, 0);
1407 if (code == SYMBOL_REF)
1408 return XSTR (x, 0) == XSTR (y, 0);
1410 /* Compare the elements. If any pair of corresponding elements
1411 fail to match, return 0 for the whole things. */
1413 fmt = GET_RTX_FORMAT (code);
1414 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1419 if (XWINT (x, i) != XWINT (y, i))
1424 if (XINT (x, i) != XINT (y, i))
1429 /* Two vectors must have the same length. */
1430 if (XVECLEN (x, i) != XVECLEN (y, i))
1433 /* And the corresponding elements must match. */
1434 for (j = 0; j < XVECLEN (x, i); j++)
1435 if (rtx_equal_for_loop_p (XVECEXP (x, i, j), XVECEXP (y, i, j), movables) == 0)
1440 if (rtx_equal_for_loop_p (XEXP (x, i), XEXP (y, i), movables) == 0)
1445 if (strcmp (XSTR (x, i), XSTR (y, i)))
1450 /* These are just backpointers, so they don't matter. */
1456 /* It is believed that rtx's at this level will never
1457 contain anything but integers and other rtx's,
1458 except for within LABEL_REFs and SYMBOL_REFs. */
1466 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to all
1467 insns in INSNS which use thet reference. */
1470 add_label_notes (x, insns)
1474 enum rtx_code code = GET_CODE (x);
1479 if (code == LABEL_REF)
1481 for (insn = insns; insn; insn = NEXT_INSN (insn))
1482 if (reg_mentioned_p (XEXP (x, 0), insn))
1483 REG_NOTES (insn) = gen_rtx (EXPR_LIST, REG_LABEL, XEXP (x, 0),
1488 fmt = GET_RTX_FORMAT (code);
1489 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1492 add_label_notes (XEXP (x, i), insns);
1493 else if (fmt[i] == 'E')
1494 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1495 add_label_notes (XVECEXP (x, i, j), insns);
1499 /* Scan MOVABLES, and move the insns that deserve to be moved.
1500 If two matching movables are combined, replace one reg with the
1501 other throughout. */
1504 move_movables (movables, threshold, insn_count, loop_start, end, nregs)
1505 struct movable *movables;
1513 register struct movable *m;
1515 /* Map of pseudo-register replacements to handle combining
1516 when we move several insns that load the same value
1517 into different pseudo-registers. */
1518 rtx *reg_map = (rtx *) alloca (nregs * sizeof (rtx));
1519 char *already_moved = (char *) alloca (nregs);
1521 bzero (already_moved, nregs);
1522 bzero (reg_map, nregs * sizeof (rtx));
1526 for (m = movables; m; m = m->next)
1528 /* Describe this movable insn. */
1530 if (loop_dump_stream)
1532 fprintf (loop_dump_stream, "Insn %d: regno %d (life %d), ",
1533 INSN_UID (m->insn), m->regno, m->lifetime);
1535 fprintf (loop_dump_stream, "consec %d, ", m->consec);
1537 fprintf (loop_dump_stream, "cond ");
1539 fprintf (loop_dump_stream, "force ");
1541 fprintf (loop_dump_stream, "global ");
1543 fprintf (loop_dump_stream, "done ");
1545 fprintf (loop_dump_stream, "move-insn ");
1547 fprintf (loop_dump_stream, "matches %d ",
1548 INSN_UID (m->match->insn));
1550 fprintf (loop_dump_stream, "forces %d ",
1551 INSN_UID (m->forces->insn));
1554 /* Count movables. Value used in heuristics in strength_reduce. */
1557 /* Ignore the insn if it's already done (it matched something else).
1558 Otherwise, see if it is now safe to move. */
1562 || (1 == invariant_p (m->set_src)
1563 && (m->dependencies == 0
1564 || 1 == invariant_p (m->dependencies))
1566 || 1 == consec_sets_invariant_p (m->set_dest,
1569 && (! m->forces || m->forces->done))
1573 int savings = m->savings;
1575 /* We have an insn that is safe to move.
1576 Compute its desirability. */
1581 if (loop_dump_stream)
1582 fprintf (loop_dump_stream, "savings %d ", savings);
1584 if (moved_once[regno])
1588 if (loop_dump_stream)
1589 fprintf (loop_dump_stream, "halved since already moved ");
1592 /* An insn MUST be moved if we already moved something else
1593 which is safe only if this one is moved too: that is,
1594 if already_moved[REGNO] is nonzero. */
1596 /* An insn is desirable to move if the new lifetime of the
1597 register is no more than THRESHOLD times the old lifetime.
1598 If it's not desirable, it means the loop is so big
1599 that moving won't speed things up much,
1600 and it is liable to make register usage worse. */
1602 /* It is also desirable to move if it can be moved at no
1603 extra cost because something else was already moved. */
1605 if (already_moved[regno]
1606 || (threshold * savings * m->lifetime) >= insn_count
1607 || (m->forces && m->forces->done
1608 && n_times_used[m->forces->regno] == 1))
1611 register struct movable *m1;
1614 /* Now move the insns that set the reg. */
1616 if (m->partial && m->match)
1620 /* Find the end of this chain of matching regs.
1621 Thus, we load each reg in the chain from that one reg.
1622 And that reg is loaded with 0 directly,
1623 since it has ->match == 0. */
1624 for (m1 = m; m1->match; m1 = m1->match);
1625 newpat = gen_move_insn (SET_DEST (PATTERN (m->insn)),
1626 SET_DEST (PATTERN (m1->insn)));
1627 i1 = emit_insn_before (newpat, loop_start);
1629 /* Mark the moved, invariant reg as being allowed to
1630 share a hard reg with the other matching invariant. */
1631 REG_NOTES (i1) = REG_NOTES (m->insn);
1632 r1 = SET_DEST (PATTERN (m->insn));
1633 r2 = SET_DEST (PATTERN (m1->insn));
1634 regs_may_share = gen_rtx (EXPR_LIST, VOIDmode, r1,
1635 gen_rtx (EXPR_LIST, VOIDmode, r2,
1637 delete_insn (m->insn);
1642 if (loop_dump_stream)
1643 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1645 /* If we are to re-generate the item being moved with a
1646 new move insn, first delete what we have and then emit
1647 the move insn before the loop. */
1648 else if (m->move_insn)
1652 for (count = m->consec; count >= 0; count--)
1654 /* If this is the first insn of a library call sequence,
1656 if (GET_CODE (p) != NOTE
1657 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1660 /* If this is the last insn of a libcall sequence, then
1661 delete every insn in the sequence except the last.
1662 The last insn is handled in the normal manner. */
1663 if (GET_CODE (p) != NOTE
1664 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1666 temp = XEXP (temp, 0);
1668 temp = delete_insn (temp);
1671 p = delete_insn (p);
1675 emit_move_insn (m->set_dest, m->set_src);
1676 temp = get_insns ();
1679 add_label_notes (m->set_src, temp);
1681 i1 = emit_insns_before (temp, loop_start);
1682 if (! find_reg_note (i1, REG_EQUAL, NULL_RTX))
1684 = gen_rtx (EXPR_LIST,
1685 m->is_equiv ? REG_EQUIV : REG_EQUAL,
1686 m->set_src, REG_NOTES (i1));
1688 if (loop_dump_stream)
1689 fprintf (loop_dump_stream, " moved to %d", INSN_UID (i1));
1691 /* The more regs we move, the less we like moving them. */
1696 for (count = m->consec; count >= 0; count--)
1700 /* If first insn of libcall sequence, skip to end. */
1701 /* Do this at start of loop, since p is guaranteed to
1703 if (GET_CODE (p) != NOTE
1704 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
1707 /* If last insn of libcall sequence, move all
1708 insns except the last before the loop. The last
1709 insn is handled in the normal manner. */
1710 if (GET_CODE (p) != NOTE
1711 && (temp = find_reg_note (p, REG_RETVAL, NULL_RTX)))
1715 rtx fn_address_insn = 0;
1718 for (temp = XEXP (temp, 0); temp != p;
1719 temp = NEXT_INSN (temp))
1725 if (GET_CODE (temp) == NOTE)
1728 body = PATTERN (temp);
1730 /* Find the next insn after TEMP,
1731 not counting USE or NOTE insns. */
1732 for (next = NEXT_INSN (temp); next != p;
1733 next = NEXT_INSN (next))
1734 if (! (GET_CODE (next) == INSN
1735 && GET_CODE (PATTERN (next)) == USE)
1736 && GET_CODE (next) != NOTE)
1739 /* If that is the call, this may be the insn
1740 that loads the function address.
1742 Extract the function address from the insn
1743 that loads it into a register.
1744 If this insn was cse'd, we get incorrect code.
1746 So emit a new move insn that copies the
1747 function address into the register that the
1748 call insn will use. flow.c will delete any
1749 redundant stores that we have created. */
1750 if (GET_CODE (next) == CALL_INSN
1751 && GET_CODE (body) == SET
1752 && GET_CODE (SET_DEST (body)) == REG
1753 && (n = find_reg_note (temp, REG_EQUAL,
1756 fn_reg = SET_SRC (body);
1757 if (GET_CODE (fn_reg) != REG)
1758 fn_reg = SET_DEST (body);
1759 fn_address = XEXP (n, 0);
1760 fn_address_insn = temp;
1762 /* We have the call insn.
1763 If it uses the register we suspect it might,
1764 load it with the correct address directly. */
1765 if (GET_CODE (temp) == CALL_INSN
1767 && reg_mentioned_p (fn_reg, body))
1768 emit_insn_after (gen_move_insn (fn_reg,
1772 if (GET_CODE (temp) == CALL_INSN)
1773 i1 = emit_call_insn_before (body, loop_start);
1775 i1 = emit_insn_before (body, loop_start);
1778 if (temp == fn_address_insn)
1779 fn_address_insn = i1;
1780 REG_NOTES (i1) = REG_NOTES (temp);
1784 if (m->savemode != VOIDmode)
1786 /* P sets REG to zero; but we should clear only
1787 the bits that are not covered by the mode
1789 rtx reg = m->set_dest;
1795 (GET_MODE (reg), and_optab, reg,
1796 GEN_INT ((((HOST_WIDE_INT) 1
1797 << GET_MODE_BITSIZE (m->savemode)))
1799 reg, 1, OPTAB_LIB_WIDEN);
1803 emit_move_insn (reg, tem);
1804 sequence = gen_sequence ();
1806 i1 = emit_insn_before (sequence, loop_start);
1808 else if (GET_CODE (p) == CALL_INSN)
1809 i1 = emit_call_insn_before (PATTERN (p), loop_start);
1811 i1 = emit_insn_before (PATTERN (p), loop_start);
1813 REG_NOTES (i1) = REG_NOTES (p);
1818 if (loop_dump_stream)
1819 fprintf (loop_dump_stream, " moved to %d",
1823 /* This isn't needed because REG_NOTES is copied
1824 below and is wrong since P might be a PARALLEL. */
1825 if (REG_NOTES (i1) == 0
1826 && ! m->partial /* But not if it's a zero-extend clr. */
1827 && ! m->global /* and not if used outside the loop
1828 (since it might get set outside). */
1829 && CONSTANT_P (SET_SRC (PATTERN (p))))
1831 = gen_rtx (EXPR_LIST, REG_EQUAL,
1832 SET_SRC (PATTERN (p)), REG_NOTES (i1));
1835 /* If library call, now fix the REG_NOTES that contain
1836 insn pointers, namely REG_LIBCALL on FIRST
1837 and REG_RETVAL on I1. */
1838 if (temp = find_reg_note (i1, REG_RETVAL, NULL_RTX))
1840 XEXP (temp, 0) = first;
1841 temp = find_reg_note (first, REG_LIBCALL, NULL_RTX);
1842 XEXP (temp, 0) = i1;
1846 do p = NEXT_INSN (p);
1847 while (p && GET_CODE (p) == NOTE);
1850 /* The more regs we move, the less we like moving them. */
1854 /* Any other movable that loads the same register
1856 already_moved[regno] = 1;
1858 /* This reg has been moved out of one loop. */
1859 moved_once[regno] = 1;
1861 /* The reg set here is now invariant. */
1863 n_times_set[regno] = 0;
1867 /* Change the length-of-life info for the register
1868 to say it lives at least the full length of this loop.
1869 This will help guide optimizations in outer loops. */
1871 if (uid_luid[regno_first_uid[regno]] > INSN_LUID (loop_start))
1872 /* This is the old insn before all the moved insns.
1873 We can't use the moved insn because it is out of range
1874 in uid_luid. Only the old insns have luids. */
1875 regno_first_uid[regno] = INSN_UID (loop_start);
1876 if (uid_luid[regno_last_uid[regno]] < INSN_LUID (end))
1877 regno_last_uid[regno] = INSN_UID (end);
1879 /* Combine with this moved insn any other matching movables. */
1882 for (m1 = movables; m1; m1 = m1->next)
1887 /* Schedule the reg loaded by M1
1888 for replacement so that shares the reg of M.
1889 If the modes differ (only possible in restricted
1890 circumstances, make a SUBREG. */
1891 if (GET_MODE (m->set_dest) == GET_MODE (m1->set_dest))
1892 reg_map[m1->regno] = m->set_dest;
1895 = gen_lowpart_common (GET_MODE (m1->set_dest),
1898 /* Get rid of the matching insn
1899 and prevent further processing of it. */
1902 /* if library call, delete all insn except last, which
1904 if (temp = find_reg_note (m1->insn, REG_RETVAL,
1907 for (temp = XEXP (temp, 0); temp != m1->insn;
1908 temp = NEXT_INSN (temp))
1911 delete_insn (m1->insn);
1913 /* Any other movable that loads the same register
1915 already_moved[m1->regno] = 1;
1917 /* The reg merged here is now invariant,
1918 if the reg it matches is invariant. */
1920 n_times_set[m1->regno] = 0;
1923 else if (loop_dump_stream)
1924 fprintf (loop_dump_stream, "not desirable");
1926 else if (loop_dump_stream && !m->match)
1927 fprintf (loop_dump_stream, "not safe");
1929 if (loop_dump_stream)
1930 fprintf (loop_dump_stream, "\n");
1934 new_start = loop_start;
1936 /* Go through all the instructions in the loop, making
1937 all the register substitutions scheduled in REG_MAP. */
1938 for (p = new_start; p != end; p = NEXT_INSN (p))
1939 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
1940 || GET_CODE (p) == CALL_INSN)
1942 replace_regs (PATTERN (p), reg_map, nregs, 0);
1943 replace_regs (REG_NOTES (p), reg_map, nregs, 0);
1948 /* Scan X and replace the address of any MEM in it with ADDR.
1949 REG is the address that MEM should have before the replacement. */
1952 replace_call_address (x, reg, addr)
1955 register enum rtx_code code;
1961 code = GET_CODE (x);
1975 /* Short cut for very common case. */
1976 replace_call_address (XEXP (x, 1), reg, addr);
1980 /* Short cut for very common case. */
1981 replace_call_address (XEXP (x, 0), reg, addr);
1985 /* If this MEM uses a reg other than the one we expected,
1986 something is wrong. */
1987 if (XEXP (x, 0) != reg)
1993 fmt = GET_RTX_FORMAT (code);
1994 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1997 replace_call_address (XEXP (x, i), reg, addr);
2001 for (j = 0; j < XVECLEN (x, i); j++)
2002 replace_call_address (XVECEXP (x, i, j), reg, addr);
2008 /* Return the number of memory refs to addresses that vary
2012 count_nonfixed_reads (x)
2015 register enum rtx_code code;
2023 code = GET_CODE (x);
2037 return ((invariant_p (XEXP (x, 0)) != 1)
2038 + count_nonfixed_reads (XEXP (x, 0)));
2042 fmt = GET_RTX_FORMAT (code);
2043 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2046 value += count_nonfixed_reads (XEXP (x, i));
2050 for (j = 0; j < XVECLEN (x, i); j++)
2051 value += count_nonfixed_reads (XVECEXP (x, i, j));
2059 /* P is an instruction that sets a register to the result of a ZERO_EXTEND.
2060 Replace it with an instruction to load just the low bytes
2061 if the machine supports such an instruction,
2062 and insert above LOOP_START an instruction to clear the register. */
2065 constant_high_bytes (p, loop_start)
2069 register int insn_code_number;
2071 /* Try to change (SET (REG ...) (ZERO_EXTEND (..:B ...)))
2072 to (SET (STRICT_LOW_PART (SUBREG:B (REG...))) ...). */
2074 new = gen_rtx (SET, VOIDmode,
2075 gen_rtx (STRICT_LOW_PART, VOIDmode,
2076 gen_rtx (SUBREG, GET_MODE (XEXP (SET_SRC (PATTERN (p)), 0)),
2077 SET_DEST (PATTERN (p)),
2079 XEXP (SET_SRC (PATTERN (p)), 0));
2080 insn_code_number = recog (new, p);
2082 if (insn_code_number)
2086 /* Clear destination register before the loop. */
2087 emit_insn_before (gen_rtx (SET, VOIDmode,
2088 SET_DEST (PATTERN (p)),
2092 /* Inside the loop, just load the low part. */
2098 /* Scan a loop setting the variables `unknown_address_altered',
2099 `num_mem_sets', `loop_continue', loops_enclosed', `loop_has_call',
2100 and `loop_has_volatile'.
2101 Also, fill in the array `loop_store_mems'. */
2104 prescan_loop (start, end)
2107 register int level = 1;
2110 unknown_address_altered = 0;
2112 loop_has_volatile = 0;
2113 loop_store_mems_idx = 0;
2119 for (insn = NEXT_INSN (start); insn != NEXT_INSN (end);
2120 insn = NEXT_INSN (insn))
2122 if (GET_CODE (insn) == NOTE)
2124 if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
2127 /* Count number of loops contained in this one. */
2130 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
2139 else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_CONT)
2142 loop_continue = insn;
2145 else if (GET_CODE (insn) == CALL_INSN)
2147 unknown_address_altered = 1;
2152 if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
2154 if (volatile_refs_p (PATTERN (insn)))
2155 loop_has_volatile = 1;
2157 note_stores (PATTERN (insn), note_addr_stored);
2163 /* Scan the function looking for loops. Record the start and end of each loop.
2164 Also mark as invalid loops any loops that contain a setjmp or are branched
2165 to from outside the loop. */
2168 find_and_verify_loops (f)
2172 int current_loop = -1;
2176 /* If there are jumps to undefined labels,
2177 treat them as jumps out of any/all loops.
2178 This also avoids writing past end of tables when there are no loops. */
2179 uid_loop_num[0] = -1;
2181 /* Find boundaries of loops, mark which loops are contained within
2182 loops, and invalidate loops that have setjmp. */
2184 for (insn = f; insn; insn = NEXT_INSN (insn))
2186 if (GET_CODE (insn) == NOTE)
2187 switch (NOTE_LINE_NUMBER (insn))
2189 case NOTE_INSN_LOOP_BEG:
2190 loop_number_loop_starts[++next_loop] = insn;
2191 loop_number_loop_ends[next_loop] = 0;
2192 loop_outer_loop[next_loop] = current_loop;
2193 loop_invalid[next_loop] = 0;
2194 loop_number_exit_labels[next_loop] = 0;
2195 current_loop = next_loop;
2198 case NOTE_INSN_SETJMP:
2199 /* In this case, we must invalidate our current loop and any
2201 for (loop = current_loop; loop != -1; loop = loop_outer_loop[loop])
2203 loop_invalid[loop] = 1;
2204 if (loop_dump_stream)
2205 fprintf (loop_dump_stream,
2206 "\nLoop at %d ignored due to setjmp.\n",
2207 INSN_UID (loop_number_loop_starts[loop]));
2211 case NOTE_INSN_LOOP_END:
2212 if (current_loop == -1)
2215 loop_number_loop_ends[current_loop] = insn;
2216 current_loop = loop_outer_loop[current_loop];
2221 /* Note that this will mark the NOTE_INSN_LOOP_END note as being in the
2222 enclosing loop, but this doesn't matter. */
2223 uid_loop_num[INSN_UID (insn)] = current_loop;
2226 /* Now scan all JUMP_INSN's in the function. If any branches into a loop
2227 that it is not contained within, that loop is marked invalid.
2229 Also look for blocks of code ending in an unconditional branch that
2230 exits the loop. If such a block is surrounded by a conditional
2231 branch around the block, move the block elsewhere (see below) and
2232 invert the jump to point to the code block. This may eliminate a
2233 label in our loop and will simplify processing by both us and a
2234 possible second cse pass. */
2236 for (insn = f; insn; insn = NEXT_INSN (insn))
2237 if (GET_CODE (insn) == JUMP_INSN)
2239 int this_loop_num = uid_loop_num[INSN_UID (insn)];
2241 mark_loop_jump (PATTERN (insn), this_loop_num);
2243 /* See if this is an unconditional branch outside the loop. */
2244 if (this_loop_num != -1
2245 && (GET_CODE (PATTERN (insn)) == RETURN
2246 || (simplejump_p (insn)
2247 && (uid_loop_num[INSN_UID (JUMP_LABEL (insn))]
2248 != this_loop_num))))
2251 rtx our_next = next_real_insn (insn);
2253 /* Go backwards until we reach the start of the loop, a label,
2255 for (p = PREV_INSN (insn);
2256 GET_CODE (p) != CODE_LABEL
2257 && ! (GET_CODE (p) == NOTE
2258 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_BEG)
2259 && GET_CODE (p) != JUMP_INSN;
2263 /* If we stopped on a JUMP_INSN to the next insn after INSN,
2264 we have a block of code to try to move.
2266 We look backward and then forward from the target of INSN
2267 to find a BARRIER at the same loop depth as the target.
2268 If we find such a BARRIER, we make a new label for the start
2269 of the block, invert the jump in P and point it to that label,
2270 and move the block of code to the spot we found. */
2272 if (GET_CODE (p) == JUMP_INSN
2273 && JUMP_LABEL (p) != 0
2275 && ! simplejump_p (p)
2276 && next_real_insn (JUMP_LABEL (p)) == our_next)
2279 = JUMP_LABEL (insn) ? JUMP_LABEL (insn) : get_last_insn ();
2280 int target_loop_num = uid_loop_num[INSN_UID (target)];
2283 for (loc = target; loc; loc = PREV_INSN (loc))
2284 if (GET_CODE (loc) == BARRIER
2285 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2289 for (loc = target; loc; loc = NEXT_INSN (loc))
2290 if (GET_CODE (loc) == BARRIER
2291 && uid_loop_num[INSN_UID (loc)] == target_loop_num)
2296 rtx cond_label = JUMP_LABEL (p);
2297 rtx new_label = get_label_after (p);
2299 /* Ensure our label doesn't go away. */
2300 LABEL_NUSES (cond_label)++;
2302 /* Verify that uid_loop_num is large enough and that
2304 if (INSN_UID (new_label) < max_uid_for_loop
2305 && invert_jump (p, new_label))
2309 /* Include the BARRIER after INSN and copy the
2311 new_label = squeeze_notes (new_label, NEXT_INSN (insn));
2312 reorder_insns (new_label, NEXT_INSN (insn), loc);
2314 /* All those insns are now in TARGET_LOOP_NUM. */
2315 for (q = new_label; q != NEXT_INSN (NEXT_INSN (insn));
2317 uid_loop_num[INSN_UID (q)] = target_loop_num;
2319 /* The label jumped to by INSN is no longer a loop exit.
2320 Unless INSN does not have a label (e.g., it is a
2321 RETURN insn), search loop_number_exit_labels to find
2322 its label_ref, and remove it. Also turn off
2323 LABEL_OUTSIDE_LOOP_P bit. */
2324 if (JUMP_LABEL (insn))
2327 r = loop_number_exit_labels[this_loop_num];
2328 r; q = r, r = LABEL_NEXTREF (r))
2329 if (XEXP (r, 0) == JUMP_LABEL (insn))
2331 LABEL_OUTSIDE_LOOP_P (r) = 0;
2333 LABEL_NEXTREF (q) = LABEL_NEXTREF (r);
2335 loop_number_exit_labels[this_loop_num]
2336 = LABEL_NEXTREF (r);
2340 /* If we didn't find it, then something is wrong. */
2345 /* P is now a jump outside the loop, so it must be put
2346 in loop_number_exit_labels, and marked as such.
2347 The easiest way to do this is to just call
2348 mark_loop_jump again for P. */
2349 mark_loop_jump (PATTERN (p), this_loop_num);
2351 /* If INSN now jumps to the insn after it,
2353 if (JUMP_LABEL (insn) != 0
2354 && (next_real_insn (JUMP_LABEL (insn))
2355 == next_real_insn (insn)))
2359 /* Continue the loop after where the conditional
2360 branch used to jump, since the only branch insn
2361 in the block (if it still remains) is an inter-loop
2362 branch and hence needs no processing. */
2363 insn = NEXT_INSN (cond_label);
2365 if (--LABEL_NUSES (cond_label) == 0)
2366 delete_insn (cond_label);
2373 /* If any label in X jumps to a loop different from LOOP_NUM and any of the
2374 loops it is contained in, mark the target loop invalid.
2376 For speed, we assume that X is part of a pattern of a JUMP_INSN. */
2379 mark_loop_jump (x, loop_num)
2387 switch (GET_CODE (x))
2400 /* There could be a label reference in here. */
2401 mark_loop_jump (XEXP (x, 0), loop_num);
2408 mark_loop_jump (XEXP (x, 0), loop_num);
2409 mark_loop_jump (XEXP (x, 1), loop_num);
2414 mark_loop_jump (XEXP (x, 0), loop_num);
2418 dest_loop = uid_loop_num[INSN_UID (XEXP (x, 0))];
2420 /* Link together all labels that branch outside the loop. This
2421 is used by final_[bg]iv_value and the loop unrolling code. Also
2422 mark this LABEL_REF so we know that this branch should predict
2425 if (dest_loop != loop_num && loop_num != -1)
2427 LABEL_OUTSIDE_LOOP_P (x) = 1;
2428 LABEL_NEXTREF (x) = loop_number_exit_labels[loop_num];
2429 loop_number_exit_labels[loop_num] = x;
2432 /* If this is inside a loop, but not in the current loop or one enclosed
2433 by it, it invalidates at least one loop. */
2435 if (dest_loop == -1)
2438 /* We must invalidate every nested loop containing the target of this
2439 label, except those that also contain the jump insn. */
2441 for (; dest_loop != -1; dest_loop = loop_outer_loop[dest_loop])
2443 /* Stop when we reach a loop that also contains the jump insn. */
2444 for (outer_loop = loop_num; outer_loop != -1;
2445 outer_loop = loop_outer_loop[outer_loop])
2446 if (dest_loop == outer_loop)
2449 /* If we get here, we know we need to invalidate a loop. */
2450 if (loop_dump_stream && ! loop_invalid[dest_loop])
2451 fprintf (loop_dump_stream,
2452 "\nLoop at %d ignored due to multiple entry points.\n",
2453 INSN_UID (loop_number_loop_starts[dest_loop]));
2455 loop_invalid[dest_loop] = 1;
2460 /* If this is not setting pc, ignore. */
2461 if (SET_DEST (x) == pc_rtx)
2462 mark_loop_jump (SET_SRC (x), loop_num);
2466 mark_loop_jump (XEXP (x, 1), loop_num);
2467 mark_loop_jump (XEXP (x, 2), loop_num);
2472 for (i = 0; i < XVECLEN (x, 0); i++)
2473 mark_loop_jump (XVECEXP (x, 0, i), loop_num);
2477 for (i = 0; i < XVECLEN (x, 1); i++)
2478 mark_loop_jump (XVECEXP (x, 1, i), loop_num);
2482 /* Nothing else should occur in a JUMP_INSN. */
2487 /* Return nonzero if there is a label in the range from
2488 insn INSN to and including the insn whose luid is END
2489 INSN must have an assigned luid (i.e., it must not have
2490 been previously created by loop.c). */
2493 labels_in_range_p (insn, end)
2497 while (insn && INSN_LUID (insn) <= end)
2499 if (GET_CODE (insn) == CODE_LABEL)
2501 insn = NEXT_INSN (insn);
2507 /* Record that a memory reference X is being set. */
2510 note_addr_stored (x)
2515 if (x == 0 || GET_CODE (x) != MEM)
2518 /* Count number of memory writes.
2519 This affects heuristics in strength_reduce. */
2522 if (unknown_address_altered)
2525 for (i = 0; i < loop_store_mems_idx; i++)
2526 if (rtx_equal_p (XEXP (loop_store_mems[i], 0), XEXP (x, 0))
2527 && MEM_IN_STRUCT_P (x) == MEM_IN_STRUCT_P (loop_store_mems[i]))
2529 /* We are storing at the same address as previously noted. Save the
2530 wider reference, treating BLKmode as wider. */
2531 if (GET_MODE (x) == BLKmode
2532 || (GET_MODE_SIZE (GET_MODE (x))
2533 > GET_MODE_SIZE (GET_MODE (loop_store_mems[i]))))
2534 loop_store_mems[i] = x;
2538 if (i == NUM_STORES)
2539 unknown_address_altered = 1;
2541 else if (i == loop_store_mems_idx)
2542 loop_store_mems[loop_store_mems_idx++] = x;
2545 /* Return nonzero if the rtx X is invariant over the current loop.
2547 The value is 2 if we refer to something only conditionally invariant.
2549 If `unknown_address_altered' is nonzero, no memory ref is invariant.
2550 Otherwise, a memory ref is invariant if it does not conflict with
2551 anything stored in `loop_store_mems'. */
2558 register enum rtx_code code;
2560 int conditional = 0;
2564 code = GET_CODE (x);
2574 /* A LABEL_REF is normally invariant, however, if we are unrolling
2575 loops, and this label is inside the loop, then it isn't invariant.
2576 This is because each unrolled copy of the loop body will have
2577 a copy of this label. If this was invariant, then an insn loading
2578 the address of this label into a register might get moved outside
2579 the loop, and then each loop body would end up using the same label.
2581 We don't know the loop bounds here though, so just fail for all
2583 if (flag_unroll_loops)
2590 case UNSPEC_VOLATILE:
2594 /* We used to check RTX_UNCHANGING_P (x) here, but that is invalid
2595 since the reg might be set by initialization within the loop. */
2596 if (x == frame_pointer_rtx || x == arg_pointer_rtx)
2599 && REGNO (x) < FIRST_PSEUDO_REGISTER && call_used_regs[REGNO (x)])
2601 if (n_times_set[REGNO (x)] < 0)
2603 return n_times_set[REGNO (x)] == 0;
2606 /* Read-only items (such as constants in a constant pool) are
2607 invariant if their address is. */
2608 if (RTX_UNCHANGING_P (x))
2611 /* If we filled the table (or had a subroutine call), any location
2612 in memory could have been clobbered. */
2613 if (unknown_address_altered
2614 /* Don't mess with volatile memory references. */
2615 || MEM_VOLATILE_P (x))
2618 /* See if there is any dependence between a store and this load. */
2619 for (i = loop_store_mems_idx - 1; i >= 0; i--)
2620 if (true_dependence (loop_store_mems[i], x))
2623 /* It's not invalidated by a store in memory
2624 but we must still verify the address is invariant. */
2628 /* Don't mess with insns declared volatile. */
2629 if (MEM_VOLATILE_P (x))
2633 fmt = GET_RTX_FORMAT (code);
2634 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2638 int tem = invariant_p (XEXP (x, i));
2644 else if (fmt[i] == 'E')
2647 for (j = 0; j < XVECLEN (x, i); j++)
2649 int tem = invariant_p (XVECEXP (x, i, j));
2659 return 1 + conditional;
2662 /* Return 1 if OTHER (a mem ref) overlaps the area of memory
2663 which is SIZE bytes starting at BASE. */
2666 addr_overlap_p (other, base, size)
2671 HOST_WIDE_INT start = 0, end;
2673 if (GET_CODE (base) == CONST)
2674 base = XEXP (base, 0);
2675 if (GET_CODE (base) == PLUS
2676 && GET_CODE (XEXP (base, 1)) == CONST_INT)
2678 start = INTVAL (XEXP (base, 1));
2679 base = XEXP (base, 0);
2683 return refers_to_mem_p (other, base, start, end);
2686 /* Return nonzero if all the insns in the loop that set REG
2687 are INSN and the immediately following insns,
2688 and if each of those insns sets REG in an invariant way
2689 (not counting uses of REG in them).
2691 The value is 2 if some of these insns are only conditionally invariant.
2693 We assume that INSN itself is the first set of REG
2694 and that its source is invariant. */
2697 consec_sets_invariant_p (reg, n_sets, insn)
2701 register rtx p = insn;
2702 register int regno = REGNO (reg);
2704 /* Number of sets we have to insist on finding after INSN. */
2705 int count = n_sets - 1;
2706 int old = n_times_set[regno];
2710 /* If N_SETS hit the limit, we can't rely on its value. */
2714 n_times_set[regno] = 0;
2718 register enum rtx_code code;
2722 code = GET_CODE (p);
2724 /* If library call, skip to end of of it. */
2725 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
2730 && (set = single_set (p))
2731 && GET_CODE (SET_DEST (set)) == REG
2732 && REGNO (SET_DEST (set)) == regno)
2734 this = invariant_p (SET_SRC (set));
2737 else if (temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
2739 this = invariant_p (XEXP (temp, 0));
2746 else if (code != NOTE)
2748 n_times_set[regno] = old;
2753 n_times_set[regno] = old;
2754 /* If invariant_p ever returned 2, we return 2. */
2755 return 1 + (value & 2);
2759 /* I don't think this condition is sufficient to allow INSN
2760 to be moved, so we no longer test it. */
2762 /* Return 1 if all insns in the basic block of INSN and following INSN
2763 that set REG are invariant according to TABLE. */
2766 all_sets_invariant_p (reg, insn, table)
2770 register rtx p = insn;
2771 register int regno = REGNO (reg);
2775 register enum rtx_code code;
2777 code = GET_CODE (p);
2778 if (code == CODE_LABEL || code == JUMP_INSN)
2780 if (code == INSN && GET_CODE (PATTERN (p)) == SET
2781 && GET_CODE (SET_DEST (PATTERN (p))) == REG
2782 && REGNO (SET_DEST (PATTERN (p))) == regno)
2784 if (!invariant_p (SET_SRC (PATTERN (p)), table))
2791 /* Look at all uses (not sets) of registers in X. For each, if it is
2792 the single use, set USAGE[REGNO] to INSN; if there was a previous use in
2793 a different insn, set USAGE[REGNO] to const0_rtx. */
2796 find_single_use_in_loop (insn, x, usage)
2801 enum rtx_code code = GET_CODE (x);
2802 char *fmt = GET_RTX_FORMAT (code);
2807 = (usage[REGNO (x)] != 0 && usage[REGNO (x)] != insn)
2808 ? const0_rtx : insn;
2810 else if (code == SET)
2812 /* Don't count SET_DEST if it is a REG; otherwise count things
2813 in SET_DEST because if a register is partially modified, it won't
2814 show up as a potential movable so we don't care how USAGE is set
2816 if (GET_CODE (SET_DEST (x)) != REG)
2817 find_single_use_in_loop (insn, SET_DEST (x), usage);
2818 find_single_use_in_loop (insn, SET_SRC (x), usage);
2821 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2823 if (fmt[i] == 'e' && XEXP (x, i) != 0)
2824 find_single_use_in_loop (insn, XEXP (x, i), usage);
2825 else if (fmt[i] == 'E')
2826 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2827 find_single_use_in_loop (insn, XVECEXP (x, i, j), usage);
2831 /* Increment N_TIMES_SET at the index of each register
2832 that is modified by an insn between FROM and TO.
2833 If the value of an element of N_TIMES_SET becomes 127 or more,
2834 stop incrementing it, to avoid overflow.
2836 Store in SINGLE_USAGE[I] the single insn in which register I is
2837 used, if it is only used once. Otherwise, it is set to 0 (for no
2838 uses) or const0_rtx for more than one use. This parameter may be zero,
2839 in which case this processing is not done.
2841 Store in *COUNT_PTR the number of actual instruction
2842 in the loop. We use this to decide what is worth moving out. */
2844 /* last_set[n] is nonzero iff reg n has been set in the current basic block.
2845 In that case, it is the insn that last set reg n. */
2848 count_loop_regs_set (from, to, may_not_move, single_usage, count_ptr, nregs)
2849 register rtx from, to;
2855 register rtx *last_set = (rtx *) alloca (nregs * sizeof (rtx));
2857 register int count = 0;
2860 bzero (last_set, nregs * sizeof (rtx));
2861 for (insn = from; insn != to; insn = NEXT_INSN (insn))
2863 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2867 /* If requested, record registers that have exactly one use. */
2870 find_single_use_in_loop (insn, PATTERN (insn), single_usage);
2872 /* Include uses in REG_EQUAL notes. */
2873 if (REG_NOTES (insn))
2874 find_single_use_in_loop (insn, REG_NOTES (insn), single_usage);
2877 if (GET_CODE (PATTERN (insn)) == CLOBBER
2878 && GET_CODE (XEXP (PATTERN (insn), 0)) == REG)
2879 /* Don't move a reg that has an explicit clobber.
2880 We might do so sometimes, but it's not worth the pain. */
2881 may_not_move[REGNO (XEXP (PATTERN (insn), 0))] = 1;
2883 if (GET_CODE (PATTERN (insn)) == SET
2884 || GET_CODE (PATTERN (insn)) == CLOBBER)
2886 dest = SET_DEST (PATTERN (insn));
2887 while (GET_CODE (dest) == SUBREG
2888 || GET_CODE (dest) == ZERO_EXTRACT
2889 || GET_CODE (dest) == SIGN_EXTRACT
2890 || GET_CODE (dest) == STRICT_LOW_PART)
2891 dest = XEXP (dest, 0);
2892 if (GET_CODE (dest) == REG)
2894 register int regno = REGNO (dest);
2895 /* If this is the first setting of this reg
2896 in current basic block, and it was set before,
2897 it must be set in two basic blocks, so it cannot
2898 be moved out of the loop. */
2899 if (n_times_set[regno] > 0 && last_set[regno] == 0)
2900 may_not_move[regno] = 1;
2901 /* If this is not first setting in current basic block,
2902 see if reg was used in between previous one and this.
2903 If so, neither one can be moved. */
2904 if (last_set[regno] != 0
2905 && reg_used_between_p (dest, last_set[regno], insn))
2906 may_not_move[regno] = 1;
2907 if (n_times_set[regno] < 127)
2908 ++n_times_set[regno];
2909 last_set[regno] = insn;
2912 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
2915 for (i = XVECLEN (PATTERN (insn), 0) - 1; i >= 0; i--)
2917 register rtx x = XVECEXP (PATTERN (insn), 0, i);
2918 if (GET_CODE (x) == CLOBBER && GET_CODE (XEXP (x, 0)) == REG)
2919 /* Don't move a reg that has an explicit clobber.
2920 It's not worth the pain to try to do it correctly. */
2921 may_not_move[REGNO (XEXP (x, 0))] = 1;
2923 if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
2925 dest = SET_DEST (x);
2926 while (GET_CODE (dest) == SUBREG
2927 || GET_CODE (dest) == ZERO_EXTRACT
2928 || GET_CODE (dest) == SIGN_EXTRACT
2929 || GET_CODE (dest) == STRICT_LOW_PART)
2930 dest = XEXP (dest, 0);
2931 if (GET_CODE (dest) == REG)
2933 register int regno = REGNO (dest);
2934 if (n_times_set[regno] > 0 && last_set[regno] == 0)
2935 may_not_move[regno] = 1;
2936 if (last_set[regno] != 0
2937 && reg_used_between_p (dest, last_set[regno], insn))
2938 may_not_move[regno] = 1;
2939 if (n_times_set[regno] < 127)
2940 ++n_times_set[regno];
2941 last_set[regno] = insn;
2947 if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN)
2948 bzero (last_set, nregs * sizeof (rtx));
2953 /* Given a loop that is bounded by LOOP_START and LOOP_END
2954 and that is entered at SCAN_START,
2955 return 1 if the register set in SET contained in insn INSN is used by
2956 any insn that precedes INSN in cyclic order starting
2957 from the loop entry point.
2959 We don't want to use INSN_LUID here because if we restrict INSN to those
2960 that have a valid INSN_LUID, it means we cannot move an invariant out
2961 from an inner loop past two loops. */
2964 loop_reg_used_before_p (set, insn, loop_start, scan_start, loop_end)
2965 rtx set, insn, loop_start, scan_start, loop_end;
2967 rtx reg = SET_DEST (set);
2970 /* Scan forward checking for register usage. If we hit INSN, we
2971 are done. Otherwise, if we hit LOOP_END, wrap around to LOOP_START. */
2972 for (p = scan_start; p != insn; p = NEXT_INSN (p))
2974 if (GET_RTX_CLASS (GET_CODE (p)) == 'i'
2975 && reg_overlap_mentioned_p (reg, PATTERN (p)))
2985 /* A "basic induction variable" or biv is a pseudo reg that is set
2986 (within this loop) only by incrementing or decrementing it. */
2987 /* A "general induction variable" or giv is a pseudo reg whose
2988 value is a linear function of a biv. */
2990 /* Bivs are recognized by `basic_induction_var';
2991 Givs by `general_induct_var'. */
2993 /* Indexed by register number, indicates whether or not register is an
2994 induction variable, and if so what type. */
2996 enum iv_mode *reg_iv_type;
2998 /* Indexed by register number, contains pointer to `struct induction'
2999 if register is an induction variable. This holds general info for
3000 all induction variables. */
3002 struct induction **reg_iv_info;
3004 /* Indexed by register number, contains pointer to `struct iv_class'
3005 if register is a basic induction variable. This holds info describing
3006 the class (a related group) of induction variables that the biv belongs
3009 struct iv_class **reg_biv_class;
3011 /* The head of a list which links together (via the next field)
3012 every iv class for the current loop. */
3014 struct iv_class *loop_iv_list;
3016 /* Communication with routines called via `note_stores'. */
3018 static rtx note_insn;
3020 /* Dummy register to have non-zero DEST_REG for DEST_ADDR type givs. */
3022 static rtx addr_placeholder;
3024 /* ??? Unfinished optimizations, and possible future optimizations,
3025 for the strength reduction code. */
3027 /* ??? There is one more optimization you might be interested in doing: to
3028 allocate pseudo registers for frequently-accessed memory locations.
3029 If the same memory location is referenced each time around, it might
3030 be possible to copy it into a register before and out after.
3031 This is especially useful when the memory location is a variable which
3032 is in a stack slot because somewhere its address is taken. If the
3033 loop doesn't contain a function call and the variable isn't volatile,
3034 it is safe to keep the value in a register for the duration of the
3035 loop. One tricky thing is that the copying of the value back from the
3036 register has to be done on all exits from the loop. You need to check that
3037 all the exits from the loop go to the same place. */
3039 /* ??? The interaction of biv elimination, and recognition of 'constant'
3040 bivs, may cause problems. */
3042 /* ??? Add heuristics so that DEST_ADDR strength reduction does not cause
3043 performance problems.
3045 Perhaps don't eliminate things that can be combined with an addressing
3046 mode. Find all givs that have the same biv, mult_val, and add_val;
3047 then for each giv, check to see if its only use dies in a following
3048 memory address. If so, generate a new memory address and check to see
3049 if it is valid. If it is valid, then store the modified memory address,
3050 otherwise, mark the giv as not done so that it will get its own iv. */
3052 /* ??? Could try to optimize branches when it is known that a biv is always
3055 /* ??? When replace a biv in a compare insn, we should replace with closest
3056 giv so that an optimized branch can still be recognized by the combiner,
3057 e.g. the VAX acb insn. */
3059 /* ??? Many of the checks involving uid_luid could be simplified if regscan
3060 was rerun in loop_optimize whenever a register was added or moved.
3061 Also, some of the optimizations could be a little less conservative. */
3063 /* Perform strength reduction and induction variable elimination. */
3065 /* Pseudo registers created during this function will be beyond the last
3066 valid index in several tables including n_times_set and regno_last_uid.
3067 This does not cause a problem here, because the added registers cannot be
3068 givs outside of their loop, and hence will never be reconsidered.
3069 But scan_loop must check regnos to make sure they are in bounds. */
3072 strength_reduce (scan_start, end, loop_top, insn_count,
3073 loop_start, loop_end)
3086 /* This is 1 if current insn is not executed at least once for every loop
3088 int not_every_iteration = 0;
3089 /* This is 1 if current insn may be executed more than once for every
3091 int maybe_multiple = 0;
3092 /* Temporary list pointers for traversing loop_iv_list. */
3093 struct iv_class *bl, **backbl;
3094 /* Ratio of extra register life span we can justify
3095 for saving an instruction. More if loop doesn't call subroutines
3096 since in that case saving an insn makes more difference
3097 and more registers are available. */
3098 /* ??? could set this to last value of threshold in move_movables */
3099 int threshold = (loop_has_call ? 1 : 2) * (3 + n_non_fixed_regs);
3100 /* Map of pseudo-register replacements. */
3104 rtx end_insert_before;
3106 reg_iv_type = (enum iv_mode *) alloca (max_reg_before_loop
3107 * sizeof (enum iv_mode *));
3108 bzero ((char *) reg_iv_type, max_reg_before_loop * sizeof (enum iv_mode *));
3109 reg_iv_info = (struct induction **)
3110 alloca (max_reg_before_loop * sizeof (struct induction *));
3111 bzero ((char *) reg_iv_info, (max_reg_before_loop
3112 * sizeof (struct induction *)));
3113 reg_biv_class = (struct iv_class **)
3114 alloca (max_reg_before_loop * sizeof (struct iv_class *));
3115 bzero ((char *) reg_biv_class, (max_reg_before_loop
3116 * sizeof (struct iv_class *)));
3119 addr_placeholder = gen_reg_rtx (Pmode);
3121 /* Save insn immediately after the loop_end. Insns inserted after loop_end
3122 must be put before this insn, so that they will appear in the right
3123 order (i.e. loop order). */
3125 end_insert_before = NEXT_INSN (loop_end);
3127 /* Scan through loop to find all possible bivs. */
3133 /* At end of a straight-in loop, we are done.
3134 At end of a loop entered at the bottom, scan the top. */
3135 if (p == scan_start)
3140 p = NEXT_INSN (loop_top);
3143 if (p == scan_start)
3147 if (GET_CODE (p) == INSN
3148 && (set = single_set (p))
3149 && GET_CODE (SET_DEST (set)) == REG)
3151 dest_reg = SET_DEST (set);
3152 if (REGNO (dest_reg) < max_reg_before_loop
3153 && REGNO (dest_reg) >= FIRST_PSEUDO_REGISTER
3154 && reg_iv_type[REGNO (dest_reg)] != NOT_BASIC_INDUCT)
3156 if (basic_induction_var (SET_SRC (set), dest_reg,
3157 &inc_val, &mult_val))
3159 /* It is a possible basic induction variable.
3160 Create and initialize an induction structure for it. */
3163 = (struct induction *) alloca (sizeof (struct induction));
3165 record_biv (v, p, dest_reg, inc_val, mult_val,
3166 not_every_iteration, maybe_multiple);
3167 reg_iv_type[REGNO (dest_reg)] = BASIC_INDUCT;
3169 else if (REGNO (dest_reg) < max_reg_before_loop)
3170 reg_iv_type[REGNO (dest_reg)] = NOT_BASIC_INDUCT;
3174 /* Past CODE_LABEL, we get to insns that may be executed multiple
3175 times. The only way we can be sure that they can't is if every
3176 every jump insn between here and the end of the loop either
3177 returns, exits the loop, or is a forward jump. */
3179 if (GET_CODE (p) == CODE_LABEL)
3187 insn = NEXT_INSN (insn);
3188 if (insn == scan_start)
3193 insn = NEXT_INSN (loop_top);
3196 if (insn == scan_start)
3200 if (GET_CODE (insn) == JUMP_INSN
3201 && GET_CODE (PATTERN (insn)) != RETURN
3202 && (! condjump_p (insn)
3203 || (JUMP_LABEL (insn) != 0
3204 && (INSN_UID (JUMP_LABEL (insn)) >= max_uid_for_loop
3205 || INSN_UID (insn) >= max_uid_for_loop
3206 || (INSN_LUID (JUMP_LABEL (insn))
3207 < INSN_LUID (insn))))))
3215 /* Past a label or a jump, we get to insns for which we can't count
3216 on whether or how many times they will be executed during each
3218 /* This code appears in three places, once in scan_loop, and twice
3219 in strength_reduce. */
3220 if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
3221 /* If we enter the loop in the middle, and scan around to the
3222 beginning, don't set not_every_iteration for that.
3223 This can be any kind of jump, since we want to know if insns
3224 will be executed if the loop is executed. */
3225 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
3226 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3227 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3228 not_every_iteration = 1;
3230 /* At the virtual top of a converted loop, insns are again known to
3231 be executed each iteration: logically, the loop begins here
3232 even though the exit code has been duplicated. */
3234 else if (GET_CODE (p) == NOTE
3235 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP)
3236 not_every_iteration = 0;
3238 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3239 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3240 or not an insn is known to be executed each iteration of the
3241 loop, whether or not any iterations are known to occur.
3243 Therefore, if we have just passed a label and have no more labels
3244 between here and the test insn of the loop, we know these insns
3245 will be executed each iteration. This can also happen if we
3246 have just passed a jump, for example, when there are nested loops. */
3248 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3249 && no_labels_between_p (p, loop_end))
3250 not_every_iteration = 0;
3253 /* Scan loop_iv_list to remove all regs that proved not to be bivs.
3254 Make a sanity check against n_times_set. */
3255 for (backbl = &loop_iv_list, bl = *backbl; bl; bl = bl->next)
3257 if (reg_iv_type[bl->regno] != BASIC_INDUCT
3258 /* Above happens if register modified by subreg, etc. */
3259 /* Make sure it is not recognized as a basic induction var: */
3260 || n_times_set[bl->regno] != bl->biv_count
3261 /* If never incremented, it is invariant that we decided not to
3262 move. So leave it alone. */
3263 || ! bl->incremented)
3265 if (loop_dump_stream)
3266 fprintf (loop_dump_stream, "Reg %d: biv discarded, %s\n",
3268 (reg_iv_type[bl->regno] != BASIC_INDUCT
3269 ? "not induction variable"
3270 : (! bl->incremented ? "never incremented"
3273 reg_iv_type[bl->regno] = NOT_BASIC_INDUCT;
3280 if (loop_dump_stream)
3281 fprintf (loop_dump_stream, "Reg %d: biv verified\n", bl->regno);
3285 /* Exit if there are no bivs. */
3288 /* Can still unroll the loop anyways, but indicate that there is no
3289 strength reduction info available. */
3290 if (flag_unroll_loops)
3291 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 0);
3296 /* Find initial value for each biv by searching backwards from loop_start,
3297 halting at first label. Also record any test condition. */
3300 for (p = loop_start; p && GET_CODE (p) != CODE_LABEL; p = PREV_INSN (p))
3304 if (GET_CODE (p) == CALL_INSN)
3307 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3308 || GET_CODE (p) == CALL_INSN)
3309 note_stores (PATTERN (p), record_initial);
3311 /* Record any test of a biv that branches around the loop if no store
3312 between it and the start of loop. We only care about tests with
3313 constants and registers and only certain of those. */
3314 if (GET_CODE (p) == JUMP_INSN
3315 && JUMP_LABEL (p) != 0
3316 && next_real_insn (JUMP_LABEL (p)) == next_real_insn (loop_end)
3317 && (test = get_condition_for_loop (p)) != 0
3318 && GET_CODE (XEXP (test, 0)) == REG
3319 && REGNO (XEXP (test, 0)) < max_reg_before_loop
3320 && (bl = reg_biv_class[REGNO (XEXP (test, 0))]) != 0
3321 && valid_initial_value_p (XEXP (test, 1), p, call_seen, loop_start)
3322 && bl->init_insn == 0)
3324 /* If an NE test, we have an initial value! */
3325 if (GET_CODE (test) == NE)
3328 bl->init_set = gen_rtx (SET, VOIDmode,
3329 XEXP (test, 0), XEXP (test, 1));
3332 bl->initial_test = test;
3336 /* Look at the each biv and see if we can say anything better about its
3337 initial value from any initializing insns set up above. (This is done
3338 in two passes to avoid missing SETs in a PARALLEL.) */
3339 for (bl = loop_iv_list; bl; bl = bl->next)
3343 if (! bl->init_insn)
3346 src = SET_SRC (bl->init_set);
3348 if (loop_dump_stream)
3349 fprintf (loop_dump_stream,
3350 "Biv %d initialized at insn %d: initial value ",
3351 bl->regno, INSN_UID (bl->init_insn));
3353 if (valid_initial_value_p (src, bl->init_insn, call_seen, loop_start))
3355 bl->initial_value = src;
3357 if (loop_dump_stream)
3359 if (GET_CODE (src) == CONST_INT)
3360 fprintf (loop_dump_stream, "%d\n", INTVAL (src));
3363 print_rtl (loop_dump_stream, src);
3364 fprintf (loop_dump_stream, "\n");
3370 /* Biv initial value is not simple move,
3371 so let it keep initial value of "itself". */
3373 if (loop_dump_stream)
3374 fprintf (loop_dump_stream, "is complex\n");
3378 /* Search the loop for general induction variables. */
3380 /* A register is a giv if: it is only set once, it is a function of a
3381 biv and a constant (or invariant), and it is not a biv. */
3383 not_every_iteration = 0;
3388 /* At end of a straight-in loop, we are done.
3389 At end of a loop entered at the bottom, scan the top. */
3390 if (p == scan_start)
3395 p = NEXT_INSN (loop_top);
3398 if (p == scan_start)
3402 /* Look for a general induction variable in a register. */
3403 if (GET_CODE (p) == INSN
3404 && (set = single_set (p))
3405 && GET_CODE (SET_DEST (set)) == REG
3406 && ! may_not_optimize[REGNO (SET_DEST (set))])
3414 dest_reg = SET_DEST (set);
3415 if (REGNO (dest_reg) < FIRST_PSEUDO_REGISTER)
3418 if (/* SET_SRC is a giv. */
3419 ((benefit = general_induction_var (SET_SRC (set),
3422 /* Equivalent expression is a giv. */
3423 || ((regnote = find_reg_note (p, REG_EQUAL, NULL_RTX))
3424 && (benefit = general_induction_var (XEXP (regnote, 0),
3426 &add_val, &mult_val))))
3427 /* Don't try to handle any regs made by loop optimization.
3428 We have nothing on them in regno_first_uid, etc. */
3429 && REGNO (dest_reg) < max_reg_before_loop
3430 /* Don't recognize a BASIC_INDUCT_VAR here. */
3431 && dest_reg != src_reg
3432 /* This must be the only place where the register is set. */
3433 && (n_times_set[REGNO (dest_reg)] == 1
3434 /* or all sets must be consecutive and make a giv. */
3435 || (benefit = consec_sets_giv (benefit, p,
3437 &add_val, &mult_val))))
3441 = (struct induction *) alloca (sizeof (struct induction));
3444 /* If this is a library call, increase benefit. */
3445 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
3446 benefit += libcall_benefit (p);
3448 /* Skip the consecutive insns, if there are any. */
3449 for (count = n_times_set[REGNO (dest_reg)] - 1;
3452 /* If first insn of libcall sequence, skip to end.
3453 Do this at start of loop, since INSN is guaranteed to
3455 if (GET_CODE (p) != NOTE
3456 && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
3459 do p = NEXT_INSN (p);
3460 while (GET_CODE (p) == NOTE);
3463 record_giv (v, p, src_reg, dest_reg, mult_val, add_val, benefit,
3464 DEST_REG, not_every_iteration, NULL_PTR, loop_start,
3470 #ifndef DONT_REDUCE_ADDR
3471 /* Look for givs which are memory addresses. */
3472 /* This resulted in worse code on a VAX 8600. I wonder if it
3474 if (GET_CODE (p) == INSN)
3475 find_mem_givs (PATTERN (p), p, not_every_iteration, loop_start,
3479 /* Update the status of whether giv can derive other givs. This can
3480 change when we pass a label or an insn that updates a biv. */
3481 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3482 || GET_CODE (p) == CODE_LABEL)
3483 update_giv_derive (p);
3485 /* Past a label or a jump, we get to insns for which we can't count
3486 on whether or how many times they will be executed during each
3488 /* This code appears in three places, once in scan_loop, and twice
3489 in strength_reduce. */
3490 if ((GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN)
3491 /* If we enter the loop in the middle, and scan around
3492 to the beginning, don't set not_every_iteration for that.
3493 This can be any kind of jump, since we want to know if insns
3494 will be executed if the loop is executed. */
3495 && ! (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p) == loop_top
3496 && ((NEXT_INSN (NEXT_INSN (p)) == loop_end && simplejump_p (p))
3497 || (NEXT_INSN (p) == loop_end && condjump_p (p)))))
3498 not_every_iteration = 1;
3500 /* At the virtual top of a converted loop, insns are again known to
3501 be executed each iteration: logically, the loop begins here
3502 even though the exit code has been duplicated. */
3504 else if (GET_CODE (p) == NOTE
3505 && NOTE_LINE_NUMBER (p) == NOTE_INSN_LOOP_VTOP)
3506 not_every_iteration = 0;
3508 /* Unlike in the code motion pass where MAYBE_NEVER indicates that
3509 an insn may never be executed, NOT_EVERY_ITERATION indicates whether
3510 or not an insn is known to be executed each iteration of the
3511 loop, whether or not any iterations are known to occur.
3513 Therefore, if we have just passed a label and have no more labels
3514 between here and the test insn of the loop, we know these insns
3515 will be executed each iteration. */
3517 if (not_every_iteration && GET_CODE (p) == CODE_LABEL
3518 && no_labels_between_p (p, loop_end))
3519 not_every_iteration = 0;
3522 /* Try to calculate and save the number of loop iterations. This is
3523 set to zero if the actual number can not be calculated. This must
3524 be called after all giv's have been identified, since otherwise it may
3525 fail if the iteration variable is a giv. */
3527 loop_n_iterations = loop_iterations (loop_start, loop_end);
3529 /* Now for each giv for which we still don't know whether or not it is
3530 replaceable, check to see if it is replaceable because its final value
3531 can be calculated. This must be done after loop_iterations is called,
3532 so that final_giv_value will work correctly. */
3534 for (bl = loop_iv_list; bl; bl = bl->next)
3536 struct induction *v;
3538 for (v = bl->giv; v; v = v->next_iv)
3539 if (! v->replaceable && ! v->not_replaceable)
3540 check_final_value (v, loop_start, loop_end);
3543 /* Try to prove that the loop counter variable (if any) is always
3544 nonnegative; if so, record that fact with a REG_NONNEG note
3545 so that "decrement and branch until zero" insn can be used. */
3546 check_dbra_loop (loop_end, insn_count, loop_start);
3548 /* Create reg_map to hold substitutions for replaceable giv regs. */
3549 reg_map = (rtx *) alloca (max_reg_before_loop * sizeof (rtx));
3550 bzero ((char *) reg_map, max_reg_before_loop * sizeof (rtx));
3552 /* Examine each iv class for feasibility of strength reduction/induction
3553 variable elimination. */
3555 for (bl = loop_iv_list; bl; bl = bl->next)
3557 struct induction *v;
3560 rtx final_value = 0;
3562 /* Test whether it will be possible to eliminate this biv
3563 provided all givs are reduced. This is possible if either
3564 the reg is not used outside the loop, or we can compute
3565 what its final value will be.
3567 For architectures with a decrement_and_branch_until_zero insn,
3568 don't do this if we put a REG_NONNEG note on the endtest for
3571 /* Compare against bl->init_insn rather than loop_start.
3572 We aren't concerned with any uses of the biv between
3573 init_insn and loop_start since these won't be affected
3574 by the value of the biv elsewhere in the function, so
3575 long as init_insn doesn't use the biv itself.
3576 March 14, 1989 -- self@bayes.arc.nasa.gov */
3578 if ((uid_luid[regno_last_uid[bl->regno]] < INSN_LUID (loop_end)
3580 && INSN_UID (bl->init_insn) < max_uid_for_loop
3581 && uid_luid[regno_first_uid[bl->regno]] >= INSN_LUID (bl->init_insn)
3582 #ifdef HAVE_decrement_and_branch_until_zero
3585 && ! reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
3586 || ((final_value = final_biv_value (bl, loop_start, loop_end))
3587 #ifdef HAVE_decrement_and_branch_until_zero
3591 bl->eliminable = maybe_eliminate_biv (bl, loop_start, end, 0,
3592 threshold, insn_count);
3595 if (loop_dump_stream)
3597 fprintf (loop_dump_stream,
3598 "Cannot eliminate biv %d.\n",
3600 fprintf (loop_dump_stream,
3601 "First use: insn %d, last use: insn %d.\n",
3602 regno_first_uid[bl->regno],
3603 regno_last_uid[bl->regno]);
3607 /* Combine all giv's for this iv_class. */
3610 /* This will be true at the end, if all givs which depend on this
3611 biv have been strength reduced.
3612 We can't (currently) eliminate the biv unless this is so. */
3615 /* Check each giv in this class to see if we will benefit by reducing
3616 it. Skip giv's combined with others. */
3617 for (v = bl->giv; v; v = v->next_iv)
3619 struct induction *tv;
3621 if (v->ignore || v->same)
3624 benefit = v->benefit;
3626 /* Reduce benefit if not replaceable, since we will insert
3627 a move-insn to replace the insn that calculates this giv.
3628 Don't do this unless the giv is a user variable, since it
3629 will often be marked non-replaceable because of the duplication
3630 of the exit code outside the loop. In such a case, the copies
3631 we insert are dead and will be deleted. So they don't have
3632 a cost. Similar situations exist. */
3633 /* ??? The new final_[bg]iv_value code does a much better job
3634 of finding replaceable giv's, and hence this code may no longer
3636 if (! v->replaceable && ! bl->eliminable
3637 && REG_USERVAR_P (v->dest_reg))
3638 benefit -= copy_cost;
3640 /* Decrease the benefit to count the add-insns that we will
3641 insert to increment the reduced reg for the giv. */
3642 benefit -= add_cost * bl->biv_count;
3644 /* Decide whether to strength-reduce this giv or to leave the code
3645 unchanged (recompute it from the biv each time it is used).
3646 This decision can be made independently for each giv. */
3648 /* ??? Perhaps attempt to guess whether autoincrement will handle
3649 some of the new add insns; if so, can increase BENEFIT
3650 (undo the subtraction of add_cost that was done above). */
3652 /* If an insn is not to be strength reduced, then set its ignore
3653 flag, and clear all_reduced. */
3655 if (v->lifetime * threshold * benefit < insn_count)
3657 if (loop_dump_stream)
3658 fprintf (loop_dump_stream,
3659 "giv of insn %d not worth while, %d vs %d.\n",
3661 v->lifetime * threshold * benefit, insn_count);
3667 /* Check that we can increment the reduced giv without a
3668 multiply insn. If not, reject it. */
3670 for (tv = bl->biv; tv; tv = tv->next_iv)
3671 if (tv->mult_val == const1_rtx
3672 && ! product_cheap_p (tv->add_val, v->mult_val))
3674 if (loop_dump_stream)
3675 fprintf (loop_dump_stream,
3676 "giv of insn %d: would need a multiply.\n",
3677 INSN_UID (v->insn));
3685 /* Reduce each giv that we decided to reduce. */
3687 for (v = bl->giv; v; v = v->next_iv)
3689 struct induction *tv;
3690 if (! v->ignore && v->same == 0)
3692 v->new_reg = gen_reg_rtx (v->mode);
3694 /* For each place where the biv is incremented,
3695 add an insn to increment the new, reduced reg for the giv. */
3696 for (tv = bl->biv; tv; tv = tv->next_iv)
3698 if (tv->mult_val == const1_rtx)
3699 emit_iv_add_mult (tv->add_val, v->mult_val,
3700 v->new_reg, v->new_reg, tv->insn);
3701 else /* tv->mult_val == const0_rtx */
3702 /* A multiply is acceptable here
3703 since this is presumed to be seldom executed. */
3704 emit_iv_add_mult (tv->add_val, v->mult_val,
3705 v->add_val, v->new_reg, tv->insn);
3708 /* Add code at loop start to initialize giv's reduced reg. */
3710 emit_iv_add_mult (bl->initial_value, v->mult_val,
3711 v->add_val, v->new_reg, loop_start);
3715 /* Rescan all givs. If a giv is the same as a giv not reduced, mark it
3718 For each giv register that can be reduced now: if replaceable,
3719 substitute reduced reg wherever the old giv occurs;
3720 else add new move insn "giv_reg = reduced_reg".
3722 Also check for givs whose first use is their definition and whose
3723 last use is the definition of another giv. If so, it is likely
3724 dead and should not be used to eliminate a biv. */
3725 for (v = bl->giv; v; v = v->next_iv)
3727 if (v->same && v->same->ignore)
3733 if (v->giv_type == DEST_REG
3734 && regno_first_uid[REGNO (v->dest_reg)] == INSN_UID (v->insn))
3736 struct induction *v1;
3738 for (v1 = bl->giv; v1; v1 = v1->next_iv)
3739 if (regno_last_uid[REGNO (v->dest_reg)] == INSN_UID (v1->insn))
3743 /* Update expression if this was combined, in case other giv was
3746 v->new_reg = replace_rtx (v->new_reg,
3747 v->same->dest_reg, v->same->new_reg);
3749 if (v->giv_type == DEST_ADDR)
3750 /* Store reduced reg as the address in the memref where we found
3752 *v->location = v->new_reg;
3753 else if (v->replaceable)
3755 reg_map[REGNO (v->dest_reg)] = v->new_reg;
3758 /* I can no longer duplicate the original problem. Perhaps
3759 this is unnecessary now? */
3761 /* Replaceable; it isn't strictly necessary to delete the old
3762 insn and emit a new one, because v->dest_reg is now dead.
3764 However, especially when unrolling loops, the special
3765 handling for (set REG0 REG1) in the second cse pass may
3766 make v->dest_reg live again. To avoid this problem, emit
3767 an insn to set the original giv reg from the reduced giv.
3768 We can not delete the original insn, since it may be part
3769 of a LIBCALL, and the code in flow that eliminates dead
3770 libcalls will fail if it is deleted. */
3771 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
3777 /* Not replaceable; emit an insn to set the original giv reg from
3778 the reduced giv, same as above. */
3779 emit_insn_after (gen_move_insn (v->dest_reg, v->new_reg),
3783 /* When a loop is reversed, givs which depend on the reversed
3784 biv, and which are live outside the loop, must be set to their
3785 correct final value. This insn is only needed if the giv is
3786 not replaceable. The correct final value is the same as the
3787 value that the giv starts the reversed loop with. */
3788 if (bl->reversed && ! v->replaceable)
3789 emit_iv_add_mult (bl->initial_value, v->mult_val,
3790 v->add_val, v->dest_reg, end_insert_before);
3791 else if (v->final_value)
3795 /* If the loop has multiple exits, emit the insn before the
3796 loop to ensure that it will always be executed no matter
3797 how the loop exits. Otherwise, emit the insn after the loop,
3798 since this is slightly more efficient. */
3799 if (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
3800 insert_before = loop_start;
3802 insert_before = end_insert_before;
3803 emit_insn_before (gen_move_insn (v->dest_reg, v->final_value),
3807 /* If the insn to set the final value of the giv was emitted
3808 before the loop, then we must delete the insn inside the loop
3809 that sets it. If this is a LIBCALL, then we must delete
3810 every insn in the libcall. Note, however, that
3811 final_giv_value will only succeed when there are multiple
3812 exits if the giv is dead at each exit, hence it does not
3813 matter that the original insn remains because it is dead
3815 /* Delete the insn inside the loop that sets the giv since
3816 the giv is now set before (or after) the loop. */
3817 delete_insn (v->insn);
3821 if (loop_dump_stream)
3823 fprintf (loop_dump_stream, "giv at %d reduced to ",
3824 INSN_UID (v->insn));
3825 print_rtl (loop_dump_stream, v->new_reg);
3826 fprintf (loop_dump_stream, "\n");
3830 /* All the givs based on the biv bl have been reduced if they
3833 /* For each giv not marked as maybe dead that has been combined with a
3834 second giv, clear any "maybe dead" mark on that second giv.
3835 v->new_reg will either be or refer to the register of the giv it
3838 Doing this clearing avoids problems in biv elimination where a
3839 giv's new_reg is a complex value that can't be put in the insn but
3840 the giv combined with (with a reg as new_reg) is marked maybe_dead.
3841 Since the register will be used in either case, we'd prefer it be
3842 used from the simpler giv. */
3844 for (v = bl->giv; v; v = v->next_iv)
3845 if (! v->maybe_dead && v->same)
3846 v->same->maybe_dead = 0;
3848 /* Try to eliminate the biv, if it is a candidate.
3849 This won't work if ! all_reduced,
3850 since the givs we planned to use might not have been reduced.
3852 We have to be careful that we didn't initially think we could eliminate
3853 this biv because of a giv that we now think may be dead and shouldn't
3854 be used as a biv replacement.
3856 Also, there is the possibility that we may have a giv that looks
3857 like it can be used to eliminate a biv, but the resulting insn
3858 isn't valid. This can happen, for example, on the 88k, where a
3859 JUMP_INSN can compare a register only with zero. Attempts to
3860 replace it with a compare with a constant will fail.
3862 Note that in cases where this call fails, we may have replaced some
3863 of the occurrences of the biv with a giv, but no harm was done in
3864 doing so in the rare cases where it can occur. */
3866 if (all_reduced == 1 && bl->eliminable
3867 && maybe_eliminate_biv (bl, loop_start, end, 1,
3868 threshold, insn_count))
3871 /* ?? If we created a new test to bypass the loop entirely,
3872 or otherwise drop straight in, based on this test, then
3873 we might want to rewrite it also. This way some later
3874 pass has more hope of removing the initialization of this
3877 /* If final_value != 0, then the biv may be used after loop end
3878 and we must emit an insn to set it just in case.
3880 Reversed bivs already have an insn after the loop setting their
3881 value, so we don't need another one. We can't calculate the
3882 proper final value for such a biv here anyways. */
3883 if (final_value != 0 && ! bl->reversed)
3887 /* If the loop has multiple exits, emit the insn before the
3888 loop to ensure that it will always be executed no matter
3889 how the loop exits. Otherwise, emit the insn after the
3890 loop, since this is slightly more efficient. */
3891 if (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
3892 insert_before = loop_start;
3894 insert_before = end_insert_before;
3896 emit_insn_before (gen_move_insn (bl->biv->dest_reg, final_value),
3901 /* Delete all of the instructions inside the loop which set
3902 the biv, as they are all dead. If is safe to delete them,
3903 because an insn setting a biv will never be part of a libcall. */
3904 /* However, deleting them will invalidate the regno_last_uid info,
3905 so keeping them around is more convenient. Final_biv_value
3906 will only succeed when there are multiple exits if the biv
3907 is dead at each exit, hence it does not matter that the original
3908 insn remains, because it is dead anyways. */
3909 for (v = bl->biv; v; v = v->next_iv)
3910 delete_insn (v->insn);
3913 if (loop_dump_stream)
3914 fprintf (loop_dump_stream, "Reg %d: biv eliminated\n",
3919 /* Go through all the instructions in the loop, making all the
3920 register substitutions scheduled in REG_MAP. */
3922 for (p = loop_start; p != end; p = NEXT_INSN (p))
3923 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
3924 || GET_CODE (p) == CALL_INSN)
3926 replace_regs (PATTERN (p), reg_map, max_reg_before_loop, 0);
3927 replace_regs (REG_NOTES (p), reg_map, max_reg_before_loop, 0);
3930 /* Unroll loops from within strength reduction so that we can use the
3931 induction variable information that strength_reduce has already
3934 if (flag_unroll_loops)
3935 unroll_loop (loop_end, insn_count, loop_start, end_insert_before, 1);
3937 if (loop_dump_stream)
3938 fprintf (loop_dump_stream, "\n");
3941 /* Return 1 if X is a valid source for an initial value (or as value being
3942 compared against in an initial test).
3944 X must be either a register or constant and must not be clobbered between
3945 the current insn and the start of the loop.
3947 INSN is the insn containing X. */
3950 valid_initial_value_p (x, insn, call_seen, loop_start)
3959 /* Only consider pseudos we know about initialized in insns whose luids
3961 if (GET_CODE (x) != REG
3962 || REGNO (x) >= max_reg_before_loop)
3965 /* Don't use call-clobbered registers across a call which clobbers it. On
3966 some machines, don't use any hard registers at all. */
3967 if (REGNO (x) < FIRST_PSEUDO_REGISTER
3968 #ifndef SMALL_REGISTER_CLASSES
3969 && call_used_regs[REGNO (x)] && call_seen
3974 /* Don't use registers that have been clobbered before the start of the
3976 if (reg_set_between_p (x, insn, loop_start))
3982 /* Scan X for memory refs and check each memory address
3983 as a possible giv. INSN is the insn whose pattern X comes from.
3984 NOT_EVERY_ITERATION is 1 if the insn might not be executed during
3985 every loop iteration. */
3988 find_mem_givs (x, insn, not_every_iteration, loop_start, loop_end)
3991 int not_every_iteration;
3992 rtx loop_start, loop_end;
3995 register enum rtx_code code;
4001 code = GET_CODE (x);
4025 benefit = general_induction_var (XEXP (x, 0),
4026 &src_reg, &add_val, &mult_val);
4028 /* Don't make a DEST_ADDR giv with mult_val == 1 && add_val == 0.
4029 Such a giv isn't useful. */
4030 if (benefit > 0 && (mult_val != const1_rtx || add_val != const0_rtx))
4032 /* Found one; record it. */
4034 = (struct induction *) oballoc (sizeof (struct induction));
4036 record_giv (v, insn, src_reg, addr_placeholder, mult_val,
4037 add_val, benefit, DEST_ADDR, not_every_iteration,
4038 &XEXP (x, 0), loop_start, loop_end);
4040 v->mem_mode = GET_MODE (x);
4046 /* Recursively scan the subexpressions for other mem refs. */
4048 fmt = GET_RTX_FORMAT (code);
4049 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4051 find_mem_givs (XEXP (x, i), insn, not_every_iteration, loop_start,
4053 else if (fmt[i] == 'E')
4054 for (j = 0; j < XVECLEN (x, i); j++)
4055 find_mem_givs (XVECEXP (x, i, j), insn, not_every_iteration,
4056 loop_start, loop_end);
4059 /* Fill in the data about one biv update.
4060 V is the `struct induction' in which we record the biv. (It is
4061 allocated by the caller, with alloca.)
4062 INSN is the insn that sets it.
4063 DEST_REG is the biv's reg.
4065 MULT_VAL is const1_rtx if the biv is being incremented here, in which case
4066 INC_VAL is the increment. Otherwise, MULT_VAL is const0_rtx and the biv is
4067 being set to INC_VAL.
4069 NOT_EVERY_ITERATION is nonzero if this biv update is not know to be
4070 executed every iteration; MAYBE_MULTIPLE is nonzero if this biv update
4071 can be executed more than once per iteration. If MAYBE_MULTIPLE
4072 and NOT_EVERY_ITERATION are both zero, we know that the biv update is
4073 executed exactly once per iteration. */
4076 record_biv (v, insn, dest_reg, inc_val, mult_val,
4077 not_every_iteration, maybe_multiple)
4078 struct induction *v;
4083 int not_every_iteration;
4086 struct iv_class *bl;
4089 v->src_reg = dest_reg;
4090 v->dest_reg = dest_reg;
4091 v->mult_val = mult_val;
4092 v->add_val = inc_val;
4093 v->mode = GET_MODE (dest_reg);
4094 v->always_computable = ! not_every_iteration;
4095 v->maybe_multiple = maybe_multiple;
4097 /* Add this to the reg's iv_class, creating a class
4098 if this is the first incrementation of the reg. */
4100 bl = reg_biv_class[REGNO (dest_reg)];
4103 /* Create and initialize new iv_class. */
4105 bl = (struct iv_class *) oballoc (sizeof (struct iv_class));
4107 bl->regno = REGNO (dest_reg);
4113 /* Set initial value to the reg itself. */
4114 bl->initial_value = dest_reg;
4115 /* We haven't seen the initializing insn yet */
4118 bl->initial_test = 0;
4119 bl->incremented = 0;
4124 /* Add this class to loop_iv_list. */
4125 bl->next = loop_iv_list;
4128 /* Put it in the array of biv register classes. */
4129 reg_biv_class[REGNO (dest_reg)] = bl;
4132 /* Update IV_CLASS entry for this biv. */
4133 v->next_iv = bl->biv;
4136 if (mult_val == const1_rtx)
4137 bl->incremented = 1;
4139 if (loop_dump_stream)
4141 fprintf (loop_dump_stream,
4142 "Insn %d: possible biv, reg %d,",
4143 INSN_UID (insn), REGNO (dest_reg));
4144 if (GET_CODE (inc_val) == CONST_INT)
4145 fprintf (loop_dump_stream, " const = %d\n",
4149 fprintf (loop_dump_stream, " const = ");
4150 print_rtl (loop_dump_stream, inc_val);
4151 fprintf (loop_dump_stream, "\n");
4156 /* Fill in the data about one giv.
4157 V is the `struct induction' in which we record the giv. (It is
4158 allocated by the caller, with alloca.)
4159 INSN is the insn that sets it.
4160 BENEFIT estimates the savings from deleting this insn.
4161 TYPE is DEST_REG or DEST_ADDR; it says whether the giv is computed
4162 into a register or is used as a memory address.
4164 SRC_REG is the biv reg which the giv is computed from.
4165 DEST_REG is the giv's reg (if the giv is stored in a reg).
4166 MULT_VAL and ADD_VAL are the coefficients used to compute the giv.
4167 LOCATION points to the place where this giv's value appears in INSN. */
4170 record_giv (v, insn, src_reg, dest_reg, mult_val, add_val, benefit,
4171 type, not_every_iteration, location, loop_start, loop_end)
4172 struct induction *v;
4176 rtx mult_val, add_val;
4179 int not_every_iteration;
4181 rtx loop_start, loop_end;
4183 struct induction *b;
4184 struct iv_class *bl;
4185 rtx set = single_set (insn);
4189 v->src_reg = src_reg;
4191 v->dest_reg = dest_reg;
4192 v->mult_val = mult_val;
4193 v->add_val = add_val;
4194 v->benefit = benefit;
4195 v->location = location;
4197 v->combined_with = 0;
4198 v->maybe_multiple = 0;
4200 v->derive_adjustment = 0;
4206 /* The v->always_computable field is used in update_giv_derive, to
4207 determine whether a giv can be used to derive another giv. For a
4208 DEST_REG giv, INSN computes a new value for the giv, so its value
4209 isn't computable if INSN insn't executed every iteration.
4210 However, for a DEST_ADDR giv, INSN merely uses the value of the giv;
4211 it does not compute a new value. Hence the value is always computable
4212 regardless of whether INSN is executed each iteration. */
4214 if (type == DEST_ADDR)
4215 v->always_computable = 1;
4217 v->always_computable = ! not_every_iteration;
4219 if (type == DEST_ADDR)
4221 v->mode = GET_MODE (*location);
4225 else /* type == DEST_REG */
4227 v->mode = GET_MODE (SET_DEST (set));
4229 v->lifetime = (uid_luid[regno_last_uid[REGNO (dest_reg)]]
4230 - uid_luid[regno_first_uid[REGNO (dest_reg)]]);
4232 v->times_used = n_times_used[REGNO (dest_reg)];
4234 /* If the lifetime is zero, it means that this register is
4235 really a dead store. So mark this as a giv that can be
4236 ignored. This will not prevent the biv from being eliminated. */
4237 if (v->lifetime == 0)
4240 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
4241 reg_iv_info[REGNO (dest_reg)] = v;
4244 /* Add the giv to the class of givs computed from one biv. */
4246 bl = reg_biv_class[REGNO (src_reg)];
4249 v->next_iv = bl->giv;
4251 /* Don't count DEST_ADDR. This is supposed to count the number of
4252 insns that calculate givs. */
4253 if (type == DEST_REG)
4255 bl->total_benefit += benefit;
4258 /* Fatal error, biv missing for this giv? */
4261 if (type == DEST_ADDR)
4265 /* The giv can be replaced outright by the reduced register only if all
4266 of the following conditions are true:
4267 - the insn that sets the giv is always executed on any iteration
4268 on which the giv is used at all
4269 (there are two ways to deduce this:
4270 either the insn is executed on every iteration,
4271 or all uses follow that insn in the same basic block),
4272 - the giv is not used outside the loop
4273 - no assignments to the biv occur during the giv's lifetime. */
4275 if (regno_first_uid[REGNO (dest_reg)] == INSN_UID (insn)
4276 /* Previous line always fails if INSN was moved by loop opt. */
4277 && uid_luid[regno_last_uid[REGNO (dest_reg)]] < INSN_LUID (loop_end)
4278 && (! not_every_iteration
4279 || last_use_this_basic_block (dest_reg, insn)))
4281 /* Now check that there are no assignments to the biv within the
4282 giv's lifetime. This requires two separate checks. */
4284 /* Check each biv update, and fail if any are between the first
4285 and last use of the giv.
4287 If this loop contains an inner loop that was unrolled, then
4288 the insn modifying the biv may have been emitted by the loop
4289 unrolling code, and hence does not have a valid luid. Just
4290 mark the biv as not replaceable in this case. It is not very
4291 useful as a biv, because it is used in two different loops.
4292 It is very unlikely that we would be able to optimize the giv
4293 using this biv anyways. */
4296 for (b = bl->biv; b; b = b->next_iv)
4298 if (INSN_UID (b->insn) >= max_uid_for_loop
4299 || ((uid_luid[INSN_UID (b->insn)]
4300 >= uid_luid[regno_first_uid[REGNO (dest_reg)]])
4301 && (uid_luid[INSN_UID (b->insn)]
4302 <= uid_luid[regno_last_uid[REGNO (dest_reg)]])))
4305 v->not_replaceable = 1;
4310 /* Check each insn between the first and last use of the giv,
4311 and fail if any of them are branches that jump to a named label
4312 outside this range, but still inside the loop. This catches
4313 cases of spaghetti code where the execution order of insns
4314 is not linear, and hence the above test fails. For example,
4315 in the following code, j is not replaceable:
4316 for (i = 0; i < 100; ) {
4317 L0: j = 4*i; goto L1;
4321 printf ("k = %d\n", k); }
4322 This test is conservative, but this test succeeds rarely enough
4323 that it isn't a problem. See also check_final_value below. */
4327 INSN_UID (p) >= max_uid_for_loop
4328 || INSN_LUID (p) < uid_luid[regno_last_uid[REGNO (dest_reg)]];
4331 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
4332 && LABEL_NAME (JUMP_LABEL (p))
4333 && ((INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start)
4334 && (INSN_LUID (JUMP_LABEL (p))
4335 < uid_luid[regno_first_uid[REGNO (dest_reg)]]))
4336 || (INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end)
4337 && (INSN_LUID (JUMP_LABEL (p))
4338 > uid_luid[regno_last_uid[REGNO (dest_reg)]]))))
4341 v->not_replaceable = 1;
4343 if (loop_dump_stream)
4344 fprintf (loop_dump_stream,
4345 "Found branch outside giv lifetime.\n");
4353 /* May still be replaceable, we don't have enough info here to
4356 v->not_replaceable = 0;
4360 if (loop_dump_stream)
4362 if (type == DEST_REG)
4363 fprintf (loop_dump_stream, "Insn %d: giv reg %d",
4364 INSN_UID (insn), REGNO (dest_reg));
4366 fprintf (loop_dump_stream, "Insn %d: dest address",
4369 fprintf (loop_dump_stream, " src reg %d benefit %d",
4370 REGNO (src_reg), v->benefit);
4371 fprintf (loop_dump_stream, " used %d lifetime %d",
4372 v->times_used, v->lifetime);
4375 fprintf (loop_dump_stream, " replaceable");
4377 if (GET_CODE (mult_val) == CONST_INT)
4378 fprintf (loop_dump_stream, " mult %d",
4382 fprintf (loop_dump_stream, " mult ");
4383 print_rtl (loop_dump_stream, mult_val);
4386 if (GET_CODE (add_val) == CONST_INT)
4387 fprintf (loop_dump_stream, " add %d",
4391 fprintf (loop_dump_stream, " add ");
4392 print_rtl (loop_dump_stream, add_val);
4396 if (loop_dump_stream)
4397 fprintf (loop_dump_stream, "\n");
4402 /* All this does is determine whether a giv can be made replaceable because
4403 its final value can be calculated. This code can not be part of record_giv
4404 above, because final_giv_value requires that the number of loop iterations
4405 be known, and that can not be accurately calculated until after all givs
4406 have been identified. */
4409 check_final_value (v, loop_start, loop_end)
4410 struct induction *v;
4411 rtx loop_start, loop_end;
4413 struct iv_class *bl;
4414 rtx final_value = 0;
4417 bl = reg_biv_class[REGNO (v->src_reg)];
4419 /* DEST_ADDR givs will never reach here, because they are always marked
4420 replaceable above in record_giv. */
4422 /* The giv can be replaced outright by the reduced register only if all
4423 of the following conditions are true:
4424 - the insn that sets the giv is always executed on any iteration
4425 on which the giv is used at all
4426 (there are two ways to deduce this:
4427 either the insn is executed on every iteration,
4428 or all uses follow that insn in the same basic block),
4429 - its final value can be calculated (this condition is different
4430 than the one above in record_giv)
4431 - no assignments to the biv occur during the giv's lifetime. */
4434 /* This is only called now when replaceable is known to be false. */
4435 /* Clear replaceable, so that it won't confuse final_giv_value. */
4439 if ((final_value = final_giv_value (v, loop_start, loop_end))
4440 && (v->always_computable || last_use_this_basic_block (v->dest_reg, v->insn)))
4442 int biv_increment_seen = 0;
4448 /* When trying to determine whether or not a biv increment occurs
4449 during the lifetime of the giv, we can ignore uses of the variable
4450 outside the loop because final_value is true. Hence we can not
4451 use regno_last_uid and regno_first_uid as above in record_giv. */
4453 /* Search the loop to determine whether any assignments to the
4454 biv occur during the giv's lifetime. Start with the insn
4455 that sets the giv, and search around the loop until we come
4456 back to that insn again.
4458 Also fail if there is a jump within the giv's lifetime that jumps
4459 to somewhere outside the lifetime but still within the loop. This
4460 catches spaghetti code where the execution order is not linear, and
4461 hence the above test fails. Here we assume that the giv lifetime
4462 does not extend from one iteration of the loop to the next, so as
4463 to make the test easier. Since the lifetime isn't known yet,
4464 this requires two loops. See also record_giv above. */
4466 last_giv_use = v->insn;
4472 p = NEXT_INSN (loop_start);
4476 if (GET_CODE (p) == INSN || GET_CODE (p) == JUMP_INSN
4477 || GET_CODE (p) == CALL_INSN)
4479 if (biv_increment_seen)
4481 if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4484 v->not_replaceable = 1;
4488 else if (GET_CODE (PATTERN (p)) == SET
4489 && SET_DEST (PATTERN (p)) == v->src_reg)
4490 biv_increment_seen = 1;
4491 else if (reg_mentioned_p (v->dest_reg, PATTERN (p)))
4496 /* Now that the lifetime of the giv is known, check for branches
4497 from within the lifetime to outside the lifetime if it is still
4507 p = NEXT_INSN (loop_start);
4508 if (p == last_giv_use)
4511 if (GET_CODE (p) == JUMP_INSN && JUMP_LABEL (p)
4512 && LABEL_NAME (JUMP_LABEL (p))
4513 && ((INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (v->insn)
4514 && INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (loop_start))
4515 || (INSN_LUID (JUMP_LABEL (p)) > INSN_LUID (last_giv_use)
4516 && INSN_LUID (JUMP_LABEL (p)) < INSN_LUID (loop_end))))
4519 v->not_replaceable = 1;
4521 if (loop_dump_stream)
4522 fprintf (loop_dump_stream,
4523 "Found branch outside giv lifetime.\n");
4530 /* If it is replaceable, then save the final value. */
4532 v->final_value = final_value;
4535 if (loop_dump_stream && v->replaceable)
4536 fprintf (loop_dump_stream, "Insn %d: giv reg %d final_value replaceable\n",
4537 INSN_UID (v->insn), REGNO (v->dest_reg));
4540 /* Update the status of whether a giv can derive other givs.
4542 We need to do something special if there is or may be an update to the biv
4543 between the time the giv is defined and the time it is used to derive
4546 In addition, a giv that is only conditionally set is not allowed to
4547 derive another giv once a label has been passed.
4549 The cases we look at are when a label or an update to a biv is passed. */
4552 update_giv_derive (p)
4555 struct iv_class *bl;
4556 struct induction *biv, *giv;
4560 /* Search all IV classes, then all bivs, and finally all givs.
4562 There are three cases we are concerned with. First we have the situation
4563 of a giv that is only updated conditionally. In that case, it may not
4564 derive any givs after a label is passed.
4566 The second case is when a biv update occurs, or may occur, after the
4567 definition of a giv. For certain biv updates (see below) that are
4568 known to occur between the giv definition and use, we can adjust the
4569 giv definition. For others, or when the biv update is conditional,
4570 we must prevent the giv from deriving any other givs. There are two
4571 sub-cases within this case.
4573 If this is a label, we are concerned with any biv update that is done
4574 conditionally, since it may be done after the giv is defined followed by
4575 a branch here (actually, we need to pass both a jump and a label, but
4576 this extra tracking doesn't seem worth it).
4578 If this is a jump, we are concerned about any biv update that may be
4579 executed multiple times. We are actually only concerned about
4580 backward jumps, but it is probably not worth performing the test
4581 on the jump again here.
4583 If this is a biv update, we must adjust the giv status to show that a
4584 subsequent biv update was performed. If this adjustment cannot be done,
4585 the giv cannot derive further givs. */
4587 for (bl = loop_iv_list; bl; bl = bl->next)
4588 for (biv = bl->biv; biv; biv = biv->next_iv)
4589 if (GET_CODE (p) == CODE_LABEL || GET_CODE (p) == JUMP_INSN
4592 for (giv = bl->giv; giv; giv = giv->next_iv)
4594 /* If cant_derive is already true, there is no point in
4595 checking all of these conditions again. */
4596 if (giv->cant_derive)
4599 /* If this giv is conditionally set and we have passed a label,
4600 it cannot derive anything. */
4601 if (GET_CODE (p) == CODE_LABEL && ! giv->always_computable)
4602 giv->cant_derive = 1;
4604 /* Skip givs that have mult_val == 0, since
4605 they are really invariants. Also skip those that are
4606 replaceable, since we know their lifetime doesn't contain
4608 else if (giv->mult_val == const0_rtx || giv->replaceable)
4611 /* The only way we can allow this giv to derive another
4612 is if this is a biv increment and we can form the product
4613 of biv->add_val and giv->mult_val. In this case, we will
4614 be able to compute a compensation. */
4615 else if (biv->insn == p)
4619 if (biv->mult_val == const1_rtx)
4620 tem = simplify_giv_expr (gen_rtx (MULT, giv->mode,
4625 if (tem && giv->derive_adjustment)
4626 tem = simplify_giv_expr (gen_rtx (PLUS, giv->mode, tem,
4627 giv->derive_adjustment),
4630 giv->derive_adjustment = tem;
4632 giv->cant_derive = 1;
4634 else if ((GET_CODE (p) == CODE_LABEL && ! biv->always_computable)
4635 || (GET_CODE (p) == JUMP_INSN && biv->maybe_multiple))
4636 giv->cant_derive = 1;
4641 /* Check whether an insn is an increment legitimate for a basic induction var.
4642 X is the source of the insn.
4643 DEST_REG is the putative biv, also the destination of the insn.
4644 We accept patterns of these forms:
4645 REG = REG + INVARIANT
4646 REG = INVARIANT + REG
4647 REG = REG - CONSTANT
4649 If X is suitable, we return 1, set *MULT_VAL to CONST1_RTX,
4650 and store the additive term into *INC_VAL.
4652 If X is an assignment of an invariant into DEST_REG, we set
4653 *MULT_VAL to CONST0_RTX, and store the invariant into *INC_VAL.
4655 Otherwise we return 0. */
4658 basic_induction_var (x, dest_reg, inc_val, mult_val)
4664 register enum rtx_code code;
4667 code = GET_CODE (x);
4671 if (XEXP (x, 0) == dest_reg)
4673 else if (XEXP (x, 1) == dest_reg)
4678 if (invariant_p (arg) != 1)
4682 *mult_val = const1_rtx;
4686 if (XEXP (x, 0) == dest_reg
4687 && GET_CODE (XEXP (x, 1)) == CONST_INT)
4688 *inc_val = GEN_INT (- INTVAL (XEXP (x, 1)));
4692 *mult_val = const1_rtx;
4695 /* Can accept constant setting of biv only when inside inner most loop.
4696 Otherwise, a biv of an inner loop may be incorrectly recognized
4697 as a biv of the outer loop,
4698 causing code to be moved INTO the inner loop. */
4701 if (invariant_p (x) != 1)
4706 if (loops_enclosed == 1)
4709 *mult_val = const0_rtx;
4720 /* A general induction variable (giv) is any quantity that is a linear
4721 function of a basic induction variable,
4722 i.e. giv = biv * mult_val + add_val.
4723 The coefficients can be any loop invariant quantity.
4724 A giv need not be computed directly from the biv;
4725 it can be computed by way of other givs. */
4727 /* Determine whether X computes a giv.
4728 If it does, return a nonzero value
4729 which is the benefit from eliminating the computation of X;
4730 set *SRC_REG to the register of the biv that it is computed from;
4731 set *ADD_VAL and *MULT_VAL to the coefficients,
4732 such that the value of X is biv * mult + add; */
4735 general_induction_var (x, src_reg, add_val, mult_val)
4745 /* If this is an invariant, forget it, it isn't a giv. */
4746 if (invariant_p (x) == 1)
4749 /* See if the expression could be a giv and get its form.
4750 Mark our place on the obstack in case we don't find a giv. */
4751 storage = (char *) oballoc (0);
4752 x = simplify_giv_expr (x, &benefit);
4759 switch (GET_CODE (x))
4763 /* Since this is now an invariant and wasn't before, it must be a giv
4764 with MULT_VAL == 0. It doesn't matter which BIV we associate this
4766 *src_reg = loop_iv_list->biv->dest_reg;
4767 *mult_val = const0_rtx;
4772 /* This is equivalent to a BIV. */
4774 *mult_val = const1_rtx;
4775 *add_val = const0_rtx;
4779 /* Either (plus (biv) (invar)) or
4780 (plus (mult (biv) (invar_1)) (invar_2)). */
4781 if (GET_CODE (XEXP (x, 0)) == MULT)
4783 *src_reg = XEXP (XEXP (x, 0), 0);
4784 *mult_val = XEXP (XEXP (x, 0), 1);
4788 *src_reg = XEXP (x, 0);
4789 *mult_val = const1_rtx;
4791 *add_val = XEXP (x, 1);
4795 /* ADD_VAL is zero. */
4796 *src_reg = XEXP (x, 0);
4797 *mult_val = XEXP (x, 1);
4798 *add_val = const0_rtx;
4805 /* Remove any enclosing USE from ADD_VAL and MULT_VAL (there will be
4806 unless they are CONST_INT). */
4807 if (GET_CODE (*add_val) == USE)
4808 *add_val = XEXP (*add_val, 0);
4809 if (GET_CODE (*mult_val) == USE)
4810 *mult_val = XEXP (*mult_val, 0);
4812 benefit += rtx_cost (orig_x, SET);
4814 /* Always return some benefit if this is a giv so it will be detected
4815 as such. This allows elimination of bivs that might otherwise
4816 not be eliminated. */
4817 return benefit == 0 ? 1 : benefit;
4820 /* Given an expression, X, try to form it as a linear function of a biv.
4821 We will canonicalize it to be of the form
4822 (plus (mult (BIV) (invar_1))
4824 with possible degeneracies.
4826 The invariant expressions must each be of a form that can be used as a
4827 machine operand. We surround then with a USE rtx (a hack, but localized
4828 and certainly unambiguous!) if not a CONST_INT for simplicity in this
4829 routine; it is the caller's responsibility to strip them.
4831 If no such canonicalization is possible (i.e., two biv's are used or an
4832 expression that is neither invariant nor a biv or giv), this routine
4835 For a non-zero return, the result will have a code of CONST_INT, USE,
4836 REG (for a BIV), PLUS, or MULT. No other codes will occur.
4838 *BENEFIT will be incremented by the benefit of any sub-giv encountered. */
4841 simplify_giv_expr (x, benefit)
4845 enum machine_mode mode = GET_MODE (x);
4849 /* If this is not an integer mode, or if we cannot do arithmetic in this
4850 mode, this can't be a giv. */
4851 if (mode != VOIDmode
4852 && (GET_MODE_CLASS (mode) != MODE_INT
4853 || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT))
4856 switch (GET_CODE (x))
4859 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
4860 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
4861 if (arg0 == 0 || arg1 == 0)
4864 /* Put constant last, CONST_INT last if both constant. */
4865 if ((GET_CODE (arg0) == USE
4866 || GET_CODE (arg0) == CONST_INT)
4867 && GET_CODE (arg1) != CONST_INT)
4868 tem = arg0, arg0 = arg1, arg1 = tem;
4870 /* Handle addition of zero, then addition of an invariant. */
4871 if (arg1 == const0_rtx)
4873 else if (GET_CODE (arg1) == CONST_INT || GET_CODE (arg1) == USE)
4874 switch (GET_CODE (arg0))
4878 /* Both invariant. Only valid if sum is machine operand.
4879 First strip off possible USE on first operand. */
4880 if (GET_CODE (arg0) == USE)
4881 arg0 = XEXP (arg0, 0);
4884 if (CONSTANT_P (arg0) && GET_CODE (arg1) == CONST_INT)
4886 tem = plus_constant (arg0, INTVAL (arg1));
4887 if (GET_CODE (tem) != CONST_INT)
4888 tem = gen_rtx (USE, mode, tem);
4895 /* biv + invar or mult + invar. Return sum. */
4896 return gen_rtx (PLUS, mode, arg0, arg1);
4899 /* (a + invar_1) + invar_2. Associate. */
4900 return simplify_giv_expr (gen_rtx (PLUS, mode,
4902 gen_rtx (PLUS, mode,
4903 XEXP (arg0, 1), arg1)),
4910 /* Each argument must be either REG, PLUS, or MULT. Convert REG to
4911 MULT to reduce cases. */
4912 if (GET_CODE (arg0) == REG)
4913 arg0 = gen_rtx (MULT, mode, arg0, const1_rtx);
4914 if (GET_CODE (arg1) == REG)
4915 arg1 = gen_rtx (MULT, mode, arg1, const1_rtx);
4917 /* Now have PLUS + PLUS, PLUS + MULT, MULT + PLUS, or MULT + MULT.
4918 Put a MULT first, leaving PLUS + PLUS, MULT + PLUS, or MULT + MULT.
4919 Recurse to associate the second PLUS. */
4920 if (GET_CODE (arg1) == MULT)
4921 tem = arg0, arg0 = arg1, arg1 = tem;
4923 if (GET_CODE (arg1) == PLUS)
4924 return simplify_giv_expr (gen_rtx (PLUS, mode,
4925 gen_rtx (PLUS, mode,
4926 arg0, XEXP (arg1, 0)),
4930 /* Now must have MULT + MULT. Distribute if same biv, else not giv. */
4931 if (GET_CODE (arg0) != MULT || GET_CODE (arg1) != MULT)
4934 if (XEXP (arg0, 0) != XEXP (arg1, 0))
4937 return simplify_giv_expr (gen_rtx (MULT, mode,
4939 gen_rtx (PLUS, mode,
4945 /* Handle "a - b" as "a + b * (-1)". */
4946 return simplify_giv_expr (gen_rtx (PLUS, mode,
4948 gen_rtx (MULT, mode,
4949 XEXP (x, 1), constm1_rtx)),
4953 arg0 = simplify_giv_expr (XEXP (x, 0), benefit);
4954 arg1 = simplify_giv_expr (XEXP (x, 1), benefit);
4955 if (arg0 == 0 || arg1 == 0)
4958 /* Put constant last, CONST_INT last if both constant. */
4959 if ((GET_CODE (arg0) == USE || GET_CODE (arg0) == CONST_INT)
4960 && GET_CODE (arg1) != CONST_INT)
4961 tem = arg0, arg0 = arg1, arg1 = tem;
4963 /* If second argument is not now constant, not giv. */
4964 if (GET_CODE (arg1) != USE && GET_CODE (arg1) != CONST_INT)
4967 /* Handle multiply by 0 or 1. */
4968 if (arg1 == const0_rtx)
4971 else if (arg1 == const1_rtx)
4974 switch (GET_CODE (arg0))
4977 /* biv * invar. Done. */
4978 return gen_rtx (MULT, mode, arg0, arg1);
4981 /* Product of two constants. */
4982 return GEN_INT (INTVAL (arg0) * INTVAL (arg1));
4985 /* invar * invar. Not giv. */
4989 /* (a * invar_1) * invar_2. Associate. */
4990 return simplify_giv_expr (gen_rtx (MULT, mode,
4992 gen_rtx (MULT, mode,
4993 XEXP (arg0, 1), arg1)),
4997 /* (a + invar_1) * invar_2. Distribute. */
4998 return simplify_giv_expr (gen_rtx (PLUS, mode,
4999 gen_rtx (MULT, mode,
5000 XEXP (arg0, 0), arg1),
5001 gen_rtx (MULT, mode,
5002 XEXP (arg0, 1), arg1)),
5011 /* Shift by constant is multiply by power of two. */
5012 if (GET_CODE (XEXP (x, 1)) != CONST_INT)
5015 return simplify_giv_expr (gen_rtx (MULT, mode,
5017 GEN_INT ((HOST_WIDE_INT) 1
5018 << INTVAL (XEXP (x, 1)))),
5022 /* "-a" is "a * (-1)" */
5023 return simplify_giv_expr (gen_rtx (MULT, mode, XEXP (x, 0), constm1_rtx),
5027 /* "~a" is "-a - 1". Silly, but easy. */
5028 return simplify_giv_expr (gen_rtx (MINUS, mode,
5029 gen_rtx (NEG, mode, XEXP (x, 0)),
5034 /* Already in proper form for invariant. */
5038 /* If this is a new register, we can't deal with it. */
5039 if (REGNO (x) >= max_reg_before_loop)
5042 /* Check for biv or giv. */
5043 switch (reg_iv_type[REGNO (x)])
5047 case GENERAL_INDUCT:
5049 struct induction *v = reg_iv_info[REGNO (x)];
5051 /* Form expression from giv and add benefit. Ensure this giv
5052 can derive another and subtract any needed adjustment if so. */
5053 *benefit += v->benefit;
5057 tem = gen_rtx (PLUS, mode, gen_rtx (MULT, mode,
5058 v->src_reg, v->mult_val),
5060 if (v->derive_adjustment)
5061 tem = gen_rtx (MINUS, mode, tem, v->derive_adjustment);
5062 return simplify_giv_expr (tem, benefit);
5066 /* Fall through to general case. */
5068 /* If invariant, return as USE (unless CONST_INT).
5069 Otherwise, not giv. */
5070 if (GET_CODE (x) == USE)
5073 if (invariant_p (x) == 1)
5075 if (GET_CODE (x) == CONST_INT)
5078 return gen_rtx (USE, mode, x);
5085 /* Help detect a giv that is calculated by several consecutive insns;
5089 The caller has already identified the first insn P as having a giv as dest;
5090 we check that all other insns that set the same register follow
5091 immediately after P, that they alter nothing else,
5092 and that the result of the last is still a giv.
5094 The value is 0 if the reg set in P is not really a giv.
5095 Otherwise, the value is the amount gained by eliminating
5096 all the consecutive insns that compute the value.
5098 FIRST_BENEFIT is the amount gained by eliminating the first insn, P.
5099 SRC_REG is the reg of the biv; DEST_REG is the reg of the giv.
5101 The coefficients of the ultimate giv value are stored in
5102 *MULT_VAL and *ADD_VAL. */
5105 consec_sets_giv (first_benefit, p, src_reg, dest_reg,
5120 /* Indicate that this is a giv so that we can update the value produced in
5121 each insn of the multi-insn sequence.
5123 This induction structure will be used only by the call to
5124 general_induction_var below, so we can allocate it on our stack.
5125 If this is a giv, our caller will replace the induct var entry with
5126 a new induction structure. */
5128 = (struct induction *) alloca (sizeof (struct induction));
5129 v->src_reg = src_reg;
5130 v->mult_val = *mult_val;
5131 v->add_val = *add_val;
5132 v->benefit = first_benefit;
5134 v->derive_adjustment = 0;
5136 reg_iv_type[REGNO (dest_reg)] = GENERAL_INDUCT;
5137 reg_iv_info[REGNO (dest_reg)] = v;
5139 count = n_times_set[REGNO (dest_reg)] - 1;
5144 code = GET_CODE (p);
5146 /* If libcall, skip to end of call sequence. */
5147 if (code == INSN && (temp = find_reg_note (p, REG_LIBCALL, NULL_RTX)))
5151 && (set = single_set (p))
5152 && GET_CODE (SET_DEST (set)) == REG
5153 && SET_DEST (set) == dest_reg
5154 && ((benefit = general_induction_var (SET_SRC (set), &src_reg,
5156 /* Giv created by equivalent expression. */
5157 || ((temp = find_reg_note (p, REG_EQUAL, NULL_RTX))
5158 && (benefit = general_induction_var (XEXP (temp, 0), &src_reg,
5159 add_val, mult_val))))
5160 && src_reg == v->src_reg)
5162 if (find_reg_note (p, REG_RETVAL, NULL_RTX))
5163 benefit += libcall_benefit (p);
5166 v->mult_val = *mult_val;
5167 v->add_val = *add_val;
5168 v->benefit = benefit;
5170 else if (code != NOTE)
5172 /* Allow insns that set something other than this giv to a
5173 constant. Such insns are needed on machines which cannot
5174 include long constants and should not disqualify a giv. */
5176 && (set = single_set (p))
5177 && SET_DEST (set) != dest_reg
5178 && CONSTANT_P (SET_SRC (set)))
5181 reg_iv_type[REGNO (dest_reg)] = UNKNOWN_INDUCT;
5189 /* Return an rtx, if any, that expresses giv G2 as a function of the register
5190 represented by G1. If no such expression can be found, or it is clear that
5191 it cannot possibly be a valid address, 0 is returned.
5193 To perform the computation, we note that
5196 where `v' is the biv.
5198 So G2 = (c/a) * G1 + (d - b*c/a) */
5202 express_from (g1, g2)
5203 struct induction *g1, *g2;
5207 /* The value that G1 will be multiplied by must be a constant integer. Also,
5208 the only chance we have of getting a valid address is if b*c/a (see above
5209 for notation) is also an integer. */
5210 if (GET_CODE (g1->mult_val) != CONST_INT
5211 || GET_CODE (g2->mult_val) != CONST_INT
5212 || GET_CODE (g1->add_val) != CONST_INT
5213 || g1->mult_val == const0_rtx
5214 || INTVAL (g2->mult_val) % INTVAL (g1->mult_val) != 0)
5217 mult = GEN_INT (INTVAL (g2->mult_val) / INTVAL (g1->mult_val));
5218 add = plus_constant (g2->add_val, - INTVAL (g1->add_val) * INTVAL (mult));
5220 /* Form simplified final result. */
5221 if (mult == const0_rtx)
5223 else if (mult == const1_rtx)
5224 mult = g1->dest_reg;
5226 mult = gen_rtx (MULT, g2->mode, g1->dest_reg, mult);
5228 if (add == const0_rtx)
5231 return gen_rtx (PLUS, g2->mode, mult, add);
5235 /* Return 1 if giv G2 can be combined with G1. This means that G2 can use
5236 (either directly or via an address expression) a register used to represent
5237 G1. Set g2->new_reg to a represtation of G1 (normally just
5241 combine_givs_p (g1, g2)
5242 struct induction *g1, *g2;
5246 /* If these givs are identical, they can be combined. */
5247 if (rtx_equal_p (g1->mult_val, g2->mult_val)
5248 && rtx_equal_p (g1->add_val, g2->add_val))
5250 g2->new_reg = g1->dest_reg;
5255 /* If G2 can be expressed as a function of G1 and that function is valid
5256 as an address and no more expensive than using a register for G2,
5257 the expression of G2 in terms of G1 can be used. */
5258 if (g2->giv_type == DEST_ADDR
5259 && (tem = express_from (g1, g2)) != 0
5260 && memory_address_p (g2->mem_mode, tem)
5261 && ADDRESS_COST (tem) <= ADDRESS_COST (*g2->location))
5271 /* Check all pairs of givs for iv_class BL and see if any can be combined with
5272 any other. If so, point SAME to the giv combined with and set NEW_REG to
5273 be an expression (in terms of the other giv's DEST_REG) equivalent to the
5274 giv. Also, update BENEFIT and related fields for cost/benefit analysis. */
5278 struct iv_class *bl;
5280 struct induction *g1, *g2;
5283 for (g1 = bl->giv; g1; g1 = g1->next_iv)
5284 for (pass = 0; pass <= 1; pass++)
5285 for (g2 = bl->giv; g2; g2 = g2->next_iv)
5287 /* First try to combine with replaceable givs, then all givs. */
5288 && (g1->replaceable || pass == 1)
5289 /* If either has already been combined or is to be ignored, can't
5291 && ! g1->ignore && ! g2->ignore && ! g1->same && ! g2->same
5292 /* If something has been based on G2, G2 cannot itself be based
5293 on something else. */
5294 && ! g2->combined_with
5295 && combine_givs_p (g1, g2))
5297 /* g2->new_reg set by `combine_givs_p' */
5299 g1->combined_with = 1;
5300 g1->benefit += g2->benefit;
5301 /* ??? The new final_[bg]iv_value code does a much better job
5302 of finding replaceable giv's, and hence this code may no
5303 longer be necessary. */
5304 if (! g2->replaceable && REG_USERVAR_P (g2->dest_reg))
5305 g1->benefit -= copy_cost;
5306 g1->lifetime += g2->lifetime;
5307 g1->times_used += g2->times_used;
5309 if (loop_dump_stream)
5310 fprintf (loop_dump_stream, "giv at %d combined with giv at %d\n",
5311 INSN_UID (g2->insn), INSN_UID (g1->insn));
5315 /* EMIT code before INSERT_BEFORE to set REG = B * M + A. */
5318 emit_iv_add_mult (b, m, a, reg, insert_before)
5319 rtx b; /* initial value of basic induction variable */
5320 rtx m; /* multiplicative constant */
5321 rtx a; /* additive constant */
5322 rtx reg; /* destination register */
5328 /* Prevent unexpected sharing of these rtx. */
5332 /* Increase the lifetime of any invariants moved further in code. */
5333 update_reg_last_use (a, insert_before);
5334 update_reg_last_use (b, insert_before);
5335 update_reg_last_use (m, insert_before);
5338 result = expand_mult_add (b, reg, m, a, GET_MODE (reg), 0);
5340 emit_move_insn (reg, result);
5341 seq = gen_sequence ();
5344 emit_insn_before (seq, insert_before);
5347 /* Test whether A * B can be computed without
5348 an actual multiply insn. Value is 1 if so. */
5351 product_cheap_p (a, b)
5357 struct obstack *old_rtl_obstack = rtl_obstack;
5358 char *storage = (char *) obstack_alloc (&temp_obstack, 0);
5361 /* If only one is constant, make it B. */
5362 if (GET_CODE (a) == CONST_INT)
5363 tmp = a, a = b, b = tmp;
5365 /* If first constant, both constant, so don't need multiply. */
5366 if (GET_CODE (a) == CONST_INT)
5369 /* If second not constant, neither is constant, so would need multiply. */
5370 if (GET_CODE (b) != CONST_INT)
5373 /* One operand is constant, so might not need multiply insn. Generate the
5374 code for the multiply and see if a call or multiply, or long sequence
5375 of insns is generated. */
5377 rtl_obstack = &temp_obstack;
5379 expand_mult (GET_MODE (a), a, b, NULL_RTX, 0);
5380 tmp = gen_sequence ();
5383 if (GET_CODE (tmp) == SEQUENCE)
5385 if (XVEC (tmp, 0) == 0)
5387 else if (XVECLEN (tmp, 0) > 3)
5390 for (i = 0; i < XVECLEN (tmp, 0); i++)
5392 rtx insn = XVECEXP (tmp, 0, i);
5394 if (GET_CODE (insn) != INSN
5395 || (GET_CODE (PATTERN (insn)) == SET
5396 && GET_CODE (SET_SRC (PATTERN (insn))) == MULT)
5397 || (GET_CODE (PATTERN (insn)) == PARALLEL
5398 && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET
5399 && GET_CODE (SET_SRC (XVECEXP (PATTERN (insn), 0, 0))) == MULT))
5406 else if (GET_CODE (tmp) == SET
5407 && GET_CODE (SET_SRC (tmp)) == MULT)
5409 else if (GET_CODE (tmp) == PARALLEL
5410 && GET_CODE (XVECEXP (tmp, 0, 0)) == SET
5411 && GET_CODE (SET_SRC (XVECEXP (tmp, 0, 0))) == MULT)
5414 /* Free any storage we obtained in generating this multiply and restore rtl
5415 allocation to its normal obstack. */
5416 obstack_free (&temp_obstack, storage);
5417 rtl_obstack = old_rtl_obstack;
5422 /* Check to see if loop can be terminated by a "decrement and branch until
5423 zero" instruction. If so, add a REG_NONNEG note to the branch insn if so.
5424 Also try reversing an increment loop to a decrement loop
5425 to see if the optimization can be performed.
5426 Value is nonzero if optimization was performed. */
5428 /* This is useful even if the architecture doesn't have such an insn,
5429 because it might change a loops which increments from 0 to n to a loop
5430 which decrements from n to 0. A loop that decrements to zero is usually
5431 faster than one that increments from zero. */
5433 /* ??? This could be rewritten to use some of the loop unrolling procedures,
5434 such as approx_final_value, biv_total_increment, loop_iterations, and
5435 final_[bg]iv_value. */
5438 check_dbra_loop (loop_end, insn_count, loop_start)
5443 struct iv_class *bl;
5448 enum rtx_code branch_code;
5451 rtx before_comparison;
5454 /* If last insn is a conditional branch, and the insn before tests a
5455 register value, try to optimize it. Otherwise, we can't do anything. */
5457 comparison = get_condition_for_loop (PREV_INSN (loop_end));
5458 if (comparison == 0)
5461 /* Check all of the bivs to see if the compare uses one of them.
5462 Skip biv's set more than once because we can't guarantee that
5463 it will be zero on the last iteration. Also skip if the biv is
5464 used between its update and the test insn. */
5466 for (bl = loop_iv_list; bl; bl = bl->next)
5468 if (bl->biv_count == 1
5469 && bl->biv->dest_reg == XEXP (comparison, 0)
5470 && ! reg_used_between_p (regno_reg_rtx[bl->regno], bl->biv->insn,
5471 PREV_INSN (PREV_INSN (loop_end))))
5478 /* Look for the case where the basic induction variable is always
5479 nonnegative, and equals zero on the last iteration.
5480 In this case, add a reg_note REG_NONNEG, which allows the
5481 m68k DBRA instruction to be used. */
5483 if (((GET_CODE (comparison) == GT
5484 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
5485 && INTVAL (XEXP (comparison, 1)) == -1)
5486 || (GET_CODE (comparison) == NE && XEXP (comparison, 1) == const0_rtx))
5487 && GET_CODE (bl->biv->add_val) == CONST_INT
5488 && INTVAL (bl->biv->add_val) < 0)
5490 /* Initial value must be greater than 0,
5491 init_val % -dec_value == 0 to ensure that it equals zero on
5492 the last iteration */
5494 if (GET_CODE (bl->initial_value) == CONST_INT
5495 && INTVAL (bl->initial_value) > 0
5496 && (INTVAL (bl->initial_value) %
5497 (-INTVAL (bl->biv->add_val))) == 0)
5499 /* register always nonnegative, add REG_NOTE to branch */
5500 REG_NOTES (PREV_INSN (loop_end))
5501 = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
5502 REG_NOTES (PREV_INSN (loop_end)));
5508 /* If the decrement is 1 and the value was tested as >= 0 before
5509 the loop, then we can safely optimize. */
5510 for (p = loop_start; p; p = PREV_INSN (p))
5512 if (GET_CODE (p) == CODE_LABEL)
5514 if (GET_CODE (p) != JUMP_INSN)
5517 before_comparison = get_condition_for_loop (p);
5518 if (before_comparison
5519 && XEXP (before_comparison, 0) == bl->biv->dest_reg
5520 && GET_CODE (before_comparison) == LT
5521 && XEXP (before_comparison, 1) == const0_rtx
5522 && ! reg_set_between_p (bl->biv->dest_reg, p, loop_start)
5523 && INTVAL (bl->biv->add_val) == -1)
5525 REG_NOTES (PREV_INSN (loop_end))
5526 = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
5527 REG_NOTES (PREV_INSN (loop_end)));
5534 else if (num_mem_sets <= 1)
5536 /* Try to change inc to dec, so can apply above optimization. */
5538 all registers modified are induction variables or invariant,
5539 all memory references have non-overlapping addresses
5540 (obviously true if only one write)
5541 allow 2 insns for the compare/jump at the end of the loop. */
5542 int num_nonfixed_reads = 0;
5543 /* 1 if the iteration var is used only to count iterations. */
5544 int no_use_except_counting = 0;
5546 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5547 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5548 num_nonfixed_reads += count_nonfixed_reads (PATTERN (p));
5550 if (bl->giv_count == 0
5551 && ! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
5553 rtx bivreg = regno_reg_rtx[bl->regno];
5555 /* If there are no givs for this biv, and the only exit is the
5556 fall through at the end of the the loop, then
5557 see if perhaps there are no uses except to count. */
5558 no_use_except_counting = 1;
5559 for (p = loop_start; p != loop_end; p = NEXT_INSN (p))
5560 if (GET_RTX_CLASS (GET_CODE (p)) == 'i')
5562 rtx set = single_set (p);
5564 if (set && GET_CODE (SET_DEST (set)) == REG
5565 && REGNO (SET_DEST (set)) == bl->regno)
5566 /* An insn that sets the biv is okay. */
5568 else if (p == prev_nonnote_insn (prev_nonnote_insn (loop_end))
5569 || p == prev_nonnote_insn (loop_end))
5570 /* Don't bother about the end test. */
5572 else if (reg_mentioned_p (bivreg, PATTERN (p)))
5573 /* Any other use of the biv is no good. */
5575 no_use_except_counting = 0;
5581 /* This code only acts for innermost loops. Also it simplifies
5582 the memory address check by only reversing loops with
5583 zero or one memory access.
5584 Two memory accesses could involve parts of the same array,
5585 and that can't be reversed. */
5587 if (num_nonfixed_reads <= 1
5589 && !loop_has_volatile
5590 && (no_use_except_counting
5591 || (bl->giv_count + bl->biv_count + num_mem_sets
5592 + num_movables + 2 == insn_count)))
5594 rtx condition = get_condition_for_loop (PREV_INSN (loop_end));
5598 /* Loop can be reversed. */
5599 if (loop_dump_stream)
5600 fprintf (loop_dump_stream, "Can reverse loop\n");
5602 /* Now check other conditions:
5603 initial_value must be zero,
5604 final_value % add_val == 0, so that when reversed, the
5605 biv will be zero on the last iteration.
5607 This test can probably be improved since +/- 1 in the constant
5608 can be obtained by changing LT to LE and vice versa; this is
5611 if (comparison && bl->initial_value == const0_rtx
5612 && GET_CODE (XEXP (comparison, 1)) == CONST_INT
5613 /* LE gets turned into LT */
5614 && GET_CODE (comparison) == LT
5615 && (INTVAL (XEXP (comparison, 1))
5616 % INTVAL (bl->biv->add_val)) == 0)
5618 /* Register will always be nonnegative, with value
5619 0 on last iteration if loop reversed */
5621 /* Save some info needed to produce the new insns. */
5622 reg = bl->biv->dest_reg;
5623 jump_label = XEXP (SET_SRC (PATTERN (PREV_INSN (loop_end))), 1);
5624 new_add_val = GEN_INT (- INTVAL (bl->biv->add_val));
5626 final_value = XEXP (comparison, 1);
5627 start_value = GEN_INT (INTVAL (XEXP (comparison, 1))
5628 - INTVAL (bl->biv->add_val));
5630 /* Initialize biv to start_value before loop start.
5631 The old initializing insn will be deleted as a
5632 dead store by flow.c. */
5633 emit_insn_before (gen_move_insn (reg, start_value), loop_start);
5635 /* Add insn to decrement register, and delete insn
5636 that incremented the register. */
5637 p = emit_insn_before (gen_add2_insn (reg, new_add_val),
5639 delete_insn (bl->biv->insn);
5641 /* Update biv info to reflect its new status. */
5643 bl->initial_value = start_value;
5644 bl->biv->add_val = new_add_val;
5646 /* Inc LABEL_NUSES so that delete_insn will
5647 not delete the label. */
5648 LABEL_NUSES (XEXP (jump_label, 0)) ++;
5650 /* Emit an insn after the end of the loop to set the biv's
5651 proper exit value if it is used anywhere outside the loop. */
5652 if ((regno_last_uid[bl->regno]
5653 != INSN_UID (PREV_INSN (PREV_INSN (loop_end))))
5655 || regno_first_uid[bl->regno] != INSN_UID (bl->init_insn))
5656 emit_insn_after (gen_move_insn (reg, final_value),
5659 /* Delete compare/branch at end of loop. */
5660 delete_insn (PREV_INSN (loop_end));
5661 delete_insn (PREV_INSN (loop_end));
5663 /* Add new compare/branch insn at end of loop. */
5665 emit_cmp_insn (reg, const0_rtx, GE, NULL_RTX,
5666 GET_MODE (reg), 0, 0);
5667 emit_jump_insn (gen_bge (XEXP (jump_label, 0)));
5668 tem = gen_sequence ();
5670 emit_jump_insn_before (tem, loop_end);
5672 for (tem = PREV_INSN (loop_end);
5673 tem && GET_CODE (tem) != JUMP_INSN; tem = PREV_INSN (tem))
5677 JUMP_LABEL (tem) = XEXP (jump_label, 0);
5679 /* Increment of LABEL_NUSES done above. */
5680 /* Register is now always nonnegative,
5681 so add REG_NONNEG note to the branch. */
5682 REG_NOTES (tem) = gen_rtx (EXPR_LIST, REG_NONNEG, NULL_RTX,
5688 /* Mark that this biv has been reversed. Each giv which depends
5689 on this biv, and which is also live past the end of the loop
5690 will have to be fixed up. */
5694 if (loop_dump_stream)
5695 fprintf (loop_dump_stream,
5696 "Reversed loop and added reg_nonneg\n");
5706 /* Verify whether the biv BL appears to be eliminable,
5707 based on the insns in the loop that refer to it.
5708 LOOP_START is the first insn of the loop, and END is the end insn.
5710 If ELIMINATE_P is non-zero, actually do the elimination.
5712 THRESHOLD and INSN_COUNT are from loop_optimize and are used to
5713 determine whether invariant insns should be placed inside or at the
5714 start of the loop. */
5717 maybe_eliminate_biv (bl, loop_start, end, eliminate_p, threshold, insn_count)
5718 struct iv_class *bl;
5722 int threshold, insn_count;
5724 rtx reg = bl->biv->dest_reg;
5726 struct induction *v;
5728 /* Scan all insns in the loop, stopping if we find one that uses the
5729 biv in a way that we cannot eliminate. */
5731 for (p = loop_start; p != end; p = NEXT_INSN (p))
5733 enum rtx_code code = GET_CODE (p);
5734 rtx where = threshold >= insn_count ? loop_start : p;
5736 if ((code == INSN || code == JUMP_INSN || code == CALL_INSN)
5737 && reg_mentioned_p (reg, PATTERN (p))
5738 && ! maybe_eliminate_biv_1 (PATTERN (p), p, bl, eliminate_p, where))
5740 if (loop_dump_stream)
5741 fprintf (loop_dump_stream,
5742 "Cannot eliminate biv %d: biv used in insn %d.\n",
5743 bl->regno, INSN_UID (p));
5750 if (loop_dump_stream)
5751 fprintf (loop_dump_stream, "biv %d %s eliminated.\n",
5752 bl->regno, eliminate_p ? "was" : "can be");
5759 /* If BL appears in X (part of the pattern of INSN), see if we can
5760 eliminate its use. If so, return 1. If not, return 0.
5762 If BIV does not appear in X, return 1.
5764 If ELIMINATE_P is non-zero, actually do the elimination. WHERE indicates
5765 where extra insns should be added. Depending on how many items have been
5766 moved out of the loop, it will either be before INSN or at the start of
5770 maybe_eliminate_biv_1 (x, insn, bl, eliminate_p, where)
5772 struct iv_class *bl;
5776 enum rtx_code code = GET_CODE (x);
5777 rtx reg = bl->biv->dest_reg;
5778 enum machine_mode mode = GET_MODE (reg);
5779 struct induction *v;
5788 /* If we haven't already been able to do something with this BIV,
5789 we can't eliminate it. */
5795 /* If this sets the BIV, it is not a problem. */
5796 if (SET_DEST (x) == reg)
5799 /* If this is an insn that defines a giv, it is also ok because
5800 it will go away when the giv is reduced. */
5801 for (v = bl->giv; v; v = v->next_iv)
5802 if (v->giv_type == DEST_REG && SET_DEST (x) == v->dest_reg)
5806 if (SET_DEST (x) == cc0_rtx && SET_SRC (x) == reg)
5808 /* Can replace with any giv that was reduced and
5809 that has (MULT_VAL != 0) and (ADD_VAL == 0).
5810 Require a constant for MULT_VAL, so we know it's nonzero. */
5812 for (v = bl->giv; v; v = v->next_iv)
5813 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
5814 && v->add_val == const0_rtx
5815 && ! v->ignore && ! v->maybe_dead
5821 /* If the giv has the opposite direction of change,
5822 then reverse the comparison. */
5823 if (INTVAL (v->mult_val) < 0)
5824 new = gen_rtx (COMPARE, GET_MODE (v->new_reg),
5825 const0_rtx, v->new_reg);
5829 /* We can probably test that giv's reduced reg. */
5830 if (validate_change (insn, &SET_SRC (x), new, 0))
5834 /* Look for a giv with (MULT_VAL != 0) and (ADD_VAL != 0);
5835 replace test insn with a compare insn (cmp REDUCED_GIV ADD_VAL).
5836 Require a constant for MULT_VAL, so we know it's nonzero. */
5838 for (v = bl->giv; v; v = v->next_iv)
5839 if (CONSTANT_P (v->mult_val) && v->mult_val != const0_rtx
5840 && ! v->ignore && ! v->maybe_dead
5846 /* If the giv has the opposite direction of change,
5847 then reverse the comparison. */
5848 if (INTVAL (v->mult_val) < 0)
5849 new = gen_rtx (COMPARE, VOIDmode, copy_rtx (v->add_val),
5852 new = gen_rtx (COMPARE, VOIDmode, v->new_reg,
5853 copy_rtx (v->add_val));
5855 /* Replace biv with the giv's reduced register. */
5856 update_reg_last_use (v->add_val, insn);
5857 if (validate_change (insn, &SET_SRC (PATTERN (insn)), new, 0))
5860 /* Insn doesn't support that constant or invariant. Copy it
5861 into a register (it will be a loop invariant.) */
5862 tem = gen_reg_rtx (GET_MODE (v->new_reg));
5864 emit_insn_before (gen_move_insn (tem, copy_rtx (v->add_val)),
5867 if (validate_change (insn, &SET_SRC (PATTERN (insn)),
5868 gen_rtx (COMPARE, VOIDmode,
5869 v->new_reg, tem), 0))
5878 case GT: case GE: case GTU: case GEU:
5879 case LT: case LE: case LTU: case LEU:
5880 /* See if either argument is the biv. */
5881 if (XEXP (x, 0) == reg)
5882 arg = XEXP (x, 1), arg_operand = 1;
5883 else if (XEXP (x, 1) == reg)
5884 arg = XEXP (x, 0), arg_operand = 0;
5888 if (CONSTANT_P (arg))
5890 /* First try to replace with any giv that has constant positive
5891 mult_val and constant add_val. We might be able to support
5892 negative mult_val, but it seems complex to do it in general. */
5894 for (v = bl->giv; v; v = v->next_iv)
5895 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
5896 && CONSTANT_P (v->add_val)
5897 && ! v->ignore && ! v->maybe_dead
5903 /* Replace biv with the giv's reduced reg. */
5904 XEXP (x, 1-arg_operand) = v->new_reg;
5906 /* If all constants are actually constant integers and
5907 the derived constant can be directly placed in the COMPARE,
5909 if (GET_CODE (arg) == CONST_INT
5910 && GET_CODE (v->mult_val) == CONST_INT
5911 && GET_CODE (v->add_val) == CONST_INT
5912 && validate_change (insn, &XEXP (x, arg_operand),
5913 GEN_INT (INTVAL (arg)
5914 * INTVAL (v->mult_val)
5915 + INTVAL (v->add_val)), 0))
5918 /* Otherwise, load it into a register. */
5919 tem = gen_reg_rtx (mode);
5920 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
5921 if (validate_change (insn, &XEXP (x, arg_operand), tem, 0))
5924 /* If that failed, put back the change we made above. */
5925 XEXP (x, 1-arg_operand) = reg;
5928 /* Look for giv with positive constant mult_val and nonconst add_val.
5929 Insert insns to calculate new compare value. */
5931 for (v = bl->giv; v; v = v->next_iv)
5932 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
5933 && ! v->ignore && ! v->maybe_dead
5941 tem = gen_reg_rtx (mode);
5943 /* Replace biv with giv's reduced register. */
5944 validate_change (insn, &XEXP (x, 1 - arg_operand),
5947 /* Compute value to compare against. */
5948 emit_iv_add_mult (arg, v->mult_val, v->add_val, tem, where);
5949 /* Use it in this insn. */
5950 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
5951 if (apply_change_group ())
5955 else if (GET_CODE (arg) == REG || GET_CODE (arg) == MEM)
5957 if (invariant_p (arg) == 1)
5959 /* Look for giv with constant positive mult_val and nonconst
5960 add_val. Insert insns to compute new compare value. */
5962 for (v = bl->giv; v; v = v->next_iv)
5963 if (CONSTANT_P (v->mult_val) && INTVAL (v->mult_val) > 0
5964 && ! v->ignore && ! v->maybe_dead
5972 tem = gen_reg_rtx (mode);
5974 /* Replace biv with giv's reduced register. */
5975 validate_change (insn, &XEXP (x, 1 - arg_operand),
5978 /* Compute value to compare against. */
5979 emit_iv_add_mult (arg, v->mult_val, v->add_val,
5981 validate_change (insn, &XEXP (x, arg_operand), tem, 1);
5982 if (apply_change_group ())
5987 /* This code has problems. Basically, you can't know when
5988 seeing if we will eliminate BL, whether a particular giv
5989 of ARG will be reduced. If it isn't going to be reduced,
5990 we can't eliminate BL. We can try forcing it to be reduced,
5991 but that can generate poor code.
5993 The problem is that the benefit of reducing TV, below should
5994 be increased if BL can actually be eliminated, but this means
5995 we might have to do a topological sort of the order in which
5996 we try to process biv. It doesn't seem worthwhile to do
5997 this sort of thing now. */
6000 /* Otherwise the reg compared with had better be a biv. */
6001 if (GET_CODE (arg) != REG
6002 || reg_iv_type[REGNO (arg)] != BASIC_INDUCT)
6005 /* Look for a pair of givs, one for each biv,
6006 with identical coefficients. */
6007 for (v = bl->giv; v; v = v->next_iv)
6009 struct induction *tv;
6011 if (v->ignore || v->maybe_dead || v->mode != mode)
6014 for (tv = reg_biv_class[REGNO (arg)]->giv; tv; tv = tv->next_iv)
6015 if (! tv->ignore && ! tv->maybe_dead
6016 && rtx_equal_p (tv->mult_val, v->mult_val)
6017 && rtx_equal_p (tv->add_val, v->add_val)
6018 && tv->mode == mode)
6023 /* Replace biv with its giv's reduced reg. */
6024 XEXP (x, 1-arg_operand) = v->new_reg;
6025 /* Replace other operand with the other giv's
6027 XEXP (x, arg_operand) = tv->new_reg;
6034 /* If we get here, the biv can't be eliminated. */
6038 /* If this address is a DEST_ADDR giv, it doesn't matter if the
6039 biv is used in it, since it will be replaced. */
6040 for (v = bl->giv; v; v = v->next_iv)
6041 if (v->giv_type == DEST_ADDR && v->location == &XEXP (x, 0))
6046 /* See if any subexpression fails elimination. */
6047 fmt = GET_RTX_FORMAT (code);
6048 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
6053 if (! maybe_eliminate_biv_1 (XEXP (x, i), insn, bl,
6054 eliminate_p, where))
6059 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6060 if (! maybe_eliminate_biv_1 (XVECEXP (x, i, j), insn, bl,
6061 eliminate_p, where))
6070 /* Return nonzero if the last use of REG
6071 is in an insn following INSN in the same basic block. */
6074 last_use_this_basic_block (reg, insn)
6080 n && GET_CODE (n) != CODE_LABEL && GET_CODE (n) != JUMP_INSN;
6083 if (regno_last_uid[REGNO (reg)] == INSN_UID (n))
6089 /* Called via `note_stores' to record the initial value of a biv. Here we
6090 just record the location of the set and process it later. */
6093 record_initial (dest, set)
6097 struct iv_class *bl;
6099 if (GET_CODE (dest) != REG
6100 || REGNO (dest) >= max_reg_before_loop
6101 || reg_iv_type[REGNO (dest)] != BASIC_INDUCT)
6104 bl = reg_biv_class[REGNO (dest)];
6106 /* If this is the first set found, record it. */
6107 if (bl->init_insn == 0)
6109 bl->init_insn = note_insn;
6114 /* If any of the registers in X are "old" and currently have a last use earlier
6115 than INSN, update them to have a last use of INSN. Their actual last use
6116 will be the previous insn but it will not have a valid uid_luid so we can't
6120 update_reg_last_use (x, insn)
6124 /* Check for the case where INSN does not have a valid luid. In this case,
6125 there is no need to modify the regno_last_uid, as this can only happen
6126 when code is inserted after the loop_end to set a pseudo's final value,
6127 and hence this insn will never be the last use of x. */
6128 if (GET_CODE (x) == REG && REGNO (x) < max_reg_before_loop
6129 && INSN_UID (insn) < max_uid_for_loop
6130 && uid_luid[regno_last_uid[REGNO (x)]] < uid_luid[INSN_UID (insn)])
6131 regno_last_uid[REGNO (x)] = INSN_UID (insn);
6135 register char *fmt = GET_RTX_FORMAT (GET_CODE (x));
6136 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
6139 update_reg_last_use (XEXP (x, i), insn);
6140 else if (fmt[i] == 'E')
6141 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6142 update_reg_last_use (XVECEXP (x, i, j), insn);
6147 /* Given a jump insn JUMP, return the condition that will cause it to branch
6148 to its JUMP_LABEL. If the condition cannot be understood, or is an
6149 inequality floating-point comparison which needs to be reversed, 0 will
6152 If EARLIEST is non-zero, it is a pointer to a place where the earliest
6153 insn used in locating the condition was found. If a replacement test
6154 of the condition is desired, it should be placed in front of that
6155 insn and we will be sure that the inputs are still valid.
6157 The condition will be returned in a canonical form to simplify testing by
6158 callers. Specifically:
6160 (1) The code will always be a comparison operation (EQ, NE, GT, etc.).
6161 (2) Both operands will be machine operands; (cc0) will have been replaced.
6162 (3) If an operand is a constant, it will be the second operand.
6163 (4) (LE x const) will be replaced with (LT x <const+1>) and similarly
6164 for GE, GEU, and LEU. */
6167 get_condition (jump, earliest)
6176 int reverse_code = 0;
6177 int did_reverse_condition = 0;
6179 /* If this is not a standard conditional jump, we can't parse it. */
6180 if (GET_CODE (jump) != JUMP_INSN
6181 || ! condjump_p (jump) || simplejump_p (jump))
6184 code = GET_CODE (XEXP (SET_SRC (PATTERN (jump)), 0));
6185 op0 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 0);
6186 op1 = XEXP (XEXP (SET_SRC (PATTERN (jump)), 0), 1);
6191 /* If this branches to JUMP_LABEL when the condition is false, reverse
6193 if (XEXP (XEXP (SET_SRC (PATTERN (jump)), 2), 0) == JUMP_LABEL (jump))
6194 code = reverse_condition (code), did_reverse_condition ^= 1;
6196 /* If we are comparing a register with zero, see if the register is set
6197 in the previous insn to a COMPARE or a comparison operation. Perform
6198 the same tests as a function of STORE_FLAG_VALUE as find_comparison_args
6201 while (GET_RTX_CLASS (code) == '<' && op1 == const0_rtx)
6203 /* Set non-zero when we find something of interest. */
6207 /* If comparison with cc0, import actual comparison from compare
6211 if ((prev = prev_nonnote_insn (prev)) == 0
6212 || GET_CODE (prev) != INSN
6213 || (set = single_set (prev)) == 0
6214 || SET_DEST (set) != cc0_rtx)
6217 op0 = SET_SRC (set);
6218 op1 = CONST0_RTX (GET_MODE (op0));
6224 /* If this is a COMPARE, pick up the two things being compared. */
6225 if (GET_CODE (op0) == COMPARE)
6227 op1 = XEXP (op0, 1);
6228 op0 = XEXP (op0, 0);
6231 else if (GET_CODE (op0) != REG)
6234 /* Go back to the previous insn. Stop if it is not an INSN. We also
6235 stop if it isn't a single set or if it has a REG_INC note because
6236 we don't want to bother dealing with it. */
6238 if ((prev = prev_nonnote_insn (prev)) == 0
6239 || GET_CODE (prev) != INSN
6240 || FIND_REG_INC_NOTE (prev, 0)
6241 || (set = single_set (prev)) == 0)
6244 /* If this is setting OP0, get what it sets it to if it looks
6246 if (SET_DEST (set) == op0)
6248 enum machine_mode inner_mode = GET_MODE (SET_SRC (set));
6250 if ((GET_CODE (SET_SRC (set)) == COMPARE
6253 && GET_MODE_CLASS (inner_mode) == MODE_INT
6254 && (GET_MODE_BITSIZE (inner_mode)
6255 <= HOST_BITS_PER_WIDE_INT)
6256 && (STORE_FLAG_VALUE
6257 & ((HOST_WIDE_INT) 1
6258 << (GET_MODE_BITSIZE (inner_mode) - 1))))
6259 #ifdef FLOAT_STORE_FLAG_VALUE
6261 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
6262 && FLOAT_STORE_FLAG_VALUE < 0)
6265 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')))
6267 else if (((code == EQ
6269 && (GET_MODE_BITSIZE (inner_mode)
6270 <= HOST_BITS_PER_WIDE_INT)
6271 && GET_MODE_CLASS (inner_mode) == MODE_INT
6272 && (STORE_FLAG_VALUE
6273 & ((HOST_WIDE_INT) 1
6274 << (GET_MODE_BITSIZE (inner_mode) - 1))))
6275 #ifdef FLOAT_STORE_FLAG_VALUE
6277 && GET_MODE_CLASS (inner_mode) == MODE_FLOAT
6278 && FLOAT_STORE_FLAG_VALUE < 0)
6281 && GET_RTX_CLASS (GET_CODE (SET_SRC (set))) == '<')
6283 /* We might have reversed a LT to get a GE here. But this wasn't
6284 actually the comparison of data, so we don't flag that we
6285 have had to reverse the condition. */
6286 did_reverse_condition ^= 1;
6292 else if (reg_set_p (op0, prev))
6293 /* If this sets OP0, but not directly, we have to give up. */
6298 if (GET_RTX_CLASS (GET_CODE (x)) == '<')
6299 code = GET_CODE (x);
6302 code = reverse_condition (code);
6303 did_reverse_condition ^= 1;
6307 op0 = XEXP (x, 0), op1 = XEXP (x, 1);
6313 /* If constant is first, put it last. */
6314 if (CONSTANT_P (op0))
6315 code = swap_condition (code), tem = op0, op0 = op1, op1 = tem;
6317 /* If OP0 is the result of a comparison, we weren't able to find what
6318 was really being compared, so fail. */
6319 if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC)
6322 /* Canonicalize any ordered comparison with integers involving equality. */
6323 if (GET_CODE (op1) == CONST_INT)
6325 HOST_WIDE_INT const_val = INTVAL (op1);
6326 unsigned HOST_WIDE_INT uconst_val = const_val;
6332 op1 = GEN_INT (const_val + 1);
6337 op1 = GEN_INT (const_val - 1);
6342 op1 = GEN_INT (uconst_val + 1);
6347 op1 = GEN_INT (uconst_val - 1);
6352 /* If this was floating-point and we reversed anything other than an
6353 EQ or NE, return zero. */
6354 if (TARGET_FLOAT_FORMAT == IEEE_FLOAT_FORMAT
6355 && did_reverse_condition && code != NE && code != EQ
6356 && GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT)
6360 /* Never return CC0; return zero instead. */
6365 return gen_rtx (code, VOIDmode, op0, op1);
6368 /* Similar to above routine, except that we also put an invariant last
6369 unless both operands are invariants. */
6372 get_condition_for_loop (x)
6375 rtx comparison = get_condition (x, NULL_PTR);
6378 || ! invariant_p (XEXP (comparison, 0))
6379 || invariant_p (XEXP (comparison, 1)))
6382 return gen_rtx (swap_condition (GET_CODE (comparison)), VOIDmode,
6383 XEXP (comparison, 1), XEXP (comparison, 0));